text
stringlengths 1
2.56M
| id
stringlengths 40
40
| metadata
dict |
|---|---|---|
\section{Introduction}\label{Introduction}
Equiangular tight frames (ETF) define regular geometrical structures in Hilbert spaces. An ETF composed of $N$ normalized vectors $\{\varphi_1,\dots,\varphi_N\}$ minimizes the highest possible coherence $\max_{j\neq k}|\langle\varphi_j|\varphi_k\rangle|$. These frames, which are also known as \emph{Grassmannian frames} \cite{SH03}, are thus typically related with optimal solutions of practical problems in communications, coding theory and sparse approximation \cite{JMF14,SH03,T04}. In particular, ETF provide an error-correcting code which is maximally robust against two erasures \cite{HP04}. ETF are also closely related to strongly regular graphs \cite{W09}, difference sets \cite{DF07,XZG74} and Steiner systems \cite{FMT12}. A special class of ETF consisting of $d^2$ vectors in $\mathbb{C}^d$ is known in quantum mechanics as \emph{Symmetric Informationally Complete Positive Operator Valued Measure} (SIC-POVM) \cite{RBSC04}. From considering rank-one SIC-POVM measurements it is possible to reconstruct any quantum state. Moreover, SIC-POVM are optimal measurements in the sense that the redundancy of information provided by any pair of outcomes of the POVM is minimized.
Along the years, many important results have been derived for real and complex ETF. A summary of the current state of the art can be found in a recent work of M. Fickus and D. Mixon \cite{FM15}.
However, the theory is still far from being complete. To illustrate this fact, let us mention that the full classification of complex ETF is only known in dimensions $d=2$ and $d=3$ \cite{S14}. The aim of the present work is to shed new light on ETF by establishing a connection with unistochastic matrices theory.
The paper is organised as follows: In Section \ref{S2} we review basic properties of equiangular tight frames. Section \ref{S3} is devoted to establishing the connection between ETF and unistochastic matrices. We study some fundamental properties and present new inequivalent classes of complex ETF in infinitely many dimensions. In Section \ref{S4} we present a method to introduce free parameters in a given real or complex ETF. In Section \ref{S5} we define an iterative procedure to find the unitary matrix underlying a unistochastic matrix. In Section \ref{S5} we summarize the results and pose some open questions. In Appendix \ref{Ap1} we explicitly derive a $6$-parametric family of complex ETF stemming from the real ETF($6,16$). Finally, Appendix \ref{Ap2} contains all the proofs of our propositions.
\section{Equiangular tight frames}\label{S2}
A complex ETF($d,N$) is a set of $N$ vectors $\{\varphi_k\}_{k=1,\dots,N}$ in $\mathbb{C}^d$ such that they are:
\begin{enumerate}
\item Normalized: $\|\varphi_j\|=1$ for every $j=1,\dots,N$.
\item Equiangular: $|\langle\varphi_j,\varphi_l\rangle|=\frac{1}{\alpha}$ for every $j\neq l=1,\dots,N$ and a fixed $\alpha>0$.
\item Tight frame: $\frac{d}{N}\sum_{j=1}^N\langle\varphi_j,\phi\rangle\varphi_j=\phi$ for every $\phi\in\mathbb{C}^d$.
\end{enumerate}
If the set of vectors forming the ETF belongs to $\mathbb{R}^d$, then we have a real ETF. It is easy to prove that the parameter $\alpha>0$ in property 2, which is called \emph{inverse coherence}, satisfies $\alpha=\sqrt{\frac{d(N-1)}{N-d}}$ (Welch bound). In quantum mechanics, the third property (tight frame) is equivalent to the fact that the rank-one projectors $\{\Pi_j=\langle\varphi_j,\cdot\rangle\,\varphi_j\}_{j=1,\dots,N}$ satisfy $\frac{d}{N}\sum_{j=1}^N\Pi_j=\mathbb{I}$, where $\mathbb{I}$ denotes the identity matrix, i.e., the projectors form a \emph{Positive Operator Valued Measure} (POVM) \cite{NC11}.
The ETF problem can be equivalently posed in terms of Gram matrices $G_{jk}=\langle\varphi_j,\varphi_k\rangle$. A Gram matrix $G\in\mathbb{C}^{N\times N}$ is associated with an ETF($d,N$) if the following properties hold:
\begin{subequations}\label{Gram}
\begin{align}
G_{jj}&=1 \quad \text{for every $j=0,\dots N-1$}; \label{i} \\
|G_{jl}|&=\frac{1}{\alpha} \quad \text{for every $j\neq l=0,\dots N-1$ and a fixed $\alpha>0$}; \label{ii} \\
\sigma(G)&\in\{0,N/d\}, \label{iii}
\end{align}
\end{subequations}
where $\sigma(G)$ is the spectrum of $G$. It is easy to prove that the eigenvalues $N/d$ and $0$ of $G$ have multiplicities $d$ and $N-d$, respectively. When property \eqref{iii} is replaced by the weaker property $\mathrm{Rank}(G)=d$, the conditions \eqref{Gram} define a set of $N$ \emph{equiangular lines} in dimension $d$ \cite{LS73}. We recall that the vectors forming an ETF can be explicitly found from the Gram matrix by considering the \emph{Cholesky decomposition}, which can be efficiently implemented (see \cite[page 52]{S13}).
\section{ETF and unistochastic matrices }\label{S3}
A bistochastic matrix $B$ is a square matrix of size $N$ having non-negative real entries such that
\begin{equation}
\sum_{i=0}^{N-1}B_{ik}=1\hspace{0.3cm}\mbox{and}\hspace{0.3cm}\sum_{j=0}^{N-1}B_{kj}=1,
\end{equation}
for every $k=0,\dots N-1$. The matrix $B$ is called unistochastic if there exists a unitary matrix $U$ such that $B_{ij}=|U_{ij}|^2$, for every $i,j=0,\dots N-1$. If $U$ is a real orthogonal matrix then $B$ is called orthostochastic. Full set of unistochastic matrices is simple to derive for $N=2$ and it has been fully characterised also for $N=3$ \cite{BEKTZ}. The problem for $N=4$ is still open. A detailed explanation on unistochastic matrices and their applications can be found e.g. in Ref. \cite{S15}.
From now on we will restrict our attention to a particular type of unistochastic matrices, denoted $\mathcal{B}_N(\theta)$. Namely, $\mathcal{B}_N(\theta)$ are unistochastic matrices such that there exists a unitary hermitian matrix $U_N(\theta)$ having a real and non-negative constant diagonal such that
\begin{equation}\label{bisto}
\mathcal{B}_N(\theta)_{ij}=|U_N(\theta)_{ij}|^2=\left\{\begin{array}{cl}
\cos^2(\theta) & \mbox{ if } i=j;\\
\frac{1}{N-1}\sin^2(\theta) & \mbox{ if } i\neq j,
\end{array}\right.
\end{equation}
where $\theta\in[0,\pi/2]$ for convenient reasons (see caption of Fig.\ref{Fig1}). We are now in position to establish the connection existing between ETF and unistochastic matrices $\mathcal{B}_N(\theta)$.
\begin{prop}\label{Prop1}
A complex ETF($d,N$) exists if and only if a unistochastic matrix $\mathcal{B}_N(\theta)$ exists for $d=N\sin^2(\theta/2)$, $\theta\in[0,\pi/2]$. In particular, the ETF($d,N$) is real if and only if $\mathcal{B}_N(\theta)$ is orthostochastic.
\end{prop}
Proofs of propositions can be found in Appendix \ref{Ap2}.
\begin{figure}[htbp]
\centering
\subfigure [Polar representation for ETF.]{\includegraphics[width=70mm]{Fig1a.jpg}}\hspace{1cm}
\subfigure[ Existing ETF for $N=4,\dots,16$.]{\includegraphics[width=60mm]{Fig1b.jpg}}
\caption{Polar representation for complex ETF induced by the existence of unistochastic matrices $\mathcal{B}_N(\theta)$. Proposition \ref{Prop1} allows us to identify coordinates $(\theta,N)$ unambiguously with ETF($d,N$), where $d=N\sin^2(\theta/2)$; E.g., orthonormal basis ($d=0,\,\theta=0$), regular simplices ($d=1,\,\sin(\theta/2)=1/\sqrt{N}$), and SIC-POVM ($d=\sqrt{N},\,\sin(\theta/2)=1/\sqrt{d}$). Note that $\theta\in[0,\pi/2]$ and $\theta\in[\pi/2,\pi]$ characterize ETF($d,N$) and ETF($N-d,N$) (Naimark complement), respectively. For example, $d=N$ and $d=N+1$ are the Naimark complements for orthonormal bases and regular simplices. Along this work we consider the convention $\theta\in[0,\pi/2]$, anti-clock wise.}
\label{Fig1}
\end{figure}
Unitary matrices having prescribed moduli of the entries have been previously studied \cite{AMM91, D94,ST08,TC15}. However, those approaches are only marginally related to our problem because hermiticity of the underlying unitary matrix and uniformity of the main diagonal are fundamental assumptions to connect ETF with unistochastic matrices.
Hermitian complex Hadamard matrices have a remarkable property, namely $U_{N_1}(\theta_{N_1})\otimes U_{N_2}(\theta_{N_2})=U_{N_1N_2}(\theta_{N_1N_2})$, where $\otimes$ denotes the Kronecker product. This property allows us to derive a wide range of inequivalent equiangular tight frames:
\begin{prop}\label{Prop2}
Let $\sqrt{N}=p_1^{r_1}\times\dots\times p_a^{r_a}$ be the prime power decomposition of $\sqrt{N}$, where $N$ is a square and $p_1,\dots,p_a$ are distinct prime numbers. Then there exist at least
\begin{equation}
\mathcal{N}=\mathcal{P}(r_1)\mathcal{P}(r_2)\dots \mathcal{P}(r_a),
\end{equation}
inequivalent ETF($(N-\sqrt{N})/2,N$), where $\mathcal{P}(r)$ denotes the number of unrestricted partitions of the integer number $r$.
\end{prop}
Our approach based on unistochastic matrices enables us to address another interesting problem in the ETF theory, namely, the possibility to have the off-diagonal entries in the Gram matrix proportional to roots of unity. In Proposition \ref{Prop3} below, we show that for a given value of $N$ only certain $m$-th roots of unity are allowed.
\begin{prop}\label{Prop3}
Consider an ETF with a Gram matrix such that its off-diagonal entries, when normalized to have absolute value $1$, are $m$-th roots of unity. Let $m=p_1^{a_1}\dots p_r^{a_r}$ be the decomposition of $m$ in prime power factors. Then there exist numbers $x_j\in\mathbb{N}\cup\{0\}$ such that $2k+N-2=x_1p_1+\cdots+x_rp_r$, where $k=\cot(\theta)\sqrt{N-1}$.
\end{prop}
Using relation $d=N\sin^2(\theta/2)$, derived in Proposition \ref{Prop1}, one can express $\cot(\theta)$ in terms of $N$ and $d$. Hence we obtain
\begin{equation}\label{k}
k=\frac{N-2d}{2}\sqrt{\frac{N-1}{d(N-d)}}.
\end{equation}
Proposition \ref{Prop3} implies in particular that if $2k$ is not an integer number, then the normalized off-diagonal entries of the Gram matrix cannot be roots of unity. Note that for a real ETF the number $k$ is always integer \cite{STDH07}.
In order to illustrate the result, let us consider $N=64$ and fourth roots of unity. According to Proposition \ref{Prop3}, for $m=2^2$
we have $2k+64-2=2x_1$. The only ETF compatible with this equation and equation \eqref{k} are ETF(32,64), ETF(28,64) and ETF(8,64). Let us note that ETF(8,64) is a special kind of SIC-POVM known as \emph{Hoggar lines} \cite{H98}. In general, the following restriction holds for SIC-POVM.
\begin{corol}\label{Corol1}
A SIC-POVM in dimension $d$ admits a Gram matrix with off-diagonal entries proportional to roots of unity if $d=2$ or $d+1$ is a square.
\end{corol}
The proof is staightforward from the fact that $2k$ has to be integer (see Prop. \ref{Prop3}) combined with equation \eqref{k}. Concerning the existence of SIC-POVM having Gram matrix composed by roots of unity up to dimension $d=100$, we have:
\begin{itemize}
\item $d=2,\hspace{0.1cm}3,\hspace{0.1cm} 8$: the existence of SIC-POVM confirmed with 4th, 6th and 4th roots of unity in the off-diagonal entries of the Gram matrix, respectively;
\item $d=15,\hspace{0.1cm} 24,\hspace{0.1cm} 35,\hspace{0.1cm} 48$: SIC-POVM exist, but the possibility of having roots of unity in the off-diagonal entries is open;
\item $d=63,\hspace{0.1cm} 80,\hspace{0.1cm} 99$: the existence of SIC-POVM is open.
\end{itemize}
\section{Free parameters for ETF}\label{S4}
In this section, we present a simple method to construct a parametric family of complex ETF stemming from a given ETF in dimension $d=(N-\sqrt{N})/2$. The construction is based on the following idea, introduced in Ref. \cite{G13}: Two vectors $v,w\in\mathbb{C}^N$ are called \emph{equivalent to real (ER) pair} if $v\circ w^*\in\mathbb{R}^N$, where the circle and asterisk denote the Hadamard (entrywise) product and complex conjugation, respectively. Let $C_A$ and $C_B$ be two columns of a complex Hadamard matrix that form an ER pair. Note that the entries of the vector $C_A\circ C_B^*$ are $\pm1$. We introduce a parameter $\alpha\in[0,2\pi)$ in vectors $C_A(\alpha)$ and $C_B(\alpha)$ as follows:
\begin{itemize}
\item If $(C_A\circ C_B^*)_j=-1$, we set $(C_A(\alpha))_j=e^{i\alpha}(C_A)_j$ and $(C_B(\alpha))_j=e^{i\alpha}(C_B)_j$;
\item if $(C_A\circ C_B^*)_j=1$, we set $(C_A(\alpha))_j=(C_A)_j$ and $(C_B(\alpha))_j=(C_B)_j$.
\end{itemize}
Columns $C_A(\alpha)$ and $C_B(\alpha)$ are orthogonal for any $\alpha\in[0,2\pi)$ and they belong to the plane defined by $C_A(0)$ and $C_B(0)$, i.e., they are orthogonal to the rest of the columns of the Hadamard matrix. Therefore, if we replace columns $C_A$ and $C_B$ in the original Hadamard matrix with $C_A(\alpha)$ and $C_B(\alpha)$, respectively, we obtain a family of complex Hadamard matrices parametrized by $\alpha$.
Note that real pairs of columns form trivially ER pairs, thus they always allow to introduce a free parameter.
In order to illustrate the procedure, let us consider the Hadamard matrix
$$
H=\left(\begin{array}{rrrr}
-1&1&1&1\\
1&-1&1&1\\
1&1&-1&1\\
1&1&1&-1
\end{array}\right).
$$
Let $C_A=(1,1,-1,1)^T$ and $C_B=(1,1,1,-1)^T$ be the third and the fourth column of $H$. Since $C_A\circ C_B^*=(1,1,-1,-1)^T$, we set $C_A(\alpha)=(1,1,-e^{i\alpha},e^{i\alpha})^T$ and $C_B(\alpha)=(1,1,e^{i\alpha},-e^{i\alpha})^T$. Then
\begin{equation}\label{Ht}
H(\alpha)=\left(\begin{array}{rrrr}
-1&1&1&1\\
1&-1&1&1\\
1&1&-e^{i\alpha}&e^{i\alpha}\\
1&1&e^{i\alpha}&-e^{i\alpha}
\end{array}\right)
\end{equation}
is a 1-parametric family of complex Hadamard matrices.
\begin{figure}[!h]
\centering
{\includegraphics[width=8cm]{Fig2.jpg}}
\caption{Introduction of a free parameter $\alpha$ in a pair of vectors of an ETF. A complex cone is generated by vectors $v_1(\alpha)$ and $v_2(\alpha)$ in such a way that the $N$ vectors $\{v_1(\alpha),v_2(\alpha),v_3,\dots,v_N\}$ define a complex ETF for every $\alpha\in[0,\pi)$. Note that $v_2(\pi)=v_1(0)$ and $v_1(\pi)=v_2(0)$.}
\label{Fig2}
\end{figure}
As we have mentioned in Section~\ref{S3}, real and complex Hadamard matrices of size $N$ are associated with real and complex ETF$((N-\sqrt{N})/2,N)$, respectively. The technique of introducing free parameters in a Hadamard matrix can be thus used for constructing parametric families of ETF$((N-\sqrt{N})/2,N)$, cf. Fig. \ref{Fig2}. It is important to remark that our method has to be applied to both columns and rows in order to obtain a Hermitian matrix. That is, the ER pair of columns $\{C_A(\alpha),C_B(\alpha)\}$ has to be complemented with the associated ER pair of rows $\{R_A(-\alpha),R_B(-\alpha)\}$, where the parameter $\alpha$ is reflected due to the requirement of hermiticity. Let us add some clarifying remarks:
\begin{itemize}
\item In the special case $N=4$, a hermitian family of complex Hadamard matrices cannot be constructed from a real symmetric Hadamard matrix with constant diagonal, because introducing rows $\{R_A(-\alpha),R_B(-\alpha)\}$ in addition to columns $\{C_A(\alpha),C_B(\alpha)\}$ in Eq.(\ref{Ht}) leads to the matrix $H(0)$. Equivalently, a family of ETF(3,4) does not exist.
\item Hermitian families exist for every real symmetric Hadamard matrix having constant diagonal 1 of even square size $N>4$, as ER pairs always exist for such cases. The first hermitian family stemming from a real Hadamard occurs for $N=16$. The 6-parametric family is presented in Appendix \ref{Ap1}. It is worth noting that this method allows us to introduce up to $N/2-1$ free parameters.
\item We can also introduce $N/2-1$ free parameters for hermitian Fourier matrices when $N$ is even \cite{G13}. For other hermitian complex Hadamard matrices of even square size $N$ the method cannot be applied for the reason that ER pairs do not exist.
\end{itemize}
Now we are ready to formulate the main result of this section, which is based on the considerations above.
\begin{prop}\label{Prop5}
There exists complex ETF$((N-\sqrt{N})/2,N)$ admitting the introduction of $N/2-1$ linearly independent free parameters for every even square value of $N>4$. For real ETF$((N-\sqrt{N})/2,N)$ the result holds if and only if a real symmetric Hadamard matrix exists.
\end{prop}
Proof follows from Theorem 3.1 in Ref.\cite{G13}. For odd values of $N$, free parameters cannot be introduced by using our method. This is so because ER pairs do not exist if $N$ is odd \cite{G13}. Given that ER pairs is the most general method to introduce free parameters in pair of columns (see \cite{G13}) we conclude that the rotation of two vectors of any ETF$((N-\sqrt{N})/2,N)$ cannot generate a family of ETF for any odd $N$.
\section{Entanglement in equiangular tight frames}\label{S5}
Equiangular tight frames have important applications in quantum mechanics, as they represent symmetric POVM quantum measurements. It is therefore interesting to study entanglement of the symmetric vectors forming a POVM, which could be important for some theoretical and experimental implementation purposes. In this section we discuss average entanglement properties of both SIC-POVM and the 6-parametric family of ETF(6,16), defined in Appendix \ref{Ap1}.
Before proceeding forward let us recall that purity of a quantum state $\rho$ is given by $P=\mathrm{Tr}(\rho^2)$. It quantifies how close is a quantum state to the surface of pure states, i.e, the surface of rank one-projectors: if $\rho$ is defined in dimension $d$, then $1/d\leq P\leq 1$, where $P=1/d$ holds for the maximally mixed state $\rho=\mathbb{I}/d$ and $P=1$ for any pure state $\rho=|\phi\rangle\langle\phi|$. For example, the entire set of quantum states for qubit systems satisfying $P=1$ determine the surface of the Bloch sphere \cite{NC11} (Poincar\'{e} sphere for mathematicians), whereas the center of the sphere represents the maximally mixed state.
An ETF($d,d^2$), i.e. SIC-POVM in dimension $d$, defined for a bipartite system $d=d_A\times d_B$ has a fixed average purity of reductions \cite{ZTE10}:
\begin{equation}\label{avpuSIC}
\frac{1}{d_Ad_B}\sum_{j=1}^{d_Ad_B}\mathrm{Tr}(\rho^2_{A_j})=(d_A + d_B)/(d_Ad_B + 1).
\end{equation}
Here, $\rho_{A_j}=\mathrm{Tr}_B(|\psi_j\rangle\langle\psi_j|)$ is the reduction to the first subsystem $A$ and $|\psi_j\rangle$ is the $j$th element of the SIC-POVM. The symbol $\mathrm{Tr}_B$ means partial trace over the second subsystem $B$. Eq.(\ref{avpuSIC}) also holds for the average purity of reductions to subsystem $B$, as the bipartite states forming the POVM are pure. The key property required to derive Eq.(\ref{avpuSIC}) is the fact that SIC-POVM are 2-designs. Other classes of ETF are only 1-designs and, therefore, it is expected that similar results do not hold.
As a novel contribution, here we show that the average purity of reductions for the 6-parametric family of ETF(6,16), presented in Appendix \ref{Ap1} depends on the parameters, and we present approximate lower and upper bounds. This family of symmetric POVM is defined for a qubit-qutrit system ($d=6=d_A\times d_B=2\times 3$). The average purity for the qubit subsystem (A) seems to lie in the range
\begin{equation}\label{bounds}
0.576737\lessapprox\frac{1}{16}\sum_{j=1}^{16}\mathrm{Tr}(\rho^2_{A_j})\lessapprox0.804885,
\end{equation}
where the approximate lower (LB) and upper (UB) bounds are attained for parameters
\begin{equation}
\vec{\alpha}_{LB}=\{0.0970, 0.0957, 0.4536, 0.7275, 0.7287, 0.2258\},
\end{equation}
and
\begin{equation}
\vec{\alpha}_{UB}=\{2.2222,2.2233,3.1401,0.4173,2.9043,2.6317\},
\end{equation}
respectively. The approximate bounds given in Eq.(\ref{bounds}) were obtained from considering numerical optimizations in Mathematica. An analytic derivation of the bounds seems to be a hard problem. As an interesting consequence of Eq.(\ref{bounds}), we noted that the 16 vectors forming the ETF(6,16) can be neither fully separable nor maximally entangled, because the average purity of reductions can attain neither the value $UB=1$ nor the value $LB=1/2$.
In general, for $N$ vectors defining an ETF in dimension $d\gg \sqrt{N}$ it is reasonable to expect weak restrictions on the average purity of reductions. Indeed, vectors forming an ETF($d,d$), i.e. an orthonormal basis in dimension $d$, can be fully separable (e.g., tensor product basis in dimension $d=2^N$ for an $N$ qubit system) or maximally entangled (e.g. generalized Bell basis in dimension $d=d'\times d'$ for a two-qudit system having $d'$ internal levels each).
\section{Algorithm to find the underlying unitary matrix}\label{S6}
In this section we present an iterative algorithm that efficiently finds the underlying unitary matrix $U$ existing behind a unistochastic matrix $B$. The algorithm works as follows:
1. [Seed]: Start from a random matrix $A^{(0)}$ of size $N$.
2. [Bistochasticity]: $A^{(0)}_{i,j}\,\rightarrow\,A^{(1)}_{i,j}=\frac{A_{i,j}}{|A_{i,j}|}\sqrt{B_{i,j}}$.
3. [Unistochasticity]: Apply the Schmidt orthogonalization to columns of $A^{(1)}$.
The iteration of steps 2 and 3 generate a sequence of matrices that converges to a unitary matrix $U$ such that $B_{ij}=|U_{ij}|^2$, for a suitable choice of the seed in step 1. Indeed, unistochastic matrices are attractive fixed points of the composed non-linear operators $T=T_3T_2$, where $T_2$ and $T_3$ are the non-linear operators associated to the mappings defined in steps 2 and 3, respectively. A solution is found when the seed belongs to the basin of attraction of a unistochastic matrix $B$. The additional steps:
$2^{\prime}$. \, [Hermiticity]: $A\,\rightarrow\,(A+A^{\dag})/2$,
$2^{\prime\prime}$. \, [Constant diagonal]: $A_{i,i}\,\rightarrow \cos(\theta)$, where $d=N\sin^2(\theta/2)$, \vspace{0.2cm}
\noindent generate the unitary hermitian matrix $U_N(\theta)$ having constant diagonal and, consequently, the unistochastic matrix $\mathcal{B}_N(\theta)$, where $d=N\sin^2(\theta/2)$. In this way, we determine a complex ETF($d,N$). A similar procedure can be applied to generate a real ETF. In the case that the imposed matrix $B$ is bistochastic but not unistochastic, the algorithm exhibit oscillations without converging.
We have implemented the above described procedure for matrices of size $N=4,\dots,22$. Up to dimension $d=19$ we were able to reproduce all known classes of complex ETF \cite{FM15} by considering $10^3$ seeds. This number of seeds was no more sufficient in dimension $d=20$, where $10^4$ seeds were required to find ETF(10,20).
In dimension $d=22$ we focused our attention on the existence of the complex ETF(11,22), which is still open. Such case would have associated an hermitian complex conference matrix $U_{22}(\pi/2)$. Our simulations considered $10^6$ random seeds and we could not find a single successful convergence of our iterative procedure, which suggests that an hermitian complex conference matrix of size 22 does not exist. On the other hand, note that a real symmetric conference matrix of size 22 does not exist, which excludes the existence of a real ETF(11,22). This is so because the number 22 is not the sum of two integer squares \cite{B50}.
\section{Conclusions}\label{S7}
We have introduced a one-to-one connection between complex equiangular tight frames (ETF) and a special kind of unistochastic matrices $\mathcal{B}_N(\theta)$, defined in Eq.(\ref{bisto}). The connection has been established in Prop. \ref{Prop1}. As a direct consequence, we have found new classes of complex ETF (see Prop. \ref{Prop2}). Furthermore, we presented new integrality restrictions for some classes of equiangular tight frames, i.e., for those having Gram matrix whose off-diagonal entries are proportional to roots of unity (see Prop. \ref{Prop3}). In particular, the only possible SIC-POVM of such kind may only exist in dimension $d=2$ or when $d+1$ is a square number (see Corol. \ref{Corol1}). We also proposed a method to introduce non-trivial free parameters in real and complex ETF (see Prop. \ref{Prop5}). To illustrate our method, we explicitly derived a 6-parametric family of complex ETF stemming from the real ETF(6,16) (see Appendix \ref{Ap1}). Moreover, we studied the average purity of reductions ($P$) for this family, which defines a symmetric POVM for a qubit-qutrit quantum system. As consequence, we have found that $P$ can be neither maximal ($P=1$ for separable states) nor minimal ($P=1/2$ for maximally entangled states). Lower and upper bounds for $P$ were derived (see Section \ref{S6}).
Furthermore, we presented an efficient algorithm to find unistochastic matrices, which is simple to implement in a computer language. By introducing a refinement of this procedure, we can also find the unitary hermitian matrices relying behind unistochastic matrices $\mathcal{B}_N(\theta)$, which allows to find an ETF($d,N$), where $d=N\sin^2(\theta/2)$. By using this procedure, we calculated all the parameters ($d,N$) for which a complex ETF($d,N$) exists, up to $N=19$. The results are consistent with known classes described in the literature \cite{FM15}. Additionally, we exhaustively studied the existence of the complex ETF(11,22), which seems not to exist. This result would imply that an hermitian complex conference matrix of size 22 does not exist.
We conclude the paper with two important open questions: (\emph{i}) Find integrality restrictions for a general complex ETF, and (\emph{ii}) Solve the Fickus conjecture \footnote{See the blog \emph{Short, fat matrices} by D. Mixon, https://dustingmixon.wordpress.com/2015/07/08/conjectures-from-sampta/}: Consider the three numbers $d$, $N-1$ and $N-d$. If a complex ETF($d,N$) exists then one of these three numbers divides the product of the other two. Our numerical simulations up to $N=19$, as well as all solutions presented in the most complete catalog of ETF, Ref. \cite{FM15}, are consistent with the conjecture.
\section*{Acknowledgements}
DG thanks to S. Friedline, M. Fickus and K. \.{Z}yczkowski for fruitful discussions on equiangular tight frames and unistochastic matrices. This work has been supported by the Polish National Science Center under the project number DEC-2011/02/A/ST1/00119 and by the Czech Science Foundation (GA\v{C}R) within the project 17-01706S.
| 58aebaccd998d5658a9e52e799040e83acf574e1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The interest in studying rare $B$ decays is immense. This is due to the circumstance that
these decays, such as $b \to (s,d) \;\gamma, \; b \to (s,d ) \; \ell^+\ell^-$, are flavour-changing-neutral-current (FCNC) processes, involving the quantum number transitions $ | \Delta B |=1, |\Delta Q |=0$. In the SM \cite{Glashow:1961tr}, they are
not allowed at the tree level, but are induced by loops and are
governed by the GIM (Glashow-Iliopoulos-Maiani) mechanism \cite{Glashow:1970gm}, which
imparts them sensitivity to higher masses, $(m_t, m_W)$.
As a consequence, they determine the CKM \cite{Cabibbo:1963yz} (Cabibbo-Kobayashi-Maskawa)
matrix elements. Of these, the elements in the third row,
$V_{td}$, $V_{ts}$ and $V_{tb}$ are of particular interest. While $\vert V_{tb}\vert$ has been measured in the
production and decays of the top quarks in hadronic collisions \cite{Agashe:2014kda}, the first two are currently not yet directly
accessible. In the SM, these CKM matrix elements have been indirectly determined from the $B^0$ - $\bar{B}^0$
and $B_s^0$ - $\bar{B}_s^0$ mixings. Rare $B$-decays provide independent measurements of the same
quantities.
In theories involving physics beyond the SM (BSM), such as the 2-Higgs doublet models
or supersymmetry, transitions involving the FCNC processes are
sensitive to the masses and couplings of the new particles. Precise experiments and
theory are needed to establish or definitively rule out the BSM effects.
Powerful calculation techniques, such as the heavy quark effective theory (HQET) \cite{Manohar:2000dt}
and the soft collinear effective theory (SCET)
\cite{Bauer:2000yr,Beneke:2002ph,Becher:2014oda}
have been developed to incorporate power $1/m_c$ and $1/m_b$ corrections to the
perturbative QCD estimates. More importantly, they enable a better theoretical description by separating the
various scales involved in $B$ decays and in establishing factorisation of the decay matrix elements.
In exclusive decays, one also needs the decay form factors and a lot of theoretical progress has been made using the lattice QCD \cite{Aoki:2016frl} and QCD sum rule techniques
\cite{Ali:1993vd,Colangelo:2000dp,Ball:2004ye,Straub:2015ica}, often complementing each other,
as they work best in the opposite $q^2$-ranges. It is this continued progress in QCD calculational framework,
which has taken us to the level of sophistication required to match the experimental advances.
In this paper, I review what, in my view, are some of the key measurements in the radiative, semileptonic and
leptonic rare $B$-decays and confront them with the SM-based calculations, carried out with the
theoretical tools just mentioned. However, this is not a comprehensive review of this subject, but the hope
is that the choice of topics reflects both the goals achieved in explaining some landmark measurements and focus on
open issues. In section 2, I review the inclusive and some exclusive radiate rare $B$-decays.There are no
burning issues in this area - at least not yet. In section 3, the corresponding inclusive and exclusive
semileptonic decays are taken up. Again, there are no open issues in the inclusive semileptonic decays, but
experimental precision is limited currently, which is bound to improve significantly at Belle II. There are, however,
a lot of open issues in the exclusive semileptonic decays, in particular in $R_K$, the ratio of the
decay widths for $B \to K \mu^+ \mu^-$ and $B \to K e^+ e^-$, hinting at the possible breakdown of lepton universality,
the linchpin of the SM, reviving the interest in low-mass leptoquarks. One should also mention here
similar issues in tree-level semileptonic decays, such as $R^{\tau/\ell}_D$ and $R^{\tau/\ell}_{D^*}$,
the ratios involving the decays $B \to D^{(*)} \tau \nu_\tau$ and $B \to D^{(*)} \ell \nu_\ell$ ($\ell= e, \mu$).
There are also other dissenting areas, which go by the name of $P_5^\prime$-anomaly, which is a certain coefficient
in the angular description of the decay $B \to K^* \mu^+ \mu^-$, which presumably need
a better theoretical (read QCD) description than is available at present. They are discussed in section 3.
In section 4, we discuss the CKM-suppressed $b \to d \ell^+ \ell^-$ decays, which is a new experimental
frontier initiated by the LHCb measurements of the branching fraction and the dimuon invariant mass distribution
in the decay $B^\pm \to \pi^\pm \mu^+ \mu^-$. Finally, the rarest $B$- and $B_s$-decays measured so far,
$B \to \mu^+ \mu^-$ and $B_s \to \mu^+ \mu^-$, are taken up in section 5. Current measurements also
show some (mild) deviations in their branching ratios versus the SM. A representative global fit of the data on the semileptonic and leptonic rare $B$-decays in terms of the Wilson coefficients from possible new physics is shown in section 6. Some concluding remarks are made in section 7.
\section{Rare Radiative $B$-decays in the SM and Experiments}
In 1993, the CLEO collaboration at the Cornell $e^+e^-$ collider measured the decay
$B \to K^* \gamma$ \cite{Ammar:1993sh},
initiating the field of rare $B$-decays, followed two years later by the measurement of
the inclusive decay $B \to X_s \gamma$ \cite{Alam:1994aw}.
The branching ratio $ {\cal}(B \to K^* \gamma) = (4.5 \pm 1.5 \pm 0.9)\times 10^{-5}$ was in agreement
with the SM estimates, but theoretical uncertainty was large. Measuring the Inclusive process
$B \to X_s \gamma$ was challenging, but as the
photon energy spectrum in this process was already
calculated in 1990 by Christoph Greub and me\cite{Ali:1990vp}, this came in handy for the
CLEO measurements \cite{Chen:2001fja}
shown in Fig. \ref{fig:belle-celo-bsgama} (left frame) and compared with the theoretical prediction \cite{Ali:1990vp},
Since then, a lot of experimental and theoretical effort has gone in the precise measurements and in performing
higher order perturbative and non-perturbative calculations. As a consequence,
$ B \to X_s \gamma$ has now become the standard candle of FCNC processes, with the measured
branching ratio and the precise higher order SM-based calculation providing valuable constraints on
the parameters of BSM physics. The impact of the $B$-factories on this measurement can be judged by the scale in Fig. \ref{fig:belle-celo-bsgama} (right frame),
which is due to the Belle collaboration \cite{Koppenburg:2004fz}.
\begin{figure}
\includegraphics[width=2.0in]{np-lhc-figs/CLEO-Spectrum.jpg}
\includegraphics[width=2.35in]{np-lhc-figs/BELLE-Spectrum.jpg}
\caption{Photon energy spectrum in the inclusive decay $B \to X_s \gamma$
measured by CLEO (left frame) \cite{Chen:2001fja} and Belle (right frame) \cite{Koppenburg:2004fz}.}
\label{fig:belle-celo-bsgama}
\end{figure}
The next frontier of rare $B$-decays involves the so-called electroweak penguins, which govern
the decays of the inclusive processes $ B \to (X_s , X_d) \ell^+ \ell^-$ and the exclusive decays
such as $B \to (K,K^*, \pi) \ell^+ \ell^-$. These processes have rather small branching ratios and hence
they were first measured at the $B$-factories. Inclusive decays remain their domain, but
experiments at the LHC, in particular, LHCb, are now at the forefront of exclusive semileptonic decays.
Apart from these, also the leptonic $B$-decays $ B_s \to \mu^+\mu^-$ and
$ B_d \to \mu^+\mu^-$ have been measured at the LHC.
I will review some of the key measurements
and the theory relevant for their interpretation. This description is anything but comprehensive, for which I refer to some recent excellent references \cite{Blake:2016olu,Descotes-Genon:2015uva,Koppenburg:2016rji,Bevan:2014iga}
and resources, such as HFAG \cite{Amhis:2014hma} and FLAG \cite{Aoki:2016frl}.
\subsection{Inclusive decays $B \to X_s \gamma$ at NNLO in the SM}
The leading order diagrams for the decay $b \to s \gamma$ are shown are shown in
Fig. \ref{fig:leading-bsgamma-diag}, including also the tree diagram for $b \to u \bar{u} s \gamma$, which yields a
soft photon. The first two diagrams are anyway suppressed due to the CKM matrix elements, as indicated.
The charm- and top- quark contributions enter with opposite signs, and the relative contributions
indicted are after including the leading order (in $\alpha_s$)
QCD effects. A typical diagram depicting perturbative QCD corrections due to the exchange of a gluon
is also shown.
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/slide3.jpg}}
\caption{Examples of the leading electroweak diagrams for $B \to X_s \gamma$ from the up,
charm, and top quarks. A diagram involving a gluon exchange
is shown in the lower figure.}
\label{fig:leading-bsgamma-diag}
\end{figure}
The QCD logarithms $\alpha_s \ln M_W^2/m_b^2$ enhance the branching ratio ${\cal B}(B \to X_s \gamma)$
by more than a factor 2, and hence such logs have to be resummed. This is done using an effective field theory
approach, obtained by integrating out the top quark and the $W^\pm$ bosons. Keeping terms up to dimension-6,
the effective Lagrangian for $B \to X_s \gamma$ and $B \to X_s \ell^+ \ell^-$ reads as follows:
\begin{eqnarray}
{\cal L} \;= \;\; {\cal L}_{QCD \times QED}(q,l)
\;\;+ \frac{4 G_F}{\sqrt{2}} V_{ts}^* V_{tb} \sum_{i=1}^{10} C_i(\mu) O_i \nonumber
\end{eqnarray}
($q=u,d,s,c,b$,~~$l = e,\mu, \tau$)
\begin{eqnarray}
O_i = \left\{ \begin{array}{lll}
(\bar{s} \Gamma_i c)(\bar{c} \Gamma'_i b),
& i=1, (2), & C_i(m_b) \sim -0.26 ~(1.02)\\[4mm]
(\bar{s} \Gamma_i b) {\Sigma}_q (\bar{q} \Gamma'_i q),
& i=3,4,5,6,~ & |C_i(m_b)| < 0.08\\[4mm]
\frac{e m_b}{16 \pi^2} \bar{s}_L \sigma^{\mu \nu} b_R F_{\mu \nu},
& i=7, & C_7(m_b) \sim -0.3\\[4mm]
\frac{g m_b}{16 \pi^2} \bar{s}_L \sigma^{\mu \nu} T^a b_R G^a_{\mu \nu},
& i=8, & C_8(m_b) \sim -0.16\\[4mm]
\frac{e^2}{16 \pi^2} (\bar{s}_L \gamma_{\mu} b_L) (\bar{l}
\gamma^{\mu} {(\gamma_5)} l),
& i=9,{(10)} & C_i(m_b) \sim 4.27~ (-4.2)
\end{array} \right.
\nonumber
\end{eqnarray}
Here, $G_F$ is the Fermi coupling constant, $V_{ij}$ are the CKM matrix elements, $O_i$ are the
four-Fermi and dipole operators, and $C_i(\mu)$ are the Wilson coefficients, evaluated
at the scale $\mu$, which is taken typically as $\mu=m_b$, and their values in the NNLO accuracy
are given above for $\mu=4.8$ GeV. Variations due to a different choice
of $\mu$ and uncertainties from the upper scale-setting $m_t/2 \leq \mu_0 \leq 2 m_t$ can be seen elsewhere
\cite{Blake:2016olu}.
There are three essential steps of the calculation:
\vspace*{-3mm}
\begin{itemize}
\item \underline{Matching}:~~Evaluating
$C_i(\mu_0)$ at $\mu_0 \sim M_W$ by requiring the equality of the SM and
the effective theory Green functions.
\item \underline{Mixing}:~~ Deriving the effective theory renormalisation group equation (RGE)
and evolving $C_i(\mu)$ from $\mu_0$ to $\mu_b \sim m_b$.
\item \underline{Matrix elements}:~~ Evaluating the on-shell
amplitudes at $\mu_b \sim m_b$.
\end{itemize}
\vspace*{-3mm}
All three steps have been improved in perturbation theory and now include the next-to-next-to-leading order effects (NLLO),
i.e., contributions up to $O(\alpha_s^2(m_b))$. A monumental theoretical effort stretched well over a decade
with the participation of a large number of theorists underlies the current theoretical precision of the branching
ratio. The result is usually quoted for a threshold photon energy to avoid experimental background from other
Bremsstrahlung processes. For the decay with $E_\gamma > 1.6$ GeV in the rest frame of the $B$ meson,
the result at NNLO accuracy is \cite{Misiak:2015xwa,Czakon:2015exa}
\begin{equation}
{\cal B}(B \to X_s \gamma)= (3.36 \pm 0.23) \times 10^{-4},
\end{equation}
where the dominant SM uncertainty is non-perturbative \cite{Benzke:2010js}. This is
to be compared with the current experimental average of the same \cite{Amhis:2014hma}
\begin{equation}
{\cal B}(B \to X_s \gamma)= (3.43 \pm 0.21 \pm 0.07) \times 10^{-4},
\end{equation}
where the first error is statistical and the second systematic, yielding a ratio $1.02 \pm 0.08$,
providing a test of the SM to an accuracy better than 10\%.
The CKM-suppressed decay $B \to X_d \gamma$ has also been calculated in the NNLO precision.
The result for $E_\gamma > 1.6$ GeV is \cite{Misiak:2015xwa}
\begin{equation}
{\cal B}(B \to X_d \gamma)= (1.73 ^{+0.12}_{-0.22}) \times 10^{-5}.
\end{equation}
This will be measured precisely at Belle II. The constraints on the CP asymmetry are not very restrictive, but the
current measurements are in agreement with the SM expectation. For further details, see HFAG \cite{Amhis:2014hma}.
\subsection{Bounds on the charged Higgs mass from ${\cal B}(B \to X_s \gamma)$}
As the agreement between the SM and data is excellent, the decay rate for $B \to X_s \gamma$ provides constraints on
the parameters of the BSM theories, such as supersymmetry and the 2 Higgs-doublet models
(2HDM). In calculating the BSM effects, depending on the model, the SM operator basis may have to be enlarged,
but in many cases one anticipates additive contributions to the Wilson coefficients in the SM basis. In the context
of $ B \to X_s \gamma$, it is customary to encode the BSM effects in the Wilson coefficients of the
dipole operators $C_7(\mu)$ and $C_8(\mu)$, and the constraints from the branching ratio
on the additive coefficients $\Delta C_7$ and $\Delta C_8$ then takes
the numerical form \cite{Misiak:2015xwa}
\begin{equation}
{\cal B}(B \to X_s \gamma) \times 10^4 = (3.36 \pm 0.23) -8.22 \Delta C_7 -1.99 \Delta C_8.
\end{equation}
To sample the kind of constraints that can be derived on the parameters of the BSM models, the 2HDM is a
good case, as the branching ratio for the decay $B \to X_s \gamma$ in this model has been derived to the same theoretical accuracy \cite{Hermann:2012fc}.The Lagrangian for the 2HDM is
\begin{equation}
{\cal L}_{H^+}= (2\sqrt{2} G_F)^{1/2}\Sigma_{i,j=1}^{3} \bar{u}_i (A_u m_{u_i}
V_{ij}P_L - A_d m_{d_j}V_{ij} P_R) d_j H^* + h.c.,
\end{equation}
where $V_{ij}$ are the CKM matrix elements and
$P_{L/R}=(1 \mp \gamma_5)/2 $. The 2HDM contributions to the Wilson coefficients are proportional to
$A_iA_j^*$, representing the contributions from the up-type $ A_u$ and down-type $ A_d $ quarks. They
are defined in terms of the ratio of the vacuum expectation values, called $\tan \beta$, and are model dependent.
\begin{itemize}
\item 2HDM of type-I: $A_u=A_d=\frac{1}{\tan \beta},$
\item 2HDM of type-II: $A_u=-1/A_d=\frac{1}{\tan \beta}$.
\end{itemize}
Examples of Feynman diagrams contributing to $B \to X_s \gamma$ in the 2HDM are shown in
Fig. \ref{fig:2HDM-bsgamma-diag}.
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/Misiak-2HDM-NNLO-fig2.png}}
\caption{Sample Feynman diagrams that matter for $B \to X_s \gamma$ in the 2HDM \cite{Hermann:2012fc}.
$H^\pm$ denotes a charged Higgs. }
\label{fig:2HDM-bsgamma-diag}
\end{figure}
Apart from $\tan \beta$, the other parameter of the 2HDM is the mass of the charged Higgs $M_H^\pm$.
As ${\cal B}(B \to X_s \gamma)$
becomes insensitive to $\tan \beta$ for larger values, $\tan \beta>2$ , the 2HDM contribution depends essentially on $M_H^\pm$. The current measurements and the SM estimates then provide constraints on $M_H^\pm$,
as shown in Fig. \ref{fig:2HDM-Steinhauser}, \footnote{I thank Matthias Steinhauser for providing this figure.}
updated using \cite{Misiak:2015xwa,Hermann:2012fc},
yielding \cite{Misiak:2015xwa}
$M_H^\pm > 480$ GeV (\@ 90\% C.L.) and $M_H^\pm > 358$ GeV (\@ 99\% C.L.) .
These constraints are
competitive to the direct searches of the $H^\pm$ at the LHC. %
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/Steinhauser-2HDM.png}}
\caption{Constraints on the charged Higgs mass $m_H^\pm$ from
${\cal B}(B \to X_s \gamma)$ in the 2HDM \cite{Misiak:2015xwa,Hermann:2012fc}. Measured branching ratio (exp) and the SM
estimates are also shown. The curves demarcate the central values and $\pm 1\sigma$ errors.}
\label{fig:2HDM-Steinhauser}
\end{figure}
\subsection{Exclusive radiative rare $B$ decays}
Exclusive radiative decays, such as $B \to V \gamma$ ($V=K^*, \rho, \omega $) and
$B_s \to \phi \gamma$, have been well-measured at the $B$ factories. In addition, they
offer the possibility of measuring CP- and isospin asymmetries, a topic I will not discuss here. Theoretically,
exclusive decays are more
challenging, as they require the knowledge of the form factors at $q^2=0$, which can not be
calculated directly using Lattice QCD. However, light-cone QCD sum rules \cite{Ball:2004ye,Straub:2015ica}
also do a good job for
calculating heavy $\to$ light form factors at low-$q^2$. In addition, the matrix elements require gluonic exchanges between the spectator quark and the active quarks (spectator-scattering),
introducing intermediate scales in the decay rates. Also long-distance
effects generated by the four-quark operators with charm quarks are present and are calculable in
limited regions \cite{Khodjamirian:2010vf}.
Thus, exclusive decays are theoretically not as precise as the inclusive decay $B \to X_s \gamma$. However, techniques embedded in HQET and SCET have led to the
factorisation of the decay matrix elements into perturbatively calculable (hard) and non-perturbative
(soft) parts, akin to the deep inelastic scattering processes. These factorisation-based approaches
are the main work-horse in this field. Renormalisation group (RG) methods then allow to
sum up large logarithms, and this program has been carried out to a high accuracy.
A detailed discussion of the various techniques requires a thorough review, which can't be carried out here.
I will confine myself by pointing to some key references, beginning from the QCD factorisation approach,
pioneered by Beneke, Buchalla, Neubert and Sachrajda \cite{Beneke:1999br}, which has been applied to
the radiative decays $B \to (K^*, \rho, \omega) \gamma$
\cite{Beneke:2001at,Ali:2001ez,Bosch:2001gv,Beneke:2004dp}. Another theoretical framework, called
pQCD \cite{Li:1994iu,Keum:2000wi}, has also been put to use in these decays \cite{Keum:2004is,Lu:2005yz}.
The SCET-based methods have also been harnessed \cite{Chay:2003kb,Becher:2005fg}.
The advantage of SCET is that it allows for an unambiguous separation of the scales, and an operator definition
of each object in the factorisation formula can be given. Following the QCD factorisation
approach, a factorisation formula for the $B \to V \gamma$ matrix element can be written in SCET as well
\begin{equation}
\langle V \gamma \vert Q_i \vert B \rangle = \Delta_i C^{A} \xi_{V_\perp}
+ \frac{ \sqrt{m_B} F f_{V_\perp}} {4} \int dw du\; \phi_+^{B} (w) \phi_\perp^{V}(u)\; t_i^{II},
\label{eq:scet-fact}
\end{equation}
where $F$ and $ f_{V_\perp} $ are meson decay constants; $\phi_+^{B} (w)$ and $\phi_\perp^{V}(u) $ are the
light-cone distribution amplitudes for the $B$- and $V$-meson, respectively.
The SCET form factor $ \xi_{V_\perp} $ is
related to the QCD form factor through perturbative and power corrections, and the perturbative hard QCD
kernels are the coefficients $ \Delta_i C^{A} $ and $t_i^{II} $. They are known to complete NLO accuracy in
RG-improved perturbation theory \cite{Becher:2005fg}.
The factorisation formula (\ref{eq:scet-fact}) has been calculated to NNLO accuracy in SCET \cite{Ali:2007sj} (except for the NNLO corrections from the spectator scattering). As far as the
decays $B \to K^* \gamma$ and $B_s \to \phi \gamma$ are concerned, the partial NNLO theory
is still the state-of-the-art.
Their branching ratios as well as the ratio of the decay rates
$ {\cal B}(B_s \to \phi \gamma/{\cal B}(B \to K^* \gamma)$ are given in Table \ref{tab:exclusive-rad-decays},
together with the current experimental averages \cite{Amhis:2014hma}. The corresponding calculations for the CKM-suppressed
decays $ B \to (\rho,\omega) \gamma$ are not yet available to the desired theoretical accuracy, due to the annihilation contributions, for which,
to the best of my knowledge, no factorisation theorem of the kind discussed above has been proven.
The results from a QCD-Factorisation based approach \cite{Ali:2001ez}
for $ B \to \rho \gamma$ are also
given in Table \ref{tab:exclusive-rad-decays} and compared with the data. The exclusive decay rates
shown are in agreement with the experimental measurements, though theoretical precision is not better than 20\%. Obviously, there is need for a better theoretical description, more so as Belle II will measure the radiative decays with greatly improved precision. I will skip a discussion of the isospin and CP asymmetries in these decays,
as the current experimental bounds \cite{Amhis:2014hma} are not yet probing the SM in these observables.
\begin{table}
\begin{center}
\tbl{Measurements [HFAG 2014] \cite{Amhis:2014hma} and SM-based estimates of
${\cal B}(B \to ( K^*, \rho) \gamma)$ and ${\cal B}(B_s \to \phi \gamma)$
in units of $ 10^{-5}$, and the ratio ${\cal B}(B_s \to \phi \gamma)/{\cal B}(B^0 \to K^{*0} \gamma)$. }
{\begin{tabular}{| l | l | l |}\hline
Decay Mode & Expt.~(HFAG)
& Theory (SM) \vphantom{$\fbox{\Big[}$} \\ \hline
$B^0 \to K^{*0} \gamma$&$ 4.33 \pm 0.15$
& $4.6 \pm 1.4 $ \vphantom{$\fbox{\Big[}$} \\
\hline
$B^+\to K^{*+} \gamma $&$4.21 \pm 0.18$ & $ 4.3 \pm 1.4$ \vphantom{$\fbox{\Big[}$} \\
\hline
$B_s\to \phi \gamma $&$ 3.59 \pm 0.36 $ & $ 4.3 \pm 1.4 $ \vphantom{$\fbox{\Big[}$} \\ \hline
$ B_s\to \phi \gamma/ B^0 \to K^{*0} \gamma $ & $0.81 \pm 0.08$ & $1.0 \pm 0.2$ \vphantom{$\fbox{\Big[}$}
\\ \hline
$B^0 \to \rho^0 \gamma $ & $0.86^{+0.15}_{-0.14}$ & $ 0.65 \pm 0.12$ \vphantom{$\fbox{\Big[}$} \\
\hline
$B^+ \to \rho^+ \gamma $ & $0.98 \pm 0.25 $ & $ 1.37 \pm 0.26$ \vphantom{$\fbox{\Big[}$} \\
\hline
\label{tab:exclusive-rad-decays}
\end{tabular}}
\end{center}
\end{table}
\section{Semileptonic $b \to s$ decays $B \to (X_s , K, K^*) \ell^+ \ell^-$}
There are two $b \to s$ semileptonic operators in SM:
\begin{equation}
O_i = \frac{e^2}{16 \pi^2} (\bar{s}_L \gamma_{\mu} b_L)
(\bar{l} \gamma^{\mu} (\gamma_5) l), \hspace{1.5cm} i=9,(10)
\nonumber
\end{equation}
Their Wilson Coefficients have the following perturbative expansion:
\begin{eqnarray}
C_9(\mu) &=& \frac{4 \pi}{\alpha_s(\mu)} C_9^{(-1)}(\mu)
~+~ C_9^{(0)}(\mu) ~+~ \frac{\alpha_s(\mu)}{4 \pi} C_9^{(1)}(\mu)
~+~ ...\nonumber\\
C_{10} &=& C_{10}^{(0)} ~+~\frac{\alpha_s(M_W)}{4 \pi} C_{10}^{(1)} ~+~ ...
\nonumber
\end{eqnarray}
The term
$C_9^{(-1)}(\mu)$ reproduces the electroweak
logarithm that originates from the photonic penguins with charm quark
loops, shown below \cite{Ghinculov:2003qd}.
\centerline{\includegraphics[width=3.5cm,angle=0]{np-lhc-figs/030103-Tobias1.jpg}}
The first two terms in the perturbative expansion of $C_9(m_b)$ are
\begin{eqnarray}
&& C_9^{(0)}(m_b) \simeq ~2.2; ~~\frac{4 \pi}{\alpha_s(m_b)} \,C_9^{(-1)}(m_b) =
\frac{4}{9}\, \ln \frac{M_W^2}{m_b^2} + {\cal O}(\alpha_s) \simeq~2.
\nonumber
\end{eqnarray}
As they are very similar in magnitude,
one needs to calculate the NNLO contribution to get reliable estimates of the decay rate.
In addition, leading power corrections in $1/m_c$ and $1/m_b$ are required.
\subsection{Inclusive semileptonic decays $B \to X_s \ell^+ \ell^-$}
A lot of theoretical effort has gone into calculating the perturbative
QCD NNLO , electromagnetic logarithms and power corrections
\cite{Ghinculov:2003qd,Asatryan:2001zw,Ali:2002jg,Huber:2005ig,Huber:2007vv}.
The B-factory experiments
Babar and Belle have measured the dilepton invariant mass spectrum $d{\cal B}(B \to X_s \ell^+ \ell^-)/dq^2$
practically in the entire kinematic region
and have also measured the so-called Forward-backward lepton asymmetry $A_{FB}(q^2)$ \cite{Ali:1991is}. They are shown in
Fig. \ref{fig:belle-babar-bsll}, and compared with the SM-based theoretical calculations.
Note that a cut of $q^2 > 0.2$ GeV$^2$ on the dilepton invariant squared mass is used.
As seen in these
figures, two resonant regions near $q^2=M_{J/\psi}^2$ and $q^2=M_{J/\psi^\prime}^2$ have to be excluded
when comparing with the short-distance contribution.
They make up what is called the long-distance contribution from the processes
$B \to X_s + (J/\psi, J/\psi^\prime) \to X_s + \ell^+ \ell^-$, whose dynamics is determined by the hadronic matrix
elements of the operators $O_1$ and $O_2$. They have also been calculated via
a dispersion relation \cite{Kruger:1996cv} and data on the measured quantity
$R_{\rm had} (s)= \sigma (e^+ e^- \to {\rm hadrons})/\sigma (e^+ e^- \to \mu^+ \mu ^-)$, and in some
analyses are also included.
As the (short-distance) contribution is expected
to be a smooth function of $q^2$, one uses the perturbative distributions in interpolating these
regions as well. The experimental distributions are in agreement with the SM, including also the zero point
of $A_{FB}(q^2)$, which is a sensitive function of the ratio of the two Wilson coefficients $C_9$ and $C_{10}$.
The branching ratio for the inclusive decay $B \to X_s \ell^+ \ell^-$ with a lower cut on the
dilepton invariant mass $q^2 > 0.2~ {\rm GeV}^2$ at NNLO accuracy is \cite{Ali:2002jg}
\begin{equation}
{\cal B}(B \to X_s \ell^+ \ell^-)= (4.2 \pm 0.7) \times 10^{-6},
\end{equation}
to be compared with the current experimental average of the same \cite{Amhis:2014hma}
\begin{equation}
{\cal B}(B \to X_s \ell^+ \ell^-)= (3.66^{+0.76}_{-0.77}) \times 10^{-6}.
\end{equation}
The two agree within theoretical and experimental errors. The experimental cuts which are imposed to
remove the $J/\psi$ and $\psi^\prime$ resonant regions are indicated in Fig. \ref{fig:belle-babar-bsll}.
The effect of logarithmic QED corrections becomes
important for more restrictive cuts on $q^2$, and they have been worked out for different choices of the
$q^2$-range in a recent paper \cite{Huber:2015sra}.
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/Babar-bsll-2013.jpg}}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/Belle-bsll-afb-2014.jpg}}
\caption{Dilepton invariant mass Distribution measured by BaBar \cite{Lees:2013nxa} (upper frame) and the Forward-backward Asymmetry $A_{\rm FB}$ measured by Belle \cite{Sato:2014pjr} (lower frame) in $B \to X_s \ell^+ \ell^-$.
The curve(above) and the band (below) are the SM expectations, discussed in the text.}
\label{fig:belle-babar-bsll}
\end{figure}
\subsection{Exclusive Decays $B \to (K,K^*) \ell^+ \ell^-$}
The $B \to K$ and $B \to K^*$ transitions involve the following weak currents:
\begin{equation}
\Gamma_\mu^1=\bar{s}\gamma_\mu (1-\gamma_5) b, ~~
\Gamma_\mu^2=\bar{s}\sigma_{\mu\nu}q^\nu (1+\gamma_5) b.
\end{equation}
Their matrix elements involve altogether 10 non-perturbative $q^2$-dependent functions (form factors),
denoted by the following functions: \footnote{As we will also discuss later the decays $ B \to \pi \ell^+ \ell^- $,
we distinguish the $B \to K$ and $B \to \pi$ form factors by a superscript.}
\begin{eqnarray}
\langle K \vert \Gamma_\mu^1 \vert B \rangle && \supset f^K_+(q^2),
f^K_-(q^2). \nonumber \\
\langle K \vert \Gamma_\mu^2 \vert B \rangle && \supset f_T^K(q^2). \nonumber \\
\langle K^* \vert \Gamma_\mu^1 \vert B \rangle && \supset V(q^2),
A_1(q^2),A_2(q^2),A_3(q^2). \nonumber \\
\langle K^* \vert \Gamma_\mu^2 \vert B \rangle && \supset
T_1(q^2),T_2(q^2),T_3(q^2).\nonumber
\end{eqnarray}
Data on $B \to K^* \gamma$ provides normalisation of $T_1(0)=T_2(0)
\simeq 0.28$. These form factors have been calculated using a number of non-perturbative
techniques, in particular the QCD sum rules \cite{Ball:2004ye,Bharucha:2010im} and Lattice QCD
\cite{Dalgic:2006dt,Bouchard:2013zda,Bailey:2015nbd}.
They are complementary to
each other, as the former are reliable in the low-$q^2$ domain and the latter can calculate
only for large-$q^2$. They are usually combined to get reliable profiles of the form factors in
the entire $q^2$ domain. However, heavy quark symmetry allows to reduce the number of independent form
factors from 10 to 3 in low-$q^2$ domain $(q^2/m_b^2 \ll 1)$.
Symmetry-breaking corrections have been evaluated \cite{Beneke:2000wa}.
The decay rate, dilepton invariant mass distribution and the
Forward-backward asymmetry in the low-$q^2$ region have been calculated for $B \to K^* \ell^+ \ell^-$
using the SCET formalism \cite{Ali:2006ew}.
Current measurements of the branching ratios in the inclusive and exclusive semileptonic decays
involving $b \to s$ transition are summarised in Table \ref{tab:exclusive-decays} and compared with the
corresponding SM estimates.
The inclusive measurements and the SM rates include a cut on the dilepton invariant mass
$M_{\ell^+\ell^-} > 0.2$ GeV.
They are in agreement with each other, though precision is
currently limited due to the imprecise knowledge of the form factors.
\subsection{Current tests of lepton universality in semileptonic $B$-decays}
%
Currently, a number of measurements in $B$ decays suggests a breakdown of the lepton $(e, \mu, \tau)$
universality in semileptonic processes. In the SM, gauge bosons couple with equal strength to all three
leptons and the couplings of the Higgs to a pair of charged lepton is proportional to the charged lepton mass,
which are negligibly small for $\ell^+ \ell^-= e^+ e^-, \mu^+ \mu^-$. Hence, if the lepton non-universality is
experimentally established, it would be a fatal blow to the SM.
We briefly summarise the experimental situation starting from the decay $ B^\pm \to K^\pm \ell^+ \ell^- $,
whose decay rates were discussed earlier.
Theoretical accuracy is vastly improved if instead of the absolute rates, ratios of the decay rates
are computed.
Data on the decays involving $K^{(*)} \tau^+ \tau^-$ is currently
sparse, but first measurements of the ratios involving the final states
$K^{(*)} \mu^+ \mu^-$ and $ K ^{(*)} e^+ e^- $ are available. In particular, a 2.6$\sigma$ deviation from the
$e$-$\mu$ universality is reported by the LHCb collaboration in the ratio involving $B^\pm \to K^\pm \mu^+ \mu^-$
and $B^\pm \to K^\pm e^+ e^-$ measured in the low-$q^2$ region,
which can be calculated rather accurately.
In the interval $ 1 \leq q^2 \leq 6$ GeV$^2$, LHCb finds \cite{Aaij:2014ora}
\begin{equation}
R_K \equiv \frac{\Gamma(B^\pm \to K^\pm \mu^+ \mu^-)} {\Gamma(B^\pm \to K^\pm e^+ e^-)}
= 0.745^{+0.090}_{-0.074} ({\rm stat}) \pm 0.035 ({\rm syst}).
\label{eq:RK-anomaly}
\end{equation}
This ratio in the SM is close to 1 to a very high accuracy \cite{Bobeth:2007dw} over the entire $q^2$ region measured by the LHCb. Thus, the measurement in (\ref{eq:RK-anomaly}) amounts to
about $2.6\sigma$ deviation from the SM. Several BSM scenarios have been proposed to account for the
$R_K$ anomaly, discussed below, including a $Z^\prime$-extension of the SM \cite{Chiang:2016qov}.
It should, however, be noted that the currently measured
branching ratios
${\cal B}(B^\pm \to K^\pm e^+ e^-)= (1.56^{+ 0.19 + 0.06}_{-0.15 -0.4})\times 10^{-7} $ and
${\cal B}(B^\pm \to K^\pm \mu^+ \mu^-)= (1.20 \pm 0.09 \pm 0.07)\times 10^{-7} $
are also lower than the SM estimates ${\cal B}^{\rm SM}(B^\pm \to K^\pm e^+ e^-)=
{\cal B}^{\rm SM} (B^\pm \to K^\pm \mu^+ \mu^-)= (1.75^{+0.60}_{-0.29}) \times 10^{-7}$, and the experimental
error on the ${\cal B}(B^\pm \to K^\pm e^+ e^-) $ is twice as large. One has to also factor in that the electrons
radiate very profusely (compared to the muons) and implementing the radiative corrections in hadronic machines
is anything but straight forward.
In coming years, this and similar ratios, which can also be calculated to high accuracy, will be measured with greatly
improved precision at the LHC and Belle II.
The other place where lepton non-universality is reported is in the ratios of the decays
$B \to D^{(*)} \tau \nu_\tau$ and $B \to D^{(*)} \ell \nu_\ell$. Defining
\begin{equation}
R^{\tau/\ell}_{D^{(*)}} \equiv \frac{{\cal B}(B \to D^{(*)} \tau \nu_\tau)/ {\cal B} ^{\rm SM}(B \to D^{(*)} \tau \nu_\tau) }
{{\cal B}(B \to D^{(*)} \ell \nu_\ell)/ {\cal B} ^{\rm SM}(B \to D^{(*)} \ell \nu_\ell)},
\label{eq:RDDS-tau}
\end{equation}
the current averages of the BaBar, Belle, and the LHCb data are \cite{Amhis:2014hma}:
\begin{equation}
R^{\tau/\ell}_D= 1.37 \pm 0.17;\hspace{3mm } R^{\tau/\ell}_{D^*}= 1.28 \pm 0.08.
\label{eq:RDDS-tau-expt}
\end{equation}
This amounts to about $3.9\sigma$ deviation from the $\tau/\ell $ ($\ell= e, \mu $) universality. Interestingly, this
happens in a tree-level charged current process. If confirmed experimentally, this would call for a
drastic contribution to an effective four-Fermi $LL$ operator
$( \bar{c} \gamma_\mu b_L)(\tau_L \gamma_\mu \nu_L )$. It is then conceivable that the
non-universality in $R_K$ (which is a loop-induced $b \to s$ process) ia also due to an $LL$ operator
$( \bar{s} \gamma_\mu b_L)(\bar{\mu}_L \gamma_\mu \mu_L )$. Several suggestions along these lines
involving a leptoquark have been made \cite{Hiller:2014yaa,Bauer:2015knc,Barbieri:2015yvd}. It is
worth recalling that leptoquarks were introduced by Pati and Salam in 1973 in an attempt to unify
leptons and quarks in $SU(4)$ \cite{Pati:1973uk,Pati:1974yy}. The lepton non-universality in $B$ decays
has revived the interest in theories with low-mass leptoquarks, discussed recently
in a comprehensive work on this topic \cite{Dorsner:2016wpm}.
\begin{table}
\begin{center}
\tbl{Measurements [PDG 2014] and SM-based estimates \cite{Ali:2002jg} of the
branching ratios ${\cal B}(B \to (X_s, K, K^*) \ell^+ \ell^-)$ in units of $ 10^{-6}$}
{\begin{tabular}{| l | l | l |}\hline
Decay Mode & Expt.~(BELLE \& BABAR)
& Theory (SM) \vphantom{$\fbox{\Big[}$} \\ \hline
$B\to K\ell^+\ell^-$&$ 0.48\pm 0.04$
& $0.35 \pm 0.12 $ \vphantom{$\fbox{\Big[}$} \\
\hline
$B\to K^*e^+e^-$&$1.19 \pm 0.20 $ & $1.58 \pm 0.49$ \vphantom{$\fbox{\Big[}$} \\
\hline
$B\to K^*\mu^+\mu^-$&$1.06 \pm 0.09 $ & $1.19 \pm 0.39$ \vphantom{$\fbox{\Big[}$} \\ \hline
$B\to X_s \mu^+ \mu^-$ & $4.3 \pm 1.2$ & $4.2\pm 0.7$ \vphantom{$\fbox{\Big[}$}
\\ \hline
$B\to X_s e^+ e^-$ & $4.7 \pm 1.3$ & $ 4.2\pm 0.7$ \vphantom{$\fbox{\Big[}$} \\
\hline
$B\to X_s \ell^+ \ell^-$ & $4.5 \pm 1.0 $ & $ 4.2\pm 0.7$ \vphantom{$\fbox{\Big[}$} \\
\hline
\label{tab:exclusive-decays}
\end{tabular}}
\end{center}
\end{table}
\subsection{Angular analysis of the decay $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$}
For the inclusive decays $ B \to X_s \ell^+ \ell^-$, the observables which have been measured are the
integrated rates, the dilepton invariant mass $d\Gamma/dq^2 $ and the FB asymmetry $A_{\rm FB}(q^2) $.
They are all found to be in agreement with the SM. In the exclusive decays such as $B \to K^* \ell^+ \ell^-$
and $B_s \to \phi \ell^+ \ell^-$, a complete angular analysis of the decay is experimentally feasible.
This allows one to measure a number of additional observables, defined below.
\begin{eqnarray}
\frac{1}{d(\Gamma + \bar\Gamma)} \frac{d^4(\Gamma + \bar\Gamma)}{dq^2 d\Omega }
&=& \frac{9}{32 \pi}\left[\frac{3}{4} (1-F_L)\sin^2\theta_K + F_L \;\cos^2\theta_K \right. \nonumber \\
&& \left.+ \frac{1}{4} (1-F_L)\sin^2\theta_K\;\cos 2\theta_\ell - F_L \cos^2\theta_K \cos 2\theta_\ell \right. \nonumber \\
&& \left.+S_3 \sin^2 \theta_K\; \sin^2\theta_\ell\;\cos 2\phi + S_4 \sin 2\theta_K \sin 2 \theta_\ell \cos\phi\right. \nonumber \\
&& \left.+S_5 \sin 2\theta_K \sin \theta_\ell \cos\phi + \frac{4}{3}\; A_{\rm FB}\sin^2\theta_K \cos \theta_\ell\right. \nonumber \\
&& \left.+ S_7 \sin 2\theta_K \sin \theta_\ell\;\sin\phi + S_8 \sin 2\theta_K \sin 2\theta_\ell\;\sin\phi\right. \nonumber \\
&& \left.+ S_9 \sin^2 \theta_K \sin^2\theta_\ell\;\sin 2\phi
\right].
\label{eq:angular-observables}
\end{eqnarray}
The three angles $\theta_K$, $ \theta_\ell $ and $ \phi $ for the decay $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$
are defined in Fig. \ref{fig:angular-analysis}.%
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/lhcb-angular-analysis.png}}
\caption{Definitions of the angles in $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$.}
\label{fig:angular-analysis}
\end{figure}
An angular analysis of the decay chains
$B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$ \cite{Aaij:2015oid} and $B_s^0 \to \phi ( \to K^+ K^-) \mu^+ \mu^-$
\cite{Aaij:2013aln} has been carried out by LHCb.
The observables in (\ref{eq:angular-observables}) are
$q^2$-dependent coefficients of the Wilson coefficients and hence they probe the underlying dynamics.
Since these coefficients have been calculated to a high accuracy, the remaining theoretical
uncertainty lies in the form factors and also from the charm-quark loops. The form factors have been
calculated using the QCD sum rules and in the high-$q^2$ region also using lattice QCD. They limit the
current theoretical accuracy. However, a number of
so-called optimised observables has been proposed \cite{Descotes-Genon:2013vna}, which reduce the dependence on the form factors. Using the
LHCb convention, these observables are defined as \cite{Aaij:2015oid}
\begin{eqnarray}
P_1 &\equiv& 2 S_3/(1- F_L);~~P_2 \equiv 2 A_{\rm FB}/ 3(1-F_L);~~P_3 \equiv -S_9/(1-F_L), \nonumber\\
&& P^\prime_{4,5,6,8} \equiv S_{4,5,7,8}/\sqrt{F_l(1-F_L)}.
\end{eqnarray}
\begin{figure}
\centerline{\includegraphics[width=4.5in]{np-lhc-figs/lhcb-max-likelihood.png}}
\caption{CP-averaged variables in bins of $q^2$ for the observables
$F_{\rm L}$, $A_{\rm FB}$,
$S_3$ and $S_5$ in $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$ measured by LHCb \cite{Aaij:2015oid} and
comparison with the SM \cite{Altmannshofer:2014rta}.}
\label{fig:lhcb-ang-analys-results-1}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=4.5in]{np-lhc-figs/lhcb-2015-fig-8.jpg}}
\caption{The optimised angular observables in bins of $q^2$ in the decay $B^0 \to K^{*0}( \to K^+ \pi^-)\; \mu^+ \mu^-$
measured by the LHCb collaboration \cite{Aaij:2015oid} and comparison with the SM \cite{Descotes-Genon:2014uoa}.}
\label{fig:lhcb-ang-analys-results-2}
\end{figure}
These angular observables have been analysed in a number of theoretical studies
\cite{Descotes-Genon:2014uoa,Jager:2012uw,Jager:2014rwa,Altmannshofer:2014rta,Straub:2015ica,Hurth:2013ssa},
which differ in the treatment of their non-perturbative input, mainly form factors.
The LHCb collaboration, which currently
dominates this field, has used these SM-based estimates and compared with their data in various
$q^2$ bins. Two representative comparisons based on the theoretical estimates from
Altmannshofer and Straub \cite{Altmannshofer:2014rta} and Descotes-Genon, Hofer, Matias and
Virto \cite{Descotes-Genon:2014uoa} are shown in Figs. \ref{fig:lhcb-ang-analys-results-1} and
\ref{fig:lhcb-ang-analys-results-2}, respectively. They are largely in agreement with the data, except for the
distributions in the observables $S_5(q^2)$ (in Fig. \ref{fig:lhcb-ang-analys-results-1}) and
$P_5^\prime(q^2)$ (in Fig. \ref{fig:lhcb-ang-analys-results-2})
in the bins around $q^2 \geq 5$ GeV$^2$.
The pull on the SM depends on the theoretical model, reaching 3.4$\sigma$ in the bin
$4.3 \leq q^2 \leq 8.68$ GeV$^2$ compared to DHMV \cite{Descotes-Genon:2014uoa}.
There are deviations of a similar nature, between 2 and 3$\sigma$, seen in
the comparison of $S_5$ and other quantities, such as the partial branching ratios in $B \to K^* \mu^+ \mu^-$,
$B _s \to \phi \mu^+ \mu^-$ and $F_L(q^2)$ \cite{Altmannshofer:2014rta}.
An analysis of the current Belle data \cite{Abdesselam:2016llu} , shown in Fig. \ref{fig:belle-ang-analys-results-2},
displays a similar pattern as the one reported by LHCb. As the Belle data has larger errors,
due to limited statistics, the resulting pull on the SM is less significant. In the interval
$4.0 \leq q^2 \leq 8.0$ GeV$^2$, Belle reports deviations of $2.3\sigma$ (compared to DHMV \cite{Descotes-Genon:2014uoa}), $1.72\sigma$ (compared to BSZ \cite{Straub:2015ica}) and $1.68 \sigma$ (compared to
JC \cite{Jager:2014rwa}). These measurements will improve greatly at Belle II.
\begin{figure}
\centerline{\includegraphics[width=5.5in]{np-lhc-figs/Belle-P5.png}}
\caption{The optimised angular observables $P_4^\prime $ and $P_5^\prime $
in bins of $q^2$ in the decay $B^0 \to K^{*0}( \to K^+ \pi^-)\; \mu^+ \mu^-$
measured by Belle \cite{Abdesselam:2016llu} and comparison with the SM \cite{Descotes-Genon:2014uoa}.}
\label{fig:belle-ang-analys-results-2}
\end{figure}
To quantify the deviation of the LHCb data from the SM estimates,
a $\Delta \chi^2$ distribution for the real part of the Wilson coefficient $ {\rm Re} C_9(m_b) $ is shown in Fig.
\ref{fig:lhcb-chi-square}. In calculating the $\Delta \chi^2$, the other Wilson coefficients are set to their SM values.
The coefficient $ {\rm Re} C_9^{\rm SM}(m_b)=4.27 $ at the NNLO accuracy in the SM is indicated by a vertical line.
The best fit of the LHCb data yields a value which is shifted from the SM, and the deviation in this coefficient
is found to be $ \Delta {\rm Re} C_9 (m_b)=-1.04 \pm 0.25 $. The deviation is tantalising, but not yet
conclusive. A bit of caution is needed here as the SM estimates used in the analysis above may have to be revised, once the residual
uncertainties are better constrained. In particular, the hadronic contributions generated by the four-quark operators with charm are difficult to estimate, especially around $q^2 \sim 4 m_c^2$, leading to an effective shift in the
value of the Wilson coefficient being discussed \cite{Ciuchini:2015qxb}.
\begin{figure}
\centerline{\includegraphics[width=3.5in]{np-lhc-figs/lhcb-2015-fig-14.jpg}}
\caption{The $\Delta \chi^2$ distribution for the real part of the Wilson coefficient $ {\rm Re} C_9(m_b) $
from a fit of the CP-averaged observables $F_{\rm L}, A_{\rm FB}, S_3, ...,S_9$ in $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$ by the LHCb collaboration \cite{Aaij:2015oid}.}
\label{fig:lhcb-chi-square}
\end{figure}
\section{CKM-suppressed $ b \to d \ell^+ \ell^- $ transitions in the SM}
Weak transitions $ b \to d \ell^+ \ell^- $ , like the radiative decays $b \to d \gamma$, are CKM suppressed and because of this the structure of the
effective weak Hamiltonian is different than the one encountered earlier for the $ b \to s \ell^+ \ell^- $
transitions.
\begin{eqnarray}
{\cal H}_{\rm eff}^{b \to d} &=& - \frac{4 G_G }{\sqrt{2}}\left[V_{tb}^* V_{td} \sum_{i=1}^{10}
C_i(\mu)(O_i(\mu)\right. \nonumber \\
&& \left. + V_{ub}^* V_{ud} \sum_{i=1}^{2} C_i(\mu)( O_i(\mu) - O_i(\mu) )\right] + {\rm h.c.}.
\end{eqnarray}
Here $O_i(\mu)$ are the dimension-six operators introduced earlier (except for the interchange $s \to d$
quark) in $ {\cal H}_{\rm eff}^{b \to s}$. As the two CKM factors
are comparable in magnitude $\vert V_{tb}^* V_{td}\vert \simeq \vert V_{ub}^* V_{ud} \vert $, and have different weak phases,
we anticipate sizeable CP-violating asymmetries in both the inclusive $ b \to d \ell^+ \ell^- $ and
exclusive transitions, such as $B \to (\pi, \rho) \ell^+ \ell^-$.
The relevant operators appearing in $ {\cal H}_{\rm eff}^{b \to d}$ are:
\underline{Tree operators}
\begin{equation}
{\cal O}_1 =
\left ( \bar d_L \gamma_\mu T^A c_L \right )
\left ( \bar c_L \gamma^\mu T^A b_L \right ), \quad
{\cal O}_2 =
\left ( \bar d_L \gamma_\mu c_L \right )
\left ( \bar c_L \gamma^\mu b_L \right ) ,
\end{equation}
\begin{equation}
{\cal O}_1^{(u)} =
\left ( \bar d_L \gamma_\mu T^A u_L \right )
\left ( \bar u_L \gamma^\mu T^A b_L \right ), \quad
{\cal O}_2^{(u)} =
\left ( \bar d_L \gamma_\mu u_L \right )
\left ( \bar u_L \gamma^\mu b_L \right ).
\end{equation}
\underline{Dipole operators}
\begin{equation}
{\cal O}_7 =
\frac{e \, m_b}{g_{\rm s}^2}
\left ( \bar d_L \sigma^{\mu \nu} b_R \right ) F_{\mu \nu}, \quad
{\cal O}_8 =
\frac{m_b}{g_{\rm s}}
\left ( \bar d_L \sigma^{\mu \nu} T^A b_R \right ) G_{\mu \nu}^A.
\end{equation}
\underline{Semileptonic operators}
\begin{equation}
{\cal O}_9 =
\frac{e^2}{g_{\rm s}^2}
\left ( \bar d_L \gamma^\mu b_L \right )
\sum_\ell \left ( \bar \ell \gamma_\mu \ell \right ), \quad
{\cal O}_{10} =
\frac{e^2}{g_{\rm s}^2}
\left ( \bar d_L \gamma^\mu b_L \right )
\sum_\ell \left ( \bar \ell \gamma_\mu \gamma_5 \ell \right ).
\end{equation}
Here, $e(g_s)$ is the QED (QCD) coupling constant. Since the inclusive decay $B \to X_d \ell^+ \ell^-$ has
not yet been measured, but hopefully will be at Belle II, we discuss the exclusive decay $B^+ \to \pi^+ \ell^+ \ell^- $, which
is the only $b \to d$ semileptonic transition measured so far.
%
\subsection{Exclusive decay $B^+ \to \pi^+ \ell^+ \ell^- $}
The decay \footnote{Charge conjugation is implied here.} $B^+ \to \pi^+ \ell^+ \ell^- $ is induced
by the vector and tensor currents and their matrix elements are defined as
\begin{equation}
\langle \pi (p_\pi) | \bar b \gamma^\mu d | B (p_B) \rangle =
f^\pi_+ (q^2) \left ( p_B^\mu + p_\pi^\mu \right ) +
\left [ f^\pi_0 (q^2) - f^\pi_+ (q^2) \right ]
\displaystyle\frac{m^2_B - m^2_\pi}{q^2} q^\mu,
\end{equation}
%
\vspace*{-6mm}
\begin{equation}
\langle \pi (p_\pi) | \bar b \sigma^{\mu \nu} q_\nu d | B (p_B) \rangle =
\displaystyle\frac{ i f^\pi_T (q^2)} {m_B + m_\pi}
\left [ \left ( p_B^\mu + p_\pi^\mu \right ) q^2 -
q^\mu \left ( m_B^2 - m_\pi^2 \right ) \right ].
\end{equation}
These form factors are related to the ones in the decay $B \to K \ell^+ \ell^-$, called $f^K_i(q^2) $,
discussed earlier, by $SU(3)_F$
symmetry. Of these, the form factors $f^\pi_+(q^2) $ and $f^\pi_0(q^2) $ are related by isospin symmetry to the corresponding ones
measured in the charged current process $B^0 \to \pi^- \ell^+ \nu_\ell $ by Babar and Belle, and they can
be extracted from the data. This has been done using several parameterisations of the form factors with all of them giving an adequate description of the data \cite{Ali:2013zfa}. Due to their analytic properties, the so-called
$z$-expansion methods, in which the form factors are expanded in a Taylor series in $z$, employed in the
Boyd-Grinstein-Lebed (BGL) parametrisation \cite{Boyd:1994tt} and the Bourrely-Caprini-Lellouch (BCL) \cite{Bourrely:2008za} parametrisation, are preferable.
The BGL parametrisation is used in working out the decay rate and the invariant dilepton mass
distribution \cite{Ali:2013zfa} for $B^+ \to \pi^+ \ell^+ \ell^- $, which is discussed below.
The BCL-parametrisation is used by
the lattice-QCD groups, the HPQCD \cite{Bouchard:2013zda,Bouchard:2013pna}
and Fermilab/MILC \cite{Bailey:2015nbd} collaborations, to determine the form factors
$f^\pi_i(q^2) $ and $f^K_i(q^2)$. In particular, the Fermilab/MILC collaboration has worked out the dilepton
invariant mass distribution in the decay of interest $B^+ \to \pi^+ \ell^+ \ell^- $, making use of their
simulation in the large-$q^2$ region and extrapolating with the BCL parametrisation.
We first discuss the low-$q^2$ region ($q^2 \ll m_b^2$). In this case,, heavy quark symmetry (HQS) relates all three
form factors of interest $f^\pi_i(q^2) $
and this can be used advantageously to have a reliable estimate of the dilepton invariant mass
spectrum in this region. Including lowest order HQS-breaking, the resulting expressions
for the form factors (for $q^2/m_b^2 \ll 1 $) are
worked out by Beneke and Feldmann \cite{Beneke:2000wa}.
Thus, fitting the form factor $f_+ (q^2)$ from the charged current data on
$B \to \pi \ell^+ \nu_\ell$ decay, and taking into account the HQS and its breaking, lead to a model-independent predictions of the differential branching ratio (dimuon mass spectrum) in the neutral current process $ B^+ \to \pi^+ \ell^+ \ell^-$ for low-$q^2$ values. However, the long-distance contribution, which arises from the
processes $B^+ \to \pi^+ (\rho^0, \omega) \to \pi^+ \mu^+ \mu^-$ are not included here.
The SM invariant dilepton mass distribution in $ B^+ \to \pi^+ \ell^+ \ell^-$
integrated over the range $1 {\rm GeV}^2 \leq q^2 \leq 8 {\rm GeV}^2$
yields a partial branching ratio
\begin{equation}
{\cal B}(B^+ \to \pi^+ \mu^+ \mu^-) = (0.57 ^{+0.07}_{-0.05}) \times 10^{-8}.
\end{equation}
Thanks to the available data on the charged current process and heavy quark symmetry, this enables an
accuracy of about 10\% for an exclusive branching ratio, comparable to the theoretical accuracy in the
inclusive decay $B \to X_s \gamma$, discussed earlier. Thus,
the decay $ B^+ \to \pi^+ \mu^+ \mu^- $ offers
a key advantage compared to the decay $ B^+ \to K^+ \ell^+ \ell^-$, in which case the
charged current process is not available.
The differential branching ratio in the entire $q^2$ region is given by
%
\begin{equation}
\frac{ d {\cal B}(B^+ \to \pi^+ \ell^+ \ell^-)}{d q^2} = C_B \vert V_{tb}V_{td}^* \vert^2 \sqrt{\lambda(q^2)}
\sqrt{1- \frac{4 m_\ell^2}{q^2}} F(q^2),
\end{equation}
where the constant $C_B=G_F^2 \alpha_{\rm em}^2 \tau_B/1024 \pi^5 m_B^3$ and $ \lambda(q^2)$ is the
usual kinematic function $\lambda(q^2)= (m_B^2 + m_\pi^2 -q^2)^2 - 4 m_B^2 m_\pi^2 $. The function
$F(q^2)$ depends on the effective Wilson coefficients, $C_7^{\rm eff}$, $C_9^{\rm eff}$, and $C_{10}^{\rm eff}$,
and the three form factors $f_+^\pi (q^2)$, $f_0^\pi (q^2)$ and $f_T^\pi (q^2)$.
A detailed discussion of the determination of the form factors, of which only $f_+^\pi (q^2)$ and $f_T^\pi (q^2)$
are numerically important for $\ell^\pm = e^\pm, \mu^\pm$ is given elsewhere \cite{Ali:2013zfa}. We recall that
$f_+^\pi (q^2)$ is constrained by the data on the charged current process in the entire
$q^2$ domain. In addition, the lattice-QCD results on the form factors in the large-$q^2$ domain
and the HQS-based relations in the low-$q^2$ region provide sufficient constraints on the form factor.
This has enabled a rather precise determination of the invariant dilepton mass distribution
in $B^+ \to \pi^+ \ell^+ \ell^-$.
Taking into account the various parametric and form-factor dependent uncertainties, this yields the following
estimate for the branching ratio for $B^+ \to \pi^+ \mu^+ \mu^- $ \cite{Aaij:2015nea}
\begin{equation}
{\cal B}_{\rm SM}(B^+ \to \pi^+ \mu^+ \mu^-) = (1.88 ^{+0.32}_{-0.21}) \times 10^{-8},
\label{eq:Bpimu}
\end{equation}
to be compared with the measured branching ratio by the LHCb collaboration \cite{Aaij:2015nea}
(based on 3${\rm fb}^{-1}$ data):
\begin{equation}
{\cal B}_{\rm LHCb}(B^+ \to \pi^+ \mu^+ \mu^-) = (1.83 \pm 0.24 \pm 0.05) \times 10^{-8},
\label{eq:BpimuLHCb}
\end{equation}
where the first error is statistical and the second systematic, resulting in excellent agreement.
The dimuon invariant mass distribution measured by the LHCb collaboration \cite{Aaij:2015nea}
is shown in Fig. \ref{fig:lhcb-Bpimumu-result}, and compared with
the SM-based theoretical prediction, called APR13 \cite{Ali:2013zfa}, and the
lattice-based calculation, called FNAL/MILC 15 \cite{Bailey:2015nbd}.
Also shown is a comparison with a calculation, called HKR \cite{Hambrock:2015wka} , which has
essentially the same short-distance contribution in the low-$q^2$ region, as discussed earlier, but additionally takes into account the contributions from the lower resonances $\rho, \omega$ and $\phi$.
This adequately describes the distribution in the $q^2$ bin, around
1 GeV$^2$.
With the steadily improving lattice calculations for the various input hadronic quantities and the
form factors, theoretical error indicated in Eq. (\ref{eq:Bpimu}) will go down considerably. Experimentally,
we expect rapid progress due to the increased statistics at the LHC, but also from Belle II, which will
measure the corresponding distributions and branching ratio also in the decays $ B^+ \to \pi^+ e^+ e^- $,
and $B^+ \to \pi^+ \tau^+ \tau^- $, providing a complementary test of
the $e$-$\mu$-$\tau$ universality in $b \to d$ semileptonic transitions.
\begin{figure}
\centerline{\includegraphics[width=4.5in]{np-lhc-figs/lhcb-fig-4.jpg}}
\caption{Comparison of the dimuon invariant mass distribution
in $B^+ \to \pi^+ \mu^+ \mu^-$ in the SM with the LHCb data \cite{Aaij:2015nea}.
Theoretical distributions shown are : APR13 \cite{Ali:2013zfa}, HKR15 \cite{Hambrock:2015wka},
and FNAL/MILC \cite{Bailey:2015nbd}. }
\label{fig:lhcb-Bpimumu-result}
\end{figure}
\section{Leptonic Rare $B$ Decays}
The final topic discussed in this write-up involves purely leptonic decays $ B_s^0 \to \ell^+ \ell^-$
and $ B^0 \to \ell^+ \ell^-$ with $\ell^+ \ell^- =e^+ e^-, \mu^+ \mu^-, \tau^+ \tau^- $. Of these,
$ {\cal B}(B_s^0 \to \mu^+ \mu^-)= (2.8^{+0.7}_{-0.6}) \times 10^{-9} $ is now well measured, and the corresponding CKM-suppressed decay $ {\cal B}(B^0 \to \mu^+ \mu^-)= (3.9^{+1,6}_{-1.4}) \times 10^{-10} $ is almost on the verge
of becoming a measurement. These numbers are from the combined CMS/LHCb
data \cite{CMS:2014xfa}. From the experimental point of view, their measurement is
a real {\it tour de force}, considering the tiny branching ratios and the formidable background at the LHC.
In the SM, these decays are dominated by the axial-vector operator
$O_{10} = (\bar{s}_\alpha \gamma^\mu P_L b_\alpha) (\bar{\ell} \gamma_\mu \gamma_5 \ell) $.
In principle, the operators $O_S= m_b (\bar{s}_\alpha \gamma^\mu P_R b_\alpha) (\bar{\ell} \ell)$ and
$O_P= m_b (\bar{s}_\alpha \gamma^\mu P_R b_\alpha) (\bar{\ell} \gamma_5 \ell)$ also contribute, but are
chirally suppressed in the SM. This need not be the case in BSM scenarios, and hence the
great interest in measuring these decays precisely. In the SM, the measurement of $ {\cal B}(B_s^0 \to \mu^+ \mu^-)$ and $ {\cal B}(B^0 \to \mu^+ \mu^-)$ provide a measurement of the
Wilson coefficient $C_{10} (m_b)$. Their ratio
$ {\cal B}(B^0 \to \mu^+ \mu^-)/{\cal B}(B_s^0 \to \mu^+ \mu^-) $ being proportional to the ratio of the
CKM matrix-elements $\vert V_{td}/V_{ts}\vert^2$ is an important constraint on the CKM unitarity triangle.
%
\begin{figure}
\centerline{\includegraphics[width=4.5in]{np-lhc-figs/LHCb-CMS-mumu.png}}
\caption{Likelihood contours in the ${\cal B}(B^0 \to \mu^+ \mu^-)$ versus ${\cal B}(B_s^0 \to \mu^+ \mu^-)$
plane. The (black) cross in (a) marks the best-fit value and the SM expectation is shown as the (red) marker.
Variations of the test statistics $-2 \Delta \ln L$ for ${\cal B}(B_s^0 \to \mu^+ \mu^-)$ (b) and
${\cal B}(B^0 \to \mu^+ \mu^-)$ (c) are shown. The SM prediction is denoted with the vertical (red) bars.
(From the combined CMS-LHCb data \cite{CMS:2014xfa}.) }
\label{fig:LHCb-CMS-mumu}
\end{figure}
%
The decay rate $\Gamma(B_s^0 \to \mu^+ \mu^-)$ in the SM can be written as
%
\begin{equation}
\Gamma(B_s^0 \to \mu^+ \mu^-)= \frac{G_F^2 M_W^2 m_{B_s}^3 f_{B_s}^2} { 8 \pi^5} \vert V_{tb}^*V_{ts}\vert^2
\frac{4 m_\ell^2}{ m_{B_s}^2} \sqrt{1- \frac{4 m_\ell^2}{ m_{B_s}^2} } \vert C_{10}\vert^2 + O(\alpha_{\rm em}).
\end{equation}
The coefficient $C_{10} $ has been calculated by taking into account the NNLO QCD corrections and
NLO electroweak corrections, but the $O(\alpha_{\rm em})$ contribution indicated above is ignored, as it is small. The SM branching
ratio in this accuracy have been obtained in \cite{Bobeth:2013uxa,Hermann:2013kca,Bobeth:2013tba}, where a careful account of the various
input quantities is presented. The importance of including the effects of the width difference $\Delta \Gamma_s$
due to the $B_s^0$ - $\bar{B}_s^0$ mixings in extracting the branching ratio for $B_s \to \mu^+ \mu^-$
has been emphasised in the literature \cite{DeBruyn:2012wk} and is included in the analysis.
The time-averaged branching ratios, which in the SM to a good approximation equals to
$\overline {\cal B} (B_s \to \mu^+ \mu^-)= \Gamma( B_s \to \mu^+ \mu^-)/\Gamma_H (B_s)$, where
$\Gamma_H (B_s) $ is the heavier mass-eigenstate total width, is given below \cite{Bobeth:2013uxa}
\begin{equation}
\overline {\cal B} (B_s \to \mu^+ \mu^-) =(3.65 \pm 0.23) \times 10^{-9}.
\label{eq:Bobeth-Bsmumu}
\end{equation}
In evaluating this, a value $f_{B_s}=227.7(4.5)$ MeV was used from the earlier FLAG average \cite{Aoki:2013ldr}. In the most recent
compilation by the FLAG collaboration \cite{Aoki:2016frl} , this coupling constant has been updated to
$f_{B_s}=224(5)$ MeV, which reduces the branching ratio to $\overline {\cal B} (B_s \to \mu^+ \mu^-) =(3.55 \pm 0.23) \times 10^{-9}$. This is compatible with the current measurements to about 1$\sigma$, with the uncertainty
dominated by the experiment.
The corresponding branching ratio $\overline {\cal B} (B^0 \to \mu^+ \mu^-)$ is
evaluated as \cite{Bobeth:2013uxa}
\begin{equation}
\overline {\cal B} (B^0 \to \mu^+ \mu^-) =(1.06 \pm 0.09) \times 10^{-10},
\label{eq:Bobeth-Bdmumu}
\end{equation}
which, likewise, has to be scaled down to $ (1.01 \pm 0.09) \times 10^{-10} $, due to the current
average \cite{Aoki:2016frl}
$f_{B}=186(4)$ MeV, compared to $f_{B}=190.5(4.2)$ MeV used in deriving the result given in
Eq. (\ref{eq:Bobeth-Bdmumu}). This is about 2$\sigma$ below the current measurement, and the ratio of the
two leptonic decays $\overline {\cal B} (B_s \to \mu^+ \mu^-)/\overline {\cal B} (B^0 \to \mu^+ \mu^-) $ is
off by about 2.3$\sigma$. The likelihood contours in the ${\cal B}(B^0 \to \mu^+ \mu^-)$ versus ${\cal B}(B_s^0 \to \mu^+ \mu^-)$ plane from the combined CMS/LHCb data are shown in Fig. \ref{fig:LHCb-CMS-mumu}.
The anomalies in the decays $ B \to K^* \mu^+ \mu^-$, discussed previously, and the deviations in
$ {\cal B}(B^0 \to \mu^+ \mu^-)$ and ${\cal B}(B_s^0 \to \mu^+ \mu^-)$, if consolidated experimentally,
would require an extension of the SM. A recent proposal based on the group $SU(3)_C \times SU(3)_L \times U(1)$
is discussed by Buras, De Fazio and Girrbach \cite{Buras:2013dea}. Lepton non-universality, if confirmed,
requires a leptoquark-type solution. A viable candidate theory to replace the SM and accounting for all the
current anomalies, in my opinion, is not in sight.
\section{Global fits of the Wilson Coefficients $C_9$ and $C_{10}$}
As discussed in the foregoing, a number of deviations from the SM-estimates are currently present in the
data on semileptonic and leptonic rare $B$-decays. They lie mostly around 2 to 3$\sigma$.
A comparison of the LHCb data on a number of angular observables
$F_{\rm L}, A_{\rm FB}, S_3, ...,S_9$ in $B^0 \to K^{*0}( \to K^+ \pi^-) \mu^+ \mu^-$ with the SM-based
estimates was shown in Fig, \ref{fig:lhcb-chi-square}, yielding a value of ${\rm Re}( C_9)$ which deviates
from the SM by about 3$\sigma$. A number of groups has undertaken similar fits of the data and the outcome
depends on a number of assumed correlations. However, it should be stressed that there are still
non-perturbative contributions present in the current theoretical estimates which are not
yet under complete quantitative control. The contributions from the charm quarks in the loops is a case in
point. Also, form factor uncertainties are probably larger than assumed in some of
these global fits.
As a representative example of the kind of constraints on the Wilson coefficients $C_9$ and $C_{10}$ that
follow from the data on semileptonic and leptonic decays of the $B$ mesons is shown in
Fig. \ref{fig:fermilab-milc-c9-10} from the Fermilab/MILC collaboration \cite{Bailey:2015nbd}. This shows that the SM point indicated by
$(0,0)$ in the ${\rm Re}(C_{9}^{NP},{\rm Re}(C_{10}^{NP})$-plane lies a little beyond 2$\sigma$. In some other
fits, the deviations are larger but still far short for a discovery of BSM effects.
As a lot of the experimental input in this and similar analysis is due to the LHCb data, this has to be confirmed
by an independent experiment. This, hopefully, will be done by Belle II. We are better advised to wait and see
if these deviations become statistically significant enough to warrant new physics. Currently, the situation is
tantalising but not conclusive.
%
\begin{figure}
\centerline{\includegraphics[width=4.5in]{np-lhc-figs/Fermilab-Fig-14.jpg}}
\caption{Present constraints on the Wilson coefficients ${\rm Re} C_{10}^{\rm NP}$ vs.
${\rm Re} C_{9}^{\rm NP}$ from the semileptonic rare $B$-decays and $B_s \to \mu^+ \mu^-$.
The SM-point is indicated.
(From Fermilab/MILC Lattice Collaboration \cite{Bailey:2015nbd}.)}
\label{fig:fermilab-milc-c9-10}
\end{figure}
\section{Concluding Remarks}
From the measurement by the CLEO collaboration of the rare decay $B \to X_s \gamma$ in 1995,
having a branching ratio of about $3 \times 10^{-4}$, to the rarest of the measured $B$ decays, $B^0 \to \mu^+ \mu^-$,
with a branching fraction of about $1 \times 10^{-10}$ by the LHCb and CMS collaborations, SM has been
tested over six orders of magnitude. This is an impressive feat, made possible by dedicated experimental
programmes carried out with diverse beams and detection techniques over a period of more than 20 years.
A sustained theoretical effort has accompanied the experiments all along, underscoring both the continued
theoretical interest in $b$ physics and an intense exchange between the two communities. With the exception of a
few anomalies, showing deviations from the SM ranging between 2 to 4$\sigma$ in statistical significance,
a vast majority of the measurements is in quantitative agreement with the SM. In particular, all quark flavour
transitions are described by the CKM matrix whose elements are now determined. The CP asymmetry
measured so far in laboratory experiments is explained by the Kobayashi-Maskawa phase. FCNC processes,
of which rare $B$-decays discussed here is a class, are governed by the GIM mechanism, with the particles
in the SM (three families of quarks and leptons, electroweak gauge bosons, gluons, and the Higgs) accounting for all the observed phenomena - so far. Whether this astounding consistence will continue will be tested in the
coming years, as the LHC experiments analyse more data, enabling vastly improved precision in some
of the key measurements discussed here. In a couple of years from now, Belle II will start taking data, providing
independent and new measurements. They will be decisive in either deepening our knowledge about the SM,
or hopefully in discovering the new frontier of physics.
\section{Acknowledgment}
I thank Harald Fritzsch for inviting me to this very stimulating conference and Prof. K.K. Phua for the
warm hospitality in Singapore. I also thank Mikolaj Misiak for reading the manuscript and helpful suggestions.
| da056c99f573ac95f1c54b0081ad17c55265353d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Topological complexity is a numerical homotopy invariant, introduced by Farber in the course of his topological study of the robot motion planning problem \cite{Far03, Far06}. For any space $X$, the number ${\sf TC}(X)$ is defined to be the sectional category of the free path fibration on $X$, and as such gives a quantitative measure of the complexity of navigation in $X$. Computation of ${\sf TC}(X)$ for a given space $X$ can be delicate, but is often achievable by combining cohomological lower bounds (in terms of the zero-divisors cup-length \cite[Theorem 7]{Far03}) with upper bounds coming from obstruction theory or the specific geometry of the space at hand.
A class of spaces for which the computation of topological complexity presents a unique challenge are the Eilenberg--Mac Lane spaces $K(\pi,1)$, for $\pi$ a torsion-free discrete group. A description of ${\sf TC}(\pi)\coloneqq {\sf TC}(K(\pi,1))$ in terms of algebraic properties of the group $\pi$ (as requested by Farber in \cite{Far06}) seems to be out of reach at present. In certain cases one can often compute the exact value of ${\sf TC}(\pi)$ using the bounds mentioned above. There are standard bounds ${\sf cd}(\pi)\leq {\sf TC}(\pi)\leq 2\, {\sf cd}(\pi)$ in terms of the cohomological dimension of $\pi$. There are also sharper bounds in terms of cohomological dimensions of certain subgroups or quotient groups of $\pi\times \pi$ (see \cite[Theorem 1.1]{GLO} and \cite[Proposition 3.7]{G}, reproduced as Theorems \ref{thm:lowerbound} and \ref{thm:upperbound} below) which can be of use when the cohomological or obstruction-theoretic bounds are either insufficient, or computationally infeasible.
In this paper we investigate the topological complexity of certain subgroups of Artin's braid groups. Recall that the full braid group $B_n$, the pure braid group $P_n$ and the symmetric group $\mathfrak{S}_n$ fit into an extension
\[
\xymatrix{
1 \ar[r] & P_n \ar[r] & B_n \ar[r]^-{\pi} & \mathfrak{S}_n \ar[r] & 1}
\]
where the projection $\pi$ sends a braid to the associated permutation of its endpoints (definitions will be given in Section \ref{Braidgroups} below). Given any subgroup $G\le \mathfrak{S}_n$, its pre-image $B_n^G\coloneqq\pi^{-1}(G)$ is a subgroup of $B_n$ containing $P_n$. The cohomological dimension of $B_n^G$ is $n-1$, and so by the standard bounds mentioned above we have $n-1\le {\sf TC}(B_n^G)\le 2n-2$. We will prove the following.
\begin{thm}\label{thm:introupper}
Suppose that $G\leq \mathfrak{S}_n$ satisfies either of the following conditions:
\begin{itemize}
\item $G\leq\mathfrak{S}_{n-k}\times \mathfrak{S}_k$ where $(n,k)=(n-1,k)=(n-1,k-1)=1$, or
\item $G\leq \mathfrak{S}_{n-2}\times \{1\}^2$.
\end{itemize}
Then we have
\[
{\sf TC}(B_n^G)\leq 2n-3.
\]
\end{thm}
\begin{thm}\label{thm:introlower}
Let $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ for $k\ge2$. Then \[{\sf TC}(B_n^G)\ge 2n-k-1.\]
Furthermore, if $G\le \mathfrak{S}_{n-1}\times \{1\}$, then \[{\sf TC}(B_n^G)\ge 2n-3.\]
\end{thm}
\begin{cor}\label{exact}
If $G\leq \mathfrak{S}_{n-2}\times \{1\}^2$, then
\[
{\sf TC}(B_n^G)=2n-3.
\]
\end{cor}
These results extend the computation of the topological complexity of pure braid groups (due to Farber and Yuzvinsky \cite{FY}) to various \emph{mixed} or \emph{coloured braid groups}. Our methods are somewhat different (in particular we use \cite[Theorem 1.1]{GLO} instead of zero-divisors cup-length).
The above results generalise to the setting of higher topological complexity. The generalised results are formally very similar and thus we do not give them in the introduction. They are discussed in Section \ref{highertc}. We should mention that the results for higher topological complexity extend the computation of the higher topological complexity of pure braid groups due to Gonz\'alez and Grant \cite{GG}. Again, our methods are somewhat different to theirs, which are in turn similar to the ones used by Farber and Yuzvinsky in \cite{FY}.
Configuration spaces of points in the plane give models for Eilenberg--Mac Lane spaces of braid groups, including those considered in this paper (see \lemref{lem:keypione} below). Therefore, our results have implications for motion planning of $n$ agents moving in a planar region avoiding collisions, where the agents are partitioned into equivalence classes according to their function.
Several of the results in this paper were inspired by corresponding results in \cite{CGJ}; we thank the authors of that paper for their correspondence and in particular for providing an algebraic proof of \lemref{lem:centre}. We also thank Jes\'us Gonz\'alez for suggesting that we generalise our results to higher topological complexity.
\section{Topological complexity of aspherical spaces
For any space $X$, let $p_X:X^I\to X\times X$ denote the free path fibration on $X$, with projection $p_X(\gamma) = (\gamma(0),\gamma(1))$. Recall that the \emph{topological complexity} of $X$, denoted ${\sf TC}(X)$, is defined to be the minimal $k$ such that $X\times X$ admits a cover by $k+1$ open sets $U_0,U_1,\ldots , U_k$, on each of which there exists a local section of $p_X$ (that is, a continuous map $s_i:U_i\to X^I$ such that $p_X\circ s_i = \mathrm{incl}_i:U_i\hookrightarrow X\times X$). Note that here we use the reduced version of ${\sf TC}(X)$, which is one less than the number of open sets in the cover.
Let $\pi$ be a discrete group. It is well-known that there exists a connected CW-complex $K(\pi,1)$ with
\[\pi_i(K(\pi,1))=\left\{\begin{array}{ll} \pi & (i=1) \\ 0 & (i\ge2). \end{array}\right.\]
Such a space is called an \emph{Eilenberg--Mac Lane space} for the group $\pi$. Furthermore, $K(\pi,1)$ is unique up to homotopy. This makes the following definition sensible.
\begin{defn}
The topological complexity of a discrete group $\pi$ is given by
\[{\sf TC}(\pi)\coloneqq{\sf TC}(K(\pi,1)).\]
\end{defn}
In the survey article \cite{Far06} Farber poses the problem of describing ${\sf TC}(\pi)$ solely in terms of algebraic properties of the group $\pi$. Very little is known about this problem in general. We remark that the corresponding question about $\mathsf{cat}(\pi):=\mathsf{cat}(K(\pi,1))$ has been completely answered: By work of Eilenberg--Ganea \cite{EG} and Stallings \cite{Sta} and Swan \cite{Swa}, we have $\mathsf{cat}(\pi)={\sf cd}(\pi)$, where ${\sf cd}$ denotes the cohomological dimension. In cases where the exact value of ${\sf TC}(\pi)$ is known, it often agrees with the standard cohomological lower bound in terms of the zero-divisors cup-length. However, one can also give potentially sharper lower bounds for ${\sf TC}(\pi)$ which take into account the subgroup structure of $\pi$.
\begin{thm}[{Grant--Lupton--Oprea \cite[Theorem 1.1]{GLO}}]\label{thm:lowerbound}
Let $\pi$ be a discrete group, and let $A$ and $B$ be subgroups of $\pi$. Suppose that $gAg^{-1} \cap B = \{1\}$
for every $g \in \pi$. Then \[{\sf TC}(\pi)\ge {\sf cd}(A \times B).\]
\end{thm}
This result has been applied in \cite{GLO} to calculate the topological complexity of Higman's group. In Section \ref{Lower bounds} of the current paper we will use it to give lower bounds for the topological complexity of mixed braid groups.
In fact, the upper bounds used in this paper are also group-theoretic in nature.
\begin{thm}[{Grant \cite[Proposition 3.7]{G}}]\label{thm:upperbound}
Let $\pi$ be a torsion-free discrete group, with centre $\mathcal{Z}(\pi)\le \pi$. Identify $\mathcal{Z}(\pi)$ with its image under the diagonal homomorphism $d:\pi\to \pi\times \pi$. Then \[{\sf TC}(\pi)\le {\sf cd}\left(\frac{\pi\times\pi}{\mathcal{Z}(\pi)}\right).\]
\end{thm}
This result was applied in \cite{G} to give upper bounds for the topological complexity of finitely generated torsion-free nilpotent groups. In Section \ref{Upper bounds} below we will apply it to various subgroups of the full braid group.
\section{Braid groups and their subgroups}\label{Braidgroups
\begin{defn}
A \emph{braid on $n$ strands} is an isotopy class of embeddings of the disjoint union of $n$ intervals in $\mathbb{R}^3$. The embeddings start at a set of $n$ distinct points in the plane $z=0$ and end at the corresponding set of points in the plane $z=1$, and are monotonically increasing in the direction of the $z$-axis. The isotopies are fixed on the boundary. Note that the start and end points of the individual strands are not required to correspond.
The \emph{full braid group} $B_n$ for $n\ge2$ is the group of braids with $n$ strands, where the group operation is given by concatenating braids.
\end{defn}
The group $B_n$ was first studied by Artin \cite{Art}, who showed (among other things) that it is generated by the braids $\sigma_i$ for $i=1,\ldots ,n-1$ which pass the $i$th strand over the $(i+1)$st strand. There is a canonical epimorphism
\[\pi\colon B_n\to\mathfrak{S}_n\]
which sends the generator $\sigma_i$ to the transposition $(i\;i+1)$ in the symmetric group $\mathfrak{S}_n$. Thus a braid $\gamma\in B_n$ gets sent to the associated permutation $\pi(\gamma)\in \mathfrak{S}_n$ of its endpoints.
The kernel $P_n\coloneqq\pi^{-1}(\{1\})$ of this projection is the \emph{pure braid group} and its elements are called \emph{pure braids}. More generally, given a subgroup $G\leq \mathfrak{S}_n$, we denote its pre-image by $B_n^G\coloneqq\pi^{-1}(G)$. Such a subgroup of $B_n$ may be called a \emph{$G$-braid group}, and its elements \emph{$G$-braids}. These are the groups whose topological complexity we are interested in computing.
\begin{ex}\label{ex:mixed}
Given an integer $1\le k\le n-1$, let $G=\mathfrak{S}_{n-k}\times \mathfrak{S}_k\leq \mathfrak{S}_n$ be the group of permutations which preserve the partition of $n$ objects into the first $n-k$ objects and the last $k$ objects. We denote the subgroup $B_n^{\mathfrak{S}_{n-k}\times \mathfrak{S}_k}$ by $B_{n-k,k}$, and refer to it as a \emph{mixed braid group}. One could imagine using two colours to distinguish the first $n-k$ braids from the last $k$ braids (hence groups of this form also are often called \emph{coloured braid groups}).
\end{ex}
In proving algebraic facts about braid groups and their subgroups, it is often easiest to argue topologically using configuration space models for their $K(\pi,1)$'s. Let
$$\mathbb{C}_m:=\mathbb{C}\setminus\{0,1,\ldots , m-1\}$$
denote the complex plane with $m$ punctures. Recall that the \emph{configuration space} of $n$ points on the plane with $m$ punctures is given by
\[F(\mathbb{C}_m,n)\coloneqq\{(x_1,\ldots,x_n)\in(\mathbb{C}_m)^n\,|\,x_i\ne x_j\text{ for }i\ne j\}.\]
\begin{lem}[Fadell--Neuwirth \cite{FN}]\label{lem:fadellneuwirth}
For each $m\ge0$ and $n\ge2$, projection onto the first coordinate gives a locally trivial fibration sequence
\begin{equation}\label{Fadell-Neuwirth}
F(\mathbb{C}_{m+1},n-1)\to F(\mathbb{C}_m,n)\to\mathbb{C}_m.
\end{equation}
Furthermore, this fibration admits a section.
\end{lem}
\begin{lem}\label{lem:keypione}
Any subgroup $G\leq \mathfrak{S}_n$ of the symmetric group acts freely on $F(\mathbb{C},n)$ by permuting the coordinates. The quotient space $F(\mathbb{C},n)/G$ under this action is an Eilenberg--Mac Lane space $K(B_n^G,1)$.
\end{lem}
\begin{proof}
It is easily seen from the definitions that $B_n^G$ is the fundamental group of $F(\mathbb{C},n)/G$.
It is also well known that $F(\mathbb{C},n)$ is aspherical. This follows from the Fadell-Neuwirth fibrations (\ref{Fadell-Neuwirth}) and induction. Explicitly, note that $F(\mathbb{C}_{n-1},1) = \mathbb{C}_{n-1}$ is homotopic to a wedge of circles and therefore aspherical, and that for $1\le k\le n-1$ we have that $F(\mathbb{C}_{n-k-1},k+1)$ fibres over an aspherical space $\mathbb{C}_{n-k-1}$ with fibre $F(\mathbb{C}_{n-k},k)$.
The spaces $F(\mathbb{C},n)/G$ are therefore aspherical, because they have $F(\mathbb{C},n)$ as a covering space.
\end{proof}
Recall that a duality group is a group whose (co)homology satisfies a generalization of Poincar\'e duality \cite{BE}. In more detail, we call a group $\pi$ a \emph{duality group of dimension $n$} if there exists a $\mathbb{Z}\pi$-module $C$ and an element $e\in H_n(\pi;C)$ such that the cap product homomorphism
\[e\cap-\colon H^k(\pi;A)\overset{\cong}{\to}H_{n-k}(\pi;C\otimes A)\]
is an isomorphism for all $k\in \mathbb{Z}$ and all $\mathbb{Z}\pi$-modules $A$. It follows that ${\sf cd}(\pi)=n$.
For instance, the fundamental group of a closed aspherical $n$-manifold is a duality group of dimension $n$.
\begin{thm}[Bieri--Eckmann \cite{BE}]\label{thm:duality}
\begin{enumerate}
\item If a torsion-free group $\pi$ has a finite index subgroup $S$ which is a duality group of dimension $n$, then $\pi$ is also a duality group of dimension $n$.
\item If $K$ and $Q$ are duality groups of dimensions $m$ and $n$ respectively which fit into an extension of groups
\[
\xymatrix{
1 \ar[r] & K \ar[r] & \pi \ar[r] & Q\ar[r] & 1,
}
\]
then $\pi$ is a duality group of dimension $m+n$.
\item Non-trivial free groups are duality groups of dimension $1$.
\end{enumerate}
\end{thm}
The following lemma tells us that the groups we are interested in are indeed duality groups.
\begin{lem}\label{lem:dualitygroups}
For every $G\le\mathfrak{S}_n$, the group $B_n^G$ is a duality group of dimension $n-1$.
Furthermore, $\tilde P_{n-2}\coloneqq\pi_1(F(\mathbb{C}_2,n-2))$ is a duality group of dimension $n-2$.
\end{lem}
\begin{proof}
The fact that $P_n=\pi_1 (F(\mathbb{C},n))$ is a duality group of dimension $n-1$ follows from the same chain of Fadell--Neuwirth fibrations used in the proof of \lemref{lem:keypione}. Indeed, by Lemma \ref{lem:fadellneuwirth} we have for each $1\le k\le n-1$ a split short exact sequence of fundamental groups
\[\xymatrix{
1 \ar[r] & \pi_1 (F(\mathbb{C}_{n-k},k)) \ar[r] & \pi_1 (F(\mathbb{C}_{n-k-1},k+1)) \ar[r] & \pi_1 (\mathbb{C}_{n-k-1}) \ar[r] & 1.}
\]
Our inductive hypothesis is that $\pi_1 (F(\mathbb{C}_{n-k},k))$ is a duality group of dimension $k$. The induction step follows from Theorem \ref{thm:duality} (2), with the base case $k=1$ given by Theorem \ref{thm:duality} (3).
Observe that in the course of the induction (two steps before the end) we have shown that $\tilde P_{n-2}\coloneqq\pi_1(F(\mathbb{C}_2,n-2))$ is a duality group of dimension $n-2$.
Finally, let $G\le\mathfrak{S}_n$. The group $B_n^G$ is torsion-free because it has a finite-dimensional $K(\pi,1)$. Now the claim follows from \thmref{thm:duality} (1) because $B_n^G$ contains the duality group $P_n$ as a finite index subgroup.
\end{proof}
In order to apply Theorem \ref{thm:upperbound} to the groups $B_n^G$, we need to identify their centres. Note that $P_n\leq B_n^G\leq B_n$. It is well known that $\mathcal{Z}(P_n)$ and $\mathcal{Z}(B_n)$ are infinite cyclic, and are equal for $n\geq 3$.
\begin{lem}\label{lem:centre}
For any subgroup $G\leq \mathfrak{S}_n$, the centre $\mathcal{Z}(B_n^G)$ is infinite cyclic. Furthermore, when $n\ge3$ we have
\[
\mathcal{Z}(P_n)=\mathcal{Z}(B_n^G)=\mathcal{Z}(B_n),
\]
and all of these groups are generated by the full twist
\[\Delta^2=((\sigma_1\ldots\sigma_{n-1})(\sigma_1\ldots\sigma_{n-2})\ldots(\sigma_1\sigma_2)\sigma_1)^2=(\sigma_1\ldots\sigma_{n-1})^n.\]
\end{lem}
\begin{proof}
The case $n=2$ is clear since all the groups will be isomorphic to $\mathbb{Z}$. For the rest of the proof assume $n\ge3$.
It is a standard result that the full twist $\Delta^2$ generates $\mathcal{Z}(P_n)=\mathcal{Z}(B_n)$ (for a proof see for instance \cite[Theorem 1.24]{KT}). Since $\Delta^2$ commutes with everything (and in particular with braids in $B_n^G$) we therefore have $\mathcal{Z}(P_n)\leq \mathcal{Z}(B_n^G)$. It just remains to be shown that $\mathcal{Z}(B_n^G)\leq P_n$, since then it follows easily that $\mathcal{Z}(B_n^G)\leq \mathcal{Z}(P_n)$. In fact, a seemingly stronger statement is true: If a braid $\gamma\in B_n$ commutes with every pure braid, then $\gamma$ is itself a pure braid.
To see this note that it is possible to assign to each pure braid an \emph{ordered} link by closing off the braid, that is, by connecting together both ends of each strand in the braid. This is illustrated in the following picture.
\begin{center}
\begin{tikzpicture}[scale=0.5]
\braid[number of strands=3,thick,style strands={1}{red},style strands={2}{green},style strands={3}{blue}] a_1a_2^{-1}a_2^{-1}a_1a_2a_2;
\braid[number of strands=3,thick,style strands={1}{red},style strands={2}{green},style strands={3}{blue},xshift=-8cm] a_1a_2^{-1}a_2^{-1}a_1a_2a_2;
\draw[thick,red,rounded corners](1,0)--(1,0.5)--(0,0.5)--(0,-7)--(1,-7)--(1,-6.5);
\draw[thick,green,rounded corners](2,0)--(2,1)--(-0.5,1)--(-0.5,-7.5)--(2,-7.5)--(2,-6.5);
\draw[thick,blue,rounded corners](3,0)--(3,1.5)--(-1,1.5)--(-1,-8)--(3,-8)--(3,-6.5);
\end{tikzpicture}
\captionof{figure}{A pure braid and its closure.}
\end{center}
Note that conjugating a pure braid by a braid $\gamma\in B_n$ before closing it off will result in the same link, but with the ordering of its components permuted by $\pi(\gamma)$. Suppose that $\gamma\in B_n$ commutes with every braid in $P_n$. Then conjugation by $\gamma$ is the identity on $P_n$, and the permutation $\pi(\gamma)$ acts trivially on the set of isotopy classes of ordered links of $n$ components. It follows that $\pi(\gamma)=1$, so $\gamma$ is a pure braid.
\end{proof}
\section{Upper bounds}\label{Upper bounds
Using the algebraic results of the previous section, we can now establish our main upper bound.
\begin{thm}\label{thm:torsionbound}
Suppose that $G\le \mathfrak{S}_n$ is such that $B_n^G/\mathcal{Z}(B_n^G)$ is torsion-free. Then
\[{\sf TC}(B_n^G)\le 2n-3.\]
\end{thm}
\begin{proof}
Consider the group extension
\begin{equation*}
\xymatrix{ 1 \ar[r] & \mathcal{Z}(B_n^G) \ar[r]^-{d} & B_n^G\times B_n^G \ar[r] & (B_n^G\times B_n^G)/\mathcal{Z}(B_n^G) \ar[r] & 1,}
\end{equation*}
where $d$ denotes the diagonal embedding. By Theorem \ref{thm:upperbound}, it will suffice to prove that the quotient group $(B_n^G\times B_n^G)/\mathcal{Z}(B_n^G)$ has cohomological dimension $2n-3$. It follows from the assumptions that this quotient is torsion-free, and by Lemma \ref{lem:centre} it contains $(P_n\times P_n)/\mathcal{Z}(P_n)$ as a subgroup of finite index. Thus we are reduced (by Theorem \ref{thm:duality}(1)) to showing that $(P_n\times P_n)/\mathcal{Z}(P_n)$ is a duality group of dimension $2n-3$.
Using the operations of addition and multiplication in $\mathbb{C}$, it is easily shown that the first two Fadell--Neuwirth fibrations used in the proofs of Lemmas \ref{lem:keypione} and \ref{lem:dualitygroups} are in fact trivial, so that we have a splitting up to homeomorphism
\[F(\mathbb{C},n)\approx F(\mathbb{C}_2,n-2)\times\mathbb{C}_1\times\mathbb{C}.\]
On fundamental groups this gives a splitting $P_n\cong\tilde{P}_{n-2}\times\mathbb{Z}$, under which the centre $\mathcal{Z}(P_n)$ corresponds to $\{1\}\times\mathbb{Z}$. We therefore find that
\[
(P_n\times P_n)/\mathcal{Z}(P_n) \cong (\tilde{P}_{n-2}\times\mathbb{Z}\times \tilde{P}_{n-2}\times\mathbb{Z})/\mathcal{Z}(P_n) \cong \tilde{P}_{n-2}\times \tilde{P}_{n-2}\times \mathbb{Z}.
\]
Since the group $\tilde{P}_{n-2}$ was shown in \lemref{lem:dualitygroups} to be a duality group of dimension $n-2$, the result now follows from Theorem \ref{thm:duality}(2).
\end{proof}
We now give examples of subgroups $G\leq \mathfrak{S}_n$ for which the assumption of the theorem holds. We continue to denote the mixed braid group $B_n^{\mathfrak{S}_{n-k}\times \mathfrak{S}_k}$ from Example \ref{ex:mixed} by $B_{n-k,k}$.
\begin{prop}\label{prop:torsionfree}
The mixed braid group modulo its centre $B_{n-k,k}/\mathcal{Z}(B_{n-k,k})$ is torsion-free if and only if $(n,k)=(n-1,k)=(n-1,k-1)=1$.
\end{prop}
\begin{proof}
The centre $\mathcal{Z}(B_{n-k,k})$ is infinite cyclic generated by the full twist $\Delta^2=(\sigma_1\ldots\sigma_{n-1})^n$, by \lemref{lem:centre}. Thus torsion elements in $B_{n-k,k}/\mathcal{Z}(B_{n-k,k})$ will be represented by braids $\alpha\in B_{n-k,k}$ such that $\alpha^m=\Delta^l$, for some powers $m,l\in\mathbb{Z}$, but $\alpha$ is not a power of $\Delta^2$ itself.
Viewing the braid group $B_n$ as the fundamental group of the configuration space $F(\mathbb{C},n)/\mathfrak{S}_n$, the full twist corresponds to a rotation of the point configuration by $2\pi$. Section~2.2 from \cite{GM} tells us that all torsion elements are represented by rotations up to conjugation. To be precise, there exist $m,l\in\mathbb{Z}$ such that $\alpha^m=\Delta^l$ if and only if there exist $\beta\in B_n$ and $p\in\mathbb{Z}$ such that either $\alpha=\beta\delta^p\beta^{-1}$ or $\alpha=\beta\epsilon^p\beta^{-1}$. Here the elements $\delta$ and $\epsilon$ can be expressed in terms of the standard generators as
\[\delta=\sigma_1\ldots\sigma_{n-1}\text{ and }\epsilon=\sigma_1(\sigma_1\ldots\sigma_{n-1}),\] and correspond to minimal rotations in $F(\mathbb{C},n)/\mathfrak{S}_n$, where in the case of $\epsilon$ one of the points in the configuration is a fixed point of the rotation.
Recall that $\pi$ denotes the canonical projection $\pi\colon B_n\to\mathfrak{S}_n$.
Observe that $\pi(\delta)$ is an $n$-cycle and $\pi(\epsilon)$ is an $(n-1)$-cycle times a 1-cycle. Therefore $\pi(\delta)^p$ is a product of $d$ $n/d$-cycles, where $d=(p,n)$ is the greatest common divisor of $p$ and $n$, which could be 1. Similarly $\pi(\epsilon)^q$ is a product of $d$ $(n-1)/d$-cycles and a 1-cycle, where $d=(q,n-1)$.
Recall that elements in the symmetric group are conjugate if and only if they have the same cycle structure. Therefore, $\pi(\delta)^p$ is conjugate to an element in $\mathfrak{S}_{n-k}\times\mathfrak{S}_k$ if and only if $n/(p,n)$ divides $k$. Similarly, $\pi(\epsilon)^q$ is conjugate to an element in $\mathfrak{S}_{n-k}\times\mathfrak{S}_k$ if and only if $(n-1)/(q,n-1)$ divides $k$ or $k-1$.
One implication in the theorem now follows from combining the above observations. If $B_{n-k,k}/\mathcal{Z}(B_{n-k,k})$ had torsion there would exist an element $\alpha\in B_{n-k,k}$ which is conjugate to $\delta^p$ or $\epsilon^q$. Therefore, there would exist an element $\pi(\alpha)\in\mathfrak{S}_{n-k}\times\mathfrak{S}_k$ conjugate to $\pi(\delta)^p$ or $\pi(\epsilon)^q$. This implies that either $(n,k)\ne1$ or $(n-1,k)\ne1$ or $(n-1,k-1)\ne1$.
The converse implication follows from a similar argument. Assume that for instance $(n,k)\ne1$. Then there exists an element $\alpha\in\mathfrak{S}_{n-k}\times\mathfrak{S}_k$ conjugate to $\pi(\delta)^p$:
\[\alpha=\beta\pi(\delta)^p\beta^{-1}.\]
Taking a lift $\sigma_\beta$ of $\beta$ yields an element $\sigma_\beta\delta^p\sigma_\beta^{-1}$ of $B_{n-k,k}$.
\end{proof}
\begin{cor}\label{cor:torsion1}
If $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ with $(n,k)=(n-1,k)=(n-1,k-1)=1$, then the group $B_n^G/\mathcal{Z}(B_n^G)$ is torsion-free.
\end{cor}
\begin{proof}
If $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$, then $B_n^G$ is a subgroup of $B_{n-k,k}$ but by \lemref{lem:centre}, the centre of both groups is the same. Therefore, if $B_{n-k,k}/\mathcal{Z}(B_{n-k,k})$ is torsion-free then so is $B_n^G/\mathcal{Z}(B_n^G)$.
\end{proof}
\begin{prop}\label{prop:torsion2}
If $G=H\times\{1\}^2\le\mathfrak{S}_n$ where $H\le \mathfrak{S}_{n-2}$, then the group $B_n^G/\mathcal{Z}(B_n^G)$ is torsion-free.
\end{prop}
\begin{proof}
In the proof of \thmref{thm:torsionbound} it was observed that there is a homeomorphism $F(\mathbb{C},n)\approx F(\mathbb{C}_2,n-2)\times\mathbb{C}_1\times\mathbb{C}$. This splitting is compatible with the action of $G=H\times\{1\}^2$ (as is clear from writing out explicit formulae, or see Lemma 3.1 and Theorem 3.3 of \cite{CGJ} for a proof). Therefore we have a homeomorphism $$F(\mathbb{C},n)/G\approx (F(\mathbb{C}_2,n-2)/H)\times \mathbb{C}_1\times \mathbb{C},$$ which on fundamental groups gives the splitting $B_n^G\cong \tilde{B}_{n-2}^H\times \mathbb{Z}$, where $\tilde{B}_{n-2}^H\coloneqq\pi_1(F(\mathbb{C}_2,n-2)/H)$. Under this splitting the centre $\mathcal{Z}(B_n^G)=\mathbb{Z}$ corresponds to $\{1\}\times\mathbb{Z}$, and therefore we have $B_n^G/\mathcal{Z}(B_n^G)\cong \tilde{B}_{n-2}^H$.
Finally we observe that $\tilde{B}_{n-2}^H$ is torsion-free because it has the finite-dimensional space $F(\mathbb{C}_2,n-2)/H$ as its $K(\pi,1)$.
\end{proof}
Putting together \corref{cor:torsion1}, \propref{prop:torsion2} and \thmref{thm:torsionbound} yields:
\begin{thm}\label{thm:upperresult}
Let $G\le \mathfrak{S}_{n-2}\times\{1\}^2\le \mathfrak{S}_n$ or $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ with $(n,k)=(n-1,k)=(n-1,k-1)=1$. Then \[{\sf TC}(B_n^G)\le 2n-3.\]
\end{thm}
\begin{rem}
In the case $G\leq \mathfrak{S}_{n-2}\times\{1\}^2$, the upper bound ${\sf TC}(B_n^G)\le 2n-3$ can be deduced from the fact that there is a splitting
\[
B_n^G \cong \left(B_n^G/\mathcal{Z}(B_n^G)\right)\times \mathcal{Z}(B_n^G),
\]
together with the product formula for topological complexity \cite[Theorem 11]{Far03}. We do not know if the upper bound ${\sf TC}(B_n^G)\le 2n-3$ for $G\leq\mathfrak{S}_{n-k}\times \mathfrak{S}_k$ can be obtained in this way. For instance, we do not know if the centre of the mixed braid group $B_{5,3}$ splits off as a direct factor.
\end{rem}
\section{Lower bounds}\label{Lower bounds
We now give lower bounds for ${\sf TC}(B_n^G)$ in certain cases, using Theorem \ref{thm:lowerbound}. Our arguments generalise the argument given in \cite[Proposition 3.3]{GLO} for pure braids.
\begin{thm}\label{lower1}
Let $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ for $k\ge2$. Then \[{\sf TC}(B_n^G)\ge 2n-k-1.\]
Furthermore, if $G\le \mathfrak{S}_{n-1}\times \mathfrak{S}_1$, then \[{\sf TC}(B_n^G)\ge 2n-3.\]
\end{thm}
\begin{proof}
Suppose $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ for $k\ge2$. To apply \thmref{thm:lowerbound}, we need to give two subgroups of $B_n^G$ which contain no non-trivial conjugates. The following subgroups satisfy this assumption:
\begin{enumerate}
\item Think of $P_{n-k+1}$ for $k\ge 2$ as a subgroup of $P_n$ by adding $k-1$ strands which do not interact with any other strand.
\item Denote by $A_n$ the subgroup of $P_n$ generated by braids of the form \[\alpha_j=\sigma_{j}\sigma_{j+1}\ldots \sigma_{n-1}\sigma_n\sigma_n\sigma_{n-1}\ldots \sigma_{j+1}\sigma_{j},\]
for $1\le j\le n-1$. In words, the generator $\alpha_j$ takes the $j$th strand over the last $n-j$ strands and back under them until it gets back to its original position, see the braid diagrams below. Looking at the braid diagrams it becomes evident that the generators commute, and therefore $A_n\cong\mathbb{Z}^{n-1}$.
\end{enumerate}
Since $P_n\leq B_n^G$, these are indeed both subgroups of $B_n^G$. It remains to show that $P_{n-k+1}\cap \gamma A_n\gamma^{-1}=\{1\}$ for all $\gamma\in B_n^G$.
Recall that closing off a pure braid yields an ordered link, and that conjugating a pure braid in $P_n$ by an element of $B_n^G$ before closing it off will result in an isotopic link up to permutation of its components by an element of $G$. In light of the above, it suffices to show that the links coming from closing off nontrivial elements of $P_{n-k+1}$ can not be obtained by permuting with an element in $G$ the components of the link coming from closing off an element of $A_n$.
For this we note that if we close a non-trivial braid in $A_n$, the last two components of the link will be linked either with some other component or with each other. This can be seen for the generators $\alpha_i$ by taking a look at the following braid diagrams.
\vspace{.3cm}
\begin{minipage}{.48\linewidth}
\begin{center}
\begin{tikzpicture}[xscale=0.7,yscale=1.25]
\braid[number of strands=4,thick,style strands={3}{red}] 1 a_3a_3 1;
\end{tikzpicture}
\captionof{figure}{The braid $\alpha_{3}\in A_4$. }
\end{center}
\end{minipage}%
\begin{minipage}{.48\linewidth}
\begin{center}
\begin{tikzpicture}[scale=0.7]
\braid[number of strands=4,thick,style strands={2}{red}] a_2a_3a_3a_2;
\end{tikzpicture}
\captionof{figure}{The braid $\alpha_{2}\in A_4$. }
\end{center}
\end{minipage}
\vspace{.3cm}
Because the generators $\alpha_i$ of $A_n$ commute with each other, every element $\sigma \in A_n$ can be written uniquely in the form
\[\sigma=\alpha_{1}^{m_{1}}\alpha_{2}^{m_{2}}\cdots\alpha_{n-1}^{m_{n-1}}\qquad(m_i\in \mathbb{Z}).\]
The exponent $m_i$ determines the linking number of the $i$th strand with all subsequent strands, from the $(i+1)$st up to the $n$th. Therefore the last two components of the link which results from closing off $\sigma$ are unlinked from the first $n-2$ and from each other if and only if all the $m_i=0$, i.e. if and only if $\sigma$ is trivial. In particular, closing off a non-trivial element of $A_n$ will yield a link where at least two of the last $k$ components are linked with something else.
On the other hand, closing braids in $P_{n-k+1}\subset P_{n}$ yields links in which at most one of the last $k$ components is linked with another component. Therefore, a link coming from a non-trivial element in $A_n$ cannot be obtained from a link coming from a braid in $P_{n-k+1}$ by permuting the components by an element of $G$. This is because, by assumption, an element of $G$ does not permute any of the last $k$ components with any of the first $n-k$.
Finally, \thmref{thm:lowerbound} yields
\[{\sf TC}(B_n^G)\ge {\sf cd}(P_{n-k+1})+{\sf cd}(A_n)=(n-k)+(n-1)=2n-k-1.\]
The proof for $G\le \mathfrak{S}_{n-1}\times \mathfrak{S}_1$ is analogous but has to be stated separately. Here we may take the subgroups $P_{n-1}$ and $A_n$, and just use the linking number with the last strand.
\end{proof}
Note that the lower bound goes down by one at each step for $k\ge2$ but it is the same for $B_{n-1,1}$ and $B_{n-2,2}$.
\begin{cor}\label{cor:lower2}
Let $G\le \mathfrak{S}_{n-2}\times\{1\}^2\le \mathfrak{S}_n$. Then \[{\sf TC}(B_n^G)\ge 2n-3.\]
\end{cor}
Note that in \corref{cor:lower2} the lower bound coincides with the upper bound from \thmref{thm:upperresult} and thus gives a complete answer in that case.
\begin{rem}\label{rem:zerodivisors}
The optimal lower bound ${\sf TC}(P_n)\geq 2n-3$ can be obtained fairly easily using zero-divisors cup-length \cite{FY}, because the rational cohomology algebra $H^*(P_n;\mathbb{Q})$ is well understood.
One could ask whether the lower bounds for ${\sf TC}(B_n^G)$ given above can be obtained similarly using rational cohomology. An application of the Cartan--Leray spectral sequence would give an algebra isomorphism
\[
H^*(B_n^G;\mathbb{Q})\cong H^*(P_n;\mathbb{Q})^G.
\]
In the cases we are interested in, however, the calculation of this ring of invariants appears not so straightforward, and does not seem to have been carried out in the literature. In the low dimensional cases in which we were able to compute the zero-divisors cup-length, it was less effective than our lower bound. One could try to replace $\mathbb{Q}$ with some other coefficient ring, but then calculations become even more involved.
For the full braid group $B_n$, both the zero-divisors cup-length and the methods given in this paper appear to be insufficient.
\end{rem}
\section{Higher topological complexity}\label{highertc}
The concept of higher topological complexity was introduced by Rudyak in \cite{Rud}. See also the subsequent paper \cite{BGRT} by Basabe, Gonz\'alez, Rudyak and Tamaki.
Recall that in Section 2 the topological complexity of a space $X$ was defined using the free path fibration $p_X:X^I\to X\times X$, with projection $p_X(\gamma) = (\gamma(0),\gamma(1))$. For any natural number $m\ge2$, the \emph{$m$th topological complexity} of $X$, denoted ${\sf TC}_m(X)$, can similarly be defined with help of the fibration $$p^m_X:X^I\to X^m$$ with projection $$p^m_X(\gamma) = (\gamma(0),\gamma(1/(m-1)),\ldots,\gamma((m-2)/(m-1)),\gamma(1)).$$Concretely, it is the minimal $k$ such that $X^m$ admits a cover by $k+1$ open sets $U_0,U_1,\ldots , U_k$, on each of which there exists a local section of the fibration $p^m_X$. Of course, the case $m=2$ corresponds to the original topological complexity.
The results for the topological complexity of mixed braid groups in the last two sections rely on \cite[Theorem 1.1]{GLO} and \cite[Proposition 3.7]{G} (denoted \thmref{thm:lowerbound} and \thmref{thm:upperbound} in this paper). Those theorems can be generalised in a straightforward way to the following statements about higher topological complexity.
\begin{thm}\label{thm:higherlowerbound}
Let $\pi$ be a discrete group, and let $A$ and $B$ be subgroups of $\pi$. Suppose that $gAg^{-1} \cap B = \{1\}$
for every $g \in \pi$. Then \[{\sf TC}_m(\pi)\ge {\sf cd}(A \times B \times \pi^{m-2}).\]
\end{thm}
\begin{thm}\label{thm:higherupperbound}
Let $\pi$ be a torsion-free discrete group, with centre $\mathcal{Z}(\pi)\le \pi$. Identify $\mathcal{Z}(\pi)$ with its image under the diagonal homomorphism $d_m:\pi\to \pi^m$. Then \[{\sf TC}_m(\pi)\le {\sf cd}\left(\frac{\pi^m}{\mathcal{Z}(\pi)}\right).\]
\end{thm}
Using this, the proofs in the earlier sections generalise to yield the following results.
\begin{thm}\label{thm:higherupperresult}
Suppose that $G\leq \mathfrak{S}_n$ satisfies either of the following conditions:
\begin{itemize}
\item $G\leq\mathfrak{S}_{n-k}\times \mathfrak{S}_k$ where $(n,k)=(n-1,k)=(n-1,k-1)=1$, or
\item $G\leq \mathfrak{S}_{n-2}\times \{1\}^2$.
\end{itemize}
Then we have
\[
{\sf TC}_m(B_n^G)\leq m(n-1)-1.
\]
\end{thm}
\begin{thm}\label{thm:higherlowerresult}
Let $G\le \mathfrak{S}_{n-k}\times \mathfrak{S}_k$ for $k\ge2$. Then \[{\sf TC}_m(B_n^G)\ge m(n-1)-k+1.\]
Furthermore, if $G\le \mathfrak{S}_{n-1}\times \{1\}$, then \[{\sf TC}(B_n^G)\ge m(n-1)-1.\]
\end{thm}
As before, in some cases the upper and lower bounds coincide to give an equality.
\begin{cor}
If $G\leq \mathfrak{S}_{n-2}\times \{1\}^2$, then
\[
{\sf TC}(B_n^G)=m(n-1)-1.
\]
\end{cor}
| 70f39af37a7df86a539c4cfb31b1e6d00e856e70 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Most studies of spin glasses use a model of the
Edwards-Anderson~\cite{edwards:75} type in which the interactions are
short-range and have random sign. However, it is argued that spin glass
behavior is more general and that the necessary ingredients are simply
\textit{randomness} and \textit{frustration}. Indeed an antiferromagnet on a
random graph (which has all interactions negative) is
found~\cite{krzakala:08a} to have spin glass behavior, in which
disorder and
frustration arise from large loops in the graph. In this paper we also
study spin glass behavior in a disordered model with only anti-ferromagnetic
interactions, but of the long-range Coulomb-type.
In addition to clarifying the
general conditions under which spin glass behavior can occur, further motivation for
our work comes from experiment. In certain highly frustrated random magnets,
Schiffer and Daruka~\cite{schiffer:97} showed that new magnetic degrees of
freedom emerge, so-called ``orphan'' spins, which could potentially
undergo a glassy transition. Subsequently, two of us and
collaborators~\cite{sen:12,rehn:15} showed that the orphan spins have Coulomb
interactions between them, due to entropic effects, and performed
numerical simulations on the resulting model. A final motivation for our
study is that
Villain~\cite{villain:77b} showed that antiferromagnetic Coulomb
interactions arise between effective Ising spins, called ``chiralities'', in an
XY (i.e.~2-component) spin model with frustration and speculated that this
could lead to a glassy transition.
The main question we address in this work is whether the
spin glass transition in the random Coulomb
antiferromagnet is in the same universality class as the Edwards-Anderson (EA)
spin glass. Microscopically they are very different. The EA model is
short-range and its interactions have random sign while the Coulomb antiferromagnet
has all interactions negative and is long-range. However,
Ref.~\cite{rehn:15} showed that there is a screening mechanism in the
random Coulomb antiferromagnet so we might expect that the interactions driving
a spin glass transition are short-range. In addition,
both the EA model and the random Coulomb antiferromagnet have
disorder and frustration, so one might imagine that
the universal behavior at
a spin glass transition could be the same. We will try to see if this is the
case
by numerical simulations. Our conclusion is that the data is
\textit{consistent} with this hypothesis, but other scenarios can not be ruled
out because there are strong corrections to finite-size scaling for the
range of sizes that we can study.
Reference \cite{rehn:15} performed Monte Carlo simulations on a Heisenberg
(i.e.~three-component) version of the random Coulomb antiferromagnet because
the orphan spins emerging in experimental frustrated quantum
magnets~\cite{schiffer:97,sen:12} are of the Heisenberg type. However, in
order to try to answer questions about the spin glass
universality class we prefer to
study the Ising (i.e.~one-component) version of the model. One reason is that
the updating algorithm is simpler and more efficient than for the Heisenberg case.
More important is that even for the EA model, the nature of the spin glass
transition in, say, three dimensions has been harder to elucidate for the
Heisenberg case than for the Ising case. This is partly because the transition
temperature is much lower and partly because there seem to be larger
corrections to finite-size scaling as well as complications due to additional
(chiral) degrees of freedom, see for example Refs.~\cite{banos:12,viet:09a}.
By contrast, the transition in the three-dimensional Ising EA spin glass is
much better understood, see Refs.~\cite{hasenbusch:08b,baity-jesietal:13}. By
using Ising rather than Heisenberg spins, and by some refinements to the Monte
Carlo method, we are able to study significantly larger sizes than in
Ref.~\cite{rehn:15}.
The plan of this paper is as follows. In Sec.~\ref{sec:model} we describe the
model and the numerical method used to simulate it. In
Sec.~\ref{sec:quantities} we explain
the finite-size scaling method used to investigate the transition, while in
Sec.~\ref{sec:results} we describe the results and interpret them for the cases of
dimension $d$ equal to 2 and 3. Our conclusions are summarized in
Sec.~\ref{sec:conclusions}.
\section{The Model}
\label{sec:model}
We study $N$ Ising spins, $S_i = \pm 1$, randomly placed on a $d-$dimensional
hypercubic lattice of size $L$ for $d=2$ and $3$. The concentration of spins is
therefore $x = N / L^d$.
The Hamiltonian is given by
\begin{equation}
\mathcal{H} = -\sum_{\langle i, j \rangle} J(\mathbf{r}_{ij}) S_i S_j \, ,
\label{Ham}
\end{equation}
where the interactions $J(\mathbf{r}_{ij})$ are given by the lattice Green function
\begin{equation}
J(\mathbf{r}_{ij}) = -{a_d \over L^d} \sum_{\mathbf{k}}
{\cos\left(\mathbf{k}\cdot\mathbf{r}_{ij}\right) \over d - \sum_{\ell = 1}^d \cos
k_\ell} \, .
\label{Jij}
\end{equation}
The factor $a_d$ is introduced so that the large-distance limit has the
Coulomb form
\begin{subequations}
\begin{align}
J(\mathbf{r}_{ij}) &= \log (r_{ij} / \mathcal{L}), \qquad(d=2),\\
&= -1/r_{ij}, \qquad\qquad\ \ \ \ (d=3),
\label{large_r}
\end{align}
\label{coulomb}
\end{subequations}
where $\mathcal{L}$ is a constant which can be chosen to be larger than any
$r_{ij}$ so the interactions are all antiferromagnetic. In fact, since we impose
the ``charge neutrality'' condition,
\begin{equation}
\sum_{i=1}^N S_i = 0,
\label{neut}
\end{equation}
the Hamiltonian is actually independent of $\mathcal{L}$. The numerical values
of $a_d$ are
\begin{subequations}
\begin{align}
a_2 &= \pi, \\
a_3 &= 2 \pi.
\end{align}
\end{subequations}
Note that, since the \textit{positions} of the spins are random, the
interactions will be different for different samples, but always
antiferromagnetic. To reduce error bars coming from sample-to-sample
fluctuations we need to average over many, typically several hundred, samples.
We simulate this model using the Metropolis Monte Carlo method, modified as
follows to incorporate the charge neutrality
condition in Eq.~\eqref{neut}. A site $i$ is
chosen, either sequentially or at random, and then one of the $z$ nearest
sites to this, $j$ say, is chosen at random. If the spins on $i$ and $j$
are antiparallel they
are both flipped with the usual Metropolis probability, and otherwise no change is
made. Repeating this procedure $N$ times corresponds to one Monte Carlo sweep. An
earlier version of the code took both spins to be random, so with high
probability they are far away, but this leads to an acceptance probability that
decreases rapidly with increasing system size.
We incorporate parallel
tempering (replica exchange)~\cite{hukushima:96,marinari:98b} to speed up
equilibration at low temperatures. In this approach, simulations are done at
several temperatures for the same set of interactions and
global moves are performed in which
entire spin configurations at neighboring temperatures are exchanged, with a
probability satisfying the detailed balance condition. We determine the
temperatures empirically by requiring that the acceptance ratios for global
moves are reasonable, typically of order $0.3$. Appropriate
temperatures can be estimated with
sufficient accuracy from relatively short test runs.
\section{Quantities Calculated and Finite-Size Scaling}
\label{sec:quantities}
The main object of interest is the spin glass order parameter $q$ defined as
the instantaneous overlap between the
spin configurations in two copies of the system with the same interactions,
\begin{equation}
q = {1 \over N} \sum_{i=1}^N S_i^{(1)} S_i^{(2)} \, ,
\end{equation}
where $``(1)"$ and $``(2)"$ refer to the copies. From measurements of $q$ we
determine the spin glass susceptibility
\begin{equation}
\chi_{SG} = N [\langle q^2 \rangle]_\text{av}
\end{equation}
and the Binder ratio
\begin{equation}
g = {1 \over 2}\, \left(3 - {[\langle q^4 \rangle]_\text{av}
\over [\langle q^2 \rangle]_\text{av}^2} \right)
\, ,
\end{equation}
in which angular brackets refer to a Monte Carlo average for
a single sample and square brackets $[\cdots]_\text{av}$ refer
to an average over samples. It is also useful to define a
wavevector-dependent spin glass susceptibility by
\begin{equation}
\chi_{SG}(\mathbf{k}) = {1 \over N}\sum_{i, j} \left[\left\langle S_i^{(1)} S_j^{(1)}
S_i^{(2)} S_j^{(2)} \right\rangle\right]_\text{av}
\cos\left(\mathbf{k}\cdot \mathbf{r}_{ij}\right) \, ,
\end{equation}
from which one can determine a correlation length $\xi_L$ according to
\begin{equation}
\xi_L = {1 \over 2 \sin(q_\text{min})} \, \left({\chi_{SG}(0) \over
\chi_{SG}(\textbf{q}_\text{min})} - 1\right)^{1/2} \, ,
\end{equation}
where $\textbf{q}_\text{min}$ is the smallest non-zero wavevector,
i.e.~$\textbf{q}_\text{min} = (2\pi/L)(1, 0)$ in $d=2$ and $\textbf{q}_\text{min} =
(2\pi/L)(1, 0, 0)$ in $d=3$.
To investigate whether or not there is a spin glass phase transition it is
essential to use finite-size scaling (FSS), see for example Refs.~\cite{privman:90,banos:12}.
If there is a transition at $T=
T_c$, in which the bulk (i.e.~infinite system-size) correlation length
diverges with an exponent $\nu$, i.e.
\begin{equation}
\xi_\infty \propto (T - T_c)^{-\nu},
\end{equation}
then the Binder ratio, being dimensionless, will have the FSS form
\begin{equation}
g = \widetilde{g}\left( L^{1/\nu} (T - T_c) \right)\, ,
\label{gscale}
\end{equation}
so curves of $g$ against $T$ for different sizes \textit{intersect} at $T_c$.
The spin glass susceptibility, however, is not dimensionless and for an
infinite system size diverges at $T_c$ with an exponent $\gamma$, i.e.
\begin{equation}
\chi_{SG} \propto (T - T_c)^{-\gamma}, \qquad (L \to \infty),
\end{equation}
so its FSS form is
\begin{equation}
\chi_{SG} = L^{2 - \eta} \, \widetilde{\chi}\left( L^{1/\nu} (T - T_c) \right)\, ,
\label{chisgscale}
\end{equation}
where $\eta$ is related to the other exponents by
\begin{equation}
\gamma = (2 -\eta) \nu \, .
\label{gamma}
\end{equation}
The correlation length $\xi_L$ divided by the system size is also
dimensionless and so has the FSS form
\begin{equation}
{\xi_L \over L} = \widetilde{X}\left( L^{1/\nu} (T - T_c) \right)\, .
\label{xiscale}
\end{equation}
\section{Results}
\label{sec:results}
\subsection{Screening}
\label{sec:screen}
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{chisg_smallbeta.eps}
\caption{
Data for the spin glass susceptibility at small $\beta \, (\equiv 1/T)$ in two
dimensions for a concentration $x = 1/16$ for $N = 16$ and $64$. The curve is
a fit to the data for $N = 64$.
The behavior is clearly linear, not
quadratic, at
small $\beta$, which is a result of screening. The intercept at $\beta = 0$
differs measurably from one for very small sizes only because of the charge
neutrality constraint, Eq.~\eqref{neut}.
\label{fig:screen}
}
\end{center}
\end{figure}
The spin glass susceptibility can be expressed as
\begin{align}
\chi_{SG} &= {1 \over N} \sum_{i,j} [C_{ij}^2]_\text{av} \nonumber \\
&= 1 + {1 \over N} \sum_{i \ne j} [C_{ij}^2]_\text{av}
\label{chi_Cij}
\end{align}
where $C_{ij} = \langle S_i S_j \rangle$ and the terms with $i = j$ give unity,
which is the result for $T = \infty$. Naively, one can obtain the second term
in Eq.~\eqref{chi_Cij} at high temperature by expanding the Boltzmann factors
in powers of $\beta \, (\equiv 1/T)$,
with the result
\begin{equation}
C_{ij} = \beta J_{ij} + O(\beta^2) \qquad (i \ne j) ,
\label{Cij}
\end{equation}
but then the sum in the second term in Eq.~\eqref{chi_Cij} diverges for the
Coulomb potential interaction in Eq.~\eqref{coulomb}. Clearly a resummation of
terms is needed to get a finite result. In fact,
Ref.~\cite{rehn:15} showed that the interactions are screened up to a length
scale $\lambda$ where
\begin{equation}
\lambda \propto \sqrt{T}.
\label{lambda}
\end{equation}
Within certain approximations, the final result of Ref.~\cite{rehn:15} in $d=2$, is
\begin{equation}
C_{ij} \propto \beta K_0(r_{ij} / \lambda) ,
\label{K0}
\end{equation}
where $K_0(x)$ is a modified Bessel functions which decays exponentially to
zero at large $x$. This is to be compared with
the naive result in Eq.~\eqref{Cij} that
$C_{ij} = \beta J(\mathbf{r}_{ij}) \propto \beta
\log(r_{ij})$. Inserting Eq.~\eqref{K0} into Eq.~\eqref{chi_Cij} one has, at
high-$T$ and in $d = 2$,
\begin{align}
\chi_{SG} &= 1 + \text{const.}\, \lambda^2 \beta^2 \\
&= 1 + \text{const.}' \, \beta \, .
\end{align}
Hence, because of screening, the leading correction to the
infinite-temperature result is of order $\beta$, rather than $\beta^2$ which
would be the case for short-range interactions. This result is clearly shown
by the numerics in Fig.~\ref{fig:screen}.
Because the interactions are screened, we might expect the universality class
of the spin glass transition in the disordered Coulomb antiferromagnet to be
the same as that of the \textit{short-range} EA spin glass. Even if this is
the case, the fact that the screening length is temperature dependent will
give rise to additional, and possibly
large, corrections to FSS which could complicate the analysis.
\subsection{Equilibration}
\label{sec:equil}
To test for equilibration we obtain data for runs of different length in which
the number of sweeps doubles for each run, and for all runs we average over the
last half of the sweeps. It is easy to see that this can actually be done in a single
run by using \textit{all} the data. We require that the data is
independent of run time within small error bars for the last two data
points. Figure \ref{fig:equil} shows an example for $N = 576$ for $d = 2$ at
the lowest temperature $T = 0.032$.
\subsection{Two-dimensions}
\label{sec:2d}
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{equil_chisg.eps}
\caption{
Equilibration plot for $N = 576, d=2$ with $x = 1/ 16$ at the lowest
temperature of $T = 0.032$. The spin glass susceptibility $\chi_{SG}$ is
plotted against Monte Carlo sweeps $t$.
For each point, averaging is carried out over the
last half of the sweeps.
\label{fig:equil}
}
\end{center}
\end{figure}
\begin{table}[htb]
\caption{
Simulation parameters for $d=2$. The concentration is $x=1/16$ where $N = x
L^2$. For each value of
size $N$, $N_{\mathrm{samp}}$ samples were run for $N_{\mathrm{sweep}}$
sweeps with
averaging performed over the last half.
Parallel tempering Monte Carlo was performed with $N_T$ temperatures distributed
between $T_{\mathrm{min}}$ and $T_{\mathrm{max}}$.
}
\begin{ruledtabular}
\begin{tabular}{rrrrrrr}
$N$ & $L$ & $N_{\mathrm{sweep}}$ & $T_{\mathrm{min}}$ &
$T_{\mathrm{max}}$ & $N_T$ & $N_{\mathrm{samp}}$ \\
\hline
64 & 32 & 2097152 & 0.025 & 1 & 15 & 256 \\
144 & 48 & 2097152 & 0.025 & 0.500 & 15 & 1024 \\
256 & 64 & 8388608 & 0.025 & 0.500 & 15 & 767 \\
400 & 80 & 16777216 & 0.025 & 0.25 & 15 & 256 \\
576 & 96 & 83886080 & 0.032 & 0.500 & 16 & 467 \\
784 & 112 & 167772160 & 0.05 & 0.25 & 11 & 96 \\
\end{tabular}
\end{ruledtabular}
\label{tab:2d}
\end{table}
Now we discuss our results for two-dimensions for which we use a fixed
concentration $x = 1/16$. The parameters of the
simulations are shown Table~\ref{tab:2d}.
For the EA model it is well established that the spin glass transition only
occurs at $T = 0$ where the correlation length diverges with an exponent $\nu
\simeq 3.4$~\cite{bray:84,hartmann:01a,fernandez:16} and the exponent
$\eta$, which is related to the
divergence of the spin glass susceptibility according to
Eqs.~\eqref{chisgscale}
and \eqref{gamma}, is
$\eta = 0$.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{g_2d.eps}
\caption{
Data for the Binder ratio in $d = 2$.
\label{fig:g_2d}
}
\end{center}
\end{figure}
Our data for the Binder ratio $g$ is shown in Fig.~\ref{fig:g_2d}. There is no
sign of any intersections and hence no indication of a finite temperature
transition. This is consistent with the EA model in $d = 2$. Figure
\ref{fig:g_2d_scale} shows a scaling plot according to Eq.~\eqref{gscale} assuming
$T_c = 0$. If there is a zero temperature transition the data should collapse,
at least for large enough sizes and low enough temperatures. We were not able
to collapse the data for
\textit{all} sizes with any choice of the correlation length
exponent $\nu$ but the data for the largest sizes collapses fairly well with
$\nu\simeq 3.5$ (as shown) consistent with results for the EA model. However,
there is a large uncertainty in this estimate; we find that any value for
$\nu$ in the range $2.7$ to $4.5$ gives a plausible fit for large sizes.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{g_2d_scale.eps}
\caption{
Scaling plot for Binder ratio in $d = 2$ according to Eq.~\eqref{gscale} with $T_c =
0$. We are not able to get \textit{all} the data to collapse for any value of
the correlation length exponent $\nu$. However the data for the largest sizes
collapses with $\nu \simeq 3.5$ (shown), consistent with the value in the EA
model. However, there are big uncertainties in our estimate (see text).
\label{fig:g_2d_scale}
}
\end{center}
\end{figure}
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{chisg_2d_scale.eps}
\caption{
The spin glass susceptibility in $d = 2$ scaled according to
Eq.~\eqref{chisgscale} with $T_c = 0$ and $\eta = 0$. No value of $\nu$
succeeds in scaling all the data though the largest sizes scale reasonably
well with $\nu \simeq 2.7$ (shown), not very different from the value for the EA model
of $3.4$. However, there are big uncertainties in our estimate (see text).
\label{fig:chisg_2d}
}
\end{center}
\end{figure}
In Fig.~\ref{fig:chisg_2d} we show scaled
data for $\chi_{SG}$ according to
Eq.~\eqref{chisgscale} assuming $T_c = 0$ and $\eta = 0$ (the latter
corresponding to a non-degenerate ground state which seems reasonable).
As with the data for $g$ in
Fig.~\ref{fig:g_2d_scale}, we can not scale all the data for any value of
$\nu$. However, the data for larger sizes scales fairly well with a value of
$2.7$ (shown) not very different from the value for the EA model of 3.4.
However, there are big uncertainties in our estimate; any value between $2.0$
and $4.0$ gives a plausible collapse for large sizes.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{xi_2d.eps}
\caption{
Correlation length divided by system size in $d = 2$.
\label{fig:xi_2d}
}
\end{center}
\end{figure}
Our attempts to scale the data for $g$ and $\chi_{SG}$ indicate the presence
of substantial corrections to FSS for the range of sizes that we can study.
This problem is even more severe for the data for $\xi_L/L$. Being
dimensionless, the data for this quantity should
intersect if there is a transition, see Eq.~\eqref{xiscale}. As shown
in Fig.~\ref{fig:xi_2d} there \textit{is} an intersection involving the
smallest size studied, $N = 64$, but not for larger sizes. Rather the data for
larger sizes seems to merge at low-$T$.
If we try to scale the data for $\xi_L/L$ we need a
large value of $\nu$ to collapse the data for large sizes. Figure
\ref{fig:xi_2d_scale} shows the result with $\nu = 5.0$.
These results indicate that the data for
$\xi_L/L$ is not at large enough sizes to give a clear prediction for the
nature of the transition.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{xi_2d_scale.eps}
\caption{
Scaling plot of the correlation length divided by system size in $d = 2$
assuming $\nu = 5.0$.
\label{fig:xi_2d_scale}
}
\end{center}
\end{figure}
\subsection{Three-dimensions}
\label{sec:3d}
\begin{table}[htb]
\caption{
Simulation parameters for $d=3$. The concentration is $x=1/64$ where $N = x
L^3$. For each value of
size $N$, $N_{\mathrm{samp}}$ samples were run for $N_{\mathrm{sweep}}$
sweeps with
averaging performed over the last half.
Parallel tempering Monte Carlo was performed with $N_T$ temperatures distributed
between $T_{\mathrm{min}}$ and $T_{\mathrm{max}}$.
}
\begin{ruledtabular}
\begin{tabular}{rrrrrrr}
$N$ & $L$ & $N_{\mathrm{sweep}}$ & $T_{\mathrm{min}}$ &
$T_{\mathrm{max}}$ & $N_T$ & $N_{\mathrm{samp}}$ \\
\hline
64 & 16 & 20480 & 0.0100 & 0.2000 & 9 & 1000 \\
216 & 24 & 1310720 & 0.0100 & 0.2000 & 12 & 400 \\
512 & 32 & 41943040 & 0.0200 & 0.2150 & 15 & 360 \\
1000 & 40 & 83886080 & 0.0215 & 0.1305 & 15 & 568 \\
\end{tabular}
\end{ruledtabular}
\label{tab:3d}
\end{table}
Next we discuss our results for three-dimensions for which we use a fixed
concentration $x = 1/64$. The parameters of the
simulations are shown Table~\ref{tab:3d}. For the EA model in $d=3$ it is
firmly established that there is a spin glass transition at non-zero
temperature~\cite{hasenbusch:08b,baity-jesietal:13}.
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{g_3d.eps}
\caption{
Binder ratio for $d = 3$.
\label{fig:g_3d}
}
\end{center}
\end{figure}
\begin{figure}[tb!]
\begin{center}
\includegraphics[width=\columnwidth]{xi_3d.eps}
\caption{
Correlation length divided by system size for $d = 3$.
\label{fig:xi_3d}
}
\end{center}
\end{figure}
Data for the dimensionless quantities $g$ and $\xi_L/L$ are shown in
Figs.~\ref{fig:g_3d} and \ref{fig:xi_3d} respectively. The results for $g$
seem to merge at low-$T$ but do not obviously cross. It should be mentioned
that even for the EA model the splaying out of the data
for $g$ below $T_c$ is only a small
effect, see for example Ref.~\cite{kawashima:96}, but is, nonetheless, observable
with good data on large sizes.
For small sizes the data for $\xi_L/L$ shows a large splaying out, but the data
for large sizes seems only to merge. Splaying out for the smallest size was
also observed in $d=2$, see Fig.~\ref{fig:xi_2d}, and was interpreted as a FSS
correction since it disappears for larger sizes. The same is presumably true
here; we should give most weight to the data for larger sizes. But the larger
size data in Fig.~\ref{fig:xi_3d}
looks very marginal, possibly suggesting that $d = 3$ is the lower
critical dimension, $d_l$. This is different from the EA model where $d_l$ is
approximately, or possibly exactly~\cite{boettcher:05}, equal to $2.5$.
\section{Conclusions}
\label{sec:conclusions}
We have studied spin glass behavior in the random Coulomb Ising
antiferromagnet in two and three dimensions by Monte Carlo simulations, with
results analyzed by FSS. Since the interactions are screened, and so are
effectively short-ranged, a natural hypothesis is that the critical behavior
is the same as that of the short-range Ising EA model.
For the latter, a transition occurs at $T=0$ in $d=2$ but at a
non-zero temperature in $d = 3$. Our results indicate a zero
temperature transition in $d=2$, with a correlation length exponent compatible
with that found for the EA model, though with big error bars. However, in
$d=3$, we do not find unambiguous evidence for a non-zero temperature
transition temperature.
Rather the data for larger sizes seems to be ``marginal''. This could
indicate that the system sizes are simply not large enough to see the
asymptotic critical behavior, or it could be that our model is \textit{not} in
the same universality class as that of the short-range Ising EA model, but
rather has a different lower critical dimension, $d_l=3$ rather than $d_l=2.5$ for
the EA model. If this is the case, the nature of the physics causing the
difference in universal behavior is unclear to us.
There are clearly large corrections to FSS for this problem, the data for the
correlation length in Figs.~\ref{fig:xi_2d} and \ref{fig:xi_3d} providing
striking examples. As well as corrections to scaling that occur for the
short-range EA model, here we have an additional contribution because the
screening length is
\textit{temperature-dependent}. In fact, according to
Ref.~\cite{rehn:15} the screening length is singular for $T \to 0$, see
Eq.~\eqref{lambda}. If this result holds down to $T=0$ it could possibly
\textit{change} the critical behavior in $d=2$, where the transition is also
at zero temperature, rather than simply giving a \textit{correction} to scaling.
In $d=3$, the transition is at finite-$T$ for the EA model, so we expect only
a correction to scaling from the $T$-dependence of $\lambda$.
We close on a historical note. The question of whether or not there is a
finite temperature transition in the $d=3$ Ising EA spin glass was
controversial for many years. It was only later, when better FSS methods
were developed and
computers became more powerful, that the question was definitely answered in
the affirmative. Perhaps, therefore, it is not surprising that this early effort
on a Coulomb spin glass does not leave to a definite conclusion, given the
extra difficulties of long-range interactions and larger corrections to
scaling.
\begin{acknowledgments}
The work of APY is supported in part by the National
Science Foundation under Grant
No.~DMR-1207036 and by
a Gutzwiller Fellowship at the Max Planck Institute
for the Physics of Complex Systems (MPIPKS) Dresden. APY also thanks
Roderich Moessner for his kind hospitality while visiting the MPIPKS.
The work in Dresden was supported by DFG under grant SFB 1143.
RM and JR are grateful to Alex Andreanov, Kedar Damle, Anto Scardicchio and
Arnab Sen for collaboration on related work.
\end{acknowledgments}
| 2920fa802d583f2c598b278200d7e5a7307b01de | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{intro}
A Carnot group (Definition \ref{Carnotdef}) is a simply connected Lie group whose Lie algebra admits a stratification. Carnot groups have translations, dilations, Haar measure and points are connected by horizontal curves (Definition \ref{horizontalcurve}), which are used to define the Carnot-Carath\'eodory (CC) distance (Definition \ref{ccdistance}). With so much structure, the study of analysis and geometry in Carnot groups is an active and interesting research area \cite{BLU07, CDPT07, Gro96, Mon02, SC16, Vit14}.
Many interesting geometric and analytic problems have been studied in the context of Carnot groups. For example, a geometric notion of intrinsic Lipschitz function between subgroups of a general Carnot group was introduced in \cite{FSC2} to study rectifiable sets \cite{FSSC2,Mag} and minimal surfaces \cite{CMPSC1,CMPSC2,SCV}. Moreover, Carnot groups have been applied to study degenerate equations, control theory and potential theory \cite{BLU07}. These problems are highly non-trivial due to the complexity of the geometry of Carnot groups. For instance, any Carnot group (except for Euclidean spaces themselves) contains no subset of positive measure that is bi-Lipschitz equivalent to a subset of a Euclidean space \cite{Sem96}. This follows from Pansu's theorem (Theorem \ref{PansuThm}), a generalization of Rademacher's theorem asserting that Lipschitz maps between Carnot groups are differentiable almost everywhere \cite{Pan89, Mag01}.
A set in a metric space is (upper) porous (Definition \ref{def_porous}) if each of its points sees nearby relatively large holes in the set on arbitrarily small scales. A set is $\sigma$-porous if it is a countable union of porous sets. Properties and applications of porous sets are surveyed in \cite{Zaj87, Zaj05}. Porous sets have applications to differentiability in the linear setting. For example, they were recently used in \cite{PS15} to show that if $n>1$ then there exists a Lebesgue null set $N\subset \mathbb{R}^{n}$ such that every Lipschitz map $f\colon \mathbb{R}^{n}\to \mathbb{R}^{n-1}$ is differentiable at a point of $N$. Applications of porosity to differentiability also exist in the non-linear setting of Carnot groups. For instance, \cite{PS16.2} showed that if $\mathbb{G}$ is a Carnot group and $f\colon \mathbb{G} \to \mathbb{R}$ is a Lipschitz map, then there exists a $\sigma$-porous set $A$ such that if $f$ is differentiable at $x\in \mathbb{G}\setminus A$ in all horizontal directions then $f$ is Pansu differentiable at $x$. Hence it is also interesting to study porous sets and their applications in Carnot groups.
In Section \ref{porousproperties} we investigate the structure of porous sets themselves in Carnot groups. Every $\sigma$-porous set in a metric space is of first category, which means it is a countable union of nowhere dense sets. In $\mathbb{R}^{n}$, $\sigma$-porous sets also have Lebesgue measure zero. However, proving a set is $\sigma$-porous gives a stronger result than proving it is of first category or has measure zero: in $\mathbb{R}^{n}$ there exists a closed nowhere dense set of Lebesgue measure zero which is not $\sigma$-porous \cite[Theorem 2.3]{Zaj05}. We show there is a natural analogue of this result in Carnot groups (Theorem \ref{poroussmall}).
Any Carnot group can be represented in coordinates as a Euclidean space $\mathbb{R}^n$ equipped with some additional structure. Hence one can compare porosity with respect to the CC and Euclidean distances. We show that, at least in the first Heisenberg group, the two notions differ: for each distance, one can construct a set which is porous with respect to the given distance but not porous with respect to the other distance (Proposition \ref{enotcc} and Proposition \ref{ccnote}). This does not follow immediately from the fact that the two distances are not Lipschitz equivalent: if $(M,d)$ is a metric space then the fractal metric $d^{\varepsilon}$, $0<\varepsilon<1$, need not be Lipschitz equivalent to $d$ but gives the same family of porous sets (Remark \ref{sameporoussets}).
In Section \ref{porousnondifferentiability} we give another connection between porosity and differentiability in Carnot groups. We adapt Euclidean arguments from \cite{HMZW97} to show that for any $\sigma$-porous set $P$ in a Carnot group $\mathbb{G}$, one can find a real-valued Lipschitz function on $\mathbb{G}$ which is not even Pansu subdifferentiable on $P$ (Proposition \ref{nonsubdiff}). As a consequence, a universal differentiability set in a Carnot group cannot be $\sigma$-porous (Corollary \ref{udsnotporous}). Universal differentiability sets are sets which contain a point of differentiability for every real-valued Lipschitz function. Such sets were investigated in \cite{DM11, DM12, DM14} in Euclidean spaces and in \cite{PS16.1} in the Heisenberg group.
In Section \ref{gradientproblem} we use the above mentioned non-differentiable functions and arguments from \cite{HMZW97} to investigate the horizontal gradient. The horizontal gradient is defined like the ordinary gradient in Euclidean spaces, but using only horizontal directional derivatives. We show that, as in the Euclidean case, the preimage of an open set under the horizontal gradient mapping is either empty or far from being porous (Theorem \ref{gradienttheorem}). In the Euclidean case this result is related to the so-called \emph{gradient problem}. Assume $n\geq 2$, $\Omega\subset \mathbb{R}^n$ is an open set and $f: \Omega\to \mathbb{R}$ is an everywhere differentiable function. The gradient problem asks whether it is true or not that fixed $\tilde\Omega\subset\mathbb{R}^n$ open, the set $\{p\in\Omega\ |\ \nabla f(p)\in \tilde\Omega\}$ is either empty or of positive $n$-dimensional Lebesgue measure. It is known that the answer is affirmative for $n=1$ \cite{Cla, D} and negative in higher dimensions \cite{Buc}. We are not aware of any result of this type in the context of Carnot groups.
\medskip
\noindent \textbf{Acknowledgement.}
Part of this work was carried out while A. P. was a postdoctoral researcher at the University of Bologna and G. S. was supported by a Taft Summer Research Fellowship from the Charles Phelps Taft Research Center. The project was finalized when A. P. visited the Department of Mathematical Sciences of the University of Cincinnati. A.P. wishes to express his gratitude to the University of Cincinnati for the stimulating atmosphere and the excellent working conditions.
The authors thank helpful referees for valuable remarks and corrections.
\section{Preliminaries}\label{prelim}
In this section we define Carnot groups and porous sets. We refer the interested reader to \cite{BLU07, SC16} for an introduction to Carnot groups and \cite{Zaj87, Zaj05} for information about porosity.
\subsection{Carnot Groups}
\begin{definition}\label{Carnotdef}
A simply connected finite dimensional Lie group $\mathbb{G}$ is a \emph{Carnot group of step $s$} if its Lie algebra $(\mathfrak{g}, [\cdot,\cdot])$ is \emph{stratified of step $s$}, which means that there exist non-zero linear subspaces $V_1, \ldots ,V_s$ of $\mathfrak{g}$ such that
\[\mathfrak{g}=V_1\oplus \cdots \oplus V_s,\]
with
\[[V_1,V_{i}]=V_{i+1} \mbox{ if }1\leq i\leq s-1, \mbox{ and } [V_1,V_s]=\{0\}.\]
Here
\[[V,W]=\mathrm{span}\{[X,Y]: X\in V,\ Y\in W\},\]
where $[X, Y]$ denotes the Lie bracket in the Lie algebra.
\end{definition}
\subsection{Coordinates on Carnot Groups}
The \emph{exponential map} $\exp \colon \mathfrak{g}\to \mathbb{G}$ is defined by $\exp(X)=\gamma(1)$, where
$\gamma \colon[0,1]\to \mathbb{G}$ is the unique solution to the initial value problem:
\[\gamma'(t)=X(\gamma(t)),\qquad \gamma(0)=0.\]
The exponential map $\exp\colon \mathfrak{g}\to \mathbb{G}$ is a diffeomorphism. Throughout the paper we fix a basis $X_{1}, \ldots, X_{n}$ of $\mathfrak{g}$ adapted to the stratification, in which a basis of $V_{i+1}$ follows a basis of $V_{i}$ for each $i$. Define $m=\dim(V_1)$ and notice $X_{1}, \ldots, X_{m}$ is then a basis of $V_1$. Any $x\in \mathbb{G}$ can be written uniquely as
\[x=\exp(x_{1}X_{1}+\ldots +x_{n}X_{n})\]
for some $x_{1},\ldots, x_{n}\in \mathbb{R}$. We identify $x$ with $(x_{1},\ldots, x_{n})\in \mathbb{R}^{n}$ and $\mathbb{G}$ with $(\mathbb{R}^{n},\cdot)$, where the group operation on $\mathbb{R}^{n}$ is determined by the Baker-Campbell-Hausdorff formula on $\mathfrak{g}$. This is known as \emph{exponential coordinates of the first kind}. With this identification the Lebesgue measure $\mathcal{L}^{n}$ on $\mathbb{R}^n$ is a \emph{Haar~measure} on $\mathbb{G}$
For any $\lambda>0$, the \emph{dilation} $\delta_{\lambda}\colon \mathfrak{g}\to \mathfrak{g}$ is defined as the linear map satisfying $\delta_{\lambda}(X)=\lambda^{i}X$ whenever $X\in V_{i}$. Dilations are extended to $\mathbb{G}$ so as to satisfy the equality $\exp \circ \delta_{\lambda} = \delta_{\lambda}\circ \exp$. Dilations satisfy the equality $\delta_{\lambda}(xy)=\delta_{\lambda}(x)\delta_{\lambda}(y)$ for $x, y\in \mathbb{G}$ and $\lambda>0$.
Define the \emph{projection} $p\colon \mathbb{R}^{n} \to \mathbb{R}^{m}$ by $p(x_1,\ldots, x_n)=(x_{1}, \ldots, x_{m})$.
\subsection{Carnot-Carath\'eodory Distance}
\begin{definition}\label{horizontalcurve}
An absolutely continuous curve $\gamma\colon [a,b]\to \mathbb{G}$ is \emph{horizontal} if there exist $u_{1}, \ldots, u_{m}\in L^{1}[a,b]$ such that for almost every $t\in [a,b]$:
\[\gamma'(t)=\sum_{i=1}^{m}u_{i}(t)X_{i}(\gamma(t)).\]
Define the horizontal length of such a curve $\gamma$ by $L(\gamma)=\int_{a}^{b}|u(t)|\; \mathrm{d} t$, where $u=(u_{1}, \ldots, u_{m})$ and $|\cdot|$ denotes the Euclidean norm on $\mathbb{R}^{m}$.
\end{definition}
\begin{definition}\label{ccdistance}
The \emph{Carnot-Carath\'eodory distance (CC distance)} between points $x, y \in \mathbb{G}$ is defined by:
\[d_{c}(x,y)=\inf \{ L(\gamma): \, \gamma \colon [0,1]\to \mathbb{G} \mbox{ horizontal joining }x\mbox{ to }y \}.\]
\end{definition}
The Chow-Rashevskii Theorem implies that any two points of $\mathbb{G}$ can be connected by horizontal curves \cite[Theorem 9.1.3]{BLU07}. It follows that the CC distance is indeed a distance on $\mathbb{G}$. The following identities hold for $x, y, z\in \mathbb{G}$ and $r>0$:
\[d_{c}(zx,zy)=d_{c}(x,y), \qquad d_{c}(\delta_{r}(x),\delta_{r}(y))=rd_{c}(x,y).\]
For brevity we write $d_{c}(x)$ instead of $d_{c}(x,0)$. Since $\mathbb{G}$ is identified with $\mathbb{R}^{n}$, we may compare the CC distance $d_{c}$ with the Euclidean distance $d_{e}$. These distances induce the same topology on $\mathbb{R}^{n}$ but are not Lipschitz equivalent.
\subsection{Pansu Differentiability and Directional Derivatives}
\begin{definition}\label{subdiff}
A map $L\colon \mathbb{G} \to \mathbb{R}$ is \emph{group linear} if the following identities hold for every $x, y\in \mathbb{G}$ and $r>0$:
\[L(xy)=L(x)+L(y), \qquad L(\delta_{r}(x))= r L(x).\]
A map $f\colon \mathbb{G} \to \mathbb{R}$ is \emph{Pansu subdifferentiable} at $x_0\in \mathbb{G}$ if there exists a group linear map $L\colon \mathbb{G} \to \mathbb{R}$ such that
\begin{equation}\label{subdiff2}\liminf_{h\to 0} \frac{f(x_{0}h)-f(x_{0})-L(h)}{d_{c}(h)}\geq 0.\end{equation}
Such a map $f$ is \emph{Pansu differentiable} at $x_0$ if there exists a group linear map $L\colon \mathbb{G}\to \mathbb{R}$ such that
\[\lim_{h\to 0} \frac{f(x_{0}h)-f(x_{0})-L(h)}{d_{c}(h)}= 0.\]
\end{definition}
If the map $L$ in the definition of Pansu differentiability exists then it is unique, called the \emph{Pansu differential}, and we denote it by $df(x_{0})$. For simplicity we now state the celebrated Pansu theorem for scalar targets, though a similar result holds when the target is any Carnot group \cite{Pan89}.
\begin{theorem}[Pansu's Theorem]\label{PansuThm}
Let $f\colon \mathbb{G} \to \mathbb{R}$ be a Lipschitz map. Then $f$ is Pansu differentiable almost everywhere.
\end{theorem}
We define horizontal directional derivatives of a Lipschitz function by composition with horizontal lines.
\begin{definition}\label{directionalderivative}
Suppose $f\colon \mathbb{G}\to \mathbb{R}$, $x\in \mathbb{G}$ and $X\in V_{1}$. We say that $f$ is \emph{differentiable at $x$ in direction $X$} if the following limit exists:
\[Xf(x)=\lim_{t\to 0}\frac{f(x\exp(tX))-f(x)}{t}.\]
\end{definition}
If $x\in \mathbb{G}$ and $X\in V_{1}$ is horizontal, the map $t\mapsto x\exp(tX)$ is Lipschitz. Hence if $f\colon \mathbb{G}\to \mathbb{R}$ is Lipschitz, the composition $t\mapsto f(x\exp(tX))$ is a Lipschitz mapping from $\mathbb{R}$ to $\mathbb{R}$ so is differentiable almost everywhere. Thus Lipschitz functions have directional derivatives along each horizontal line, at almost every point and in the direction of the line. It follows from the definition of $d_{c}$ that if $f\colon \mathbb{G} \to \mathbb{R}$ is Lipschitz then $|X_{i}f(x)|\leq \mathrm{Lip}(f)$ whenever $1\leq i\leq m$ and $X_{i}f(x)$ exists. Here we recall that $X_{1},\ldots, X_{m}$ is our fixed basis of $V_1$, which we used to define horizontal length and CC distance.
\begin{definition}
Suppose $f\colon \mathbb{G}\to \mathbb{R}$, $x\in \mathbb{G}$ and $X_{i}f(x)$ exists for every $1\leq i\leq m$. The \emph{horizontal gradient} of $f$ at $x$ is defined by
\[\nabla_{H}f(x)=(X_1f(x), \ldots ,X_{m}f(x)) \in \mathbb{R}^{m}.\]
\end{definition}
As observed in \cite[Remark 3.3]{MS01}, if $f$ is Pansu differentiable at $x_0$ then $X_{i}f(x)$ exists for every $1\leq i\leq m$ and
\begin{equation}\label{gradform}df(x_0)(h)=\langle \nabla_{H}f(x_0), p(h) \rangle.\end{equation}
Here $\langle \cdot, \cdot \rangle$ denotes the Euclidean inner product on $\mathbb{R}^{m}$.
\subsection{Porous Sets}
When a metric space is clear from the context, we denote the open ball of centre $x$ and radius $r>0$ by $B(x,r)$.
\begin{definition}\label{def_porous}
Let $(M, d)$ be a metric space and $E\subset M$.
Given $\lambda\in (0,1)$ and $a\in E$, we say $E$ is \emph{$\lambda$-porous at $a$} if there is a sequence of points $x_{n}\in M$ with $x_{n}\to a$ such that
\[B(x_n,\lambda d(x_n,a))\cap E=\varnothing \qquad \mbox{ for every }n\in \mathbb{N}.\]
The set $E$ is \emph{porous at $a$} if it is $\lambda$-porous at $a$ for some $\lambda\in (0,1)$.
The set $E$ is \emph{porous} if there is $\lambda \in (0,1)$ such that $E$ is $\lambda$-porous at $a$ for every point $a\in E$, with $\lambda$ independent of the point $a$.
A set is \emph{$\sigma$-porous} if it is a countable union of porous sets.
\end{definition}
Porous sets in metric spaces are nowhere dense and $\sigma$-porous sets in metric spaces are of first category, which means they can be written as a countable union of nowhere dense sets. However, in any non-empty topologically complete metric space without isolated points, there exists a closed nowhere dense set which is not $\sigma$-porous \cite[Theorem 2.4]{Zaj05}.
Unless otherwise stated, porosity in a Carnot group will mean porosity with respect to the CC distance (or a Lipschitz equivalent distance).
\section{Geometry of Porous Sets in Carnot Groups}\label{porousproperties}
In this section we compare CC porosity with Euclidean porosity and other notions of smallness of sets.
\subsection{Measure and Porosity in Carnot Groups}
We begin by observing that porous sets in Carnot groups have measure zero. The simple proof does not depend on the structure of Carnot groups; it is known that porous sets have measure zero in any metric space equipped with a doubling measure.
\begin{proposition}\label{porousnull}
Porous sets in Carnot groups have measure zero.
\end{proposition}
\begin{proof}
Haar measure $\mathcal{L}^{n}$ on $\mathbb{G}$ equipped with the CC distance is doubling, so the Lebesgue differentiation theorem holds \cite[Theorem 1.8]{Hei01}. Given $P\subset \mathbb{G}$ Lebesgue measurable, applying the Lebesgue differentiation theorem to the characteristic function of $P$ implies that the density
\[\lim_{r\downarrow 0} \frac{\mathcal{L}^{n}(P\cap B(x,r))}{\mathcal{L}^{n}(B(x,r))}\]
is equal to $1$ for almost every $x\in P$ and equal to $0$ for almost every $x\notin P$.
Recall that a set is of class $G_{\delta}$ if it can be expressed as a countable intersection of open sets. We claim that to prove the theorem it suffices to show that every Lebesgue measurable porous set has Lebesgue measure zero. This follows from the fact that every porous set is contained in a porous set of class $G_{\delta}$. To see this fact, suppose $P$ is porous with porosity constant $c$. Let $H=(\mathbb{G} \setminus P)^{\circ}$ and define
\[Q=(\mathbb{G}\setminus H)\cap\bigcap_{n=1}^{\infty}Q_{n},\]
where
\[Q_{n}= \{x \in \mathbb{G}: \exists \, z \in H, \, d_{c}(z,x)<\frac{1}{n}, \, \overline{B(z,(c/2)d(z,x))}\subset H\}.\]
The sets $Q_{n}$ are open while $\mathbb{G}\setminus H$ is closed (hence $G_{\delta}$) so $Q$ is $G_{\delta}$. It is easy to check that $Q$ is a porous set containing $P$.
Now suppose $P$ is a Lebesgue measurable porous set with $\mathcal{L}^{n}(P)>0$. By the Lebesgue differentiation theorem, there exists $x\in P$ such that
\begin{equation}\label{pnpf}
\lim_{r\downarrow 0} \frac{\mathcal{L}^{n}(P\cap B(x,r))}{\mathcal{L}^{n}(B(x,r))}=1.
\end{equation}
Since $P$ is porous, there are $0<\lambda<1$ and $x_{n}\to x$ such that:
\begin{equation}\label{emptyint}
B(x_{n},\lambda d(x_{n},x))\cap P=\varnothing \mbox{ for every }n\in \mathbb{N}.
\end{equation}
Using \eqref{pnpf} with radii $2d(x_{n},x)\to 0$ and observing
\[B(x_{n},\lambda d(x_{n},x))\subset B(x,2d(x_{n},x)),\]
equation \eqref{emptyint} implies:
\[ \lim_{n\to \infty} \frac{\mathcal{L}^{n}(B(x_{n},\lambda d(x_{n},x)))}{\mathcal{L}^{n}(B(x,2d(x_{n},x)))}=0.\]
This contradicts the fact that $\mathcal{L}^{n}$ is doubling. Hence $\mathcal{L}^{n}(P)=0$, which proves the proposition.
\end{proof}
Proposition \ref{porousnull} implies that the collection of $\sigma$-porous sets is contained in the collection of first category measure zero sets. In $\mathbb{R}^{n}$ there exists a closed nowhere dense set of Lebesgue measure zero which is not $\sigma$-porous \cite[Theorem 2.3]{Zaj05}. We now prove a similar statement in Carnot groups.
\begin{theorem}\label{poroussmall}
There exists a closed nowhere dense set in $\mathbb{G}$ which has measure zero but is not $\sigma$-porous.
\end{theorem}
Before proving Theorem \ref{poroussmall} we prove the following lemma.
\begin{lemma}\label{puc}
Let $(x,y)\in \mathbb{R}\times \mathbb{R}^{n-1}=\mathbb{G}$. Then for every $t\in \mathbb{R}$, there exists $\tau\in \mathbb{R}^{n-1}$ such that
\[d_{c}((t,\tau),(x,y))=|t-x|.\]
\end{lemma}
\begin{proof}[Proof of Lemma \ref{puc}]
Recall that we are using exponential coordinates of the first kind. By \cite[Proposition 2.2.22]{BLU07} we can write the group law of $\mathbb{G}$ as
\begin{align}\label{grouplaw}
(p\cdot q)_i=p_i+q_i+\mathcal{R}_i(p,q)\quad \mbox{ for every }p,q\in\mathbb{G} \mbox{ and } i=1,\ldots, n,
\end{align}
where $\mathcal{R}_i(p,q)$ is a polynomial function depending only on $p_k$ and $q_k$ with $k<i$ and $\mathcal{R}_i\equiv 0$ for all $i\leq m$. Let $y=(y_2,\ldots, y_n)$. We will define $\tau=(\tau_2,\ldots, \tau_n)$ by induction. Let $\tau_2:=y_2$ and for $3\leq i\leq n$ define:
\[
\tau_i:=y_i+\mathcal{R}_{i}((t, \tau)^{-1},(x,y)).
\]
Such a definition is justified because, since $p^{-1}=-p$ for $p\in \mathbb{G}$, the formula for $\tau_i$ depends only on $\tau_j$ for $j<i$.
Using \eqref{grouplaw}, it is easy to see that $(t,\tau)^{-1}\cdot (x,y)=(x-t,0,\ldots, 0)$. The conclusion follows by noticing that
\[d_c((t,\tau), (x,y))=d_c(x-t,0,\ldots, 0)=|x-t|.\]
\end{proof}
\begin{proof}[Proof of Theorem \ref{poroussmall}]
Using \cite[Theorem 2.3]{Zaj05}, choose a closed nowhere dense set $N\subset \mathbb{R}$ which has Lebesgue measure zero but is not $\sigma$-porous. Then
\[N\times \mathbb{R}^{n-1}\subset \mathbb{R}^{n}=\mathbb{G}\]
is closed and nowhere dense in $\mathbb{G}$ (since the distances $d_{c}$ and $d_{e}$ are topologically equivalent) and clearly has Lebesgue measure zero. It remains to show that $N\times \mathbb{R}^{n-1}$ is not $\sigma$-porous with respect to $d_{c}$.
Arguing as in \cite[Lemma 3.4]{Zaj76}, it suffices to check that if $F\subset \mathbb{R}$ and $F\times \mathbb{R}^{n-1}$ is porous in $\mathbb{G}$ then $F$ is porous in $\mathbb{R}$. Fix $a\in F$. Since $F\times \mathbb{R}^{n-1}$ is porous in $\mathbb{G}$ at $(a,0)$, there exists a constant $0<c<1$ (independent of $a$) and sequences $x^{k}\in \mathbb{R}$, $y^{k}\in \mathbb{R}^{n-1}$, $r^{k}>0$ such that
\begin{enumerate}
\item $B((x^{k},y^{k}),r^{k})\cap (F\times \mathbb{R}^{n-1})=\varnothing$ for all $k\in\mathbb{N}$.
\item $r^{k}> cd_{c}((a,0),(x^{k},y^{k}))$ for all $k\in\mathbb{N}$.
\item $x^{k}\to a$ and $y^{k}\to 0$.
\end{enumerate}
The definition of $d_{c}$ implies $d_{c}((a,0),(x^{k},y^{k}))\geq |x^{k}-a|$. Hence (2) implies $r^{k}>c|x^{k}-a|$. If $t\in \mathbb{R}$ and $|t-x^{k}|<r^{k}$ then by Lemma \ref{puc}, there exists $\tau\in \mathbb{R}^{n-1}$ such that $(t,\tau)\in B((x^{k},y^{k}),r^{k})$. Hence (1) implies $t\notin F$, so $(x^{k}-r^{k},x^{k}+r^{k})\cap F=\varnothing$. Combining these observations with (3) shows that $F$ is porous and concludes the proof.
\end{proof}
\subsection{Comparison of Euclidean and CC Porosity}
\begin{remark}\label{sameporoussets}
If $d_{1}$ and $d_{2}$ are Lipschitz equivalent distances then a set is porous with respect to $d_{1}$ if and only if it is porous with respect to $d_{2}$. Notions of porosity may be the same even if the two distances are not Lipschitz equivalent: it is easy to show that if $(M,d)$ is a metric space then $d$ and any snowflaked metric $d^{\varepsilon}$, $0<\varepsilon<1$, give the same porous sets.
\end{remark}
Recall that $\mathbb{G}$ is identified with $\mathbb{R}^{n}$ in coordinates, so on $\mathbb{G}$ we can consider both the CC distance and the Euclidean distance. We next show that if $\mathbb{G}$ is the first Heisenberg group then these two distances give incomparable families of porous sets.
\begin{definition}\label{Heisenberg}
The first Heisenberg group $\mathbb{H}^1$ is $\mathbb{R}^{3}$ equipped with the non-commutative group law:
\begin{equation}\label{Hgplaw}
(x,y,t) (x',y',t') = (x+x', \ y+y', \ t+t'-2( xy' - yx')).
\end{equation}
The Koranyi distance on $\mathbb{H}^{1}$ is defined by:
\begin{equation}\label{Koranyi}
d_{k}(a,b)=\|a^{-1}b\|_{k}, \mbox{ where } \|(x,y,t)\|_{k}=((x^{2}+y^{2})^{2}+t^{2})^{1/4}.
\end{equation}
For brevity we write $d_{k}(a)$ instead of $d_{k}(a,0)$.
\end{definition}
The Koranyi distance is Lipschitz equivalent to the CC distance on $\mathbb{H}^{1}$, so a set is porous with respect to the Koranyi distance if and only if it is porous with respect to the CC distance. When necessary, we use the notation $B_{e}(a,r)$ and $B_{k}(a,r)$ to distinguish between balls in $\mathbb{R}^{3}$ with respect to the Euclidean or Koranyi distance. The following estimate is clear from \eqref{Koranyi}:
\begin{equation}\label{simpleest}
d_{k}(x,y,t)\geq \max \left( |(x,y)|,\sqrt{|t|} \right).
\end{equation}
Let $C$ be the middle third Cantor set. Note $C$ is $(1/3)$-porous as a subset of $[0,1]$ with the Euclidean distance. If $\alpha, \beta \in \mathbb{R}$ and $S\subset \mathbb{R}$, then we define $\alpha +\beta S=\{\alpha+\beta s \colon s\in S\}$. If $S\subset \mathbb{R}$ is porous then $\alpha +\beta S$ is also porous, with the same porosity constant as $S$.
\begin{lemma}\label{intobs}
For $n, k\in \mathbb{N}$ with $0\leq k\leq 2^{n-1}$, let $A_{n,k}$ be the translated dilated Cantor set:
\begin{equation}\label{Ank}
A_{n,k}:=(2^{-n} + k2^{-2n}) + 2^{-2n}C.
\end{equation}
Then:
\begin{enumerate}
\item The intersection $A_{n,k}\cap A_{m,l}$ is at most one point if $(n,k)\neq (m,l)$.
\item Each set $A_{n,k}$ is $(1/3)$-porous as a subset of the interval
\[[2^{-n} + k2^{-2n}, 2^{-n} + (k+1)2^{-2n}].\]
\item For every $0<t<1$:
\[[t, t+4t^{2}]\cap \bigcup_{n=1}^{\infty} \bigcup_{k=0}^{2^n - 1}A_{n,k}\neq \varnothing.\]
\end{enumerate}
\end{lemma}
\begin{proof}
The first two assertions are clear from the definition of $A_{n,k}$. To prove the third, let $t \in (0,1)$ then choose $n\in \mathbb{N}$ such that $2^{-n}\leq t<2^{-(n-1)}$. Using \eqref{Ank}, we see that the interval $[t,t+2^{-2(n-1)}]$ intersects $A_{n,k}$ or $A_{n+1,k}$ for some $k$. Since
\[2^{-2(n-1)}=4\cdot 2^{-2n}\leq 4t^2,\]
we deduce that $[t,t+4t^{2}]$ intersects $A_{n,k}$ or $A_{n+1,k}$. This proves the lemma.
\end{proof}
\begin{proposition}\label{enotcc}
Define the cone $\Lambda$ by
\[\Lambda:=\{(x,y,t)\in \mathbb{H}^1\ \colon |t| \leq |(x,y)|\}.\]
Define
\[P_{e}:=\Lambda \cap \Big(\{0\}\cup \bigcup_{n=1}^{\infty} \bigcup_{k=0}^{2^n - 1} \{(x,y,t)\in\mathbb{H}^1 \colon |(x,y)|\in A_{n,k} \}\Big). \]
The set $P_{e}$ is porous with respect to the Euclidean distance, but not porous at the point $0$ with respect to the CC distance.
\end{proposition}
We prove Proposition \ref{enotcc} in Claim \ref{enotccclaim1} and Claim \ref{enotccclaim2}.
\begin{claim}\label{enotccclaim1}
The set $P_{e}$ is porous with respect to the Euclidean distance.
\end{claim}
\begin{proof}
We first verify that $P_{e}$ is porous with respect to the Euclidean distance at $0$. Since $P_{e}\subset \Lambda$, it suffices to show
\begin{equation}\label{eucat0}
B_{e}((0,0,1/n),1/3n)\cap \Lambda = \varnothing \mbox{ for every }n\in \mathbb{N}.
\end{equation}
Let $(x,y,t)\in B_{e}((0,0,1/n),1/3n)$. Then $|(x,y)|<1/3n$ and $|t|>2/3n$, in particular $|t|>|(x,y)|$. Hence $(x,y,t)\notin \Lambda$, proving \eqref{eucat0}.
Now we show Euclidean porosity at the remaining points of $P_{e}$: suppose that $(x,y,t) \in P_{e}\setminus \{0\}$. Then there exists $p\in \mathbb{N}$ and $0\leq k\leq 2^{p}-1$ such that $|(x,y)|\in A_{p,k}$. Since $A_{p,k}$ is $(1/3)$-porous in $[2^{-p} + k2^{-2p}, 2^{-p} + (k+1)2^{-2p}]$ at $|(x,y)|$ and the sets $A_{p,k}$ only meet at their endpoints, we may find $r_{n}>0$ with $r_{n}\to |(x,y)|$ such that:
\begin{equation}\label{porosityAnk}
B_{e}(r_{n}, |r_{n}-|(x,y)||/3)\cap (A_{m,l} \cup \{0\})=\varnothing
\end{equation}
for every $m\in \mathbb{N}$ and every $0\leq l\leq 2^{m}-1$.
Define $(x_{n},y_{n})=r_{n}(x,y)/|(x,y)|$. Notice
\begin{equation}\label{july91}
|(x_{n},y_{n})|=r_{n} \quad \mbox{and} \quad |(x_{n},y_{n})-(x,y)|=|r_{n}-|(x,y)||.
\end{equation}
Since $r_{n}\to |(x,y)|$, we have $(x_{n}, y_{n}, t)\to (x,y,t)$. We claim that for $n\in \mathbb{N}$:
\begin{equation}\label{P1porosity}
B_{e}((x_{n},y_{n},t), d_{e}((x_{n},y_{n},t),(x,y,t))/3)\cap P_{e}=\varnothing.
\end{equation}
For this, suppose:
\[(a,b,c)\in B_{e}((x_{n},y_{n},t), d_{e}((x_{n},y_{n},t),(x,y,t))/3).\]
Using \eqref{july91}, we have
\[d_{e}((x_{n},y_{n},t),(x,y,t))=|(x_{n},y_{n})-(x,y)|=|r_{n}-|(x,y)||.\]
Hence, using the above equality, the triangle inequality, and \eqref{july91} again,
\[ ||(a,b)|-r_{n}| \leq |(a,b)-(x_{n},y_{n})|\leq |r_{n}-|(x,y)||/3.\]
This implies
\[|(a,b)|\in B_{e}(r_{n}, |r_{n}-|(x,y)||/3).\]
Using \eqref{porosityAnk}, we deduce that $|(a,b)|\neq 0$ and $|(a,b)|\notin A_{m,l}$ for any choice of $m\in \mathbb{N}$ and $0\leq l\leq 2^{m}-1$. Hence $(a,b,c)\notin P_{e}$, verifying \eqref{P1porosity}. This shows that $P_{e}$ is Euclidean porous at $(x,y,t)$ and completes the proof.
\end{proof}
\begin{claim}\label{enotccclaim2}
The set $P_{e}$ is not porous with respect to the CC distance at $0$.
\end{claim}
\begin{proof}
Let $\lambda>0$. It suffices to show that if $(x,y,t) \in \mathbb{H}^{1}$ and $d_{k}(x,y,t)$ is sufficiently small, then:
\begin{equation}\label{nonCCporous}
B_{k}( (x, y ,t), \lambda d_{k}(x, y, t))\cap P_{e} \neq \varnothing.
\end{equation}
\vspace{0.3cm}
\textit{Case 1:} Suppose $(x,y,t)\notin \Lambda$, so $|(x,y)| < |t|$. For each $s>0$, define:
\[p_{s}:=\left(x+\frac{sx}{|(x,y)|}, y+\frac{sy}{|(x,y)|}, t-\sqrt{15}s^2\right).\]
Without loss of generality we assume $t>0$. In the case $t<0$, one should instead choose $t+\sqrt{15}s^2$ as the final coordinate in the definition of $p_{s}$.
We first notice that $d_{k}(p_{s},(x,y,t))=2s$. Therefore for every $0<s<\lambda\sqrt{t}/2$ we have $p_s\in B_k((x,y,t), \lambda d_{k}(x, y, t))$. Since
\[\left| \left( x+\frac{sx}{|(x,y)|}, y+\frac{sy}{|(x,y)|} \right) \right| = |(x,y)|+s,\]
we see $p_s\in \Lambda$ if and only if
\[
t-\sqrt{15}s^2\leq |(x,y)|+s
\]
so it suffices to choose $s$ such that
\[
t-\sqrt{15}s^2\leq s.
\]
If $d_k(x,y,t)$ is sufficiently small, which forces $t$ to be small, this holds for every $s \in (\lambda\sqrt{t}/8, \lambda\sqrt{t}/2)$.
We now show that $p_s\in P_{e}$ for some such $s$. This holds provided the interval
\[ \left(|(x,y)|+ \frac{\lambda\sqrt{t}}{8},\ |(x,y)|+ \frac{\lambda\sqrt{t}}{2} \right)\]
intersects some set $A_{n,k}$. By making $d_{k}(x,y,t)$ small, we may assume that $|(x,y)|+ \frac{\lambda\sqrt{t}}{8}$ lies in $(0,1)$. By Lemma \ref{intobs}, it then suffices to prove that
\begin{align}\label{yy}
\left( |(x,y)|+\frac{\lambda \sqrt{t}}{2} \right) - \left(|(x,y)|+\frac{\lambda\sqrt{t}}{8}\right)-5\left(|(x,y)|+\frac{\lambda\sqrt{t}}{8}\right)^2\geq 0.
\end{align}
The factor 5 instead of 4 in \eqref{yy} takes into account that Lemma \ref{intobs} requires a closed interval $[\theta, \theta+4\theta^{2}]$ rather than an open one. To prove \eqref{yy}, let $\delta=3\lambda/320$ and assume $|(x,y)|<\delta$, which implies $|(x,y)|^2<\delta |(x,y)|$. Using also the inequality $(a+b)^2\leq 2a^{2}+2b^{2}$, $t, \lambda \in (0,1)$ and $|(x,y)| < |t|$, we estimate as follows:
\begin{align*}
&\left(|(x,y)|+\frac{\lambda\sqrt{t}}{2}\right) - \left(|(x,y)|+\frac{\lambda\sqrt{t}}{8}\right)-5\left(|(x,y)|+\frac{\lambda\sqrt{t}}{8}\right)^2\\
&\qquad > \frac{7\lambda\sqrt{t}}{32}-10\delta |(x,y)|\\
&\qquad > \frac{7\lambda\sqrt{t}}{32}-10\delta \sqrt{t}\\
&\qquad \geq 0.
\end{align*}
This verifies \eqref{yy} so $p_{s}\in P_{e}$ for some $s$, verifying \eqref{nonCCporous} and completing the proof in this case.
\vspace{0.3cm}
\textit{Case 2:} Suppose $(x,y,t)\in \Lambda$, so $|t|\leq |(x,y)|$. For each $s>0$, define
\begin{equation}\label{psform}
p_{s}= \left(x+\frac{sx}{|(x,y)|}, y+\frac{sy}{|(x,y)|}, t\right).
\end{equation}
It is easy to show that $d_{k}(p_{s},(x,y,t))=s$ so $p_{s}\in B_{k}( (x, y ,t), \lambda d_{k}(x, y, t))$ whenever $0<s<\lambda |(x,y)|$. Clearly also
\begin{equation}\label{psprojnorm}
\left| \left( x+\frac{sx}{|(x,y)|}, y+\frac{sy}{|(x,y)|} \right) \right| = |(x,y)|+s.
\end{equation}
Since $s>0$ and $|t|\leq |(x,y)|$, we deduce that $p_{s}\in \Lambda$.
For sufficiently small $|(x,y)|$, the interval $(|(x,y)|, |(x,y)|+\lambda |(x,y)|)$ will contain a subinterval of the form $[\theta, \theta + 4\theta^{2}]$ for some $\theta \in (0,1)$, which by Lemma \ref{intobs} necessarily meets some set $A_{n,k}$. Using \eqref{psform} and \eqref{psprojnorm}, this yields \eqref{nonCCporous} and proves the claim in this case.
\end{proof}
\begin{proposition}\label{ccnote}
Define the cusp $\Upsilon$ by
\[\Upsilon:=\{(x,y,t)\in \mathbb{H}^1 \colon |t| \geq 2|(x,y)|^{2}\}.\]
Define
\[P_{c}:=\Upsilon \cap \Big(\{0\}\cup \bigcup_{n=1}^{\infty} \bigcup_{k=0}^{2^n - 1} \{(x,y,t)\in\mathbb{H}^1 \colon |(x,y)|\in A_{n,k} \}\Big).\]
The set $P_{c}$ is porous with respect to the CC distance, but not porous at the point $0$ with respect to the Euclidean distance.
\end{proposition}
We prove Proposition \ref{ccnote} in Claim \ref{ccnoteclaim1} and Claim \ref{ccnoteclaim2}.
\begin{claim}\label{ccnoteclaim1}
The set $P_{c}$ is porous with respect to the CC distance.
\end{claim}
\begin{proof}
We first verify that $P_{c}$ is porous with respect to the Koranyi distance at $0$. Since $P_{c}\subset \Upsilon$, it suffices to prove that
\begin{equation}\label{july11eq}
B_{k}((1/n,0,0), 1/3n)\cap \Upsilon=\varnothing \quad \mbox{ for every }n\in\mathbb{N}.
\end{equation}
If $(x,y,t)\in B_{k}((1/n,0,0), 1/3n)$ then
\[\Big|(x-\frac{1}{n},y)\Big|\leq \frac{1}{3n}\quad \mbox{ and }\quad \Big|t+\frac{2y}{n}\Big|\leq \frac{1}{9n^2}.\]
In particular:
\[\frac{2}{3n} \leq x\leq \frac{4}{3n},\quad |y|\leq \frac{1}{3n}\quad \mbox{and}\quad \Big|t+\frac{2y}{n}\Big|\leq \frac{1}{9n^2}.\]
Therefore:
\begin{align*}
|t|\leq \Big|t+\frac{2y}{n}\Big|+\frac{2|y|}{n} \leq \frac{1}{9n^2}+\frac{2}{3n^2}\leq \frac{x^2}{4}+\frac{3x^2}{2}< 2 |(x,y)|^2.
\end{align*}
We conclude that $(x,y,t)\notin \Upsilon$, verifying \eqref{july11eq}.
We now show CC porosity at the remaining points of $P_{c}$. Suppose that $(x,y,t) \in P_{c}\setminus \{0\}$. Using the definition of $P_{c}$, there exists $p\in \mathbb{N}$ and $0\leq k\leq 2^{p}-1$ such that $|(x,y)|\in A_{p,k}$. Since $A_{p,k}$ is $(1/3)$-porous as a subset of the interval $[2^{-p} + k2^{-2p}, 2^{-p} + (k+1)2^{-2p}]$ and the sets $A_{p,k}$ only meet at their endpoints, we may find $r_{n}>0$ with $r_{n}\to |(x,y)|$ such that:
\begin{equation}\label{usefulpc}
B_{e}(r_{n}, |r_{n}-|(x,y)||/3)\cap (A_{m,l} \cup \{0\})=\varnothing
\end{equation}
for every $m\in \mathbb{N}$ and every $0\leq l\leq 2^{m}-1$.
Define $(x_{n},y_{n})=r_{n}(x,y)/|(x,y)|$. Notice that $|(x_{n},y_{n})|=r_{n}$ and
\[|(x_{n},y_{n})-(x,y)|=|r_{n}-|(x,y)||.\]
Clearly $(x_{n}, y_{n}, t)\to (x,y,t)$. We claim that for every $n\in \mathbb{N}$:
\begin{equation}\label{usefulpc2}
B_{k}((x_{n},y_{n},t), d_{k}((x_{n},y_{n},t),(x,y,t))/3)\cap P_{c}=\varnothing.
\end{equation}
For this, suppose
\[(a,b,c)\in B_{k}((x_{n},y_{n},t), d_{k}((x_{n},y_{n},t),(x,y,t))/3).\]
Using \eqref{Hgplaw}, we see:
\[d_{k}((x_{n},y_{n},t),(x,y,t))=|(x-x_{n}, y-y_{n})|=|r_n-|(x,y)||.\]
Hence
\begin{align*}
|(a,b)-(x_{n},y_{n})| &\leq d_{k}((x_{n},y_{n},t),(x,y,t))/3\\
&= |r_n-|(x,y)||/3.
\end{align*}
We deduce that
\[|(a,b)|\in B_{e}(r_{n}, |r_{n}-|(x,y)||/3).\]
Using \eqref{usefulpc}, we deduce that $|(a,b)|\neq 0$ and $|(a,b)|\notin A_{m,l}$ for any choice of $m\in \mathbb{N}$ and $0\leq l\leq 2^{m}-1$. Hence $(a,b,c)\notin P_{c}$, verifying \eqref{usefulpc2} which proves that $P_{c}$ is porous with respect to the Koranyi metric.
\end{proof}
\begin{claim}\label{ccnoteclaim2}
The set $P_{c}$ is not porous with respect to the Euclidean distance at the point $0$.
\end{claim}
\begin{proof}
Let $\lambda>0$. It suffices to show that for any $(x,y,t) \in \mathbb{H}^{1}$ with $d_{e}(x,y,t)$ sufficiently small:
\begin{equation}\label{nonEporous1}
B_{e}( (x, y ,t), \lambda d_{e}(x, y, t))\cap P_{c} \neq \varnothing.
\end{equation}
Without loss of generality assume $t\geq 0$. It will be clear from the proof that a similar argument works if $t<0$. For $0\leq s<|(x,y)|$, let
\begin{equation}\label{pcdefqs}
q_{s}:=\left( x-\frac{sx}{|(x,y)|}, y-\frac{sy}{|(x,y)|}, t+s \right).
\end{equation}
If $t<0$, one can instead choose $t-s$ as the final coordinate in the definition of $q_{s}$. Clearly $d_{e}(q_{s},(x,y,t))=s\sqrt{2}$. Hence $s< \lambda |(x,y)|/\sqrt{2}$ implies
\[q_{s}\in B_{e}( (x, y ,t), \lambda d_{e}(x, y, t)).\]
Next notice
\begin{equation}\label{usefula}
\left| \left(x-\frac{sx}{|(x,y)|}, y-\frac{sy}{|(x,y)|}\right) \right|=|(x,y)|-s.
\end{equation}
Using the definition of $\Upsilon$, \eqref{pcdefqs} and \eqref{usefula}, we see $q_{s}\in \Upsilon$ if and only if
\begin{equation}\label{tosatisfy1}
t+s\geq 2(|(x,y)|-s)^2.
\end{equation}
Since $t\geq 0$, \eqref{tosatisfy1} holds whenever
\begin{equation}\label{tosatisfy2}
s\geq 2(|(x,y)|-s)^2.
\end{equation}
Ensure $(x,y,t)$ is chosen with $d_{e}(x,y,t)$ small enough so that \eqref{tosatisfy2} holds if $s=\lambda |(x,y)|/2\sqrt{2}$. Then \eqref{tosatisfy2} holds, and hence $q_{s}\in \Upsilon$, whenever $s$ satisfies
\begin{equation}\label{srange}
\frac{\lambda |(x,y)|}{2\sqrt{2}} \leq s< \frac{\lambda |(x,y)|}{\sqrt{2}}.
\end{equation}
Finally we observe that if $d_{e}(x,y,t)$ is sufficiently small, then the interval
\[\left( |(x,y)|- \frac{\lambda |(x,y)|}{\sqrt{2}},\ |(x,y)| - \frac{\lambda |(x,y)|}{2\sqrt{2}} \right)\]
will contain a subinterval of the form $[\theta, \theta + 4\theta^{2}]$ for some $\theta \in (0,1)$, hence by Lemma \ref{intobs} intersect some set $A_{n,k}$. Using \eqref{pcdefqs}, \eqref{usefula} and the definition of $P_{c}$, this implies $q_{s}\in P_{c}$ for some $s$ satisfying \eqref{srange}. For such $s$, we have
\[q_{s}\in B_{e}( (x, y ,t), \lambda d_{e}(x, y, t))\cap P_{c},\]
which proves \eqref{nonEporous1}.
\end{proof}
\begin{remark}
We expect that in any Carnot group there exist sets which are porous in each distance (CC or Euclidean) but not even $\sigma$-porous in the other distance. Such constructions and their justifications may be complicated, since it is harder to show a set is not $\sigma$-porous than to show it is not porous: one may need to use Foran systems or similar techniques \cite[Lemma 4.3]{Zaj87}.
\end{remark}
\section{Non-differentiability on a $\sigma$-Porous Set}\label{porousnondifferentiability}
In this section we construct a Lipschitz function which is subdifferentiable at no point of a given $\sigma$-porous set in a Carnot group (Theorem \ref{nonsubdiff}). We first give basic properties of (Pansu) subdifferentiability (Definition \ref{subdiff}), which are simple adaptations of similar statements in Banach spaces \cite{HMZW97}.
\begin{proposition}\label{basicproperties}
The following statements hold:
\begin{enumerate}
\item $f\colon \mathbb{G}\to \mathbb{R}$ is Pansu differentiable at $a\in \mathbb{G}$ if and only if $f$ and $-f$ are both Pansu subdifferentiable at $a$.
\item Suppose $f$ and $g$ are Pansu subdifferentiable at $a$ and $\lambda>0$. Then $f+g$ and $\lambda f$ are Pansu subdifferentiable at $a$.
\item If
\[\limsup_{h\to 0} \frac{f(ah)+f(ah^{-1})-2f(a)}{d_{c}(h)}>0\]
then $-f$ is not Pansu subdifferentiable at $a$.
\item $f\colon \mathbb{R} \to \mathbb{R}$ is subdifferentiable at $a$ if and only if $f_{+}(a)\geq f^{-}(a)$, where
\[f_{+}(a)=\liminf_{t\to 0+} \frac{f(a+t)-f(a)}{t}\]
denotes the lower right Dini derivative, and
\[f^{-}(a)=\limsup_{t\to 0-} \frac{f(a+t)-f(a)}{t}\]
denotes the upper left Dini derivative.
\item Suppose $L\colon \mathbb{G}\to \mathbb{R}$ is a non-zero group linear map. If $H\colon \mathbb{R}\to \mathbb{R}$ is not subdifferentiable at $L(a)$, then $H\circ L$ is not Pansu subdifferentiable at $a$.
\item If $f\colon \mathbb{G}\to \mathbb{R}$ attains a local minimum at a point $x_0 \in\mathbb{G}$, then $f$ is Pansu subdifferentiable at $x_0$.
\end{enumerate}
\end{proposition}
\begin{proof}
For (1), clearly Pansu differentiability of $f$ implies subdifferentiability of $f$ and $-f$. We check the opposite implication. Suppose $f$ and $-f$ are Pansu subdifferentiable. Then there exist group linear maps $L_{1}, L_{2}\colon \mathbb{G} \to \mathbb{R}$ such that
\begin{equation}\label{fsub} \liminf_{h\to 0} \frac{f(x_{0}h)-f(x_{0})-L_{1}(h)}{d_{c}(h)}\geq 0,\end{equation}
\begin{equation}\label{mfsub} \liminf_{h\to 0} \frac{f(x_{0})-f(x_{0}h)-L_{2}(h)}{d_{c}(h)}\geq 0.\end{equation}
Adding \eqref{fsub} and \eqref{mfsub} yields
\begin{equation}\label{used1} \liminf_{h\to 0} \frac{-L_{1}(h)-L_{2}(h)}{d_{c}(h)}\geq 0.\end{equation}
For each $v\in \mathbb{G}$, let $h=\delta_{t}(v)$. Group linearity of $L_{1}$ and $L_{2}$ implies
\begin{equation}\label{used2}\frac{-L_{1}(h)-L_{2}(h)}{d_{c}(h)}=\frac{-L_{1}(v)-L_{2}(v)}{d_{c}(v)}.\end{equation}
Letting $t\to 0$ and using \eqref{used1} and \eqref{used2} yields $-L_{1}(v)-L_{2}(v)\geq 0$. Hence $L_{1}(v)+L_{2}(v)\leq 0$ for every $v\in \mathbb{G}$. Replacing $v$ by $v^{-1}$ yields the opposite inequality $L_{1}(v)+L_{2}(v)\geq 0$ for every $v\in \mathbb{G}$. Hence $L_{2}=-L_{1}$. Pansu differentiability of $f$ follows directly from this equality, \eqref{fsub} and \eqref{mfsub}.
Statement (2) is trivial.
Suppose the condition in (3) holds but $-f$ is Pansu subdifferentiable at $a$ with corresponding map $L$. Since $L$ is group linear, $L(h^{-1})=-L(h)$. Hence
\begin{align*}
&\frac{-f(ah)+f(a)-L(h)}{d_{c}(h)}+\frac{-f(ah^{-1})+f(a)-L(h^{-1})}{d_{c}(h)}\\
&\qquad =-\frac{f(ah)+f(ah^{-1})-2f(a)}{d_{c}(h)}.
\end{align*}
Consequently either
\[ \liminf_{h\to 0} \frac{-f(ah)+f(a)-L(h)}{d_{c}(h)}<0\]
or
\[ \liminf_{h\to 0} \frac{-f(ah^{-1})+f(a)-L(h^{-1})}{d_{c}(h)}<0.\]
This contradicts Pansu subdifferentiability of $-f$ at $a$, proving (3).
Statement (4) is exactly as stated in \cite{HMZW97}.
We now prove (5). Since $H$ is not subdifferentiable at $L(a)$, we know by (4) that $H_{+}(L(a))<H^{-}(L(a))$. By \cite[Theorem 19.2.1]{BLU07}, every element of $\mathbb{G}$ is a product of elements of the form $\exp(X)$ with $X\in V_1$. Hence, since $L$ is a non-zero group linear map, there exists $X\in V_{1}$ such that $L(\exp(X)) \neq 0$. To show $H\circ L$ is not Pansu subdifferentiable at $a$, we show $f\colon \mathbb{R}\to \mathbb{R}$ given by $f(t)=H(L(a\exp(tX)))$ is not subdifferentiable at $0$. Notice:
\begin{align*}
\frac{f(t)-f(0)}{t}&=\frac{H(L(a\exp(tX)))-H(L(a))}{t}\\
&= \frac{H(L(a)+tL(\exp(X)))-H(L(a))}{tL(\exp(X))}\cdot L(\exp(X)).
\end{align*}
If $L(\exp(X))>0$ then
\[f_{+}(0)=H_{+}(L(a))L(\exp(X))\]
and
\[f^{-}(0)=H^{-}(L(a))L(\exp(X)).\]
Hence $f_{+}(0)<f^{-}(0)$, so $f$ is not subdifferentiable at $0$. If $L(\exp(X))<0$, then
\[f_{+}(0)=H^{-}(L(a))L(\exp(X))\]
and
\[f^{-}(0)=H_{+}(L(a))L(\exp(X))),\]
so again $f_{+}(0)<f^{-}(0)$. We conclude that $f$ is not subdifferentiable at $0$. Hence (5) holds.
We now verify (6). If $f$ has a local minimum at $x_0$, then $f(x_0 h)\geq f(x_0)$ for all $h\in\mathbb{G}$ with $d_{c}(h)$ sufficiently small. Therefore:
\[\frac{f(x_0 h)-f(x_0)}{d_{c}(h)}\geq 0\quad \mbox{ for } h\in\mathbb{G}\setminus\{0\} \mbox{ with }d_{c}(h)\mbox{ sufficiently small},\]
and (6) follows.
\end{proof}
The following lemma \cite{HMZW97} will be used in Section \ref{gradientproblem}. There is no similar statement if $\mathbb{R}$ is replaced by $\mathbb{R}^n$ ($n>1$) or $\mathbb{H}^{n}$: these spaces admit measure zero sets containing points of (Pansu) differentiability for every real-valued Lipschitz function \cite{Pre90, PS16.1, LDPS17}.
\begin{lemma}\label{nullnonsubdiff}
Let $Z\subset \mathbb{R}$ have Lebesgue measure zero. Then there exists a Lipschitz function on $\mathbb{R}$ which is subdifferentiable at no point of $Z$.
\end{lemma}
In Banach spaces which admit a suitable bump function, one can construct a Lipschitz function which is differentiable at no point of any given $\sigma$-porous set. We show that the same is true in a general Carnot group. Our proof is a modification of \cite[Lemma 2]{HMZW97}. The next lemma is \cite[Lemma 3]{HMZW97}, where the set is assumed to be `uniformly porous'. In the present paper `uniformly porous' simply means `porous', since our definition of porosity already requires the relative size of holes (i.e. the parameter $\lambda$ in Definition \ref{def_porous}) to be uniform over the porous set. We denote the closed ball of center $x$ and radius $r>0$ in a metric space by $\overline{B}(x,r)$.
\begin{lemma}\label{cover}
Let $M$ be a metric space and $E$ be a closed porous subset of $M$. Then there exists $C>1$ and $S\subset M\times (0,1)$ such that the family $\mathcal{B}=\{ \overline{B}(x,r)\colon (x,r)\in S\}$ is disjoint, $\bigcup \mathcal{B} \cap E=\varnothing$, and for each $\delta>0$:
\[ \bigcup \mathcal{B} \cup \bigcup \{\overline{B}(x,Cr)\colon (x,r)\in S,\ r<\delta \}=M.\]
\end{lemma}
\begin{definition}
A \emph{bump function} $b\colon \mathbb{G}\to \mathbb{R}$ is a Lipschitz function which is everywhere Pansu differentiable, has compact support, is non-negative and is not identically zero.
\end{definition}
\begin{lemma}
There exists a bump function $b\colon \mathbb{G} \to \mathbb{R}$ with $\mathrm{Lip}(b)\leq 1$.
\end{lemma}
\begin{proof}
Let $b\colon \mathbb{R}^{n}\to \mathbb{R}$ be a non-negative $C^{1}$ function (in the classical sense) with compact support which is not identically zero. We can choose $b$ to have Lipschitz constant as small as we desire, without changing the support of $b$. Classically $C^{1}$ functions are Pansu differentiable with continuously varying derivative \cite[Remark 5.9]{FSSC2}, so $b$ is also Pansu differentiable.
The CC distance is bounded below by a multiple of the Euclidean distance on compact sets \cite{Monti}, so $b$ is also Lipschitz with respect to the CC distance in the domain. Since the Euclidean Lipschitz constant could be made arbitrarily small, we can ensure the CC Lipschitz constant is at most 1.
Finally, the CC distance and the Euclidean distance induce the same topology \cite{Monti}, so $b$ has compact support also with respect to the CC distance.
\end{proof}
\begin{theorem}\label{nonsubdiff}
Let $E$ be a $\sigma$-porous subset of $\mathbb{G}$. Then there is a Lipschitz function $f\colon \mathbb{G}\to \mathbb{R}$ which is Pansu subdifferentiable at no point of $E$.
\end{theorem}
\begin{proof}
Let $b$ be a bump function on $\mathbb{G}$. By composing with group translations and dilations if necessary, we may assume that $b(0)=\beta>0$, $b$ is supported in $B(0,1)$ and $b$ is $1$-Lipschitz. Write $E=\cup_{i=1}^{\infty}E_{i}$ where each set $E_{i}$ is porous. We apply Lemma \ref{cover} to each set $E_{i}$ considered as a subset of the metric space $M_{i}=\mathbb{G}\setminus (\overline{E}_{i}\setminus E_{i})$. Then $E_{i}$ is porous and closed in $M_{i}$. Choose $S_{i}$ and $C_{i}>1$ corresponding to $E_{i} \subset M_{i}$ using Lemma \ref{cover}.
For each $i$, the family $\mathcal{B}_{i}^{\ast}=\{B(x,r)\colon (x,r)\in S_{i}\}$ is disjoint, where the balls are defined in $\mathbb{G}$ (not the subspace $M_{i}$). Indeed, if two members of $\mathcal{B}_{i}^{\ast}$ did intersect in $\mathbb{G}$ then, using Lemma \ref{cover} and the definition of $M_{i}$, this intersection must be in $\overline{E_{i}}$. Since the intersection is open, it would contain a point of $E_{i} \subset M$ which is impossible.
Lemma \ref{cover} implies that for every $\delta>0$,
\[\bigcup \{\overline{B}(x,C_{i}r)\colon (x,r)\in S_{i},\ r<\delta \}\supset E_{i}.\]
For each $i$, define $f_{i}\colon \mathbb{G}\to \mathbb{R}$ by:
\[f_{i}(x) = \begin{cases} 0 & \mbox{if } x\notin \bigcup \mathcal{B}_{i}^{\ast}, \\ rb(\delta_{1/r}(y^{-1}x)) & \mbox{if } x\in B(y,r),\ (y,r)\in S_{i}. \end{cases}\]
Since the CC distance is invariant under left translations, compatible with dilations and $b$ is $1$-Lipschitz, it follows $x\mapsto rb(\delta_{1/r}(y^{-1}x))$ is also $1$-Lipschitz. Each map $f_{i}$ is a supremum of $1$-Lipschitz functions, hence $1$-Lipschitz. Since also $S_{i}\subset \mathbb{G}\times (0,1)$, we have $0\leq f_{i}\leq 2$. The maps $f_{i}$ are Pansu differentiable on $\bigcup \mathcal{B}_{i}^{\ast}$ because $b$ is Pansu differentiable on $\mathbb{G}$. Clearly also $f_{i}(x)=r\beta$ for $(x,r)\in S_{i}$.
Suppose $x\in E_{i}$. Then for arbitrarily small $r>0$, we find $(z,r)\in S_{i}$ such that $x\in \overline{B}(z,C_{i}r)$. Hence $d_{c}(x,z)\leq C_{i}r$ and $f_{i}(z)=r\beta$, which implies that $f_{i}(z)/d_{c}(x,z)\geq \beta/C_{i}$. We can write $z=xh$, where $h=x^{-1}z$ and $d_{c}(h)\leq C_{i}r$. Since $E_{i}\cap \bigcup \mathcal{B}_{i}^{\ast}=\varnothing$, we know $f_{i}(x)=0$. Letting $r\to 0$ gives:
\begin{equation}\label{july11a}\limsup_{h \to 0} \frac{f_{i}(xh)+f_{i}(xh^{-1})-2f_{i}(x)}{d_{c}(h)}\geq \frac{\beta}{C_{i}}>0.\end{equation}
If $x\in \bigcup \mathcal{B}_{i}^{\ast}$ then Pansu differentiability of $f_{i}$ at $x$ implies:
\begin{equation}\label{july11b}\lim_{h\to 0} \frac{f_{i}(xh)+f_{i}(xh^{-1})-2f_{i}(x)}{d_{c}(h)}=0.\end{equation}
If $x\notin \bigcup \mathcal{B}_{i}^{\ast}$ then $f_{i}(x)=0$ implies
\begin{equation}\label{july11c}\liminf_{h\to 0} \frac{f_{i}(xh)+f_{i}(xh^{-1})-2f_{i}(x)}{d_{c}(h)}\geq 0.\end{equation}
Define $f\colon \mathbb{G}\to \mathbb{R}$ by
\[f(x)=\sum_{i=1}^{\infty} \frac{f_{i}(x)}{2^{i}}.\]
Since each $f_{i}$ is $1$-Lipschitz, $f$ is $1$-Lipschitz. Let $j\in \mathbb{N}$ and $x\in E_{j}$. Then, for any $J>j$,
\begin{align*}
&\limsup_{h\to 0} \frac{f(xh)+f(xh^{-1})-2f(x)}{d_{c}(h)}\\
&\qquad \geq \limsup_{h\to 0} \frac{1}{2^j} \frac{f_{j}(xh)+f_{j}(xh^{-1})-2f_{j}(x)}{d_{c}(h)}\\
&\qquad \qquad + \liminf_{h\to 0} \sum_{\substack{1\leq i\leq J\\ i\neq j}} \frac{1}{2^i} \frac{f_{i}(xh)+f_{i}(xh^{-1})-2f_{i}(x)}{d_{c}(h)}\\
&\qquad \qquad -\sum_{i=J+1}^{\infty} \frac{2}{2^i}\\
&\qquad \geq \frac{\beta}{C_{j}}-\frac{1}{2^{J-1}}\\
&\qquad >0
\end{align*}
for $J$ sufficiently large, using \eqref{july11a}, \eqref{july11b}, \eqref{july11c} and that each $f_{i}$ is $1$-Lipschitz. Proposition \ref{basicproperties}(3) then asserts that $-f$ is not Pansu subdifferentiable on $E$, which proves the theorem.
\end{proof}
A universal differentiability set in $\mathbb{G}$ is a subset $A\subset \mathbb{G}$ such that every Lipschitz function $f\colon \mathbb{G}\to \mathbb{R}$ is Pansu differentiable at a point of $A$. Theorem \ref{nonsubdiff} gives the following corollary showing that, as in the Euclidean case, universal differentiability sets are far from being porous.
\begin{corollary}\label{udsnotporous}
A universal differentiability set in a Carnot group cannot be $\sigma$-porous.
\end{corollary}
\section{The Horizontal Gradient}\label{gradientproblem}
We next use Theorem \ref{nonsubdiff} and arguments from \cite{HMZW97} to prove Theorem \ref{gradienttheorem}. Roughly, this states that preimages of open sets under the horizontal gradient are far from being $\sigma$-porous. We first prove a useful lemma. Recall that $m$ is the dimension of the horizontal layer $V_{1}$ of $\mathbb{G}$.
\begin{lemma}\label{usefullemma}
Let $B(z,r)\subset \mathbb{G}$ be an open ball and $E\subset B(z,r)$. Fix parameters $v,r, \rho, \theta >0$ such that $8m\theta<\rho r v$. Let $b\colon \mathbb{G} \to \mathbb{R}$ be a 1-Lipschitz bump function satisfying $b(0)=v$ and supported in $B(0,1)$. Suppose there exists a continuous function $F \colon \overline{B}(z,r) \to \mathbb{R}$ such that:
\begin{itemize}
\item $|F(x)|\leq \theta$ for each $x\in B(z,r)$.
\item The horizontal gradient $\nabla_{H} F(x)\in \mathbb{R}^{m}$ exists at each $x\in B(z,r)$, and $|\nabla_{H} F(x)|>\rho$ for each $x\in B(z,r)\setminus E$.
\end{itemize}
Then every Lipschitz function $h\colon \mathbb{G}\to \mathbb{R}$ is Pansu subdifferentiable at a point of $E$.
\end{lemma}
\begin{proof}
Let $\widetilde{b}(x):=b(\delta_{1/r}(z^{-1}x))$. Clearly $\widetilde{b}(z)=v$ and $\widetilde{b}$ is $(1/r)$-Lipschitz and supported in $B(z,r)$. Since $8m\theta<\rho r v$, we can choose $\eta$ satisfying $4\theta/v<\eta <\rho r/2m$. Define
\[G(x):=F(x)-\eta \widetilde{b}(x).\]
If $x\in \partial B(z,r)$ then $\widetilde{b}(x)=0$. Using also $|F(x)|\leq \theta$, we see
\begin{equation}\label{ppp}
|G(x)|=|F(x)-\eta \widetilde{b}(x)|\leq \theta \quad \mbox{ for }x\in \partial B(z,r).
\end{equation}
Moreover,
\begin{equation}\label{pppp}
G(z)=F(z)-\eta \widetilde{b}(z)\leq \theta-\eta v\leq -3\theta.
\end{equation}
Let $h\colon \mathbb{G}\to \mathbb{R}$ be a Lipschitz function. We intend to prove that $h$ is Pansu subdifferentiable at a point of $E$, for which we may assume that $h$ is not identically zero on $B(z,r)$. Define
\[a:=\sup_{\overline{B}(z,r)} |h|, \qquad c:=\min\left\{\frac{\theta}{2a},\frac{\rho}{4 m \mathrm{Lip}(h)}\right\}.\]
Then $\widetilde{h}(x):= c h(x)$ satisfies
\begin{align}\label{ineq1}
&|\widetilde{h}(x)|\leq \theta/2 \quad \mbox{for } x\in \overline{B}(z,r),\\
\label{ineq2}
&\widetilde{h}\ \mbox{is}\ (\rho/4m)\mbox{-Lipschitz on}\ \overline{B}(z,r).
\end{align}
Let $ H(x):=G(x)+\widetilde{h}(x)$ for $x\in \overline{B}(z,r)$. By \eqref{ppp} and \eqref{ineq1}, we see
\begin{align}\label{p}
H(x)\geq -3\theta/2\quad \mbox{for } x\in\partial B(z,r).
\end{align}
Using \eqref{pppp} and \eqref{ineq1}, we get
\begin{align}\label{q}
H(z)\leq -5\theta/2.
\end{align}
Putting together \eqref{p} and \eqref{q} we infer that $H$ attains its minimum at a point $x_0\in B(z,r)$. By Proposition \ref{basicproperties}(6), $H$ is Pansu subdifferentiable at $x_0$. Since $G$ is Pansu differentiable at $x_0$, $\widetilde{h}$ is Pansu subdifferentiable at $x_0$.
We claim that $x_0\in E$. Suppose $x_0\notin E$. Then by the properties of $F$ in the statement of the lemma, we know $|\nabla_{H}F(x_0)|>\rho$. By the definition of the horizontal gradient, there is $1\leq i\leq m$ such that $|X_i F(x_0)|>\rho/m$. Without loss of generality we suppose $X_i F(x_0)<-\rho/m$: otherwise in what follows consider directional derivatives in direction $-X_{i}$ instead of in direction $X_{i}$. Since $\widetilde{b}$ is $(1/r)$-Lipschitz and $4\theta/v<\eta <\rho r/2m$, we have
\[|X_i(\eta \widetilde{b})(x_0)|\leq \mathrm{Lip}(\eta \widetilde{b})<\rho/2m.\]
Thus, since $G=F-\eta \widetilde{b}$,
\[X_iG(x_0)=X_iF(x_0)-X_i(\eta \widetilde{b})(x_0)<-\rho/2m.\]
To conclude, notice $H=G+\widetilde{h}$, $X_iG(x_0)<-\rho/2m$ and $\mathrm{Lip}(\widetilde{h})\leq \rho/4m$. This implies that $H$ does not attain its minimum at $x_0$, a contradiction. Hence $x_{0}\in E$, so $h$ is Pansu subdifferentiable at a point of $E$.
\end{proof}
\begin{theorem}\label{gradienttheorem}
Let $D\subset \mathbb{G}$ be an open set and $f\colon D\to \mathbb{R}$ be Pansu differentiable. Denote $g(x)=\nabla_{H} f(x)$ for $x\in D$ and suppose $G\subset \mathbb{R}^{m}$ is an open set such that $g^{-1}(G)\neq \varnothing$. Then the following statements hold:
\begin{enumerate}
\item $g^{-1}(G)$ is porous at none of its points.
\item If $T\subset \mathbb{G}$ is open and $T\cap g^{-1}(G)\neq \varnothing$, then $L(T\cap g^{-1}(G))$ has positive Lebesgue measure for every non-zero group linear $L~\colon~\mathbb{G}~\to~\mathbb{R}$. In particular, the one-dimensional Hausdorff measure of $T\cap g^{-1}(G)$ with respect to the CC metric is positive.
\item If $T\subset \mathbb{G}$ is open and $T\cap g^{-1}(G)\neq \varnothing$, then $T\cap g^{-1}(G)$ is not $\sigma$-porous.
\end{enumerate}
\end{theorem}
\begin{remark}\label{remark}
Properties (1)--(3) of Theorem \ref{gradienttheorem} hold if and only if the following is true:
\begin{enumerate}
\item[(4)] Suppose $a\in g^{-1}(G)$ and $B(z_{n},r_{n})$ is a sequence of open balls such that $z_{n}\to a$ and $r_{n}>cd_{c}(z_{n},a)$ for some $c>0$ and all $n$. Then there exists $n_{0}$ such that for all $n\geq n_{0}$:
\begin{enumerate}
\item[(a)] $L(g^{-1}(G)\cap B(z_{n},r_{n}))$ has positive Lebesgue measure for every non-zero group linear map $L\colon \mathbb{G}\to \mathbb{R}$.
\item[(b)] $g^{-1}(G)\cap B(z_{n},r_{n})$ is not $\sigma$-porous.
\end{enumerate}
\end{enumerate}
\end{remark}
\begin{proof}[Proof of Remark \ref{remark}]
We first assume (4) and prove (1)--(3). Suppose that $a\in g^{-1}(G)$, $z_{n}\to a$ and $r_{n}>cd(z_{n},a)$ for some fixed $c$ and every $n\in \mathbb{N}$. Then (4b) asserts that $g^{-1}(G)\cap B(z_n,r_n)$ is not $\sigma$-porous for all sufficiently large $n$, in particular it is non-empty. Hence $g^{-1}(G)$ cannot be porous at $a$. This proves (1).
Now let $T\subset\mathbb{G}$ be open and $T\cap g^{-1}(G)\neq \varnothing$. Choose $a\in T\cap g^{-1}(G)$, a sequence $z_{n}\to a$ with $z_{n}\neq a$ and let $r_{n}=d_{c}(z_n,a)/2$. Then (4) asserts that for sufficiently large $n$, $L(g^{-1}(G)\cap B(z_n,r_n))$ has positive Lebesgue measure for every group linear map $L$ and $g^{-1}(G)\cap B(z_n,r_n)$ is not $\sigma$-porous. Since $T$ is open, we have
\[ g^{-1}(G)\cap B(z_n,r_n) \subset g^{-1}(G)\cap T\]
for sufficiently large $n$. This yields (2) and (3). The statement about Hausdorff measure in (2) follows because group linear maps are Lipschitz, so the image of a set of one-dimensional Hausdorff measure zero would have one-dimensional Hausdorff measure zero.
We now assume (1)--(3) and prove (4). Suppose $a\in g^{-1}(G)$ and $B(z_{n},r_{n})$ is a sequence of open balls such that $z_{n}\to a$ and $r_{n}>cd_{c}(z_{n},a)$ for some $c>0$ and all $n$. By (1), $B(z_{n},r_{n})\cap g^{-1}(G)\neq \varnothing$ for all sufficiently large $n$. Properties (2) and (3) applied with $T=B(z_n,r_n)$ then give (4a) and (4b) respectively.
\end{proof}
\begin{proof}[Proof of Theorem \ref{gradienttheorem}]
Suppose $a, z_{n}, r_{n}, c$ are as in Remark \ref{remark}(4). Since making $r_{n}$ smaller makes the statement stronger, we may assume $r_{n}\to 0$. Choose a $1$-Lipschitz bump function $b\colon \mathbb{G} \to \mathbb{R}$ supported in $B(0,1)$ which satisfies $b(0)=v$ for some $v>0$. Since $G$ is open and $g(a)\in G$, there exists $\rho>0$ such that $|g(x)-g(a)|>\rho$ for any $x\not\in g^{-1}(G)$. Using Pansu differentiability of $f$ at $a$ and \eqref{gradform}, which expresses the Pansu derivative in terms of the horizontal gradient, we may find $\delta>0$ such that $d_{c}(x,a)<\delta$ implies
\[ |f(x)-f(a)-\langle g(a), p(a^{-1}x)\rangle |\leq \frac{\rho v d_{c}(x,a)}{16m(1+1/c)}.\]
Let
\[F(x):=f(x)-f(a)-\langle g(a), p(a^{-1}x)\rangle.\]
Choose $n_{0}$ such that $B(z_{n},r_{n})\subset B(a,\delta)$ for $n>n_{0}$. Fix $n>n_{0}$. For every $x\in B(z_{n},r_{n})$ we have, using $r_{n}>cd_{c}(z_{n},a)$ as in Remark \ref{remark}(4),
\begin{align*}
|F(x)| \leq \frac{\rho v d_{c}(x,a)}{16m(1+1/c)} &\leq \frac{\rho v(r_{n}+d_{c}(z_{n},a))}{16m(1+1/c)}\\
&< \frac{\rho v(r_{n}+r_{n}/c)}{16m(1+1/c)}\\
&=\frac{\rho vr_{n}}{16m}.
\end{align*}
For $x\notin g^{-1}(G)$ we have $|\nabla_{H} F(x)|=|g(x)-g(a)|>\rho$. Now the assumptions of Lemma \ref{usefullemma} hold with $z=z_{n}, r=r_{n}$, $\theta=\rho vr_{n}/16m$ and $E=B(z_{n},r_{n})\cap g^{-1}(G)$. Hence every real-valued Lipschitz map on $\mathbb{G}$ is Pansu subdifferentiable at a point of $E$.
To prove Remark \ref{remark}(4a), suppose $L(E)$ has measure zero for some non-zero group linear map $L\colon \mathbb{G}\to \mathbb{R}$. Using Lemma \ref{nullnonsubdiff} we can choose a Lipschitz function $H\colon \mathbb{R} \to \mathbb{R}$ which is subdifferentiable at no point of $L(E)$. Hence $h=H\circ L$ is a Lipschitz function which, by Proposition \ref{basicproperties}(5), is not Pansu subdifferentiable at any point of $E$. This contradicts Lemma \ref{usefullemma}.
To prove Remark \ref{remark}(4b), suppose $E$ is $\sigma$-porous. Then by Theorem \ref{nonsubdiff} there exists a Lipschitz function which is Pansu subdifferentiable at no point of $E$. This again contradicts Lemma \ref{usefullemma}.
\end{proof}
| 640f023529b6e7a0741e62417ddee80d262b190d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}\vspace{-0.2cm}
We study asynchronous parallel block-descent methods for the following class of nonconvex nonsmooth minimization problems: \vspace{-0.2cm}
\begin{equation}
\begin{array}{rl}
\underset{\mathbf{x}\triangleq (\mathbf{x}_1,\ldots, \mathbf{x}_N)}{\min} & F(\mathbf{x}) \triangleq f(\mathbf{x}) + \sum\limits_{i=1}^N g_i({\mathbf{x}}_i)\\[0.3em]
&\mathbf{x}_i \in \mathcal X_i, \qquad i = 1, \dots , N,
\end{array}\tag{P}\vspace{-0.1cm}
\label{eq:optimization_problem}
\end{equation}
where $f$ is a {smooth, possibly nonconvex function, $g_i$ are possibly nonsmooth, convex functions, and $\mathcal{X}_i\subseteq \Re^{n_i}$ is a closed, possibly nonconvex set. }
Instances of Problem \eqref{eq:optimization_problem} arise in
many fields, including compressed sensing, machine learning, data mining, and genomics,
just to name a few.
Typically, in data-driven applications $f$ might measure the misfit between the observations and the postulated model, parametrized on ${\mathbf{x}}$, while the regularizers $g_i$ encode structural constraints on the solution, such as sparsity.
Many of the aforementioned applications give rise to extremely
large-scale problems, which naturally call for \emph{asynchronous, parallel} solution methods. In fact, well suited to
modern
computational architectures, asynchronous methods reduce the idle times of workers,
mitigate communication and/or memory-access congestion, and make
algorithms
more fault-tolerant.
In this paper, we introduce a general asynchronous block-descent algorithm for finding stationary solutions of Problem \eqref{eq:optimization_problem}.
We consider a generic multi-worker architecture (e.g., shared memory system, message passing-based system, cluster computer, cloud federation) wherein multiple workers,
continuously
and without coordination with each other, update a block-variable by solving a strongly convex block-model of Problem \eqref{eq:optimization_problem}. More specifically, at iteration $k$, a worker updates a block-variable ${\mathbf{x}}^k_{i^k}$ of $\mathbf{x}^k$ to ${\mathbf{x}}^{k+1}_{i^k}$, with $i^k$ in the set $\mathcal N\triangleq \{1,\ldots, N\}$, thus generating the vector ${\mathbf{x}}^{k+1}$.
When updating block $i^k$, in general, the worker does not have access to the current vector ${\mathbf{x}}^k$, but it will use instead the
local estimate
${\mathbf{x}}^{k-\mathbf{d}^k} \triangleq (x_1^{k-d_1^k}, x_2^{k-d^k_2}, \ldots, x_N^{k-d^k_N})$, where $\mathbf{d}^k \triangleq (d_1^k, d_2^k, \ldots, d_N^k)$ is the
``vector of
delays'', whose components $d_i^k$ are
nonnegative integers. Note that ${\mathbf{x}}^{k-
\mathbf{d}^k}$ is
nothing else but a combination of delayed,
block-variables.
The way each worker forms its own estimate ${\mathbf{x}}^{k-\mathbf{d}^k}$ depends on the particular
architecture under consideration
and it is
immaterial to the {analysis} of the algorithm.
We only observe here that if all delays ${d}^k_i$ are zeros, the model reduces to a standard synchronous one. \newline
Given ${\mathbf{x}}^{k-\mathbf{d}^k}$ and $i^k$, block ${\mathbf{x}}^k_{i^k}$ is updated
by solving the following {\em strongly
convex} block-approximation of Problem~\eqref{eq:optimization_problem}: \vspace{-0.1cm}
\begin{equation}\label{eq:local model with ik}
\hat {\mathbf{x}}_{i^k}({\mathbf{x}}^{k-\mathbf{d}^k})
\triangleq \mathop{\rm argmin}_{{\mathbf{x}}_{i^k} \in \tilde{\mathcal{X}}_{i^k}({\mathbf{x}}^{k-\mathbf{d}^k}) }
\tilde f_{i^k}({\mathbf{x}}_{i^k};
{{\mathbf{x}}}^{k-\mathbf{d}^k}) + g_{i^k}({\mathbf{x}}_{i^k}),\vspace{-0.1cm}
\end{equation}
and then setting\vspace{-0.3cm}
\begin{equation}
{\mathbf{x}}^{k+1}_{i^k} = {\mathbf{x}}^{k}_{i^k} + \gamma\, \left(\hat {\mathbf{x}}_{i^k}({\mathbf{x}}^{k-\mathbf{d}^k}) -
{\mathbf{x}}^{k}_{i^k}\right).
\label{update} \vspace{-0.1cm}
\end{equation}
In \eqref{eq:local model with ik}, $\tilde f_{i^k}$ and $\tilde{\mathcal{X}}_{i^k}$ represent a
strongly convex surrogate of $f$ and a convex set obtained replacing the nonconvex functions
defining $\mathcal{X}_{i^k}$ by suitably chosen upper convex approximations, respectively;
both $\tilde f_{i^k}$ and $\tilde{\mathcal{X}}_{i^k}$ are built using the out-of-sync information $
{\mathbf{x}}^{k-\mathbf{d}^k}$. {If the set $\mathcal{X}_{i^k}$ is convex, then we will
always take $\tilde{\mathcal{X}}_{i^k}=\mathcal{X}_{i^k}$.}
In (\ref{update}), $\gamma\in (0,1]$ is the stepsize. Note that, in the above asynchronous
model, the worker that is in charge of the computation \eqref{eq:local model with ik} and the
consequent update (\ref{update}) is immaterial.
\smallskip
\noindent \textbf{Major contributions:}
{Our main contributions are:}
\smallskip
\noindent 1. {\em A new probabilistic model for asynchrony fixing some unresolved issues}:
Almost all modern asynchronous algorithms for convex and nonconvex problems are modeled in
a probabilistic way.
We put forth a novel probabilistic model describing the statistics of $(i^k,\mathbf{d}^k)$ that differs markedly from existing ones.
This new model allows us not only to fix some important theoretical issues that mar most of the papers
in the field (see discussion below on related work), but it also lets us analyze for the first time in a
sound way several {practically used and effective} computing settings and new
asynchronous algorithms. For instance, it is
widely accepted that in shared-memory systems, the best performance are obtained by first
partitioning the variables among cores, and then letting each core update in an asynchronous fashion their own
block-variables, according to some randomized cyclic rule.
To the best of our knowledge, this is the first work proving convergence of such
practically effective methods in an asynchronous setting.
\smallskip
\noindent 2. {\em The ability to effectively deal with nonconvex constraints}: All the works in the literature but \cite{davis2016asynchronous,DavisEdmundsUdell} can deal only with unconstrained or convex constrained problems.
On the other hand, the algorithms in \cite{davis2016asynchronous,DavisEdmundsUdell} require at each iteration the computation of the global optimal solution of nonconvex subproblems, which, except in few special cases, can be as difficult as solving the original nonconvex problem.
Our method is the first asynchronous method that allows one to deal (under adequate assumptions)
with nonconvex constraints while solving only strongly convex subproblems.\smallskip
\noindent 3. {\em The possibility to leverage potentially complex, but effective subproblems
\eqref{eq:local model with ik}}:
Asynchronous methods so far are all built around a proximal linearization method, which
corresponds, in our framework, to setting
\[
\tilde f_{i}({\mathbf{x}}_{i^k};{{\mathbf{x}}}^{k-\mathbf{d}^k}) = \nabla_{{\mathbf{x}}_{i^k}} f\left({{\mathbf{x}}}^{k-\mathbf{d}^k}\right)^\text{T}
\left({\mathbf{x}}_{i^k} -
{\mathbf{x}}_{i^k}^{k-d_{i^k}^k}\right) + \beta \,\| {\mathbf{x}}_{i^k} - {\mathbf{x}}_{i^k}^{k-d_{i^k}^k}\|^2_2,
\]
for some constant $\beta>0$.
This choice often leads to efficient solution methods and, in some cases,
even to subproblems that admit a solution in closed-form. For instance, it has been shown to be very efficient on composite
quadratic problems, like LASSO. However, moving to more nonlinear
problems, one may want to use more complex/higher order models. In fact, the more sophisticated the subproblem \eqref{eq:local model with ik}, the better the
overall behavior of
the algorithm (at least in terms of iterations) is. This happens at the price of computationally more expensive
subproblems. But in asynchronous and distributed methods, the bottleneck
is often given by the communication cost. In
these cases, it might be desirable to reduce the communication overhead at the price of more complex subproblems to solve. Furthermore, there are many application for which one can define subproblems
that, while not being proximal linearizations, still admit closed-form solutions (see. e.g., \cite{scutari2014decomposition,daneshmand2015hybrid}). Overall, the ability to use more complex subproblems is an additional degree of freedom
that one may want to exploit to improve the performance of the algorithm.\vspace{0.1cm} \newline
\noindent 4. {\em Almost sure convergence and complexity analysis}: We prove i) almost sure convergence to stationary solutions of Problem \eqref{eq:optimization_problem}; and ii) convergence to $\epsilon$-stationary solutions in an $\mathcal O(\epsilon^{-1})$ number of iterations.
We remark that our convergence results match similar ones in the literature \cite{liu2015asynchronous,liu2015asyspcd,davis2016asynchronous,DavisEdmundsUdell,peng2016arock}, which however were obtained in a simplified setting (e.g., only for unconstrained or convex constrained problems) and under unrealistic probabilistic assumptions on the pair index-delay $(i^k,\mathbf{d}^k)$ (see discussion on related work). Our analysis builds on an induction technique based on our probabilistic model and a novel Lyapunov function that properly combines variable dynamics and their delayed versions.
\vspace{0.01cm} \newline
\noindent 5. {\em A theoretical almost linear speed-up for a wide range of number of cores}: The holy grail of asynchronous methods is the ideal linear speed-up (with respect to the number of workers). This theoretical limit is not achievable in practice; in fact, as the number of workers increases, the effective speedup is always limited by associated overheads (communication costs, conflicts, etc.), which make the linear growth impossible to achieve for arbitrarily large number of workers.
By using the number of iterations needed to achieve an $\epsilon$-stationary solution as a proxy for the computational time and leveraging our new Lyapunov function, we are able to show almost linear speed-up
in many settings of practical interest.
This is the first theoretical result on speedup, based on a realistic probabilistic model for asynchrony (see discussion in contribution 1).
\noindent
{\bf Related work.} Although asynchronous block-methods have a long history (see, e.g.,
\cite{baudet1978asynchronous,chazan1969chaotic, Bertsekas_Book-Parallel-Comp, frommer2000asynchronous,tseng1991rate}), their revival and probabilistic analysis
have taken place only in recent years; this is mainly due to the
current trend towards huge scale
optimization and the availability
of ever more complex computational architectures that call for efficient and resilient algorithms.
Indeed, asynchronous parallelism has been applied to many state-of-the-art optimization algorithms,
including stochastic gradient methods
\cite{nedic2001distributed,Hogwild!,lian2015asynchronous,huo2016asynchronous,
Mania_et_al_stochastic_asy2016,leblond2017Asaga,Pedregosa2017breaking} and ADMM-like schemes
\cite{hong2014distributed,wei20131,iutzeler2013asynchronous}.
Block-Coordinate Descent (BCD) methods are part of the folklore in optimization;
more recently, they
have been proven to be particularly effective in solving
very large-scale problems arising, e.g., from data-intensive applications. Their asynchronous counterpart has been introduced and studied in the seminal work
\cite{liu2015asyspcd}, which motivated and oriented much of subsequent research in the field, see e.g.
\cite{liu2015asynchronous,davis2016asynchronous,DavisEdmundsUdell,peng2016arock,peng2016convergence}.
We refer the interested reader to
{\cite{Wright2015} and references therein for a detailed overview of BCD methods.
There are several differences between the above methods and the framework proposed in this
paper, as detailed next.
\newline\noindent $\bullet$ \textit{On the probabilistic model:} All current probabilistic models for asynchronous BCD methods are
based on the (implicit or explicit) assumption that the random variables $i^k$ and $\mathbf{d}^k$ are {\em independent};
this greatly simplifies the convergence analysis.
However, in reality there is a strong dependence of the delays $\mathbf{d}^k$ on the updated block $i^k$. Consider the setting where the variables are partitioned among two workers and each worker updates only its own block-variables; let $\mathcal N_1$ and $\mathcal N_2$ be the index set of the blocks controlled by worker $1$ and $2$, respectively. It is clear that in the updates of worker $1$ it will always be $d_i^k=0$, for all $i\in \mathcal N_1$ and $k$, while (at least some) delays $d_i^k$ associated with the blocks $i\in \mathcal N_2$ will be positive; the opposite happens to worker two. The independence assumption is unrealistic also in settings where all the workers share all the variables. Blocks that are updated less frequently than others, when updated, will have larger associated delays. This happens, for instance, in problems where i) some blocks are more expensive to update than others, because they are larger, bear more nonzero entries, or data retrieval requires longer times; or ii) the updates are carried by heterogeneous workers (e.g., some are faster or busier than others).
We tested this assumption, performing an asynchronous algorithm on two different architectures and measuring the average delay corresponding to different blocks updated. The experiments were performed on a shared-memory system with 10 cores of an Intel E5-2699Av4 processor. An asynchronous algorithm was applied to a LASSO problem {\cite{tibshirani1996regression}} with 10000 variables, partitioned uniformly into 100 contiguous blocks; the Hessian matrix was generated with high sparsity on several rows.
All the cores can update any block, selected uniformly at random. {We found that} blocks associated with the sparse rows of the Hessian have delays $\mathbf{d}^k$ with components between 0 and 3, while the delays of the other blocks were all bigger than 20.
Even when the computing environment is homogeneous and/or the block updates have the same cost, the aforementioned dependence persists. We simulated a message-passing system on Purdue Community Cluster Snyder; we used two nodes of the cluster, each of them equipped with 10 cores of an Intel Xeon-E5 processors and its own shared memory. Every node can update every block, selected uniformly at random. We ran an asynchronous algorithm on the same LASSO problem described above but now with a dense Hessian matrix. The blocks updated by node 1 have an average delay of 12 while those updated by node 2 experience an average delay of 22.
This can be due to several uncontrollable factors, like operation system and memory schedulers, buses controllers, etc., which are hard to rigorously model and analyze.
Another unrealistic assumption often made in the literature \cite{Hogwild!,liu2015asynchronous,liu2015asyspcd,davis2016asynchronous}
is that the block-indices $i^k$ are selected \emph{uniformly} at random. While this assumption simplifies the
convergence analysis, it limits the applicability of the model; see Examples 4 and 5 in Section~\ref{sec:Examples}.
In a nutshell, this assumption may be
satisfied only if all workers have
the same computational power and have access to all variables.
We conclude the discussion on probabilistic models underlying asynchronous
algorithms mentioning the line of work dealing with stochastic gradient
methods. Stochastic gradient methods are similar to block-descent approaches in that at each
iteration sampling is performed to determine the nature of the update, but sampling is
done among functions in an optimization problem minimizing the sum of functions,
as opposed to block variables.
A related, albeit different, issue of
independence in the probabilistic models used in stochastic gradient methods was first noted in the technical
report \cite{Mania_et_al_stochastic_asy2016}, see also \cite{leblond2017Asaga,Pedregosa2017breaking} for
further developments. These papers circumvent the issue by enforcing independence
(a) using a particular manner of
labeling iterations as well as
(b) reading the entire vector of variables regardless of the sparsity
pattern among the summand functions in the objective.
However, the analysis in \cite{Mania_et_al_stochastic_asy2016,leblond2017Asaga,Pedregosa2017breaking} is (c) only performed in the context of strongly
convex unconstrained problems, (d) involves uniform sampling and (e) is only applicable
for the shared memory setting.
Thus, while the analysis and procedures described in the references above are interesting,
on the whole requirements (b)-(e) make these proposals
of marginal interest in the context of block-descent methods
(even assuming they can actually be adapted to our setting).
\begin{comment}
{We conclude the discussion on probabilistic models underlying asynchronous algorithms mentioning the line of work dealing with stochastic gradient methods. A related, albeit different, issue of independence in the probabilistic models used therein was first noted in the technical report
\cite{Mania_et_al_stochastic_asy2016}, see also \cite{leblond2017Asaga,Pedregosa2017breaking} for further
developments.
The approach in these papers is to modify the algorithm to \emph{enforce} the ``desired independence'', by (a)
changing the way iterations are labeled and (b) by forcing the reading of (at least) the \emph{entire} vector of optimization variables at each iteration.
Furthermore, even assuming the approach in \cite{Mania_et_al_stochastic_asy2016,leblond2017Asaga,Pedregosa2017breaking} can be adapted to our BCD setting, to prove convergence, they would require that (c) the indices $i^k$ are selected uniformly at random; (d) a shared memory system is considered; and
(e) the optimization problem is strongly convex and unconstrained. On the whole, these requirements make the very interesting proposals in
\cite{Mania_et_al_stochastic_asy2016, leblond2017Asaga,Pedregosa2017breaking} not practical and of marginal interest in the context of block-descent methods.}
\end{comment}
\newline\indent Differently from the aforementioned works, our more general and sophisticated probabilistic model neither postulates the independence between $i^k$ and $\mathbf{d}^k$ nor requires artificial changes in the algorithm [e.g., costly unnecessary readings, as in (b)] to enforce it; it handles instead the potential dependency among variables directly. By doing so, one can establish convergence without
requiring any of the restrictive conditions (b)-(e), and significantly enlarge the class of computational architecture falling within the model [e.g., going beyond (d) and (e)]$-$see Section~\ref{sec:Examples} for several examples.
{The necessity of
a new probabilistic model of asynchrony in BCD methods was first observed in our conference works \cite{cannelli2016parallel,cannelli2017asynchronous} while the foundations of our approach were
presented in our technical reports \cite{companion2,companion} along with some numerical results. %
Here we improve the analysis of \cite{companion2,companion} by relaxing
considerably the assumptions for convergence and tightening the complexity bounds. }\newline
\noindent $\bullet$ \textit{Nonconvex constraints}: Another important feature of our algorithm is the ability to handle nonconvex objective functions
and nonconvex constraints by an algorithm that only needs to solve, at each iteration, a strongly convex optimization subproblem. Almost all asynchronous methods cited above can handle only convex
optimization problems or, in
the case of fixed point problems, nonexpansive mappings.
The exceptions are \cite{lian2015asynchronous,yun2014nomad} and, more relevant to our setting,
\cite{davis2016asynchronous,DavisEdmundsUdell} that
study
unconstrained and constrained nonconvex optimization problems, respectively. However, the papers dealing with constrained problems, i.e.
\cite{davis2016asynchronous,DavisEdmundsUdell}, propose algorithms that require, at each iteration, the global
solution of nonconvex subproblems.
Except for few cases,
the subproblems could be hard to solve and potentially as difficult as the original one.
\newline
\noindent $\bullet$ \textit{Successive Convex Approximation}:
All the asynchronous algorithms described so far use proximal linearization to define
subproblems. As already pointed out, this is the first paper where subproblem models able to capture more structure of the objective functions are
considered. This offers more freedom and flexibility to tailor the minimization algorithm to the problem
structure, in order to obtain more efficient solution methods.
\newline \textbf{Notation}: We use the following notation for random variables and their realizations: underlined
symbols denote
random variables, e.g., $\underline{\mathbf{x}}^k$, $\mathbf{x}^{k-\underline{\mathbf{d}}^k}$, whereas the
same symbols with no underline
are the corresponding realizations.\vspace{-0.6cm}
\section{Asynchronous Algorithmic Framework}
\label{sec:model}\vspace{-0.2cm}
In this section we introduce the assumptions on Problem \eqref{eq:optimization_problem} along with the formal description of the proposed algorithm. For simplicity of presentation, we begin studying \eqref{eq:optimization_problem}
assuming that there are only convex constraints, i.e., all ${\cal X}_i$ are convex.
This unnecessary
assumption will be removed in Section \ref{nonconvex}.\smallskip
\noindent \textbf{Assumption A (On Problem (\ref{eq:optimization_problem})).}\vspace{-0.1cm}
\begin{description}
\item[(A1)] Each set $\mathcal{X}_i\subseteq\mathbb{R}^{n_i}$ is nonempty, closed, and convex;\smallskip
\item[(A2)] $f:\mathcal{O}\rightarrow\mathbb{R}$ is $C^1$, where $\mathcal{O}$ is an open set containing $\mathcal{X}\triangleq \mathcal{X}_1\times \cdots \times \mathcal{X}_N$;\smallskip
\item[(A3)] $\nabla_{\mathbf{x}_i} f$ is $L_{f}$-Lipschitz continuous on $\mathcal{X}$;
\smallskip
\item[(A4)] Each $g_i:\mathcal{O}_i\rightarrow\mathbb{R}$ is convex, possibly nonsmooth, and $L_g$-Lipschitz continuous on $\mathcal{X}_i$, where $\mathcal{O}_i$ is an open set containing $\mathcal{X}_i$;
\smallskip
\item[(A5)] $F$ is coercive on $\mathcal{X}$, i.e., $\underset{ \mathbf{x} \in \mathcal{X}, \|\mathbf{x}\| \to \infty}{\lim}F(\mathbf{x})=+\infty$.
\end{description}
\noindent
These assumptions are rather standard. For example, A3 holds trivially if $\mathcal{X}$ is bounded and $\nabla f$ is locally Lipschitz. We remark that in most practical cases the
$g_i$'s are norms or polyhedral functions and A4 is readily satisfied. Finally, A5 guarantees the existence of a solution.
We introduce now our algorithmic asynchronous framework. The asynchronous iterations performed by the workers are given in \eqref{eq:local model with ik} and \eqref{update} [cf.~Section \ref{sec:intro}]. However, the analysis of the algorithm based directly on \eqref{eq:local model with ik}-\eqref{update} is not a simple task.
The key idea is then
to introduce a ``global view'' of \eqref{eq:local model with ik}-\eqref{update}
that captures through a unified, general, probabilistic model several specific computational architectures/systems and asynchronous modus operandi.
The iteration $k\to k+1$ is triggered when a block-component $i^k$ of the current ${\mathbf{x}}^k$ is updated by some worker using (possibly) delayed information ${\mathbf{x}}^{k-\d^k}$, thus generating the new vector ${\mathbf{x}}^{k+1}$. Note that, in this model, the worker that performs the update is immaterial. Given \eqref{eq:local model with ik} and \eqref{update}, it is clear that the update ${\mathbf{x}}^k\to {\mathbf{x}}^{k+1}$ is fully determined once $i^k$ and $\d^k$ are specified. In several asynchronous methods, the index $i^k$ is chosen randomly. Even when this is not the case, the values of $i^k$ and $\d^k$ are difficult to preview beforehand, because they depend on several factors which are hard to model mathematically, such as the computational architecture, the specific hardware, the communication protocol employed by the workers, possible hardware failures, etc..
Therefore, we model the sequence of pairs $\{(i^k, \d^k)\}_{k\in \mathbb N_+}$ generated by the algorithmic process as a realization of a stochastic process; the probabilistic space associated to this stochastic process is formally introduced in Section \ref{sec:probabilistic}. The proposed general asynchronous model is summarized in
Algorithm~\ref{alg:global}, which we term Asynchronous FLexible ParallEl Algorithm (AsyFLEXA).
\begin{algorithm}[t]
\caption{Asynchronous FLexible ParallEl Algorithm (AsyFLEXA)}
\label{alg:global}
\begin{algorithmic}
\STATE{\textbf{Initialization:}} $k=0$, $\mathbf{x}^0\in\mathcal{X}$, $\gamma\in(0;1]$.
\WHILE{a termination criterion is not met}
\STATE{\texttt{(S.1)} The random variable $(\underline{i}^k,\underline{\mathbf{d}}^k)$ is realized as $(i^k,
\mathbf{d}^k)$;}
\STATE{\texttt{(S.2)}} $\hat{\mathbf{x}}_{i^k}(\mathbf{x}^{k-\mathbf{d}^k})$ is computed: \vspace{-0.2cm}
\begin{equation}\label{eq:best-response-convex}
\hat{{\mathbf{x}}}_{i^k}({\mathbf{x}}^{k-\mathbf{d}^k})
\triangleq \mathop{\rm argmin}_{{\mathbf{x}}_{i^k} \in {\mathcal{X}}_{i^k}}
\tilde f_{i^k}({\mathbf{x}}_{i^k};{\mathbf{x}}^{k-\mathbf{d}^k}) + g_{i^k}({\mathbf{x}}_{i^k}),\vspace{-0.2cm}
\end{equation}
\STATE{\texttt{(S.3)}} $\mathbf{x}^k_{i^k}$ is acquired;
\STATE{\texttt{(S.4)} The block $i^k$ is updated:\vspace{-0.2cm}
\begin{equation}\label{update-Algo1}\mathbf{x}_i^{k+1}=\begin{cases}\mathbf{x}^k_i+\gamma(\hat{\mathbf{x}}_i
(\mathbf{x}^{k-\mathbf{d}^k})-\mathbf{x}_i^k),&\text{if }i=i^k\\
\mathbf{x}_i^k&\text{if }i\neq i^k\end{cases}\vspace{-0.2cm}\end{equation}}
\STATE{\texttt{(S.5)} $k \leftarrow k+1;$} \ENDWHILE
\end{algorithmic}
\end{algorithm}
\noindent \textbf{Discussion on Algorithm~\ref{alg:global}.} Several comments are in order.
\begin{asparaenum} \item \emph{On the generality of the model:} Algorithm~\ref{alg:global} represents a gamut of asynchronous schemes and architectures, all captured in an abstract and unified way by the stochastic process modeling the specific mechanism of generation of the delay vectors $\mathbf{d}^k$ and indices $i^k$ of the blocks to updates.
For concreteness, we show next how Algorithm~\ref{alg:global} customizes when modeling asynchrony in shared-memory and message passing-based architectures. \smallskip
\noindent\textit{Example 1: Shared-memory systems.} Consider a shared-memory system wherein multiple cores update in an asynchronous fashion blocks of the vector ${\mathbf{x}}$, stored in a shared memory. An iteration $k \to k+1$ of Algorithm~\ref{alg:global} is triggered when a core writes the (block) update ${\mathbf{x}}_{i^k}^{k+1}$ in the shared memory (Step 4). Note that the cores need not know the global iteration index $k$. No memory lock is assumed, implying that components
of the variables may be written by some cores while other components are simultaneously
read by others. This \emph{inconsistent read} produces vectors ${\mathbf{x}}^{k-\mathbf{d}^k} = (x_1^{k-d_1^k}, x_2^{k-d^k_2}, \ldots, x_N^{k-d^k_N})$, to be used in the computation of $\hat{\mathbf{x}}_{i^k}$ (Step 2), whose (block) component ${\mathbf{x}}_i^{k-d_i^k}$ is a (possibly) delayed version of block $i$ read by the core that is going to perform the update.
Note that, while ${\mathbf{x}}_i^{k-d_i^k}$ existed in the shared memory at some point in time, the entire delayed vector ${\mathbf{x}}^{k-\mathbf{d}^k}$ might have not at any time. Also, in Step 4, it is tacitly assumed that the update of a block is {\em atomic} (the block is written in the shared memory as a whole) and while a core is writing that block no other core can modify the same block. This is minor requirement, which can be easily enforced in modern architectures either by a block-coordinate look or using a dual-memory writing approach, see \cite[Section 1.2.1]{peng2016arock}.
\begin{figure}
\begin{center}
\includegraphics[width=11cm]{figure-cores}
\end{center}\vspace{-0.2cm}
\caption{AsyFLEXA modeling block asynchronous updates in a shared-memory system: three cores, vector variables ${\mathbf{x}}\in\mathbb{R}^3$, scalar blocks ($n_i=1$, for all $i$)
}
\label{fig:inconsistent_read}\vspace{-0.4cm}
\end{figure}
Figure~\ref{fig:inconsistent_read} shows few iterations of the algorithm dynamics in the asynchronous setting described above. The (continuous) time when operations (reading, writing, computation) are performed is indicated in the top horizontal axes whereas the global (discrete) iteration counter is reported in the bottom axes.
The asynchronous updates happen as follows. At iteration $k=3$, Core $3$ writes $x_1^3$; therefore, ${\mathbf{x}}^3$ differs from ${\mathbf{x}}^2$
in the first component. Core $2$ locks $x_3^2$ to quickly read it and perform the linear combination with $\hat x_3({\mathbf{x}}^{3-\mathbf{d}^3})$ [cf. (\ref{update-Algo1})], and updates $x_3$; therefore ${\mathbf{x}}^4$ differs from ${\mathbf{x}}^3$ in just the 3rd component. Note that core $2$ reads $x^2_3$ which is equal to $x^3_3$, so $d_3^3=0$; this is because between the lock and the writing of core $2$, no other cores wrote $x_3$ (core $3$ updates $x_1$). At iteration $k=5$, core $3$ writes $x_2^5$. In this case ${\mathbf{x}}^{4-\mathbf{d}^4}$, used to compute $x^5_2 = (1-\gamma)x^4_2+\gamma\hat x_2({\mathbf{x}}^{4-\mathbf{d}^4})$, is exactly equal to ${\mathbf{x}}^4$, since core $3$ reads the vector entirely after the last update, so $\mathbf{d}^4=\mathbf{0}$. A different situation happens when iteration $k=6$ is triggered by core $1$: the vector used by the core to perform its update is ${{\mathbf{x}}}^{5-\mathbf{d}^5}$, which is such that
$x^{5-\mathbf{d}^5}_1 = x^2_1 \neq x^3_1$, $x^{5-\mathbf{d}^5}_2 = x^2_2$, and $x^{5-\mathbf{d}^5}_3 = x^4_3\neq x^2_3$; therefore, ${\mathbf{x}}^{5-\mathbf{d}^5}$ never existed in the shared memory at any time. It can be seen that the vector of delays at $k=6$ reads $\d^5 = (3,1,0)^{\text{T}}$.
Note that the delay vector $\mathbf{d}^k$ used at a given iteration may not be unique: different values for the components $d_i^k$ may produce the same delayed vector $\mathbf{x}^{k-\mathbf{d}^k}$. For instance, in the example above, the vector $(2,1,0)^{\text{T}}$ could have been used in place of $\mathbf{d}^5$.
\smallskip
\noindent\textit{Example 2: Message passing systems.} Consider a message passing-based system: multiple computational units (e.g., clouds, cluster computers) are connected through a (directed) graph, modeling the communication pattern among the units. Suppose that every worker has in charge the update of a set of block variables, partitioned among all the workers. In this setting, Algorithm~\ref{alg:global} still models asynchronous updates and communications. There is no shared memory; every worker updates its own variables writing its own local memory, and then broadcasts its updates to its neighbors, according to a given protocol, which specifies, for instance, how ofter the communications will happen and with whom. In this setting, ${\mathbf{x}}^{k-\d^k}$ corresponds to the most recent information a worker has received from the others at the time of its update.\smallskip
\item \textit{On the surrogate functions $\tilde f_{i}$.}
A degree of freedom offered by the proposed framework is the choice of the surrogate function used in the subproblems (\ref{eq:best-response-convex}) solved by the workers at each iteration. We consider the following general class of surrogate functions (we denote by $\nabla \tilde f_{i}$
the partial gradient of $\tilde f_i$ with respect to the first argument).
\noindent \textbf{Assumption B (On the surrogate functions $\tilde{f}_i$'s).} Given $\tilde f_{i}: \mathcal{X}_i\times \mathcal{X}\to \mathbb{R}$, with $i\in\mathcal{N}$, we assume:
\begin{description}
\item[ (B1)] $\tilde f_{i} (\bullet; \mathbf{y})$ is $C^1$ on an open set containing $\mathcal{X}_i$, and $c_{\tilde{f}}$-strongly
convex
on $ \mathcal{X}_i$, for all $\mathbf{y}\in \mathcal{X}$
\smallskip
\item[ (B2)] $\nabla \tilde f_{i} (\mathbf{y}_i;\mathbf{y}) = \nabla_{\mathbf{y}_i} f(\mathbf{y})$, for all $\mathbf{y} \in \mathcal{X}$;
\smallskip
\item[ (B3)] $\nabla \tilde f_{i} (\mathbf{y}_i;\bullet)$
is $L_B$-Lipschitz continuous
on $ \mathcal{X}$,
for all $\mathbf{y}_i \in \mathcal{X}_i$;\smallski
\item [ (B4)]$\nabla \tilde f_{i} (\bullet;\mathbf{y})$
is $L_E$-Lipschitz continuous
on $ \mathcal{X}_i$, for all $\mathbf{y} \in \mathcal{X}$.
\end{description}\smallskip
\noindent The surrogate $\tilde f_i(\bullet;\mathbf{x}^k)$ should be regarded as a (simple) strongly convex
local approximation of $f$ around $\mathbf{x}^k\in \mathcal X$, that
preserves the first order properties of $f$.
Finding a surrogate $\tilde f_i$ that satisfies Assumption B is in general non difficult; in
any case, one can always choose $\tilde f_{i} (\mathbf{x}_i; \mathbf{x}^k)
= \nabla_{{\mathbf{x}}_i} f(\mathbf{x}^k)^\text{T}({\mathbf{x}}_i -\mathbf{x}_i^k) + \beta\| {\mathbf{x}}_i - \mathbf{x}_i^k\|^2_2$, where $\beta$ is a
positive constant, which leads to the classical proximal-gradient update. However, having the possibility to use a different $\tilde f_i$ may be
useful to exploit some potential structure in the problem; of course, a trade-off is expected: the more complex the
$\tilde f_i$, the more information will be retained in $\hat {\mathbf{x}}_i(\mathbf{x}^k) $, but also the more intensive its computation is
expected to be. On the other hand, the solution of more complex subproblems will in general decrease the number of
information exchanges in the system, which may be a key advantage in many applications.
Some valid instances of $\tilde f_i$'s going beyond the proximal-gradient choice are discussed next; we refer the interested reader to \cite{FLEXA,facchinei2016feasible} for further examples.
\smallskip
\noindent $\bullet$ If $f(\mathbf{x}_1,\ldots,\mathbf{x}_N)$ is block-wise uniformly convex, instead of linearizing $f$
one can exploit a second-order approximation and set $\tilde{f}_i(\mathbf{x}_i;
\mathbf{x}^k)=f(\mathbf{x}^k)+\nabla_{\mathbf{x}_i}f(\mathbf{x}^k)^\text{T}(\mathbf{x}_i-\mathbf{x}^k_i)+\frac{1}{2}
(\mathbf{x}_i-\mathbf{x}^k_i)^\text{T}\nabla^2_{\mathbf{x}_i\mathbf{x}_i}f(\mathbf{x}^k)
(\mathbf{x}_i-\mathbf{x}^k_i)+ \beta\|\mathbf{x}_i-\mathbf{x}_i^k\|^2_2$;\smallskip
\noindent$\bullet$ In the same setting as above, one can also better preserve the partial convexity of $f$ and set
$\tilde{f}_i(\mathbf{x}_i;\mathbf{x}^k)=f(\mathbf{x}_i,\mathbf{x}_{-i}^k)+ \beta\|\mathbf{x}_i-\mathbf{x}_i^k\|^2_2$,
where $\mathbf{x}_{-i}\triangleq (\mathbf{x}_{1},\ldots, \mathbf{x}_{i-1},
\mathbf{x}_{i+1},\ldots ,\mathbf{x}_{N})$;\smallskip
\noindent$\bullet$ As a last example, suppose that $f$ is the difference of two convex functions $f^{(1)}$ and
$f^{(2)}$, i.e., $f(\mathbf{x})=f^{1}(\mathbf{x})-f^{2}(\mathbf{x})$, one can preserve the partial convexity in $f$ setting
$\tilde{f}_i(\mathbf{x}_i;\mathbf{x}^k)=f^{(1)}(\mathbf{x}_i, \mathbf{x}_{-
i}^k)-\nabla_{\mathbf{x}_i}f^2(\mathbf{x}^k)^\text{T}(\mathbf{x}_i-\mathbf{x}^k_i)+{ \beta}\,\|
\mathbf{x}_i-\mathbf{x}_i^k\|^2_2$.\vspace{-0.4cm}
\end{asparaenum}
\section{AsyFLEXA: Probabilistic Model} \label{sec:probabilistic}\vspace{-0.2cm}
In this section, we complete the description of AsyFLEXA, introducing the probabilistic model underlying the generation of the pairs index-delays.
Given Problem~\eqref{eq:optimization_problem} and an initial point ${\mathbf{x}}^0$, the pair $(i^k,\d^k)$ in Step 1 of Algorithm~\ref{alg:global}, for each $k$, is a realization of a random vector $\underline{\boldsymbol{\omega}}^k\triangleq (\underline{i}^k,
\underline{\mathbf{d}}^k)$, taking values on $\mathcal N \times \mathcal D$,
where ${\cal D}$ is the set of all possible delay vectors.
We anticipate that all the delays $d_i^k$ are assumed to be bounded (Assumption C below), i.e., $d^k_i \leq \delta$, for all $k$ and $i$.
Hence, ${\cal D}$ is contained in the set of all possible $N$-length vectors whose components are
integers between $0$ and $\delta$.
Let $\Omega$ be the sample space of all the sequences $\omega\triangleq\{ (i^k,
\mathbf{d}^k)\}_{k\in\mathbb{N}_+}$.\@ifstar\footnote@star\footnote@nostar{With a slight abuse of notation,
we denote by $\boldsymbol{\omega}_k$ the $k$-th element of the sequence $\omega\in \Omega$, and by
$\boldsymbol{\omega}^k$ the value taken by the random variable $\underline{\boldsymbol{\omega}}^k$ over $\omega$,
i.e.
$\underline{\boldsymbol{\omega}}^k(\omega) = \boldsymbol{\omega}^k$.} We will use the following
shorthand notation: we set
$\underline{\boldsymbol{\omega}}^{0:k} \triangleq (\underline{\boldsymbol{\omega}}^0,\underline{\boldsymbol{\omega}}^1, \ldots,
\underline{\boldsymbol{\omega}}^k)$ (the first $k+1$ random variables);
${\boldsymbol{\omega}}^{0:k} \triangleq ({\boldsymbol{\omega}}^0,{\boldsymbol{\omega}}^1, \ldots,{\boldsymbol{\omega}}^k)$ ($k+1$ possible values
for the random variables $\underline{\boldsymbol{\omega}}^{0:k}$); and
${\boldsymbol{\omega}}_{0:k} \triangleq ({\boldsymbol{\omega}}_0,{\boldsymbol{\omega}}_1, \ldots,{\boldsymbol{\omega}}_k)$ (the first $k+1$
elements of $\omega$). We introduce next the probability space that will be used to build our probabilistic model.
The sample space is $\Omega$. To define a $\sigma$-algebra on $\Omega$, we consider, for
$k\geq 0$ and every $\boldsymbol{\omega}^{0:k}\in {\cal N}\times {\cal D}$,
the cylinder\vspace{-0.1cm}
$$C^k(\boldsymbol{\omega}^{0:k}) \triangleq \{\omega\in \Omega: \boldsymbol{\omega}_{0:k} =
\boldsymbol{\omega}^{0:k}
\},\vspace{-0.1cm}$$ i.e., $C^k(\boldsymbol{\omega}^{0:k})$ is the subset of $\Omega$ of all sequences $\omega$ whose first $k$
elements are $\boldsymbol{\omega}^0, \ldots \boldsymbol{\omega}^k$. Let us
denote by ${\cal C}^k$ the set of all possible $C^k(\boldsymbol{\omega}^{0:k})$ when $\boldsymbol{\omega}^t$, $t=0,
\ldots, k$, takes all possible values;
note, for future reference, that ${\cal C}^k$ is a partition of $\Omega$. Denoting by $\sigma\left({\cal C}^k\right)$ the $\sigma$-algebra generated by ${\cal C}^k$, define for all $k$, \vspace{-0.1cm}
\begin{equation}\label{eq:sub-sigma}
{\cal F}^k\triangleq \sigma\left({\cal C}^k\right) \qquad \mbox{\rm and}\qquad
{\cal F}\triangleq \sigma\left(\cup_{t=0}^\infty{\cal C}^t\right).
\end{equation}
We have ${\cal F}^k\subseteq {\cal F}^{k+1} \subseteq {\cal F}$ for all $k$. The latter inclusion is obvious, the former
derives easily from the fact that any cylinder in ${\cal C}^{k-1}$ can be obtained as a finite union of cylinders in ${\cal
C}^k$.
\newline The desired probability space is fully defined once $\mathbb{P}(C^k(\boldsymbol{\omega}^{0:k}))$, the probabilities
of all cylinders, are given. These probabilities should satisfy some very natural, minimal consistency
properties, namely:
(i) the probabilities of the union of a finite number of disjoint cylinders should be equal to the sum of the probabilities
assigned to each cylinder; and (ii) suppose that a cylinder $C^k(\boldsymbol{\omega}^{0:k})$ is contained in the union
$U$ of a countably infinite number of other cylinders, then $\mathbb{P}(C^k(\boldsymbol{\omega}^{0:k})) \leq P(U)$.
Suppose now that such a $\mathbb P$ is given.
Classical results
(see, e.g., \cite[Theorem 1.53]{klenke2013probability})
ensure that one can extend these probabilities to a probability measure $P$ over $(\Omega, {\cal
F})$, thus defining our working probability space $A \triangleq
(\Omega, {\cal F}, P)$. By appropriately choosing the probabilities of the cylinders, we can model
in a unified way many cases of practical interest; several examples are given in Section~\ref{sec:Examples}.
Given $A$, we can finally define the discrete-time, discrete-value stochastic process $\underline{\boldsymbol{\omega}}$,
where $\{\underline{\boldsymbol{\omega}}^k(\omega)\}_{k\in \mathbb{N}_+}$ is a sample path of the process. The
$k$-th entry
$\underline{\boldsymbol{\omega}}^k(\omega)$ of $\underline{\boldsymbol{\omega}}(\omega)-$the $k$-th element of
the sequence $\omega-$is a realization of the random vector $\underline{\boldsymbol{\omega}}^k= (\underline{i}^k,
\underline{\mathbf{d}}^k):\Omega \mapsto \mathcal N\times \mathcal D$.
This process fully describes the evolution of Algorithm 1. Indeed, given an instance of Problem (\ref{eq:optimization_problem}) and a starting point, the trajectories
of the variables ${\mathbf{x}}^k$
and $ {\mathbf{x}}^{k-\mathbf{d}^k}$ are completely determined once a sample path $\{ (i^k, \mathbf{d}^k)\}_{k\in
\mathbb{N}_+}$ is drawn from $\underline{\boldsymbol{\omega}}$.
Note that the joint probability
$$
p_{\underline{\boldsymbol{\omega}}^{0:k}}(\boldsymbol{\omega}^{0:k})\triangleq \mathbb{P}
(\underline{\boldsymbol{\omega}}^{0:k}=
\boldsymbol{\omega}^{0:k})
$$
is simply the probability of the corresponding cylinder: $C^k(\boldsymbol{\omega}^{0:k})$.
We will often need to consider the conditional probabilities $p((i,\mathbf{d})\,|\, \boldsymbol{\omega}^{0:k})\triangleq
\mathbb{P}(\underline{\boldsymbol{\omega}}^{k+1}=(i,\mathbf{d})|
\underline{\boldsymbol{\omega}}^{0:k}={\boldsymbol{\omega}}^{0:k})$.
Note that we have
\begin{equation}\label{eq:conditional prob}
p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k}) = \frac{\mathbb{P}(C^{k+1}(\boldsymbol{\omega}^{0:k+1}))}
{\mathbb{P}(C^{k}(\boldsymbol{\omega}^{0:k}))},
\end{equation}
where we tacitly assume $ p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k})=0$, if $\mathbb{P}(C^{k}(\boldsymbol{\omega}^{0:k}))=0$.
We remark that these probabilities need not be known in practice to implement the algorithm. They are instead determined
based on
the particular system (hardware architecture, software implementation, asynchrony, etc.) one is interested to model. Here, we make only some minimal assumptions on these probabilities and stochastic model, as stated next.
\noindent \textbf{Assumption C (On the probabilistic model).} Given Algorithm 1 and the stochastic process
$\underline{\boldsymbol{\omega}}$, suppose that\vspace{-0.2cm}
\begin{description}
\item[(C1)] There exists a $\delta\in \mathbb{N}_+$, such that $d^k_{i} \leq \delta$, for all $i$ and $k$;\smallskip
\item[(C2)] For all $i\in\mathcal{N}$ and $\omega\in\Omega$, there exists at least one $t\in[0,\ldots,T]$, with $T>0$, such that $$\sum_{\mathbf{d}\in
\mathcal D}p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k+t-1})\geq p_{\min},\quad\text{if}\quad p_{\underline{\boldsymbol{\omega}}^{0:k+t-1}}(\boldsymbol{\omega}^{0:k+t-1})>0,$$ for some $p_{\min}>0$;\smallskip
\item[(C3)] $d^k_{i^k} =0$,
for any
$k\geq0$.
\end{description}
\noindent
These are quite reasonable assumptions, with very intuitive interpretations. C1 just limits the age of the old
information used in the updates. Condition C2 guarantees that every $T$ iterations each block-index $i$ has a
non negligible positive probability to be updated. These are minimal requirements that are satisfied in {practically all computational environments.}
The condition
$d^k_{i^k} =0$ means that when a worker updates the $i^k$-th block, it uses the most
recent value of that block-variable. This assumption
is automatically satisfied, e.g., in a message passing-based system or in a shared memory-based architecture if the
variables are partitioned and
assigned to different cores (see Example 6 in Section \ref{sec:Examples}).
If instead all the cores can update all variables,
$d^k_{i^k} =0$ can be simply enforced by a
software lock on the ${i^k}-$th block of the shared memory: once a core $c$ has read a
block-variable
${\mathbf{x}}_{i^k}$, no other core can change it, until $c$ has performed its update. Note that in practice it is very unlikely that this lock
affects
the performance of the
algorithm, since usually the number of cores is much smaller than the number
of block-variables.
Actually, in some systems, this lock can bring in some benefits. For instance, consider two cores sharing all variables, with one core much faster than the other. A lock on ${\mathbf{x}}_{i^k}^k$ will prevent potentially much older information to overwrite most recent updates of the faster core.
Note also that conditions similar to C3 are required by all block asynchronous methods in the literature but \cite{peng2016arock}: they take the form of locking the variable to update before performing a prox
operation
\cite{liu2015asyspcd}.
\begin{remark} \label{rem:key1} \rm
The knowledge of the probability space $A$ is by no means required from the workers to
perform the updates. One need not even specify explicitly the probability distribution; it is
sufficient to show that a probability space $A$ satisfying Assumption C exists for the specific
system (e.g., computational architecture, asynchronous protocol, etc.) under consideration. We
show next how to do so for several schemes of practical interest. \vspace{-0.6cm}
\end{remark}
\subsection{Examples and special cases}\label{sec:Examples} \vspace{-0.2cm} The proposed model encompasses a gamut of practical schemes, some of which are discussed next.
{It is of note that our framework allows us to analyze in a unified way not only randomized methods, but also deterministic algorithms.}
\noindent
{\bf 1. Deterministic sequential cyclic BCD:}
In a deterministic, cyclic method there is only one core that cyclically updates all block-variables; for simplicity we assume the natural order, from 1 to $N$. Since there is only one core, the reading is always consistent and there are no delays: $\cal D=\{\mathbf{0}\}$. To represent the cyclic choice it is now enough to assign probability 1 too all cylinders of the type\vspace{-0.2cm}
$$
C^k = \{ \omega: \omega_0 = (1,\mathbf{0}), \omega_1 = (2,\mathbf{0}), \ldots, \omega_k = ((k \mod N) +1,\mathbf{0}) \}\vspace{-0.2cm}
$$
and probability zero to all others. It is easy to see that Assumption C is satisfied. This can be seen as a probabilistic model of the deterministic algorithm in \cite{scutari2014decomposition}. The consequence however is that, by Theorem~\ref{Theorem_convergence}, convergence can be claimed only in a probabilistic sense (a.s.).
This is not surprising, as we are describing a deterministic algorithm as limiting case of a probabilistic model.
\noindent
{\bf 2. Randomized sequential BCD:}
Suppose now that there is only one core selecting at each iteration randomly an index $i$, with a positive probability.
Therefore, at each
iteration, ${\mathbf{x}}^{k-\mathbf{d}^k} = {\mathbf{x}}^k$ or, equivalently, $\cal
D=\{\mathbf{0}\}$. This scheme can be described by a stochastic process, where the cylinders are assigned arbitrary
probabilities but satisfying all the conditions given in previous subsection.
\noindent
{\bf 3. Randomized parallel BCD:}
Suppose that there are $C$ cores and the
block-variables are partitioned in $C$ groups $I_1, I_2, \ldots,
I_C$; each set $I_c$ is assigned to one core only, say $c$.
Hence, if core $c$ performs the update at iteration $k$, all variables $i\in I_c$ satisfy $d_i^k=0$.
Denote by $\boldsymbol{0}(c)$, $\boldsymbol{1}(c)$, \ldots, and
$(\boldsymbol{C-1})(c)$, $c=1,\ldots,C$, the $N$-length vectors whose components are zeros in the positions of the block-variables in the set $I_c$ and all $0, 1,\ldots, C-1$ in the other positions, respectively. Set $\mathcal{D}=\{\boldsymbol{0}(c), \boldsymbol{1}(c), \ldots,
(\boldsymbol{C-1})(c),\,\,c=1\ldots,C\}$, and denote by $c^k$ the core performing the update at iteration $k$.
Assign to the cylinders the following probabilities: $\forall i^0,i^1,\ldots,i^{2C-1},\ldots \in\mathcal{N}$, \vspace{-0.1cm}
\begin{align*}
&\mathbb{P}(C^0((i^0,\boldsymbol{0}(c^0) ))) = 1/N,\smallskip\\
& \mathbb{P}(C^1((i^0,\boldsymbol{0}(c^0) ), (i^1,\boldsymbol{1}(c^1) ))) = 1/N^2,\\
\ldots\\
&\mathbb{P}(C^{C-1}((i^0,\boldsymbol{0}(c^0) ), (i^1,\boldsymbol{1}(c^1) ),\ldots, (i^{C-1},\boldsymbol{C-1}(c^{C-1}) ))) = 1/N^C,\smallskip\\
&\mathbb{P}(C^C( (i^0,\boldsymbol{0}(c^0) ), (i^1,\boldsymbol{1}(c^1) ),\ldots, (i^{C-1},\boldsymbol{C-1}(c^{C-1}) ),
(i^C,\boldsymbol{0}(c^C) ))) = 1/N^{C+1},\smallskip\\
&\mathbb{P}(C^{C+1} ( (i^0,\boldsymbol{0}(c^0) ), (i^1,\boldsymbol{1}(c^1) ),\ldots, (i^{C},\boldsymbol{0}(c^C) ),
(i^{C+1},\boldsymbol{1}(c^{C+1}) ))) = 1/N^{C+2},\smallskip\\
&\ldots\smallskip\\
&\mathbb{P}(C^{2C-1}
( (i^0,\boldsymbol{0}(c^0) ), (i^1,\boldsymbol{1}(c^1) ),\ldots,
(i^{2C-1},\boldsymbol{C-1}(c^{2C-1}) ))) = 1/N^{2C},\\
&\ldots
\end{align*}
In words,
in the first $C$ iterations (from $k=0$ to $k=C-1$), all updates are performed using
the same vector ${\mathbf{x}}^{k-\mathbf{d}^k} = {\mathbf{x}}^0$; and at each iteration any index has uniform probability
to be selected.
This situation is then repeated for the next $C$ iterations, this time using ${\mathbf{x}}^{k-\mathbf{d}^k} = {\mathbf{x}}^C$, and so on.
This model clearly corresponds to a randomized parallel block-coordinate descent method wherein $C$ cores update $C$ block-variables chosen uniformly at random. Note that Assumption C is trivially satisfied.
The example above clearly shows that defining probabilities by using the cylinders can be quite tedious even in simple
cases. Using \eqref{eq:conditional prob} we can equivalently define the probabilistic model by specifying the conditional
probabilities $p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k})$, which is particularly convenient when at every iteration
$k$ the probability that $\underline{\boldsymbol{\omega}}^k$ takes value $(i,\mathbf{d})$
is independent of $ \boldsymbol{\omega}^{0:k-1}$. We exemplify this alternative approach in the following examples.
\noindent
{\bf 4. Asynchronous BCD in shared memory systems:}
Consider a
generic shared memory system, under Assumption C3. Then,
the set $\cal D$ is given by all the $N$-length vectors whose components are non negative integers between 0 and $\delta$.
Suppose that, at every $k$, all cores select an index uniformly at random, but the probabilities associated with the delays can be different. Then, for every
$k\geq 0
$, given $\boldsymbol{\omega}^{0:k}$, and $i\in {\cal N}$, we have
\vspace{-0.1cm}
$$
\sum_{\d\in {\cal D}} p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k}) = \frac{1}{N}.\vspace{-0.1cm}
$$
This setting is consistent with the one studied in \cite{lian2015asynchronous,davis2016asynchronous,DavisEdmundsUdell,liu2015asyspcd}.
Our probabilistic model however is more general than that of \cite{lian2015asynchronous,davis2016asynchronous,DavisEdmundsUdell,liu2015asyspcd}. For instance, differently from \cite{lian2015asynchronous,davis2016asynchronous,DavisEdmundsUdell,liu2015asyspcd}, we can easily model scenarios wherein $
\sum_{\d\in {\cal D}} p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k})
$ are not uniform and/or depend
on the iteration and/or
on the history of the algorithm.
This
possibility has important
ramifications, since the assumption that the indices are selected uniformly at random is
extremely strong and unrealistic. In fact, it is
satisfied only if all cores have the
same computational power and have access to all variables.
This is not the case, in most of practical settings. For instance, consider a computational architecture composed of two CPUs sharing all the variables, with one CPU much faster than the other.
If the recent history exhibits iterations with a small
value of $\|\d^k\|_\infty$, then it is more likely that the slower core will perform the next update, and vice versa. Similar situations are expected also in other common settings, such as shared memory systems with variable partitioning (see Example 5 below) and message passing-based architectures.
This clearly shows that our model captures realistic architectures more faithfully.
\noindent
{\bf 5. Asynchronous BCD in shared memory systems with variable
partitioning:} Consider the setting as in Example 4, but now partition the variables across cores, as described in Example 3.
This is the configuration most often used in numerical experiments, since it has proven
to be most effective in practice; it also models a message passing architecture.
In order to satisfy C3, it is enough to set, for all $\boldsymbol{\omega}^{0:k}$ and $i\in I_c$, \vspace{-0.1cm}
$$
p((i,\mathbf{d})\,|\, {\boldsymbol{\omega}}^{0:k}) = 0, \; \hbox{\rm if some $\d_j\neq \mathbf{0}$, $j\in I_c$.}
$$
A variant of this setting is the {\em without replacement} updating scheme considered in the numerical experiments of \cite{liu2015asyspcd}: the block-variables are partitioned among the cores and, at each ``epoch", variables in each partition are first randomly shuffled and then updated cyclically by the core. This choice of the updates was shown to be numerically very effective. While \cite{liu2015asyspcd} cannot provide any theoretical analysis of such a scheme, we can easily cover this case by just merging this example with Example 2.
\noindent \textbf{Other examples:} Several other examples can be considered, which we omit because of space limitation. Here we only mention that it is quite straightforward to
analyze by our model also ``hybrid'' systems, which combine somehow two or more examples described above. {For instance, consider a cluster computer system
wherein the optimization variables are partitioned across the machines; let $I_m$ be the set of variables controlled by machine $m$ and stored in its internal shared memory. The update of the variables in $I_m$ is performed by the processors/cores of machine $m$ according to some shared memory-based asynchronous scheme (e.g., subject to inconsistent read). The information on the variables not in $I_m$ is instead updated through communication with the other processors (message passing)}
\vspace{-0.4cm}
\section{AsyFLEXA: Convergence Results}\label{sec:convergence}\vspace{-0.2cm}
We present now our main convergence theorem, under Assumptions A-C. The extension to the case of nonconvex constraints is addressed in Section \ref{nonconvex}.
We will use $\|M_F(\mathbf{x})\|_2$ as a measure of optimality, with \vspace{-0.1cm}
\begin{equation}
M_F(\mathbf{x})\triangleq \mathbf{x}-\underset{\mathbf{y}\in\mathcal{X}}{\text{arg min}}\left\{\nabla f(\mathbf{x})^\text{T}(\mathbf{y}-\mathbf{x})+g(\mathbf{y})+\frac{1}{2}\|\mathbf{y}-\mathbf{x}\|^2_2\right\}.\vspace{-0.1cm}
\end{equation}
This is a valid measure of stationarity because $M_F(\mathbf{x})$ is continuous and $\|M_F(\mathbf{x})\|_2={0}$ if
and only if $\mathbf{x}$ is a stationary solution of Problem \eqref{eq:optimization_problem}.
To state our major convergence result, we need to introduce first the following intermediate definitions.
Recalling the definition of
$T$ as in Assumption $C2$, let
$\underline{\mathcal{K}}_i^k$ be the (random) set of iterations between $k-\delta$ and $k+T-1$ at which the block-variable $i$ has been updated,
$\mathcal{K}_i^k \triangleq\{t\in[k-\delta;k+T-1]\,|\,\underline{i}^t=i\}$, while
$\bar{\underline{\mathcal{K}}}^k_i$ is the subset of $\underline{\mathcal{K}}_i^k$ containing only the elements of $\underline{\mathcal{K}}_i^k$ (iterations) between $k-
\delta$ and $k-1$. Our convergence results leverage a Lyapunov function $\tilde{F}$ that suitably combines present and past iterates, and it is defined as
\begin{equation}
\tilde{F}(\mathbf{x}^k,\dots,\mathbf{x}^{k-\delta})
= F(\mathbf{x}^k)+\delta\frac{L_f}{2}\,\left(\sum\limits_{l=k-\delta}^{k-1}\left(l-(k-1)+\delta\right)\,\|\mathbf{x}^{l+1}-\mathbf{x}^l\|_2^2\right),
\label{lyapunov}\end{equation}
where it is understood that ${\mathbf{x}}^l = {\mathbf{x}}^0$, if $l<0$; therefore, $\tilde{F}$ is well defined for any $k\geq 0$. Note that, by this convention, $\tilde{F}(\mathbf{x}^0,\dots,\mathbf{x}^{0-\delta})= F({\mathbf{x}}^0)$. Furthermore,
we also have
$F^* \triangleq \underset{{\mathbf{x}} \in {\cal X}}{\min}\; F({\mathbf{x}})\leq\underset{\mathbf{x}^k,\dots,\mathbf{x}^{k-
\delta}\in {\cal X}}{\min}\;\tilde F (\mathbf{x}^k,\dots,\mathbf{x}^{k-\delta})$. We are now ready to state our major convergence result.
\begin{theorem}\label{Theorem_convergence} Let Problem \eqref{eq:optimization_problem} be given, along with Algorithm 1 and the stochastic process $\underline{\boldsymbol{\omega}}$. Let $\{\underline{\mathbf{x}}^k\}_{k\in\mathbb{N}_+}$ be the
sequence generated by the algorithm, given
${\mathbf{x}}^0\in {\cal X}$. Suppose that Assumptions A-C hold true and that
\begin{equation}\label{eq:max gamma}
\gamma<\frac{c_{\tilde{f}}}{L_f+\frac{\delta^2L_f}{2}}.
\end{equation}
Define $K_\epsilon$ to be the first iteration such that $\mathbb{E}\left(\|M_F(\underline{\mathbf{x}}^k)\|_2^2\right)\leq\epsilon$. Then:
\noindent
(a) Every limit point of $\{\underline{{\mathbf{x}}}^k\}_{k\in\mathbb{N}_+}$ is a stationary solution of \eqref{eq:optimization_problem}
a.s.;
\noindent
(b) The sequence of objective function values $\{ F(\underline{\mathbf{x}}^k)\}_{k\in\mathbb{N}_+}$ converges a.s.;\vspace{-0.3cm}
\begin{align}\label{eq:compKbound} (c)\;\;\; &
K_\epsilon\, \leq \, \frac{C_1(\gamma, \delta)(T+1)(F(\mathbf{x}^0)-F^*)}{\epsilon}\\&+\frac{C_2(\gamma, \delta)\gamma^2}{\epsilon}
\underbrace{
\sum\limits_{k=0}^{K_\epsilon}\mathbb{E}\left(\sum\limits_{i=1}^N\underline{M}_i^k\sum\limits_{t\in\underline{\mathcal{K}}_i^k}\left(\tilde{F}(\underline{\mathbf{x}}^t,\ldots,\underline{\mathbf{x}}^{t-\delta})-\tilde{F}(\underline{\mathbf{x}}^{t+1},\ldots,\underline{\mathbf{x}}^{t+1-\delta})\right)\right)}_\text{B},\nonumber
\end{align}
where: \vspace{-0.2cm}
\begin{align}\label{eq:C1}
&C_1(\gamma, \delta)\triangleq\frac{2\left(1+(1+L_E)(1+L_B+L_E)+\gamma^2Np_{\text{min}}\alpha^{-1}
(1+(L_f+1)^2)\right)}{\gamma\left(c_{\tilde{f}}-\gamma\left(L_f+\frac{\delta^2L_f}{2}\right)\right)
\left(p_{\text{min}}-p_{\text{min}}\alpha\right)},\\[0.2cm]
&C_2(\gamma, \delta)\triangleq\frac{2TL_B(1+L_B+L_E)}{\gamma\left(c_{\tilde{f}}-\gamma\left(L_f+\frac{\delta^2L_f}{2}
\right)\right)\left(p_{\text{min}}-p_{\text{min}}\alpha\right)},
\label{eq:C2} \end{align}
$\alpha$ is an arbitrary
fixed value in $(0;1)$, $\underline{M}_i^k\triangleq\underset{l=k,\ldots,k+T}{\max}|\bar{\underline{\mathcal{K}}}_i^l|$ .\vspace{-0.2cm}
\end{theorem}
\begin{proof}
See Appendix.\vspace{-0.2cm}
\end{proof}
\noindent
The theorem states that convergence to stationary points occurs a.s. (the objective function values converge too); it also gives an estimate of the number of
iterations $K_\epsilon$ necessary to enforce $\mathbb{E}\left(\|M_F(\mathbf{x}^k)\|_2^2\right)\le \epsilon$. Convergence is guaranteed if, in particular, the step-size is sufficiently small; the bound \eqref{eq:max gamma} makes this precise. Note that if the method is synchronous, $\delta =0$,
the bound in \eqref{eq:max gamma}, going like the inverse of the Lipschitz constant, becomes the renowned conditions used in many synchronous (proximal-gradient-like) schemes. The term $\delta^2/2$ in the denominator of \eqref{eq:max gamma} should then be seen as the price to pay for asynchrony: the larger the possible delay $\delta$, the smaller $\gamma$ should be to tolerate such delays. Roughly speaking, this means that the more chaotic the computational environment, the more conservative the step should be, and consequently the smaller the steps of the algorithm are.
The interpretation of the bound $\eqref{eq:compKbound}$ is not immediate, because of the presence of the term $B$; we now elaborate on it.
If there exists a (deterministic) bound $C$ on $\underline{M}_i^k$,
i.e., $\underline{M}_i^k \leq C$ for all $k$ and $i$, then one can write
\begin{align*}
B \, &\leq \, C
\mathbb{E}\left(\sum\limits_{k=0}^{K_\epsilon}\sum\limits_{i=1}^N\sum\limits_{t\in\underline{\mathcal{K}}_i^k}\left(\tilde{F}(\underline{\mathbf{x}}^t,\ldots,\underline{\mathbf{x}}^{t-\delta})-\tilde{F}(\underline{\mathbf{x}}^{t+1},\ldots,\underline{\mathbf{x}}^{t+1-\delta})\right)\right)\\
&\leq\, C (T+\delta)(F({\mathbf{x}}^0)- F^*).
\end{align*}
Therefore, \eqref{eq:compKbound} can be upper bounded as\vspace{-0.2cm}
\begin{equation}
K_\epsilon\, \leq\, \Big[ C_1(\gamma, \delta)\cdot(T+1)+C_2(\gamma, \delta)\cdot\gamma^2\cdot C\cdot(T+\delta) \Big]\, \frac{F(\mathbf{x}^0)-F^*}{\epsilon}.
\label{eq:compKbound_worst}
\end{equation}
Recalling the definition of $\underline{M}_i^k$ and that $|\bar{\underline{\mathcal{K}}}_i^k|$
is a random variable counting the number of times the index $i$ has been updated in the
iteration window $[k-\delta, k-1]$, $\underline{M}_i^k \leq \delta$ always holds; therefore, one can always take $C=\delta$.
Of course this is a very rough approximation: it is hard to expect that in a
given time window always the same variable, $\bar i$, is updated and, even if this were the
case, all other $\underline{M}^k_i$, $i\neq \bar i$, would be $0$ and not $\delta$. Consider for example the commonly analyzed ``uniform case'' where the processing of every block-variable requires the same time.
In this case one can reasonably take $C=1$ in \eqref{eq:compKbound_worst} {\em independently of the number of workers.}
This intuition is corroborated by our experiments, which are summarized in Table \ref{tabled}. AsyFLEXA was ran on two different architectures, namely: a shared-memory system with 10 cores, and a message passing architecture composed of two nodes, with 10 cores each. Two LASSO problems with $10,000$ variables each were considered, and the variables were equally partitioned across the workers. In the first LASSO instance, the Hessian matrix was a dense matrix, which models situations where the workload is equally distributed across the workers. In the second LASSO problem, the Hessian matrix had many sparse rows, to create some unbalancedness in the workers' workload. Table~\ref{tabled} shows the empirical average delay (the average is taken over the components of the delay vector and time) and the maximum delay $\delta$, estimated in $500$ epochs (one epoch is triggered when all blocks have been updated once). As expected,
$\delta$ is much larger than the experienced average delay, confirming that \eqref{eq:compKbound_worst} with $C=\delta$ is a very conservative bound.
\begin{table}
\begin{center}
\begin{tabular}{|*{4}{c|}}
\hline
&Average Delay&$\delta$&\# of cores\\
\hline
Multi-core Machine: Balanced Workload &1.11&3&10\\
\hline
Multi-core Machine: Unbalanced Workload &2.58&28&10\\
\hline
Message Passing System: Balanced Workload&1.87&30&10 per node\\
\hline
Message Passing System: Unbalanced Workload&3.01&36&10 per node\\
\hline
\end{tabular}
\caption{{Average delay and maximum delay $\delta$ for AsyFLEXA, ran on a multi-core machine and on a message passing system.
}}\label{tabled}
\end{center}\vspace{-0.8cm}
\end{table}
While $C$ can always be pessimistically upper bounded by $\delta$, a tighter value can be found by tailoring the analysis to the specific problem and architecture under consideration.
Finally, we remark the importance of the use, in the complexity analysis, of the $\underline{M}^i_k$, counting the number of times the index $i$ has been updated in a certain iteration window. The use of these variables seems to be a new feature of our analysis.
While getting a sharp estimate for the upper bound $C$ may be difficult in practice, the bound
\eqref{eq:compKbound} gives a good insight into the elements that really influence the algorithm, showing that what really matters, in some expressions appearing in \eqref{eq:compKbound}, is not $\delta$, but the usually much smaller number of times the blocks are actually updated. The use of these variables allows us to get a sharper bound with respect to the case in which one sets $C=\delta$.
From this point of view, we believe that typical upper bounds, as those obtained in
\cite{liu2015asynchronous,liu2015asyspcd,davis2016asynchronous,DavisEdmundsUdell,peng2016arock}, where $\delta$ is the only considered ``delay", do not give an accurate description of the actual
worst-case scenario.}
\noindent \textit{Almost linear speedup:} To study the speedup achievable by the proposed method,
we make two simplifying assumptions, consistent with those made in the literature, namely: (a) {$\delta$ is proportional to the number of workers}, which is reasonable in ``symmetric'' situations; and (b)
$K_{\epsilon}$ is a good proxy for the number of iterations performed by
the algorithm to reach the desired accuracy.
Choose the stepsize $\gamma$ to be small enough so that
\eqref{eq:max gamma} is always satisfied in the range of values of $\delta$ under consideration; then $C_1(\gamma, \delta)$ and $C_2(\gamma, \delta)$ can be taken to be
constants. Consider now the two summands in square brackets in
\eqref{eq:compKbound_worst}. Without the second term, one would have
ideal linear speed-up.
However, since one can expect the second term to be much smaller than the
first (at least
when $\delta$ is not large), an almost linear speed-up can be anticipated. In fact, by \eqref{eq:C1} and \eqref{eq:C2}, the second term is smaller than the first one, if $\gamma$ is sufficiently small.
In practice, of course, the speed up and in particular the range of the number of workers for which
linear speedup is expected, will be problem and architecture dependent.\vspace{-0.5cm}
\section{Nonconvex Constraints} \label{nonconvex}\vspace{-0.2cm}
In this section, we remove the assumption that
all constraints are convex, and study the following more general nonconvex constrained optimization problem:\vspace{-0.1cm}
\begin{equation}
\begin{aligned}
& \underset{\mathbf{x}}{\min} & & F(\mathbf{x})= f(\mathbf{x})
+ \sum_{i=1}^N g_i({\mathbf{x}}_i)\\[0.3em]
&&&\begin{rcases}\mathbf{x}_i \in \mathcal X_i, \qquad i = 1, \dots , N, \\[0.3em]
c_{1}(\mathbf{x}_1)\leq0,\ldots, c_{N}(\mathbf{x}_N)\leq0, \end{rcases}\triangleq\mathcal{K}
\end{aligned}
\label{ncc_1}\tag{P$^\prime$}\vspace{-0.1cm}
\end{equation}
where $c_{i}(\mathbf{x}_i)
\leq0$ are nonconvex private constraints, with $c_i: \mathcal{O}_i\rightarrow \mathbb{R}^{m_i}$, and $\mathcal{O}_i$ denoting an open set containing $\mathcal{X}_i$; let also define $\mathcal{K}_i\triangleq\{\mathbf{x}_i
\in\mathcal{X}_i:c_{i}(\mathbf{x}_i)\leq 0,\}$.
Note that $c_{i}(\mathbf{x}_i)$ is a vector function, whose individual component is denoted by $c_{i,j}$, with $j =1, \ldots, m_i$.
Problem (\ref{ncc_1}) is motivated by several applications in
signal processing, machine learning, and networking; see, \cite{scutari2016parallel}
and references therein for some concrete examples.
To deal with nonconvex constraints, we need some
\textit{regularity} of the constraint functions. Anticipating that all $c_i$ are assumed to be $C^1$ on $\mathcal X_i$, we will use the Mangasarian-Fromovitz Constraint
Qualification (MFCQ). \vspace{-0.1cm}
\begin{definition} A point $\bar{\mathbf{x}}\in\mathcal{K}$ satisfies the MFCQ if the following implication is
satisfied:\vspace{-0.1cm}
\begin{equation}
\begin{rcases}
\mathbf{0}\in\sum\limits_{i=1}^N\sum\limits_{j\in\bar{J}_i}\mu_{i,j}\nabla_{\mathbf{x}}c_{i,j}(\bar{\mathbf{x}}_i)+N_\mathcal{X}
(\bar{\mathbf{x}})\\
\mu_{i,j}\geq 0,\,\forall j\in\bar{J}_i,\,\forall i\in\mathcal{N}
\end{rcases}
\Rightarrow\mu_{i,j}=0,\,\forall j\in\bar{J}_i,\,\forall i\in\mathcal{N},
\end{equation}
where $N_\mathcal{X}(\bar{\mathbf{x}})\triangleq\{\mathbf{z}\in\mathcal{X}:\mathbf{z}^\text{T}(\mathbf{y}-\bar{\mathbf{x}})\leq0,\,\forall\mathbf{y}\in\mathcal{X}\}$ is the normal cone to $\mathcal{X}$ at $\bar{\mathbf{x}}$, and $\bar{J}_i\triangleq\{j:c_{i,j}(\bar{\mathbf{x}}_i)=0\}$ is the index set of nonconvex constraints that are active at $\bar{\mathbf{x}}_i$. \end{definition}
We study Problem~\eqref{ncc_1} under the following assumptions. \smallskip
\noindent \textbf{Assumption A$^\prime$ (On the problem model).} Suppose that
\begin{description}[topsep=-2.0pt,itemsep=-2.0pt]
\item[(A1$^\prime$)] Each set $\mathcal{X}_i\subseteq\mathbb{R}^{n_i}$ is nonempty, closed, and convex;\smallskip
\item[(A2$^\prime$)] $f:\mathcal{O}\rightarrow\mathbb{R}$ is $C^1$, where $\mathcal{O}$ is an open set containing $\mathcal{K}$;\smallskip
\item[(A3$^\prime$)] $\nabla_{\mathbf{x}_i} f$ is $L_{f}$-Lipschitz continuous
on $\mathcal{K}$;\smallskip
\item[(A4$^\prime$)] Each $g_{i}:\mathcal{O}_i\rightarrow\mathbb{R}$ is convex, possibly nonsmooth, and $L_g$-Lipschitz continuous on $\mathcal{X}_i$, where $\mathcal{O}_i$ is an open set containing $\mathcal{X}_i$;\smallskip
\item[(A5$^\prime$)] $\mathcal{K}$ is a compact set;\smallskip
\item[(A6$^\prime$)] Each $c_{i,j}: \mathcal{O}_i\rightarrow \mathbb{R}$ is $C^1$;
\smallskip
\item[(A7$^\prime$)] All feasible points of problem \eqref{ncc_1} satisfy the MFCQ.
\end{description}
\smallskip
\noindent Assumptions A1$^\prime$-A4$^\prime$ are a duplication of A1-A4, repeated here for ease of reference; A5$^\prime$ is stronger than A5, and made here for the sake of
simplicity (one could relax it with A5); and A6$^\prime$ is a standard differentiability assumption on the non convex constraints $c_{i,j}$. \smallskip
\noindent \textbf{AsyFLEXA-NCC:} We are now ready to introduce our asynchronous algorithmic framework for (\ref{ncc_1}), termed AsyFLEXA-NCC (where NCC stands for Non Convex Constraints).
The method is still given by Algorithm~\ref{alg:global}, with the only difference that now
also the nonconvex constraints are replaced by suitably chosen convex approximations; the probabilistic model concerning the choice of the the pair index-delays is the same as the one we used in the case of convex constraints, see Section
\ref{sec:probabilistic}.
More specifically, AsyFLEXA-NCC is given by Algorithm~\ref{alg:global} wherein the subproblem (\ref{eq:best-response-convex}) in Step 2 is replaced by
\begin{equation}
\hat{\textbf{x}}_{i^k}(\textbf{x}^{k-\mathbf{d}^k})\triangleq \underset{\textbf{x}_{i^k} \in \mathcal{K}_{i^k}(\mathbf{x}^k_{i^k})}{\arg\min}
\left\{{\tilde{F}_{i^k}(\textbf{x}_i;\textbf{x}^{k-\mathbf{d}^k})}\triangleq \tilde f_{i^k}({\mathbf{x}}_i; {{\mathbf{x}}}^{k-\mathbf{d}^k}) + g_{i^k}({\mathbf{x}}_i)\right\},
\label{ncc_2}
\end{equation}
where $\tilde f_{i^k}$ is defined as in (\ref{eq:best-response-convex});
$\mathcal{K}_{i^k}(\mathbf{x}_{i^k}^k)$ is a convex approximation of $\mathcal{K}_{i^k}$ at $\mathbf{x}^{k-\mathbf{d}^k}$, defined as $$\mathcal{K}_{i^k}(\mathbf{x}^k_{i^k})\triangleq\{\mathbf{x}_i\in\mathcal{X}_{i^k}:\tilde{c}_{i^k,j}(\mathbf{x}_i;\mathbf{x}^k_{i^k})\leq 0,\, j=1, \ldots, m_{i^k}\};$$
and $\tilde{c}_{i^k,j}:\mathcal{X}_{i^k}\times\mathcal{K}_{i^k}\rightarrow\mathbb{R}$ is a suitably chosen surrogate of ${c}_{i^k,j}$. Note that $K_{i^k}$ depends on $\mathbf{x}_{i^k}^{k}$ and not on $\mathbf{x}_{i^k}^{k-d_{i^k}^k}$, because of Assumption C3 ($d_{i^k}^k=0$).
The surrogate functions $\tilde{c}_{i^k,j}$ can be chosen according to the following assumptions
($\nabla \tilde{c}_{i,j}$ below denotes the partial gradient of $\tilde{c}_{i,j}$ with respect to the first argument).
\smallskip
\noindent \textbf{Assumption D (On the surrogate functions $\tilde{c}_{i,j}$'s).}
\begin{description}[topsep=-2.0pt,itemsep=-2.0pt]
\item[ (D1)] Each $\tilde{c}_{i,j}(\bullet;\mathbf{y})$ $C^1$ on an open set containing $\mathcal{X}_i$, and convex on $\mathcal{X}_i$ for all
$\mathbf{y}\in\mathcal{K}_i$;\smallskip
\item[ (D2)] $\tilde{c}_{i,j}(\mathbf{y};\mathbf{y})=c_{i}(\mathbf{y})$, for all $\mathbf{y}\in\mathcal{K}_i$;
\smallskip
\item[ (D3)] $c_{i,j}(\mathbf{z})\leq\tilde{c}_{i,j}(\mathbf{z};\mathbf{y})$ for all
$\mathbf{z}\in\mathcal{X}_i$ and $\mathbf{y}\in\mathcal{K}_i$;\smallskip
\item[ (D4)] $\tilde{c}_{i,j}(\bullet;\bullet)$ is continuous on $\mathcal{X}_i\times\mathcal{K}_i$;\smallskip
\item[ (D5)] $\nabla_{\mathbf{y}_i} c_{i,j}(\mathbf{y})=\nabla\tilde{c}_{i,j}(\mathbf{y};\mathbf{y})$, for all
$\mathbf{y}\in\mathcal{K}_i$;\smallskip
\item[ (D6)] $\nabla\tilde{c}_{i,j}(\bullet;\bullet)$ is continuous on $\mathcal{X}_i\times\mathcal{K}_i$;
\smallskip
\item[ (D7)] Each $\tilde{c}_{i,j}(\bullet;\bullet)$ is Lipschitz continuous on
$\mathcal{X}_i\times\mathcal{K}_i$.\medskip
\end{description}
Roughly speaking, Assumption D requires $\tilde c_{i,j}$ to be an upper convex approximation of $c_{i,j}$ having the same gradient of $c_{i,j}$ at the base point $\bf y$. Finding such approximations is less difficult than it might seem at a first sight. Two examples are given below, while we refer the reader to \cite{facchinei2016feasible,scutari2016parallel} for a richer list.
\smallskip
\noindent
$\bullet$
Suppose $c_{i,j}$ has
a $L_{\nabla c_{i,j}}$-Lipschitz continuous gradient on the (compact) set $\mathcal{K}_i$. By the Descent Lemma \cite[Proposition A32]{Bertsekas_Book-Parallel-Comp}, the following convex approximation satisfies Assumption D:\vspace{-0.2cm}
\[
\tilde{c}_{i,j}({\mathbf{x}};{\mathbf{y}})\triangleq c_{i,j}({\mathbf{y}})+\nabla_{{\mathbf{x}}}c_{i,j}({\mathbf{y}})^{\scriptscriptstyle T}({\mathbf{x}}-{\mathbf{y}})+
\frac{L_{\nabla c_{i,j}}}{2}\|{\mathbf{x}}-{\mathbf{y}}\|^{2}_2\ge c_{i,j}({\mathbf{x}}).
\]
\noindent
$\bullet$
Suppose that $c_{i,j}$ has a DC structure, that is,
$
c_{i,j}({\mathbf{x}})=c_{i,j}^{+}({\mathbf{x}})-c_{i,j}^{-}({\mathbf{x}}),
$
$c_{i,j}^{+}$ and $c_{i,j}^{-}$ are two convex and continuously differentiable functions.
By linearizing the concave part $-c_{i,j}^{-}$
and keeping the convex part $c_{i,j}^{+}$ unchanged, we obtain the
following convex upper approximation of $c_{i,j}$ that satisfies Assumption D:
\[
\tilde{c}_{i,j}({\mathbf{x}};{\mathbf{y}})\triangleq c_{i,j}^{+}({\mathbf{x}})-c_{i,j}^{-}({\mathbf{y}})-\nabla_{{\mathbf{x}}}c_{i,j}^{-}({\mathbf{y}})
^{\scriptscriptstyle T}({\mathbf{x}}-{\mathbf{y}})\ge c_{i,j}({\mathbf{x}}).\
\]
Note that {the former example is quite general and in principle can be applied to
practically all constraints, even if it could be numerically undesirable if $L_{\nabla c_{i,j}}$ is too large;
the latter example covers, in a possibly more suitable way, the case of concave constraints.}
\smallskip
\noindent \textbf{AsyFLEXA-NCC: Convergence.} In order to gauge convergence, we
redefine the stationarity measure $M_F$, to account for the presence of
nonconvex constraints. We use $\|M_F^c(\mathbf{x})\|_2$, with \vspace{-0.2cm}
\[
M_F^c(\mathbf{x})=\mathbf{x}-\underset{\mathbf{y}\in\mathcal{K}_1(\mathbf{x}_1)\times
\ldots\times\mathcal{K}_N(\mathbf{x}_N)}{\text{arg min}}\{\nabla f(\mathbf{x})^\text{T}
(\mathbf{y}-\mathbf{x})+g(\mathbf{y})+\frac{1}{2}\|\mathbf{y}-\mathbf{x}\|^2_2\}.
\]
It is a valid merit function: $\|M_F^c(\mathbf{x})\|_2$ is continuous and is zero only at
stationary solutions of \eqref{ncc_1} \cite{scutari2014distributed}.
\begin{theorem}\label{th:ncc}
Let Problem \eqref{ncc_1} be given, along with AsyFLEXA-NCC and the stochastic process $\underline{\boldsymbol{\omega}}$. Let $\{\underline{\mathbf{x}}^k\}_{k\in\mathbb{N}_+}$ be the
sequence generated by the algorithm, given
${\mathbf{x}}^0\in {\cal K}$. Suppose that Assumptions A',B-D hold and that
$
\gamm
$ is chosen as in (\ref{eq:max gamma}).
Define $K_\epsilon$ to be the first iteration such that $\mathbb{E}\left(\|M_F^c(\underline{\mathbf{x}}^k)\|_2^2\right)\leq\epsilon$. Then: i) $\underline{\mathbf{x}}^k\in\mathcal{K}_1(\underline{\mathbf{x}}_1^k)\times\ldots\times\mathcal{K}_N(\underline{\mathbf{x}}^N)\subseteq\mathcal{K}$ for all $k\geq0$ (iterate feasibility); and ii) all results in
Theorem \ref{Theorem_convergence} hold
with $M_F$ replaced by $M_F^c$.
\end{theorem} \vspace{-0.4cm}
\begin{proof} See Appendix~\ref{proof_theorem_2}. \end{proof}
We are aware of only one other BCD-asynchronous method \cite{davis2016asynchronous,DavisEdmundsUdell} able to deal with nonconvex
constraints. This method requires the ability to
find global minima of {\em nonconvex} subproblems while our scheme does not suffer from this drawbacks, {as} it only calls for the
solution of strongly convex subproblems. On the other hand, it needs a feasible starting
point and the ability to build approximations $\tilde c_{i,j}$ satisfying Assumption D. While our
requirements are easier to be met in practice (and our analysis is based on a grounded probabilistic model), we think that the two approaches complement
each other and may cover different applications.\vspace{-0.4cm}
\section{Conclusions}\vspace{-0.2cm}
We proposed a novel model for the parallel block-descent asynchronous minimization of the sum of a
nonconvex smooth function and a convex nonsmooth one, subject to nonconvex constraints.
Our model captures the essential features of modern multi-core architectures by providing
a more realistic probabilistic description of asynchrony that that offered by the state of the art. Building on our new probabilistic model, we proved sublinear convergence rate of our
algorithm and a near linear speedup when the number of workers is not too large.
{While we performed some simple numerical tests to validate some of our theoretical findings, extensive simulations are beyond the scope of this paper, and will be the subject of a subsequent work. Some preliminary numerical results can be found in \cite{cannelli2017asynchronous}.}
\vspace{-0.4cm}
\bibliographystyle{plain}
\addcontentsline{toc}{chapter}{Bibliography}
| 9f992332219e28ac0b307e8ae96ba0747bc1ca73 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and Summary}
{}Industry classifications such as GICS, BICS, ICB, NAICS, SIC, etc.\footnote{\, Hereinafter we will refer to these as ``fundamental" industry classifications (see below).} are widely used in quantitative trading. They group stocks into baskets, e.g., industries, i.e., based on some kind of a similarity criterion. On general grounds one then expects (or hopes) that stocks within such baskets on average should be relatively highly correlated. This is valuable information and can be used in various ways. E.g., one can build a simple mean-reversion statistical arbitrage strategy whereby one assumes that stocks in a given industry move together, cross-sectionally demeans stock returns within said industry, shorts stocks with positive residual returns and goes long stocks with negative residual returns, with some generally nonuniform weights.\footnote{\, More generally, one employs a weighted regression instead of demeaning, and there are various ways of fixing the aforesaid weights. For a pedagogical discussion, see, e.g., \cite{MeanRev}.} Industries can also be used as risk factors in multifactor risk models.\footnote{\, For a discussion and literature on multifactor risk models, see, e.g., \cite{GrinoldKahn}.}
{}The aforementioned ``fundamental" industry classifications are based on grouping companies together based on fundamental/economic data (see Section \ref{sec2}), which is expected to add value on longer holding horizons. What about shorter holding horizons relevant to quantitative trading strategies? Other than a large number of market players using such industry classifications to arbitrage mispricings,\footnote{\, This very relevant reason should not to be underestimated, despite its ``behavioral" nature.} how do we know that they are competitive with purely statistical methods at short horizons?
{}It is no secret that modern quantitative trading heavily relies on statistical methods such as data mining, machine learning, clustering algorithms, etc. However, after all, quantitative trading is a secretive field and resources on how things are done in practice are at best scarce.\footnote{\, Thus, we are unaware of another paper discussing the material herein at short horizons.} The purpose of these notes is to discuss a systematic quantitative framework -- in what is intended to be a ``pedagogical" fashion -- for building what we refer to as {\em statistical industry classifications}, solely based on stock returns and no additional extraneous data. Under the hood we have clustering algorithms. However, picking a clustering algorithm -- and we will see that some work better than others -- is insufficient. E.g., what should we cluster? Correlations? Returns? The answer turns out to be neither and stems from quantitative trading intuition, which is not something one expects to find in machine learning books. We discuss various nuances in constructing statistical industry classifications, and it is those nuances that make a sizable difference. Quant trading is all about detail.
{}One motivation for considering statistical industry classifications -- apart from the evident, to wit, the fact that they differ from ``fundamental" industry classifications and are widely used in quant trading -- is scenarios where ``fundamental" industry classifications are unavailable (or are of subpar quality). This could be in emerging or smaller markets, or even in the U.S. if the underlying trading portfolios are relatively small and a ``fundamental" industry classification produces too fragmented a grouping. However, perhaps an equally -- if not more -- important motivation is application of these methods to returns for ``instruments" other than stocks, e.g., quantitative trading alphas, for which there is no analog of a ``fundamental" industry classification \cite{Billion}. We will keep this in mind below.\footnote{\, Optimizing weights in alpha portfolios has its own nuances \cite{Billion}; however, the methods we discuss here are readily portable to alpha returns as they are purely statistical. Here we backtests them (see below) on stock returns as the historical data is readily available. Alpha return time series are highly proprietary, so publishing backtests is not feasible.}
{}In Section \ref{sec2} we briefly review some generalities of (binary) ``fundamental" industry classifications to set up the framework for further discussion. Next, in Section \ref{sec.stat.clust} we address the issue of what to cluster. We discuss why clustering correlations is suboptimal, and why so is directly clustering returns. We argue that returns should be normalized before clustering and give an explicit prescription for such normalization. We then discuss how to construct single-level and multilevel (hierarchical -- e.g., BICS has 3 levels: sectors, industries and sub-industries) statistical industry classifications together with some tweaks (e.g., cross-sectionally demeaning returns at less granular levels). Many clustering algorithms such as k-means are not deterministic. This can be a nuisance. We give an explicit prescription for aggregating classifications from multiple samplings, which in fact improves stability and performance. We discuss algorithms for ``bottom-up" (most granular to least granular level), ``top-down" (least granular to most granular level) and ``relaxation" (hierarchical agglomerative) clustering, together with their ``pros" and ``cons".
{}In Section \ref{sec.backtests} we discuss detailed backtests of the various algorithms in Section \ref{sec.stat.clust} and subsequent sections utilizing the intraday alphas and backtesting procedure described in \cite{Het} by using the resultant multilevel statistical industry classifications for building heterotic risk models. The backtests unequivocally suggest that there is structure in the return time series beyond what is captured by simple principal component analysis and clustering adds value. However, clustering still cannot compete with ``fundamental" industry classifications in terms of performance due to inherent out-of-sample instabilities in any purely statistical algorithm.
{}In Section \ref{sec.dyn} we take it a step further and give a prescription for fixing the number of clusters at each level using the methods discussed in \cite{StatRM}, including eRank (effective rank) defined in \cite{RV}. We also discuss a heuristic for fixing the number of levels, albeit we empirically observe that the number of levels is not as influential as the number of clusters, at least in our backtests. We take this even further in Section \ref{sec.hybrid}, where we give an algorithm for improving a ``fundamental" industry classification via further clustering large sub-industries (using BICS nomenclature) at the most granular level via statistical industry classification algorithms we discuss here thereby increasing granularity and improving performance. We briefly conclude in Section \ref{sec.concl} and outline some ideas.
{}We give the R source code for our algorithms in Appendix \ref{app.A} (multilevel ``bottom-up" clustering, dynamical cluster numbers), Appendix \ref{app.B} (multilevel ``top-down" clustering) and Appendix \ref{app.C} (``relaxation" clustering). Appendix \ref{app.D} contains legalese.
\section{Industry Classification}\label{sec2}
{}An industry classification is based on a similarity criterion: stocks' membership in ``groups" or ``clusters" such as sectors, industries, sub-industries, etc. -- the nomenclature varies from one industry classification scheme to another. Commonly used industry classifications such as GICS, BICS, ICB, NAICS, SIC, etc., are based on fundamental/economic data (such as companies' products and services and more generally their revenue sources, suppliers, competitors, partners, etc.). Such industry classifications are essentially independent of the pricing data and, if well-built, tend to be rather stable out-of-sample as companies seldom jump industries.\footnote{\, However, there is variability in the performance of different industry classifications.}
{}An industry classification can consist of a single level: $N$ tickers labeled by $i=1,\dots,N$ are grouped into $K$ ``groups" -- let us generically call them ``clusters" -- labeled by $A=1,\dots,K$. So, we have a map $G:\{1,\dots,N\}\mapsto\{1,\dots,K\}$ between stocks and ``clusters".\footnote{\, Here we are assuming that each stock belongs to one and only one ``cluster". Generally, this assumption can be relaxed thereby allowing for ``conglomerates" that belong to multiple sub-industries, industries, sectors, etc. However, this is not required for our purposes here.} More generally, we can have a hierarchy with multiple levels. We can schematically represent this via: Stocks $\rightarrow$ Level-1 ``Clusters" $\rightarrow$ Level-2 ``Clusters" $\rightarrow \dots \rightarrow$ Level-$P$ ``Clusters". Let us label these $P$ levels by $\mu = 1,\dots, P$. Level-1 is the most granular level with $N$ stocks grouped into $K_1$ ``clusters". The Level-1 ``clusters" are in turn grouped into $K_2$ Level-2 ``clusters", where $K_2 < K_1$, and so on, Level-$P$ being least granular.\footnote{\, The branches in this hierarchy tree are assumed to have equal lengths. More generally, we can have branches of nonuniform lengths. However, shorter branches can always be extended to the length of the longest branch(es) by allowing single-element (including single-stock) ``clusters".} Thus, consider
BICS\footnote{\, Bloomberg Industry Classification System.} as an illustrative example, which has a 3-level hierarchy: Stocks $\rightarrow$ Sub-industries $\rightarrow$ Industries $\rightarrow$ Sectors. (Here ``Sub-industries" is the most granular level, while ``Sectors" is the least granular level.) So, we have: $N$ stocks labeled by $i=1,\dots,N$; $K$ sub-industries labeled by $A=1,\dots,K$; $F$ industries labeled by $a=1,\dots,F$; and $L$ sectors labeled by $\alpha=1,\dots,L$. Let $G$ be the map between stocks and sub-industries, $S$ be the map between sub-industries and industries, and $W$ be the map between industries and sectors:
\begin{eqnarray}\label{G.map}
&&G:\{1,\dots,N\}\mapsto\{1,\dots,K\}\\
&&S:\{1,\dots,K\}\mapsto\{1,\dots,F\}\label{S.map}\\
&&W:\{1,\dots,F\}\mapsto\{1,\dots,L\}\label{W.map}
\end{eqnarray}
The beauty of such ``binary" industry classifications (generally, with $P$ levels) is that the ``clusters" (in the case of BICS, sub-industries, industries and sectors) can be used to identify blocks (sub-matrices) in the sample correlation matrix $\Psi_{ij}$ of stock returns.\footnote{\, And this is useful in constructing risk models for portfolio optimization \cite{Het}.} E.g., for sub-industries the binary matrix $\delta_{G(i), A}$ defines such blocks.
\section{Statistical Clustering}\label{sec.stat.clust}
{}What if we do not have access to industry classifications based on fundamental data\footnote{\, Commercially available industry classifications such as GICS and ICB come at nontrivial cost. The underlying SIC data is available from SEC for free, albeit only by company names, not by ticker symbols. It takes considerable effort to download this data and transform it into an actual industry classification. Alternatively, it can be purchased from commercial providers.} or one is unavailable for the stock universe we wish to trade? Can we build an industry classification from pricing data, i.e., directly from stock returns? After all, intuitively, the time series of returns contains information about how correlated the stocks are. Can we extract it and transform it into an industry classification?
{}The answer is yes, but it is tricky. The key issue is that correlations between stocks typically are highly unstable out-of-sample. A naive attempt at constructing an industry classification based on stock returns may produce an industry classification with subpar performance. Our goal here is to discuss how to mitigate the out-of-sample instability by building statistical industry classifications based on clustering quantities other than returns. But first let us discuss clustering itself.
\subsection{K-means}
{}A popular clustering algorithm is k-means \cite{Steinhaus}, \cite{Lloyd1957}, \cite{Forgy}, \cite{MacQueen}, \cite{Hartigan}, \cite{HartWong}, \cite{Lloyd1982}. The basic idea behind k-means is to partition $N$ observations into $K$ clusters such that each observation belongs to the cluster with the nearest mean. Each of the $N$ observations is actually a $d$-vector, so we have an $N \times d$ matrix $X_{is}$, $i=1,\dots,N$, $s=1,\dots,d$. Let $C_a$ be the $K$ clusters, $C_a = \{i| i\in C_a\}$, $a=1,\dots,K$. Then k-means attempts to minimize
\begin{equation}\label{k-means}
g = \sum_{a=1}^K \sum_{i \in C_a} \sum_{s=1}^d \left(X_{is} - Y_{as}\right)^2
\end{equation}
where
\begin{equation}\label{centers}
Y_{as} = {1\over n_a} \sum_{i\in C_a} X_{is}
\end{equation}
are the cluster centers (i.e., cross-sectional means),\footnote{\, Throughout this paper ``cross-sectional" refers to ``over the index $i$".} and $n_a = |C_a|$ is the number of elements in the cluster $C_a$. In (\ref{k-means}) the measure of ``closeness" is chosen to be the Euclidean distance between points in ${\bf R}^d$, albeit other measures are possible.
{}One ``drawback" of k-means is that it is not a deterministic algorithm. Generically, there are copious local minima of $g$ in (\ref{k-means}) and the algorithm only guarantees that it will converge to a local minimum, not the global one. Being an iterative algorithm, k-means starts with a random or user-defined set of the centers $Y_{as}$ at the initial iteration. However, as we will see, this ``drawback" actually adds value.
\subsection{What to Cluster?}
{}So, what should we cluster to construct statistical industry classifications? I.e., what should we pick as our matrix $X_{is}$ in (\ref{k-means})? It is tempting to somehow use pair-wise stock correlations. However, the sample correlation matrix $\Psi_{ij}$ computed based on the time series of stock returns is highly unstable out-of-sample.\footnote{\, The sample correlation matrix contains less information than the underlying time series of returns. Thus, it knows nothing about serial means of returns, only deviations from these means.} So, what if we identify $X_{is}$ with the time series of the underlying stock returns? Let $R_{is}$ be these stock returns, where $s = 1,\dots,d$ now is interpreted as labeling the observations in the time series (e.g., trading days). Further, for definiteness, let $s=1$ correspond to the most recent observation. Now we can build a statistical industry classification by applying k-means to $X_{is} = R_{is}$. Intuitively this makes sense: we are clustering stocks based on how close the returns are to the centers (i.e., within-cluster cross-sectional means) of the clusters they belong to. However, this is a suboptimal choice.
{}Indeed, this can be understood by observing that, in the context of stock returns, a priori there is no reason why the centers $Y_{as}$ in (\ref{centers}) should be computed with equal weights. We can think of the clusters $C_a$ as portfolios of stocks, and $Y_{as}$ as the returns for these portfolios. Therefore, based on financial intuition, we may wish to construct these portfolios with nonuniform weights. Furthermore, upon further reflection, it become evident that clustering returns make less sense than it might have appeared at first. Indeed, stock volatility is highly variable, and its cross-sectional distribution is not even quasi-normal but highly skewed, with a long tail at the higher end -- it is roughly log-normal. Clustering returns does not take this skewness into account and inadvertently we might be clustering together returns that are not at all highly correlated solely due to the skewed volatility factor.
{}A simple solution is to cluster the normalized returns ${\widetilde R}_{is} = R_{is} / \sigma_i$, where $\sigma_i^2 = \mbox{Var}(R_{is})$ is the serial variance. This way we factor out the skewed volatility factor. Indeed, $\mbox{Cov}({\widetilde R}_i, {\widetilde R}_j) = \mbox{Cor}(R_i, R_j) = \Psi_{ij}$ (we suppress the index $s$ in the serial covariance Cov and correlation Cor) is the sample correlation matrix with $|\Psi_{ij}| \leq 1$. However, as we will see below, clustering ${\widetilde R}_{is}$, while producing better results than clustering $R_{is}$, is also suboptimal. Here are two simple arguments why this is so.
{}Clusters $C_a$ define $K$ portfolios whose weights are determined by what we cluster. When we cluster $X_{is} = R_{is}$, the centers are $Y_{as} = \mbox{Mean}(R_{is}|i\in C_a)$, i.e., we have equal weights $\omega_i\equiv 1$ for the aforesaid $K$ portfolios, and we group $R_{is}$ (at each iterative step in the k-means algorithm) by how close these returns are to these equally-weighted portfolios. However, equally-weighted portfolios themselves are suboptimal. So are portfolios weighted by $\omega_i\equiv 1/\sigma_i$, which is what we get if we cluster $X_{is} = {\widetilde R}_{is}$, where the centers are $Y_{as} = \mbox{Mean}(R_{is}/\sigma_i|i\in C_a)$. Thus, portfolios that maximize the Sharpe ratio \cite{Sharpe1994} are weighted by inverse variances:\footnote{\, More precisely, this is the case in the approximation where the sample covariance matrix is taken to be diagonal. In the context of clustering it makes sense to take the diagonal part of the sample covariance matrix as the full sample covariance matrix is singular for clusters with $n_a > d-1$. Even for $n_a \leq d-1$ the sample covariance matrix, while invertible, has highly out-of-sample unstable off-diagonal elements. In contrast, the diagonal elements, i.e., sample variances $\sigma_i^2$, are much more stable, even for short lookbacks. So it makes sense to use them in defining $\omega_i$.} $\omega_i=1/\sigma_i^2$. We get such portfolios if we cluster $X_{is} = {\widehat R}_{is}$, where ${\widehat R}_{is} = R_{is}/\sigma_i^2$, so the centers are $Y_{as} = \mbox{Mean}(R_{is}/\sigma_i^2|i\in C_a)$. Clustering ${\widehat R}_{is}$, as we will see, indeed outperforms clustering ${\widetilde R}_{is}$. Can we understand this in a simple, intuitive fashion?
{}By clustering ${\widetilde R}_{is} = R_{is}/\sigma_i$, we already factor out the volatility dependence. So, why would clustering ${\widehat R}_{is} = R_{is}/\sigma_i^2$ work better? Clustering ${\widetilde R}_{is}$ essentially groups together stocks that are (to varying degrees) highly correlated in-sample. However, there is no guarantee that they will remain as highly correlated out-of-sample. Intuitively, it is evident that higher volatility stocks are more likely to get uncorrelated with their respective clusters. This is essentially why suppressing by another factor or $\sigma_i$ in ${\widehat R}_{is}$ (as compared with ${\widetilde R}_{is}$) leads to better performance: inter alia, it suppresses contributions of those volatile stocks into the cluster centers $Y_{is}$.
\subsubsection{A Minor Tweak}\label{sub.norm.ret}
{}So, we wish to cluster ${\widehat R}_{is} = R_{is}/\sigma_i^2$. There is a potential hiccup with this in practice. If some stocks have very low volatilities, we could have large ${\widehat R}_{is}$ for such stocks. To avoid any potential issues with computations, we can ``smooth" this out via (MAD = mean absolute deviation):\footnote{\, This is one possible tweak. Others produce similar results.}
\begin{eqnarray}\label{tweak}
&&{\widehat R}_{is} = {R_{is} \over {\sigma_i u_i}}\\
&&u_i = {\sigma_i\over v_i}\\
&&v_i = \exp(\mbox{Median}(\ln(\sigma_i)) - 3~\mbox{MAD}(\ln(\sigma_i)))
\end{eqnarray}
and for all $v_i < 1$ we set $v_i \equiv 1$. This is the definition of ${\widehat R}_{is}$ we use below (unless stated otherwise). Furthermore, Median($\cdot$) and MAD($\cdot$) above are cross-sectional.
\subsection{Multilevel Clustering}
{}If we wish to construct a single-level statistical industry classification, we can simply cluster ${\widehat R}_{is}$ defined in (\ref{tweak}) into $K$ clusters via k-means. What if we wish to construct a multilevel statistical industry classification (see Section \ref{sec2})? We discuss two approaches here, which we can refer to as ``bottom-up" and ``top-down".\footnote{\, W.r.t. classification levels; ``bottom-up" should not be confused with agglomerative clustering.}
\subsubsection{Bottom-Up Clustering}\label{sub.bottomup}
{}Say we wish to construct a $P$-level classification. We can construct it as a sequence: $K_1 \rightarrow K_2 \rightarrow \dots \rightarrow K_P$ ($K_1 > K_2 > \dots > K_P$), where we first construct the most granular level with $K_1$ clusters, then we cluster these $K_1$ clusters into fewer $K_2$ clusters and so on, until we reach the last and least granular level with $K_P$ clusters. Given\footnote{\, We will discuss what these cluster number ``should" be below.} the integers $K_1,\dots,K_P$, the question is what to use as the returns at each step. Let these returns be $[R(\mu)]_{i(\mu), s}$ (i.e., we cluster $[R(\mu)]_{i(\mu), s}$ into $K_\mu$ clusters via k-means), where $\mu =1,\dots,P$, $i(\mu) = 1,\dots,K_{\mu-1}$, and we have conveniently defined $K_0 = N$, so $i(1)$ is the same index as $i$. As above, we can take $[R(1)]_{is} = {\widehat R}_{is}$. What about $[R(\mu)]_{i(\mu), s}$ at higher levels $\mu > 1$? We have some choices here. Let $C_{a(\mu)} = \{i(\mu)| i(\mu)\in C_{a(\mu)}\}$, $a(\mu)=1,\dots,K_\mu$ be the clusters at each level $\mu$. I.e., the index $a(\mu)$ is the same as the index $i(\mu+1)$ for $0 < \mu < P$. Then we can take (in the second line below $2 < \mu \leq P$)
\begin{eqnarray}
&&[R(2)]_{i(2), s} = \mbox{Mean}(R^\prime_{is}| i \in \{1,\dots,N\})\\
&&[R(\mu)]_{i(\mu), s} = \mbox{Mean}([R^\prime(\mu-1)]_{i(\mu-1),s}| i(\mu-1) \in C_{a(\mu - 1)})\label{Rmu}
\end{eqnarray}
where we can take (i) $R^\prime_{is} = R_{is}$ and $[R^\prime(\mu)]_{i(\mu),s} = [R(\mu)]_{i(\mu),s}$, or (ii) $R^\prime_{is} = {\widehat R}_{is}$ and $[R^\prime(\mu)]_{i(\mu),s} = [{\widehat R}(\mu)]_{i(\mu),s}$, where (Var($\cdot$) below is the serial variance)
\begin{eqnarray}
&&[{\widehat R}(\mu)]_{i(\mu),s} = {[R(\mu)]_{i(\mu),s} \over \sigma_{i(\mu)}^2}\\
&&\sigma_{i(\mu)}^2 = \mbox{Var}([R(\mu)]_{i(\mu),s})
\end{eqnarray}
These two definitions produce very similar results in our backtests (see below).
\subsubsection{Another Minor Tweak}\label{sub.demean}
{}In the bottom-up clustering approach we just discussed above, the higher level clusters tend to be highly correlated with each other. I.e., the corresponding cluster returns have a prominent ``market" (or ``overall") mode\footnote{\, See, e.g., \cite{CFM}, \cite{Billion}.} component in them. That is, averages of pair-wise ($i(\mu)\neq j(\mu)$) serial correlations $[\Psi(\mu)]_{i(\mu),j(\mu)} = \mbox{Cor}([R(\mu)]_{i(\mu), s}, [R(\mu)]_{j(\mu), s})$ at higher levels $\mu > 1$ are substantial.\footnote{\, Consequently, there is a large gap between the first $[\lambda(\mu)]^{(1)}$ and higher $[\lambda(\mu)]^{(p)}$, $p>1$, eigenvalues of $[\Psi(\mu)]_{i(\mu),j(\mu)}$; the eigenvalues are ordered decreasingly: $[\lambda(\mu)]^{(1)} > [\lambda(\mu)]^{(2)} > \dots$} To circumvent this, we can simply cross-sectionally demean the returns at higher levels, i.e., for $\mu > 1$ we substitute $[R(\mu)]_{i(\mu), s}$ by $[R(\mu)]_{i(\mu), s} - \mbox{Mean}([R(\mu)]_{i(\mu), s} | i(\mu)\in C_a(\mu))$. However, cross-sectional demeaning at level-1 ($\mu = 1$) leads to worse performance. Intuitively, we can understand this as follows. Demeaning at the most granular level removes the ``market" mode.\footnote{\, This essentially drops the 1st principal component from the spectral decomposition of $\Psi_{ij}$.} Unlike higher-level returns $[R(\mu)]_{i(\mu), s}$, $\mu > 1$, the level-1 returns are not all that highly correlated with each other, so it pays to keep the ``market" mode intact as, e.g., high-beta stocks statistically are expected to cluster together, while low-beta stocks are expected to cluster differently. So, the upshot is that we demean the returns at higher levels, but not level-1 returns.
\subsubsection{Aggregating Multiple Samplings}\label{sub.aggr}
{}As mentioned above, k-means is not a deterministic algorithm. Unless the initial centers are preset, the algorithm starts with random initial centers and converges to a different local minimum in each run. There is no magic bullet here: trying to ``guess" the initial centers is not any easier than ``guessing" where, e.g., the global minimum is. So, what is one to do? One possibility is to simply live with the fact that every run produces a different answer. The question then one must address in a given context is whether the performance in an actual application is stable from one such random run to another, or if it is all over the place. As we will see below, in our backtests, happily, the performance is extremely stable notwithstanding the fact that each time k-means produces a different looking industry classification.
{}So, this could be the end of the story here. However, one can do better. The idea is simple. What if we {\em aggregate} different industry classifications from multiple runs (or samplings) into one? The question is how. Suppose we have $M$ runs ($M \gg 1$). Each run produces an industry classification with $K$ clusters. Let $\Omega^r_{ia} = \delta_{G^r(i),a}$, $i=1,\dots,N$, $a=1,\dots,K$ (here $G^r:\{1,\dots,N\} \mapsto \{1,\dots,K\}$ is the map between the stocks and the clusters),\footnote{\, For terminological definiteness here we focus on the level-1 clusters; it all straightforwardly applies to all levels. Also, the superscript $r$ in $\Omega^r_{ia}$ and $G^r(i)$ is an index, not a power.} be the binary loadings matrix from each run labeled by $r=1,\dots,M$. Here we are assuming that somehow we know how to properly order (i.e., align) the $K$ clusters from each run. This is a nontrivial assumption, which we will come back to momentarily. However, assuming, for a second, that we know how to do this, we can aggregate the loadings matrices $\Omega^r_{ia}$ into a single matrix ${\widetilde \Omega}_{ia} = \sum_{r=1}^M \Omega^r_{ia}$. Now, this matrix does not look like a binary loadings matrix. Instead, it is a matrix of occurrence counts, i.e., it counts how many times a given stock was assigned to a given cluster in the process of $M$ samplings. What we need to construct is a map $G$ such that one and only one stock belongs to each of the $K$ clusters. The simplest criterion is to map a given stock to the cluster in which ${\widetilde\Omega}_{ia}$ is maximal, i.e., where said stock occurs most frequently. A caveat is that there may be more than one such clusters. A simple criterion to resolve such an ambiguity is to assign said stock to the cluster with most cumulative occurrences (i.e., we take $q_a = \sum_{i=1}^N {\widetilde\Omega}_{ia}$ and assign this stock to the cluster with the largest $q_a$, if the aforesaid ambiguity occurs). In the unlikely event that there is still an ambiguity, we can try to do more complicated things, or we can simply assign such a stock to the cluster with the lowest value of the index $a$ -- typically, there is so much noise in the system that dwelling on such minutiae simply does not pay off.
{}However, we still need to tie up a loose end, to wit, our assumption that the clusters from different runs were somehow all aligned. In practice each run produces $K$ clusters, but i) they are not the same clusters and there is no foolproof way of mapping them, especially when we have a large number of runs; and ii) even if the clusters were the same or similar, they would not be ordered, i.e., the clusters from one run generally would be in a different order than clusters from another run.
{}So, we need a way to ``match" clusters from different samplings. Again, there is no magic bullet here either. We can do a lot of complicated and contrived things with not much to show for it at the end. A simple pragmatic solution is to use k-means to align the clusters from different runs. Each run labeled by $r=1,\dots,M$, among other things, produces a set of cluster centers $Y^r_{as}$. We can ``bootstrap" them by row into a $(KM) \times d$ matrix ${\widetilde Y}_{{\widetilde a}s} = Y^r_{as}$, where ${\widetilde a} = a + (r - 1)K$ takes values ${\widetilde a}=1,\dots,(KM)$. We can now cluster ${\widetilde Y}_{{\widetilde a}s}$ into $K$ clusters via k-means. This will map each value of ${\widetilde a}$ to $\{1,\dots,K\}$ thereby mapping the $K$ clusters from each of the $M$ runs to $\{1,\dots,K\}$. So, this way we can align all clusters. The ``catch" is that there is no guarantee that each of the $K$ clusters from each of the $M$ runs will be uniquely mapped to one value in $\{1,\dots,K\}$, i.e., we may have some empty clusters at the end of the day. However, this is fine, we can simply drop such empty clusters and aggregate (via the above procedure) the smaller number of $K^\prime < K$ clusters. I.e., at the end we will end up with an industry classification with $K^\prime$ clusters, which might be fewer than the target number of clusters $K$. This is not necessarily a bad thing. The dropped clusters might have been redundant in the first place. Another evident ``catch" is that even the number of resulting clusters $K^\prime$ is not deterministic. If we run this algorithm multiple times, we will get varying values of $K^\prime$. However, as we will see below, the aggregation procedure improves performance in our backtests and despite the variability in $K^\prime$ is also very stable from run to run. In Appendix \ref{app.A} we give the R source code for bottom-up clustering with various features we discuss above, including multilevel industry classification, the tweaks, and aggregation.\footnote{\, The source code in Appendix \ref{app.A}, Appendix \ref{app.B} and Appendix \ref{app.C} hereof is not written to be ``fancy" or optimized for speed or in any other way. Its sole purpose is to illustrate the algorithms described in the main text in a simple-to-understand fashion. See Appendix \ref{app.D} for some legalese.}
\subsubsection{Top-Down Clustering}\label{sub.topdown}
{}Above we discussed bottom-up clustering. We can go the other way around and do top-down clustering. I.e., we can construct a $P$-level classification as a sequence $K_P \rightarrow K_{P-1} \rightarrow \dots \rightarrow K_2 \rightarrow K_1$ (as before, $K_1 > K_2 > \dots > K_P$). More conveniently, we start with the entire universe of stocks and cluster ${\widehat R}_{is}$, $i=1,\dots,N$, into $L_P = K_P$ clusters. At level-$(P-1)$, we cluster each level-$P$ cluster $C_{a(P)} = \{i|i\in C_{a(P)}\}$, $a(P) = 1,\dots,K_P$, into $L_{P-1}$ clusters. We do this by clustering the returns ${\widehat R}_{is}$, $i\in C_{a(P)}$ via k-means into $L_{P-1}$ clusters.\footnote{\, More generally, we can nonuniformly cluster each level-$P$ cluster with its own $[L(a(P))]_{P-1}$.} At level-$(P-2)$, we cluster each level-$(P-1)$ cluster $C_{a(P-1)} = \{i|i\in C_{a(P-1)}\}$, $a(P-1) = 1,\dots,K_{P-1}$, into $L_{P-2}$ clusters. We do this by clustering the returns ${\widehat R}_{is}$, $i\in C_{a(P-1)}$ via k-means into $L_{P-2}$ clusters. And so on.\footnote{\, Note that, in contrast to bottom-up clustering, because here we are going ``backwards", it is convenient to label the elements of each cluster at each level using the index $i$, which labels stocks.} In the zeroth approximation, $K_{P-1} = L_{P-1} K_P$, $K_{P-2} = L_{P-2} K_{P-1}$, and so on, so $K_1 = K_* = \prod_{\mu=1}^P L_\mu$. However, if at some level-$\mu$ we have some cluster $C_{a(\mu)}$ with $n_a(\mu) \leq L_\mu$, then we leave this cluster intact and do not cluster it, i.e., we ``roll it" forward unchanged. Therefore, we can have $K_1 < K_*$ at the most granular level-1. Also, instead of simply clustering via a single-sampling k-means, as above we can aggregate multiple samplings. Then at any level-$\mu$ we can end up clustering a given cluster $C_{a(\mu)}$ into $L_\mu$ or fewer clusters. Note, since here we work directly with the returns ${\widehat R}_{is}$, in contrast to the bottom-up approach, no cross-sectional demeaning is warranted at any level. In Appendix \ref{app.B} we give the R source code for top-down clustering, including with aggregation over multiple samplings.
\subsubsection{Relaxation Clustering}\label{relax}
{}Instead of k-means, which is nondeterministic, we can use other types of clustering, e.g., hierarchical agglomerative clustering. Let us focus on a 1-level classification here as we can always generalize it to multilevel cases as above. So, we have $N$ stocks, and we wish to cluster them into $K$ clusters. If $K$ is not preset, we can use SLINK \cite{SLINK}, etc. (see, e.g., \cite{HAC}). If we wish to preset $K$, then we can use a similar approach, except that it must be tweaked such that all observations are somehow squeezed into $K$ clusters. We give the R code for one such algorithm in Appendix \ref{app.C}. Basically, it is a relaxation algorithm which, as above, clusters ${\widehat R}_{is}$ (not $R_{is}$). The distance $D(i,j)$ between two $d$-vectors ${\widehat R}_{is}$ and ${\widehat R}_{js}$ is simply the Euclidean distance in ${\bf R}^d$. The initial cluster contains $i_1$ and $j_1$ with the smallest distance. If some $i_2$ and $j_2$ (such that $i_2\neq i_1$, $i_2\neq j_1$, $j_2\neq i_1$ and $j_2\neq j_1$) are such that $D(i_2,j_2)$ is smaller than the lesser of $D(i_1, \ell)$ and $D(j_1, \ell)$ for all $\ell$ ($\ell\neq i_1$ and $\ell\neq j_1$), then $i_2$ and $j_2$ form the second cluster. Otherwise $\ell$ that minimizes $D(i_1, \ell)$ or $D(j_1, \ell)$ is added to the first cluster. This is continued until there are $K$ clusters. Once we have $K$ clusters, we can only add to these clusters.\footnote{\, A brute force algorithm where at each step rows and columns are deleted from the matrix $D(i,j)$ is too slow. The R source code we give in Appendix \ref{app.C} is substantially more efficient than that. However, it is still substantially slower than the k-means based algorithms we discuss above.}
\section{Backtests}\label{sec.backtests}
{}Let us backtest the above algorithms for constructing statistical industry classification by utilizing the same backtesting procedure as in \cite{Het}. The remainder of this subsection very closely follows most parts of Section 6 thereof.\footnote{\, We ``rehash" it here not to be repetitive but so that our presentation here is self-contained.}
\subsection{Notations}
{}Let $P_{is}$ be the time series of stock prices, where $i=1,\dots,N$ labels the stocks, and $s=1,2,\dots$ labels the trading dates, with $s=1$ corresponding to the most recent date in the time series. The superscripts $O$ and $C$ (unadjusted open and close prices) and $AO$ and $AC$ (open and close prices fully adjusted for splits and dividends) will distinguish the corresponding prices, so, e.g., $P^C_{is}$ is the unadjusted close price. $V_{is}$ is the unadjusted daily volume (in shares). Also, for each date $s$ we define the overnight return as the previous-close-to-open return:
\begin{equation}\label{c2o.ret}
E_{is} = \ln\left({P^{AO}_{is} / P^{AC}_{i,s+1}}\right)
\end{equation}
This return will be used in the definition of the expected return in our mean-reversion alpha. We will also need the close-to-close return
\begin{equation}\label{c2c.ret}
R_{is} = \ln\left({P^{AC}_{is} / P^{AC}_{i,s+1}}\right)
\end{equation}
An out-of-sample (see below) time series of these returns will be used in constructing the risk models. All prices in the definitions of $E_{is}$ and $R_{is}$ are fully adjusted.
{}We assume that: i) the portfolio is established at the open\footnote{\, This is a so-called ``delay-0" alpha: the same price, $P^O_{is}$ (or adjusted $P^{AO}_{is}$), is used in computing the expected return (via $E_{is}$) and as the establishing fill price.} with fills at the open prices $P^O_{is}$; ii) it is liquidated at the close on the same day -- so this is a purely intraday alpha -- with fills at the close prices $P^C_{is}$; and iii) there are no transaction costs or slippage -- our aim here is not to build a realistic trading strategy, but to test {\rm relative} performance of various statistical industry classifications. The P\&L for each stock
\begin{equation}
\Pi_{is} = H_{is}\left[{P^C_{is}\over P^O_{is}}-1\right]
\end{equation}
where $H_{is}$ are the {\em dollar} holdings. The shares bought plus sold (establishing plus liquidating trades) for each stock on each day are computed via $Q_{is} = 2 |H_{is}| / P^O_{is}$.
\subsection{Universe Selection}\label{sub.univ}
{}For the sake of simplicity,\footnote{\, In practical applications, the trading universe of liquid stocks typically is selected based on market cap, liquidity (ADDV), price and other (proprietary) criteria.} we select our universe based on the average daily dollar volume (ADDV) defined via (note that $A_{is}$ is out-of-sample for each date $s$):
\begin{equation}\label{ADDV}
A_{is}= {1\over m} \sum_{r=1}^m V_{i, s+r}~P^C_{i, s+r}
\end{equation}
We take $m=21$ (i.e., one month), and then take our universe to be the top 2000 tickers by ADDV. To ensure that we do not inadvertently introduce a universe selection bias, we rebalance monthly (every 21 trading days, to be precise). I.e., we break our 5-year backtest period (see below) into 21-day intervals, we compute the universe using ADDV (which, in turn, is computed based on the 21-day period immediately preceding such interval), and use this universe during the entire such interval. We do have the survivorship bias as we take the data for the universe of tickers as of 9/6/2014 that have historical pricing data on http://finance.yahoo.com (accessed on 9/6/2014) for the period 8/1/2008 through 9/5/2014. We restrict this universe to include only U.S. listed common stocks and class shares (no OTCs, preferred shares, etc.) with BICS (Bloomberg Industry Classification System) sector assignments as of 9/6/2014.\footnote{\, The choice of the backtesting window is intentionally taken to be exactly the same as in \cite{Het} to simplify various comparisons, which include the results therefrom.} However, as discussed in detail in Section 7 of \cite{MeanRev}, the survivorship bias is not a leading effect in such backtests.\footnote{\, Here we are after the {\em relative outperformance}, and it is reasonable to assume that, to the leading order, individual performances are affected by the survivorship bias approximately equally as the construction of all alphas and risk models is ``statistical" and oblivious to the universe.}
\subsection{Backtesting}\label{sub.back}
{}We run our simulations over a period of 5 years (more precisely, 1260 trading days going back from 9/5/2014, inclusive). The annualized return-on-capital (ROC) is computed as the average daily P\&L divided by the intraday investment level $I$ (with no leverage) and multiplied by 252. The annualized Sharpe Ratio (SR) is computed as the daily Sharpe ratio multiplied by $\sqrt{252}$. Cents-per-share (CPS) is computed as the total P\&L in cents (not dollars) divided by the total shares traded.
\subsection{Optimized Alphas}\label{sub.opt}
{}The optimized alphas are based on the expected returns $E_{is}$ optimized via Sharpe ratio maximization using heterotic risk models \cite{Het} based on statistical industry classifications we are testing.\footnote{\, In \cite{Het} BICS is used for the industry classification. Here we simply plug in the statistical industry classification instead of BICS. In the case of a single-level industry classification, we can either add the second level consisting of the ``market" with the $N\times 1$ unit matrix as the loadings matrix; or, equivalently, we can use the option {\tt{\small{mkt.fac = T}}} in the R function {\tt{\small{qrm.het()}}} in Appendix B of \cite{Het}, which accomplishes this internally.} We compute the heterotic risk model covariance matrix $\Gamma_{ij}$ every 21 trading days (same as for the universe). For each date (we omit the index $s$) we maximize the Sharpe ratio subject to the dollar neutrality constraint:
\begin{eqnarray}
&&{\cal S} = {\sum_{i=1}^N H_i~E_i\over {\sqrt{\sum_{i,j=1}^N {\Gamma}_{ij}~H_i~H_j}}} \rightarrow \mbox{max}\\
&&\sum_{i=1}^N H_i = 0\label{d.n.opt}
\end{eqnarray}
In the absence of bounds, the solution is given by
\begin{equation}\label{H.opt}
H_i = -\eta \left[\sum_{j = 1}^N {\Gamma}^{-1}_{ij}~E_j - \sum_{j=1}^N {\Gamma}^{-1}_{ij}~{{\sum_{k,l=1}^N {\Gamma}^{-1}_{kl}~E_l}\over{\sum_{k,l = 1}^N {\Gamma}^{-1}_{kl}}}\right]
\end{equation}
where ${\Gamma}^{-1}$ is the inverse of ${\Gamma}$, and $\eta > 0$ (mean-reversion alpha) is fixed via (we set the investment level $I$ to \$20M in our backtests)
\begin{equation}
\sum_{i=1}^N \left|H_i\right| = I
\end{equation}
Note that (\ref{H.opt}) satisfies the dollar neutrality constraint (\ref{d.n.opt}).
{}In our backtests we impose position bounds (which in this case are the same as trading bounds as the strategy is purely intraday) in the Sharpe ratio maximization:
\begin{equation}\label{liq}
|H_{is}| \leq 0.01~A_{is}
\end{equation}
where $A_{is}$ is ADDV defined in (\ref{ADDV}). In the presence of bounds computing $H_i$ requires an iterative procedure and we use the R code in Appendix C of \cite{Het}.
\subsection{Simulation Results}
{}Table \ref{table.try1.100.30.10} summarizes simulation results for 11 independent runs for the ``bottom-up" 3-level statistical industry classification with $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ (see Subsection \ref{sub.bottomup}). Despite the nondeterministic nature of the underlying k-means algorithm, pleasantly, the backtest results are very stable. Table \ref{table.try100.100} summarizes simulation results for 11 independent runs for the ``bottom-up" single-level statistical industry classification with the target number of clusters $K=100$ based on aggregating 100 samplings (so the actual number of resultant clusters $K^\prime$ can be smaller than $K$ -- see Subsection \ref{sub.aggr}). Again, the backtest results are very stable. Table \ref{table.try100.100.30.10} summarizes simulation results for 23 independent runs for the ``bottom-up" 3-level statistical industry classification with the target number of clusters $K_1=100$, $K_2 = 30$ and $K_3 = 10$ based on aggregating 100 samplings (so the actual number of resultant clusters $K_\mu^\prime$ can be smaller than $K_\mu$, $\mu=1,2,3$ -- see Subsection \ref{sub.aggr}). The first 15 (out of 23) runs correspond to {\tt{\small{norm.cl.ret = F}}} (this corresponds to choice (i) after Equation (\ref{Rmu}) in Subsection \ref{sub.bottomup}), while the other 8 runs correspond to {\tt{\small{norm.cl.ret = T}}} (this corresponds to choice (ii) after said Equation); see the function {\tt{\small{qrm.stat.ind.class.all()}}} in Appendix \ref{app.A}. The aforesaid stability persists to these cases as well. Table \ref{table.num.clust} summarizes the number of actual clusters in a statistical industry classification obtained via aggregating 100 samplings. The target numbers of clusters in a 3-level hierarchy are $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$, as in Table \ref{table.try100.100.30.10}.
{}Table \ref{table.try1.topdown} summarizes simulation results for ``top-down" 3-level statistical industry classifications obtained via a single sampling in each run, with 3 runs for each $L_\mu$. The 3-vector $L_\mu$, $\mu=1,2,3$, is defined in Subsection \ref{sub.topdown}. Recall that in the zeroth approximation the number of clusters at the most granular level-1 is $K_1 = L_1 L_2 L_3$; however, the actual value can be lower due to the reasons explained in Subsection \ref{sub.topdown}. Here too we observe substantial stability. Table \ref{table.try100.topdown} summarizes simulation results for ``top-down" 3-level statistical industry classifications obtained via aggregating 100 samplings in each run, with 3 runs for each $L_\mu$. Stability persists.
{}From the above results it is evident that aggregating multiple samplings on average improves both performance and stability. Furthermore, not surprisingly, decreasing granularity worsens the Sharpe ratio. 3-level classifications outperform single-level classifications.\footnote{\, Also, ``bottom-up" by construction uses more information than and outperforms ``top-down".} Above we mentioned that clustering ${\widehat R}_{is} = R_{is}/\sigma_i^2$ outperforms clustering ${\widetilde R}_{is} = R_{is}/\sigma_i$, which in turn outperforms clustering $R_{is}$. Thus, a random run for the ``bottom-up" 3-level classification with $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ based on clustering $R_{is}$ using a single sampling produced a typical performance with ROC = 41.885\%, SR = 15.265 and CPS = 1.889 (cf. Table \ref{table.try1.100.30.10}). A random run for the ``bottom-up" 3-level classification with $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ based on clustering ${\widetilde R}_{is}$ using a single sampling produced a typical performance with ROC = 42.072\%, SR = 15.840 and CPS = 1.973 (cf. Table \ref{table.try1.100.30.10}).\footnote{\, Table \ref{table.try1.100.30.10} is based on clustering ${\widehat R}_{is}$ defined via (\ref{tweak}). However, clustering ${\widehat R}^*_{is} = R_{is}/\sigma^2_i$ produces essentially the same results. Thus, a random run for the ``bottom-up" 3-level classification with $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ based on clustering ${\widehat R}^*_{is}$ via aggregating 100 samplings produced a typical performance with ROC = 41.707\%, SR = 16.220 and CPS = 2.091 (cf. Table \ref{table.try100.100.30.10}).}
{}In contrast to nondeterministic k-means based algorithms, the relaxation algorithm (Subsection \ref{relax}) is completely deterministic. We run it using the code in Appendix \ref{app.C} to obtain a 3-level classification with the target numbers of clusters $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ (as in the ``bottom-up" cases, we cross-sectionally demean the level-2 and level-3 returns, but not the level-1 returns). The simulation results are sizably worse than for k-means based algorithms: ROC = 41.266\%, SR = 15.974 and CPS = 1.990. How come? Intuitively, this is not surprising. All such relaxation mechanisms (hierarchical agglomerative algorithms) start with a ``seed", i.e., the initial cluster picked based on some criterion. In Subsection \ref{relax} this is the first cluster containing the pair $(i_1,j_1)$ that minimized the Euclidean distance. However, generally this choice is highly unstable out-of-sample, hence underperformance. In contrast, k-means is much more ``statistical", especially with aggregation.
\section{How to Fix Cluster Numbers?}\label{sec.dyn}
{}Thus far we have picked the number of clusters $K_\mu$ as well as the number of levels $P$ ``ad hoc".\footnote{\, Here we focus on the k-means based ``bottom-up" and ``top-down" algorithms. As discussed above, the relaxation algorithm underperforms the k-means based algorithms.} Can we fix them ``dynamically"? If we so choose, here we can do a lot of complicated things. Instead, our approach will be based on pragmatism (rooted in financial considerations) and simplicity.\footnote{\, A variety of methods for fixing the number of clusters have been discussed in other contexts. See, e.g., \cite{Rousseeuw}, \cite{Goutte}, \cite{Sugar}, \cite{Lleiti}, \cite{DeAmorim}.} As can be surmised from Tables \ref{table.try100.100} and \ref{table.try100.100.30.10}, the number of levels does not make it or break it in our context. What is more important is the number of clusters. So, suppose we have a given number of levels $P > 1$. Let us start by asking, what should $K_1$ (most granular level) and $K_P$ (least granular level) be? In practice, the number of stocks $N>d-1$, so the sample correlation matrix $\Psi_{ij}$ is singular. (In fact, in most practical applications $N\gg d-1$.) We can model it via statistical risk models \cite{StatRM}. These are factor models obtained by truncating the spectral decomposition of $\Psi_{ij}$
\begin{equation}
\Psi_{ij} = \sum_{a = 1}^{d - 1} \lambda^{(a)}~V_i^{(a)}~V_j^{(a)}
\end{equation}
via the first $d - 1$ principal components $V_i^{(a)}$ (only $d - 1$ eigenvalues $\lambda^{(a)}$ are positive, $\lambda^{(1)} > \lambda^{(2)} > \dots, \lambda^{(d-1)} > 0$, while the rest of the eigenvalues $\lambda^{(a)} \equiv 0$, $a\geq d$) to the first $F$ principal components ($F < d - 1$) and compensating the deficit on the diagonal (as $\Psi_{ii}\equiv 1$) by adding diagonal specific (idiosyncratic) variance $\xi_i^2$:
\begin{equation}
\Gamma_{ij} = \xi_i^2~\delta_{ij} + \sum_{a = 1}^F \lambda^{(a)}~V_i^{(a)}~V_j^{(a)}
\end{equation}
I.e., we approximate $\Psi_{ij}$ (which is singular) via $\Gamma_{ij}$ (which is positive-definite as all $\xi_i^2 > 0$ and are fixed from the requirement that $\Gamma_{ii} \equiv 1$). The question then is, what should $F$ be? One simple (``minimization" based) algorithm for fixing $F$ is given in \cite{Het}. Another, even simpler algorithm recently proposed in \cite{StatRM}, is based on eRank (effective rank) defined below.\footnote{\, For prior works on fixing $F$, see, e.g., \cite{Connor} and \cite{Bai}.}
\subsection{Effective Rank}
{}Thus, we simply set (here $\mbox{Round}(\cdot)$ can be replaced by $\mbox{floor}(\cdot) = \lfloor\cdot\rfloor$)
\begin{equation}\label{eq.eRank}
F = \mbox{Round}(\mbox{eRank}(\Psi))
\end{equation}
Here $\mbox{eRank}(Z)$ is the effective rank \cite{RV} of a symmetric semi-positive-definite (which suffices for our purposes here) matrix $Z$. It is defined as
\begin{eqnarray}
&&\mbox{eRank}(Z) = \exp(H)\\
&&H = -\sum_{a=1}^L p_a~\ln(p_a)\\
&&p_a = {\lambda^{(a)} \over \sum_{b=1}^L \lambda^{(b)}}
\end{eqnarray}
where $\lambda^{(a)}$ are the $L$ {\em positive} eigenvalues of $Z$, and $H$ has the meaning of the (Shannon a.k.a. spectral) entropy \cite{Campbell60}, \cite{YGH}.
{}The meaning of $\mbox{eRank}(Z)$ is that it is a measure of the effective dimensionality of the matrix $Z$, which is not necessarily the same as the number $L$ of its positive eigenvalues, but often is lower. This is due to the fact that many returns can be highly correlated (which manifests itself by a large gap in the eigenvalues) thereby further reducing the effective dimensionality of the correlation matrix.
\subsection{Fixing $K_\mu$}\label{sub.dyn}
{}There is no magic bullet here. It just has to make sense. Intuitively, it is natural to identify the number of clusters $K_P$ at the least granular level with the number of factors $F$ in the context of statistical risk models.\footnote{\, The number of factors $F$ essentially measures the effective number of degrees of freedom in the underlying time series of returns $R_{is}$. Hence identification of $K_P$ with this number.} In the following, we will therefore simply take
\begin{equation}
K_P = \mbox{Round}(\mbox{eRank}(\Psi))
\end{equation}
Adding more granular levels explores deeper substructures in the time series of returns based on the closeness criterion. In this regard, we can fix the number of clusters $K_1$ at the most granular level as follows. The average number of stocks per cluster at level-1 is $N_1 = N/K_1$ (we are being cavalier with rounding). Assume for a second that the number of stocks in each cluster is the same and equal $N_1$. If $N_1 > d-1$, then the sub-matrices $\Psi_{ij}$, $i,j\in C_{a(1)}$ (recall that $C_{a(1)}$, $a(1) = 1,\dots,K_1$, are the level-1 clusters) are singular. For $N_1 \leq d - 1$ they are nonsingular. Therefore, intuitively, it is natural to fix $K_1$ by requiring that $N_1 = d - 1$. Restoring rounding, in the following we will set
\begin{equation}
K_1 = \mbox{Round}(N / (d-1))
\end{equation}
What about $K_\mu$, $1 < \mu < P$? Doing anything overly complicated here would be overkill. Here is a simple prescription (assuming $K_1 > K_P$):\footnote{\, I.e., $K_\mu$ are (up to rounding) equidistant on the log scale. For $P=3$ the ``midpoint" $K_2 = \sqrt{K_1 K_P}$ is simply the geometric mean. With this prescription, we can further fix $P$ via some heuristic, e.g., take maximal $P$ such that the difference $K_{P-1} - K_P\geq \Delta$, where $\Delta$ is preset, say, $\Delta = K_P$. For $K_1 = 100$ and $K_P = 10$, this would give us $P=4$ with $K_2 = 46$ and $K_3 = 22$.}
\begin{equation}
K_\mu = \left[K_1^{P - \mu}~K_P^{\mu - 1}\right]^{1\over{P-1}},~~~\mu=1,\dots,P
\end{equation}
We give the R source code for building ``bottom-up" statistical industry classifications using this prescription in Appendix \ref{app.A}. Table \ref{table.dyn} summarized simulation results for $P =2,3,4,5$. It is evident that the number of levels is not a driver here. The results are essentially the same as for $K_1 = 100$ (recall that $N = 2000$ and $d = 21$ in our case) in Tables \ref{table.try100.100} and \ref{table.try100.100.30.10}. Table \ref{table.try100.K} isolates the $K$ dependence and suggests that the performance peaks around $K = 100$. Again, there is no magic bullet here.\footnote{\, Note from Table \ref{table.try100.K} that too little granularity lowers the Sharpe ratio due to insufficient coverage of the risk space, while too much granularity lowers cents-per-share due to overtrading.}
\subsection{Comparisons}
{}Let us compare the (very stable) results we obtained for statistical industry classifications with two ``benchmarks": statistical risk models \cite{StatRM} and heterotic risk models with BICS used as the industry classification \cite{Het}. More precisely, statistical risk models in \cite{StatRM} were built based on the sample correlation matrix $\Psi_{ij}$, which is equivalent to basing them on normalized returns ${\widetilde R}_{is} = R_{is}/\sigma_i$. If we use the eRank based algorithm for fixing the number of statistical risk factors $F$, then the performance is ROC = 40.777\%, SR = 14.015 and CPS = 1.957 \cite{StatRM}.\footnote{\, In \cite{StatRM} rounding is to 2 decimals, while here we round to 3 decimals.} However, as we argued above, it makes more sense to build models using ${\widehat R}_{is} = R_{is}/\sigma_i^2$. So, we should compare our results here with the statistical risk models based on ${\widehat R}_{is}$. To achieve this, we can simply replace the line {\tt{\small{tr <- apply(ret, 1, sd)}}} in the R function {\tt{\small{qrm.erank.pc(ret, use.cor = T)}}} in Appendix A of \cite{StatRM} by {\tt{\small{tr <- apply(ret, 1, sd) / apply(qrm.calc.norm.ret(ret), 1, sd)}}}, where the R function {\tt{\small{qrm.calc.norm.ret()}}} is given in Appendix \ref{app.A} hereof. The performance is indeed better: ROC = 40.878\%, SR = 14.437 and CPS = 2.018. So, the k-means based clustering algorithms still outperform statistical risk models, which implies that going beyond the $F$ statistical factors adds value, i.e., there is more structure in the data than is captured by the principal components alone. However, statistical industry classifications still sizably underperform heterotic risk models based on BICS \cite{Het}:\footnote{\, Here we use the results from \cite{HetPlus}, which slightly differ from those in \cite{Het}, where rounding down (as opposed to simply rounding) was employed.} ROC = 49.005\%, SR = 19.230 and CPS = 2.365. Clearly, statistical industry classifications are not quite on par with industry classifications such as BICS, which are based on fundamental/economic data (such as companies' products and services and more generally their revenue sources, suppliers, competitors, partners, etc.). Such industry classifications are essentially independent of the pricing data and, if well-built, tend to be rather stable out-of-sample as companies seldom jump industries. In contrast, statistical industry classifications by nature are less stable out-of-sample. However, they can add substantial value when ``fundamental" industry classifications are unavailable, including for returns other than for stocks, e.g., quantitative trading alphas \cite{Billion}.
{}Finally, before we close this section, let us discuss the ``top-down" classifications with dynamically determined numbers of clusters $K_\mu$. More precisely, recall that in this case we work with the vector $L_\mu$ (see Subsection \ref{sub.topdown}). The code we used in the ``bottom-up" case (Appendix \ref{app.A}) can be used in this case as well (via a parameter choice). A random (and typical) run with $P=3$ gives ROC = 41.657\%, SR = 15.897 and CPS = 2.079, while another such run with $P=4$ gives ROC = 41.683\%, SR = 15.697 and 2.073. These results are in line with our results in Table \ref{table.try100.topdown}.
\section{Hybrid Industry Classification}\label{sec.hybrid}
{}One application of a statistical industry classification is to use it as a means for improving a ``fundamental" industry classification such as BICS, GICS, etc. Thus, a ``fundamental" classification at the most granular level can have overly large sub-industries, using the BICS nomenclature for definiteness. One way to deal with such large sub-industries is to further cluster them using statistical industry classification methods discussed above. Let us illustrate this using BICS as an example.
{}Table \ref{table.bics.summary} summarizes top 10 most populous (by stock counts) sub-industries in one of our 2000 stock backtesting portfolios. For comparison, the stock count summary across all 165 sub-industries in this sample is Min = 1, 1st Qu. = 3, Median = 8, Mean = 12.12, 3rd Qu. = 15, Max = 94, StDev = 14.755, MAD = 8.896 (see Table \ref{table.num.clust} for notations). So, we have some ``large" sub-industries, which are outliers.
{}We can further split these large sub-industries into smaller clusters using our ``bottom-up" clustering algorithm. In fact, it suffices to split them using a single-level algorithm. We give the R code for improving an existing ``fundamental" industry classification using our statistical industry classification algorithm in Appendix \ref{app.A}. The idea is simple. Let us label the sub-industries (the most granular level) in the ``fundamental" industry classification via $A = 1,\dots, K_*$. Let $N_A$ be the corresponding stock counts. Let
\begin{equation}
\kappa_A = \mbox{Round}(N_A / (d-1))
\end{equation}
We then split each sub-industry with $\kappa_A > 1$ into $\kappa_A$ clusters. Table \ref{table.improve} summarizes the simulation results for 14 runs. This evidently improves performance. Table \ref{table.ind.counts} gives summaries of top 10 most populous sub-industries before and after statistical clustering based on 60 datapoints at the end of each 21-trading-day interval in our backtests (recall that we have $1260 = 60 \times 21$ trading days -- see Section \ref{sec.backtests}). The average numbers of sub-industries are 165.45 before and 184.1 after clustering.
\section{Concluding Remarks}\label{sec.concl}
{}In this paper we discuss all sorts of nuances in constructing statistical industry classifications. Under the hood we have clustering algorithms. However, it is precisely those nuances that make a sizable difference. E.g., if we naively cluster $R_{is}$, we get a highly suboptimal result compared with clustering ${\widetilde R}_{is} = R_{is}/\sigma_i$, which in turn underperforms clustering ${\widehat R}_{is} = R_{is}/\sigma_i^2$. In this regard, let us tie up a ``loose end" here: what if we cluster ${\overline R}_{is} = R_{is}/\sigma_i^3$? It underperforms clustering ${\widehat R}_{is}$. Thus, a typical run for a 3-level ``bottom-up" classification with target cluster numbers $K_1 = 100$, $K_2 = 30$ and $K_3 = 10$ based on clustering ${\overline R}_{is}$ and aggregating 100 samplings produces the following: ROC = 40.686, SR = 15.789 and CPS = 2.075.
{}So, suppressing returns $R_{is}$ by $\sigma_i^2$ indeed appears to be optimal -- for the intuitive reasons we discussed above. We saw the same in the context of statistical risk models. In this regard, it would be interesting to explore this avenue in the context of heterotic risk models \cite{Het} and the more general (heterotic CAPM) construction of \cite{HetPlus}. In the latter framework, it would be interesting to utilize an aggregated (non-binary) matrix ${\widetilde \Omega}_{ia}$ (see Subsection \ref{sub.aggr}). These ideas are outside of the scope hereof and we hope to return to them elsewhere.
| 77a7fc1347ee6eabede2aa3eb8b4eb325a684451 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and Statement of Results}
Let $f \in S_k(\Gamma_0(N),\chi)$ be a newform of weight $k$, level $N$, and nebentypus $\chi$. Associated to $f$ is an $L$-function $L(s, f)$, which can be normalized so that the completed $L$-function \begin{equation}
\Lambda(s, f) := N^{s/2} \int_0^\infty f(iy)y^{s-1} dy = \left(\frac{\sqrt{N}}{2\pi}\right)^s \Gamma(s) L(s, f)
\end{equation}
satisfies the functional equation \begin{equation} \label{fe} \Lambda(s, f) = \epsilon(f) \Lambda(k-s, \bar{f}) \end{equation} for some $\epsilon(f) \in \mathbb C$ with $|\epsilon(f)| = 1.$
The \emph{period polynomial} associated to $f$ is the degree $k-2$ polynomial defined by
\begin{equation}\label{defnperiod}
r_f(z) \defeq \int_0^{i \infty} f(y)(y-z)^{k-2} dy.
\end{equation}
By the binomial theorem, we have
\begin{align*}
r_f(z) &= i^{k-1}N^{-\frac{k-1}{2}} \sum_{n = 0}^{k-2} \binom{k-2}{n}(\sqrt{N}iz)^n \Lambda(k-1-n, f), \\
&= -\frac{(k-2)!}{(2\pi i)^{k-1}}\sum_{n = 0}^{k-2} \frac{(2\pi iz)^n}{n!} L(k-1-n,f).
\end{align*}
Thus, $r_f(z)$ is the generating function for the special values $L(1,f), L(2,f) \ldots, L(k-1,f)$ of the $L$-function associated to $f$. For background on period polynomials, we refer the reader to \cites{period1,period2,period3,period4,period5}.
When $k \ge 3$, the period polynomial $r_f(z)$ is nonconstant, so one can consider where the roots of $r_f(z)$ are located. To this end, we use the functional equation \eqref{fe} to observe that
\[
\overline{r_f(z)} = -(\sqrt{N}i\overline z)^{k-2} \epsilon(f)^{-1}r_f\left(\frac{1}{N\overline{z}}\right).
\]
Thus, if $\rho$ is a root of $r_f(z)$, then $\frac{1}{N\bar\rho}$ is also a root. Much like the behavior of the nontrivial zeroes of $L(s, f)$ predicted by the Generalized Riemann Hypothesis, one can consider whether all the roots of $r_f(z)$ lie on the curve of symmetry of the roots: in this case, the circle $|\rho| = 1/\sqrt N$. It is natural to expect the following conjecture, which is supported by extensive numerical evidence.
\begin{conjecture*}[``Riemann Hypothesis" for period polynomials]
Let $f \in S_k(\Gamma_0(N),\chi)$ be a newform. Then, the roots of $r_f(z)$ all lie on the circle $|\rho| = 1/\sqrt N$.
\end{conjecture*}
El-Guindy and Raji~\cite{elg} proved this for Hecke eigenforms on $SL_2(\mathbb Z)$ with full level ($N=1$, for which the circle of symmetry is $|z| = 1$). They were inspired by the work of Conrey, Farmer, and \.{I}mamo\u{g}lu~\cite{farmer}, who showed an analogous result for the odd parts of these period polynomials, again with full level.
Recent work by Jin, Ma, Ono, and Soundararajan~\cite{ono} proved the conjecture for all newforms of even weight $k \ge 4$ and trivial nebentypus. They also showed that the roots of $r_f(z)$ are equidistributed on the circle of symmetry for sufficiently large $N$ or $k$. Using similar methods, L{\"o}brich, Ma, and Thorner~\cite{thorner} proved an analogous result for polynomials generating special values of $L(s,\mathcal M)$ for a sufficiently well-behaved class of motives $\mathcal M$ with odd weight and even rank.
In this paper, we generalize the methods of \cite{ono} to prove the conjecture for all but possibly finitely many newforms.
\begin{theorem}\label{thm:main}
The ``Riemann Hypothesis'' for period polynomials holds for all but possibly finitely many newforms with weight $k \ge 3$ and nontrivial nebentypus.
\end{theorem}
\begin{remark} Note that for $k < 3$, the period polynomial is a constant. Therefore, Theorem~\ref{thm:main} is essentially the best result for which one could hope, since an effective computation can check that Theorem~\ref{thm:main} also holds for the finitely many possible exceptions. We denote the set of these finitely many newforms as $\mathcal S$, which consists of the following:
\begin{enumerate}
\item For $k=5$, all newforms with level $N \le 10331$.
\item For $k \ge 6$, all newforms with level $N \le C(k)$, where $C(k)$ is a constant given by tables at the end of Section \ref{table1} and \ref{table2}.
\end{enumerate}
We know of no counterexamples to Theorem \ref{thm:main}.
\end{remark}
We also show that the roots of $r_f(z)$ are equidistributed on the circle of symmetry for sufficiently large $N$ or $k$.
\begin{theorem}
\label{thm:equi}
Let $f \in S_k(\Gamma_0(N),\chi)$ be a newform of weight $k \ge 4$, level $N$, and nebentypus $\chi$ such that $f \notin \mathcal S$. Then, the following are true:
\begin{enumerate}[label={\upshape(\roman*)}]
\item Suppose that $k = 4$, and let $z_1, z_2$ denote the roots of $r_f(z)$. Then for any real $\varepsilon > 0$,
\[
\arg z_1 -\arg z_2 \equiv \pi + O_\varepsilon(N^{-\frac{1}{4} + \varepsilon}) \pmod{2\pi},
\]
where the implied constant depends only on $\varepsilon$ and is effectively computable.
\item Suppose that $k=5$. There exists $c_f \in \mathbb R$ such that the arguments of the roots of $r_f(z)$ can be written as
\[
c_f + \theta_{\ell} + O_{\varepsilon}\left(\frac{1}{N^{\frac{1}{2} -\varepsilon}}\right) \pmod{2\pi}, \hspace{40pt} 0 \le \ell \le 2,
\]
where $\theta_{\ell}$ denotes the unique solution $\on{mod}$ $2\pi$ of
\[
\frac{k-2}{2}\theta_{\ell} - \frac{2\pi}{\sqrt N} \sin \theta_\ell = \ell \pi,
\]
and the implied constant depends only on $\varepsilon$ and is effectively computable.
\item Suppose that $k >5$. There exists $c_f \in \mathbb R$ such that the arguments of the roots of $r_f(z)$ can be written as
\[
c_f + \theta_{\ell} + O\left(\frac{1}{2^{k/2}\sqrt N}\right) \pmod{2\pi}, \hspace{40pt} 0 \le \ell \le k-3,
\]
Here, $\theta_{\ell}$ is the unique solution $\on{mod}$ $2\pi$ to the equation
\[
\frac{k-2}{2}\theta_{\ell} - \frac{2\pi}{\sqrt N} \sin \theta_\ell = \ell \pi,
\]
and the implied constant is absolute and effectively computable.
\end{enumerate}
\end{theorem}
In Section~\ref{sec:prelims}, we introduce notation and lemmas that we will be using in our proof. In Section~\ref{sec:lowweight}, we will prove our main results for $k=3, 4,$ and $5$ using ad hoc arguments. For larger $k$, we prove Theorem~\ref{thm:main} in Section~\ref{sec:evenweight} (the case of $k$ even) and Section~\ref{sec:oddweight} (the case of $k$ odd), and we prove Theorem~\ref{thm:equi} in Section~\ref{sec:equi}. Finally, in Section~\ref{sec:numerical}, we detail our Sage computations suggesting that the roots of the period polynomial of the newform
\[
f(\tau) = q+10q^3+64q^4+74q^5+O(q^6) \in S_7\left(\Gamma_0(11),\left(\frac{-11}{\bullet}\right)\right)
\]
are all on the circle $|z|=1/\sqrt{11}$. This newform $f$ is in our finite set $\mathcal S$ of possible exceptions, which suggests that Theorem~\ref{thm:main} should be true even for newforms in $\mathcal S$.
\section{Preliminaries}\label{sec:prelims}
Throughout this section, we assume that $f \in S_k(\Gamma_0(N), \chi)$ is a newform of weight $k \ge 3$, level $N$, and arbitrary nebentypus $\chi.$ We note that the nebentypus character will be essentially invisible throughout our proof, other than the fact that it determines the level of $f$. We now define some notation related to $r_f(z)$ and prove lemmas about the values of $\Lambda(s,f)$ and $L(s,f)$ along the real line. The lemmas will be very similar in spirit to those proven in \cite{ono}.
Define $\delta$ to satisfy $\delta^2 = \epsilon(f)^{-1}.$ Now, define \begin{equation} \label{tpoly} t_f(z) = i^{-k+1}N^\frac{k-1}{2}z^{-\frac{k-2}{2}}\delta \cdot r_f\left(\frac{z}{i\sqrt{N}}\right) = \sum_{n = 0}^{k-2} \binom{k-2}{n}z^{n-\frac{k-2}{2}} \delta\Lambda(k-1-n, f), \end{equation} where $z^\frac{1}{2}$ denotes $r^\frac{1}{2}e^{\theta i/2}$ for $z = re^{\theta i}$ and $0 \le \theta < 2\pi.$ Using \eqref{fe}, one can compute
\[
\overline{t_f(z)} = t_f\left(\frac{1}{\bar{z}}\right).
\]
Therefore, if $t_f(z) = 0$, then $t_f\left(\frac{1}{\bar{z}}\right) = 0.$ Additionally, for $|z| = 1$, we also have $\overline{t_f(z)} = t_f(z)$, so $t_f(z)$ is real for $|z| = 1.$ Note that $t_f(z) = 0$ if and only if $r_f(\frac{z}{i\sqrt{N}}) = 0.$ Therefore, to prove Theorems \ref{thm:main} and \ref{thm:equi}, it suffices to show that all roots of $t_f(z)$ lie on the circle $|z| = 1$ and are equidistributed.
We will require the following monotonicity result.
\begin{lemma} \label{mono} We have \[ \Big|\Lambda\Big(\frac{k}{2}, f\Big)\Big| < \Big|\Lambda\Big(\frac{k}{2}+1, f\Big)\Big| < \dots < \Big|\Lambda\Big(\frac{k}{2}+j, f\Big)\Big| < \cdots \] Also, for all $0 < a < b$, \[ \Big| \Lambda\Big(\frac{k+1}{2} + a, f\Big) \Big| < \Big| \Lambda\Big(\frac{k+1}{2} + b, f\Big) \Big|. \] \end{lemma}
\begin{proof} As $\Lambda(s, f)$ is entire of order $1$, we apply the Hadamard factorization theorem to write
\[ \Lambda(s, f) = e^{A+Bs} \prod_\rho (1-s/\rho)e^{s/\rho}, \]
where the sum is taken over all roots $\rho$ of $\Lambda(s, f)$. By \cite[Proposition 5.7(3)]{ink}, we have that
\[
\on{Re}(B) = -\sum_{\rho} \on{Re}\Big({\frac{1}{\rho}}\Big).
\]
Note that $\frac{k-1}{2} < \on{Re}{\rho} < \frac{k+1}{2}$. This implies that $|1 - \frac{s}{\rho}|$ is increasing for $s \ge \frac{k+1}{2}$ and $|1 - \frac{k/2}{\rho} | < |1 - \frac{k/2+1}{\rho}|$, from which the lemma follows.
\end{proof}
We also prove a useful inequality on ratios of $L$-function values.
\begin{lemma} \label{lratio} For all $0 < a < b$, we have \[ \left|\frac{L(\frac{k+1}{2}+a, f)}{L(\frac{k+1}{2}+b, f)} - 1\right| \le \frac{\zeta(1+a)^2}{\zeta(1+b)^2}-1. \] \end{lemma}
\begin{proof}
We have that
\[ \left|\frac{L(\frac{k+1}{2}+a, f)}{L(\frac{k+1}{2}+b, f)} - 1\right| = \left| \exp\left(-\int_a^b \frac{L'(\frac{k+1}{2}+s, f)}{L(\frac{k+1}{2}+s, f)} ds \right)-1 \right|. \]
If we express \[ -\frac{L'(s, f)}{L(s, f)} = \sum \frac{\Lambda_f(n)}{n^s}, \] then Deligne's bound on the eigenvalues of the Hecke operators on $S_k(\Gamma_0(N),\chi)$ states that
\begin{equation}\label{deligne23}
|\Lambda_f(n)| \le 2n^\frac{k-1}{2}\Lambda(n)
\end{equation}
where $\Lambda(n)$ denotes the von Mangoldt function; for a reference, see \cite[Theorem 2.32]{onobook}. Therefore, we have that
\begin{equation}\label{deligne}\left| \int_a^b \frac{L'(\frac{k+1}{2} + s, f)}{L(\frac{k+1}{2} + s, f)} \right| \le -2 \int_a^b \frac{\zeta'(1+s)}{\zeta(1+s)} ds = 2 \log \frac{\zeta(1+a)}{\zeta(1+b)}. \end{equation} Now the lemma follows from the inequality $|e^x-1| \le e^{|x|}-1.$
\end{proof}
Finally, we show a lemma that serve as our main means of proving Theorem \ref{thm:main} for period polynomials.
\begin{lemma}
\label{ivt}
Let $\on{sgn}(r)$ equal $-1$ for negative real numbers $r$, $1$ for positive real numbers $r$, and $0$ for $r = 0.$ If there exist real numbers $0 \le \theta_1 < \theta_2 < \dots < \theta_k < 2\pi$ such that either \[ \on{sgn}(t_f(e^{i\theta_j})) = (-1)^j \text{ for all } 1 \le j \le k, \] or \[ \on{sgn}(t_f(e^{i\theta_j})) = (-1)^{j+1} \text{ for all } 1 \le j \le k, \] then all solutions to $t_f(z) = 0$ satisfy $|z| = 1.$
\end{lemma}
\begin{proof}
First, $t_f(z)$ is real for $|z| = 1$, so $\on{sgn}(t_f(e^{i\theta_j}))$ is well defined. Now, by the Intermediate Value Theorem, there exist $\theta \in (\theta_j, \theta_{j+1})$ such that $t_f(e^{i\theta}) = 0$ for all $1 \le j \le k-1.$ This gives us $k-1$ roots of $t_f(z)$ that lie on $|z| = 1.$ When $k$ is even, we also get a root in the range $(\theta_k, \theta_1+2\pi)$ by the Intermediate Value Theorem. When $k$ is odd, we may redefine the square root in order to move the discontinuity into an interval outside of $[\theta_k, \theta_1 + 2\pi]$. This would only affect the sign of $t_f(z)$. By the intermediate value theorem, this shows the existence of a zero with argument in the range $[\theta_k, \theta_1 + 2\pi]$ as desired. As $t_f(z) = 0$ for at most $k$ values of $z$, the above argument shows that we have found all of them.
\end{proof}
\section{Proof for Weights 3, 4, and 5}\label{sec:lowweight}
Here we prove Theorem \ref{thm:main} and \ref{thm:equi} for $k = 3, 4,$ and $5.$
\subsection{The weight 3 case}
For $k = 3,$ (\ref{tpoly}) gives that \[ t_f(z) = \delta z^{-\frac{1}{2}}(\Lambda(2, f) + z\Lambda(1, f)). \] By \eqref{fe}, we know that \[ |\Lambda(2, f)| = |\Lambda(1, f)|, \] so the root of $t_f(z)$ lies on the unit circle.
\subsection{The weight 4 case}
For $k = 4$, we have \[ t_f(z) = \delta(z^{-1}\Lambda(3, f) + 2\Lambda(2, f) + z\Lambda(1, f)). \] Now, note that for $|z| = 1$, it follows that \[ \frac{1}{2}t_f(z) = \on{Re}(\delta(\Lambda(2, f) + z\Lambda(1, f))) = \on{Re}(\delta\Lambda(2, f)) + \on{Re}(\delta z\Lambda(1, f)).\] By Lemma \ref{mono}, we have that \[ |\on{Re}(\delta\Lambda(2, f))| \le |\Lambda(2, f)| < |\Lambda(3, f)| = |\Lambda(1, f)|, \] so there exist $2$ values of $z$ with $|z| = 1$ such that \[ \on{Re}(\delta z\Lambda(1, f)) = -\on{Re}(\delta\Lambda(2, f)), \] as desired.
In order to prove Theorem~\ref{thm:equi}(iii), we need to bound $|\Lambda(2, f)| / |\Lambda(1, f)|$. First, note that
\[
|\Lambda(1, f)| = |\Lambda(3, f)| \gg N^{\frac{3}{2}}.
\]
In order to bound $|\Lambda(2, f)|$, we appeal to the Phragm{\'e}n-Lindel{\"o}f Principle; specifically, see \cite[Lemma 5.2, Theorem 5.53]{ink} and apply \eqref{deligne23}. This allows to obtain for any $\epsilon > 0$
\[
|\Lambda(2, f)| \le \max_{t \in \mathbb{R}} |\Lambda(5/2 + \epsilon + it, f)| = \max_{t \in \mathbb{R}} N^{\frac{5}{4} +\half\epsilon}|L(5/2 + \epsilon + it, f)| \ll_{\epsilon} N^{\frac{5}{4}+\epsilon},
\]
Thus, we have that
\[
\frac{|\Lambda(2, f)|}{|\Lambda(1, f)|} \ll N^{-\frac{1}{4}+\epsilon}
\]
and the values of $z$ satisfying $t_f(z) = 0$ satisfy
\[
\arg z = \pm \frac{\pi}{2} + \arg(\overline{\delta \Lambda(1, f)}) + O(N^{-\frac{1}{4}+\epsilon})
\]
\subsection{The weight 5 case}
For $k = 5$, we have \[ t_f(z) = \delta(z^{-\frac{3}{2}}\Lambda(4, f) + 3z^{-\frac{1}{2}}\Lambda(3, f) + 3z^\frac{1}{2}\Lambda(2, f) + z^\frac{3}{2}\Lambda(1, f)). \]
Once again, for $|z| = 1$, we have
\[
\frac{1}{2}t_f(z) = \on{Re}(\delta(3z^\frac{1}{2}\Lambda(2, f) + z^\frac{3}{2}\Lambda(1, f))) = \on{Re}(3\delta z^\frac{1}{2}\Lambda(2, f)) + \on{Re}(\delta z^\frac{3}{2}\Lambda(1, f)).
\]
There exist three reals $0 \le \theta_1 < \theta_2 < \theta_3 < 2\pi$ such that $|\on{Re}(\delta (e^{i\theta_j})^\frac{3}{2} \Lambda(1, f))| = |\Lambda(1, f)|$ for $1 \le j \le 3$, and $\on{Re}(\delta (e^{i\theta_j})^\frac{3}{2} \Lambda(1, f))$ alternates in sign. Thus, by Lemma \ref{ivt}, we are done if we are able to show that
\[ |\on{Re}(3\delta z^\frac{1}{2}\Lambda(2, f))| < |\Lambda(1, f)|,
\]
which is equivalent to proving
\[
\frac{|\Lambda(3, f)|}{|\Lambda(4, f)|} < \frac{1}{3}.
\]
Let $0 < \epsilon < 1.$ By Lemmas \ref{mono} and \ref{lratio}, it follows that
\[ \frac{|\Lambda(3, f)|}{|\Lambda(4, f)|} \le \frac{|\Lambda(3+\epsilon, f)|}{|\Lambda(4, f)|} = \frac{|L(3+\epsilon, f)|}{|L(4, f)|} \left(\frac{2\pi}{\sqrt{N}}\right)^{1-\epsilon} \frac{\Gamma(3+\epsilon)}{\Gamma(4)} \le \frac{\zeta(1+\epsilon)^2}{\zeta(2)^2} \left(\frac{2\pi}{\sqrt{N}}\right)^{1-\epsilon} \frac{\Gamma(3+\epsilon)}{\Gamma(4)}.
\]
Choosing $\epsilon = 2/5$, the last expression is less than $\frac{1}{3}$ for $N \ge 10332$, which completes the proof for $k = 5.$
To show the desired equidistribution property, define $\theta_1$ and $\theta_2$ as above. Now let $\theta_{\pm} = (\theta_1 + \theta_2)/2 \pm \varepsilon$, for $\varepsilon > 0$ to be chosen later. Then, we see that
\[
|\on{Re} (\delta (e^{\theta_{\pm} i})^{\frac{3}{2}} \Lambda(1, f)| = |\Lambda(1, f)| \sin \Big(\frac{3}{2}\varepsilon\Big)
\]
with the sign of $\on{Re} (\delta (e^{i\theta_{\pm}})^{\frac{3}{2}} \Lambda(1, f)$ being different for $\varepsilon > 0$ and $\varepsilon < 0$.
If we can show that
\[
|\on{Re} (3\delta z^{\half} \Lambda(2, f)| \le \Big|\sin \Big(\frac{3}{2}\varepsilon\Big)\Big|\cdot|\Lambda(1, f)|
\]
then Lemma \ref{ivt} will show that the root has an argument lying between $\theta_{-}$ and $\theta_{+}$. By the bounding above, we only require
\[
\frac{\zeta(1+\epsilon)^2}{\zeta(2)^2} \left(\frac{2\pi}{\sqrt{N}}\right)^{1-\epsilon} \frac{\Gamma(3+\epsilon)}{\Gamma(4)} < \frac{1}{3} \sin \Big(\frac{3}{2}\varepsilon\Big).
\]
Choosing $\varepsilon = O(N^{-\half + \epsilon})$ suffices.
\section{Proof for Remaining Even Weights}\label{sec:evenweight}
In this section, we will show Theorem \ref{thm:main} for all even weights $k \ge 6.$ Throughout the section, we will restrict our attention to those $z$ such that $|z| = 1.$ For simplicity, let $m = \frac{k-2}{2}$, and define \begin{align} P_f(z) &= \frac{1}{2} \binom{2m}{m} \Lambda(m+1, f) + \sum_{n = 0}^{m-1} \binom{k-2}{n} z^{m-n} \delta \Lambda(n+1, f) \nonumber \\ \label{pfeven} &= \frac{1}{2} \binom{2m}{m} \Lambda(m+1, f) + (2m)!\left(\frac{\sqrt{N}}{2\pi}\right)^{2m+1} \delta^{-1} z^m \sum_{n = 0}^{m-1} \frac{1}{n!}\left(\frac{2\pi}{z\sqrt{N}}\right)^n \overline{L(2m+1-n, f)}. \end{align}
This satisfies
\[
t_f(z) = P_f(z) + \overline{P_f(z)} = 2 \on{Re}(P_f(z)).
\]
Next, define
\begin{align*} Q_f(z) &= \frac{1}{(2m)!} \left(\frac{2\pi}{\sqrt{N}}\right)^{2m+1}P_f(z) \\
&= \frac{1}{2(m!)^2}\left(\frac{2\pi}{\sqrt{N}}\right)^{2m+1} \Lambda(m+1, f) + \delta^{-1}z^m \sum_{n = 0}^{m-1}\frac{1}{n!}\left(\frac{2\pi}{z\sqrt{N}}\right)^n \overline{L(2m+1-n, f)}. \end{align*}
Note that $\on{sgn}(Q_f(z)) = \on{sgn}(P_f(z)).$ As in \cite{ono}, rewrite
\[
Q_f(z) = \delta^{-1} \overline{L(2m+1, f)} z^m \exp\left(\frac{2\pi}{z\sqrt{N}}\right) + S_1(z) + S_2(z) + S_3(z),
\]
where we define
\begin{align*}
&S_1(z) = \delta^{-1} \overline{L(2m+1, f)} z^m \sum_{n = 0}^{m-1} \frac{1}{n!} \left(\frac{2\pi}{z\sqrt{N}}\right)^n \left(\overline{\left(\frac{L(2m+1-n, f)}{L(2m+1, f)}\right)}-1 \right) \\
&S_2(z) = -\delta^{-1} \overline{L(2m+1, f)} z^m \sum_{n \ge m} \frac{1}{n!} \left(\frac{2\pi}{z\sqrt{N}}\right)^n \\
&S_3(z) = \frac{1}{2(m!)^2} \left(\frac{2\pi}{\sqrt{N}}\right)^{2m+1} \Lambda(m+1, f).
\end{align*}
For $z = e^{i\theta}$, note that
\begin{equation} \label{defC}
\arg\left(\delta^{-1} \overline{L(2m+1, f)} z^m \exp\left(\frac{2\pi}{z\sqrt{N}}\right)\right) = C + m\theta - \frac{2\pi \sin\theta}{\sqrt{N}},
\end{equation}
where $C$ is a fixed constant depending on $\delta$ and $L(2m+1, f).$ Therefore, we can pick $k$ values of $z$ on the circle $|z| = 1$ such that the previous expression has argument $\ell\pi$ for integers $\ell.$ The value of $Q_f(z)$ at these points have alternating positive and negative real part with magnitude at least $|L(2m+1, f)|\exp\left(-\frac{2\pi}{\sqrt{N}}\right)$. By Lemma \ref{ivt}, it suffices to show that
\[
|S_1(z)| + |S_2(z)| + |S_3(z)| < |L(2m+1, f)|\exp\left(-\frac{2\pi}{\sqrt{N}}\right).
\]
To bound $S_1(z)$, we use Lemma \ref{lratio} in the form $\Big|\frac{L(2m+1-n, f)}{L(2m+1, f)} \Big| \le \zeta(\frac{1}{2}+m-n)^2-1.$ This gives \[ |S_1(z)| \le |L(2m+1, f)| \sum_{n = 1}^{m-1} \frac{1}{n!} \left(\frac{2\pi}{\sqrt{N}}\right)^n (\zeta(1/2+m-n)^2-1). \] For the term $n = m-1$ in the above expression, we use the bound $\zeta(3/2)^2-1 \le 35/6$. For $0 \le n \le m-2$, note that $2^x(\zeta(1/2+x)^2-1)$ is decreasing for $x \ge 2.$ Therefore, for $0 \le n \le m-2$, we find that \[ \zeta(1/2 + m-n)^2-1 \le 2^{n-m}\cdot 4(\zeta(5/2)^2-1) \le \frac{16}{5}2^{n-m}. \] Now, we combine the above estimates with $S_2(z)$ to obtain \begin{align} \frac{|S_1(z)| + |S_2(z)|}{|L(2m+1, f)|} &\le \frac{16}{5}\sum_{n = 1}^{m-1} \frac{1}{n!} \left(\frac{2\pi}{\sqrt{N}}\right)^n \frac{2^n}{2^m} + \frac{17}{4} \frac{1}{(m-1)!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} + \sum_{n \ge m} \frac{1}{n!} \left(\frac{2\pi}{\sqrt{N}}\right)^n \nonumber \\ \label{s12} &\le \frac{16}{5}2^{-m} \left(\exp\left(\frac{4\pi}{\sqrt{N}}\right) - 1\right) + \frac{17}{4} \frac{1}{(m-1)!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1}. \end{align}
To finish, we estimate $|S_3(z)|$ using Lemma \ref{mono} and then \ref{lratio}. \begin{align} |S_3(z)| &\le \frac{1}{2(m!)^2} \left(\frac{2\pi}{\sqrt{N}}\right)^{2m+1} |\Lambda(m+1, f)| \le \frac{1}{2(m!)^2} \left(\frac{2\pi}{\sqrt{N}}\right)^{2m+1} |\Lambda(m+2, f)| \nonumber \\ &\le \frac{m+1}{2m!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} |L(m+2, f)| \le \frac{m+1}{2m!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} |L(2m+1, f)| \zeta(3/2)^2 \nonumber \\ \label{s3} &\le \frac{7}{2} \frac{m+1}{m!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} |L(2m+1, f)|. \end{align}
By using (\ref{s12}) and (\ref{s3}), it suffices to verify \begin{equation} \label{finaleven} \frac{16}{5}2^{-m} \left(\exp\left(\frac{4\pi}{\sqrt{N}}\right) - 1\right) + \frac{17}{4} \frac{1}{(m-1)!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} + \frac{7}{2} \frac{m+1}{m!} \left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} < \exp\left(-\frac{2\pi}{\sqrt{N}}\right). \end{equation}
For each value of $m$ in the first row on the following table, the value $N(m)$ is such that inequality \eqref{finaleven} holds for all $N\ge N(m)$. Note that the case $m = 1$ was done in Section \ref{sec:lowweight}.
\begin{center} \label{table1}
\begin{tabular} { |c || c | c |c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$m$& 29& 21& 18& 16& 14 & 13& 12& 11& 10 & 9&8&7&6&5&4&3&2 \\
\hline
$N(m)$& 1& 2& 3& 4& 5& 6& 7& 9& 11&14& 19& 27&41&69&142&433&5875 \\
\hline
\end{tabular}
\end{center}
Therefore, for all $m \ge 29$, $N(m) = 1$. This completes our proof of Theorem \ref{thm:main} for $k$ even.
\section{Proof for Remaining Odd Weights}\label{sec:oddweight}
In this section, we will show Theorem \ref{thm:main} for all odd weights $k \ge 7.$ As in the above section, we will restrict our attention to those $z$ such that $|z| = 1.$ For simplicity, let $m = \frac{k-3}{2}$, and define \begin{align} P_f(z) &= \sum_{n = 0}^{m} \binom{k-2}{n} z^{m-n+\frac{1}{2}} \delta \Lambda(n+1, f) \nonumber \\ \label{pfodd} &= (2m+1)! \left(\frac{\sqrt{N}}{2\pi}\right)^{2m+2}\delta^{-1} z^{m+\frac{1}{2}} \sum_{n = 0}^m \frac{1}{n!}\left(\frac{2\pi}{z\sqrt{N}}\right)^n \overline{L(2m+2-n, f)}, \end{align} so $t_f(z) = 2\on{Re}(P_f(z))$ As in the above section, define
\begin{align*}
Q_f(z) &= \frac{1}{(2m+1)!} \left(\frac{2\pi}{\sqrt{N}}\right)^{2m+2} P_f(z) = \delta^{-1} z^{m+\frac{1}{2}} \sum_{n = 0}^m \frac{1}{n!}\left(\frac{2\pi}{z\sqrt{N}}\right)^n \overline{L(2m+2-n, f)} \\
&= \overline{L(2m+2, f)}\delta^{-1}z^{m+\frac{1}{2}} \exp\left(\frac{2\pi}{z\sqrt{N}}\right) + S_1(z) + S_2(z) + S_3(z),
\end{align*}
where $S_1(z), S_2(z)$, and $S_3(z)$ are defined as follows.
\[
S_1(z) = \overline{L(2m+2, f)}\delta^{-1}z^{m+\frac{1}{2}} \sum_{n = 0}^{m-1} \frac{1}{n!}\left(\frac{2\pi}{z\sqrt{N}}\right)^n\left(\overline{\left(\frac{L(2m+2-n, f)}{L(2m+2, f)}\right)} - 1\right),
\]
\[
S_2(z) = -\overline{L(2m+2, f)}\delta^{-1}z^{m+\frac{1}{2}} \sum_{n \ge m} \frac{1}{n!} \left(\frac{2\pi}{z\sqrt{N}}\right)^n,
\]
\[
S_3(z) = \delta^{-1}z^{m+\frac{1}{2}} \frac{1}{m!} \left(\frac{2\pi}{z\sqrt{N}}\right)^m \overline{L(m+2, f)}.
\]
As in Section \ref{sec:evenweight}, it suffices to show that
\[
|S_1(z)| + |S_2(z)| + |S_3(z)| < |L(2m+2, f)|\exp\left(-\frac{2\pi}{\sqrt{N}}\right).
\]
The proof of this will proceed in a very similar way to that of the above section. Note that the function $2^x(\zeta(1+x)^2-1)$ is decreasing for $x \ge 1.$ By Lemma \ref{lratio}, for $0 \le n \le m-1$, we can bound \[ \Big|\frac{L(2m+2-n, f)}{L(2m+2, f)} - 1 \Big| \le \zeta(1+m-n)^2-1 \le 2^{n-m} \cdot 2(\zeta(2)^2-1) \le 2^{n-m} \cdot \frac{7}{2}. \] By Lemma \ref{lratio}, we have \begin{equation} \label{s12odd} \frac{|S_1(z)| + |S_2(z)|}{|L(2m+2, f)|} \le \frac{7}{2}2^{-m} \sum_{n = 1}^{m-1} \frac{1}{n!} \left(\frac{4\pi}{\sqrt{N}}\right)^n + \sum_{n \ge m} \frac{1}{n!} \left(\frac{2\pi}{\sqrt{N}}\right)^n \frac{2^n}{2^m} \le \frac{7}{2}2^{-m} \left(\exp\left(\frac{4\pi}{\sqrt{N}}\right) - 1\right). \end{equation} Now we use Lemma \ref{mono} to bound $|L(m+2, f)|$. \[ |L(m+2, f)| \le \frac{1}{(m+1)!}\left(\frac{2\pi}{\sqrt{N}}\right)^{m+2} |\Lambda(m+2, f)| \le (m+2) \left(\frac{2\pi}{\sqrt{N}}\right)^{-1} |L(m+3, f)|. \] Therefore, we have that \begin{equation} \label{s3odd} |S_3(z)| \le \frac{m+2}{m!}\left(\frac{2\pi}{\sqrt{N}}\right)^{m-1}|L(m+3, f)| \le \frac{m+2}{m!}\left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} \zeta(2)^2|L(2m+2, f)|, \end{equation} after applying Lemma $\ref{lratio}.$ Finally, by using (\ref{s12odd}) and (\ref{s3odd}), it suffices to show that \begin{equation} \label{finalodd} \frac{7}{2}2^{-m} \left(\exp\left(\frac{4\pi}{\sqrt{N}}\right) - 1\right) + \frac{m+2}{m!}\left(\frac{2\pi}{\sqrt{N}}\right)^{m-1} \zeta(2)^2 < \exp\left(-\frac{2\pi}{\sqrt{N}}\right). \end{equation}
For each value of $m$ in the first row on the following table, the value $N(m)$ is such that inequality \eqref{finalodd} holds for all $N\ge N(m)$. Note that the cases $m = 0, 1$ was done in Section \ref{sec:lowweight}.
\begin{center} \label{table2}
\begin{tabular}
{ |c || c | c |c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$m$& 31& 23& 19& 17& 16 & 15& 14& 13& 12 & 11 & 10 & 9&8&7&6&5&4&3&2 \\
\hline
$N(m)$& 1& 2& 3& 4& 5& 6& 7& 8& 10&13& 16& 22&31&47&76&137&285&766&5258 \\
\hline
\end{tabular}
\end{center}
Therefore, for $m \ge 31$, $N(m) = 1$. This completes the proof of Theorem \ref{thm:main} for $k$ odd.
\section{Equidistribution of Roots for Large Weights}\label{sec:equi}
Let $k \ge 6.$ For even $k$, set $m = \frac{k-2}{2}$. Then, the arguments in the previous sections show that for $z = e^{i\theta}$, then \[ Q_f(z) = |L(2m+1, f)| \left(\exp\left(i(m\theta+C) + \frac{2\pi}{\sqrt{N}}e^{-i\theta}\right) + O\Big(\frac{1}{2^m \sqrt{N}}\Big) \right) \] for some real constant $C$ is defined in \eqref{defC}. Therefore, \[ \on{Re}(Q_f(z)) = |L(2m+1, f)|\left(\exp\left(\frac{2\pi}{\sqrt{N}} \cos \theta\right) \cos\left(m\theta+C - \frac{2\pi}{\sqrt{N}} \sin \theta \right) + O\Big(\frac{1}{2^m \sqrt{N}}\Big) \right). \] Consider the $\theta_\ell$ such that $m\theta_\ell+C - \frac{2\pi}{\sqrt{N}} \sin \theta_\ell = \frac{\pi}{2} + \ell \pi.$ Then it is simple to verify that for some constant $D$, the two values $\theta_\ell \pm \frac{D}{2^m \sqrt{N}}$, $\on{Re}(Q_f(z))$ has different signs. This completes the proof for even $k$.
For odd $k$, set $m = \frac{k-3}{2}.$ Then, the arguments in the previous section show that for $z = e^{\theta i}$,
\[
Q_f(z) = |L(2m+2, f)| \left(\exp\left(i\Big(\Big(m+\frac{1}{2}\Big)\theta+C\Big) + \frac{2\pi}{\sqrt{N}}e^{-\theta i}\right) + O\Big(\frac{1}{2^m \sqrt{N}}\Big) \right).
\]
Therefore, it follows that
\[
\on{Re}(Q_f(z)) = |L(2m+2, f)| \left(\exp\left(\frac{2\pi}{\sqrt{N}} \cos \theta\right)\cos\left(\Big(m+\frac{1}{2}\Big)\theta + C - \frac{2\pi}{\sqrt{N}} \sin \theta\right) + O\left(\frac{1}{2^m \sqrt{N}}\right) \right).
\]
Now, consider the values $\theta_\ell$ such that $(m+1/2)\theta_\ell + C - \frac{2\pi}{\sqrt{N}} \sin \theta_\ell = \frac{\pi}{2} + \ell\pi.$ Once again, one can verify that for the values $\theta_\ell \pm \frac{D}{2^m \sqrt{N}},$ $\on{Re}(Q_f(z))$ has opposite signs. This completes the proof of Theorem \ref{thm:equi}.
\section{A Numerical Example}\label{sec:numerical}
Consider the newform $f \in S_7\left(\Gamma_0(11),\left(\frac{-11}{\bullet}\right)\right)$ whose $q$-series is given by
\[
q+10q^3+64q^4+74q^5+O(q^6).
\]
All the coefficients of $f$ are real, and we have $\epsilon(f) = 1$. In light of the functional equation $L(s,f)=L(k-1-s,f)$, we can use Sage to compute the critical values of $L(s,f)$ and thereby obtain $r_f(z)$. We calculate that the roots of $r_f(z)$ are
\[
z_1 \approx -0.294570496142963 - 0.0643219535709181 i,
\]
\[
z_2 \approx -0.204098252273756 + 0.221930156418385i,
\]
\[
z_3 \approx 0.301511344577764 i,
\]
\[
z_4 \approx 0.204098252273756 + 0.221930156418385i,
\]
and
\[
z_5 \approx 0.294570496142963 - 0.0643219535709181i.
\]
All five roots have absolute value $\approx 0.301511344577764 \approx 1/\sqrt{11}$, as expected given the statement of Theorem~\ref{thm:main}.
\section*{Acknowledgments}
\noindent This research was supervised by Ken Ono at the Emory University Mathematics REU and was supported by the National Science Foundation (grant number DMS-1557960). We would like to thank Ken Ono and Jesse Thorner for offering their advice and guidance and for providing many helpful discussions and valuable suggestions on the paper.
\bibliographystyle{amsxport}
| 940bbbc02176c67865b3e73b66aac934602bdadc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The production of top-quark pairs plays a key role in the physics program of
the LHC. On the one hand, this process can be exploited
for detailed studies of top-quark properties and interactions, for
precision tests of the Standard Model~(SM), and for measurements of
fundamental parameters such as the top-quark mass.
On the other hand, it represents a challenging background in many SM
studies and searches of physics beyond the Standard Model~(BSM).
The sensitivity of such analyses can depend in a critical way
on the precision of theoretical simulations,
and given that any experimental measurement is
performed at the level of top-decay products,
precise theoretical predictions are needed for the full process of
$\ensuremath{t \bar t}\xspace$ production and decay, including, if possible, also irreducible
backgrounds and interference effects.
This is especially
important in the context of precision measurements of the top-quark
mass.
After the discovery of the Higgs boson and the measurement of its mass, the
allowed values of the $W$-boson and top-quark masses are strongly correlated,
and a precise determination of both parameters would lead to a SM test of
unprecedented precision~\cite{Agashe:2014kda}. At present there is some
tension, at the $1.6\,\sigma$ level, between the indirect top-mass
determination from electroweak precision data~($177\pm 2.1$~GeV) and the
combination of direct measurements at the Tevatron and the LHC~($173.24\pm
0.95$~GeV).
The precise value of the top-quark mass is particularly crucial to the issue
of vacuum stability in the Standard Model~\cite{Degrassi:2012ry}. At high
scales, the Higgs quartic coupling $\lambda$ evolves to increasingly small
values as $m_t$ grows, and it is remarkable that above about $m_t=171$~GeV,
i.e.~very close to the present world average, $\lambda$ becomes negative at
the Planck scale, rendering the electroweak vacuum meta-stable, while for $m_t>176$~GeV
the electroweak vacuum becomes unstable.
The most precise top-mass measurements are based upon fits of $m_t$-dependent
Monte Carlo predictions to certain kinematic distributions.
For a precise $m_t$ determination, it is crucial to rely on Monte Carlo
generators that describe $\ensuremath{t \bar t}\xspace$ production and decay, including the shape
of top resonances, on the basis of higher-order scattering amplitudes. These
are given in terms of a theoretically well-defined top-mass parameter in an
unambiguous way, and can provide more reliable estimates of perturbative
theoretical uncertainties.
Perturbative predictions for inclusive $\ensuremath{t \bar t}\xspace$ production are available up
to next-to-next-to leading order~(NNLO) in QCD~\cite{Czakon:2013goa,
Czakon:2015owf}, and the next-to-leading order~(NLO) electroweak
corrections are also known~\cite{Beenakker:1993yr, Bernreuther:2006vg,
Bernreuther:2008md, Kuhn:2006vh, Hollik:2011ps, Kuhn:2013zoa,
Pagani:2016caq}. Calculations at NLO QCD exist also for $\ensuremath{t \bar t}\xspace$
production in association with one~\cite{Dittmaier:2007wz} or
two~\cite{Bredenstein:2009aj, Bredenstein:2010rs, Bevilacqua:2009zn,
Bevilacqua:2010ve, Bevilacqua:2011aa} extra jets. The present state-of-the
art accuracy of $\ensuremath{t \bar t}\xspace$ generators is NLO QCD, and inclusive generators
matching NLO QCD matrix elements to parton showers~(NLO+PS, from now on) have
been available for quite some time: in~\citere{Frixione:2003ei}, based
upon the \MCatNLO~\cite{Frixione:2002ik} method, and
in~\citere{Frixione:2007nw}, based upon the \POWHEG{}
method~\cite{Nason:2004rx, Frixione:2007vw}. In the following we will refer
to the latter as the \hvq{} generator.\footnote{ \hvq{} is the name of the
corresponding directory in the \POWHEGBOX{} package. The \hvq{} code is also
available under the \VTWO{} package.} More recent generators can provide NLO
QCD precision also for $\ensuremath{t \bar t}\xspace$ production in association with up to one or
two additional jets~\cite{Kardos:2011qa, Alioli:2011as, Frederix:2012ps,
Kardos:2013vxa, Cascioli:2013era, Hoeche:2013mua, Hoeche:2014qda}.
Top-quark decays are known at NNLO QCD~\cite{Brucherseifer:2013iv}, but so
far they have always been implemented at lower precision in complete
calculations of top-pair production and decay. The vast majority of such
calculations rely on the narrow-width approximation~(NWA), where matrix
elements for $\ensuremath{t \bar t}\xspace$ production and decay factorise.
Various generators based on the NWA
approximation~\cite{Frixione:2003ei, Frixione:2007nw, Kardos:2011qa,
Alioli:2011as, Frederix:2012ps, Kardos:2013vxa, Cascioli:2013era,
Hoeche:2013mua, Hoeche:2014qda} apply NLO QCD corrections only to $\ensuremath{t \bar t}\xspace$
production and include finite-width effects and spin correlations
in an approximate way using the method of~\citere{Frixione:2007zp}.
The best
available NWA fixed-order calculations implement NLO QCD corrections to the
production and decay parts with exact spin
correlations~\cite{Bernreuther:2004jv, Melnikov:2009dn, Campbell:2012uf}.
The \ttNLOdec{}\footnote{The name \ttNLOdec{} refers to the corresponding
directory in the \VTWO{} package.} generator of~\citere{Campbell:2014kua}
implements the results of~\citere{Campbell:2012uf} using
the \POWHEG{} method~\cite{Nason:2004rx, Frixione:2007vw}. Finite width and
interference effects are implemented in an approximate way, using LO
$pp\to \ensuremath{W^+W^- b \bar b}\xspace$ matrix elements. Thus, in
the resonance region it provides NLO corrections to both production and decay,
including NLO corrections to $W$ hadronic decays, and implements full spin
correlations. In addition, it can be operated both in the five-flavour number
scheme~(5FNS) and in the four-flavour number scheme~(4FNS).
A complete description of $\ensuremath{t \bar t}\xspace$ production and decay beyond the NWA
requires the calculation of the full set of Feynman diagrams that
contribute to the production of $\ensuremath{W^+W^- b \bar b}\xspace$ final states, including also
leptonic or hadronic $W$-boson decays. The existing predictions at
NLO QCD~\cite{Bevilacqua:2010qb, Denner:2010jp, Denner:2012yc,
Heinrich:2013qaa, Frederix:2013gra, Cascioli:2013wga} deal with the
different-flavour dilepton channel, $pp\toe^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$.
Besides an exact NLO treatment of spin correlations and off-shell effects
associated with the top-quark and $W$-boson resonances, such calculations
account for non-factorisable NLO effects~\cite{Beenakker:1999ya,
Melnikov:1995fx, Falgari:2013gwa} and provide an exact NLO description
of the top resonance, including quantum corrections to the top propagator.
Moreover, in addition to doubly-resonant topologies of $\ensuremath{t \bar t}\xspace$ type, also
genuine non-resonant effects stemming from topologies with less than two top
or $W$-propagators are included, as well as quantum interferences between
different topologies.
The first NLO calculations of the $pp\toe^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$
process~\cite{Bevilacqua:2010qb, Denner:2010jp, Denner:2012yc,
Heinrich:2013qaa} have been performed in the 5FNS, where $b$ quarks are
treated as massless particles.
In the meanwhile, NLO QCD predictions in the 5FNS are
available also for $e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$ production in association with
one extra jet~\cite{Bevilacqua:2015qha}.
Due to the presence of collinear $g\to \ensuremath{b \,\bar b}\xspace$ singularities, the
applicability of these calculations in the 5FNS is limited to observables
that involve at least two hard $b$ jets. This restriction can be
circumvented through NLO calculations\footnote{For a discussion at LO
see~\citere{Kauer:2001sp}.} in the 4FNS, where $b$ quarks are treated as
massive partons~\cite{Frederix:2013gra, Cascioli:2013wga}.
In addition to a more reliable description of $b$-quark kinematics, these
calculations give access to the full $e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$ phase space, including
regions where one or both $b$ quarks become unresolved. This is crucial in
order to describe top backgrounds in presence of jet vetoes.
Moreover, inclusive $e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$ calculations in the 4FNS guarantee a
consistent theoretical treatment of single-top $\ensuremath{W t}\xspace$ production at
NLO.
In the 5FNS, $\ensuremath{W t}\xspace$ and $\ensuremath{t \bar t}\xspace$ production and decay involve partonic
channels of type $gb\to \ensuremath{W^+W^- b}\xspace$ and $gg\to \ensuremath{W^+W^- b \bar b}\xspace$, respectively. The $gg\to
\ensuremath{W^+W^- b \bar b}\xspace$ channel at LO is part of the NLO radiative corrections to the $gb\to
\ensuremath{W^+W^- b}\xspace$ one,
thus yielding a NLO correction that, being $t\bar{t}$ mediated, is much
larger than the Born term. This led to the proposal of various
methods~\cite{Zhu:2001hw, Campbell:2005bb, Frixione:2008yi, White:2009yt} to
define single top cross sections not including the resonant $t\bar{t}$
contribution.
However, the separation of $tW$ and $t\bar{t}$ production
breaks gauge invariance and
does not allow for a consistent treatment of interference effects.
On the other hand, in the 4FNS the $pp\to e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$
calculations provide a unified NLO description of $\ensuremath{t \bar t}\xspace$ and $\ensuremath{W t}\xspace$
production, with a fully consistent treatment of their quantum
interference~\cite{Cascioli:2013wga}. Single-top production in the 4FNS is
described by topologies with a single top propagator and a collinear $g\to
\ensuremath{b \,\bar b}\xspace$ splitting in the initial state.
The fact that $g\to \ensuremath{b \,\bar b}\xspace$ splittings are accounted for by the matrix
elements guarantees a more precise modelling of the spectator $b$ quark,
while the simultaneous presence of $\ensuremath{W t}\xspace$ and $\ensuremath{t \bar t}\xspace$ channels, starting from
LO, ensures a perturbatively stable description of both contributions, as
well as a NLO accurate prediction for their interference.
A generator based on the \POWHEG{} method and $pp\to e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$ matrix
elements at NLO in the 5FNS has been presented in~\citere{Garzelli:2014dka}.
However, the matching of parton showers to matrix elements that involve
top-quark resonances poses nontrivial technical and theoretical
problems~\cite{Jezo:2015aia} that have not been addressed
in~\citere{Garzelli:2014dka} and which cannot be solved within the original
formulations of the \POWHEG{} or \MCatNLO{} methods.
The problem is twofold. On the one hand, when interfacing a generator to a
shower, if we do not specify which groups of final-state particles arise from
the decay of the same resonance, the recoil resulting from shower emissions
leads to arbitrary shifts of the resonance invariant masses, whose magnitude
can largely exceed the top-quark width, resulting in unphysical distortions
of the top line shape~\cite{Jezo:2015aia}.
On the other hand, in the context of the infrared-subtraction and matching
procedures, the standard mappings that connect the Born and real-emission
phase spaces affect the top resonances in a way that drastically deteriorates
the efficiency of infrared~(IR) cancellations and jeopardises the consistency
of the matching method~\cite{Jezo:2015aia}.
A general NLO+PS matching technique that allows for a consistent treatment of
resonances has been introduced, and applied to $t$-channel single-top
production, in~\citere{Jezo:2015aia}. This approach will be referred to as
resonance-aware matching. It is based on the \POWHEG\footnote{A related
approach within the \MCatNLO{} framework has been presented and also
applied to $t$-channel single-top production in~\citere{Frederix:2016rdc}.}
method and is implemented in the \RES{} framework, which represents an
extension of the \POWHEGBOX~\cite{Alioli:2010xd}. In this framework each
component of the cross section~(i.e.~Born, virtual and real) is separated
into the sum of contributions that are
dominated by well-defined resonance histories,
such that in the narrow-width limit
each parton can be uniquely attributed either to the
decay products of a certain resonance or to the production subprocess.
Within each contribution the subtraction procedure is organized in
such a way that the off-shellness of resonant $s$-channel propagators is preserved,
and resonance information on the final-state particles can be communicated to the shower program that
handles further radiation and hadronization. This avoids uncontrolled
resonance distortions, ensuring a NLO accurate description of the top line
shape. The resonance-aware approach
also improves the efficiency of infrared subtraction and phase-space integration
in a dramatic way.
In this paper we present a NLO+PS generator, that we dub \bbfourl{} in the
following, based on NLO matrix elements for $pp\toe^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$ in the 4FNS
matched to \PythiaEight{}~\cite{Sjostrand:2007gs, Sjostrand:2014zea} using
the resonance-aware \POWHEG{} method. This new generator combines, for the
first time, the following physics features:
\begin{itemize}
\item[-] consistent NLO+PS treatment of top resonances, including quantum
corrections to top propagators and off-shell top-decay chains;
\item[-] exact spin correlations at NLO, interference between NLO radiation
from top production and decays, full NLO accuracy in $\ensuremath{t \bar t}\xspace$ production
and decays;
\item[-] unified treatment of $\ensuremath{t \bar t}\xspace$ and $\ensuremath{W t}\xspace$ production with interference
at NLO;
\item[-] improved modelling of $b$-quark kinematics thanks to $b$-quark mass
effects;
\item[-] access to phase-space regions with unresolved $b$ quarks and/or jet
vetoes.
\end{itemize}
These improvements are of particular interest for precision top-mass
measurements, for $Wt$ analyses, and for top backgrounds in the presence of
jet vetoes or in the off-shell regime.
Technically, the \bbfourl{} generator is based on
{\tt OpenLoops}\xspace{}~\cite{OLhepforge} matrix elements. To this end we have
developed a general and fully-flexible \POWHEGBOX{}+{\tt OpenLoops}\xspace{} interface,
which allows one to set up NLO+PS generators for any desired process.
The paper is organized as follows.
In \refse{sec:reminder} we briefly review the
resonance-aware matching method.
In \refse{sec:BOXRES} we discuss new developments in the
\RES{} framework that have been relevant for the present work.
In \refse{sec:description} we discuss various aspects of the \bbfourl{}
generator, including scope, usage, interface to \PythiaEight{}, and
consistency checks.
In \refse{sec:setup} we detail the setup employed for the
phenomenological studies presented in the subsequent sections. There we
compare the \bbfourl{} generator to the previously available \POWHEG{}
generators, the \hvq{} and \dec{} ones, and we present technical studies that
show the impact of the resonance-aware matching and of other improvements
implemented in \bbfourl{}.
Specifically, in \refse{sec:ttbarPhenomenology} we consider observables
that are directly sensitive to top-quark resonances and top-decay products,
while in \refse{sec:WtPhenomenology} we investigate the $e^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$
cross section in the presence of jet vetoes that enhance its single-top
content.
Our conclusions are presented in \refse{sec:conc}.
The \RES{} framework together with the \bbfourl{} generator can be downloaded at \url{http://powhegbox.mib.infn.it}.
\section{Resonance-aware subtraction and matching}
\label{sec:reminder}
In the following we recapitulate the problems that arise in
processes where intermediate narrow resonances can radiate as they decay,
and summarize the ideas and methodology behind the
resonance-aware algorithm of~\citere{Jezo:2015aia}.
We refer the reader to the original publication
for the description of the method in full detail.
Commonly used IR subtraction methods for the calculation of NLO
corrections~\cite{Frixione:1995ms,Catani:1996vz,Catani:2002hc}
are based upon some procedure of momentum reshuffling for the construction of collinear and
infrared counterterms. More specifically, given the kinematics of the
real-emission process, and having specified a particular collinear
region~(i.e.~a pair of partons that are becoming collinear), there is a
well-defined mapping that constructs a Born-like kinematic
configuration~(called the ``underlying Born'' configuration) as a function of
the real one. The mapping is such that, in the strict collinear limit, the Born
configuration is obtained from the real one by appropriately merging the
collinear partons. In the traditional methods, these mappings do not
necessarily preserve the virtuality of possible intermediate $s$-channel
resonances. If we consider the collinear region of two partons arising from
the decay of the same $s$-channel resonance, the typical difference in the
resonance virtuality between the real kinematics and the underlying-Born one
is of order $m^2/E$, where $m$ is the mass of the two-parton system, and $E$
is its energy. Because of this, the cancellation between the real
contribution and the subtraction term becomes effective only if $m^2/E <
\Gamma$, where $\Gamma$ is the width of the resonance.
As long as $\Gamma$ is above zero, the traditional NLO calculations do
eventually converge, thanks to the fact that in the strict collinear limit
the cancellation takes place. However, convergence becomes more problematic
as the width of the resonance decreases.
The presence of radiation in resonance decays causes even more severe
problems in NLO+PS frameworks. In \POWHEG{}, radiation is generated
according to the formula
\begin{eqnarray}
\mathrm{d} \sigma & = & \bar{B} (\Phi_{\mathrm{B}}) \,\mathrm{d} \Phi_{\mathrm{B}} \left[
\Delta (q_{\tmop{cut}}) + \sum_{\alpha} \Delta (k^{\alpha}_{\sss T})
\frac{R_{\alpha} (\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}}))}{B
(\Phi_{\mathrm{B}})} \,\mathrm{d} \Phi_{\tmop{rad}} \right]. \label{eq:powheg}
\end{eqnarray}
The first term in the square bracket corresponds to the probability that no
radiation is generated with hardness above an infrared cutoff
$q_{\tmop{cut}}$, and its kinematics corresponds to the Born one. Each
$\alpha$ in the sum labels a collinear singular region of the real cross
section. The full real matrix element is
decomposed into a sum of terms
\begin{equation}
R = \sum_{\alpha} R_{\alpha}\,,
\end{equation}
where each $R_{\alpha}$ is singular only in the region labelled by $\alpha$.
The real phase space $\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})$
depends upon the singular region $\alpha$ and is given as a function of the
Born kinematics $\Phi_{\mathrm{B}}$ and three radiation variables
$\Phi_{\tmop{rad}}$. The inverse of $\Phi_{\alpha}$
implements the previously mentioned mapping of the real kinematics into an
underlying Born one. Thus,
for a given $\Phi_{\mathrm{B}}$ and $\Phi_{\tmop{rad}}$,
each term in the sum inside the square bracket in~\refeq{eq:powheg}
is associated with a
different real phase-space point. For each $\alpha$, $k^{\alpha}_{\sss T}$ is
defined as the hardness of the collinear splitting characterized by the
kinematics $\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})$. It
usually corresponds to the relative transverse momentum of the two collinear
partons.
The Sudakov form factor, $\Delta$, is such that the square bracket in~\refeq{eq:powheg}, after performing the integrals in $\mathrm{d}
\Phi_{\tmop{rad}}$, becomes exactly
equal to one~(a property sometimes called {\it unitarity}
of the real radiation).
In general we have
\begin{eqnarray}
\Delta (q) & = & \prod_{\alpha} \Delta_{\alpha} (q)\,,
\end{eqnarray}
with
\begin{eqnarray}
\Delta_{\alpha} (q) & = & \exp \left[ - \int_{k^{\alpha}_{\sss T} > q}
\frac{R_{\alpha} (\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}}))}{B
(\Phi_{\mathrm{B}})}\, \mathrm{d} \Phi_{\tmop{rad}} \right] . \label{eq:Sudakov}
\end{eqnarray}
In order to achieve NLO accuracy, the $\bar{B} (\Phi_{\mathrm{B}})$ factor must
equal the NLO inclusive cross section at given underlying Born kinematics,
\begin{eqnarray}
\bar{B} (\Phi_{\mathrm{B}}) & = & B (\Phi_{\mathrm{B}}) + V (\Phi_{\mathrm{B}}) +
\sum_{\alpha} \int R_{\alpha} (\Phi_{\alpha} (\Phi_{\mathrm{B}}\,,
\Phi_{\tmop{rad}})) \,\mathrm{d} \Phi_{\tmop{rad}}\,,
\end{eqnarray}
where both the second and third term on the right hand side are infrared
divergent, but the sum, being an inclusive cross section, is finite. The
cancellation of singularities is achieved with the usual subtraction
techniques.
We are now in a position to discuss the problems that arise in processes with
radiation in decays of resonances. In order to do this, we focus on the $W^- W^+ b
\bar{b}$ production process. As an example of the problem, we consider a
real emission contribution where a gluon $g$ is
radiated, such that the mass of the $W^+ b g$ and $W^-\bar{b}$ systems are very
close to the top nominal mass.
We call $\alpha_b$ the singular region
corresponding to $b$ and $g$, and $\alpha_{\bar{b}}$ the region corresponding
to the $\bar{b}$ and $g$ becoming collinear, respectively.
If we consider the
case when the $b$ and $\bar b$ partons are relatively close in direction, as
$g$ becomes collinear to the $b$ or the $\bar{b}$ parton, two components will
dominate the real cross section, $R_{\alpha_b}$ and $R_{\alpha_{\bar{b}}}$,
in a proportion that is determined by how close the gluon is to the $b$ or to
the $\bar{b}$ partons. If the gluon is not much closer to the $b$ region with
respect to the $\bar{b}$ one, the $R_{\alpha_{\bar{b}}}$ contribution will be
comparable or larger than the $R_{\alpha_b}$ one. We now observe that, for
the same real kinematic configuration, we have two singular regions and two
corresponding underlying-Born configurations. In the $\alpha_b$ singular
region, the underlying Born is obtained by merging the $b g$ system into a
single $b$, while in the $\alpha_{\bar b}$ region it is the ${\bar b} g$
system that is merged into a single $\bar{b}$. It is therefore clear that, in
the $\alpha_b$ merging, the resonance virtualities are nearly preserved in
the underlying Born, while in the $\alpha_{\bar b}$ one the resonances will
be far off-shell. The $R_{\alpha_{\bar{b}}} / B$ terms appearing both in~\refeq{eq:powheg}
and~(\ref{eq:Sudakov}) will become very large, the top
resonances being on-shell in the numerator and off-shell in the
denominator. However, in the \POWHEG{} framework, these ratios
should be either small~(of order $\alpha_s$) or should approach the
Altarelli-Parisi splitting functions for the method to work.
It is thus clear that, if resonances are present, the traditional
decomposition into singular regions must be revised.
In particular, each $\alpha$ should become associated
to a specific resonance structure of the event, such that
collinear partons originate from the same resonance. Furthermore,
the phase space mapping $\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})$
should preserve the virtuality of the intermediate resonances.
This is, in brief, what was done in~\citere{Jezo:2015aia}.
The resonance-aware formalism also offers the opportunity to modify and
further improve the \POWHEG{} radiation formula. We make, for the moment, the
assumption that each decaying resonance has only one singular region, and the
radiation not originating from a resonance decay also has only one singular
region. This is the case, for example, for the resonance structure of the
process $g g \rightarrow (t \rightarrow W^+ b) (\bar{t} \rightarrow W^-
\bar{b})$, since in \POWHEG{} the initial-state-radiation~(ISR) regions are
combined into a single one. We consider the formula
\begin{eqnarray}
\mathrm{d} \sigma & = & \bar{B} (\Phi_{\mathrm{B}}) \,\mathrm{d} \Phi_{\mathrm{B}}
\prod_{\alpha = \alpha_b, \alpha_{\bar{b}}, \alpha_{\rm\tiny ISR}} \left[
\Delta_{\alpha} (q_{\tmop{cut}}) + \Delta_{\alpha} (k^{\alpha}_{\sss T})
\frac{R_{\alpha} (\Phi_{\alpha} (\Phi_{\mathrm{B}},
\Phi^{\alpha}_{\tmop{rad}}))}{B (\Phi_{\mathrm{B}})} \,\mathrm{d}
\Phi^{\alpha}_{\tmop{rad}} \right], \label{eq:allrad}
\end{eqnarray}
where, by writing $\Phi^{\alpha}_{\tmop{rad}}$, we imply that the radiation
variables are now independent for each singular region. By expanding the
product, we see that we get a term with no emissions at all, as in
\refeq{eq:powheg}, plus terms with multiple~(up to three) emissions. It can be shown
that, as far as the hardest radiation is concerned, formula~(\ref{eq:allrad}) is
equivalent to formula~(\ref{eq:powheg}). To this end, one begins by
rewriting~\refeq{eq:allrad} as a sum of three terms, with appropriate
$\theta$ functions such that each term represents the case where the hardest
radiation comes from one of the three regions. It is easy then to integrate
in each term all radiations but the hardest, thus recovering the full Sudakov
form factor appearing in the second term in the square bracket of~\refeq{eq:powheg}.
The \bbfourl{} generator can generate radiation using the improved
multiple-radiation scheme of formula~(\ref{eq:allrad}) or the conventional
single-radiation approach of~\refeq{eq:powheg}. In events generated with
multiple emissions included, the hardest radiation from all sources
(i.e.~production, $t$ and $\bar{t}$ decays) may be present. The
\POWHEG{} generated event is then completed by a partonic shower Monte
Carlo program that attaches further radiation to the event.
The interface to the shower must be such that the shower does not generate
radiation in production, in $t$ decay and in $\bar{t}$ decay that is harder
than the one generated by \POWHEG{} in production, $t$ and $\bar{t}$ decay,
respectively.\footnote{
We note that this method guarantees full NLO accuracy, including exact spin correlations,
only at the level of each individual emission, while
correlation effects between multiple QCD emissions are handled in approximate form.
Nevertheless it should be clear that~\refeq{eq:allrad}
represents a significant improvement with respect
to pure parton showering after the first emission.}
\section{The \RES{} framework}
\label{sec:BOXRES}
In this section we illustrate features that have been added to the \RES{}
package since the publication of~\citere{Jezo:2015aia}, and discuss some
issues that were not fully described there.
\subsection*{Automatic generation of resonance histories}
In the \RES{} implementation of~\citere{Jezo:2015aia}, the initial
subprocesses and the associated resonance structures were set up by hand. We
have now added an algorithm for the automatic generation of all relevant
resonance histories for a given process at a specified perturbative order.
Thanks to this feature, the user only needs to provide a list of
subprocesses, as was the case in the \VTWO{} package. This is a considerable
simplification, in view of the fact that, when electroweak processes are
considered, the number of resonance histories can increase substantially.
Details of this feature are given in~\refapp{app:resonance_histories}.
\subsection*{Colour assignment}
Events that are passed to a shower generator for subsequent
showering must include colour-flow information in the limit of large
number of colours. In the \VTWO{} framework, colours are assigned with a
probability proportional to the corresponding component of the colour flow
decomposition of the amplitude.
The extension of this approach to the \RES{} framework requires some care
due to possible inconsistencies between the
colour assignment and the partitioning
into resonance histories.
This issue and its systematic solution are discussed in detail in~\refapp{app:colourassignment}.
\subsection*{\POWHEG{}+{\tt OpenLoops}\xspace interface}
All tree and one-loop amplitudes implemented in the \bbfourl{} generator are based on the
{\tt OpenLoops}\xspace program ~\cite{OLhepforge} in combinations with \Collier~\cite{Denner:2016kdg}
or {\tt CutTools}\xspace~\cite{Ossola:2007ax} and {\tt OneLOop}\xspace \cite{vanHameren:2010cp}.
In the framework of the present work a new
general process-independent interface between the \POWHEGBOX{} and {\tt OpenLoops}\xspace
has been developed. It allows for a straightforward implementation of
a multitude of NLO multi-leg processes matched to parton showers including QCD
and, in the future, also NLO electroweak corrections~\cite{Kallweit:2014xda,
Kallweit:2015dum}. Technical details and a brief documentation of this new
interface can be found in~\refapp{sec:openloops_interface}.
\section{Description of the generator}
\label{sec:description}
The implementation of combined off-shell \ensuremath{t \bar t}\xspace and \ensuremath{W t}\xspace production in the
\RES{} framework presented in this paper is based on all possible Feynman
diagrams contributing to the process \ensuremath{\ensuremath{pp \,\to \,e^+\nu_e\mu^-\bar\nu_\mu\, \bbbar}\xspace + X}\xspace at NLO accuracy in QCD,
i.e.~up to order $\as^3 \aem^4$.
All bottom-mass effects have been fully taken into
account and for the consistent treatment of top-, $W$-, and $Z$-resonances at
NLO we rely on the automated implementation of the complex-mass
scheme~\cite{Denner:1999gp, Denner:2005fg} within {\tt OpenLoops}\xspace.
\subsection{Resonance histories}\label{sec:reshists}
The Born level resonance structure for \ensuremath{pp \,\to \,e^+\nu_e\mu^-\bar\nu_\mu\, \bbbar}\xspace{} at $O(\as^2 \aem^4)$
is actually very simple. Indeed, it is sufficient to consider
two kinds of resonance histories. In~\reffi{fig:born_res_hist}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.7\textwidth]{diagrams/born_res_hist}
\end{center}
\caption{Sample Feynman graphs corresponding to the two resonance histories
relevant for $pp \to \mu^+ \nu_\mu e^- \bar{\nu}_e \,\ensuremath{b \,\bar b}\xspace$ production.}
\label{fig:born_res_hist}
\end{figure}
we show two corresponding Feynman diagrams for the process $pp \to \mu^+
\nu_\mu e^- \bar{\nu}_e \,\ensuremath{b \,\bar b}\xspace$ .
Internally, according to the \RES{} conventions~\cite{Jezo:2015aia},
the resonance histories are described by the arrays
\begin{verbatim}
flav_1 = [i, j, 6, -6, 24, -24, -13, 14, 11, -12, 5, -5],
flavres_1 = [0, 0, 0, 0, 3, 4, 5, 5, 6, 6, 3, 4],
flav_2 = [i, j, 23, 24, -24, -13, 14, 11, -12, 5, -5],
flavres_2 = [0, 0, 0, 3, 3, 4, 4, 5, 5, 0, 0],
\end{verbatim}
for all
relevant
choices of initial parton flavours \verb!i,j!. In
\verb!flav! we store the identities of the initial- and final-state
particles, with intermediate resonances, if they exist, labelled according to
the Monte Carlo numbering scheme~(gluons are labeled by zero in the
\POWHEGBOX). In \verb!flavres!, for each particle, we give the position of
the resonance from which it originates. For partons associated with the production subprocess
\verb!flavres! is set to zero.
The resonance structures that differ only by the external parton flavours are
collected into resonance groups, so that, in the present case, we have only
two resonance groups.
We remark that there is no need of a unique correspondence between resonance structures and possible
combinations of resonant propagators in individual Feynman diagrams.
What is required is that all resonances present in any given Feynman graph are also
present in an associated resonance structure, but not vice versa.
For example, in the present implementation of the \bbfourl{} generator
the consistent treatment of single-top topologies like the one in~\reffi{fig:born_Wt}
is guaranteed through resonance histories of $\ensuremath{t \bar t}\xspace$ type
(\verb!flav_1,flavres_1!),
which involve an additional $\bar{t}\to {\bar b} W^-$ resonance.
This does not lead to any problems, since the corresponding
subtraction kinematics, which preserves the mass of the ${\bar b} W^-$ system,
is perfectly adequate also for single-top topologies.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.35\textwidth]{diagrams/born_res_hist_Wt}
\end{center}
\caption{Representative Born diagram for $Wt$ production.}
\label{fig:born_Wt}
\end{figure}
The \RES{} code automatically recognizes resonance histories that can be
collected into the same resonance group. It also includes a subroutine for
the automatic generation of an
adequate phase-space sampling for each resonance group.
In this context, rather than relying upon standard Breit-Wigner
sampling, care is taken that also the off-shell regions are
adequately populated.
This is essential in resonance histories of the kind
shown in the right graph of~\reffi{fig:born_res_hist}, where the
generation of the $W$ virtualities according to their Breit-Wigner shape
would well probe the region where an off-shell $Z$ decays into two on-shell
$W$'s, but not the regions where an on-shell $Z$ decays into an on-shell $W$
and an off-shell one. It also guarantees that cases like the diagram in~\reffi{fig:born_Wt}
are properly sampled.
The interested reader can find more technical details by inspecting the code
itself.
\subsection{The complex-mass scheme}
\label{sec:cms}
In our calculation all intermediate massive particles are consistently
treated in the complex-mass scheme~\cite{Denner:1999gp, Denner:2005fg},
where the widths of unstable particles are absorbed into
the imaginary part of the corresponding mass parameters,
\begin{equation}
\label{eq:complexmasses1}
\mu^2_i=M_i^2-\mathrm{i}\Gamma_iM_i \qquad\mbox{for}\ \ i=\ensuremath{W}\xspace,\ensuremath{Z}\xspace,\ensuremath{t}\xspace,\ensuremath{H}\xspace.
\end{equation}
This choice implies a complex-valued weak mixing angle,
\begin{equation}
\label{eq:defsintheta1}
\sin\theta_W^2=1-\cos\theta_W^2=1-\frac{\mu_\ensuremath{W}\xspace^2}{\mu_\ensuremath{Z}\xspace^2}\,,
\end{equation}
and guarantees gauge invariance at NLO~\cite{Denner:2005fg}.
\subsection{The decoupling and \MSB{} schemes}
\label{sect:decoupling}
When performing a fixed-order calculation with massive quarks, one can define
two consistent renormalization schemes that describe the same physics: the
usual \MSB{} scheme, where all flavours are treated on equal footing, and a
mixed scheme~\cite{Collins:1978wz}, that we call decoupling scheme, in
which the $\nlf$ light flavours are subtracted in the \MSB{} scheme, while
heavy-flavour loops are subtracted at zero momentum. In this scheme, heavy
flavours decouple at low energies.
In the calculation of the $\fourl\,\ensuremath{b \,\bar b}\xspace$
hard scattering cross section we treat the bottom quark
as massive and, correspondingly, $\nlf$ is equal to four.
The renormalization
of the virtual contributions is performed in the decoupling scheme with a
four-flavour running $\as$.
For consistency, the evolution of parton distribution
functions~(PDFs) should
be performed with four active flavours, so
that, in particular, no bottom-quark density is present and no bottom-quark
initiated processes have to be considered.
However, given that the process at hand is characterised by typical
scales far above the $b$-quark threshold, it is more convenient
to convert our results to the $\MSB$ scheme
in such a way that they can be expressed in terms of the $\MSB$
strong coupling constant, running with five active flavours, and also with five-flavour PDFs.
The procedure for such a switch of schemes is well known, and was discussed
in~\citere{Cacciari:1998it}.
For $\fourl \,\ensuremath{b \,\bar b}\xspace$ production, we need to transform
the $q\bar q$ and $gg$ squared Born amplitudes ${\cal B}_{qq}$ and ${\cal B}_{gg}$,
computed in the decoupling scheme, in the following way
\begin{eqnarray}
{\cal B}_{qq}&\rightarrow& \left[1
- \frac{4}{3}\,\TF \,\frac{\as}{2\pi}\, \log \(\frac{\mur^2}{\mb^2}\)\right] \, {\cal B}_{qq}\,,\\
{\cal B}_{gg}&\rightarrow&
\left[1+\frac{4}{3}\,\TF \,\frac{\as}{2\pi}\, \log \(\frac{\muf^2}{\mur^2}\)\right] \, {\cal
B}_{gg}\,.
\end{eqnarray}
where
$\mur$ and $\muf$ are the renormalization
and factorization scales, respectively, and $\mb$ is the bottom-quark
mass. The contribution of the $b$ parton densities, that are present in the
five-flavour scheme, should not be included in this context.
\subsection{The virtual corrections}
The virtual contributions have been generated using the new interface of the
\POWHEGBOX{} with the {\tt OpenLoops}\xspace amplitude generator, as described in~\refapp{sec:openloops_interface}.
While {\tt OpenLoops}\xspace guarantees a very fast evaluation of
one-loop matrix elements, the overall efficiency of the
generator can be significantly improved by minimising the
number of phase space points that require the calculation
of virtual contributions.
As detailed in \refapp{app:virtuals},
this is is achieved by evaluating the virtual and real-emission contributions with
independent statistical accuracies optimised according to the respective relative weights.
Moreover, when generating events, a reweighting method can be used in order
to restrict virtual evaluations to the small fraction of phase space points that
survive the unweighting procedure.
\subsection{Interface to the shower}
The generator presented in this work shares many common features with the one
of~\citere{Campbell:2014kua}. In particular, in both generators, Les
Houches events include resonance information, and an option for a multiple
radiation scheme is implemented, denoted as \verb!allrad! scheme, according to the
corresponding {\tt powheg.input} flag. As
explained in~\citere{Campbell:2014kua} and reviewed in~\refse{sec:reminder},
when this scheme is activated, the mechanism of
radiation generation is modified. Rather than keeping only the hardest
radiation arising from all singular regions, the program stores several
``hardest radiations'': one that takes place at the production stage, and one
for the decay of each resonance that can radiate. All these radiations are
assembled into a single Les Houches event. Thus, for example, in events with
the $t$ and $\bar{t}$ resonances, one can have up to three radiated partons:
one coming from the initial-state particles, one arising from the $b$ in the
$t$-decay, and one from the $\bar{b}$ in the $\bar{t}$-decay.
When generating fully showered events, the hardness\footnote{Here and in the
following by hardness we mean the relative transverse momentum of two
partons arising from a splitting process, either in initial- or in
final-state radiation.} of the shower must be limited in a way that depends
upon the origin of the radiating parton. If the radiating parton is not son
of a resonance, the hardness of the shower arising from it must be limited by
the hardness of the Les Houches radiation that arises in
production.\footnote{%
By radiation in production we mean any radiation that
does not arise from a decaying resonance. This can be initial-state
radiation, but also radiation from final-state partons, as in the right
diagram in~\reffi{fig:born_res_hist} and the one in~\reffi{fig:born_Wt},
where the $b$'s do not arise from a decaying
resonance.}
Radiation arising from partons originating from a resonance must have their
hardness limited by the hardness of the parton radiated from the resonance in
the Les Houches event. This requires a shower interface that goes beyond the
Les Houches approach. In~\citere{Campbell:2014kua} a suitable procedure
has been conceived and implemented in \PythiaEight{}~\cite{Sjostrand:2007gs,
Sjostrand:2014zea}. The interested reader can find all details in the
Appendix~A of ~\citere{Campbell:2014kua}. In essence, the procedure was to
examine the showered event, compute the transverse momentum of \PythiaEight{}
radiation in top decays, and veto it if higher than the corresponding
\POWHEG{} one. Vetoing is performed by rejecting the showered event, and
generating a new \PythiaEight{} shower, initiated by the same Les Houches
event. This procedure was iterated until the showered event passes the veto.
In the present work, we have adopted this procedure in order to
make a more meaningful comparison with the results of~\citere{Campbell:2014kua}.
However, we have
also verified that, by using \PythiaEight{} internal mechanism for vetoing
radiation from resonance decay, we get results that are fully compatible with
our default approach.\footnote{An interface to
\Herwigseven{}~\cite{Bellm:2015jjp} is now under development.} This
aspect and the comparison among the two methods are shown in~\refapp{sec:veto}.
\subsection{Traditional NLO+PS matching}
It is possible to run our new generator in a way that is fully equivalent to a
standard \POWHEG{} matching algorithm~(as implemented in the \VTWO{})
ignoring the resonance structure of the processes. This is
achieved by including the line \verb!nores 1! in the \verb!powheg.input!
file.\footnote{In this mode, our generator becomes similar to the
implementation~\citere{Garzelli:2014dka}, except for our use of the four
flavours scheme.} Such an option is implemented only for the purpose of
testing the new formalism with respect to the old one.
It turns out that, in the \verb!nores 1! mode, the program has much worse
convergence properties, most likely because of the less effective cancellation
of infrared singularities mentioned in~\refse{sec:reminder}. We
find, for example, that in runs with equal statistics~(with about 15 million
calls) the absolute error in the \verb!nores 1! case is roughly seven times
larger than in the \verb!nores 0!~(default) case. The generation of
events also slows down by a similar factor.
We stress again that, in the limit of small widths,
the NLO+PS results
obtained in the
\verb!nores 1!
mode are bound to become inconsistent, as
discussed in~\refse{sec:reminder} and, more extensively, in~\citere{Jezo:2015aia}.
\subsection{Consistency checks}\label{sec:Checks}
At the level of fixed-order NLO calculations, the traditional machinery of the
\POWHEGBOX{} is well tested and we trust corresponding results to be correct.
On the other hand, the NLO subtraction procedure implemented in the \RES{}
code is substantially different and still relatively new. As was done
in~\citere{Jezo:2015aia} for $t$-channel single-top production, also for the
$\fourl\,\ensuremath{b \,\bar b}\xspace$ production presented here, we systematically validated the
fixed-order NLO results obtained with the \RES{} implementation by switching
on and off the generation of resonance structures. We found perfect agreement
between the two calculations.
Additionally, we performed a detailed comparison against the fixed-order NLO
results of~\citere{Cascioli:2013wga} and found agreement at the permil
level. Furthermore, via a numerical scan in the limit of the top width going
to zero, $\Gamma_t \to 0$, we verified that any
$\as\log\left(\Gamma_t\right)$ enhanced terms in the soft-gluon limit
successfully cancel between real and virtual contributions. This last test
was performed for various light- and $b$-jet exclusive distributions which
are subject to sizable non-resonant/off-shell corrections.
\section{Phenomenological setup}
\label{sec:setup}
In this section we document the input parameters, acceptance cuts and
generator settings that have been adopted for
the numerical studies presented in~\refse{sec:ttbarPhenomenology}.
Moreover we introduce a systematic labelling scheme
for the various NLO+PS approximations that are going to be compared.
\subsection{Input parameters}
Masses and widths are assigned the following values
\begin{align}
m_{Z} &= 91.188 \;\text{GeV}\xspace\,, & \Gamma_{Z} &= 2.441 \;\text{GeV}\xspace\,,\\
m_{W} &= 80.419 \;\text{GeV}\xspace\,, & \Gamma_{W} &= 2.048 \;\text{GeV}\xspace\,, \\
m_{H} &= 125 \;\text{GeV}\xspace\,, & \Gamma_{H} &= 4.03\times10^{-3}\;\text{GeV}\xspace\,, \\
m_{t} &= 172.5 \;\text{GeV}\xspace\,, & \Gamma_{t} &= 1.329 \;\text{GeV}\xspace\,, \\
m_{b} &= 4.75 \;\text{GeV}\xspace\,. &
\end{align}
The electroweak couplings are derived from the gauge-boson masses and the
Fermi constant, ${G_{\sss\mu}}=1.16585\times10^{-5}~\text{GeV}\xspace^{-2}$, in the
$G_{\mu}$-scheme, via
\begin{equation}
\aem=\sqrt{2}\, \frac{G_\mu}{\pi} \left|\mu_{\sss W}^2\(1-\frac{\mu_{\sss
W}^2}{\mu_{\sss Z}^2}\)\right|=\frac{1}{132.50698}\,,
\end{equation}
where $\mu_{\sss W}$ and $\mu_{\sss Z}$ are complex masses given by~\refeq{eq:complexmasses1}.
The value of the
top-quark width we use is consistently calculated at NLO from all other input
parameters by computing the three-body decay widths $\Gamma(t\to f \bar f' b)$
into any pair of light fermions $f$ and $\bar f'$ and a massive $b$ quark. To this
end, we employ a numerical routine of the {\tt MCFM}\xspace implementation of~\citere{Campbell:2012uf}.
As parton distributions
we have adopted the five-flavour MSTW2008NLO PDFs~\cite{Martin:2009iq}, as
implemented in the~\citere{Buckley:2014ana},
with the corresponding five-flavour strong
coupling constant, and for their consistent combination with
four-flavour scheme parton-level cross sections
the scheme transformation of \refse{sect:decoupling} was applied.
In the evaluation of the matrix elements, only the bottom
and the top quarks are massive. All the other quarks are treated as
massless. In addition, the Cabibbo-Kobayashi-Maskawa matrix is assumed to be
diagonal.
When generating events we adopt the following scale choice:
\begin{itemize}
\item For resonance histories with a top pair we use
\begin{equation}
\label{eq:ttscale}
\mur=\muf=\lq \(m_t^2+p_{{T},t}^2\)\(m_{\bar t}^2+p_{{T},{\bar t}}^2\)\rq ^{\frac{1}{4}}\;,
\end{equation}
where the (anti)top masses and transverse momenta are defined in the underlying Born phase space
in terms of final state (off-shell) decay products.
\item For resonance histories with an intermediate $Z$ we use
\begin{equation}
\label{eq:zscale}
\mur=\muf=\frac{\sqrt{p_{Z}^2}}{2}\;,
\end{equation}
where $p_Z=p_{\ell^+}+p_{\nu_\ell}+p_{l^-}+p_{\bar\nu_l}$.
\end{itemize}
In addition, we set the value of the \POWHEGBOX{} parameter \verb!hdamp! to
the mass of the top quark. This setting yields a transverse-momentum
distribution of the top pair that is more sensitive to scale variations and
more consistent with data at large transverse momenta. It only affects
initial-state radiation. For a detailed description of this parameter, we
refer the reader to~\citere{Alioli:2008tz}.
\subsection{\PythiaEight{} settings}
We interface our \POWHEG{} generator to
\PythiaEightPone{},\footnote{An interface to \PythiaEightPtwo{} is also
available, but was not used for the present work.}
as illustrated in
Appendix~A of~\citere{Campbell:2014kua}, and so we perform the following
\PythiaEight{} calls:
\begin{verbatim}
pythia.readString("SpaceShower:pTmaxMatch = 1");
pythia.readString("TimeShower:pTmaxMatch = 1");
pythia.readString("PartonLevel:MPI = off");
pythia.readString("SpaceShower:QEDshowerByQ = off");
pythia.readString("SpaceShower:QEDshowerByL = off");
pythia.readString("TimeShower:QEDshowerByQ = off");
pythia.readString("TimeShower:QEDshowerByL = off");
\end{verbatim}
The first two calls are required when interfacing \PythiaEight{} to NLO+PS
generators. The third call switches off multi-parton interactions and it is
only invoked for performance reasons: in fact, the shower of the events is
faster when multi-parton interactions are not simulated.
The remaining calls switch off the electromagnetic radiation in
\PythiaEight{}. This makes it easier to reconstruct the $W$ boson momentum,
since we do not need to dress the charged lepton, from vector boson decay,
with electromagnetic radiation. These settings are appropriate in the
present context since we do not make any comparison with data.
\PythiaEight{} provides by default matrix-element corrections~\cite{Norrbin:2000uu}.
In our case, they are relevant for radiation in the top decays, which are corrected
using $t\to Wb g$ tree level
matrix elements. These corrections are also applied in subsequent
emissions in order to better model radiation from heavy flavours in general.
If not explicitly stated otherwise, we include the following setup calls
\begin{verbatim}
pythia.readString("TimeShower:MEcorrections = on");
pythia.readString("TimeShower:MEafterFirst = on");
\end{verbatim}
These corrections never modify the Les Houches event weight. They only affect
the radiation generated by the shower. Thus, leaving these flags on does not
lead to over-counting. If the second flag is off, matrix-element corrections
are applied only to the first shower emission. If it is on, they are also
applied to subsequent radiation. In fact, even if these corrections cannot
fully account for the structure of the matrix elements, they at least better
account for mass effects arising in radiation from the off-shell top quarks
and from the massive final-state $b$'s.
In our analysis, we keep $B$ hadrons stable, performing the corresponding
\PythiaEight{} setup calls. Aside from these, all remaining settings are left
to the defaults of \PythiaEightPone{}.
\subsection{Generators and labels}
In~\refse{sec:ttbarPhenomenology} we compare three different generators that
implement an increasingly precise treatment of $\ensuremath{t \bar t}\xspace$
production and decay:
\begin{itemize}
\item the \hvq{} generator of~\citere{Frixione:2007nw};
\item the \DEC{} generator of~\citere{Campbell:2014kua};
\item the new \bbfourl{} generator, which we consider as our best prediction.
\end{itemize}
\begin{table
\begin{center}
\begin{tabular}{l|ccc}
label & \TTBAR & \TTBARDEC & \BBFLRES \\ \hline\hline
generator & \hvq~\cite{Frixione:2007nw} & \dec~\cite{Campbell:2014kua} & \bbfourl{} \\
framework & \POWHEGBOX{} & \VTWO{} & \RES{} \\
NLO matrix elements & $ t\bar t$ & $ t(\to \ell^+\nu_{\sss \ell} b)\,\bar t(\to l^-\bar{\nu}_{\sss l} \bar b)$ & $\fourl\,\ensuremath{b \,\bar b}\xspace$ \\
decay accuracy & LO+PS & NLO+PS & NLO+PS \\
NLO radiation & single & multiple & multiple \\
spin correlations & approx. & exact & exact \\
off-shell $\ensuremath{t \bar t}\xspace$ effects & BW smearing & LO $ b\bar b 4\ell$ reweighting & exact \\
\ensuremath{W t}\xspace \& non-resonant effects & no & LO $ b\bar b 4\ell$ reweighting & exact \\
$ b$-quark massive & yes & yes & yes \\
\hline
\end{tabular}
\end{center}
\caption{Labels and characteristic features of the three generators considered in this paper.}
\label{tab:generators}
\end{table}
The main physics features of the various generators
and the labels that will be used to identify the corresponding predictions
are listed in~\refta{tab:generators}.
All generators are run with their default settings and
are interfaced to
\PythiaEightPone{}.
The \bbfourl{} generator implements
the scale choice
of~\refeqs{eq:ttscale}{eq:zscale}, while in \DEC{} and \hvq{} a scale
corresponding to~\refeq{eq:ttscale} is used.
In order to quantify the impact of various aspects of the
resonance-aware approach, in~\refse{sec:ttbarPhenomenology} we will
compare various settings of the \bbfourl{} generator where
some resonance-aware improvements are turned on and off or are replaced by
certain approximations. Specifically, the following settings will be considered:
\begin{itemize}
\item[(a)] the resonance-aware formalism is switched on
with default settings;
\item[(b)] the resonance-aware formalism
is switched off,
which corresponds to using the traditional \POWHEG{} approach;
\item[(c)] the resonance-aware formalism is switched off, but a resonance
assignment is guessed based on the kinematic structure of the events,
according to the method described in~\refapp{app:kinematic_guess};
\item[(d)] the resonance-aware formalism is switched on, but,
instead of applying the multiple-radiation scheme of~\refeq{eq:allrad},
only a single radiation is generated with \POWHEG{} according to~\refeq{eq:powheg};
\item[(e)] same as (d), but the resonance
information is stripped off in the \POWHEG{} Les Houches event file
before passing it to the showering program.
\end{itemize}
The various \bbfourl{} settings and corresponding labels are summarised in~\refta{tab:bb4loptions}.
\begin{table
\begin{center}
\begin{tabular}{lc||cccc}
&\BBFLRES{} setting & resonance-aware & radiation in & flags in the \\
\phantom{xx}& label &matching & production and decay & {\tt powheg.input}\\ \hline
(a)&{\bbfourlPYdefault} & yes & multiple & 1, 0, 0, 0 \\
(b)&{\noresPY} & no & single & 0, 0, 0, 1 \\
(c)&{\noresiPY} & no (kinematic guess) & single & 0, 0, 1, 1 \\
(d)&{\bbfourlPYnoallrad} & yes & single & 0, 0, 0, 0 \\
(e)&{\stripresPY} & yes (stripped off) & single & 0, 1, 0, 0\\
\hline
\end{tabular}
\end{center}
\caption{Labels for the various \BBFLRES{} predictions that are considered
and compared in this paper. In the last column we list the values of the
\POWHEGBOX{} flags {\tt allrad}, {\tt stripres}, {\tt guessres}, {\tt
nores}, to be specified in the {\tt powheg.input} file.}
\label{tab:bb4loptions}
\end{table}
\subsection{Physics objects}
In the subsequent sections we study various observables defined in terms of the following physics objects.
\begin{enumerate}[(a)]
\item \label{item:Bhad} We denote as $B$ and $\bar{B}$ hadron the hardest
$b$-flavoured and $\bar b$-flavoured hadron in the event.
\item Final-state hadrons are recombined into jets using the
{\tt FastJet} implementation~\cite{Cacciari:2011ma} of the anti-$\ensuremath{k_{\sss T}}\xspace$ jet
algorithm~\cite{Cacciari:2008gp}
with $R=0.5$.
\item We denote as $b$-jet~($\ensuremath{{j_{\rm \sss B}}\xspace}$) and
anti-$b$-jet~($\ensuremath{j_{\rm \sss {\bar B}}}\xspace$) the jet that contains the hardest $B$ and $\bar{B}$
hadron, respectively.
When examining results
obtained with the hadronization switched off, jets are $b$-tagged based on
$b$ quarks rather than $B$ hadrons.
\item Leptons, neutrinos and missing transverse energy are identical to their
corresponding objects at matrix-element level, since we switched off QED
radiation and hadron decays in \PythiaEight{}.
\item Reconstructed $W^+$ and $W^-$ bosons are identified with the
corresponding off-shell lepton-neutrino pairs in the hard matrix
elements.\footnote{Similarly as for top resonances, also $W$ resonances are
identified with their off-shell decay products according to the resonance
history of the event at hand. This information is written in the shower
record and propagated through the shower evolution. In this way, possible
QED radiation off charged leptons is included into the $W$-boson momentum
at Monte Carlo truth level. However, since electromagnetic radiation
from \PythiaEight{} is turned off in our analysis, each $W$ boson
coincides with a bare lepton--neutrino system.}
\item Reconstructed top and anti-top quarks are defined as off-shell $W^+\ensuremath{{j_{\rm \sss B}}\xspace}$
and $W^-\ensuremath{j_{\rm \sss {\bar B}}}\xspace$ pairs, respectively, i.e.~$b$-jets and $W$-bosons are matched
based on charge and $b$-flavour information at Monte-Carlo truth level. The
same approach is used for $\ell^+\ensuremath{{j_{\rm \sss B}}\xspace}$ and $\l^-\ensuremath{j_{\rm \sss {\bar B}}}\xspace$ pairs.
\end{enumerate}
Unless stated otherwise, in kinematic distributions we always
perform an average over the $t$ and $\bar{t}$ case~(thus also on
lepton--antilepton, $b$--anti-$b$, etc.).
The top-pair observables in \refses{sec:ttNLOdec}{sec:RES_HVQ}
are computed by requiring the presence of a $b$ and a $\bar b$ jet with
\begin{eqnarray}
\label{eq:jet_cuts}
\ensuremath{p_{\sss T}}\xspace^{j} > 30~\text{GeV}\xspace, \qquad \qquad |\eta^{j}|<2.5\,,
\end{eqnarray}
and applying the following leptonic cuts,
\begin{eqnarray}
\label{eq:leptonic_cuts}
\ensuremath{p_{\sss T}}\xspace^{l} > 20~\text{GeV}\xspace, \qquad |\eta^{l}|<2.5\,, \qquad \ensuremath{p_{\sss T}}\xspace^{\sss \mathrm{miss}} > 20~\text{GeV}\xspace,
\end{eqnarray}
where $l=\ell^+, l^-$ and $\ensuremath{p_{\sss T}}\xspace^{\sss\mathrm{miss}}$ is obtained from
the vector sum of the transverse momentum of the neutrinos in the final
state.
\section{Top-pair dominated observables}\label{sec:ttbarPhenomenology}
Here we present numerical predictions for
$pp\to e^+\nu_e\mu^-\bar\nu_\mu\ensuremath{b \,\bar b}\xspace+X$ at $\sqrt{s}=$8\,TeV.
In particular, we study various observables that are sensitive
to the shape of top resonances.
\subsection{Comparison with traditional NLO+PS matching}
\label{sec:nores}
In the following, we compare nominal \bbfourl{} predictions,
generated with default settings, with results
obtained by switching off the resonance-aware formalism~(i.e.~setting the
flag \verb!nores! to 1). In this way we get results that are fully equivalent
to a \VTWO{}~(or ``traditional'') implementation.
For this comparison we do not impose any cuts,
i.e.~we perform a fully inclusive analysis that involves, besides $t\bar t$ production, also
significant contributions from $\ensuremath{W t}\xspace$ single-top production.
Events generated with the traditional implementation do not contain any
information whatsoever about their resonance structures. We label the curves
obtained by showering these events as \nores{}. Because the resonance
information is not available, the shower generator will not preserve the
virtualities of the resonances.
In order to further explore the usability of
the \nores{} results, we also consider the possibility of reconstructing the
resonance information of the Les Houches event on the basis of its kinematic
proximity to one of the possible resonant configurations.
Specifically, we perform an educated guess of the resonance structure of the event, assigning
it to a $t{\bar t}$ or to a $Z$ resonance configuration~(see~\refse{sec:reshists}),
and assigning the radiation either to the initial
state or to the outgoing $b$'s.
The curves obtained this way are labelled as \noresi{} and the procedure for
reconstructing the resonance information from the event kinematics is
detailed in~\refapp{app:kinematic_guess}.
We first consider, in~\reffi{fig:m_w_jbot-res-nores},
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_m_w_jbot-cut1-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_m_l_jbot-cut1-rebin-pyMEC1}
\end{center}
\caption{NLO+PS predictions for the invariant mass of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$~(left) and of the $l\ensuremath{{j_{\rm \sss B}}\xspace}$~(right)
systems obtained with the new \bbfourl generator. We compare our default resonance-aware predictions
(\resdefault) against the ``traditional'', i.e. resonance-unaware, implementation (\nores{}) and
a prediction where the event-by-event resonance information is obtained from a guess based on kinematics.
In the ratio plot we illustrate relative deviations with respect to \resdefault.}
\label{fig:m_w_jbot-res-nores}
\end{figure}
the invariant mass of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$ and of the $l\ensuremath{{j_{\rm \sss B}}\xspace}$ systems.
In the \nores{} case, we observe that
the reconstructed mass peak has a wider shape. This is
expected, since neither the \POWHEGBOX{} nor the shower program preserve the
virtuality of the top resonances. In the \noresi{} case
the width of the peak is diminished, although not quite at the level of
the resonance-aware prediction, labelled as \resdefault{}. We also
observe a mild shift in the peak in the \noresi{} case,
which improves the agreement with the \default{} result.
The distribution in the mass of the lepton-\ensuremath{{j_{\rm \sss B}}\xspace}{} system
also shows marked differences in shape in the region that is most
relevant for a top-mass determination, with more pronounced differences in
the \nores{} case.
The above findings suggest that the width of the peak is determined both by the shower
generator being aware of the resonances in the Les Houches event, and by the
hardest radiation generation being performed in a way that is consistent with
the resonance structure.
In order to assess the effects
that originate solely from resonance-aware matching and showering
in a more accurate way, in~\reffi{fig:m_w_jbot-res-nores-more}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-stripres_nores_PY8_m_w_jbot-cut1-pyMEC1}
\end{center}
\caption{Invariant mass of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$ system obtained with the \bbfourl generator.
We compare our resonance-aware predictions without employing the multiple radiation scheme (\bbfourlNoallrad{})
against the ``traditional'', i.e. resonance-unaware, implementation (\nores{}) and
a prediction where any resonance information is stripped off the Les Houches event file (\stripres{}).
In the ratio plot we illustrate relative deviations with respect to \bbfourlNoallrad{}.}
\label{fig:m_w_jbot-res-nores-more}
\end{figure}
we disable the multiple radiation scheme
of~\refeq{eq:allrad} (by setting {\tt allrad 0})
and compare the resulting resonance-aware predictions (\bbfourlNoallrad{})
against the cases where resonance information
is removed from the Les Houches event before
showering~(\stripres{}) or the case where the resonance-aware system is
completely switched off (\nores{}).
We find that the \stripres{} result lies between the \bbfourlNoallrad{}
and the \nores{} ones, somewhat
closer to the latter, and the differences between
the various predictions are considerable.
Therefore, we conclude that the observed widening of the
peak in~\reffis{fig:m_w_jbot-res-nores}{fig:m_w_jbot-res-nores-more} can be
attributed to both
shortcomings of a resonance unaware parton
shower matching: the parton shower reshuffling not
preserving the resonance masses, and the
uncontrolled effects of resonances at the level of
the first emission in the traditional \POWHEG{} approach.
In~\reffi{fig:bjet-allrad-nores} we display
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_bjet-m-cut1-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_bprofile-cut1-integ-pyMEC1}
\end{center}
\caption{Mass~(left) and profile~(right) of the $b$-jet $\ensuremath{{j_{\rm \sss B}}\xspace}$. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-res-nores}.}
\label{fig:bjet-allrad-nores}
\end{figure}
the $\ensuremath{{j_{\rm \sss B}}\xspace}$ mass and profile,
defined as
\begin{equation}\label{eq:profile}
\Etbj(\Delta_{\sss\mathrm R})= \int {\mathrm d}\sigma\;\frac{\sum_j
\ensuremath{p_{\sss T}}\xspace^{\sss j}\,
\theta\!\(\Delta_{\sss \mathrm R}-\Delta_{\sss\mathrm R}^{(j,\ensuremath{{j_{\rm \sss B}}\xspace})}\)}{\ensuremath{p_{\sss T}}\xspace^\ensuremath{{j_{\rm \sss B}}\xspace}}\,.
\end{equation}
This observable corresponds to the cross section weighted by the fraction of the total
hadronic transverse momentum of the particles contained in a given cone
around the jet axis, with respect to the transverse momentum of the $b$-jet.
Again we observe marked differences among the \default{} and the \nores{}
results, and, to a lesser extent, between the \default{} and \noresi{}
ones. Both plots suggest that in the \nores{} case there is less activity
around the $B$ hadron, leading to smaller jet masses and to a slightly
steeper jet profile. The particularly pronounced shape distortion of the $\ensuremath{{j_{\rm \sss B}}\xspace}$
mass plot near 10~GeV in the \noresi{} case can be tentatively attributed to
the transition from the region where radiation~(generated with the
traditional method) does not change the mass of the resonance by an amount
comparable or larger than its width, to the region where it does, so that we
see the difference between the \noresi{} and \default{} results grow with
larger jet masses.
In~\reffi{fig:bfrag-allrad-nores}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_bfrag-avg-cut1-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nores_PY8_bptdec-avg-cut1-rebin-pyMEC1}
\end{center}
\caption{$B$ fragmentation function and $B$-hadron transverse momentum in the top
decay frame. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-res-nores}.}
\label{fig:bfrag-allrad-nores}
\end{figure}
we compare the $B$ fragmentation function and the $B$-hadron transverse momentum
computed in the reconstructed top-decay rest frame. The $\xB$ variable is
defined as the $B$ energy in the reconstructed top rest frame normalized to
the maximum value that it can attain at the given top virtuality, while
\ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{} is the transverse momentum of the $B$ relative to the recoiling $W$
in the same frame. We find marked differences also for these
distributions. While in the case of the $\ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace$ variable we see
a reasonable consistency between the \noresi{} and \default{} results,
the agreement deteriorates in the case of the fragmentation function.
We conclude that the consistent treatment of resonances implemented in
the \bbfourl{} generator yields a narrower peak for the reconstructed top distribution
with respect to a traditional (resonance-blind) NLO+PS matching approach.
Furthermore, a large part of the difference is not related to the
lack of resonance information at the level of the shower generator, and thus
cannot be reduced by using a more sophisticated interface to the shower
based on a resonance-guessing approach of kinematic nature.
\subsection{Comparison with the {\tt ttb\_NLO\_dec} generator}
\label{sec:ttNLOdec}
In this section we compare the \bbfourl{} generator
against the {\tt ttb\_NLO\_dec} generator of~\citere{Campbell:2014kua}.
The standard $\ensuremath{t \bar t}\xspace$ cuts
of~\refeqs{eq:jet_cuts}{eq:leptonic_cuts} are applied throughout.
We examined a large set of distributions, but here we
only display the most relevant ones, and those that show the largest
discrepancies.
We begin by showing in~\reffi{fig:m_w_jbot-res-dec}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_m_w_jbot-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_m_l_jbot-cut3-rebin-pyMEC1}
\end{center}
\caption{Invariant mass of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$~(left) and of the $l\ensuremath{{j_{\rm \sss B}}\xspace}$~(right)
systems. Comparison of NLO+PS predictions obtained with the \bbfourl~(\BBFLRES) and the
\DEC~(\TTBARDEC) generators. In the ratio plot we illustrate relative deviations with respect to the \BBFLRES{} prediction. }
\label{fig:m_w_jbot-res-dec}
\end{figure}
the invariant mass distribution of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$ and $l\ensuremath{{j_{\rm \sss B}}\xspace}$ systems. We
observe remarkable agreement between the \bbfourl{} and \DEC{} generators,
especially in the description of the reconstructed top peak and of the
shoulder in the lepton-\ensuremath{{j_{\rm \sss B}}\xspace}{} invariant mass. This agreement is quite
reassuring. In fact, in the \DEC{} generator, the separation of radiation in
production and resonance decay is unambiguous, while in \bbfourl{} it is
based on a probabilistic approach according to a kinematic proximity criterion.
Thus, in the light of \reffi{fig:m_w_jbot-res-dec},
the former generator supports the method of separation of resonance
histories adopted by the latter. On the other hand, off-shell
and non-resonant effects are implemented in
the \DEC{} generator in LO approximation, by reweighting the on-shell
result.
Thus the \bbfourl{} results support the validity of this
approximation in the \DEC{} implementation.
As an indicative estimate of the potential implications
for precision $m_t$ determination, we have determined that
in a window of $\pm 30$~GeV around the peak of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$ distributions, the
average $W\ensuremath{{j_{\rm \sss B}}\xspace}$ mass computed with the \dec{} generator is roughly 0.1~GeV smaller
than the one from \bbfourl{}.
The NLO distribution in the mass of the reconstructed top was also examined
in~\citere{Campbell:2014kua} (sec.~3.2, Fig.~3). There, the
\DEC{} fixed-order NLO result was compared to the fixed-order NLO result
of~\citere{Denner:2012yc}, and the former was found to be enhanced by about
10\%{} in a region of roughly 1~GeV around the peak. This comparison was
carried out with massless $b$ quarks, since mass effects were not available
in~\citere{Denner:2012yc}. We computed the same distribution and carried
out the same NLO comparison, using however the \bbfourl{} generator instead
of the result of~\citere{Denner:2012yc} and taking into account $b$-mass effects.
Again, we find the same enhancement in the \DEC{} NLO
result. However, in the fully showered result we see instead a small
suppression of the peak in the \DEC{} relative to the \bbfourl{} generator,
suggesting that the NLO difference tends to be washed out by showering effects.
We examined several distributions involving $b$-jets (here again we average over
the $b$- and $\bar{b}$-jet contributions). We found no appreciable difference
for the $b$-jet transverse momentum, while we did find significant differences
in the jet mass and the jet profile, displayed in~\reffi{fig:bjet-allrad-dec}.
\begin{figure}[bt]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_bjet-m-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_bprofile-cut3-integ-pyMEC1}
\end{center}
\caption{Mass~(left) and profile~(right) of the $\ensuremath{{j_{\rm \sss B}}\xspace}$. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-res-dec}.}
\label{fig:bjet-allrad-dec}
\end{figure}
Both plots indicate that the \bbfourl{} generator yields slightly wider
$b$-jets as compared to the \DEC{} one.
In~\reffi{fig:8TeV_fragb} we plot
the $B$ fragmentation function and the \ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{} observables.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_bfrag-avg-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_dec_PY8_bptdec-avg-cut3-rebin-pyMEC1}
\end{center}
\caption{The $B$ fragmentation function and transverse-momentum distribution of the
\ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{} observable. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-res-dec}.}
\label{fig:8TeV_fragb}
\end{figure}
We find that the fragmentation function is slightly harder, and the \ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{}
distribution is slightly softer in the \bbfourl{} case.
Again, this is consistent with the
observation of slightly reduced radiation from $b$'s in the \bbfourl{} case.
We have verified that this feature persists also when hadronization is
switched off in \PythiaEight{}.
Although the differences in the $b$-jet structure are quite significant,
they are not sufficient to induce an observable shift in the reconstructed mass peak. This
could only happen if the difference in the jet profile caused a consistent
difference in the jet energy, due to energy loss outside the jet-cone. This does
not seem to be the case since the jet profiles become similar in the two
generators already for $\Delta_{\rm R}< 0.5$.
\subsection{Comparison with the {\tt hvq} generator}
\label{sec:RES_HVQ}
In this section we compare the \bbfourl{} generator against the
\hvq{} generator of~\citere{Frixione:2007vw}, which is based on on-shell NLO matrix
elements for \ensuremath{t \bar t}\xspace{} production.
Again the standard $\ensuremath{t \bar t}\xspace$ cuts of~\refeqs{eq:jet_cuts}{eq:leptonic_cuts} are applied throughout.
The $W\ensuremath{{j_{\rm \sss B}}\xspace}$ and $l\ensuremath{{j_{\rm \sss B}}\xspace}$~mass distributions, shown in~\reffi{fig:m_w_jbot-allrad-hvq},
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_m_w_jbot-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_m_l_jbot-cut3-rebin-pyMEC1}
\end{center}
\caption{Invariant mass of the $W\ensuremath{{j_{\rm \sss B}}\xspace}$~(left) and of the $l\ensuremath{{j_{\rm \sss B}}\xspace}$~(right)
systems. Comparison of NLO+PS predictions obtained with the \bbfourl~(\BBFLRES) and the
\hvq~(\TTBAR) generators. In the ratio plot we illustrate relative deviations with respect to the \BBFLRES{} prediction.}
\label{fig:m_w_jbot-allrad-hvq}
\end{figure}
show reasonably good agreement
between the two generators as far as the shape of the
$W\ensuremath{{j_{\rm \sss B}}\xspace}$ peak and of the $l\ensuremath{{j_{\rm \sss B}}\xspace}$ shoulder are concerned.
However, for large top virtualities, i.e. in the tails of both distributions, sizable differences can be appreciated.
As we will see below, such differences originate from the fact that,
in this region, the \bbfourl{} generator tends to radiate considerably less,
which results in narrower b-jets as compared to the \hvq{} generator.
We note that the observed deviations
with respect to the \hvq{} generator are more drastic than the ones
observed in section~\ref{sec:ttNLOdec} for the \DEC{} generator.
The $m_{W\ensuremath{{j_{\rm \sss B}}\xspace}}$ distribution on the left of \reffi{fig:m_w_jbot-allrad-hvq}
additionally suggests a non-negligible shift in the reconstructed top mass between the two generators.
In fact, we determined that in a window of $\pm 30$~GeV
around the peak of the $m_{W\ensuremath{{j_{\rm \sss B}}\xspace}}$ distributions, the average $W\ensuremath{{j_{\rm \sss B}}\xspace}$ mass computed
with the \hvq{} generator is roughly 0.5~GeV smaller than with the \bbfourl{} one.
In~\reffi{fig:bjet-allrad-hvq}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_bjet-m-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_bprofile-cut3-integ-pyMEC1}
\end{center}
\caption{Mass~(left) and profile~(right) of $\ensuremath{{j_{\rm \sss B}}\xspace}$. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-allrad-hvq}.}
\label{fig:bjet-allrad-hvq}
\end{figure}
we show distributions in the $b$-jet mass and profile, as defined in \refeq{eq:profile}.
Both plots indicate significantly narrower $b$-jets in the predictions
obtained with the \bbfourl{} generator.
Similarly, as
shown in~\reffi{fig:8TeV_fragb-allrad-hvq},
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_bfrag-avg-cut3-rebin-pyMEC1}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_hvq_PY8_bptdec-avg-cut3-rebin-pyMEC1}
\end{center}
\caption{The $B$ fragmentation function and \ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{} distribution. Absolute predictions and ratios as
in~\reffi{fig:m_w_jbot-allrad-hvq}.}
\label{fig:8TeV_fragb-allrad-hvq}
\end{figure}
the \bbfourl{} generator yields a harder $B$ fragmentation function and a softer \ensuremath{p^{\rm\sss B}_{\sss \rm T,dec}}\xspace{}
distribution.
The pattern we observe for the structure of $b$-jets is consistent with the
fact that the \bbfourl{} generator has a reduced radiation in $b$-jets with
respect to \PythiaEight{}. In the \hvq{} generator, radiation from the $b$'s
is handled exclusively by \PythiaEight{}, while, in the \bbfourl{} generator,
the hardest radiation from the $b$ is handled by \POWHEG{}. It should be
stressed, however, that the $B$ fragmentation function has a considerable
sensitivity to the hadronization parameters. It would therefore be desirable
to tune these parameters to $B$ production data in $e^+ e^-$ annihilation,
within the \POWHEG{} framework, in order to perform a meaningful comparison.
In~\reffi{fig:8TeV_res-allrad_dec_hvq_PY8_m_w_jbot} we show a summary of
the shape of the reconstructed top peak comparing each of the available \POWHEG{}
generators for $\ensuremath{t \bar t}\xspace$ production: \bbfourl{}, \DEC{} and \hvq{}. We notice a
fair consistency between the \bbfourl{} generator and the \DEC{} one, while larger deviations
are observed comparing against \hvq{}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.6\textwidth]{plots/8TeV_res-allrad_dec_hvq_PY8_m_w_jbot-cut3-pyMEC1}
\end{center}
\caption{The $W\ensuremath{{j_{\rm \sss B}}\xspace}$ mass distribution near the top peak for the three
generators \bbfourl{}~(\BBFLRES), \DEC{}~(\TTBARDEC) and \hvq{}~(\TTBAR). In the ratio plot we illustrate relative deviations with respect to the \BBFLRES{} prediction.}
\label{fig:8TeV_res-allrad_dec_hvq_PY8_m_w_jbot}
\end{figure}
\section{Jet vetoes and single-top enriched observables}
\label{sec:WtPhenomenology}
In this section we investigate the behaviour of the $\bbfourl$ generator in
the presence of $b$-jet and light-jet vetoes.
Such kinematic restrictions are widely used in order to reduce top
backgrounds in $H\to\ensuremath{W^+W^-}\xspace$ studies and in many other analyses that involve
charged leptons and missing energy.
Also, jet vetoes play an essential role for experimental studies of $Wt$ single-top
production~\cite{Chatrchyan:2014tua,Aad:2015eto}.
In particular, the separation of \ensuremath{W t}\xspace{} and $t\bar{t}$ production typically
relies upon the requirement that one large transverse-momentum $b$-jet is
missing in the first process.
From the theoretical point of view, the separation of \ensuremath{W t}\xspace{} and $t\bar{t}$
production is not a clear cut one, since the two processes interfere. As
pointed out in the introduction, in the $\bbfourl$ generator this problem is
solved by providing a unified description of $\ensuremath{t \bar t}\xspace$ and \ensuremath{W t}\xspace{} production
and decay, where also interference effects are included at NLO. Thus jet
vetoes are expected to enrich the relative single-top content of $\bbfourl$
samples, resulting in significant differences with respect to other
generators that do not include $\ensuremath{W t}\xspace$ contributions and interferences at NLO.
The \bbfourl{} generator is particularly well-suited for the study of jet
vetoes also because it includes $b$-mass effects, NLO radiation in
top-production and -decay subprocesses, as well as resummation of multiple
QCD emissions and hadronization effects as implemented in the parton shower.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth, trim=0 0 0 0,clip]{plots/8TeV_res-dec-hvq-NLO_PY8_bjet-pt-cut1} \\
\vspace{-0.099cm}
\includegraphics[width=0.49\textwidth]{plots/8TeV_res-allrad_nlo_PY8_bjet-pt-cut1}
\end{center}
\caption{Distribution in the $b$-jet transverse momentum: comparison of
NLO+PS predictions obtained with the three generators \bbfourl~(\BBFLRES),
\DEC~(\TTBARDEC) and \hvq (\TTBAR). The middle frame illustrates relative
NLO+PS deviations with respect to the \BBFLRES{} prediction, while the
lower frame compares \BBFLRES{} versus corresponding fixed-order NLO results.}
\label{fig:res-dec-hvq-lo-bjpt}
\end{figure}
A first picture of the $b$-jet activity in the three generators, \bbfourl{},
\DEC{} and \hvq{}~(labelled according to~\refta{tab:generators} as
\BBFLRES{}, \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} and \ensuremath{t \bar t}\xspace{} respectively), is provided
by~\reffi{fig:res-dec-hvq-lo-bjpt}, where we compare NLO+PS distributions in
the transverse momentum of the $b$-jet.
More precisely, the plotted observable
corresponds to the sum of the $b$- and \mbox{$\bar b$-jet} spectra and was computed
in absence of any acceptance cut.
Thus it involves potentially enhanced contributions from single-top topologies, which
can lead to significant deviations between the \ensuremath{t \bar t}\xspace{} prediction\footnote{%
In order to make sure that, apart form the absence of $\ensuremath{W t}\xspace$ contributions,
the \ensuremath{t \bar t}\xspace{} predictions are internally consistent, we have checked that
off-shell top contributions~(which are modelled through an heuristic
Breit--Wigner smearing approach in \hvq{}) play only a marginal role for
the observable at hand. To this end we have applied cuts to the $t$ and
$\bar{t}$ virtualities, imposing that they should not differ from the $t$
pole mass by more than 15~GeV. The effect of such cuts was found to be
negligible.} and the ones that implement off-shell $\fourl\,\ensuremath{b \,\bar b}\xspace$ matrix
elements.
At large transverse momentum, the various predictions have rather similar
shape, but the \ensuremath{t \bar t}\xspace{} result features a clear deficit of about 10\% with
respect the \BBFL{} and \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} ones. This can be attributed to the
missing single-top contributions in the \hvq{} generator.
At high $\ensuremath{p_{\sss T}}\xspace$, thanks to the implementation of $\ensuremath{W t}\xspace$ contributions via exact
Born matrix elements for $\ensuremath{pp \,\to \,e^+\nu_e\mu^-\bar\nu_\mu\, \bbbar}\xspace$, the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} prediction is found
to be in good agreement with the \BBFL{} one.
At small transverse momenta, the relative weight of $\ensuremath{W t}\xspace$ production becomes
more important, and the deficit of the \ensuremath{t \bar t}\xspace{} prediction grows rather
quickly, reaching up to 50\% for very small transverse momenta. The
\ensuremath{t \bar t(b\bar b4\ell)}\xspace{} and \BBFL{} predictions remain in good agreement down to
$p_{{\rm T},\ensuremath{{j_{\rm \sss B}}\xspace}}\simeq 10$~GeV, but at smaller transverse momenta the
\ensuremath{t \bar t(b\bar b4\ell)}\xspace{} one develops a deficit that grows up to about 25\%. This can
be attributed, at least in part, to the increased importance of $\ensuremath{W t}\xspace$
channels combined with the fact that these channels are not supplemented by
an appropriate NLO correction in the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} predictions. We also note
that the discrepancy at hand can be interpreted as a kinematic shift of a few
GeV only, while the enhancement of the resulting correction can be attributed
to the pronounced steepness of the absolute $p_{{\rm T},\ensuremath{{j_{\rm \sss B}}\xspace}}$ distribution in the
soft region. Its sign is consistent with the fact that radiation arising
from $\fourl\,\ensuremath{b \,\bar b}\xspace$ NLO matrix elements is expected to be rather soft in
the presence of single-top contributions with initial-state collinear $g\to
\ensuremath{b \,\bar b}\xspace$ splittings, while in the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{}{} generator radiation is
always emitted as if all $b$-quarks would arise from top decays, which
results in a harder emission spectrum.
The lower frame of~\reffi{fig:res-dec-hvq-lo-bjpt} illustrates the
relative importance of matching and shower effects in the \bbfourl{}
generator, comparing against corresponding fixed-order NLO predictions.
Again we observe nontrivial shape effects in the soft region. While they are
not directly related to the differences observed in the middle frame, such
effects highlight the importance of a consistent treatment of radiation and
shower effects at small $b$-jet $\ensuremath{p_{\sss T}}\xspace$. On the other hand the good agreement
between the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{}{} and \BBFL{} predictions down to 10~GeV suggests
that matching and pure shower effects are reasonably well under control in
the bulk of the phase space.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_dec_hvq_PY8_0b_excl_pt_thresh}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_dec_hvq_PY8_0j_excl_pt_thresh}\\
\vspace{-0.068cm}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_nlo_PY8_0b_excl_pt_thresh}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_nlo_PY8_0j_excl_pt_thresh}
\caption{Integrated cross sections at 8~TeV in jet bins with zero $b$-jets
as a function of the jet-$\pT$ threshold. The left plot is inclusive with
respect to extra jet radiation~($\ensuremath{n_{j}}\xspace{}\ge\ensuremath{n_{b}}\xspace{}=0$), while the right plot is
exclusive ~($\ensuremath{n_{j}}\xspace=\ensuremath{n_{b}}\xspace=0$). Absolute predictions and ratios as
in~\reffi{fig:res-dec-hvq-lo-bjpt}. }
\label{fig:8TeV_zerob}
\end{center}
\end{figure}
Jet-binning and jet-veto effects are studied
in~\reffis{fig:8TeV_zerob}{fig:8TeV_oneb}. For this analysis we
apply again the lepton selection cuts of~\refeq{eq:leptonic_cuts} and,
at variance with the $b$-jet definition in~\refse{sec:ttbarPhenomenology},
we identify as $b$-jets those
jets containing at least a $b$- or \mbox{$\bar b$-flavoured} hadron, irrespectively of its hardness.%
\footnote{At fixed-order NLO, jet clustering and
$b$-jet tagging are applied at parton level.}
Events are categorised
according to the number of~(light \textit{or} heavy-flavour) jets,
\ensuremath{n_{j}}\xspace{}, and to the number of $b$-jets, \ensuremath{n_{b}}\xspace{},
in the rapidity range $|\eta|<2.5$,
while we vary the jet transverse-momentum threshold \pTthr{} that defines
jets.
In~\reffi{fig:8TeV_zerob}, to investigate the effect of a $b$-jet veto,
the integrated cross sections is plotted versus the jet-veto threshold,
\pTthr{}. In the left plot the veto acts only on \mbox{$b$-jets} \mbox{($\ensuremath{n_{j}}\xspace{}\ge\ensuremath{n_{b}}\xspace{}=0$),}
while in the right plot a veto against light and $b$-jets is
applied~\mbox{($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=0$).}
For $\pTthr{} \gtrsim 80$~GeV the vetoed cross section is dominated by $t
\bar t$ production and quickly converges towards the inclusive result. In
this region we observe few-percent level agreement between the
\ensuremath{t \bar t(b\bar b4\ell)}\xspace{}{} and \BBFL{} predictions, while the on-shell \ensuremath{t \bar t}\xspace{} prediction
features a 10\% deficit due to the missing single-top topologies.
Reducing the jet-veto scale increases this deficit up to $-30\%$ in the case
of the inclusive $\ensuremath{n_{b}}\xspace{}=0$ cross section. This finding is well consistent
with the size of finite-width effects reported in~\citere{Cascioli:2013wga}.
In the case of the exclusive zero-jet cross section~($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=0$, shown
on the right) the deficit of the \ensuremath{t \bar t}\xspace{} prediction
is even more pronounced and reaches up to $-50\%$
at $\pTthr{}=10$~GeV.
Also the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} results feature a similar, although less pronounced,
deficit as the \ensuremath{t \bar t}\xspace{} ones in the soft region.
This can be attributed to the fact that initial-state radiation in both, the
\hvq{} and \DEC{}, generators is computed with on-shell tops, and thus
overestimates the radiation produced near the single-top kinematic region.
Matching and pure shower effects are illustrated in the lower frames
of~\reffi{fig:8TeV_zerob}. Both in the inclusive~($\ensuremath{n_{j}}\xspace{}>\ensuremath{n_{b}}\xspace{}=0$) and
exclusive~($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=0$) case we observe that, down to 20~GeV, NLO+PS
predictions feature an increasingly strong enhancement with respect to
fixed-order ones. This can be attributed to shower-induced losses of $b$-jet
transverse momentum. In the exclusive case~($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=0$) this
enhancement is somewhat milder, which we tentatively attribute to the
interplay of parton shower radiation with the additional light-jet veto.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_dec_hvq_PY8_1b_excl_pt_thresh}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_dec_hvq_PY8_1j1b_excl_pt_thresh}\\
\vspace{-0.065cm}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_nlo_PY8_1b_excl_pt_thresh}
\includegraphics[width=0.49\textwidth]{single-top-plots/8TeV_res-allrad_nlo_PY8_1j1b_excl_pt_thresh}
\caption{Integrated cross sections at 8~TeV in jet bins with one $b$-jet as
a function of the jet-$\pT$ threshold. The left plot is inclusive with
respect to extra jet radiation~($\ensuremath{n_{j}}\xspace{}\ge\ensuremath{n_{b}}\xspace{}=1$), while the right plot is
exclusive~($\ensuremath{n_{j}}\xspace=\ensuremath{n_{b}}\xspace=1$). Absolute predictions and ratios as
in~\reffi{fig:res-dec-hvq-lo-bjpt}. }
\label{fig:8TeV_oneb}
\end{center}
\end{figure}
In~\reffi{fig:8TeV_oneb} we plot the cross section with exactly one
$b$-jet above the threshold \pTthr{}, i.e.~we veto additional $b$-jets above
this threshold. Again, inclusive results ($\ensuremath{n_{j}}\xspace{}\ge \ensuremath{n_{b}}\xspace{}=1$, shown on the
left) are compared with exclusive ones ($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=1$, shown on the right).
The one-$b$-jet bin is typically used in $Wt$ single-top analyses. Similarly
as for the zero-$b$-jet case, the difference between the \BBFL{} and \ensuremath{t \bar t}\xspace{}
results points to an increasingly important single-top contribution at small
$\pTthr{}$. Its quantitative impact is consistent with the fixed-order
results of~\citere{Cascioli:2013wga}, and at $\pTthr{}=30$~GeV it amounts to
about 10\% and $20\%$, respectively, in the inclusive and exclusive cases.
Similarly as for the zero-$b$-jet case, \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} predictions feature a
qualitatively similar but quantitatively less pronounced deficit with respect
to the \BBFL{} predictions. Matching and shower effects turn out to be
rather mild in the inclusive case, probably due to the fact that the absolute
distribution is not particularly steep in the limit of small transverse
momentum. In contrast, the exclusive one-jet cross
section~($\ensuremath{n_{j}}\xspace{}=\ensuremath{n_{b}}\xspace{}=1$) is much more sensitive to the jet-veto scale,
which leads to sizable matching and shower
effects at small \pTthr{}.
In summary, jet-vetoed cross sections can involve enhanced single-top
contributions that are completely missing in the \ensuremath{t \bar t}\xspace{} predictions
obtained with the \hvq{} generator while they are significantly
underestimated in the \ensuremath{t \bar t(b\bar b4\ell)}\xspace{} predictions
of the \DEC{}
generator, where single-top contributions are
implemented via LO reweighting~\cite{Campbell:2014kua}. In practice such a reweighting approach
ceases to work in phase-space regions far away from the double-resonant
region.
\section{Conclusions}\label{sec:conc}
In this paper we have presented the first Monte Carlo generator that provides
a fully consistent NLO+PS simulation of $\ensuremath{t \bar t}\xspace$ production and decay in the
different-flavour dilepton channel, including all finite-width and
interference effects.
This new generator, dubbed $\bbfourl$, is based on the full NLO matrix
elements for the process $pp\toe^+\nu_e\mu^-\bar\nu_\mu\,\ensuremath{b \,\bar b}\xspace$. This guarantees NLO accuracy
in $\ensuremath{t \bar t}\xspace$ production and decay, as well as the exact treatment of spin
correlations and off-shell effects in top decay.
Top resonances are dressed with quantum corrections, and also
non-factorisable corrections associated with the interference of radiation in
production and decays are taken into account.
Bottom-quark masses are consistently included, which is quite important for
the accurate modelling of $b$-quark fragmentation. Moreover, finite
$b$-quark masses permit to avoid collinear singularities from initial- or
final-state $g\to \ensuremath{b \,\bar b}\xspace$ splittings. This allows for $\ensuremath{W^+W^- b \bar b}\xspace$ simulations in
the full phase space, including regions with unresolved $b$ quarks, which are
indispensable for the simulation of top backgrounds in the presence of jet
vetoes. It moreover provides a unified NLO description of $\ensuremath{t \bar t}\xspace$ and
single-top $\ensuremath{W t}\xspace$ production, including their quantum interference.
The technical problems that arise from infrared subtractions and
NLO+PS matching in the presence of top-quark resonances
are addressed by means of the
fully general resonance-aware matching method
that was proposed in~\citere{Jezo:2015aia} and implemented in the \RES{}
framework. This framework, besides allowing for a consistent matching to
shower Monte Carlo generators, also ameliorates the efficiency of infrared
subtraction and phase-space integration in a drastic way, and allows for a
factorised treatment of NLO radiation in off-shell top production and decays.
This represents a significant improvement (especially for what concerns top
decays) with respect to the case where NLO+PS matching is applied to a
single QCD emission.
Technically, the $\bbfourl$ generator was realised by implementing
{\tt OpenLoops}\xspace{} matrix elements in the \POWHEGBOX{} framework. To this
end we have developed a new and fully flexible interface, which allows
one to set up \POWHEGBOX{}+{\tt OpenLoops}\xspace{} NLO+PS generators for any
desired process in a rather straightforward way.
We have carried out a thorough study of the impact of the resonance-aware
method.
To this end, we have compared our results with those
obtained after disabling the
resonance-aware formalism
in such a way that
the \bbfourl{} generator becomes fully equivalent to a traditional \VTWO{}
implementation.
On the one hand we observed that ignoring resonance structures
can deteriorate the performance of the generator
up to the point of rendering it unusable. On the other hand,
we observed considerable distortions in the
reconstructed mass of the top resonances with respect to the full resonance-aware
result. In essence, the mass distribution becomes wider around the peak and
slightly shifted. We were able to track the origin of these effects to two
competing causes: the generation of radiation performed by \POWHEG{}, that
is considerably modified in the resonance-aware method, and the generation of
radiation in the shower stage, where the shower Monte Carlo, being unaware of
which groups of particles arise from the same resonance, tends to widen the
resonance peaks. We have also shown that it does not seem to be possible to remedy
this last problem by reconstructing the resonance structure on the basis
of simple kinematic guesses.
Much attention was dedicated to the comparison of the new $\bbfourl$
generator and the \dec{} generator of~\citere{Campbell:2014kua}. Both
are capable of handling NLO spin correlations and radiation in top
decays. However, off-shell effects are only computed at LO in \dec{} by
reweighting the NLO cross section using the ratio of the full off-shell Born
cross section divided by its zero-width approximation.
These
two generators are expected to provide similar results in the vicinity of
top resonances.
In fact, in this
region, we find only modest
differences between the two.
In particular, the
top virtuality distribution and distributions involving $b$ jets are in
reasonably good agreement. Slightly larger differences are found in
distributions involving $B$ hadrons, like for example, the $B$ fragmentation
function, in the top-decay frame.
A section of this work was dedicated to a comparison
against the \hvq{} generator, which has been heavily used by the LHC experimental
collaborations for the generation of $t\bar{t}$ samples in both Run~I and
Run~II.
Close to the mass peak, \bbfourl{} and \hvq{} predictions are fairly consistent,
but the agreement is quickly spoiled
as one moves towards off-shell regions.
Furthermore, the ratio of the \hvq{} to the
\bbfourl{} results exhibits a negative slope across the resonance peak, and
we found that the average virtuality of the top resonance in a window of
$\pm 30$~GeV around the peak differs by about 0.5~GeV for the two generators.
This calls for dedicated studies of the implications of
resonance-aware matching in the context of precision $m_t$-measurements.
More sizable differences have been
observed in the structure of the associated $b$-jets, the \bbfourl{}
generator leading consistently to narrower jets and a harder fragmentation
function for the associated $B$ hadron.
The above findings should be interpreted
by keeping in mind that within the \hvq{}
generator radiation in top decays is solely handled by \PythiaEight{}, with
matrix-element corrections turned on by default.
These matrix-element corrections should improve the
overall agreement between \hvq{} and \bbfourl{}, and we have verified that
disabling them leads to
much more pronounced differences between the two generators.
We have included in this work an indicative comparative study of jet-veto
effects when using the \bbfourl{}, \dec{} and \hvq{} generators.
In the presence of jet vetoes,
the \hvq{} generator alone is clearly not adequate,
since it misses the essential component of associated \ensuremath{W t}\xspace
production. Perhaps surprisingly, it turns out
that also the \dec{} generator does
not perform sufficiently well.
Since $\ensuremath{W t}\xspace$ production effects are
included in this generator only at the level of a leading-order reweighting,
we are led to conclude that the lack of
NLO accuracy in the simulation of
$\ensuremath{W t}\xspace$ contributions limits the usability of the \dec{} generator
in single-top enriched regions.
We stress however that the issue of jet-veto effects is
complex, and deserves a dedicated future study.
The theoretical improvements implemented in the $\bbfourl$ generator
are relevant for phenomenological studies and experimental analysis that depends on the
kinematic details of top-decay products. In particular, this new
generator is ideally suited for precision determinations of the
top-quark mass, for measurements of $\ensuremath{W t}\xspace$ production, and for analyses
where $\ensuremath{t \bar t}\xspace$ and $\ensuremath{W t}\xspace$ production are subject to jet vetoes. The
exact treatment of off-shell and non-resonant effects is also
important for top backgrounds in Higgs and BSM studies based on
kinematic selections with high missing energy or boosted $b\bar{b}$
pairs.
\acknowledgments We thank P.~Maierh\"ofer for valuable help and discussion
and important improvements in {\tt OpenLoops}\xspace. We thank P.~Skands for useful
exchanges about the \PythiaEight{} interface. We also wish to thank
A.~Denner, S.~Dittmaier and L.~Hofer for providing us with pre-release
versions of the one-loop tensor-integral library \Collier. This research was
supported in part by the Swiss National Science Foundation~(SNF) under
contracts BSCGI0-157722 and PP00P2-153027, by the Research Executive Agency
of the European Union under the Grant Agreement PITN--GA--2012--316704~({\it
HiggsTools}), and by the Kavli Institute for Theoretical Physics through
the National Science Foundation's Grant No. NSF PHY11-25915. PN and CO would
like to express a special thanks to the Mainz Institute for Theoretical
Physics~(MITP) for its hospitality and support while part of this work was
carried out.
| 42dfcebe1e914267a5d99f041bf5999d22b20079 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Background}
\label{background}
\subsection{Notation}
Let $ I_t : \mathbb R^2\mapsto \mathbb R $ refer to an image captured at time $t$ treated as a function of real values using sub pixel interpolation \cite{Dame10_mi_thesis} for non integral locations.
The patch corresponding to the tracked object's location in $ I_t $ is denoted by $\mathbf{I_t}(\mathbf{x_t}) \in \mathbb R^N $ where $\mathbf{x_t}=[\mathbf{x_{1t}},..., \mathbf{x_{Nt}}]$ with $\mathbf{x_{kt}}=[x_{kt}, y_{kt}]^T \in \mathbb R^2$ being the Cartesian coordinates of pixel $ k $.
Further, $\mathbf{w}(\mathbf{x}, \mathbf{p_s}) : \mathbb{R}^2 \times \mathbb{R}^S\mapsto \mathbb{R}^2$ denotes a warping function of $ S $ parameters
that represents the set of allowable image motions of the tracked object by specifying the deformations that can be applied to $\mathbf{x_0}$ to align
$\mathbf{I_t}(\mathbf{w}(\mathbf{x_0},\mathbf{p_{st}}))$
with
$ \mathbf{I_0}(\mathbf{x_0}) $.
Finally $f(\mathbf{I^*}, \mathbf{I^c},\mathbf{p_a} ) : \mathbb{R}^N \times \mathbb{R}^N \times \mathbb{R}^A\mapsto \mathbb{R}$ is a function of $A$ parameters that measures the similarity between two
patches - the reference or template patch $\mathbf{I^*}$
and a candidate patch $ \mathbf{I^c} $.
For brevity, $ \mathbf{I_t}(\mathbf{w}(\mathbf{x_0}, \mathbf{p_{st}}))$ and $\mathbf{I_0}(\mathbf{x_0})$ will also be denoted as $ \mathbf{I_t} $ and $ \mathbf{I_0} $ respectively.
\subsection{Decomposing registration based tracking}
\label{decompose_registration_tracking}
Using this notation, RBT can be formulated as a search problem whose aim is to find the optimal parameters $\mathbf{p_t}=[\mathbf{p_{st}}, \mathbf{p_{at}}]\in\mathbb{R}^{S+A}$
that maximize the similarity, given by $f$, between $\mathbf{I^*} = \mathbf{I_0}$
and $\mathbf{I^c}=\mathbf{I_t}$, that is,
\begin{equation}
\begin{aligned}
\label{reg_tracking}
\mathbf{p_t} = \underset{\mathbf{p_s},\mathbf{p_a} } {\mathrm{argmax}} ~f(\mathbf{I_0}(\mathbf{x_0}),\mathbf{I_t}(\mathbf{w}(\mathbf{x_0},\mathbf{p_s})), \mathbf{p_a})
\end{aligned}
\end{equation}
This formulation gives rise to an intuitive way to decompose the tracking task into three sub modules - the similarity metric $ f $, the warping function $ \mathbf{w} $ and the optimization approach. These can be designed to be semi independent in the sense that any given optimizer can be applied unchanged to several combinations of methods for the other two modules which in turn interact only through a well defined and consistent interface.
In this work, these sub modules are respectively referred to as AM, SSM and SM. A more detailed description with examples follows.
\subsubsection{Appearance Model}
\label{appearanceModel}
This is the image similarity metric defined by $f$ in Eq. \ref{reg_tracking} that the SM uses to compare different warped patches from $ I_t $ to find the closest match with $\mathbf{I^*}$.
It may be noted that $\mathbf{I^*}$ is not constrained to be fixed but may be updated during tracking as in \cite{Ross08ivt}.
Examples of $f$ with $A=0$ include sum of squared differences (\textbf{SSD}) \cite{Lucas81lucasKanade,Baker04lucasKanade_paper,Benhimane07_esm_journal}, sum of conditional variance (\textbf{SCV}) \cite{Richa11_scv_original}, reversed SCV (\textbf{RSCV}) \cite{Dick13nn}, normalized cross correlation (\textbf{NCC}) \cite{Scandaroli2012_ncc_tracking}, mutual information (\textbf{MI}) \cite{Dowson08_mi_ict,Dame10_mi_ict} and cross cumulative residual entropy (\textbf{CCRE}) \cite{Wang2007_ccre_registration,Richa12_robust_similarity_measures}.
There has also been a recent extension to SCV called \textbf{LSCV} \cite{Richa14_scv_constraints} that claims to handle localized intensity changes better.
Finally, it has been shown \cite{Ruthotto2010_thes_ncc_equivalence} that SSD performs better when applied to z-score \cite{Jain2005_ncc_z_score} normalized images which makes it equivalent to NCC. The resultant formulation is considered here as a distinct AM called Zero Mean NCC (\textbf{ZNCC}).
These AMs can be divided into 2 categories - those that use some form of the L2 norm as $ f $ - SSD, SCV, RSCV, LSCV and ZNCC - and those that do not - MI, CCRE and NCC. The latter are henceforth called robust AMs after \cite{Richa12_robust_similarity_measures}. The two AMs introduced here - SSIM and SPSS - too fall into this category.
To the best of our knowledge, the only AMs with $A\neq 0$ introduced thus far in literature are those with an illumination model (\textbf{ILM}), where $f$ is expressed as $f(\mathbf{I^*}, g(\mathbf{I^c},\mathbf{p_a}))$ with $g : \mathbb{R}^N \times \mathbb{R}^A \mapsto \mathbb{R}^N$ accounting for differences in lighting conditions under which $I_0$ and $I_t$ were captured. These include gain \& bias (\textbf{GB}) \cite{Bartoli2006,Bartoli2008} and piecewise gain \& bias (\textbf{PGB}) \cite{Silveira2007_esm_lighting,Silveira2009_ilm_rgb_vs,Silveira2010_esm_ilm_rgb,Silveira14_esm_ilm_omni} with the latter comprising surface modeling with radial basis function (\textbf{RBF}) too. Though there have also been AMs that incorporate online learning \cite{Jepson2003_online_am,Ross08ivt,Firouzi2014227_online_am}, they are not compatible with all SMs used here and so have not been tested.
\subsubsection{State Space Model}
\label{stateSpace}
This is the warping function $ \mathbf{w} $ that represents the deformations that the tracked patch can undergo. It therefore defines the set of allowable image motions of the corresponding object and can be used to place constraints on the search space of $ \mathbf{p_s} $ to make the optimization more robust or efficient. Besides the DOF of allowed motion, SSM also includes the actual parameterization of $ \mathbf{w} $. For instance, even though both represent 8 DOF motion, $\mathbb{SL}(3)$ \cite{Benhimane07_esm_journal, Kwon2014_sl3_aff_pf} is considered as a different SSM than standard homography \cite{Lucas81lucasKanade,Baker04lucasKanade_paper} that uses actual entries of the corresponding matrix.
This work uses 7 SSMs including 5 from the standard hierarchy of geometrical transformations \cite{Hartley04MVGCV,Szeliski2006_fclk_extended} - translation, isometry, similitude, affine and homography - along with two alternative parameterizations of homography - $\mathbb{SL}(3)$ and corner based (i.e. using x,y coordinates of the bounding box corners). More complex SSMs to handle non rigid objects have also been proposed in literature like thin plate splines \cite{Bookstein1989_tps}, basis splines \cite{Szeliski1997_spline} and quadric surfaces \cite{Shashua2001_q_warp}. However, these are not tested here as the datasets used only feature rigid objects so these cannot be fairly evaluated.
Extensions like incorporation of 3D pose \cite{Cobzas2009_ic_3d_tracking} and camera parameters \cite{Trummer2008_gklt} to $ \mathbf{w} $ are also excluded.
\subsubsection{Search Method}
\label{searchMethod}
This is the optimization procedure that searches for the warped patch in $ I_t $ that best matches $\mathbf{I^*}$. There have been two main categories of SMs in literature - gradient descent (GD) and stochastic search. The former category includes the four variants of the classic Lucas Kanade (LK) tracker \cite{Lucas81lucasKanade} - forward additive (\textbf{FALK}) \cite{Lucas81lucasKanade}, inverse additive (\textbf{IALK}) \cite{Hager98parametricModels}, forward compositional (\textbf{FCLK}) \cite{Szeliski2006_fclk_extended} and inverse compositional (\textbf{ICLK}) \cite{Baker01ict} - that have been shown \cite{Baker01ict,Baker04lucasKanade_paper} to be equivalent to first order terms. Here, however, we show experimental results (Sec. \ref{res_sm}) proving that they perform differently in practice.
A more recent approach of this type is the Efficient Second order Minimization (\textbf{ESM}) \cite{Benhimane07_esm_journal} technique that uses gradients from both $ I_0 $ and $ I_t $ to make the best of ICLK and FCLK.
Several extensions have also been proposed to these SMs to handle specific challenges like motion blur \cite{Park2009_esm_blur}, resolution degradation \cite{Ito2011_res_degrade}, better Taylor series approximation \cite{Keller2008} and optimal subset selection \cite{Benhimane2007}. Though these can be fit within our framework, either as distinct SMs or as variants thereof, they are not considered here for lack of space.
There are three main stochastic SMs in literature - approximate nearest neighbor (\textbf{NN}) \cite{Gu2010_ann_classifier_tracker,Dick13nn}, particle filters (\textbf{PF}) \cite{Isard98condensation,Kwon2009_affine_ois,Li2012_pf_affine_self_tuning_journal,Firouzi2014227_online_am,Kwon2014_sl3_aff_pf} and random sample consensus (\textbf{RANSAC}) \cite{Brown2007_ransac_stitching,Zhang2015_rklt}. Though less prone to getting stuck in local maxima than GD, their performance depends largely on the number and quality of random samples used and tends to be rather jittery and unstable due to the limited coverage of the search space. One approach to obtain better precision is to combine them with GD methods \cite{Dick13nn,Zhang2015_rklt} where results from the former are used as starting points for the latter. Three such combinations have been tested here as examples of hybrid or composite SMs - NN+ICLK (\textbf{NNIC}), PF+FCLK (\textbf{PFFC}) and RANSAC+FCLK (\textbf{RKLT}).
\section{Conclusions}
\label{conclusions}
We presented two new similarity measures for registration based tracking and also formulated a novel way to decompose trackers in this domain into three sub modules. We tested many different combinations of methods for each sub module to gain interesting insights into their strengths and weaknesses. Several surprising results were obtained that proved previously published theoretical analysis to be somewhat inaccurate in practice, thus demonstrating the usefulness of this approach.
Finally, the open source tracking framework used for generating these results was made publicly available so these can be easily reproduced.
\section{Introduction}
\label{introduction}
Visual tracking is an important field in computer vision with diverse application domains including robotics, surveillance, targeting systems, autonomous navigation and augmented reality. Registration based tracking (\textbf{RBT}),
also known in literature as direct visual tracking \cite{Silveira10_esm_ext2, Scandaroli2012_ncc_tracking, Richa12_robust_similarity_measures, Richa14_scv_constraints},
is a
sub field thereof where the object is tracked by warping each image in a sequence to align the object patch with the template. Trackers of this type are especially popular in robotics and augmented reality applications both because they estimate the object pose with greater precision and are significantly faster than online learning and detection based trackers (\textbf{OLT}s) \cite{singh16_modular_results}.
However, OLTs are better suited to general purpose tracking as RBTs are prone to failure when the object undergoes significant appearance changes due to factors like occlusions and lighting variations or when completely novel views of the object are presented by deformations or large pose changes.
As a result, OLTs are more popular in literature and have been the subject of several recent studies \cite{Liu14_olt_survey, Smeulders2014_tracking_survey, Kristan2016_vot16}. The scope of such studies is usually limited to finding the trackers that work best under
challenging conditions by testing them on representative sequences with little to no analysis provided regarding \textit{why} specific trackers work better than others for given challenges. This is understandable since such trackers differ widely in design and have little in common that may be used to relate them to each other and perform comparative analysis from a design perspective.
As we show in this work, however, RBTs do not have this drawback and can be decomposed into three sub modules - appearance model (\textbf{AM}), state space model (\textbf{SSM}) and search method (\textbf{SM}) (Sec. \ref{decompose_registration_tracking}) - that makes their systematic analysis feasible. Though this decomposition is somewhat obvious and indeed has been observed before \cite{Szeliski2006_fclk_extended, Richa12_robust_similarity_measures}, it has never been explored systematically or used to improve the study of this paradigm of tracking. It is the intent of this work to fill this gap by unifying the myriad of contributions made in this domain since the original Lucas Kanade (LK) tracker was introduced \cite{Lucas81lucasKanade}.
Most of these works have presented novel ideas for only one of these submodules while using existing methods, often selected arbitrarily, for the rest.
For instance, Hager \& Belheumer \cite{Hager98parametricModels}, Shum \& Szeliski\cite{Shum00_fc}, Baker \& Matthews\cite{Baker01ict} and Benhimane \& Malis\cite{Benhimane04esmTracking} all introduced new variants of the Newton type SM used in \cite{Lucas81lucasKanade} but only tested these with SSD\footnote{refer Sec. \ref{decompose_registration_tracking} for acronyms} AM.
Similarly, Dick et. al \cite{Dick13nn} and Zhang et. al \cite{Zhang2015_rklt} only combined their stochastic SMs with RSCV and SSD respectively.
Conversely, Richa et. al \cite{Richa11_scv_original}, Scandaroli et. al \cite{Scandaroli2012_ncc_tracking} and Dame et. al \cite{Dame10_mi_ict} introduced SCV, MI and NCC as AMs but combined these only with a single SM - ESM in the former and ICLK in the other two.
Even more recent works that use illumination models (\textbf{ILM}) (Sec. \ref{appearanceModel}), including Bartoli \cite{Bartoli2006,Bartoli2008}, Silvera \& Malis \cite{Silveira2007_esm_lighting,Silveira2009_ilm_rgb_vs,Silveira2010_esm_ilm_rgb} and Silvera \cite{Silveira14_esm_ilm_omni}, have combined their respective ILMs with only a single SM in each case.
Finally, most SMs and AMs have been tested with only one SSM - either homography \cite{Benhimane04esmTracking,Benhimane07_esm_journal,Dame10_mi_ict} or affine \cite{Baker01ict,Ross08ivt}.
In fact, Benhimane \& Malis \cite{Benhimane04esmTracking} mentioned that their SM only works with $\mathbb{SL}(3)$ SSM though experiments (Sec. \ref{results}) showed that it works equally well with others.
Such limited testing might not only give false indications about a method's capability but might also prevent it from achieving its full potential. For instance, an AM that outperforms others with a given SM might not do so with other SMs and an SM may perform better with an AM other than the one it is tested with. In such cases, our decomposition can be used to experimentally find the optimal combination of methods for any contribution while also providing novel insights about its workings. We demonstrate its practical applicability by comparing several existing methods for these sub modules not only with each other but also with two new AMs based on structural similarity (\textbf{SSIM}) \cite{Wang04_ssim_original} that we introduce here and fit within this framework.
SSIM is a popular measure for evaluating the quality of image compression algorithms by comparing the compressed image with the original one.
Since it measures the information loss in the former - essentially a slightly distorted or damaged version of the latter - it makes a suitable metric for comparing candidate warped patches with the template to find the one with the minimum loss and thus most likely to represent the same object. Further, it has been designed to capture the perceptual similarity of images and is known to be robust to illumination and contrast changes \cite{Wang04_ssim_original}. It is reasonable, therefore, to expect it to perform well for tracking under challenging conditions too. As such, it has indeed been used for tracking before with particle filters \cite{Loza2009_ssim_pf,Wang2013_ssim_bc}, gradient ascent \cite{Loza11_ssim_lk_tracking} and hybrid \cite{Loza2009a_ssim_hybrid} SMs. All of these trackers, however, used imprecise SSMs with low degrees of freedom (DOF) - estimating only translation and scaling of the target patch. To the best of our knowledge, no attempt has been made to use SSIM for high DOF RBT within the LK framework \cite{Lucas81lucasKanade}. This work aims to fill this gap too.
To summarize, following are the main contributions of this work:
\begin{itemize}
\item Adapt a popular image quality measure - SSIM - for high precision RBT and introduce a simpler but faster version called SPSS (Sec. \ref{sec_ssim}).
\item Evaluate these models comprehensively by comparing against 8 existing AMs using 11 SMs and 7 SSMs. Experiments are done using 4 large datasets with over 100,000 frames in all to ensure their statistical significance.
\item Compare low DOF RBTs against state of the art OLTs to validate the suitability of the former for fast and high precision tracking applications.
\item Provide an open source tracking framework\footnote{available at \url{http://webdocs.cs.ualberta.ca/~vis/mtf/}}
called MTF \cite{singh16_mtf} using which all results can be reproduced and which, owing to its efficient C++ implementation, can also serve as a practical tracking solution.
\end{itemize}
Rest of this paper is organized as follows - section \ref{background} describes the decomposition, section \ref{sec_ssim} presents details about SSIM and section \ref{experiments} provides results and analysis.
\section{Results and Analysis}
\label{experiments}
\subsection{Datasets}
Following four publicly available datasets
have been used to analyze the trackers:
\begin{enumerate}
\item Tracking for Manipulation Tasks (\textbf{TMT}) dataset \cite{Roy2015_tmt} that contains videos of some common tasks performed at several speeds and under varying lighting conditions. It has 109 sequences with 70592 frames.
\item Visual Tracking Dataset provided by \textbf{UCSB} \cite{Gauglitz2011_ucsb} that has 96 short sequences with a total of 6889 frames. These are more challenging than TMT but also somewhat artificial as they were created to represent specific challenges rather than realistic tasks.
\item \textbf{LinTrack} dataset \cite{Zimmermann2009_lintrack} that has 3 long sequences with a total of 12477 frames. These are more realistic than those in UCSB but also more difficult to track.
\item A collection of 28 challenging planar tracking sequences from several significant works in literature \cite{Kwon2014_sl3_aff_pf,Silveira2007_esm_lighting,Silveira2009_ilm_rgb_vs,Silveira2010_esm_ilm_rgb,Silveira14_esm_ilm_omni,Crivellaro2014_dft,Nebehay2015_cmt}. We call this the \textbf{PAMI} dataset after \cite{Kwon2014_sl3_aff_pf} from where several of the sequences originate. There are 16511 frames in all.
\end{enumerate}
All of these datasets except PAMI have full pose (8 DOF) ground truth data which makes them suitable for evaluating high precision trackers that are the subject of this study. For PAMI, this data was generated using a combination of very high precision tracking and manual refinement \cite{Roy2015_tmt}.
\subsection{Evaluation Measures}
\textbf{Alignment Error} ($E_{AL}$) \cite{Dick13nn} has been used as the metric to compare tracking result with the ground truth since it accounts for fine misalignments of pose better than other measures like center location error and Jaccard index.
A tracker's overall accuracy is measured through its \textbf{success rate} (SR) which is defined as the fraction of total frames where $E_{AL}$ is less than a threshold of $t_p$ pixels. Formally,
$ SR = |S|/|F| $
where $ S = \{f^{i} \in F : E_{AL}^{i} < t_p \}$, $F$ is the set of all frames and $E_{AL}^{i}$ is the error in the $i^{th}$ frame $f^{i}$.
Since there are far too many sequences to present results for each, an overall summary of performance is reported instead by averaging the SR over all sequences in the four datasets. In addition, to better utilize frames that follow a tracker's first failure in any sequence, we initialize trackers at $ 10 $ different evenly spaced frames in each sequence. Therefore the SR plots represent accumulated tracking performance over a total of $ |F| = 589380 $ frames, out of which $ 106469 $ are unique.
Finally, the SR is evaluated for several values of $ t_p $ ranging from 0 to 20 and the resulting SR vs. $ t_p $ plot is studied to get an overall idea of how precise and robust a tracker is.
As an alternative measure for robustness, reinitialization tests \cite{Kristan2016_vot16} were also conducted where a tracker is reinitialized after skipping 5 frames every time its $ E_{AL} $ exceeds 20 and the number of such reinitialization is counted.
Due to space constraints, these results are presented in the supplementary (available on MTF website).
\subsection{Parameters Used}
All results were generated using a fixed sampling resolution of $ 50{\times}50 $ irrespective the tracked object's size. Input images were smoothed using a Gaussian filter with a $ 5{\times}5 $ kernel before being fed to the trackers. Iterative SMs were allowed to perform a maximum of $ 30 $ iterations per frame but only as long as the L2 norm of the change in bounding box corners in each iteration exceeded $ 0.0001 $. Implementation details of NN, PF and RANSAC were taken from \cite{Dick13nn}, \cite{Kwon2014_sl3_aff_pf} and \cite{Zhang2015_rklt} respectively.
NN was run with 1000 samples and PF with 500 particles.
For RANSAC and RKLT, a $ 10{\times} 10 $ grid of sub patches was used and each sub patch was tracked by a 2 DOF FCLK tracker with a sampling resolution of $ 25{\times} 25 $. As in \cite{Wang04_ssim_original}, SSIM parameters are computed as $ C_1=(K_1L)^2 $ and $ C_2=(K_2L)^2 $ with $ K_1 =0.01$, $ K_2 =0.03$ and $ L=255$. Learning based trackers (Sec. \ref{res_ssm}) were run using default settings provided by their respective authors. Speed tests were performed on a 4 GHz Intel Core i7-4790K machine with 16 GB of RAM.
\begin{figure*}[!htbp]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{fc_ic_esm.eps}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{fa_ia_pf.eps}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{pffc_nn_nnic.eps}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=\textwidth]{ransac_rklt_ssm.eps}
\end{subfigure}
\caption{Success rates for AMs using different SMs with Homography as well as for different SSMs using SSIM with ESM (bottom right). Plot legends indicate areas under the respective curves. Best viewed on a high resolution screen.}
\label{fig_am}
\end{figure*}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{speed_esm_pf.eps}
\caption{Speeds of ESM and PF with homography. Solid and dotted lines show the means and standard deviations respectively.}
\label{fig_speed_esm_pf}
\end{figure}
\subsection{Results}
\label{results}
The results presented in this section are organized into three sections corresponding to the three sub modules. In each, we present and analyze results comparing different methods for the respective sub module.
\subsubsection{Appearance Models}
\label{res_am}
Fig. \ref{fig_am} shows the SR curves for all AMs using different SMs and homography SSM with plot legends also indicating the areas under the respective curves which are equivalent \cite{vCehovin2014_eval} to the average SR over this range of $ t_p $. Fig. \ref{fig_speed_esm_pf} shows the speeds of these AMs with ESM and PF as representatives of GD and stochastic SMs respectively.
LSCV is excluded to reduce clutter as it was found to perform very similarly to SCV. Its results are in the supplementary instead along with those for the ILMs. Further, RANSAC and RKLT do not include CCRE and MI results since both are far too slow for 100 of them to be run simultaneously in real time. Also, both perform \textit{very} poorly with these SMs, probably because the small sub patches used there \cite{Zhang2015_rklt} do not have sufficient information to be represented well by the joint histograms employed by these AMs.
Several observations can be made here. Firstly, NCC is the best performer with all SMs except IALK and NN - IALK performs poorly with all robust AMs (Sec. \ref{res_sm}) while NN performs best with MI. Nevertheless, SSIM is usually equivalent to NCC or a close second showing that the latter is among the best AMs known. Also, as expected, SSIM is much better than SPSS with all SMs, with the latter only managing to match the performance of SSD on average.
Further, though ZNCC is claimed to be equivalent to NCC \cite{Ruthotto2010_thes_ncc_equivalence} and also has a wider basin of convergence due to its SSD like formulation, it usually does \textit{not} perform as well as NCC.
Secondly, in spite of being the most sophisticated and computationally expensive AM, CCRE is the worst performer with GD SMs and even MI is only slightly better than SSD on average. However, MI and CCRE are actually the best performing AMs with NN and MI is so with NNIC too. This shows that their poor performance with GD SMs is likely to be due to their narrower basin of convergence rather than an inherent weakness in the AMs themselves. This also explains MI's significant lead over CCRE with these SMs though the two differ only in the latter using a cumulative joint histogram. It seems likely that the additional complexity of CCRE along with the resultant invariance to appearance changes further \textit{reduces} its basin of convergence \cite{Dame10_mi_ict}. This discrepancy in performance between GD and stochastic SMs demonstrates the inadequacy of evaluating an AM with only one SM.
Thirdly, SCV outperforms RSCV with both inverse GD SMs - ICLK and IALK - though the reverse is true with both the forward ones - FCLK and FALK. This pattern also holds for stochastic and composite SMs - SCV is better with NN/NNIC where samples are generated from $ I_0 $ but worse with PF/PFFC where these come from $ I_t $. Also, their performance is very similar with ESM where information from both $ I_0 $ and $ I_t $ is used. This is probably because SCV performs likelihood substitution with $ \mathbf{I_0} $ \cite{Richa11_scv_original} while RSCV does so with $ \mathbf{I_t} $ \cite{Dick13nn}. Fourthly, the separation between AMs is narrower with RANSAC than other SMs as this SM depends more on the \textit{number} of sub patches used than on the tracker used for each. Conversely, PF shows maximum variation between AMs, thus indicating its strong reliance on $ f $.
Finally, SPSS is not much faster than SSIM with either SM though it has lower computational complexity. This is partly due to SSIM finding convergence in fewer iterations with GD SMs and partly due to the way Eigen \cite{eigenweb} optimizes matrix multiplications, many of which are used for computing $ f_{ssim} $ and its derivatives while those of $ f_{spss} $ have to be computed pixel by pixel. The same holds for SSD too.
\subsubsection{State Space Models}
\label{res_ssm}
\begin{figure*}[!htbp]
\begin{center}
\includegraphics[width=\textwidth]{lt_with_ssim_2r.eps}
\caption{Success Rates for SSIM using translation as well as for 8 OLTs. The former are shown with \textbf{solid} lines and the latter in \textbf{dashed} lines. The speed plot on the right has \textbf{logarithmic scaling} on the x axis for clarity though actual figures are also shown.}
\label{fig_lt_with_ssim_2r}
\end{center}
\end{figure*}
Results in this section follow a slightly different format from the other two due to the difference in the motivations for using low DOF SSMs - they are more robust due to reduced search space and faster due to the gradients of $ \mathbf{w} $ being less expensive to compute.
Limiting the DOF also makes them directly comparable to OLTs which too have low DOF. As a result, 2 DOF RBTs were also tested with 8 state of the art OLTs \cite{Kristan2016_vot16} - DSST \cite{Danelljan2014_dsst}, KCF \cite{henriques2015high_kcf}, CMT \cite{Nebehay2015_cmt}, RCT \cite{Zhang2012_rct}, TLD \cite{Kalal12tld}, Struck \cite{hare2011struck}, FRG \cite{Adam2006_fragtrack} and GOTURN \cite{Held2016_goturn}.
Lastly, in order to make the evaluations fair, \textit{lower DOF ground truth} has been used for all accuracy results in this section. This was generated for each SSM using least squares optimization to find the warp parameters that, when applied to the initial bounding box, will produces a warped box whose $E_{AL}$ with respect to the 8 DOF ground truth is as small as it is possible to achieve given the constraints of that SSM.
Fig. \ref{fig_lt_with_ssim_2r} shows the results of these tests in terms of both accuracy and speed.
As expected, all the OLTs have low SR for smaller $ t_p $ since they are less precise in general \cite{Kristan2016_vot16}. What is more interesting, however, is that none of them, with the exception of DSST and perhaps Struck, managed to surpass the best RBTs even for larger $ t_p $.
The superiority of DSST and Struck over other OLTs is consistent with results on VOT dataset \cite{Kristan2016_vot16}. However, the very poor performance of GOTURN \cite{Held2016_goturn}, which is one of the best trackers on that dataset, indicates a fundamental difference in the challenges involved in the two paradigms of tracking.
The speed plot shows another reason why OLTs are not suitable for high speed tracking applications - they are $ 10 $ to $ 30 $ times slower than RBTs except PF and RANSAC that are not implemented efficiently yet.
It is not surprising that tracking based SLAM systems like SVO \cite{Forster2014_svo} use registration based trackers as they need to track hundreds of patches per frame.
To conclude the analysis in this section, we tested the performance of different SSMs against each other and the results are reported in the bottom right subplot of Fig. \ref{fig_am} using ESM with SSIM. Contrary to expectations, lower DOF SSMs are not better except perhaps affine and similitude that have slightly higher SR than homography for larger $ t_p $.
In fact, the SR curves of all low DOF SSMs approach those of homography as $ t_p $ increases which does indicate their higher robustness - though not as precise as homography, they tend to be more resistant to complete failure.
Also, all three parameterizations of homography have almost identical performance, with their plots showing near perfect overlap. This suggests that the theoretical justification \cite{Benhimane07_esm_journal} for using ESM only with $\mathbb{SL}(3)$ has little practical significance.
\subsubsection{Search Methods}
\label{res_sm}
SR plots comparing different SMs for each AM are presented in the supplementary to save space as they contain the same data as Fig. \ref{fig_am}.
First fact to observe is that the four variants of LK do not perform identically. Though FCLK and FALK are indeed evenly matched, both ICLK and IALK are significantly worse, with ICLK being notably better than IALK. This is especially true for CCRE where it outperforms both additive SMs. This contradicts the equivalence between these variants that was reported in \cite{Baker04lucasKanade_paper} and justified there using both theoretical analysis and experimental results. The latter, however, were only performed on synthetic images and even the former used several approximations. Therefore, it is perhaps not surprising that this equivalence does not hold under real world conditions.
Secondly, ESM fails to outperform FCLK for any AM except SCV and ZNCC and even there it does not lead by much. This fact too emerges in contradiction to the theoretical analysis in \cite{Benhimane07_esm_journal} where ESM was shown to have second order convergence and so should be better than first order methods like FCLK.
Thirdly, both additive LK variants, and especially IALK, fare worse against compositional ones when using robust AMs compared to SSD-like AMs. This is probably because the Hessian after convergence approach used for extending Newton's method to these AMs does not make as much sense for additive formulations \cite{Dame10_mi_thesis}.
Lastly, PFFC is the best performing SM followed by NNIC and RKLT which proves the efficacy of composite SMs.
\section{Structural Similarity}
\parskip 0pt
\label{sec_ssim}
SSIM was originally introduced \cite{Wang04_ssim_original} to assess the loss in image quality incurred by compression methods like JPEG. It has been very popular in this domain since it closely mirrors the approach adopted by the human visual system to subjectively evaluate the quality of an image.
SSIM between $ \mathbf{I_0} $ and $\mathbf{I_t}$ is defined as a product of 3 components:
{\footnotesize
\begin{align}
\label{eq_ssim_complete}
f_{ssim}
=
\left(
\dfrac{2\mu_t\mu_0+C_1}{\mu_t ^2+\mu_0^2+C_1}\right)^\alpha
\left(\dfrac{2\sigma_{t}\sigma_{0}+C_2}{\sigma_t ^2 + \sigma_0^2 + C_2}\right)^\beta
\left(\dfrac{\sigma_{t0}+C_3}{\sigma_{t}\sigma_{0}+C_3}\right)^\gamma
\end{align}
}
where
$\mu_t$ is the mean and $ \sigma_{t}$ is the sample standard deviation of $ \mathbf{I_t} $ while
$ \sigma_{t0} $ is the sample \textit{covariance} between $ \mathbf{I_t} $ and $ \mathbf{I_0} $. The three components of $ f_{ssim} $ from left to right are respectively used for luminance, contrast and structure comparison between the two patches. The positive constants $ \alpha,\beta,\gamma$ are used to assign relative weights to these components while $ C_1 $, $ C_2 $, $ C_3 $ are added to ensure their numerical stability with small denominators. Here, as in most practical implementations \cite{Wang04_ssim_original,Loza2009_ssim_pf,Loza2009a_ssim_hybrid,Loza11_ssim_lk_tracking,Wang2013_ssim_bc}, it is assumed that $ \alpha = \beta = \gamma = 1 $ and $ C_3 = \dfrac{C_2}{2} $ so that Eq. \ref{eq_ssim_complete} simplifies to:
\begin{align}
\label{eq_ssim}
f_{ssim}
=
\dfrac{
\left(2\mu_t\mu_0+C_1\right)
\left(2\sigma_{t0}+C_2\right)
}
{
\left(\mu_t ^2+\mu_0^2+C_1\right)
\left(\sigma_t ^2 + \sigma_0^2 + C_2\right)
}
\end{align}
\subsection{Newton's Method with SSIM}
\parskip 0pt
\label{newton_ssim}
Using SSIM with NN and PF is straightforward since these only need $ f_{ssim} $ to be computed between candidate patches. However, GD SMs also require its derivatives as they solve Eq \ref{reg_tracking} by estimating an incremental update $ \Delta\mathbf{p_t} $ to $ \mathbf{p_{t-1}} $ using some variant of Newton's method as:
\begin{align}
\label{eq_nwt_method}
\Delta\mathbf{p}_t=-\hat{\mathbf{H}}^{-1}\hat{\mathbf{J}}^T
\end{align}
where $ \hat{\mathbf{J}} $ and $ \hat{\mathbf{H}} $ respectively are estimates for Jacobian $ \mathbf{J} = \partial f/\partial \mathbf{p}$ and Hessian $ \mathbf{H} = \partial^2 f/\partial \mathbf{p}^2$ of $ f $ w.r.t. $ \mathbf{p} $.
These can be further decomposed using chain rule as:
\begin{gather}
\label{eq_jac_basic}
\mathbf{J}
=
\dfrac{\partial f}{\partial \mathbf{I}}
\dfrac{\partial I}{\partial \mathbf{p}}
=
\dfrac{\partial f}{\partial \mathbf{I}}
\nabla \mathbf{I}
\dfrac{\partial \mathbf{w}}{\partial \mathbf{p}}\\
\label{eq_hess_basic}
\mathbf{H}
=
\dfrac{\partial \mathbf{I}}{\partial \mathbf{p}}^T
\dfrac{\partial^2 f}{\partial \mathbf{I}^2}
\dfrac{\partial \mathbf{I}}{\partial \mathbf{p}}
+
\dfrac{\partial f}{\partial \mathbf{I}}
\dfrac{\partial^2 \mathbf{I}}{\partial \mathbf{p}^2}
\end{gather}
Of the right hand side terms in Eqs. \ref{eq_jac_basic} and \ref{eq_hess_basic}, only $\partial f/\partial \mathbf{I} $ and $\partial^2 f/\partial \mathbf{I}^2 $ depend on $ f $ so the relevant expressions for $ f_{ssim} $ are presented below - general formulations corresponding to $ \mathbf{J} $ and $ \mathbf{H} $ in Eqs. \ref{eq_ssim_grad_final} and \ref{eq_ssim_hess_final} respectively, followed by specializations for $ \hat{\mathbf{J}} $ and $ \hat{\mathbf{H}} $ used by specific SMs. Detailed derivations are presented in the supplementary.
{\footnotesize
\begin{align}
\label{eq_ssim_grad_final}
\dfrac{\partial f_{ssim}}{\partial \mathbf{I_t}}
=
\mathbf{f'}
=
\dfrac{2}{cd}
\left[
\left(\dfrac{a\mathbf{\bar{I}_0}-cf\mathbf{\bar{I}_t}}{N-1} + \dfrac{\mu_0b - \mu_tfd}{N} \right)
\right]
\end{align}
\begin{align}
\label{eq_ssim_hess_final}
\dfrac{\partial^2 f_{ssim}}{\partial \mathbf{I_t}^2}
&=
\dfrac{2}{cd}
\bigg[
\dfrac{1}{N}
S_N
\left(
\dfrac{4}{N-1}
\left(
\mu_0\mathbf{\bar{I}_0}- \mu_tf\mathbf{\bar{I}_t}
\right)
-
\dfrac{3\mu_td}{2}\mathbf{f'}
-
\dfrac{fd}{N}
\right)\notag\\
&-
\dfrac{c}{N-1}
\left(
\dfrac{3}{2}\mathbf{f'}^T\mathbf{\bar{I}_t}
+
f\mathbb{I}
\right)
\bigg]
\end{align}
}
with $ \mathbf{\bar{I}_t} = \mathbf{I_t}-\mu_t, a=2\mu_t\mu_0+C_1 $, $ b=2\sigma_{t0}+C_2 $, $ c=\mu_t ^2+\mu_0^2+C_1, d=\sigma_t ^2 + \sigma_0^2 + C_2, f=f_{ssim} $ and $ S_n\left(\mathbf{K}\right) $ denoting an $ n{\times} k $ matrix formed by stacking the $ 1\times k $ vector $ \mathbf{K} $ into rows.
The form of $\partial f/\partial \mathbf{I}$ used by the four variants of LK is identical except that ICLK requires the differentiation to be done w.r.t. $ \mathbf{I_0} $ instead of $ \mathbf{I_t} $ - the expressions for this are trivial to derive since SSIM is symmetrical.
ESM was originally \cite{Benhimane07_esm_journal} formulated as using the mean of $\nabla \mathbf{I_0}$ and $\nabla \mathbf{I_t}$ to compute $ \mathbf{J} $ but, as this formulation is only applicable to SSD, a generalized version \cite{Brooks10_esm_ic,Scandaroli2012_ncc_tracking} is considered here that uses the \textit{difference} between FCLK and ICLK Jacobians.
It is generally assumed \cite{Baker04lucasKanade_paper,Benhimane07_esm_journal} that the second term of Eq. \ref{eq_hess_basic} is too costly to compute and too small near convergence to matter and so is omitted to give the Gauss Newton Hessian:
\begin{align}
\label{eq_hess_basic_gn}
\mathbf{H}_{gn}
=
\dfrac{\partial \mathbf{I}}{\partial \mathbf{p}}^T
\dfrac{\partial^2 f}{\partial \mathbf{I}^2}
\dfrac{\partial \mathbf{I}}{\partial \mathbf{p}}
\end{align}
Though $ \mathbf{H}_{gn} $ works very well for SSD (and in fact even better than $ \mathbf{H}$ \cite{Baker04lucasKanade_paper,Dame10_mi_thesis}), it is well known \cite{Dame10_mi_thesis,Scandaroli2012_ncc_tracking} to \textit{not} perform well with other AMs like MI, CCRE and NCC and we can confirm that the latter is true for SSIM and SPSS too. For these AMs, an approximation to $ \mathbf{H} $ \textit{after convergence} has to be used instead by assuming perfect alignment between the patches or $\mathbf{I^c}=\mathbf{I^*} $ . This approximation is here referred to as the \textbf{Self Hessian} and, as this substitution can be made by setting either $ \mathbf{I^c} = \mathbf{I_0}$ or $ \mathbf{I^*} = \mathbf{I_t} $, we get two forms which are respectively deemed to be ICLK and FCLK Hessians:
{\footnotesize
\begin{gather}
\label{eq_hess_iclk}
\hat{\mathbf{H}}_{ic}
=
\mathbf{H}_{self}^\mathbf{*}
=
\dfrac{\partial \mathbf{\mathbf{I_0}}}{\partial \mathbf{p}}^T
\dfrac{\partial^2 f(\mathbf{I_0}, \mathbf{I_0})}{\partial \mathbf{I}^2}
\dfrac{\partial \mathbf{\mathbf{I_0}}}{\partial \mathbf{p}}
+
\dfrac{\partial f(\mathbf{I_0}, \mathbf{I_0})}{\partial \mathbf{I}}
\dfrac{\partial^2 \mathbf{\mathbf{I_0}}}{\partial \mathbf{p}^2}\\
\label{eq_hess_fclk}
\hat{\mathbf{H}}_{fc}
=
\mathbf{H}_{self}^\mathbf{c}
=
\dfrac{\partial \mathbf{\mathbf{I_t}}}{\partial \mathbf{p}}^T
\dfrac{\partial^2 f(\mathbf{I_t}, \mathbf{I_t})}{\partial \mathbf{I}^2}
\dfrac{\partial \mathbf{\mathbf{I_t}}}{\partial \mathbf{p}}
+
\dfrac{\partial f(\mathbf{I_t}, \mathbf{I_t})}{\partial \mathbf{I}}
\dfrac{\partial^2 \mathbf{\mathbf{I_t}}}{\partial \mathbf{p}^2}
\end{gather}
}
It is interesting to note that $ \mathbf{H}_{gn} $ has the exact same form as $ \mathbf{H}_{self} $ for SSD (since $ \partial f_{ssd}(\mathbf{I_0}, \mathbf{I_0})/\partial \mathbf{I} = \partial f_{ssd}(\mathbf{I_t}, \mathbf{I_t})/\partial \mathbf{I} = \mathbf{0}$) so it seems that interpreting Eq. \ref{eq_hess_basic_gn} as the first order approximation of Eq. \ref{eq_hess_basic} for SSD as in \cite{Baker04lucasKanade_paper,Dowson08_mi_ict,Dame10_mi_thesis} is incorrect and it should instead be seen as a special case of $ \mathbf{H}_{self} $.
Setting $ \mathbf{I_0} = \mathbf{I_t} $ simplifies Eq. \ref{eq_ssim_hess_final} to:
\begin{align}
\dfrac{\partial^2 f_{ssim}(\mathbf{I_t}, \mathbf{I_t})}{\partial \mathbf{I_t}^2}
=
\dfrac{-2}{\bar{c}\bar{d}}
\left[
\dfrac{\bar{d}}{N^2}
+
\dfrac{\bar{c}}{N-1}\mathbb{I}
\right]
\end{align}
with $ \bar{c}=2\mu_t^2 + C_1 $ and $ \bar{d}=2\sigma_t^2 + C_2 $. Finally, FALK and IALK use the same form of $ \partial^2 f/\partial \mathbf{I}^2 $ as FCLK while ESM uses the \textit{sum} of $ \hat{\mathbf{H}}_{fc} $ and $\hat{\mathbf{H}}_{ic}$.
\subsection{Simplifying SSIM with pixelwise operations}
\parskip 0pt
In the formulation described so far, SSIM has been computed over the \textit{entire} patch - i.e. $ \mu_t $, $ \sigma_t $ and $ \sigma_{t0} $ have been computed over all $ N $ pixels in $ \mathbf{I_t} $ and $ \mathbf{I_0} $. In its original form \cite{Wang04_ssim_original}, however, the expression in Eq. \ref{eq_ssim} was applied to several corresponding \textit{sub windows} within the two patches - for instance $ 8\times 8 $ sub windows that are moved pixel-by-pixel over the entire patch - and the \textit{mean} of all resultant scores taken as the overall similarity score. For tracking applications, such an approach is not only impracticable from speed perspective, it presents another issue for GD SMs - presence of insufficient texture in these small sub windows may cause Eq. \ref{eq_nwt_method} to become ill posed if $ \mathbf{J} $ and $ \mathbf{H} $ are computed for each sub window and then averaged.
As a result, the formulation used here considers only one end of the spectrum of variation of size and number of sub windows - a single sub window of the same size as the patch. Now, if the \textit{other} end of the spectrum is considered - $ N $ sub windows of size $ 1\times 1 $ each - then a different AM is obtained that may provide some idea about the effect of window wise operations while also being much simpler and faster. The resultant model is termed as \textbf{Sum of Pixelwise Structural Similarity} or \textbf{SPSS}. When considered pixel wise, $ \sigma_t $ and $ \sigma_{t0} $ become null while $ \mu_t$ becomes equal to the pixel value itself so that Eq. \ref{eq_ssim} simplifies to:
\begin{align}
\label{eq_spss}
f_{spss}
=
\sum\limits_{i=1}^{N}
\dfrac{
2\mathbf{I_t}(\mathbf{x}_\mathbf{it})
\mathbf{I_0}(\mathbf{x}_\mathbf{i0}) + C_1
}
{
\mathbf{I_t}(\mathbf{x}_\mathbf{it})^2
+
\mathbf{I_0}(\mathbf{x}_\mathbf{i0})^2
+
C_1
}
\end{align}
Similar to SSD, contributions from different pixels to $ f_{spss} $ are independent of each other so that each entry of $ \partial f_{spss}/\partial \mathbf{I_t}$ has contribution only from the corresponding pixel. This also holds true for each entry of the principal diagonal of $\partial^2 f_{spss}/\partial \mathbf{I_t}^2 $ (which is a diagonal matrix). Denoting the contributions of the $ i^{th} $ pixel to $ f_{spss} $, $ \partial f_{spss}/\partial \mathbf{I_t}$ and $\partial^2 f_{spss}/\partial \mathbf{I_t}^2 $ respectively as $ f_i $, $ f'_i $ and $ f''_i $, we get:
\begin{align}
\label{eq_spss_grad}
f'_i
&=
\dfrac{
2
(\mathbf{I_0}(\mathbf{x}_\mathbf{i0})
-
\mathbf{I_t}(\mathbf{x}_\mathbf{it})
f_{i}
)
}
{
\mathbf{I_t}(\mathbf{x}_\mathbf{it})^2
+
\mathbf{I_0}(\mathbf{x}_\mathbf{i0})^2
+
C_1
}\\
\label{eq_spss_hess}
f''_i
&=
\dfrac{
-2(
f_{i} + 3\mathbf{I_t}(\mathbf{x}_\mathbf{it})f'_{i}
)
}
{
\mathbf{I_t}(\mathbf{x}_\mathbf{it})^2
+
\mathbf{I_0}(\mathbf{x}_\mathbf{i0})^2
+
C_1
}\\
\label{eq_spss_self_hess}
f''_i(\mathbf{I_t}, \mathbf{I_t})
&=
\dfrac{-2}
{
2\mathbf{I_t}(\mathbf{x}_\mathbf{it})^2 + C_1
}
\end{align}
| 9b5e0249e32f073a5952e4c6c8115fbd7396c384 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this paper, we study the boundary value problem of the fourth order curl operator of type
\begin{equation}\label{eq:bvpA}
{\bf(A)}\quad\left\{
\begin{array}{rl}
(\nabla\times)^4\undertilde{u}=\undertilde{f},&\mbox{in}\ \Omega;
\\
\nabla\cdot\undertilde{u}=0 & \mbox{in}\ \Omega;
\\
\undertilde{u}\times\mathbf{n}=\undertilde{0},\ (\nabla\times \undertilde{u})\times \mathbf{n}=\undertilde{0} & \mbox{on}\ \partial\Omega,
\end{array}
\right.
\end{equation}
where ${\rm div}\undertilde{f}=0$, and of variant type
\begin{equation}\label{eq:bvpB}
{\bf(B)}\quad\left\{
\begin{array}{rl}
(\nabla\times)^4\undertilde{u}+\undertilde{u}=\undertilde{f},&\mbox{in}\ \Omega;
\\
\undertilde{u}\times\mathbf{n}=\undertilde{0},\ (\nabla\times \undertilde{u})\times \mathbf{n}=\undertilde{0} & \mbox{on}\ \partial\Omega.
\end{array}
\right.
\end{equation}
For \eqref{eq:bvpB}, it is not necessary that ${\rm div}\undertilde{f}=0$. But evidently, ${\rm div}\undertilde{u}=0$ when ${\rm div}\undertilde{f}=0$.
The boundary value problem of fourth order curl operator $(\nabla\times)^4$ arises in different applications, like in magnetohydrodynamics(MHD) and in the inverse electromagnetic scattering theory. In MHD, $(\nabla\times)^4\bf{B}$ is involved in the resistive system where $\bf{B}$ is the magnetic field as a primary variable\cite{Zheng.B;Hu.Q;Xu.J2011}, and in the inverse electromagnetic scattering theory, $(\nabla\times)^4$ appears in computing the transmission eigenvalue\cite{Cakoni.F;Haddar.H2007}. Some more applications of $(\nabla\times)^4$ can be found in the sequel works. The divergence free condition and the boundary conditions as in \eqref{eq:bvpA} are generally used. The two types of equations as above were discussed in, e.g., \cite{Sun.J2016,Nicaise.S2016} and \cite{Zheng.B;Hu.Q;Xu.J2011,Hong.Q;Hu.J;Shu.S;Xu.J2012}, respectively.
There have been works devoted to the discretization of the model problems. For the primal variational formulation, two kinds of not-conforming discretizations are discussed in literature, including a nonconforming element constructed in Zheng-Hu-Xu \cite{Zheng.B;Hu.Q;Xu.J2011}, and discretizations given in Hong-Hu-Shu-Xu \cite{Hong.Q;Hu.J;Shu.S;Xu.J2012} with standard high order Nedelec elements in the framework of discontinuous Galerkin method. So far, no curl-curl-conforming other than the $H^2$-conforming finite element is known to us. An alternative approach is to introduce and deal with mixed/order-reduced formulation. It is natural to consider possibly the operator splitting technique which introduces an intermediate variable and then reduce the original problem to a system of second order equations. This is the way adopted by Sun \cite{Sun.J2016}. The associated eigenvalue problem is also discussed therein. Beyond these discussions, few results on the discretisation are known to us.
The operator $(\nabla\times)^4$ is of fourth order and not completely symmetric; this makes the model problems bear complicated intrinsic structure. Primarily, the multiple high stiffness effects the property of the problems. Very recently, Nicaise \cite{Nicaise.S2016} studies the boundary value problem \eqref{eq:bvpA}, and proves that the solution does not generally belong to $\undertilde{H}^3(\Omega)$ on polyhedons, and the $\undertilde{H}^2(\Omega)$ regularity of the solution of \eqref{eq:bvpA} is still open even on convex polyhedrons. The multiple high stiffness also makes the concentrative construction of finite element functions difficult. Moreover, the structure of the finite element spaces are very complicated, which makes designing optimal solvers/multilevel methods difficult; there has been no discussion along this line. The order-reduced discretisation scheme by \cite{Sun.J2016} enables to utilise the existing edge element to solve the original problem; this scheme can be viewed as an analogue of the Ciarlet-Raviart's scheme \cite{Ciarlet.P;Raviart.P1974} for biharmonic equation in the context of fourth order curl problem. However, the structure has not become friendlier with this formulation. The stability analysis has not been presented in \cite{Sun.J2016}, and thus the intrinsic topology is not clear and the convergence analysis is constructed in quite a technical way there. Further, the internal structure of the finite element space is not yet clear either, and thus designing optimal solvers/multilevel methods for the scheme is also difficult.
At the current situation, in this paper, we introduce new mixed formulations to figure out and utilise a clear and amiable structure. By bringing in auxiliary variables for the problems {\bf(A)} and {\bf(B)}, we present mixed formulations which are stable in Babu\v{s}ka-Brezzi's sense on the spaces of $L^2$, $H({\rm curl})$ and $H^1$ types. Also, we establish the regularity results for the mixed formulations on convex polyhedral domains. As the mixed formulations are equivalent to the primal ones, the $\undertilde{H}{}^2(\Omega)$ regularity of $\undertilde{u}$ and $\nabla\times\undertilde{u}$ are confirmed for \eqref{eq:bvpA} and \eqref{eq:bvpB} on convex polyhedrons, and the assumptions adopted in \cite{Sun.J2016} and \cite{Zheng.B;Hu.Q;Xu.J2011} are confirmed. The mixed formulations admit amiable discretisation with finite element spaces corresponding to $L^2$, $H({\rm curl})$ and $H^1$ under some mild conditions, and the theoretical convergence analysis can be done in a standard friendly way. Several finite element examples are presented which can exploit the regularity of the solutions to an optimal extent. As the structures of the dicsretized $L^2$, $H({\rm curl})$ and $H^1$ spaces have been well-studied, the newly-developed discretisation scheme can be solved optimally by the aid of some existing optimal preconditioners\cite{Hiptmair.R;Xu.J2007,Xu.J1992,Xu.J2010,Ruesten.T;Winther.R1992}. Moreover, it is easy to find finite element spaces that are nested on nested grids, both algebraically and topologically with respect to the mixed formulation; this can bring convenience in designing further high-efficiency algorithms.
We would emphasize the new variational problems \eqref{eq:nvpA}(for \eqref{eq:bvpA}) and \eqref{eq:nvpB}(for \eqref{eq:bvpB}) are the starting point of what we are going to do and what we are able to do. These new primal formulations arise from configurating the essential boundary conditions that should be satisfied by the solutions, and they differ from traditional ones, like ones discussed in \cite{Hong.Q;Hu.J;Shu.S;Xu.J2012}, \cite{Nicaise.S2016} or \cite{Sun.J2016}. The variational formulation \eqref{eq:nvpB} is similar to the one used in \cite{Zheng.B;Hu.Q;Xu.J2011}, but the original boundary condition discussed in \cite{Zheng.B;Hu.Q;Xu.J2011} is different from that of \eqref{eq:bvpB}. The new variational formulations possess enough capacity for the essential boundary conditions, and make the sequel analysis smoother.
The remaining of the paper is organised as follows. In Section \ref{sec:pre}, we present some preliminaries and the model problems in the primal formulation. We will particularly figure out the appropriate spaces of the model problem by clarifying the boundary conditions and specify the space whose capacity is big enough for the boundary condition and the variational form. In Section \ref{sec:mf}, the mixed formulation of the model problems are given with stability analysis. Section \ref{sec:dis} is then devoted to the discretizations, including general discussion on the conditions to be satisfied, and also some specific examples. Finally in Section \ref{sec:con}, concluding remarks are given.
\section{Model problems: New primal formulations}
\label{sec:pre}
\subsection{Preliminaries: Sobolev spaces and finite elements}
Let $\Omega\subset\mathbb{R}^3$ be a simply connected polyhedral domain, with boundary $\Gamma=\partial\Omega$, and unit outward norm vector $\mathbf{n}$. In this paper, we use the bold symbol for a vector in $\mathbb{R}^3$, and a subscript $\undertilde{~}$ for a vector valued function. We assume $\Gamma$ is also connected. We use $L^2(\Omega)$ and $H^t_{(0)}(\Omega)$ for $t=1,2,\dots$ for the standard Lebesque space and Sobolev spaces. Denote
\begin{eqnarray}
H^s({\rm curl},\Omega):=\{\undertilde{v}\in (L^2(\Omega))^3:{\rm curl}^j\undertilde{v}\in (L^2(\Omega))^3,\ 1\leqslant j\leqslant t\},\ t=1,2,\dots,
\end{eqnarray}
equipped with the inner product $(\undertilde{u},\undertilde{v})_{H^t({\rm curl},\Omega)}=(\undertilde{u},\undertilde{v})+\sum_{j=1}^t({\rm curl}^j\undertilde{u},{\rm curl}^j\undertilde{v}),$ and the corresponding norm $\|\cdot\|_{H^t({\rm curl},\Omega)}$. Particularly, $H^1({\rm curl},\Omega)=H({\rm curl},\Omega)$, and $\|\cdot\|_{H({\rm curl},\Omega)}=\|\cdot\|_{{\rm curl},\Omega}$. Similarly, define
\begin{equation}
H({\rm curl}^2,\Omega):=\{\undertilde{v}\in (L^2(\Omega))^3:{\rm curl}\curl\undertilde{v}\in(L^3(\Omega))^2\},
\end{equation}
equipped with the inner product $(\undertilde{u},\undertilde{v})_{H({\rm curl}^2,\Omega)}=(\undertilde{u},\undertilde{v})+({\rm curl}\curl\undertilde{u},{\rm curl}\curl\undertilde{v})$ and the corresponding norm. Corresponding to the boundary condition, define $H^2_0({\rm curl},\Omega):=\{\undertilde{v}\in H^2({\rm curl},\Omega):\undertilde{v}\times\mathbf{n}=\undertilde{0}\ \mbox{and}\ ({\rm curl}\undertilde{v})\times\mathbf{n}=\undertilde{0}\ \mbox{on}\ \Gamma\}$. Define $H({\rm div},\Omega)=\{\undertilde{v}\in (L^2(\Omega))^3:{\rm div}\undertilde{v}\in L^2(\Omega)\}$, and $H_0({\rm div},\Omega)=\{\undertilde{v}\in H({\rm div},\Omega):\undertilde{v}\cdot\mathbf{n}=0\,\mbox{on}\,\Gamma\}$. In the sequel, we use $\nabla\times$ for ${\rm curl}$ in equations.
\begin{lemma}\label{lem:picurl}
There exists a constant $C$, such that it holds for $\undertilde{v}\in H_0({\rm curl},\Omega)$ and ${\rm div}\undertilde{v}=0$ that
\begin{equation}
\|\undertilde{v}\|_{0,\Omega}\leqslant C\|\undertilde{v}\|_{{\rm curl},\Omega}.
\end{equation}
\end{lemma}
\begin{lemma}\label{lem:hongetal}[Lemma 2.1 of \cite{Hong.Q;Hu.J;Shu.S;Xu.J2012}]
$H^2_0({\rm curl},\Omega)$ is the closure of $(\mathcal{C}^\infty_0(\Omega))^3$ in $H({\rm curl}^2,\Omega)$, and
$$
\|\nabla\times\undertilde{v}\|_{0,\Omega}\leqslant \frac{1}{2}(\|\nabla\times\nabla\times\undertilde{v}\|_{0,\Omega}+\|\undertilde{v}\|_{0,\Omega})\ \ \mbox{on}\ \,H^2_0({\rm curl},\Omega).
$$
\end{lemma}
Let $\Omega$ be subdivided to tetrahedrons which form a grid $\mathcal{G}_h$. We impose the shape regularity assumption on $\mathcal{G}_h$. On the grid can finite element spaces be constructed. We refer to \cite{GiraultRaviart1986,Monk.P2003,Ciarlet.P1978} for the context of finite element methods. We only recall these familiar finite element spaces:
\begin{itemize}
\item continuous Lagrangian element space of $k$-th degree: subspace of $H^1(\Omega)$, consist of piecewise $k$-th-degree polynomials; denoted by $\mathcal{L}^k_{h(0)}$, without or with respect to the $H^1_0(\Omega)$ boundary condition;
\item Nedelec edge element of first family of $k$-th degree: subspace of $H({\rm curl},\Omega)$, consist of piecewise polynomials of the form $\undertilde{u}+\undertilde{v}$, with $\undertilde{u}\in (P_{k-1})^3$ and $\undertilde{v}\in \undertilde{x}\times(\hat{P}_{k-1})^3$, where $P_{k-1}$ is the space of $(k-1)$-th degree polynomials, and $\hat{P}_{k-1}^3$ is the space of homogeneous $(k-1)$-th degree polynomials; denoted by $N^k_{h(0)}$, without or with respect to the $H_0({\rm curl},\Omega)$ boundary condition.
\end{itemize}
Particularly, $\mathcal{L}^0_{h}$ denotes the space of piecewise constant, and $\mathcal{L}^0_{h0}=\mathcal{L}^0_{h}\cap L^2_0(\Omega)$. The fact below is well known.
\begin{lemma}
$\nabla \mathcal{L}^k_{h(0)}=\{\undertilde{v}\in N^k_{h(0)}:\nabla\times\undertilde{v}=\undertilde{0}\}$.
\end{lemma}
\subsection{New primal formulation}
The variational formulations of the boundary value problem are used in literature:
\begin{itemize}
\item for {\bf (A)} (\cite{Sun.J2016}, e.g.,): given $\undertilde{f}$ with ${\rm div}\undertilde{f}=0$, find $\undertilde{u}\in H^2_0({\rm curl},\Omega)$ and ${\rm div}\undertilde{u}=0$, such that
\begin{equation}\label{eq:vfA}
(\nabla\times\nabla\times \undertilde{u},\nabla\times\nabla\times\undertilde{v})=(\undertilde{f},\undertilde{v}),\ \forall\,\undertilde{v}\in H^2_0({\rm curl},\Omega);
\end{equation}
\item for {\bf (B)} (\cite{Hong.Q;Hu.J;Shu.S;Xu.J2012}, e.g.,): given $\undertilde{f}$, to find $\undertilde{u}\in H^2_0({\rm curl},\Omega)$, such that
\begin{equation}\label{eq:vfB}
(\nabla\times\nabla\times \undertilde{u},\nabla\times\nabla\times\undertilde{v})+(\undertilde{u},\undertilde{v})=(\undertilde{f},\undertilde{v}),\ \forall\,\undertilde{v}\in H^2_0({\rm curl},\Omega).
\end{equation}
\end{itemize}
The well-posedness of the two variational problems is guaranteed by Lemmas \ref{lem:picurl} and \ref{lem:hongetal}.
Note that $\nabla\times\undertilde{u}\in H_0({\rm div},\Omega)$ for $\undertilde{u}\in H_0({\rm curl},\Omega)$, and thus the boundary control of $\nabla\times\undertilde{u}$ is more than the capacity of boundary condition of $H_0({\rm curl},\Omega)$. This way, we set up the variational form on another Sobolev space. We begin with the fact below.
\begin{lemma}\label{lem:h=cd}[Lemma 2.5, \cite{GiraultRaviart1986}]
If $\Omega\subset\mathbb{R}^3$ is bounded, simply connected with Liptschiz-continuous boundary, then
$$
\undertilde{H}{}^1_0(\Omega):=(H^1_0(\Omega))^3=H_0({\rm curl},\Omega)\cap H_0({\rm div},\Omega).
$$
and
$$
(\nabla\undertilde{u},\nabla\undertilde{v})=(\nabla\times\undertilde{u},\nabla\times\undertilde{v})+({\rm div}\undertilde{u},{\rm div}\undertilde{v})\ \ \mbox{for}\ \undertilde{u},\undertilde{v}\in \undertilde{H}{}^1_0(\Omega).
$$
\end{lemma}
Now, define $\undertilde{H}{}^1_0({\rm curl},\Omega):=\{\undertilde{v}\in H_0({\rm curl},\Omega):\nabla\times\undertilde{v}\in \undertilde{H}{}^1_0(\Omega)\}$. The observation below is crucial.
\begin{lemma}
$\undertilde{H}{}^1_0({\rm curl},\Omega)=H^2_0({\rm curl},\Omega)$.
\end{lemma}
\begin{proof}
Evidently, $\undertilde{H}{}^1_0({\rm curl},\Omega)\subset H^2_0({\rm curl},\Omega)$. On the other hand, given $\undertilde{v}\in H^2_0({\rm curl},\Omega)$, $\nabla\times\undertilde{v}\in H_0({\rm curl},\Omega)\cap H_0({\rm div},\Omega)=\undertilde{H}{}^1_0(\Omega)$. This finishes the proof.
\end{proof}
Moreover, $(\nabla\nabla\times\undertilde{u},\nabla\nabla\times\undertilde{v})=(\nabla\times\nabla\times\undertilde{u},\nabla\times\nabla\times\undertilde{v})$ on $\undertilde{H}{}^1_0({\rm curl},\Omega)$. Therefore, we establish the variational form of the primal model problems as:
\begin{enumerate}[{\bf(A$'$)}]
\item Given $\undertilde{f}$ with ${\rm div}\undertilde{f}=0$, find $\undertilde{u}\in \undertilde{H}{}^1_0({\rm curl},\Omega)$, ${\rm div}\,\undertilde{u}=0$, such that
\begin{equation}\label{eq:nvpA}
(\nabla\nabla\times\undertilde{u},\nabla\nabla\times \undertilde{v})=(\undertilde{f},\undertilde{v}),\ \ \forall\,\undertilde{v}\in \undertilde{H}{}^1_0({\rm curl},\Omega).
\end{equation}
\item Given $\undertilde{f}$, find $\undertilde{u}\in \undertilde{H}{}^1_0({\rm curl},\Omega)$, such that
\begin{equation}\label{eq:nvpB}
(\nabla\nabla\times\undertilde{u},\nabla\nabla\times \undertilde{v})+(\undertilde{u},\undertilde{v})=(\undertilde{f},\undertilde{v}),\ \ \forall\,\undertilde{v}\in \undertilde{H}{}^1_0({\rm curl},\Omega).
\end{equation}
\end{enumerate}
\begin{lemma}
The variational problems {\bf(A$'$)} and {\bf(B$'$)} are well-posed. They are equivalent to \eqref{eq:vfA} and \eqref{eq:vfB}, respectively.
\end{lemma}
\begin{remark}
The variational problem on $\undertilde{H}{}^1_0({\rm curl},\Omega)$ has indeed been used in \cite{Zheng.B;Hu.Q;Xu.J2011}, where the original boundary condition is $\undertilde{u}\times\mathbf{n}=\nabla\times\undertilde{u}=\undertilde{0}$. In this paper here, we show that even when the boundary condition is simplified, the space $\undertilde{H}{}^1_0({\rm curl},\Omega)$ is still the appropriate one.
\end{remark}
\section{Mixed formulation of model problems}
\label{sec:mf}
In this section, we present mixed problems that are equivalent to {\bf (A$'$)} and {\bf (B$'$)}, thus to {\bf (A)} and {\bf (B)}, respectively. The stability and regularity of the mixed problems are given.
\subsection{Mixedization of Problem {\bf(A $'$)}}
We start with the observations below. Let $\undertilde{u}$ be the solution of {\bf (A$'$)} (thus {\bf (A)}). Define $\undertilde{\varphi}:=\nabla\times\undertilde{u}$, then $\undertilde{\varphi}\in\undertilde{H}{}^1_0(\Omega)$, and ${\rm div}\undertilde{\varphi}=0$. On the other hand, given $\undertilde{\varphi}\in\undertilde{H}{}^1_0(\Omega)$, $\undertilde{\varphi}={\rm curl}\undertilde{u}$ for $\undertilde{u}\in H_0({\rm curl},\Omega)$ iff $({\rm div}\undertilde{\varphi},q)=0$ for any $q\in L^2_0(\Omega)$ and $(\undertilde{\varphi},\nabla\times\undertilde{s})=(\nabla\times\undertilde{u},\nabla\times\undertilde{s})$ for any $\undertilde{s}\in H_0({\rm curl},\Omega)$. Also, $\undertilde{u}$ is uniquely determined by the divergence free condition, which reads equivalently $(\undertilde{u},\nabla h)=0$ for any $h\in H_0(\Omega)$. Now define
\begin{equation}
V:=H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times\undertilde{H}{}^1_0(\Omega)\times L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega).
\end{equation}
The mixed formulation of {\bf (A)} is to find $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)\in V$, such that for $\forall\,(n,\undertilde{v},\undertilde{\psi},q,\undertilde{s},h)\in V$,
\begin{equation}\label{eq:mxdvfA'}
\left\{
\begin{array}{ccccccll}
&&&&(\undertilde{r},\nabla n)&& = 0,
\\
&&&&-(\nabla\times\undertilde{r},\nabla\times\undertilde{v})&+(\nabla g,\undertilde{v})& =(\undertilde{f},\undertilde{v}),
\\
&&(\nabla\undertilde{\varphi},\nabla\undertilde{\psi})&+(p,{\rm div}\undertilde{\psi}) &+(\nabla\times\undertilde{r},\undertilde{\psi})& & =0,
\\
&&({\rm div}\undertilde{\varphi},q)&&&&=0,
\\
(\nabla m,\undertilde{s})&-(\nabla\times\undertilde{u},\nabla\times\undertilde{s})& (\undertilde{\varphi},\nabla\times\undertilde{s}) & &&&=0,
\\
&(\undertilde{u},\nabla h)&&&&&=0.
\end{array}
\right.
\end{equation}
\begin{lemma}\label{lem:wellposeA'}
Given $\undertilde{f}\in \undertilde{L}{}^2(\Omega)$, the problem \eqref{eq:mxdvfA'} admits a unique solution $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)\in V$. Moreover,
\begin{equation}
\|m\|_{1,\Omega}+\|\undertilde{u}\|_{{\rm curl},\Omega}+\|\undertilde{\varphi}\|_{1,\Omega}+\|p\|_{0,\Omega}+\|\undertilde{r}\|_{{\rm curl},\Omega}+\|g\|_{1,\Omega}\leqslant C\|\undertilde{f}\|_{(H_0({\rm curl},\Omega))'}.
\end{equation}
\end{lemma}
\begin{proof}
We are going to verify conditions by Brezzi's theory. Define
\begin{equation}\label{eq:aforA}
a((m,\undertilde{u},\undertilde{\varphi}),(n,\undertilde{v},\undertilde{\psi})):=(\nabla\undertilde{\varphi},\nabla\undertilde{\psi}),
\end{equation}
and
\begin{equation}\label{eq:bforA}
b((m,\undertilde{u},\undertilde{\varphi}),(q,\undertilde{s},h)):=({\rm div}\undertilde{\varphi},q)+(\nabla m,\undertilde{s})-(\nabla\times\undertilde{u},\nabla\times\undertilde{s})+(\undertilde{\varphi},\nabla\times\undertilde{s})+(\undertilde{u},\nabla h).
\end{equation}
Then $a(\cdot,\cdot)$ and $b(\cdot,\cdot)$ are continuous on $[H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H{}^1_0(\Omega)}]^2$ and $[H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega)]\times[L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega)]$, respectively. Define $Z:=\{(m,\undertilde{u},\undertilde{\varphi})\in [H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega)]: b((m,\undertilde{u},\undertilde{\varphi}),(q,\undertilde{s},h))=0,\ \forall\,(q,\undertilde{s},h)\in [L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega)]\}$. Then it remains for us to verify the coercivity of $a(\cdot,\cdot)$ on $Z$ and inf-sup condition: given nonzero $(q,\undertilde{s},h)\in L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega)$,
\begin{equation}
\sup_{(m,\undertilde{u},\undertilde{\varphi})\in H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega)}\frac{b((m,\undertilde{u},\undertilde{\varphi}),(q,\undertilde{s},h))}{\|m\|_{1,\Omega},\|\undertilde{u}\|_{{\rm curl},\Omega},\|\undertilde{\varphi}\|_{1,\Omega}}\geqslant C(\|q\|_{0,\Omega}+\|\undertilde{s}\|_{{\rm curl},\Omega}+\|h\|_{1,\Omega}).
\end{equation}
Given $(m,\undertilde{u},\undertilde{\varphi})\in Z$, then $m=0$. Since $(\undertilde{u},\nabla h)=0$ for any $h\in H^1_0(\Omega)$, we have $\|\undertilde{u}\|_{{\rm curl},\Omega}\leqslant C\|\nabla\times\undertilde{u}\|_{0,\Omega}$. Since $(\nabla\times\undertilde{u},\nabla\times\undertilde{u})=(\undertilde{\varphi},\nabla\times\undertilde{u})$, we have $\|\nabla\times\undertilde{u}\|_{0,\Omega}\leqslant \|\undertilde{\varphi}\|_{0,\Omega}\leqslant C\|\nabla\undertilde{\varphi}\|_{0,\Omega}$. This confirms the coercivity of $a(\cdot,\cdot)$ on $Z$.
Given $(q,\undertilde{s},h)\in L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega)$, firstly, we decomose $\undertilde{s}=\undertilde{s}{}_1+\undertilde{s}{}_2$, such that $\undertilde{s}{}_1\in\nabla H^1_0(\Omega)$, and $\undertilde{s}{}_2\in (\nabla H^1_0(\Omega))^\perp$. Set $\undertilde{\varphi}$ to be such that $({\rm div}\undertilde{\varphi},q)=(q,q)$ and $\|\undertilde{\varphi}\|_{1,\Omega}\leqslant C\|{\rm div}\undertilde{\varphi}\|_{0,\Omega}$, $m$ to be such that $\nabla m=\undertilde{s}{}_1$. Further, $\undertilde{u}$ is chosen to be $\undertilde{u}{}_1+\nabla h$, such that $(\undertilde{u}{}_1,\nabla g)=0$ for any $g\in H^1_0(\Omega)$ and $(\undertilde{\varphi}-\nabla\undertilde{u}{}_1,\nabla\times\undertilde{v})=(\nabla\times\undertilde{s},\nabla\times\undertilde{v})$ for any $\undertilde{v}\in H_0({\rm curl},\Omega)$. Then
$$
b((m,\undertilde{u},\undertilde{\varphi}),(q,\undertilde{s},h))=(q,q)+(\undertilde{s}{}_1,\undertilde{s}{}_1)+({\rm curl}\undertilde{s}{}_2,{\rm curl}\undertilde{s}{}_2)+(\nabla h,\nabla h)\geqslant C(\|q\|_0^2+\|\undertilde{s}\|_{{\rm curl},\Omega}^2+\|\nabla h\|_{0,\Omega}^2).
$$
Meanwhile, $\|\nabla m\|_{0,\Omega}=\|\undertilde{s}{}_1\|_{0,\Omega}\leqslant \|\undertilde{s}\|_{{\rm curl},\Omega}$, $\|\undertilde{\varphi}\|_{1,\Omega}\leqslant C\|q\|_{0,\Omega}$, and $\|\undertilde{u}\|_{{\rm curl},\Omega}\leqslant C(\|\nabla h\|_{0,\Omega}+\|\undertilde{s}\|_{{\rm curl},\Omega})$. This constructs the inf-sup condition and finishes the proof.
\end{proof}
\begin{lemma}\label{lem:equiv}
The problem \eqref{eq:mxdvfA'} is equivalent to the variational problem {\bf (A$'$)}.
\end{lemma}
\begin{proof}
Let $\undertilde{u}\in \undertilde{H}{}^1_0({\rm curl},\Omega)$ such that $(\nabla\nabla\times\undertilde{u},\nabla\nabla\times\undertilde{v})=(\undertilde{f},\undertilde{v})$ for any $\undertilde{v}\in \undertilde{H}{}^1_0({\rm curl},\Omega)$, then $(0,\undertilde{u},\nabla\times\undertilde{u},0,\nabla\times\undertilde{u},0)$ solves the system \eqref{eq:mxdvfA'} with the same $\undertilde{f}$. On the other hand, let $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)$ be the solution of \eqref{eq:mxdvfA'}. Then ${\rm div}\undertilde{u}=0$, $\nabla\times\undertilde{u}=\undertilde{\varphi}$, and it can be proved that $(\nabla\undertilde{\varphi},\nabla\undertilde{\psi})=(\undertilde{f},\undertilde{v})$ for any $\undertilde{\psi}\in\undertilde{H}{}^1_0(\Omega)$ and $\undertilde{\psi}={\rm curl}\undertilde{v}$. The proof is completed.
\end{proof}
\begin{lemma}\label{lem:decomp}
The problem \eqref{eq:mxdvfA'} can be decomposed to the three subsystems and solved sequentially:
\begin{enumerate}
\item given $\undertilde{f}$, solve for $\undertilde{r}\in H_0({\rm curl},\Omega)$ and $g\in H^1_0(\Omega)$ that
\begin{equation}\label{eq:subA-1}
\left\{
\begin{array}{ccll}
(\nabla\times\undertilde{r},\nabla\times\undertilde{v}) & -(\nabla g,\undertilde{v}) &=-(\undertilde{f},\undertilde{v}) &\undertilde{v}\in H_0({\rm curl},\Omega)
\\
(\nabla n,\undertilde{r})& &=0&\forall\,n\in H^1_0(\Omega);
\end{array}
\right.
\end{equation}
\item with $\undertilde{r}$ obtained, solve for $\undertilde{\varphi}\in \undertilde{H}{}^1_0(\Omega)$ and $p\in L^2_0(\Omega)$ that
\begin{equation}\label{eq:subA-2}
\left\{
\begin{array}{ccll}
(\nabla\undertilde{\varphi},\nabla\undertilde{\psi}) & +(p,{\rm div}\undertilde{\psi}) &=-(\nabla\times\undertilde{r},\undertilde{\psi})&\forall\,\undertilde{\psi}\in \undertilde{H}{}^1_0(\Omega)
\\
({\rm div}\undertilde{\varphi},q)&&=0&\forall\,q\in L^2_0(\Omega);
\end{array}
\right.
\end{equation}
\item with $\undertilde{\varphi}$ obtained, solve for $\undertilde{u}\in H_0({\rm curl},\Omega)$ and $m\in H^1_0(\Omega)$ that
\begin{equation}\label{eq:subA-3}
\left\{
\begin{array}{ccll}
(\nabla\times\undertilde{u},\nabla\times\undertilde{s})&-(\nabla m,\undertilde{s}) &=(\undertilde{\varphi},\nabla\times\undertilde{s})&\forall\,\undertilde{s}\in H_0({\rm curl},\Omega)
\\
(\nabla h,\undertilde{u})&&=0&\forall\,h\in H^1_0(\Omega).
\end{array}
\right.
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
We only have to show that the three subproblems are all well-posed, which can be verified by the stable Helmholtz decomposition of $H_0({\rm curl},\Omega)$. The proof is finished.
\end{proof}
The theorem below constructs the regularity of the mixed system.
\begin{theorem}\label{thm:regA}
Let $\Omega$ be a convex polyhedron, and $\undertilde{f}\in (L^2(\Omega))^3$ such that ${\rm div}\undertilde{f}=0$. Let $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)$ be the solution of \eqref{eq:mxdvfA'}. Then
\begin{eqnarray}
&m=0,
\\
&\undertilde{u}\in \undertilde{H}{}^2(\Omega)\cap H_0({\rm curl},\Omega),\ {\rm curl}\undertilde{u}\in \undertilde{H}{}^2(\Omega),
\\
&\undertilde{\varphi}\in \undertilde{H}{}^2(\Omega)\cap \undertilde{H}{}^1_0(\Omega),\label{eq:regAphi}
\\
&p\in H^1(\Omega)\cap L^2_0(\Omega)\label{eq:regAp}
\\
&\undertilde{r}\in \undertilde{H}{}^2(\Omega)\cap H_0({\rm curl},\Omega),\label{eq:regAr}
\\
&g=0.\label{eq:regAg}
\end{eqnarray}
\end{theorem}
We postpone the proof of Theorem \ref{thm:regA} after some technical lemmas.
\begin{lemma}(Sections 3.4 and 3.5 of \cite{GiraultRaviart1986})\label{lem:GRembd}
Let $\Omega$ be a convex polyhedron. Then
\begin{equation}
H_0({\rm div},\Omega)\cap H({\rm curl},\Omega)\subset\undertilde{H}{}^1(\Omega),\ \ H_0({\rm curl},\Omega)\cap H({\rm div},\Omega)\subset\undertilde{H}{}^1(\Omega).
\end{equation}
\end{lemma}
The lemma below is a smoothened analogue of Lemma \ref{lem:GRembd}, and we adopt the version in \cite{Chen.L;Wu.Y;Zhong.L;Zhou.J2016}. From this point onwards, $\lesssim$, $\gtrsim$, and $\raisebox{-1.5mm}{$\;\stackrel{\raisebox{-3.9mm}{=}}{{\sim}}\;$}$ respectively denote $\leqslant$, $\geqslant$, and $=$ up to a constant. The hidden constants depend on the domain, and, when triangulation is involved, they also depend on the shape-regularity of the triangulation, but they do not depend on $h$ or any other mesh parameter.
\begin{lemma}\label{lem:CWZZlem3.2}(Lemma 3.2 of \cite{Chen.L;Wu.Y;Zhong.L;Zhou.J2016})
For functions $\undertilde{\psi}\in H({\rm div},\Omega)\cap H_0({\rm curl},\Omega)$ or $H_0({\rm div},\Omega)\cap H({\rm curl},\Omega)$ satisfying ${\rm curl}\undertilde{\psi}\in \undertilde{H}{}^1(\Omega)$ and ${\rm div}\undertilde{\psi}\in \undertilde{H}{}^1(\Omega)$. Then $\undertilde{\psi}\in \undertilde{H}{}^2(\Omega)$ and $\|\undertilde{\psi}\|_2\lesssim \|{\rm curl}\undertilde{\psi}\|_{1,\Omega}+\|{\rm div}\undertilde{\psi}\|_{1,\Omega}$.
\end{lemma}
\paragraph{\bf Proof of Theorem \ref{thm:regA}}
By Lemma \ref{lem:decomp}, we will show the regularity result by dealing with the systems \eqref{eq:subA-1}, \eqref{eq:subA-2} and \eqref{eq:subA-3} sequentially.
Since ${\rm div}\undertilde{f}=0$, by \eqref{eq:subA-1}, it holds that $g=0$, and $(\nabla\times\undertilde{r},\nabla\times\undertilde{v})=(\undertilde{f},\undertilde{v})$ for any $\undertilde{v}\in H_0({\rm curl},\Omega)$. Thus $\nabla\times\nabla\times\undertilde{r}=\undertilde{f}\in \undertilde{L}^2(\Omega)$, namely $\nabla\times\undertilde{r}\in H({\rm curl},\Omega)$. Also, $\nabla\times\undertilde{r}\in H_0({\rm div},\Omega)$, and it follows that $\nabla\times\undertilde{r}\in H({\rm curl},\Omega)\cap H_0({\rm div},\Omega)\subset \undertilde{H}{}^1(\Omega)$. Note that by \eqref{eq:subA-1}, ${\rm div}\undertilde{r}=0$, thus by Lemma \ref{lem:CWZZlem3.2}, $\undertilde{r}\in \undertilde{H}^2(\Omega)$. This proves \eqref{eq:regAr} and \eqref{eq:regAg}.
Substitute $\nabla\times\undertilde{r}$ into the system \eqref{eq:subA-2}, standardly we obtain the estimate \eqref{eq:regAphi} and \eqref{eq:regAp}. We refer to \cite{Nicaise.S1997} for the regularity analysis of the 3D Stokes problem.
Further, substitute $\undertilde{\varphi}$ into the system \eqref{eq:subA-3}, we obtain $m=0$, ${\rm div}\undertilde{u}=0$, $\undertilde{u}\in \undertilde{H}{}^1(\Omega)\cap H_0({\rm curl},\Omega)$ and ${\rm curl}\undertilde{u}\in \undertilde{H}{}^1(\Omega)$. By Lemma \ref{lem:CWZZlem3.2} again, it holds that $\undertilde{u}\in\undertilde{H}{}^2(\Omega)$. Moreover, $\nabla\times\undertilde{u}=\undertilde{\varphi}$; this leads to $\nabla\times\undertilde{u}\in\undertilde{H}{}^2(\Omega)$.
Summing all above completes the proof. \qed
\begin{remark}
This theorem constructs the regularity result of the boundary value problem {\bf(A)} in forms \eqref{eq:bvpA} and \eqref{eq:vfA}, and confirms the regularity assumption of \cite{Sun.J2016} in the mid of Page 190 there. Moreover, by the same virtue, if further $\undertilde{f}\in\undertilde{H}{}^1(\Omega)$, by Lemma \ref{lem:CWZZlem3.2}, we will obtain ${\rm curl}\undertilde{r}\in\undertilde{H}{}^2(\Omega)$.
\end{remark}
\subsection{Mixedization of Problem (B$'$)}
We simply repeat the procedure for Problem {\bf (A$'$)}. Define
$$
U:=H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega)\times H_0({\rm curl},\Omega)\times L^2_0(\Omega).
$$
The mixed variational problem is to find $(m,\undertilde{u},\undertilde{\varphi},\undertilde{u},p)\in U$, such that
\begin{equation}\label{eq:mxdvpB'}
\left\{
\begin{array}{ccccccll}
&&&(\undertilde{r},\nabla n)&&& = 0, & n\in H^1_0(\Omega)
\\
&(\undertilde{u},\undertilde{v})&&-(\nabla\times\undertilde{r},\nabla\times\undertilde{v})&& &=(\undertilde{f},\undertilde{v}),& \undertilde{v}\in H_0({\rm curl},\Omega)
\\
&&(\nabla\undertilde{\varphi},\nabla\undertilde{\psi})&+(\nabla\undertilde{r},\undertilde{\psi})& &+(p,{\rm div}\undertilde{\psi}) & =0, & \undertilde{\psi}\in \undertilde{H}{}^1_0(\Omega)
\\
(\nabla m,\undertilde{s})&-(\nabla\times\undertilde{u},\nabla\times\undertilde{s})& (\undertilde{\varphi},\nabla\times\undertilde{s}) & &&&=0, & \undertilde{s}\in H_0({\rm curl},\Omega)
\\
&&({\rm div}\undertilde{\varphi},q)&&&&=0, &q\in L^2_0(\Omega).
\end{array}
\right.
\end{equation}
\begin{lemma}
Given $\undertilde{f}\in \undertilde{L}{}^2(\Omega)$, the problem \eqref{eq:mxdvpB'} admits a unique solution $(m,\undertilde{u},\undertilde{\varphi},\undertilde{r},p)\in U$. Moreover,
\begin{equation}
\|m\|_{1,\Omega}+\|\undertilde{u}\|_{{\rm curl},\Omega}+\|\undertilde{\varphi}\|_{1,\Omega}+\|\undertilde{r}\|_{{\rm curl},\Omega}+\|p\|_{0,\Omega}\leqslant C\|\undertilde{f}\|_{(H_0({\rm curl},\Omega))'}.
\end{equation}
\end{lemma}
\begin{proof}
Again, we are going to verify Brezzi's conditions. Define $a((m,\undertilde{u},\undertilde{\varphi}),(n,\undertilde{v},\undertilde{\psi})):=(\undertilde{u},\undertilde{v})+(\nabla\undertilde{\varphi},\nabla\undertilde{\psi})$, and $b((m,\undertilde{u},\undertilde{\varphi}),(\undertilde{s},q)):=(\nabla m,\undertilde{s})-(\nabla\times\undertilde{u},\nabla\times\undertilde{s})+(\undertilde{\varphi},\nabla\times\undertilde{s})+({\rm div}\undertilde{\varphi},q)$. The continuity of $a(\cdot,\cdot)$ and $b(\cdot,\cdot)$ associated with $U$ follow immediately.
Define $Z:=\{(m,\undertilde{u},\undertilde{\varphi})\in H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega):b((m,\undertilde{u},\undertilde{\varphi}),(\undertilde{s},q))=0,\forall\,(\undertilde{s},q)\in H_0({\rm curl},\Omega)\times L^2_0(\Omega)\}$. Given $(m,\undertilde{u},\undertilde{\varphi})\in Z$, we have $m=0$ and $\undertilde{\varphi}=\nabla\times\undertilde{u}$. Thus the coercivity of $a(\cdot,\cdot)$ on $Z$ follows.
Given $(\undertilde{s},q)\in H_0({\rm curl},\Omega)\times L^2_0(\Omega)$, decompose $\undertilde{s}=\undertilde{s}{}_1+\undertilde{s}{}_2$ with $\undertilde{s}{}_1\in \nabla H^1_0(\Omega)$, and $\undertilde{s}{}_2\in(\nabla H^1_0(\Omega))^\perp$. Set $\undertilde{\varphi}\in\undertilde{H}{}^1_0(\Omega)$, such that $({\rm div}\undertilde{\varphi},q)=(q,q)$ and $\|\undertilde{\varphi}\|_1\leqslant C\|q\|_0$, $\nabla m=\undertilde{s}{}_1$, and $\undertilde{u}\in (\nabla H^1_0(\Omega))^\perp$ such that $(\undertilde{\varphi}-\nabla\times \undertilde{u},\nabla\times\undertilde{v})=(\nabla\times\undertilde{s},\nabla\times\undertilde{v})$ for any $\undertilde{v}\in H_0({\rm curl},\Omega)$. Then $b((m,\undertilde{u},\undertilde{\varphi}),(\undertilde{s},q))=(\undertilde{s}{}_1,\undertilde{s}{}_1)+(\nabla\times\undertilde{s},\nabla\times\undertilde{s})+(q,q)$, and $\|m\|_1+\|\undertilde{u}\|_{{\rm curl}}+\|\undertilde{\varphi}\|_{1,\Omega}\leqslant C(\|\undertilde{s}\|_{{\rm curl},\Omega}+\|q\|_{0,\Omega})$. This leads to the inf-sup condition and completes the proof.
\end{proof}
Similar to Lemma \ref{lem:equiv}, we can obtain the equivalence result below.
\begin{lemma}
The problem \eqref{eq:mxdvpB'} is equivalent to the primal problem {\bf (B$'$)}.
\end{lemma}
\begin{theorem}\label{thm:regB}
Let $\Omega$ be a convex polyhedron and $(m,\undertilde{u},\undertilde{\varphi},\undertilde{r},p)$ be the solution of \eqref{eq:mxdvpB'}. Then $m=0$, ${\rm curl}\undertilde{u}\in \undertilde{H}{}^2(\Omega)$, $\undertilde{\varphi}\in \undertilde{H}{}^2(\Omega)\cap \undertilde{H}{}^1_0(\Omega)$, $\undertilde{r}\in \undertilde{H}{}^1(\Omega)\cap H_0({\rm curl},\Omega)$, ${\rm curl}\undertilde{r}\in \undertilde{H}{}^1(\Omega)$ and $p\in L^2_0(\Omega)\cap H^1(\Omega)$. Further, if ${\rm div}\undertilde{f}=0$, then $\undertilde{u}\in \undertilde{H}{}^2(\Omega)$.
\end{theorem}
\begin{proof}
By the stability of the system and since ${\rm div}\undertilde{r}=0$, we obtain $\undertilde{r}\in \undertilde{H}{}^1(\Omega)$. As $(\nabla\times\undertilde{r},\nabla\times\undertilde{v})=-(\undertilde{f}-\undertilde{u},\undertilde{v})$ for any $\undertilde{v}\in H_0({\rm curl},\Omega)$, we have $\nabla\times\nabla\times\undertilde{r}=\undertilde{f}-\undertilde{u}$, thus $\nabla\times\undertilde{r}\in H({\rm curl},\Omega)\cap H_0({\rm div},\Omega)\subset \undertilde{H}{}^1(\Omega)$. Then, standardly, we obtain $\undertilde{\varphi}\in\undertilde{H}{}^2(\Omega)\cap \undertilde{H}{}^1_0(\Omega)$, and $p\in H^1(\Omega)\cap L^2_0(\Omega)$. By the second last line of the system, $m=0$ and $\nabla\times\undertilde{u}=\undertilde{\varphi}$; thus $\nabla\times\undertilde{u}\in\undertilde{H}{}^2(\Omega)$. Further, if ${\rm div}\undertilde{f}=0$, then ${\rm div}\undertilde{u}=0$; this combined with that ${\rm curl}\undertilde{u}\in\undertilde{H}{}^1_0(\Omega)$ leads to that $\undertilde{u}\in\undertilde{H}{}^2(\Omega)$. The proof is completed.
\end{proof}
\begin{remark}
Theorem \ref{thm:regB} constructs regularity of {\bf (B)} in forms \eqref{eq:bvpB} and \eqref{eq:mxdvpB'}. It also confirms the validity of assumptions in Lemma 3.8 and Theorem 3.12 of \cite{Zheng.B;Hu.Q;Xu.J2011} by showing that $\undertilde{u}\in\undertilde{H}{}^2(\Omega)$ and $\nabla\times\undertilde{u}\in\undertilde{H}{}^2(\Omega)$ when $\undertilde{f}\in \undertilde{L}{}^2(\Omega)$ with ${\rm div}\undertilde{f}=0$.
The argument of this proof can be repeated onto that of Theorem \ref{thm:regA}. The difference between the two proofs is that we do not try to decompose \eqref{eq:mxdvpB'} to subsystems sequentially.
\end{remark}
\section{Finite element discretizations of the mixed formulations}
\label{sec:dis}
For simplicity, we only consider the conforming finite element discretizations. Namely, we choose finite element spaces $H^1_{h0}\subset H^1_0(\Omega)$, $H_{h0}({\rm curl})\subset H_0({\rm curl},\Omega)$, $\undertilde{H}{}^1_{h0}\subset \undertilde{H}{}^1_0(\Omega)$ and $L^2_{h0}\subset L^2_0(\Omega)$, and use them to replace the respective Sobolev spaces when existing in the mixed variational form to generate a discretisation scheme. Particularly, the spaces $H^1_0(\Omega)$ and $H_0({\rm curl},\Omega)$ may appear more than once in the mixed formulation; this hints us to use different $H^{1,a}_{h0}$ and $H^{1,b}_{h0}$ and different $H^a_{h0}({\rm curl})$ and $H^b_{h0}({\rm curl})$ as their respective discretisation when convenient. In this section, we present some conditions of the well-posed-ness of the discretised system, construct generally its convergence analysis, and give some examples.
\subsection{Discretize Problem (A$'$)}
Let $H^{1,a}_{h0}\times H^a_{h0}({\rm curl})$ and $H^{1,b}_{h0}\times H^b_{h0}({\rm curl})$, identical or not, be two finite element subspaces of $H^1_0(\Omega)\times H_0({\rm curl},\Omega)$, and finite element space $\undertilde{H}{}^1_{h0}\times L^2_{h0}\subset \undertilde{H}{}^1_0(\Omega)\times L^2_0(\Omega)$. Define $V_h:=H^{1,a}_{h0}\times H^a_{h0}({\rm curl})\times \undertilde{H}{}^1_{h0}\times L^2_{h0}\times H^b_{h0}({\rm curl})\times H^{1,b}_{h0}$, and $V_h':=H^{1,b}_{h0}\times H^b_{h0}({\rm curl})\times \undertilde{H}{}^1_{h0}\times L^2_{h0}\times H^a_{h0}({\rm curl})\times H^{1,a}_{h0}$. The discretized formulation of {\bf (A$'$) } is to find $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,p_h,\undertilde{r}{}_h,g_h)\in V_h$, such that, for any $(n_h,\undertilde{v}{}_h,\undertilde{\psi}{}_h,q_h,\undertilde{s}{}_h,h_h)\in V_h'$,
\begin{equation}\label{eq:dismxdvfA'
\left\{
\begin{array}{lll}
a(( m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h),( n_h,\undertilde{v}{}_h,\undertilde{\psi}{}_h))&+b(( p_h,\undertilde{r}{}_h, g_h),(n_h,\undertilde{v}{}_h,\undertilde{\psi}{}_h)) &= (\undertilde{f},\undertilde{v}{}_h),
\\
b(( m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h), ( q_h,\undertilde{s}{}_h, h_h)) &&=0,
\end{array}
\right.
\end{equation}
where $a(\cdot,\cdot)$ and $b(\cdot,\cdot)$ follows the definitions \eqref{eq:aforA} and \eqref{eq:bforA}. The system is symmetric indefinite when $V_h=V_h'$, and unsymmetric otherwise.
For the well-posedness of the discretised system, we set up some assumptions below.
\begin{description}
\item[A1] The exact relation holds: $\nabla H^{1,\alpha}_{h0}=\{\undertilde{s}{}_h\in H_{h0}^\alpha({\rm curl}):{\rm curl}\undertilde{s}{}_h=0\},\ \alpha=a,b.$
\item[A2] The Poincar\'e inequality holds: $\|\undertilde{s}{}_h\|_{0,\Omega}\leqslant C\|\nabla\times\undertilde{s}{}_h\|_{0,\Omega}$ for $\undertilde{s}{}_h\in (\nabla H^{1,\alpha}_{h0})^\perp$. Here $ (\nabla H^{1,\alpha}_{h0})^\perp$ is the orthogonal completion of $(\nabla H^{1,\alpha}_{h0})$ in $H_{h0}^\alpha({\rm curl})$ in $L^2$ inner product, $\alpha=a,b$.
\item[A3] The inf-sup condition holds: $\displaystyle\inf_{q_h\in L^2_{h0}\setminus\{0\}}\sup_{\undertilde{\psi}{}_h\in \undertilde{H}{}^1_{h0}\setminus\{\undertilde{0}\}}\frac{({\rm div}\undertilde{\psi}{}_h,q_h)}{\|q_h\|_{0,\Omega}\|\undertilde{\psi}{}_h\|_{1,\Omega}}\geqslant C$.
\end{description}
\begin{remark}
The assumptions {\bf A1} and {\bf A2} are imposed on the space pair $H^{1,\alpha}_{h0}$ and $H_{h0}^\alpha({\rm curl})$, $\alpha=a,b$, and the assumption {\bf A3} is imposed on the space pair $\undertilde{H}{}^1_{h0}$ and $L^2_{h0}$. This allows us to choose the three space pairs independently. Provided {\bf A3} itself, the system \eqref{eq:dismxdvfA'} can be decomposed to three subproblems and solved sequentially.
\end{remark}
\begin{lemma}
Provided assumptions \textbf{A1}$\sim$\textbf{A3}, the problem \eqref{eq:dismxdvfA'} is well-posed on $V_h$.
\end{lemma}
\begin{proof}
When $H^{1,a}_{h0}=H^{1,b}_{h0}$ and $H^a_{h0}({\rm curl})=H^b_{h0}({\rm curl})$, namely $V_h=V_h'$, the proof is essentially the same as that of Lemma \ref{lem:wellposeA'}; otherwise, we can decompose the system to three subproblems and analyse the subproblems one by one, and the result can be proved.
\end{proof}
The convergence of the scheme is surveyed in the lemma below.
\begin{lemma}\label{lem:errA}
Let $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)\in V$ and $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,p_h,\undertilde{r}{}_h,g_h)\in V_h$ be the solutions of \eqref{eq:mxdvfA'} and \eqref{eq:dismxdvfA'}, respectively. Provided assumptions {\bf A1}$\sim${\bf A3}, it holds that
\begin{enumerate}
\item $\displaystyle\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}\leqslant C\inf_{\undertilde{s}{}_h\in H^b_{h0}({\rm curl})}\|\undertilde{r}-\undertilde{s}{}_h\|_{{\rm curl},\Omega};$
\item $\displaystyle \|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{1,\Omega}+\|p-p_h\|_{0,\Omega}\leqslant C\big[\inf_{\undertilde{\psi}{}_h\in\undertilde{H}{}^1_{h0},q_h\in L^2_{h0}}(\|\undertilde{\varphi}-\undertilde{\psi}{}_h\|_{1,\Omega}+\|p-q_h\|_{0,\Omega})+\|\undertilde{r}-\undertilde{r}{}_h\|_{0,\Omega}\big]$;
\item $\displaystyle \|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega} \leqslant C\big[\inf_{\undertilde{v}{}_h\in H^a_{h0}({\rm curl})}\|\undertilde{u}-\undertilde{v}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}\big]$;
\item $m_h=0=m$, $g_h=0=g$.
\end{enumerate}
\end{lemma}
\begin{proof}
Similar to Lemma \ref{lem:decomp}, we can decompose \eqref{eq:dismxdvfA'} to three subproblems and solve them sequentially. Then the lemma follows from the Ce\'a lemma and the second Strang lemma directly. We only have to note that, as ${\rm div}\undertilde{f}=0$, it follows that $g_h=0$ and $m_h=0$ by {\bf A1}, and since $g=0$ and $m=0$, $\inf_{h_h\in H^1_{h0}}\|g-h_h\|_{1,\Omega}=0$ and $\inf_{n_h\in H^1_{h0}}\|m-n_h\|=0$.
\end{proof}
\begin{remark}
By Lemma \ref{lem:errA}, the error $\|m-m_h\|_{1,\Omega}+\|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega}$ could be comparable to $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}$. This implies that when $H^{1,a}_{h0}$ and $H^a_{h0}({\rm curl})$ are chosen appropriately, a higher order convergence rate of $\undertilde{u}$ may be expected. Indeed, for general $\undertilde{f}$ (not assuming $\undertilde{f}\in \undertilde{H}{}^1(\Omega)$), that we choose $H^a_{h0}({\rm curl})$ bigger than $H^b_{h0}({\rm curl})$ (meanwhile $H^{1,a}_{h0}$ bigger than $H^{1,b}_{h0}$) coincides with the higher smoothness of $\undertilde{u}$ than that of $\undertilde{r}$ on convex polyhedrons. However, if $H^{1,a}_{h0}\neq H^{1,b}_{h0}$, $V_h\neq V_h'$ and the problem is no longer symmetric, which may bring some extra difficulty. This is why we still discuss the choice $V_h=V_h'$ in, e.g., eigenvalue computation and other applications.
\end{remark}
\subsection{Discretize Problem (B$'$)}
Define $U_h:=H^1_{h0}\times H_{h0}({\rm curl})\times\undertilde{H}{}^1_{h0}\times H_{h0}({\rm curl})\times L^2_{h0}$. The discretized mixed formulation is to find $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,\undertilde{r}{}_h,p_h)\in U_h$, such that, for any $(n_h,\undertilde{v}{}_h,\undertilde{\psi}{}_h,\undertilde{s}{}_h,q_h)\in U_h$,
\begin{equation}\label{eq:dismxdvpB'}
\left\{
\begin{array}{ccccccll}
&&&(\undertilde{r}{}_h,\nabla n_h)&&& = 0 &
\\
&(\undertilde{u}{}_h,\undertilde{v}{}_h)&&-(\nabla\times\undertilde{r}{}_h,\nabla\times\undertilde{v}{}_h)&& &=(\undertilde{f},\undertilde{v}{}_h)
\\
&&(\nabla\undertilde{\varphi}{}_h,\nabla\undertilde{\psi}{}_h)&+(\nabla\times\undertilde{r}{}_h,\undertilde{\psi}{}_h)& &+(p_h,{\rm div}\undertilde{\psi}{}_h) & =0 &
\\
(\nabla m_h,\undertilde{s}{}_h)&-(\nabla\times\undertilde{u}{}_h,\nabla\times\undertilde{s}{}_h)& (\undertilde{\varphi}{}_h,\nabla\times\undertilde{s}{}_h) & &&&=0 &
\\
&&({\rm div}\undertilde{\varphi}{}_h,q_h)&&&&=0.
\end{array}
\right.
\end{equation}
The stability and the convergence of the scheme are surveyed in the two lemmas below.
\begin{lemma}
Provided assumptions \textbf{A1}$\sim$\textbf{A3}, the problem \eqref{eq:dismxdvpB'} is well-posed on $U_h$.
\end{lemma}
\begin{lemma}\label{lem:errB}
Let $(m,\undertilde{u},\undertilde{\varphi},\undertilde{r},p)$ and $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,\undertilde{r}{}_h,p_h)$ be the solutions of \eqref{eq:mxdvpB'} and \eqref{eq:dismxdvpB'}, respectively.
\begin{enumerate}
\item $m_h=0$;
\item $\displaystyle\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{1,\Omega}+\|p-p_h\|_{0,\Omega}
\\
\leqslant C\inf_{(n_h,\undertilde{v}{}_h,\undertilde{\psi}{}_h,\undertilde{s}{}_h,q_h)\in U_h}[\|\undertilde{u}-\undertilde{v}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{\varphi}-\undertilde{\psi}{}_h\|_{1,\Omega}+\|\undertilde{r}-\undertilde{s}{}_h\|_{{\rm curl},\Omega}+\|p-q_h\|_{0,\Omega}]$;
\item $\displaystyle\|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}
\leqslant C\Big[
\inf_{\undertilde{v}{}_h,\undertilde{s}{}_h\in H_{h0}({\rm curl})}(\|\undertilde{u}-\undertilde{v}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{r}-\undertilde{s}{}_h\|_{{\rm curl},\Omega})+\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}\Big]
$
\end{enumerate}
\end{lemma}
\begin{proof}
The first item follows from {\bf A1}, and the second follows from the Cea lemma. Note that $(m,\undertilde{u},\undertilde{r})$ solves the problem
\begin{equation}\label{eq:reducedB}
\left\{
\begin{array}{cccll}
&&(\undertilde{r},\nabla n) & = 0 &\forall\,n\in H^1_{0}(\Omega)
\\
& (\undertilde{u},\undertilde{v})& -(\nabla\times \undertilde{r},\nabla\times\undertilde{v})& =(\undertilde{f},\undertilde{v})&\forall\,\undertilde{v}\in H_0({\rm curl},\Omega)
\\
(\nabla m,\undertilde{s})&-(\nabla\times\undertilde{u},\nabla\times\undertilde{s})&&=-(\undertilde{\varphi},\nabla\times\undertilde{s})&\forall\,\undertilde{s}\in H_0({\rm curl},\Omega),
\end{array}
\right.
\end{equation}
and $(m_h,\undertilde{u}{}_h,\undertilde{r}{}_h)$ solves the finite element problem:
\begin{equation}\label{eq:subproblemBdis}
\left\{
\begin{array}{cccll}
&&(\undertilde{r}{}_h,\nabla n_h) & = 0 &\forall\,n_h\in H^1_{h0}
\\
& (\undertilde{u}{}_h,\undertilde{v}{}_h)& -(\nabla\times\undertilde{r}{}_h,\nabla\times\undertilde{v}{}_h)& =(\undertilde{f},\undertilde{v}{}_h)&\forall\,\undertilde{v}{}_h\in H_{h0}({\rm curl})
\\
(\nabla m_h,\undertilde{s}{}_h)&-(\nabla\times\undertilde{u}{}_h,\nabla\times\undertilde{s}{}_h)&&=-(\undertilde{\varphi}{}_h,\nabla\times\undertilde{s}{}_h)&\forall\,\undertilde{s}{}_h\in H_{h0}({\rm curl}).
\end{array}
\right.
\end{equation}
The third item follows from the Ce\'a lemma and the second Strang lemma.
\end{proof}
Lemma \ref{lem:errB} reveals that the error $\|\undertilde{u}-\tilde{\undertilde{u}}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{r}-\tilde{\undertilde{r}}{}_h\|_{{\rm curl},\Omega}$ can be comparable with $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_0$. This hints us to use some bigger $H_{h0}({\rm curl})$ to expect higher accuracy of $\undertilde{u}$ and $\undertilde{r}$ than that of $\undertilde{\varphi}$ and $p$ for smooth $\undertilde{u}$ and $\undertilde{r}$.
\subsection{Examples of finite element quartos}
\subsubsection{Problem {\bf (A$'$)}}
For problem {\bf (A$'$)}, we choose
\begin{equation}
H^{1,a}_{h0}:=\mathcal{L}^2_{h0},\, H^a_{h0}({\rm curl}):=N^2_{h0},\, \undertilde{H}{}^1_{h0}:=(\mathcal{L}^2_{h0})^2,\, H^{1,b}_{h0}:=\mathcal{L}^1_{h0},\, H^b_{h0}({\rm curl}):=N^1_{h0},\, L^2_{h0}:=\mathcal{L}^1_{h}\cap L^2_0(\Omega).
\end{equation}
The assumptions {\bf A1}$\sim${\bf A3} can be verified. Particularly, {\bf A1} and {\bf A3} can be found in \cite{Boffi.D;Brezzi.F;Fortin.M2013}, and {\bf A2} can be found in \cite{Hiptmair.R2002,AFW2006,Kikuchi.F1989}. Its convergence then follows from Lemma \ref{lem:errA}. We here present a specific estimate on convex polyhedrons.
\begin{lemma}\label{lem:errestA}
Let $\Omega$ be a convex polyhedron and $\undertilde{f}\in\undertilde{L}^2$, ${\rm div}\undertilde{f}=0$. Let $(m,\undertilde{u},\undertilde{\varphi},p,\undertilde{r},g)\in V$ and $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,p_h,\undertilde{r}{}_h,g_h)\in V_h$ be the solutions of \eqref{eq:mxdvfA'} and \eqref{eq:dismxdvfA'}, respectively. Then
\begin{enumerate}
\item $g_h=0=g$ and $m_h=0=m$;
\item $\displaystyle\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}\leqslant Ch(\|\undertilde{r}\|_{1,\Omega}+\|{\rm curl}\undertilde{r}\|_{1,\Omega})\leqslant Ch\|\undertilde{f}\|_{0,\Omega}$;
\item $\displaystyle \|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{1,\Omega}+\|p-p_h\|_{0,\Omega}\leqslant Ch(\|\undertilde{\varphi}\|_{2,\Omega}+\|p\|_{1,\Omega}+\|\undertilde{r}\|_{1,\Omega}+\|{\rm curl}\undertilde{r}\|_{1,\Omega})\leqslant Ch\|\undertilde{f}\|_{0,\Omega}$;
\item $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_0\leqslant Ch^2\|\undertilde{f}\|_{0,\Omega}$;
\item $\displaystyle \|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega}\leqslant C(h^2(\|\undertilde{u}\|_{2,\Omega}+\|{\rm curl}\undertilde{u}\|_{2,\Omega})+\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}) \leqslant Ch^2\|\undertilde{f}\|_{0,\Omega}$.
\end{enumerate}
\end{lemma}
\begin{proof}
We only prove the estimate $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}$ by dual argument, and the remaining follows from Lemma \ref{lem:errA} directly.
Define $\hat{V}_h:=H^{1,b}_{h0}\times H^b_{h0}({\rm curl})\times \undertilde{H}{}^1_{h0}\times L^2_{h0}\times H^b_{h0}({\rm curl})\times H^{1,b}_{h0}$, and let $(\hat{m}_h,\hat{\undertilde{u}}{}_h,\hat{\undertilde{\varphi}}{}_h,\hat{p}_h,\hat{\undertilde{r}}{}_h,\hat{g}_h)\in \hat{V}_h$ be so that
\begin{equation}\label{eq:dualsystem}
\left\{
\begin{array}{lll}
a((\hat m_h,\hat\undertilde{u}{}_h,\hat\undertilde{\varphi}{}_h),(\hat n_h,\hat\undertilde{v}{}_h,\hat\undertilde{\psi}{}_h))&+b((\hat p_h,\hat\undertilde{r}{}_h,\hat g_h),(\hat n_h,\hat\undertilde{v}{}_h,\hat\undertilde{\psi}{}_h)) &= (\undertilde{f},\hat\undertilde{v}{}_h),
\\
b((\hat m_h,\hat\undertilde{u}{}_h,\hat\undertilde{\varphi}{}_h), (\hat q_h,\hat\undertilde{s}{}_h,\hat h_h)) &&=0.
\end{array}
\right.
\end{equation}
for any $(\hat n_h,\hat\undertilde{v}{}_h,\hat\undertilde{\psi}{}_h,\hat q_h,\hat\undertilde{s}{}_h,\hat h_h)\in\hat{V}_h$, where $a(\cdot,\cdot)$ and $b(\cdot,\cdot)$ follows the definitions \eqref{eq:aforA} and \eqref{eq:bforA}.Then $\hat{m}_h=\hat{g}_h=0$, and $(\hat{\undertilde{\varphi}}{}_h,\hat{p}_h,\hat{\undertilde{r}}{}_h)=({\undertilde{\varphi}}{}_h,{p}_h,{\undertilde{r}}{}_h)$. Let $(\tilde m,\tilde\undertilde{u},\tilde\undertilde{\varphi},\tilde p,\tilde\undertilde{r},\tilde g)\in V$ solve the variational problem
\begin{equation}\label{eq:dualsystem}
\left\{
\begin{array}{lll}
a((\tilde m,\tilde\undertilde{u},\tilde\undertilde{\varphi}),(n,\undertilde{v},\undertilde{\psi}))&+b((\tilde p,\tilde\undertilde{r},\tilde g),(n,\undertilde{v},\undertilde{\psi})) &= (\undertilde{\varphi}-\undertilde{\varphi}{}_h,\undertilde{\psi}),
\\
&& \quad (n,\undertilde{v},\undertilde{\psi})\in H^1_0(\Omega)\times H_0({\rm curl},\Omega)\times \undertilde{H}{}^1_0(\Omega)
\\
b((\tilde m,\tilde\undertilde{u},\tilde\undertilde{\varphi}), (q,\undertilde{s},h)) &&=0
\\
&&\quad (q,\undertilde{s},h)\in L^2_0(\Omega)\times H_0({\rm curl},\Omega)\times H^1_0(\Omega).
\end{array}
\right.
\end{equation}
Then $\tilde\undertilde{r}=\undertilde{0}$, $\tilde g=0$, $\tilde m=0$, and, by the same virtue as that of Theorem \ref{thm:regA},
$$
\|\tilde\undertilde{\varphi}\|_{2,\Omega}+\|\tilde p\|_{1,\Omega}+\|\tilde \undertilde{u}\|_{2,\Omega}+\|{\rm curl}\tilde\undertilde{u}\|_{2,\Omega}\leqslant C\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}.
$$
Thus substituting $(n,\undertilde{v},\undertilde{\psi},q,\undertilde{s},h)=(m-\hat m_h,\undertilde{u}-\hat\undertilde{u}{}_h,\undertilde{\varphi}-\undertilde{\varphi}{}_h,p-p_h,\undertilde{r}-\undertilde{r}{}_h,g-\hat g_h)$ into \eqref{eq:dualsystem}, we have
\begin{multline*}
\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}^2=a((\tilde m,\tilde\undertilde{u},\tilde\undertilde{\varphi}),(m-\hat m_h,\undertilde{u}-\hat\undertilde{u}{}_h,\undertilde{\varphi}-\undertilde{\varphi}{}_h))
\\
+b((\tilde p,\tilde\undertilde{r},\tilde g),(p-\hat p_h,\undertilde{r}-\undertilde{r}{}_h,g-\hat g_h))+b((\tilde m,\tilde\undertilde{u},\tilde\undertilde{\varphi}), (p-\hat p_h,\undertilde{r}-\undertilde{r}{}_h,g-\hat g_h)).
\end{multline*}
Further, for any $(\hat n_h,\hat\undertilde{v}{}_h,\hat\undertilde{\psi}{}_h,\hat q_h,\hat\undertilde{s}{}_h,\hat h_h)\in\hat{V}_h$, the orthogonality holds that
\begin{multline*}
\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}^2
=a((\tilde m-\hat n_h,\tilde\undertilde{u}-\hat\undertilde{v}{}_h,\tilde\undertilde{\varphi}-\hat\undertilde{\psi}{}_h),(m-\hat m_h,\undertilde{u}-\hat\undertilde{u}{}_h,\undertilde{\varphi}-\undertilde{\varphi}{}_h))+
\\
b((\tilde p-\hat q_h,\tilde\undertilde{r}-\hat\undertilde{s}{}_h,\tilde g-\hat h_h),(p-\hat p_h,\undertilde{r}-\undertilde{r}{}_h,g-\hat g_h))+b((\tilde m-\hat n_h,\tilde\undertilde{u}-\hat\undertilde{s}{}_h,\tilde\undertilde{\varphi}), (p-\hat p_h,\undertilde{r}-\undertilde{r}{}_h,g-\hat g_h)).
\end{multline*}
Thus
\begin{multline*}
\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}^2\leqslant C(\|\tilde m-\hat m_h\|_{0,\Omega}+\|\tilde\undertilde{u}-\hat\undertilde{u}{}_h\|_{{\rm curl},\Omega}+\|\tilde\undertilde{\varphi}-\hat\undertilde{\varphi}{}_h\|_{1,\Omega}+\|\tilde p-\hat p_h\|_{0,\Omega}+\|\tilde\undertilde{r}-\hat\undertilde{r}{}_h\|_{{\rm curl},\Omega})
\\
\times \inf_{(\hat n_h,\hat\undertilde{v}{}_h,\hat\undertilde{\psi}{}_h,\hat q_h,\hat\undertilde{s}{}_h,\hat h_h)\in\hat{V}_h}
(\|\tilde m-\hat n_h\|_{0,\Omega}+\|\tilde\undertilde{u}-\hat\undertilde{v}{}_h\|_{{\rm curl},\Omega}+\|\tilde\undertilde{\varphi}-\hat\undertilde{\psi}{}_h\|_{1,\Omega}+\|\tilde p-\hat q_h\|_{0,\Omega}+\|\tilde\undertilde{r}-\hat\undertilde{s}{}_h\|_{{\rm curl},\Omega}).
\end{multline*}
Then by finite element estimate,
\begin{equation}
\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}^2\leqslant Ch\|\undertilde{f}\|_{0,\Omega}\cdot h\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega},
\end{equation}
which leads to that $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}\leqslant Ch^2\|\undertilde{f}\|_{0,\Omega}$. This completes the proof.
\end{proof}
\begin{remark}
The convergence analysis with respect to the regularities of $\undertilde{u}$, $\undertilde{r}$, and etc. follows directly from Lemma \ref{lem:errA}. We here construct a convergence analysis with respect to $\|\undertilde{f}\|_{0,\Omega}$, which can exploit the regularity of solution functions to a full extent with economical complexity. When further $\undertilde{f}\in\undertilde{H}{}^1$, we can set $H^{1,b}_{h0}:=\mathcal{L}^2_{h0}$ and $H^b_{h0}({\rm curl}):=N^2_{h0}$, and obtain higher accuracy of $\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}$.
\end{remark}
\subsubsection{Problem {\bf (B$'$)}}
For problem {\bf (B$'$)}, we choose
\begin{equation}
H^1_{h0}:=\mathcal{L}^2_{h0},\, H_{h0}({\rm curl}):=N^2_{h0},\, \undertilde{H}{}^1_{h0}:=(\mathcal{L}^2_{h0})^2,\, L^2_{h0}:=\mathcal{L}^1_{h}\cap L^2_0(\Omega).
\end{equation}
The assumptions {\bf A1}$\sim${\bf A3} can be verified. For the convergence rate, we have the lemma below.
\begin{lemma}
Let $\Omega$ be a convex polyhedron and $\undertilde{f}\in\undertilde{L}^2$. Let $(m,\undertilde{u},\undertilde{\varphi},\undertilde{r},p)\in U$ and $(m_h,\undertilde{u}{}_h,\undertilde{\varphi}{}_h,\undertilde{r}{}_h,p_h)\in U_h$ be the solutions of \eqref{eq:mxdvpB'} and \eqref{eq:dismxdvpB'}, respectively. Then
\begin{enumerate}
\item $m_h=0=m$;
\item $\displaystyle \|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{1,\Omega}+\|p-p_h\|_{0,\Omega}\leqslant Ch\|\undertilde{f}\|_{0,\Omega}$, $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_0\leqslant Ch^2\|\undertilde{f}\|_{0,\Omega}$;
\item $\displaystyle\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega}\leqslant Ch\|\undertilde{f}\|_{0,\Omega}$;
\item if $\undertilde{f}\in \undertilde{H}{}^1(\Omega)$, then $\displaystyle \|\undertilde{u}-\undertilde{u}{}_h\|_{{\rm curl},\Omega}+\|\undertilde{r}-\undertilde{r}{}_h\|_{{\rm curl},\Omega}\leqslant Ch^2\|\undertilde{f}\|_{1,\Omega}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Similar to Lemma \ref{lem:errestA}, we only have to construct the estimate $\|\undertilde{\varphi}-\undertilde{\varphi}{}_h\|_{0,\Omega}$, which can be done by repeating the same dual argument as that in the proof of Lemma \ref{lem:errestA}, and we omit it here. The remaining follows from Lemma \ref{lem:errB} directly. The proof is completed.
\end{proof}
\section{Concluding remarks}
\label{sec:con}
In this paper, we study the fourth order curl problem, and develop for them amiable mixed schemes. We construct equivalent mixed formulation of the two variant model problems; the mixed formulations are stable on standard $H({\rm curl})$ and $H^1$ spaces. Regularity results are constructed for the mixed formulations, and then the primal ones. Existing finite element quartos that satisfy quite mild assumptions can then lead to stable and convergent discretisation scheme of the model problem. Some finite element examples which are optimal with respect to the regularity are given. The newly developed schemes are friendly in building, analysing and designing optimal solvers. The scheme can be implemented with various familiar finite element packages.
In the derivation of the equivalent formulations, we figure out the boundary condition that $\nabla\times\undertilde{u}$ has to satisfy. Particularly, $(\nabla\times\undertilde{u})\cdot\mathbf{n}=0$ is a condition which should be satisfied by the variable $\nabla\times\undertilde{u}$ essentially; also the condition can not be imposed weakly with the duality operated on $H({\rm curl},\Omega)$. This forces us to bring $\undertilde{H}{}^1_0(\Omega)$ which has bigger capacity for boundary conditions into discussion, and Lemma \ref{lem:h=cd} guarantees the equivalence.
As only spaces of low smoothness and of low degrees are involved, it is possible to construct finite element discretizations that are nested algebraically and topologically on nested grids; this will provide convenience in designing multilevel methods, and can be utilised in practice (c.f., e.g.,\cite{Chen.L;Hu.X;Wang.M;Xu.J2015,Chen.L2015,Zhang.S;Xi.Y;Ji.X2016,Li.Z;Zhang.S2016}). In this paper, we only consider two familiar variants boundary value problems. In many contexts, second order operators also appear in the boundary value problem, and parameters of various scale may appear in front of operators of different orders; see the model problem in \cite{Zheng.B;Hu.Q;Xu.J2011}. Designing parameter-robust discretisation is interesting and practically important, and will be discussed in future. Also, we focus ourselves on source problems in the present paper. The utilisation of the mixed scheme presented in this paper onto eigenvalue computation and analysis, especially in designing multilevel algorithm (c.f. \cite{Zhang.S;Xi.Y;Ji.X2016}), will be discussed in future.
| d17d0426eaa84b71b9b3078b761a47bf9eaaf9b2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{introduction}
Communication for omniscience (CO) is a problem proposed in \cite{Csiszar2004}. It is assumed that there is a group of users in the system and each of them observes a component of a discrete memoryless multiple source in private. The users can exchange their information over lossless broadcast channels so as to attain \textit{omniscience}, the state that each user obtains the total information in the entire multiple source in the system. The CO problem in \cite{Csiszar2004} is based on an asymptotic source model, where the communication rates could be real or fractional. Meanwhile, coded cooperative data exchange (CCDE) problem proposed in \cite{Roua2010} can be considered a non-asymptotic CO problem where the communication rates are required to be integral. By incorporating the idea of packet-splitting, the CCDE problem can be easily extended to an asymptotic setting.
Determining a rate vector that achieves omniscience with the minimum sum-rate is a fundamental problem in CO. Although the non-asymptotic CO problem has been frequently studied in the literature, there still does not exist an efficient algorithm for the asymptotic setting. The reasons are explained as follows. The submodularity of the CO problem has been shown in \cite{ChanMMI,ChanSuccessive,ChanSuccessiveIT,MiloIT2015,CourtIT2014,Ding2015ICT,Ding2015ISIT}. By designating a sum-rate, a submodular function minimization (SFM) algorithm can check whether the sum-rate is achievable for CO and/or return an achievable rate vector with the given sum-rate. Since the SFM algorithm completes in strongly polynomial time, the remaining problem is how to adapt the sum-rate to the minimum. This problem is not difficult for non-asymptotic setting since every adaptation should be integral. For example, the authors in \cite{MiloIT2015,CourtIT2014} proposed efficient adaptation algorithms for non-asymptotic CO problem, the complexity of which only grows logarithmically in the total amount of information in the system.
However, when considering the asymptotic setting, it is not clear how to choose the step size in each adaptation (Improper step sizes may result in an infinite loop). More specifically, even if we know that a sum-rate is over/below the optimum, it is not sure how much we should decrease/increase from the current estimation. On the other hand, the authors in \cite{MiloDivConq2011} proposed a divide-and-conquer (DV) algorithm for the asymptotic setting by repetitively running a decomposition algorithm (DA) in \cite{MinAveCost}. The idea is to first find the fundamental partition \cite{ChanMMI}, the one corresponds to the minimum sum-rate, and then iteratively break each non-singleton element into singletons so that each tuple in the optimal rate vector is determined. However, the DA algorithm is able to not only determine the fundamental partition but also return an optimal rate vector, which we will explain in this paper. Therefore, those further divisions of the fundamental partition in the DV algorithm are not necessary.
In this paper, we propose a modified decomposition algorithm (MDA) for solving the asymptotic CO problem based on the DA algorithm in \cite{MinAveCost}. The MDA algorithm starts with a lower estimation of the minimum sum-rate. In each iteration, the step size is determined based on the finest/minimum partition of a Dilworth truncation problem. We prove the optimality of the output rate vector and show that the estimation sequence converges monotonically upward to the minimum sum-rate. In addition, we propose a fusion method implementation of the coordinate-wise saturation capacity algorithm (CoordSatCapFus) for solving the Dilworth truncation problem. In the CoordSatCapFus algorithm, the SFM in each iteration is done over a fused user set with a cardinality smaller than the original one. We show that the MDA algorithm can reduce the cubic calls of SFM (in the DV algorithm) to quadratic calls of SFM. We do an experiment to show that the fusion method in the CoordSatCapFus algorithm contributes to a considerable reduction in computation complexity when the number of users grows. We also discuss how to solve the non-asymptotic CO problem by one more run of the CoordSatCapFus algorithm. Finally, we show how to choose a proper linear ordering to solve the minimum weighted sum-rate problem.
\section{System Model}
\label{sec:system}
Let $V$ with $\Card{V}>1$ be the finite set that contains the indices of all users in the system. We call $V$ the \textit{ground set}. Let $\RZ{V}=(\RZ{i}:i\in V)$ be a vector of discrete random variables indexed by $V$. For each $i\in V$, user $i$ can privately observe an $n$-sequence $\RZ{i}^n$ of the random source $\RZ{i}$ that is i.i.d.\ generated according to the joint distribution $P_{\RZ{V}}$. We allow users exchange their sources directly so as to let all users $i\in V$ recover the source sequence $\RZ{V}^n$. We consider both asymptotic and non-asymptotic models. In the asymptotic model, we will characterize the asymptotic behavior as the \emph{block length} $n$ goes to infinity. In non-asymptotic model, the communication rates are required to be integer-valued.
Let $\rv_V=(r_i:i\in V)$ be a rate (vector). We call $\rv_V$ an achievable rate if omniscience is possible by letting users communicate with the rates designated by $\rv_V$. Let $r$ be the function associated with $\rv_V$ such that $r(X)=\sum_{i\in X} r_i,\forall X \subseteq V$ with the convention $r(\emptyset)=0$. For $X,Y \subseteq V$, let $H(\RZ{X})$ be the amount of randomness in $\RZ{X}$ measured by Shannon entropy \cite{YeungITBook}
and $H(\RZ{X}|\RZ{Y})=H(\RZ{X \cup Y})-H(\RZ{Y})$ be the conditional entropy of $\RZ{X}$ given $\RZ{Y}$. In the rest of this paper, we simplify the notation $\RZ{X}$ to $X$. It is shown in \cite{Csiszar2004} that an achievable rate must satisfy the Slepian-Wolf constraints:
\begin{equation} \label{eq:SWConstrs}
r(X) \geq H(X|V\setminus X), \quad \forall X \subset V.
\end{equation}
The interpretation of the Slepian-Wolf constraint on $X$ is: To achieve CO, the total amount of information sent from user set $X$ should be at least complementary to total amount of information that is missing in user set $V \setminus X$. The set of all achievable rate vectors is
$$ \RRCO(V)=\Set{ \rv_V\in\Real^{|V|} \colon r(X) \geq H(X|V\setminus X),\forall X \subset V }. $$
\subsection{Asymptotic and No-asymptotic Models}
In an asymptotic CO model, the minimum sum-rate can be determined by the following linear programming (LP)
\begin{equation} \label{eq:MinSumRate}
\RACO(V)=\min\Set{r(V) \colon \rv_V\in \RRCO(V)}
\end{equation}
and the set of all optimal rates is
$$ \RRACO^*(V)=\Set{\rv_V\in \RRACO(V) \colon r(V)=\RACO(V)}. $$
In a non-asymptotic CO model, $H(X) \in \Z_+$ for all $X \subseteq V$ and the minimum sum-rate can be determined by the integer linear programming (ILP)
$\RNCO(V)=\min\Set{r(V) \colon \rv_V\in \RRCO(V) \cap \Z^{|V|} }$.
The optimal rate set is
$ \RRNCO^*(V)=\Set{\rv_V \in \RRCO(V) \cap \Z^{|V|} \colon r(V)=\RNCO(V)}$.
\subsection{Corresponding CCDE Systems}
CCDE is an example of CO, where the asymptotic model corresponds to the CCDE system that allows packet-splitting, while the non-asymptotic model corresponds to the CCDE system that does not allow packet-splitting. In CCDE, $\RZ{i}$ is the packet set that is obtained by user $i$, where each packet $\Has{j}$ belongs to a field $\F_q$. The users are geographically close to each other so that they can transmit linear combinations of their packet set via lossless wireless channels to help the others recover all packets in $\RZ{V}=\cup_{i\in V} \RZ{i}$. In this problem, the value of $H(X)$ can be obtained by counting the number of packets in $\RZ{X}$, i.e., $H(X)=|\RZ{X}|$ and $H(X|Y)=|\RZ{X \cup Y}|-|\RZ{Y}|$.
\begin{example} \label{ex:main}
Let $V=\{1,\dotsc,5\}$. Each user observes respectively
\begin{equation}
\begin{aligned}
\RZ{1} & = (\RW{a},\RW{c},\RW{e},\RW{f}), \\
\RZ{2} & = (\RW{a},\RW{d},\RW{h}), \\
\RZ{3} & = (\RW{b},\RW{c},\RW{e},\RW{f},\RW{g},\RW{h}), \\
\RZ{4} & = (\RW{a},\RW{c},\RW{f},\RW{g},\RW{h}), \\
\RZ{5} & = (\RW{b},\RW{d},\RW{f}),
\end{aligned} \nonumber
\end{equation}
where $\RW{j}$ is an independent uniformly distributed random bit. The users exchange their private observations to achieve the omniscience of $\RZ{V}=(\RW{a},\dotsc,\RW{h})$. In this system, $\RACO(V)= \frac{11}{2}$ and $\RNCO(V) = 6$. $\rv_V = (0,\frac{1}{2},2,\frac{5}{2},\frac{1}{2})$ is an optimal rate in $\RRACO^*(V)$ for asymptotic model, while $\rv_V=(0,1,2,3,0)$ is the optimal rate in $\RRNCO^*(V)$ for non-asymptotic model. The method to implement rate $\rv_V = (0,\frac{1}{2},2,\frac{5}{2},\frac{1}{2})$ is to let users divide each packets into two chunks of equal length and transmit according to rate $(0,1,4,5,1)$ with each tuple denotes the number of packet chunks. $(0,\frac{1}{2},2,\frac{5}{2},\frac{1}{2})$ and $\frac{11}{2}$ are the normalized rate and sum-rate, respectively.
\end{example}
\section{Preliminaries}
In this section, we list some existing results derived previously in \cite{ChanMMI,ChanSuccessive,ChanSuccessiveIT,CourtIT2014,MiloIT2015,Ding2015ICT,Ding2015Game,Ding2015ISIT,Ding2015NetCod,FujishigePolyEntropy,Fujishige2005} for CO.
\subsection{Submodularity and Nonemptiness of Base Polyhedron}
It is shown in \cite{FujishigePolyEntropy,Fujishige2005} that the entropy function $H$ is the rank function of a polymatroid, i.e., it is (a) normalized: $H(\emptyset) = 0 $; (b) monotonic: $H(X) \geq H(Y)$ for all $X,Y \subseteq V$ such that $Y \subseteq X$; (c) submodular:
\begin{equation} \label{eq:SubMIneq}
H(X) + H(Y) \geq H(X \cap Y) + H(X \cup Y)
\end{equation}
for all $X,Y \subseteq V$. For $\alpha\in\RealP$, define the set function $\Fu{\alpha}$ as
$$ \Fu{\alpha}(X)=\begin{cases} H(X|V\setminus X) & X \subset V \\ \alpha & X=V \end{cases}. $$
Let $ \FuHash{\alpha}(X)=\Fu{\alpha}(V)-\Fu{\alpha}(V \setminus X)=\alpha-\Fu{\alpha}(V \setminus X), \forall X \subseteq V$
be the \textit{dual set function} of $\Fu{\alpha}$. It is shown in \cite{ChanMMI,Ding2015ISIT,Ding2015NetCod} that $\FuHash{\alpha}$ is intersecting submodular, i.e., $\FuHash{\alpha}(X) + \FuHash{\alpha}(Y) \geq \FuHash{\alpha}(X \cap Y) + \FuHash{\alpha}(X \cup Y)$ for all $X,Y \subseteq V$ such that $X \cap Y \neq \emptyset$. The polyhedron and base polyhedron of $\FuHash{\alpha}$ are respectively
\begin{equation}
\begin{aligned}
& P(\FuHash{\alpha},\leq) = \Set{\rv_V\in\Real^{|V|} \colon r(X) \leq \FuHash{\alpha}(X),\forall X \subseteq V}, \\
& B(\FuHash{\alpha},\leq) = \Set{\rv_V \in P(\FuHash{\alpha},\leq) \colon r(V) = \FuHash{\alpha}(V)}. \\
\end{aligned} \nonumber
\end{equation}
It is shown in \cite{Ding2015ICT,Ding2015ISIT,Ding2015Game} that $B(\FuHash{\alpha},\leq) = \Set{ \rv_V \in \RRCO(V) \colon r(V) = \alpha}$, i.e., $B(\FuHash{\alpha},\leq)$ denotes the set of all achievable rates with sum-rate equal to $\alpha$, and $B(\FuHash{\alpha},\leq) \neq \emptyset$ if and only if $\alpha \geq \RACO(V)$. In addition, $B(\FuHash{\RACO(V)},\leq) = \RRACO^*(V)$ and $B(\FuHash{\RNCO(V)},\leq) \cap \Z^{|V|} = \RRNCO^*(V)$.
Denote $\Pi(V)$ the set that contains all possible partitions of $V$ and $\Pi'(V)=\Pi(V)\setminus\{V\}$. For $\Pat \in \Pi(V)$, let $\FuHash{\alpha}[\Pat] = \sum_{X \in \Pat} \FuHash{\alpha}(X)$. The Dilworth truncation of $\FuHash{\alpha}$ is \cite{Dilworth1944}
\begin{equation} \label{eq:Dilworth}
\FuHashHat{\alpha}(X) = \min_{\Pat\in \Pi(X)} \FuHash{\alpha}[\Pat] , \quad \forall X \subseteq V.
\end{equation}
If $\alpha \geq \RACO(V)$, $\FuHashHat{\alpha}$ is submodular with $\FuHashHat{\alpha}(V)=\alpha$ and $B(\FuHashHat{\alpha},\leq)=B(\FuHash{\alpha},\leq)$ \cite[Lemma IV.7]{Ding2015Game}.
\subsection{Minimum Sum-rate and Fundamental Partition}
The authors in \cite{Ding2015ISIT,Ding2015Game} show that
\begin{equation} \label{eq:RACO}
\RACO(V) = \max_{\Pat \in \Pi'(V)} \sum_{X \in \Pat} \frac{ H(V) - H(X) }{|\Pat|-1}
\end{equation}
and $\RNCO(V) = \lceil \RACO(V) \rceil$. Meanwhile, in the studies on secrecy capacity in \cite{ChanMMI,ChanSuccessive,ChanSuccessiveIT}, it is shown that maximum secrecy capacity in $V$ equals to the multivariate mutual information (MMI) $I(V)$, which has a dual relationship with $\RACO(V)$: $\RACO(V)=H(V)-I(V)$, and the finest/minimal maximizer of \eqref{eq:RACO} is called the \textit{fundamental partition} and denoted by $\Pat^*$.
\begin{algorithm} [t]
\label{algo:MDA}
\small
\SetAlgoLined
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\SetKwFor{For}{for}{do}{endfor}
\SetKwRepeat{Repeat}{repeat}{until}
\SetKwIF{If}{ElseIf}{Else}{if}{then}{else if}{else}{endif}
\BlankLine
\Input{the ground set $V$, an oracle that returns the value of $H(X)$ for a given $X \subseteq V$ and a linear ordering $\Phi = (\phi_1,\dotsc,\phi_{|V|})$}
\Output{$\rv_V$ which is a rate vector in the base polyhedron $B(\FuHashHat{\RACO(V)},\leq)$, $\Pat^*$ which is the fundamental partition and $\alpha$ which equals to $\RACO(V)$ }
\BlankLine
initiate $\Pat \leftarrow \{\Set{i} \colon i\in V\}$ and $\alpha \leftarrow \sum_{X \in \Pat}\frac{ H(V) - H(X) }{ |\Pat| - 1 } $ \;
$(\rv_V,\Pat^*) \leftarrow \text{CoordSatCapFus}(V,H,\alpha,\Phi)$ \;
\While{$\Pat^* \neq \Pat$}{
update $\Pat \leftarrow \Pat^*$ and $\alpha \leftarrow \sum_{X \in \Pat^*}\frac{ H(V) - H(X) }{ |\Pat^*| - 1 } $\;
$(\rv_V,\Pat^*) \leftarrow \text{CoordSatCapFus}(V,H,\alpha,\Phi)$ \;
}
return $\rv_V$, $\Pat^*$ and $\alpha$\;
\caption{Modified Decomposition Algorithm (MDA) }
\end{algorithm}
\section{Algorithm} \label{sec:algo}
In this section, we propose a MDA algorithm, the modified version of the DA algorithm in \cite{MinAveCost}, in Algorithm 1 for solving the asymptotic CO problem and show how to extend it to solve the non-asymptotic one. The MDA algorithm starts with $\alpha$, a lower estimation of $\RACO(V)$, and iteratively updates it by the minimal/finest minimizer of the Dilworth truncation problem $\FuHashHat{\alpha} = \min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ until it reaches the optimal one. The finest minimizer of the Dilworth truncation problem and a rate vector in the base polyhedron $B(\FuHashHat{\alpha},\leq)$ are determined by the CoordSatCapFus algorithm in Algorithm 2. The CoordSatCapFus algorithm is a fusion method to implement the coordinate-wise saturation capacity (CoordSatCap) algorithm that is proposed in \cite{Fujishige2005} and adopted in \cite{MinAveCost} for the Dilworth truncation problem. We list the notations in Algorithms 1 and 2 below.
Let $\chi_X$ be the characteristic vector of the subset $X \subseteq V$. We shorten the notation $\chi_{\Set{i}}$ to $\chi_i$ for a singleton subset of $V$. Let $ \Phi = (\phi_1,\dotsc,\phi_{|V|})$ be a linear ordering of $V$. For example, $\Phi = (2,3,1,4)$ is a linear ordering of $V = \Set{1,\dotsc,4}$. In Section~\ref{sec:MinWeightSumRate}, we will show that by choosing a proper linear ordering of $V$ the output rate $\rv_V$ of Algorithm 1 also minimizes a weighted sum-rate objective function. For $U \subseteq \Pat$ where $\Pat$ is some partition in $\Pi(V)$, denote $\tilde{U}=\cup_{X \in U} X$, i.e., $U$ is a fusion of all the subsets in $U$ into one subset of $V$. For example, for $U = \Set{\Set{1,3},\Set{2,4},\Set{5},\Set{6}} \subset \Set{\Set{1,3},\Set{2,4},\Set{5},\Set{6},\Set{7}} \in \Pi({\Set{1,\dotsc,7}})$, we have $\tilde{U} = \Set{1,\dotsc,6}$. By using these notations, we propose the MDA algorithm for the asymptotic CO problem and show that they can be easily extended to solve the non-asymptotic CO problem as follows.
\begin{algorithm} [t]
\label{algo:PolyRank}
\small
\SetAlgoLined
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\SetKwFor{For}{for}{do}{endfor}
\SetKwRepeat{Repeat}{repeat}{until}
\SetKwIF{If}{ElseIf}{Else}{if}{then}{else if}{else}{endif}
\BlankLine
\Input{the ground set $V$, an oracle that returns the value of $H(X)$ for a given $X \subseteq V$, $\alpha$ which is an estimation of $\RACO(V)$ and a linear ordering $\Phi = ( \phi_1,\dotsc,\phi_{|V|})$}
\Output{$\rv_V$ which is a rate vector in $B(\FuHashHat{\alpha},\leq)$ and $\Pat^*$ which is the minimal/finest minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ }
\BlankLine
$\rv_V \leftarrow (\alpha - H(V)) \chi_V$ \tcp*{$\rv \in P(\FuHash{\alpha},\leq)$ by doing so.}
initiate $ r_{\phi_1} \leftarrow \FuHash{\alpha}(\Set{\phi_1})$ and $\Pat^* \leftarrow \Set{\Set{\phi_1}}$\;
\For{$i=2$ \emph{\KwTo} $|V|$}{
determine the saturation capacity
$$ \hat{\xi} \leftarrow \min\Set{ \FuHash{\alpha}(\Set{\phi_i} \cup \tilde{U}) - r(\Set{\phi_i} \cup \tilde{U}) \colon U \subseteq \Pat^*} $$
and the minimal/smallest minimizer $U^*$\;
$U_{\phi_i}^* \leftarrow U^* \cup \Set{\phi_i}$\;
$ \rv_V \leftarrow \rv_V + \hat{\xi} \chi_{\phi_i}$\;
\tcc{ merge/fuse all subsets in $\Pat^*$ that intersect with $\tilde{U}_{\phi_i}^*$ into one subset $\tilde{U}_{\phi_i}^* \cup \tilde{\X}$}
$\X \leftarrow \Set{X \in \Pat^* \colon X \cap \tilde{U}_{\phi_i}^* \neq \emptyset}$\;
$\Pat^* \leftarrow (\Pat^* \setminus \X) \cup \Set{ \tilde{U}_{\phi_i}^* \cup \tilde{\X} }$\;
}
return $\rv_V$ and $\Pat^*$\;
\caption{Coordinate-wise Saturation Capacity Algorithm by Fusion Method (CoordSatCapFus)}
\end{algorithm}
\subsection{Asymptotic Model}
The optimality of the MDA algorithm for the asymptotic setting is summarized in the following theorem with the proof in Appendix~\ref{app:theo:main}, where every step in the CoordSatCapFus algorithm is explained.
\begin{theorem} \label{theo:main}
The MDA algorithm outputs the minimum sum-rate $\RACO(V)$, the fundamental partition $\Pat^*$ and an optimal rate $\rv_V \in \RRACO(V)$. The estimation of $\RACO(V)$, $\alpha$, converges monotonically upward to $\RACO(V)$.
\end{theorem}
\begin{example} \label{ex:AlgorithmACO}
For the system in Example~\ref{ex:main}, we start the MDA algorithm with singleton partition $\Pat=\Set{\Set{1},\dotsc,\Set{5}}$ and $\alpha = \sum_{i \in V}\frac{ H(V) - H(\Set{i}) }{ |V| - 1 } = \frac{19}{4}$. Let the linear ordering be $\Phi = (4,3,2,5,1)$. By calling the CoordSatCapFus algorithm, we have the following results.
We initiate $\rv_V = (\alpha - H(V)) \chi_V = (-\frac{13}{4},\dotsc,-\frac{13}{4})$ and set $\Pat^*=\Set{\Set{4}}$ and $r_4 = \FuHash{19/4}(\Set{4}) = \frac{7}{4}$ so that $\rv_V = (-\frac{13}{4},-\frac{13}{4},-\frac{13}{4},\frac{7}{4},-\frac{13}{4})$.
\begin{itemize}
\item For $\phi_2=3$, the values of $\FuHash{\alpha}(\Set{\phi_2} \cup \tilde{U}) - r(\Set{\phi_2} \cup \tilde{U})$ for all $U \subseteq \Pat^* = \Set{\Set{4}}$ are
$$ \FuHash{19/4}(\Set{3}) - r(\Set{3}) = 6 , \FuHash{19/4}(\Set{3,4}) - r(\Set{3,4}) = 21/4.$$
So, the saturation capacity $\hat{\xi} = 21/4$, the minimal minimizer $U^* = \Set{\Set{4}}$ and $U_4^* = \Set{\Set{3},\Set{4}}$. We update to $r_3 = -\frac{13}{4} + \frac{21}{4} = 2$ so that $\rv_V = (-\frac{13}{4},-\frac{13}{4},2,\frac{7}{4},-\frac{13}{4})$. We have only one element $\Set{4} \in \Pat^*$ such that $\tilde{U}_{4}^* \cap \Set{4} \neq \emptyset$. So, $\X = \Set{\Set{4}}$ and $\tilde{U}_{\phi_i}^* \cup \tilde{\X} = \Set{3,4}$. We update to $\Pat^* = \Set{\Set{3,4}}$.
\item For $\phi_3=2$, the values of $\FuHash{\alpha}(\Set{\phi_3} \cup \tilde{U}) - r(\Set{\phi_3} \cup \tilde{U})$ for all $U \subseteq \Pat^* = \Set{\Set{3,4}}$ are
\begin{equation}
\begin{aligned}
&\FuHash{19/4}(\Set{2}) - r(\Set{2}) = 3 , \\ &\FuHash{19/4}(\Set{2,3,4}) - r(\Set{2,3,4}) = 17/4.
\end{aligned} \nonumber
\end{equation}
We have $\hat{\xi} = 3$ and $U_{2}^* = \Set{\Set{2}}$. We update to $\rv_V = (-\frac{13}{4},-\frac{1}{4},2,\frac{7}{4},-\frac{13}{4})$. Since $\tilde{U}_{2}^* \cap X = \emptyset, \forall X \in \Pat$, we have $\X = \emptyset$ and $\Pat^* = \Set{\Set{3,4},\Set{2}}$.
\item For $\phi_4 = 5$, we have $\hat{\xi}=3$, $U_{5}^*=\Set{\Set{5}}$ and $\X=\emptyset$. We update to $\rv_V = (-\frac{13}{4},-\frac{1}{4},2,\frac{7}{4},-\frac{1}{4})$ and $\Pat^* = \Set{\Set{3,4},\Set{2},\Set{5}}$.
\item For $\phi_5=1$, we have $\hat{\xi} = \frac{13}{4}$, $U_1^* = \Set{\Set{3,4},\Set{1}}$ and $\X = \Set{\Set{3,4}}$. Therefore, the CoordSatCapFus algorithm terminates with $\rv_V = (0,-\frac{1}{4},2,\frac{7}{4},-\frac{1}{4})$ and $\Pat^*=\Set{\Set{1,3,4},\Set{2},\Set{5}}$.
\end{itemize}
Since $\Pat \neq \Pat^*$, we continue the iteration in the MDA algorithm. In the second iteration, we have $\Pat=\Set{\Set{1,3,4},\Set{2},\Set{5}}$ and $\alpha = \frac{11}{2}$. The CoordSatCapFus algorithm returns $\rv_V = (0,\frac{1}{2},2,\frac{5}{2},\frac{1}{2})$ and $\Pat^* = \Set{\Set{1,3,4},\Set{5},\Set{2}}$. The MDA algorithm terminates since $\Pat = \Pat^*$. One can show that the outputs $\rv_V$, $\Pat^*$ and $\alpha$ are respectively an optimal rate in $\RRACO^*(V)$, the fundamental partition and the minimum sum-rate $\RACO(V)$ for asymptotic model. We plot the value of $\alpha$ in each iteration, or the estimation sequence of $\RACO(V)$, in Fig.~\ref{fig:Converge}. It can be shown that $\alpha$ converges monotonically upward to $\RACO(V)$.
\end{example}
\begin{figure}[tbp]
\centering
\scalebox{0.7}{\input{figures/Converge.tex}}
\caption{The estimation sequence of $\RACO(V)$, i.e., the value of $\alpha$ in each iteration, when the MDA algorithm is applied to the system in Example~\ref{ex:AlgorithmACO}.}
\label{fig:Converge}
\end{figure}
The CoordSatCap algorithm is one of the standard tools for solving the Dilworth truncation problem in the literature, e.g., \cite{MinAveCost}. It is also used in \cite{MiloIT2015,CourtIT2014} to determine an optimal rate vector in $\RRNCO^*(V)$ and/or checking whether a sum-rate $\alpha$ is achievable for non-asymptotic setting.\footnote{If the sum-rate $\alpha$ is not achievable, the rate $\rv_V \in B(\FuHashHat{\alpha},\leq)$ returned by the CoordSatCap algorithm has $r(V)$ strictly less than $\alpha$.} But, in these works, the CoordSatCap algorithm is implemented on the original user set instead of a fused one. For example, in \cite{MiloIT2015,CourtIT2014}, the the saturation capacity $\hat{\xi}$ is determined by the SFM problem
\begin{equation} \label{eq:SatCapOld}
\min\Set{ \FuHash{\alpha}(X) - r(X) \mid \phi_i \in X \subseteq V_i},
\end{equation}
where $V_{i} = \Set{\phi_1,\dotsc,\phi_i}$. Problem \eqref{eq:SatCapOld} can be solved in $O(\SFM(|V_{i-1}|))$ time, where $\SFM(|V|)$ denotes the complexity of an SFM algorithm for a set function defined on $2^{V}$. On the contrary, the corresponding SFM problem
\begin{equation} \label{eq:SatCapFus}
\min\Set{ \FuHash{\alpha}(\Set{\phi_i} \cup \tilde{U}) - r(\Set{\phi_i} \cup \tilde{U}) \colon U \subseteq \Pat^*}
\end{equation}
in step 4 in the CoordSatCapFus algorithm is done over $\Pat^*$, a fused/merged user sets of $V_{i-1}$ that is obtained by steps 8 and 9 in the previous iterations. Here, the objective function in \eqref{eq:SatCapFus} is submodular on $2^{\Pat^*}$. Problem \eqref{eq:SatCapFus} can be solved in $\SFM(|\Pat^*|)$ time. Since $|\Pat^*| \leq |V_2|$, \eqref{eq:SatCapFus} is less complex than \eqref{eq:SatCapOld}. For example, in the first iteration of the MDA algorithm when $\phi_3 = 2$ in Example~\ref{ex:AlgorithmACO}, We have $\Pat^*=\Set{\Set{3,4}}$ and $V_2 = \Set{3,4}$ such that $|\Pat^*|<|V_2|$. Problem~\eqref{eq:SatCapFus} completes in $O(\SFM(1))$ time, while problem \eqref{eq:SatCapOld} completes in $O(\SFM(2))$ time.\footnote{In the case when $|V|=1$, SFM reduces to comparison between two possible sets, empty and ground sets, i.e., it is not necessary to call the SFM algorithm. This example just shows the difference in complexity. } See the experimental results in Section~\ref{sec:Complexity}.
\subsection{Non-asymptotic Model} \label{sec:NCO}
The algorithms in \cite{MiloIT2015,CourtIT2014} for non-asymptotic CO model can adjust $\alpha$ on the nonnegative integer grid until it finally reaches $\RNCO(V)$, where the CoordSatCap can be replaced by the CoordSatCapFus algorithm which is less complex. See experimental results in Section~\ref{sec:Complexity}.
In fact, the value of $\RNCO(V)$ and an optimal rate in $\RRNCO^*(V)$ can be determined by one more call of the CoordSatCapFus algorithm after solving the asymptotic CO problem. Let $\RACO(V)$ be the asymptotic minimum sum-rate determined by the MDA algorithm. We know automatically $\RNCO(V) = \lceil \RACO(V) \rceil$. By calling the CoordSatCapFus algorithm with input $\alpha = \RNCO(V)$, we can determine the value of an optimal rate in $B(\FuHashHat{\RNCO(V)},\leq) \cap \Z^{|V|} = \RRNCO^*(V)$. The integrality of this optimal vector is shown in Section~\ref{sec:MinWeightSumRate}.
\begin{example} \label{ex:AlgorithmNCO}
Assume that we get $\RACO(V) = \frac{11}{2}$ in Example~\ref{ex:AlgorithmACO}. Then, $\RNCO(V) = \lceil \RACO(V) \rceil = 6$. By calling
$$ (\rv_V,\Pat^*) \leftarrow \text{CoordSatCapFus}(V,H,\RNCO(V),\Phi),$$
we have $\rv_V=(0,1,2,3,0)$ for linear ordering $\Phi = (4,3,2,5,1)$ and $\Pat^* = \Set{\Set{1,2,3,4,5}}$,\footnote{For $\RNCO(V) > \RACO(V) $, the minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\RNCO(V)}[\Pat]$ is uniquely $\Set{V}$ \cite{Narayanan1991PLP}.} where $\rv_V$ is an optimal rate in $\RRNCO^*(V)$ for non-asymptotic model.
\end{example}
\section{Minimum Weighted Sum-rate Problem} \label{sec:MinWeightSumRate}
Let $\wv_V = (w_i \colon i \in V) \in \RealP^{|V|}$ and $\wv_V^{\intercal} \rv_V = \sum_{i \in V} w_i r_i$. We say that $\Phi = (\phi_1,\dotsc,\phi_{|V|})$ is a linear ordering that is consistent with $\wv_V$ if $w_{\phi_1} \leq w_{\phi_2} \leq \dotsc \leq w_{\phi_{|V|}}$.
\begin{theorem}
Let $\Phi$ be the linear ordering consistent with $\wv_V$. The optimal rate $\rv_V$ returned by the MDA algorithm for asymptotic model is the minimizer of $\min \Set{ \wv_V^{\intercal} \rv_V \colon \rv_V \in \RRACO^*(V) }$; The optimal rate $\rv_V$ returned by $\text{CoordSatCapFus}(V,H,\lceil \RACO(V) \rceil,\Phi)$ for asymptotic model is the minimizer of $ \min \Set{ \wv_V^{\intercal} \rv_V \colon \rv_V \in \RRNCO^*(V) } $.
\end{theorem}
\begin{IEEEproof}
In the last iteration of the MDA algorithm, we call the CoordSatCapFus algorithm by inputting $\alpha=\RACO(V)$. The Dilworth truncation $\FuHashHat{\RACO(V)}$ is a polymatroid rank function \cite{ChanMMI}. Let $\text{EX}(\FuHashHat{\RACO(V)})$ be the set that contains all extreme points, or vertices, of the base polyhedron $B(\FuHashHat{\RACO(V)},\leq)$. We have the initial point $\rv_V = (\alpha - H(V)) \chi_V \leq \rv_V^\prime, \forall \rv_V^\prime \in \text{EX}(\FuHashHat{\RACO(V)})$.\footnote{$\rv_V \leq \rv_V^\prime, \forall \rv_V^\prime \in \text{EX}(f)$ is a tighter condition than $\rv_V \in P(f,\leq)$.} So, the CoordSatCapFus algorithm necessarily returns an extreme point in $B(\FuHashHat{\RACO(V)},\leq)$ which minimizes $\min \Set{ \wv_V^{\intercal} \rv_V \colon \rv_V \in \RRACO^*(V) }$\cite{Fujishige2005}. In the same way, we can prove the claim for the non-asymptotic model. In addition, $\FuHash{\RNCO(V)}$ is integer-valued. So is $\FuHashHat{\RNCO(V)}$. Therefore, all extreme points in $B(\FuHashHat{\RNCO(V)},\leq)$ are integral.
\end{IEEEproof}
For example, one can show that $\rv_V = (0,\frac{1}{2},2,\frac{5}{2},\frac{1}{2})$ in Example~\ref{ex:AlgorithmACO} and $\rv_V=(0,1,2,3,0)$ in Example~\ref{ex:AlgorithmNCO} are the minimum weighted sum-rate vector in $\RRACO^*(V)$ and $\RRNCO^*(V)$, respectively, where the weight $\wv_V$ is the one that linear ordering $\Phi = (4,3,2,5,1)$ is consistent with, e.g., $\wv_V = (4,0.5,0.5,0.3,3.3)$.
Note, any linear ordering is consistent with $\wv_V=(1,\dotsc,1)$, i.e., if the problem is just to determine the minimum sum-rate and an optimal rate vector, the linear ordering can be arbitrarily chosen.
\section{Complexity} \label{sec:Complexity}
The authors in \cite{MiloDivConq2011} proposed a divide-and-conquer (DV) algorithm for the asymptotic CO problem. The idea is to directly apply the DA algorithm in \cite{MinAveCost} to determine the fundamental partition and iteratively break each non-singleton subsets in it into singletons to determine each tuple in the optimal rate. Since the DA algorithm completes in $O(|V|^2\cdot\SFM(|V|))$ time, the complexity of the DV algorithm is upper bounded by $O(|V|^3\cdot\SFM(|V|))$. The complexity of the MDA algorithm is upper bounded by $O(|V|^2\cdot\SFM(|V|))$,\footnote{The complexity of the CoordSatCapFus algorithm based on~\eqref{eq:SatCapFus} in the worst case is the same as the CoordSatCap algorithm based on~\eqref{eq:SatCapOld}. The worst case is when $\Pat^*=\Set{\Set{\phi_1},\dotsc,\Set{\phi_i}}$ for all $i$ in the CoordSatCapFus algorithm. In the DA algorithm in \cite{MinAveCost}, the CoorSatCap algorithm is implemented for solving the Dilworth truncation problem. Therefore the complexity of MDA algorithm is upper bounded by $O(|V|^2\cdot\SFM(|V|))$. } which is lower than the DV algorithm.
Let $|V|$ be the size of the SFM problem with complexity $\SFM(|V|)$. As aforementioned, although the numbers of calls of SFM algorithm are the same, the size of each SFM problem in the CoordSatCapFus algorithm based on \eqref{eq:SatCapFus} is less than that in the CoordSatCap algorithm based on \eqref{eq:SatCapOld} in general. We do an experiment to compare the complexity of these two algorithms. Let $H(V)$ be fixed to $50$ and $|V|$ vary from $5$ to $30$. For each value of $|V|$, we repeat the procedure for 20 times: (a) randomly generate a CO system; (b) apply the MDA algorithm twice, one calls the CoordSatCapFus algorithm and the other calls the CoordSatCap algorithm. We record overall/summed size of the SFM algorithm in each run of the MDA algorithm and average over the repetitions. The results are shown in Fig.~\ref{fig:Complexity}. It can be seen that by implementing the CoordSatCapFus algorithm, there is a considerable reduction in complexity when the size of user set $|V|$ grows.
\begin{figure}[tbp]
\centering
\scalebox{0.7}{\input{figures/Complexity.tex}}
\caption{The size of SFM problem over repetitions in the experiment in Section~\ref{sec:Complexity}, where $H(V)$ is fixed to $50$ and $|V|$ varies from $5$ to $50$.}
\label{fig:Complexity}
\end{figure}
\section{Conclusion}
We proposed an MDA algorithm for determining the minimum sum-rate and a corresponding optimal rate for the asymptotic CO problem. The MDA algorithm mainly proposed an idea on how to update the minimum sum-rate estimation: A closer estimation to the optimum could be obtained by the minimal/finest minimizer of a Dilworth truncation problem based on the current estimation. We also proposed a CoordSatCapFus algorithm to solve the Dilworth truncation problem which was less complex than the original CoordSatCap algorithm. We discussed how to extend the MDA algorithm to solve the non-asymptotic problem and how to choose a proper linear ordering of the user set to solve a minimum weighted sum-rate problem.
\appendices
\section{Proof of Theorem~\ref{theo:main}} \label{app:theo:main}
In \cite{Narayanan1991PLP,MinAveCost}, the authors proposed a DA for determining the principal partition sequence (PSP) for a clustering problem. Since the fundamental partition is one of the partitions in PSP \cite{ChanMMI,ChanInfoCluster}, we adapt DA to MDA to just determine the fundamental partition. A similar approach can be found in \cite{ChanInfoCluster}. Based on the studies in \cite{Narayanan1991PLP,ChanInfoCluster}, if the CoordSatCapFus algorithm is able to determine the minimum and the minimal/finest minimizer of the Dilworth truncation problem $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ for a given value of $\alpha$, the MDA algorithm outputs $\RACO(V)$, the fundamental partition and an optimal rate $\rv_V \in \RRACO^*(V) = B(\FuHashHat{\RACO(V)},\leq)$. In addition, the value of $\alpha$ of the MDA algorithm converges monotonically upward to $\RACO(V)$.
Now, we show that CoordSatCapFus algorithm determines the finest minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$. For $\FuHash{\alpha}$, Consider the (original/general) CoordSatCap algorithm \cite{Fujishige2005}:
\begin{enumerate}[step 1:]
\item Initiate $\rv_V$ such that $\rv_V \in P(\FuHash{\alpha},\leq)$;
\item For each dimension $i \in \Set{1,\dotsc,|V|}$, do $ \rv \leftarrow \rv + \hat{\xi} \chi_{\phi_i}$, where $\hat{\xi}$ is the \text{saturation capacity}
\begin{equation} \label{eq:SatCap}
\hat{\xi} = \min\Set{\FuHash{\alpha}(X) - r(X) \colon \phi_i \in X \subseteq V}.
\end{equation}
\end{enumerate}
$\hat{\xi}$ in \eqref{eq:SatCap} is the maximum increment in $r_{\phi_i}$ such that the resulting $\rv_V$ is still in $P(\FuHash{\alpha},\leq)$, hence the name saturation capacity. Due to the intersecting submodularity of $\FuHash{\alpha}$, \eqref{eq:SatCap} is an SFM problem and the CoordSatCap algorithm finally updates $\rv_V$ to a vector/rate in $B(\FuHashHat{\alpha},\leq)$ with $r(V) = \FuHashHat{\alpha}(V)$.
The minimal minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ is determined as follows. Let $\hat{X}_{\phi_i}$ be the minimal minimizer of \eqref{eq:SatCap} for dimension $\phi_i$. By iteratively merging dimensions $\phi_i,\phi_j \in V$ such that $\phi_i \in \hat{X}_{\phi_j}$ until there is no such pair left, we can determine the finest partition in $\Pi(V)$ that minimizes $\FuHash{\alpha}[\Pat]$ \cite{Bilxby1985,Fujishige2005,MinAveCost}.\footnote{The minimal minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ corresponds to the minimal separators of a submodular system with the rank function being $\FuHashHat{\alpha}$. Define the partial order $\preceq$ as $\phi_i \preceq \phi_j$ if $\phi_i \in \hat{X}_{\phi_j}$. Let $G(V,E)$ be the digraph with the edge set constituted by edges $e_{\phi_i,\phi_j} \in E$ if $\phi_i \preceq \phi_j$. The minimal separators are the strongly connected components of the underlying undirected graph of $G(V,E)$. The procedure that updates $\Pat^*$ in Appendix~\ref{app:theo:main} is exactly the one that determines these minimal separators. For more details, we refer the reader to \cite{Bilxby1985,Fujishige2005}.} The implementation is as follows. Initiate $\Pat^*=\Set{\Set{\phi_i} \colon i \in V}$ at the beginning of the CoordSatCap algorithm. After obtaining each $\hat{X}_{\phi_i}$ for $i$ in step 2, do the followings:
\begin{itemize}
\item find all elements in $\Pat^*$ that intersect with $\hat{X}_{\phi_i}$, i.e., determine $\X = \Set{X \in \Pat^* \colon X \cap \hat{X}_{\phi_i} \neq \emptyset}$;
\item merge all the elements in $\X$ to form a single element in $\Pat^*$ by $\Pat^* = (\Pat^* \setminus \X) \cup \tilde{\X}$.
\end{itemize}
$\Pat^*$ is the minimal minimizer of $\min_{\Pat\in \Pi(V)} \FuHash{\alpha}[\Pat]$ at the end of the CoordSatCap algorithm. It is easy to see that by letting $\rv_V = (\alpha-H(V))\chi_V$ we have $\rv_V \in P(\FuHash{\alpha},\leq)$ initially. Let $\Phi$ be any linear ordering of $V$. We have
\begin{multline} \label{eq:PolyRank1}
\min\Set{ \FuHash{\alpha}(X) - r(X) \colon \phi_i \in X \subseteq V} \\ = \min\Set{ \FuHash{\alpha}(X) - r(X) \colon \phi_i \in X \subseteq V_i}
\end{multline}
where $V_{i} = \Set{\phi_1,\dotsc,\phi_{i}}$ due to the monotonicity of the entropy function $H$ \cite{Fujishige2005}.\footnote{This property has also been used in \cite{MiloIT2015,CourtIT2014} for solving the non-asymptotic CO problem.}
\begin{lemma} \label{lemma:PolyRank2}
Let $\Pat^*$ be the partition that is updated in each iteration of the CoordSatCap algorithm as described above,
\begin{multline}
\min\Set{\FuHash{\alpha}(X) - r(X) \colon \phi_i \in X \subseteq V} \\ = \min\Set{\FuHash{\alpha}(\tilde{U}) - r(\tilde{U}) \colon \Set{\phi_i} \in U \subseteq \Pat^*}. \nonumber
\end{multline}
Let $\hat{X}_{\phi_i}$ and $U_{\phi_i}^*$ be the minimal minimizer of the LHS and RHS, respectively, of the equation above. Then, $\hat{X}_i=\tilde{U}_i^*$.
\end{lemma}
\begin{IEEEproof}
For any $X \subseteq V$, let $\Y=\Set{Y \in \Pat \colon Y \cap X \neq \emptyset}$. We have
\begin{equation}
\begin{aligned}
& \quad \FuHash{\alpha}(X) - r(X) - \FuHash{\alpha}(\tilde{\Y}) + r(\tilde{\Y}) \\
& = \FuHash{\alpha}(X) - \FuHash{\alpha}(\tilde{\Y}) + r( \tilde{\Y} \setminus X ) \\
& = \FuHash{\alpha}(X) - \FuHash{\alpha}(\tilde{\Y}) + \sum_{Y \in \Y} \big( \FuHash{\alpha}(Y) - \FuHash{\alpha}(Y \cap X) \big) \geq 0,
\end{aligned} \nonumber
\end{equation}
where the last inequality is obtained by applying submodular inequality \eqref{eq:SubMIneq} inductively over intersecting subsets. The minimality of $\tilde{U}_i^*$ over all $X \subseteq V$ such that $\phi_i \in X$ can also be seen by induction. So, $\hat{X}_i=\tilde{U}_i^*$.
\end{IEEEproof}
Based on \eqref{eq:PolyRank1} and Lemma~\ref{lemma:PolyRank2}, we can implement the CoordSatCap algorithm by a fusion method as in the CoordSatCapFus algorithm, where steps 8 and 9 are equivalent to the procedure that updates $\Pat^*$ as described above.
\bibliographystyle{ieeetr}
| 27560135805e70ad1b0e6c5e69651927fa414f50 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\large Supplementary Information }
\section{Method and experimental setup}
The target WS$_2$ flake was identified with an optical microscope and then transferred to cover part of the long graphene channel. To promote adhesion, the wafer was annealed in O$_2$ at $300\celsius$ for 3 hours. Standard electron-beam lithography and electron-beam evaporation were used to connect the graphene with multiple $80 \rm{nm}$ thick Au electrodes. The electrodes allow independent four-terminal resistivity measurements in the covered and uncovered areas. After fabrication of electrodes, an h-BN flake was transferred to cover the whole device to serve as a top gate dielectric, followed by the top Au gate metal fabrication with similar electron-beam lithography procedures. No additional annealing was performed thereafter.
\section{Reproducible WAL in a single-layer graphene on WS$_2$ device}
Here we present a second device that has the similar characteristics as the first one shown in the main text. Instead of transferring WS$_2$ onto graphene, in this second sample we transferred graphene onto WS$_2$, in order to show that the absence of density saturation is independent of the transfer sequence and robustness of WAL exists. Figure~\ref{fig:s1}(a) shows the conductivity of graphene versus gate voltage. This device exhibits almost the same properties as the device in the main text, despite its lower mobility ($\sim 3000\rm{cm}^{-1}\rm{s}^{-1}\rm{V}^{-1}$ on the hole side, and $\sim 2000\rm{cm}^{-1}\rm{s}^{-1}\rm{V}^{-1}$ on the electron side), which is mainly due to the bubbles in the device. As clearly seen in the inset, bubbles (small black dots) appear across the graphene flake. We intentionally did not choose a bubble-free area in order to minimize the UCFs. Figure~\ref{fig:s1}(b) shows the WAL features observed in this device at different gate voltages. Compared to the first sample, MC is smallerhere. We symmetrize the data to show clearer temperature dependence, as is shown in Figure~\ref{fig:s1}(c). Just as in the first device, WAL is present on both hole and electron sides and disappears quickly as temperature increases.
\begin{figure}[H]
\includegraphics[width=\textwidth]{sup_1v2.pdf}
\caption{(a) Conductivity versus gate voltage. Inset: SEM image of graphene on WS$_2$ before electrode fabrication. Bubbles are visible after graphene is transferred onto WS$_2$. (b) MC and its dependence on the temperature (c) at the trace corresponding to the black dotted line in (b).
}
\label{fig:s1}
\end{figure}
\section{Temperature dependence of MC}
Figure~\ref{fig:s3}(a) shows the temperature dependence of the MC data of the WS$_2$-covered graphene. Compared with the spin diffusion length and the SOC scattering rate \cite{Han11}, the inelastic dephasing rate $\tau_{\phi}^{-1}$ is much more sensitive to temperature at low temperatures; thus the dramatic decrease of the WAL signal can be primarily attributed to the significantly increased inelastic dephasing rate. The extracted dephasing rate $\tau_{\phi}^{-1}$ as a function of temperature appears in Fig.~\ref{fig:s3}(b). We find that $\tau_{\phi}^{-1}$ obeys approximately a linear temperature dependence, which can be explained by the electron-electron scattering in the diffusive regime \cite{Altshuler82},
\begin{equation}
\tau_{\phi}^{-1}= \alpha \frac{k_B T}{2 E_F\tau_0}\ln\left(\frac{E_F\tau_0}{\hbar}\right),
\label{eq.phase_rate}
\end{equation}
where $\alpha$ is a correction coefficient equal to $2.4$. If the graphene mobility is extremely high, the sample will reach the ballistic regime ($k_B T \tau_0/\hbar \gg1$) and the temperature dependence of $\tau_{\phi}^{-1}$ will turn parabolic \cite{Narozhny02}. Under this circumstance, Eq.~\eqref{eq:MC_SOC} will also be rendered invalid since it is developed for the diffusive regime. However, since our device has a moderate mobility of $\sim 4,000\, \rm{cm}^2\rm{s}^{-1}\rm{V}^{-1}$, it is well in the diffusive transport regime ($k_B T \tau_0/\hbar \ll 1$); therefore Eq.~\eqref{eq:MC_SOC} is applicable. At low temperatures, the dephasing time $\tau_{\phi}^{-1}$ appears to start deviating from the straight line. In principle, at low temperatures the electron-electron scattering may not be the dominant inelastic scattering mechanism, as compared with electron-phonon interactions, the spin-flip scattering of electrons from localized spins \cite{McCann12}, etc.
\begin{figure}[H]
\centering
\includegraphics[width=0.7\linewidth]{sup_3.pdf}
\caption{(a) Temperature dependence of MC for WS$_2$-covered graphene at carrier density $n = 5\times10^{12}\rm{cm}^{-2}$. Solid lines are fits assuming a temperature-independent SOC rate. The dephasing rate extracted from the fitting is shown in (b). The temperature dependence of the dephasing rate (black dots) is approximately linear. The red dotted line is the dephasing rate calculated from Eq. \eqref{eq.phase_rate} with $\alpha = 2.4$. }
\label{fig:s3}
\end{figure}
\section{Universal conduction fluctuations (UCF)}
Universal conductance fluctuations can be extracted by removing the WAL background in the MC, as shown in Fig.~\ref{fig:sup_UCF}(a). The WAL curve is fitted by Eq.~\ref{eq:MC_SOC} and describes the experimental data quite well up to 50 mT. In addition to the reproducibility of the MC curves, the nearly symmetric fluctuations in conductance as a function of magnetic field provides further evidence that for UCF.
To calculate the phase coherence length $l_{\phi}$ from the UCF, we utilized the autocorrelation function $F(\Delta B)= \langle \delta \sigma (B+\Delta B)\delta \sigma(B) \rangle$ to find the characteristic magnetic field $B_{\phi}$ $(B_{\phi} l_{\phi}^2=h/2e)$, which is determined by $F(B_{\phi} )= \frac{1}{2} F(0)$. Figure~\ref{fig:sup_UCF}(b) shows the normalized autocorrelation function at different gate voltages. When the device approaches a higher carrier density, the characteristic field $B_{\phi}$ clearly decreases, indicating an increase in the phase coherence length $l_{\phi}$ due to the weaker electron-electron interaction than the Dirac region. The $l'_{\phi}$ s extracted from UCF agree reasonably well with those extracted from WAL, as shown in Fig.~\ref{fig:sup_UCF}(c), suggesting the validity of Eq.~\eqref{eq:MC_SOC} and thus the extracted spin-orbit scattering rates.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{sup_UCFv2.pdf}
\caption{(a) UCF extracted from MC by removing the WAL background. (b) Normalized autocorrelation functions at different gate voltages. $F(B_{\phi} )/F(0)=1/2$ gives the characteristic magnetic field $B_{\phi}$. (c) Phase coherence length extracted from UCF and WAL. }
\label{fig:sup_UCF}
\end{figure}
\section{DFT Computational details}
Our first-principles calculations were carried out using the projected augmented plane-wave method \cite{Kresse99, Blochl94} as implemented in the Vienna ab initio simulation package. The valence electron configuration for C, W, and S is $2s^2 2 2p^2$, $5d^4 6s^2$, and $3s^23p^4$, respectively. Generalized gradient approximation of Perdew-Burke-Ernzerhof type was used for the description of exchange-correlation interactions among electrons \cite{Perdew96}. Spin-orbit coupling was included in the self-consistent calculation level. We employed $5\times5\times1$ $K$-point grid for Graphene/WS$_2$ heterostructure containing $162$ C, $49$ W and $98$ S atoms. A large supercell was adopted to minimize the lattice mismatch; most of our simulations correspond to a supercell with $9\times 9$ graphene on $7\times 7$ WS$_2$ (see, however, the next section). Slab structures were separated by $\sim 13$ \AA~of vacuum along the surface normal. The energy cutoff for the plane-wave-basis expansion was set to $400 \rm{eV}$. Positions of all atoms were fully relaxed until the convergence of total energies became better than $0.1 \rm{meV}$. In order to treat the van der Waals force properly, we adopted the DFT-D3 method suggested by Ref.~\onlinecite{Grimme10}.
\section{Effect of the supercell size on the low-energy states}
\label{app.size_effect}
With a $9\times 9$ graphene unit cell, the Dirac cones are folded to $\Gamma$ point. Consequently, spurious interaction between the valleys may appear in our DFT calculations. To verify that the induced SOC of graphene is not sensitive to the chosen cell size, we repeated the simulations for a different supercell. Specifically, we adopted $5\times5$ graphene in contact with a $4 \times 4$ WS$_2$ monolayer so that the $K$ and $K'$ valleys reside at distinct momenta. The lattice mismatch increases to $2.49 \%$ whereas the separation between graphene and WS$_2$ decreases to $3.21$ \AA. Although we used a different supercell, band energies near the Dirac point remain essentially the same as shown in Fig.~\ref{fig:sup_band_structure}(a). We also found that the Dirac cones are very sensitive to the SOC strength. Figure~\ref{fig:sup_band_structure}(b) shows that the band order becomes inverted at the $K$ and $K'$ points as SOC is turned on from zero to its physical value (we checked that the same band inversion occurs in the $9\times 9$ graphene supercell). We can therefore conclude that the enhancement of SOC is only weakly affected by the specific atomic configuration and that inter-valley coupling provides a rather minor effect even when the Dirac cones are shifted to the same momentum.
\begin{figure}[h]
\includegraphics[width=0.5\columnwidth]{DFT_Supple1.pdf}
\caption{(a) Calculated band structure for the slab structure of WS$_2$ monolayer ($4\times4$) on graphene ($5\times 5$). (b) Eenergy gap of the Dirac cone as a function of the spin-orbit coupling strength, which ranges from zero to the true value corresponding to 1 on the horizontal axis. }
\label{fig:sup_band_structure}
\end{figure}
\section{Fitting the model Hamiltonian to first-principles calculations}
By matching energies from the model Hamiltonian in Eq.~\eqref{eq:hamiltonian} with DFT calculations, we determined the following parameters for a $9\times 9$ ($5\times 5$) graphene supercell: $M=0.79$ $(0.69) \rm{meV}$, $\lambda_{\rm{R}}=0.03$ $(0.08) \rm{meV}$, $\lambda_{\rm{VZ}}=0.96$ $(1.47) \rm{meV}$ and $\lambda_{\rm{I}}\sim 0$ $(0)\rm{meV}$ for the $9\times 9$ ($5\times 5$) supercell. From Fig.~\ref{fig:sub_DFT_fitting} we see that high-quality fits are possible for either interlayer separation.
\begin{figure}[h]
\includegraphics[width=0.5\columnwidth]{sub_DFT_fittingv2.pdf}
\caption{Calculated band structure from DFT and the model Hamiltonian for (a) $9\times9$ and (b) $5\times 5$ supercell.}
\label{fig:sub_DFT_fitting}
\end{figure}
\end{document}
| 9ff7b0217c2f75c9acdfa2b50b15d22596c0626f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Nonequilibrium systems can display phase transitions, with a number of well understood universality classes \cite{hinr00,henk08,odor08}.
Some features not observed in equilibrium systems can occur if detailed balance is not fulfilled, e.g., correlations with a power law decay far
from criticality and phase transitions with short range interactions in one dimension. Examples of nonequilibrium phase transitions include boundary induced phase transitions \cite{krug91,blyt07},
phase transitions into absorbing states \cite{hinr00}, real space condensation \cite{evan05}, and a transition to collective motion in active systems \cite{tama95,solo15}.
The production of entropy is a signature of systems out of equilibrium. This entropy production can be defined
for many nonequilibrium models, which can be related to quite different phenomena. Such a feature makes the
investigation of the critical behavior of entropy production appealing. It is intriguing to wonder
whether nonequilibrium phase transitions can be classified with respect to the critical behavior of the entropy production.
The critical behavior of the entropy production rate has been analyzed in the majority vote model \cite{croc05}, in a two-dimensional Ising model in contact
with two heat baths \cite{deol11,tome12}, and in a model for nonequilibrium wetting \cite{bara12}. For the first two models, the first derivative of the
entropy production rate with respect to the control parameter was found to diverge at the critical point. For the third model the first derivative of the entropy production rate
was found to be discontinuous at criticality. This discontinuity in the first derivative was also found within a mean-field approximation of the first model \cite{croc05}.
Furthermore, the entropy production rate of a model for population dynamics with a non-equilibrium phase transition has also been analyzed in \cite{andr10}.
In this paper we investigate the critical behavior of the entropy production rate in an Ising model driven by
a magnetic field that oscillates deterministically in time. This model displays a phase transition characterized by an order
parameter given by the magnetization integrated over a period \cite{tome90,chak99}. For the mean-field model, the phase
transition can be either first or second-order, depending on the field amplitude and frequency. For the two dimensional
model, the same kind of phase diagram with first and second-order phase transitions has been observed \cite{chak99}. However, a more careful
numerical analysis indicates that for the two dimensional model the transition is always second-order \cite{korn02}, with critical
exponents compatible with the Ising universality class \cite{korn01,buen08,fuji01}.
In contrast to the models listed in the previous paragraph, this model is driven by an external protocol
and, therefore, reaches a periodic steady state \cite{sini07a,raha08,astu11,raz16}.
We analyze both a mean-field Ising model where all spins interact with each other and a two-dimensional Ising model with nearest neighbors interactions.
Within the mean-field approximation, the entropy production rate is found to have a kink at the critical point for a second-order phase transition
and is found to be discontinuous at criticality for a first-order phase transition. For the two-dimensional model
the first derivative of the entropy production rate is found to diverge at the critical point, independent of the the frequency and amplitude of the field, which is
in agreement with the absence of a first-order phase transition in two dimensions.
An Ising model with a field that changes at stochastic time-intervals between two values is also considered.
The critical behavior of the entropy production rate does not change in relation to the one observed
in the model with a deterministic field. However, this model allows for a further
perspective related to the relation between information and thermodynamics \cite{saga12,horo13,mand12,bara14,bara14b,hart14,horo14,parr15,bara14a,hart16}.
The field can be seen as a signal and the system as a sensor that follows this signal. An information theoretic observable that quantifies the rate at which
the system obtains information about the field is the learning rate \cite{bara14a,hart16}. This learning rate appears in a second law inequality for bipartite systems \cite{hart14,horo14},
being bounded by the thermodynamic entropy production rate that quantifies heat dissipation \cite{bara14a}.
We study the critical behavior of the learning rate. Specifically, we introduce a lower bound on the learning rate that can be calculated in
numerical simulations. Its first derivative with respect to the temperature is found to be discontinuous at the critical point.
The paper is organized in the following way. In Sec. \ref{sec2} we calculate the entropy production rate for the mean-field model. The two-dimensional model
with a deterministic field is analyzed in Sec. \ref{sec3}. In Sec. \ref{sec4} we introduce the model with a field that changes at stochastic time-intervals
and investigate the critical behavior of the learning rate. We conclude in Sec. \ref{sec5}.
\section{Mean-field approximation}
\label{sec2}
\subsection{Model definition and phase diagram}
We consider a Curie-Weiss mean-field Ising model of $L^2$ spins $s_i=\pm 1$ that is subjected to a time-dependent external field of strength $h(t)$. The time-dependent Hamiltonian
is given by
\begin{equation}
H(t)\equiv -\frac{J}{L^2}\sum_{ ij} s_is_j-h(t)\sum_i{s_i},
\end{equation}
where the first term on the right hand side involves a sum over all spins. The size-dependent pre-factor in this first term makes the Hamiltonian extensive. The external field
varies periodically with time as
\begin{equation}
h(t)\equiv h_0 \cos(\omega t),
\label{eqfield}
\end{equation}
i.e., the field oscillates with an amplitude $h_0$ and a frequency $\omega$.
We consider a model with Markovian dynamics. The transition rate from configuration $(s_1,\ldots,s_i,\ldots,s_{L^2})$
to configuration $(s_1,\ldots,-s_i,\ldots,s_{L^2})$ is denoted by $w_i(s_i)$. The probability of an state $\mathbf{s}=(s_1,\ldots,s_i,\ldots,s_{L^2})$
at time $t$ is denoted $P_\mathbf{s}(t)$. The average magnetization at time $t$ is given by
\begin{equation}
m(t)\equiv \sum_\mathbf{s} s_iP_\mathbf{s}(t),
\end{equation}
where the above definition is independent of $i$ due to homogeneity.
Even though the model does not reach an equilibrium state due to the periodic variation of the external protocol, the transition rates at a fixed
time $t$ fulfill the detailed balance condition
\begin{equation}
\frac{w_i(s_i)}{w_i(-s_i)}= \textrm{e}^{-2 h_i(t)s_i/T},
\end{equation}
where $T$ is the temperature, Boltzmann constant is $k_B=1$ throughout, and
\begin{equation}
h_i(t)\equiv \frac{J}{L^2}\sum_js_j+h(t).
\end{equation}
Assuming Glauber transition rates, i.e.,
\begin{equation}
w_i(s_i)= \frac{1}{2\tau}[1-s_i \tanh( h_i/T)],
\end{equation}
where $\tau$ sets the time-scale for a spin flip, it is possible to show that the magnetization follows the equation \cite{tome90}
\begin{equation}
\tau \frac{dm}{dt}= -m(t)+\langle\tanh[h_i(t)/T]\rangle.
\label{eqmag1}
\end{equation}
In the thermodynamic limit $L\to \infty$ this equation is simplified by the relation $\sum_js_j/L^2\to m(t)$.
Hence, in this limit, Eq. (\ref{eqmag1}) becomes
\begin{equation}
\tau \frac{dm}{dt}= -m(t)+\tanh[Jm(t)/T+h(t)/T].
\label{eqmag2}
\end{equation}
\begin{figure}
\centering\includegraphics[width=90mm]{figure1.eps}
\vspace{-3mm}
\caption{Phase diagram of the mean-field model. The parameters
in Eq. (\ref{eqmag2}) are set to $J=\tau=1$ and $\omega= 0.2\pi$. In the region marked by the dots the transition is first-order and
in the solid line the transition is second-order. Above (below) the critical line the order parameter is $M=0$ ($M\neq0$).}
\label{fig1}
\end{figure}
The solution of Eq. (\ref{eqmag2}) reaches a periodic steady state, i.e., $m(t)=m(t+2\pi/\omega)$, that is independent of the initial
condition.
This model has a phase transition at a critical temperature $T_c$ that depends on the amplitude of the field $h_0$ and the
frequency $\omega$. The order parameter of this transition is the magnetization integrated over a period in the periodic steady state
\begin{equation}
M\equiv \frac{\omega}{2\pi}\int_{0}^{2\pi/\omega}m(t)dt.
\label{magdef}
\end{equation}
Below (above) the critical temperature the magnetization is $M\neq0$ ($M=0$). A phase diagram obtained with numerical integration of Eq. (\ref{eqmag2}) is shown in
Fig. \ref{fig1}. This phase diagram has been obtained in \cite{tome90} and is shown here for illustrative purposes.
Depending on $h_0$ and $\omega$ the phase transition can be first-order with a discontinuity in $M$, or second-order.
\subsection{Heat and Work}
Taking the time derivative of the internal energy per spin $u(t)\equiv H/L^2$, we obtain
\begin{equation}
\frac{du}{dt}= \frac{du}{dm}\frac{dm}{dt}+\frac{du}{dh}\frac{dh}{dt}.
\end{equation}
Following the standard definition of work in stochastic thermodynamics \cite{seif12}, we identify
the rate of work done on the system as
\begin{equation}
\dot{w}(t)\equiv \frac{du}{dh}\frac{dh}{dt}.
\label{eqmfw}
\end{equation}
The expression for the dissipated heat follows from the first law
\begin{equation}
\dot{q}(t)\equiv \dot{w}(t)-\frac{du}{dt}= -\frac{du}{dm}\frac{dm}{dt}.
\end{equation}
Since $u(t)$ is periodic, we obtain
\begin{equation}
\int_0^{2\pi/\omega}\dot{w}(t)dt= \int_0^{2\pi/\omega}\dot{q}(t)dt,
\label{eqmfwq}
\end{equation}
i.e., the average work done on the system in one period equals the average dissipated heat in one period.
The entropy production rate in the periodic steady state is defined as
\begin{equation}
\sigma\equiv \frac{1}{T}\frac{\omega}{2\pi}\int_{0}^{2\pi/\omega}\dot{q}(t)dt=\frac{1}{T}\frac{\omega}{2\pi}\int_{0}^{2\pi/\omega}\dot{w}(t)dt,
\label{eqmfsig}
\end{equation}
where we used the first law (\ref{eqmfwq}) in the second equality.
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure2a.eps}\label{fig2a}}
\subfigure[]{\includegraphics[width=75mm]{figure2b.eps}\label{fig2b}}
\vspace{-3mm}
\caption{Critical behavior of the entropy production rate for the mean-field model. The vertical lines mark the critical points. The parameters
in Eq. (\ref{eqmag2}) are set to $J=\tau=1$ and $\omega= 0.2\pi$. (a) $h_0=0.25$, corresponding to a second-order phase transition. (b)
$h_0=0.55$, corresponding to a first-order phase transition.
}
\label{fig2}
\end{figure}
We solved equation (\ref{eqmag2}) numerically and calculated the entropy production rate $\sigma$ with Eqs. (\ref{eqmfw}) and (\ref{eqmfsig}). The results
are shown in Fig. \ref{fig2}, where we plot $\sigma$ as a function of the temperature $T$ for two different values of the amplitude $h_0$ and fixed frequency $\omega$.
If $h_0$ is such that the phase transition is second-order, the entropy production rate has a kink at the critical point, indicating that the first derivative of $\sigma$
with respect to $T$ has a discontinuity at criticality. If the phase transition is first-order, the entropy production rate itself is discontinuous at the critical point.
Hence, within the mean-field model the critical behavior of the entropy production rate is different in the two different regions of the phase diagram.
We note that a discontinuity in the first derivative of $\sigma$ with respect to the control parameter has also been
observed in a mean-field approximation of the majority vote model \cite{croc05} and in a model for nonequilibrium wetting \cite{bara12}.
\section{Two-dimensional Ising model}
\label{sec3}
\subsection{Model definition}
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{fignewa.eps}\label{fignewa}}
\subfigure[]{\includegraphics[width=77mm]{fignewb.eps}\label{fignewb}}
\vspace{-3mm}
\caption{Binder cumulant $U_L$ as a function of the temperature $T$ for different system sizes. The frequency is $\omega= 0.04\pi$ and $J=1$.
(a) $h_0=1.0$, where the crossing point gives $T_c=1.37(1)$. (b) $h_0=2.8$, where the crossing point gives $T_c=0.33(1)$. The minimum
in (b) that decreases with system size is not due to a first-order phase transition but rather a finite-size effect \cite{korn02}.
}
\label{fignew}
\end{figure}
For the two-dimensional Ising model with nearest neighbors interactions, a configuration $\mathbf{s}=(s_1,s_2,\ldots,s_{L^2})$ at time $t$ has energy
\begin{equation}
E_{\mathbf{s}}(t)\equiv -J\sum_{ \langle ij\rangle} s_is_j-h(t)\sum_i{s_i},
\end{equation}
where the first sum is over nearest neighbors, $h(t)$ is given by (\ref{eqfield}), and we consider periodic boundary conditions.
The Binder cumulant is defined as \cite{land14}
\begin{equation}
U_L\equiv1-\langle M^4\rangle/(3\langle M^2\rangle^2),
\end{equation}
where $M$ represents the magnetization integrated over a period and the brackets denote an average over stochastic trajectories.
We calculated this Binder Cumulant with numerical simulations, which are explained below. The critical temperatures are determined
from the crossing points of the Binder Cumulant in Fig. \ref{fignew}. The lack of a minimum of the Binder cumulant that crosses from 2/3 to 0 in Fig. \ref{fignewa} is
an indicator of a second-order phase transition. The minimum of the Binder cumulant at a negative value in Fig. \ref{fignewb} is an indicator of a first-order
phase transition, and this transition has been interpreted to be first-order from this kind of numerical
result \cite{chak99}. However, this minimum has been shown to be finite size effect with a solid theoretical argument and extensive
numerical simulations that show that the minimum disappears for large enough systems \cite{korn02} (see Fig. \ref{fignewb}). The basic idea of the
theoretical argument is that for small systems the temperature at which the system
crossover from the multiple droplet to the single droplet regime is above the critical temperature, leading to major differences in the
probability distribution of the order parameter \cite{korn02}. Hence, convincing numerical evidence supports that in two dimensions the transition
is always second-order, independent of $h_0$ and $\omega$. Our numerical results in Fig. \ref{fignewb} indicate a minimum that decreases with system size, in agreement
with the results from \cite{korn02}.
The system reaches a periodic steady state characterized by the probability $P_\mathbf{s}(t)$, which has a period $2\pi/\omega$.
The entropy production rate per spin in this periodic steady state is defined as \cite{seif12}
\begin{equation}
\sigma_L\equiv \frac{1}{L^2}\frac{\omega}{2\pi}\int_0^{2\pi/\omega}dt\left(\sum_{\mathbf{s}\mathbf{s}'}w_{\mathbf{s}\mathbf{s}'}(t)P_\mathbf{s}(t)\frac{E_{\mathbf{s}}(t)-E_{\mathbf{s}'}(t) }{T}\right),
\end{equation}
where $w_{\mathbf{s}\mathbf{s}'}(t)$ is the transition rate from state $\mathbf{s}$ to state $\mathbf{s}'$ at time $t$. These transition rates are nonzero only if the configurations $\mathbf{s}$ and
$\mathbf{s}'$ differ by one spin flip and they fulfill the detailed balance relation
$w_{\mathbf{s}\mathbf{s}'}(t)/w_{\mathbf{s}'\mathbf{s}}(t)=\textrm{e}^{[E_{\mathbf{s}}(t)-E_{\mathbf{s}'}(t)]/T}$. The factor $L^{-2}$ makes $\sigma_L$ finite in the thermodynamic limit, where
\begin{equation}
\sigma\equiv\lim_{L\to \infty}\sigma_L.
\end{equation}
The rate of dissipated heat is $\dot{q}=T\sigma$, which is equal to the rate of work done on the system due to the first law.
Numerical simulations were performed with the following procedure. The initial condition is
a random configuration of spins, corresponding to $T\to \infty$. The time $t$ is discretized with the integer variable $n$, in such a way
that for $n=L^2$ we have $t=1$, i.e., the time $t$ is in units of Monte Carlo steps. We use the standard
metropolis rule for flipping a spin \cite{newm99}. A randomly chosen spin $s_i$ may flip depending on the energy difference
\begin{equation}
\Delta E_i(n/L^2)= 2s_i[J\sum_js_j+h(n/L^2)],
\end{equation}
where the sum in $j$ is over the four nearest neighbors. If this energy difference is negative the spin flips with probability one, and if it
is positive, the spin flips with probability $\textrm{e}^{-\Delta E_i(n/L^2)/T}$.
The entropy production rate $\sigma_L$ was computed in the following way. After a
certain transient the system reaches a periodic steady state and we compute the change in the entropy of the external medium
$\Delta S$ from time $0$, after the transient, to time $\mathcal{T}$, which corresponds
to several periods $2\pi/\omega$. If the system jumps from a configuration $\mathbf{s}$ to a configuration $\mathbf{s}'$ the entropy $\Delta S$ changes by an amount
$[E_{\mathbf{s}}(t)-E_{\mathbf{s}'}(t)]/T$. The entropy production rate per spin is then given by
\begin{equation}
\sigma_L=\Delta S/(\mathcal{T}L^2).
\label{entsim}
\end{equation}
\subsection{Critical behavior of the entropy production}
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure3a.eps}\label{fig3a}}
\subfigure[]{\includegraphics[width=77mm]{figure3b.eps}\label{fig3b}}
\vspace{-3mm}
\caption{Entropy production rate $\sigma_L$ as a function of the temperature $T$. The frequency is $\omega= 0.04\pi$ and $J=1$.
(a) $h_0=1.0$ and (b) $h_0=2.8$. The vertical lines indicate the critical temperatures given in the caption of Fig. \ref{fignew}.
}
\label{fig3}
\end{figure}
In Fig. \ref{fig3} we plot the entropy production rate $\sigma_L$ as a function of the temperature $T$ for two different values of $h_0$. In both cases the entropy production rate
has a maximum above $T_c$. At the critical point, the entropy production rate seems to have an inflection, indicating a divergence
of the first derivative of $\sigma_L$ at the critical point, i.e.,
\begin{equation}
d_T\sigma\sim |T-T_c|^{-\alpha}.
\end{equation}
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure4a.eps}\label{fig4a}}
\subfigure[]{\includegraphics[width=75mm]{figure4b.eps}\label{fig4b}}
\vspace{-3mm}
\caption{The first derivative $d_T\sigma_L$ as a function of the temperature $T$. The frequency is $\omega= 0.04\pi$ and $J=1$.
(a) $h_0=1.0$ and (b) $h_0=2.8$. The vertical lines indicate the critical points given in the caption of Fig. \ref{fignew}.
}
\label{fig4}
\end{figure}
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure5a.eps}\label{fig5a}}
\subfigure[]{\includegraphics[width=75mm]{figure5b.eps}\label{fig5b}}
\vspace{-3mm}
\caption{Maximum of the first derivative $d_T\sigma_L^{\textrm{max}}$ as a function of the system size $L$. The frequency is $\omega= 0.04\pi$ and $J=1$.
(a) $h_0=1.0$, where the transition is second-order. (b) $h_0=2.8$, where the transition is first-order.
In both cases this maximum behaves as $d_T\sigma_L^{max}\sim \ln L$.
}
\label{fig5}
\end{figure}
Direct numerical evaluation of the exponent $\alpha$ with off-critical simulations turn out to be difficult. As shown in Fig. \ref{fig4}, the
first derivative $d_T\sigma_L$ for a finite system has a maximum that increases with $L$, in agreement with the expectation that $d_T\sigma$ diverges at criticality.
Plotting this maximum $d_T\sigma_L^{\textrm{max}}$ as a function of
system size in Fig \ref{fig5}, we obtain
\begin{equation}
d_T\sigma_L^{\textrm{max}}\sim \ln L.
\label{entcri}
\end{equation}
The exponent $\alpha$ is related to an exponent $\psi$, defined by $d_T\sigma_L^{\textrm{max}}\sim L^{\psi}$, through the scaling relation $\alpha=\psi\nu$, where
$\nu$ is the critical exponent characterizing the divergence of the correlation length at criticality ($\nu\simeq 1.1$ \cite{chak99}). Relation (\ref{entcri}) implies $\psi=0$, and, therefore,
$\alpha=0$.
The critical behavior of the entropy production rate is the same in both plots shown in Fig. \ref{fig5}. This result provides further support for a
second-order phase transition also for the amplitude $h_0=2.8$, for which the Binder cumulant for small systems shows
a minimum in Fig \ref{fignewb}, in the sense that with a first-order phase transition the system would explore different regions of the
phase-space, which could lead to a discontinuity in the entropy production at $T_c$. However, the lack of a jump in $\sigma$ cannot be taken
as a demonstration that the transition is not first-order: a relation connecting the average magnetization integrated over a period with the
entropy production is not known, and, therefore, a discontinuity in the order parameter does not necessarily imply a discontinuity in the entropy production.
The same kind of critical behavior of the first derivative of the entropy production, characterized by a logarithmic divergence,
has been observed in a majority vote model \cite{croc05} and in a Ising model in contact with two heat baths \cite{tome12}.
\section{Critical behavior of the learning rate}
\label{sec4}
\subsection{Stochastic external field}
We now consider a two-dimensional Ising model with a magnetic field that changes at stochastic times. A similar model has been considered in \cite{acha98}.
This magnetic field changes at a rate $\Gamma$ between the values $h_0$ and $-h_0$. The system and
external field together form a bipartite Markov process \cite{hart14,horo14}, which has $2\times 2^{L^2}$ states.
A state of this bipartite process is characterized by the vector $\mathbf{s}$ and a binary variable $x=\pm 1$ that
indicates whether the external field is $h_0$ or $-h_0$. The transition rate from an state $(\mathbf{s},x)$ to a
state $(\mathbf{s}',x')$ is
\begin{equation}
w_{\mathbf{s}\mathbf{s}'}^{xx'}\equiv
\cases{
w^{xx'}_{\mathbf{s}}=\Gamma\qquad\textrm{if $x\neq x'$ and $\mathbf{s}= \mathbf{s}'$ }\\
w^{x}_{\mathbf{s}\mathbf{s}'}=\chi_{\mathbf{s}\mathbf{s}'}/(1+\textrm{e}^{(E^x_{\mathbf{s}'}-E^x_\mathbf{s})/T})\qquad\textrm{if $x= x'$ and $\mathbf{s}\neq \mathbf{s}'$ }\\
0 \qquad\textrm{if $x\neq x'$ and $\mathbf{s}\neq \mathbf{s}'$,}
}
\end{equation}
where $\chi_{\mathbf{s}\mathbf{s}'}$ is $1$ if both configurations differ by a single spin flip and $0$ otherwise. The energy $E^x_\mathbf{s}$ is given by
\begin{equation}
E^x_\mathbf{s}\equiv -J\sum_{ \langle ij\rangle} s_is_j-xh_0\sum_i{s_i}.
\end{equation}
A bipartite Markov process has two kinds of jumps, internal ones that lead to a spin flip and external ones that change the external field.
The dissipated heat is related to the internal jumps, whereas the external jumps are related to work. Hence, only internal jumps appear
in the entropy production rate per spin $\sigma_L$, which is defined as \cite{seif12}
\begin{equation}
\sigma_L\equiv\frac{1}{L^2}\sum_x\sum_{\mathbf{s}\mathbf{s}'}P_\mathbf{s}^{x}w^x_{\mathbf{s}\mathbf{s}'}\frac{(E^x_\mathbf{s}-E^x_{\mathbf{s}'})}{T},
\end{equation}
where $P_\mathbf{s}^{x}$ is the stationary distribution.
We have performed continuous-time Monte Carlo simulations of this model, using a method related to the method introduced in \cite{bort75}. The
main difference is that we also have to account for jumps that lead to a change in the magnetic field. In our algorithm, at each jump,
there is a probability $1-p^{spin}$, which depends on the state of the system, that the magnetic field changes. A spin flip happens with
probability $p^{spin}$, and is executed with the procedure explained in \cite{bort75}.
The parameter $\Gamma$ is written as $\Gamma=\gamma L^2$, where for different system sizes $\gamma$ is kept fixed. The
parameter $\Gamma$ must scale as $L^2$ for the following reason. The escape rate
of an state $(\mathbf{s},x)$ is $r_{\mathbf{s}x}=r^{spin}_{\mathbf{s}x}+\Gamma$, where $r^{spin}_{\mathbf{s}x}$ is the sum of all
transition rates that lead to a spin flip. Since there are $L^2$ spins that can be flipped, the parameter $\Gamma$ has to scale with $L^2$ for the probability of
a change in the field $1-p^{spin}=\Gamma/(r^{spin}_{\mathbf{s}x}+\Gamma)$ to be conserved with a change in system size.
The entropy production is calculated by adding
$(E^x_\mathbf{s}-E^x_{\mathbf{s}'})/T$ to $\Delta S$ every time a jump from $(x,\mathbf{s})$ to $(x,\mathbf{s}')$ occurs. If the simulation runs
for a time $\mathcal{T}$, after some transient, the entropy production rate per spin is calculated with expression (\ref{entsim}).
The critical point is again determined with the Binder cumulant
\begin{equation}
U_L=1-\langle m^4\rangle/(3\langle m^2\rangle^2),
\end{equation}
where the brackets denote an average over stochastic trajectories and $m= \sum_is_i/L^2$.
The Binder cumulant in Fig. \ref{fig6a} indicates a second-order phase transition. The critical behavior
of the entropy production rate is the same as in the previous model, as shown in Fig. \ref{fig6b}, where the maximum
of the first derivative of the entropy production rate follows the behavior in Eq. (\ref{entcri}).
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure6a.eps}\label{fig6a}}
\subfigure[]{\includegraphics[width=75mm]{figure6b.eps}\label{fig6b}}
\vspace{-3mm}
\caption{Critical behavior of $\sigma$ for the model with a stochastic field. The parameters are $\gamma=0.002$, $J=1$, and $h_0=0.5$.
(a) Binder Cumulant for different systems sizes indicating $T_c=2.19(1)$. (b) Maximum of the first derivative of the entropy production rate $d_T\sigma_L^{\textrm{max}}$
as a function of $L$.
}
\label{fig6}
\end{figure}
\subsection{Learning rate}
The Ising model with a stochastic field allows us to consider a further aspect related to information theory.
The external field can be interpreted as a stochastic signal and the system of spins as a sensor that follows the signal.
It turns out that there is a quantity, called learning rate \cite{bara14a,hart16}, that characterizes the rate at which the system obtains information
about the signal. Technically, this learning rate is a time derivative of the mutual information between system and external field.
In the stationary state, the learning rate (per spin) $\lambda_L$ is given by \cite{bara14a,hart16}
\begin{equation}
\lambda_L= \frac{\Gamma}{L^2}\sum_{\mathbf{s}}(P^{+1}_\mathbf{s}-P^{-1}_\mathbf{s})\ln\frac{P^{+1}_\mathbf{s}}{P^{-1}_\mathbf{s}}.
\end{equation}
The second law for a sensor following a signal reads \cite{hart14,horo14,bara14a}
\begin{equation}
\lambda_L\le \sigma_L.
\label{inelearn}
\end{equation}
The learning rate $\lambda_L$ that quantifies how much information the system obtains about the signal is bounded by $\sigma_L$, which quantifies heat dissipation.
This inequality allows for the definition of the informational efficiency $\eta= \lambda_L/\sigma_L$.
\begin{figure}
\subfigure[]{\includegraphics[width=75mm]{figure7a.eps}\label{fig7a}}
\subfigure[]{\includegraphics[width=75mm]{figure7b.eps}\label{fig7b}}
\vspace{-3mm}
\caption{Exact learning rate $\lambda_L$ and entropy production rate $\sigma_L$ for small systems. The parameters are $\gamma=0.002$, $J=1$, and $h_0=0.5$.
(a) $L=2$. (b) $L=3$.
}
\label{fig7}
\end{figure}
In Fig. \ref{fig7} we plot the learning rate $\lambda_L$ and the entropy production rate $\sigma_L$ as functions of temperature $T$ for $L=2,3$. They both
have maxima at some intermediate values of $T$. These figures were obtained with the exact calculation of the eigenvector of
the stochastic matrix that is associated with the eigenvalue $0$, which lead to the stationary distribution $P^x_{\mathbf{s}}$.
For larger values of $L$ we have to use Monte Carlo simulations. The problem that arises for the calculation of $\lambda_L$ in simulations is that
the increment in the learning rate after a jump depends on the nonequilibrium stationary probability $P^{x}_{\mathbf{s}}$, which is not known.
We propose the following lower bound on the learning rate. Instead of the microscopic configuration $\mathbf{s}$ we consider some mesoscopic variable $a$. In particular,
we consider a variable that gives the ``class'' of the spin in the continuous-time simulation \cite{bort75}. This variable takes the orientation of a spin and its nearest neighbors
into account and has $10$ possible outcomes: for each spin orientation the number of nearest neighbors with $s_j=1$ can go from $0$ to $4$. The lower bound is then written as
\begin{equation}
\lambda_L^{lb}\equiv \frac{\Gamma}{L^2}\sum_{a}(P^{+1}_a-P^{-1}_a)\ln\frac{P^{+1}_a}{P^{-1}_a}.
\label{learnlb}
\end{equation}
This lower bound fulfills $\lambda_L^{lb}\le \lambda_L$, which is illustrated in Fig. \ref{fig7b}, due to the log sum inequality \cite{cove06}.
The probability $P^{x}_a$ can be calculated in a Monte Carlo simulation by calculating the density of spins in each one of the ten classes in the steady state.
\begin{figure}
\centering\includegraphics[width=99mm]{figure8.eps}
\vspace{-3mm}
\caption{Critical behavior of the lower bound on the learning rate $\lambda_L^{lb}$. The parameters are $\gamma=0.002$, $J=1$, and $h_0=0.5$. The critical temperature $T_c=2.19(1)$
is indicated by the dotted vertical line.}
\label{fig8}
\end{figure}
It turns out that the lower bound $\lambda_L^{lb}$ scales as $L^{-4}$, going to zero in the limit of infinite system size.
This scaling can be understood with the following heuristic argument. The probability of changing the magnetic field in a transition, instead of
flipping a spin, is conserved as $L$ is increased. Therefore, after a change in the field, the average number of spin flips
before the next change in the external field is conserved for increasing $L$. However, the number of spins flips to equilibrate the system
is proportional to $L^2$. As the number of spin flips is constant and the number of spin flips necessary for equilibration scale as $L^2$, it
is reasonable to expect that the difference $|P^{+1}_a-P^{-1}_a|\sim L^{-2}$, which we have confirmed numerically.
From the expression (\ref{learnlb}) we obtain $\lambda_L^{lb}\sim L^{-4}$.
In Fig. \ref{fig8} we plot the scaled learning rate $\lambda_L^{lb}L^{4}$ as a function of the temperature $T$.
For large $L$, this lower bound seems to have a kink, in the form of a local minimum, at the critical point. Hence, our results indicated that the
first derivative of $\lambda_L^{lb}$ with respect to $T$ is discontinuous in the limit $L\to \infty$.
The critical behavior of the lower bound on the learning rate is then different from that of the entropy production.
We note that the lower bound on the efficiency $\lambda_L^{lb}/\sigma_L\le\eta$ goes to zero in the thermodynamic limit due to the different scaling of
$\lambda_L^{lb}$ in relation to $\sigma_L$.
\section{Conclusion}
\label{sec5}
We have analyzed the critical behavior of the entropy production rate of a nonequilibrium Ising
model subjected to a time-dependent periodic field. For the mean-field model, this entropy production rate
is found to have a jump at criticality if the transition is first-order. However, if the transition is second-order, the entropy production rate
is continuous but its first derivative is discontinuous at criticality. For the two-dimensional model, for which the transition is second-order, the first
derivative of the entropy production rate has a logarithmic divergence at the critical point. The novelty of our results in relation to previous studies
on the critical behavior of entropy production rate \cite{croc05,tome12,deol11,bara12} are the following. The models analyzed here are driven
by an external periodic protocol, in contrast to previous studies that consider models driven by
a fixed thermodynamic force; the entropy production rate was found to have a jump at criticality for
the mean-field model in the region with a first-order phase transition, which is a critical behavior
that has not been observed in previous studies. Furthermore, the results for the deterministic field support
the lack of a first-order phase transition in two-dimensions.
We have also investigated the critical behavior of the learning rate for the model with
an external field that changes at stochastic time-intervals between two values. It turns
out that the calculation of the learning rate within numerical simulations would require
the unknown nonequilibrium stationary distribution. We introduced a lower bound
on the learning rate that can be calculated within numerical simulations. Our numerics
indicates that the critical behavior of this lower bound is different from the one of the
entropy production: it has a local minimum at the critical point and its first
derivative seems to be discontinuous.
Our results on the Ising model with a stochastic field offers two fresh perspectives. First,
most studies on the relation between information and thermodynamics consider small systems.
However, the inequalities for bipartite processes \cite{hart14,horo14} are also valid for macroscopic
systems with a large number of states, as explicitly illustrated here. For example, it would
be interesting to build a model of a macroscopic Maxwell's demon using the framework for bipartite systems.
Second, this model allows for the definition of an informational efficiency. However,
the lower bound on the efficiency, which is the quantity we could calculate, turned out to go to zero in
the thermodynamic limit. Analyzing the critical behavior of efficiency in nonequilibrium models is also
an interesting perspective.
This work and the few previous studies on the critical behavior of entropy production demonstrate that the
average entropy production can be a useful observable to determine the
critical point of a generic nonequilibrium phase transition. As an interesting direction for future work, higher
order moments of the fluctuating entropy production could be even more effective for a precise determination of the critical point.
While the entropy production has been found to display three distinct behaviors at the critical line, i.e., a logarithmic divergence on its first derivative,
a discontinuity on its first derivative, and a discontinuity, the deeper question whether entropy production can be used to classify nonequilibrium phase
transitions in a meaningful way remains open.
{\noindent \textbf{Acknowledgements}}\newline We thank Shamik Gupta
for carefully reading the manuscript and Per Arne Rikvold for pointing out \cite{korn02}.
\section*{References}
| b70fd233b971540e73ae3ad3d63298ea8f984c95 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A difficult question regarding the simple modules over a simple, simply connected algebraic group $G$ over an algebraically closed field of positive characteristic $\Bbbk$ is to find an explicit formula for their characters. A formula involving the action of the corresponding affine Weyl group was proposed by Lusztig \cite{L - problems} in 1980. Subsequently this formula was shown to hold in large characteristic by the combined efforts of Kazhdan-Lusztig, Kashiwara-Tanisaki, Lusztig and Andersen-Jantzen-Soergel. More recently Williamson \cite{W - counterexample} found many counter-examples to the expected bounds in this conjecture. A different approach to character formula with emphasis in the Steinberg decomposition for algebraic groups is given in \cite{L - character}.
Around 1990 Lusztig started to study quantum groups $U_\lambda(\g)$ at a primitive root of unity $\lambda$ of order $\ell$ in order to have algebras over the complex numbers whose representation theory resembles those of simply connected semisimple algebraic groups over algebraically
closed fields of positive characteristic. In particular he conjectured a similar formula for the character of simple modules \cite{L - mod rep}, which holds in this case by a hard proof of Kazhdan-Lusztig. A remarkable fact about $U_\lambda(\g)$ is that it fits into a Hopf algebra extension of the corresponding small quantum group $\u_\lambda(\g)$ by the enveloping algebra $U(\g)$; each simple module satisfies a kind of Steinberg decomposition: it is written as a tensor product of a simple module of $\u_\lambda(\g)$ with a simple module $U(\g)$, viewed as $U_\lambda(\g)$-module via a (kind of) Frobenius map.
A fundamental difference, however, between the representation theory of the algebraic group and the corresponding quantum group at a root of unity is the form of the Steinberg (resp. Lusztig) tensor product theorem: for the algebraic group the theorem involves an arbitrary number of iterations of the Frobenius twist, whereas for the quantum group only one Frobenius twist occurs. It has been proposed by Soergel and Lusztig that there might exist analogues of the quantum group which parallel to a greater and greater extent the representation theory of the algebraic group. Such an object has the potential to deepen our understanding of the representation theory of algebraic groups \cite{W - notes}.
The purpose of this paper is to propose such an object for $\Sl_2$. More precisely, we introduce a family of finite dimensional algebras $\{\Dl{N}{\Sl_2}\}_{N\in\N_0}$ over the complex numbers
to mimic the filtration of the algebra of distributions of $\SL_2$ as a filtration by finite-dimensional subalgebras. This filtration is deeply motivated by the approach proposed in \cite{L - character}. The main objective is to
find a $\C$-algebra whose representation category \emph{behaves} as that of simple, simply connected algebraic groups over algebraically
closed fields of positive characteristic, even more similar than $U_\lambda(\g)$.
\begin{itemize}[leftmargin=*]\renewcommand{\labelitemi}{$\diamond$}
\item Each algebra $\Dl{N}{\Sl_2}$ is presented by generators and relations; the relations in Definition \ref{defn:Dlambda} resemble those defining finite dimensional subalgebras of the algebra of distributions of $\SL_2$ \cite{Tak - hyperalgebras}.
\item The first step corresponds to the small quantum group: $\Dl{0}{\Sl_2}\simeq \u_\lambda(\Sl_2)$. If $M<N$, then $\Dl{M}{\Sl_2}$ is a subalgebra of $\Dl{N}{\Sl_2}$, and at the same time there exists a surjective map $\Dl{N}{\Sl_2} \twoheadrightarrow \Dl{M}{\Sl_2}$, see Lemma \ref{lemma:proj over uqsl2}. Thus there exists a surjective map $\Dl{N}{\Sl_2} \twoheadrightarrow \u_\lambda(\Sl_2)$, a kind of \emph{Frobenius map}.
\item For the algebra of distributions, there exist extensions of Hopf algebras between consecutive terms of a filtration by (finite dimensional) Hopf subalgebras, see Proposition \ref{prop:extn Hopf algs}. In this case, $\Dl{N-1}{\Sl_2} \subset \Dl{N}{\Sl_2}$ is a
$\u_\lambda(\Sl_2)$-cleft extension for all $N\in\N$, see Proposition \ref{prop:cleft-ext}.
\item Each $\Dl{N}{\Sl_2}$ admits a \emph{triangular decomposition} into a positive, a zero and a negative part, see Proposition \ref{prop:basis, triang decomp}. Reasonably each simple module is a \emph{highest weight module}, see Proposition \ref{prop:simple-modules}.
\item Each simple $\Dl{N}{\Sl_2}$-module admits a \emph{Steinberg decomposition} as the tensor product of a simple $\Dl{N-1}{\Sl_2}$-module and a simple $\u_\lambda(\Sl_2)$-module as stated in Theorem \ref{thm:simple-modules-tensor-decomp}.
\end{itemize}
The next step is to define families of algebras $\Dl{N}{\g}$ for any semisimple Lie algebra $\g$ \cite{An}. This will be the content of a forthcoming paper, where the first step includes the definition of \emph{Lusztig's isomorphisms at level $N$}: that is, to consider Lusztig's isomorphisms for small quantum groups, which induce maps for $\Dl{N}{\g}$. Hence $\Sl_2$ is a key step to prove the existence of a PBW basis labeled by the roots of $\g$.
As the simple modules of these algebras satisfy a Steinberg tensor product decomposition, we hope to attack the modular case from the approach established by Lusztig in \cite{L - character}.
\subsection{Notation}
Let $H$ be a Hopf algebra with counit $\epsilon$ and antipode $\cS$. $H^+$ is the augmentation ideal, i.-e. the kernel of $\epsilon$.
The left adjoint action of $H$ on itself is
$\Ad(a)b=a_1 b \cS(a_2)$, $a,b\in H$. A Hopf subalgebra $A$ is (left) normal if it is stable by
the (left) adjoint action. Given $\pi:H\to K$ a Hopf algebra map,
\begin{align*}
H^{\co \pi} &= \{h\in H: (\id \ot \pi)\Delta(h)=h\ot 1\}, \
^{\co \pi} H = \{h\in H: (\pi \ot \id)\Delta(h)=1\ot h\},
\end{align*}
are the sets of left, respectively right, coinvariant elements.
\medskip
Let $A$ be a right $H$-comodule algebra; that is, an $H$-comodule such that the coaction map $\rho:A\to A\ot H$ is an algebra map. Let
\begin{align*}
B:=A^{\co H} &= \{a\in A: \rho(h)=a\ot 1\},
\end{align*}
the subalgebra of coinvariant elements. Then $B\subset A$
is a cleft extension if there exists an $H$-colinear convolution-invertible map
$\gamma:H\to A$; we refer to \cite[Section
7]{Mo} for more information.
A sequence of Hopf algebra maps
\begin{align*}
\xymatrix@C=9.5pt{\Bbbk \ar@{->}[rr] & & A \ar@{->}[rr]^{\iota} & & C \ar@{->}[rr]^{\pi} & & B \ar@{->}[rr] & & \Bbbk }
\end{align*}
is \emph{exact} \cite{And - ext,ADev} if the following conditions hold:
\begin{align*}
\iota&\text{ is injective,} & \pi&\text{ is surjective,}
& \ker\pi &= C \iota(A^{+}), & \iota(A) &= C^{\co \pi}.
\end{align*}
\section{Algebras of Distributions of reductive groups}\label{section:dist alg}
Let $\Bbbk$ be an algebraically closed field, $p=\ch \Bbbk$. Let $G$ be a simply connected semisimple algebraic group with Cartan matrix $A=(a_{ij})_{1\le i,j \le\theta}$,
$\g=\operatorname{Lie} G$.
Following \cite[Chapter 7]{Jan - book} we recall some definitions concerning algebraic groups. Then we give some results which illustrate those results we want to mimic for the quantum counterpart studied later.
\subsection{The algebra of distributions}
Let $I_e=\{f\in\Bbbk[G]| f(e)=0\}$. A \emph{distribution} on $G$ (with support on $e$) of order $n\in\N_0$ is a linear map $\mu:\Bbbk[G]\to\Bbbk$ such that
$\mu_{|I_e^{n+1}}\equiv 0$. Let $\Dis{n}{G}$ be the set of all distributions of order $n$, which is a $\Bbbk$-vector space. Now
\begin{align*}
\Di{G}= \bigcup_{n\geq 0} \Dis{n}{G} = \{ \mu \in \Bbbk[G]^{\ast} | \mu_{|I_e^{n+1}}\equiv 0 \text{ for some } n\in\N\}
\end{align*}
is the set of all distributions, which is a $\Bbbk[G]$-module. Then $\Di{G}$ is a Hopf subalgebra of $\Bbbk[G]^{\ast}$, called the \emph{algebra of distributions} (or the \emph{hyperalgebra}) of $G$. As algebra and coalgebra, it is filtered:
\begin{align*}
\Dis{m}{G} \cdot \Dis{n}{G} &\subset \Dis{m+n}{G} &
\text{for all } & m,n \in\N_0.
\end{align*}
We describe now two basic examples. We refer to \cite[7.8]{Jan - book} for more details.
\begin{example}\label{ex:additive-group}
Let $G=G_a$ (the additive group, which is the spectrum of $\Bbbk[t]$). For each $n\in\N_0$ we set $\gamma_n\in \Bbbk[G_a]^{\ast}$ as the function such that $\gamma_n(t^m)=\delta_{n,m}$ for all $m\in\N_0$. Then $(\gamma_n)_{n\in\N_0}$ is a basis of $\Di{G_a}$, and
\begin{align*}
\gamma_m \gamma_n & = \binom{m+n}{m} \gamma_{m+n}, & \text{for all } & m,n \in\N_0.
\end{align*}
In particular, $\gamma_1^p =p! \gamma_p=0$.
\end{example}
\begin{example}\label{ex:multiplicative-group}
Let $G=G_m$ (the multiplicative group, which is the spectrum of $\Bbbk[t,t^{-1}]$). For each $n\in\N_0$ we set $\varpi_n\in \Bbbk[G_a]^{\ast}$ as the function such that $\varpi_n((t-1)^m)=\delta_{n,m}$ for all $m\leq n$, $\varpi_n(I_e^{n+1})=0$. Then $(\varpi_n)_{n\in\N_0}$ is a basis of $\Di{G_m}$, and the multiplication satisfies that
\begin{align*}
\varpi_m \varpi_n & = \sum_{i=0}^{\min\{m,n\}} \frac{(m+n-i)!}{(m-i)!(n-i!i!)} \varpi_{m+n-1}, & \text{for all } & m,n \in\N_0.
\end{align*}
\end{example}
\medskip
By \cite{Tak - hyperalgebras} the algebra $\Di{G}$ is presented by generators $H_i\at{n}$, $X_i\at{n}$, $Y_i\at{n}$, $1\le i\le\theta$,
$n\in\N_0$, where $H_i\at{0} = X_i\at{0}=Y_i\at{0}=1$, and relations
\begin{align}\label{eq:rels dist-alg -1}
H_i(t) H_i(u) &= H_i(t+u+tu), \\ \label{eq:rels dist-alg -2}
H_i(t) H_j(u) &= H_j(u)H_i(t), \\ \label{eq:rels dist-alg -3}
X_i(t) X_i(u) &= X_i(t+u), \\ \label{eq:rels dist-alg -4}
Y_i(t) Y_i(u) &= Y_i(t+u), \\ \label{eq:rels dist-alg -5}
X_i(t) Y_i(u) &= Y_i\left(\frac{u}{1+tu}\right) H_i(tu) X_i\left(\frac{t}{1+tu}\right), \\ \label{eq:rels dist-alg -6}
X_i(t) Y_j(u) &= Y_j(u)X_i(t), \\ \label{eq:rels dist-alg -7}
H_i(t) X_j(u) &= X_j\left( (1+t)^{a_{ij}} u \right) H_i(t), \\ \label{eq:rels dist-alg -8}
H_i(t) Y_j(u) &= Y_j\left( (1+t)^{-a_{ij}} u \right) H_i(t), \\ \label{eq:rels dist-alg -9}
\ad \left( X_i\at{n}\right)\left( X_j\at{m}\right) &= \sum_{k=0}^n (-1)^k X_i\at{n-k}X_j\at{m}X_i\at{k}=0, & n &>-ma_{ij}, \\ \label{eq:rels dist-alg -10}
\ad \left( Y_i\at{n}\right)\left( Y_j\at{m}\right) &= \sum_{k=0}^n (-1)^k Y_i\at{n-k}Y_j\at{m}Y_i\at{k}=0, & n &>-ma_{ij}.
\end{align}
for $1\le i\neq j \le \theta$, where we consider the following elements of $\Di{G}[[t]]$:
\begin{align*}
H_i(t) &= \sum_{n=0}^\infty t^n H_i\at{n}, & X_i(t) &= \sum_{n=0}^\infty t^n X_i\at{n}, & Y_i(t) &= \sum_{n=0}^\infty t^n Y_i\at{n}.
\end{align*}
From \eqref{eq:rels dist-alg -2} we have $H_i\at{m}H_j\at{n} = H_j\at{n}H_i\at{m}$ for $i\neq j$, and from \eqref{eq:rels dist-alg -1},
\begin{align}\label{eq:reln Hn Hm}
H_i\at{m}H_i\at{n} &= \sum\limits_{\ell=0}^{\min\{m,n\} } \binom{m+n-\ell}{m}\binom{m}{\ell} H_i\at{m+n-\ell}.
\end{align}
From \eqref{eq:rels dist-alg -3} and \eqref{eq:rels dist-alg -4},
\begin{align}\label{eq:reln Xn Xm}
X_i\at{m}X_i\at{n} &= \binom{m+n}{m} X_i\at{m+n}, & Y_i\at{m}Y_i\at{n} &= \binom{m+n}{m} Y_i\at{m+n}.
\end{align}
From these formulas, the $H_i\at{n}$'s generate a copy of $\Di{G_m}$, while the $X_i\at{n}$'s, respectively the $Y_i\at{n}$'s, generate a copy of $\Di{G_a}$, see Examples \ref{ex:additive-group} and \ref{ex:multiplicative-group}.
From \eqref{eq:rels dist-alg -6} we have $X_i\at{m}Y_j\at{n} = Y_j\at{n}X_i\at{m}$ for $i\neq j$, and from \eqref{eq:rels dist-alg -5},
\begin{align*}
\sum_{n,m} t^n u^m X_i\at{n}Y_i\at{m} &= \sum_{a,b,c} \frac{u^{a+b}t^{b+c}}{(1+tu)^{a+c}} Y_i\at{a}H_i\at{b}X_i\at{c} \\
&= \sum_{a,b,c,d} (-1)^d \binom{a+c+d}{d} u^{a+b+d}t^{b+c+d} Y_i\at{a}H_i\at{b}X_i\at{c}.
\end{align*}
Thus,
\begin{align}\label{eq:reln Xn Ym}
X_i\at{n}Y_i\at{m} = \sum_{\ell=0}^{\min\{m,n\} } \sum_{k=0}^\ell (-1)^{\ell-k} \binom{m+n-\ell-k}{\ell-k} Y_i\at{m-\ell}H_i\at{k}X_i\at{n-\ell}.
\end{align}
A particular case of this formula is the following
\begin{align}\label{eq:reln Xn Ym, case}
[X_i\at{p^n},Y_i\at{p^m}] &= \sum_{\ell=1}^{\min \{p^m, p^n\}} Y_i\at{p^m-\ell} \left( \sum_{k=0}^\ell \binom{\ell+k}{\ell-k} H_i\at{k} \right) X_i\at{p^n-\ell}.
\end{align}
From \eqref{eq:rels dist-alg -7} and \eqref{eq:rels dist-alg -8} we have
\begin{align}\label{eq:reln Hn Xm}
[H_i\at{p^m}, X_j\at{p^n}] & = \delta_{n,m} a_{ij} X_i\at{p^n}, &
[H_i\at{p^m}, Y_j\at{p^n}] & = -\delta_{n,m} a_{ij} Y_i\at{p^n}.
\end{align}
\medskip
\subsection{Hopf algebra extensions from $\Di{G}$}
Let $\cD_n G:=\Dis{p^n}G$. As a consequence of these formulas we have the following result.
\begin{lemma}\label{lemma:normal Hopf subalg}
For all $n\in\N$, $\cD_{n}{G}$ is a normal Hopf subalgebra of $\cD_{n+1}{G}$.
\end{lemma}
\pf
$\cD_{n+1}{G}$ is generated as an algebra by $X_i\at{p^k}$, $Y_i\at{p^k}$, $H_i\at{p^k}$, $0\le k\le n$, so it is enough to prove
that $\cD_{n}{G}$ is stable by the adjoint action of $X_i\at{p^n}$, $Y_i\at{p^n}$, $H_i\at{p^n}$ since the remaining generators
belong to $\cD_{n}{G}$, and $\cD_{n}{G}$ is a Hopf subalgebra.
As $X_i\at{p^n}$ is primitive, $\Ad X_i\at{p^n} = \ad X_i\at{p^n}$. If $m<n$,
then $\ad (X_i\at{p^n})Y_i\at{p^m}$, $\ad (X_i\at{p^n})H_i\at{p^m}\in \cD_{n}{G}$, and $\ad (X_i\at{p^n})X_i\at{p^m}=0$. Let $j\neq i$. Note that
$\ad (X_i\at{p^n})X_j\at{p^m}=\ad (X_i\at{p^n})Y_j\at{p^m}=0$ since they commute, and from \eqref{eq:reln Hn Xm}, $\ad (X_i\at{p^n})H_j\at{p^m}=0$.
Therefore $\ad (X_i\at{p^n})\cD_{n}{G}\subset \cD_{n}{G}$
Analogous computations show that $\ad (Y_i\at{p^n})\cD_{n}{G}$, $\ad (H_i\at{p^n})\cD_{n}{G}\subset \cD_{n}{G}$.
\epf
Now define $\pi_k:\cD_{k+1}{G}\to \cD_{1}{G}=\Ur(\g)$ as follows
\begin{align}\label{eq:defn pi n}
\pi_k(X_i\at{n}) &= \begin{cases} X_i\at{n'}, & \mbox{if }n=p^k n', \\ 0, & \mbox{otherwise}, \end{cases} &
\pi_k(H_i\at{n}) &= \begin{cases} H_i\at{n'}, & \mbox{if }n=p^k n', \\ 0, & \mbox{otherwise}. \end{cases} \\ \notag
\pi_k(Y_i\at{n}) &= \begin{cases} Y_i\at{n'}, & \mbox{if }n=p^k n', \\ 0, & \mbox{otherwise}. \end{cases}
\end{align}
\begin{remark}\label{rem:comb numbers non zero}
Let $n<p^{k+1}$, $0<t<p$. If $p^k$ does not divide $n$, then $\binom{p^k t}{n}=0$, otherwise $n=p^k n'$ and $\binom{p^k t}{n}=\binom{t}{n'}$
by Lucas' Theorem.
\end{remark}
\begin{lemma}\label{lemma:surj Hopf alg map}
$\pi_k$ is a surjective Hopf algebra map.
\end{lemma}
\pf
First we have to check that $\pi_k$ is well defined; i.-e. that the map defined from the free algebra on generators $X_i\at{n}$,
$Y_i\at{n}$ and $H_i\at{n}$ annihilates the defining relations. We check easily that for all $i\neq j$, and $m,n\in \N_0$,
\begin{align*}
\pi_k(H_i\at{m}H_j\at{n} -H_j\at{n}H_i\at{m})&=\pi_k(X_i\at{m}Y_j\at{n} - Y_j\at{n}X_i\at{m})=0.
\end{align*}
For \eqref{eq:reln Hn Hm}, $\pi_k$ annihilates both sides of the equation if $p^k$ does not divides $m$ since either
$\pi_k(H_i\at{m+n-\ell})=0$ or else $\binom{m+n-\ell}{m}=0$. Now set $m=p^k m'$, $n=p^k n'$
\begin{align*}
H_i\at{m}H_i\at{n} &= \sum\limits_{\ell=0}^{\min\{m,n\} } \binom{m+n-\ell}{m}\binom{m}{\ell} H_i\at{m+n-\ell} \\
&= \sum\limits_{\ell'=0}^{\min\{m',n'\} } \binom{p^k(m'+n'-\ell')}{p^km'}\binom{p^k m'}{p^k\ell'} H_i\at{p^k(m'+n'-\ell')} \\
&= \sum\limits_{\ell'=0}^{\min\{m',n'\} } \binom{m'+n'-\ell'}{m'}\binom{m'}{\ell'} H_i\at{p^k(m'+n'-\ell')},
\end{align*}
since $\binom{m}{\ell}=0$ when $p^k$ does not divide $\ell$, so $\pi_k$ applies \eqref{eq:reln Hn Hm} to 0.
For \eqref{eq:reln Xn Xm}, if $p^k$ does not divide $m+n$, then both sides of the equality are annihilated by $\pi_k$.
If $p^k$ divides $m+n$ but does not divide $m$, then again $\pi_k$ annihilates both sides
of \eqref{eq:reln Xn Xm} since $\binom{m+n}{m}\equiv 0 \, (\mod p)$. Finally, if $p^k$ divides $m$ and $n$, then $m=p^km'$, $n=p^kn'$ and
\begin{align*}
\pi_k\Big( X_i\at{m}X_i\at{n} &- \binom{m+n}{m}X_i\at{m+n} \Big)= X_i\at{m'}X_i\at{n'} - \binom{p^k(m'+n')}{p^km'}X_i\at{m'+n'} \\
&= X_i\at{m'}X_i\at{n'} - \binom{m'+n'}{m'}X_i\at{m'+n'}=0.
\end{align*}
The proof for the $Y_i\at{m}$'s is analogous.
Notice that $\pi_k(X_i\at{m}Y_j\at{n}-Y_j\at{n}X_i\at{m})=0$ if $i\neq j$. For \eqref{eq:reln Xn Ym}, it is enough
to verify that \eqref{eq:reln Xn Ym, case} is annihilated since $X_i\at{M}$, $Y_i\at{N}$ can be written as products
of $X_i\at{p^m}$, $Y_j\at{p^n}$. If either $m<k$ or else $n<k$, then $\pi_k$ annihilates both sides of the equality.
Let $m=n=k$. Then
\begin{align*}
\pi_k\Big( [X_i\at{p^k},Y_i\at{p^k}] &-\sum_{\ell=1}^{p^k} Y_i\at{p^k-\ell} \left( \sum_{t=0}^\ell \binom{\ell+t}{\ell-t} H_i\at{t} \right) X_i\at{p^k-\ell}\Big) \\
&= [X_i,Y_i]-H_i=0.
\end{align*}
Now $\pi_k$ annihilates both equations of \eqref{eq:reln Hn Xm} by direct computation.
For \eqref{eq:rels dist-alg -9}, $\pi_k$ annihilates the left hand side if $p^k$ does not divide either $m$ or else $n$.
If $m=p^km'$, $n=p^kn'$ with $n>-ma_{ij}$, then $n'>-m'a_{ij}$ and
\begin{align*}
\pi_k \left( \ad \left( X_i\at{n}\right)\left( X_j\at{m}\right)\right) = \ad \left( X_i\at{n'}\right)\left( X_j\at{m'}\right)=0.
\end{align*}
Finally \eqref{eq:rels dist-alg -10} follows analogously. Hence $\pi_k$ is an algebra map.
To see that $\pi_k$ is a Hopf algebra map, it remains to prove that $\pi_k$ is a coalgebra map. But it follows
since the elements $X_i\at{p^j}$, $Y_i\at{p^j}$, $H_i\at{p^j}$, $0\le j\le k$, which are primitive elements and generate $\cD_{k+1}{G}$
as an algebra, are applied to primitive elements of $\cD_{1}{G}$.
\epf
The map $\pi_k$ fits in an exact sequence of Hopf algebras.
\begin{prop}\label{prop:extn Hopf algs}
The sequence of Hopf algebras
\begin{align}\label{eq:ex seq Dist}
\xymatrix{ \Bbbk \ar@{->}[r] & \cD_{k}{G} \ar@{^{(}->}[r] & \cD_{k+1}{G} \ar@{->>}[r]^{\pi_k} & \cD_{1}{G} \ar@{->}[r] & \Bbbk }
\end{align}
is exact.
\end{prop}
\pf
By Lemmas \eqref{lemma:normal Hopf subalg} and \eqref{lemma:surj Hopf alg map}, it remains to prove that
\begin{itemize}
\item $\ker \pi_k= \cD_{k+1}{G}(\cD_{k}{G})^+$, and
\item $\cD_{k}{G}=\cD_{k+1}{G}^{\co \pi_k}=\{x\in\cD_{k+1}{G}: (\id\otimes\pi_k)\Delta(x)=x\otimes 1 \}$.
\end{itemize}
Note that $\cD_{k+1}{G}(\cD_{k}{G})^+ \subseteq \ker\pi_k$ since $(\cD_{k}{G})^+$ is spanned by $X_i\at{k}$, $Y_i\at{k}$,
$H_i\at{k}$, $1\le k\le p^{n-1}$; the equality follows because both subspaces have the
same dimension, $\dim\cD_{k+1}{G}-\dim\cD_{k}{G}$. Now $\cD_{k+1}{G}^{\co \pi_k} \supseteq \cD_{k}{G}$, and the
equality follows by \cite[Theorem 3.4]{Tak - HopfGalois}.
\epf
\subsection{Steinberg decomposition for simple modules}
The purpose of this section is to introduce the Steinberg's tensor product Theorem. We will prove an analogous result for our quantum version of algebra of distributions of $SL_2$. In order to state this result, we fix some notation \cite{Jan - book}.
\medspace
Let $T\leq G$ be a maximal split torus and $X = X(T)$ be the group of characters of $T$.
$R$ is the associated root system, $S$ is a fixed basis of $R$ and $R^+$ is the set of positive roots corresponding to $S$.
For each $\alpha\in R$, let $\alpha^{\vee}$ be the associated coroot, and
$\langle\beta,\alpha^{\vee}\rangle$ denotes the natural pairing, with the normalization $\langle \alpha, \alpha^{\vee} \rangle = 2$ for all $\alpha\in S$.
We consider the following subsets of $X$:
\begin{align*}
X_+ &= \{\lambda\in X\mid \langle \lambda,\alpha^{\vee} \rangle \geq 0, \ \forall \alpha\in R^+\},
\mbox{ the set of dominant weights,} \\
X_r &= \{\lambda\in X_+\mid \langle\lambda,\alpha^{\vee}\rangle < p^r, \ \forall \alpha\in S\},
\text{ the set of $r$-restricted weights}, \ r\geq 1.
\end{align*}
Recall that the assignment $\lambda\mapsto L(\lambda)$ establishes a bijection between $X^+$ and the simple $G$-modules up to isomorphism.
Let $B\leq G$ be the Borel subgroup containing $T$ corresponding to $R^-=-R^+$. Given $\lambda\in X_+$ and $M$ a (rational) $G$-module, we set
\begin{itemize}[leftmargin=*]
\item $M_{\lambda} = \{m\in M\mid t.m = \lambda(t)m\mbox{ for all }t\in T\}$ is the $\lambda$-weight space;
\item $\nabla(\lambda) = \ind_B^G(\lambda)$, the costandard module of highest weight $\lambda$;
\item $L(\lambda) = \soc_G\nabla(\lambda)$, the simple module with highest weight $\lambda$.
\end{itemize}
Let $\cF: G\to G$ be the Frobenius morphism: it arises from the map $\Bbbk\to \Bbbk$, $x\mapsto x^p$. Then $M^{[r]}$ is the $G$-module over the underlying additive group $M$ with $G$-action obtained up to compose the original $G$-action with $\cF^r$.
\begin{theorem}\label{thm:Steinberg}
\cite[Proposition II.3.16]{Jan - book}
Let $r\in\N$, $\lambda\in X_r$, $\mu\in X_+$. Then
\begin{align*}
L(\lambda+p^r \mu)& \simeq L(\lambda)\otimes L(\mu)^{[r]}.
\end{align*}
\end{theorem}
\section{Some cleft extensions of $\u_\lambda(\Sl_2)$}
We introduce the algebras $\Dl{N}{\Sl_2}$, $N\in\N_0$, and prove some properties about their algebra structure which mimic \S \ref{section:dist alg}.
\subsection{$q$-numbers}
We use the following $q$-numbers as in \cite{L - book},
\begin{align}\label{eq:def qbinomial}
[m]_\lambda&:= \frac{\lambda^m-\lambda^{-m}}{\lambda-\lambda^{-1}}, &
\fac{m} &= (m)_\lambda (m-1)_\lambda \dots (1)_\lambda, \\
\bic{m}{n} &:= \prod_{j=1}^n \frac{\lambda^{m-j+1}-\lambda^{-m+j-1}}{\lambda^j-\lambda^{-j}}, &
0 & \leq n <\ell.
\end{align}
Let $\lambda$ a primitive root of unity of order $\ell$; we assume that $\ell>1$ is odd.
Now we need $q$-\emph{binomial numbers} associated to the $\ell$-expansion. Set
\begin{align}\label{eq:def gen qbinomial}
\bil{m}{n} &:= \prod_{i\geq 0} \bic{m_i}{n_i}, & m &=\sum_{i\geq0}m_i \ell^i, \, n=\sum_{i\geq0}n_i \ell^i, \, 0\leq m_i,n_i<\ell.
\end{align}
\begin{lemma}
Let $m,n,p\geq 0$. Then
\begin{align} \label{eq:prop qbil-1}
\bil{m+n}{m}&=\bil{m+n}{n}, \\ \label{eq:prop qbil-2}
\bil{m+n}{n}\bil{m+n+p}{p}&=\bil{n+p}{n}\bil{m+n+p}{m}.
\end{align}
\end{lemma}
\pf
For \eqref{eq:prop qbil-1}, if $m_i+n_i<\ell$ for all $i$, then $(m+n)_i=m_i+n_i$ and
$$ \bil{m+n}{m} = \prod_{i\geq 0} \bic{m_i+n_i}{n_i} = \bil{m+n}{n}. $$
Otherwise there exists $i\geq 0$ such that $m_i+n_i\geq\ell$, we assume $i$ is minimal with this property.
Thus $(m+n)_i=m_i+n_i-\ell< m_i,n_i$, and both sides are 0.
For \eqref{eq:prop qbil-2}, if $m_i+n_i+p_i<\ell$ for all $i$, then $(m+n+p)_i=m_i+n_i+p_i$ and
$$ \bil{m+n}{n}\bil{m+n+p}{p} = \prod_{i\geq 0} \frac{\fac{m_i+n_i+p_i}}{\fac{m_i}\fac{n_i}\fac{p_i}} = \bil{n+p}{n}\bil{m+n+p}{m}. $$
Otherwise there exists $i\geq 0$ such that $m_i+n_i+p_i\geq\ell$, we assume $i$ is minimal with this property.
\begin{itemize}
\item If $m_i+n_i, n_i+p_i\geq \ell$, then $\bil{m+n}{n}=\bil{n+p}{n}=0$ since $(m+n)_i=m_i+n_i-\ell<n_i$, $(n+p)_i=n_i+p_i-\ell<n_i$.
\item If $m_i+n_i\geq \ell>n_i+p_i$, then $\bil{m+n}{n}=\bil{m+n+p}{m}=0$ since $(m+n)_i=m_i+n_i-\ell<n_i$, $(n+p)_i=n_i+p_i$, $(m+n+p)_i=m_i+n_i+p_i-\ell<m_i$.
\item Finally, if $m_i+n_i, n_i+p_i<\ell$, then $\bil{m+n+p}{p}=\bil{m+n+p}{m}=0$.
\end{itemize}
In all the cases, both sides of \eqref{eq:prop qbil-2} are 0.
\epf
\subsection{The Hopf algebra $\u_\lambda(\Sl_2)$}
Throughout this work $\u_\lambda(\Sl_2)$ denotes the algebra presented by generators $E$, $K$, $F$, and relations
\begin{align}\label{eq:defn-rels-uqsl2-1}
& K^\ell=1, & & KF=\lambda^{-2} \, FK, & & \\ \label{eq:defn-rels-uqsl2-2}
&E^\ell=F^\ell=0, & & KE=\lambda^2 \, EK, & & EF-FE = \frac{K-K^{-1}}{\lambda-\lambda^{-1}}.
\end{align}
It is slightly different from the small quantum group appearing in \cite{L - mod rep}. Then $\u_\lambda(\Sl_2)$ is a Hopf algebra with coproduct:
\begin{align*}
\Delta(K) &= K\ot K, & \Delta(E)&=E\ot 1+K\ot E, &
\Delta(F)=F\ot K^{-1} + 1\ot F.
\end{align*}
\medbreak
Let $\u_\lambda^+(\Sl_2)$, respectively $\u_\lambda^0(\Sl_2)$, $\u_\lambda^-(\Sl_2)$, be the subalgebra spanned
by $E$, respectively $K$, $F$. Then $\u_\lambda^0(\Sl_2) \simeq \Bbbk \Z/\ell\Z$ while
$\u_\lambda^{\pm}(\Sl_2)$ are isomorphic to $\Bbbk[x]/<x^\ell>$. The multiplication induces a linear isomorphism
$$ \u_\lambda^-(\Sl_2) \otimes \u_\lambda^0(\Sl_2) \otimes \u_\lambda^+(\Sl_2) \simeq \u_\lambda(\Sl_2) . $$
Thus $\{F^a K^b E^c | 0\le a,b,c<\ell \}$ is a basis of $\u_\lambda(\Sl_2)$ and $\dim \u_\lambda(\Sl_2)=\ell^3$.
To simplify the notation in forthcoming computations, let
\begin{align*}
E\at{a} &= \frac{E^a}{\fac{a}}, & F\at{a}&= \frac{F^a}{\fac{a}}, &
\bic{K;s}{a}&= \prod_{j=1}^a \frac{\lambda^{s-j+1}K-\lambda^{-s+j-1}K^{-1} }{\lambda^j- \lambda^{-j}}.
\end{align*}
By direct computation,
\begin{align}\label{eq:conmut-En-Fn}
E\at{m}F\at{n} &= \sum_{i=0}^{\min\{ m,n\} } F\at{n-i} \bic{K;2i-m-n}{i} E\at{m-i},
& 0\le & m,n <\ell.
\end{align}
Let $\u_\lambda^{\ge 0}(\Sl_2)$ be the subalgebra spanned by $E$ and $K$. For each $0\le z<\ell$ the 1-dimensional representation
$\Bbbk_z$ of $\u_\lambda^0(\Sl_2) \simeq \Bbbk \Z/\ell\Z$ given by $K\mapsto \lambda^z$ can be extended to $\u_\lambda^{\ge 0}(\Sl_2)$ by
$E\mapsto 0$. Let $\cM(z)= \u_\lambda(\Sl_2)\otimes_{\u_\lambda^{\ge 0}(\Sl_2)} \Bbbk_z$: it is a $\u_\lambda(\Sl_2)$-module with basis
$v_j:= F\at{j}\otimes 1$, $0\le j<\ell$, such that for all $0\le m,n<\ell$
\begin{align}\label{eq:verma-uqsl2-Fm-vn}
F\at{m} \cdot v_n &= \bic{m+n}{m} v_{m+n}, & K\cdot v_n &=\lambda^{z-2n} v_n,
\\ \label{eq:verma-uqsl2-Em-vn}
E\at{m} \cdot v_n &= \bic{z+m-n}{m} v_{n-m}.
\end{align}
Here $v_n=0$ if either $n<0$ or $n\ge \ell$. Each module $\cM(z)$ has a maximal proper submodule $\cN(z)$. The quotient $\cL(z)=\cM(z)/\cN(z)$ is simple, and has dimension $z+1$: indeed $(v_i)_{0\le i\le z}$ is a basis of $\cL(z)$. Moreover, the family $\{\cL(z)\}_{0\le z<\ell}$ is a set of representatives of the classes of simple modules up to isomorphism.
\subsection{The cleft extensions $\Dl{N}{\Sl_2}$}
We mimic the definition by generators and relations of the algebra of distributions, but in a \emph{quantized} context.
\begin{definition}\label{defn:Dlambda}
Let $\Dl{N}{\Sl_2}$ be the algebra defined by generators $E\atb{i}$, $F\atb{i}$, $K\atb{i}$, $0\le i\le N$ and relations
\begin{align}
K\atb{i} K\atb{j} &= K\atb{j}K\atb{i}, & \left(K\atb{i}\right)^{\ell}&=1; \label{eq:rel D - comm K} \\
K\atb{i} E\atb{j} &=\lambda^{2\delta_{ij}} E\atb{j}K\atb{i}, & K\atb{i} F\atb{j} &=\lambda^{-2\delta_{ij}} F\atb{j}K\atb{i};
\label{eq:rel D - act K over E,F}\\
E\atb{i} E\atb{j} &= E\atb{j}E\atb{i}, & F\atb{i} F\atb{j} &= F\atb{j}F\atb{i}; \label{eq:rel D - comm E - comm F}\\
\left(E\atb{i}\right)^{\ell} &= \left(F\atb{i}\right)^{\ell}=0; &
E\atb{i}F\atb{j}&= F\atb{j}E\atb{i}, \quad j\neq i; \label{eq:rel D - powers E,F}
\end{align}
\begin{equation}\label{eq:rel D - bracket E,F}
E\atb{j}F\atb{j} = \sum_{t=0}^{\ell^j} F\at{\ell^j-t} \bik{K;2t-2\ell^j}{t} E\at{\ell^j-t}.
\end{equation}
Here,
$K\atb{-i}:=(K\atb{i})^{-1}$; for $m=\sum_{i=0}^N m_i\ell^i$, $s=\sum_{i=0}^N s_i\ell^i$, $t=\sum_{i=0}^N t_i\ell^i$, $0\le m_i,s_i, t_i<\ell$,
\begin{align*}
E\at{m} &:= \prod_{i=0}^N \frac{\left(E\atb{i}\right)^{m_i}}{\fac{m_i}},
&
\bik{K;s}{t} &=\prod_{i=0}^N \bic{K\atb{i};s_i}{t_i},
&
F\at{m} &:= \prod_{i=0}^N \frac{\left(F\atb{i}\right)^{m_i}}{\fac{m_i}}.
\end{align*}
\end{definition}
\medbreak
\begin{remark}
$\Dl{N}{\Sl_2}$ is $\Z$-graded, with
\begin{align*}
\deg E\atb{i}&=-\deg F\atb{i}=\ell^i, & \deg K\atb{i}&=0, & 0\le & i\le N.
\end{align*}
\end{remark}
\begin{lemma}\label{lemma:proj over uqsl2}
For each pair $M<N$, there exists a surjective algebra map $\pi_{M,N}:\Dl{N}{\Sl_2} \to \Dl{M}{\Sl_2}$ such that
\begin{align*}
\pi_N(X\atb{i})&=
\begin{cases} X\atb{i-N+M}, & i\ge N-M, \\ 0 & i<N-M, \end{cases}
& X\in\{E,F,K\}.
\end{align*}
In particular, there exists a surjective algebra map $\pi_N:\Dl{N}{\Sl_2} \to \u_\lambda(\Sl_2)$,
\begin{align*}
\pi_N(E\atb{i})&=\delta_{iN}E, & \pi_N(F\atb{i})&=\delta_{iN}F, & \pi_N(K\atb{i})&=K^{\delta_{iN}}, & 0\le & i\le N.
\end{align*}
\end{lemma}
\pf
Straightforward.
\epf
Let $\Dp{N}{\Sl_2}$, resp. $\Dm{N}{\Sl_2}$, $\Dz{N}{\Sl_2}$, be the subalgebras generated by $E\atb{i}$, resp. $F\atb{i}$, $K\atb{i}$, $0\le i\le N$.
Let $\Dge{N}{\Sl_2}$, resp. $\Dle{N}{\Sl_2}$, be the subalgebras generated by $E\atb{i}$ and $K\atb{i}$, resp. $F\atb{i}$ and $K\atb{i}$.
\begin{remark}
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item There exists an algebra antiautomorphism $\phi_N$ of $\Dp{N}{\Sl_2}$ such that $\phi_N(E\atb{i})=F\atb{i}$, $\phi_N(F\atb{i})=E\atb{i}$,
$\phi_N(K\atb{i})=K\atb{i}$, $0\le i\le N$.
\item There exists an algebra map $\iota_N:\Dl{N-1}{\Sl_2} \to \Dl{N}{\Sl_2}$ which identifies the corresponding generators. Clearly, $\phi_N \circ \iota_N=\iota_N \circ \phi_{N-1}$.
\end{enumerate}
\end{remark}
\begin{lemma}\label{lemma:Verma-modules}
Let $z=\sum_{i=0}^N z_i\ell^i$, $0\le z_i<\ell$. There exists a $\Dl{N}{\Sl_2}$-module $M(z)$ with basis $(v_t)_{0\le t\le \ell^{N+1}-1}$ such that
\begin{align}\label{eq:Verma-modules-action-1}
E\atb{i}\cdot v_t&=\left\{ \begin{array}{ll} \bra{z_i+1-t_i} v_{t-\ell^i}, & t_i>0, \\ 0, & t_i=0; \end{array} \right. &
K\atb{i}\cdot v_t&=\lambda^{z_i-2t_i} v_t, \\ \label{eq:Verma-modules-action-2}
F\atb{i}\cdot v_t&= \bra{t_i+1} v_{t+\ell^i}, & 0\le & i\le N.
\end{align}
\end{lemma}
\pf
We simply check that $E\atb{i}, F\atb{i}, K\atb{i}\in\End M(z)$, $0\le i\le N$,
satisfy relations \eqref{eq:rel D - comm K}--\eqref{eq:rel D - bracket E,F}. First equation of \eqref{eq:rel D - comm K} holds since
$K\atb{i} K\atb{j}\cdot v_t=\lambda^{z_i+z_j-2t_i-2t_j} v_t$, while the second follows since $\lambda^\ell=1$.
For the first relation in \eqref{eq:rel D - act K over E,F}, both sides annihilate $v_t$ if $t_j=0$; for $t_j\neq 0$, $(t-\ell^j)_i=t_i-\delta_{ij}$, so
\begin{align*}
K\atb{i} E\atb{j}\cdot v_t &= \bra{z_j+1-t_j} \lambda^{z_i-2(t-\ell^j)_i} v_{t-\ell^j}= \lambda^{2\delta_{ij}}
E\atb{j}K\atb{i} \cdot v_t.
\end{align*}
For the second relation, both sides annihilate $v_t$ if $t_j=\ell-1$; for $t_j<\ell-1$,
\begin{align*}
K\atb{i} F\atb{j}\cdot v_t &= \bra{t_j+1} \lambda^{z_i-2(t+\ell^j)_i} v_{t+\ell^j}= \lambda^{-2\delta_{ij}}
F\atb{j}K\atb{i} \cdot v_t,
\end{align*}
since $(t+\ell^j)_i=t_i+\delta_{ij}$. For the first equation in \eqref{eq:rel D - comm E - comm F}, if $t_it_j\neq 0$, $i\neq j$, then
\begin{align*}
E\atb{i} E\atb{j}\cdot v_t &= \bra{z_j-t_j+1} \bra{z_i-(t-\ell^j)_i+1} v_{t-\ell^i-\ell^j} \\
& = \bra{z_j-t_j+1} \bra{z_i-t_i+1} v_{t-\ell^i-\ell^j} = E\atb{j} E\atb{i} \cdot v_t,
\end{align*}
while for $t_it_j=0$, both sides are $0$. The second equation follows similarly.
For the first part of \eqref{eq:rel D - powers E,F}, $(E\atb{i})^{t_i+1}\cdot v_t=(F\atb{i})^{\ell-t_i}\cdot v_t=0$, so
$(E\atb{i})^{\ell}$, $(F\atb{i})^{\ell}$ are 0 as operators on $M(z)$. For the second equality, fix $i\neq j$. If $t_i=0$,
then either $(t+\ell^j)_i=t_i$ or else $t_j=\ell-1$; in any case, $E\atb{i}F\atb{j}\cdot v_t=0=F\atb{j}E\atb{i}\cdot v_t$.
If $t_j=\ell-1$, then again both sides are $0$. Finally set $t_i\neq 0$, $t_j\neq \ell-1$. Hence,
\begin{align*}
E\atb{i}F\atb{j}\cdot v_t &= \bra{t_j+1} \bra{z_i-(t+\ell^j)_i+1} v_{t-\ell^i+\ell^j} \\
& = \bra{(t-\ell^i)_j+1} \bra{z_i-t_i+1} v_{t-\ell^i+\ell^j} = F\atb{j}E\atb{i} \cdot v_t,
\end{align*}
It remains to consider \eqref{eq:rel D - bracket E,F}, which can be written as
\begin{equation}\label{eq:rel D - bracket E,F-v2}
E\atb{j}F\atb{j}-F\atb{j}E\atb{j} -\frac{K\atb{j}-K\atb{-j}}{\lambda-\lambda^{-1}} = \sum_{s=1}^{\ell^j-1} F\at{\ell^j-s} \bik{K;2s}{s} E\at{\ell^j-s}.
\end{equation}
If $1\le s\le \ell^j-1$, then there exists $i<j$ such that $s_i\neq 0$. If $t_i\geq \ell-s_i$, then
\begin{align*}
\bic{K\atb{i};2s_i}{s_i} (E\atb{i})^{\ell-s_i}\cdot v_t &= \prod_{k=1}^{\ell-s_i} \bra{z_i-t_i+k} \prod_{k=1}^{s_i} \bra{z_i-t_i+k} v_{t-(\ell-s_i)\ell^i}=0.
\end{align*}
If $t_i<\ell-s_i$, then $(E\atb{i})^{\ell-s_i}\cdot v_t=0$. In any case, $\bik{K;2s}{s} E\at{\ell^j-s} \cdot v_t=0$, so the right-hand side of \eqref{eq:rel D - bracket E,F-v2} acts by 0 on each $v_t$. For the left-hand side,
\begin{multline*}
\left( E\atb{j}F\atb{j}-F\atb{j}E\atb{j} -\frac{K\atb{j}-K\atb{-j}}{\lambda-\lambda^{-1}}\right) \cdot v_t \\
= \left( \bra{t_j+1}\bra{z_j-t_j} - \bra{z_j-t_j+1}\bra{t_j}-\bra{z_j-2t_j} \right) v_t=0,
\end{multline*}
when $t_j\neq 0,\ell-1$. If $t_j=0$, then
\begin{multline*}
\left( E\atb{j}F\atb{j}-F\atb{j}E\atb{j} -\frac{K\atb{j}-K\atb{-j}}{\lambda-\lambda^{-1}}\right) \cdot v_t = E\atb{j}\cdot v_{t+\ell^j}-0-\bra{z_j} v_t=0,
\end{multline*}
and finally if $t_j=\ell-1$, then
\begin{align*}
\left( E\atb{j}F\atb{j}-F\atb{j}E\atb{j} -\frac{K\atb{j}-K\atb{-j}}{\lambda-\lambda^{-1}}\right) \cdot v_t
=-\bra{z_j+2}\left( F\atb{j}\cdot v_{t-\ell^j}+ v_t\right)=0.
\end{align*}
In any case, the left-hand side of \eqref{eq:rel D - bracket E,F-v2} also acts by 0 on each $v_t$.
\epf
\begin{lemma}\label{lemma:comodule alg}
There exists an algebra map $\rho_N: \Dl{N}{\Sl_2} \to \u_\lambda(\Sl_2)\otimes \Dl{N}{\Sl_2}$,
\begin{align*}
\rho_N(E\atb{i}) &= 1\otimes E\atb{i}, \quad i<N, & \rho_N(E\atb{N}) &= E \otimes 1 + K\otimes E\atb{N}, \\
\rho_N(F\atb{i}) &= 1\otimes F\atb{i}, \quad i<N, & \rho_N(F\atb{N}) &= F \otimes K\atb{-N} + 1\otimes F\atb{N}, \\
\rho_N(K\atb{i}) &= 1\otimes K\atb{i}, \quad i<N, & \rho_N(K\atb{N}) &= K\otimes K\atb{N}.
\end{align*}
Moreover $\Dl{N}{\Sl_2}$ is a left $\u_\lambda(\Sl_2)$-comodule algebra with this map.
\end{lemma}
\pf
Let $\F$ be the free algebra generated by $E\atb{i}$, $F\atb{i}$, $K\atb{i}$, and $\widetilde\rho_N:\F\to \u_\lambda(\Sl_2)\otimes \Dl{N}{\Sl_2}$
the map defined on the generators as $\rho_N$. We check that $\widetilde\rho_N$ annihilates each defining relation of $\Dl{N}{\Sl_2}$ so it induces the algebra map $\rho_N$. For each relation $\mathtt{r}$ involving only generators $E\atb{i}$, $F\atb{i}$, $K\atb{i}$, $0\le i< N$ we have that $\widetilde\rho_N(\mathtt{r})=1\otimes \mathtt{r}=0$, so we consider those relations involving at least one of the generators $E\atb{N}$, $F\atb{N}$, $K\atb{N}$.
For \eqref{eq:rel D - comm K}, $\widetilde\rho_N \left( (K\atb{N}) ^{\ell} \right)= K^{\ell}\otimes (K\atb{N})^{\ell} = 1 \otimes 1$ and for $i<N$,
$$ \widetilde\rho_N \left( K\atb{i} K\atb{N} -K\atb{N}K\atb{i} \right)= K\otimes \left( K\atb{i} K\atb{N} -K\atb{N}K\atb{i} \right)=0. $$
For \eqref{eq:rel D - act K over E,F} and \eqref{eq:rel D - comm E - comm F}, if $i<N$, then
\begin{align*}
\widetilde\rho_N (K\atb{i} E\atb{N} &- E\atb{N}K\atb{i})= K\otimes (K\atb{i} E\atb{N} - E\atb{N}K\atb{i})=0, \\
\widetilde\rho_N (K\atb{N} E\atb{i} &- E\atb{i}K\atb{N})= K\otimes (K\atb{N} E\atb{i} - E\atb{i}K\atb{N})=0, \\
\widetilde\rho_N (K\atb{N} E\atb{N} &-\lambda^2 E\atb{N}K\atb{N})= (KE-\lambda^2 EK)\otimes K\atb{N} \\
& \qquad + K^2\otimes (K\atb{N} E\atb{N} -\lambda^2 E\atb{N}K\atb{N})=0, \\
\widetilde\rho_N(E\atb{i} E\atb{N} &- E\atb{N}E\atb{i})= K\otimes (E\atb{i} E\atb{N} - E\atb{N}E\atb{i})=0.
\end{align*}
The formulas with $F$ in place of $E$ follow analogously.
For \eqref{eq:rel D - powers E,F},
\begin{align*}
\widetilde\rho_N \left( (E\atb{N})^{\ell}\right) &= \sum_{j=0}^\ell \bil{\ell}{j} E^{\ell-j} K^j\otimes (E\atb{N})^{j}=0,
\end{align*}
and analogously $\widetilde\rho_N \left( (F\atb{N})^{\ell}\right) =0$.
Finally, for \eqref{eq:rel D - bracket E,F} set $\mathtt{r}_{N}$ as the difference between the two sides of this equation, see also
\eqref{eq:rel D - bracket E,F-v2}. By direct computation,
\begin{align*}
\widetilde\rho_N ( \mathtt{r}_{N} ) & = K \otimes \mathtt{r}_{N}+ \Big(EF-FE- \frac{K-K^{-1}}{\lambda-\lambda^{-1}} \Big) \otimes K\atb{-N} =0.
\end{align*}
Then $\rho_N$ is a well defined algebra map, and gives a left $\u_\lambda(\Sl_2)$-coaction.
\epf
\begin{prop}\label{prop:cleft-ext}
Let $\rho_N$ as above. Then $\iota_N(\Dl{N-1}{\Sl_2})= ^{\co\rho_N} \Dl{N}{\Sl_2}$, and $\iota_N(\Dl{N-1}{\Sl_2}) \subset \Dl{N}{\Sl_2}$ is a
$\u_\lambda(\Sl_2)$-cleft extension.
\end{prop}
\pf
Let $\gamma:\u_\lambda(\Sl_2)\to \Dl{N}{\Sl_2}$ be the linear map such that
\begin{align}\label{eq:section}
\gamma(F\at{a} K^b E\at{c})&=\frac{(F\atb{N})^a}{\fac{a}} (K\atb{N})^b \frac{(E\atb{N})^c}{\fac{c}}, & 0\le & a,b,c<\ell.
\end{align}
By direct computation,
\begin{align*}
(\id\otimes\gamma)\circ \Delta(F\at{a} K^b E\at{c}) &= \sum_{i,j} F\at{a-i} K^{b+i+j} E\at{c-j}\otimes \frac{(F\atb{N})^i}{\fac{i}} (K\atb{N})^b \frac{(E\atb{N})^j}{\fac{j}} \\
&= \rho\circ\gamma (F^a K^b E^c),
\end{align*}
so $\gamma$ is map of $\u_\lambda(\Sl_2)$-comodules. We claim that $\gamma$ is convolution invertible. By \cite[Lemma 5.2.10]{Mo},
it is enough to restrict $\gamma$ to the coradical of $\u_\lambda(\Sl_2)$, that is, to $\u^0_\lambda(\Sl_2)$.
Now $\kappa:\u^0_\lambda(\Sl_2)\to \Dl{N}{\Sl_2}$, $\kappa(K^b)= (K\atb{N})^{-b}$, $0\le b<\ell$ is the inverse of
$\gamma_{|\u^0_\lambda(\Sl_2)}$ and the claim follows.
Let $B_N:=\{ F\at{m} K\at{n} E\at{p} | 0\le m,n,p < \ell^{N+1} \}$. We claim that $\Dl{N}{\Sl_2}$ is spanned by $B_N$\footnote{In Proposition
\ref{prop:basis, triang decomp} we shall prove that $B_N$ is indeed a basis of $\Dl{N}{\Sl_2}$}.
Let $I$ be the subspace spanned by $B_N$. Note that $I$ is a left ideal, since it is stable by
left multiplication by $F\atb{n}$, $K\atb{n}$ and $E\atb{n}$ by \eqref{eq:rel D - comm K}-\eqref{eq:rel D - bracket E,F}.
Thus $I=\Dl{N}{\Sl_2}$ since $1\in I$, so $\Dl{N}{\Sl_2}$ is spanned by $B_N$. As
$$ F\at{m} K\at{n} E\at{p}=F\at{m'} K\at{n'} E\at{p'} \frac{(F\atb{N})^{m_N}}{\fac{m_N}} (K\atb{N})^{n_N} \frac{(E\atb{p_N})^c}{\fac{p_N}}, $$
where $0\le m'=m-m_N\ell^N, n'=n-n_N\ell^N, p'=p-p_N\ell^N <\ell^N$, and $F\at{m'} K\at{n'} E\at{p'}\in \iota_N(\Dl{N-1}{\Sl_2})$, we have that
$$ \dim \Dl{N}{\Sl_2}\le \dim \iota_N(\Dl{N-1}{\Sl_2}) \ell^3. $$
As we have a cleft extension, $\Dl{N}{\Sl_2}\simeq ^{\co\rho_N} \Dl{N}{\Sl_2}\otimes \u_\lambda(\Sl_2)$; using this fact and that
$\iota_N(\Dl{N-1}{\Sl_2}) \subset ^{\co\rho_N} \Dl{N}{\Sl_2}$ since $\iota_N$ sends each generator of $\Dl{N-1}{\Sl_2}$ to a coinvariant element, we have that
$$ \dim \Dl{N}{\Sl_2} = \dim ^{\co\rho_N} \Dl{N}{\Sl_2}\ell^3 \ge \dim \iota_N(\Dl{N-1}{\Sl_2}) \ell^3. $$
Hence $\dim ^{\co\rho_N} \Dl{N}{\Sl_2} = \dim \iota_N(\Dl{N-1}{\Sl_2})$, which means that these two subalgebras of $\Dl{N}{\Sl_2}$ coincide.
\epf
\begin{prop}\label{prop:basis, triang decomp}
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item\label{item:pos-neg-zero-parts} There exist algebra isomorphisms
\begin{align*}
\Dz{N}{\Sl_2} & \simeq \Bbbk(\Z_\ell)^{N+1}, & \Dge{N}{\Sl_2} &\simeq \left(\u_\lambda^{\ge 0}(\Sl_2)\right)^{N+1}, \\
\Dpm{N}{\Sl_2} & \simeq \left(\u_\lambda^{\pm}(\Sl_2)\right)^{N+1}, & \Dle{N}{\Sl_2} &\simeq \left(\u_\lambda^{\le 0}(\Sl_2)\right)^{N+1}.
\end{align*}
\item\label{item:PBW-basis} $B_N:=\{ F\at{m} K\at{n} E\at{p} | 0\le m,n,p < \ell^{N+1} \}$ is a basis of $\Dl{N}{\Sl_2}$.
\item\label{item:triang-decomp} The multiplication induces a linear isomorphism
$$ \Dm{N}{\Sl_2} \otimes \Dz{N}{\Sl_2} \otimes \Dp{N}{\Sl_2} \simeq \Dl{N}{\Sl_2} .$$
\end{enumerate}
\end{prop}
\pf
The algebra $\left(\u_\lambda(\Sl_2)\right)^{N+1}$ is generated by $\Et_i$, $\Ft_i$, $\Kt_i$, $0\le i\le N$,
where each 3-uple $\Et_i$, $\Ft_i$, $\Kt_i$ satisfy \eqref{eq:defn-rels-uqsl2-1}, \eqref{eq:defn-rels-uqsl2-2},
and generators with different subindex commute. There are algebra maps $\Phi^{\ddagger}:\left(\u_\lambda^{\ddagger}(\Sl_2)\right)^{N+1}\to
\mathcal{D}_{\lambda,N}^{\ddagger}(\Sl_2)$, $\ddagger\in\{\pm,0,\ge 0,\le 0\}$, where $\Et_i\mapsto E\atb{i}$, $\Ft_i\mapsto F\atb{i}$,
$\Kt_i\mapsto K\atb{i}$, depending on each case.
For $0\le z<\ell^{N+1}$, let $\Psi_z:\Dl{N}{\Sl_2}\to\End M(z)$ be the algebra map of Lemma \ref{lemma:Verma-modules}.
Notice that $\Psi_z\Phi^-$ is injective, and then $\Phi^-$ is so; thus $\Dm{N}{\Sl_2} \simeq \left(\u_\lambda^{-}(\Sl_2)\right)^{N+1}$.
The map $\Phi^{0}:\Bbbk(\Z_\ell)^{N+1} \to \Dz{N}{\Sl_2}$, $\alpha_i\mapsto K_i$ is surjective. The action of $\Bbbk(\Z_\ell)^{N+1}$ over $v_0$
is given by character $K_i\mapsto \lambda^{z_i}$. Thus $\Bbbk(\Z_\ell)^{N+1} \simeq \Dz{N}{\Sl_2}$. From here we derive that $\Phi^{\le 0}$ is also
an isomorphism. The remaining isomorphisms in \ref{item:pos-neg-zero-parts} follow by using the antiautomorphism $\phi$.
For \ref{item:PBW-basis}, we have to prove that $B_N$ is linearly independent since we have proved that $\Dl{N}{\Sl_2}$ is spanned by $B_N$
in the proof of Proposition \ref{prop:cleft-ext}. We invoke Diamond Lemma \cite[Theorem 1.2]{B - diamond}. Indeed, the lexicographical order
for words written with letters $\{F\atb{i}, K\atb{i},E\atb{i}\}_{0\le i\le N}$ such that
$$ F\atb{0}< \dots< F\atb{N} < K\atb{0}< \dots< K\atb{N} <E\atb{0}< \dots< E\atb{N} $$
is \emph{compatible} (in the notation of loc. cit.) with the reduction system. Each element of $B_N$ is \emph{irreducible}, so $B_N$ is contained
in a basis of $\Dl{N}{\Sl_2}$. Thus $B_N$ is a linearly independent set.
Finally \ref{item:triang-decomp} follows \ref{item:pos-neg-zero-parts} and \ref{item:PBW-basis}.
\epf
\begin{definition}
By Proposition \ref{prop:basis, triang decomp} \ref{item:PBW-basis} each $\iota_N$ is injective. Hence we may consider $\Dl{N-1}{\Sl_2}$ as a subalgebra of $\Dl{N}{\Sl_2}$. Moreover we can consider the inclusions $\iota_{M,N}:\Dl{M}{\Sl_2}\to \Dl{N}{\Sl_2}$ for $M\leq N$, where
\begin{align*}
\iota_{N,N}&=\id_{\Dl{N}{\Sl_2}}, & \iota_{M,N}&=\iota_{M}\iota_{M+1}\dots \iota_{N-1}\mbox{ for }M<N.
\end{align*}
Then we define
\begin{align}\label{eq:defn-Dlambda}
\Dli{\Sl_2} := \lim_{\rightarrow} \Dl{N}{\Sl_2}.
\end{align}
\end{definition}
\section{Finite-dimensional irreducible $\Dl{N}{\Sl_2}$-modules}
Next we study simple modules for the algebras $\Dl{N}{\Sl_2}$. We prove that they are highest weight modules as we can expect, and obtain a decomposition related with the inclusion $\iota_{N-1,N}:\Dl{N-1}{\Sl_2}\to \Dl{N}{\Sl_2}$ and the \emph{Frobenius map} $\pi_{N}:\Dl{N}{\Sl_2}\to \u_\lambda(\Sl_2)$. The tensor product decomposition can be seen as an analogous of Steinberg decomposition, c.f. Theorem \ref{thm:Steinberg}.
\subsection{Highest weight modules}
Now we mimic what is done for simple modules of quantum groups, e. g. \cite[\S 6 \& 7]{L - mod rep}. For the sake of completeness we include the proofs.
\medbreak
Let $V$ be a finite dimensional $\Dl{N}{\Sl_2}$-module. As $\Dz{N}{\Sl_2}$ is the group algebra of $\Z_\ell^{N+1}$, $V$ decomposes as the direct sum of eigenspaces: each $K\atb{i}$ acts by a scalar $\lambda^{p_i}$, $0\le p_i<\ell$. Hence we may encode the data saying that $V=\oplus_{0\le p<\ell^{N+1}} V_p$, where
\begin{align}\label{eq:defn-weight-sp}
V_p&:= \{ v\in V | K\atb{i}\cdot v=\lambda^{p_i} v \mbox{ for all } 0\le i\le N \}, & p&=\sum_{i=0}^{N}p_i\ell^i.
\end{align}
\begin{definition}
We say that $v\in V$ is a \emph{primitive vector} of weight $p$ if $v\in V_p$ and $E\atb{i}\cdot v=0$ for all $0\le i\le N$. $V$ is called a \emph{highest weight module} if it is generated (as $\Dl{N}{\Sl_2}$-module) by a primitive vector $v$, which is called a \emph{highest weight vector}; its weight $p$ is called a \emph{highest weight}.
\end{definition}
\smallbreak
Given $0\leq p<\ell^{N+1}$, let $\Bbbk_p$ be the 1-dimensional representation of $\Dge{N}{\Sl_2}\simeq \left(\u_\lambda^{\ge 0}(\Sl_2)\right)^{N+1}$ such that $K\atb{i}\cdot 1 =\lambda^{p_i}$ and $E\atb{i}\cdot 1=0$. Let
\begin{align*}
\verma{N}{p} &=\Ind{\Dge{N}{\Sl_2}}{\Dl{N}{\Sl_2}} \Bbbk_p \simeq \Dl{N}{\Sl_2}\otimes_{\Dge{N}{\Sl_2}} \Bbbk_p.
\end{align*}
Notice that $v_0:= 1\otimes 1\in \verma{N}{p}$ is a primitive vector, and moreover $\verma{N}{p}$ is a highest weight module with highest weight $p$.
\begin{remark}\label{rem:Verma-modules}
Let $v_t=F\at{t}v_0\in \verma{N}{p}$. Then $(v_t)_{0\le t<\ell^{N+1}}$ is a basis of $\verma{N}{p}$, and $\verma{N}{p}$
is isomorphic the module $M(p)$ in Lemma \ref{lemma:Verma-modules}.
Moreover the action on the basis $(v_t)_{0\le t<\ell^{N+1}}$ is given by formulas \eqref{eq:Verma-modules-action-1} and \eqref{eq:Verma-modules-action-2}.
Indeed there is a $\Dge{N}{\Sl_2}$-linear map $\Bbbk_p\to M(p)$ such that $1\mapsto v_0$; it induces a $\Dl{N}{\Sl_2}$-linear map $\verma{N}{p}\to M(p)$, which is surjective by direct computation, and both modules have dimension $\ell^{N+1}$.
\end{remark}
\begin{remark}
Let $V$ a highest weight module of weight $p$. Then each proper submodule is contained in $\oplus_{t\neq p} V_p$; hence $V$ has a maximal proper submodule $\widehat{V}$ and $V/\widehat{V}$ is a simple $\Dl{N}{\Sl_2}$-module, and at the same time a highest weight module of highest weight $p$.
\end{remark}
\begin{definition}
Let $\irr{N}{p}:= \verma{N}{p}/\widehat{\verma{N}{p}}$; that is, the simple highest weight module obtained as a quotient of $\verma{N}{p}$.
\end{definition}
\begin{prop}\label{prop:simple-modules}
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item\label{item:trivial-action-E} Let $0\le p<\ell^{N+1}$. Then
$$ \{v\in\irr{N}{p}|E\atb{i}v=0\mbox{ for all }0\le i\le N\}=\Bbbk v_0 .$$
\item\label{item:classif-simple-modules} There exists a bijection between $\{p|0\le p<\ell^{N+1}\}$ and the finite-dimensional simple modules of $\Dl{N}{\Sl_2}$ given by $p\mapsto \irr{N}{p}$.
\end{enumerate}
\end{prop}
\pf
\ref{item:trivial-action-E} Let $v\in\irr{N}{p}-0$ be such that $E\atb{i}v=0$ for all $0\le i\le N$.
We may assume that $v$ has weight $t$ for some $0\le t<\ell^{N+1}$, since $E\atb{i}$ applies each eigenspace of the $\Dz{N}{\Sl_2}$ to another.
Thus $v=a\, v_n$ for some $a\in\Bbbk^\times$ and some $0\le n<\ell^{N+1}$, since each 1-dimensional summand in the decomposition
$\verma{N}{p}=\oplus_{0\le n<\ell^{N+1}} \Bbbk v_n$ corresponds to a different eigenspace for the action of
$\Dz{N}{\Sl_2}\simeq \Bbbk (\Z_\ell)^{N+1}$. As $\irr{N}{p}$ is simple, $\irr{N}{p}=\Dl{N}{\Sl_2}v$, but
\begin{align*}
\Dl{N}{\Sl_2} v = \Dle{N}{\Sl_2} v = \Dle{N}{\Sl_2} v_n \subseteq \oplus_{n\le m<\ell^{N+1}} \Bbbk v_m.
\end{align*}
Hence $n=p$ and the claim follows.
\medskip
\ref{item:classif-simple-modules} Let $\cL$ be a simple $\Dl{N}{\Sl_2}$-module. As a $\Dz{N}{\Sl_2}$-module, $\cL=\oplus \cL_t$.
We pick $v\in\cL_t-0$. We may assume that $E\atb{i}v=0$ for all $0\le i\le N$. Indeed, if $E\atb{j}v=0$ for $j=0,\dots, i-1$ but
$E\atb{i}v\neq 0$, let $n\geq 0$ be such that $w:=(E\atb{i})^n v\neq 0$, $(E\atb{i})^{n+1}v=0$. Then $n<\ell$ since $(E\atb{i})^\ell=0$,
and $w$ satisfies $E\atb{j}w=0$ for $j=0,\dots, i$ since $E\atb{j}E\atb{i}=E\atb{i}E\atb{j}$.
Now there exists a $\Dge{N}{\Sl_2}$-linear map $\widetilde{\phi}:\Bbbk_t\to\cL$, $1\mapsto v$, which induces
a $\Dl{N}{\Sl_2}$-linear map $\phi:\verma{N}{t}\to\cL$ such that $1\mapsto v$. As $\cL$ is simple, $\Dl{N}{\Sl_2}v=\cL$,
so $\phi$ is surjective. Hence $\ker \phi\neq 0$ is a proper submodule of $\verma{N}{t}$ and $\cL\simeq \verma{N}{t}/\ker\phi$ is simple.
Thus $\cL\simeq \irr{N}{t}$.
By \ref{item:trivial-action-E}, $\irr{N}{p}\not\simeq\irr{N}{t}$ if $p\neq t$, and the claim follows.
\epf
\subsection{A tensor product decomposition}
\begin{prop}\label{prop:simple-module-trivial}
\begin{enumerate}[leftmargin=*,label=\rm{(\alph*)}]
\item\label{item:restriction-simple-pN=0} Let $0\le p<\ell^{N}$. Then
\begin{align}\label{eq:extension-N-1-to-N}
E\atb{N}\cdot v&=F\atb{N}\cdot v=0, & K\atb{N}\cdot v &=v, &
\mbox{for all }v &\in \irr{N}{p}.
\end{align}
Moreover, $\irr{N}{p}\simeq \irr{N-1}{p}$ as $\Dl{N-1}{\Sl_2}$-modules.
\item\label{item:extension-simple-pN=0} Reciprocally $\irr{N-1}{p}$ may be endowed of an $\Dl{N}{\Sl_2}$-action by extending the $\Dl{N-1}{\Sl_2}$-action via \eqref{eq:extension-N-1-to-N}, and $\irr{N-1}{p} \simeq \irr{N}{p}$ as $\Dl{N}{\Sl_2}$-modules.
\end{enumerate}
\end{prop}
\pf
\ref{item:restriction-simple-pN=0} By the first equation of \eqref{eq:Verma-modules-action-1}, $E\atb{i}v_{\ell^N}=0$
for all $0\le i\le N$, so $\Dl{N}{\Sl_2}v_{\ell^N}=\Dle{N}{\Sl_2}v_{\ell^N}=\oplus_{n\ge \ell^N}\Bbbk v_n$ is a proper submodule
of $\verma{N}{p}$. Hence $v_n=0$ in $\irr{N}{p}$ for all $n\ge\ell^N$, and $\irr{N}{p}$ is spanned by (the image of) $(v_m)_{0\le m<\ell^N}$.
Thus \eqref{eq:extension-N-1-to-N} follows by this fact and \eqref{eq:Verma-modules-action-1}-\eqref{eq:Verma-modules-action-2}.
By \eqref{eq:extension-N-1-to-N}, $W\subset \irr{N}{p}$ is a $\Dl{N-1}{\Sl_2}$-submodule if and only if $W$ is a $\Dl{N}{\Sl_2}$-submodule.
Hence $\irr{N}{p}$ is simple as $\Dl{N-1}{\Sl_2}$-module and the last statement follows.
\medbreak
\ref{item:extension-simple-pN=0} We have to check all the defining relations \eqref{eq:rel D - comm K}-\eqref{eq:rel D - bracket E,F}
of $\Dl{N}{\Sl_2}$. Those not involving $E\atb{N}$, $F\atb{N}$, $K\atb{N}$ follow since $\irr{N-1}{p}$ is a $\Dl{N-1}{\Sl_2}$-module,
and relations $E\atb{N}$, $F\atb{N}$, $K\atb{N}$ follow easily except \eqref{eq:rel D - bracket E,F} for $j=N$.
It is equivalent to \eqref{eq:rel D - bracket E,F-v2}, whose left-hand side acts by $0$ on each $v_t$. For the right-hand side, if
$1\le s\le \ell^N-1$, then there exists $i<N$ such that $s_i\neq 0$, and as in the proof of Lemma \ref{lemma:Verma-modules},
$\bik{K;2s}{s} E\at{\ell^j-s} \cdot v_t=0$, so the right-hand side of \eqref{eq:rel D - bracket E,F-v2} acts by 0 on each $v_t$.
Now $\irr{N-1}{p}$ is a highest weight module as $\Dl{N}{\Sl_2}$-module, with highest weight $p$, and simple at the same time, so
$\irr{N-1}{p} \simeq \irr{N}{p}$ as $\Dl{N}{\Sl_2}$-modules.
\epf
\begin{remark}\label{rem:scalar-ext-ul-Dl}
Thanks to the algebra map $\pi_N:\Dl{N}{\Sl_2} \to \u_\lambda(\Sl_2)$, every $\u_\lambda(\Sl_2)$-module is canonically a $\Dl{N}{\Sl_2}$-module.
In particular each simple $\u_\lambda(\Sl_2)$-module $\cL(p)$, $0\le p<\ell$, is a $\Dl{N}{\Sl_2}$-module.
\end{remark}
\begin{lemma}\label{lem:simple-module-trivial}
Let $p=p_N\ell^N$, $0\le p_N<\ell$. Then $\irr{N}{p}\simeq \cL(p_N)$.
\end{lemma}
\pf
As $\pi_N$ is surjective, $W$ is a $\Dl{N}{\Sl_2}$-submodule of $\cL(p_N)$ if and only if $W$ is a $\u_\lambda(\Sl_2)$-submodule.
Thus $\cL(p_N)$ is a simple $\Dl{N}{\Sl_2}$-module. Now
\begin{align*}
E\atb{i}v_0&=0, & K\atb{i}v_0&=\lambda^{p_N\delta_{iN}}v_0, & \mbox{for all }& 0\leq i\le N.
\end{align*}
Hence $v_0\in\cL(p_N)-0$ is a highest weight vector of weight $p=p_N\ell^N$ and the Lemma follows by Proposition \ref{prop:simple-modules}
\epf
\begin{remark}\label{rem:Dl-ul-tensor-modules}
Recall that $\Dl{N}{\Sl_2}$ is an $\u_\lambda(\Sl_2)$-comodule algebra, so the category of $\Dl{N}{\Sl_2}$-modules is a module category over the category of $\u_\lambda(\Sl_2)$-modules: Given a $\u_\lambda(\Sl_2)$-module $\cM$ and a $\Dl{N}{\Sl_2}$-module $\cN$, $\cM\otimes \cN$ is naturally a $\Dl{N}{\Sl_2}$-module via $\rho$.
\end{remark}
Finally we use Remark \ref{rem:Dl-ul-tensor-modules} to describe a \emph{tensor product decomposition} of simple $\Dl{N}{\Sl_2}$-modules.
\begin{theorem}\label{thm:simple-modules-tensor-decomp}
Let $p=p_N\ell^N+\widehat{p}$, where $0\le \widehat{p}<\ell^N$, $0\le p_N<\ell$. Then
\begin{align*}
\irr{N}{p} & \simeq \cL(p_N)\otimes \irr{N}{\widehat{p}} & \mbox{as }\Dl{N}{\Sl_2}-\mbox{modules.}
\end{align*}
\end{theorem}
\pf
Let $v_0'$, $v_0''$ be highest weight vectors of $\cL(p_N)$, $\irr{N}{\widehat{p}}$, respectively.
We denote $L=\cL(p_N)\otimes \irr{N}{\widehat{p}}$. As $\cL(p_N)$ is generated by $\{v_t'|0\le t<\ell\}$ as in \eqref{eq:verma-uqsl2-Fm-vn}, and $\irr{N}{\widehat{p}}$ is generated by $\{v_t''=F\atb{t}v_0''|0\le t<\ell^N\}$, see Proposition \ref{prop:simple-module-trivial},
$L$ is generated by $\{v_t=v_{t_N}'\otimes v_{\widehat{t}}'' | 0\le t=\widehat{t}+t_N\ell^N <\ell^{N+1}\}$.
Given $F\at{m} K\at{n} E\at{p}\in B_N$, $0\le m,n,p < \ell^{N+1}$,
we may write
\begin{align*}
F\at{m} K\at{n} E\at{p}&=F\at{m_N\ell^N} K\at{n_N\ell^N} E\at{p_N\ell^N}F\at{m'} K\at{n'} E\at{p'}, & 0\le &m',n',p'<\ell^N.
\end{align*}
Here, $F\at{m'} K\at{n'} E\at{p'}\in \Dl{N-1}{\Sl_2}= ^{\co\rho_N} \Dl{N}{\Sl_2}$, cf. Proposition \ref{prop:cleft-ext}. Thus
\begin{align*}
F\at{m} K\at{n} E\at{p}(y\otimes z) & = F\at{m_N}K\at{n_N}E\at{p_N}y \otimes F\at{m'} K\at{n'} E\at{p'} z,
\end{align*}
for all $y\in \cL(p_N)$, $z\in\irr{N}{\widehat{p}}$, where we use \eqref{eq:extension-N-1-to-N}. From here,
$v_0=v_{0}'\otimes v_{0}''$ is a primitive vector, and $L$ is a highest weight module of highest weight $p$. Thus
it suffices to prove that $L$ is simple. Let $W$ be a submodule of $L$. In particular, $W$ is a $\Dz{N}{\Sl_2}$-submodule,
so it decomposes as a direct sum of eigenspaces; each $v_t$, $0\le t<\ell^{N+1}$, spans the eigenspace of weight $t$, so
we may assume that $v_t\in W$ for some $t$. Let $t$ be minimal. Hence
\begin{align*}
0&=E\atb{N}v_t=E \, v_{t_N}'\otimes v_{\widehat{t}}'', & 0&=E\atb{j}v_t=v_{t_N}'\otimes E\atb{j}v_{\widehat{t}}'', \, 0\le j<N,
\end{align*}
so $E \, v_{t_N}'=0=E\atb{j}v_{\widehat{t}}''$, $0\le j<N$. From here, $t_N=\widehat{t}=0$, and then $W=L$.
\epf
\begin{remark}
$\Dl{N}{\Sl_2}$ is an \emph{augmented algebra} via the map $\eps: \Dl{N}{\Sl_2}\to \Bbbk$,
\begin{align}\label{eq:DlN-augmented}
\epsilon(E\atb{j})&=\epsilon(F\atb{j})=0, & \epsilon(K\atb{j})&=1, &
\mbox{for all }0 &\le j\le N.
\end{align}
Thence $\Bbbk$ is a $\Dl{N}{\Sl_2}$-module and $\Bbbk\simeq \irr{N}{0}$ via $\epsilon$, so $\irr{N}{p}\simeq \cL(p_N)\otimes \Bbbk$
if $p=p_N\ell^N$, $0\le p_N<\ell$.
\end{remark}
\section*{Acknowledgement} The main part of this paper was written during my visit to the Max-Planck-Institute in Bonn as an Alexander von Humboldt Fellow. I would like to thank the institute for its excellent working environment and support. I would like to thank especially to Geordie Williamson for all the discussions and the guidance.
| 733da6b0862d6bec4fea71c1b27bff30575c7363 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Knowledge of quantum rovibrational states near the dissociation threshold is mandatory for the understanding
of the molecular dynamics of formation and depletion processes. In this respect the ozone molecule is a
particular interesting subject for both fundamental molecular
physics~\cite{SCH06:625, MAR13:17703,IVA13:17708,SUN10:555,LI14:081102,RAO15:633,GRE07:2044,TYU14:143002,GAR06:163005,MAU16:054107}
and various applications owing to the well-known role that this molecule plays in atmospheric physics and
climate processes~\cite{LU09:118501,BOY09:6255}.
Despite of the significant progress made over past decades in the study of ozone
spectroscopy~\cite{TYU14:143002,flaud1990atlas,MIK96:227,CAM06:1,BAR13:172,MON12:840,BAB14:169}
and dynamics~\cite{SCH06:625, MAR13:17703,IVA13:17708,SUN10:555,LI14:081102,RAO15:633,MAU16:054107,XIA89:6086,GAO01:259,
GAO02:137,CHA04:2700,XIE05:131,WIE97:745,SCH03:1966,LIN06:5305,VAN07:2866,VET07:138301,GRE09:181103,DAW11:081102}
many aspects of this molecule as well as of the $\rm O_2 + O$ complex in high energy states are not yet fully understood.
One of the major motivations for recent investigations of excited ozone has been the discovery of the mass-independent
fractionation reported by Mauersberger {\it et al.}~\cite{MAU81:935,KRA96:1324,JAN03:34},
Thiemens {\it et al.}~\cite{THI83:1073}, Hippler {\it et al.}~\cite{HIP90:6560},
in laboratory and atmospheric experiments: for most molecules, the isotope enrichment scales
according to relative mass differences, but the case of ozone shows an extremely marked deviation from this rule.
This has been considered as a ``milestone in the study of isotope effects''~\cite{MAR13:17703} and
a ``fascinating and surprising aspect $\dots$ of selective enrichment of heavy ozone isotopomers''~\cite{DAW11:081102}.
On the theoretical side, many efforts have been devoted to the interpretation of these findings,
in the research groups of Gao and Marcus~\cite{GAO01:259,GAO02:137}, Troe {\it et al.}~\cite{LUT05:2764,HIP90:6560},
Grebenshchikov and Schinke~\cite{GRE09:241,GRE09:181103},
Babikov {\it et al.}~\cite{VET07:138301,BAB03:2577}, Dawes {\it et al.}~\cite{DAW11:081102}
and in many other studies, see~\cite{SCH06:625, MAR13:17703,IVA13:17708,SUN10:555,LI14:081102,RAO15:633,CHA04:2700,XIE05:131,WIE97:745,SCH03:1966,LIN06:5305,VAN07:2866,JAN01:951,GUI15:12512,NDE15:12043,NDE15:7712}
and references therein.
Several fundamental issues raised by the ozone studies could have an impact on the understanding
of important phenomena in quantum molecular physics and of the complex energy transfer dynamics near the dissociation threshold.
It has been recognized that a non-trivial account of the symmetry properties~\cite{RAO15:633,JAN01:951},
efficient variational methods for the nuclear motion calculations and an accurate determination of the
full-dimensional ozone potential energy surfaces (PES) are prerequisites for an adequate description
of related quantum states and processes in the high energy range. The ozone molecule exhibits a complex
electronic structure and represents a challenge for accurate ab initio
calculations~\cite{GRE07:2044,GAR06:163005,XAN91:8054,SIE01:1795,SIE02:9749,KAL08:054312,HOL10:9927,SHI11:184104}.
Earlier 1D PES studies predicted an ``activation barrier'' at the transition state (TS) along the
minimum energy path (MEP)~\cite{ATC97:176,HAY82:862,BAN93:155}.
Later on more advanced electronic structure calculations have
suggested that the MEP shape could have a ``reef''-like structure~\cite{HER02:478,SCH04:5789,FLE03:610}
with a submerged barrier below the dissociation limit. Following preliminary estimations of
Fleurat-Lessard {\it et al.}~\cite{FLE03:610}, this ``reef''
feature was incorporated into a so-called ``hybrid PES'' by Babikov {\it et al.}~\cite{BAB03:2577} by introducing
a 1D semi-empirical correction to the three-dimensional Siebert-Schinke-Bittererova (SSB)~\cite{SIE01:1795,SIE02:9749} PES
with empirical adjustments to match the experimental dissociation energy.
This modified mSSB surface containing a shallow van der Waals (vdW) minimum along the dissociation
reaction coordinate around $r_1 \sim 4.5 - 5.0 \, a_0$ has been used to study the metastable
states~\cite{BAB03:2577} and also suggested the existence of van der Waals
bound states~\cite{GRE03:6512,JOY05:267,BAB03:6554,LEE04:5859}.
A detailed review of ozone investigations up to this stage has been presented in the
``Status report of the dynamical studies of the ozone isotope effect'' by Schinke {\it et al.}~\cite{SCH06:625}
who concluded that the calculated rate constants were about 3-5 times smaller than the measured ones and had
a wrong temperature dependence.
Recently Dawes {\it et al.}~\cite{DAW11:081102,DAW13:201103} have argued that an accurate account of several interacting
electronic states in the TS region should result in a ground state potential function without the
``reef'' feature found in previous ab initio calculations. Since this work,
and based on scattering studies~\cite{LI14:081102,XIE15:064308,SUN15:174312},
the ``reef structure'' was considered a ``deficiency''~\cite{NDE16:074302} of the SSB PES~\cite{SIE01:1795,SIE02:9749}
and its modified mSSB versions~\cite{BAB03:2577,AYO13:164311},
and was thought a plausible reason for the disagreement in rate constant calculations~\cite{SCH06:625,DAW11:081102}.
Ndengue {\it et al.}~\cite{NDE16:074302} have reported energies of $J = 0$ and $J = 1$ bound rovibrational ozone states
below $D_0$ using the Dawes {\it et al.} \cite{DAW13:201103} PES. Variational calculations of the 100 lowest bound vibrational states
using that PES resulted in a root-mean-square (RMS) obs-calc error~\cite{NDE16:074302}
of $\sim 20 \, \ensuremath{\rm cm^{-1}}\xspace$ with respect to the experimentally observed band centers of $\rm (^{16}O)_3$.
In 2013 Tyuterev {\it et al.}~\cite{TYU13:134307} have proposed a new analytical representation for the ozone PES accounting
for its complicated shape on the way towards the dissociation limit. They constructed two
PES versions based on extended ab initio calculations. Both PESs were computed at a high level of
electronic structure theory with the largest basis sets ever used for ozone, $MRCI(+Q) / AVXZ$ with $X = 5, 6$ and
extrapolation to the complete basis set limit. The first PES, referred to as R\_PES (``reef\_PES'' )
has been obtained including a single electronic state in the orbital optimization. It possesses the
``reef'' TS feature, as most published potentials do. The second one accounts for Dawes {\it et al.}'s correction~\cite{DAW11:081102}
which considers interaction with excited states. This latter potential is referred to as NR\_PES
(``no\_reef\_PES''). Both PESs have very similar equilibrium configurations in the bottom of the main
$C_{2v}$ potential well and give the same dissociation threshold: the theoretical value for both of
them, $D_0 = 1.0548 \, {\rm eV} \approx 8508 \, \ensuremath{\rm cm^{-1}}\xspace$, lies between two experimental dissociation energies
with a deviation of only 0.6\% from the most recent experimental value of Ruscic~\cite{RUS06:6592,ruscic2014updated}
(as cited in~\cite{HOL10:9927}).
Vibrational calculations, using the NR\_PES by Tyuterev {\it et al.}~\cite{TYU13:134307}, of all $\rm (^{16}O)_3$ band centers observed in rotationally resolved spectroscopy experiments
have resulted in an (RMS) obs-calc error of only $\sim 1 \, \ensuremath{\rm cm^{-1}}\xspace$
without any empirical adjustment.
Metastable ozone states above the dissociation threshold are expected to play a key role in the two-step
Linderman mechanism~\cite{BAB03:2577} of ozone formation at low pressures. They have been studied by Babikov {\it et al.}~\cite{BAB03:2577}
and by Grebenschikov and Schinke~\cite{GRE09:241,GRE09:181103} involving also lifetime calculations. Both investigations
are based on SSB or mSSB potential surfaces~\cite{BAB03:2577} exhibiting the ``reef''-structure features.
Assignment of recent very sensitive cavity-ring-down laser experiments in the TS energy range
(from 70\% to 93\% of $D_0$) have been possible~\cite{TYU14:143002,MON13:49,STA13:104,DEB13:24,BAR14:51,STA14:211,CAM15:84,STA15:203}
due to ro-vibrational predictions using the NR\_PES
that changed the shape of the bottleneck range along the MEP and transformed the reef into a kind of smooth shoulder.
The predictions of bound states with this latter PES in the TS energy range (from 70\% to 93\% of $D_0$)
exhibit average errors of only $1 - 2 \, \ensuremath{\rm cm^{-1}}\xspace$ for six ozone isotopologues, 666, 668, 686, 868, 886 and 888~\footnote{Here we
use a common abbreviation for the ozone isotopologues: $666 \equiv \rm (^{16}O)_3$, $668 \equiv \rm {^{16}O^{16}O^{18}O}$,
$686 \equiv \rm {^{16}O^{18}O^{16}O}$, etc.}.
This clearly demonstrated~\cite{TYU14:143002} that the NR\_PES by Tyuterev {\it et al.}~\cite{TYU13:134307}
is much more accurate than other available surfaces for the description of all experimental spectroscopic data,
at least up to 8000~$\rm cm^{-1}$,
that is, for bound states up to at least 93\% of the dissociation threshold.
In the original publication of Ref.~\cite{TYU13:134307}, bound states have been computed in the $C_{2v}$ symmetry of the main
potential well. To our knowledge no systematic studies of metastable ozone states with this NR\_PES~\cite{TYU13:134307}
have been published so far.
In the present work we report the first calculations of resonance state energies, corresponding wave functions
and lifetimes using this PES. Furthermore, bound states near the dissociation threshold are investigated
in full $D_{3h}$ symmetry, accounting for possible permutation of identical nuclei over the three potential wells.
\section{Symmetry considerations: Stationary approach}
\label{sec:theory}
In the electronic ground state, the ozone molecule has $C_{2v}$ symmetry at equilibrium such that the global
potential energy surface has three relatively deep minima, corresponding to three possible arrangements of the oxygen atoms known as ``open configurations''.
As the barriers between two wells are very high, low-lying rovibrational states of the homonuclear ozone isotopologues,
such as $\rm (^{16}O)_3$, which we study in the present article,
may be characterized by irreducible representations (irreps) of the
molecular symmetry group $C_{2v}(M)$, which is isomorphic with the $C_{2v}$ point group.
In the terminology of Longuet-Higgins~\cite{longuet1963symmetry,bunker98}, transformations between the three possible
arrangements of three oxygen atoms in ozone are not feasible at low energies.
For weakly bound rovibrational states, however, for which tunneling of the barrier becomes noticeable,
and for continuum states of ozone above the barrier, the transformation between arrangements
becomes feasible: The description of the dynamics of such states cannot be restricted to one potential well.
In this situation, the complete molecular symmetry group must be employed to classify nuclear motion.
This group is the three-particle permutation inversion group,
$S_3 \times I$. It is isomorphic with the point group $D_{3h}$ and hence may also be designated
$D_{3h}(M)$, where $M$ stands for molecular symmetry group~\cite{bunker98}.
Dissociation of the ozone molecule on the electronic ground state surface leads to an oxygen atom
and a dioxygen molecule, both in their electronic ground states, i.e. O($^3P$) + O$_2(X^3\Sigma_g^-)$.
The symmetry group of the oxygen atom is just the inversion group $I$, while that of the
oxygen molecule is the two-particle permutation inversion group $S_2 \times I$. The latter may
be designated $D_{\infty h}(M)$ in order to retain the $D_{\infty h}$ nomenclature for the irreducible
representations~\footnote{It is important here to note that
no degenerate representations are included in the molecular symmetry group $D_{\infty h}(M)$,
in contrast to the point group $D_{\infty h}$, which contains representations of type
$\Pi$, $\Delta$ etc. The order of the group $D_{\infty h}(M)$ is four, while that of the
point group $D_{\infty h}$ is infinite.}.
In the asymptotic channel, exchange of identical nuclei between the atom
and the diatomic molecule becomes unfeasible as their distance goes to infinity. It is clear
from this discussion that the molecular symmetry groups $C_{2v}(M)$ and $D_{\infty h}(M)$
are equivalent and just provide different sets of labels for the four irreducible representations.
They are two manifestations of the $S_2 \times I$ group.
To make this paper self-contained we give the characters and symmetry labels in Table~\ref{tab_C2v}. Of the symmetry elements of the point group $D_{\infty h}$ only those are
retained for the molecular symmetry group $D_{\infty h}(M)$ that correspond to a permutation inversion operation. This excludes
symmetry elements such as $2C(\phi)$ which leave all nuclei on their place.
The molecule is placed in the $xz$ plane, which is the convention normally
used in ozone spectroscopy.~\footnote{This choice is different from that of Bunker \& Jensen~\cite{bunker98}.
As a result, the symmetry labels $B_1$ and $B_2$ are interchanged.}
The correspondence of the axes
is thus ($x \rightarrow b, y \rightarrow c, z \rightarrow a$).
The transformation properties of the $p$ orbitals, which are needed in the discussion of the
asymptotic states, are indicated in the last column of the table.
\begin{table}[t]
\caption{Character table of the point groups $C_{2v}$, $D_{\infty h}$ (excerpts) and the
permutation inversion group $S_2 \times I$ using the nomenclatures
of $C_{2v}(M)$ and $D_{\infty h}(M)$ for the irreducible representations.
}
\label{tab_C2v}
\begin{tabular}{cc|@{~~~}c@{~~~}c@{~~~}c@{~~~}c | c}
\hline \hline
\multicolumn{2}{c|@{~~~}}{$C_{2v}$} & E & $C_{2b}$ & $\sigma_{ab}$ & $\sigma_{bc}$ & \\
\multicolumn{2}{c|@{~~~}}{$D_{\infty h}$} & E & $\infty C_2'$ & $\infty \sigma_v$ & $i$ & \\ \hline
\multicolumn{2}{c|@{~~~}}{$S_2 \times I$} & $E$ & (12) & $E^{*}$ & $(12)^{*}$ & \\
$C_{2v}(M)$ & $D_{\infty h}(M)$ & & & & & \\ \hline
$A_1$ & $\Sigma_g^+$ & 1 & 1 & 1 & 1 & $p_b$ ($p_x$)\\
$B_1$ & $\Sigma_u^+$ & 1 & -1 & 1 & -1 & $p_a$ ($p_z$)\\
$A_2$ & $\Sigma_u^-$ & 1 & 1 & -1 & -1 & \\
$B_2$ & $\Sigma_g^-$ & 1 & -1 & -1 & 1 & $p_c$ ($p_y$) \\
\hline \hline
\end{tabular}
\end{table}
Classification of states in $S_2 \times I$ is convenient for rovibrational states situated deep in the wells,
and for the dissociating resonances. We now wish to relate them with the symmetry species of
the complete permutation inversion group $S_3 \times I$, or $D_{3h}(M)$. These correlations are
shown in Table~\ref{tab_D3h}. In addition to the symmetry elements of $S_2 \times I$,
which are the identity, $E$, the pair permutation, $(12)$, the inversion of the spatial coordinate
system, $E^{*}$, and the combination $(12)^{*} = (12) \times E^{*} = E^{*} \times (12)$, a
new class appears, the cyclic permutations, $\{(123), (132)\}$, as well as the class built up by
its combination with the inversion of the coordinate system, \{$(123)^{*}$, $(132)^{*}$\}.
These new operations describe the exchange between the three localized structures.
The correlation presented in Table~\ref{tab_D3h} is obtained by matching the characters of the
common operators, i.e. the identity operation, pair permutations and the inversion of the coordinate system.
\begin{table}[t]
\caption{Character table of the group $S_3 \times I$ and the relation with
the irreducible representations of the group $S_2 \times I$ using the nomenclatures
$C_{2v}(M)$ and $D_{\infty h}(M)$.}
\label{tab_D3h}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|cccccc|c|c}
\hline \hline
$S_3 \times I$ & E & \{(123), (132)\} & \{(12), (23), (13)\} & $E^{*}$ & \{$(123)^{*}$, $(132)^{*}$\} & \{$(12)^{*}$, $(23)^{*}$, $(13)^{*}$\} & \multicolumn{2}{c}{$S_2 \times I$} \\
$D_{3h}(M)$ & & & & & & & $C_{2v}(M)$ & $D_{\infty h}(M)$ \\ \hline
$A'_1$ & 1 & ~1 & ~1 & ~1 & ~1 & ~1 & $A_1$ & $\Sigma_g^+$ \\
$A'_2$ & 1 & ~1 & -1 & ~1 & ~1 & -1 & $B_1$ & $\Sigma_u^+$ \\
$E'$ & 2 & -1 & ~0 & ~2 & -1 & ~0 & $A_1 + B_1$ & $\Sigma_g^+ + \Sigma_u^+$ \\
$A''_1$ & 1 & ~1 & ~1 & -1 & -1 & -1 & $A_2$ & $\Sigma_u^-$ \\
$A''_2$ & 1 & ~1 & -1 & -1 & -1 & ~1 & $B_2$ & $\Sigma_g^-$ \\
$E''$ & 2 & -1 & ~0 & -2 & ~1 & ~0 & $A_2 + B_2$ & $\Sigma_u^- + \Sigma_g^-$ \\
\hline \hline
\end{tabular}
}
\end{table}
The rovibrational states of ozone may now be classified in the $D_{3h}(M)$ group, allowing for
tunneling between the three wells. They can be considered superpositions of the three states localized in their wells,
which give rise to a one-dimensional representation and a two-dimensional representation, just as in the case of
triplet $\rm H_3^+$ which has been discussed before~\cite{alijah15}. The energy difference between the one and the two-dimensional
representations is called tunneling splitting. Purely vibrational states have
positive parity, i.e. belong to either $A'_1$, $A'_2$ or $E'$, while both prime and double prime states exist
for rotationally excited states.
The localized vibrational states to be superimposed may be classified in $C_{2v}(M)$
by the approximate normal mode quantum numbers
$| v_1 \, v_2 \, v_3 \rangle$ of the symmetric stretching vibration, $v_1$, the bending vibration, $v_2$, and the antisymmetric stretching vibration, $v_3$.
Since these transform as $A_1$, $A_1$ and $B_1$, respectively, the symmetry of $| v_1 \, v_2 \, v_3 \rangle$
is $A_1$ for $v_3$ even and $B_1$ for $v_3$ odd. In $D_{3h}(M)$, they give rise to the pairs
($A'_1, E'$), ($A'_1, E'$) and ($A'_2, E'$), referring to the one and two-dimensional representations.
Only those vibrational states that have $A'_1$ symmetry are allowed for the isotopologue $\rm (^{16}O)_3$ as can
be seen from the following analysis:
The $^{16}$O isotope is a boson, with zero nuclear spin, i.e. the total wave function of $\rm (^{16}O)_3$
must be symmetric under exchange of any two $^{16}$O nuclei and transform as $A_1'$ or $A_1''$ in $D_{3h}(M)$.
The nuclear spin function transforms as $A'_1$.
Likewise, the electronic wave function of the ground state,
$X \, ^1A_1$ in spectroscopic notation, since the open structure minima have $C_{2v}$ symmetry, is totally symmetric with respect
to all nuclear permutations.
It means that the rovibrational part of $(^{16}$O$)_3$ should also be symmetric under an exchange of any two oxygen nuclei,
i.e. should transform as the $A_1'$ or the $A_1''$ irreducible representation. Purely vibrational states have positive parity
and thus symmetry $A_1'$, the other symmetry species are not allowed.
We note in particular that the degenerate tunneling component has zero statistical weight, giving rise to ``missing levels''
in spectroscopic language. As a consequence, tunneling splitting of the purely vibrational states cannot be observed.
The calculations of the present article were performed in hyperspherical coordinates, as they permit
straightforward implementation of the full permutation inversion symmetry.
The rovibrational wave function $\Psi^{Jm}_{v}$ of tunneling ozone can be written as an expansion over products of
rotational ${\cal R}_{Jkm}(\Omega)$ and vibrational factors $\psi_{v}^{Jk}({\cal Q})$
\begin{equation}
\label{eq:expansion}
\Psi^{Jm}_{v}(\Omega,{\cal Q})=\sum_{k} {\cal R}_{Jkm}(\Omega)\psi_{v}^{Jk}({\cal Q})\,,
\end{equation}
where ${\cal R}_{Jkm}(\Omega)$ are symmetric top rotational wave functions proportional to the Wigner functions $D^J_{mk}$
\begin{equation}
{\cal R}_{Jkm}(\Omega)=\sqrt \frac{2J+1}{8\pi^2}\left[D^J_{mk}(\Omega)\right]^*\,,
\end{equation}
and depending on the three Euler angles $\Omega$. The vibrational part of the wave function depends on the internal projection
$k$ of the angular momentum onto the axis perpendicular to the molecular plane, denoted the $y$-axis in Table~\ref{tab_C2v}.
Note that no decomposition is made here in terms of the $C_{2v}(M)$ normal modes, which
would be an approximation.
Each product in expansion (\ref{eq:expansion}) should have the same symmetry in the $D_{3h}(M)$ group as the
total rovibrational wave function, i.e. $A_1'$ or $A_1''$. The symmetry
$\Gamma^{r}$ of the rotational functions ${\cal R}_{Jkm}(\Omega)$ in $D_{3h}(M)$ is well known
(see, for example, \cite{bunker98,kokoouline03b}). It imposes restrictions on the possible irreducible representations
of the vibrational factors $\psi_{v}^{Jk}({\cal Q})$: The rotational and vibrational wave functions should be of the same species,
both $A_1$, or both $A_2$, or both $E$. Parities of the wave functions are not restricted. The parity of the
vibrational functions is always positive, the parity of the rotational function is positive for even $k$ and
negative for odd $k$. Examples of the irreducible representations of rotational and vibrational functions
are given in Table \ref{tab:irreps} for $J \le 3$.
\begin{table}
\caption{Allowed combinations of irreducible representations of the rotational and vibrational factors in the expansion of
Eq.~(\ref{eq:expansion}). \footnote{The symmetrized combinations of functions with $k=\pm 3$ transform as the $A_1''$ and $A_2''$
representations in $D_{3h}(M)$. The direct products, $\Gamma^r \times \Gamma^v$, of two $E$ representations
yield $A_1 + A_2 + E$ and contain the $A_1$ representation. Only the latter is listed in the last line of the table.
The parity is given by the usual rule $' \times '= '$, $'' \times ''= '$, $' \times ''= ''$, $'' \times '= ''$.}
}
\label{tab:irreps}
\begin{tabular}{c p{1cm} p{1cm} p{1cm} p{1cm} p{1cm} p{1cm} p{1cm} p{1cm} p{1cm} p{1.3cm} }
\hline \hline
$J$ &0 &1 &1 &2 &2&2&3&3&3&3 \\
\hline
$k$ &0 &0 &$\pm$1 &0 &$\pm$1&$\pm$2&0&$\pm$1&$\pm$2&$\pm$3$^*$\\
$\Gamma^r$ &$A_1'$ &$A_2'$ & $E''$&$A_1'$&$E''$&$E'$&$A_2'$&$E''$&$E'$&$A_1''$, $A_2''$\\
$\Gamma^v$ &$A_1'$ &$A_2'$ & $E'$&$A_1'$&$E''$&$E'$&$A_2'$&$E'$&$E'$&$A_1'$, $A_2'$\\ \hline
$\Gamma^r \times \Gamma^v$ & $A_1'$ & $A_1'$ & $A_1''$ & $A_1'$ & $A_1'$ & $A_1'$ & $A_1'$ & $A_1''$ & $A_1'$ & $A _1''$, $A _1''$ \\
\hline
\hline
\end{tabular}
\end{table}
Let us now turn to the symmetry classification of the wave functions of the decaying resonance states.
The lowest dissociation limit of ozone produces the oxygen atom, O$\, (^3P)$,
and the oxygen molecule, O$_2\, (X^3\Sigma_g^-)$, in their electronic ground states.
The orbital degeneracy of the atomic $P$ state is three. One orbital is oriented perpendicular to the
plane spanned by the three nuclei, denoted as $p_c$ in Table~\ref{tab_C2v}. According to Table~\ref{tab_D3h},
it transforms as $A_2''$ in $D_{3h}(M)$. The two in-plane orbitals transform as $E'$. On the other hand,
the electronic symmetry of the di-oxygen molecule is $\Sigma_g^-$ in $D_{\infty h}(M)$, or $A_2''$ in $D_{3h}(M)$.
At large distances, the electronic ground state, $X \, ^1A_1$, of ozone correlates with the
perpendicular ($p_c$) component of the atomic $P$ state plus the diatomic $\Sigma_g^-$ state, which have both
$A_2''$ symmetry in $D_{3h}(M)$ such that their product is indeed $A_1'$.
The electronic ground state of $\rm O_2$ is antisymmetric with respect to an exchange of the two nuclei.
Since the vibrational states of $\rm O_2$ are totally symmetric,
this implies that the rotational functions must be antisymmetric to yield a symmetric nuclear
wave function. The rotational functions of $\rm O_2$ transform as $\Sigma_g^+$ for even values of $j$ and
as $\Sigma_g^-$ for odd values. Rotational states of $\rm ^{16}O_2$ must therefore have
odd rotational angular momentum, $j$,
and the lowest rovibrational state is ($v=0, j=1$).
Let us now analyze the asymptotic wave function in the exit channel $\kappa$ with $\kappa = 1, 2, 3$.
It can be expanded as
\begin{equation}
\label{eq:scattering}
\Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa}) \approx \frac{1}{r_{\kappa} R_{\kappa}} \varphi^{el}_a\varphi^{el}_d
\chi_{v_dj}(r_{\kappa}) {\cal Y}^{Jm}_{jl}(\hat{r}_{\kappa}, \hat{R}_{\kappa}) e^{i (k R_{\kappa} - l \pi/2)}\,,
\end{equation}
where $\exp(i (k R_{\kappa} - l \pi/2))$ is the scattering function of the outgoing wave
and $\chi_{v_dj}(r_{\kappa})$ the vibrational wave function of the $\rm O_2$ molecule;
$r_{\kappa}$ and ${R}_{\kappa}$ are the true, not mass-scaled, distances in the Jacobi coordinate system $\kappa$. Functions $\varphi^{el}_a$ and $\varphi^{el}_d$ represent electronic states of the O$\,(P)$ atom and the O$_2\,(X^3\Sigma_g^+)$ molecule.
Angular momenta of the atom-diatom relative motion, $l$, and of the rotation of the oxygen molecule, $j$, must be coupled
to yield the total angular momentum, $J$, which is taken care of by the bipolar harmonics, ${\cal Y}^{Jm}_{jl}$.
They are defined as
\begin{equation}
{\cal Y}^{Jm}_{jl}(\hat{r}_{\kappa}, \hat{R}_{\kappa}) = \sum_{m_l, m_j} C_{j m_j l m_l}^{Jm} Y_{j m_j}(\hat{r}_{\kappa}) Y_{l m_l}(\hat{R}_{\kappa})\,,
\end{equation}
where the $Y$ are spherical harmonics and $C$ are Clebsch-Gordan coefficients. The scattering function in Eq.~(\ref{eq:scattering}) is not symmetric with respect to permutation of three bosonic nuclei and, therefore, cannot be correlated in this form with the short-distance form of Eq.~(\ref{eq:expansion}), which does have correct symmetry behavior (for the combinations of quantum numbers given in Table \ref{tab:irreps}). To bring the function of Eq.~(\ref{eq:scattering}) to the form satisfying the permutationl symmetry of three bosons, in the language of group theory, one has to apply projectors of the $D_{3h}(M)$ group of the two allowed irreducible representations, $A_1'$ or $A_1''$. An efficient way to perform it is to use a general approach of Ref.~\cite{douguet08a} applicable to a three-body system with arbitrary total nuclear spin. Equations (19) of that reference do not take into account the electronic part of the total wave function. The electronic wave function of the dioxygen $\varphi^{el}_d$ changes sign under permutation of the two atoms and under the inversion operation, and the atomic $\varphi^{el}_a$ changes sign under the inversion only. Therefore, Eqs.~(19) of Ref.~\cite{douguet08a} take the following form for the present case
\begin{eqnarray}
\label{eq:12}
(12) \Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa}) =(-1)^{j+1}\Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa})\nonumber\\
E^* \Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa}) =(-1)^{l+j}\Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa})
\end{eqnarray}
With these properties, the projectors $P_{\Gamma}$ take the form (see Eqs.~(20) of Ref.~\cite{douguet08a})
\begin{eqnarray}
\label{eq:projector_21adapted2}
P_{\Gamma}\Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa})=\nonumber\\
\left(1+ \chi^{\Gamma}_{23}(23)+\chi^{\Gamma}_{31}(31)\right) \left(1 + (-1)^{j+1}\chi^{\Gamma}_{12}\right)\left(1+ (-1)^{l+j}\chi^{\Gamma}_{E^*}\right)\Psi^{Jm}_{\kappa v_d j l}(\vec{r}_{\kappa}, \vec{R}_{\kappa})\,,
\end{eqnarray}
for any of the $D_{3h}(M)$ representations.
Here, $\chi^{\Gamma}$ are characters of the representation $\Gamma$ given in Table~\ref{tab_D3h}. From the expression in the second parentheses on the right side of the equation above, it is clear that for the allowed representations $A_1'$ and $A_1''$, if $j$ is even, the projectors are identically zero, ${P}_{A_1'} = 0$, ${P}_{A_1''} = 0$. It is simply means that a free molecule $^{16}$O$_2 \, (X^3\Sigma_g^+)$ can only have odd rotational angular momentum $j$. The expression in the third parentheses means that if the quantum numbers $l$ and $j$ have different parity, the projectors again give identically zero for $A_1'$ (but not for $A_1''$). In particular, it implies that dissociative states of $^{16}$O$_3$ with rotational angular momentum $J=0$ do not exist within the adiabatic approximation.
\section{Nuclear dynamics}
The present, stationary theoretical approach to describe nuclear dynamics was developed previously by
Kokoouline {\it et al.}~\cite{kokoouline06,blandon07,blandon09,alijah15}. It is based on the two-step procedure of solving the stationary Schr\"odinger equation in hyperspherical coordinates~\cite{johnson80,johnson83a,johnson83b}. Although the method was previously applied to several three-body problems, it has never been applied to a system with large masses of the three particles and so many bound states: In Ref.~\cite{blandon07} the method was developed and tested on a benchmark system of a three-boson nucleus with a very shallow potential supporting only one bound state and one resonance. In~\cite{blandon09}, the method was employed to calculate resonances in three-body collisions of hydrogen atoms. The lowest H$_3$ potential energy surface has two coupled sheets without any bound state but with many resonances. The method was also routinely used to represent the vibrational continuum in studies of dissociate recombination of isotopologues of H$_3^+$~\cite{kokoouline04a,santos07}. An important difference of the present study with the previous ones is that the number of bound states is large, which requires a significantly larger basis to represent the vibrational dynamics near and above the dissociation.
We briefly summarize the main elements of the approach. To solve the Schr\"odinger equation
\begin{equation}
\label{eq:Schr1}
\left[T(\rho,\theta,\phi)+V(\rho,\theta,\phi)\right]\Phi_v(\rho,\theta,\phi)=E_v\Phi_v(\rho,\theta,\phi)
\end{equation}
for three particles interacting through the potential $V(\rho,\theta,\phi)$ in the hyperspherical coordinates $\rho,\theta$, and $\phi$, first, the adiabatic hyperspherical curves $U_a(\rho)$ and the corresponding hyperangular eigenstates $\varphi_a(\rho_i;\theta,\phi)$ (hyperspherical adiabatic states -- HSA) are obtained by solving the equation in the two-dimensional space of the hyperangles $\theta$ and $\phi$ for several fixed values of the hyper-radius $\rho_j\ (j=1,2,\cdots)$, i.e. the following equation is solved
\begin{equation}
\label{eq:Had}
\left[\hbar^2\frac{\Lambda^2+\frac{15}{4}}{2\mu \rho_j^2}+V(\rho_i;\theta,\phi)\right] \varphi_a(\rho_j;\theta,\phi)=U_a(\rho_i)\varphi_a(\rho_j;\theta,\phi).
\end{equation}
In the above equation, $\Lambda^2$ is the grand angular momentum squared~\cite{WHI68:1103,johnson83b} and $\mu$ is the three-particle reduced mass: For identical oxygen atoms with mass $m_O$, one has $\mu=m_O/\sqrt{3}$. The equation is solved using the approach described in~\cite{esry97}. Solution of Eq.~(\ref{eq:Had}) yields adiabatic curves $U_a(\rho)$ and eigenfunctions $\varphi_a(\rho;\theta,\phi)$, defining a set of HSA channels $a$. The HSA states are then used to expand the wave function $\Phi_v$ in Eq.~(\ref{eq:Schr1})
\begin{equation}
\label{eq:vibr_func_SVD}
\Phi_{v}({\cal Q})=\sum_a \psi_a(\rho_j) \varphi_a(\rho_j;\theta,\phi)\,.
\end{equation}
The expansion coefficients $\psi_a(\rho_i)$ depend on hyper-radius $\rho$. Following the original idea of Ref. \cite{tolstikhin96} the hyper-radial wave functions $\psi_a(\rho_i)$ are then expanded in the discrete variable representation (DVR) basis $\pi_j(\rho)$
\begin{equation}
\psi_a(\rho)=\sum_j c_{j,a}\pi_j(\rho).
\end{equation}
Inserting the two above expansions into the initial Schr\"odinger equation (\ref{eq:Schr1}), one obtains
\begin{equation}
\label{eq:gener_eigen2}
\sum_{j',a'}\left[\langle\pi_{j'} |-\frac{\hbar^2}{2\mu}\frac{d^2}{d\rho^2}|\pi_j\rangle{\cal O}_{j'a',ja}+U_a(\rho_j)\delta_{j',j}\delta_{a'a}\right]c_{j'a'}=E\sum_{a'}{\cal O}_{ja',ja}c_{ja'}\,
\end{equation}
with
\begin{equation}
{\cal O}_{j'a',ja}=\langle\varphi_{a'}(\rho_{j'};\theta,\phi)|\varphi_a(\rho_j;\theta,\phi)\rangle .
\end{equation}
In the above equation, the matrix elements of the second-oder derivative with respect to $\rho$ is calculated analytically (see, for example, \cite{tuvi1997hermiticity,kokoouline99} and references therein).
The described approach of solving the Schr\"odinger equation using the adiabatic (HSA) basis replaces the usual form of non-adiabatic couplings in terms of derivatives with respect to $\rho$ with overlaps between adiabatic states $\varphi_a(\rho,\theta,\phi)$ evaluated at different values of $\rho$.
The approach is particularly advantageous here, since the adiabaticity of the hyper-radial motion,
when separated from hyperangular motion, is not satisfactory,
so that multiple avoided crossings between HSA energies $U_a(\rho)$ occur.
This is the usual situation in three-body dynamics.
Representing non-adiabatic couplings by derivatives $\langle\varphi_{a'}\lvert \partial/\partial\rho\rvert\varphi_{a}\rangle$ and $\langle\varphi_{a'}\lvert \partial^2/\partial\rho^2\rvert\varphi_{a}\rangle$ near the avoided crossings would require a very small grid step in $\rho$. The use of overlaps between HSA states reduces significantly the number of grid points along $\rho$ required for accurate representation of vibrational dynamics.
\begin{figure}[ht]
\vspace{2cm}
\includegraphics[width=15cm]{fig1.pdf}
\caption{Ozone potential energy surface, NR\_PES of Ref.~\cite{TYU13:134307}, as a function of the two hyperangles for several values of the hyper-radius. In the plots, the hyperangles are represented in a polar coordinate system (see Fig.~6 of Ref.~\cite{kokoouline03b}): $\theta$ increases from the center of each plot to its edge; $\phi$ is a cyclic variable (polar angle) changing from 0 to $2\pi$. The minimum of PES, situated near $\rho=4.2~a_0$, is chosen as origin. The electronic energy of dissociation to the atom and the diatomic molecule at equilibrium is at 9150 cm$^{-1}$.}
\label{fig:pes}
\end{figure}
In Ref.~\cite{TYU13:134307}, the main features of the PES were demonstrated in internal coordinates.
In the present study, the NR\_PES of Ref.~\cite{TYU13:134307}, which had been originally defined
in the $C_{2v}$ wells, was symmetrized according to the nuclear permutations and converted in the
hyperspherical coordinates~\cite{johnson80,johnson83a,johnson83b}.
Fig.~\ref{fig:pes} shows the PES as a function of the two hyper-angles for several values of the hyper-radius. As evident from the plot at $\rho=5.4$ bohr the potential barrier between the wells is situated at energies 9000 cm$^{-1}$, i.e. very close to the dissociation threshold. The passage between the wells occurs at geometries beyond the ``shoulder'' of the ozone potential. Therefore, one expects weakly bound low-energy resonances delocalized between the three potential wells. To represent nuclear dynamics of such near-dissociation levels, one needs to take into account the three potential wells simultaneously.
The energy $D_0$ of dissociation to the ${\rm ^{16}O} \, (^3P)$ and
${\rm ^{16}O_2} \, (X \,{^3\Sigma}_g^- [v_d = 0, j = 0])$ products is 8555 cm$^{-1}$ above the ground rovibrational level of $\rm ^{16}O_3$.
\begin{figure}[ht]
\includegraphics[width=12cm]{fig2.pdf}
\caption{Hyperspherical adiabatic curves $U_a(\rho)$ of the $A_1$ irreducible representation and $J=0$ as a function of hyper-radius obtained for $^{16}$O$_3$. In this figure, the energy origin is chosen at the ground vibrational level of $^{16}$O$_3$, situated 1443.524 cm$^{-1}$ above the PES minimum. The $v_d=0,1,2$ labels indicate energies of the ${\rm O \, + \, O_2} \, (v_d)$ asymptotic vibrational channels. Multiple HSA curves between the vibrational channels correspond to various rotational channels $j$ of the dissociating oxygen molecule. For the $A_1$ vibrational states, only even $j$ are allowed. Because $J=0$, the partial wave in each asymptotic channel $(v_d,j,l)$ is determined simply as $l=j$.}
\label{fig:adiabatic}
\end{figure}
A convenient way of analyzing nuclear dynamics of three atoms is given by HSA curves, which could be viewed in a way similar to Born-Oppenheimer curves for diatomic molecules, except that the adiabatic and dissociation coordinate in the HSA curves is the hyper-radius, not the inter-atomic distance. In contrast to the case of Born-Oppenheimer separation between electronic and vibrational motion for diatomic molecules, non-adiabatic coupling between HSA states is almost always strong and cannot be neglected. Nevertheless, many key features of the dynamics can easily be identified and qualitatively studied. The HSA curves obtained for $A_1$ vibrational symmetry and $J=0$ are shown in Fig.~\ref{fig:adiabatic}. At small values of hyper-radius, near $\rho=4.2$, the lowest HSA curves have a minimum, which corresponds to the O$_3$ equilibrium. Each of the lowest HSA curves near the minimum represents approximately a particular combination of $v_2$ and $v_3$ vibrational modes of O$_3$. The $v_1$ mode near the O$_3$ equilibrium is represented by the continuous variable $\rho$, which is at this first step not quantized in the space of HSA coordinates. Therefore, the lowest HSA curve $U_a(\rho)$ ($a=1$) near $\rho=4.2$ is an adiabatic representation of the set $(\rho,0,0)$ of vibrational modes of O$_3$ corresponding to the normal mode quantum numbers $v_2 = v_3 = 0$, the second and third HSA curves are $(v_2 = 1, v_3 = 0)$ and $(v_2 = 2, v_3 = 0)$, the fourth one is $(v_2 = 0, v_3 = 2)$, etc. Odd $v_3$ are not present in $A_1$ vibrational symmetry.
At energies near and higher than 6000 \ensuremath{\rm cm^{-1}}\xspace above the $(0,0,0)$ level, the normal modes are significantly mixed and the mode assignment becomes more difficult. However, the HSA curves at large energies, above the energy of dissociation, and at large $\rho$, provide a convenient description of dissociation dynamics. At large $\rho$, each adiabatic curve converges to a particular asymptotic channel represented by a rovibrational level $(v_d,j)$ of O$_2$ and the partial wave of relative motion of O$_2$ and O. As one can see, there are multiple very sharp avoided crossings, especially in the zone of transition from short to large $\rho$.
\section{Bound states near $\bm{D_0}$ and predissociated resonances}
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=13cm]{fig3.pdf}
\caption{The largest of the three rotational constants for $A_1$ vibrational states in the $D_{3h}$ group. Almost linear dependence of the rotational constant $A_v$ on the energy of the vibrational states permits an assignment of normal modes for low energy levels. Normal mode quantum numbers are specified for a few levels.
Note than at high energy the normal mode assignment becomes ``nominative'' and is to be taken with
caution because of strong anharmonic basis state mixing.}
\label{fig:Bv_A1}
\end{figure}
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=11cm]{fig4.pdf}
\caption{Comparison of the energies of band centers obtained in this study with the previous calculation \cite{TYU13:134307} and experimental data \cite{TYU14:143002,CAM06:1,BAR13:172,MON12:840,BAB14:169,MON13:49,STA13:104,DEB13:24,BAR14:51,STA14:211,CAM15:84} for two vibrational symmetries ($A_1$ and $A_2$ in the $D_{3h}$ group employed here, $A_1$ and $B_1$ in the $C_{2v}$ group employed in Ref.~\cite{TYU13:134307}). The difference between the present results and the previous calculation and experimental data is labeled as $D_{3h}-C_{2v}$ and exp$-D_{3h}$ respectively.}
\label{fig:theory_exp}
\end{figure}
A series of calculations with different parameters of the numerical approach were performed to assess the uncertainty of the obtained energies with respect to the numerical procedure. The final results for $A_1$ and $A_2$ vibrational levels were obtained with 60 HSA states. The number of $B$-splines used for each of the hyperspherical angles $\theta$ and $\phi$ was 120. Similar to previous work by Alijah and Kokoouline
on the $\rm H_3^+$ molecule~\cite{alijah15}, the interval of variation of $\varphi$ was from $\pi/6$ to $\pi/2$ in calculations of $A_1$ and $A_2$ levels. The variation interval of $\rho$ was from $2.9$ to $16$, a variable step width \cite{kokoouline99,kokoouline06,blandon09} along the $\rho$ grid was used with 192 grid points. The estimated uncertainty due to the employed numerical method is better than $0.001$~\ensuremath{\rm cm^{-1}}\xspace for low vibrational levels and about 0.01~\ensuremath{\rm cm^{-1}}\xspace for levels at around 7500~\ensuremath{\rm cm^{-1}}\xspace above the ground vibrational level.
This convergence error is significantly lower than the uncertainties of the ozone PES.
Figure~\ref{fig:theory_exp} compares the energies of $^{16}$O$_3$ band centers up to 8000~\ensuremath{\rm cm^{-1}}\xspace obtained in this study with the previous calculation \cite{TYU13:134307} and experimental data \cite{TYU14:143002,CAM06:1,BAR13:172,MON12:840,BAB14:169,MON13:49,STA13:104,DEB13:24,BAR14:51,STA14:211,CAM15:84}. The RMS deviation between the calculation of Ref.~\cite{TYU13:134307} in $C_{2v}$ symmetry and the present $D_{3h}$ calculations is of 0.03~\ensuremath{\rm cm^{-1}}\xspace only up to this energy cut-off. This confirms a good nuclear basis set convergence of both methods. The RMS (obs.-calc.) deviation for all vibrational band centers directly observed in high-resolution spectroscopy experiments is 1~\ensuremath{\rm cm^{-1}}\xspace. This is by one order of magnitude better than the accuracy of vibrational calculations using other ozone PESs available in the literature. The uncertainty in the determination of resonance energies depends on their widths and is roughly 10\% of the respective width. The uncertainty in calculated widths is better than 20\% for most of the resonances.
The assignment of vibrational bands is simplified by using the vibrational dependence of rotational
constants predicted from the PES and derived from ro-vibrational spectra analyses as described
in Refs.~\cite{BAR13:172,MON13:49,STA13:104,DEB13:24,BAR14:51,STA14:211,STA15:203}.
The largest rotational constant, $A_v$, corresponding to the ``linearization'' $z$-axis,
is given by the following expression in hyper-spherical coordinates \cite{kokoouline04b}
\begin{eqnarray}
A_v=\langle\Psi^{00}_{v} \lvert \frac{1}{\mu\rho^2(1-\sin\theta)}\rvert\Psi^{00}_{v}\rangle\,.
\end{eqnarray}
At low vibrational excitations, the rotational constant $A_v$ has nearly linear behavior with respect to the normal mode quantum numbers, $v_1, v_2$, or $v_3$, with proportionality coefficients different for each mode. This can be seen in Fig.~\ref{fig:Bv_A1}. For example, when $v_1=v_3=0$, the levels $(0,v_2,0)$ form almost a straight line in the $A_v(E_v)$ plot. The same is true for other series, $(v_1,0,0)$, $(0,0,v_3)$, $(1,v_2,0)$, etc. Near the dissociation limits, the normal mode approximation is not valid any more and the series become mixed, although, the $(v_1,0,0)$ and $(0,v_2,0)$ series survive even above the dissociation.
Such states cannot dissociate into $\rm O_2 \, + \, O$ unless mixed with the antisymmetric vibrational mode.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig5.pdf}
\caption{Wave functions of the $(7,0,0)$ (upper panels) and the $(0,11,0)$ (lower panels) levels as functions of the Jacobi coordinates $R, r,$ and $\gamma$. On the left panels, the dependence on $R$ and $\gamma$ is shown for a fixed value of $r=2.28\ a_0$, which is the equilibrium
nuclear distance in the $\rm O_2$ molecule. The right panels show the $R, r$-dependence for fixed $\gamma=40^\circ$. }
\label{fig:Jacobi_A1_v1_v2}
\end{figure}
Figures~\ref{fig:Jacobi_A1_v1_v2}, \ref{fig:Jacobi_A1_v3} and the upper panel of Fig.~\ref{fig:Jacobi_A1_v2_diss} show wave functions of five bound vibrational levels of $A_1$ vibrational symmetry in terms of Jacobi coordinates $R, r$, and $\gamma$, where $r$ is the distance between two oxygen nuclei of a chosen pair, $R$ is the distance from the center of mass of this pair to the third nucleus, and $\gamma$ is the angle between the vectors along $r$ and $R$. The left panels of the figures demonstrate the dependence of the wave functions on $R$ and $\gamma$. The interval of variation of $\gamma$ is from $0^\circ$ to $180^\circ$, such that it covers two of the three possible equivalent arrangements (permutations) of the three nuclei, i.e. it represents two of the three potential wells of the ozone potential. As evident, the obtained wave functions are symmetric with respect to an exchange between the two wells. Since the calculations were performed in hyper-spherical coordinates, the wave functions are also symmetric with respect to the exchange involving the third well, but the Jacobi coordinates cannot easily represent such a symmetry.
To demonstrate the nature of wave functions of different normal modes, the functions chosen in Figs.~\ref{fig:Jacobi_A1_v1_v2}, \ref{fig:Jacobi_A1_v3}, and \ref{fig:Jacobi_A1_v2_diss} represent ``pure'' vibrational modes: $(7,0,0)$, $(0,11,0)$, $(0,12,0)$, $(0,0,4)$, and $(0,0,6)$. It is easy to identify the pure $v_1$ (symmetric stretching) and $v_2$ (bending) modes by counting nodes in the Jacobi coordinates, but the behavior of the antisymmetric stretching mode $v_3$ is more complicated in Jacobi coordinates.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig6.pdf}
\caption{As Fig.~\ref{fig:Jacobi_A1_v1_v2}, but for the pure antisymmetric stretching modes $(0,0,4)$ (upper panel) and $(0,0,6)$ (lower panel).}
\label{fig:Jacobi_A1_v3}
\end{figure}
For the calculation of states above the dissociation threshold $D_0$, a complex absorbing potential (CAP) and variable grid step along $\rho$ adapted to the local de~Broglie wave length were used as described in Ref.~\cite{blandon07}. The parameters of the CAP were chosen to absorb the outgoing dissociation flux for the interval of energies approximately between 100 and 4000 \ensuremath{\rm cm^{-1}}\xspace. When the method of CAP is used, the spectrum of the Hamiltonian matrix for energies above the dissociation limit contains not only the relatively long living resonance states but also non-physical ``box states''. Real and imaginary parts of box state eigenvalues depend on the CAP and grid parameters. A manual separation of resonances and box states is difficult for this case because of a large number of resonances. Several calculations with variable parameters, such as CAP, the number of grid pints along $\rho$, the number of the HSA states, the number of $B$-splines in the HSA calculations, were performed. Spectra obtained with different sets of parameters were compared, allowing us to separate the box-states from the resonances, as the latter no not
depend on the numerical parameters in a converged calculation.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig7.pdf}
\caption{As Fig.~\ref{fig:Jacobi_A1_v1_v2}, but for the $(0,12,0)$ bound (upper panel) and $(0,13,0)$ resonance (lower panel) vibrational states. The interval or variation of $R$ in the left panel plots is larger than in Figs.~\ref{fig:Jacobi_A1_v1_v2} and \ref{fig:Jacobi_A1_v3} in order to demonstrate the long-range tail of the wave functions. Note that the state shown in the lower figure exists only if rotation is excited. Thus, the wave functions are just the vibrational parts of the total rovibrational functions.
}
\label{fig:Jacobi_A1_v2_diss}
\end{figure}
The lower panel of Fig.~\ref{fig:Jacobi_A1_v2_diss} gives an example of a resonance wave function of $A_1$ vibrational symmetry. As discussed above, such levels are not allowed for $^{16}$O$_3$, but we will consider them because the same analysis can be applied to other isotopologues of O$_3$,
and also because a similar behavior can be exhibited by $A_1$ vibrational factors of rotationally excited $A_2$ states which are allowed for $^{16}$O$_3$. At short distances, the resonance is mainly described by the $(0,13,0)$ normal mode contribution. Its wave function looks very similar to that of the $(0,12,0)$ level. It is still bound but has one more node along the $v_2$ coordinate.
The outgoing dissociative flux is clearly visible in the $R,\gamma$ plot. The contrast in the $R,r$ plot is not quite sufficient to see the flux clearly.
The vibrational resonance $(0,13,0)$ corresponds to large-amplitude bending motion of ozone. The energy of such bending oscillations is above dissociation, but the system does not dissociate fast, because the O$_2 \, + \, O$ dissociation implies that two of the three internuclear distances should become very large and the third distance should stay small, whereas when the molecule oscillates in the $v_2$ or the $v_1$ modes, all three internuclear distances increase simultaneously.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=13cm]{fig8.pdf}
\caption{The $A_v$ rotational constants for $A_2 (D_{3h})$ vibrational states. Normal mode quantum numbers are specified for a few levels. Above the dissociation threshold (vertical dotted red line), the vibrational levels are predissociated.}
\label{fig:Bv_A2}
\end{figure}
Figure~\ref{fig:Bv_A2} shows the vibrational dependence of the rotational constants $A_v$ obtained for $A_2$ vibrational symmetry in the $D_{3h}(M)$ group. The energy origin of the figure is the same as in Fig.~\ref{fig:Bv_A1}, i.e. the energy of the ground rovibrational level of ozone $(0,0,0),J=0$. The same three families of vibrational levels corresponding to the three normal modes, are easily identified. The figure also includes some of the low-energy predissociated resonances above the dissociation limit. Figure \ref{fig:resonances_A2} shows widths of the $A_2$ vibrational levels situated above the dissociation threshold. Most of the resonances shown in the figure have widths between 2 and 70~\ensuremath{\rm cm^{-1}}\xspace (lifetimes between 0.08 and 2~ps) with a few outliers having significantly smaller widths. These outliers are the levels highly-excited in the $v_1$ mode, as demonstrated in Figs.~\ref{fig:Jacobi_A2_v1_diss} and Figs.~\ref{fig:Jacobi_A2_diss}.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig9.pdf}
\caption{Widths, $\Gamma$, of resonances of the $A_2 (D_{3h})$ vibrational levels. Vertical dotted lines indicate threshold energies for dissociation channels with different excitation of the oxygen molecule $v_d = 0, 1,$ and 2. Numerically, lifetimes $\tau$ in ps are related to the widths in \ensuremath{\rm cm^{-1}}\xspace as $\tau[\mathrm{ps}]=(2\pi c \Gamma[\mathrm{cm^{-1}}])^{-1}$, where $c$ is the speed of light in units of cm/ps, $c=0.0299792458$ cm/ps. Wave functions of the encircled levels are shown in Figs.~\ref{fig:Jacobi_A2_v1_diss} and~\ref{fig:Jacobi_A2_diss}. }
\label{fig:resonances_A2}
\end{figure}
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig10.pdf}
\caption{Same as Fig.~\ref{fig:Jacobi_A1_v1_v2}, but for the $(7,0,1)$ and $(0,10,1)$ wave functions, of $A_2(D_{3h})$ overall vibrational symmetry.}
\label{fig:Jacobi_A2_v1_v2}
\end{figure}
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig11.pdf}
\caption{Vibrational part of the wave functions of the $(8,0,1)$ (upper panels) and $(9,0,1)$ (lower panels) levels of $A_2(D_{3h})$ vibrational symmetry. The calculated widths are $\Gamma=0.016$~\ensuremath{\rm cm^{-1}}\xspace for the $(8,0,1)$ level and and 1.7~\ensuremath{\rm cm^{-1}}\xspace for $(9,0,1)$. }
\label{fig:Jacobi_A2_v1_diss}
\end{figure}
Figures~\ref{fig:Jacobi_A2_v1_v2}, \ref{fig:Jacobi_A2_v1_diss}, \ref{fig:Jacobi_A2_diss} shows some of the bound and resonance vibrational levels of $A_2$ vibrational symmetry of the $D_{3h}$ group. The vibrational levels $(v_1,v_2,v_3)$ with odd $v_3$ have overall $A_2$ symmetry. As mentioned above, continuum states (including dissociative states) of ozone $^{16}$O$_3$ can only be of $A_2$ vibrational symmetry.
Figure~\ref{fig:Jacobi_A2_v1_diss} demonstrates two resonance wave functions from the $(v_1,0,1)$ series. Although excitation of the $v_1$ mode differs for these two levels only by one quantum, their lifetimes are very different, 330~ps for the $(8,0,1)$ and 3.1~ps for the $(9,0,1)$ level. Figure~\ref{fig:Jacobi_A2_diss} shows two examples of wave functions for levels where all three modes are excited and mixed.
\begin{figure}[ht]
\vspace{1cm}
\includegraphics[width=12cm]{fig12.pdf}
\caption{Vibrational part of the wave functions of two highly-excited levels of the $A_2(D_{3h})$ symmetries. The calculated widths are $\Gamma=0.36$~\ensuremath{\rm cm^{-1}}\xspace for the function shown in the upper panels and 0.8 ~\ensuremath{\rm cm^{-1}}\xspace for the function shown in the lower panels. }
\label{fig:Jacobi_A2_diss}
\end{figure}
Figure \ref{fig:energy_levels} shows the energies of symmetry-allowed levels for the two lowest values of the angular momentum, $J=0$ and 1. The standard notation notation $\{J_{K_aK_c}\}$ for rotational states of an asymmetric top molecule is used for bound states within the well. The irreducible representation in the $S_3$ permutation group of vibrational part of the total wave function is also specified. Only states of $A_2 (D_{3h})$ vibrational irrep can dissociate and, therefore, only resonances of this symmetry are shown in the figure.
\begin{figure}[h]
\vspace{1cm}
\includegraphics[width=12cm]{fig13.pdf}
\caption{Energies of vibrational levels for $J=0$ and $J=1$. In addition to the symbol $\{J_{K_aK_c}\}$ of the rotational states of an asymmetric top molecule, irreducible representations of the vibrational part ($\Gamma^v$, below the graph)
and rotation part ($\Gamma^r$, above the graph) of the wave function are specified.
The corresponding resonances are shown in blue color.}
\label{fig:energy_levels}
\end{figure}
\section{Conclusion}
In this study, energies, widths, and wave functions of $^{16}$O$_3$ vibrational resonances were determined for levels up to about 3000 \ensuremath{\rm cm^{-1}}\xspace above the dissociation threshold. The predissociated resonances have lifetimes between 0.08 and 2~ps with a few long-living levels. These outliers are levels with the highly-excited $v_1$ and $v_2$ modes. An example of a long living state is (9,0,0) $J$=1 level with the lifetime of 330~ps. Energies of bound states of the ozone isotopologue $^{16}$O$_3$ up to the dissociation threshold were also computed. The total permutation inversion symmetry $S_3\times I$ of the three oxygen atoms was taken into account using hyper-spherical coordinates. The effect of the symmetry is negligible for the levels deep in the ozone potential, but vibrational levels near the dissociation threshold cannot be represented correctly within one potential well and, therefore, the complete permutation symmetry group should be used.
Symmetry properties of allowed rovibrational levels of ozone (applicable to $^{16}$O$_3$ and $^{18}$O$_3$) as well as correlation diagrams between the bound-state and dissociation regions were derived and discussed. The correlation diagrams are not trivial because ozone dissociates to (or is formed from)
a $P$-state oxygen atom and an O$_2$ molecule of symmetry $^3\Sigma^-_g$.
Within the employed model including only the lowest PES of ozone, the purely vibrational states, i.e. $J=0$ states, of ozone $^{16}$O$_3$ (and $^{18}$O$_3$)
cannot dissociate to the fragments allowed by symmetry of the electronic ground state of the O$_2$ molecule. Note that excited rotational states with $J>0$ satisfying Eq.~(\ref{eq:scattering}) do exist. Examples of such resonances are shown in Figs.~\ref{fig:Jacobi_A2_v1_diss} and \ref{fig:Jacobi_A2_diss}. We would like to stress here, that the single electronic PES model neglects the coupling of the angular momentum of the molecular frame, $\boldsymbol R$, with the electronic angular momentum, $\boldsymbol L$, which is not zero.
In general, the total (but without nuclear spin, $\boldsymbol I$) angular momentum $\boldsymbol J$ can be written as
$\boldsymbol J = \boldsymbol R + \boldsymbol L + \boldsymbol S + \boldsymbol \Pi$, where $\boldsymbol S$ is the electronic spin and $\boldsymbol \Pi$
the vibrational angular momentum. From this we obtain the approximate quantum number of the rotation of the molecular frame as $\boldsymbol R = \boldsymbol J - \boldsymbol L - \boldsymbol S - \boldsymbol \Pi$.
Neglecting the effect of $\boldsymbol L$ and $\boldsymbol S$ in the rovibrational problem,
$\boldsymbol R \approx \boldsymbol J$ is a ``good'' quantum number. Our rovibrational energies have been calculated within this approximation, as have been those obtained by other workers in the field.
However, the importance of the electronic angular momentum is evident from asymptotic behavior of Eq.~(\ref{eq:scattering}):
At large distances between O and $\rm O_2$, the electronic angular momentum is clearly not zero. In a more accurate model, the electronic momentum should be
accounted for and coupled to the angular momentum of the nuclear frame, due to the cross-terms generated by $\boldsymbol R^2$, conserving the total angular momentum, $\boldsymbol J$.
In such a more accurate model,
the continuum vibrational spectrum for $J \approx R = 0$ is allowed (since $R$ is not ``conserved'' any more). The corresponding vibrational resonances should have relatively long lifetimes because they can only decay due to non-Born-Oppenheimer and Coriolis couplings involving the three PES's converging to the same dissociation limit, with the oxygen atom being in the triply-degenerate electronic state. Such long-living states above the dissociation threshold, for example, $(9,0,0)$ and $(10,0,0)$ have indeed been observed in experiments. Therefore, an accurate theoretical determination of lifetimes of $J=0$ resonance levels would involve three potential energies surfaces.
The above discussion did not take into account spin-orbit coupling. For even more realistic description of the nuclear motion states in the continuum, one has to consider
the effect of coupling of the electronic singlet state with electronic triplet states, of symmetry $B_1$ and $A_2$ in $C_{2v}(M)$, or $A_2''$ and $A_1''$ in $D_{3h}(M)$,
or $A''$ in $C_s(M)$, that approach the same asymptotic dissociation limit, O$(^3P)$
and $\rm O_2$($X^3\Sigma_g^-$), as the electronic ground state.
Rosmus, Palmieri and Schinke~\cite{ROS02:4871} have determined the spin-orbit coupling elements with
all relevant triplet states in the asymptotic channel.
The matrix elements are of the order of $\langle X\, ^1A_1' | H_{so} | ^3A_2'' \rangle \approx 60 \, \ensuremath{\rm cm^{-1}}\xspace$. The long-range behavior of the potential energy surfaces accounting for spin-orbit coupling was discussed in Ref.~\cite{lepers2012long}.
To study the effect of spin-orbit coupling, the total nuclear-electronic wave function
must be expanded, including for simplicity just one generic triplet state, as
\begin{equation}
\label{eq:so-expansion}
\Psi_{vJm}(\Omega,{\cal Q})=\sum_k \left[ c_1 \, |^1A_1'\rangle \psi_{vJk; ^1A_1'}({\cal Q})
+ c_2 \, |^3A_2''\rangle \psi_{vJk; ^3A_2''}({\cal Q}) \right] {\cal R}_{Jkm}(\Omega)\,.
\end{equation}
As before, the product of electronic and nuclear motion functions must have the same symmetry
as the rotational function, except for their parity. However, the nuclear motion component of
the $^3A_2''$ electronic state is antisymmetric and can therefore correlate with the
asymptotic $\rm O_2 \, + \, O$ wave function. A full treatment of the nuclear dynamics of ozone accounting for the spin-orbit coupling would involve solving the rovibrational Schr\"odinger equation on several coupled potential energy surfaces, which is hardly possible at present. However, an adiabatic approach with respect to the spin-orbit coupling should also be accurate and could be used in a future study. In the approach, the first step would be to construct the matrix of the potential energy. The matrix would include the lowest three Born-Oppenheimer PES's, mentioned above, and the spin-orbit coupling such as described in Refs.~\cite{ROS02:4871,lepers2012long}. The matrix then should be diagonalized for each geometry, which will produce adiabatic potential surfaces accounting for the spin-orbit coupling. Because the lowest Born-Oppenheimer PES ($X^1A_1$) does not cross the two other PES's at energies near or below the dissociation threshold, after the diagonalization, the lowest obtained PES will be very similar to the original $X^1A_1$ Born-Oppenheimer PES, except that the dissociation limit will be shifted down. Low-energy rovibrational states obtained with the new adiabatic PES will be almost identical to the ones discussed in this study. States near the dissociation threshold (approximately 60~\ensuremath{\rm cm^{-1}}\xspace above and below $D_0$) will have energies somewhat different compared to the ones obtained with the PES without the spin-orbit coupling. However, qualitatively the structure of rovibrational levels near the dissociation will stay the same.
\section*{Acknowledgments}
This work is supported by the CNRS through an invited professor position for V. K. at the GSMA,
the National Science Foundation, Grant No PHY-15-06391
and the ROMEO HPC Center at the University of Reims Champagne-Ardenne. The supports from Tomsk State University Academic D. Mendeleev
funding Program, from French LEFE Chat program of the Institut National des Sciences de l'Univers (CNRS) and from Laboratoire International Franco-Russe SAMIA are acknowledged.
\clearpage
| 043939c49ecf2670497167ddeed27ee497368bee | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Background}
This section will very briefly review definitions from Baker and Bowler's paper (\cite{BB17}), but only as a reminder to the reader who has already read the more complete treatment there.
\begin{nota} $E$ denotes a finite set. $[n]$ denotes $\{1,2,\ldots,n\}$. $F$ denotes a set $G\cup\{0\}$, where $(G, N_G)$ is a tract (Definition~\ref{def-tract}).
$\mathbf 0$ denotes the vector in $F^E$ with all components $0$.
The {\bf support} of $X\in F^E$ is $\underline{X}:=\{e\in E: X(e)\neq 0\}$.
The {\bf zero set} of $X$ is $X^0:=\{e\in E: X(e)= 0\}$.
The letter $i$ will always denote $\sqrt{-1}$ in $\mathbb C$.
If $X\in F^E$ and $e\in E$ then $X\backslash e$ denotes the restriction of $X$ to $E-\{e\}$. If $f\not\in E$ and $\alpha\in F$ then $Xf^\alpha$ denotes the extension of $X$ to $E\cup\{f\}$ with $X(f)=\alpha$.
For any $S\subseteq F^E$, $\mathrm{Minsupp}(S)$ denotes the set of elements of $S$ of minimal support.
\end{nota}
\subsection{Tracts and hyperfields}
\begin{defn}\label{def-tract} (\cite{BB17}) A {\bf tract} is a multiplicative group $G$ together with a subset $N_G$ of the group semiring ${\mathbb N}[G]$ satisfying all of the following.
\begin{enumerate}
\item the zero element of ${\mathbb N}[G]$ belongs to $N_G$.
\item the identity element $1$ of $G$ is not in $N_G$.
\item There is a unique element $\eta$ of $G$ with $1+\eta\in N_G$.
\item $N_G$ is closed under the action of $G$ on $N_G$.
\end{enumerate}
\end{defn}
We often refer to the set $F=G\cup\{0\}$ as the tract, and for $g\in G$ we often denote $\eta g$ as $-g$. If $a_1, \ldots, a_k\in F$ then $a_1\boxplus\cdots \boxplus a_k$ (or $\bighplus_{j=1}^k a_j$) denotes $\{b\in F: -b+\sum\limits_{j=1}^k a_j\in N_G\}$. The canonical example of a tract is when $F$ is a field, $G=F-\{0\}$, and $N_G$ is the set of all formal sums of elements of $F$ which add to 0 in $F$. In this case the set $\bighplus_{j=1}^k a_j$ has exactly one element, the actual sum of the $a_j$.
A particularly interesting class of tracts arise from {\em hyperfields}, which we now define.
\begin{defn} A {\bf hyperoperation} on a set $R$ is a map $\boxplus$ from $R\times R$ to the set of nonempty subsets of $R$.
If $a,b,c\in R$ then $(a\boxplus b)\boxplus c$ is defined to be ${\bigcup_{x\in a\boxplus b} x\boxplus c}$. Likewise define $a\boxplus(b\boxplus c)$.
\end{defn}
As in~\cite{BB17}, all hyperoperations in this paper will be commutative and associative, with identity $0$. Thus for any finite $S\subseteq R$ the sum ${\bighplus_{s\in S} s}$ is well-defined, with ${\bighplus_{s\in \emptyset} s}$ defined to be $\{0\}$.
Note that the $\boxplus$ defined for a tract is not necessarily a hyperoperation, since for elements $a,b$ in a tract, $a\boxplus b$ might be the empty set.
\begin{defn} [\cite{BB17})]\label{defn:hypergroup} A commutative {\bf hypergroup} is a tuple $(G,\boxplus,0)$, where $\boxplus$ is a commutative and associative hyperoperation on $G$, such that
\begin{enumerate}
\item $0\boxplus x=\{x\}$ for all $x\in G$.
\item For every $x\in G$ there is a unique element of $G$, denoted $-x$, such that $0\in x\boxplus -x$.
\item For all $x$, $y$, and $z$, $x\in y\boxplus z$ if and only if $z\in x\boxplus (-y)$.
\end{enumerate}
A commutative {\bf hyperfield} is a tuple $(R,\cdot,\boxplus,1,0)$ such that $0\neq 1$ and
\begin{enumerate}
\item $(R-\{0\},\cdot, 1)$ is a commutative group.
\item $(R, \boxplus, 0)$ is a commutative hypergroup.
\item $0\cdot x=x\cdot 0=0$ for all $x\in R$.
\item $a\cdot(x\boxplus y)=(a\cdot x)\boxplus(a\cdot y)$ for all $a,x,y\in R$
\end{enumerate}
\end{defn}
If $F$ is a hyperfield then let $G=F-\{0\}$ and let $N_G=\{\sum\limits_{j=1}^k a_j: 0\in\bighplus_{j=1}^k a_j\}$. Then $G$ and $N_G$ define a tract whose associated operation $\boxplus$ coincides with the hyperaddition in $F$.
If $F$ is a tract and $X,Y\in F^E$ then we define $$X\boxplus Y:=\{Z: \forall e\ Z(e)\in X(e)\boxplus Y(e)\}.$$ $F$ acts on $F^E$ by componentwise multiplication. (If $F$ is a hyperfield then $F^E$ is an $F$-module in the sense of~\cite{BB17}.)
\begin{defn} [\cite{BB17}] A homomorphism $(G, N_G)\to (G', N_{G'})$ of tracts is a group homomorphism $f:G\to G'$ together with a map $\tilde{f}:{\mathbb N}[G]\to{\mathbb N}[G']$ such that $\tilde{f}(\sum a_jg_j)=\sum a_jf(g_j)$ for all $a_j\in\mathbb N$ and $g_j\in G$ and $F(N_G)\subseteq N_{G'}$.
\end{defn}
We'll also use $f$ to denote the extension $(F=G\cup\{0\})\to (F'=G'\cup\{0\})$ sending 0 to 0, and we'll use $f$ to denote the componentwise map $F^E\to (F')^E$.
If $F$ and $F'$ are hyperfields then a homomorphism of the corresponding tracts amounts to a function $f: F\to F'$ such that $f(xy)=f(x)f(y)$ and $f(x\boxplus y)\subseteq f(x)\boxplus f(y)$ for all $x$ and $y$.
\begin{defn} [\cite{BB17}] Let $R$ be a tract equipped with an automorphism $c$ such that $c^2=\mathrm{id}$, which we call {\bf conjugation}. Denote the image of $x\in R$ under $c$ by $x^c$. For $X,Y\in R^E$, the {\bf inner product} is defined as
$$X\cdot Y:=\sum_{e\in E} X(e)\cdot Y(e)^c\in {\mathbb N}[G].$$
We say $X$ and $Y$ are {\bf orthogonal}, denoted $X\perp Y$, if $X\cdot Y\in N_G$. If $S\subseteq R^E$, we denote by $S^\perp$ the set of all $X\in R^E$ such that $X\perp Y$ for all $Y\in S$.
\end{defn}
We will always view a tract as being equipped with $c$. When no $c$ is specified, we will take $c=\mathrm{id}$.
\begin{example} \label{ex:tracts}
Viro's paper~\cite{Viro1} provides an excellent introduction to and motivation for hyperfields. Several of the following hyperfields were first introduced there.
Notice that several of these hyperfields have operations defined in terms of the usual field operations on $\mathbb R$ and $\mathbb C$. The symbol $+$, when denoting a binary operation, always denotes the usual addition.
\begin{enumerate}
\item The field $\mathbb R$ is a hyperfield, with $c=\mathrm{id}$.
\item The field $\mathbb C$ is a hyperfield, with $c$ being conjugation.
\item The {\bf Krasner hyperfield} $\mathbb K$ on elements $\{0,1\}$ is the unique two-element hyperfield for which $1\boxplus 1=\{0, 1\}$. The automorphism $c$ is the identity.
\item \label{def:signhyp} The {\bf sign hyperfield} $\mathbb S$ on elements $\{0,+,-\}$ has addition $0\boxplus x=\{x\}$ for all $x$, $x\boxplus x=\{x\}$ for all $x$, and $\boxplus -=\{0,+,-\}$. Multiplication is given by $0\odot x=0$ for all $x$, $+\odot +=-\odot -=+$, and $+\odot -=-$. The automorphism $c$ is the identity.
\item \label{phasetracts} Let $S^1$ denote the unit circle in $\mathbb C$. For $x\in\mathbb C$, the {\bf phase} of $x$ is
$$\operatorname{ph}(x)=\left\{\begin{array}{ll}
0&\mbox{ if $x=0$}\\
\frac{x}{|x|}&\mbox{ otherwise}
\end{array}\right.$$
The {\bf phase hyperfield} $\mathbb P$ on elements $S^1\cup\{0\}$ has addition $x\boxplus y:=\{\operatorname{ph}(ax+by): a,b\in\mathbb R_{>0}\}$ and has multiplication inherited from $\mathbb C$. The automorphism $c$ is conjugation. ( For nonzero $x$ we have $x^c=x^{-1}$, and in~\cite{AD} the inner product is described in this way.)
This is not the hyperfield which gets the name ``phase hyperfield" in~\cite{Viro1}: see~(\ref{tropicalphase}) below.
\item (\cite{Viro1}) The {\bf triangle hyperfield} $\triangle$ on elements $\mathbb R_{\geq 0}$ has addition $x\boxplus y:=\{z: |x-y|\leq z\leq x+y\}$ and has multiplication inherited from $\mathbb R$. The automorphism $c$ is the identity.
\item (\cite{Viro1}) The {\bf tropical real hyperfield} $\mathbb T\mathbb R$ on elements $\mathbb R$ has addition
$$x\boxplus y:=\left\{\begin{array}{ll}
x&\mbox{ if $|x|>|y|$ or $x=y$}\\
y&\mbox{ if $|x|<|y|$}\\
\{z: |z|\leq |x|\}&\mbox{ if $x=-y$}
\end{array}\right.$$
and has multiplication inherited from $\mathbb R$. The automorphism $c$ is the identity.
\item (\cite{Viro1}) The {\bf tropical complex hyperfield} $\mathbb T\mathbb C$ on elements $\mathbb C$ has addition
$$x\boxplus y:=\left\{\begin{array}{ll}
x&\mbox{ if $|x|>|y|$}\\
y&\mbox{ if $|x|<|y|$}\\
\{|x|\operatorname{ph}(ax+by):a,b\in\mathbb R_{\geq 0}\}&\mbox{ if $|x|=|y|$ and $x\neq -y$}\\
\{z: |z|\leq |x|\}&\mbox{ if $x=-y$}
\end{array}\right.$$
and has multiplication inherited from $\mathbb C$. The automorphism $c$ is conjugation.
\item \label{tropicalphase} (\cite{Viro1}) The {\bf tropical phase hyperfield} ${\mathbb T}\P$ on elements $S^1\cup\{0\}$ (called the {\em phase hyperfield} in~\cite{Viro1}) has addition
$$x\boxplus y:=\left\{\begin{array}{ll}
S^1\cup\{0\}&\mbox{ if $x=-y\neq 0$}\\
\{\operatorname{ph}(ax+by): a,b\in\mathbb R_{\geq0}\}&\mbox{otherwise}
\end{array}\right.$$
and has multiplication inherited from $\mathbb C$. The automorphism $c$ is conjugation.
\item (\cite{Viro1}) The {\bf ultratriangle hyperfield} $\mathbb T\triangle$ on elements $\mathbb R_{\geq 0}$ has addition
$$x\boxplus y=\left\{\begin{array}{ll}
\max(x,y)&\mbox{ if $x\neq y$}\\
\{z:z\leq x\}&\mbox{ if $x=y$}
\end{array}\right.$$
and has multiplication inherited from $\mathbb R$. The automorphism $c$ is the identity.
\end{enumerate}
\end{example}
\begin{example} Here are some examples of conjugation-preserving morphisms of tracts.
\begin{enumerate}
\item The inclusion $\mathbb R\to\mathbb C$ is a morphism of tracts.
\item For every tract $F$ the function $\kappa:F\to \mathbb K$ sending each nonzero element of $F$ to $ 1$ is a morphism of tracts.
\item The function $\operatorname{ph}: \mathbb C\to\mathbb P$ introduced in Example~\ref{ex:tracts}.\ref{phasetracts} is a morphism of tracts. It restricts to a morphism $\mathrm{sign}: \mathbb R\to\mathbb S$.
\item Other morphisms which are interesting but will not be considered further here include
\begin{enumerate}
\item $\mathbb C\to\triangle$ taking each $x\in\mathbb C$ to $|x|$
\item $\mathrm{sign}:\mathbb T\mathbb R\to \mathbb S$
\item $\operatorname{ph}: \mathbb T\mathbb C\to{\mathbb T}\P$
\item $\mathbb T\mathbb C\to\mathbb T\triangle$ taking each $x\in\mathbb T\mathbb C$ to $|x|$
\end{enumerate}
\end{enumerate}
\end{example}
\subsection{Matroids over tracts}
\begin{defn}\cite{BB17} Let $L$ be a finite lattice with minimal element $\hat{0}$. An element $a\in L$ is an {\bf atom} if $a\neq\hat{0}$ and there is no $b\in L$ such that $\hat{0}<b<a$. A set $S$ of atoms in $L$ is a {\bf modular family} if the height of $S$ in $L$ is $|S|$.
If $\mathcal{C}$ is a subset of $F^E$ then $S\subseteq\mathcal{C}$ is called a {\bf modular family} if the set of supports of elements of $S$ is a modular family in the lattice of unions of supports of elements of $\mathcal{C}$.
\end{defn}
\begin{defn}[\cite{BB17}] \label{strongcircaxs} Let $E$ be a nonempty finite set, $F$ a tract, and $\mathcal{C}\subseteq F^E$. We say $\mathcal{C}$ is the set of {\bf $F$-circuits of a strong $F$-matroid on $E$} if $\mathcal{C}$ satisfies all of the following axioms.
\begin{description}
\item [Nontriviality] $\mathbf 0\not\in \mathcal{C}$.
\item [Symmetry] If $X\in \mathcal{C}$ and $\alpha\in F-\{0\}$ then $\alpha X\in \mathcal{C}$.
\item [Incomparability] If $X,Y\in \mathcal{C}$ and $\underline{X}\subseteq\underline{Y}$ then there exists $\alpha\in F-\{0\}$ such that $X=\alpha Y$.
\item [Strong modular elimination] Let $\{X_1, \ldots, X_k, X\}\subseteq\mathcal{C}$ be a modular family of size $k+1$ such that $\underline{X}\not\subseteq\cup_{j=1}^k\underline{X_j}$, and for each $1\leq j\leq k$ let $e_j\in (\underline{X}\cap \underline{X_j})\backslash\cup_{l\neq j} \underline{X_l}$ be such that $X(e_j)=-X_j(e_j)$. Then there is an $F$-circuit $Z\in\mathcal{C}$ such that $Z(e_j)=0$ for each $1\leq j\leq k$ and $Z\in X\boxplus \bighplus_{j=1}^k X_j$.
\end{description}
\end{defn}
\begin{remark} Baker and Bowler also defined {\em weak $F$-matroids}, which only require Strong Modular Elimination for modular families of size 2.
If $F$ is a field, $F=\mathbb K$, or $F=\mathbb S$, then every weak $F$-matroid is a strong $F$-matroid. If $F$ is a tract for which strong and weak $F$-matroids coincide, we will refer simply to {\bf $F$-matroids}.
\end{remark}
The following is part of Theorem 2.24 in~\cite{BB17}.
\begin{thm} Let $\mathcal{C}$ be the $F$-circuit set of a strong $F$-matroid. Then $\mathcal{C}^*:=\mathrm{Minsupp}(\mathcal{C}^\perp-\{\mathbf 0\})$ is also the $F$-circuit set of a strong $F$-matroid, and $(\mathcal{C}^*)^*=\mathcal{C}$.
\end{thm}
A {\bf strong $F$-matroid} is the information encoded in a set $\mathcal{C}$ satisfying Definition~\ref{strongcircaxs}. Thus a strong $F$-matroid $\mathcal{M}$ has an associated pair $(\mathcal{C}(\mathcal{M}),\mathcal{C}^*(\mathcal{M}))$. $\mathcal{C}(M)$ is the set of {\bf $F$-circuits of $\mathcal{M}$}, and $\mathcal{C}^*(\mathcal{M})$ is the set of {\bf $F$-cocircuits of $\mathcal{M}$}. The strong $F$-matroid with $F$-circuit set $\mathcal{C}^*(\mathcal{M})$ and $F$-cocircuit set $\mathcal{C}(\mathcal{M})$ is called the {\bf dual} to $\mathcal{M}$ and is denoted $\mathcal{M}^*$.
\begin{remark} Baker and Bowler give another characterization of $F$-matroids as well, via {\em Grassmann-Pl\"ucker functions}, which generalize chirotopes of oriented matroids.
\end{remark}
\begin{example} Matroids in the usual sense are essentially $\mathbb K$-matroids. Specifically, if $S\subseteq E$ let $X_S\in \mathbb K^E$ be the indicator function for $S$ (so $X_S(e)= 1$ if and only if $e\in S$). Then $C\subseteq 2^E$ is the set of circuits of a matroid in the usual sense if and only if $\{X_S: S\in C\}$ is the set of $\mathbb K$-circuits of a $\mathbb K$-matroid (Corollary 1 in~\cite{Del11}).
When we refer to a {\em matroid} without modification, we mean a matroid in the traditional sense -- an object with circuits which are subsets of $E$. When we wish to use the language of matroids over tracts, we will talk about $\mathbb K$-matroids.
\end{example}
\begin{example} $\mathcal{C}\subseteq\mathbb S^E$ is the set of $\mathbb S$-circuits of an $\mathbb S$-matroid if and only if $\mathcal{C}$ is the set of signed circuits of an oriented matroid. This follows from Corollary 1 in~\cite{Del11} together with Theorem 3.6.1 in~\cite{BLSWZ}.
\end{example}
\begin{example} If $K$ is a field,
then $\mathcal{C}^*\subseteq K^E$ satisfies the $K$-circuit axioms if and only if $\mathcal{C}^*=\mathrm{Minsupp}(V-\{\mathbf 0\})$ for some linear subspace $V$ of $K^E$ (\cite{BB17}). We call the $K$-matroid with $K$-cocircuit set $\mathcal{C}^*$ the {\bf $K$-matroid corresponding to $V$.}
\end{example}
\begin{prop}[\cite{BB17}] \label{prop:morph} A conjugation-preserving morphism $f:F\to F'$ of tracts induces a map $f_*$ from (weak resp.\ strong) $F$-matroids to (weak resp.\ strong) $F'$-matroids so that
$$\mathcal{C}(f_*(\mathcal{M}))=\{\alpha f(X): X\in\mathcal{C}(\mathcal{M}), \alpha\in F'-\{0\}\}$$
and
$$\mathcal{C}^*(f_*(\mathcal{M}))=\{\alpha f(X): X\in\mathcal{C}^*(\mathcal{M}), \alpha\in F'-\{0\}\}.$$
\end{prop}
\begin{example} If $\mathcal{M}$ is an $F$-matroid then the ordinary matroid corresponding to $\kappa_*(\mathcal{M})$ is called the {\bf underlying matroid} of $\mathcal{M}$.
Since underlying matroids are well-defined, we can refer to classical matroid properties of an $F$-matroid, e.g.\ bases and circuits. (For instance, a circuit of an $F$-matroid $\mathcal{M}$ is defined to be a circuit of the underlying matroid, hence is the support of an $F$-circuit of $\mathcal{M}$.)
\end{example}
\begin{example} If $V$ is a linear subspace of $\mathbb R^n$ and $\mathcal{M}_V$ is the $\mathbb R$-matroid associated to $V$ then $\mathrm{sign}_*(\mathcal{M})$ is the oriented matroid associated to $V$.
Similar comments hold for subspaces of $\mathbb C^n$ and strong $\mathbb P$-matroids.
\end{example}
\begin{cor}\label{cor:underlying} If $\mathcal{M}$ is a (weak or strong) $F$-matroid then there is a bijection $G X\to\underline{X}$ from the set of $G$-orbits of circuits of $\mathcal{M}$ to the set of circuits of the underlying matroid.
\end{cor}
\section{Main Theorem: $F$-vectors}
This section introduces our new axiomatization of strong $F$-matroids.
The idea is to generalize the following vision of subspaces of a vector space: let $K$ be a field and $V$ the row space of a matrix $M$ over $K$ with columns indexed by $E$. A subset $B$ of $E$ is a basis of the $K$-matroid corresponding to $V$ if and only if the set of columns indexed by $B$ is a basis for the column space. In this case $M$ is equivalent by row operations to a matrix $M_B$ such that the submatrix of $M_B$ with columns indexed by $B$
is an identity matrix. We call $M_B$ a {\em reduced row-echelon form for $M$ with respect to $B$}. A necessary
condition for an element $X$ of $K^E$ to be in $V$ is that $X$ must be a linear combination of the rows of $M_B$. (Of course, because $K$ is a field this is also a sufficient condition, but for general tracts, and even hyperfields, this will take more thought.)
Let $\{R_j:j\in B\}$ denote the set of rows of $M_B$, where $R_j$ is the unique element of $V$ such that for all $k\in B$, $R_j(k)=\delta_{jk}$. (Here
$\delta_{jk}$ denotes the Kronecker delta.)
The above condition for $X$ to be in $V$ can be rephrased as
\begin{equation}\label{eqn:rref}
X=\sum_{j\in B} X(j)R_j
\end{equation}
Our generalization of rows in reduced-row-echelon-form matrices is $F$-cocircuits, and our definition of $F$-covectors is vectors satisfying Equation~(\ref{eqn:rref}) (with hypersum $\boxplus$ in place of sum) with respect to every basis.
In order to get a stand-alone definition of $F$-vectors and $F$-covectors, we first need to define bases without reference to circuits or cocircuits.
\begin{defn} Let $F$ be a tract, $E$ a finite set, and $\mathcal{W}\subseteq F^E$.
A {\bf support basis} of $\mathcal{W}$ is a minimal $B\subseteq E$ such that $B\cap\underline{X}\neq\emptyset$ for every $X\in\mathcal{W}-\{\mathbf 0\}$.
\end{defn}
\begin{lemma}\label{lem:supportbasis1} 1. For any $\mathcal{W}\subseteq F^E$, the set of support bases of $\mathcal{W}$ is the set of support bases of $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$.
2. If $\mathcal{M}$ is a (weak or strong) $F$-matroid on $E$ and $B\subseteq E$ then $B$ is a basis for $\mathcal{M}$ if and only if $B$ is a support basis for $\mathcal{C}^*(\mathcal{M})$.
\end{lemma}
\begin{proof} (1) is clear. (2) follows from Corollary~\ref{cor:underlying} together with Proposition 2.1.16 in~\cite{Oxley}.
\end{proof}
\begin{defn} Let $B$ be a support basis for $\mathcal{W}$.
A {\bf nearly reduced row-echelon form} for $\mathcal{W}$ with respect to $B$ is a subset $\{S_j:j\in B\}$ of $\mathcal{W}$ such that $\underline{S_j}\cap B=\{j\}$ for each $j\in B$. A {\bf reduced row-echelon form} for $\mathcal{W}$ with respect to $B$ is a nearly reduced row-echelon form $\{R_j:j\in B\}$ such that $R_j(k)=\delta_{jk}$ for each $j,k\in B$.
\end{defn}
In other words, a reduced row-echelon form is a nearly reduced row-echelon form $\{S_j:j\in B\}$ such that $S_j(j)=1$ for every $j$. A reduced row-echelon form with respect to $B=\{b_1, \ldots, b_r\}$ is a subset $\{R_{b_j}:j\in[r]\}$ of $\mathcal{W}$ such that, if $(b_1, \ldots, b_r)$ is extended to an ordering $(b_1, \ldots, b_n)$ of $E$, then $\{R_{b_j}:j\in[r]\}$ of $\mathcal{W}$ is the set of rows of a matrix with columns indexed by $E$ of the form $(I|A)$.
\begin{lemma}\label{lem:minrref} Let $\mathcal{W}\subseteq F^E$ and $B$ a support basis.
\begin{enumerate}
\item $\mathcal{W}$ has at least one nearly reduced row-echelon form with respect to $B$.
\item If $\mathcal{W}$ is closed under multiplication by $F-\{0\}$ then $\mathcal{W}$ has at least one reduced row-echelon form with respect to $B$.
\item $\{S_j:j\in B\}\subseteq \mathcal{W}$ is a (nearly) reduced row-echelon form for $\mathcal{W}$ if and only if it is also a (nearly) reduced row-echelon form for $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$.
\end{enumerate}
\end{lemma}
The proof is straightforward.
\begin{lemma}\label{lem:circrref} If $\mathcal{M}$ is a (weak or strong) $F$-matroid then, for every basis $B$, $\mathcal{C}^*(\mathcal{M})$ has a unique reduced row-echelon form with respect to $B$.
\end{lemma}
This follows from the following lemma, which introduces the idea of {\bf fundamental $F$-circuit $FC(e,B)$} and {\bf fundamental $F$-cocircuit $FC^*(e,B)$}. The reduced row-echelon form promised by Lemma~\ref{lem:circrref} is $\{FC^*(j,B):j\in B\}$.
\begin{lemma}\label{lem:fundcirc} Let $\mathcal{M}$ be a (weak or strong) $F$-matroid and $B$ a basis for $\mathcal{M}$.
\begin{enumerate}
\item If $e\in E-B$ then there is a unique element of $\mathcal{C}(\mathcal{M})$, denoted $FC(e,B)$, such that $\underline{FC(e,B)}\subseteq B\cup\{e\}$ and $FC(e,B)(e)=1$.
\item If $j\in B$ then there is a unique element of $\mathcal{C}^*(\mathcal{M})$, denoted $FC^*(j,B)$, such that $\underline{FC^*(j,B)}\subseteq (E-B)\cup\{j\}$ and $FC^*(j,B)(e)=1$.
\item If $X\in\mathcal{C}(\mathcal{M})$ and $e\in\underline{X}$ then there is a basis $B'$ for $\mathcal{M}$ such that $X$ is a scalar multiple of $FC(e,B')$.
\end{enumerate}
\end{lemma}
\begin{proof} 1. An elementary result of matroid theory (cf. 1.2.6 in~\cite{Oxley}) says that
there is a unique circuit $C(e,B)$ of the underlying matroid of $\mathcal{M}$ such that $e\in C(e,B)\subseteq B\cup\{e\}$.
By Proposition~\ref{prop:morph} there is an element of $\mathcal{C}(\mathcal{M})$ with support $C(e,B)$, and by Symmetry and Incomparability there is a unique such element with value 1 on $e$.
The proof of (2) is similar.
The proof of (3) follows from a standard matroid theory result: $\underline{X}-\{e\}$ is independent in $\mathcal{M}$, hence extends to a basis $B'$. Uniqueness of $C(e,B')$ implies that $\underline{X}=C(e,B')$, and so the result follows from Corollary~\ref{cor:underlying}.
\end{proof}
\begin{defn}
Let $X, Y_1, \ldots, Y_k\in F^E$. We say $X$ is a {\bf linear combination of $\{Y_1, \ldots, Y_k\}$} if there are $\alpha_1,\ldots,\alpha_k\in F$ such that $X\in\bighplus_{j=1}^k \alpha_j Y_j$.
\end{defn}
The following is clear but is important enough to be a lemma:
\begin{lemma}\label{lem:coeffs} If $B\subseteq E$ and $\{R_j: j\in B\}\subseteq F^E$ satisfies $R_j(k)=\delta_{jk}$ for all $k\in B$ then $X\in F^E$ is a linear combination of $\{R_j: j\in B\}$ if and only if, for each $e\in E$,
$$X(e)\in\bighplus_{j\in B} X(j)R_j(e)=\bighplus_{j\in B\cap\underline{X}} X(j)R_j(e).$$
\end{lemma}
We now arrive at the axiomatization of $F$-vectors of $F$-matroids.
\begin{defn} [\bf Tract Vector Axiom, First Form] \label{def:vectorax} Let $F$ be a tract, $E$ a finite set, and $\mathcal{W}\subseteq F^E$. $\mathcal{W}$ is an {\bf $F$-vector set} if $\mathcal{W}$ is exactly the set of all $X\in F^E$ such that, for every support basis $B$ and every nearly reduced row echelon form $\{S_j:j\in B\}$ with respect to $B$, $X$ is a linear combination of $\{S_j:j\in B\}$.
\end{defn}
We make this the definition for the sake of brevity, but it is easy to verify the following equivalent characterization:
\begin{prop}[\bf Tract Vector Axiom, Second Form] Let $F$ be a tract, $E$ a finite set, and $\mathcal{W}\subseteq F^E$. $\mathcal{W}$ is an $F$-vector set if
and only if
\begin{enumerate}
\item $\mathcal{W}$ has a reduced row-echelon form $\{R_j:j\in B\}$ with respect to each support basis $B$, and
\item $\mathcal{W}$ is exactly the set of $X\in F^E$ such that, for every support basis $B$, $X$ is a linear combination of $\{R_j:j\in B\}$.
\end{enumerate}
\end{prop}
\begin{prop} \label{prop:urref} When $\mathcal{W}$ is an $F$-vector set and $B$ is a support basis then the reduced row-echelon form $\{R_j:j\in B\}$ is unique.
\end{prop}
\begin{proof} If $S_j\in\mathcal{W}$ satisfies $S_j(k)=\delta_{jk}$ for all $k\in B$, then by Lemma~\ref{lem:coeffs}, for every $e$,
$S_j(e)\in\bighplus_{k\in B\cap\underline{S_j}} S_j(k)R_k(e)=\{R_j(e)\}$. Thus $S_j=R_j$.
\end{proof}
For any subset $\mathcal S$ of $F^E$ for which reduced-row-echelon forms are unique, we will use $\{R_j:j\in B\}$ to denote the reduced row-echelon form for $\mathcal S$ with respect to a basis $B$.
\begin{example} Corollary~\ref{cor:OM} will show that, in the case $F=\mathbb S$, Definition~\ref{def:vectorax} coincides with the standard definition of signed vectors of an oriented matroid (cf.\ 3.7.5 in~\cite{BLSWZ}). That is, $\mathcal{V}\subseteq \mathcal S^E$ is an $\mathbb S$-vector set if and only if $\mathbf 0\in\mathcal{V}$ and $\mathcal{V}$ satisfies {\em Symmetry, Composition,} and {\em Elimination}.
For general $F$, if $\mathcal{W}\subseteq F^E$ satisfies the tract vector axioms then $\mathbf 0\in\mathcal{W}$ and $\mathcal{W}$ satisfies an analog to the Symmetry Axiom for oriented matroids (Lemma~\ref{lem:symmetry}). In general $\mathcal{W}$ need not satisfy the Elimination or Composition Axioms (see Sections~\ref{sec:elim} and~\ref{sec:flatcompos}).
\end{example}
We now begin to relate Definition~\ref{def:vectorax} to $F$-matroids.
\begin{lemma}\label{lem:suppbasesarebases} Let $\mathcal{W}\subseteq F^E$ satisfy the Tract Vector Axiom, and let $\mathcal{B}$ be the set of support bases of $\mathcal{W}$. Then $\mathcal{B}$ is the set of bases of a matroid. A set $\{e_1, \ldots, e_k\}\subseteq E$ is independent in this matroid if and only if for each $j\in[k]$ there is a $Y_j\in\mathcal{W}$ such that $\underline{Y_j}\cap\{e_1,\ldots, e_k\}=\{e_j\}$.
\end{lemma}
\begin{proof} We first prove that if $B$ is a support basis, $e\in E-B$, and $Y(e)\neq 0$ for some $Y\in\mathcal{W}$ then there is an $f\in B$ such that $(B-\{f\})\cup\{e\}$ is a support basis. Since $Y$ is a linear combination
of $\{R_j: j\in B\}$ and $Y(e)\neq 0$, there is some $f\in B$ such that $R_f(e)\neq 0$. Fix such an $f$.
If $Z$ is an element of $\mathcal{W}$ and $\underline{Z}\cap (B-\{f\})=\emptyset$, then since $Z$ is a linear combination of $\{R_j: j\in B\}$ we have that $Z$ is a multiple of $R_f$. Thus $e\in\underline{Z}$. Thus $(B-\{f\})\cup\{e\}$ is a support basis.
Thus the support bases of $\mathcal{W}$ satisfy the Basis Exchange Axiom for matroids: if $B'$ is a basis containing $e$, then certainly there is a $Y\in\mathcal{W}$ with $Y(e)\neq 0$, and so the above argument gives the desired exchange.
To see the second statement, first note that if $\{e_1,\ldots, e_k\}$ is independent, then it is contained in a basis, which implies the existence of the $Y_j$. To see the converse, we induct on $k$. Since $Y_j(e_j)\neq 0$, by the above argument each $e_j$ is in a basis. If $B$ is a basis containing $\{e_1, \ldots, e_{k-1}\}$, then consider the reduced row-echelon form $\{R_j:j\in B\}$.
Then $Y_k=\bighplus_{j\in B} \alpha_j R_j$, for some values $\alpha_j$, but since $Y_k(e_l)=0$ for all $l<k$, we have $\alpha_{e_l}=0$ for all $l<k$. Thus there is some $j\in B-\{e_1, \ldots, e_{k-1}\}$ such that $R_j(e_k)\neq 0$, and the above argument shows that $(B-\{j\})\cup\{e_k\}$ is a basis.
\end{proof}
The following is obvious, but useful enough to state as a lemma.
\begin{lemma} \label{lem:orbitreps} Let $\mathcal{C}\subset F^E$, and let $S\subseteq \mathcal{C}$ contain at least one representative from each $G$ orbit in $\mathcal{C}$. Then $\mathcal{C}^\perp=S^\perp$.
\end{lemma}
In particular (see Lemma~\ref{lem:fundcirc}), for an $F$-matroid $\mathcal{M}$,
to show that an $X\in F^E$ is in $\mathcal{C}(\mathcal{M})^\perp$, it is enough to show that $X\perp FC(e,B)$ for every $B$ and $e$.
\begin{defn} Let $\mathcal{M}$ be an $F$-matroid and $X\in F^E$. We say $X$ is {\bf consistent} with $\mathcal{C}^*(\mathcal{M})$ if, for each basis $B$, $X$ is a linear combination of $\{R_j:j\in B\}$.
\end{defn}
\begin{lemma} \label{lem:cstarcperp} Let $\mathcal{M}$ be a (strong or weak) $F$-matroid and $X\in F^E$. Then $X\in\mathcal{C}(\mathcal{M})^\perp$ if and only if $X$ is consistent with $\mathcal{C}^*(\mathcal{M})$.
\end{lemma}
\begin{proof} We will show, for every basis $B$ and every $e\in E-B$, that $X\perp FC(e,B)$ if and only if $X(e)\in \bighplus_{j\in B\cap\underline{X}}X(j)R_j(e)$. It then follows from Lemma~\ref{lem:orbitreps} that $X\in\mathcal{C}(\mathcal{M})^\perp$ if and only if, for every basis $B$, $X\in \bighplus_{j\in B\cap\underline{X}}X(j)R_j$, and this is equivalent to $X$ being a linear combination of $\{R_j:j\in B\}$ by Lemma~\ref{lem:coeffs}.
Let $X\in F^E$ and $Y=FC(e,B)\in\mathcal{C}(\mathcal{M})$ for some basis $B$ and some $e$. Consider $\{R_j: j\in B\}\subseteq \mathcal{C}^*(\mathcal{M})$, the reduced row-echelon form with respect to $B$. Since $R_j\perp Y$ and $\underline{R_j}\cap\underline{Y}\subseteq \{j,e\}$, we have that if $j\in\underline{Y}$ then
$$R_j(e)Y(e)^c+ R_j(j)Y(j)^c=R_j(e)Y(e)^c+ Y(j)^c\in N_G$$ and so $Y(j)^c=-R_j(e)Y(e)^c$.
Further, if $j\in B-\underline{Y}$ then since $Y(e)\neq 0$ we have that $R_j(e)=0$.
Thus
\begin{align*}
X\cdot Y&=\sum_{j\in\underline{X}\cap\underline{Y}} X(j)Y(j)^{c}\\[12pt]
&=X(e)Y(e)^{c}+\sum_{j\in B\cap\underline{X}\cap\underline{Y}}-X(j)R_j(e)Y(e)^{c}\\[12pt]
&=(X(e)+\sum_{j\in B\cap\underline{X}\cap\underline{Y}}-X(j)R_j(e))Y(e)^{c}
\end{align*}
and so $X\perp Y$ if and only if $X(e)+\sum_{j\in B\cap\underline{X}\cap\underline{Y}}-X(j)R_j(e)\in N_G$. This is true if and only if $X(e)\in \bighplus_{j\in B\cap\underline{X}\cap\underline{Y}}X(j)R_j(e)$. Since $R_j(e)=0$ for every $j\in B-\underline{Y}$, we have that $X\perp Y$ if and only if $X(e)\in \bighplus_{j\in B\cap\underline{X}}X(j)R_j(e)$.
\end{proof}
\begin{defn} Let $\mathcal{M}$ be a strong $F$-matroid.
The set $\mathcal{V}^*(\mathcal{M})$ of {\bf $F$-covectors} of $\mathcal{M}$ is $\mathcal{C}(\mathcal{M})^\perp$. Equivalently, $\mathcal{V}^*(\mathcal{M})$ is the set of all elements of $F^E$ which are consistent with $\mathcal{C}^*(\mathcal{M})$.
The set $\mathcal{V}(\mathcal{M})$ of {\bf $F$-vectors} of $\mathcal{M}$ is $(\mathcal{C}^*(\mathcal{M}))^\perp$. Equivalently, $\mathcal{V}(\mathcal{M})$ is the set of all elements of $F^E$ which are consistent with $\mathcal{C}(\mathcal{M})$.
\end{defn}
\begin{thm}\label{thm:cryptom} If $\mathcal{M}$ is a strong $F$-matroid then $\mathcal{V}(\mathcal{M})$ and $\mathcal{V}^*(\mathcal{M})$ are $F$-vector sets (in the sense of Definition~\ref{def:vectorax}).
Further, every $F$-vector set is $\mathcal{V}^*(\mathcal{M})$ for some strong $F$-matroid $\mathcal{M}$, and
$$\mathcal{C}^*(\mathcal{M})=\mathrm{Minsupp}(\mathcal{V}^*(\mathcal{M})-\{\mathbf 0\})$$
$$\mathcal{C}(\mathcal{M})=\mathrm{Minsupp}((\mathcal{V}^*(\mathcal{M}))^\perp-\{\mathbf 0\}).$$
\end{thm}
The characterization of strong $F$-matroids by $F$-covectors solidifies the assertion from the introduction that ``the concept of matroids over a tract generalizes the concept of linear subspaces of a vector space". Previously this assertion rested on Grassmann-Pl\"ucker functions (\cite{BB17}), but the discussion at the beginning of this section justifies the following.
\begin{prop} Let $K$ be a field and $V$ a rank $r$ subspace of $K^E$. Then there is a strong $K$-matroid $\mathcal{M}$ such that
\begin{enumerate}
\item $V=\mathcal{V}^*(\mathcal{M})$
\item $V^\perp=\mathcal{V}(\mathcal{M})$, and
\item The projective coordinates of the Pl\"ucker embedding for $V$ constitute a Grassmann-Pl\"ucker function for $\mathcal{M}$.
\end{enumerate}
Further, every $K$-matroid arises in this way.
\end{prop}
Thus when $K$ is a field the Grassmann-Pl\"ucker function $\varphi_\mathcal{M}$ of a $K$-matroid $\mathcal{M}$ can be viewed as an element of $\mathbb P(\bigwedge^r K^{E\choose r})$ of the form $[v_1\wedge\cdots\wedge v_r]$, where $\{v_1, \ldots, v_r\}$ is a basis for the vector space $\mathcal{V}^*(\mathcal{M})$. The algebraic structure of the exterior algebra $\bigwedge^r K^{E\choose r}$ gives a lovely relationship between $\varphi_\mathcal{M}$ and the vector space $\mathcal{V}^*(\mathcal{M})$:
$$X\in\mathcal{V}^*(\mathcal{M})\Leftrightarrow X\wedge \varphi_\mathcal{M}=0.$$
This begs for generalization to other tracts, and such a generalization has been accomplished for tracts arising from an idempotent semifield by Jeffrey and Noah Giansiracusa~\cite{gian}. The tracts $\mathbb K$ and $\mathbb T\triangle$ are examples. Their paper is not written in the language of matroids over tracts, but the characterization of $F$-matroids via Grassmann-Pl\"ucker functions in~\cite{BB17} allows one to characterize Proposition 4.2.1 in \cite{gian} as: for a tract $F$ arising from an idempotent semifield and a rank $r$ $F$-matroid $\mathcal{M}$, there is an associated graded ``tropical Grassmann algebra" $\bigwedge\mathcal{M}$ such that $\varphi_\mathcal{M}\in\mathbb P(\bigwedge^r\mathcal{M})$ and, for all $X\in F^E$,
$$X\in\mathcal{C}(\mathcal{M})^\perp\Leftrightarrow X\wedge\varphi_\mathcal{M}=0.$$
\begin{remark} \label{rem:simpledef} The role of reduced row-echelon forms in Definition~\ref{def:vectorax} may seem like overkill. One might hope for a definition that says, for every maximal linearly independent subset $S$ of $\mathcal{W}$, every element of $\mathcal{W}$ has a unique expression as a linear combination of $S$. Sadly, this is not equivalent to Definition~\ref{def:vectorax}, as will be illustrated by an example in Section~\ref{sec:lincomb}.
\end{remark}
\begin{remark} Baker and Bowler (\cite{BB17}, Section 3) define the vectors and covectors of an $F$-matroid $\mathcal{M}$ to be $\mathcal{C}^*(\mathcal{M})^\perp$ resp.\ $\mathcal{C}(\mathcal{M})^\perp$, without giving an independent axiomatization of vectors and covectors.
\end{remark}
\section{Proof of Theorem~\ref{thm:cryptom} }
We begin by proving that if $\mathcal{W}$ is an $F$-vector set then $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ satisfies the $F$-circuit axioms.
\begin{lemma}\label{lem:symmetry} [Symmetry for $F$-vector sets] If $\mathcal{W}\subseteq F^E$ satisfies the tract vector axioms, $X\in\mathcal{W}$, and $\alpha\in F$ then $\alpha X\in\mathcal{W}$.
\end{lemma}
\begin{proof} If $B$ is a support basis and $X\in\bighplus_{j\in B} \beta_j R_j$ then $\alpha X\in\bighplus_{j\in B} \alpha\beta_j R_j$.
\end{proof}
\begin{lemma} \label{lem:existbasis} Let $\mathcal{W}$ be an $F$-vector set.
1. If $X\in\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ and $b\in\underline{X}$ then there is a support basis $B$ containing $b$ such that $X$ is a multiple of $R_b$.
2. If $\{X_1, \ldots, X_k\}\subseteq\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ is a modular family and, for each $j\in[k]$, $e_j\in \underline{X_j}-\bigcup_{l\neq j}\underline{X_l}$ then there is a basis $B\supseteq\{e_1, \ldots, e_k\}$ such that each $X_j$ is a multiple of $R_{e_j}$.
3. If $\{X_1, \ldots, X_k,X\}\subseteq\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ is a modular family, for each $j\in[k]$, $e_j\in \underline{X_j}-\bigcup_{l\neq j}\underline{X_l}$, and $f\in\underline{X}-\bigcup_{j=1}^k\underline{X_j}$ then there is a basis $B\supseteq\{e_1, \ldots, e_k,f\}$ such that each $X_j$ is a multiple of $R_{e_j}$ and $X$ is a linear combination of $\{X_1, \ldots, X_k, R_f\}$.
\end{lemma}
\begin{proof} 1. There is no $Z\in\mathcal{W}$ such that $X^0\cup\{b\}\subseteq Z^0$, since otherwise $X$ does not have minimal support. Thus $X^0\cup\{b\}$ contains a basis $B$ which has $b$ as an element, and uniqueness of reduced row echelon forms (Proposition~\ref{prop:urref}) implies $R_b=X(b)^{-1}X$.
2. We first claim that $(\bigcap_{j=1}^k X_j^0)\cup\{e_1, \ldots, e_k\}$ has nontrivial intersection with the support of each $Z\in\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$. If $\underline{Z}\cap\bigcap_{j=1}^k X_j^0=\emptyset$ then $\underline{Z}\subseteq \bigcup_{j=1}^k\underline{X_j}$. Let $l$ be minimal such that $\underline{Z}\subseteq \bigcup_{j=1}^l\underline{X_j}$. Then either $l=1$, in which case $\underline{Z}=\underline{X_1}$, and hence $e_1\in Z$, or
$$\bigcup_{j=1}^{l-1}\underline{X_j}\subset \bigcup_{j=1}^{l-1}\underline{X_j}\cup\underline{Z}\subseteq\bigcup_{j=1}^{l}\underline{X_j}.$$ By modularity the second inclusion must in fact be an equality, and so $e_l\in\underline{Z}$.
Thus $(\bigcap_{j=1}^k X_j^0)\cup\{e_1, \ldots, e_k\}$ contains a basis $B$, and certainly $\{e_1, \ldots, e_k\}\subseteq B$. Uniqueness of reduced row echelon forms then implies that each $X_j$ is a multiple of $R_{e_j}$.
3. By the same argument as in (2) we see that $(X^0\cap \bigcap_{j=1}^k X_j^0)\cup\{e_1, \ldots, e_k,f\}$ contains a basis $B$ so that $\{e_1, \ldots, e_k\}\subseteq B$. Then
$$X\in\bighplus_{b\in B} X(b)R_b=\bighplus_{b\in B\cap\{e_1, \ldots, e_k, f\}} X(b)R_b.$$
By modularity $\underline{X}\not\subseteq\bigcup_{j=1}^k\underline{X_j}$, and so $X\not\in\bighplus_{b\in\{e_1,\ldots, e_k\}} X(b)R_b$. Thus $f\in B$ and we have our desired result.
\end{proof}
\begin{lemma}\label{lem:easycircaxs} If $\mathcal{W}$ is an $F$-vector set then $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ satisfies Symmetry and Incomparability.
\end{lemma}
\begin{proof} Symmetry follows from Lemma~\ref{lem:symmetry}.
To see Incomparability: let $X,Y\in\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ such that $\underline{X}\subseteq\underline{Y}$, and let $b\in\underline{X}$. By definition $\underline{X}=\underline{Y}$. Consider a basis $B$ contained in $X^0\cup\{b\}$. This exists by Lemma~\ref{lem:existbasis}, and also by Lemma~\ref{lem:existbasis} $X=X(b) R_b$ and $Y=Y(b)R_b$. Thus $X=X(b)Y(b)^{-1}Y$.
\end{proof}
\begin{lemma} \label{cor:modelim} If $\mathcal{W}$ is an $F$-vector set then $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ satisfies the strong $F$-circuit axioms.
\end{lemma}
\begin{proof} Lemma~\ref{lem:easycircaxs} proved everything except Strong Modular Elimination. Let $\{X_1, \ldots, X_k, X\} \subseteq \mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$ and $\{e_1, \ldots, e_k\}$ be as in the hypothesis of Strong Modular Elimination. Let $f\in \underline{X}\backslash\cup_{j=1}^k\underline{X_j}$. By Lemma~\ref{lem:existbasis}.3 we get a basis $B\supseteq\{e_1, \ldots, e_k, f\}$ such that $$X\in X(f) R_f\boxplus\bighplus_{j=1}^k X(e_j) R_{e_j}.$$ But $X_j=X_j(e_j)R_{e_j}=-X(e_j)R_{e_j}$, and so
\begin{align}
X\in&X(f)R_f\boxplus\bighplus_{j=1}^k -X_j\\
X(f)R_f\in&X\boxplus\bighplus_{j=1}^k X_j
\end{align}
and so $X(f)R_f$ is our desired elimination $Z$.
\end{proof}
We now complete the proof of Theorem~\ref{thm:cryptom}.
\begin{proof} Let $\mathcal{M}$ be an $F$-matroid. Then $\mathcal{V}^*(\mathcal{M})=\mathcal{C}(\mathcal{M})^\perp$, and so $\mathrm{Minsupp}(\mathcal{V}^*(\mathcal{M})-\{\mathbf 0\})=\mathrm{Minsupp}(\mathcal{C}(\mathcal{M})^\perp-\{\mathbf 0\})$, which is $\mathcal{C}^*(\mathcal{M})$ by definition of $\mathcal{C}^*$.
$\mathcal{V}^*(\mathcal{M})$ has a reduced row-echelon form with respect to every basis, by Lemmas~\ref{lem:minrref}.3 and~\ref{lem:circrref}, and by definition $\mathcal{V}^*(\mathcal{M})$ is the set of all $X$ that are linear combinations of $\{R_j: j\in B\}$ for all $B$. Thus $\mathcal{V}^*(\mathcal{M})$ satisfies the second form of the tract Vector Axiom. By duality, $\mathcal{V}(\mathcal{M})=\mathcal{V}^*(\mathcal{M}^*)$ does as well.
Lemma~\ref{cor:modelim} showed that if $\mathcal{W}$ is an $F$-vector set then $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})=\mathcal{C}^*(\mathcal{M})$ for some $\mathcal{M}$. Thus, by Lemma~\ref{lem:supportbasis1}, the set of support bases of $\mathcal{W}$ is exactly the set of bases of $\mathcal{M}$, and so the definition of $\mathcal{W}$ tells us that $\mathcal{W}$ is exactly the set of elements of $F^E$ that are consistent with $\mathcal{C}^*(\mathcal{M})$. Thus by definition $\mathcal{W}=\mathcal{V}^*(\mathcal{M})$.
Finally, to see that $\mathcal{C}(\mathcal{M})=\mathrm{Minsupp}(\mathcal{W}^\perp-\{\mathbf 0\})$: since $\mathcal{W}=\mathcal{C}(\mathcal{M})^\perp$, we have $\mathcal{C}(\mathcal{M})\subseteq\mathcal{W}^\perp-\{\mathbf 0\}$. Thus it suffices to show that if $X\in\mathcal{C}(\mathcal{M})$, $Y\in\mathcal{W}^\perp-\{\mathbf 0\}$, and $\underline{Y}\subseteq\underline{X}$, then $Y=\alpha X$ for some $\alpha\in F-\{0\}$. Let $e\in\underline{Y}$ and $f\in\underline{X}-\{e\}$. Then Lemma~\ref{lem:fundcirc} says that $X$ is a multiple of $FC(f,B)$ for some basis $B$. This $B$ contains $e$, and so the reduced row-echelon form $\{R_j:j\in B\}\subseteq\mathcal{C}^*(\mathcal{M})$ contains an element $R_e$. Note $\underline{R_e}\cap\underline{X}=\{e,f\}$. Thus
$e\in\underline{Y}\cap\underline{R_e}\subseteq\{e,f\}$, and so orthogonality of $Y$ and $R_e$ implies that $Y(e)=Y(e)R_e(e)^c=-Y(f)R_e(f)^c$. Likewise $X(e)=X(e)R_e(e)^c=-X(f)R_e(f)^c$, and so
$$-\frac{1}{R_e(f)^c}=\frac{X(f)}{X(e)}=\frac{Y(f)}{Y(e)}$$
$$Y(f)=\frac{Y(e)}{X(e)}X(f)$$
and so $Y=\frac{Y(e)}{X(e)}X$.
\end{proof}
\section{Basic properties}
\subsection{Duality}\label{sec:duality}
Our results so far show that $\mathcal{V}^*(\mathcal{M})=\mathcal{C}(\mathcal{M})^\perp\supseteq\mathcal{V}(\mathcal{M})^\perp$. If $F$ is a field then the inclusion $\mathcal{V}^*(\mathcal{M})\supseteq\mathcal{V}(\mathcal{M})^\perp$ is an equality: this is just ordinary orthogonality of vector spaces. This is also the case if $F=\mathbb S$ (the oriented matroid case) -- see Example~\ref{ex:OMcomp}.
{\em However, in general $\mathcal{V}(\mathcal{M})^\perp$ might be a proper subset of $\mathcal{V}(\mathcal{M})^*$.} This will be shown by an example with $F=\mathbb P$ in Section~\ref{sec:dualitycounterex}.
\begin{defn} (\cite{BB17}, see also \cite{DWperfect} ) A tract is {\bf perfect} if $\mathcal{V}(\mathcal{M})^\perp=\mathcal{V}^*(\mathcal{M})$ for every $F$-matroid $\mathcal{M}$.
\end{defn}
In~\cite{BB17} Baker and Bowler give various conditions which guarantee that a tract is perfect.
\subsection{Deletion and contraction}\label{sec:delcontr}
Deletion and contraction of $F$-matroids are defined in a way consistent with matroid and oriented matroid definitions:
\begin{thm}[\cite{BB17}, 2.29] \label{bakerdelcontr}
Let $\mathcal{M}$ be an $F$-matroid on $E$, and let $A\subseteq E$. Then $\mathrm{Minsupp}(\{X\backslash A: X\in\mathcal{C}(\mathcal{M})\}-\{\mathbf 0\})$ is the set of circuits of an $F$-matroid $\mathcal{M}/A$, called the {\bf contraction} of $\mathcal{M}$ with respect to $A$, and $\{X\backslash A: X\in\mathcal{C}(\mathcal{M}), \underline{X}\cap A=\emptyset\}$ is the set of circuits of an $F$-matroid $\mathcal{M}\backslash A$, called the {\bf deletion} of $\mathcal{M}$ with respect to $A$. Moreover, $(M/A)^*=M\backslash A$ and $(M\backslash A)^*=M/A$.
\end{thm}
It turns out that the covector interpretations of deletion and contraction are more problematic. In contrast to the situation with matroids and oriented matroids, {\em it is not always true that $\mathcal{V}^*(\mathcal{M}\backslash e)=\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M})\}$, nor that $\mathcal{V}^*(M/e)=\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M}), X(e)=0\}$.} This is illustrated by examples in Sections~\ref{sec:nodel} and~\ref{sec:nocontr}. The best we can say is the following two propositions.
\begin{prop} \label{prop:contraction} $\{X: Xe^0\in\mathcal{V}^*(\mathcal{M})\}\subseteq\mathcal{V}^*(\mathcal{M}/e)$
\end{prop}
\begin{proof} By Theorem~\ref{thm:cryptom}, $\mathcal{V}^*(\mathcal{M})=\mathcal{C}(\mathcal{M})^\perp$. Thus if $Xe^0\in\mathcal{V}^*(\mathcal{M})$ then
$ Xe^0\perp Y$ for all $Y\in\mathcal{C}(\mathcal{M})$, and so $X\perp (Y\backslash e)$ for all such $Y$. Since $\mathcal{C}(\mathcal{M}/e)\subseteq \{Y\backslash e:Y\in\mathcal{C}(\mathcal{M})\}$, we have $X\in\mathcal{V}^*(\mathcal{M}/e)$.
\end{proof}
\begin{prop} $\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M})\}\subseteq \mathcal{V}^*(\mathcal{M}\backslash e)$.
\end{prop}
\begin{proof} By Theorem~\ref{bakerdelcontr}, $\mathcal{C}(\mathcal{M}\backslash e)=\{X\backslash e: X\in\mathcal{C}(\mathcal{M}), X(e)=0\}$. Thus if $Y\in\mathcal{C}(\mathcal{M})^\perp$ then $Y\backslash e\in\mathcal{C}(\mathcal{M}\backslash e)^\perp$.
\end{proof}
It would be interesting to find direct characterizations of $\mathcal{V}^*(\mathcal{M}\backslash e)$ and $\mathcal{V}^*(\mathcal{M}/e)$ in terms of $\mathcal{V}^*(\mathcal{M})$ for arbitrary matroids over tracts, and to characterize the tracts $F$ such that, for all $F$-matroids $\mathcal{M}$ and elements $e$ of $\mathcal{M}$,
$\mathcal{V}^*(\mathcal{M}\backslash e)=\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M})\}$ and $\mathcal{V}^*(M/e)=\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M}), X(e)=0\}$.
\subsection{Morphisms of tracts}\label{sec:morph}
\begin{prop} \label{prop:pushforward} If $f:F\to F'$ is a morphism of tracts and $\mathcal{M}$ is a strong $F$-matroid then $\mathcal{V}(f_*(\mathcal{M}))\supseteq\{\alpha f(X): X\in\mathcal{V}(\mathcal{M}), \alpha\in F'-\{0\}\}$.
\end{prop}
\begin{proof} By Proposition~\ref{prop:morph}, $\mathcal{C}^*(f_*(\mathcal{M}))=\{\alpha f(Y): Y\in\mathcal{C}^*(\mathcal{M}), \alpha\in F'-\{0\}\}.$ Also, if $X,Y\in F^E$ and $X\perp Y$ then $f(X)\perp f(Y)$. Thus
\begin{align*}
(\mathcal{C}^*(f_*(\mathcal{M})))^\perp&=\{ \alpha f(Y): Y\in\mathcal{C}^*(\mathcal{M}), \alpha\in F'-\{0\} \}^\perp\\
&\supseteq \{\alpha f(X):X\in\mathcal{C}^*(\mathcal{M})^\perp, \alpha\in F'-\{0\}\}\\
&=\{\alpha f(X):X\in\mathcal{V}(\mathcal{M}), \alpha\in F'-\{0\}\}\qedhere
\end{align*}
\end{proof}
In general, equality will not hold. For instance, consider the inclusion $\eta:\mathbb R\to\mathbb C$ and an $\mathbb R$-matroid $\mathcal{M}$ with $\mathcal{V}(\mathcal{M})$ a two-dimensional subspace of $\mathbb R^n$ spanned by $\{v,w\}$. Then $\{\alpha\eta(X): X\in\mathcal{V}(\mathcal{M}), \alpha\in \mathbb C-\{0\}\}$ has topological dimension 3 in $\mathbb C^n$, but $\mathcal{V}(\eta_*(\mathcal{M}))$ is the complexification of $\mathcal{V}(\mathcal{M})$, hence is a rank 2 linear subspace of $\mathbb C^n$, hence has topological dimension 4.
A bit more surprisingly, equality need not hold even when the morphism of tracts is surjective:
\begin{example} \label{ex:morph}
Consider the tract morphism $\operatorname{ph}:\mathbb C\to \mathbb P$ and the $\mathbb C$-matroids $\mathcal{M}_1$ and $\mathcal{M}_2$ with
$$\mathcal{V}^*(\mathcal{M}_1)=\mathop{\rm Row}\left(\begin{array}{cccc}
1&1+i&1&0\\
1+i&3i&0&1
\end{array}\right)$$
and
$$\mathcal{V}^*(\mathcal{M}_2)=\mathop{\rm Row}\left(\begin{array}{cccc}
1&1+i&1&0\\
1+i&4i&0&1
\end{array}\right).$$
As shown in~\cite{AD},
$$\{\operatorname{ph}(X): X\in\mathcal{C}^*(\mathcal{M}_1)\}=\{\operatorname{ph}(X): X\in\mathcal{C}^*(\mathcal{M}_2)\}$$
but
$$\{\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_1)\}\not\subseteq\{\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_2)\}.$$
Thus $\operatorname{ph}_*(\mathcal{M}_1)=\operatorname{ph}_*(\mathcal{M}_2)$ and
\begin{align*}
\mathcal{V}^*(\operatorname{ph}_*(\mathcal{M}_1))&\supseteq\{\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_1)\}\cup\{\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_2)\}\mbox{ by Proposition~\ref{prop:pushforward}}\\
&\supsetneq\{\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_2)\}\\
&=\{\alpha\operatorname{ph}(X): X\in\mathcal{V}^*(\mathcal{M}_2),\alpha\in\mathbb P-\{0\}\}.
\end{align*}
\end{example}
\subsection{Loops and coloops}
\begin{defn} A {\bf loop}\ resp.\ {\bf coloop}\ of an $F$-matroid is a loop resp.\ coloop of the underlying matroid.
\end{defn}
\begin{lemma} \label{lem:coloop} 1. $e$ is a loop of $\mathcal{M}$ if and only if $X(e)=0$ for every $X\in\mathcal{V}^*(\mathcal{M})$.
2. $e$ is a coloop of $\mathcal{M}$ if and only if $\mathcal{V}^*(\mathcal{M})=\{Xe^\alpha: X\in\mathcal{V}^*(\mathcal{M}\backslash e), \alpha\in F\}$.
\end{lemma}
\begin{proof} 1. $e$ is a loop of $\mathcal{M}$ if and only if $\{e\}$ is a circuit of the underlying matroid, hence if and only if there is a $Y\in\mathcal{C}(\mathcal{M})$ with $\underline{Y}=\{e\}$. But then, for every $X\in\mathcal{V}^*(\mathcal{M})$, $X\perp Y$ implies that $X(e)=0$.
2. Recall $\mathcal{C}(\mathcal{M}\backslash e)=\{Y\backslash e: Y\in\mathcal{C}(\mathcal{M})\}$. Also, $e$ is a coloop of $\mathcal{M}$ if and only if $e$ is not in any circuit of the underlying matroid. Thus
\begin{align*}
\mbox{$e$ is a coloop of $\mathcal{M}$}&\Leftrightarrow Y(e)=0\ \forall Y\in\mathcal{C}(\mathcal{M})\\
&\Leftrightarrow\mathcal{C}(\mathcal{M})^\perp=\{Xe^\alpha: X\in\mathcal{C}(\mathcal{M}\backslash e)^\perp, \alpha\in F\}\\
&\Leftrightarrow\mathcal{V}^*(\mathcal{M})=\{Xe^\alpha: X\in\mathcal{V}^*(\mathcal{M}\backslash e), \alpha\in F\}.\qedhere
\end{align*}
\end{proof}
\section{Examples}
\subsection{Rank 1 $F$-matroids}
Rank 1 weak $F$-matroids are also strong $F$-matroids.
\begin{prop}\label{prop:rank1} Let $F$ be a tract and $E$ a finite set.
\begin{enumerate}
\item Every nonzero $\varphi\in F^E$ is a Grassmann-Pl\"ucker function for a rank 1 $F$-matroid.
\item If $\mathcal{M}$ is a rank 1 $F$-matroid with Grassmann-Pl\"ucker function $\varphi$ then
\begin{enumerate}
\item $\mathcal{C}^*(\mathcal{M})=\mathcal{V}^*(\mathcal{M})-\{\mathbf 0\}=G \varphi$.
\item $\mathcal{C}(\mathcal{M})=\{\alpha X_{e,f}:\varphi(e)\varphi(f)\neq 0, \alpha\in G\}\cup\{\alpha Y_e:\varphi(e)=0, \alpha\in G\}$, where
\begin{align*}
X_{e,f}(e)^c&=\varphi(f)^{-1} \\
X_{e,f}(f)^c&=-\varphi(e)^{-1} \\
X(g)&=0\qquad\mbox{if $g\not\in\{e,f\}$, }
\end{align*}
and
\begin{align*}
Y_e(e)&=1\\
Y_e(g)&=0\qquad\mbox{ for all $g\neq 0$.}
\end{align*}
\item If $e$ is a nonloop of $\mathcal{M}$, then $\mathcal{V}(\mathcal{M})=\{\varphi\}^\perp$.
\end{enumerate}
\end{enumerate}
\end{prop}
This follows immediately from the definitions of the various objects and the crytomorphisms in~(\cite{BB17}).
\subsection{Matroids}
\begin{prop}\label{prop:matroidvect} Let $\mathcal{M}$ be a $\mathbb K$-matroid. Then
$$\mathcal{V}^*(\mathcal{M})=\{X\in\mathbb K^E: \underline{X}\mbox{ is a union of cocircuits of $\mathcal{M}$}\}.$$
\end{prop}
\begin{proof} If $\underline{X}=\bigcup_{j=1}^k X_j$ with each $X_j\in\mathcal{C}^*(\mathcal{M})$, then consider $Y\in\mathcal{C}(\mathcal{M})$. Note that two elements $A,B$ of $\mathbb K^E$ are orthogonal if and only if $|\underline{A} \cap\underline{B}|\neq 1$. Since $\underline{X}\cap\underline{Y}=\bigcup_{j=1}^k(\underline{X_j}\cap\underline{Y})$ and $|\underline{X_j}\cap\underline{Y}|\neq 1$ for every $j$, we have that $X\perp Y$. Hence $X\in\mathcal{C}(\mathcal{M})^\perp=\mathcal{V}^*(\mathcal{M})$.
To see the converse, we induct on the number of nonloops in $X^0$ for $X\in\mathcal{V}^*(\mathcal{M})$. If $X^0$ contains only loops, then $\underline{X}$ is the union of all cocircuits of $\mathcal{M}$. Otherwise, let $e\in X^0$ be a nonloop. Then
$X\backslash e\in\mathcal{V}^*(\mathcal{M}/e)$
by Proposition~\ref{prop:contraction}, and by our induction hypothesis
$\underline{X\backslash e}=\underline{X_1}\cap\cdots\cap\underline{X_k}$ for some $\mathbb K$-cocircuits $X_1,\ldots, X_k$ of $\mathcal{M}\backslash e$. But then $X_1e^0,\ldots, X_ke^0\in\mathcal{C}^*(\mathcal{M})$ by Theorem 2.29 of~\cite{BB17},
and $\underline{X}$ is the union of the supports of these.
\end{proof}
Another explanation for this equality can be given by the Inflation Property (Section~\ref{sec:absorp}).
\subsection{Oriented matroids}
\begin{cor}\label{cor:OM} $\mathcal{W}\subseteq \mathbb S^E$ is the set of $\mathbb S$-vectors of an $\mathbb S$-matroid with $\mathbb S$-circuit set $\mathcal{C}$ if and only $\mathcal{W}$ is the set of signed vectors of an oriented matroid with signed circuit set $\mathcal{C}$.
\end{cor}
\begin{proof} For $F=\mathbb S$ the $F$-circuit axioms coincide with the usual signed circuit axioms (Corollary 1 in~\cite{Del11} and Theorem 3.6.1 in~\cite{BLSWZ}). Further, $\mathcal{W}$ is the set of signed vectors of an oriented matroid if and only if $\mathcal{W}=(\mathcal{C}^*)^\perp$ for some signed circuit set $\mathcal{C}^*$ (Proposition 3.7.12 in~\cite{BLSWZ}). In either the usual oriented matroid context of the $\mathbb S$-matroid context, the set of signed circuits/$\mathbb S$-circuits corresponding to $\mathcal{W}$ is $\mathrm{Minsupp}(\mathcal{W}-\{\mathbf 0\})$.
\end{proof}
\subsection{Phased matroids}
Phased matroids (i.e., $\mathbb P$-matroids) seem to be where conjectures on this subject go to die. This section will give examples to show that various properties satisfied by $F$-covector sets when $F$ is a field or $F=\mathbb S$ fail to hold when $F=\mathbb P$.
\subsubsection{Topological closure}
Each of the tracts of Example~\ref{ex:tracts} has a topology as subspaces of $\mathbb R$ or $\mathbb C$, and hence the $F$-covector sets of matroids over each of these $F$ have topologies as subspaces of $F^E$.
In contrast to the situation in topological vector spaces, $\mathcal{V}^*(\mathcal{M})=\mathcal{C}(\mathcal{M})^\perp$ does {\em not} imply that $\mathcal{V}^*(\mathcal{M})$ is topologically closed. For instance, consider $\mathbb P$ with topology as a subspace of $\mathbb C$. If $\mathcal{M}$ is the $\mathbb C$-matroid with
$$\mathcal{V}^*(\mathcal{M})=\mathop{\rm Row}\left(\begin{array}{ccc}
1&0&i\\
0&1&1
\end{array}\right)$$
and $\mathcal{M}'=\operatorname{ph}_*(\mathcal{M})$
then it is a straightforward exercise to check that $\mathcal{V}^*(\mathcal{M}')=\{(\alpha, \beta, \gamma)\in\mathbb P^3:\gamma\in (i\cdot\alpha)\boxplus \beta\}$. Note that $i\boxplus 1=\{\exp(i\theta):0<\theta<\pi/2\}$. Thus
$(1,1,i)$ is in the topological closure of $\mathcal{V}^*(\mathcal{M}')$ but is not in $\mathcal{V}^*(\mathcal{M}')$.
\subsubsection{ Linear independence counterexample}\label{sec:lincomb}
Let $\mathcal{M}_1$ and $\mathcal{M}_2$ be as in Example~\ref{ex:morph}, and let $\mathcal{M}=\operatorname{ph}_*(\mathcal{M}_1)=\operatorname{ph}_*(\mathcal{M}_2)$. This $\mathcal{M}$ will shoot down the hope for a simpler characterization of $F$-vector sets expressed in Remark~\ref{rem:simpledef}.
\begin{defn}[cf. \cite{BB17}] $\{X_1, \ldots, X_k\}\subseteq F^E$ is {\bf linearly dependent} if there exist elements $c_1, \ldots, c_k$ of $F$, not all $0$, such that $\mathbf 0\in\bighplus_{j=1}^k c_j X_j$.
A subset of $F^E$ is {\bf linearly independent} if it is not linearly dependent.
\end{defn}
Linear independence behaves badly. Not all maximal linearly independent subsets of a $\mathbb P$-vector set need have the same size. In our example, $(2+i, 1+4i, 1,1)\in\mathcal{V}^*(\mathcal{M}_1)$ and
$(2+i, 1+5i, 1,1)\in\mathcal{V}^*(\mathcal{M}_2)$, and so $X_1:=\operatorname{ph}(2+i, 1+4i, 1,1)$ and $X_2:=\operatorname{ph}(2+i, 1+5i, 1,1)$ are elements of $\mathcal{V}^*(\mathcal{M})$. Each reduced row-echelon form for $\mathcal{V}^*(\mathcal{M})$ is a maximal linearly independent set of size 2, but also $\mathcal{V}^*(\mathcal{M})$ contains $S:=\{X_1, X_2, \operatorname{ph}(1, 1+i, 1,0)\}$, which is linearly independent.
Further, the element $\operatorname{ph}(1+i, 3i,0,1)$ of $\mathcal{V}^*(\mathcal{M})$ can be expressed as a linear combination of $S$ both as
$$\operatorname{ph}(1+i, 3i,0,1)\in (-1)X_1\boxplus 0X_2\boxplus \operatorname{ph}(1, 1+i, 1,0)$$
and as
$$\operatorname{ph}(1+i, 3i,0,1)\in 0X_1\boxplus (-1)X_2\boxplus \operatorname{ph}(1, 1+i, 1,0)$$
thus killing the simpler characterization of $F$-vector sets hoped for in Remark~\ref{rem:simpledef}.
\subsubsection{Phase diagrams}
This section introduces a visualization tool that will be helpful in later counterexamples.
We will depict $X\in (S^1\cup\{0\})^n$ by labelled points on a picture of $S^1$. (If $X(f)=0$ then the label $f$ is not used.) We call this the {\bf phase diagram} for $X$. Thus, for instance, the leftmost circle in Figure~\ref{fig:vperpex} depicts $(1,0,\exp(i\pi/2),\exp(-i\pi/2))$. It is easy to see from the phase diagram, for instance, whether $X\in \mathbb P^n$ or $X\in{\mathbb T}\P^n$ is orthogonal to $(1,1,\ldots, 1)$:
\begin{enumerate}
\item In $\mathbb P^n$, $X\neq\mathbf 0$ is orthogonal to $(1,1,\ldots, 1)$ if and only either the phase diagram has only two points (possibly with multiple labels), which are antipodal to each other, or the points in the phase diagram for $X$ do not all lie in a common closed half-circle.
\item In ${\mathbb T}\P^n$, $X\neq\mathbf 0$ is orthogonal to $(1,1,\ldots, 1)$ if and only if the points in the phase diagram for $X$ do not all lie in a common open half-circle.
\end{enumerate}
\subsubsection{Duality counterexample}\label{sec:dualitycounterex}
Now we give the example promised in Section~\ref{sec:duality}, showing that for some $\mathbb P$-matroids $\mathcal{M}$, $\mathcal{V}^*(\mathcal{M})\neq\mathcal{V}(\mathcal{M})^\perp$.
Let
$$V=\mathop{\rm Row}\left(\begin{array}{cccc}
1&0&1+i&1-i\\
0&1&1-i&1+i
\end{array}\right)
=\mathop{\rm Row}\left(\begin{array}{cccc}
1-i&1+i&1&0\\
1+i&1-i&0&1
\end{array}\right). $$
Let $\mathcal{M}_V$ denote the $\mathbb C$-matroid with $\mathcal{V}^*(\mathcal{M}_V)=V$, and let $\mathcal{M}:=\operatorname{ph}_*(\mathcal{M}_V)$.
We can read off our four $S^1$-orbits of $\mathbb P$-cocircuits of $\mathcal{M}$ from these four row vectors: one representative from each orbit is shown in Figure~\ref{fig:vperpex}.
\begin{figure}
\begin{tikzpicture}
\draw (5,5) circle (1cm);
\draw [fill] (6,5) circle (.1cm); \node at (6.4,5) {$1$};
\draw [fill] (5.7,5.7) circle (.1cm); \node at (6.1,5.8) {$3$};
\draw [fill] (5.7,4.3) circle (.1cm); \node at (6,4.2) {$4$};
\draw (8,5) circle (1cm);
\draw [fill] (9,5) circle (.1cm); \node at (9.4,5) {$2$};
\draw [fill] (8.7,5.7) circle (.1cm); \node at (9.1,5.8) {$4$};
\draw [fill] (8.7,4.3) circle (.1cm); \node at (9,4.2) {$3$};
\draw (11,5) circle (1cm);
\draw [fill] (12,5) circle (.1cm); \node at (12.4,5) {$3$};
\draw [fill] (11.7,5.7) circle (.1cm); \node at (12.1,5.8) {$2$};
\draw [fill] (11.7,4.3) circle (.1cm); \node at (12,4.2) {$1$};
\draw (14,5) circle (1cm);
\draw [fill] (15,5) circle (.1cm); \node at (15.4,5) {$4$};
\draw [fill] (14.7,5.7) circle (.1cm); \node at (15.1,5.8) {$1$};
\draw [fill] (14.7,4.3) circle (.1cm); \node at (15,4.2) {$2$};
\end{tikzpicture}
\caption{$\mathbb P$-cocircuits of $\mathcal{M}$ for Section \ref{sec:dualitycounterex}\label{fig:vperpex}}
\end{figure}
From the figure it's clear that any element of $(S^1)^4$ sufficiently close to $(1,1,-1,-1)$ will be orthogonal to each of these $\mathbb P$-cocircuits, hence will be in $\mathcal{V}(\mathcal{M})$. For a concrete example, $X:=(1,1,\exp({i(\pi+.01)}),\exp({i(\pi+.01)}))\in\mathcal{V}(\mathcal{M})$. But also $(1,1,2,2)\in V=\mathcal{V}^*(\mathcal{M}_V)$, and so by Proposition~\ref{prop:pushforward} $\operatorname{ph}(1,1,2,2)=(1,1,1,1)\in\mathcal{V}^*(\mathcal{M})$. But $X\not\perp (1,1,1,1)$.
\subsubsection{Deletion counterexample}\label{sec:nodel}
Here is the horrible example promised in Section~\ref{sec:delcontr}, showing that $\mathcal{V}^*(\mathcal{M}\backslash e)$ need not be $\{X\backslash e: X\in\mathcal{V}^*(\mathcal{M})\}$.
\begin{example}\label{ex:el:bad} (A slight variation on this example arose in~\cite{AD}.) Let
\[ V:=\mathop{\rm Row}
\left(\begin{array}{ccccccc}
0 & -1 & 0 & 0 & i& 1-i &1\\
-1& 0 &-1 & 0 &-i& 3+i &2\\
0 & -i & 0 & 2i&-i& -2i &-i\\
0 & 0 &-i &i+1& 0& -2 &-1\\
\end{array}\right)\]
and let $\mathcal{M}_V$ be the $\mathbb C$-matroid with $\mathcal{V}^*(\mathcal{M}_V)=V$.
Let $\mathcal{M}=\operatorname{ph}_*(\mathcal{M}_V)$. We will find an element of $\mathcal{V}^*(\mathcal{M}\backslash 7)$ which is not in $\{X\backslash 7: X\in\mathcal{V}^*(\mathcal{M})\}$.
By definition, $\mathcal{C}(\mathcal{M}_V\backslash 7)= \mathrm{Minsupp}(\{X\backslash 7: X\in V^\perp, X(7)=0\}-\{\mathbf 0\})$. We list a representative from each
$(\mathbb C-\{0\})$-orbit below.
$$\begin{array}{rccccccl}
\big(&-1+i,&-i,&1,&\frac{1}{2}-\frac{i}{2},&1,&0&\big)\\ [12pt]
\big(&2-i,&1+i,&1,&\frac{3}{2}+\frac{i}{2},&0,&1&\big)\\ [12pt]
\big(&5-5i,&1+3i,&-2-2i,&0,&-3-i,&1-i&\big)\\ [12pt]
\big(&3-2i,&1+2i,&0,&1+i,&-1,&1&\big)\\ [12pt]
\big(&3+4i,&0,&5,&\frac{7}{2}+\frac{i}{2},&3 -i,&2+i&\big)\\ [12pt]
\big(&0,&7-4i&13,&\frac{25}{2}-\frac{5i}{2},&8+i,&5-i&\big)\end{array}$$
Let $Z_0=(1,1,-1,1,1,-1)\in\mathbb P^6$. It is easy to check, by drawing phase diagrams, that any element of $(S^1)^6$ sufficiently close to $Z_0$ is orthogonal to $\operatorname{ph}(X)$ for each representative $X$ above, hence is in $\mathcal{V}^*(\mathcal{M}\backslash 7)$. In particular, if we let $\alpha=\exp(i\epsilon)$ with $\epsilon>0$ small, then the function $Z_1=(\alpha,1,-1,1,\alpha,-1)$ is in $\mathcal{V}^*(\mathcal{M}\backslash 7)$.
On the other hand, notice that
$Y_1:=(1,1,1,1,0,0,1)$ and $Y_2:=(0,0,1,1,1,1,-1)$ are both in $\mathrm{Minsupp}(V^\perp-\{\mathbf 0\}))= \mathcal{C}(\mathcal{M}_V)$. Since $Y_1=\operatorname{ph}(Y_1)$ and $Y_2=\operatorname{ph}(Y_2)$, we have $Y_1, Y_2\in\mathcal{C}(\mathcal{M})$. Consider a $Z\in\mathbb P^7$ such that $Z\backslash 7=Z_1$. Then
$$Z\cdot Y_1=\alpha+1+-1+ Z(7)$$
and $$Z\cdot Y_2=-1+1+ \alpha+ -Z(7).$$
Thus
$Z$ is orthogonal to $Y_1$ if and only if $Z(7)$ in the lower open half-circle of $S^1$, while $Z$ is orthogonal to $Y_2$ if and only if $Z(7)$ in the upper open half-circle of $S^1$. Thus $Z_1\in\mathcal{V}^*(\mathcal{M}\backslash 7)- \{Z\backslash 7: Z\in\mathcal{V}^*(\mathcal{M})\}$.
\end{example}
\subsubsection{Contraction counterexample}\label{sec:nocontr}
This section gives an example of a rank 3 $\mathbb P$-matroid $\mathcal{M}'$ on elements $[6]$ such that $\mathcal{V}^*(\mathcal{M}'/6)\supsetneq \{X\backslash 6: X\in\mathcal{V}^*(\mathcal{M}'), X(6)=0\}$.
Let
\[ W:=\mathop{\rm Row}
\left(\begin{array}{cccccc}
3&0&0&1&1&-3\\
0&3&0&1&1&3+3i\\
0&0&3&1&1&3-3i
\end{array}\right).\]
Thus
\[W^\perp:=\mathop{\rm Row}
\left(\begin{array}{cccccc}
1&-1+i&-1-i&0&0&1\\
1&1&1&0&-3&0\\
1&1&1&-3&0&0
\end{array}\right).\]
Let $\mathcal{M}_W$ be the $\mathbb C$-matroid with $\mathcal{V}^*(M_W)=W$, and thus $\mathcal{C}(\mathcal{M}_W)=\mathrm{Minsupp}(W^\perp-\{\mathbf 0\})$. From the second matrix above we can find a representative from each $(\mathbb C-\{0\})$ orbit of $\mathcal{C}(\mathcal{M}_W)$, as follows.
$$\begin{array}{rccccccl}
(&1,&-1+i,&-1-i,&0,&0,&1&)\\
(&2+i,&2i,&0,&-3-3i,&0,&1&)\\
(&2-i,&0,&-2i,&-3+3i,&0,&1&)\\
(&0,&2-i,&2+i,&-3,&0,&-1&)\\
(&2+i,&2i,&0,&0,&-3-3i,&1&)\\
(&2-i,&0,&-2i,&0,&-3+3i,&1&)\\
(&0,&2-i,&2+i,&0,&-3,&-1&)\\[12pt]
(&1,&1,&1,&0,&-3,&0&)\\
(&1,&1,&1,&-3,&0,&0&)\\
(&0,&0,&0,&1,&-1,&0&)\\
\end{array}$$
Let $\mathcal{M}'=\operatorname{ph}(\mathcal{M}_W)$. Thus the phases of the elements of the above list give us a list of representatives for the $(\mathbb C-\{0\})$ orbits in $\mathcal{C}(\mathcal{M}')$. Since $\mathcal{C}(\mathcal{M}'/6)=\mathrm{Minsupp}\{X\backslash 6: X\in\mathcal{C}(\mathcal{M})\}$, the phases of the first seven elements of the above list, with their sixth components removed, gives us the elements of $\mathcal{C}(\mathcal{M}'/6)$. By drawing phase diagrams we check that any element of $(S^1)^5$ sufficiently close to $(1,1,1,1,1)$ is orthogonal to each of these 7 elements of $\mathcal{C}(\mathcal{M}'/6)$, hence is in $\mathcal{V}^*(\mathcal{M}'/6)$. For instance, if $\beta\in S^1$ is close to but not equal to 1, then $(1,1,1,1,\beta)\in\mathcal{V}^*(\mathcal{M}/6)$. But certainly $(1,1,1,1,\beta,0)\not\in\mathcal{V}^*(\mathcal{M}')$, since $(1,1,1,1,\beta,0)$ is not orthogonal to the element $(0,0,0,1,-1,0)$ of $\mathcal{C}(\mathcal{M}')$.
\section{Sum properties}\label{sec:elim}
When $K$ is a field and $\mathcal{M}$ is a $K$-matroid, then $\mathcal{V}^*(\mathcal{M})$ is a vector space, hence is closed under (hyper)addition. In contrast, even for an $\mathbb S$-matroid $\mathcal{M}$, if $X,Y\in\mathcal{V}^*(\mathcal{M})$ then $X\boxplus Y$ is not necessarily a subset of $\mathcal{V}^*(\mathcal{M})$. However, matroids over fields, $\mathbb K$, and $\mathbb S$ all satisfy the following weaker versions of additive closure.
\begin{wcproperty} If $X,Y\in \mathcal{V}^*(\mathcal{M})$ then $(X\boxplus Y)\cap \mathcal{V}^*(\mathcal{M})\neq \emptyset$.
\end{wcproperty}
\begin{eproperty}If $X,Y\in \mathcal{V}^*(\mathcal{M})$ and $X(e)=-Y(e)$ then there is a $Z\in (X\boxplus Y)\cap \mathcal{V}^*(\mathcal{M})\neq \emptyset$
such that $Z(e)=0$.
\end{eproperty}
\begin{acproperty} If $X,Y\in \mathcal{V}^*(\mathcal{M})$ and $\alpha\in X(e)\boxplus Y(e)$ then there is a $Z\in (X\boxplus Y)\cap \mathcal{V}^*(\mathcal{M})\neq \emptyset$
such that $Z(e)=\alpha$.
\end{acproperty}
It would be interesting to characterize the tracts whose matroids satisfy each of these properties.
If $F$ is a tract and $x,y\in F$ with $x\boxplus y=\emptyset$, then a rank 1 $F$-matroid with 1 element will not satisfy the Weak Closure Property. Examples of such tracts are given in~\cite{BB17}.
\begin{conj} The Weak Closure Property holds for all matroids over hyperfields.
\end{conj}
Chris Eppolito has recently found an example of a matroid over a hyperfield violating the Elimination Property.
As a small contribution to the study of $X\boxplus Y$, we have the following.
\begin{prop} \label{prop:tropicalclosure} Let $F\in\{\mathbb T\mathbb R, \mathbb T\mathbb C, \mathbb T\triangle\}$. Let $\mathcal{M}$ be an $F$-matroid and $X_1, X_2\in\mathcal{V}^*(\mathcal{M})$. Assume there is at most one value $e$ such that $|X_1(e)|=|X_2(e)|\neq 0$ and $X_1(e)\neq X_2(e)$. Then $X_1\boxplus X_2\subseteq \mathcal{V}^*(\mathcal{M})$.
\end{prop}
\begin{proof} Let $Z\in\mathcal{C}(\mathcal{M})$. If $\underline{Z}\cap\underline{X_1}=\emptyset$ then for every $X'\in X_1\boxplus X_2$ we have $X'\cdot Z=X_2\cdot Z$, and so $X'\perp Z$. Similarly if $\underline{Z}\cap\underline{X_2}=\emptyset$ then every $X'\in X_1\boxplus X_2$ is in $Z^\perp$.
Now assume $\underline{Z}\cap\underline{X_1}\neq\emptyset$ and $\underline{Z}\cap\underline{X_2}\neq\emptyset$. Thus, by Lemma~\ref{lem:tropadd} there are constants $c_1$, $c_2$ and nonempty subsets $S_1$, $S_2$ of $E$ such that for $j\in\{1,2\}$,
\begin{enumerate}
\item for all $e\in E$, $|X_j(e)Z(e)^c|\leq c_j$, with equality if and only if $e\in S_j$, and
\item $\mathbf 0\in\bighplus_{e\in S_j} X_j(e)Z(e)^c$.
\end{enumerate}
If $c_1\neq c_2$, without loss of generality assume $c_2>c_1$. Then for each $f_1\in S_2$ and each $X'\in X_1\boxplus X_2$ we have $X'(f_1)=X_2(f_1)$. Also, for every $f_2\in E- S_2$ we have $|X'(f_2) Z(f_2)^c|<c_2$. Thus
\begin{align*}
X'\cdot Z&= \sum_{e\in E} X'(e)Z(e)^c\\
&= \sum_{e\in S_2} X'(e)Z(e)^c\\
&= \sum_{e\in S_2} X_2(e)Z(e)^c\in N_G.
\end{align*}
If $c_1=c_2$, then for each $X'\in X_1\boxplus X_2$, $|X'(e)Z(e)^c|$ is maximized at each $e\in S_1\cup S_2$, and further, by our hypothesis, $X'(e)=X_1(e)=X_2(e)$ for all but at most one element $e_0$ of $S_1\cap S_2$.
Thus
\begin{align*}
X'\cdot Z&= \sum_{e\in E} X'(e)Z(e)^c\\
&= \sum_{e\in S_1\cup S_2} X'(e)Z(e)^c
\end{align*}
and for $j\in\{1,2\}$
\begin{align*}
X_j(e_0)Z(e_0)^c+\sum_{e\in S_1\cup S_2-\{e_0\}} X_j(e)Z(e)^c\in N_G\\[12pt]
-X_j(e_0)Z(e_0)^c\in\bighplus_{e\in S_1\cup S_2-\{e_0\}} X_j(e)Z(e)^c
=\bighplus_{e\in E-\{e_0\}} X'(e)Z(e)^c
\end{align*}
and so by Lemma~\ref{lem:tropclos}
\begin{align*}
-X'(e_0)Z(e_0)^c&\in\bighplus_{e\in E-\{e_0\}} X'(e)Z(e)^c\\
\sum_{e\in E} X'(e)Z(e)^c&\in N_G\qedhere
\end{align*}
\end{proof}
\begin{lemma}\label{lem:tropclos} $F\in\{\mathbb T\mathbb R, \mathbb T\mathbb C, \mathbb T\triangle\}$. Let $S$ be a subset of $F$ and $a,b\in F$. If $\{a,b\}\subseteq \bighplus_{s\in S} s$ then $a\boxplus b\subseteq \bighplus_{s\in S} s$.
\end{lemma}
This follows immediately from Lemma~\ref{lem:tropadd}.
\section{Flats and Composition}\label{sec:flatcompos}
When we think of oriented matroids as matroids with extra structure, the signed covector set of an oriented matroid can be thought of as extra structure on the lattice of flats of the underlying matroid. That every flat underlies some signed covector follows from the Composition Axiom for oriented matroids.
For general $F$-matroids, this falls apart: for instance, for an $\mathbb F_2$-matroid $\mathcal{M}$, not every flat need arise as the underlying flat of an element of $\mathcal{V}^*(\mathcal{M})$. Section~\ref{sec:flats} will sketch the relationship between flats and $F$-covectors, and Section~\ref{sec:compos} will propose a general notion of a ``composition operation", defined for a particular tract $F$, so that existence of a composition operation implies the same relationship between flats and $F$-covector sets as we have in the case of oriented matroids. We will then explore composition operations on some particular tracts.
\subsection{Flats}\label{sec:flats}
\begin{defn} A {\bf hyperplane} of a matroid is the complement of a cocircuit.
A {\bf flat} of a matroid is an intersection of hyperplanes.
\end{defn}
As always, when we refer to matroid properties of an $F$-matroid we mean properties of the underlying matroid.
\begin{prop}\label{prop:flat1} Let $F$ be a tract and $\mathcal{M}$ an $F$-matroid.
\begin{enumerate}
\item The set of hyperplanes of $\mathcal{M}$ is $\{X^0: X\in\mathcal{C}^*(\mathcal{M})\}$.
\item $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is a subset of the set of flats of $\mathcal{M}$.
\item If $F$ is an infinite field or $F\in\{\mathbb K,\mathbb S\}$ then $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is exactly the set of flats of $\mathcal{M}$.
\item If $F$ is an infinite tract satisfying the Weak Closure Property then $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is exactly the set of flats of $\mathcal{M}$.
\item If $F$ is a finite field then there is an $F$-matroid $\mathcal{M}$ such that $\emptyset$ is a flat and $\emptyset\not\in\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$.
\end{enumerate}
\end{prop}
\begin{proof} (1) follows from Proposition~\ref{prop:morph}, and (2) follows from Proposition~\ref{prop:pushforward} applied to $\kappa: F\to \mathbb K$ and Proposition~\ref{prop:matroidvect}.
If $F=\mathbb K$ then by Proposition~\ref{prop:matroidvect} $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is exactly the set of flats of $\mathcal{M}$. If $F=\mathbb S$ then the Composition Axiom for oriented matroids implies that $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is exactly the set of flats of $\mathcal{M}$.
Since a field is an example of a tract satisfying the Weak Closure Property, we prove the remainder of (3) by proving (4). Let $X, Y\in F^E$. For every $e\in E$ there is at most one $\alpha_e\in F$ such that $X(e)=-\alpha_eY(e)$. Thus there are only finitely many values $\alpha\in F$ such that $X\boxplus\alpha Y$ contains an element $Z$ such that $\underline{Z}\neq\underline{X}\cup\underline{Y}$. Proceeding inductively, we see that for any $X_1, \ldots X_k\in F^E$, there are only finitely many values $\alpha_2, \ldots, \alpha_k\in F$ such that $X_1\boxplus\bighplus_{j=2}^k\alpha_j X_j$ contains an element $Z$ such that $\underline{Z}\neq\bigcup_{j=1}^k\underline{X_j}$. Any flat has the form $\bigcap_{j=1}^k X_j^0$ for cocircuits $X_1, \ldots, X_k$, so by taking an appropriate linear combination of these $X_j$ we get $Z\in\mathcal{V}^*(\mathcal{M})$ with $Z^0=\bigcap_{j=1}^k X_j^0$.
To prove (5), let $F-\{0\}=\{a_1,\ldots, a_n\}$, and let $\mathcal{M}$ be the rank 2 $F$-matroid on $E=[n+2]$ with
$$\mathcal{V}^*(M)=\mathop{\rm Row}
\left(\begin{array}{cccccc}
1 & 0 & 1 & 1 & \cdots& 1\\
0& 1 &a_1 & a_2 &\cdots&a_n
\end{array}\right)$$
Then the hyperplanes of $\mathcal{M}$ are exactly the single-element subsets of $E$, and so the empty set is a flat. However, any element of $\mathcal{V}^*(\mathcal{M})$ (i.e., any linear combination of the two rows) has a 0 coordinate.
\end{proof}
It would be interesting to characterize the tracts for which $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is not always the set of flats of an $F$-matroid $\mathcal{M}$, and to better understand the extent to which these tracts ``behave like finite fields".
This is one motivation for looking at {\em composition operations}, the subject of Section~\ref{sec:compos}. Proposition~\ref{prop:compflat} will prove that if a tract $F$ admits a composition operation then $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is the set of flats of $\mathcal{M}$ for every $F$-matroid $\mathcal{M}$. We'll then find composition operations for all of the tracts introduced in Example~\ref{ex:tracts} except $\mathbb P$.
We do not know a composition operation for $\mathbb P$, and we do not know if every flat of a $\mathbb P$-matroid is the 0 set of a $\mathbb P$-covector.
\subsection{Composition operations}\label{sec:compos}
The usual axioms for signed vectors of oriented matroids includes a {\em Composition Property}, which says that
if $X,Y\in \mathcal{V}(\mathcal{M})$ then $X\circ Y\in \mathbb S^E$ defined by $$X\circ Y(e)=\left\{\begin{array}{ll}
X(e)&\mbox{ if $X(e)\neq 0$}\\
Y(e)&\mbox{ otherwise}
\end{array}\right.$$
is also in $\mathcal{V}(\mathcal{M})$.
\begin{defn} \label{def:composition} A {\bf composition operation} on a tract $F$ is a hyperoperation $\circ_F$ defined on $F^E$ for all finite $E$ such that
\begin{enumerate}
\item For every $X_1$ and $ X_2$ and every $Y\in X_1\circ_F X_2$, $\underline{Y}=\underline{ X_1}\cup\underline{ X_2}$, and
\item If $X_1\in Z^\perp$ and $ X_2\in Z^\perp $ then $X_1\circ_F X_2\subseteq Z^\perp$.
\end{enumerate}
\end{defn}
In other words, a composition operation associates to each $X,Y\in F^E$ a nonempty subset of $(X^\perp\cap Y^\perp)^\perp\cap\{Z: \underline{Z}=\underline{X}\cup\underline{Y}\}$.
When $| X_1\circ_F X_2|=1$ for all $X_1$ and $X_2$, we will often treat $\circ_F$ as a binary operation -- as we already do for the usual composition for oriented matroids.
\begin{remark} \label{rem: compos} This definition is chosen to give us the hypotheses needed to prove the results of this section. All of the examples we will consider, including ordinary composition of oriented matroids, are, in addition, associative. Further, all of them except $\epsilon$-composition and its inspiration, Example~\ref{ex:realcomp}, satisfy the condition that $X\circ_F Y\subseteq X\boxplus Y$ for all $X$ and $Y$.
Both of these additional properties align
with the geometric motivation for composition from oriented matroids, and both come up frequently in oriented matroid proofs. Thus Definition~\ref{def:composition} should not necessarily be taken as, well, definitive.
\end{remark}
\begin{example} \label{ex:realcomp} The geometric motivation behind the definition of composition for oriented matroids is the observation that, if $X,Y\in\mathbb R^E$, then for all sufficiently small $\epsilon>0$, $\mathrm{sign}(X+\epsilon Y)=\mathrm{sign}(X)\circ\mathrm{sign}(Y)$, where the composition on the right-hand side is the usual oriented matroid composition.
Specifically, given $X,Y\in \mathbb R^E$, let $$\epsilon_0=\mathrm{min}\left(\left|\frac{X(e)}{Y(e)}\right|: X(e)Y(e)< 0\right).$$
(If there is no $e$ such that $X(e)Y(e)< 0$ then let $\epsilon_0=\infty$.) Then
$X\circ_\mathbb R Y:=\{X+\epsilon Y: \epsilon<\epsilon_0\}$ is a composition operation, with the additional property that, for every $Z\in X\circ_\mathbb R Y$, $\mathrm{sign} (Z)=\mathrm{sign}(X)\circ \mathrm{sign}(Y)$.
\end{example}
\begin{defn} 1. For $X_1, \ldots, X_k\in F^E$, define $X_1\circ_F\cdots \circ_F X_k$ recursively:
$$X_1\circ_F\cdots \circ_F X_k=\bigcup_{Y\in X_1\circ_F\cdots \circ_F X_{k-1}} Y\circ_F X_k.$$
2. For ${\mathcal S}\subseteq F^E$, let ${\mathcal S}_\circ$ denote the union of all compositions $X_1\circ_F\cdots\circ_F X_k$, where $k\in\mathbb N$ and $X_1, \ldots X_k\in {\mathcal S}$.
\end{defn}
\begin{prop} \label{prop:compflat} Let $F$ be a tract admitting a composition operation $\circ_F$, and let $\mathcal{M}$ be an $F$-matroid.
\begin{enumerate}
\item If $X, Y\in \mathcal{V}(\mathcal{M})$ then $X\circ_F Y\subseteq\mathcal{V}(\mathcal{M})$.
\item $(\mathcal{C}^*(\mathcal{M})_\circ)^\perp= \mathcal{V}(\mathcal{M})$.
\item $\{X^0:X\in\mathcal{V}^*(\mathcal{M})\}$ is exactly the set of flats of the underlying matriod of $\mathcal{M}$.
\end{enumerate}
\end{prop}
\begin{proof} The first two statements follow immediately from $\mathcal{V}(\mathcal{M})=\mathcal{C}^*(\mathcal{M})^\perp$.
By Proposition~\ref{prop:matroidvect} $A\subseteq E$
is a flat if and only if $E-A$ is a union of cocircuits of the underlying matroid. For any set of cocircuits, Corollary~\ref{cor:underlying} promises the existence of $F$-cocircuits of $\mathcal{M}$ with these supports. The composition of these cocircuits is a subset of $\mathcal{V}^*(\mathcal{M})$ consisting of elements with support $E-A$. Thus every flat of $\mathcal{M}$ is $X^0$ for some $X\in\mathcal{V}^*(\mathcal{M})$. The converse is given by Proposition~\ref{prop:flat1}.
\end{proof}
\begin{example} \label{ex:OMcomp} Every signed vector of an oriented matroid is a composition of signed circuits (cf. 3.7.2 in~\cite{BLSWZ}). Thus if $F=\mathbb S$ then $\mathcal{V}(\mathcal{M})=\mathcal{C}(\mathcal{M})_\circ$, and so
$\mathcal{V}(\mathcal{M})^\perp=\mathcal{V}^*(\mathcal{M})$. (Compare to Section~\ref{sec:duality}.)
\end{example}
It would be interesting to characterize the tracts $F$ with some notion of composition (possibly with the additional constraints of Remark~\ref{rem: compos}) such that every $F$-vector of an $F$-matroid $\mathcal{M}$ is contained in a composition of $F$-circuits of $\mathcal{M}$.
by Proposition~\ref{prop:compflat}.2, such tracts are perfect.
\subsubsection{The Inflation Property}\label{sec:absorp}
\begin{defn} A tract is said to have the {\bf Inflation Property} if, whenever $\sum_{j=1}^k a_j\in N_G-\{0\}$ we have $b+\sum_{j=1}^k a_j\in N_G$ for all $b\in G$.
\end{defn}
\begin{prop}\label{prop:inflate} For any tract $F$, the following are equivalent.
\begin{enumerate}
\item If $a\in F-\{0\}$ then $a\boxplus -a=F$.
\item $1\boxplus -1=F$.
\item For every $a\in F-\{0\}$ and $b\in F$, $a\in a\boxplus b$.
\end{enumerate}
Further any hyperfield having these properties satisfies the Inflation Property.
\end{prop}
\begin{proof} (1) is equivalent to (2) because $N_G$ is invariant under multiplication by elements of $G$.
For every $a, b\in F$, $b\in a\boxplus -a$ if and only if $a\in a\boxplus b$. Thus (2) is equivalent to (3).
If $F$ is a hyperfield satisfying these properties, $a_1\in F-\{0\}$, and $\sum_{j=1}^k a_j\in N_G$ then $-a_1\in \bighplus_{j=2}^k a_j$, and so $\bighplus_{j=1}^k a_j=(a_1\boxplus -a_1)\cup(a_1\boxplus(\bighplus_{j=2}^k a_j-\{-a_1\})=F$.
\end{proof}
\begin{example} $\mathbb K$, $\mathbb S$, and ${\mathbb T}\P$ all satisfy the Inflation Property.
.
\end{example}
\begin{prop} If $F$ satisfies the Inflation Property then the composition $\circ$ defined by
$$X\circ Y(e)=\left\{\begin{array}{ll}
X(e)&\mbox{ if $X(e)\neq 0$}\\
Y(e)&\mbox{ otherwise}
\end{array}\right.$$
is a composition operation.
\end{prop}
\begin{proof} Let $X,Y,Z\in F^E$ such that $X\perp Z$ and $Y\perp Z$.
If $X\cdot Z=\{0\}$, then $(X\circ Y)\cdot Z=Y\cdot Z$, thus $X\circ Y\perp Z$.
Otherwise, $$(X\circ Y)\cdot Z=\sum_{j\in\underline{X}} X(j)Z(j)^c + \sum_{j\in X^0}Y(j)Z(j)^c$$ and since $\sum_{j\in\underline{X}} X(j)Z(j)^c\in N_G$ we have $(X\circ Y)\cdot Z\in N_G$.
\end{proof}
\subsubsection{$\mathbb T$-composition}
Throughout the remaining sections $+$ and $\sum$ denote ordinary addition in $\mathbb R$.
\begin{prop} \label{prop:t-composition}
Let $F\in\{\triangle, {\mathbb T}\P, \mathbb T\mathbb R, \mathbb T\mathbb C, \mathbb T\triangle\}$, and let $\mathcal{M}$ be an $F$-matroid. Then the operation
$$X\circ_{\mathop{\rm max}} Y(e)=\left\{\begin{array}{ll}
X(e)&\mbox{ if $|X(e)|\geq |Y(e)|$}\\
Y(e)&\mbox{ otherwise.}
\end{array}\right.$$
is a composition operation on $F$.
\end{prop}
\begin{lemma} \label{lem:trianglesum} Let $\{s_1, \ldots, s_k\}$ be a finite subset of $\triangle$ with $s_1\leq\cdots\leq s_k$. Then $0\in\bighplus_{j=1}^k s_j$ if and only if $s_k\leq\sum_{j=1}^{k-1} s_j$.
\end{lemma}
\begin{proof} $0\in(\bighplus_{j=1}^{k-1} s_j)\boxplus s_k$ if and only if $s_k\in\bighplus_{j=1}^{k-1} s_j$. Induction on $k$ shows that the smallest element of $\bighplus_{j=1}^{k-1} s_j$ is $\max(0, s_{k-1}-s_{k-1}-\cdots-s_1)$, while the largest element of $\bighplus_{j=1}^{k-1} s_j$ is $\sum_{j=1}^{k-1} s_j$. Thus
$$s_k\in\bighplus_{j=1}^{k-1} s_j\Leftrightarrow \max(0, s_{k-1}-s_{k-1}-\cdots-s_1)\leq s_k\leq \sum_{j=1}^{k-1} s_j$$
But the first inequality above is vacuous: for any $s_k\in \triangle$ we have $0\leq s_k$, and by hypothesis $s_k\geq s_{k-1}\geq s_{k-1}-s_{k-1}-\cdots-s_1$.
\end{proof}
\begin{lemma}\label{lem:tropadd} Let $F\in\{{\mathbb T}\P, \mathbb T\mathbb R, \mathbb T\mathbb C, \mathbb T\triangle\}$, and let $(s_j:j\in J)$ be a sequence in $F$. Then $0\in\bighplus_{j\in J} s_j$ if and only if there is a $J'\subseteq J$ so that
\begin{enumerate}
\item $|s_j|\leq|s_{j'}|$ for all $j\in J$ and $j'\in J'$ and
\item $0\in\bighplus_{j\in J'} s_j$.
\end{enumerate}
Also,
\begin{enumerate}
\item if $F={\mathbb T}\P$ or $F=\mathbb T\mathbb C$ then $0\in\bighplus_{j\in J'} s_j$ if and only if $\{s_j: j\in J'\}$ is not contained in an open half circle,
\item if $F=\mathbb T\mathbb R$ then $0\in\bighplus_{j\in J'} s_j$ if and only if there exists $j_1, j_2\in J'$ such that $s_{j_1}=-s_{j_2}$, and
\item if $F=\mathbb T\triangle$
then $0\in\bighplus_{j\in J'} s_j$ if and only if either $s_j=0$ for all $j$ or there exists $j_1\neq j_2\in J'$ such that $s_{j_1}=s_{j_2}$.
\end{enumerate}
\end{lemma}
The proof is easy.
Now we can prove Proposition~\ref{prop:t-composition}:
\begin{proof}
We first give the argument for $F=\triangle$. Let $X,Y,Z\in \triangle^E$ such that $X\perp Z$ and $Y\perp Z$. Consider an $f$ with $|(X\circ_{\mathop{\rm max}} Y)(f)Z(f)^c|$ as large as possible. Then
\begin{align*}
\sum_{e\neq f} (X\circ_{\mathop{\rm max}} Y)(e)Z(e)^c&\geq \sum_{e\neq f} X(e)Z(e)^c\\[12pt]
&\geq X(f)Z(f)^c\qquad\mbox{ since $0\in X\cdot Z$, by Lemma~\ref{lem:trianglesum}}
\end{align*}
and similarly $\sum_{e\neq f} (X\circ_{\mathop{\rm max}} Y)(e)Z(e)^c\geq Y(f)Z(f)^c$. Since $(X\circ_{\mathop{\rm max}} Y)(f)\in\{X(f), Y(f)\}$, the result follows.
For $F\in \{{\mathbb T}\P, \mathbb T\mathbb R, \mathbb T\mathbb C, \mathbb T\triangle\}$,
let $X,Y,Z\in F^E$ such that $X\perp Z$ and $Y\perp Z$. Consider an $f$ with $|(X\circ_{\mathop{\rm max}} Y)(f)Z(f)^c|$ as large as possible. Thus for every $e$
\begin{align*}
|(X\circ_{\mathop{\rm max}} Y)(f)Z(f)^c|&\geq|(X\circ_{\mathop{\rm max}} Y)(e)Z(e)^c|\\
&=|(X\circ_{\mathop{\rm max}} Y)(e)||Z(e)^c|\\
&\geq|X(e)Z(e)^c|.
\end{align*}
Similarly $|(X\circ_{\mathop{\rm max}} Y)(f)Z(f)^c|\geq |Y(e)Z(e)^c|$.
If $(X\circ Y)(f)=X(f)$, then orthogonality of $X$ and $Z$ implies the existence of $E_0\subseteq E$ with
$|X(e)Z(e)^c|=|X(f)Z(f)^c|$ for all $e\in E_0$, and $0\in\bighplus_{e\in E_0} X(e)Z(e)^c$.
Since for $e\in E_0$
\begin{align*}
|X(e)Z(e)^c|&=|(X\circ_T Y)(f)Z(f)^c|\\
&\geq|Y(e)Z(e)^c|
\end{align*}
we have $|X(e)|\geq |Y(e)|$, and so $(X\circ_T Y)(e)=X(e)$. Thus for all $e'\in E$ and $e\in E_0$, $$|(X\circ_T Y)(e')Z(e')^c|\leq |(X\circ_T Y)(e)Z(e)^c|$$ and $$0\in\bighplus_{e\in E_0} (X\circ_T Y)(e)Z(e)^c$$ so $0\in (X\circ_T Y)\cdot Z$.
A similar argument covers the case when $(X\circ Y)(f)=Y(f)$.
\end{proof}
\subsubsection{ $\epsilon$-composition}
This section gives a $\mathbb T$-analog to the operation of Example~\ref{ex:realcomp}.
\begin{defn} Let $F\in \{\triangle,\mathbb T\mathbb R,\mathbb T\mathbb C,\mathbb T\triangle\}$, and let $X,Y\in F^E$.
For each real number $\epsilon>0$, define $X\circ_\epsilon Y\in F^E$ to be the set of $Z$ such that, for some $\eta$ with $0<\eta<\epsilon$:
$$Z(e)=\left\{\begin{array}{ll}
X(e)&\mbox{ if $X(e)\neq 0$}\\
\eta Y(e)&\mbox{ if $X(e)= 0$.}\end{array}\right.$$
\end{defn}
\begin{prop} Let $F\in \{\triangle,\mathbb T\mathbb R,\mathbb T\mathbb C,\mathbb T\triangle\}$, let $\mathcal{M}$ be an $F$-matroid, and let $X,Y\in\mathcal{V}^*(\mathcal{M})$. Then there is an $\epsilon>0$ such that $X\circ _{\epsilon}Y\subseteq\mathcal{V}^*(\mathcal{M})$.
\end{prop}
\begin{proof} Let $\{Z_1, \ldots, Z_k\}$ be a choice of one $F$-circuit from each $G$-orbit of $\mathcal{C}(\mathcal{M})$. For each $j$ such that $\underline{X}\cap\underline{Z_j}\neq\emptyset$, let $ d_j=\max(|X(e)||Z_j(e)|:e\in E)$, and let $\epsilon_j=\mathrm{min}(|\frac{ d_j}{Z_j(f)}|:f\in\underline{Z}-\underline{X})$. Then let $\epsilon$ be the minimum over all $\epsilon_j$.
For all $j$ such that $\underline{X}\cap\underline{Z_j}=\emptyset$ and for all $X'\in X\circ_\epsilon Y$ we have $X'\cdot Z_j=\eta Y\cdot Z_j$, for some $\eta$, and so $X'\perp Z_j$.
If $\underline{X}\cap\underline{Z_j}\neq\emptyset$, we have slightly different arguments for $F=\triangle$ and for $F\in\{\mathbb T\mathbb R,\mathbb T\mathbb C,\mathbb T\triangle\}$.
If $F=\triangle$ and $X'\in X\circ_\epsilon Y$, consider an $f$ with $|X'(f)Z_j(f)^c|$ as large as possible.
Then
\begin{align*}
\sum_{e\neq f} X'(e)Z_j(e)^c&\geq \sum_{e\neq f} X(e)Z_j(e)^c\\[12pt]
&\geq X(f)Z_j(f)^c\qquad\mbox{ since $0\in X\cdot Z_j$, by Lemma~\ref{lem:trianglesum}}
\end{align*}
Also
\begin{align*}
\sum_{e\neq f} X'(e)Z_j(e)^c&\geq \sum_{e\neq f} \eta Y(e)Z_j(e)^c\\[12pt]
&\geq\eta Y(f)Z_j(f)^c\qquad\mbox{ since $0\in Y\cdot Z_j$.}
\end{align*}
Since $X'(f)\in\{X(f), \eta Y(f)\}$, the result follows.
If $F\in \{\mathbb T\mathbb R,\mathbb T\mathbb C,\mathbb T\triangle\}$, then $X\cdot Z_j=\bighplus_{e\in\underline{X}} X(e)Z(e)^c=\{\alpha\in F: |\alpha|\leq d_j\}$. Denote this set $I_{ d_j}$. Thus
for every $X'\in X\circ_\eta Y$
we have
\begin{align*}
X'\cdot Z_j&=\bighplus_{e\in\underline{X}} X(e)Z_j(e)^c \boxplus\bighplus_{f\in\underline{Z_j}-\underline{X}}\eta Y(f)Z_j(f)^c\\
&=I_{ d_j}\boxplus\bighplus_{f\in\underline{Z_j}-\underline{X}}\eta Y(f)Z_j(f)^c\\
&=I_{ d_j}.
\end{align*}
Thus $X'\in\mathcal{C}(\mathcal{M})^\perp$.
\end{proof}
\noindent{\bf Acknowledgements.} Many thanks to Matt Baker, Emanuele Delucchi, and Thomas Zaslavsky for helpful discussions. Chris Eppolito contributed the idea of the Inflation Property. Thanks also to Ting Su for comments and corrections.
\bibliographystyle{amsalpha}
| 9151536c5c40136418682612f344a07826277cbf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Background}
In the theory of flows on homogeneous spaces, Ratner's theorems on measure rigidity, topological rigidity, and orbit equidistribution \cite{ratner_raghunathans_1991, ratner_raghunathans_1991_1} play a major role. Their applications go far beyond the realm of dynamical systems and include results in number theory and mathematical physics \cite{ElkiesMcM04, shah_equidistribution_2009, marklof_frobenius_2010, marklof_strombergsson_free_path_length_2010}; thorough expositions and comprehensive references
may be found in \cite{morris_ratner_2005}.
In the last decade there has been an increased interest in obtaining \emph{effective} versions of
Ratner's results, such as giving a rate of convergence of measures in the measure rigidity theorem. There are
two general situations where effective results may be
proved: when the group generating the flow is horospherical, or when it is ``large'' in an
appropriate sense (cf.~\cite[Sec.~1.5.2]{einsiedler_effective_2009}).
Recently, rates of convergence were obtained for several settings where the corresponding group is neither horospherical nor large. Green and Tao \cite{green_tao_quantative_2012} proved effective
equidistribution of polynomial orbits on nilmanifolds, while Einsiedler, Margulis, and Venkatesh \cite{einsiedler_effective_2009}
proved effective equidistribution for closed orbits of semisimple groups on
general homogeneous
spaces. Str\"ombergsson \cite{strombergsson_effective_2013} and Browning and the author \cite{browning_vinogradov_2016} gave rates for the convergence of measures on special horocycle lifts; the present paper further explores this direction by giving a rate of convergence for measures on general horocycle lifts.
\subsection{Results}
For $x\in \R$ and $y>0$, let
\begin{align}
n(x)&=\begin{pmatrix}1 & x\\0 & 1\end{pmatrix},& a(y)&=\begin{pmatrix}\sqrt y& 0\\ 0 & 1/\sqrt y\end{pmatrix}.
\end{align}
It is a fundamental result in homogeneous dynamics that long closed horocycles $\{n(x)\colon \strut x \in [-\frac12,\frac12)\}$ on $X=\sltz\quot \sltr$ equidistribute under the geodesic flow $a(y)$ as $y\to 0$. That is, for every bounded continuous $f\colon X\to \R$,
\beq
\lim_{y\to 0}\int_{-\frac 12}^{\frac 12} f(n(x)a(y))\, dx = \int_X f(g) \, d\mu_X(g),
\eeq
where $\mu_X$ is the Haar probability measure on $X$. This can be proved using thickening followed by applying the mixing property of $a(y)$ \cite{margulis_some_aspects_2004}, which is a general approach when the integral is taken over all unstable directions of a flow. The rate of convergence was given in \cite{zagier_eisenstein_1979, sarnak_asymptotic_1981} and is related to the zero-free region for the Riemann zeta function. It is proved that for $f\in C^{\infty}_0 (X)$,
\beq
\int_{-\frac 12}^{\frac 12} f(n(x)a(y))\, dx = \int_X f(g) \, d\mu_X(g) + o_f(y^{1/2}),
\eeq
where the error term depends on the error term in the Prime Number Theorem.
In the present paper we establish a similar result for certain horocycle lifts.
Let $G=\ASL(2,\R)=\sltr\ltimes \R^2$, and set $\Gamma=\ASL(2,\Z)=\SL(2,\Z)\ltimes \Z^2$, which is a lattice in $G$. We view elements of $G$ as ordered pairs $(M,\bm x)$ with $M\in \sltr$ and $\bm x\in \R^2$, and multiply them following the rule
\beq
(M,\bm x)(M',\bm x')=(MM',\bm x M' + \bm x')
\eeq
thinking of $\bm x, \bm x'$ as row vectors in $\R^2$. Writing $Y=\Gamma\quot G$, we equip this homogeneous space with the Haar probability measure that we denote $\mu_Y$. When no confusion can arise we shorten $(M,\bm 0)$ to $M$.
It is important to note the relationship between $X$ and $Y$, the latter being a bundle over the former with two-dimensional torus fiber. The space $X$ parametrizes unimodular lattices in $\R^2$, while $Y$ is the space of lattice translates in $\R^2$. Thus, each point in $X$ corresponds to a lattice $\Lambda\subset \R^2$, and a choice of $\bm x\in \Lambda\quot \R^2$ determines the translated lattice $\Lambda + \bm x\subset \R^2$, which corresponds to a point of $Y$.
For a continuous function $\bm \xi =(\xi_1,\xi_2)\colon \R \to \R^2$ define
\beq
\tilde n(x) = (\mathbbm{1},\bm \xi(x)) n(x);
\eeq
we call this is a \emph{lift} of a horocycle from $X$ to $Y$. Let $\rho\colon \R\to \R$ be nonnegative, continuously differentiable, supported on a compact interval (without loss of generality, $\supp\rho\subset (-1,0]$), and of integral $1$. It is natural to ask whether the lifted measures $\nu_y$ defined by
\beq
\int_{\R} f(\tilde n(x) a(y))\, \rho(x) dx = \int_Y f(g) \, d\nu_y(g) \label{eq:measures}
\eeq
have a weak-* limit as $y\to0$.
Using Ratner's Theorem \cite{ratner_raghunathans_1991}, Elkies and McMullen \cite{ElkiesMcM04} established a condition on $\vecxi$ under which $\nu_y$ converges to $\mu_Y$. Let $\Xi(x)=x\xi_1(x)+\xi_2(x)$. A horocycle lift is called \emph{rationally linear} if for some $\alpha,\beta\in\Q$,
\beq
\leb\{x\colon \Xi(x)=\alpha x+\beta\}>0.\label{eq:rational}
\eeq
\begin{theorem}[{\cite[Th.~2.2]{ElkiesMcM04}}] Suppose that a horocycle lift $\tilde n$ is \emph{not} rationally linear, that $\Xi$ is Lipschitz, and that $\xi_1$ is continuous. Then, $\nu_y\to \mu_Y$ in the weak-* topology as $y\to 0$.
\end{theorem}
In the present paper, we make convergence in this theorem effective, which requires an effective version of rational nonlinearity. We say $\Xi$ is \emph{$D$-nice} for some $D\ge 2$ if $\Xi$ is twice continuously differentiable and there exist $x_0 \in\R$ and $C_1, C_2>0$ such that
\begin{align}
C_1|x-x_0|^{D-2} \le |\Xi''(x)| \le C_2|x - x_0|^{D-2}
\end{align}
for every $x$ in the support of $\rho$. For such a lift, set $C = \max\left\{C_1^{-1/2}, C_2^{1/2}\right\}.$ We say that $\tilde n$ is \emph{$D$-nice} if the corresponding $\Xi$ is $D$-nice.
\begin{theorem}\label{th:main}
Fix a density function $\rho$ as before, and let $\tilde n$ be $D$-nice.
Assume that $f$, $\rho$, and $\tilde n$ are such that all norms in \eqref{eq:constant} are finite. Then for every $\eps>0$ there exists a constant $C(\eps, f,\rho,\tilde n)$ such that
\beq
\left\lvert \nu_y(f) - \mu_Y(f) \right\rvert \le
C(\eps,f,\rho,\tilde n)\, y^{\min\left\{\frac1{16},\frac1{2D}\right\}-\eps}
\eeq
for all $y\in(0,1).$ Moreover, we can take
\begin{align}\label{eq:constant}
C(\eps,f,\rho,\tilde n) = A(\eps,\eta) C \|f\|_{C^8_{\text b}}
(C + \|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty})
\|\rho\|_{W^{1,1}}^{1-\eta} \|\rho\|_{W^{2,1}}^\eta
\end{align}
for any $\eta\in(0,1)$, and the function $A$ is universal.
\end{theorem}
We observe that the $\eps$-loss in the error term can be replaced by a logarithmic loss with a slightly more tedious computation as in \cite{browning_vinogradov_2016}. The overall error would then be a constant times $y^{\min\left\{\frac1{16},\frac1{2D}\right\}} \log^{\kappa}(2+1/y)$ for some $\kappa>0$. The norms used to define $C(\eps, f,\rho, \tilde n)$ are rigorously defined in \eqref{eq:norm}, \eqref{eq:sobolev_norm}.
\subsection{Discussion}
This paper stands in the series of works that prove effective equidistribution results in the setting of a sequence on measures supported on the unstable manifold of a diagonal flow. The first and by now classical is \cite{sarnak_asymptotic_1981}; it gives the optimal rate of equidistribution of long closed horocycles on quotients of the $\SL(2,\R)$ by using Eisenstein series to relate this question to the zero-free region of the Riemann zeta function. The case of non-uniform measure on the horocycle was treated by Str\"ombergsson \cite{strombergsson_uniform_2004}; this work also proves effective equidistribution for horocycle pieces of optimal intermediate length (length of piece can be nearly as short as the square root of the length of the horocycle).
Horocycle lifts to $Y$ were first studied in \cite{ElkiesMcM04} where an ineffective equidistribution theorem for general lifts is proved. Str\"ombergsson \cite{strombergsson_effective_2013} used number-theoretic techniques similar to those employed in the present paper to give a rate for equidistribution of linear irrational lifts (in our notation, these correspond to $\vecxi$ being a constant that is not in $\Q^2$ and our results do not apply as $\Xi'' = 0$). The rate in this setup depends on the Diophantine properties of $\vecxi$. The method of Str\"ombergsson was further developed in \cite{browning_vinogradov_2016} to treat the case of rational quadratic lifts, with the case $\vecxi(x) = (x/2, - x^2/4)$ being the most interesting. The treatment of this particular lift yields a rate for the convergence of the gap distribution of the sequence $\sqrt n \imod 1$.
The present paper completes the effectivization of equidistribution theorems for lifts. The powers of $y$ (up to $y^{\eps}$) that appear in error terms in the aforementioned theorems are
\begin{align}
y^{1/2}, & \text{ Sarnak \cite{sarnak_asymptotic_1981},} \label{eq:sarnak_rate}\\
y^{1/4}, & \text{ Str\"ombergsson \cite{strombergsson_effective_2013}, assuming best Diophantine condition,}\\
y^{1/4}, & \text{ Browning, V. \cite{browning_vinogradov_2016},}\\
y^{1/16}, & \text{ present work, assuming best lift}.
\end{align}
We also remark that $y^{3/4}$ in \eqref{eq:sarnak_rate} would be equivalent to the Riemann hypothesis. The novelty of the present paper is that it does not make use of quadratic niceties of \cite{browning_vinogradov_2016} or linear simplicity of \cite{strombergsson_effective_2013}, allowing for the treatment of general lifts.
The key result is Proposition \ref{prop:cancellations}, which establishes cancellations in a certain Kloosterman-like exponential sum \eqref{eq:sum_l} for all values of the indices involved.
In addition to the theorems mentioned above, we must also mention the recent result of Ubis \cite{ubis_effective_2016}, who used the ``Fourier method'' on $\R^d$ to prove effective equidistribution of certain manifolds on $(\Gamma\quot \SL(2,\R))^d$. Fix $\tilde a(y) = (a(y),\dots, a(y)) \in \sltr^d$ and consider its $d$-dimensional unstable manifold. Then, given a submanifold that is ``totally curved'' and has positive codimension, Ubis gives a rate of equidistribution of this submanifold under the action of $\tilde a(y)$. Although the method of this paper is different from that of the present work, the setup is quite similar, which gives hope that other equidistribution statements of this flavor (for example, results of Shah on equidistribution of curves \cite{shah_limiting_2009, shah_asymptotic_2009, shah_equidistribution_2009}) will be effectivized in the near future.
Related results on effective equidistribution for $\sltr$ include papers of Tanis and Vishe \cite{tanis_uniform_2015} and Flaminio, Forni, and Tanis \cite{flaminio_effective_2015} on period integrals, both building on the work on the seminal paper of Flaminio and Forni \cite{flaminio_invariant_2003} on invariant distributions for the horocycle flow. Effective equidistribution of ``relatively large'' orbits is proven by Einsiedler, Margulis, and Venkatesh \cite{einsiedler_effective_2009}.
Another direction of refining convergence in Ratner's theorem is extending weak-* convergence to unbounded test functions, known as the problem of \emph{convergence of moments}.
This question was answered affirmatively in certain situations, relating to theta functions and their application to values of inhomogeneous quadratic forms \cite{marklof_pair_correlation_2003};
the pair correlation function of the sequence $\sqrt n$ modulo $1$ \cite{EMV_two_point};
directions of Euclidean lattice points \cite{EMV_directions_2013};
directions of hyperbolic lattice points \cite{directions_hyperbolic}.
The question of convergence of moments is open for the main result of this paper, Theorem \ref{th:main}, as is the question of the rate of equidistribution of the unipotent flow $\{\tilde n(x) : x\in\R\}$ with $\vecxi(x) = (x/2, -x^2/4).$ We hope to return to these questions in future work.
\subsection{Plan of paper}
Section \ref{sec:number} contains an application of number-theoretic techniques to control a special exponential sum. In Section \ref{sec:fourier}, we single out the main term from the integral in the statement of Theorem \ref{th:main}; we then bound the error term in Section \ref{sec:error}.
\subsection{Notation}
Given functions $f,g\colon S\rightarrow \R$, with
$g$ positive, we will write $f\ll g$ if
there exists a constant $c$ such that
$|f(s)|\leq cg(s)$ for all $s\in S$.
\subsection{Acknowledgements}
The author would like to thank Trevor Wooley for useful discussions and Tim Browning and Jens Marklof for comments on earlier versions of this paper.
The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP/2007--2013) / ERC Grant Agreement n.\ 291147.
\section{Special exponential sums\label{sec:number}}
In this section we make a detailed examination of the
exponential sum
\begin{align}
\label{eq:sum_l}
S_c(k,l,n) = S = \sum_{\substack{(c,d) = 1\\ 0\le d < c}} e\left( \frac{l\bar d - kd}c - nc \Xi\left( - \frac dc\right)\right).
\end{align}
We distinguish two cases that require different treatment, according as $l=0$ or $l\ne 0$.
Consider first the case $l=0$; the cancellations in the sum $S$ must come from analytic properties of $\Xi$. Writing
\begin{align}
S = \sum_{h\mid c} \mu(h) \sum_{d=1}^{c/h} e\left(-nc \Xi\left(-\frac{dh}c\right) - \frac{kdh}c\right) = \sum_{h\mid c} \mu(h) S\left(\frac ch\right),
\end{align}
we massage the inner sum over $d$.
We have
\begin{align}
S(x) = \sum_{d=1}^x e\left( -hnx\Xi\left(-\frac dx\right) - \frac{kd}x\right) = \sum_{d=1}^x e(w(d)).
\end{align}
Since $w''(d) = - \frac{hn}{x} \Xi''\left(-\frac dx\right)$, our assumption on $\Xi$ implies that
$|w''(d)| \asymp_{C_1,C_2} \frac{hn}x \lvert-\frac dx - z_0\rvert ^{D-2}.$ Let $\delta>0$. Using the van der Corput estimate (cf.\ \cite[p.~8]{graham_kolesnik_1991}) when $\lvert-\frac dx - z_0\rvert > \delta$ and the trivial estimate otherwise, we get the bound
\begin{align}
|S(x)| \ll \delta x + C_2^{1/2} h^{1/2} n^{1/2} x^{1/2} \delta^{\frac{D-2}2} + \frac{x^{1/2}}{C_1^{1/2} h^{1/2}n^{1/2}} \delta^{-\frac{D-2}2}.
\end{align}
The optimal choice for $\delta$ is $x^{-1/D}$, giving the bound $C h^{1/2}n^{1/2} x^{1-1/D}$ for $S(x)$, where $C = \max\left\{C_2^{1/2}, C_1^{-1/2}\right\}.$
The contribution of the case $l=0$ is thus
\begin{align}
S & \ll C \sum_{h\mid c} |\mu(h)| h^{1/2} n^{1/2} \left(\frac ch \right)^{1-1/D}\\
& \ll C n^{1/2} c^{1-1/D} \sum_{h\mid c} |\mu(h)| h^{1/D+1/2-1}\\
& \ll_\eps C n^{1/2} c^{1-1/D+\eps}.
\end{align}
Consider now the case $l\ne 0$. We adopt the convention that the range of summation includes only those values of the indices for which the summands are defined, allowing us to drop the coprimality condition. We begin by applying the Weyl-van der Corput inequality \cite[eq.~(2.3.5)]{graham_kolesnik_1991} for some $H\in[1,c]$ to be chosen later, which gives
\begin{align}
|S|^2 & \ll \frac{c^2} H + \frac cH \sum_{1\le h\le H} \bigg\lvert\sum_{d = 1}^c e\left(\frac{l(\overline{d+h} - \overline d)}c - nc \left( \Xi\left( - \frac{d+h}c\right) - \Xi \left( - \frac dc\right)\right) \right)\bigg\rvert.
\end{align}
Writing
\begin{align}
a_d & = e\left(\frac{l(\overline{d+h} - \overline d)}c\right),
&
b_d &
=e\left(- nc \left( \Xi\left( - \frac{d+h}c\right) - \Xi \left( - \frac dc\right)\right)\right),
\end{align}
we set
\begin{align}
T = \sum_{d=1}^c a_d b_d;
\end{align}
here $T = T_c(h,l,n)$ depends on $c$, $h$, $l$, and $n$; and we follow the convention that the terms with $a_d$ undefined are assumed to be zero. Now we seek to get cancellations in the sum $T$. Summing by parts, we can write
\begin{align}
\label{eq:T_by_parts}
T &
=
b_c\sum_{q=1}^c a_d + \sum_{d=1}^{c-1} \sum_{q=1}^d a_q (b_d - b_{d+1}) ,
\end{align}
provided $c\ge 2$ (when $c=1$, the bound $T \ll 1$ is satisfactory).
Set $A_d = \sum_{q = 1}^d a_q$ and $B_d = b_d -b_{d+1}$;
we need to bound $A_d$ and $B_d$. For the first, we use smoothing to write the sum as a complete sum modulo $c$ followed by standard estimates for exponential sums; for the second, we rely on Taylor's theorem and smoothness of $\Xi$.
Let $\delta\in (0,1)$ be a number we will choose later depending on $c$, and let $I_\alpha \colon [0,1] \to \{0,1\}$ be the indicator of $[0,\alpha]$ for $\alpha\in [0,1]$. Let $\psi\colon \R \to\R$ be smooth, of integral $1$, supported on $[-1,1]$. Then, $\psi_\delta(x) = \frac1\delta \psi\left(\frac x \delta\right)$ is smooth, of integral $1$, supported on $[-\delta,\delta]$. Set
\begin{align}
I^{\delta}_{\alpha} (x) =
\begin{cases}
I_{\alpha + \delta} * \psi_\delta(x),& \alpha+\delta\le 1\\
I_{1}(x), & \alpha+\delta >1.
\end{cases}
\end{align}
Using the notation where $e_c(\cdot) = e(\tfrac{\cdot}c)$, we can write
\begin{align}
A_d
& = \sum_{q=1}^c e_c(l(\overline{q+h} - \overline q)) I_{d/c}(q/c)\\
& = \sum_{q=1}^c e_c(l(\overline{q+h} - \overline q)) I_{d/c}^\delta (q/c) + O(\delta c).
\end{align}
We introduce quantities
\begin{align}
c_{k,d}^\delta & = \int_0^1 I_{d/c}^\delta (x) e(-kx) \, dx, &
U_c(h, k,l)& = \sum_{q=1}^c e_c(l(\overline{q+h} - \overline q) + kq).
\end{align}
Then, we can write
\begin{align}
A_d = \sum_{k\in\Z} c^\delta_{k,d} U_c(h, k,l) +O(\delta c).\label{eq:need_num_theory}
\end{align}
Now the sum $U_c(h,k,l)$ can be treated using results of Bombieri \cite{bombieri} for $c$ a prime, generalized by Cochrane and Zheng \cite{cochrane-zheng} for $c$ a prime power.
We begin by recording the easy multiplicative property
\begin{equation}\label{eq:mult}
U_{q_1q_2} (h, k,l) = U_{q_1}(h, k\bar q_2, l\bar q_2, ) U_{q_2} (h, k\bar q_1, l\bar q_1)
\end{equation}
whenever $q_1,q_2\in \N$ are coprime and $\bar q_1,\bar q_2\in \Z$ satisfy $q_1\bar q_1+q_2\bar q_2=1.$
This renders it sufficient to study $U_{p^m}(h, k,l)$ for a prime power $p^m$.
We may write $U_{p^m}(h,k,l)$
in the form
\begin{equation}\label{eq:fall}
\sideset{}{^\star} \sum_{\substack{q\imod{p^m}}} e_{p^m}\left(\frac{f_1(q)}{f_2(q)}
\right),
\end{equation}
where $f_1(q) = kq^2(q+h) - hl$ and $f_2(q)=q(q+h)$.
The symbol
$\sum^\star$ emphasizes the fact that $q$ is only taken over values for which $q\nmid f_2(q)$,
in which scenario $f_1(q)/f_2(q)$ means
$f_1(q)\overline{f_2(q)}$.
We proceed by establishing the following result, which is far form optimal, but sufficient for our needs.
\begin{lemma}\label{lem:primes}
Let $p$ be a prime and $m\in \N$. Then we have
\begin{align}
U_{p^m}(h,k,l)
\ll
\begin{cases}p^{1/2}(p,(k,hl))^{1/2} , & m=1,
\\
p^{2m/3} (p^m,(k,lh))^{m/3}, & m>1.
\end{cases}
\end{align}
\end{lemma}
\begin{proof}
When $m=1$, the result follows from \cite[eq.~(1.2)]{cochrane-zheng}, which is a restatement of Bombieri's result \cite{bombieri}.
When $m>1$, we use \cite[Cor.~3.2]{cochrane-zheng}. In their notation, we have $d(f_1) = 3$, $d(f_2) = 2$, $d(f) = 5$, $d^*(f) = 3$,
\begin{align}
d_p(f) =
\begin{cases}
0, & p\mid k, p\mid hl,\\
2, & p\mid k, p\nmid hl,\\
1, & p\nmid k, p\mid hl,\\
5, & p\nmid k, p\nmid hl.\\
\end{cases}
\end{align}
In the last three cases, $d^*_p(f) = 2,1,3$, respectively, so that, by \cite[Cor.~3.2]{cochrane-zheng}, $U_{p^m}(h,k,l) \ll p^{2m/3}$, which is satisfactory. In the first case, we choose $t\in\N$ so that $p^t\| (k,hl)$. If $t\ge m$, the trivial bound on $U_{p^m}(h,k,l)$ is satisfactory. If $t < m$, we write $t = t_1 + t_2$, where $p^{t_1} \mid h$ and $p^{t_2} \mid l$,
\begin{align}
U_{p^m}(h,k,l) = p^t U_{p^{m-t}}(hp^{-t_1}, kp^{-t}, lp^{-t_2}) \ll p^t p^{2(m-t)/3} = p^{2m/3+t/3},
\end{align}
which is satisfactory.
\end{proof}
We write $c=uv$, where $u$ is square-free and $v$ is square-full. That is, $p\mid u$ implies $p^2 \nmid u$ and $p\mid v$ implies $p^2 \mid v$.
Using the multiplicativity property \eqref{eq:mult}, we may
apply Lemma \ref{lem:primes} for different primes
to arrive at the following result.
\begin{lemma}\label{lem:cancellations}
Let $c\in \mathbb{N}$ and let $h, k, l\in \mathbb{Z}$.
Then for every $\eps>0$ we have
\begin{align}
U_{c}(h,k,l)
\ll_\eps
c^{\eps}
u^{1/2} (u, (k,hl))^{1/2}
v^{2/3} (v, (k,hl))^{1/3}.
\end{align}
\end{lemma}
We substitute this bound into \eqref{eq:need_num_theory} together with the bound for Fourier coefficients
\begin{align}
c_{k,d}^\delta \ll_{\gamma,\psi} \frac1{1+k} \cdot \left(\frac{1}{k\delta + 1}\right)^\gamma
\end{align}
for $\gamma \ge 0$. Choosing $\gamma = \eps$, we get
\begin{align}
A_d & \ll_{\eps} \sum_{k\in\Z}
c^{\eps}
u^{1/2} (u, (k,hl))^{1/2}
v^{2/3} (v, (k,hl))^{1/3}
\frac1{1+k} \cdot \left(\frac{1}{k\delta + 1}\right)^\eps + O(\delta c)
\\
&
\ll \sum_{k\in\Z}
c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3}
\frac{(c,k)^{1/2}}{1+k} \cdot \left(\frac{1}{k\delta + 1}\right)^{\eps} + O(\delta c).
\end{align}
Now we observe that
\begin{align}\label{eq:gcd_av}
\sum_{k=1}^K (c, k)^{1/2} & \le \sum_{s\mid c} s^{1/2} \sum_{\substack{k\le K \\ s\mid k}} 1
\ll
\sum_{s\mid c} s^{1/2} \frac{K}{s}\\
&
\ll K\sum_{s\mid c} s^{-1/2}
\ll K \tau(c) \ll_\eps K c^\eps.
\end{align}
Summing by parts, we conclude that
\begin{align}A_d\ll_\eps c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3} \delta^{-\eps} + c\delta
\ll_\eps
c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3},
\end{align}
choosing $\delta = c^{-1/2}$. Combining this deduction with $B_d \ll \frac{nH}c \sup|\Xi''|$ (and the trivial bound for $b_c$ in the boundary term of \eqref{eq:T_by_parts}), we get
\begin{align}T &
\ll
c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3}+\sum_{d=1}^{c-1} c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3}
\frac{nH}c \sup|\Xi''|
\\
&
\ll
(1+\sup|\Xi''| ) nH c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3}.
\end{align}
We finally get
\begin{align}
|S|^2 & \ll \frac{c^2}H + \frac cH \sum_{1\le h\le H}(1+ \sup|\Xi''|) nH c^{\eps}
u^{1/2} (u, hl)^{1/2}
v^{2/3} (v, hl)^{1/3}\\
&\ll
\frac{c^2}H + c^{1+\eps} H n u^{1/2} (u, l)^{1/2}
v^{2/3} (v, l)^{1/3}(1+\sup|\Xi''|),
\end{align}
using
\eqref{eq:gcd_av} with $h$ in place of $k$. The optimal choice for $H$ is $[c^{1/4}]$, so that
\begin{align}
S \ll_\eps (1+ \sup |\Xi''|^{1/2}) c^{5/8+\eps} u^{1/4} v^{1/3} n^{1/2} (u, l)^{1/4}(v, l)^{1/6}.\label{eq:S_l_not_0}
\end{align}
Note that we are not concerned with the value of $\eps$, and thus don't distinguish between $\eps$ and $\eps/2$.
We have thus proved the following proposition.
\begin{proposition} \label{prop:cancellations}
Let $S = S_c(k,l, n)$ be the sum defined in \eqref{eq:sum_l}. Write $c = uv$ with $u$ square-free and $v$ square-full. Then, we have
\begin{align}
S_c(k,0,n) & \ll_\eps C n^{1/2} c^{1-1/D+\eps}\\
S_c(k,l,n) & \ll_\eps (1+ \sup |\Xi''|^{1/2}) c^{5/8+\eps} u^{1/4} v^{1/3} n^{1/2} (u, l)^{1/4}(v, l)^{1/6}.
\end{align}
\end{proposition}
\section{Fourier decomposition\label{sec:fourier}}
In this section we develop the tools necessary to prove Theorem \ref{th:main} and decompose $f$ into a Fourier series on the torus. We proceed exactly as in \cite{strombergsson_effective_2013, browning_vinogradov_2016}.
To begin with we note that
\begin{align}
f((1,\bm \xi)M)=f((1,\bm \xi+\bm n)M)
\end{align}
for $\bm n \in \Z^2$.
So for $M$ fixed, $f$ is a well defined function on $\R^2/\Z^2$ and we can expand it into a Fourier series as
\beq \label{eq:fourier}f((1,\bm \xi)M)=\sum_{\bm m\in\Z^2}\hat f(M,\bm m) e(\bm m\bm \xi),\eeq
where
\begin{align}\hat f(M,\bm m)=\int_{\T^2}f((1,\bm \xi')M)e(-\bm m \bm \xi')d\bm \xi'.
\end{align}
Note that
\beq\hat f(TM,\bm m)=\hat f(M,\bm m (T^{-1})^t),
\label{eq:transpose}\eeq
for $T\in\sltz.$ Set $\tilde f_n(M)=\hat f(M,(n,0))$.
These functions of $M\in\sltr$ are left-invariant under the group $\left(\begin{smallmatrix}1&\Z\\0&1\end{smallmatrix}\right)$ by \eqref{eq:transpose}.
Now it follows from
\eqref{eq:transpose} that
\begin{align}
\tilde f_n\left(\abcd M\right)=\hat f \left(\abcd M,(n,0)\right)
&=\hat f \left(M,(n,0)\begin{pmatrix} d & -c\\ -b & a\end{pmatrix}\right)\\
&=\hat f(M,(nd,-nc)).
\end{align}
Therefore we can rewrite \eqref{eq:fourier} with $\bm \xi= (\xi_1(x),\xi_2(x))$ and $M=\big(\begin{smallmatrix}\sqrt y&x/\sqrt y\\0&1/\sqrt y\end{smallmatrix}\big)$ as
\beq f\left(\left(1,\bm\xi\right)M\right)=\tilde f_0(M)+\sum_{n\ge 1}\sum_{(c,d)=1} \tilde f_n\left(\begin{pmatrix}*&*\\c&d\end{pmatrix}M\right)e\left(n\left(d\xi_1(x) - c\xi_2(x)\right)\right),\label{eq:afterfourier}\eeq
where $\left(\begin{smallmatrix}*&*\\c&d\end{smallmatrix}\right)=\ensuremath{\left(\begin{smallmatrix}a & b\\c& d\end{smallmatrix}\right)}$ is any matrix in \sltz\ with $c$ and $d$ in the second row as specified.
Integrating \eqref{eq:afterfourier} over $x$, we obtain
\beq
\int_{\R} f(\tilde n(x)a(y))\,\rho(x)dx=
M(y)+E(y),
\eeq
where
\beq\label{eq:main_term}
M(y)=
\int_{\R} \tilde f_0\begin{pmatrix}\sqrt y&x/\sqrt y\\0&1/\sqrt y\end{pmatrix} \,\rho(x)\,dx
\eeq
and
\beq
E(y)=\sum_{\substack{n\ge1\\ (c,d) =1}} \int_{\R} e\left(n\left(d\xi_1(x) - c\xi_2(x)\right)\right)\tilde f_n\left(\begin{pmatrix}*&*\\c&d\end{pmatrix} \begin{pmatrix}\sqrt y&x/\sqrt y\\0&1/\sqrt y\end{pmatrix}\right)\,\rho(x)dx.\label{eq:errorterms}
\eeq
The main term in this expression is $M(y)$ and, as is well-known
(cf.~\cite{flaminio_invariant_2003, strombergsson_deviation_2013}), we have
\begin{align}M(y)=\int_X f\, d\mu\int_\R \rho(x)dx +O_\eps(\|f\|_{C_{\mathrm{b}}^4} \|\rho\|_{W^{1,1}} y^{1/2-\eps})\end{align}
for every $\eps>0$, where the Sobolev norm of $\rho$ is defined in \eqref{eq:sobolev_norm}.
This statement is nothing more than effective equidistribution of horocycles under the geodesic flow on $\sltz\quot\sltr.$ We need not seek the best error term for this problem, since there will be larger contributions to the error term in Theorem \ref{th:main}.
It remains to estimate $E(y)$ as $y\to 0$, which we do in Section \ref{sec:error}.
\medskip
We end this section with several technical results that will help us to estimate $E(y)$. First, however, we give a precise definition of $\|\cdot\|_{C_{\mathrm{b}}^m}$ and $\|\cdot\|_{W^{k,p}}$ for functions on $G$ and hence also on $X$. Following \cite{strombergsson_effective_2013}, we let $\mathfrak g = \Sl(2,\R)\oplus \R^2$ be the Lie algebra of $G$ and fix
\begin{align}\begin{aligned}
X_1&=\left(\left(\begin{smallmatrix}0&1\\0&0\end{smallmatrix}\right),\bm 0\right),\quad
X_2=\left(\left(\begin{smallmatrix}0&0\\mathbbm{1}&0\end{smallmatrix}\right),\bm 0\right),\quad
X_3=\left(\left(\begin{smallmatrix}1&0\\0&-1\end{smallmatrix}\right),\bm 0\right),\\
X_4&=\left(\left(\begin{smallmatrix}0&0\\0&0\end{smallmatrix}\right),(1,0)\right),\quad
X_5=\left(\left(\begin{smallmatrix}0&0\\0&0\end{smallmatrix}\right),(0,1)\right)
\end{aligned}\end{align}
to be a basis of $\mathfrak g$. Every element of the universal enveloping algebra $U(\mathfrak g)$ corresponds to a left-invariant differential operator on functions on $X$. We define
\beq\label{eq:norm}
\|f\|_{C_{\mathrm{b}}^m} = \sum_{\deg D\le m} \|Df\|_{L^\infty},
\eeq
where the sum runs over monomials in $X_1,\dots,X_5$ of degree at most $m$.
We also Sobolev norms of functions on \R. For $1\le p< \infty$ and a positive integer $k$, set
\beq\label{eq:sobolev_norm}
\|\rho\|_{W^{k,p}} = \sum_{s=0}^k \|\rho^{(s)}\|_{L^p} = \sum_{s=0}^k \left( \int_\R\nolimits \lvert \rho^{(s)}(x) \rvert^p dx\right)^{1/p}.
\eeq
The following result is
\cite[Lemma~4.2]{strombergsson_effective_2013}.
\begin{lemma}\label{lem:andreas_fourier1}
Let $m\ge0$ and $n>0$ be integers. Then
\begin{align}\tilde f_n\abcd \ll_m \frac{\|f\|_{C_{\mathrm{b}}^m}}{n^m(c^2+d^2)^{m/2}}, \quad \forall \abcd \in \SL(2,\R). \end{align}
\end{lemma}
Passing to Iwasawa coordinates in \sltr, we write
\beq\label{eq:iwasawa}\tilde f_n(u,v,\theta)=\tilde f_n\left(
\begin{pmatrix}
1&u\\
0&1
\end{pmatrix}
\begin{pmatrix}
\sqrt v&0\\
0&1/\sqrt v
\end{pmatrix}
\begin{pmatrix}
\cos\theta &-\sin\theta\\
\sin\theta &\cos\theta
\end{pmatrix}\right).
\eeq
for $u\in \mathbb{R}, v>0$ and $\theta\in \mathbb{R}/2\pi\mathbb{Z}$.
The following is
\cite[Lemma~4.4]{strombergsson_effective_2013}.
\begin{lemma}\label{lem:andreas_fourier2}
Let $m, k_1, k_2, k_3\ge 0$ and $n>0$ be integers, and let $k=k_1+k_2+k_3$. Then
\begin{align}\d_u^{k_1}\d_v^{k_2} \d_\theta^{k_3} \tilde f_n(u,v,\theta)\ll_{m,k} \|f\|_{C_{\mathrm{b}}^{m+k}}n^{-m} v^{m/2-k_1-k_2}.\end{align}
\end{lemma}
As a consequence of Lemma \ref{lem:andreas_fourier1}, we get the bound
\begin{align}\label{eq:fn_control}\tilde f_n\left(u,\frac{\sin^2 \theta}{c^2 y},\theta\right) \ll_m \|f\|_{C^m_{\text b}} \min\left\{1,\left(\frac{\lvert \sin\theta\rvert }{nc\sqrt y}\right)^m\right\}\end{align} for every integer $m\ge 0$. We also note that for $a, A, B > 0$ and $B-A> -1$ we have
\beq
\int_{-\pi}^\pi \frac{d\theta}{\lvert \sin \theta\rvert ^A} \min\left\{1,\left(\frac{\lvert \sin \theta\rvert}{a}\right)^B\right\} \ll \min\{a^{-B},a^{-A+1}\}.\label{eq:theta_integral}
\eeq
\section{Error terms\label{sec:error}}
The purpose of this section is to estimate $E(y)$ in
\eqref{eq:errorterms}.
We begin with the case $c=0$. Then $d=\pm1$ by coprimality, and \cite[eq.~(25)]{strombergsson_effective_2013} yields
\begin{align}\label{eq:total_c_0}
E_{c= 0}(y)
=
\int_{\R} \tilde f_n\left(\pm \begin{pmatrix}\sqrt y& x/\sqrt y\\0&1/\sqrt y\end{pmatrix}\right)\, \rho(x)dx
\ll
\|f\|_{C_{\mathrm{b}}^2}\frac{y}{n^2}
.
\end{align}
After summing over $n$, the contribution from this term is clearly much smaller than that claimed in Theorem~\ref{th:main}.
The remaining contribution ($c\ne 0$) to the error term $E(y)$ in \eqref{eq:errorterms} is
\beq
E_{c\ne 0}(y)=\sum_{\substack{c\ne0 \\ n\ge1\\ (c,d) =1}}
\int_{\R} e\left(n\left(d\xi_1(x) - c\xi_2(x)\right)\right)\tilde f_n\left(\begin{pmatrix}*&*\\c&d\end{pmatrix} \begin{pmatrix}\sqrt y&x/\sqrt y\\0&1/\sqrt y\end{pmatrix}\right)\rho(x)dx.
\eeq
Now we proceed to the change of variables, following \cite[Lemma~6.1]{strombergsson_effective_2013}. Writing the argument of $\tilde f_n$ in Iwasawa coordinates \eqref{eq:iwasawa}, we get
\begin{align}
\int_\R e\left(n\left(d\xi_1(x) - c\xi_2(x)\right)\right) \tilde f_n \left(\begin{pmatrix}a&b\\c&d\end{pmatrix} \begin{pmatrix}\sqrt y&x/\sqrt y\\0& 1/\sqrt y\end{pmatrix}\right) \rho(x)\, dx = \int_0^\pi g(\theta) d\theta,
\end{align}
for $c>0$, where
\begin{align}
g(\theta) & = e\left(n\left( d\xi_1\left(-\frac dc+y\ctg \theta\right) -c\xi_2\left(-\frac dc+y\ctg \theta\right)\right)\right)\label{eq:exp_to_simplify}
\\
& \quad \times \tilde f_n\left(\frac{a}c -\frac{\sin 2\theta}{2c^2 y}, \frac{\sin^2\theta}{c^2y},\theta\right)
\rho\left(-\frac dc + y\ctg\theta\right)
\frac{y}{\sin^2\theta}.\notag
\end{align}
We have the same integral with limits $-\pi$ and 0 if $c<0$. Combining terms with positive and negative $c$ gives
\begin{align}
E_{c\ne 0}(y)=
\sum_{\substack{ c,n\ge 1}} \sum_{(c,d)=1} \int_{-\pi}^\pi g(\theta)d\theta.
\end{align}
Let
\beq
\tilde g(\theta) = \rho\left(-\tfrac dc \right) \tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
e\left ( - n c\Xi\left(-\tfrac dc\right)\right)
\frac{y}{\sin^2\theta},
\eeq
where $\Xi(z)=z\xi_1(z)+\xi_2(z)$.
\begin{lemma}
For every $\eps>0$, we have
\begin{equation}\sum_{c,n\ge 1} \sum_{\substack{(c,d)=1\\d \in \Z}} \int_{\theta=-\pi}^\pi (g(\theta)-\tilde g(\theta))d\theta \ll_\eps
\|f\|_{C^4_{\text b}} (1+\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty})\|\rho\|_{W^{1,1}}
y^{1/2 -\eps}
\label{eq:toanalyze}
\end{equation}
for $0<y<1$.
\end{lemma}
\begin{proof}
We write
\begin{align}
g(\theta) - \tilde g(\theta) & =
\tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
\frac{y}{\sin^2\theta}
\\
\notag
&\quad\times
\big[e\left(n\left( d\xi_1\left(-\tfrac dc+y\ctg \theta\right) -c\xi_2\left(-\tfrac dc+y\ctg \theta\right)\right)\right) \rho\left(-\tfrac dc +y\ctg \theta\right) \\
&\quad\quad - e\left ( - n c\Xi\left(-\tfrac dc\right)\right) \rho\left(-\tfrac dc\right)
\big]
\notag
\\
&
=
\tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
\frac{y}{\sin^2\theta}e\left(- nc\Xi \left(-\tfrac dc\right)\right)
\\
&
\quad\times
\left[e\left(O((\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty}) ncy \lvert \ctg \theta \rvert\right) \rho\left(-\tfrac dc +y\ctg \theta\right) - \rho\left( -\tfrac dc \right)
\right].\notag
\end{align}
When $ncy \lvert \ctg \theta \rvert < 1,$ we use Taylor expansion of the exponential; in the complementary case we bound it trivially.
The contribution of the first option comes in two parts since $e(z) = 1 + O(z)$.
The first part is controlled by using \eqref{eq:fn_control} with $m=2$ and \eqref{eq:theta_integral} with $A = 2$ and $B=2$, together with elementary inequalities.
We have the bound
\begin{align}
\sum_{c,d,n} & \int_{\theta=-\pi}^\pi \tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
\frac{y\, d\theta}{\sin^2\theta}e\left(- nc\Xi \left(-\tfrac dc\right)\right)
\\\notag
&\quad\times
\left[\rho\left(-\tfrac dc +y\ctg \theta\right) - \rho\left( -\tfrac dc \right)
\right] \mathbbm{1}_{ncy\lvert \ctg \theta\rvert <1 }\\
&\ll \|f\|_{C^2_\text{b}} \int_{-\pi}^\pi\limits \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^2\right\}
\\\notag
&\quad \times
\sum_{\mathclap{(c,d) = 1}} \left|\rho\left( -\tfrac dc\right) - \rho\left( - \tfrac dc + y\ctg\theta\right)\right| \mathbbm{1}_{ncy\lvert \ctg \theta\rvert <1 }\\
&\ll \|f\|_{C^2_\text{b}} \int_{-\pi}^\pi\limits \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^2\right\}
\\
\notag
&\quad \times
\sum_{h\mid c} |\mu(h)| \sum_{d\in \Z} \int_{-\frac{dh}{c}}^{-\frac{dh}{c} + y\ctg \theta} |\rho'(t)|\, dt \, \mathbbm{1}_{y\lvert \ctg \theta\rvert <\frac1{nc} }.
\end{align}
Now we use the condition $y\lvert \ctg \theta \rvert <\frac1{nc}$ to recast the sum in $d$ and the integral in $t$ to a single integral over the real line:
\begin{align}
&
\ll
\|f\|_{C^2_\text{b}} \int_{-\pi}^\pi \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^2\right\} \sum_{h\mid c} |\mu(h)| \int_{-\infty}^\infty |\rho'(t) | \, dt\\
&
\ll_\eps \|f\|_{C^2_\text{b}} \|\rho\|_{W^{1,1}} \int_{-\pi}^\pi \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^2\right\} c^\eps
\end{align}
At this step we take advantage of \eqref{eq:theta_integral} with $A = 2 = B$, giving the bound
\begin{align}
&\ll \|f\|_{C^2_\text{b}} \|\rho\|_{W^{1,1}} \sum_{n,c} yc^\eps \min\left\{(nc\sqrt y)^{-2}, (nc\sqrt y)^{-1}\right\}
\\
&
\ll \|f\|_{C^2_\text{b}} \|\rho\|_{W^{1,1}} \sum_{k=1}^\infty y k^\eps \min\left\{(k\sqrt y)^{-2}, (k\sqrt y)^{-1}\right\}
\\
& \ll \|f\|_{C^2_\text{b}} \|\rho\|_{W^{1,1}} y^{1/2 -\eps}.
\end{align}
The second part is controlled by using \eqref{eq:fn_control} with $m=4$ followed by \eqref{eq:theta_integral} with $A = 3$ and $B=4$, together with elementary inequalities.
The bound in this case is
\begin{align}
\sum_{c,d,n} &\int_{\theta=-\pi}^\pi\tilde f_n \left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
\frac{y\, d\theta}{\sin^2\theta}e\left(- nc\Xi \left(-\tfrac dc\right)\right)
\\\notag
&\quad \times
\left[ O((\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty}) ncy \lvert \ctg \theta \rvert \rho\left(-\tfrac dc +y\ctg \theta\right)
\right] \mathbbm{1}_{ncy\lvert \ctg \theta\rvert <1 }\\
&\ll \|f\|_{C^4_\text{b}} (\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty}) \int_{-\pi}^\pi \frac{y\, d\theta}{\sin^2\theta} \sum_{n,c} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^4\right\} ncy \lvert \ctg\theta \rvert
\\
&
\quad
\times\sum_{h\mid c} |\mu(h)| \sum_{d\in\Z} \rho\left(-\frac{dh}c + y\ctg \theta\right)\mathbbm{1}_{ncy\lvert \ctg \theta\rvert <1 }\notag
\\
&\ll \|f\|_{C^4_\text{b}} (\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty}) \int_{-\pi}^\pi \frac{ncy^2\, d\theta}{\lvert \sin \theta \rvert^3} \sum_{n,c} \min\left\{ 1, \left(\frac{\sin \theta}{nc\sqrt y} \right)^4\right\}
\\
&
\quad
\times \sum_{h\mid c} |\mu(h)| \left( \frac ch \int_{-\infty}^\infty \rho(t) \, dt + \int_{-\infty}^\infty |\rho'(t)|\,dt \right).\notag
\end{align}
Now the last line is at most a constant times $c^{1+\eps}\|\rho\|_{W^{1,1}}$, since $\rho$ is of integral $1$ and is supported on an interval of length $1$. We get
\begin{align}
&\ll \|f\|_{C^4_\text{b}} (\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty})\|\rho\|_{W^{1,1}}
\sum_{n,c}y^2 n c^{2+\eps} \min\{(nc\sqrt y)^{-2}, (nc\sqrt y)^{-4}\}\\
&\ll_\eps \|f\|_{C^4_\text{b}} (\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty})\|\rho\|_{W^{1,1}}
\sum_{k=1}^\infty y^2 k^{2+\eps}\min\{(k\sqrt y)^{-2}, (k\sqrt y)^{-4}\}\\
&\ll_\eps \|f\|_{C^4_\text{b}} (\|\Xi\|_{C^1_{\text b}} + \|\xi_1\|_{L^\infty})\|\rho\|_{W^{1,1}} y^{1/2 - \eps}.
\end{align}
Now we peruse the second option, $ncy\lvert \ctg \theta \rvert \ge 1$. Again, we distinguish two subcases, $1\ge ncy \ge \lvert \tg \theta\rvert$ and $1 <ncy \ge \lvert \tg \theta\rvert.$ The first subcase is dealt with using \eqref{eq:fn_control} with $m=3$ followed by elementary estimates for the integral in $\theta$:
\begin{align}
\sum_{c,d,n} & \int_{\theta=-\pi}^\pi \left\lvert\tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)\right\rvert
\frac{y\, d\theta}{\sin^2\theta}
\\\notag
&
\quad \times
\left[\rho\left(-\tfrac dc +y\ctg \theta\right) + \rho\left( -\tfrac dc \right)
\right] \mathbbm{1}_{1\ge ncy \ge \lvert \tg \theta\rvert}\\
&\ll
\|f\|_{C^3_{\text b}} \int_{-\pi}^\pi\limits \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{1 , \left(\frac{\lvert \sin \theta\rvert }{nc\sqrt y}\right)^3\right\}
\\\notag
&
\quad \times
\sum_{\mathclap{(c,d)=1}} \left( \rho\left( - \tfrac dc + y\ctg \theta\right) + \rho\left( -\tfrac dc \right) \right)
\mathbbm{1}_{1\ge ncy \ge \lvert \tg \theta\rvert}\\
&\ll
\|f\|_{C^3_{\text b}} \int_{-\pi}^\pi\limits \sum_{n,c} \frac{y\, d\theta}{\sin^2 \theta} \min\left\{1 , \left(\frac{\lvert \sin \theta \rvert}{nc\sqrt y}\right)^3\right\} c^{1+\eps}\|\rho\|_{W^{1,1}} \mathbbm{1}_{1\ge ncy \ge \lvert \tg \theta\rvert}
\end{align}
Using the same reasoning as before convert the sum over $d$ into an integral, we arrive at the bound
\begin{align}
&
\ll
\|f\|_{C^3_{\text b}}\|\rho\|_{W^{1,1}}\sum_{n,c} yc^{1+\eps}\int_{0\le \theta \ll ncy} \frac{\theta \, d\theta}{n^3c^3 y^{3/2}} \mathbbm{1}_{1\ge ncy}
\\
&
\ll
\|f\|_{C^3_{\text b}}\|\rho\|_{W^{1,1}} \sum_{n,c} y^{-1/2} n^{-1} c^{\eps} y^2\mathbbm{1}_{1\ge ncy}\\
&
\ll
\|f\|_{C^3_{\text b}}\|\rho\|_{W^{1,1}} y^{1/2 -\eps}.
\end{align}
In the second subcase, we only keep the condition $ncy > 1$ to get the bound
\begin{align}
\sum_{c,d,n} & \int_{\theta=-\pi}^\pi\limits \left\lvert\tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)\right\rvert
\frac{y\, d\theta}{\sin^2\theta}
\left[\rho\left(-\tfrac dc +y\ctg \theta\right) + \rho\left( -\tfrac dc \right)
\right] \mathbbm{1}_{1< ncy }\\
&\ll \|f\|_{C^3_{\text b}} \|\rho\|_{W^{1,1}}\sum_{n,c} \int_{-\pi}^\pi \frac{y \, d\theta}{\sin^2\theta} \min\left\{1 , \left(\frac{\lvert \sin \theta\rvert }{nc\sqrt y}\right)^3\right\} c^{1+\eps} \mathbbm{1}_{1< ncy }\\
&
\ll
\|f\|_{C^3_{\text b}}\|\rho\|_{W^{1,1}} \sum_{n,c} \int_{-\pi}^\pi
\frac{\lvert \sin\theta \rvert y c^{1+\eps}d\theta}{(nc\sqrt y)^3}\mathbbm{1}_{1< ncy }\\
&
\ll
\|f\|_{C^3_{\text b}} \|\rho\|_{W^{1,1}} y^{-1/2}\sum_{k>1/y} \frac{k^{1+\eps}}{k^3}
\\
&
\ll
\|f\|_{C^3_{\text b}}\|\rho\|_{W^{1,1}} y^{1/2-\eps}.
\end{align}
\end{proof}
\bigskip
We need to analyze
\begin{align}
\tilde E_{c\ne 0}(y) = \sum_{c,n\ge 1} \int_{\theta=-\pi}^\pi\limits \sum_{\substack{(c,d) = 1\\ d\in\Z}}
\tilde f_n\left(\frac{\bar d}c -\frac{\sin 2\theta}{2c^2y},\frac{\sin^2 \theta}{c^2y},\theta\right)
e\left( - nc\Xi\left(-\frac dc\right)\right) \rho\left(-\frac dc\right)\frac{y \,d\theta}{\sin^2\theta}.
\end{align}
Define Fourier coefficients
\begin{align}
\label{eq:bdef} b_l^{(n,c)}(\theta)& = \int_{0}^1 \tilde f_n\left (u, \frac{\sin^2\theta}{c^2y},\theta\right) e(-lu) \, du,\\
\label{eq:adef} a_k^{(n,c)} & = \int_0^1 \rho(u) e(-ku) du.
\end{align}
Then, we can use the fact that $\rho$ is supported within $(-1, 0]$ to write
\begin{align}
\tilde E_{c\ne 0}(y) & = \sum_{\substack{c,n\ge 1\\ k,l\in\Z}} \int_{\theta=-\pi}^\pi\limits \sum_{\substack{(c,d) = 1\\ 0\le d < c}}
b_l^{n,c}(\theta)
a_k^{(n,c)}
e\left(\frac{l\bar d}c -\frac{l\sin 2\theta}{2c^2y}\right)
\\
&
\notag
\quad \times
e\left(-nc\Xi\left(-\frac dc\right)-\frac{kd}c\right)
\frac{y \,d\theta}{\sin^2\theta}\\
&\le
\sum_{\substack{c,n\ge 1\\ k,l\in\Z}} \int_{\theta=-\pi}^\pi
|b_l^{n,c}(\theta)
a_k^{(n,c)}|
\left\lvert\sum_{\substack{(c,d) = 1\\ 0\le d < c}}
e\left(\frac{l\bar d - kd}c -
nc\Xi\left(-\frac dc\right)
\right)
\right\rvert
\frac{y \,d\theta}{\sin^2\theta}.
\end{align}
Our objective is to get savings for the sum over $d$ and use the bounds
\begin{align}\label{eq:fourierbound} b_l^{(n,c)}(\theta)\ll \begin{cases}\|f\|_{C_{\mathrm{b}}^m}\min\left\{1, \left(\dfrac{\lvert\sin\theta\rvert}{nc\sqrt y}\right)^m\right\}& \text{for any }m\ge 0,\\
l^{-2}\|f\|_{C_{\mathrm{b}}^{m+2}} n^{-4}\min\left\{1, \left(\dfrac{\lvert \sin\theta\rvert}{nc\sqrt y}\right)^{m-4}\right\}& \text{for any }m\ge 4.
\end{cases}
\end{align}
and
\begin{align}
\label{eq:measure_fourier}
a_k\ll_\eta (1+|k|)^{-1-\eta} \|\rho\|_{W^{1,1}}^{1-\eta} \|\rho\|_{W^{2,1}}^\eta, \text{ for } \eta\in(0,1).
\end{align}
The bound \eqref{eq:fourierbound} is taken from \cite[Lemma~4.1]{browning_vinogradov_2016}, while the bound \eqref{eq:measure_fourier} follows from \eqref{eq:adef} and integration by parts.
We use the first bound with $m=2$ and $m=6$, and note that
\begin{align}\label{eq:strominequality}
\int_{-\pi}^\pi \min\left\{1, \left(\frac{\lvert\sin\theta\rvert}a\right)^2\right\} \frac{d\theta}{\sin^2\theta}
\ll \frac{1}{a(1+a)},
\end{align}
for $a>0$, following \eqref{eq:fn_control}. This inequality will be applied with $a = nc\sqrt y$.
We write
\begin{align}
S_c(k,l,n) = S = \sum_{\substack{(c,d) = 1\\ 0\le d < c}}
e\left(\frac{l\bar d - kd}c -
nc\Xi\left(-\frac dc\right)
\right).
\end{align}
Cancellations in the exponential sum are proved in Section \ref{sec:number}, where Proposition \ref{prop:cancellations} is established,
distinguishing two cases, when $l=0$ and when $l\ne 0$.
Combining contributions of these two cases, we control $\tilde E_{c\ne 0}(y)$ by
\begin{align}
\tilde E_{c\ne 0}(y) &\ll C \sum_{\substack{c,n\ge 1\\ k\in\Z}} |a_k| \int_{-\pi}^\pi |b_0(\theta)|\frac{y \, d\theta}{\sin^2 \theta} n^{1/2} c^{1-1/D+\eps} \\ \notag
&
\quad + (1+C_2^{1/2}) \sum_{\substack{c, n\ge 1\\ k\in\Z\\ l\ne 0}} |a_k| \int_{-\pi}^\pi |b_l(\theta)|\frac{y \, d\theta}{\sin^2 \theta}
c^{\frac58+\eps} u^{\frac14} v^{\frac13} n^{\frac12} (u, l)^{\frac14}(v, l)^{\frac16}
\\
& = E_{l=0}(y) + E_{l\ne0}(y).
\end{align}
For $E_{l=0}(y)$ we use the first bound from \eqref{eq:fourierbound} with $m=2$, and \eqref{eq:measure_fourier} followed by \eqref{eq:strominequality}. After bringing out factors of $F = C\|f\|_{C^2_{\text b}} \|\rho\|_{W^{1,1}}^{1-\eta} \|\rho\|_{W^{2,1}}^\eta$, we get
\begin{align}
E_{l = 0} &
\ll_{\eta,\eps} F \sum_{c,n,k} (1+|k|)^{-1-\eta} \frac{y}{nc\sqrt y(1+nc\sqrt y)} n^{1/2} c^{1-1/D+\eps} \\
&
\ll F \sum_{c, n} \frac{\sqrt y c^{\eps}}{\sqrt n c^{1/D}(1+cn\sqrt y)} \\
&
\ll F
\sum_{n=1}^\infty \sqrt{\frac yn} \left[ \sum_{c\ge \frac{1}{n\sqrt y}} \frac{\mathbbm{1}_{n\sqrt y \le 1}}{c^{1+\frac 1D-\eps} n\sqrt y} + \sum_{c\ge 1} \frac{\mathbbm{1}_{n\sqrt y \ge 1}}{c^{1+\frac 1D-\eps} n\sqrt y} + \sum_{\mathclap{c\le \frac1{n\sqrt y}}} \frac{\mathbbm{1}_{n\sqrt y \le 1}}{ c^{\frac1D-\eps}} \right].
\end{align}
Bounding sums over $c$ gives
\begin{align}
E_{l = 0} &\ll F \bigg[ \sum_{n=1}^\infty \sqrt{\frac yn} \frac{(n\sqrt y)^{1/D-\eps}}{n\sqrt y} \mathbbm{1}_{n\sqrt y \le 1}
+ \sum_{n=1}^\infty \sqrt {\frac yn} \frac1{n\sqrt y}\mathbbm{1}_{n\sqrt y \ge 1}
\\ \notag
&\quad + \sum_{n=1}^\infty \sqrt{\frac yn} \left(\frac1{n\sqrt y}\right)^{1-\frac 1D +\eps} \mathbbm{1}_{n\sqrt y \le 1} \bigg]
\\
&
\ll F \bigg[
\sum_{n\le \frac1{\sqrt y}} n^{-\frac32 + \frac 1D -\eps} y^{\frac1{2D} -\eps} + \sum_{n \ge \frac1{\sqrt y}} n^{-3/2}
+ \sum_{n\le \frac1{\sqrt y}} \frac{y^{\frac1{2D}-\eps}}{n^{\frac32 - \frac1D+\eps}} \bigg]
\\
&
\ll F
y^{\frac{1}{2D}-\eps}.
\label{eq:total_l_0}
\end{align}
When $l\ne 0$, we use \eqref{eq:measure_fourier}, \eqref{eq:fourierbound} with $m=6,$ and \eqref{eq:strominequality}. Abbreviating $H = \|f\|_{C^8_{\text b}} \allowbreak \|\rho\|_{W^{1,1}}^{1-\eta} \|\rho\|_{W^{2,1}}^\eta (1+C_2^{1/2})$, we have
\begin{align}
E_{l\ne 0} & \ll H \sum_{c,l,n} \frac{l^{-2} n^{-4} y}{nc\sqrt y(1 + nc\sqrt y)} (n^{1/2} c^{5/8+\eps} u^{1/4} v^{1/3} (uv, l)^{1/4})\\
&\ll
H \sum_{c,n} \frac{ n^{-4} y}{nc\sqrt y(1 + nc\sqrt y)} (n^{1/2} c^{5/8+\eps} u^{1/4} v^{1/3} )
\end{align}
Now we divide the sum over $c = uv$ into dyadic intervals $[2^{j-1}, 2^j)$, $j\in \N$. This gives
\begin{align}
E_{l\ne 0}& \ll H \sum_{n,j} n^{-5} y^{1/2} \sum_{\substack{ v \le 2^j\\ \text{sq.-full}}} \sum_{\substack{2^{j-1}/v \le u\le 2^j/v \\ \text{sq.-free}}} \frac{u^{-1/8+\eps} v^{-1/24+\eps}}{1+nc\sqrt y}\\
&
\ll
H\sum_{n,j} n^{-5} y^{1/2} \sum_{\substack{ v \le 2^j\\ \text{sq.-full}}} \frac{v^{-1/24+\eps}}{1+n2^{j-1}\sqrt y} \sum_{u\le 2^j/v} u^{-1/8 +\eps}\\
&
\ll H\sum_{n,j} \frac{n^{-5} y^{1/2} 2^{7j/8+\eps}}{1+n2^{j-1}\sqrt y} \sum_{\substack{ v \le 2^j\\ \text{sq.-full}}} {v^{-1/24+\eps} v^{-7/8+\eps}}.
\end{align}
The sum over $v$ is convergent as square-full numbers have square-root density. The remaining sum gives the bound \begin{align}H y^{1/16-\eps} ,\label{eq:total_l_not_0}\end{align} as needed.
The total error contribution is
\begin{align}
H y^{1/16-\eps} +
F y^{\frac{1}{2D}-\eps} +
\|f\|_{C^4_{\text b}} y^{1/2-\eps}
(\|\xi_1\|_{L^\infty} +\|\Xi\|_{C^1_{\text b}}+1)\|\rho\|_{W^{1,1}},
\end{align}
coming from \eqref{eq:total_l_not_0}, \eqref{eq:total_l_0}, and \eqref{eq:toanalyze}, which is majorized by the expression in the statement of the theorem.
\bibliographystyle{plain}
| c7d983ebdfaa24b32da16e14d7586567546915c0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Nyikos' Manifold Problem}
For us, a \emph{manifold} is simply a locally Euclidean
topological space. Mary Ellen Rudin proved that
$\mathrm{MA}+{\sim}\mathrm{CH}$ implies every perfectly normal manifold is
metrizable~\cite{Rudin1979}. Hereditary normality ($T_5$) is a natural
weakening of perfect normality; Peter Nyikos noticed that, although
the Long Line and Long Ray are hereditarily normal non-metrizable
manifolds, and indeed the only $1$-dimensional non-metrizable
connected manifolds \cite{Nyikos1993}, it is difficult to find examples of
dimension $>1$ (although one can do so with
$\largediamond$~\cite{Rudin1979} or $\mathrm{CH}$~\cite{RudinZenor}). He
therefore raised the problem of whether it was consistent that there
weren't any \cite{Nyikos1983}, \cite{Nyikos1993}. In a series of
papers~\cite{Nyikos2002,Nyikos2003,Nyikos2004a,Nyikos2004} he was
finally able to prove this from the consistency of a supercompact
cardinal, if he also assumed that the manifolds were hereditarily
collectionwise Hausdorff. We will demonstrate that neither of
these extra assumptions is necessary:
\begin{theorem}
It is
consistent that every hereditarily normal manifold of dimension $>1$
is metrizable.
\end{theorem}
For a coherent Souslin tree $S$ (see \S 2)
PFA($S$) is the statement \cite[\S 4]{Todorcevic}:
If $\mathcal P$ is a proper poset that preserves $S$
and if $\mathcal D_\alpha (\alpha < \omega_1)$ is a sequence of
dense open subsets of $\mathcal P$ there is a filter $\mathcal
G\subset \mathcal P$ such that $\mathcal G\cap \mathcal D_\alpha \neq
\emptyset $ for all $\alpha <\omega_1$. The notation PFA(S)[S] is
adopted in \cite{Larson} to abbreviate that we are in a forcing
extension by $S$ of a model in which $S$ was a coherent Souslin tree
and in which PFA(S) held.
\begin{theorem}
It is a consequence of PFA(S)[S] that every hereditarily normal
manifold of dimension greater than 1 is metrizable.
\end{theorem}
We will isolate some known (quotable) consequences of PFA(S)[S].
The first, rather easy,
is that the bounding number $\mathfrak b$ is greater than
$\omega_1$ \cite{Larson1999}. The next is the important P-ideal
dichotomy.
\begin{defn}
A collection $\mathcal{I}$ of countable subsets of a set $X$ is a
\textrm{\bf $\mathbf{P}$-ideal} if each subset of a member of
$\mathcal{I}$ is in $\mathcal{I}$, finite unions of members of $\mathcal{I}$ are in
$\mathcal{I}$, and whenever $\{ I_n : n\in\omega\}\subseteq\mathcal{I}$, there is
a $J\in\mathcal{I}$ such that $I_n - J$ is finite for all
$n$.
\end{defn}
\begin{quotation}
\textit{\textbf{$\mathbf{PID}$ is the statement:}
For every
$P$-ideal $\mathcal{I}$ of countable subsets of some uncountable
set $A$ either
\begin{enumerate}
\item[(i)] there is an uncountable $B\subset A$ such that
$[B]^{\aleph_0}\subset \mathcal{I}$, or else
\item [(ii)] the set $A$ can be decomposed into countably many
sets, $\{ B_n : n\in \omega\}$, such that
$[B_n]\cap \mathcal{I}=\emptyset$ for each $n\in \omega$.
\end{enumerate}
}
\end{quotation}
The consistency of $\mathbf{PID}$ does have large cardinal strength
but for P-ideals on $\omega_1$ it does not --
see the discussion at the bottom of page 6 in
\cite{Todorcevic}. A statement similar to the $\mathbf{PID}$ for
ideals on $\omega_1$ is the one we need;
it also does not have large cardinal strength and is
weaker than the $\omega_1$ version of the statement
in \cite[6.2]{Todorcevic}. The statement $\mathbf{P}_{22}$
was introduced in \cite{EisworthNyikos}.
For completeness, and to introduce the ideas we will need
for another consequence of PFA(S)[S], we include a proof
in \S 2 that it is a consequence of PFA(S)[S].
\begin{quotation}
\textbf{$\mathbf{P_{22}}$ is the statement:\/}
Suppose $\mathcal{I}$ is a $P$-ideal on a stationary subset $B$ of
$\omega_1$. Then either
\begin{enumerate}
\item[(i)]there is a stationary $E\subseteq B$ such that every
countable subset of $E$ is in $\mathcal{I}$,
\item[or (ii)]there is a stationary $D\subseteq B$ such that
$[D]^{\aleph_0}\cap \mathcal{I}$ is empty.
\end{enumerate}
\end{quotation}
A space $X$ is said to be $\aleph_1$-collectionwise Hausdorff
if the points of any closed discrete subset of cardinality at most
$\aleph_1$ can be surrounded by pairwise disjoint open sets
(separated). If a separable
space is hereditarily $\aleph_1$-collectionwise
Hausdorff, then it can have no uncountable discrete subsets (known as
having countable spread).
The next consequence of PFA(S)[S] is:
\begin{quotation}
\textit{\textbf{$\mathbf{CW}$:} Normal, first countable spaces are
$\aleph_1$-collectionwise Hausdorff.}
\end{quotation}
$\mathbf{CW}$
was first shown to be consistent in~\cite{Tall1977};
it was derived
from $V=L$ in~\cite{Fleissner1974},
and was shown to be a consequence of
PFA(S)[S]
in~\cite{Larson}. In fact, it is shown in \cite{Larson}
that simply forcing with any Souslin tree will
produce a model of $\mathbf{CW}$.
Let us note now that $\mathbf{CW}$ implies that any
hereditarily normal manifold is hereditarily $\aleph_1$-collectionwise
Hausdorff. Therefore $\mathbf{CW}$ implies
that each separable hereditarily normal manifold has countable spread.
Our next axiom is our crucial new additional consequence of PFA(S)[S]:
\begin{quotation} \textit{$\mathbf{PPI}^+$: every sequentially compact
non-compact regular space contains an uncountable free sequence.
Additionally, if the space is
first countable, then it contains a copy of the ordinal space
$\omega_1$.}
\end{quotation}
Let $\mathbf{GA}$ denote the group (or conjunction)
of hypotheses:
$\mathfrak b>\omega_1$, $\mathbf{CW}$,
$\mathbf{PPI}^+$ and $\mathbf{P}_{22}$. We
have, or show,
that each is a consequence of PFA(S)[S], and also establish
the desired theorem. We show in \S \ref{nolarge} that $\mathbf{GA}$ is
consistent (not requiring any large cardinals).
\begin{theorem} $\mathbf{GA}$ implies that all hereditarily normal
manifolds of
dimension greater than one are metrizable.
\end{theorem}
We acknowledge some other historical connections.
The statement $\mathbf{PPI}^+$ is a strengthening of
\begin{quotation}
\textit{$\mathbf{PPI}$: Every first countable perfect pre-image of
$\omega_1$ includes a copy of $\omega_1$.}
\end{quotation}
$\mathbf{PPI}$ was proved from $\mathrm{PFA}$ by
Fremlin~\cite{Fremlin1988}, see also e.g.~\cite{Dow1992}. Another
consequence of PFA(S)[S] relevant to this proof is
\medskip
\begin{quotation}
\textit{$\mathbf{\mathop{\pmb{\sum}}}^-$: In a compact $T_2$, countably tight space,
locally countable subspaces of size $\aleph_1$ are
$\sigma$-discrete.}
\end{quotation}
$\mathbf{\mathop{\pmb{\sum}}}^-$ was proved from $\mathrm{MA}+{\sim}\mathrm{CH}$ by
Balogh~\cite{Balogh1983}, extending work
of~\cite{Szentmiklossy1978}.
$\mathbf{\mathop{\pmb{\sum}}}^-$ implies $\mathfrak b > \omega_1$;
this follows from the result in \cite[2.4]{ppit}
where it is shown that $\mathfrak
b = \aleph_1$ implies there is a
compact hereditarily separable space which is not Lindelof,
since $\mathbf{\mathop{\pmb{\sum}}}^-$ implies there is no such space.
$\mathbf{\mathop{\pmb{\sum}}}^-$ was shown to be a consequence of PFA(S)[S] in
\cite{FTT}.
We will need the following consequence of $\mathbf{GA}$
which is a weaker statement
than $\mathbf{\mathop{\pmb{\sum}}}^-$. The key fact that PFA(S)[S] implies
compact, separable, hereditarily normal spaces are
hereditarily Lindel\"of was first proven
in \cite[10.6]{Todorcevic}.
\begin{lem}
$\mathbf{GA}$ implies\label{metric} that if $X$ is a hereditarily
normal manifold\label{noTypeII}
then separable subsets of $X$ are Lindelof and metrizable.
\end{lem}
\begin{proof}
Let $Y$ be any separable subset of $X$ and assume
that $Y$ is not Lindelof. Recursively choose points
$y_\alpha$, together with open sets $U_\alpha$, so that
$y_\alpha \in Y\setminus \bigcup_{\beta<\alpha} U_\beta$,
$y_\alpha \in U_\alpha$, and $\overline{U_\alpha}$ is separable and
compact. Define an ideal $\mathcal{I}$
of countable subset $a$ of $\omega_1$
according to the property that $a\in \mathcal{I}$ providing
$\{ y_\alpha : \alpha \in a\}\cap U_\beta$ is finite for all
$\beta\in \omega_1$. Since $\mathfrak b > \omega_1$
we have that $\mathcal{I}$ is a P-ideal (see \cite[6.4]{Todorcevic}). To check
this, assume that $\{ a_n : n \in \omega\}$ are pairwise disjoint
infinite members of $\mathcal{I}$. For each $n$, fix an enumerating
function
$e_n $ from $\omega$ onto $a_n$. For each $\beta\in \omega_1$,
there is a function $f_\beta\in \omega^\omega$ so that, for each
$n\in \omega$ and each $m > f_\beta(n)$, $y_{e_n(m)}\notin U_\beta$.
Using $\mathfrak b>\omega_1$, there is an $f\in \omega^\omega$ such
that $f_\beta < ^* f$ for each $\beta\in \omega_1$.
For each $n$, let $F_n = \{ e_n(m) : m<f(n)\}$.
It follows that
$a = \bigcup \{a_n \setminus F_n : n\in \omega\}$ meets each
$U_\beta$ in a finite set. Thus $a\in \mathcal{I}$ and mod finite
contains $a_n$ for each $n$.
If $B$ is any subset of $\omega_1$ such that $[B]^{\aleph_0}\subset
\mathcal I$, then $D= \{ y_\beta : \beta \in B\}$ is discrete
since $D\cap U_\beta$ is finite for each $\beta \in B$. By
$\mathbf{P}_{22}$
we must then have that there is an uncountable $B\subset \omega_1$
satisfying
that $[B]^{\aleph_0}\cap \mathcal{I}$ is empty. Now let $A$ be the closure
(in $X$) of $\{ y_\beta : \beta\in B\}$. We check that $A$ is
sequentially compact. Let $\{ x_n : n\in \omega\}$ be any infinite
subset of $A$, we show that there is a limit point in $A$.
Since $X$ is first countable this shows that $A$ is sequentially
compact.
If $\{x_n : n\in \omega\}\cap \{y_\beta : \beta\in B\}$ is infinite,
let $b\in [B]^{\aleph_0}$ be chosen so that $\{x_n : n\in \omega\}
\supset \{ y_\alpha : \alpha \in b\}$. Since $b\notin \mathcal{I}$, there is
a $\beta\in \omega_1$ such that $\{y_\alpha : \alpha \in b\}\cap
U_\beta$ is infinite, and so has a limit point in the compact set
$\overline{U_\beta}$. Otherwise, we may suppose that, for each $n$,
there is an infinite $a_n\subset B$ such that $\{y_\alpha : \alpha \in
a_n\}$ converges to $x_n$. Again, using that $\mathfrak b>\omega_1$,
similar to the verification that $\mathcal{I}$ is a P-ideal,
there must be a $\beta\in \omega_1$ such that $U_\beta \cap
\{y_\alpha : \alpha \in a_n\}$ is infinite for infinitely many $n$.
For any such $\beta$, there are infinitely many $n$ with
$x_n\in \overline{U_\beta}$. It again follows that
$\overline{U_\beta}$ contains a limit of the sequence
$\{ x_n : n\in \omega\}$.
To finish the proof, we apply $\mathbf{PPI}^+$ to conclude
that either $A$ is compact or it contains a copy of
$\omega_1$. Since $\omega_1$ contains uncountable discrete sets and
$Y$ is separable, we must have that $A$ is compact. However, the
final contradiction is that $A$ is not hereditarily Lindelof and so
it cannot be covered by finitely many Euclidean open subsets of $X$.
\end{proof}
The literature on non-metrizable manifolds has identified two main types
of non-Lindel\"of manifolds, literally called Type I and Type II. A manifold is
Type II if it is separable and non-Lindel\"of. Lemma \ref{noTypeII} shows
that there are no hereditarily normal Type II manifolds if $\mathbf{GA}$ holds.
A manifold is said to be Type I, e.g. the Long Line,
if it can be written as an increasing
$\omega_1$-chain,
$\{ Y_\alpha : \alpha\in \omega_1\}$, where each $Y_\alpha$ is
Lindel\"of, open, and
contains the closure of each $Y_\beta$ with $\beta<\alpha$.
In this next definition, we use the set-theoretic notion of countable
elementary submodels to help make a more strategic choice of
a representation of our Type I manifolds. For a cardinal $\theta$,
the notation $H(\theta)$
denotes the standard set-theoretic notion of the set of all sets
that are hereditarily of cardinality less than $\theta$. These
are commonly used as stand-ins for the entire set-theoretic
universe
to avoid issues with G\"odel's famous incompleteness
theorems in arguments and constructions using elementary submodels.
We refer the reader to any advanced book on set-theory for information
about the properties of $H(\theta)$.
\begin{defn}
Suppose that $X$ is a non-metrizable
manifold with dimension $n$. Let $\mathcal B_X$ denote the collection
of compact subsets of $X$ that are homeomorphic to the closed
Euclidean
$n$-ball $\mathbb B^n$. A family $\{ M_\alpha : \alpha\in \omega_1\}$
is an elementary chain for $X$ if there is a regular
cardinal $\theta$ with $\mathcal B_X\in H(\theta)$ so that for each
$\alpha\in \omega_1$, $M_\alpha$ is a countable elementary submodel
of $H(\theta)$ such that
$\mathcal B$ and each $M_\beta$ (
$\beta<\alpha$) are members of $M_\alpha$. The chain is said to be a
continuous chain if for each limit $\alpha\in \omega_1$,
$M_\alpha = \bigcup_{\beta < \alpha } M_\beta$.
Whenever $\{ M_\alpha
: \alpha \in \omega_1\}$ is an elementary chain for $X$,
let $X(M_\alpha)$ denote the union of the collection $\mathcal B_X
\cap M_\alpha$.
\end{defn}
Here is the main reason for our preference to use elementary submodels
in this proof. Again the main ideas are from \cite{Nyikos2004}, but the
proof using elementary submodels
is much simpler. Throughout the paper the term
\textit{
component\/} refers to
the
standard notion of connected component.
\begin{lem}
Suppose that $X$ is a non-metrizable hereditarily normal
manifold of dimension $n>1$. Let $\theta$ be
a large enough regular
cardinal $\theta$ so that\label{elementary}
$\mathcal B_X\in H(\theta)$
and let $M$ be a countable elementary submodel
of $H(\theta)$ such that
$\mathcal B_X$ is a member of $M$.
Then $X(M) = \bigcup (M\cap \mathcal B_X)$ is
an open Lindelof subset of $X$ with the property
that every
component of the non-empty boundary,
$\partial X(M)$, is non-trivial.
\end{lem}
\begin{proof}
We let $\mathcal{B}_X$ denote the family of all homeomorphic copies
of the closed unit ball of~$\mathbb{R}^n$ in~$X$.
As $X$~is a manifold this family is such that whenever $O$ is open in~$X$
and $x\in O$ there is a $B\in\mathcal{B}_X$ such that $x$ is in the interior
of~$B$ and $B\subseteq O$.
Let $Y$ denote the set $X(M)$. Since $Y$ is metrizable, and
$X$ is not, $Y$ is a proper subset of $X$.
Each member of $\mathcal B_X\cap M$ is separable and hence
$B\cap M$ is dense in~$B$ whenever $B\in\mathcal B_X\cap M$;
it follows that $Y\cap M$ is a dense subset of~$Y$.
We also note that $Y$ is open since if $B\in \mathcal B_X\cap M$,
then $B$ is compact and so is contained in the interior of a finite union
of members of $\mathcal B_X$. By elementarity,
there is a such a finite set in $\mathcal B_X\cap M$.
Similarly, if $\mathcal B'$ is a finite subset of $\mathcal B_X\cap M$,
then, by elementarity, each Lindel\"of component of
$X\setminus \mathcal B'$ that meets $Y\cap M$ will be
a subset of~$Y$. More precisely, if $C$ is such a component
and if $y\in C\cap M$, then $M$ will witness
that there is a countable collection of members
of $\mathcal B_X$ that covers the component of $y$ in $X\setminus
\mathcal B'$.
Also, we have that $X$ itself must have non-Lindel\"of
components since Lindel\"of
subsets of any manifold are metrizable while $X$, being locally connected, is
the free union of its components.
Now we choose any $x$ in $\partial X(M) = \partial Y =
\overline{Y} \setminus Y$.
Take any $B\in \mathcal B_X$ with $x$ in its interior.
We assume, working towards a contradiction,
that the component of $x$ in $\partial Y$ is $\{x\}$.
Since $\partial Y\cap B$ is compact and $\{x\}$ is a
component of $\partial Y\cap B$, we can split the latter set into two
relatively clopen sets $C$ and $D$, where $C$~is the union of all components
of $\partial Y\cap B$ that meet the boundary of~$B$ and $D$~is its complement.
For now we allow for the possibility that $C=\emptyset$ but $D$~is not empty
as it contains~$x$.
We choose $W$ with $D\subseteq W$ and such that $\overline {W}$
is contained in the interior of~$B$ and disjoint from~$C$.
Note that $W$ and $\overline{W}$ are Lindel\"of because $B$ is compact and
hereditarily Lindel\"of, being homeomorphic to the unit ball of~$\mathbb{R}^n$.
Since $\partial W$ and $\partial Y$ are disjoint the set
$\partial W\cap Y$ is closed and hence compact.
There is a finite subfamily~$\mathcal B_1$ of $M\cap \mathcal B_X$ whose
union contains $\partial W\cap Y$.
The complement $W\setminus \bigcup\mathcal B_1$ is a neighbourhood of~$x$,
so it meets $Y\cap M$.
The component, $E$, of $x$ in this complement is Lindel\"of but not contained
in $Y$, therefore $E$ is not a component of $X\setminus \bigcup\mathcal B_1$
which implies that $E\setminus \overline{Y}$ is not empty.
Since $\dim E>1$, $x$ can not be a cut-point of $E$ and so it
follows that $x$ is not the only point of $E\cap \partial Y
\subset W\cap \partial Y$.
This means that we can choose disjoint open subsets of $W$,
say $O_1$ and $ O_2$, each also having compact non-empty intersection with
$\partial Y$ and whose boundaries miss~$\partial Y$.
Fix points $z_1\in O_1\cap \partial Y$ and $z_2\in O_2\cap \partial Y$.
Now $Y\cap(\partial O_1 \cup \partial O_2 )$ is compact and again can be
covered by some $K$ where $K$ is equal to a union of some finite
subfamily~$\mathcal B_2$ of~$M\cap \mathcal B_X$.
Also since $Y\cap (\partial O_1\cup \partial O_2)$ is disjoint
from the boundary of $B$, we can ensure that $K$ is disjoint
from the boundary of $B$.
Since $K$ is a compact subset of $Y$, each component of
$O_1\setminus K$ and $O_2\setminus K$ meets $Y\cap M$;
so choose points $y_1$ and $y_2$ in $Y\cap M$ that are in the components
in $O_1\setminus K$ and $O_2\setminus K$ of $z_1$ and $z_2$ respectively.
Let $C_1$ and $C_2$ be the components in $X\setminus K$ of $ y_1$
and $ y_2$ respectively.
Neither component is contained in $Y$ and so neither is Lindel\"of.
Thus, neither is contained in $B$ and so they both meet the (arcwise)
connected boundary of~$B$.
Since components of $B\setminus K$ are path-connected, there is a
path in $X\setminus K$ from $y_1$ to~$y_2 $.
By elementarity there is such a path in~$M$ and such a path would lie
completely within~$Y$
(the path is covered by a finite subfamily of~$\mathcal{B}_X$ and one such
family should be in~$M$).
However, $Y\cap (O_1\setminus K)$ is clopen in $Y$ so there is no path
in~$Y$ that connects~$y_1$ and~$y_2$.
This contradiction finishes the proof.
\end{proof}
This next corollary is the representation as a Type I sub-manifold that we
require.
\begin{corollary}
Suppose that $X$ is a non-metrizable hereditarily normal manifold of
dimension greater than 1. Then there is an increasing chain
$\{ Y_\alpha : \alpha \in \omega_1\}$ of open Lindelof subsets
satisfying that\label{component}
\begin{enumerate}
\item for each $\alpha$, the boundary $\partial Y_\alpha$ is non-empty
and contained in $Y_{\alpha+1}$,
\item for each $\alpha$, each component of $\partial
Y_\alpha$ is non-trivial,
\item for limit $\alpha$, $Y_\alpha = \bigcup \{Y_\beta : \beta\in
\alpha\}$.
\end{enumerate}
Additionally, the union $ \bigcup\{ Y_\alpha : \alpha\in
\omega_1\}$ is closed (and open) in $X$.
\end{corollary}
\begin{proof}
Fix a continuous elementary chain $\{ M_\alpha : \alpha\in
\omega_1\}$ for $X$. Fix any $\alpha\in \omega_1$.
By Lemma \ref{elementary}, $Y_\alpha = X(M_\alpha)$ is Lindelof with non-empty
boundary, $\partial X(M_\alpha)$, and
each
component in $\partial X(M_\alpha)$
is non-trivial.
By Lemma \ref{metric}, $\overline{X(M_\alpha)}$
is Lindelof, and so by elementarity,
$M_{\alpha+1}\cap \mathcal B_X$ is a cover
of $\overline{X(M_\alpha)}$. Finally, $\bigcup \{ Y_\alpha : \alpha
\in \omega_1\} $
is closed
because any $x\in X$ that is in the closure will be in
$\overline{Y_\alpha}\subset Y_{\alpha+1}$ for some $\alpha\in
\omega_1$.
\end{proof}
Now we are ready to give a proof of the main theorem. The
clever topological ideas of the
proof are taken from \cite[p189]{Nyikos2002}.
A sketch of this proof appears in~\cite{Tall2005}. The main idea of
the proof is to use $\mathbf{PPI}^+$ to find copies of
$\omega_1$ and, combined with Lemma \ref{component}, to show
that, in fact, there
are copies of the Tychonoff plank in the space. It is easily
shown that the Tychonoff plank is not hereditarily normal.
\begin{theorem} The statement $\mathbf{GA}$ implies that
each hereditarily normal manifold of dimension
greater than 1 is metrizable.
\end{theorem}
\begin{proof}
Assume that $X$ is a non-metrizable hereditarily normal
manifold of dimension greater
than 1. Let $\{ Y_\alpha : \alpha \in \omega_1\}$ be chosen
as in Corollary \ref{component}.
For each $\alpha\in \omega_1$, choose any
point $x_\alpha\in \partial Y_\alpha$.
It is immediate that $\overline{\{ x_\alpha :
\alpha\in \omega_1\}}$ is nowhere dense in $X$.
Also let
$\{ U_\alpha : \alpha \in \omega_1\}\subset \mathcal B_X$ be
any selection so that $U_\alpha\subset Y_{\alpha+1}$
and $x_\alpha$ is in the interior of $U_\alpha$.
We first show that
if $E\subset \omega_1$ is stationary, then
$D= \{ x_\alpha : \alpha \in E\}$ is not discrete.
For each limit $\alpha$, using item (3) of Corollary \ref{component},
there is a $\beta_\alpha < \alpha$ such that $U_\alpha\cap
Y_{\beta_\alpha}\setminus \overline{D}$
is not empty. By the pressing down lemma,
there is a fixed $\beta$ such that $\beta = \beta_\alpha$ for
uncountably many $\alpha\in E$. Since $Y_\beta\setminus \overline{D}$
is separable,
there are $\alpha, \alpha'\in E$ such that $U_{\alpha}\cap
U_{\alpha'}\cap Y_\beta$ is not empty.
The choice
of the sequence of $U_\alpha$'s was (basically) arbitrary,
and so it follows that $D$ can not be separated
Since $D\cup (X\setminus \overline{D})$ is first countable,
and thus $\aleph_1$-collectionwise Hausdorff,
this shows that $D$ cannot be discrete.
Define the ideal $\mathcal{I}$ by $a\in \mathcal{I}$ if $a\in
[\omega_1]^{\aleph_0}$ and, for all $\beta\in \omega_1$,
$\{ x_\alpha : \alpha \in a\}\cap U_\beta$ is finite. As before,
$\mathcal{I}$ is a P-ideal on $\omega_1$. If $A\subset \omega_1$ satisfies
that
$[A]^{\aleph_0}\subset \mathcal{I}$, then $D = \{ x_\alpha : \alpha \in
A\}$
is discrete. Therefore there is no such stationary $A$, and
so by $\mathbf{P}_{22}$, there is a
stationary subset $A$ of $\omega_1$ such that
$ [A]^{\aleph_0}\cap \mathcal{I}$ is empty. It again follows that
$X_A =
\overline{\{x_\alpha : \alpha \in A\}}$ is sequentially compact. Let
us choose, by
applying $\mathbf{PPI}^+$, a copy $W$
of $\omega_1$ contained in $X_A$. Let
$W = \{ w_\xi : \xi\in\omega_1\}$ be the homeomorphic indexing of
$W$.
For each $\alpha\in \omega_1$, Lemma \ref{metric} implies that
$\overline{Y_\alpha}$ is Lindelof and, by elementarity,
contained in $Y_{\alpha+1}$.
Therefore, we have that, for each $\alpha$,
$W\cap Y_\alpha$ is countable, and its closure is contained in
$Y_{\alpha+1}$. It follows that there is a cub
$C\subset\omega_1$ satisfying that for each
$\gamma<\delta$ both in $C$, the set $\{ w_\beta : \gamma\leq
\beta<\delta\}$ is contained in $Y_\delta\setminus
Y_\gamma$. Therefore
$\{ w_\gamma : \gamma\in C\}$ is another copy of $\omega_1$ with
the property that $w_\gamma \in \partial Y_\gamma$ for each
$\gamma\in C$.
For each $\gamma\in C$, apply Lemma \ref{component}, so as to choose
infinite compact connected $K_\gamma\subset \partial Y_\gamma$ with
$w_\gamma\in K_\gamma$. Make another selection $y_\gamma\in
K_\gamma\setminus W $ arbitrarily. Now choose,
for each $\gamma\in C$,
a basic
set $V_\gamma\in \mathcal B_X$ so that $y_\gamma$ is in the interior of
$V_\gamma$ and $V_\gamma \subset X\setminus W$. Proceeding as we did
with the sequence of $\{ x_\alpha : \alpha\in \omega_1\}$, there is a
stationary set $A_1\subset C$ so that
$\{ y_\alpha : \alpha \in A_1\}$ has sequentially compact closure.
Since $\overline{Y_{\gamma}}$ is Lindelof and contains
$\{ y_\alpha : \alpha \in A_1\cap \gamma\}$ for each $\gamma\in C$,
it follows then that the closure of
$\{ y_\alpha : \alpha \in A_1\cap \gamma\}$ is compact and disjoint
from $W$ for each $\gamma\in C$. Since $X$ is first countable, this
also implies that the closure of the entire set
$\{ y_\alpha : \alpha \in A_1\}$ is disjoint from the closed set $W$.
Since $X$ is normal, there is a continuous function $f$ from $X$ into
$[0,1]$
such that $f[W]=\{1\}$ and
$f(y_\alpha) = 0$ for all $\alpha\in A_1$. Note that $f[K_\alpha] =
[0,1]$
for each $\alpha\in A_1$. Finally, using $f$ we will show there is a
non-normal subspace for our contradiction. For each $\alpha\in A_1$,
choose, yet another, point $z_\alpha \in K_\alpha$ in such a way
that the map $f$ restricted to $\{ z_\alpha : \alpha\in A_1\}$ is
one-to-one. Repeating the steps above, there is a stationary set
$A_2\subset A_1$ so that the closure of each countable subset
of $\{ z_\alpha : \alpha \in A_2\}$ is compact. Let $Z$ denote
the closure of the set $\{ z_\alpha : \alpha \in A_2\}$,
and for each $r\in [0,1]$, let
$Z_r = f^{-1}(r)\cap Z$. We will use the following property of these
subsets of $Z$.
Consider any open set $U$ of
$X$ that contains $Z_r\cap \partial Y_\gamma$ for any $r\in [0,1]$
and $\gamma\in C_\omega$. Since $Z_r \cap Y_\gamma$ has compact
closure, there is a $\beta <\gamma$ such that
$Z_r\setminus Y_\beta$ is contained in $U$. By the pressing down
lemma, given any open $U$ containing $Z_r\cap Y_\gamma$ for all
any stationary set of $\gamma\in \omega_1$, there is a $\beta\in
\omega_1$ such that $Z_r\setminus Y_\beta$ is contained in $U$.
Choose any $r\in [0,1]$
such that $r$ is a complete accumulation point of $\{ f(z_\alpha) :
\alpha \in A_2\}$.
Choose any sequence
$\{r_n : n\in \omega\}$ converging to $r$ so that each $r_n$
is also a complete accumulation point of $\{ f(z_\alpha) : \alpha \in
A_2\}$. There is a common cub $C_\omega$ such
that $Z_{r_n}\cap \partial Y_\gamma$
and $Z_r\cap \partial Y_\gamma$ is not empty for each
$n\in \omega$ and $\gamma \in C_\omega$.
Let $Z_r(C_\omega') = \{ Z_r\cap \partial Y_\gamma : \gamma \in
C_\omega'\}$
where $C_\omega'$ is the set of relative limit points of
$C_\omega$. Since
$Z_r$ is closed in $Z$, it follows that
$Z_r\setminus Z_r(C_\omega')$ is a closed subset of $Z\setminus
Z_r(C_\omega')$. We also note that $H = Z \cap
\bigcup \{ \partial Y_\gamma : \gamma\in C_\omega'\}$ is a closed
subset of $Z$, and so $H\setminus Z_r(C_\omega')$ is a closed
subset of $Z\setminus Z_r(C_\omega')$. We show that $Z_r\setminus
Z_r(C_\omega')$ and $H\setminus Z_r(C_\omega')$ can not be separated by
disjoint open subsets of $Z\setminus Z_r(C_\omega')$. Since
$Z_r\setminus Z_r(C_\omega')$ and $H\setminus Z_r(C_\omega')$ are disjoint,
this will complete the proof.
Suppose that $U$ is an open subset of $Z\setminus Z_r(C_\omega')$
that contains $H\setminus Z_r(C_\omega')$. By the above mentioned
property of each $Z_{r_n}$, we have that there is a $\beta\in
\omega_1$
such that $Z_{r_n}\setminus Y_\beta$ is contained in $U$ for each
$n\in \omega$. Choose any $\beta < \gamma\in C_\omega \setminus
C_\gamma'$. For each $n$, choose $z_n \in Z_{r_n}\cap \partial
Y_\gamma$. Since $Z\cap \partial Y_\gamma$ is compact, let
$z$ be any limit point of $\{ z_n : n\in \omega\}$. By the continuity
of $f$, $f(z) = r$ and so $z\in Z_r\cap \partial Y_\gamma$. In
other words, $z\in Z_r\setminus Z_r(C_\omega')$, completing
the proof that $H\setminus Z_r(C_\omega')$ and $Z_r\setminus
Z_r(C_\omega')$ can not be separated by open sets.
\end{proof}
\section{on $\mathbf{P}_{22}$}
As usual $S$ is a coherent Souslin tree. For us, it will
be a full branching downward closed subtree of $\omega^{<\omega_1}$.
Naturally it is a Souslin tree (no uncountable antichains) and
has the additional property
\begin{quotation}
for each $s\in S$ and $t\in \omega^{<\omega_1}$ with $\dom(t) =
\dom(s)$, $t$ is in $S$ if and only if
$\{ \xi \in \dom(s) : s(\xi)\neq t(\xi)\}$ is finite.
\end{quotation}
In a forcing argument using $S$ as the forcing poset, we
will still use $s < s'$ to mean that $s \subset s'$, and so,
$s'$ is a stronger condition. We will also use
the more compact notation $o(s)$ to
denote the order-type of $\dom(s)$ for $s\in S$.
Now we give a proof that our statement $\mathbf{P}_{22}$ is a
consequence of PFA(S)[S] following \cite[6.1]{Todorcevic}.
Here is a simple standard fact about forcing with a Souslin tree
that we will need repeatedly.
\begin{lemma} Suppose that
$S$ is a Souslin tree and $S\in M$ for some countable
elementary\label{basiclemma}
submodel of any $H(\theta)$ ($\theta \geq {\omega_2}$).
If $\dot x, \dot X \in M$ are Souslin names, and $s\in S\setminus M$,
then there is an $s'< s$ with $s'\in M$ such that
\begin{enumerate}
\item $s\Vdash \dot X = \emptyset$ if and only if $s' \Vdash \dot X
=\emptyset$,
\item $s\Vdash \dot x \in \dot X$ if and only if $s'\Vdash \dot x\in
\dot X$.
\end{enumerate}
\end{lemma}
\begin{proof}
The second item follows from the first (by simply considering
the set $\dot X \cap \{\dot x\}$) so we consider any $\dot X$ in $M$.
Since $S$ is a ccc forcing and the set of conditions that decide
the statement ``$\dot X = \emptyset$'' is dense and open, there is a
$\gamma\in M\cap\omega_1$ such that each element of $S_\gamma$ decides
this statement. Therefore $s\restriction \gamma$ decides the statement
and, since $s$ is a stronger condition than $s\restriction \gamma$,
they assign the same truth value to the statement.
\end{proof}
Note, for example, Lemma \ref{basiclemma} can be used to show that
if $\dot E\in M$ is an $S$-name of a subset of
$\omega_1$
and $s\Vdash M\cap \omega_1\in \dot E$,
then $s\Vdash \dot E $ is stationary.
To see this we can let $\dot X$ denote the set of
(ground model)
cub subsets of $\omega_1$ that are disjoint from $\dot E$.
Then, if $s\Vdash M\cap \omega_1 \in \dot E$, we have
that for all
cub $C$ in $M$, $s\Vdash C\cap \dot E$ is not empty.
So, if $s'<s$ is in $M$, we have that $s'$ forces
that $\dot X$ is empty, and $\dot E$ is stationary.
\begin{proposition} Assume PFA(S) then $S$ forces
that $\mathbf{P}_{22}$ holds.
\end{proposition}
\begin{proof}
Let $\dot {\mathcal I}$ be the $S$-name for a P-ideal on a stationary
subset $B$ of $\omega_1$
and assume that some $s_0\in S$ forces that $\dot {\mathcal I}\cap
[E]^{\aleph_0}\neq \emptyset$ for all stationary sets $E$.
If $s_0$ also forces that $\dot {\mathcal I}$ is a counterexample to
$\mathbf{P}_{22}$, then using that $S$ is homogeneous
and the forcing maximum principle, we can assume
that $s_0$ is the root of $S$ and just show that $\dot {\mathcal I}$
is not a counterexample.
Fix any well-ordering
$\prec$ of $H(\aleph_2)$.
\begin{claim}
For each countable elementary submodel
$M$ of $(H(\aleph_2),\prec)$ and each $s \in S_{M\cap \omega_1}$,
there is a set $a(s,M)$ such that
$s\Vdash a(s,M)\in \dot {\mathcal
I}$ and $s\Vdash a \subset^* a(s,M)$ for all $a\in
M\cap \dot{\mathcal M}$.
\end{claim}
Proof of Claim 1: Since $s$ forces that $\dot{\mathcal I}$ is a
P-ideal, there is a $\prec$-minimal name $\dot a$ such that
$1$ forces that each member of
$M\cap \dot{\mathcal I}$ is a subset mod finite of $\dot a$. Since $S$
is ccc, there is a countable maximal
antichain $\{ s_n : n\in \omega\}$ and a countable family
$\{ a_n : n\in \omega\}$ of countable subset of $\omega_1$ such
that, for each $n$, $s_n\Vdash \dot a =a_n $. Furthermore, $s$ forces
a value on each member of $M\cap \dot {\mathcal I}$. Let
$\mathcal J$ denote the countable family of sets forced by
$s$ to be members of $M\cap \dot{\mathcal I}$. Note that every member
of $\mathcal J$ is mod finite contained in every member of
$\{ a_n :n\in \omega\}$. We may choose $a(s,M)$ to be the
$\prec$-minimal set that splits this $(\omega,\omega)$-gap.
\bigskip
One change from \cite{Todorcevic} is that we
begin with a partition
$\mathcal E = \{ E_s : s\in S\}$ of
$\omega_1$ by stationary sets
so that, in addition, $ E_s\subset B$ for each
$s\in S$ other than the root $\emptyset$.
Thus $\bigcup \{ E_s : s\in S \setminus \{\emptyset\}\}$ contains
$\omega_1\setminus B$.
We also require that $\dom(s) < \delta$ for all
limit $\delta\in E_s$.
Then we
let $\mathcal P$ be the collection of all mappings of the form
$p:\mathcal M_p \rightarrow S$, where
\begin{enumerate}
\item $\mathcal M_p $ is a finite $\in$-chain of countable elementary
submodels of $(H(\aleph_2),\mathcal E, \prec)$
\item $M\in \mathcal M_p$ and $\delta = M\cap \omega_1\in E_s$
implies $s< p(M)\in S_{\delta}$,
\item $M\in N\in \mathcal M_p$ implies $a(p(M),M)\in N$.
\end{enumerate}
We let $p\leq q$ if,
\begin{enumerate}
\addtocounter{enumi}{3}
\item $\mathcal M_p\supset \mathcal M_q$ and $q = p\restriction
\mathcal M_q$,
\item $
N\cap\omega_1\in a(q(M),M)$ whenever $N\cap\omega_1\notin E_{\emptyset}$,
$p(N) < q(M)$
with $M\in N\in \mathcal M_q$,
and $M\in \mathcal M_p\setminus
\mathcal M_q$.
\end{enumerate}
In order to apply PFA(S) to $\mathcal P$,
we have to show that $\mathcal P$ is a proper poset that preserves
that $S$ is Souslin. Once we do, we let $\mathcal G$ be a filter on
$\mathcal P$ that meets sufficiently many (no more than $\omega_1$)
dense subsets to
ensure that there is a cub $C\subset \omega_1$ such that for each
$\delta\in C$, there is a $p_\delta\in \mathcal G$ and an $M_\delta\in
\mathcal M_{p_\delta}$ with $M_\delta\cap \omega_1 = \delta$. The
role of the family $\mathcal E$ is to ensure the next Claim.
\begin{claim}
Each $s_0\in S$ forces that
the $S$-name $\dot E = \{ \delta\in B : p_\delta(M_\delta) \in \dot g\}$ is
a stationary subset of $\omega_1$, where $\dot g$
is the $S$-name of the generic branch through $S$.
\end{claim}
Proof of Claim 2: It suffices to show
that $s_0$ does not force that $\dot E$ is not stationary
by finding an extension that forces $\dot E$ is stationary.
Choose any $\delta \in C\cap E_{s_0}$. We have that $s = p_\delta(M_\delta)$
forces that $\delta = M_\delta\cap \omega_1$ is in $\dot E$.
Also, since $\delta \in E_{s_0}$, we have,
from the definition of $\mathcal P$, that $s_0 < s$.
By Lemma \ref{basiclemma}, as explained in
the discussion immediately following it, we have that
$s$ forces that $\dot E$ is stationary.
\bigskip
\begin{claim} Each $s\in S$ forces that
$[\dot E]^{\aleph_0} \subset \dot{\mathcal I}$, where
$\dot E$ is defined in Claim 2.
\end{claim}
Proof of Claim 3: It suffices to show that if $\gamma\in \omega_1$
and $s\in S_\gamma$, then $s\Vdash \dot E\cap \gamma \in \dot{\mathcal
I}$. Recall that there is a $\delta > \gamma$ such that
$s<p_\delta(M_\delta)$. By the definition of the ordering on
$\mathcal P$ (item (5)) we have that $\{ \gamma \in \dot E :
p_\gamma(M_\gamma) < p_\delta(M_\delta) \ \mbox{and}\
M_\gamma\notin \mathcal M_{p_\delta}\}$ is contained in
$a( p(M_\delta), M_\delta)$. Therefore, $p(M_\delta)$ forces
that $\dot E\cap \delta \in \mathcal I$.
\bigskip
We finish the proof of the Proposition by proving that $S\times \mathcal
P$ is proper. Let $M$ be any countable elementary submodel of
$H(\kappa)$ for some regular $\kappa > \omega_2$. We show that any pair
$(s^\dagger, q)$ where $s^\dagger\in S\setminus M$ and $M\cap
H(\aleph_2)\in \mathcal M_{p_0}$ is an $M$-generic condition for
$S\times \mathcal
P$. Consider any dense open set $D$ of $S\times \mathcal P$ that is a
member of $M$. By extending the condition $(s^\dagger,q)$ we can assume
that $(s^\dagger,q)$ is in $D$ and that there is some countable
elementary submodel of $H(\kappa)$ containing $q$ but not $s^\dagger$.
It is useful to regard $D$ as an $S$-name $\dot D$
of a dense open subset of
$\mathcal P$ in the sense that if $(t,p)\in D$, then we interpret
this as $t\Vdash p\in \dot D$.
It is evident from conditions (4) and (5) of the definition of
$\mathcal P$ that $q_0 = q\restriction M\in M$ and
that $q$ is an extension of $q_0$. Let $\delta = M\cap \omega_1$.
Let $\{ M_1, \ldots , M_{\ell}\}$
be an increasing enumeration of $\mathcal M_q\setminus M$. Of course
$M_1 = M\cap H(\aleph_2)$. Let $\{ s_0, \ldots , s_m\}$ be any
one-to-one list of the set $\{ s^\dagger \restriction \delta,
q(M_1)\restriction \delta, \ldots , q(M_{\ell})\restriction
\delta\}$
so that $s_0 = s^\dagger\restriction \delta$. For each $1\leq j\leq \ell$,
let $m_j $ denote the value such that $s_{m_j} = q(M_j)\restriction
\delta$.
Let $J$ denote those $1\leq j\leq\ell$ such that
$q(M_j)\restriction [\delta, M_j\cap \omega_1) \subset s^\dagger$.
To avoid trivialities, we can assume that we extended $(s^\dagger, q)$
if necessary,
so as to have that $J$ is not empty.
Since $S$ is a coherent Souslin tree, there is a
$\bar\delta\in M$ such that $s_0\restriction [\bar\delta, \delta) =
s_i \restriction [\bar\delta, \delta)$ for each $i\leq m$. By
increasing $\bar\delta$ we can also ensure that $\bar M\cap \omega_1 <
\bar \delta$ for each $\bar M\in \mathcal M_{q_0}$.
Let $\bar s_i = s_i\restriction \bar \delta$ for $i\leq m$, and
notice that $\{ \bar s_0,\ldots, \bar s_m\}\in M_0$.
For each $s\in S$ with $\bar\delta\leq \dom(s)$, let
$\bar s_0 \oplus s$ denote the function
$\bar s_0\cup s\restriction [\bar \delta, \dom(s))$; since $S$ is a
coherent Souslin tree $\bar s\oplus s\in S$.
Note that $J = \{ j <\ell : \bar s_0\oplus p(M_j) < s^\dagger \}$.
Also, define $J_B$ to be the set $\{ j\in J : M_j\cap \omega_1\notin
E_\emptyset\}$.
Say that $(t,p)\in D$ is \underbar{like} $(s^\dagger,q)$ providing
\begin{enumerate}
\item there is a $M^p_0\in \mathcal M_p$ such that $\bar\delta \in
M^p_0$ and $q_0 = p \restriction M^p_0$,
\item $\mathcal M_p\setminus M^p_0 $ has size $\ell$, enumerated as
$\{ M^p_0,\ldots M^p_{\ell-1}\}$ in increasing order,
\item $\bar s_{i_j}< p(M^p_j)$ for $j<\ell$
\item $J = \{ j < \ell : \bar s_0 \oplus p(M^p_j) < t \}$,
\item $J_B = \{ j\in J : M^p_j\cap \omega_1 \notin E_{\emptyset}\}$.
\end{enumerate}
Our proof that
$S\times \mathcal P$ is proper will depend on finding some
$(t,p)\in D\cap M$ that is \underbar{like} $(s^\dagger,q)$
and, in addition, is compatible with $(t_0,q)$. Of course this
requires that $t< s^\dagger $,
but what else? Since $\mathcal M_p\in M_0$
and $p<q_0$ we automatically have that $\mathcal M_p\cup \mathcal M_q$
is an $\in $-chain. The most difficult (and remaining)
requirement is to ensure
that if $p(M^p_j) < q(M_k)$ then $M^p_j\cap \omega_1 $
must be in $
a(q(M_k),M_k)$ if $M^p_j\notin E_{\emptyset}$.
Interestingly, the values of $1\leq j\leq \ell$ that
we will have to worry about are exactly those values in $J_B$
(in most proofs it would be all values of $J$). This is
because we must have that $p(M^p_j) < s_{i_k}$ and so $\bar s_0 \oplus
p(M^p_j) < s_0 < s^\dagger$. Since also, $t<s^\dagger$ and $t\in M_0$, we
have that $\bar s_0\oplus p(M^p_j) < t$, which is the requirement
that $j\in J$. One frequently troublesome aspect to these proofs is
that the values of $k$ for which $p(M^p_j) < q(M_k)$ will be
all $k$ such that $i_k = i_j$, not just values of $k$ in $J$.
For easier reference in the remaining proof, let
$a_k = a(q(M_k) , M_k)$ for $1\leq k \leq m$.
The set $\mathbf{L}\subset D$ consisting of those pairs $(t,p)$ that
are
\underbar{like} $(s^\dagger,q)$ is an element of $M$.
For each
$(t,p)\in \mathbf{L}$, let $T_{t,p} = \langle t_0, t_1, \ldots, t_\ell\rangle$
be a re-naming of $\langle t, p(M^p_1), p(M^p_2),\ldots,
p(M^p_\ell)\rangle$. Let $\mathcal T(\mathbf {L})$ denote the
set $\{ T_{t,p} : (t,p)\in \mathbf{L}\}$, and for each $1\leq j\leq
m$, let $\mathcal T(\mathbf {L})_j =
\{ \vec t \restriction j : \vec t
= \langle t_0, t_1, \ldots, t_\ell\rangle \in \mathcal
T(\mathbf{L})\}$.
Of course $\mathcal T(\mathbf{L})_\ell $ is equal to $\mathcal
T(\mathbf{L})$. Since $D$ is an open subset of
$S\times\mathcal P$, let us note that if $
\langle t_0, t_1, \ldots, t_{j-1}\rangle \in \mathcal
T(\mathbf{L})_j$,
then $ \langle \bar t_0, t_1, \ldots, t_{j-1}\rangle \in \mathcal
T(\mathbf{L})_j$ for all $\bar t_0 > t_0$.
Now we want to use $\mathcal T(\mathbf {L})$ to define an
$S$-name of a subset of $[\omega_1]^{\leq\ell}$.
For $\vec t = \langle t_0, t_1,\ldots, t_{j-1}\rangle
\in \mathcal T(\mathbf{L})_j$ ($1\leq j\leq \ell$),
let $\Delta_{\vec t}$ be the sequence $ \langle \delta_1, \ldots,
\delta_{j-1}\rangle$ where $\delta_i = \dom(t_i)$.
We define $\dot F_\ell$ to be the $S$-name consisting of all
pairs $(t_0, \langle \delta_1, \ldots , \delta_\ell \rangle)$
for which there is a $\vec t = \langle t_0, \ldots, t_\ell\rangle$
in $\mathcal T(\mathbf{L})$ such that $\Delta_{\vec t} =
\langle \delta_1, \ldots , \delta_\ell\rangle$. In saying
that $\dot F_\ell$ is an $S$-name we are adopting
the standard abuse of notation that
an element of the ground model can be used as an $S$-name for itself.
By reverse
induction on $\ell > k \geq 1$, we define $\dot F_k$.
Having defined $\dot F_{k+1}$, we define $\dot F_k$.
If $k+1\notin J_B$, then $\dot F_k = \dot F_{k+1}$.
If $j = k+1$ is in $J_B$, then
$(t_0, \langle \delta_1, \ldots, \delta_{k} \rangle)$
is in $\dot F_k$ providing $t_0$ forces that the set
$$\dot F_j( \langle \delta_1, \ldots, \delta_{k} \rangle) =
\{ \gamma : (\exists \bar t_0)(\exists \vec{\delta}~)~~
(t_0,\vec \delta~) \in \dot F_j \ \mbox{and}\
\langle \delta_1, \ldots, \delta_k, \gamma\rangle =
\vec\delta\restriction j \}$$ is stationary.
The next, somewhat standard, step is to prove that,
for each $k<\ell$ with $k+1\in J$,
$s^\dagger \Vdash \Delta_{T_{s^\dagger,q}} \restriction k \in
\dot F_k$.
Again, this is by reverse induction on $\ell>k\geq 0$.
Let $\vec\gamma = \Delta_{T_{s^\dagger, q}} = \langle \gamma_1,\ldots ,
\gamma_\ell\rangle$. Certainly, $s^\dagger \Vdash
\vec\gamma \in \dot F_\ell$. We again take note
of the fact that $\dot F_k\in M_0$ for each $0\leq k\leq \ell$.
Let $J_B=\{j_1, \ldots , j_{\bar \ell}\}$
be listed in increasing order. For $j_{\bar \ell}\leq k\leq \ell$,
we have that $s^\dagger\Vdash \dot F_k = \dot F_\ell$.
Now let $j = k+1 = j_{\bar \ell}$ and observe that
$\dot F_{j}(\vec \gamma\restriction k)$ is a member of the model
$M_{j}$, and that $\gamma_{j} = M_{j}\cap \omega_1 $
is forced by $s^\dagger$ to be an element of
$\dot F_{j}(\vec \gamma\restriction k)$. We show that
this means that $s^\dagger$
forces that
$\dot F_{j}(\vec \gamma\restriction k)$ is stationary.
Within $M_j$, there is a maximal antichain (in fact a level) of $S$
with the property that each member of the antichain decides
whether or not
$\dot F_{j}(\vec \gamma\restriction k)$ is stationary. For each such
node that decides that it is not stationary,
there is a cub in $M_j$ that is forced to be disjoint.
Since $\gamma_j$ is in every cub from $M_j$ and since
$s^\dagger$ forces that $\gamma_j$ is in $\dot F_j(\vec
\gamma\restriction k)$, we have that it is forced to be stationary.
This completes the inductive step that $s^\dagger$ forces
that $\vec\gamma\restriction k$ is in $\dot F_k$.
To complete the proof, we work our way back up
from $\min(J_B)$ to $\max (J_B)$
in order to pick a suitable $(t,p)\in D\cap M$ that is compatible with
$(s^\dagger, q)$. Recall that the main requirement, once we know
that $(t,p)\in \mathbf{L}\cap M$, is to have that
$\delta_j \in a_k$ for each $j\in J_B$
and each $1\leq k\leq \ell$ with $i_j = i_k$, where
$\Delta_{T_{t,p}} = \langle \delta_1, \ldots, \delta_\ell\rangle$.
We begin with $j_0 = \min(J)$, and we note that $s^\dagger$ forces
that $\dot F_{j_0-1} \in M_0$ is non-empty. By Lemma
\ref{basiclemma},
there is an $t_0\in M\cap S$ with $t_0<s^\dagger$
that also forces $\dot F_{j_0-1}$ is not empty. By elementarity,
there is a sequence $\vec \delta_0\in M_0$ such that
$t_0 \Vdash \vec\delta_0 \in \dot F_{j_0-1}$.
By definition, $t_0 \Vdash \dot F_{j_0}(\vec \delta_0) $ is
stationary.
Now we use our assumptions on $\dot {\mathcal I}$
in order to find a member of $\dot F_{j_0}(\vec\delta_0)$
that is in $a_{j_0}$.
This next step can seem a bit like sleight of hand. We have
that $t_0\Vdash \dot F_{j_0}(\vec \delta_0)$ is stationary,
and so there is an extension (in $M_0$) of $t_0$ and an infinite
set $a$ that is forced to be contained in $\dot F_{j_0}(\vec \delta_0)$
and to be a member of $\dot {\mathcal I}$. However,
$t_0$ may be
incomparable with $s_{i_{j_0}}$ and so $a$ is of no help in
choosing a suitable element of $a_{j_0}$. The solution is to use that
$S$ is coherent. Let $g$ be a generic filter for $S$ with
$s^\dagger\in g$. Since $S$ is coherent, the collection
$s_{i_{j_0}}\oplus g = \{ s\in S : (\exists t\in g)~~ s < (s_{i_{j_0}} \oplus
t )\} $ is also a generic filter for $S$ since it is an $\omega_1$-branch.
The ideal $\mathcal I(s_{i_{j_0}}) $ we get by interpreting
the name $\dot {\mathcal I}$ using the filter $s_{i_{j_0}}\oplus B$,
is a P-ideal satisfying that $[E]^{\aleph_0}\cap \mathcal I(s_{i_{j_0}})$
is non-empty for all stationary sets $E$. Also, the set
$E = \val_{g}(\dot F_{j_0}(\vec \delta_0))$ is a stationary set.
By elementarity, there is an infinite set $a\in M_0$
such that $a\in [E]^{\aleph_0}$ and $a\in \mathcal I(s_{i_{j_0}})$.
Again by elementarity, and Lemma \ref{basiclemma},
there is a condition $t_1\in M\cap g$ extending
$t_0$ and
satisfying that $t_1 \Vdash a \subset
\dot F_{j_0}(\vec\delta_0)$
and $s_{i_{j_0}}\oplus t_1 \Vdash a \in \dot {\mathcal I}$.
Let us note that $a\subset a_k$ for each $1\leq k\leq m$
such that $i_{j_0} = i_k$. Therefore, we may choose
a $\delta_{j_0}
\in a\cap \bigcap \{ a_k : i_k = i_{j_0}\}$. Next choose
any sequence $\vec\delta_1\in M_0$
such that, by further extending $t_1$, we have that
$t_1 \Vdash \vec\delta_1\in \dot F_{j_0}$ and
witnesses that $\delta_{j_0}\in \dot F_{j_0}(\vec \delta_0)$.
We proceed in the same way to choose $\delta_{j_1}$ and
an extension $t_2$ of $t_1$ so that $\delta_{j_1} \in a_k$
for each $k$ with $i_k = i_{j_1}$ and so that $t_2$ forces
that there is a $\vec\delta_2\in \dot F_{j_2}$ witnessing
that $\delta_{j_1}\in \dot F_{j_2}(\vec\delta_1)$. Proceeding
in this way we succeed in choosing $t_{\bar \ell}$ in
$M_0$ with $t_{\bar\ell}\subset
s^\dagger$ and a sequence $\vec\delta_{\bar\ell}$
satisfying that there is a $p\in M_0$
such that $(t_{\bar\ell}, p) \in \mathbf {L}$,
$\Delta_{T_{t,p}} = \langle \delta_1, \ldots , \delta_\ell\rangle
= \vec\delta_{\bar\ell}$,
and that
$\delta_{j_n} \in a_k$
for each $1\leq n \leq \bar \ell$ and
$1\leq k\leq \ell$ such that
$i_k = i_{j_n}$. Of course this means that $(t_{\bar \ell}, p)\in D\cap M$
and $(t_{\bar \ell}, p) \not\perp (s^\dagger, q)$ as required.
\end{proof}
\section{on $\mathbf{PPI}^+$}
This first result is a reformulation of a classic result of
Sapirovskii.
\begin{lemma} Assume that $X$ is a\label{itsomega1}
sequentially compact non-compact space.
Then either $X$
has a countable subset with non-compact closure
or $X$ has an $\aleph_1$-sized subset $E$
and an open set $\mathcal W$ containing the sequential closure
of $E$ and such that $E$
has no complete accumulation point in $\mathcal W$.
\end{lemma}
\begin{proof}
We may as well assume that countable subsets of $X$ have
compact closure. Since $X$ is sequentially compact
and not compact, it is not Lindel\"of.
Let $\mathcal U$ be any open cover of $X$ that has no
countable subcover and satisfies that the closure of each
member of $\mathcal U$ is contained in some other member
of $\mathcal U$. We begin an inductive construction by choosing
any countable subset $\mathcal U_0$ of $ \mathcal U$ and
any point $x_0 \in X\setminus \bigcup \mathcal U_0$. Suppose
$\lambda < \omega_1$, and that we have chosen,
for each $\alpha < \lambda$, a countable collection
$\mathcal U_\alpha\subset \mathcal U$ and a point $x_\alpha
\in X \setminus \bigcup \mathcal U_\alpha$,
so that $\overline{\{x_\beta: \beta <\alpha\}}\cup
\bigcup_{\beta<\alpha} \mathcal U_\beta \subset \mathcal U_\alpha$.
Since $\mathcal U$ has no countable subcover, this induction continues
for $\omega_1$-many steps. We let $E$ be the
sequence $\{x_\beta : \beta\in \omega_1\}$ and let
$\mathcal W$
be the union of the collection $\bigcup\{ \mathcal U_\alpha :
\alpha \in \omega_1\}$.
By construction we have that the closure
of every countable subset of $E$ is
contained in $\mathcal W$. But also, for each $y\in \mathcal W$,
we have
that there is an $\alpha\in \omega_1$ such that
$y\in \bigcup\mathcal U_\alpha$ while
$\bigcup\mathcal U_\alpha\cap E = \{ x_\beta : \beta < \alpha\}$ is
countable.
\end{proof}
For the remainder of the section we have an
$S$-name of a sequentially compact non-compact
space $\dot X$ which we may assume has base set $\theta$.
According to Lemma \ref{itsomega1}, we will assume
that either
$\omega\subset \theta$ is forced (by 1) to be dense in $X$,
or, that
the sequential closure
of the points $\omega_1\subset \theta$ are forced (by 1)
to have no
complete accumulation point in some open neighborhood
$\mathcal W$. In particular then, the sequential closure of
$\omega_1$ itself
contains no complete accumulation point of $\omega_1$.
In the first case, our application
of PFA(S) will be simplified if we use
the method sometimes called the
\textit{cardinal collapsing trick\/}.
This is to show that we may again assume that we have
an uncountable set, denoted $\omega_1$, so that the
sequential closure is contained in an open set $\mathcal W$
in which $\omega_1$ has no complete accumulation point.
It will be easier to remember if we call this the
separable case.
The simple countably closed poset $2^{<\omega_1}$ is $S$-preserving.
We will work, for the separable case,
in the forcing extension by $2^{<\omega_1}$ -- a model in which CH
holds. Just as we have in Lemma \ref{itsomega1}, we would like
to show that there is an uncountable set $E$ and an open set
$\mathcal W$
so that the sequential closure of $E$ is contained in
$\mathcal W$ while
having no complete accumulation points in $\mathcal W$.
We certainly
have that the forcing $2^{<\omega_1}$ preserves that $\dot X$
is forced by $S$ to be sequentially compact and not compact.
We briefly work in the forcing extension by $2^{<\omega_1}\times S$.
Let $X$ denote the space obtained from the name $\dot X$.
If the base set $\theta$ for $ X$
is equal to $\mathfrak c$ then, since it is forced to be countably
compact, it is forced that $ X$ has an uncountable set with no
complete accumulation point at all. On the other hand if
$ X$ has cardinality greater than $\mathfrak c$, we can fix
any point $ z$ of $X$ that is not in the sequential closure of
$\omega$. Since $ X$ is regular and $\omega$ is dense,
the point $ z$ has
character $\omega_1$. Let $\{ W_\alpha : \alpha \in \omega_1\}$
enumerate a neighborhood base for $z$ satisfying that the closure
of $W_{\alpha+1}$ is contained in $W_\alpha$ for each $\alpha\in
\omega_1$. For each $\alpha$, we may choose a point $x_\alpha$ from
the sequential closure of $\omega$ so that $x_\alpha$ is in $W_\beta$
for all $\beta \leq \alpha$. Now we have that the uncountable set $E =
\{ x_\alpha : \alpha\in \omega_1\}$ satisfies that its sequential
closure is contained in the open set $\mathcal W
= X\setminus \{z\}$ and has no
complete accumulation point in $\mathcal W$.
\bigskip
Next we choose an assignment of $S$-names of neighborhoods $\{ \dot
U(x,n) : x\in \theta, n\in \omega \}$, each of which is forced
to have closure contained in $\mathcal W$.
We may assume that 1 forces that these
are regular descending and that $\omega_1\cap \dot U(x,0)$ is countable
for all $x\in \theta$. These are chosen in the generic extension
by $2^{<\omega_1}$ in the separable case.
If we are also assuming that $\dot X$ is forced to be first countable,
then we assume that $\{ \dot U(x,n) : n\in \omega\}$ is forced
to form a neighborhood base for $x$.
\subsection{the sequential structure}
Since $S$ is ccc, it follows that if $\{ \dot x_n : n \in \omega\}$ is
a sequence of $S$-names and $1\forces \dot x_n \in X$ for each $n$,
then there is an infinite $L\subset \omega$ such that $1\forces
\{\dot x_n : n \in L\}$ is a converging sequence in $X$. To see this,
recursively choose a
mod finite descending sequence $\{ L_\alpha :\alpha \in \gamma\}$
and conditions $\{ s_\alpha : \alpha\in \gamma\}$ satisfying that
$s_\alpha$ forces that $\{ \dot x_n : n\in L_\beta \}$ (for
$\beta < \alpha$) is not converging, while
$\{ \dot x_n : n\in L_\alpha \}$ is. Since the family
$\{ s_\alpha : \alpha \in \gamma\}$ is an antichain, this process
must end.
\begin{definition} Say that a sequence $\{ \dot x_n : n\in L \}$ is an
$S$-converging sequence in $\dot X$
providing $1\Vdash \{\dot x_n : n\in L\}$
is a converging sequence (which includes, for example, constant
sequences).
\end{definition}
There is a well-known space in the study of sequential spaces, namely
the space $S_\omega$ from \cite{ArhFran}. This is the strongest
sequential topology on the set of finite sequences of integers,
$\omega^{<\omega}$, in which, for each $t\in \omega^{<\omega}$, the
set of immediate successors, $\{ t^\frown n : n\in \omega\}$,
converges to $t$. If $T$ is any subtree of $\omega^{<\omega}$, we will
consider $T$ to be topologized as a subspace of $S_\omega$. As
usual, for $t\in T$, $T_t$ will denote the subtree with root $t$ and
consisting of all $t'\in T$ which are comparable with $t$.
Also for $t\in T$, let $T_t^+$ denote the tree
$\{ t' \in \omega^{<\omega} : t^\frown t' \in T_t\}$
i.e. the canonically isomorphic tree with root $\emptyset$.
Of particular use will be those
$T \subset \omega^{<\omega}$ that are
well-founded (that is, contain no infinite branch).
Let $\mathbf{WF}$ denote those downward closed
well-founded trees $T$ with the property that every branching
node has a full set of immediate successors.
Such a tree will have a root, $\operatorname{root}(T)$
(which need not be the root of $\omega^{<\omega}$)
which is either the minimal branching node or,
if there are no branching nodes, the maximum member
of $T$.
When discussing the topology on $T\in \mathbf{WF}$ we ignore
the nodes strictly below the root of $T$.
The meaning
of the rank of $T$ will really be the rank of $T_t$ where $t$ is the
root of $T$. We use $\operatorname{rk}(T)$ to denote the ordinal $\alpha\in
\omega_1$ which is the rank of $T$. If $t\in T$ is a maximal node,
then $\operatorname{rk}(T_t) = 0$, and if
$\operatorname{root}(T)\subset t\in T$, then
$\operatorname{rk}(T_t) = \sup \{ \operatorname{rk}(T_{t'})+1
: t < t'\in T_t\}$.
We let $\mathbf{WF}(\alpha) = \{ T\in \mathbf{WF} :
\operatorname{rk}(T) \leq \alpha\}$ and
$\mathbf{WF}({<}\alpha) = \bigcup_{\beta<\alpha} \mathbf{WF}(\beta)$.
If we have a Hausdorff space $X$ on a base set
containing the set $ \omega_1$ and we have a point
$x$ in the sequential closure of $\omega_1$, then
there is a $T\in \mathbf{WF}$ and a function $y$
from $\max(T)$ into $\omega_1$ such that there is a continuous
extension of $y$ to all of $T$
such that $y(\operatorname{root}(T)) = x$
(it does not matter what value $y(t)$ takes for
$t<\operatorname{root}(T)$).
Since our space $\dot X$ is forced to be sequentially compact, we will
be working with points in the sequential closure of $\omega_1$.
In fact, we will only work with such function pairs
$y, T$ that are forced by 1 to extend continuously to all of $T$.
The goal is to try to make choices of points in $\dot X$ that
are, in a strong sense, not dependent on the generic filter for $S$.
For each $\alpha\in \omega_1$, let $\mathbf{Y}_\alpha$
denote the set of all functions $y$ into $\omega_1$
where $\dom(y)$ is the set of all maximal nodes of some $T\in
\mathbf{WF}(\alpha)$.
We put $y\in \mathbf{Y}_\alpha$
in $Y_\alpha$ providing $1$ forces
that $y$ extends continuously to all of $T_y$ as a function into
$\dot X$. We let $Y=\bigcup_{\alpha\in \omega_1} Y_\alpha$ and
for $y\in Y$, we will
abuse notation by letting $y$ also denote the name of the
unique continuous extension of $y$ to all of
$\{ t\in T_y : \operatorname{root}(T) \subset t\}$.
More precisely, if needed, for each $t\in T_y\setminus \max(T)$ with
$\operatorname{root}(T)\subset t$, $y(t)$ can be used to denote
the name that has the form $\{ (s,\xi_s) : s\in S_{\gamma}\}$
where $\gamma$ is minimal such that each $s$ in $S_\gamma$
decides the value of $y(t')$ for each $t\leq t'\in T_y$
in the continuous extension of $y$
and, of course, $s$ forces $y(t) = \xi_s\in \theta$ for each $s\in S_\gamma$.
The minimality of $\gamma$ makes this choice canonical.
Thus for $y\in Y$ and $\operatorname{root}(T_y)\subset t\notin
\max(T_y)$,
the sequence $\{ y({t^\frown n}) : n\in \omega \}$ is an $S$-converging
sequence that is forced to converge to $y(t)$. Note also that if
$y\in Y$ and $t\in T_y$, then $y\restriction (T_y)_t \in Y$.
\begin{definition} Say that
$y_1$ and $y_2$ in ${Y}$ are equivalent, denoted $y_1\approx
y_2$, providing $T_{y_1}^+ = T_{y_2}^+$, and for each maximal $t\in
T_{y_1}^+$, $y_1(\operatorname{root}(T_{y_1})^\frown t) $
equals $y_2(\operatorname{root}(T_{y_2})^\frown t) $.
\end{definition}
Clearly if $y_1\approx y_2$, then $y_1(\operatorname{root}(T_{y_1}))$
is the same name as $y_2(\operatorname{root}(T_{y_2}))$.
Now that we have identified our structure
$Y$ we extend the notion to define a closure
operator on any given finite power of $Y$ which will help
us understand points in the sequential closure of $\omega_1$ in $\dot
X$. If $y\in Y$, we use $e(y)$ as an alternate notation
for $y(\operatorname{root}(T_y))$. Similarly, if $\vec y\in Y^n$ (for
some $n\in \omega$), we will use $e(\vec y)$ to denote the point
$\langle e(\vec y_0), e(\vec y_1), \ldots, e(\vec y_{n-1})\rangle$.
\begin{definition}
For\label{seql}
each integer $n>0$, and subset $B$ of $Y^n$ we similarly define
the hierarchy $\{ B^{(\alpha)} : \alpha\in \omega_1\}$
by recursion. In addition, we again (recursively) view
each $ \vec b\in B^{(\alpha)}$ as naming a point in $X^n$. The
set $B$ will equal $B^{(0)}$. Naturally the point $e(\vec b)$
named
is the point of $X^n$ named coordinatewise by $\vec b$.
For limit $\alpha$, $B^{(\alpha)}$
(which could also be denoted as $B^{(<\alpha)}$) will equal
$\bigcup_{\beta<\alpha} B^{(\beta)}$. The members of $B^{(\alpha+1)}$
for any $\alpha$, will consist of the union of
$B^{(\alpha)}$ together with all
those $\vec b = \langle y_i : i\in n\rangle \in (Y_{\alpha+1})^n$ such
that there is
a sequence $\langle \vec b_k : k\in \omega \rangle$ so that
\begin{enumerate}
\item for each $k\in \omega$, $\vec b_k$ is a member of $B^{(\alpha)}\cap
(Y_\alpha)^n$,
\item for each $i\in n$ and $k\in \omega$, $(\vec b_k)_i\in Y$ is equivalent to
$y_i\restriction (T_{y_i})_{t_i^\frown k}$, where $t_i$ is the root
of $T_{y_i}$.
\end{enumerate}
When $\vec b$ is constructed from a sequence $\{ \vec b_k : k\in \omega\}$
as in this construction, we can abbreviate this by saying that
$\{ \vec b_k : k\in \omega\}$ $Y$-converges to $\vec b$. Also if
we say that $\{ \vec b_k : k\in L\}$ Y-converges to $\vec b$ for some
infinite set $L$, we just mean by a simple re-enumeration of
$\{ \vec b_k : k\in L\}$.
\end{definition}
For $n>1$ we may view $Y^n$ as an $S$-sequential structure
and for any $A\subset Y^n$, we say that $A^{(\omega_1)}$ is the
sequential closure and is sequentially closed.
Notice that this $S$-sequential structure on $Y^n$ is defined in the
ground model.
The next lemma should be obvious.
\begin{lemma} For each $A\subset Y$, $1$ forces that $e[
A^{(\omega_1)}]$
is a sequentially compact subset of $X$.
\end{lemma}
\begin{definition} For each $S$-name $\dot A$ and $s\forces \dot A\subset
Y^n$, we define
the $S$-name $(\dot A)^{(\omega_1)}$ according to the
property that for each $s<t$ and $t\forces \vec y\in (\dot
A)^{(<\omega_1)}$, there is a countable $B\subset Y^n$ such that
$t\forces B\subset \dot A$ and $\vec y\in B^{(<\omega_1)}$.
\end{definition}
For an $S$-name $\dot A$ and $s\forces \dot A\subset Y^n$, we will
also interpret
$e[(\dot A)^{(\omega_1)}]$
in the forcing extension in the natural way
as a subset of $\dot X^n$. This may need some further
clarification.
\begin{lemma}
Suppose that $\vec y$ is a member of $B^{(\alpha)}$ for some
$B\subset Y^n$ and some $\alpha<\omega_1$.
Also suppose that $\{ s_i : i <\ell \}\subset S$ and that $\dot W$ is
an S-name for a
neighborhood\label{cohere} of
$e({\vec y}\,)$. Then there is a $\vec b \in B$ such that for each $i<\ell$,
there is an $s_i' \supset s_i$ forcing that $e({\vec b}\,)\in \dot W$.
\end{lemma}
\begin{proof}
We may suppose that $B$ is a countable set and we may proceed
by induction on $\alpha$. Let $y_i $ be equal to $\vec y_i$ for
$i<n$, and let $t_i$ denote the root of $T_{y_i}$.
By the definition of $B^{(\alpha)}$,
there is a sequence
$\langle \vec y_k : k\in \omega\rangle$
with the property that $\vec y_k\in B^{(<\alpha)}$ for each $k$,
such that, for each $i<n$ and each $k\in \omega$,
$(\vec y_i)\restriction (T_{y_i})_{t_i^\frown k}$ is
equivalent to $(\vec b_k)_i$.
For each $i<\ell$, choose $s_i' \supset s_i$
so that $s_i'$ forces a value, $W_i$, on
$\dot W\cap e[
B\cup\{\vec y_k : k\in \omega\}]$.
Since this sequence, $\{ e(\vec y_k ) : k\in \omega\}$
is assumed to be $S$-converging to a point
in $W_i$, there is a $k$ such that the point $e(\vec y_k)$
is in each $W_i$. Of course the result now follows by the induction
hypothesis.
\end{proof}
Note that the
members of ${Y}_0$ have a singleton domain
and for
each $\alpha\in \omega_1$, let $e^{-1}(\alpha)$ denote the
member
of $Y_0$ that sends the minimal tree to the singleton $\alpha$.
Our assumption that $\omega_1$ has no complete accumulation points
in its sequential closure implies that no point is a member
of every member of the family
$\{~ \left(\{e^{-1}(\alpha):\alpha > \delta \}\right)^{(\omega_1)} :
\delta\in \omega_1\}$. That is,
this family is a
free filter of $S$-sequentially
closed subsets of $Y$. By Zorn's Lemma, we
can extend it to a maximal free filter, $\mathcal F_0$,
of $S$-sequentially closed subsets of $Y$.
\subsection{A new idea in PFA(S)}
Now we discuss again the special forcing
properties that a coherent Souslin tree will have.
Assume that $g$ is (the) generic filter on $S$ viewed as a
cofinal branch. For each $s\in S$, $o(s)$ is the level
(order-type of domain) of $s$ in $S$. For any $t\in S$, define
$s \oplus t$ to be the function $s\cup t\restriction[o(s),o(t))$.
Of course when $o(t) \leq o(s)$, $s\oplus t$ is simply $s$.
One of the
properties of $S$ ensures that $s \oplus t\in S$ for all $s,t\in S$. We
similarly define $s \oplus g$ to be the branch $\{ s \oplus t : t\in
g\}$.
\begin{definition} Let $bS$ denote the set of $\omega_1$-branches of
$S$.
\end{definition}
\begin{lemma} In the extension $V[g]$, $bS = \{ s \oplus g : s\in
S\}$. Furthermore, for each $s\in S$, $V[s \oplus g] = V[g]$.
\end{lemma}
The filter $\mathcal F_0$ may not generate a maximal filter in the
extension
$V[g]$ and so we will have to extend it. Looking ahead to the PFA(S)
step, we would like (but probably can't have) this (name of)
extension to give
the same filter in $V[s \oplus g]$ as it does in $V[g]$. We adopt a new
approach. We will define a filter (of $S$-sequentially closed) subsets
of the product structure $Y^{bS}$. We try to make this filter somehow
symmetric.
We introduce some notational conventions. Let $S^{<\omega}$ denote
the set of finite tuples
$\langle s_i : i< n\rangle$ for which there is a
$\delta$ such that each $s_i\in S_\delta$. Our convention will be that
they are distinct elements. We let $\Pi_{\langle s_i : i<n\rangle}$ denote the
projection from $Y^{bS}$ to $Y^n$ (which we identify with
the product $Y^{ \{ s_i \oplus g : i< n\}}$).
\begin{definition} Suppose that $\dot A$ is an $S$-name of a subset
of $Y^n$ for some $n$, in particular, that some $s$ forces
this. Let
$s'$ be any other member of $S$ with $o(s')=o(s)$. We define a new
name $\dot A^{s}_{s'}$ (the $(s,s')$-transfer perhaps) which is
defined by the property that for all $\langle y_i \rangle_{i<n}\in
Y^n$ and $s<t\in S$ such that $t\forces \langle y_i \rangle_{i<n}
\in \dot A$, we have that $s' \oplus t \forces \langle y_i
\rangle_{i<n} \in \dot A^s_{s'}$.
\end{definition}
\begin{lemma} For any generic $g\subset S$,
$\val_{s \oplus g} (\dot A) = \val_{s' \oplus g} (\dot A^s_{s'})$.
\end{lemma}
\begin{theorem} There is a family $\mathcal F$ =
$\{
(s^\alpha,\{s^\alpha_i : i<n_\alpha\}, \dot F_\alpha ) : \alpha \in
\lambda\}$ where,
\begin{enumerate}
\item for each $\alpha\in \lambda$, $
\{ s^\alpha_i : i<
n_\alpha\}\in S^{<\omega}$, $s^\alpha\in S$, $ o(s^\alpha_0)\leq o(s^\alpha)$,
\item $\dot F_\alpha$ is an $S$-name such that $s^\alpha\Vdash \dot
F_\alpha = (\dot F_\alpha)^{(\omega_1)} \subset Y^{n_\alpha}$
\item for each $s\in S$ and $F\in \mathcal F_0$, $(s,\{s\}, \check F)\in
\mathcal F$,
\item for each $s\in S_{o(s^\alpha)}$, $(s,\{s^\alpha_i :i<
n_\alpha\}, (\dot F_\alpha)^{s^\alpha}_{s})\in \mathcal F$,
\item for each generic\label{previousitem}
$g\subset S$, the family
$ \{ \Pi_{\langle s^\alpha_i : i<n_\alpha\rangle}^{-1}( \val_{g}(\dot F_\alpha)) :
s^\alpha \in g\}$ is finitely directed;
we let $\dot {\mathcal F}_1$ be the $S$-name for the filter base it
generates,
\item for each generic $g\subset S$ and each $\langle s_i : i< n\rangle\in
S^{<\omega}$, the family \\
$\{ \val_g (\dot F_\alpha) : s^\alpha\in g \ \mbox{and}\
\{ s_i \oplus g : i< n\} = \{ s^\alpha_i \oplus g : i<n_\alpha\} \}$ is
a maximal filter on the family of $S$-sequentially closed subsets of
$Y^n$.
\end{enumerate}
\end{theorem}
\begin{proof}
Straightforward recursion or Zorn's Lemma argument over the family
of ``symmetric'' filters (those satisfying (1)-(5)).
\end{proof}
\begin{definition}
For any $\langle s_i : i<\ell\rangle\in S^{<\omega}$, let
$\dot {\mathcal F}_{\langle s_i:i<\ell\rangle}$ denote the filter on
$Y^{\ell}$ induced by $ \Pi_{\langle s_i : i<\ell\rangle}(\dot{\mathcal F}_1)$.
\end{definition}
\begin{definition} Let $\mathcal A$ denote the family\label{needA}
of all
$(s,\langle s_i : i< \ell\rangle, \dot A)$ satisfying that $o(s)\geq
o(s_0)$,
$\langle s_i :
i< \ell\rangle \in S^{<\omega}$, and
$s\Vdash
\dot A \in \dot
{\mathcal F}_{\langle s_i:i<\ell\rangle}^+$. As usual,
for a family $\mathcal G$ of set,
$\mathcal G^+$ denotes the family of sets that meet each member of
$\mathcal G$.
\end{definition}
\begin{lemma} For each $(s,\langle s_i: i<n\rangle, \dot A)\in \mathcal A$,
the object $(s,\langle s_i:i<n\rangle , \dot
A^{(\omega_1)})$ is in the list $\mathcal F$.
\end{lemma}
In this next Lemma it is crucial that there are no dots on the sequence
$\langle y^M(s) : s\in S_\delta\rangle$. The significance of there
being no dots is that, regardless of the generic $g\subset S$,
we will have that $e_g(y^M(g\cap M))$ is the limit of the same
sequence from within $M$.
\begin{lemma} Suppose that
$M\prec H(\kappa)$ (for suitably big $\kappa$) is a
countable elementary submodel\label{Melement}
containing $Y, \mathcal
A$. Let $M\cap \omega_1 = \delta$.
There is a sequence $\langle y^M(s) : s\in S_\delta\rangle$ such
that for every $(\bar s,\{s_i : i<n\}, \dot A)\in \mathcal A\cap M$,
and every $s \in S_\delta$ with $\bar s < s$, there is a
$B\subset Y^n\cap M$ such that
$ \langle y^M(s_i \oplus s) : i< n\rangle \in
B^{(\delta+1)}$ and
$s\forces B\subset \dot A$.
\end{lemma}
\begin{proof}
Let $\{ (s^m,\{s^m_i : i<n_m\} , \dot A_m ): m\in \omega\}$ enumerate
the family $\mathcal A\cap M$. Also, fix an enumeration, $\{
s^\delta_m : m \in \omega\}$, of $S_\delta$. Let $\{\alpha_m : m\in
\omega\}$ be an increasing cofinal sequence in $\delta$.
At stage $m$, we let $\beta_m$ be large enough
so that $s_0^\delta
\restriction [\beta_m, \delta) = s^\delta_i \restriction [\beta_m,
\delta)$ for
all $i< m$. Replace
the list $\{ (s^j , \langle s^j_i : i< \ell_j\rangle, \dot A_j) :
j<m\}$
with (abuse of notation of course) $\{
(s^j , \langle s^j_i : i< \ell_j\rangle, \dot A_j) : j<L_m\}$
so that for all $i,j<m$ with $s^j< s^\delta_i$ (from original list) the new list
includes $(s^\delta_i\restriction\beta_m , \langle s^j_i \oplus
s^\delta_i \restriction \beta_m :
i<\ell_j\rangle, \dot A_j)$; and so that for all $j<L_m$ in the new list
$s^j$ and $s^j_i$ are all in $ S_{\beta_m}$. Nothing else is added to the new list
(in particular, the new list is contained in $\mathcal A\cap M$).
Now we have $s^j_k \oplus s^\delta_0 = s^j_k \oplus s^\delta_i$ for
all $i<m$, $j<L_m$
and suitable $k<\ell_j$. Also, whenever $s^j< s^\delta_i$
($s^\delta_i$
for $i<m$ is unique) we have that
$s^\delta_i\forces \dot A_j\in \dot {\mathcal F}^+_{\langle s^j_i : i<
\ell_j\rangle}$;
and so we also have that
$s^\delta_0\forces (\dot A_j)^{s_i^\delta}_{s^\delta_0}
\in \dot {\mathcal F}^+_{\langle s^j_i : i< \ell_j\rangle}$
(because they are essentially the same sets).
Let $\Sigma = \{\sigma_k : k\in K\}$ lex-enumerate $
\{ (s^j_i
\oplus s^\delta_0)\restriction \beta_m : j<L_m, i<\ell_j\}\in M$.
Let
$\Pi_\Sigma $ be the projection map from $Y^{bS}$ onto $Y^\Sigma$.
Let $\Pi^\Sigma_{\langle s^j_i : i< \ell_j\rangle}$ be defined
by the equation $\Pi^\Sigma_{\langle s^j_i : i< \ell_j\rangle} \circ \Pi_\Sigma
=
\Pi_{\langle s^j_i : i< \ell\rangle}$.
We consider the filter (name) $\dot {\mathcal F}_{\Sigma}$. For each
$j<L_m$ and $i<m$ such that $s^j<s^\delta_i$,
it is forced
by $s^\delta_0$ that the set $\left(
\Pi_{\langle s^j_i
: i< \ell_j\rangle}^\Sigma\right)^{-1}((\dot
A_j)^{s^\delta_i}_{s^\delta_0})^{(<\omega_1)})$ is
a member of $\dot {\mathcal F}_{\Sigma}$ and all are in $M$. Select
any
$\vec y_m \in M\cap Y^\Sigma$ with the property that
$\Pi^\Sigma_{\langle s^j_i : i< \ell_j\rangle}(\vec y_m)\in \dot
A_j^{(<\omega_1)}$ for all $j< L_m$.
Choose a sequence $\{ B_j : j<L_m\}$ of countable subsets of
$Y^{<\omega}$ (in fact $B_j\subset Y^{\ell_j}$)
which are in $M$ and satisfy that, for each $j< L_m$,
$s^\delta_0\forces B_j\subset
(\dot A_j)^{s_i}_{s_0}$ (where $s^j<s^\delta_i$), and so that
$\Pi^\Sigma_{\langle s^j_i : i< \ell_j\rangle}(\vec y_m)\in
B_j^{(<\delta)}$.
Note that if $s^j<s^\delta_i$, then $s^\delta_i \forces
B_j\subset \dot A_j$.
If we now return to the ``original'' list, we
have that for all $i,j<m$ and $s^j<s^\delta_i$,
$s^\delta_i \forces
\vec y_m\restriction
\langle s^j_k\oplus s^\delta_0 \restriction \beta_m : k<\ell_j\rangle \in
B_j^{(<\delta)}$.
Now suppose
we have so chosen $\vec y_m$ for each $m\in \omega$.
We assert the existence of an infinite set $L\subset \omega$ with
the property that for all $j,i\in \omega$, $s^\delta_i$ forces that
the sequence $\{ \vec y_m(s^\delta_i\restriction \alpha_m\oplus
s^\delta_j) :
m\in L\}$ is defined and $S$-converging on a cofinite set.
For each $i$, $ y^M(s^\delta_i)$ is the $S$-name in $Y$
which is equal to the limit of this $S$-converging sequence.
\end{proof}
\subsection{$S$-preserving proper forcing}
Now we are ready to define our poset $\mathcal P$. Recall that
we have a fixed assignment $\{ \dot U(x,n) : x\in \theta, n\in \omega\}$
of $S$-names of neighborhoods (regular descending for each $x$).
\begin{definition}
A condition $p\in \mathcal P$ consists of $({\mathcal M}_p, S_p, m_p)$
where $\mathcal M_p$ is a finite $\in$-chain of countable elementary
submodels of
$(H(\kappa), \{ \dot U(x,m): x\in \theta, m\in \omega\})
$ for some suitable $\kappa$. We let $M_p$
denote the maximal
element of $ {\mathcal M}_p$ and let
$\delta_p = M_p\cap \omega_1$. We require that
$m_p$ is a positive integer and
$S_p$ is a finite subset of $S_{\delta_p}$.
For $s\in S_p$
and $M\in \mathcal M_p$, we use both
$s\restriction M$ and $s\cap M$ to
denote $s\restriction (M\cap \omega_1)$.
We require that the sequence $\{ y^M(s) : s\in S_{M\cap \omega_1}\}$
is in each $M'$
whenever $M\in M'$ are both in $\mathcal M_p$. This can be made
automatic if we use a fixed well-ordering of $H(\mathfrak c)$ and
define
$\{ y^M(s) : s\in S_{M\cap \omega_1}\}$ to be the minimal sequence
satisfying Lemma \ref{Melement}.
It is helpful to simultaneously think of $S_p$ as inducing a finite
subtree,
$S_p^\downarrow$,
of $S$ equal to $\{ s\restriction M : s\in S_p, \ \mbox{and}\
M\in \mathcal M_p\}$.
For each $s\in S_p$ and each $M\in {\mathcal M}_p\setminus M_p$
we define an $S$-name
$\dot W_p(s\restriction M)$ of a neighborhood of $e(y^M(s\restriction M))$.
It is defined as the name of the
intersection of all sets of the form $\dot U(e(y^{M'}(
s'\restriction M')),m_p)$
where $s'\in S_p$, $M'\in \mathcal M_p\cap M_p$, and
$s\restriction M\subset s'\restriction M'$
and $e(
y^M(s\restriction M)) \in \dot U(e(y^{M'}(s'\restriction
M')),m_p)$. We adopt
the convention that $\dot W_p(s\cap M)$ is all of $X$ if
$s\cap M\notin S_p^\downarrow$.
\medskip
The definition of $p< q$ is that $\mathcal M_q\subset \mathcal M_p$,
$m_q\leq m_p$,
$S_q \subset S_p^\downarrow$
and for
each $s'\in S_p$
and
$s\in S_q$,
we have that $s'$ forces that
$e( y^M(s\restriction M))\in \dot W_q(s\restriction M')$
whenever $M\in \mathcal M_p\setminus \mathcal M_q$
and $M'$ is the minimal member of $M_q\cap (\mathcal M_q\setminus M)$.
It is a notational convenience, and worth noting,
that we make no requirements on sets of
the form $\dot U(s, m_q)$ for $s\in S_q$.
\end{definition}
Before we develop the important properties of $\mathcal P$ let us
check that
\begin{proposition}
If $\mathcal P$ is $S$-preserving, then PFA(S) implies that
$S$ forces\label{3.9}
that $\dot X$
\begin{enumerate}
\item contains a free $\omega_1$-sequence, and
\item if $\dot X$ is first countable, contains a copy of $\omega_1$.
\end{enumerate}
\end{proposition}
\begin{proof}
Let us first consider the easier non-separable case.
For any condition $q\in \mathcal P$, let $\mathcal M(q)$ denote the
collection of all $M$ such that there exists a $p<q$ such
that $M\in \mathcal M_p$.
For each $\beta < \alpha\in \omega_1$, $s\in S_\alpha$,
and $m\in \omega$, let
\begin{align*}
D(\beta,\alpha,s,m) =& \{ p\in \mathcal P : (\exists \bar s\in S_p )~~
s < \bar s, m < m_p, \mbox{and} \\
& (\exists M\in \mathcal M_p)~ (\beta\in M, \alpha\notin M)\ \ \mbox{or}\\
& (\forall M\in \mathcal M(p)) (\beta\in M \Rightarrow \alpha\in M)\}~.
\end{align*}
It is easily shown that each $D(\beta,\alpha,s,m)$ is a dense open
subset of $\mathcal P$.
Consider the family
$\mathcal D$ of all such $D(\beta,\alpha,s,m)$, and
let $G$ be a $\mathcal D$-generic filter.
Let $\mathcal M_G = \{ M : (\exists p\in G)~ M \in
\mathcal M_{p}\}$
and let $C = \{ M\cap \omega_1 : M\in \mathcal M_G\}$.
Let $g\subset S$ be a
generic filter.
For each $\gamma\in C$ and $M\in \mathcal M_G$ with $M\cap\omega_1 =
\gamma$, let $x_\gamma = e(y^{M}(g\cap M))$ (we omit the trivial
proof that there is exactly one such $M$ for each $\gamma\in C$).
We show that the set
$W = \{ x_\gamma : \gamma \in C\}$ contains an uncountable free
sequence of $X$, and that, if $X$ is first countable, $W$
is homeomorphic to the
ordinal $\omega_1$.
If $\gamma<\delta$ are both in $C$ then $x_\gamma$
and $x_\delta$ are distinct. To see this, let us note that
since $x_\gamma\in M_{\delta}$ there is a $\beta\in
M_\delta$ such that $U(x_\gamma,0)\cap \omega_1\subset \beta$.
Also, the closure of
$U(x_\gamma,1)$ was assumed to be contained in $U(x_\gamma,0)$.
But now, $x_\delta = e(y^{M_\delta}(g\cap M_\delta))$ was chosen so as
to be in the closure of $e[\dot F\cap M_\delta]$ for each $\dot F\in
M_\delta$. In particular, $x_\delta$ is in the closure of
$\omega_1\setminus \beta$, and so it is not in $U(x_\gamma,1)$.
We may now define the map $f$ sending $x_\gamma$ to the
ordinal $o.t.(C\cap \gamma)$ (the order type).
It is certainly 1-to-1 and onto. Let $\{ \xi_n : n\in \omega\}\subset
\omega_1$ be strictly increasing with supremum $\xi$. For each $n$,
let $f(x_{\gamma_n}) = \xi_n$ and $f(x_\gamma) = \xi$. Fix any $m\in
\omega$,
set $s=g\cap S_\gamma$ and choose
any $p\in G\cap D(0,\gamma,s,m)$. Since $\gamma\in C$, we may assume
by extending $p$, that there is an $M_\gamma \in \mathcal M_p$ with
$M_\gamma\cap \omega_1 = \gamma$. Let $\beta$ be the maximum element
of $\{ M\cap \omega_1 : M\in \mathcal M_p\cap M_\gamma\}$. We have
that $f(x_\beta)<\xi$ and so there is an $n_0$ such that
$\xi_n > f(x_\beta)$ for all $n>n_0$. Now for any $r<p$ with $r\in G$,
and $M\in \mathcal M_r$ with $\beta < M\cap \omega_1 < \gamma$
we have that $s\Vdash e(y^M(s\cap M)) \in \dot U(s\cap M_\gamma,
m)$. From this it follows that $x_{\gamma_n} \in U(x_\gamma,m)$ for
all $n>n_0$. This shows that each limit point of $\{ x_{\gamma_n}
:n\in \omega\}$ is in $U(x_\gamma,m)$. In fact more is true:
the set $\{ x_\alpha : \beta <\alpha < \gamma\}$ is
contained in $U(x_\gamma,m)$ and, because
$U(x_\gamma,0)$
is in $M_{\gamma+1}$,
$U(x_\gamma,0)\cap \{x_\alpha : \gamma < \alpha <\omega_1\}$ is
empty. This shows that in all of $X$,
$\{ x_\alpha : \beta < \alpha \leq \gamma \}$ and
$\{ x_\alpha : \gamma<\alpha\}$ have disjoint closures.
Since $\gamma$ was an arbitrary member of $C$, this shows
that the closure of the full initial segment
$\{ x_\alpha : \alpha \leq \gamma\}$ is disjoint from
the closure of
$\{ x_\alpha : \gamma<\alpha\}$.
Now we assume that $X$ is first countable. We have just given
a proof using sequences that,
since $\omega_1$ is also first countable,
the inverse map, $f^{-1}$, is continuous.
Since $\dot X$ is forced to be Hausdorff, $f^{-1}$ is also a
homeomorphism.
\bigskip
Now we consider the separable case.
We are working in the forcing extension by $2^{<\omega_1}$.
For each
$\alpha\in \omega_1$, let $E_\alpha$ denote the dense open subset of
$2^{<\omega_1}$ of conditions that decide which member of $\dot
X$ is equal to the chosen ordinal $\alpha$ of the copy of $\omega_1$
that is forced to have no complete accumulation points
in $\mathcal W$. In particular,
for each $x\in \theta$ and $n\in \omega$,
let $E(x,n)$ denote the dense open set of conditions in $2^{<\omega_1}$
that decide
on the value of the name $\dot U(x,n)$ and that decide the countable
set $\dot U(x,n)\cap \omega_1$. Fix a $2^{<\omega_1}$-name,
$\dot {\mathcal P}$ for our poset
$\mathcal P$ as defined above.
We are assuming that
$2^{<\omega_1}$ forces that $\dot{\mathcal P}$
is proper and $S$-preserving. By
\cite{Tadatoshi} (
and see \cite[4.1]{Todorcevic}) it follows that the iteration
$2^{<\omega_1} * \dot {\mathcal P}$ is proper and $S$-preserving.
The rest of the proof proceeds
just as in the non-separable case.
\end{proof}
We prove a kind of density lemma.
\begin{lemma}
If $\mathcal P\in M$ for some countable $M\prec H(\mu)$, then
for each $p\in \mathcal P$ and each $\alpha \in M\cap \omega_1$,
such that $M\cap H(\kappa)\in \mathcal M_p$, there is an\label{singleM}
$M'\in M$ such that $\alpha\in M'$,
$r= (\mathcal M_p\cup\{M'\} , S_p, m_p)\in
\mathcal P$, and $r<p$.
\end{lemma}
\begin{proof}
Let $M_0 = M\cap H(\kappa)$ and
let $ S_0 = \{ s_i : i< \ell\}$ enumerate the set
$\{ s \cap M : s\in S_p\}$ in the lexicographic order. In this proof
we adopt the convention that we will enumerate $S_q$ for any condition
$q$ in increasing lexicographic order. Let $M^\dagger$ be the maximum
element of $\mathcal M_p\cap M$ and set $S^\dagger = \{ s\cap
M^\dagger : s\in S_p\}$. We define $p\restriction M$ to be
$(\mathcal M_p\cap M, S^\dagger, m_p)$. It is routine to verify that
$p< p\restriction M$.
By increasing $\alpha$ we
may assume that $ s_i \restriction [\alpha, M\cap \omega_1)
= s_j \restriction [\alpha, M\cap \omega_1)$ for all
$i,j< \ell$ and that $\omega_1\cap
\bigcup
(\mathcal M_p\cap M) < \alpha$.
For each $i< \ell$, let $\bar s_i =
s_i\restriction\alpha$ and let
$\bar S = \{ \bar s_i : i< \ell\}$.
It is easily checked that $r = (\{M_0\} \cup (\mathcal M_p \cap M)
,
S_0, m_p)$ is in $\mathcal P$ and is an extension of $p\restriction M $.
Notice that this implies that
$S_r$ is equal to $\bar S \oplus s_0 = \{ \bar s_i\oplus s_0 :
i<\ell\}$.
Define the $S$-name $\dot A$ as
\[ \{(s^q_0, \langle { y}^{M_q} (s^q_i) : i< \ell\rangle) :
q< p\restriction M , M_q\cap \mathcal M_q = \mathcal M_p\cap M,
m_q = m_p, S_q = \bar S\oplus s^q_0 \}\]
This set $\dot A$ is a member of $M_0$ and, by virtue of $r$,
$(s_0,
\langle { y}^{M_0}( s_i) : i< \ell\rangle)$ is an element
of $\dot A$.
We show, by a density and elementary submodel argument, that
we have that
$ s_0$ forces that $\dot A$ is in
$\dot {\mathcal F}^+_{\langle \bar s_i:i<\ell \rangle}$.
First of all, there is a dense subset of $S$ each member
of which decides the statement ``$(\exists F \in \dot{\mathcal
F}_{\langle \bar s_i : i < \ell\rangle}) ~~(\dot A\cap F = \emptyset)$''.
Since this dense set is in $M_0$, there is an $\dot F\in M_0$ such that
$s_0 \Vdash \dot F\in
\dot {\mathcal F}_{\langle \bar s_i : i<
\ell\rangle}
$, and either $s_0 \Vdash \dot A \in \dot {\mathcal
F}_{\langle \bar s_i : i< \ell\rangle}$ or $s_0 \Vdash \dot A\cap
\dot F = \emptyset$. However, it is clear that there is a $\beta\in
M_0$ such that $s_0\restriction \beta \Vdash \dot F\in
\dot {\mathcal F}_{\langle \bar s_i : i<
\ell\rangle}$ and so, by Lemma \ref{Melement}, $\langle
y^{M_0}(s_i) : i<\ell\rangle \in \dot F$.
It follows
then that there is an $s\in M_0$
below $s_0$ which also forces that
$ \dot A$ is in
$\dot {\mathcal F}_{\langle \bar s_i :i<\ell\rangle}^+$.
Well, what all this proves is that
$(s,\{\bar s_i : i<\ell\}, \dot A)$
is a member of the collection $\mathcal A$ and is in $M\cap H(\kappa)$.
Apply Lemma
\ref{Melement} and select $B\subset M\cap Y^\ell$ so that
$s_0\Vdash B\subset \dot A$ and $s_0\Vdash
\langle y^M( \bar s_i \oplus s_0) :
i<\ell\rangle \in B^{(\delta+1)}$.
What this actually means
(see Definition \ref{seql})
is that there is a $\langle \vec b_k : k\in \omega\rangle$
of elements of
$B^{(\delta)}$ which is $S$-converging coordinatewise
to this element. Now, by Lemma \ref{cohere}, there is a
$\vec b\in B$ satisfying that each $s\in S_p$ forces
that $e(\vec b) $ is in the product neighborhood
$\dot W_p(s_0) \times \cdots\times\dot W_p(s_\ell)$. This $\vec b\in B$
is of course equal to
$ \langle {y}^{M_q} (s^q_i) : i< \ell\rangle$ for
some $q$ as in the definition of $\dot A$. It follows
that $ M_q$ is the desired value for $M'$.
\end{proof}
That was a warm-up. What we really should have proven is
\begin{lemma}
For all $(s,r)\in S\times \mathcal P$ such that $s\notin M_r$, and
any\label{forproper} $M_0\in \mathcal M_r$ and
$(s^j,\langle s^j_i : i<\ell_j\rangle, \dot A_j)\in \mathcal A
\cap M_0$ with $s^j<s$, there is an $\vec a \in Y^{\ell_j}\cap
M_0$ such that $s \forces \vec a \in \dot A_j$ and for each
$s'\in S_r$ and $i<\ell_j$,
$s' \forces e(\vec a_i) \in \dot W_r( (s^j_i \oplus s)\cap M_0)$.
\end{lemma}
\begin{proof}
Let $s_0 = s\cap M_0$.
By definition of $\mathcal A$, we have that $s^j \forces
\dot A_j\in \dot {\mathcal F}^+_{
\langle s^j_i : i<\ell_j\rangle}$. By Lemma \ref{Melement},
there is a countable $B\subset M_0$ such that
$s_0\forces B\subset \dot A_j$ and
$s_0\forces \langle y^{M_0}(s^j_i\oplus s_0) : i<\ell_j\rangle
\in B^{(\delta_0+1)}$, where $\delta_0=M_0\cap \omega_1$.
Apply Lemma \ref{cohere} to conclude there is a
$\vec b\in B$ satisfying that each $s\in S_r$ forces
that $e(\vec b) $ is in the product neighborhood
$\dot W_r(s^j_0\oplus s_0) \times \cdots\times\dot W_r(s^j_{\ell-1}
\oplus s_0)$.
\end{proof}
All we have to do now is to prove that
\begin{theorem}
The poset\label{ppithm}
$S\times \mathcal P$ is proper.
\end{theorem}
\begin{proof}
As usual, we assume that $M\prec H(\mu)$ (for some suitably large $\mu$) is countable,
and that $M_0 = M\cap H(\kappa)\in {\mathcal M}_p$ for some condition $p$. Let
$D\in M$ be a dense subset of $S\times \mathcal P$ and assume that
$(s^\dagger,r)\in D$.
We may assume that
there is some elementary submodel $M_r$
including $r$,
that $s^\dagger \notin M_r$,
and that $s^\dagger\cap M_r\in S_r$. This means that for all
$s' \in \tilde M$ and $x\in M_r\cap \theta$, $s'\oplus s^\dagger$
forces
a value on $\dot U(x,m_r) \cap M_r$. Let $\langle M_i : i\in
\bar\ell\rangle $ enumerate $\mathcal M_r\setminus M$ in increasing
order. By Lemma \ref{singleM} we can assume that
$\mathcal M_r \cap M$ is not empty, and let
$M^\dagger $ denote the maximum element. Let $\alpha = M^\dagger\cap \omega_1$
and $\delta_0 = M_0\cap \omega_1$. We may additionally assume
that
$s\restriction
[\alpha,\delta_0) = s'\restriction [\alpha,\delta_0)$ for all $s,s'
\in S_r$.
The plan now is to find $q\in M\cap \mathcal P$ so that
$(s^\dagger ,q)\in D$ and $q\not\perp r$. Usual arguing will arrange
that $q< r\restriction M$ and, loosely speaking, that, for
some easily chosen expansion $S_{r,q}$ of $S_r$,
$(\mathcal M_q\cup \mathcal M_r, S_{r,q}, m_r)$ will be a condition
in $\mathcal P$ extending $q$.
The challenge is to ensure
that such a condition is also an extension of $r$, which requires
that,
for each $\tilde M\in \mathcal M_q$, we have that
$s\Vdash e(y^{\tilde M}(s'\cap \tilde M)) \in \dot
W_r(s' \cap M)$ for all $s,s'\in S_r$.
Some standard elementary submodels as side-conditions reasoning, together
with Lemma \ref{forproper} do the trick.
We have applied Lemma
\ref{singleM} above so
we also have that $s_i = \bar s_i\oplus s_0$
for each $i< \ell$. One thing we have gained is that in checking
if $q\not\perp r$, we need only check on membership in
sets of the form $\dot W_r(\bar s_i\oplus s_0 )$.
Let us say that $q$ is \underbar{\textbf{like}} $ r$ (or $q\equiv r$)
providing
\begin{enumerate}
\item $\mathcal M_r \cap M_0$ is an
initial segment of $\mathcal M_q$,
\item $\mathcal M_q \setminus \mathcal M_r = \{ M^q_i : i< \bar \ell\}$
has cardinality $\bar \ell = |\mathcal M_r\setminus M|$,
\item the tree structure $(S_q^\downarrow,<,\oplus)$ is
isomorphic to $( S_r^\downarrow,<,\oplus)$,
\item $ \{ s^q_i : i < \ell\} = \{ s\cap M^q_0 : s\in S_q\}$ is also equal
to $\{ \bar s_i \oplus s^q_0 : i < \ell\}$,
\end{enumerate}
For $q\equiv r$, and $k<\bar \ell$,
let $\langle s^{q,k}_i : i < \ell_{q,k}\rangle $ be
the set
$\{ s\cap M^q_k : s\in S_q\}$ ordered lexicographically.
Also let $\vec y_k\supy q $ denote the $\ell_{q,k}$-tuple
$\langle y^{M^q_k} (s^{q,k}_i ) : i<\ell_{q,k}\rangle$.
We have to do this because we want these to be members
of $Y^{\ell_{q,k}}$.
Thus we have the $\bar \ell$-tuple $\langle \vec y_k\supy q
: k< \bar \ell\rangle$ associated
with each $q\equiv r$.
Of course, $\ell_{q,k}$ is equal to $\ell_{r,k}$
for $q\equiv r$.
Also, for each $k<\bar \ell$, let $i^\dagger_k$
denote the index $i$ with the property that $s^{r,k}_i < s^\dagger$.
\bigskip
Recursively define a collection of sets and names.
First we have the $S$-name:
\[ \dot {\mathcal Y}_{\bar \ell} =
\{ (s, \langle \vec y_k\supy q : k< \bar \ell\rangle) :
(s,q)\in D \ \mbox{and}\
q\equiv r
\}~.\]
As usual, we have that $\dot {\mathcal Y}_{\bar \ell}\in M_0$
($q\equiv r$ can be described within $M$). Now define,
for $k\in \{ \bar\ell-1, \bar \ell-2, \ldots, 0\}$ (in that order)
\begin{multline}
\dot A(q,k) = \{ (\bar s, \vec y_k\supy {\bar q}) :
(\bar s, \langle \vec y_j\supy {\bar q} : j\leq k \rangle) \in
\dot {\mathcal
Y}_{k+1} \ , \\
s^q_{i^\dagger_{k}} < \bar s\ \ \mbox{and}\ \
\langle \vec y_j\supy {\bar q} : j < k\rangle =
\langle \vec y_j\supy q : j < k\rangle \}
\end{multline}
and let (for $k>0$)
\[ \dot {\mathcal Y}_{ k} = \{
(s,\langle \vec y_m\supy q : m< k \rangle) : s\forces \dot
A(q,k) \in
\dot {\mathcal F}^+_{\langle s^{q,k}_i : i< \ell_{q,k}\rangle}\}~.\]
Thus $\dot A(q,\bar \ell -1)$ contains the ``top'' element of the sequence
$\langle \vec y_{k}\supy q : k< \bar \ell\rangle$.
Of course
$s\forces \dot A(q,k) \in
\dot {\mathcal F}^+_{\langle s^{q,k}_i : i< \ell_{q,k}\rangle}$
is equivalent to $(s,\langle s^{q,k}_i : i< \ell_{q,k}\rangle,
\dot A(q,k))$ being a member of $ \mathcal A$. We use the notation
$\dot A(q,k)$ rather than the more cumbersome
$\dot A(\langle \vec y_j\supy q : j < k\rangle )$ but let
us note that the definition depends only on the parameters
$\langle \vec y_j\supy q : j < k\rangle $ and
$\dot {\mathcal Y}_{k+1}$ in $M^q_k$
(and the latter is
an element of $M_0$).
Clearly $(s^\dagger, \langle \vec y_k\supy r :k <\bar \ell\rangle)\in
\dot {\mathcal Y}_{\bar \ell}$ and, for readability,
let $k=\bar \ell-1$.
Recall that $i= i^\dagger_{k}$
was defined so that $s^r_{i} = s^\dagger \cap
M^r_{k}$.
We then have that
$s^\dagger \Vdash \vec y_{k}\supy r \in \dot A(r,k)$.
Now we
show that $(s^\dagger, \langle s^{r,k}_i : i < \ell_{r,k} \rangle,
\dot A(r,k)\,)$ is in $\mathcal A$. First choose any $\xi_k\in
M^r_k \setminus M^r_{k-1}$ large enough so that all members
of $\{ s^{r,k}_i : i <\ell_{r,k}\}$ agree on the interval
$[\xi_k, M^r_k\cap \omega_1)$. Let ${\bar s}^r_i =
s^r_i\restriction \beta_k$ for each $i<\ell_{r,k}$.
As discussed
above, we have that $\dot A(r,k)$ is an element of $M^r_k$.
By Lemma \ref{basiclemma}, there is a $\gamma\in M^r_k\cap \omega_1$
such
that each $s\in S_\gamma$ decides the statement
$\dot A(r,k)\in \dot{\mathcal F}^+_{\langle {\bar s}^r_i :
i<\ell_{r,k}\rangle} $.
It follows from Lemma \ref{Melement}
that $s^\dagger$ forces that $\vec y_k\supy r$ is a witness to the
fact that $\dot A\in \dot{\mathcal F}^+_{\langle {\bar s}^r_i :
i<\ell_{r,k}\rangle}$. Then, by applying Lemma \ref{basiclemma}
again, there is a $\beta_k\in M^r_k$ such that the tuple
$( s^\dagger\restriction \beta_k,
\langle {\bar s}^r_i :
i<\ell_{r,k}\rangle, \dot A(r, k)\,)$ is in $\mathcal A\cap
M^r_k$. This now shows that (for this value of $k$)
$(s^\dagger\restriction\beta_k , \langle
\vec y_m\supy r : m < k\rangle)$ is a member of $\dot
{\mathcal Y}_{k}$.
Continuing this standard argument, walking down from $s^\dagger$,
shows that, for each $k<\bar \ell$, there is a $\xi_k\in M^r_{k}$
such that
$(s^\dagger\restriction\xi_k , \langle
\vec y_m\supy r : m < k\rangle)$ is a member of $\dot
{\mathcal Y}_{k+1}$.
Now we have that there is some $\beta_0\in M_0$ such that
$(s^\dagger\restriction\beta_0 , \emptyset)$ is a member of $\dot
{\mathcal Y}_{1}$; and more importantly that
$(s^\dagger\restriction \beta_0, \emptyset, \dot A(r, 0 ))\in
\mathcal A\cap M_0$. By Lemma \ref{forproper}, there is a
$\vec y_0\in M_0$ such that $s^\dagger\forces
\vec y_0 \in \dot A(r,0 )$ and, for each
$s'\in S_r$ and $i<\ell_{r,0}$,
$s'\forces e(~(\vec y_0)_i~ ) \in \dot W_r((
s_i \oplus s^\dagger)\restriction M_0)$. Now, we do not ``really'' mean
$(\vec y_0)_i$ but rather $y^{M^q_0}(s^0_i)$ for a suitable
$q_0\equiv r$ such that $\vec y_0 = \vec y_0\supy {q_0}$. If
we let $\gamma_0 = M^{q_0}_0\cap \omega_1$, then we have
that $s^{q_0}_{i_0^\dagger} = s^\dagger \restriction \gamma_0$, and
for each $i<\ell_{r,0}$,
$$s^0_i = s^{q_0}_i = \bar q_i \restriction \gamma_0 \oplus
s^{q_0}_{i^\dagger_0}~.$$
By elementarity, there is a $\alpha_1\in M_0$ such that
$$(s^\dagger\restriction\alpha_1, \langle
s^0_i : i < \ell_{r,0}\rangle, \dot A(q_0, 1))\in \mathcal A~.$$
We apply Lemma \ref{forproper} again and obtain
$r\equiv q_1 \in M_0$ such that
$s^\dagger\forces
\vec y_1\supy {q_1} \in \dot A(
q_0, 1 )$ and, for each
$s'\in S_r$ and $i<\ell_{r,1}$,
$s'\forces e((\vec y_1)_i\,) \in
\dot W_r((
s^{q_1}_i \oplus s^\dagger)\restriction M_0)$.
Unlike in the first step, it may
happen that $(s^{q_1}_i\oplus s^\dagger)\restriction M_0$
is not in $S_r^\downarrow$, which poses no problem since
then $\dot W_r((
s^{q_1}_i \oplus s^\dagger)\restriction M_0)$
is all of $\theta$.
If
we let $\gamma_1 = M^{q_1}_1\cap \omega_1$, then we have
that $s^{q_1}_{i_1^\dagger} = s^\dagger \restriction \gamma_1$;
and for each $i<\ell_{r,1}$~, let
$s^1_i = s^{q_1}_i $. So we again have
that $(\,s^\dagger, \langle s^1_i : i< \ell_{r_0} \rangle , \dot
A(q_1,2)\,)$ is in $\mathcal A$.
Well, we just repeat this argument for $\bar \ell$ steps until
we find $q = q_{\bar\ell} \equiv r$
with the property that $(s^\dagger, q)\in
D$ and, for each $k<\bar\ell$,
and for each
$s'\in S_r$ and $i<\ell_{r,k}$,
$s'\forces e(
\vec y_k\supy q ((s^q_{i}\oplus s^\dagger)\cap M^q_k)\,)
\in \dot W_r((s^q_i \oplus
s^\dagger)\restriction M_0)$. Again, for larger values of $k$, it may
happen
that
$(s^k_i \oplus s^\dagger) \cap M_0$ is not in $S_r^\downarrow$, and so
$\dot W_r((s^k_i \oplus s^\dagger) \cap M_0)$ would simply equal $X$
or $\theta$.
\end{proof}
\section{On\label{nolarge}
the consistency of $\mathbf{GA}$}
Shelah has defined the
$\kappa $-p.i.c. (for ``proper isomorphism condition'') . The reason for
this is that a
countable support iteration of length
$\omega_2$ of $\aleph_2$-p.i.c.proper posets will
(under CH) satisfy the
$ \aleph_2$-chain condition, while just assuming that the
factors themselves satisfy the
$\aleph_2$-chain condition does not guarantee the
iteration
will. A diamond sequence on $\omega_2$ will help us decide
which proper
$\aleph_2$-p.i.c.posets of size $\aleph_2$ to use in such an iteration. The
resulting iteration
will have cardinality $\aleph_2$ and the objects which we
want
to consider in the extension will also have cardinality $\aleph_2$.
We have to
show that given
any such reflected object there is an appropriate $\aleph_2$-p.i.c.
proper poset
of cardinality $\aleph_2$ which will introduce the required set and
that this set is still appropriate in the final model.
For this we use the method of Todorcevic \cite{matrices} in which side
conditions are finite sets (or matrices) of elementary submodels
rather than the more common method in which side conditions are
simple finite chains of elementary submodels. It is this change which
is the key in making the posets strongly $\aleph\sb2$-cc (and
iterable), thus removing the need for large cardinals to prove the
results.
\begin{defn}[{\cite[Ch. VIII]{proper}}]
\label{3.2}
A poset $ P $ is said to satisfy the $ \kappa $-pic
if whenever we have a sufficiently large cardinal $ \lambda $,
a well-ordering $\prec $
of $ H(\lambda )$, $i < j < \kappa $,
two countable elementary submodels $ N_i $ and $N_j $
of $\langle H(\lambda ), \prec , \in \rangle $
such that $ \kappa $ and $ P $ are in $ N_i \cap
N_j $ , $ i \in N_i $ , $j \in N_j$,
$N_i \cap i = N_j \cap j $,
and suppose further that we are given $ p \in
N_i $ and an
isomorphism $ h : N_i \rightarrow N_j$
such that $ h(i) = j $ and $ h $ is the identity on
$N_i \cap N_j $ then there is a $ q \in P $ such that :
\begin{enumerate}
\item
$ q < p$ , $q < h(p) $ and $ q $ is both $ N_i $ and $ N_j $ generic ,
\item if $ r \in N_i \cap P $ and $ q' < q $
there is a $ q'' < q' $ so that \\
$q'' < r $ if and only if $ q'' < h(r) $ .
\end{enumerate}
\end{defn}
\begin{proposition} A countable\label{3.3} support iteration
of length at most $ \omega_1$ of
$\aleph_2$-p.i.c. proper posets is again $ \aleph_2$-p.i.c..
Furthermore if CH holds and the
iteration has length at most $\aleph_2\ $ then the iteration satisfies the
$\aleph_2$-cc .
\end{proposition}
\begin{proposition}
A proper poset of cardinality $ \aleph_1 \ $ satisfies the
$ \aleph_2$-pic.
\end{proposition}
\begin{lemma}
[{[CH]}] If $ P $ is a\label{chaincondition}
proper $ \aleph_2$-p.i.c.
poset and $ G $ is $P$-generic
over $ V $ then $ V[G]\models \mathfrak c = \omega_1 \ $ .
\end{lemma}
Following \cite{matrices}, for a
countable
elementary submodel $N$ of $H(\aleph_2)$,
we let $\overline{N}$ be the transitive collapse,
and we let $h_N : N \rightarrow \overline{N}$ be the collapsing map,
i.e. $h_N(x) = \{ h_N(y) : y\in x\cap N \}$.
\begin{lemma}
Suppose that $N_1,N_2$ are countable\label{homega1}
elementary submodels of $H(\aleph_2)$ such that
$\overline{N_1} = \overline{N_2}$, and let
$h_{N_1,N_2} $ denote the map $h_{N_2}^{-1} \circ h_{N_1}$. Then
$h_{N_1,N_2}$ is the identity on $H(\aleph_1)\cap N_1$ and
for each $A\in N_1$ with $A\subset H(\aleph_1)$,
$A\cap N_1 = h_{N_1,N_2}(A) \cap N_2$.
\end{lemma}
\begin{proof}
It follows by $\in$-induction that each
$x\in N_1\cap H(\aleph_1)$, $x\subset N_1$
and so $x\in \overline{N_1}$. Therefore we also have, by
$\in$-induction, that $h_{N_1}(x) = x = h_{N_2}^{-1}(x)$.
\end{proof}
A family $[\mathcal{ N} ] $ is
an elementary matrix if, for some integer $n>0$,
\begin{enumerate}
\item $[\mathcal{ N}] = \{\mathcal N_1,\ldots,\mathcal N_n \}$
\item for each $1\leq i\leq n$, $\mathcal N_i$
is
a finite set of countable \ elementary \ submodels \ \ of
$H(\omega_2 )$
\item for each $1\leq i\leq n$, $\overline{N_1} = \overline{N_2}$
for each pair $ N_1,N_2 \in \mathcal N_i$,
\item for each $1\leq i< j\leq n$ and each
$ N_i \in \mathcal N_i$, there is an
$N_j \in \mathcal N_j$ with $N_i\in N_j $.
\end{enumerate}
It will be convenient to let $N \in [\mathcal{ N}]$,
for an elementary submodel $N$ of $H(\aleph_2)$,
be an abbreviation for $ N \in \mathcal N $
for some $\mathcal N \in [\mathcal N]$.
\begin{lemma}[CH] If $\dot{\mathcal I}$ is an\label{p22pic}
$S$-name of a $P$-ideal on $\omega_1$ such
that 1 forces that $\dot{ \mathcal I}\cap [E]^{\aleph_0}$ is not empty for
all stationary sets $E\subset\omega_1$, then
there is an
$S$-preserving $\aleph_2$-p.i.c.proper poset $\mathcal P$ of cardinality
$2^{\omega_1} $ such that $\mathcal P$ forces that
there is an
$S$-name $\dot E$ of a stationary set
with $1\Vdash [\dot E]^{\aleph_0}\subset \dot
{\mathcal I}$.
\end{lemma}
\begin{lemma}[CH] If $\dot X$ is an $S$-name of a\label{ppipic}
sequentially compact
non-compact first countable space, then
there is an
$S$-preserving $\aleph_2$-p.i.c. proper poset $\mathcal P$ of cardinality
$2^{\aleph_1} $ such that $\mathcal P$ forces that
there is an $S$-name
$\{ \dot x_\gamma : \gamma\in \omega_1\}\subset \dot X$
that is forced to contain an uncountable free sequence,
and, if $\dot X$ is first countable, to
be a homeomorphic copy of $\omega_1$.
\end{lemma}
The proofs are very similar with the same underlying idea in that we
replace elementary chains from the original proofs with
elementary matrices. The usage of elementary
matrices is the device to make the poset satisfy the
$\aleph_2$-p.i.c. The proof that the modified poset is proper and
$S$-preserving relies on the fact that CH guarantees that
the key combinatorics take
place within $H(\aleph_1)$ and so, by Lemma \ref{homega1}
no new arguments or constructions
are required.
Since it is newer, we sketch the
proof of Lemma \ref{ppipic} and leave the proof of Lemma \ref{p22pic}
to the interested reader. In actual fact, this method isn't really
needed for the consistency of $\mathbf{P}_{22}$ because the
needed poset can be chosen to have cardinality $\mathfrak c$.
The reason this is not true for $\mathbf{PPI}^+$ is that we must
utilize the construction of the maximal filter of $S$-sequentially closed sets
which may have cardinality $2^{\omega_1}$.
We simply indicate the modifications needed
to the proof of Lemma \ref{ppithm}.
\begin{proof}
Let $\dot X$ be the $S$-name as formulated in the Lemma.
By Lemma \ref{itsomega1}, we can pass to a subspace and assume
that either $\dot X$ is separable or that the set
$\omega_1\times\{0\}$ is a subset and is
forced
to not have a complete accumulation point. Since we are assuming CH,
we can, in either case, pass to an $S$-name of
cardinality $\aleph_1$ for a subspace that is still sequentially
compact and not compact. With this space having cardinality
$\aleph_1$ it is clear that also in the separable case, we can assume
that $\omega_1\times \{0\}$ is a subspace with no complete accumulation point.
We can now assume that the base set for $\dot X$ is
$\omega_1\times \omega_1$ (i.e. any set that is a
subset of $H(\aleph_1)$.
The entire topology $\dot \tau$ for $\dot X$ can be coded as a subset
of $S\times \omega\times \omega_1^4$ where
$(s,m,\alpha,\beta,\gamma,\delta)\in \dot \tau$ codes the fact
that $s$ forces that $(\gamma,\delta)$ is in $\dot U(~(\alpha,\beta)~,
m)$. The family $\bigcup \{ Y_\alpha : \alpha\in \omega_1\}$ and
$\mathbf{WF}$
as defined in \S 3.1 are already subsets of $H(\aleph_1)$.
We also fix a well-order $\prec_{\omega_1}$ of $H(\aleph_1)$.
Finally, with no changes, the family $\mathcal A$ as defined in
Definition \ref{needA} is a subset of $H(\aleph_2)$. This family
$\mathcal A$ was the key parameter in defining our poset $\mathcal
P$. Lemma \ref{Melement} holds for any $M\prec
(H(\aleph_2),\tau, \mathcal A)$ (meaning $\tau\in M$ and $\mathcal A$
is a new term in the language).
The choice of the sequence $\{ y^M(s) : s\in S_{M \cap \omega_1}\}$
from Lemma \ref{Melement} will be the $\prec_{\omega_1}$-minimal
such sequence.
\begin{claim}
Consider any set $\mathcal N$
of pairwise isomorphic countable elementary submodels of
$(H(\aleph_2), \dot \tau, \mathcal A)$; i.e.
$\overline{N} = \overline{N'}$ for $N,N'\in \mathcal N$. Let $\delta = N\cap
\omega_1$ for any $N\in \mathcal N$.\label{claimsame}
Let $N_1, N_2$ be elements of $\mathcal N$.
We then have that the two sequences
$\langle y^{N_1}(s) : s \in S_\delta\rangle$ and
$\langle y^{N_2}(s) : s \in S_\delta\rangle$
are the same.
\end{claim}
\bgroup
\defProof of Claim \ref{itspic}:{Proof of Claim \ref{claimsame}:}
\begin{proof}
To prove the claim, let $(\bar s,\{s_i : i < n \}, \dot A)$ be any member of
$\mathcal A\cap N_1$ and assume that $\bar s < s\in S_\delta$.
Choose $B\subset Y^n \cap N_1$ such that $s\Vdash
B\subset \dot A$ and
$s \Vdash \langle y^{N_1}(s\oplus s_i) : i < n\rangle \in
B^{(\delta+1)}$. By Lemma \ref{homega1},
$h_{N_1,N_2}((\bar s, \{ s_i : i< n\}, \dot A))$ is in $\mathcal A
\cap N_2$. Since $\dot A\subset H(\aleph_1)$, we also have
by Lemma \ref{homega1}, that $h_{N_1,N_2}(\dot A)\cap N_2$
is equal to $\dot A \cap N_1$. Therefore, we have
that $s$ also forces that $B$ is a subset of
$h_{N_1,N_2}(\dot A)$. Well this shows that
$\langle y^{N_1} (s\oplus s_i) : i < n \rangle$ satisfies
this particular requirement of
$\langle y^{N_2} (s\oplus s_i) : i < n \rangle$
with respect to $h_{N_1,N_2}((\bar s, \{ s_i : i< n\}, \dot A))$.
Since $h_{N_1,N_2}$ is an isomorphism, this shows
that $\langle y^{N_1}(s) : s\in S_\delta\rangle$ works as a
choice for
$\langle y^{N_2}(s) : s\in S_\delta\rangle$, and so, indeed,
they are the same.
\end{proof}
\egroup
\bigskip
A condition $p\in \mathcal P$ consists of $([{\mathcal N}_p], S_p, m_p)$
where $[\mathcal N_p]$ is a
elementary matrix of
submodels of
$(H(\aleph_2), \prec_{\omega_1},\dot \tau, \mathcal A)$.
We let
$\delta_p = N \cap \omega_1$
for any maximal $N\in \mathcal N_p$. We require that
$m_p$ is a positive integer and
$S_p$ is a finite subset of $S_{\delta_p}$.
For each $s\in S_p$ and each
non-maximal $N\in [{\mathcal N}_p] $
we define an $S$-name
$\dot W_p(s\restriction N)$ of a neighborhood of $e(y^N(s\restriction N))$.
It is defined as the name of the
intersection of all sets of the form $\dot U(s'\restriction N',m_p)$
where $s'\in S_p$, non-maximal $N'\in [\mathcal M_p] $, and
$s\restriction N\subset s'\restriction N'$
and $e(
y^N(s\restriction N)) \in \dot U(s'\restriction N')$. We adopt
the convention that $\dot W_p(s\cap N)$ is all of $X$ if
$s\cap N\notin S_p^\downarrow$.
\medskip
The definition of $p< q$ is that
each $N\in [\mathcal N_q]$ is a member of $[ \mathcal N_p]$,
$m_q\leq m_p$,
$S_q \subset S_p^\downarrow$
and for
each $s'\in S_p$
and
$s\in S_q$,
we have that $s'$ forces that
$e( y^N(s\restriction N))\in \dot W_q(s\restriction N')$
whenever $N\in [\mathcal N_p] $ and $N\notin [ \mathcal N_q]$
and $N'$ is a minimal member of
$[\mathcal N_q]\setminus N $, which is itself not a maximal member
of $[\mathcal N_q]$.
Again we note
that we make no requirements on sets of
the form $\dot U(s, m_q)$ for $s\in S_q$.
Because of Claim \ref{claimsame}, the proof that $\mathcal P$ is proper
and $S$-preserving proceeds exactly as in the proof of
Theorem \ref{ppithm}.
\begin{claim}
$\mathcal P$ satisfies\label{itspic}
the $\aleph_2$-p.i.c.
\end{claim}
\bgroup
\defProof of Claim \ref{itspic}:{Proof of Claim \ref{itspic}:}
\begin{proof}
Let $\lambda$ be a
sufficiently large cardinal, fix
a well-ordering $\prec $
of $ H(\lambda )$, let $i < j < \omega_2 $ be such
that there are
two countable elementary submodels $ N_i $ and $N_j $
of $\langle H(\lambda ), \prec , \in \rangle $
such that $\mathcal P $ is in $ N_i \cap
N_j $ , $ i \in N_i $ , $j \in N_j$,
$N_i \cap i = N_j \cap j $,
and suppose further that we are given $ p \in \mathcal P\cap
N_i $ and an
isomorphism $ h : N_i \rightarrow N_j$
such that $ h(i) = j $ and $ h $ is the identity on
$N_i \cap N_j $. We must show that
there is a $ q \in P $ such that :
\begin{enumerate}
\item
$ q < p$ , $q < h(p) $ and $ q $ is both $ N_i $ and $ N_j $ generic ,
\item if $ r \in N_i \cap P $ and $ q' < q $
there is a $ q'' < q' $ so that \\
$q'' < r $ if and only if $ q'' < h(r) $ .
\end{enumerate}
We first show that since $\mathcal P\in N_i\cap N_j$, we also
have that $\{\prec_{\omega_1},\dot \tau, \mathcal A\}\in N_i\cap N_j$.
The reason is that the collection $\{ N : (\exists p\in \mathcal P)
N\in [\mathcal N_p]\}$ is in $N_i\cap N_j$.
It follows that $N_i' = (N_i\cap H(\aleph_2), \prec_{\omega_1},
\dot \tau , \mathcal A )$ is an elementary submodel of
$(H(\aleph_2), \prec_{\omega_1}, \dot \tau, \mathcal A)$.
$N_j'$ defined similarly is as well. The definition of
the $[\mathcal N_q]$ for $q$ is canonical. Given
that $[\mathcal N_p] = \{ \mathcal N_1, \mathcal N_2,
\ldots , \mathcal N_n\}$, we set
$[\mathcal N_q ] = \{ \mathcal N_1\cup h(\mathcal N_1),
\ldots, \mathcal N_n \cup h(\mathcal N_n) ,
\{ N_i', N_j'\}\} $. The existence of $h$ ensures
that $\overline{N_i'} = \overline{N_j'}$. Since
$[\mathcal N_p]\in N_i'$ and $
h([\mathcal N_p])= [\mathcal N_{h(p)}]\in N_j'$
we have that $[\mathcal N_q]$ is an elementary matrix.
Choose $S_q \subset S_{N_i\cap \omega_1}$ to be any finite
set such that $S_p= h(S_p) \subset S_q^\downarrow$.
We already know that $q$ is both $N_i$ and $N_j$ generic
from the arguments in Theorem \ref{ppithm}.
Finally let $r\in N_i\cap \mathcal P$ and $q'<q$ with $q'\in \mathcal P$.
We may assume, by symmetry, that there is a $q'' <q'$ that is
also below $r$. Let $[\mathcal N_{q''} ] $ be listed
as $\{ \mathcal N_1, \ldots , \mathcal N_k \}$ and let
$1<\ell\leq k$ be chosen so that $ N_i' \in \mathcal N_\ell$.
For $1\leq m< \ell$, let $\mathcal N_m^i = \mathcal N_m\cap N_i$.
Of course we have that $N\in N_j'$ for each
$1\leq i<\ell$ and each $N\in h(\mathcal N_m^i)$.
It is the easily verified that
$$[\mathcal N_{\tilde q} ] =
\{ h(\mathcal N_1^i)\cup \mathcal N_1, \ldots ,
h(\mathcal N_{\ell-1}^i)\cup \mathcal N_{\ell-1},
\mathcal N_{\ell},\ldots, \mathcal N_k \}$$
is an elementary matrix,
and so
$\tilde q \in \mathcal P$ where
$( [\mathcal N_{\tilde q}], S_{q''}, m_{q''})$, where
$$[\mathcal N_{\tilde q} ] =
\{ h(\mathcal N_1^i)\cup \mathcal N_1, \ldots ,
h(\mathcal N_{\ell-1}^i)\cup \mathcal N_{\ell-1},
\mathcal N_{\ell},\ldots, \mathcal N_k \}~~.$$
It is immediate that $\tilde q < q''$, and so $\tilde q < q',r$.
We just have to show that $\tilde q$ is also below $h(r)$.
Since $q'' < r$, we have that $[\mathcal N_r]\in N_i$ is a submatrix
of $\{ N_i\cap \mathcal N_1,\ldots, N_i\cap \mathcal N_{\ell-1}\}$,
and so $[\mathcal N_{h(r)}]$ is a submatrix of
$[\mathcal N_{\tilde q}]$.
\end{proof}
\egroup
This completes the proof.
\end{proof}
\begin{definition} The stationary set of ordinals $\lambda \in \omega_2$
with uncountable cofinality is denoted
as $S^2_1$.
The principle $\diamondsuit(S^2_1)$ is the statement:
There is a family $\{ X_\lambda : \lambda \in S^2_1\}$ such that
\begin{enumerate}
\item for each $\lambda\in S^2_1$, $X_\lambda\subset\lambda$,
\item for each $X\subset\omega_2$, the set $E_X = \{ \lambda \in
S^2_1 : X\cap \lambda = X_\lambda\}$ is stationary.
\end{enumerate}
\end{definition}
\begin{theorem}
Assume $2^{\aleph_0}=\aleph_1$ and $\diamondsuit(S^2_1)$.
There is a proper poset $\mathbb P$ so that in the forcing extension
by $\mathbb P$ there is a coherent Souslin tree $S$ such that,
in the full forcing extension by $\mathbb P * S$,
the statement $\mathbf{GA}$ holds.
\end{theorem}
\begin{proof}
We construct a countable support iteration sequence
$\langle \mathbb P_\alpha , \dot{\mathbb Q}_\beta :
\alpha \leq \omega_2, \beta < \omega_2\rangle$.
By induction, we assume that $\mathbb P_\alpha$
is proper, has cardinality at most $\aleph_2$, and that
$$\Vdash_{\mathbb P_\alpha} \dot{\mathbb Q}_\alpha
\ \ \mbox{satisfies the}\ \
\aleph_2\mbox{-p.i.c}\ \ .$$
Note that by Lemmas \ref{3.3} and \ref{chaincondition} we will
have that, for each $\alpha < \omega_2$, CH holds
in the forcing extension by $\mathbb P_{\alpha}$.
We may assume that $\dot {\mathbb Q}_0$, and
therefore $\mathbb P_1$ is constructed so that
there is a $\mathbb P_1$-name, $\dot S$ of a coherent Souslin
tree (henceforth we suppress the dot on the $S$).
We further demand of our induction that, for $\alpha\geq 1$
$$\Vdash_{\mathbb P_\alpha} \dot{\mathbb Q}_\alpha
\ \ \mbox{is}\ \ S\mbox{-preserving}~.$$
For each ordinal $0<\alpha \in \omega_2\setminus S^2_1$, we
let $\dot{\mathbb Q}_\alpha$ denote the $\mathbb P_\alpha$-name
of the standard Hechler poset for adding a dominating real. This ensures
that $\mathfrak b = \omega_2$ in the forcing extension by $\mathbb P_{\omega_2}$.
For the rest of the construction, fix any function $h$ from $\omega_2$
onto $H(\aleph_2)$.
Also let $\{ X_\lambda : \lambda \in S^2_1\}$ be
a $\diamondsuit(S^2_1)$-sequence.
Now consider $\lambda\in S^2_1$ and let
$x_\lambda = h[ X_\lambda]$. We define $\dot{\mathbb Q}_\lambda$ according to
cases:
\begin{enumerate}
\item if $x_\lambda$ is the $\mathbb P_\lambda*S$-name of a P-ideal on $\omega_1$
such that $$1\Vdash [E]^{\aleph_0}\cap x_\lambda \ \ \mbox{is not empty for all stationary sets \ } E\subset\omega_1$$
then $\dot{\mathbb Q}_\lambda$ is the $\mathbb P_\lambda$-name of the poset
from Theorem \ref{p22pic},
\item if $x_\lambda$ is the $\mathbb P_\lambda*S$-name of a
subset of $\lambda\times\lambda\times\lambda $ so that
if we define, for $\xi,\eta\in \lambda$,
$\dot U(\xi,\eta) $ to be the $\mathbb P_\lambda*S$-name of
the subset of $\lambda$ such that $$\{ (\xi,\eta)\}\times
\dot U(\xi,\eta) = x_\lambda \cap (\{(\xi,\eta)\}\times \lambda )\ \ ,$$
i.e. for $( (p,s), (\xi,\eta,\gamma) )$ in the set $x_\lambda$,
$( (p,s), \gamma )$ is in the name $\dot U(\xi,\eta)$,
and $\mathbb P_\lambda* S$
forces that the family $\{ \dot U(\xi,\eta) : \eta\in \lambda\}$ is a
local base for $\xi$ in a
sequentially
compact regular topology on $\lambda$, and that no finite subset of
$\{ \dot U(\xi,\eta) : \xi,\eta\in \lambda\}$ covers $\lambda$,
then $\dot {\mathbb Q}_\lambda$ is the $\mathbb P_\lambda$-name of
the poset from Theorem \ref{ppipic}.
\item in all other cases, $\dot {\mathbb Q}_\lambda$ is the
$\mathbb P_\lambda$-name of the
Cohen poset $2^{<\omega}$.
\end{enumerate}
Assume that $\dot{\mathcal I}$ is a $\mathbb P_{\omega_2}*S$-name
of a P-ideal on $\omega_1$ satisfying that there is some
$(p_0,s_0)\in \mathbb P_{\omega_2}*S$ forcing that
$[E]^{\aleph_0}\cap \dot{\mathcal I}$ is not empty for all stationary
sets $E\subset \omega_1$.
The ideal of all countable subsets of $\omega_1$ is also such an
ideal, so we can find a $\mathbb P_{\omega_2}*S$-name
$\dot{\mathcal J}$ such that $(p,s) \Vdash \dot{\mathcal J}
= \dot{\mathcal I}$, and $1$ forces that
$[E]^{\aleph_0}\cap \dot{\mathcal I}$ is not empty for all stationary
sets $E\subset \omega_1$.
Let $X\subset \omega_2 $
be chosen to be the set of all $\xi\in \omega_2$
with the property that
there is a $\mu<\omega_2$ such that
$h(\xi)$ is a $\mathbb P_\mu*S$-name
with
$1\Vdash h(\xi)\in \dot{\mathcal J}$.
There is a cub $C\subset\omega_2$ such that for each
$\mu<\mu' \in C$:
\begin{enumerate}
\item
the collection $\{ h(\xi) : \xi \in X\cap \mu\}$
is a collection of $\mathbb P_{\mu'}*S$-names,
\item for each countable subset
$\{\xi_n : n\in \omega\}$ of $X\cap \mu $ there
is a $\xi<\mu'$ such that 1 forces that $h(\xi_n)$
is almost contained $h(\xi)$ for
each $n$,
\item
every $\mathbb P_\mu*S$-name
that is forced by 1 to be a member of $\dot{\mathcal J}$
is equivalent to a name in $\{ h(\xi) : \xi \in X\cap \mu'\}$.
\end{enumerate}
Therefore, there is a $\lambda\in E_X\cap C$ such
that $X_\lambda = X\cap \lambda$. We may of course
assume that $p_0\in \mathbb P_\lambda$.
Routine checking now shows that $x_\lambda$ satisfies
clause (1) in the definition of $\dot{\mathbb Q}_\lambda$.
It follows that $\mathbb P_{\omega_2}*S$ is a model
of $\mathbf{P}_{22}$.
Now suppose that we have a $\mathbb P_{\omega_2}*S$-name
of a sequentially compact non-compact space. We note
that $\mathbb P_{\omega_2}*S$ forces that $2^{\aleph_0} = \aleph_2$.
Therefore, by Lemma \ref{itsomega1}, we can pass to a name of a
sequentially
compact non-compact subspace which has cardinality
at most $\omega_2$. In fact, we can assume this space
has cardinality exactly $\omega_2$ by taking the free union
with the Cantor space.
Let $\dot Z$ denote
the $\mathbb P_{\omega_2}*S$-name of this space.
Again, by Lemma \ref{itsomega1}
we can assume that each point of the space has a
separable neighborhood. This means that, with re-indexing,
we can assume
that the base set for the space is the ordinal
$\omega_2$ and
$\{ \dot U(\xi,\eta) : \xi ,\eta\in \omega_2\}$ is
the list of $\mathbb P_{\omega_2}*S$-names of the neighborhood
bases of the points, and that no finite subcollection covers.
We define $X$ to be the set of all those
$\alpha\in \omega_2$ such that $h(\alpha)$ is
a tuple of the form $( (p,s) , (\xi,\eta,\gamma) )$,
i.e. a $\mathbb P_{\omega_2}*S$-name of a member of
$\omega_2\times\omega_2\times \omega_2$,
where $(p,s) \Vdash \gamma \in \dot U(\xi,\eta)$.
We again want to choose a $\lambda$ in $E_X \cap C$
for some special cub set $C$ and in this case it is much
simpler to make use of uncountable elementary submodels.
Let $\kappa$ be any regular cardinal greater than $2^{\omega_2}$,
and let $\{ M_\alpha : \alpha \in \omega_2\}$ be chosen so that,
for each $\alpha\in \omega_2$:
\begin{enumerate}
\item $X, h$ and $\mathbb P_{\omega_2}*S$ are in $ M_\alpha$
\item $\omega_1\subset M_\alpha$ and $M_\alpha$ has cardinality $\aleph_1$,
\item for each $\beta<\alpha$, every countable subset of $M_\beta$
is an element of $M_\alpha$,
\item $M_\alpha$ is an elementary submodel of $H(\kappa)$,
\item if $\alpha$ is a limit ordinal, then $M_\alpha = \bigcup\{
M_\beta : \beta < \alpha\}$.
\end{enumerate}
Items (2) and (4) guarantee that $M_\alpha\cap \omega_2$ is
an initial segment of $\omega_2$ -- hence an ordinal.
The chain $\{ M_\alpha : \alpha \in \omega_2\}$ is a continuous chain
because of item (5), and so $C = \{ M_\alpha \cap \omega_2 : \alpha
\in \omega_2\}$ is a closed and unbounded subset of $\omega_2$.
Now we choose $\lambda \in E_X\cap C$, we can also choose
$\lambda$ so that it is $M_\lambda$ with $M_\lambda\cap \omega_2 $
being $\lambda$.
Using items (1), (3) and (4) and the fact that $\lambda \in S^2_1$,
it is now easy to show that
$x_\lambda$ will satisfy the requirement (2) in the construction
of $\dot{\mathbb Q}_\lambda$. It then follows,
as in the proof of Lemma \ref{3.9},
that
$\mathbb P_{\lambda+1}*S$ will force the existence of the necessary
$\omega_1$-sequence showing that $\dot Z$ is not a counterexample
to $\mathbf{PPI}^+$.
\end{proof}
\def$'${$'$}
| 70f6718812ad749ca4f8ff6176e4fd9a30701557 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
\label{sec:introduction}
Medical imaging provides a mechanism to study \textit{in vivo} physiological properties of the human body, and thus plays an important role in the diagnosis, prognosis, and assessment of treatment response of different diseases.
To facilitate decision-making in clinical practice, quantitative imaging (QI), i.e., the extraction of numerical or statistical features from medical images, is being actively investigated \cite{sullivan2015metrology,rosenkrantz2015clinical}.
QI has demonstrated substantial promise in multiple clinical applications.
These include the quantification of metabolic tumor volume from oncological positron emission tomography (PET) for predicting clinical outcomes \cite{ohri2015pretreatment}, quantification of dopamine transporter uptake from single-photon emission computed tomography (SPECT) to assess the severity of Parkinson disease \cite{filippi2005123i,moon2020physics}, and quantification of regional uptake from PET and SPECT for dosimetry in targeted radionuclide therapy \cite{flux2006impact,ljungberg20023,dewaraja2012mird,li2021projection}.
Given the significant interest in QI, multiple methods have been and are being developed for QI.
For clinical translation of QI, it is essential that the measurements made by those methods are reliable.
Thus, there is an important need for objective evaluation of the reliability of measurements obtained using QI methods.
Typically, such evaluation requires the presence of either the true value of the quantitative parameter or a reference standard.
Such true values or reference standards can be available in realistic simulation and physical phantom studies \cite{du2005partial,jin2013evaluation,ouyang2006fast,he2005monte,liu2021bayesian}.
While these studies are important for the initial development of QI methods, there is an important need for techniques that can perform objective evaluation of QI methods directly with patient data.
Such evaluation then requires the presence of gold-standard quantitative values.
These are typically time-consuming, expensive, and tedious to obtain.
Further, even when an approximate gold standard is available, it could suffer from the lack of reliability.
Thus, techniques that can objectively evaluate QI methods in the absence of a gold standard are much needed.
To objectively evaluate QI methods without the knowledge of a gold standard, a regression-without-truth (RWT) technique was proposed in a set of seminal papers \cite{hoppin2002objective,kupinski2002estimation}.
The RWT technique assumes that the true and measured values are linearly related by a slope, bias, and Gaussian-distributed noise term.
It was demonstrated that even in the absence of a gold standard, the values of the slope, bias, and the standard deviation of the noise term for all the considered QI methods can be estimated using a maximum-likelihood (ML) approach.
These estimated parameters can then be used to rank different QI methods on the basis of precision.
The efficacy of the RWT technique was demonstrated in evaluating segmentation methods on the task of estimating the apparent diffusion coefficient from diffusion-weighted magnetic resonance imaging (MRI) scans \cite{jha2010evaluating,jha2012task}, and on the task of estimating the left ventricular ejection fraction from cardiac cine MRI sequences \cite{lebenberg2012nonsupervised}.
The RWT technique was then advanced further and the efficacy of the resultant no-gold-standard evaluation (NGSE) technique \cite{jha2016no} was demonstrated in objectively evaluating reconstruction methods for SPECT on the task of quantifying regional uptake \cite{jha2015objective,jha2016no}.
Further, the technique was applied to clinical oncological PET images to evaluate segmentation methods on the task of measuring metabolic tumor volume \cite{jha2017practical}. While the findings from these studies are encouraging, an important assumption in these existing evaluation techniques is that the noise between measurements obtained using different QI methods is independent of each other.
The noise with different QI methods arises in the process of measuring the same true value, and then can be correlated.
Thus, this assumption is often violated.
To address this issue, we propose an advanced NGSE technique that accounts for the presence of such correlated noise.
We start by presenting the theory of this technique.
\section{METHODS}
\subsection{Theory}
\label{sec:methods(theory)}
Consider a clinical scenario where a total of $P$ patients are being scanned by an imaging system.
From the acquired data of each patient, a set of $K$ QI methods are used to measure certain quantitative values.
Such quantitative values can be the mean activity concentration within different organs of interest.
Our objective is to estimate the parameters that can describe the relationship between the true and measured values, without access to a gold standard.
These estimated parameters can then be used to rank the QI methods.
We assume that there exists a linear stochastic relationship between the true and measured values.
This relationship is parameterized by a slope, bias, and Gaussian-distributed noise term.
We note that this noise arises in the process of measuring the same true values but with different methods.
Thus, the noise is expected to be correlated.
We model this correlated noise by a zero-mean multi-variate Gaussian distribution denoted by $\mathcal{N}(0, \mathbf{C})$.
Specifically, the diagonal elements of $\mathbf{C}$, i.e., $\{ \sigma_k^2 \}$, denote the variance of the noise of each method.
The off-diagonal elements of $\mathbf{C}$, i.e., $\{ \sigma_{k,k'} \}$, denote the covariance of the noise between methods $k$ and $k'$.
For the $p^\mathrm{th}$ patient, denote the true value by $a_p$ and the estimated value using the $k^\mathrm{th}$ method by $\hat{a}_{p,k}$.
Additionally, denote the slope and bias of the $k^\mathrm{th}$ method by $u_k$ and $v_k$, respectively.
For the $p^\mathrm{th}$ patient, we can then write the relationship between the true and measured values as
\begin{align}
\left[
\begin{array}{c}
\hat{a}_{p,1} \\ \hat{a}_{p,2} \\ \vdots \\ \hat{a}_{p,K}
\end{array}
\right]
=
\begin{bmatrix}
u_1 & v_1 \\ u_2 & v_2 \\ \vdots & \vdots \\ u_K & v_K
\end{bmatrix}
\left[
\begin{array}{c}
a_p \\ 1
\end{array}
\right]
+
\mathcal{N}(0, \mathbf{C}).
\label{eq: rel. btw. true and meas. in matrix notation}
\end{align}
Denote the vector $\left[ \hat{a}_{p,1}, \hat{a}_{p,2}, \dots, \hat{a}_{p,K} \right]^T$ by $\boldsymbol{\hat{A}}_p$, the matrix containing $\{u_k\}$ and $\{v_k\}$ by $\boldsymbol{\Theta}$, and the vector $\left[a_p, 1\right]^T$ by $\boldsymbol{A}_p$.
Based on Eq.~\eqref{eq: rel. btw. true and meas. in matrix notation}, obtaining the probability of observing $\boldsymbol{\hat{A}}_p$ given the knowledge of $\{\boldsymbol{A}_p, \boldsymbol{\Theta}, \mathbf{C}\}$ depends on the true values, which are unknown.
To address this issue, we next assume that the true values are sampled from a four-parameter beta distribution (FPBD) parameterized by a vector $\boldsymbol{\mathit{\Omega}}$ \cite{jha2016no}. This FPBD incorporates the fact that the true values lie within a certain range.
Additionally, the FPBD provides the capability to model a wide variety of the ranges and shapes of the true distribution.
Let $\boldsymbol{\mathcal{\hat{A}}} = \{\boldsymbol{\hat{A}}_p, \ p = 1, 2, \dots, P\}$ denote the collection of measurements made by all the $K$ methods from all the $P$ patients.
The NGSE technique uses an ML approach to estimate the values of $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$ that maximize the probability of observing $\boldsymbol{\mathcal{\hat{A}}}$.
The ML estimate of $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$ is given by
\begin{align}
\left\{\hat{\boldsymbol{\Theta}},\hat{\mathbf{C}},\hat{\boldsymbol{\mathit{\Omega}}}\right\}_{\mathrm{ML}} =
\underset{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}}{\mathrm{arg~max}}
\left\{
\mathrm{pr}(\boldsymbol{\mathcal{\hat{A}}} | \boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}})
\right\},
\label{eq: ML estimates}
\end{align}
where $\mathrm{pr}(\boldsymbol{\mathcal{\hat{A}}} | \boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}})$ denotes the probability of observing $\boldsymbol{\mathcal{\hat{A}}}$ given the knowledge of $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$.
We note that obtaining this probability does not require any knowledge of the true values.
The expression for $\mathrm{pr}(\boldsymbol{\mathcal{\hat{A}}} | \boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}})$ was determined to obtain the ML estimates using a constrained optimization technique based on the interior-point algorithm \cite{byrd1999interior}.
From the estimated parameters, we used the slope terms $\{\hat{u}_k\}$ and noise standard deviation terms $\{\hat{\sigma}_k\}$, i.e., the square root of the diagonal elements of the covariance matrix, to compute the noise-to-slope ratio (NSR) for each method.
For the $k^\mathrm{th}$ method, the NSR is given by
\begin{align}
\text{NSR}_k =
\frac{\hat{\sigma}_k}{\hat{u}_k}.
\end{align}
The NSR evaluates QI methods on the basis of precision \cite{kupinski2002estimation,hoppin2002objective,jha2016no}, and a lower value indicates a more precise estimation performance.
\subsection{Evaluating the NGSE technique using numerical experiments}
\label{sec:methods(evaluation)}
We evaluated the performance of the NGSE technique using multiple numerical experiments. In each experiment, $P = 200$ true values were sampled from a known FPBD.
From these true values, noisy measured values were generated for $K = 3$ hypothetical QI methods.
Each method yielded outputs that were linearly related to the true values by a slope of $u_k$ and bias of $v_k$.
The variance of the noise of each method was characterized by the diagonal elements of the covariance matrix $\mathbf{C}$.
Additionally, the covariance of the noise between different methods were characterized by the off-diagonal elements of $\mathbf{C}$.
These noisy measurements were then input to the NGSE technique to estimate $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$.
From these estimated parameters, we used the slope terms $\{\hat{u}_k\}$ and the noise standard deviation terms $\{\hat{\sigma}_k\}$ to compute the NSR for all methods to rank them based on precision, as described in Sec.~\ref{sec:methods(theory)}.
In this evaluation, we sampled the $200$ true values from FPBD for $4$ combinations of $\boldsymbol{\mathit{\Omega}}$ such that different ranges and shapes of the true distribution were modeled.
To evaluate the sensitivity of the NGSE technique to correlated noise, we generated two sets of QI methods for each combination of $\boldsymbol{\mathit{\Omega}}$.
The first set of methods had lower correlated noise with $\{\sigma_{1,2}, \sigma_{1,3}, \sigma_{2,3}\} = \{0.004,0.008,0.012\}$.
In contrast, the second set of methods had higher correlated noise with $\{\sigma_{1,2}, \sigma_{1,3}, \sigma_{2,3}\} = \{0.015,0.02,0.03\}$.
For both sets, the values of slope $\{u_k\}$, bias $\{v_k\}$, and variance of the noise $\{\sigma_k^2\}$ of the three methods were set to $\{1.1,0.9,1.05\}$, $\{0.1,0.2,0.3\}$, and $\{0.04, 0.09, 0.2025\}$, respectively.
Finally, for each combination of $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$, we repeated the experiment for $20$ different noise realizations.
Thus, we evaluated the performance of the NGSE technique for a total of $4\times2\times20=160$ trials.
\section{Results}
We first present the performance of the NGSE technique for the set of hypothetical QI methods that had lower correlated noise.
The means and standard deviations of the estimated slope $\{\hat{u}_k\}$, noise standard deviation $\{\hat{\sigma}_k\}$, and resultant NSR for all considered methods are reported in Table~\ref{tab: Results (numerical): low correlated noise}.
As described in Sec.~\ref{sec:methods(evaluation)}, these statistics were computed from a total of $80$ trials.
In each trial, we considered either a different combination of $\{\boldsymbol{\Theta}, \mathbf{C}, \boldsymbol{\mathit{\Omega}}\}$ or a different noise realization of the synthetic measurements given the true values.
We observe that the NGSE technique reliably estimated the slope, noise standard deviation, and consequently the NSR values.
From the estimated NSR values, the NGSE technique accurately ranked the methods for $78\%$ of the $80$ trials.
Further, the technique correctly identified method $1$ as the most precise method for $97\%$ of the $80$ trials.
\begin{table*}[h]
\centering
\captionsetup{justification=centering}
\caption{The means and standard deviations of slope, noise standard deviation, and resultant NSR estimated using the NGSE technique for the set of methods that had lower correlated noise.}
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Method index & True slope & Estimated slope & True noise standard deviation & Estimated noise standard deviation & True NSR & Estimated NSR \\ \hline
1 & 1.10 & 1.11~$\pm$~0.08 & 0.20 & 0.17~$\pm$~0.09 & 0.18 & 0.16~$\pm$~0.08\\ \hline
2 & 0.90 & 0.89~$\pm$~0.07 & 0.30 & 0.30~$\pm$~0.06 & 0.33 & 0.35~$\pm$~0.09\\ \hline
3 & 1.05 & 1.06~$\pm$~0.09 & 0.45 & 0.44~$\pm$~0.04 & 0.43 & 0.42~$\pm$~0.07\\ \hline
\end{tabular}
}
\label{tab: Results (numerical): low correlated noise}
\end{table*}
We then present in Table~\ref{tab: Results (numerical): high correlated noise} the results for the set of methods that had higher correlated noise.
We again observe that the NGSE technique reliably estimated the slope, noise standard deviation, and resultant NSR.
For $87\%$ of the $80$ trials, the technique yielded accurate rankings of the considered methods.
Further, the technique correctly identified that method $1$ was the most precise for $97\%$ of the $1,600$ trials.
\begin{table*}[h]
\centering
\captionsetup{justification=centering}
\caption{The means and standard deviations of slope, noise standard deviation, and resultant NSR estimated using the NGSE technique for the set of methods that had higher correlated noise.}
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Method index & True slope & Estimated slope & True noise standard deviation & Estimated noise standard deviation & True NSR & Estimated NSR \\ \hline
1 & 1.10 & 1.13~$\pm$~0.09 & 0.20 & 0.17~$\pm$~0.08 & 0.18 & 0.15~$\pm$~0.08\\ \hline
2 & 0.90 & 0.91~$\pm$~0.07 & 0.30 & 0.30~$\pm$~0.06 & 0.33 & 0.34~$\pm$~0.08\\ \hline
3 & 1.05 & 1.07~$\pm$~0.09 & 0.45 & 0.44~$\pm$~0.05 & 0.43 & 0.42~$\pm$~0.07\\ \hline
\end{tabular}
}
\label{tab: Results (numerical): high correlated noise}
\end{table*}
\section{Discussion and Conclusion}
For clinical translation of QI, there is an important need for techniques that can objectively evaluate QI methods with patient data.
In this context, existing statistical techniques assume that the noise between measurements obtained using different methods is independent of each other.
However, this assumption can often be violated since the noise arises in the process of measuring the same true value.
To address this issue, we developed an NGSE technique that models this correlated noise by a multi-variate Gaussian distribution.
Our results from the numerical experiments (Tables~\ref{tab: Results (numerical): low correlated noise} and \ref{tab: Results (numerical): high correlated noise}) showed that the NGSE technique yielded reliable estimates of slope, noise standard deviation, and consequently the NSR for all hypothetical QI methods.
Additionally, the technique yielded accurate rankings of these methods for $83\%$ of the total $160$ trials.
Further, the technique was able to identify the most precise method for $97\%$ of the cases.
This observation is especially important since when evaluating different QI methods, the objective is typically to find the method that yields the most reliable performance \cite{jha2017practical}.
All these results demonstrated that in controlled settings, where the true and measured values were linearly related by design, the NGSE technique was able to reliably rank QI methods without access to any knowledge of the ground truth.
The results motivate further evaluation of the proposed technique with QI methods that are developed for clinical applications.
These include methods developed for reconstruction, post-reconstruction processing, segmentation, and quantification.
Often, such methods are evaluated using strategies that rely on the availability of a ground truth.
Further, this ground truth may not be relevant to the clinical task.
For example, segmentation methods are evaluated using metrics such as the Dice score and Hausdorff distance, which quantify spatial overlap and shape similarity, respectively, between the estimated segmentation and a certain ground-truth segmentation.
Such evaluation then requires access to the true segmentation, which is typically unavailable.
Usually manual segmentations are used as a surrogate for the ground truth, but these can be erroneous and suffer from intra- and inter-reader variability \cite{leung2020physics}. Similarly, denoising methods for low-dose images are evaluated by comparing the denoised image to a certain normal-dose image using metrics of structural similarity index and root mean square error. However, the normal-dose image is also noisy, and thus provides a limited measure of ground truth.
More importantly, it is unclear whether the evaluation based on those conventional metrics correlates with the clinical task \cite{yu2020ai,zhu2021comparing,liu2022need}.
Thus, these methods should preferably be evaluated based on clinical-task performance \cite{jha2021objective}.
The proposed NGSE technique provides a mechanism to perform evaluation on clinically relevant quantitative tasks and without access to the ground truth.
One limitation of the proposed technique is that the true and measured values are assumed to be linearly related.
This linear relationship is desirable in QI since it ensures that the measured quantitative value is linearly related to the biological effect.
However, this assumption of linearity may not always hold true in QI.
To address this issue, one strategy is to check whether the measurements made by different methods are linearly related to each other.
This will increase the confidence that the measured values are also linearly related to the ground truth \cite{jha2017practical}.
A second limitation is that the NGSE technique requires many patient images since multiple parameters need to be estimated.
One way to reduce the required number of input images is to incorporate the prior information of the parameters to be estimated \cite{jha2015incorporating}.
Thus, extending the proposed technique to incorporate such prior knowledge is an important research direction.
In conclusion, our study demonstrated the ability of the proposed NGSE technique to accurately rank different QI methods in the presence of correlated noise, and without the need for any knowledge of the ground truth.
The results motivate further evaluation of the technique with realistic simulation studies and patient data.
\section*{ACKNOWLEDGEMENTS}
Financial support for this work was provided by the National Institute of Biomedical Imaging and Bioengineering R01-EB031051, R56-EB028287, and R21-EB024647 (Trailblazer Award).
| 0a218adb31e98b3772d7f94753d3dbbfdb8da6f8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Related work}
\noindent \textbf{Dense prediction problem} We focus on three dense prediction subtasks: such as monocular depth estimation, semantic segmentation and image super-resolution.
Learning-based monocular depth estimation was first introduced by Saxena et al.~\cite{saxena2006learning}. Later studies improved accuracy by using large network architectures~\cite{chen2019structure,eigen2015predicting,eigen2014depth,Hu2018RevisitingSI,laina2016deeper} or integrating semantic information~\cite{jiao2018look} and surface normals~\cite{qi2018geonet}. Fu et al.~\cite{fu2018deep} formulated depth estimation as an ordinal regression problem, while~\cite{chen2016single,lee2019monocular} estimated relative instead of metric depth. Facil et al.~\cite{facil2019cam} proposed to learn camera calibration from the images for depth estimation. Recent approaches further improve the performance by exploiting monocular priors such as planarity constraints~\cite{liu2019planercnn,liu2018planenet,Yin2019enforcing,huynh2020guiding,lee2019big} or occlusion~\cite{ramamonjisoa2019sharpnet}. Gonzalez and Kim~\cite{gonzalezbello2020forget} estimated depth by synthesizing stereo pairs from a single image, while~\cite{yang2021transformers} and~\cite{Ranftl2021} applied vision-transformer for depth prediction. For resource-limited hardware, it is more desirable to not only have a fast but also accurate model. A simple alternative is employing lightweight architectures such as MobileNet~\cite{howard2019searching,howard2017mobilenets,sandler2018mobilenetv2,wofk2019fastdepth}, GhostNet~\cite{han2020ghostnet}, and FBNet~\cite{tu2020efficient}. A popular approach is utilizing network compression techniques, including quantization~\cite{han2015deep}, network pruning~\cite{yang2018netadapt}, knowledge distillation~\cite{yucel2021real}. Another approach employs well-known pyramid networks or dynamic optimization schemes~\cite{aleotti2021real}.
Deep neural networks have thrived on semantic segmentation, with early works~\cite{girshick2014rich,hariharan2014simultaneous} proposed to classify region proposals to perform this task. Later on, fully convolutional neural network methods~\cite{long2015fully,dai2015boxsup} are widely adopted to process arbitrary-sized input images and train the network end-to-end. Atrous convolution-based approach~\cite{liang2015semantic} generated the middle-score map that was later refined using the dense conditional random field (CRF)~\cite{krahenbuhl2011efficient} to mitigate the low-resolution prediction problem. Chen et al.~\cite{chen2017deeplab} then implemented atrous spatial pyramid pooling for segmenting objects at different scales, while~\cite{chen2017rethinking} and~\cite{chen2018encoder} employed atrous separable convolution and an efficient decoder module to capture sharp object boundaries. Zheng et al.~\cite{zheng2015conditional} enabled end-to-end training of dense CRF by implementing recurrent layers. Other approaches~\cite{badrinarayanan2017segnet,noh2015learning} utilized transposed convolution to generate the high-resolution prediction. Long et al.~\cite{long2015fully} produced multi-resolution prediction scores and took the average to generate the final output. Hariharan et al.~\cite{hariharan2015hypercolumns} fused mid-level features and trained dense classification layers at multiple stages. Badrinarayanan et al.~\cite{badrinarayanan2017segnet} and Ronneberger et al.~\cite{ronneberger2015u} implemented transposed convolution with skip-connections to exploit mid-level features. Wang et al.~\cite{wang2020deep} utilized multi-scale parallel inter-connected layers to further exploit learned features from pre-trained ImageNet models. Lightweight approaches~\cite{orsic2019defense,li2019dfanet,yu2018bisenet} employed pre-trained backbones as a decoder and a simple decoder to perform fast segmentation. Zhao et al.~\cite{zhao2018icnet} modified the cascade architecture of~\cite{zhao2017pyramid} to shrink the model size and speed-up inference.
Image super-resolution task has also been immensely improved using deep neural networks. Dong et al.~\cite{dong2015image} first proposed a shallow but bulky network that~\cite{dong2016accelerating} later shrunk down utilizing transposed convolution. Kim et al.~\cite{kim2016accurate} implemented a deeper architecture with skip connection while Zhang et al.~\cite{zhang2018image} proposed channel attention to improve the performance. Kim et al.~\cite{kim2016deeply} introduced recursive layers that Tai et al.~\cite{tai2017image} later extended by adding the local residual connection. Likewise, many studies attempted to enhance model efficiency. Ahn et al.~\cite{ahn2018fast} suggested using cascade architecture and group convolution to reduce the number of parameters. Hui et al.~\cite{hui2018fast,hui2019lightweight} proposed the information distillation module for creating an efficient model. Lai et al.~\cite{lai2017deep} applied the Laplacian pyramid network to gradually increase the spatial resolution while downsizing the model.
However, state-of-the-art methods mainly focus on increasing accuracy at the cost of model complexity that is infeasible in resource-limited settings, while manually designed lightweight architecture is a tedious task, requires much trial-and-error, and usually leads to architectures with low accuracy.
\noindent \textbf{Neural Architecture Search} There has been increasing interest in automating network design using neural architecture search. Most of these methods focus on searching high-performance architecture using reinforcement learning~\cite{baker2016designing,liu2018progressive,pham2018efficient,zoph2016neural,zoph2018learning}, evolutionary search~\cite{real2019regularized}, differentiable search~\cite{liu2018darts}, or other learning algorithms~\cite{luo2018neural}. However, these methods are usually very slow and require huge resources for training. Other studies~\cite{dong2018dpp,elsken2018multi,hsu2018monas} also attempt to optimize multiple objectives like model size and accuracy. Nevertheless, their search
process optimizes only on small tasks like CIFAR. In contrast,
our proposed method targets real-world data such as NYU, KITTI, and Cityscapes on multiple dense prediction tasks.
\section{Lightweight Dense Precition (LDP)}
\noindent We propose the LDP framework to search for accurate and lightweight monocular depth estimation architectures utilizing a pre-defined backbone that has been successful for dense prediction in the past.
The proposed framework takes in a dataset as input to search for the best possible model. This model can be deployed for depth estimation on hardware-limited devices. The first subsection defines the search space while the remaining two describe our multi-objective exploration and search algorithm.
\subsection{Search Space}
\noindent Previous neural architecture search (NAS) studies demonstrated the significance of designing a well-defined search space. A common choice of NAS is searching for a small set of complicated cells from a smaller dataset~\cite{zoph2018learning,liu2018progressive,real2019regularized}. These cells are later replicated to construct the entire architecture that hindered layer diversity and suffered from domain differences~\cite{tan2019mnasnet}. On the other hand, unlike classification tasks, dense prediction problems involve mapping a feature representation in the encoder to predictions at larger spatial resolution in the decoder.
To this end, we build our search space upon a pre-defined backbone that is shown as the set of green blocks in Figure~\ref{fig:search_space_ver2}. The backbone is divided into multi-scale pyramid networks operating at different spatial resolutions. Each network scale consists of two encoder blocks, two decoder blocks, a refinement block, a downsampling and upsampling block (except for scale 1). Each block is constructed from a set of identical layers (marked as orange in Figure~\ref{fig:search_space_ver2}). Inspired by~\cite{tan2019mnasnet}, we search for the layer from a pool of operations and connections, including:
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=0.99\linewidth]{figures/search_algorithm.pdf}
\end{center}
\vspace{-0.35cm}
\caption{The flowchart of our architecture search that utilizes the Assisted Tabu Search (ATS) with mutation to search for accurate and lightweight monocular depth estimation networks.} \vspace{-0.55cm}
\label{fig:search_algorithm}
\end{figure*}
\begin{itemize}[noitemsep,topsep=0pt,parsep=1pt,partopsep=1pt]
\item The number of resolution scales $S$.
\item The number of layers for each block $N_{i,j}$.
\item Convolutional operations (ConvOps): vanilla 2D convolution, depthwise convolution, inverted bottleneck convolution and micro-blocks~\cite{li2021micronet}.
\item Convolutional kernel size (KSize): $3 \times 3$, $5 \times 5$.
\item Squeeze and excitation ratio (SER): $0$, $0.25$.
\item Skip connections (SOps): residual or no connection.
\item The number of output channels: $F_{i,j}$.
\end{itemize}
\noindent Here $i$ indicates the resolution scale and $j$ is the block index at the same resolution. Internal operations such as ConvOps, KSize, SER, SOps, $F_{i,j}$ are utilized to construct the layer while $N_{i,j}$ determines the number of layers that will be replicated for block$_{i,j}$. In other words, as shown in Figure~\ref{fig:search_space_ver2}, layers within a block (e.g. layers 1 to N$_{1,2}$ of Encoder Block 1,2 are the same) are similar while layers of different blocks (e.g. Layer 1 in Refine Block 1,5 versus Layer 1 in Upsample Block S,7) can be different.
We also perform layer mutation to further diversifying the network structure during the architecture search process. The mutation operations include:
\begin{itemize}[noitemsep,topsep=0pt,parsep=1pt,partopsep=1pt]
\item Swapping operations of two random layers with compatibility check.
\item Modifying a layer with a new valid layer from the predefined operations.
\end{itemize}
Moreover, we also set computational constraints to balance the kernel size with the number of output channels. Therefore, increasing the kernel size of one layer usually results in decreasing output channels of another layer.
Assuming we have a network of $S$ scales, and each block has a sub-search space of size $M$ then our total search space will be $M^{5 + [(S - 1) * 7]}$. Supposedly, a standard case with $M=192$, $S=5$ will result in a search space of size $\sim 2 \times 10^{75}$.
\subsection{Multi-Objective Exploration}
\noindent We introduce a multi-objective search paradigm seeking for both accurate and compact architectures. For this purpose, we monitor the \textit{validation grade} $\mathcal{G}$ that formulates both accuracy and the number of parameter of the trained model. It is defined by
\begin{equation}
\label{eq:validation_grade}
\mathcal{G}(m) = \alpha \times A(m) + (1 - \alpha) \times \bigg[\frac{P}{P(m)}\bigg]^{r}
\end{equation}
\noindent where $A(m)$ and $P(m)$ are validation accuracy and the number of parameters of model $m$. $P$ is the target compactness, $\alpha$ is the balance coefficient, and $r$ is an exponent with $r=0$ when $P(m) \leq P$ and otherwise $r=1$. The goal is to search for an architecture $m^{*}$ where $G(m^{*})$ is maximum.
However, computing $G$ requires training for every architecture candidate, resulting in considerable search time. To mitigate this problem, Mellor et al.~\cite{mellor2021neural} suggested to score an architecture at initialisation to predict its performance before training. For a network $f$, the \textit{score(f)} is defined as:
\begin{equation}
\label{eq:jacob_cov_score}
score(f) = log|K_{H}|
\end{equation}
\noindent where $K_{H}$ is the kernel matrix.
Let us assume that the mapping of model $f$ from a batch of data $X = \{x_i\}^{N}_{i=1}$ is $f(x_i)$.
By assigning binary indicators to every activation units in $f$, a linear region $x_i$ of data point $i$ is represented by the binary code $c_i$. The kernel matrix $K_{H}$ is defined as:
\begin{equation}
\label{eq:kernel_matrix}
K_{H} = \begin{pmatrix}
N_{A} - d_{H}(c_1, c_1) & \dots & N_{A} - d_{H}(c_1, c_N) \\
\vdots & \ddots & \vdots \\
N_{A} - d_{H}(c_N, c_1) & \dots & N_{A} - d_{H}(c_N, c_N)
\end{pmatrix}
\end{equation}
\noindent where $N_A$ is the number of activation units, and $d_{H}(c_i, c_j)$ is the Hamming distance between two binary codes. Inspired by this principle, we generate and train a set of different architectures for various dense prediction tasks. We evaluate the performance of these models and visualize the results against the \textit{score} that in our case is the mapping of depth values within image batches.
Leveraging this observation, we 1) utilize the \textit{score} in our initial network ranking, and 2) define the mutation exploration reward $\mathcal{R}$ as:
\begin{equation}
\label{eq:mutation_exploration_reward}
\mathcal{R}(m_i,m_j) = \alpha \times \frac{score(m_j)}{score(m_i)} + (1 - \alpha) \times \bigg[\frac{P}{P(m_j)}\bigg]^{r}
\end{equation}
\noindent where $m_j$ is a child network that is mutated from $m_i$ architecture.
\subsection{Search Algorithm}
\noindent The flowchart of our architecture search is presented in Figure~\ref{fig:search_algorithm}. We first randomly generate $60K$ unique parent models and create the initial network ranking based on the \textit{score} in Eq.~\ref{eq:jacob_cov_score}. We then select \textit{six} architectures in which \textit{three} are the highest-ranked while the other \textit{three} have the highest score of the networks with the size closest to the target compactness.
\begin{table*}[t!]
\caption{\label{tab:eval_nyuv2}Evaluation on the NYU-Depth-v2 dataset. Metrics with $\downarrow$ mean lower is better and $\uparrow$ mean higher is better. Type column shows the exploration method used to obtain the model. RL, ATS, and manual, refer to reinforcement learning, assisted tabu search, and manual design, respectively.}
\centering
\small
\begin{tabular}{@{}llrcccccccc@{}}
\hline
\multicolumn{2}{c}{\textbf{Architecture}} & \textbf{\#params} & \textbf{Type} & \textbf{Search Time} & \textbf{REL$\downarrow$} & \textbf{RMSE$\downarrow$} & \(\boldsymbol{\delta_{1}}\)$\uparrow$ & \(\boldsymbol{\delta_{2}}\)$\uparrow$ & \(\boldsymbol{\delta_{3}}\)$\uparrow$ \\ \hline
AutoDepth-BOHB-S & Saikia et al.'19~\cite{saikia2019autodispnet} & 63.0M & RL & 42 GPU days & 0.170 & 0.599 & - & - & - \\ \hline
EDA & Tu et al.'21~\cite{tu2020efficient} & 5.0M & Manual & - & 0.161 & 0.557 & 0.782 & 0.945 & 0.984 \\ \hline
Ef+FBNet & Tu \& Wu et al.~\cite{tu2020efficient,wu2019fbnet} & 4.7M & Manual & - & 0.149 & 0.531 & 0.803 & 0.952 & 0.987 \\ \hline
FastDepth & Wofk et al.'19~\cite{wofk2019fastdepth} & 3.9M & Manual & - & 0.155 & 0.599 & 0.778 & 0.944 & 0.981 \\ \hline
SparseSupNet & Yucel et al.'21~\cite{yucel2021real} & 2.6M & Manual & - & 0.153 & 0.561 & 0.792 & 0.949 & 0.985 \\ \hline
LiDNAS-N & Huynh et al.'21~\cite{huynh2021lightweight} & 2.1M & ATS & 4.3 GPU days & \textbf{0.132} & 0.487 & 0.845 & 0.965 & \textbf{0.993} \\ \hline
LDP-Depth-N & Ours & \textbf{2.0M} & ATS & 4.3 GPU days & \textbf{0.132} & \textbf{0.483} & \textbf{0.848} & \textbf{0.967} & \textbf{0.993} \\ \hline
\end{tabular} \vspace{-0.45cm}
\end{table*}
Starting from these initial networks, we strive for the best possible model utilizing Assisted Tabu Search (ATS), inspired by Tabu search (TS)~\cite{glover1986future} that is a high level procedure for solving multicriteria optimization problems. It is an iterative algorithm that starts from some initial feasible solutions and aims to determine better solutions while being designed to avoid traps at local minima.
We propose ATS by applying Eq.~\ref{eq:validation_grade} and~\ref{eq:mutation_exploration_reward} to TS to speed up the searching process. Specifically, we mutate numerous children models ($m_1$, $m_2$, .., $m_n$) from the current architecture ($m_c$). The mutation exploration reward $\mathcal{R}(m_c, m_i)$ is calculated using Eq.~\ref{eq:mutation_exploration_reward}.
ATS then chooses to train the mutation with the highest rewards (e.g. architecture $m_i$ as demonstrated in Figure~\ref{fig:search_algorithm}). The validation grade of this model $\mathcal{G}(m_i)$ is calculated after the training. The performance of the chosen model is assessed by comparing $\mathcal{G}(m_i)$ with $\mathcal{G}(m_c)$. If $\mathcal{G}(m_i)$ is larger than $\mathcal{G}(m_c)$, then $m_i$ is a good mutation, and we opt to build the next generation upon its structure. Otherwise, we swap to use the best option in the tabu list for the next mutation. The process stops when reaching a maximum number of iterations or achieving a terminal condition. The network ranking will be updated, and the search will continue for the remaining parent architectures.
\subsection{Implementation Details} \label{implementation_detail}
\noindent For searching, we directly perform our architecture exploration on the training samples of the target dataset. We set the target compactness parameter $P$ using the previously published compact models as a guideline. We set the maximum number of exploration iterations to 100 and stop the exploration procedure if a better solution cannot be found after 10 iterations. The total search time required to find optimal architecture is $\sim 4.3$ GPU days.
\begin{table}[b!]
\caption{\label{tab:eval_kitti}Evaluation on the KITTI dataset. Metrics with $\downarrow$ mean lower is better and $\uparrow$ mean higher is better.}
\centering
\small
\adjustbox{width=\columnwidth}{\begin{tabular}{@{}lrccccc@{}}
\hline
\textbf{Method} &\textbf{\#params} & \textbf{REL$\downarrow$} & \textbf{RMSE$\downarrow$} & \(\boldsymbol{\delta_{1}}\)$\uparrow$ & \(\boldsymbol{\delta_{2}}\)$\uparrow$ & \(\boldsymbol{\delta_{3}}\)$\uparrow$ \\ \hline
FastDepth~\cite{wofk2019fastdepth} & 3.93M & 0.156 & 5.628 & 0.801 & 0.930 & 0.971 \\ \hline
PyD-Net~\cite{poggi2018towards} & 1.97M & 0.154 & 5.556 & 0.812 & 0.932 & 0.970 \\ \hline
EQPyD-Net~\cite{cipolletta2021energy} & 1.97M & 0.135 & 5.505 & 0.821 & 0.933 & 0.970 \\ \hline
DSNet~\cite{aleotti2021real} & 1.91M & 0.159 & 5.593 & 0.800 & 0.932 & 0.971 \\ \hline
LiDNAS-K & 1.78M & \textbf{0.133} & 5.157 & 0.842 & \textbf{0.948} & 0.980 \\ \hline
LDP-Depth-K & \textbf{1.74M} & \textbf{0.133} & \textbf{5.155} & \textbf{0.844} & \textbf{0.948} & \textbf{0.981} \\ \hline
\end{tabular}} \vspace{-0.25cm}
\end{table}
\begin{table}[b!]
\vspace{-0.15cm}
\caption{\label{tab:eval_cityscapes}Segmentation results on Cityscapes dataset. (MIoU)}
\centering
\small
\begin{tabular}{@{}lrccc@{}}
\hline
{\textbf{Model}} & \textbf{\#params} & \textbf{resolution} & \textbf{val}$\uparrow$ & \textbf{test}$\uparrow$ \\ \hline
BiSeNetV1 B~\cite{yu2018bisenet} & 49.0M & $768 \times 1536$ & 74.8 & 74.7 \\ \hline
SwiftNet\cite{orsic2019defense} & 11.8M & $512 \times 1024$ & 70.2 & - \\ \hline
DFANet~\cite{li2019dfanet} & 7.8M & $512 \times 1024$ & 70.8 & 70.3 \\ \hline
BiSeNetV1 A~\cite{yu2018bisenet} & 5.8M & $768 \times 1536$ & 69.0 & 68.4 \\ \hline
MobileNeXt~\cite{zhou2020rethinking} & 4.5M & $1024 \times 2048$ & 75.5 & 75.2 \\ \hline
BiSeNetV2~\cite{yu2021bisenet} & 4.3M & $1024 \times 2048$ & 75.8 & 75.3 \\ \hline
HRNet-W16~\cite{wang2020deep} & 2.0M & $512 \times 1024$ & 68.6 & 68.1 \\ \hline
Lite-HRNet~\cite{yu2021lite} & 1.8M & $512 \times 1024$ & \textbf{76.0} & 75.3 \\ \hline \hline
FasterSeg~\cite{chen2019fasterseg} & 4.4M & $1024 \times 2048$ & 73.1 & 71.5 \\ \hline
MobileNetV3~\cite{howard2019searching} & \textbf{1.5M} & $1024 \times 2048$ & 72.4 & 72.6 \\ \hline
LDP-Seg-Ci & 1.7M & $512 \times 1024$ & 75.8 & \textbf{75.5} \\ \hline \hline
\end{tabular}
\vspace{-0.25cm}
\end{table}
\begin{table}[b!]
\caption{\label{tab:eval_coco}Segmentation results on COCO-Stuff dataset.}
\centering
\small
\begin{tabular}{@{}lrccc@{}}
\hline
{\textbf{Model}} & \textbf{\#params} & \textbf{PixAcc}(\%)$\uparrow$ & \textbf{MIoU}(\%)$\uparrow$ \\ \hline
BiSeNetV1 B~\cite{yu2018bisenet} & 49.0M & 63.2 & 28.1 \\ \hline
ICNet~\cite{zhao2018icnet} & 12.7M & - & 29.1 \\ \hline
BiSeNetV1 A~\cite{yu2018bisenet} & 5.8M & 59.0 & 22.8 \\ \hline
BiSeNetV2~\cite{yu2021bisenet} & \textbf{4.3M} & 63.5 & 28.7 \\ \hline
LDP-Seg-CO & \textbf{4.3M} & \textbf{64.2} & \textbf{29.3} \\ \hline \hline
\end{tabular}
\end{table}
For training, we use the Adam optimizer~\cite{kingma2014adam} with $(\beta_1, \beta_2, \epsilon) = (0.9, 0.999, 10^{-8})$. The initial learning rate is $7*10^{-4}$, but from epoch 10 the learning is reduced by $5\%$ per $5$ epochs. We use a batch size of 256 and augment the input RGB and ground truth depth images using random rotations ([-5.0, +5.0] degrees), horizontal flips, rectangular window droppings, and colorization (RGB only).
\section{Experiments}
We deploy the LDP framework on dense prediction tasks: monocular depth estimation, semantic segmentation, and image super-resolution. Experiments show that LDP improved performance while using only a fraction of the number of parameters needed by the competing approaches.
\subsection{Monocular Depth Estimation}
We first demonstrate our method for monocular depth estimation that can be formulated as a dense regression problem. The main goal is to infer continuous pixel-wise depth values from a single input image. For this task, we apply LDP on the NYU-Depth-v2 \cite{silberman2012indoor} and KITTI~\cite{geiger2013vision} datasets. NYU-Depth-v2 contains $\sim120K$ RGB-D images obtained from 464 indoor scenes. From the entire dataset, we use 50K images for training and the official test set of 654 images for evaluation. KITTI is an outdoor driving dataset, where we use the standard Eigen split~\cite{eigen2015predicting,eigen2014depth} for training (39K images) and testing (697 images). We report the mean absolute relative error (REL), root mean square error (RMSE), and thresholded accuracy ($\delta_i$) as our performance metrics.
\begin{table*}[t!]
\caption{\label{tab:sup_res_x2}Image Super-Resolution with scaling factor $\times 2$.}
\centering
\begin{tabular}{@{}lccccccccc@{}}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{ \textbf{Method} }} &
\multicolumn{1}{c}{\multirow{2}{*}{ \textbf{\#params} }} &
\multicolumn{2}{c}{ \textbf{Set5} } &
\multicolumn{2}{c}{ \textbf{Set14} } &
\multicolumn{2}{c}{ \textbf{BSD100} } &
\multicolumn{2}{c}{ \textbf{Urban100} } \\ \cmidrule(l){3-10}
\multicolumn{2}{c}{} & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ \\ \midrule
Bicubic & - & 33.66 & 0.9299 & 30.24 & 0.8688 & 29.56 & 0.8431 & 26.91 & 0.8425 \\
CARN~\cite{ahn2018fast} & 1.59M & 37.76 & 0.9590 & 33.52 & 0.9166 & 32.09 & 0.8978 & 31.92 & 0.9256 \\
LFFN~\cite{yang2019lightweight} & 1.52M & 37.95 & 0.9597 & 32.45 & 0.9142 & 32.20 & 0.8994 & 32.39 & 0.9299 \\
OISR-LF-s~\cite{he2019ode} & 1.37M & 38.02 & 0.9605 & 33.62 & 0.9178 & 32.20 & 0.9000 & 32.21 & 0.9290 \\
CBPN~\cite{zhu2019efficient} & 1.04M & 37.90 & 0.9590 & 33.60 & 0.9171 & 32.17 & 0.8989 & 32.14 & 0.9279 \\
OverNet~\cite{behjati2021overnet} & 0.90M & 38.11 & 0.9610 & 33.71 & 0.9179 & 32.24 & 0.9007 & 32.44 & 0.9311 \\
LapSRN~\cite{lai2017deep} & 0.81M & 37.52 & 0.9590 & 33.08 & 0.9130 & 31.80 & 0.8950 & 30.41 & 0.9100 \\
IMDN~\cite{hui2019lightweight} & 0.69M & 38.00 & 0.9605 & 33.63 & 0.9177 & 32.19 & 0.8996 & 32.17 & 0.9283 \\
MemNet~\cite{tai2017memnet} & 0.67M & 37.78 & 0.9597 & 33.28 & 0.9142 & 32.08 & 0.8978 & 31.31 & 0.9195 \\
VDSR~\cite{kim2016accurate} & 0.66M & 37.53 & 0.9587 & 33.05 & 0.9127 & 31.90 & 0.8960 & 30.77 & 0.9141 \\
IDN~\cite{hui2018fast} & 0.57M & 37.85 & 0.9598 & 33.58 & 0.9178 & 32.11 & 0.8989 & 31.95 & 0.9266 \\
CARN-M~\cite{ahn2018fast} & 0.41M & 37.53 & 0.9583 & 33.26 & 0.9141 & 31.92 & 0.8960 & 31.23 & 0.9194 \\
DRRN~\cite{tai2017image} & 0.29M & 37.74 & 0.9591 & 33.23 & 0.9136 & 32.05 & 0.8973 & 31.23 & 0.9188 \\
SRFBN~\cite{li2019feedback} & 0.28M & 37.78 & 0.9597 & 33.35 & 0.9156 & 32.00 & 0.8970 & 31.41 & 0.9207 \\
FSRCNN~\cite{dong2016accelerating} & \textbf{0.12M} & 37.00 & 0.9558 & 32.63 & 0.9088 & 31.53 & 0.8920 & 29.88 & 0.9020 \\ \hline
DeCoNASNet~\cite{ahn2021neural} & 1.71M & 37.96 & 0.9594 & 33.63 & 0.9175 & 32.15 & 0.8986 & 32.03 & 0.9265 \\
FPNet~\cite{esmaeilzehi2021fpnet} & 1.61M & \textbf{38.13} & \textbf{0.9616} & \textbf{33.83} & \textbf{0.9198} & \textbf{32.29} & 0.9018 & 32.04 & 0.9278 \\
FALSR-A~\cite{chu2021fast} & 1.03M & 37.82 & 0.9595 & 33.55 & 0.9168 & 32.12 & 0.8987 & 31.93 & 0.9256 \\
LDP-Sup-x2 & 1.02M & 38.11 & 0.9612 & \textbf{33.83} & 0.9196 & \textbf{32.29} & \textbf{0.9019} & \textbf{32.49} & \textbf{0.9314} \\
\hline \hline
\end{tabular}
\end{table*}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.99\linewidth]{figures/qualitative_cityscapes_v1.pdf}
\end{center}
\vspace{-0.35cm}
\caption{Comparison on the Cityscapes validation set.
(a) input image,
(b) ground truth,
(c) LDP-Seg-Ci,
(d) Lite-HRNet~\cite{yu2021lite},
and (e) MobileNetV3~\cite{howard2019searching}.}
\label{fig:qualitative_cityscapes} \vspace{-0.35cm}
\end{figure*}
In the case of NYU-Depth-v2, we set the target compactness $P=1.8M$ with the balance coefficient $\alpha=0.6$ to search for the optimized model on NYU-Depth-v2. We then select the best performance model (LDP-Depth-N) and compare its results with lightweight state-of-the-art methods~\cite{tu2020efficient,wofk2019fastdepth,wu2019fbnet,yucel2021real} along with their numbers of parameters. As shown in Table~\ref{tab:eval_nyuv2}, LDP-Depth-N outperforms the baseline while containing the least amount of parameters. Comparing with the best-performing approach, the proposed model improves the REL, RMSE, and $\theta_{1}$ by $11.4\%$, $8.2\%$, and $6.8\%$ while compressing the model size by $55\%$. Our method produces high-quality depth maps with sharper details as presented in Figure~\ref{fig:qualitative_nyu}. However, we observe that all methods still struggle in challenging cases, such as the scene containing Lambertian surfaces as illustrated by the example in the third column of Figure~\ref{fig:qualitative_nyu}. Moreover, the proposed method improves REL and RMSE by $22.3\%$ and $18.7\%$ while using only $3\%$ of the model parameters comparing to the state-of-the-art NAS-based disparity and depth estimation approaches~\cite{saikia2019autodispnet}. In addition, our method requires $90\%$ less search time than~\cite{saikia2019autodispnet} and outperforms~\cite{huynh2021lightweight} in almost all metrics.
\begin{table*}[t!]
\caption{\label{tab:sup_res_x4}Image Super-Resolution with scaling factor $\times 4$.}
\centering
\begin{tabular}{@{}lccccccccc@{}}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{ \textbf{Method} }} &
\multicolumn{1}{c}{\multirow{2}{*}{ \textbf{\#params} }} &
\multicolumn{2}{c}{ \textbf{Set5} } &
\multicolumn{2}{c}{ \textbf{Set14} } &
\multicolumn{2}{c}{ \textbf{BSD100} } &
\multicolumn{2}{c}{ \textbf{Urban100} } \\ \cmidrule(l){3-10}
\multicolumn{2}{c}{} & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ & PSNR$\uparrow$ & SSIM$\uparrow$ \\ \midrule
Bicubic & - & 28.42 & 0.8104 & 26.01 & 0.7027 & 25.96 & 0.6675 & 23.17 & 0.6585 \\
s-LWSR64~\cite{li2020s} & 2.27M & 32.28 & 0.8960 & 28.34 & 0.7800 & 27.61 & 0.7380 & 26.19 & 0.7910 \\
CARN~\cite{ahn2018fast} & 1.59M & 32.13 & 0.8937 & 28.60 & 0.7806 & 27.58 & 0.7349 & 26.07 & 0.7837 \\
LFFN~\cite{yang2019lightweight} & 1.53M & 32.15 & 0.8945 & 28.32 & 0.7810 & 27.52 & 0.7377 & 26.24 & 0.7902 \\
OISR-LF-s~\cite{he2019ode} & 1.52M & 32.14 & 0.8947 & 28.63 & 0.7819 & 27.60 & 0.7369 & 26.17 & 0.7888 \\
CBPN~\cite{zhu2019efficient} & 1.19M & 32.21 & 0.8944 & 28.63 & 0.7813 & 27.58 & 0.7356 & 26.14 & 0.7869 \\
LapSRN~\cite{lai2017deep} & 0.81M & 31.54 & 0.8850 & 28.19 & 0.7720 & 27.32 & 0.7280 & 25.21 & 0.7560 \\
IMDN~\cite{hui2019lightweight} & 0.72M & 32.21 & 0.8948 & 28.58 & 0.7811 & 27.56 & 0.7353 & 26.04 & 0.7838 \\
MemNet~\cite{tai2017memnet} & 0.67M & 31.74 & 0.8893 & 28.26 & 0.7723 & 27.40 & 0.7281 & 25.50 & 0.7630 \\
VDSR~\cite{kim2016accurate} & 0.66M & 31.33 & 0.8838 & 28.02 & 0.7678 & 27.29 & 0.7252 & 25.18 & 0.7525 \\
IDN~\cite{hui2018fast} & 0.60M & 31.99 & 0.8928 & 28.52 & 0.7794 & 27.52 & 0.7339 & 25.92 & 0.7801 \\
s-LWSR32~\cite{li2020s} & 0.57M & 32.04 & 0.8930 & 28.15 & 0.7760 & 27.52 & 0.7340 & 25.87 & 0.7790 \\
SRFBN~\cite{li2019feedback} & 0.48M & 31.98 & 0.8923 & 28.45 & 0.7779 & 27.44 & 0.7313 & 25.71 & 0.7719 \\
CARN-M~\cite{ahn2018fast} & 0.41M & 31.92 & 0.8903 & 28.42 & 0.7762 & 27.44 & 0.7304 & 25.62 & 0.7694 \\
DRRN~\cite{tai2017image} & 0.29M & 31.68 & 0.8888 & 28.21 & 0.7720 & 27.38 & 0.7284 & 25.44 & 0.7638 \\
FSRCNN~\cite{dong2016accelerating} & \textbf{0.12M} & 30.71 & 0.8657 & 27.59 & 0.7535 & 26.98 & 0.7150 & 24.62 & 0.7280 \\ \hline
FPNet~\cite{esmaeilzehi2021fpnet} & 1.61M & \textbf{32.32} & 0.8962 & \textbf{28.78} & \textbf{0.7856} & 27.66 & \textbf{0.7394} & 26.09 & 0.7850 \\
LDP-Sup-x4 & 1.09M & 32.30 & \textbf{0.8963} & 28.54 & 0.7836 & \textbf{27.67} & \textbf{0.7394} & \textbf{26.25} & \textbf{0.7927} \\
\hline \hline
\end{tabular}
\end{table*} \vspace{-0.15cm}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.99\linewidth]{figures/qualitative_super_v1.pdf}
\end{center}
\label{fig:qualitative_super}
\vspace{-0.35cm}
\caption{Comparison on the Urban100 and BSD100 dataset.} \vspace{-0.35cm}
\end{figure*}
For KITTI, we aim at the target compactness of $P=1.45M$ with $\alpha=0.55$. We train our candidate architectures with the same self-supervised procedure proposed by~\cite{godard2019digging} and adopted by the state-of-the-art approaches~\cite{aleotti2021real,cipolletta2021energy,poggi2018towards,wofk2019fastdepth}. After the search, we pick the best architecture (LDP-Depth-K) to compare with the baselines and report the performance figures in Table~\ref{tab:eval_kitti}. The LDP-Depth-K model yields competitive results with the baselines while also being the smallest model. We observe that our proposed method provides noticeable improvement from PyD-Net and EQPyD-Net. Examples from Figure~\ref{fig:qualitative_kitti} show that the predicted depth maps from LDP-Depth-K are more accurate and contain fewer artifacts.
\subsection{Semantic Segmentation}
We then deploy LDP for dense classification tasks such as semantic segmentation that aims to predict discrete labels of image pixels. We employ the same backbone structure as in monocular depth estimation experiments and utilize the cross-entropy loss. For this problem, we evaluate our method on the Cityscapes~\cite{cordts2016cityscapes} and COCO-stuff~\cite{caesar2018coco} datasets. Cityscapes is an outdoor dataset containing images of various urban scenarios. The dataset consists of 5K high-quality annotated frames that 19 classes are used for semantic segmentation. Following the standard procedure, we used 2975, 500, and 1525 images for training, validation, and testing, respectively. COCO-stuff was created by annotating dense stuffs (e.g., sky, ground, walls) from the COCO dataset. This dataset can be utilized for image understanding as it contains 91 more stuffs classes compared to the original dataset. For fair comparison, we also use the COCO-Stuff-10K with 9K and 1K for training and testing purposes. We utilize both the mean Intersection-over-Union (MIoU) as well as the pixel accuracy (pixAcc) to assess the performance of our models.
To perform searching on Cityscapes dataset, we set the target compactness $P=1.25M$ with the balance coefficient $\alpha=0.6$ using input image resolution of $512 \times 1024$ for all experiments. We then compare the best-performing model (LDP-Seg-Ci) with recent approaches~\cite{yu2018bisenet,li2019dfanet,orsic2019defense,yu2021lite,chen2019fasterseg,howard2019searching}. Results in Table~\ref{tab:eval_cityscapes} suggest that LDP-Seg-Ci performs on par with state-of-the-art while using fewer parameters than most methods. Although operating at a lower resolution, our generated model outperforms FasterSeg~\cite{chen2019fasterseg} while being $61\%$ more compact in terms of the number of parameters. Moreover, LDP-Seg-Ci also shows clear improvements compared to the MobileNetV3~\cite{howard2019searching} model. Qualitative results in Figure~\ref{fig:qualitative_cityscapes} also show that our model tends to produce more clean with sharper object boundaries and less cluttering than state-of-the-art approaches.
In the case of the COCO-stuff, we aim at the target compactness of $P=4.0M$ with the balance coefficient $\alpha$ set to $0.6$. During searching and testing, we crop the input into $640 \times 640$ resolution. We evaluate the performance of our best architecture (LDP-Seg-CO) with current state-of-the-art methods~\cite{yu2018bisenet,zhao2018icnet,yu2021bisenet} and report the results in Table~\ref{tab:eval_coco}. Our method also achieves good performance for semantic segmentation on the COCO-stuff dataset while using much fewer parameters than competing approaches.
\subsection{Image Super-resolution}
To further assess the applicability of our proposed framework for dense prediction problems, we apply LDP framework for the image super-resolution task. We also employ a similar scheme as in previous experiments with added upsample blocks between decoder and refinement blocks to increase spatial dimension within the network scale. We then perform architecture search and training on the DIV2K~\cite{agustsson2017ntire} dataset. DIV2K is a high-quality image super-resolution dataset consisting of 800 samples for training, 100 for validation, and 100 for testing purposes. After that, we test our generated models on standard benchmarks such as Set5~\cite{bevilacqua2012low}, Set14~\cite{zeyde2010single}, B100~\cite{arbelaez2010contour}, and Urban100~\cite{huang2015single}. The results of our method are evaluated using the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) metrics on the Y channel of the YCbCr color space.
\begin{table}[!b]
\vspace{-0.45cm}
\caption{\label{tab:runtime_comparison}Average runtime comparison of the proposed method and other lightweight models. Runtime values are measured using a Pixel 3a phone with input image resolution ($640 \times 480$).}
\centering
\small
\begin{tabular}{lc}
\hline
\textbf{Architecture} & \textbf{CPU(ms)} \\ \hline
Ef+FBNet~\cite{tu2020efficient,wu2019fbnet} & 852 \\ \hline
FSRCNN~\cite{dong2016accelerating} & 789 \\ \hline
FastDepth~\cite{wofk2019fastdepth} & 458 \\ \hline
VDSR~\cite{kim2016accurate} & 425 \\ \hline
PyD-Net~\cite{poggi2018towards} & 226 \\ \hline
Lite-HRNet~\cite{yu2021lite} & 217 \\ \hline
LiDNAS-K~\cite{huynh2021lightweight} & 205 \\ \hline
LDP-Seg-Ci & 207 \\ \hline
LDP-Depth-K & 204 \\ \hline
LDP-Sup-x2 & \textbf{203} \\ \hline
\end{tabular}
\end{table}
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=0.9\linewidth]{figures/searching_scenarios.pdf}
\end{center}
\vspace{-0.35cm}
\caption{Trade-off between accuracy vs. the number of parameters of best models trained with different searching scenarios on NYU-Depth-v2, Cityscapes and testing on Urban100 dataset.} \vspace{-0.45cm}
\label{fig:searching_scenarios} %
\end{figure*}
We search for optimized models at super-resolution scales $\times 2$ and $\times 4$ on the DIV2K dataset. In both cases, we determine the target compactness of $P=0.8$ with the balance coefficient $\alpha = 0.55$. We then compare the best-performing architectures (LDP-Sup-x2 and LDP-Sup-x4) with recent methods~\cite{esmaeilzehi2021fpnet,ahn2021neural,ahn2018fast,hui2019lightweight,chu2021fast,lai2017deep} on various testing benchmarks. Tables 10 and 11 show that our models produce more competitive results than several state-of-the-art image super-resolution methods while being relatively compact. Our models perform on par with FPNet while being at least $32\%$ smaller in terms of network size. Figure 10 provides visual comparisons on BSD100 and Urban100 benchmarks. The proposed method yields more accurate results than VDSR~\cite{kim2016accurate}, DRRN~\cite{tai2017image}, FSCRNN~\cite{dong2016accelerating} and more precise details than SRCNN~\cite{dong2014learning}, IMDN~\cite{hui2019lightweight}, FPNet~\cite{esmaeilzehi2021fpnet} methods.
\begin{figure}[!b]
\begin{center}
\includegraphics[width=0.72\linewidth]{figures/qualitative_kitti_v2.pdf}
\end{center}
\vspace{-0.35cm}
\caption{Comparison on the Eigen split of KITTI.
(a) input image,
(b) LDP-Depth-K,
(c) LiDNAS-K~\cite{huynh2021lightweight},
(d) DSNet~\cite{aleotti2021real},
(e) PyD-Net~\cite{poggi2018towards},
and (f) FastDepth~\cite{wofk2019fastdepth}.
Images in the right column presented zoom-in view for better visualization.}
\label{fig:qualitative_kitti} \vspace{-0.35cm}
\end{figure}
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=0.84\linewidth]{figures/convergence_v2.pdf}
\end{center}
\vspace{-0.35cm}
\caption{The progress of different searching scenarios on the NYU-Depth-v2, Cityscapes and testing on Urban100 dataset. Charts show the metrics including thresholded accuracy, Mean IoU, peak-signal-to-noise-ratio (PSNR) and the number of parameters vs. the number of searching iterations.} \vspace{-0.25cm}
\label{fig:convergence_v2}
\end{figure*}
\begin{figure}[!b]
\begin{center}
\includegraphics[width=0.99\linewidth]{figures/grid_search.pdf}
\end{center}
\vspace{-0.25cm}
\caption{Grid search using randomly subsampled sets from the training data to look for good balance coefficient values on NYU-Depth-v2 (left) and KITTI (right).}
\label{fig:grid_search} \vspace{-0.25cm}
\end{figure}
\subsection{Runtime Measurement}
We also compare the runtime of our models with state-of-the-art lightweight methods on an Android device using the app from the Mobile AI benchmark developed by Ignatov et al.~\cite{ignatov2021fast}. To this end, we utilize the pre-trained models provided by the authors (Tensorflow~\cite{poggi2018towards}, PyTorch~\cite{wofk2019fastdepth}), convert them to \textit{tflite} and measure their runtime on mobile CPUs.
The results in Table~\ref{tab:runtime_comparison} suggest that the proposed approaches produce competing performance, with the potential of running real-time on mobile devices with further optimization.
\subsection{Exploration Convergence}
We experiment with various settings for the multi-objective balance coefficient ($\alpha$) to assess its effect on the performance. For this purpose, we perform the architecture search with $\alpha$ set to $0.0$, $0.4$, $0.5$, $0.6$, and $1.0$ while the target compactness $P=2.0M$. Figure~\ref{fig:convergence_v2} presents the searching progress for accuracy (left), the number of parameters (center), and validation grade (right) from one parent architecture on NYU-Depth-v2. We observe that, scenario with $\alpha=0.0$ quickly becomes saturated as it only gives reward to the smallest model. Searching with $\alpha=0.4$ favors models with compact size but also with limited accuracy. The case with $\alpha=0.5$ provides a more balanced option, but accuracy is hindered due to fluctuation during searching. The exploration with $\alpha=1.0$ seeks for the network with the best accuracy yet producing significantly larger architecture while the case where $\alpha=0.6$ achieves promising accuracy although with slightly bigger model than the target compactness.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.935\linewidth]{figures/qualitative_nyu_v1.pdf}
\end{center}
\vspace{-0.5cm}
\caption{Comparison on the NYU test set.
(a) input image,
(b) ground truth,
(c) LDP-Depth-N,
(d) LiDNAS-N~\cite{huynh2021lightweight},
(e) Ef+FBNet~\cite{tu2020efficient,wu2019fbnet},
and (f) FastDepth~\cite{wofk2019fastdepth}.}
\label{fig:qualitative_nyu} \vspace{-0.5cm}
\end{figure*}
\subsection{Searching Scenarios}
To further analyze the outcome of different searching scenarios, we perform architecture searches for \textit{six} parent networks in five settings with $\alpha=0.0, 0.4, 0.5, 0.6, 1.0$ and $P=2.0M$ on NYU-Depth-v2. Results in Figure~\ref{fig:searching_scenarios} show that the best performance models in the case of $\alpha=0.5$ are more spread out, while training instances with $\alpha=0.6$ tend to produce both accurate and lightweight architectures. This, in turn, emphasizes the trade-off between validation accuracy and the network size.
\subsection{Balance Coefficient Search}
Determining a good balance coefficient value $(\alpha)$ for Eq.~\ref{eq:validation_grade} and~\ref{eq:mutation_exploration_reward} is crucial as it greatly affects the search performance. To this end, we perform grid search on randomly subsampled sets from the training data seeking for the optimized $\alpha$ value. The pattern in Figure~\ref{fig:grid_search} shows that, for NYU-Depth-v2 [4] and KITTI [2], approximately good $\alpha$ values range from $0.5$ to $0.6$. Additionally, the grid search is much faster (only requires $\sim 15$ hours on one dataset), enabling finding good $\alpha$ values when deploying to different datasets.
\section{Conclusion}
\noindent This paper proposed a novel NAS framework to construct lightweight dense prediction architectures using Assisted Tabu Search and employing a well-defined search space for balancing layer diversity and search volume. The proposed method achieves competitive accuracy on diverse datasets while running faster on mobile devices and being more compact than state-of-the-art handcrafted and automatically generated models. Our work provides a potential approach towards optimizing the accuracy and the network size for dense prediction without the need for manual tweaking of deep neural architectures.
{\small
\bibliographystyle{ieee_fullname}
| 9b2be2fab67bb4c36d2cf978330385b999bad0d3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Conclusion}
This paper presents a data collection system that is portable and enables large-scale data collection. Our systems offers better utility for pedestrian behavior research because our systems consists of human verified labels grounded in the metric space, a combination of both top-down views and perspective views, and a human-pushed cart that approximates naturalistic human motion with a socially-aware ``robot". We further couple the system setup with a semi-autonomous labeling process that easily produces human verified labels in order to meet the demands of the large-scale data collected by our hardware. Lastly, we present the TBD pedestrian dataset we have collected using our system, which not only surpasses the quantity of similar datasets, but also offers unique pedestrian interaction behavior that adds to the qualitative diversity of pedestrian interaction data.
A key concern about our current data collection setup is that our sensors consist purely of cameras. For better labeling accuracy, we are exploring whether adding a LiDAR will aid the autonomous tracking of pedestrians and produce more accurate labels. We also plan to continue making improvements to our software system and underlying methods. Although the semi-autonomous labeling process speeds up the labeling of pedestrians significantly, the bottleneck to produce huge quantities of data still lies in correcting the few erroneous tracking outcomes of the automatic tracking procedures. A centralized user interface is under development to better document these tracking errors and to provide intuitive tools to fix them.
As mentioned earlier, our approach enables additional data collection in a wide range of locations and constraints. Additional data collection and public updates to this initial dataset are planned. In particular, we would like to collect additional data from the same atrium to increase the current sample size and possibly see more variability in behavior due to population shifts (university populations are constantly changing).
Our goal is to increase usability by others and inspire more datasets to be generated using our approach. Interested parties should note that local ethics regulations may require care and limits on what can be released. Our dataset was collected under Institutional Review Board (IRB) oversight, including aspects related to public data sharing. For example, we posted signs at all entry points indicating recording was in progress and suggested alternate routes for those who did not wish to be filmed. This may be less necessary in locations where there is less expectation of privacy (e.g., extensive security cameras, locations with high frequency of social media recording, very public settings, etc).
In closing, this paper documents a new method for collecting naturalistic pedestrian behavior. A novel dataset is also provided to illustrate how this technique provides value over existing datasets and so that other groups can advance their own research. We hope this effort enables many new discoveries.
\section{Evaluation} \label{sec:evaluation}
\subsection{Comparison with Existing Datasets} \label{sec:eval-compare}
Compared to existing datasets collected in pedestrian natural environments, our TBD pedestrian dataset contains three components that greatly enhances the dataset's utility. These components are:
\begin{enumerate}
\item \textbf{Human verified labels grounded in metric space.} As mentioned in section \ref{sec:related-dsetuse}, ETH \cite{ETH} and UCY \cite{UCY} datasets are very popular and are the only datasets to be included during the evaluation of various research models in many papers. This is largely because the trajectory labels in these datasets are human verified, unlike \cite{edinburgh}, \cite{cff}, \cite{grandcentral}, and \cite{atc} that solely rely on automatic tracking to produce labels. These trajectory labels are also grounded in metric space rather than image space (e.g. \cite{stanforddrone} and \cite{towncentre} only contain labels in bounding boxes). Having labels grounded in metric space eliminates the possibility that camera poses might have an effect on the scale of the labels. It also makes the dataset useful for robot navigation related research because robots plan in the metric space rather than image space.
\item \textbf{Combination of top-down views and perspective views.} Similar to datasets with top-down views, we use top-down views to obtain ground truth trajectory labels for every pedestrian present in the scene. Similar to datasets with perspective views, we gather perspective views from a ``robot" to imitate robot perception of human crowds. A dataset that contains both top-down views and perspective views will be useful for research projects that rely on perspective views. This allows perspective inputs to their models, while still having access to ground truth knowledge of the entire scene. Examples include pedestrian motion prediction given partial observation of the scene and robot navigation research projects that only have onboard sensors as inputs to navigation models.
\item \textbf{Naturalistic human behavior with the presence of a ``robot".} Unlike datasets such as \cite{lcas} or \cite{jrdb}, the ``robot" that provides perspective view data collection is a cart being pushed by human. As mentioned in section \ref{sec:hardware}, doing so reduces the novelty effects from the surrounding pedestrians. Having the ``robot" being pushed by humans also ensures safety for the pedestrians and its own motion has more natural human behavior. As such, the pedestrians also react naturally around the robot by treating it as another human agent.
\end{enumerate}
\begin{table}[ht]
\caption{A survey of existing pedestrian datasets on how they incorporate the three components in section \ref{sec:eval-compare}. For component 1, a ``No" means either not human verified or not grounded in metric space. For component 2, TD stands for ``top-down view" and ``P" stands for ``perspective view".}
\label{tab:survey}
\begin{center}
\begin{tabular}{c||ccc}
\toprule
Datasets & Comp. 1 & Comp. 2 & Comp. 3\\
& (metric labels) & (views) & (``robot") \\
\hline
TBD (Ours) & Yes & TD + P & Human + Cart \\
ETH \cite{ETH} & Yes & TD & N/A \\
UCY \cite{UCY} & Yes & TD & N/A \\
Edinburgh Forum \cite{edinburgh} & No & TD & N/A \\
VIRAT \cite{virat} & No & TD & N/A \\
Town Centre \cite{towncentre} & No & TD & N/A \\
Grand Central \cite{grandcentral} & No & TD & N/A \\
CFF \cite{cff} & No & TD & N/A \\
Stanford Drone \cite{stanforddrone} & No & TD & N/A \\
L-CAS \cite{lcas} & No* & P & Robot\\
WildTrack \cite{wildtrack} & Yes & TD & N/A\\
JackRabbot \cite{jrdb} & Yes & P & Robot\\
ATC \cite{atc} & No & TD & N/A\\
TH\"OR \cite{thor} & Yes & TD + P & Robot\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
As shown in Table \ref{tab:survey}, current datasets only contain at most two of the three components\footnote{*L-CAS dataset does provide human verified labels grounded in the metric space. However, its pedestrian labels do not contain trajectory data, which means this dataset has limited usage in pedestrian behavior research, so we put ``No" here.}. A close comparison is the TH\"OR dataset \cite{thor}, but its perspective view data are collected by a robot. Additionally, unlike all other datasets in Table \ref{tab:survey}, the TH\"OR dataset is collected in a controlled lab setting rather than in the wild. This injects artificial factors into human behavior, making them unnatural.
\subsection{Dataset Statistics} \label{sec:eval-stats}
\begin{table}[ht]
\caption{Comparison of statistics between our dataset and other datasets that provide human verified labels grounded in the metric space. For total time length, 51 minutes of our dataset includes the perspective view data.}
\label{tab:stats}
\begin{center}
\begin{tabular}{c||ccc}
\toprule
Datasets & Time length & \# of pedestrians & Label freq (Hz)\\
\hline
\multirow{2}{*}{TBD (Ours)} & 133 mins & \multirow{2}{*}{1416} & \multirow{2}{*}{60} \\
& (51 mins) & & \\
ETH \cite{ETH} & 25 mins & 650 & 15 \\
UCY \cite{UCY} & 16.5 mins & 786 & 2.5 \\
WildTrack \cite{wildtrack} & 200 sec & 313 & 2\\
JackRabbot \cite{jrdb} & 62 mins & 260 & 7.5\\
TH\"OR \cite{thor} & 60+ mins & 600+ & 100\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
Table \ref{tab:stats} demonstrates the benefit of a semi-automatic labeling pipeline. With the aid of an autonomous tracker, humans only need to verify and make occasional corrections the tracking outcomes instead of locating the pedestrians on every single frame. The data we have collected so far already surpassed all other datasets that provide human verified labels in the metric space in terms of total time, number of pedestrians and labeling frequency. We will continue this effort and collect more data for future works.
It is worth noting that the effect of noise becomes larger with higher labeling frequency. We provide high frequency labeling so that more information and details can be available on the trajectories. When using our data, we recommend downsampling so that noise will have a lesser effect on pedestrian behavior modeling.
\subsection{Qualitative Pedestrian Behavior}
\begin{figure}[thpb]
\centering
\includegraphics[scale=0.68]{imgs/qual.jpg}
\caption{Example scenes from the TBD pedestrian dataset. a) a dynamic group. b) a static conversational group. c) a large tour group with 14 pedestrians. d) a pedestrian affecting other pedestrians' navigation plans by asking them to come to the table. e) pedestrians stop and look at their phones. f) two pedestrians change their navigation goals and turn towards the table. g) a group of pedestrians change their navigation goals multiple times. h) a crowded scene where pedestrians are heading towards different directions.}
\label{fig:qual}
\end{figure}
Due to the nature of the environment where we collected the data as described in Section \ref{sec:hardware}, we observe a mixture of corridor and open space pedestrian behavior, many of which are rarely seen in other datasets. As shown in Figure \ref{fig:qual}, we observe both static conversation groups and dynamic walking groups. In one instance, a tour group of 10+ pedestrians entered our scene. We also observe that some pedestrians naturally change goals mid-navigation, which results in turning behavior. Due to the timing of our data collection, we also observe ongoing activities where several students set up tables and engage people passing by. This activity produces additional interesting pedestrian interaction analogous to sellers touting and buyers browsing.
\section{Introduction}
Pedestrian datasets are essential tools for designing socially appropriate robot behaviors, recognizing and predicting human actions, and studying pedestrian behavior. A generally accepted assumption for these datasets is that real-world pedestrians are experts in analyzing and navigating human crowds because they are proficient at behaving in accordance to social interaction norms. Behavioral or practical research related to pedestrian motion likely involves constructing a model that captures these social interactions and movements. In general, existing datasets have been collected in support specific research questions, leading to inadvertent limitations on utility towards certain research questions. This paper describes our efforts to collect and create a dataset that supports a larger array of research questions.
\begin{figure}[h]
\centering
\includegraphics[scale=0.31]{imgs/combo.jpg}
\caption{Our dataset consists of human verified labeling in metric space, a combination of top-down views and perspective views, and a cart to imitate socially appropriate robot behavior. This set of images represent the same moment recorded from multiple sensors: a) Top-down view image taken by a static camera with ground truth pedestrian trajectory labels shown. b) Perspective-view image from a 360 camera that captures high definition videos of nearby pedestrians. c) Perspective-view RBG and depth images from a stereo camera mounted on a cart that is used to imitate onboard robot sensors.}
\label{fig:intro}
\end{figure}
For example, researchers may use these data to predict future pedestrian motions, including forecasting their trajectories \cite{Alahi1, Gupta1, Sophie, Social-STGCNN, ivanovic-trajectron}, and/or navigation goals \cite{kitani-2012, liang2020garden}. In social navigation, datasets can also be used to model interactions. For example, a key problem researchers have tried to address is the \textit{freezing robot problem} \cite{Trautman1}, in which the robot becomes stuck in dense, crowded situations while trying to be deferential to human movements for safety or end user acceptance reasons. Researchers have attributed this problem to robot's inability to model interactions \cite{sun2021move}. In other words, most current navigation algorithms do not consider pedestrian reactions and assume a non-cooperative environment. Some works \cite{nishimura2020risk} have used datasets to show that modeling the anticipation of human reactions to the robot's actions enables the robot to deliver a better performance.
However, interactions are diverse and can be rare occurrences in human crowds. Although robotic systems typically have access to each pedestrian's basic properties (e.g., position and velocity), inter-pedestrian interactions are less frequent because interactions require the presence of two or more pedestrians that usually need to be in close proximity of each other. While data documenting interactions is more limited, some work has made progress on this front. For example, Sch\"oller et al. \cite{constant-velocity-model} has shown that a linear acceleration based method can perform comparably with deep learning based models in pedestrian trajectory prediction settings. This implies that pedestrians mostly walk in linear fashion, a default behavior when not interacting with other pedestrians. Additionally, pedestrian interactions can be very diverse, especially in certain contexts. Some categories of interactions that researchers have devised include collision avoidance, grouping \cite{wang-split-merge}, and leader-follower \cite{kothari2021human}. The details of these types of interactions can further be diversified by the environment (e.g. an open plaza or a narrow corridor). Mavrogiannis et al. \cite{mavrogiannis_etal2021-core-challenges} provides more details on interaction types.
In order to better capture and model interactions to improve the performance of various pedestrian-related algorithms, considerably more data is needed across a variety of environments. To this end, we have constructed a data collection system that can achieve these two requirements: large quantity and environment diversity. First, we ensure that our equipment is completely portable and easy to set up. This allows collecting data in a variety of locations with limited lead time. Second, we address the challenge of labeling large quantities of data using a semi-autonomous labelling pipeline. We employ a state-of-the-art deep learning based \cite{zhang2021bytetrack} tracking module combined with various post-processing procedures to automatically produce high quality ground truth pedestrian trajectories in metric space.
As mentioned earlier, current datasets tend to be focused on specific pedestrian research questions. In contrast, our dataset approach offers various improvements and aims to accommodate a wide variety of pedestrian behavior research. Specifically, we include three important characteristics: (1) ground truth labeling in metric space, (2) perspective views from a moving agent, and (3) natural human motion. To the best of our knowledge, publicly available datasets only have two of these characteristics, but not all three.
To achieve this, we use multiple static cameras to ensure greater labelling accuracy. We offer both top-down and perspective views with the perspective-views supplied by cameras mounted on a cart. We use a cart pushed by one of our researchers to imitate a robot navigating through the crowd. Using a cart instead of a robot reduces the novelty effect from pedestrians \cite{brvsvcic2015escaping}, thereby capturing more natural pedestrian reactions, and increases the naturalness of the perspective-view ego motion.
In this paper we also demonstrate our system through a dataset collected in a large indoor space: the TBD pedestrian dataset\footnote{\href{https://tbd.ri.cmu.edu/tbd-social-navigation-datasets}{https://tbd.ri.cmu.edu/tbd-social-navigation-datasets}}. Our dataset contains scenes that with a variety of crowd densities and pedestrian interactions that are unseen in other datasets. This dataset can be used to complement existing datasets by injecting a new data environment and more pedestrian behavior distribution into existing dataset mixtures, such as \cite{kothari2021human}.
In summary, our contribution are as follows:
\begin{itemize}
\item We implement a novel data collection system that is portable and allows large-scale data collection. Our system also contains a pushed cart with mounted cameras to simulate robot navigation. This allows naturalistic data to be collected from a perspective view on a dynamic agent, thereby enabling model performance validation for robots lacking overhead views from infrastructure.
\item We devise a semi-autonomous labeling pipeline that enables convenient grounding of pedestrians. This pipeline consists of a deep learning-based pipeline to track pedestrians and downstream procedures to generate pedestrian trajectories in the metric space.
\item We provide a high quality large-scale pedestrian dataset. The data are collected both from overhead and perspective views and are labelled in both pixel space and the metric space for more practical use (e.g., in a social navigation setting).
\end{itemize}
\section{Related Work}
\subsection{Pedestrian Data in Research} \label{sec:related-dsetuse}
As is expected from the explosion of data-hungry machine learning methods in robotics, demand for pedestrian datasets has been on the rise in recent years. One popular category of research in this domain is human trajectory prediction (e.g., \cite{Alahi1, Gupta1, Sophie, Social-STGCNN, ivanovic-trajectron, kitani-2012, liang2020garden, wang-split-merge}). Much of this research utilizes selected mechanisms to model pedestrian interactions in hopes for better prediction performance (e.g., pooling layers in the deep learning frameworks \cite{Alahi1, Gupta1} or graph-based representations \cite{Social-STGCNN}). Rudenko et al. \cite{rudenko2019-predSurvey} provides a good summary into this topic. While the state-of-the art performance keeps improving with the constant appearance of newer models, it is often unclear how well these models can generalize in diverse environments. As shown in \cite{rudenko2019-predSurvey}, many of these models only conduct their evaluation on the relatively small-scale ETH \cite{ETH} and UCY \cite{UCY} datasets.
Another popular demand for pedestrian datasets comes from social navigation research. Compared to human motion prediction research, social navigation research focuses more on planning. For example, social navigation research uses learning-based methods to identify socially appropriate motion for better robot behavior, such as deep reinforcement learning \cite{Everett18_IROS, chen2019crowd, Chen-gaze-learn} or inverse reinforcement learning \cite{okal-IRL, Tai-IRL}. Due to the lack of sufficiently large datasets, these models often train in simulators that lack realistic pedestrian behavior. Apart from training, datasets are also increasing in popularity in social navigation evaluation due to their realistic pedestrian behavior \cite{gao2021evaluation}. Social navigation methods are often evaluated in environments using pedestrian data trajectory playback (e.g., \cite{trautmanijrr, cao2019dynamic, sun2021move, wang2022group}). However, similar to human motion prediction research, these evaluations are typically only conducted on the ETH \cite{ETH} and UCY \cite{UCY} datasets as shown by \cite{gao2021evaluation}. These two datasets only use overhead views, and therefore lack the perspective view used by most robots. Comparisons between an intial dataset from our data collection system and existing datasets can be found in section \ref{sec:eval-compare}.
\subsection{Simulators and Pedestrian Datasets} \label{sec:related-sim-dset}
Simulators can fill in the role of datasets for both training and evaluation. Simulators such as PedSIM \cite{gloor2016pedsim}, CrowdNav \cite{chen2019crowd}, SocNavBench \cite{biswas2021socnavbench} and SEAN \cite{tsoi2020sean} are in use by the research community. However, sim-to-real transfer is an unsolved problem in robotics. Apart from lack of fidelity in visuals and physics, pedestrian simulators in particular entail the additional paradox of pedestrian behavior realism \cite{mavrogiannis2021core}: If pedestrian models are realistic enough for use in simulators, why don't we apply the same model to social navigation?
In contrast, naturalistic datasets provide realistic pedestrian behavior. Unfortunately, datasets are limited in quantity, unlike simulators that can generate infinite pedestrian scenes. As mentioned in section \ref{sec:related-dsetuse}, most research is still limited to only the ETH and UCY datasets, which are small in scale and lack perspective views. Therefore, such datasets have an additional downside in that pedestrians do not react to the robot. While perspective views can be simulated using inferred laser scans from point perspectives (e.g., \cite{wang-split-merge}), this does not fill the need for camera data from perspective views. Also note that tying the simulated laser scanner location to a moving pedestrian in the data set will likely have unwanted noise in the human tracking.
\section{System Description}\label{sec:system}
In this work, we introduce a data collection system that is portable and easy to setup that will allow easy collection of large quantities of data. The data collection setup also contains a cart that provides data on naturalistic pedestrian reactions to the robot from a typical perspective view.
\subsection{Hardware Setup}\label{sec:hardware}
\begin{figure}[tbhp]
\centering
\includegraphics[scale=0.35]{imgs/cameras.jpg}
\caption{Sensor setup used to collect the TBD pedestrian dataset. (left) one of three nodes used to used to capture top-down RGB views. Each node is self contained with an external battery and communicates wirelessly with other nodes.
(right) cart used to capture sensor views from the mobile robot perspective during data collection. The cart is powered by an onboard power bank and laptop for time synchronization}
\label{fig:camera}
\end{figure}
As shown in Figure~\ref{fig:hdc_setup}, we positioned three FLIR Blackfly RGB cameras (Figure~\ref{fig:camera}) surrounding the scene on the upper floors overlooking the ground level at roughly 90 degrees apart from each other. Compared to a single overhead camera, multiple cameras ensure better pedestrian labeling accuracy. This is achieved by labeling the pedestrians from cameras that have the highest image resolution of the pedestrians (i.e., closest to pedestrians). The RGB cameras are connected to portable computers powered by lead-acid batteries. We also positioned three more units on the ground floor, but did not use them for pedestrian labeling.
In addition to the RGB cameras, we pushed a cart through the scene (Figure~\ref{fig:camera}), which was equipped with a ZED stereo camera to collect both perspective RGB views and depth information of the scene. A GoPro Fusion 360 camera for capturing high definition 360 videos of nearby pedestrians was mounted above the ZED. Data from the on-board cameras are useful in capturing pedestrian pose data and facial expressions. The ZED camera was powered by a laptop with a power bank. Our entire data collection hardware system is portable and does not require power outlets, thereby allowing data collection outdoors or in areas where wall power is not convenient.
Cart data was collected multiple times during each data collection session. We pushed the cart from one end of the scene to another end, while avoiding pedestrians and obstacles along the way in a natural motion similar to a human pushing a delivery cart. The purpose of this cart was to represent a mobile robot traversing through the human environment. However, unlike other datasets such as \cite{lcas} or \cite{jrdb} that use a Wizard-of-Oz controlled robot, we used a manually pushed cart. This provided better trajectory control, increased safety, and reduced the novelty effect from pedestrians, as curious pedestrians may intentionally block robots or display other unnatural movements \cite{brvsvcic2015escaping}.
The first batch of our data collection occurred on the ground level in a large indoor atrium area (Figure~\ref{fig:hdc_setup}). Half of the atrium area had fixed entry/exit points that led to corridors, elevators, stairs, and doors to the outside. The other half of the atrium was adjacent to another large open area and was unstructured with no fixed entry/exit points. We collected data around lunch and dinner times to ensure higher crowd densities (there was a food court in the neighboring open area). More data will be collected in the future in locations such as transit stations.
\subsection{Post-processing and Labeling}
A summary of our post processing pipeline is summarized in Figure \ref{fig:system-flowchart}.
\label{sec:postprocessing}
\begin{figure}[thpb]
\centering
\includegraphics[scale=0.27]{imgs/flowchart.jpg}
\caption{Flowchart for our post-processing pipeline. Blue blocks are preparation procedures and orange blocks are labeling procedures. The green block transforms all trajectory labels onto the ground plane $z=0$.}
\label{fig:system-flowchart}
\end{figure}
\subsubsection{Time synchronization and Calibration}\label{sec:calibration}
To ensure time synchronization across the captured videos, we employed Precision Time Protocol over a wireless network to synchronize each of the clocks of the computers powering the cameras, which allows for sub-microsecond synchronization. For redundancy, we held an LED light at a location inside the field of view of all the cameras and switched it on and off at the beginning of each recording session. We then checked for the LED light signal during the post-processing stage to synchronize the starting frame of all the captured videos for each recording session. We observed very little time drift in the individual recording computer clocks throughout the duration of each recording session, meaning that one synchronization point at the beginning of the recording sufficed.
Due to the portable nature of our system and the long distances between the cameras and the scene, we used scene reconstruction techniques to retrieve the intrinsics and poses of the cameras. We used Colmap \cite{schonberger2018robust} to perform a 3D reconstruction of the scene and estimated the static camera poses and intrinsics by additionally supplying it with dozens of static pictures of the atrium taken from a smartphone. The effectiveness of obtaining the camera parameters this way may also be applied to future work. For example, it may be possible to use crowdsourced approaches to collect such data when trying to repeat our effort with other camera deployments (e.g., a building atrium with multiple security cameras) since hundreds of images and videos may be available in populous areas.
\subsubsection{Ground plane identification}\label{sec:ground}
After the 3D reconstruction, the ground plane was not always $z=0$, but $z=0$ usually is the assumption for pedestrian datasets. We first defined an area on the ground plane and selected all the points inside the area $\mathcal{P}$. We then used RANSAC \cite{RANSAC} for maximum accuracy to identify a 2D surface $G$ within $\mathcal{P}$.
\begin{equation}
G = \text{RANSAC}(\mathcal{P})
\end{equation}
Where $G$ is expressed as $g_ax + g_by + g_cz + g_d = 0$. Once the ground plane was identified, it was then trivial to apply simple geometry to identify the homography matrix that transforms the coordinates on $G$ to $G': z=0$.
\subsubsection{Cart localization}\label{sec:cart_loc}
After the cameras were synchronized and calibrated, the next step was to localize the cart in the scene. This was achieved by first identifying the cart on the static camera videos and then applying the camera matrices to obtain the metric coordinates. We attempted multiple tracking models such as a deep learning based tracking model \cite{chen2019siammaske} on the static camera videos, but the tracking outcomes were unsatisfactory. We later attached a poster-sized AprilTag \cite{olson2011apriltag} on top of the cart for automatic pose estimation of the cart. We also explored other localization methods (e.g., wireless triangulation) and will continue to track progress on large-space localization. For the first batch of data included in our dataset, we manually labeled the locations of the cart.
\subsubsection{Pedestrian tracking and labeling}\label{sec:ped}
Similar to cart localization, we first tracked the pedestrians on the static camera videos and then identified their coordinates on the ground plane $G$. We found ByteTrack \cite{zhang2021bytetrack} to be very successful in tracking pedestrians in the image space. Upon human verification over our entire first batch of data, ByteTrack successfully aided the trajectory labeling of $91.8\%$ of the pedestrians automatically.
\begin{figure}[thpb]
\centering
\includegraphics[scale=0.215]{imgs/noise.jpg}
\caption{Smoothing of noise in auto-generated pedestrian trajectories by applying 3D correction. (left)Left: Raw tracking results from ByteTrack \cite{zhang2021bytetrack} (pixel space). Some noise is present due to human body motion. (right) Accounting for noise in 3D results in more accurate labeling.}
\label{fig:noise}
\end{figure}
Once we obtained the automatically tracked labels in pixel space, we needed to convert them into metric space. However, the process to do so was different from cart localization in section \ref{sec:cart_loc}, where the cart is either manually or automatically tracked (attached AprilTag). For the automatic tracking of pedestrians, the pedestrian's body motions while walking created significant noise, as shown in Figure \ref{fig:noise}. Therefore, the tracking noise was in 3D and assumptions that the noise solely exists on $G$ may result in large labeling inaccuracies.
We addressed this issue by estimating 3D metric coordinates from two cameras instead of assuming the metric coordinates to be on the 2D plane $G$ and obtaining these coordinates from a single camera. For each camera, we had a $3\times4$ camera matrix $P$.
\begin{equation}
P=\begin{bmatrix}
\mbox{---}\boldsymbol{p}_1\mbox{---} \\
\mbox{---}\boldsymbol{p}_2\mbox{---} \\
\mbox{---}\boldsymbol{p}_3\mbox{---}
\end{bmatrix}
\end{equation}
Where we had $P_1, P_2, P_3$ for the three cameras respectively. For a given 2D point coordinate $\boldsymbol{x}$ we wanted to estimate its corresponding 3D coordinate $\boldsymbol{X}$, so we had $\boldsymbol{x} = \alpha P\boldsymbol{X}$. We could then apply the cross product technique to eliminate the scalar $\alpha$. This gave us $\boldsymbol{x} \times P\boldsymbol{X} = \boldsymbol{0}$, or more precisely
\begin{equation}
\begin{bmatrix}
y\boldsymbol{p}_3^\top - \boldsymbol{p}_2^\top \\
\boldsymbol{p}_1^\top - x\boldsymbol{p}_3^\top
\end{bmatrix}\boldsymbol{X}=\boldsymbol{0}
\end{equation}
With two cameras $P_i, P_j| i\neq j, \; (i, j) \in \{1, 2, 3\}$, their corresponding 2D image points $(x_i, y_i), (x_j, y_j)$, and the constraint that the 3D coordinates should be on the ground plane $G$, we could construct the following system of equations to estimate the 3D coordinates.
\begin{equation}
A\boldsymbol{X}=\begin{bmatrix}
y_i\boldsymbol{p}_{i,3}^\top - \boldsymbol{p}_{i,2}^\top \\
\boldsymbol{p}_{i,1}^\top - x_i\boldsymbol{p}_{i,3}^\top \\
y_j\boldsymbol{p}_{j,3}^\top - \boldsymbol{p}_{j,2}^\top \\
\boldsymbol{p}_{j,1}^\top - x_j\boldsymbol{p}_{j,3}^\top \\
g_a, g_b, g_c, g_d
\end{bmatrix}\boldsymbol{X}=\boldsymbol{0}
\end{equation}
We could then perform singular value decomposition (SVD) on $A$ to obtain the solution.
With ByteTrack, each camera video contained a set of tracked trajectories in the image space $T_i=\{t_1,...,t_n\}, i\in\{1,2,3\}$. We estimated the 3D trajectory coordinates for each pair of 2D trajectories $(t_i, t_j)| t_i\in T_i, t_j\in T_j, i\neq j$ and the set of estimated coordinates that resulted in the lowest reprojection error were selected to be the final trajectory coordinates in the metric space. We then projected these 3D coordinates onto the ground plane $G$ and transformed them to $G'$ to obtain the final metric coordinates.
Finally, we performed human verification over the entire tracking output, fixing any errors observed during the process. We also manually identified pedestrians that were outside our target tracking zone but had interactions with the pedestrians inside the tracking zone and included them as part of our dataset.
\section{INTRODUCTION}
This template provides authors with most of the formatting specifications needed for preparing electronic versions of their papers. All standard paper components have been specified for three reasons: (1) ease of use when formatting individual papers, (2) automatic compliance to electronic requirements that facilitate the concurrent or later production of electronic products, and (3) conformity of style throughout a conference proceedings. Margins, column widths, line spacing, and type styles are built-in; examples of the type styles are provided throughout this document and are identified in italic type, within parentheses, following the example. Some components, such as multi-leveled equations, graphics, and tables are not prescribed, although the various table text styles are provided. The formatter will need to create these components, incorporating the applicable criteria that follow.
\section{PROCEDURE FOR PAPER SUBMISSION}
\subsection{Selecting a Template (Heading 2)}
First, confirm that you have the correct template for your paper size. This template has been tailored for output on the US-letter paper size.
It may be used for A4 paper size if the paper size setting is suitably modified.
\subsection{Maintaining the Integrity of the Specifications}
The template is used to format your paper and style the text. All margins, column widths, line spaces, and text fonts are prescribed; please do not alter them. You may note peculiarities. For example, the head margin in this template measures proportionately more than is customary. This measurement and others are deliberate, using specifications that anticipate your paper as one part of the entire proceedings, and not as an independent document. Please do not revise any of the current designations
\section{MATH}
Before you begin to format your paper, first write and save the content as a separate text file. Keep your text and graphic files separate until after the text has been formatted and styled. Do not use hard tabs, and limit use of hard returns to only one return at the end of a paragraph. Do not add any kind of pagination anywhere in the paper. Do not number text heads-the template will do that for you.
Finally, complete content and organizational editing before formatting. Please take note of the following items when proofreading spelling and grammar:
\subsection{Abbreviations and Acronyms} Define abbreviations and acronyms the first time they are used in the text, even after they have been defined in the abstract. Abbreviations such as IEEE, SI, MKS, CGS, sc, dc, and rms do not have to be defined. Do not use abbreviations in the title or heads unless they are unavoidable.
\subsection{Units}
\begin{itemize}
\item Use either SI (MKS) or CGS as primary units. (SI units are encouraged.) English units may be used as secondary units (in parentheses). An exception would be the use of English units as identifiers in trade, such as Ò3.5-inch disk driveÓ.
\item Avoid combining SI and CGS units, such as current in amperes and magnetic field in oersteds. This often leads to confusion because equations do not balance dimensionally. If you must use mixed units, clearly state the units for each quantity that you use in an equation.
\item Do not mix complete spellings and abbreviations of units: ÒWb/m2Ó or Òwebers per square meterÓ, not Òwebers/m2Ó. Spell out units when they appear in text: Ò. . . a few henriesÓ, not Ò. . . a few HÓ.
\item Use a zero before decimal points: Ò0.25Ó, not Ò.25Ó. Use Òcm3Ó, not ÒccÓ. (bullet list)
\end{itemize}
\subsection{Equations}
The equations are an exception to the prescribed specifications of this template. You will need to determine whether or not your equation should be typed using either the Times New Roman or the Symbol font (please no other font). To create multileveled equations, it may be necessary to treat the equation as a graphic and insert it into the text after your paper is styled. Number equations consecutively. Equation numbers, within parentheses, are to position flush right, as in (1), using a right tab stop. To make your equations more compact, you may use the solidus ( / ), the exp function, or appropriate exponents. Italicize Roman symbols for quantities and variables, but not Greek symbols. Use a long dash rather than a hyphen for a minus sign. Punctuate equations with commas or periods when they are part of a sentence, as in
$$
\alpha + \beta = \chi \eqno{(1)}
$$
Note that the equation is centered using a center tab stop. Be sure that the symbols in your equation have been defined before or immediately following the equation. Use Ò(1)Ó, not ÒEq. (1)Ó or Òequation (1)Ó, except at the beginning of a sentence: ÒEquation (1) is . . .Ó
\subsection{Some Common Mistakes}
\begin{itemize}
\item The word ÒdataÓ is plural, not singular.
\item The subscript for the permeability of vacuum ?0, and other common scientific constants, is zero with subscript formatting, not a lowercase letter ÒoÓ.
\item In American English, commas, semi-/colons, periods, question and exclamation marks are located within quotation marks only when a complete thought or name is cited, such as a title or full quotation. When quotation marks are used, instead of a bold or italic typeface, to highlight a word or phrase, punctuation should appear outside of the quotation marks. A parenthetical phrase or statement at the end of a sentence is punctuated outside of the closing parenthesis (like this). (A parenthetical sentence is punctuated within the parentheses.)
\item A graph within a graph is an ÒinsetÓ, not an ÒinsertÓ. The word alternatively is preferred to the word ÒalternatelyÓ (unless you really mean something that alternates).
\item Do not use the word ÒessentiallyÓ to mean ÒapproximatelyÓ or ÒeffectivelyÓ.
\item In your paper title, if the words Òthat usesÓ can accurately replace the word ÒusingÓ, capitalize the ÒuÓ; if not, keep using lower-cased.
\item Be aware of the different meanings of the homophones ÒaffectÓ and ÒeffectÓ, ÒcomplementÓ and ÒcomplimentÓ, ÒdiscreetÓ and ÒdiscreteÓ, ÒprincipalÓ and ÒprincipleÓ.
\item Do not confuse ÒimplyÓ and ÒinferÓ.
\item The prefix ÒnonÓ is not a word; it should be joined to the word it modifies, usually without a hyphen.
\item There is no period after the ÒetÓ in the Latin abbreviation Òet al.Ó.
\item The abbreviation Òi.e.Ó means Òthat isÓ, and the abbreviation Òe.g.Ó means Òfor exampleÓ.
\end{itemize}
\section{USING THE TEMPLATE}
Use this sample document as your LaTeX source file to create your document. Save this file as {\bf root.tex}. You have to make sure to use the cls file that came with this distribution. If you use a different style file, you cannot expect to get required margins. Note also that when you are creating your out PDF file, the source file is only part of the equation. {\it Your \TeX\ $\rightarrow$ PDF filter determines the output file size. Even if you make all the specifications to output a letter file in the source - if your filter is set to produce A4, you will only get A4 output. }
It is impossible to account for all possible situation, one would encounter using \TeX. If you are using multiple \TeX\ files you must make sure that the ``MAIN`` source file is called root.tex - this is particularly important if your conference is using PaperPlaza's built in \TeX\ to PDF conversion tool.
\subsection{Headings, etc}
Text heads organize the topics on a relational, hierarchical basis. For example, the paper title is the primary text head because all subsequent material relates and elaborates on this one topic. If there are two or more sub-topics, the next level head (uppercase Roman numerals) should be used and, conversely, if there are not at least two sub-topics, then no subheads should be introduced. Styles named ÒHeading 1Ó, ÒHeading 2Ó, ÒHeading 3Ó, and ÒHeading 4Ó are prescribed.
\subsection{Figures and Tables}
Positioning Figures and Tables: Place figures and tables at the top and bottom of columns. Avoid placing them in the middle of columns. Large figures and tables may span across both columns. Figure captions should be below the figures; table heads should appear above the tables. Insert figures and tables after they are cited in the text. Use the abbreviation ÒFig. 1Ó, even at the beginning of a sentence.
\begin{table}[h]
\caption{An Example of a Table}
\label{table_example}
\begin{center}
\begin{tabular}{|c||c|}
\hline
One & Two\\
\hline
Three & Four\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[thpb]
\centering
\framebox{\parbox{3in}{We suggest that you use a text box to insert a graphic (which is ideally a 300 dpi TIFF or EPS file, with all fonts embedded) because, in an document, this method is somewhat more stable than directly inserting a picture.
}}
\caption{Inductance of oscillation winding on amorphous
magnetic core versus DC bias magnetic field}
\label{figurelabel}
\end{figure}
Figure Labels: Use 8 point Times New Roman for Figure labels. Use words rather than symbols or abbreviations when writing Figure axis labels to avoid confusing the reader. As an example, write the quantity ÒMagnetizationÓ, or ÒMagnetization, MÓ, not just ÒMÓ. If including units in the label, present them within parentheses. Do not label axes only with units. In the example, write ÒMagnetization (A/m)Ó or ÒMagnetization {A[m(1)]}Ó, not just ÒA/mÓ. Do not label axes with a ratio of quantities and units. For example, write ÒTemperature (K)Ó, not ÒTemperature/K.Ó
\section{CONCLUSIONS}
A conclusion section is not required. Although a conclusion may review the main points of the paper, do not replicate the abstract as the conclusion. A conclusion might elaborate on the importance of the work or suggest applications and extensions.
\addtolength{\textheight}{-12cm}
\section*{APPENDIX}
Appendixes should appear before the acknowledgment.
\section*{ACKNOWLEDGMENT}
The preferred spelling of the word ÒacknowledgmentÓ in America is without an ÒeÓ after the ÒgÓ. Avoid the stilted expression, ÒOne of us (R. B. G.) thanks . . .Ó Instead, try ÒR. B. G. thanksÓ. Put sponsor acknowledgments in the unnumbered footnote on the first page.
References are important to the reader; therefore, each citation must be complete and correct. If at all possible, references should be commonly available publications.
\section*{ACKNOWLEDGMENT}
This work was supported by grants (IIS-1734361 and IIS-1900821) from the National Science Foundation.
{
\bibliographystyle{IEEEtranS}
| f50aa74478f707120aaed4c27ad5e542fff9fd45 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Experimental biologists and clinicians seek a deeper understanding of biological processes and their link with disease phenotypes by characterizing cell behavior. Gene expression offers a fruitful avenue for insights into cellular traits and changes in cellular state. Advances in technology that enable the measurement of RNA levels for individual cells via Single-cell RNA sequencing (scRNA-seq) significantly increase the potential to advance our understanding of the biology of disease by capturing the heterogeneity of expression at the cellular level \citep{Haque:2017}. Gene differential expression analysis, which contrasts the marginal expression levels of genes between groups of cells, is the most commonly used mode of analysis to interrogate cellular heterogeneity.
By contrast, the relational patterns of gene expression have received far less attention. The most intuitive relational effect is gene co-expression, a synchronization between gene expressions, which can vary dramatically among cells. Converging evidence has revealed the importance of co-expression among genes. When looking at a collection of highly heterogeneous cells, such as cells from multiple cell types, significant gene co-expression may indicate rich cell-level structure. Alternatively, when looking at a batch of highly homogeneous cells, gene co-expression could imply gene cooperation through gene co-regulation \citep{raj2006stochastic,Emmert-Streib:2014}.
Biochemistry offers a complementary motivation for the advantages of studying co-expression in addition to marginal expression levels of genes. The biological system of a cell is generally described by a non-linear dynamical system in which gene expression is variable \citep{raj2006stochastic}. Therefore, the observed gene expression level varies by time and condition, even within the same cell, while the cooperation between genes is more stable over time and condition. For this reason, it can be argued that co-expression may more reliably characterize the biological system or state of the cell \citep{dai2019cell}. scRNA-seq,
allows us to investigate gene co-expression at different resolutions, to understand not only how genes interact with each other within different cells, but also how the interactions relate to cell heterogeneity.
The recent work by \cite{dai2019cell} attempts an ambitious task: characterizing the gene co-expression at a single cell level (termed ``cell-specific network'' CSN). Specifically, for a pair of genes and a target cell, \citet{dai2019cell} construct a 2-way $2\times2$ contingency table test by binning all the cells based on whether they are in the marginal neighborhoods of the target cell and assigning the test results as a binary indicator of gene association in the target cell. Viewed over all gene pairs, the result is a cell-specific gene network.
Forgoing interpretation of the detected associations, they utilize the CSN to obtain a data transformation. Specifically, they replace the transcript counts in the gene-by-cell matrix with the degree sequence of each cell-specific network.
Although this data transformation shows encouraging success in various downstream tasks, such as cell clustering, it remains unclear what the detected ``cell-specific'' gene association network really represents.
The implementation details and interpretation of the results are presented at a heuristic level, making
it difficult for others to appreciate and generalize this line of work.
In a follow-up paper, \cite{wang2021constructing} take the first steps to capitalize on the CSN approach by redirecting the concept to obtain an estimator of co-expression. Specifically,
they propose averaging the ``cell specific" gene association indicators over cells in a class to recover a global measure of gene association (avgCSN). The resulting measure performs remarkably well in certain simulations and detailed empirical investigations of brain cell data. Compared to Pearson's correlation, the avgCSN gene co-expression appears less noisy and provides more accurate edge estimation in simulations. It is also more powerful in a test to uncover differential gene networks between diseased and control brain cells. Finally, it provides biologically meaningful gene networks in developing cells.
The empirical success of avgCSN likely lies in the nature of gene expression data: often noisy, sparse and heterogeneous, meaning not all cells exhibit co-expression at all times due to cellular state and conditions. For this reason, a successful method must be robust and sensitive to local patterns of dependencies.
{\color{black} Being an average of a series of binary local contingency table tests, the error in each entry of avgCSN is limited, meanwhile the non-negative summands ensure that local patterns are not cancelled out. By contrast, measures like Pearson's correlation can have both negative and positive summands, and therefore the final value can be small even if the dependence structure is clear for a subset of the cells.} To make the method more stable, \cite{wang2021constructing} proposed some heuristic and practical techniques to compute avgCSN, for which we would like to have more principled insights. Examples are the choice of window size in defining neighborhoods in the local contingency table test, the choice of thresholding in constructing an edge, and the range of cells to aggregate over. Many natural questions emerge: how does avgCSN relate to other gene co-expression measures and the full range of general univariate dependence measures, and why does it perform well in practice? Through theoretical analysis and extensive experimental evaluations, we address these questions, revealing that avgCSN is an empirical estimator of a new dependency measure, which enjoys various advantages over the existing measures.
For comparison, we briefly review the related work in gene co-expression measures and general univariate dependence.
Since the work by \citet{eisen1998cluster}, Pearson's correlation has been the most popular gene co-expression measure for its simple interpretation and fast computation. However, Pearson's correlation fails to detect non-linear relationships and is sensitive to outliers. Another class of co-expression methods is based on mutual information (MI) \citep{bell1962mutual,steuer2002mutual,daub2004estimating}. The computation of MI involves discretizing the data and tuning parameters, and the dependence measure does not have an interpretable scale. \citet{reshef2011detecting} proposed the maximal information coefficient (MIC) as an extension of MI, but MIC was shown to be over-sensitive in practice. More comparisons of different co-expression measures and the constructed co-expression networks can be found in \cite{song2012comparison,allen2012comparing}.
In the broader statistical literature, the problem of finding gene co-expression is closely related to that of detecting univariate dependence between two random variables. Specifically, for a pair of univariate random variables $X, Y$, how to measure the dependence between them has been a long-standing problem. The problem is often described as finding a function $\delta(X,Y)$, which measures the discrepancy between the joint distribution $F_{XY}$ and product of marginal distribution $F_{X}F_{Y}$. Numerous solutions to this problem have been provided: include the Renyi correlation \citep{renyi1959measures} measuring the correlation between two variables after suitable transformations; various regression-based techniques; Hoeffding’s D \citep{hoeffding1948non}, distance correlation (dCor) \citep{szekely2007measuring}, kernel-based measure like HSIC \citep{gretton2005measuring} and rank based measure like Kendall's $\tau$ and the refinement later, $\tau^\star$ \citep{bergsma2014consistent}. Most of these methods have not yet been widely adopted in genetics applications.
Aside from avgCSN, the methods mentioned so far do not specifically target dependence relationships that are local and often assume the data are random samples from a common distribution (in contrast with a mixture distribution) in the theoretical analysis. However, real gene interactions may change as the intrinsic cellular state varies and may only exist under specific cellular conditions. Furthermore, with data integration now being a routine approach to combat the curse of dimensionality, samples from different experimental conditions or tissue types are likely to possess different gene relationships and thus create more complex situations for detecting gene interactions. In this setting, much like avgCSN, an ideal measure accumulates subtle local dependencies, possibly only observed in a subset of the cells. A co-expression measure that aims to detect local patterns, developed by \cite{wang2014gene}, counts the proportion of matching patterns of local expression ranks as the measure of gene co-expression. Specifically, they aggregate the gene interactions across all subsamples of size $k$. However, despite its promising motivation, it has low power to detect non-monotone relationships. MIC \citep{reshef2011detecting} and HHG \citet{heller2013consistent} are also measures that attempt to account for local patterns of dependencies.
In this paper, we first give a detailed review of the related methods in \secref{aLDGintr}. Then in \secref{pop}, we show that avgCSN is indeed an empirical estimate of a valid dependence measure, which we define as averaged Local Density Gap (aLDG). In \secref{robpop} and \secref{emp}, we formally establish its statistical properties, including estimation consistency and robustness. We also investigate data-adaptive hyperparameter selection to justify and refine the heuristic choices in application in \secref{chooset}. Finally, we provide a systematic comparison of aLDG and its competitors via both simulation and real data examples in \secref{aLDGcompare}.
\section{A brief review of dependence and association measures}\label{sec:aLDGintr}
Before starting on the description of the various dependence measures, let us remark that \citet{renyi1959measures} proposed that a measure of dependence between two stochastic variables $X$ and $Y$, $\delta(X,Y)$, should ideally have the following properties:
\begin{enumerate}
\item[(i)] $\delta(X,Y)$ is defined for any $X,Y$ neither of which is constant with probability $1$.
\item[(ii)] $\delta(X,Y)$=$\delta(Y,X)$.
\item[(iii)] $0\leq \delta(X,Y) \leq 1$.
\item[(iv)] $\delta(X,Y)=0$ if and only if $X$ and $Y$ are independent.
\item[(v)] $\delta(X,Y)=1$ if either $X=g(Y)$ or $Y=f(X)$, where $f$ anf $g$ are measurable functions.
\item[(vi)] If the Borel-measurable functions $f$ and $g$ map the real axis in a one-to-one way to itself, then $\delta(f(X),g(Y))=\delta(X,Y)$.
\end{enumerate}
Particularly, a measure satisfying (iv) is called a strong dependence measure.
Apart from the above properties, there are two more properties that are particularly useful in single-cell data analysis. Single-cell data often contain a significant amount of noise, among which outliers account for a non-negligible fraction. Therefore \emph{robustness} is a desirable property in a dependence measure. Specifically, keeping with previous literature \citep{dhar2016study}, by robustness we mean that the value of the measure does not change much when a small contamination point mass, far away from the main population, is added. A formal description and corresponding evaluation metric will be described later. Another often overlooked property is \emph{locality}, which is a relatively novel concept and has not been properly defined to the best of our knowledge. Nevertheless, this concept has been catching attention over the recent decade \citep{reshef2011detecting,heller2013consistent,heller2016consistent,wang2014gene}, especially in work motivated by genetic data analysis. \emph{Locality} targets a special kind of dependence relationship that is generally restricted to a particular neighborhood in the sample space. A natural example is dependence that occurs in some, but not necessarily all of the components in a finite mixture. Another is dependence within a moving time window in a time series. Generally speaking, the interactions change as the hidden condition varies, or only exist under a specific hidden condition. A dependence measure that is \emph{local} should be able to accumulate dependence in the local regions.
No measure has all of the properties mentioned above, as far as we know. Our new measure possesses all but properties (v) and (vi). In the following, we review a selected list of univariate dependence measures in more details.
\subsection{Moment based measures}
The first class of methods is based on various moment calculations. The main advantage is fast computation and minimum tuning, while the main drawback is non-robustness to outliers from their moment-based nature.
\paragraph*{Pearson's correlation} The simplest measure is the classical Pearson’s correlation:
\begin{equation}
\text{Pearson's}\ \rho(X,Y):= \frac{\text{Cov}(X,Y)}{\sqrt{\text{Var}(X)\text{Var}(Y)}}.
\end{equation}
Plugin the sample estimation of covariance and variance, consistency and asymptotic normality can be proven using law of large numbers and the central limit theorem, respectively. Pearson's $\rho$ has been, and probably still is, the most extensively employed measure in statistics, machine learning, and real-world applications, due to its simplicity. However, it is known to detect only linear relationships. Also, as is the case for regression, it is well known that the product-moment estimator is sensitive to outliers: even just a single outlier may have substantial impact on the measure.
\paragraph*{Maximal correlation}
The maximal correlation (MC) is based on Pearson's $\rho$. It is constructed to avoid the problem that Pearson's $\rho$ can easily be zero even if there is strong dependence. \citet{gebelein1941statistische} first propose MC as
\begin{equation}
\text{MC}(X,Y) := \sup_{f,g} \rho(f(X),g(X)).
\end{equation}
Here the supremum is taken over all Borel-measurable functions $f,g$ with finite and positive variance for $f(X)$ and $g(Y)$. The measure MC can detect non-linear relationships, and in fact, it is a strong dependence measure. However, often MC cannot be evaluated explicitly except in special cases, because there does not always exist functions $f_0$ and $g_0$ such that $\text{MC} = \rho(f_0(X), g_0(Y))$. Also, it has been found to be overly ``sensitive'', i.e. it gives high value for distributions arbitrarily ``close'' to independence in practice.
\paragraph*{Distance correlation} A recent surge of interests has been placed on using distance metrics to achieve consistent independence testing against all dependencies. A notable example is the distance correlation (dCor) proposed by \citet{szekely2007measuring}:
\begin{align}
\text{dCor} (X,Y)& := \frac{V(X,Y)}{\sqrt{V(X,X)V(Y,Y)}}, \\
& \quad \text{where } V(X,Y)=\mathbb{E}{|X-X'||Y -Y'|}
+ \mathbb{E}{|X - X'|}\mathbb{E}{|Y - Y'|}\\
& \quad \quad \quad \quad \quad \quad \quad \quad \quad
- 2\mathbb{E}_{X,Y}\Big[{\mathbb{E}_{X'}|X-X'| \mathbb{E}_{Y'}|Y - Y'|}\Big], \nonumber
\end{align}
with $(X',Y')$ an i.i.d copy of $(X,Y)$. The distance correlation enjoys universal consistency against any joint distribution of finite second moments; however, in practice, it does not work well for non-monotone relationship \citep{shen2020distance}. Also, it is not robust from its moment based nature, as proven by \citet{dhar2016study}.
\paragraph*{HSIC} Recall the definition and formula for the maximal correlation, about which we mentioned it is difficult to compute since it requires the supremum of the correlation $\rho(f(X),g(Y))$ taken over Borel-measurable $f$ and $g$. In the framework of reproducing kernel Hilbert spaces (RKHS), it is possible to pose this problem and compute an analogue of MC quite easily. A state-of-the-art method in this direction is the so-called Hilbert-Schmidt Independence Criterion (HSIC) \citep{gretton2005measuring}. Denote the support of $X$ and $Y$ as $\mathcal{X}$ and $\mathcal{Y}$ respectively, HSIC considers $f,g$ to be in RKHS $\mathcal{F}$ and $\mathcal{G}$ of functionals on sets $\mathcal{X}$ and $\mathcal{Y}$ respectively. Then HSIC is defined to be the Hilbert-Schmidt (HS) norm of a Hilbert-Schmidt operator. We refer the reader to \cite{gretton2005measuring} for detailed description. What might be of interest is that, in many cases, HSIC is equivalent to dCor.
\subsection{Rank based measure}
Another line of work based on ordinal statistics is developed in parallel to the moment-based methods. A random variable $X$ is called ordinal if its possible values have an ordering, but no distance is assigned to pairs of outcomes. Ordinal data methods are often applied to data in order to achieve robustness.
\paragraph*{Spearman's $\rho_S$, Kendall's $\tau$ and $\tau^\star$} The two most popular measures of dependence for ordinal random variables $X$ and $Y$ are Kendall’s $\tau$ and Spearman’s $\rho_S$. Both Kendall's $\tau$ and Spearman's $\rho_S$ are proportional to sign versions of the ordinary covariance, which can be seen from the following expressions for the covariance:
\begin{align*}
\text{Cov}(X,Y) &= \frac12\EE{(X - X')(Y - Y')} \propto \text{Kendall} \\
& =\EE{(X' -X'')(Y' -Y''')} \propto \text{Spearman},
\end{align*}
where $(X',Y'), (X'',Y''), (X''', Y''')$ are i.i.d replications of $(X,Y)$. Note that Kendall's $\tau$ is simpler than Spearman's $\rho_S$ in the sense that it can be defined using only two rather than three independent replications of $(X, Y)$, so often Kendall's $\tau$ is preferred. A concern from certain applications is that Kendall's $\tau$ and Spearman's $\rho_S$ are not \emph{strong} dependence measures, so tests based on them are inconsistent for the alternative of a general dependence. In fact, it is often observed that they have difficulty detecting nonmonotone relationship. Later, an extension $\tau^\star$ \citep{bergsma2014consistent} mitigates such deficiency by modifying Kendall's $\tau$ to a strong measure.
\paragraph*{Hoeffding's D and BKR}
Related to the ordinal statistics-based methods, another class of methods start from the cumulative distribution function (CDF), some of which are equivalent to ordinal forms due to the relationship between CDF and ranks. The oldest example is the Hoeffing's D proposed by \citet{hoeffding1948non}:
\[
\text{Hoeffing's D} := \mathbb{E}_{X,Y} \Big[(F_{X,Y} - F_{X}F_{Y})^2\Big],
\]
where $F_X$, $F_Y$, $F_{X,Y}$ are the CDF of $X$, $Y$, $(X,Y)$ respectively. Still, Hoeffing's D is not a strong measure, while its modified version BKR \citep{blum1961distribution}:
\[
\text{BKR} := \mathbb{E}_{X}\mathbb{E}_{Y} \Big[(F_{X,Y} - F_{X}F_{Y})^2\Big]
\]
is. It turns out Hoeffding's D belongs to a more general family of coefficients, which can be formulated as
\[
\text{C}_{gh} := \int g(F_{X,Y} - F_{X}F_{Y}) d h(F_{XY})
\]
for some $g$ and $h$. We will abbreviate Hoeffding's D as HoeffD in the figures in the remainder of paper.
\subsection{Dependence measures aware of local patterns}
Most of the methods mentioned so far do not specifically target dependence relationships that can be local in nature. In the following, we describe a few measures that were designed to capture complex relationships, whether local or not.
\paragraph*{Maximal Information Coefficient} The idea behind the Maximal Information Coefficient (MIC,\cite{reshef2011detecting} statistic consists in computing the mutual information locally over a grid in the data set and then take as statistic the maximum value of these local information measures over a suitable choice of grid. However, several examples were given in \citet{simon2014comment} and \citet{gorfine2012comment} where MIC is clearly inferior to dCor.
\paragraph*{HHG}
\citet{heller2013consistent} pointed out another way to account for local patterns: that is, looking at dependence locally and then aggregating the dependence over the local regions. The local regions is simply defined as bins via partitioning the sample space.
Additionally, HHG takes a multi-scale approach: multiple sample space partitions are conducted, and results are aggregated over all of them. This results in a provably consistent permutation test. However, the cost of implementation is significantly longer computation time than its competitors: it takes $O(n^3)$ computation time while its competitors normally take at most $O(n^2)$.
\paragraph*{Matching ranks} Another method that developed specifically for accounting local pattern
is proposed by \citep{wang2014gene}. Given $n$ pair of observations of $(X,Y)$, $\{(x_i,y_i)\}_{i=1}^n$, they propose to count the number of size $k$ subsequences $(x_{i_1}, x_{i_2}, \dots x_{i_k})$ and $(y_{i_1}, y_{i_2}, \dots y_{i_k})$ such that their rank is matched. We refer to this measure as MR (Matching Ranks). Specifically, we write the scaled version of MR such that it is in range [0,1]:
\begin{align*}
\text{MR} := \frac{1}{2{n\choose k}}\sum_{1\leq i_1<i_2\dots<i_k \leq n} & \Big(\mathbf{1}\{ rank(x_{i_1}, x_{i_2}, \dots x_{i_k}) = rank(y_{i_1}, y_{i_2}, \dots y_{i_k})\} \\
& + \mathbf{1}\{ rank(x_{i_1}, x_{i_2}, \dots x_{i_k}) = rank(-y_{i_1}, -y_{i_2}, \dots -y_{i_k})\}\Big),
\end{align*}
where $rank(a_1,\dots,a_k) = (r(a_1),\dots,r(a_k))$ where $r(a_i)$ is the rank of element $a_i$ within the sequences $(a_1,\dots,a_k)$, and the equality inside the indicator function applies element-wisely. Though claimed to be able to detect complex relationship, this measure is inferior to others in some non-monotone dependence case like quadratic relationship.
\section{Our method: averaged Local Density Gap}\label{sec:aLDGdef}
First, we elaborate on the origin of our work, which was inspired by gene co-expression analysis using single-cell data. In the context of gene co-expression analysis, the pair of random variables $X,Y$ represents the expression level of a pair of genes, and the goal is to find the relationship between them. Pearson's correlation is one commonly used metric for this task. In light of the many shortcomings of this global measure of dependence, \citet{dai2019cell} proposed to characterize the gene relationships for every cell. Their method takes the following approach: for the gene pair $(X,Y)$, and a target cell $j$, partition the $n$ samples based on whether $|X_{\cdot} - X_j| < h_x$ and $|Y_{\cdot} - Y_j| < h_y$, where $h_x$ and $h_y$ are predefined window sizes. This partition can be summarized as a $2 \times 2$ contingency table (\tabref{contingency}). Then evidence against independence in this $2\times 2$ table can be quantified by a general contingency table test statistic. \citet{dai2019cell} uses
\begin{equation}
S_{X,Y}^{(j)} := \frac{ \sqrt{n} \left(n_{x,y}^{(j)}n - n_{x,\cdot}^{(j)} n_{\cdot,y}^{(j)}\right)}{\sqrt{n_{x,\cdot}^{(j)} n_{y}^{(j)} (n-n_{x,\cdot}^{(j)}) (n-n_{\cdot,y}^{(j)})}},
\end{equation}
and conducts a one-sided $\alpha$ level test based on its asymptotic normality, that is
\begin{equation}\label{contingency_test}
I_{XY}^{(j)}:= \mathbb{I}\{S_{X,Y}^{(j)} > \Phi^{-1}
(1-\alpha)\}.
\end{equation}
\begin{table}[H]
\centering
\begin{tabular}{c|c|c|c}
& $|Y_{\cdot} - Y_j| \leq h_y$ & $|Y_{\cdot} - Y_j| > h_y$ & \\ \hline
$|X_{\cdot} - X_j| \leq h_x$ & $n_{x,y}^{(j)} $ & & $n_{x,\cdot}^{(j)}$ \\ \hline
$|X_{\cdot} - X_j| > h_x$ & & & \\ \hline
& $n_{\cdot,y}^{(j)}$ & & $n$
\end{tabular}
\caption{The $2\times 2$ contingency table based on distance from $j$-th sample.}
\label{tab:contingency}
\end{table}
{\color{black} \citet{dai2019cell} claim that $I_{XY}{(j)}$ indicates whether or not gene pairs $X$ and $Y$ are dependent in cell $j$, and refer to the detected dependence as \emph{local dependence}. Though interesting as a novel concept, it lacks rigor and interpretability. Alternatively we propose to define $X$ and $Y$ as being \emph{locally independent} at position $(x,y)$ as
\begin{equation}
f_{XY}(x,y) = f_{X}(x)f_{Y}(y),
\end{equation}
then $I_{XY}$ provides a way of assessing \emph{local independence}.
Specifically, as a one-sided test, $I_{XY}(j)$ assesses whether or not $f_{XY}(x,y) > f_{X}(x)f_{Y}(y)$, at position $(x,y)$ marked by cell $j$. To assess global independence, aggregation, as proposed by \citet{wang2021constructing}, is needed. Their empirical measure can be formally written as:
\begin{equation}\label{avgcsn}
\text{avgCSN} := \frac{1}{n} \sum_{i=1}^n I_{XY} ^{(j)}.
\end{equation}
}
{\color{black}Some simple approximations gives us a population correspondence of avgCSN.} Assume the variables $X,Y$ have joint density $f_{XY}$, and marginal densities, $f_{X}$ and $f_{Y}$, that have common support. Let $\widehat{f}_{XY}, \widehat{f}_{X}, \widehat{f}_{Y}$ be the estimated densities given observations of $(X,Y)$.
Under the assumption that the bandwidth $h_x, h_y\to 0$ and $ \sqrt{h_x h_y n}\to\infty$, with some simple algebra (see \appref{derive} for detailed derivation), we see that
\begin{align}\label{deriveavgcsn}
\text{avgCSN} &\approx \frac{1}{n} \sum_{i=1}^n \mathbf{1}\left\{ \frac{\widehat{f}_{X,Y}(x_{i}, y_{i}) - \widehat{f}_{X}(x_{i}) \widehat{f}_{Y}(y_{i}) }{\sqrt{ \widehat{f}_{X}(x_{i})\widehat{f}_{Y}(y_{i})}} \geq t_n \right\},\quad \text{where } t_n = \frac{\Phi^{-1}(1-\alpha)}{\sqrt{n h_x h_y}},
\end{align}
and $\alpha\in[0,1]$ is some hyperparameter related to the test level of the local contingency test (usually $\alpha$ is set to 0.05 or 0.01). Because $t_n \downarrow 0$ as $n$ goes to infinity, we naturally think of the following population dependence measure:
\begin{equation*}
\text{Pr}_{X,Y}\left\{ \frac{f_{X,Y}(X,Y) - f_X(X) f_{Y}(Y)}{\sqrt{f_{X}(X)f_{Y}(Y)}} > 0\right\}.
\end{equation*}
In the remainder of this section, we formally define a generalized version of this measure in \secref{pop}, along with its properties on the population level. Then we discuss consistent and robust estimation in \secref{emp} and provide guidance on hyper-parameter selection in \secref{chooset}. Finally, we comment on the relationship between our measure and some of the previous work in \secref{relation}.
\subsection{Definition and basic properties}\label{sec:pop}
\begin{definition}(averaged Local Density Gap)
Consider a pair of random variables $X,Y$ whose joint and marginal densities both exist, and denote $f_{XY}, f_{X}, f_{Y}$ as their joint and marginal densities. The averaged Local Density Gap (aLDG) measure is then defined as
\begin{equation}
\text{aLDG}_t := \text{Pr}_{X,Y}\left\{ T(X,Y) > t\right\}, \quad \text{where } T(X,Y):= \frac{f_{X,Y}(X,Y) - f_X(X) f_{Y}(Y)}{\sqrt{f_{X}(X)f_{Y}(Y)}}
\end{equation}
and $t\geq 0$ is a tunable hyper-parameter.
\end{definition}
\noindent
From the definition, one can immediately realize the following lemma.
\begin{lemma}\label{lem:simpleprop}
For a pair of random variables $X,Y$ whose joint and marginal densities both exist, we have
\begin{enumerate}
\item $X \perp Y \Longleftrightarrow \text{aLDG}_0 =0 $;
\item if $t>0$, then $X \perp Y \Longrightarrow \text{aLDG}_t =0$;
\item $\text{aLDG}_t \text{ is non-increasing with regard } t$ for all $t \geq 0$;
\item $\text{aLDG}_t \in [0,1]$;
\item $\text{aLDG}_t(X,Y) = \text{aLDG}_t(Y,X)$;
\end{enumerate}
\end{lemma}
As a concrete example of the $\text{aLDG}$ measure, the left plot of \figref{aLDGpop} displays $\text{aLDG}$, given different $t$ for a bivariate Gaussian with different choices of correlation. We can see that (1) $\text{aLDG}_t$ is non-increasing with regard $t$ as our \lemref{simpleprop} suggests; (2) $\text{aLDG}_t$ equals zero at independence for all $t\geq 0$, while $\text{aLDG}_0$ equals zero if and only if there is no dependence, as our \lemref{simpleprop} suggests; (3) $\text{aLDG}_t$ increases with the dependency level, indicating that it is a sensible dependence measure.
Note that, from \lemref{simpleprop}, $\text{aLDG}_0$ is a \emph{strong}\footnote{Recall that a measure of dependence between a pair of random variable $X,Y$ is \emph{strong} if it equals zero if and only if $X$ and $Y$ are independent.} measure of dependence. While being strong is a desirable feature of a dependence measure, for $\text{aLDG}$ type of measure, we find that it comes with the sacrifice of robustness under independence (\propref{indeprob}). On the other hand, setting $t>0$ could result in insensitivity under weak dependence, but with a provable guarantee of robustness (\thmref{aLDGrobpop}). In summary, the hyper-parameter $t$ serves as a trade-off between robustness and sensitivity. In \secref{chooset} we will discuss the practical choice of $t$ in more detail. For now, we treat it as a predefined non-negative constant.
\subsection{Robustness analysis}\label{sec:robpop}
In the following, we present a formal robustness analysis. An important tool to measure the robustness of a statistical measure is the influence function (IF). It measures the influence of an infinitesimal amount of contamination at a given value on the statistical measure. The Gross Error Sensitivity (GES) summarizes IF in a single index by measuring the maximal influence an observation could have.
\begin{definition}[Influence function (IF) and Gross Error Sensitivity (GES)] Assume that the bivariate random variable $(X,Y)$ follows a distribution $F$, the influence function of a statistical functional $R$ at $F$ is defined as
\begin{align}\label{if}
\text{IF}\big((x,y), R, F\big) := \lim_{\epsilon \to 0} \frac{R\big((1-\epsilon) F + \epsilon \delta_{(x,y)}\big) - R(F)}{\epsilon}
\end{align}
where $\delta_{(x,y)}$ is a Dirac measure putting all its mass at $(x,y)$. The Gross Error Sensitivity (GES) summarizes IF in a single index by measuring the maximal influence over all possible contamination locations, which is defined as
\begin{equation}
\text{GES}(R, F) := \sup_{(x,y)} \mid \text{IF}\big((x,y), R, F\big)\mid.
\end{equation}
An estimator is called $B$-robust if its GES is bounded.
\end{definition}
Among the related work we have mentioned, only the robustness of $\tau$, $\tau^\star$, and $\text{dCor}$ have been theoretically investigated to the best of our knowledge. \citet{dhar2016study} proved that $\text{dCor}$ is not robust while $\tau$ and $\tau^\star$ are. Their evaluation criteria is a bit different from ours. We investigate the limit of the ratio when the contamination mass goes to zero. They investigate the ratio limit when the contamination position goes far away, given fixed contamination mass. We argue that our analysis aligns better with the main statistical literature. In the following, we show that $\text{aLDG}_t$ with $t>0$ is $B$-robust, under some reasonable regularity conditions.
\begin{theorem}\label{thm:aLDGrobpop}
Consider $t>0$, and a bivariate distribution $F$ of variable $(X,Y)$ whose joint and marginal densities exist as $f_{XY}$, $f_{X}$, $f_{Y}$, and satisfy
\begin{equation}
f_{\text{max}}:=||\sqrt{f_{X}f_{Y}}||_{\infty} <\infty; \quad \quad |\text{aLDG}_{t - \epsilon} - \text{aLDG}_{t}| \leq L\epsilon,\ \forall \ \epsilon >0;
\end{equation}
then we have
\begin{equation}
\text{GES}(\text{aLDG}_t, F) \leq L f_{\text{max}} +1 < \infty.
\end{equation}
\end{theorem}
The proof of \thmref{aLDGrobpop} is in \appref{aLDGrobpop}.
The first assumption about the boundness of density is common in density based statistical analysis. The second assumption about the $\text{aLDG}_{t}$ smoothness may look less familiar, however after a transformation, it is no more than a CDF-smoothness assumption: recall that $T(X,Y) := \frac{f_{XY}(X)-f_{X}(X)f_{Y}(Y)}{\sqrt{f_{X}(X)f_{Y}(Y)}}$, then
\begin{align}
|\text{aLDG}_{t-\epsilon} - \text{aLDG}_{t}|< L\epsilon \Longleftrightarrow \mathbb{P}\{|T(X,Y)-t|\leq \epsilon\} \leq L\epsilon,
\end{align}
that is, the CDF of random variable $T(X,Y)$ is L-lipschitz around $t$ for $t>0$. In \figref{smooth} we show the empirical density of $T(X,Y)$ for bivariate Gaussian of different correlation, which is generally bounded by some constant $L$ at positive values.
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{plots/densityT.pdf}
\caption{The empirical density of statistics $T$. The underlying bivariate distribution is Gaussian, and the value of $T$ is calculated using the true Gaussian density. We can see that, as the correlation increases, the density of $T$ near zero (annotated by the red dashed line) is smaller.}
\label{fig:smooth}
\end{figure}
In the following, we show that $\text{aLDG}_0$ is not robust under independence.
\begin{proposition}\label{prop:indeprob}
For any distribution $F$ over a pair of independent random variables $(X,Y)$ whose joint and marginal density exists and are smooth almost everywhere, we have
\begin{equation}
\text{GES}(\text{aLDG}_0, F) = \infty
\end{equation}
if and only if X is independent of Y.
\end{proposition}
The proof of \propref{indeprob} is in \appref{indeprob}.
The right plot in \figref{aLDGpop} provides some empirical evidence of the non-robustness of $\textnormal{aLDG}_0$ under independence. Specifically, we plot the population value of the ratio inside limitation \eqref{if}, under bivariate Gaussian with small enough contamination proportion $\epsilon$, to approximately show that the IF value of $\textnormal{aLDG}_t$ at independence indeed goes to infinity as $t$ goes to zero.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{plots/popvalue.pdf}
\caption{\textbf{(Left)} The true $\textnormal{aLDG}_t$ value for bivariate Gaussian with different levels of correlation under different choices of $t$. \textbf{(Right)} The influence function value approximated by setting the contamination proportion very small ($\epsilon = 10^{-6}$).}
\label{fig:aLDGpop}
\end{figure}
\subsection{Consistent and robust estimation}\label{sec:emp}
In this section we investigate estimation of $\text{aLDG}_t$ given finite samples. One natural way to estimate $\text{aLDG}_t$ is using the following plug-in estimator: recall that $\widehat{f}_{XY}, \widehat{f}_{X}, \widehat{f}_{Y}$ are the estimated joint and marginal densities, then given $n$ observations $\{(x_1,y_1),\dots,(x_n,y_n)\}$ of $(X,Y)$, $\text{aLDG}_t$ can be estimated by
\begin{align}
\label{eq:aLDGemp}
\widehat{\text{aLDG}}_t & := \frac{1}{n}\sum_{i=1}^n \mathbf{1}\left\{ \widehat{T}(x_i,y_i) \geq t \right\},\quad \text{where } \widehat{T}(x_i,y_i):=\frac{\widehat{f}_{X,Y}(x_{i}, y_{i}) - \widehat{f}_{X}(x_{i}) \widehat{f}_{Y}(y_{i}) }{\sqrt{ \widehat{f}_{X}(x_{i})\widehat{f}_{Y}(y_{i})}}
\end{align}
In the following, we establish the non-asymptotic high probability bound of the estimation error using the above simple plug-in estimator $\widehat{\text{aLDG}}_t$. The error rate is determined by the density estimation error for variable $X, Y$, as well as the probability estimation error for $T(X,Y)$.
\begin{theorem}\label{thm:aLDGconsist}
Consider $t>0$, and a bivariate distribution $F$ of variable $(X,Y)$ whose joint and marginal densities exist as $f_{XY}$, $f_{X}$, $f_{Y}$, and satisfy
\begin{align*}
&\inf_{x,y}f_{XY}(x,y),\ \inf_{x}f_X(x) \inf_{y}f_Y(y) \geq c_{\min},\\
& \sup_{x,y}f_{XY}(x,y),\ \sup_{x}f_X(x) \sup_{y}f_Y(y) \leq c_{\max},
\end{align*}
and for some $\eta_n$ with $\lim_{n\to\infty}\eta_n \to 0$, with probability at least $1-\frac{1}{n}$
\begin{equation}
||\widehat{f}_{XY}-f_{XY}||_{\infty}, ||\widehat{f}_{X}-f_{X}||_{\infty}, ||\widehat{f}_{Y}-f_{Y}||_{\infty} \leq \eta_n;
\end{equation}
and for some constant $0<L<\infty$,
\begin{equation}
|\text{aLDG}_{t-\epsilon}-\text{aLDG}_t| \leq L\epsilon\quad \text{for all} \ \epsilon>0.
\end{equation}
Then we have, with probability at least $1-\frac{2}{n}$, we have
\begin{equation}
\left|\widehat{\text{aLDG}}_t - \text{aLDG}_t\right| \leq LC\eta_n + \sqrt{\frac{2\log{n}}{n}},
\end{equation}
where $C$ depends only on $c_{\min}, c_{\max}$.
\end{theorem}
\thmref{aLDGconsist} is flexible in the sense that one can plug-in any kind of density estimator and its error rate to obtain the error rate of the corresponding $\widehat{\text{aLDG}}$ estimator. The proof of \thmref{aLDGconsist} is in \appref{pfconsist}. Though \thmref{aLDGconsist} was for fixed $t$, we also provide similar result that holds true uniformly over all possible $t$ in \appref{uniform}.
As for a concrete example, we provide explicit results for a special class of bivariate density and a simple density estimator. Specifically, we consider the true marginal density $f_X$, $f_Y$ that are L-Lipschitz, and the joint density $f_{XY}$ that are simply the product of $f_X$, $f_Y$; we also consider the following density estimator\footnote{The density estimator used here is not chosen to be minimax optimal. We instead design it to align the best with the practical methods \citet{dai2019cell} and \citet{wang2021constructing}, such that we can better justify and refine their heuristic choices of hyperparameter by theory.}:
\begin{align}\label{densest}
&\widehat{f}_{X}(\cdot) = \frac{1}{n} \sum_{j=1}^n K_{h_n}(\cdot, x_j) , \quad \widehat{f}_{Y}(\cdot) = \frac{1}{n} \sum_{j=1}^n K_{h_n}(\cdot, y_j), \nonumber\\
& \quad \widehat{f}_{XY}(\cdot, \cdot) = \frac{1}{n} \sum_{j=1}^n K_{h_n}(\cdot, x_j)K_{h_n}(\cdot, y_j),
\end{align}
where $K_{h_n}(\cdot,u):=\mathbf{1}\{|\cdot-u|\leq h_n\}/(2 h_n)$ is one-dimensional boxcar kernel smoothing function with bandwidth $h_n$. From \propref{densest} in \appref{densest}, the uniform estimation error rate $\eta_n$ in this setting is $O(n^{-1/6}\sqrt{\log{n}})$, given the asymptotic near-optimal bandwidth $h = O(n^{-1/6})$. Therefore, applying \thmref{aLDGconsist} gives us estimation error rate of $O(n^{-1/6}\sqrt{\log{n}})$ for $\textnormal{aLDG}_t$.
We also include robustness analysis of $\widehat{\text{aLDG}}_t$ in \appref{emprob}. Specifically, we consider an empirical contamination model that is commonly encountered in single-cell data analysis: a small proportion of the sample points are replaced by ``outliers'' far away from the rest samples. We show that $\widehat{\text{aLDG}}_t$ with and without outliers are close as long as the outlier proportion is small. This suggests that the estimator of $\text{aLDG}_t$ preserves its robust nature.
\subsection{Selection of hyper-parameter $t$}\label{sec:chooset}
In this section, we propose two methods for selecting $t$, each of which has merit. We also provide guidance on which one is preferable in different practice settings.
\paragraph*{Uniform error method} From the results in the previous section, we learn that $\text{aLDG}_0$ is not robust under independence. To prevent $\widehat{\text{aLDG}}_t$ from approaching $\text{aLDG}_0$ under independence, it is sufficient to make sure that the estimation error of $T$ under independence is uniformly dominated by $t$ with high-probability. To compute the uniform estimation error of $T$ under independence, we first manually construct the independence case via random shuffle. Given $n$ samples $\{(x_i,y_i)\}_{i=1}^n$ of $(X,Y)$, denote the corresponding empirical joint distribution as $\widehat{F}_{XY}$, and marginal joint distribution as $\widehat{F}_{X}$ and $\widehat{F}_{Y}$. Applying the random shuffle function $\pi$ on indices of one dimension (i.e. $Y$), we have
\begin{equation}
\{(x_i,y_{\pi(i)})\}_{i=1}^n \sim \widehat{F}_{X}\widehat{F}_{Y},
\end{equation}
that is the shuffled samples $\{(x_i,y_{\pi(i)})\}$ now come from a different joint distribution where $(X,Y)$ are independent.
We can then use the shuffled samples to compute the uniform estimation error of $T$ under independence. Note that $T$ under independence is exactly zero, therefore its uniform estimation error is just the uniform upper bound of its estimation. To stabilize the estimation of such upper bound, we use the median of estimated upper bound from $\max\{\lfloor 1000/n \rfloor, 5\}$ different random shuffles as the final estimation. We call this $t$ selection method the \emph{uniform error} method.
\paragraph*{Asymptotic norm method} When using $\textnormal{aLDG}_t$ in large-scale data analysis, choosing $t$ using the above data-dependent choice may be undesirable because it requires additional computations. In extensive simulations we observe that a simple alternative also performs fine in terms of maintaining consistency, power and robustness:
\begin{equation}\label{choosetnorm}
t = \Phi^{-1}\left(1-\frac{1}{n}\right)\Big/\left(\sqrt{\sigma_{X}\sigma_Y} n^{1/3}\right).
\end{equation}
This choice is motivated by the following heuristic. Recall our derivation of aLDG statistics from avgCSN around \eqref{deriveavgcsn}: as the sample size $n$ goes to infinity, and $h_x, h_y \to 0$, $h_xh_yn\to\infty$, the empirical estimation of $\text{aLDG}_t$ using the boxcar kernel cioncide with avgCSN. Therefore, $t_n$ in \eqref{deriveavgcsn} could serve as a natural choice for $t$, but one need to be extra careful about $\alpha$, which is the test level of local contingency test \eqref{contingency_test} in definition towards avgCSN. We specically modify $\alpha$ to decrease with $n$ instead of a fixed value like $0.05$ since we desire consistency: i.e. $\text{aLDG}_t$ under independence should goes to zero as $n$ goes to infinity. Finally, plugging in our choice of bandwidth $h_x = \sigma_X n^{-1/6}$, $h_y = \sigma_Y n^{-1/6}$ together with the new $\alpha_n$ in place of $\alpha$ into $t_n$ \eqref{deriveavgcsn}, we get \eqref{choosetnorm}. We call this $t$ selection method the \emph{asymptotic norm} method.
Empirically we find that the \emph{asymptotic norm} method is often too conservative given the small sample size (which is expected since it is based on the asymptotic normality of a contingency table test statistic). In practice, we recommend people use \emph{uniform error} over \emph{asymptotic norm} when the sample size is not too big (e.g., no bigger than 200). When the sample size is big enough (e.g., bigger than 200), and the computation budget is limited, we recommend the \emph{asymptotic norm} method. In the rest of the paper, we use the \emph{uniform error} method when the sample size is no bigger than 200 and the \emph{asymptotic norm} method when the sample size is bigger than 200. We admit that there could be other promising ways of selecting $t$, for example, a geometry way we provided in \appref{chooset}. Here we only present the methods that we found working the best after a careful evaluation (see \appref{chooset}).
\subsection{Relationships to HHG}\label{sec:relation}
The method that is most similar to aLDG is HHG (\cite{heller2013consistent}).
Like aLDG, HHG \citep{heller2013consistent} is based on aggregation of multiple contrasts between the local joint and marginal distributions
\begin{align*}
HHG := \sum_{i\neq j} M(i,j),\quad M(i,j) := (n-2) \frac{\Big(p_{XY}(B_{XY}^{i,j})-p_{X}(B_{X}^{i,j})p_{Y}(B_Y^{ij})\Big)^2}{p_{X}(B_{X}^{i,j})\Big(1-p_{X}(B_{X}^{i,j})\Big)p_{Y}(B_Y^{ij})\Big(1-p_{Y}(B_Y^{ij})\Big)},
\end{align*}
with $B_{X}^{i,j} = \{x: |x-x_i|\leq |x_i - x_j|\}$, $B_{Y}^{i,j} = \{y: |y-y_i|\leq |y_i - y_j|\}$ and $B_{XY}^{i,j} = B_{X}^{i,j} \otimes B_{Y}^{i,j} $, $p_{XY}, p_X, p_Y$ are joint probability function for $(X,Y)$ and marginal probability function for $X$ and $Y$ respectively. While the two measures appear quite similar, they differ in two critical aspects.
\paragraph*{The efficiency of single scale bandwidth} One notable difference between HHG and aLDG is that the former relies on a multi-scale choice of bandwidth for each sample point. Specifically, it utilizes multiple ($O(n)$) bandwidths for each data point. This results in a provably consistent permutation test; however, the cost of implementation is significantly longer computation time than its competitors. aLDG takes a single-scale approach, which considerably improves the computation efficiency. Moreover, the aLDG formulation provides a direct analogy to a density functional, which allows us to exploit existing work in density estimation to determine an appropriate bandwidth. This single-scale approach, though may not optimal, achieves comparable power to HHG, as shown in the upcoming simulation studies.
\paragraph*{The merit of thresholding} Another difference is that empirically aLDG aggregates over thresholded summands, see \eqref{eq:aLDGemp}. It turns out thresholding brings implicit robustness to noise. By contrast, consider the non-thresholded version of aLDG:
\begin{equation}
\text{aLDG}_{non}:= \EE{T(X,Y)}.
\end{equation}
Even with slight departures from independence, $\text{aLDG}_{non}$ can go to infinity. For example, consider the following joint and marginal distribution that admits a kernel product density mixture:
\begin{align*}
& f_{XY}(x,y) = \alpha k_{0,r}(x)k_{0,r}(y) + (1-\alpha)k_{0,1}(x)k_{0,1}(y),\\
& f_{X}(x) = \alpha k_{0,r}(x) + (1-\alpha)k_{0,1}(x), \quad f_{Y}(y) = \alpha k_{0,r}(y) + (1-\alpha)k_{0,1}(y)
\end{align*}
where $\alpha \in (0,1)$, $0<r\ll1$ and $k_{\mu,r}(\cdot):=\frac{1}{r}k(\frac{\cdot-\mu}{r})$, with $k$ as the density of 1-dim uniform distribution supported on $[-1,1]$.
Note that as $\alpha\to0$ and $r\to 0$, the model is essentially an independence case contaminated with a small point mass. Additionally with $\alpha/r \to \infty$, we can show that (see \appref{thred} for details)
\begin{equation}\label{nonthred}
\EE{T(X,Y)} \approx \frac{\alpha}{r} \to \infty,
\end{equation}
that is the non-thresholded version of $\text{aLDG}$ is very large under such simple case of small departure from independence, therefore is problematic. With thresholding, however, $\text{aLDG}$ is guaranteed to be approximately $\alpha$, which goes to zero for small perturbations, as one would desire.
\section{Empirical evaluation}\label{sec:aLDGcompare}
\subsection{Single-cell data application}\label{sec:real}
In this section, we evaluate aLDG among the other measures using scRNA-seq data from two studies.
\paragraph*{Chu dataset} This dataset \citep{chu2016single} contains 1018 cells
of human embryonic stem cell-derived lineage-specific progenitors. The seven cell types, including H1 embryonic stem cells (H1), H9 embryonic stem cells (H9), human foreskin fibroblasts (HFF), neuronal progenitor cells (NPC), definitive endoderm cells (DEC), endothelial cells (EC), and trophoblast-like cells (TB), were identified by fluorescence-activated cell sorting (FACS) with their respective markers. On
average, 9600 genes are measured per cell. In the following, we show some special gene pairs that exhibit strong, weak, or no relational patterns and the corresponding dependence values produced by different measures. We find that only aLDG gives a high value for strong relational patterns no matter how complex the pattern composition is; maintains near-zero values for known independent cases; and avoids a spurious relationship skewed by technical noise and sparsity (Figure \ref{fig:realbi}).
\begin{figure}[H]
\centering
\includegraphics[width=0.8
\linewidth]{plots/newrealpair.pdf}
\caption{Example of gene pair scatter plots from the Chu dataset, which has 1018 cells from 7 cell types. Gene expression is recorded as counts per million (CPM) and $\log_2$ transformed. In each plot, we show the scatter plot of $\log_2(\text{CPM}+1)$ for a pair of genes and provide the corresponding estimated dependence values using different methods to the right of the plots. \textbf{(a)} aLDG gives a much higher value than the others in these scenarios which appear to illustrate a strong mixture dependence pattern, even when the signal is predominantly in one cell type. \textbf{(b)} aLDG produces a high value for the obvious three mixture relationship in the first subplot. By contrast, in the second subplot, the cell identity are randomly shuffled for each gene pair, resulting in a constructed case of independence. Most measures, including aLDG, give near-zero values in this setting. The exception is MIC, which gives a misleadingly high value. \textbf{(c)} This example illustrate performance when there is a high level of sparsity: MIC and the moment-based methods like Pearson, dCor, and HSIC provide estimates that are greatly overestimated, while aLDG, TauStar, and Hoeffding's D are not influenced by this phenomenon. \textbf{(d)} This gene pair combines the challenge of sparsity with considerable noise: aLDG is still able to capture the less noisy, local cluster pattern in the upper left corner. }
\label{fig:realbi}
\end{figure}
\paragraph*{Autism Spectrum Disorder (ASD) Brain dataset}
\citet{velmeshev2019single} includes scRNA-seq data
from an ASD study that collected 105 thousand nuclei from cortical samples taken from 22 ASD and 19 control samples. Samples were matched for age, sex, RNA integrity number, and postmortem interval. In the following, we compare control and ASD groups by testing for differences in their gene co-expression matrices using the sparse-Leading-Eigenvalue-Driven (sLED) test \citep{zhu2017testing}. sLED takes the gene co-expression matrices for both control and ASD groups as input, and outputs a $p$-value indicating the significance of their difference. This method is particularly designed to detect differential signals attributable to a small fraction of the genes. To emphasize the contrast with differentially expressed genes, \cite{wang2021constructing} call these differential network genes.
Here we compare the power of the test for various co-expression measures. We use cells classified as L2/3 excitatory neurons (414 cells from ASD samples and 358 from control samples) and a set of 50 genes chosen randomly among the top 500 genes deferentially expressed between ASD and control samples. In addition, we manually add noise by randomly swapping 10\% of the control and ASD labels in the original data to see which measures detect the signal in the presence of greater noise. We omit HHG for this task as it requires too much computation time. Boxplots of $p$-values from sLED test across 10 independent trials (different random swapping each trial) are shown for all the remaining measures (\figref{vel23power}). Among the remaining measures, we find that HSIC, $\tau^\star$, Hoeffding's D, MIC, and aLDG perform well compared to Pearson, Spearman, Kendall, MRank and dCor. A visualization of the corresponding control versus ASD co-expression differences is displayed in \figref{vel23}, showing that the winners produce difference matrices with a few dominating entries, which is favored by the sLED test, while the others produce relatively flat and noisy patterns.
\begin{figure}[H]
\centering
\includegraphics[width=0.6\linewidth]{plots/newpval.pdf}
\caption{The estimated $p$-values obtained using sLED permutation tests for different dependency measures. We manually added noise by randomly swapping 10\% of the control and ASD labels in the original data to see which measures detect the signal in the presence of greater noise. Boxplots show the results from 10 independent repetitions. }
\label{fig:vel23power}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{plots/newrealcormat.png}
\caption{Estimated co-expression differences matrices (i.e. the absolute differences of the dependency matrices for control samples and ASD samples) obtained for different dependency measures.}
\label{fig:vel23}
\end{figure}
\subsection{Simulation results}\label{sec:simu}
In this section, we consider simulations that resembling single-cell data to gain insights underlying the behavior of aLDG relative to the other methods. Specifically, we investigate scenarios where the bivariate relationship is (1) finite mixture; (2) linear or nonlinear; (3) monotone or non-monotone. See \figref{data} for all the synthetic data distributions we considered. We evaluate each dependence measure from the following perspective: (1) ability to capture complex relationship; (2) ability to accumulate subtle local dependence; (3) interpretation of strength of dependence in common sense; (4) power as an independence test; and (5) computation time. In the following, we focus on one perspective in each subsection, showing selective examples that inform our conclusions, relegating other examples to supplementary materials.
\paragraph
{Detecting nonlinear, non-monotone relationships} By construction, aLDG is expected to detect any non-negligible deviation from independence. Though many existing measures, such as HSIC, Hoeffding's D, dCor, $\tau^\star$, claim to be sensitive to nonlinear, non-monotone relationships, some approaches are known to perform poorly under certain circumstances. By contrast, aLDG outperforms most of its competitors in the following standard evaluation experiment. \figref{nonlinear} illustrates three points: (1) at independence, except for dCor, HHG, and MIC, most measures produce negligible values, as desired; (2) for linear and monotone relationship, all measures produce high values as expected; and (3) for nonlinear non-monotone relationships only aLDG, dCor, HHG and MIC produce high values consistently. In conclusion, only aLDG can effectively detect various types of dependency relationships while maintaining near-zero value at independence. dCor, HHG, and MIC are known to be sensitive to small, artificial deviations from independence, and these simulations reveal that they are indeed too sensitive as they often produce high values at independence. A big portion of scRNA-seq data are collected over time; therefore, nonlinear, non-monotone and specifically oscillatory relationships are expected to happen. Therefore it is desirable to have a measure that is sensitive to dependence while remaining near zero of true independence, even under small perturbations.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{plots/value.pdf}
\caption{Empirical dependency estimates obtained for different data distributions for a variety of relationships between a pair of variables. For the visualization of different data distributions, see \figref{data}. Here we show the corresponding dependence level given by different measures using 200 samples (averaged over 50 trials).}
\label{fig:nonlinear}
\end{figure}
\paragraph
{Accumulating subtle local dependencies} aLDG detects the subset of the sample space that shows a pattern of dependence. In \figref{mix}, we simulated data as a bivariate Gaussian mixture consisting of three components with a varying proportion of highly dependent components and estimated the corresponding dependence level. We find that aLDG, together with other dependence measures designed to capture local dependence (HHG and MIC) increase with the proportion of highly correlated components, indicates that these global dependence measures can also detect subtle local dependence structure. Similar results are obtained for Negative Binomial mixtures \figref{nbmix}. As the finite mixture relationship is a common choice of model for scRNA-seq data, this suggests that measures able to accumulate dependencies across individual components could considerably benefit scRNA-seq data analysis.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{plots/valuemix_gauss.pdf}
\caption{Empirical aLDG value for Gaussian mixtures. In each plot we show the dependence level given by different measures for 200 samples (averaged over 50 trials). The data are generated as a three-component Gaussian mixture. From left to right, there are 0, 1, 2 and 3 out of 3 components with correlation of 0.8, while the remaining components have correlation 0, i.e., the dependence level increases from left to right. For the visualization of these different data distributions, see \figref{data}.
}
\label{fig:mix}
\end{figure}
\paragraph
{Degree of dependencies} While it is hard to define the relative dependence level in general, we argue that when one random variable is a function of the other, $Y=h(X)$, then the pair should be regarded as having the perfect dependence (and be assigned of dependence level $1$). Moreover, the dependence level should decrease as independent noise is added. That is, for $Y_\epsilon = h(X) + \epsilon$, where $\epsilon \perp X$, one should expect the dependence measure $\delta$ to satisfy $\delta(Y_{\epsilon},X) < \delta(Y,X)$. We checked this monotonicity property by simulating data with several bivariate relationships and varying levels of noise (\figref{mono}). Specifically, we simulate the noise $\epsilon$ to be standard normal, and $Y = h(X) + c\epsilon$ where $c\in[0,1]$ indicates the noise level. We find that aLDG, HSIC, MIC, dCor, and HHG all show a clear decreasing pattern as the noise level increases; however, aLDG shows the most consistent monotonic drop from perfect dependence as the noise level increased.
\begin{figure}[H]
\includegraphics[width=1\linewidth]{plots/mono.pdf}
\caption{Empirical dependence measure versus noise levels for different bivariate relationships. For the visualization of different data distributions, see \figref{data}. The results are shown for 100 samples (averaged over 50 trials). We claim that the higher the noise level is, the lower the estimated degree of dependence should be. Compared with other measures, aLDG decreases significantly as the noise level increases, and hence correctly infers the relative degree of dependence. }
\label{fig:mono}
\end{figure}
\paragraph
{Power as an independence test} Dependence measures are natural candidates for tests of independence. In this context, most existing dependence measures rely on bootstrapping or permutation to determine significance; hence we adopt this practice for all the dependence measures under comparison. \figref{non-linearpower} shows the empirical power under test level 0.05 for various types of data distribution and sample size, where we do 200 repetitions of permutations to estimate the null distribution. We observe the following outcomes: (1) almost all tests have controlled type-I error under independence; (2) Pearson's $\rho$, Spearman's $\rho_S$ and Kendall's $\tau$ are powerless for testing nonlinear and non-monotone relationships; (3) aLDG, HHG, and HSIC are consistently among the top three most powerful approaches for testing both linear and nonlinear, monotone and non-monotone relationships. Similar observations can be made for tests based on Gaussian mixtures \figref{gaussmixpower} and Negative Binomial mixtures \figref{nbmixpower}.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{plots/power.pdf}
\caption{The empirical power of permutation test at level 0.05, based on different dependency measures under different data distributions and sample sizes. For the visualization of different data distributions, see \figref{data}. The power is estimated using 50 independent trials.}
\label{fig:non-linearpower}
\end{figure}
\paragraph
{Computational comparisons}
Theoretically speaking, aLDG requires $O(n^2)$ in time of computation (where $n$ is the number of samples), which is comparable to reported requirements for most dependence measures that can detect complex relationships. This empirically confirmed in a comparison of the computation time of aLDG with all its competitors. In \figref{time} we plot the time of computation versus sample size $n$ for different dependence measures\footnote{The time include some constant wrapper function loading time, therefore, might be longer than a direct function call; however, the relative scale is still correct.}. In previous evaluations, we saw that HHG as a method motivated from capturing local dependence structure, was indeed a strong competitor to aLDG: it has high power as an independence test across almost all the data distribution we considered; however, it requires $O(n^3)$ time of computation, and \figref{time} shows this large discrepancy from all the other methods, which normally takes $O(n^2)$ time.
\section{Conclusion and Discussion}
In this paper, we formalize the idea of
averaging the \emph{cell-specific gene association} \citep{dai2019cell,wang2021constructing} under a general statistical framework. We show that this approach produces a novel univariate dependence measure, called aLDG, that can detect nonlinear, non-monotone relationships between a pair of variables. We then develop the corresponding theoretical properties of this estimator, including robustness and consistency. We also provide several hyper-parameter choices that are more justifiable and effective. Extensive simulations, motivated by expected scRNA-seq gene co-expression relationships and real data applications, show that this measure outperforms existing independence measures in various aspects: (1) it accumulates subtle local dependence over sub-populations; (2) it successfully interprets the relative strength of a monotonic function of dependence in the presence of noise better than many other measures that arose from independence test; (3) it is sensitive to complex relationships while robustly maintaining near-zero value at true independence, while several other measures are often overly sensitive to slight perturbations from independence and noise; (4) it computes comparatively rapidly compared to other dependence measures designed to capture complex relationships. Other measures perform well in some settings but fail in others that are highly relevant to the single-cell setting. For instance, MIC performed well as part of the sLED test for differences in co-expression matrices, but this measure tends to produce a high estimate of dependence even when the variables are independent, or nearly so (Figure \ref{fig:nonlinear} and Figure \ref{fig:mono}). The moment-based methods like Pearson, dCor, and HSIC perform poorly when the expression values are sparse, producing false indications of correlation (Figure \ref{fig:realbi}), and yet sparsity is the norm in most single cell data. Our method is implemented in the R package aLDG\footnote{\url{https://github.com/JINJINT/aLDG}}, where we also include all the other methods that we have compared with.
The aLDG method does have some practical challenges: as a measure based on density estimation, the hyperparameter choices such as bandwidth can affect the performance of the measure. Though we provide some asymptotically optimal choices of those hyperparameters, in practice, they can fail due to the small sample size. For any given setting, the hyperparameters can be adjusted based on realistic simulations of the actual data and a solid understanding of the scRNA-seq data distribution. Similarly, due to the reliance on density estimation, it is hard to extend this measure to a multivariate setting. The sample size required for accurate estimation grows exponentially with the dimension. In practice, this limitation has little practical importance because gene co-expression studies focus on bivariate relationships.
\paragraph{Acknowledgments}
The authors would like to thank Xuran Wang for helpful comments.
\paragraph{Funding}
This project is funded by National Institute of Mental Health (NIMH) grant R01MH123184 and NSF DMS-2015492.
\bibliographystyle{unsrtnat}
| cf1d967a690f5ec91d266e8a4fc97452e1f0107e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{#1}}
\renewcommand{\theequation}{\arabic{equation}}
\renewcommand{\thetable}{\Roman{table}}
\newcommand\scalemath[2]{\scalebox{#1}{\mbox{\ensuremath{\displaystyle #2}}}}
\newcommand{\textcolor{blue}}{\textcolor{blue}}
\newcommand{\textcolor{cyan}}{\textcolor{cyan}}
\newcommand{\textcolor{BrickRed}}{\textcolor{BrickRed}}
\definecolor{MyDarkBlue}{rgb}{0.1, 0.3, 0.8}
\definecolor{SBlue}{rgb}{0.2, 0.4, 0.4}
\definecolor{MyLightBlue}{rgb}{0.22,0.51,0.99}
\definecolor{MyGreen}{rgb}{0.0, 0.5, 0.3}
\definecolor{BrickRed}{rgb}{0.8, 0.25, 0.33}
\usepackage[colorlinks=true,linkcolor=blue,citecolor=MyDarkBlue,
urlcolor=BrickRed,bookmarksnumbered=true,bookmarksopen]{hyperref}
\hypersetup{colorlinks, citecolor=SBlue,linkcolor=MyGreen, urlcolor=BrickRed}
\newcommand{\MS}[1]{\textcolor{red}{MS: #1}}
\newcommand{\Y}[1]{\textcolor{blue}{Y: #1}}
\begin{document}
\vspace*{-0.2in}
\begin{flushright}
\end{flushright}
\begin{center}
{\Large \bf
Exploiting a future galactic supernova to probe\\\vspace{0.05 in} neutrino magnetic moments
}
\end{center}
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
\begin{center}
{
{}~\textbf{Sudip Jana,$^1$}\footnote{ E-mail: \textcolor{MyDarkBlue}{sudip.jana@mpi-hd.mpg.de}}
{}~\textbf{Yago P Porto-Silva,$^2$}\footnote{ E-mail: \textcolor{MyDarkBlue}{yporto@ifi.unicamp.br}}
{}~\textbf{Manibrata Sen$^1$}\footnote{ E-mail: \textcolor{MyDarkBlue}{manibrata@mpi-hd.mpg.de}}
}
\vspace{0.3cm}
{
\\\em$^1$Max-Planck-Institut f{\"u}r Kernphysik, Saupfercheckweg 1, 69117 Heidelberg, Germany
\\
$^2$Instituto de F{\'i}sica Gleb Wataghin - UNICAMP, 13083-859, \\ Campinas, S\~ao Paulo, Brazil
}
\end{center}
\renewcommand{\thefootnote}{\arabic{footnote}}
\setcounter{footnote}{0}
\thispagestyle{empty}
\begin{abstract}
A core-collapse supernova (SN) offers an excellent astrophysical laboratory to test non-zero neutrino magnetic moments. In particular, the neutronization burst phase, which lasts for few tens of milliseconds post-bounce, is dominated by electron neutrinos and can offer exceptional discovery potential for transition magnetic moments. We simulate the neutrino spectra from the burst phase in forthcoming neutrino experiments like the Deep Underground Neutrino Experiment (DUNE), and the Hyper-Kamiokande (HK), by taking into account spin-flavour conversions of supernova neutrinos caused by interactions with ambient magnetic fields. We find that the neutrino transition magnetic moments which can be explored by these experiments for a galactic SN are an order to several orders of magnitude better than the current terrestrial and astrophysical limits. Additionally, we also discuss how this realization might provide light on three important neutrino properties: (a) the Dirac/Majorana nature, (b) the neutrino mass ordering, and (c) the neutrino mass-generation mechanism.
\end{abstract}
\newpage
\setcounter{footnote}{0}
{
\hypersetup{linkcolor=black}
\tableofcontents
}
\newpage
\section{Introduction}\label{SEC-01}
The discovery of neutrino oscillations, driven by non-zero neutrino masses and mixing, provided the first robust evidence of physics beyond the Standard Model (SM)~\cite{ParticleDataGroup:2020ssz}. In order to account for the tiny neutrino masses and mixing, the SM must be extended. In these extensions, a generic outcome of mechanisms of neutrino mass generation is the existence of a non-zero neutrino magnetic moment $(\mu_{\nu})$ through quantum loop corrections. Thereby, a careful theoretical and experimental study of neutrino electromagnetic interactions may help shed some light on the underlying theory. Within the SM minimally extended to contain Dirac neutrino masses, neutrino magnetic moments are $\mu_{\nu}\lesssim 10^{-19}$ $\mu_B$ ~\cite{PhysRevLett.45.963}, where $\mu_B=e/(2 m_e)$ is the Bohr magneton, $e$ is the electron charge, and $m_e$ is the electron mass. On the other hand, the strength of transition neutrino magnetic moments, in case of Majorana scenario in the standard seesaw mechanism, is even smaller and of the order of $\sim 10^{-23}~\mu_B$~\cite{Pal:1981rm}.
Such a small values $\mu_\nu$ is far beyond the reach of current experimental capabilities.
New symmetries and interactions \cite{Voloshin:1987qy, Babu:1989wn, Babu:2020ivd, Barr:1990um, Georgi:1990za} can change the above picture, and allow for larger values of $\mu_\nu$, without getting into unrealistic values of neutrino masses (for a review, see Ref.~\cite{Giunti:2014ixa}). Neutrino magnetic moments can either be flavour-diagonal and/or flavour off-diagonal, depending on the nature of the neutrino. Dirac neutrinos can have an intrinsic magnetic moment that gives rise to spin precession for a given flavour $\alpha$, i.e. $\nu_{\alpha L} \rightarrow \nu_{\alpha R}$ (where $L,R$ refer to left-handed helicity and right-handed helicity respectively) in the presence of sufficiently strong magnetic fields~\cite{PhysRevLett.45.963}, as well as a transition magnetic moment (TMM) that induces flavour precession in addition to rotation of the spin, i.e., $\nu_{\alpha L} \rightarrow \nu_{\beta R}$~\cite{Pal:1981rm,Shrock:1982sc,Bilenky:1987ty}. For Dirac neutrinos, the latter processes would cause the resultant neutrinos to be invisible since they will have the wrong helicity to be detected through weak interactions. Within the minimally extended SM, such TMMs of Dirac neutrinos are small due to a cancellation as a result of the Glashow-Iliopolus-Maiani (GIM) mechanism. On the other hand, Majorana neutrinos can have only transition moments, and the resultant spin-flavour precession would convert a neutrino into an antineutrino of a different flavour, and vice-versa~\cite{Shrock:1982sc}. Furthermore, coherent forward scattering of neutrinos off the surrounding matter background can resonantly enhance these spin-flavour conversions, thereby termed as the resonant spin-flavour precession (RSFP)~\cite{Akhmedov:1987nc, Lim:1987tk, Akhmedov:1988uk}. This is very similar to matter enhancement of neutrino oscillations due to the Mikheyev-Smirnov-Wolfenstein (MSW) effect~\cite{PhysRevD.17.2369, Mikheev:1986gs}.
Bounds on neutrino magnetic moments can be placed using neutrinos from a variety of terrestrial as well as astrophysical sources.
The search for the neutrino magnetic moments began seven decades ago \cite{Cowan:1954pq}, even before the neutrino was discovered. However, it became a very popular topic three decades ago when the Chlorine experiment found an apparent time variation of solar neutrino flux in anti-correlation with the Sun-spot activity \cite{Davis:1988gd, Davis:1990fb}. After that, several neutrino experiments (reactor-based experiments like KRASNOYARSK \cite{Vidyakin:1992nf}, ROVNO \cite{Derbin:1993wy}, MUNU \cite{Daraktchieva:2005kn}, TEXONO \cite{Deniz:2009mu}, GEMMA \cite{Beda:2012zz} and CONUS \cite{Bonet:2022imz}; solar neutrino experiment like Borexino \cite{Borexino:2017fbd}; accelerator-based experiments like LAPMF \cite{Allen:1992qe} and LSND experiment~\cite{LSND:2001akn}) investigated neutrino magnetic moments by looking at $\nu_e-e$ scattering in general. The best constraint from these lab-based experiments, coming from Borexino, sets $\mu_{\nu} < 2.8 \times 10^{-11}~\mu_B$ \cite{Borexino:2017fbd}. The GEMMA \cite{Beda:2012zz} and CONUS \cite{Bonet:2022imz} collaborations placed somewhat less stringent constraints on neutrino magnetic moments of the order of $\mathcal{O} (10^{-11})~ \mu_B.$ Other terrestrial experiments mentioned above have shown sensitivity to neutrino magnetic moments of the order of $\mathcal{O} (10^{-10})~ \mu_B$. The topic of neutrino magnetic moments started receiving renewed attention when the XENON1T experiment~\cite{Aprile:2020tmw} observed an excess of low energy electron recoil events, which can be explained by a neutrino transition magnetic moment $\mu_{\nu} \in(1.4,2.9) \times 10^{-11} \mu_{B}$ at $90 \%$ C.L.
Strong bounds on neutrino magnetic moment can arise from astrophysical setups as well~\cite{Akhmedov:1992ea, Athar:1995cx,Nunokawa:1998vh}. The presence of a non-zero neutrino magnetic moment allows for a direct coupling between neutrinos and photons, thereby allowing for neutrino radiative decays, as well as plasmon decays to neutrino-antineutrino pairs. The strongest bounds usually arise from globular cluster stars, where plasmon decay can delay helium ignition, leading to anomalous cooling of stars. Absence of any such observational evidence leads to $\mu_{\nu}\leq 3\times 10^{-12} \mu_B$~\cite{Raffelt:1999tx}. For an updated bound, check~\cite{Capozzi:2020cbu}. Similar cooling bounds also exist from observations of neutrinos from SN1987A. The presence of a non-zero neutrino magnetic moment allows for $\nu_L\rightarrow \nu_R$ conversions, leading to further cooling, and hence a shorter duration of neutrino emission than observed. Constraining the amount of $\nu_R$ produced, one arrives at $\mu_\nu \leq 10^{-11}\, \mu_B$~\cite{PhysRevLett.61.27}. Cosmological constraints on neutrino magnetic moments arise from observation of the primordial abundances of elements during Big Bang Nucleosynthesis (BBN), which sets constraints of the order of $\mathcal{O} (10^{-10}) ~ \mu_B$~\cite{Vassh:2015yza}.
The neutrino flux from a core-collapse supernova (SN) can be an excellent tool to constrain neutrino magnetic moments. Neutrino flavour evolution inside a SN is currently an unsolved problem in the realm of particle physics. The dense matter environment, coupled with strong magnetic fields, makes a SN ideal for studying flavour transitions due to RSFP~\cite{Lim:1987tk, Akhmedov:1988uk}. This is specially relevant for Majorana neutrinos, where non-zero TMMs can lead to $\nu_e \rightarrow \bar{\nu}_{x}$, where $\nu_x$ is linear combination of the non-electron flavours, $\nu_\mu$ and $\nu_\tau$. Related studies were performed in~\cite{Ando:2002sk, Ando:2003pj, Ando:2003is, Akhmedov:2003fu}, where this mechanism was shown to mix different neutrino flavours, causing spectral changes. A simple, effective two-flavour analysis of neutrino flavour evolution in the presence of non-zero TMMs identifies, in addition to the usual MSW resonances, two extra resonances: (i) the RSFP-H, associated with the atmospheric mass-squared difference $(\Delta m^2_{31})$, occurring at high densities, and (ii) the RSFP-L, associated with the solar mass-squared difference $(\Delta m^2_{21})$, occurring at lower densities. Ref.~\cite{Akhmedov:2003fu} studied the impact of such RSFP in a three-flavour setup and discussed an additional type of resonance, the RSFP-E (electron-type), which is present in a pure three-flavour analysis. The RSFP-E can result in converting $\nu_e \rightarrow \bar{\nu}_x$ in the normal mass ordering (NO), while in the inverted mass ordering (IO), it would result in $\bar{\nu}_e \rightarrow \nu_x$. In all these cases, it was found that all the resonances associated with spin-flavour precession depend very sensitively on the electron number fraction per nucleon, denoted by $Y_e$. These results can be further complicated due to the presence of collective flavour oscillations, arising out of the non-linearity of the equations of motion associated with neutrino flavour evolution within a SN~\cite{deGouvea:2012hg,deGouvea:2013zp}.
However, flavour transitions due to RSFP will not be very effective if the spectra of different neutrino flavours are similar. This is mostly true of the cooling phase -- a period existing for $\mathcal{O}(1-10)$s post-bounce. However, the first $\mathcal{O}(30)$ms post-bounce is dominated by the deleptonization epoch, also known as the neutronization burst (described in detail in the next section). This epoch is characterized by a pure $\nu_e$ flux, with very little contamination of $\bar{\nu}_e$ and $\nu_{\mu,\tau}$. This phase is a robust feature of all hydrodynamic simulations, and hence acts as an excellent laboratory to test new neutrino properties. As a result, any new physics mechanism that ends up converting $\nu_e$s would cause a dramatic change in the $\nu_e$ spectra. In particular, it was discussed in Ref~\cite{Akhmedov:2003fu,Ando:2003is} that combination of MSW and RSFP transitions can result in $\nu_e \rightarrow \bar{\nu}_e$ conversions during this epoch. A smoking-gun signal of such conversion would be a suppression of the $\nu_e$ flux while leading to an enhancement in the $\bar{\nu}_e$ channel, which can be detected in experiments sensitive to neutrinos from a SN.
In this work, we expand on this idea to utilize the neutronization burst from a future galactic SN to put some of the strongest constraints on neutrino TMMs. We simulate the neutrino spectra from the burst phase in upcoming neutrino experiments such as the Deep Underground Neutrino Experiment (DUNE)~\cite{Acciarri:2016crz,DUNE:2020zfm}, and the Hyper-Kamiokande (HK)~\cite{Abe:2018uyc}. Since DUNE has maximum sensitivity to the $\nu_e$ spectra, while HK can measure the $\overline{\nu}_e$, both these experiments can be used to constrain TMMs of neutrinos using the neutronization burst spectra. The large sizes of these detectors and their detection channels make them ideal for probing the effect of TMM on the burst spectra. We find that the values of neutrino TMM that can be probed by these experiments for a SN occurring at 10$\,$kpc are two or three orders of magnitude better than the current terrestrial as well as astrophysical constraints. We perform a parameterized study of the effect of neutrino TMM in this paper. In order to relate it to the fundamental parameters of a given model, we consider a simple model based on $SU(2)_{ H}$ horizontal symmetry \cite{Babu:2020ivd}, which can give rise to large neutrino magnetic moments. We translate the bounds obtained from our analysis onto the model parameters.
The paper is structured as follows: in the next section, we briefly discuss the supernova neutronization burst. Then we analyze in detail the neutrino flavour conversion in the presence of neutrino magnetic moments. Subsequently, we analyze the discovery potential of neutrino magnetic moments in forthcoming experiments, and finally, we discuss the implications on neutrino properties before we conclude.
\section{Supernova neutronization burst}\label{SEC-03}
The supernova neutronization burst is a period of deleptonization of the SN core, which lasts for about 30~ms after core-bounce~\cite{Janka:2006fh}. During collapse, as the core stiffens and reaches nuclear densities, a shockwave is launched, which travels outwards dissociating the surrounding nuclei on its way into corresponding nucleons. Electron capture on protons leads to a large burst of $\nu_e$, which dominates the spectral content during this period, with subdominant contributions from $\bar{\nu}_e$ and $\nu_{\mu,\tau}$.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{initial-fluxes.pdf}
\caption{Time-integrated neutrino number fluxes (upto $50\,$ms) for the neutronization burst epoch. Note that $\nu_e$ dominates over other flavours upto $E\lesssim 25\,$MeV.}
\label{fig:Neut_initial_flux}
\end{figure}
The neutrino spectra from a SN can be parameterized using the ``alpha-fit'' spectra~\cite{Keil:2002in}:
\begin{equation}
F^0_{\nu}(E)=\frac{L_{\nu}}{\langle E_\nu \rangle^2}\frac{(\alpha+1)^{(\alpha+1)}}{\Gamma (\alpha+1)}\left(\frac{E}{\langle E_\nu \rangle}\right)^{\alpha}{\rm exp}\left[-(\alpha+1) \frac{E}{\langle E_\nu \rangle}\right]\,,
\label{spectra_ch1}
\end{equation}
where $L_\nu$ is the neutrino luminosity, $\langle E_\nu\rangle$ is the average energy of the neutrino, $\Gamma(z)$ denotes the Euler gamma function, and the pinching parameter, $\alpha$, defined as,
\begin{equation}
\frac{1}{1+\alpha}=\frac{\langle E^2\rangle-\langle E\rangle^2}{\langle E\rangle^2}\,.
\label{eq:alpha}
\end{equation}
is related to the spectral width.
Fig.~\ref{fig:Neut_initial_flux} shows the time-integrated (upto $t=50\,$ms) neutrino spectra during this period, obtained from a spherically symmetric simulation of a $15\,M_\odot$ progenitor~\cite{Garching}. Evidently, this period is dominated by the $\nu_e$ spectra for $E\leq 25\,$MeV, while the non-electron flavours dominate above that energy.
Deep inside the SN, the matter densities are high enough that the $\nu_e$ is produced as the instantaneous Hamiltonian eigenstate associated with the largest eigenvalue. During evolution, the flavour eigenstates decohere, and hence can be treated as an incoherent superposition of mass eigenstates. Assuming adiabatic evolution inside the SN, the $\nu_e$ flux at the Earth can be written as~\cite{Dighe:1999bi}
\begin{equation}
F_{\nu_e}(E) = \frac{1}{4\pi R^2}\left[|{\rm U}_{eh}|^2\, F^0_{\nu_e}(E) +
(1-|{\rm U}_{eh}|^2)\, F^0_{\nu_{\mu,\tau}}(E) \right]
\label{eq:NueEarth}
\end{equation}
where $U_{eh}$ is the corresponding flavour element of the PMNS matrix, and $h$ denotes the heaviest mass eigenstate. In the NO, the $\nu_e$ produced inside the SN comes out of it as a $\nu_3$, whereas in the IO, it comes out as a $\nu_2$. Since the other neutrino flavours are sub-dominant during this epoch, the flux in the NO is suppressed with respect to that in the IO by a factor of $\sim |{\rm U}_{e3}|^2/ |{\rm U}_{e2}|^2 \simeq 0.1 $. In principle, this suppression of the flux in the NO as compared to the IO can be utilized to probe the neutrino mass ordering~\cite{Dighe:1999bi}. For antineutrinos, since the sign of the matter potential inverts, a $\bar{\nu}_e$ produced will exit the SN as the lightest mass eigenstate. However, this entire picture changes as soon as neutrinos acquire a non-trivial magnetic moment, as we will discuss in the next section.
\section{Neutrino flavour Conversion in the presence of magnetic moments}\label{SEC-04}
Neutrinos can have Dirac and/or Majorana magnetic moments, which can be parameterized by the following interaction Lagrangians,
\begin{gather}
\mathcal{L}_{D} \supset \mu_{D} \overline{\nu_{L}} \sigma_{\eta \delta} \nu_{R} F^{\eta \delta}\\
\mathcal{L}_{M} \supset \mu_{\alpha\beta} \nu_{\alpha L}^{T} C \sigma_{\eta \delta} \nu_{\beta L}^{} F^{\eta \delta}
\end{gather}
where $F^{\eta \delta}$ is the electromagnetic field-strength tensor, and $\mu_{D}$ and $\mu_{\alpha\beta}$ denote the Dirac (or intrinsic), and Majorana (or transition) magnetic moment operators. Here $\nu_\alpha$ and $\nu_\beta$ denote neutrino flavours $\alpha$ and $\beta$ respectively, since transition magentic moments (TMM) connect neutrinos of different flavours. In the presence of a magnetic field in the exploding star, such TMM can lead to $\nu_{\alpha} \rightarrow \bar{\nu}_{\beta}$ transition, which violates lepton number by 2 units. On the other hand, the intrinsic magnetic moment (IMM) can only allow for $\nu_{\alpha L} \rightarrow \nu_{\alpha R}$ to occur, without changing the neutrino flavour.
For Majorana neutrinos, only the TMMs $(\alpha\neq \beta)$ are non-zero, while the IMMs $\mu_{\alpha\alpha}$ are zero\footnote{Due to $CPT$ invariance, if $\alpha = \beta $, the Majorana moment will be exactly zero. A particle's magnetic moment should be equal and opposite to that of its anti-particle, but since a Majorana particle is its own anti-particle, its diagonal moments must disappear. This does not preclude off-diagonal transition magnetic moments.}.
In this section we focus on TMMs of Majorana neutrinos, and study their flavour evolution as they propagate radially out in the magnetic fields of the exploding star. In such situations, neutrinos undergo several resonances driven by the matter enhanced MSW, as well as spin-flavour precession transitions. The interplay between RSFP and subsequent MSW conversions along the SN profile brings an interesting possibility where the $\nu_e$ flux during the neutronization burst can be efficiently converted into $\bar{\nu}_e$ giving a very distinctive signal in experiments sensitive to SN neutrinos.
\subsection{Evolution equation}
We consider a three-flavour setup, and study the evolution of neutrinos and antineutrinos in the presence of a finite transverse magnetic field intensity $B_{\perp}$. Ignoring temporal variations, the Equation of Motion (EoM) in matrix form is given by,
\begin{equation} \label{evolution}
i \frac{d}{d r}\left(\begin{array}{c}\nu \\ \bar{\nu}\end{array}\right)=\left(\begin{array}{cc}H_{\nu} & B_{\perp} M \\ -B_{\perp} M & H_{\bar{\nu}}\end{array}\right)\left(\begin{array}{c}\nu \\ \bar{\nu}\end{array}\right),
\end{equation}
where $r$ is the radial coordinate in the SN. The neutrino flavour dynamics is governed by the Hamiltonian,
\begin{equation}
H_{\nu}=\frac{1}{2 E} U\left(\begin{array}{ccc}0 & 0 & 0 \\ 0 & \Delta m_{21}^{2} & 0 \\ 0 & 0 & \Delta m_{31}^{2}\end{array}\right) U^{\dagger}+\left(\begin{array}{ccc}V_{\nu_e} & 0 & 0 \\ 0 & V_{\nu_\mu} & 0 \\ 0 & 0 & V_{\nu_\tau}\end{array}\right),
\end{equation}
where $U$ is the PMNS matrix, $E$ is neutrino energy and $\Delta m^2_{ij}$ gives the mass square difference between vacuum eigenstates $\nu_i$ and $\nu_j$, and $V$ denotes the forward scattering potential due to coherent scattering of neutrinos on electrons and nucleons. Assuming equal number densities of electron and protons ($n_e=n_p$), the potentials, at leading order, have the following form,
\begin{equation}
V_{\nu_e}=\sqrt{2}G_F(n_e-\frac{1}{2}n_n) \hspace{0.5cm} \text{and} \hspace{0.5cm} V_{\nu_{\mu,\tau}}=-\frac{1}{2}\sqrt{2}G_Fn_n,
\end{equation}
where $n_e$ and $n_n$ represent the electron and neutron number densities, respectively. The corresponding Hamiltonian for the antineutrinos, $H_{\bar{\nu}}$, only differs in the sign of the potential, such that $V_{\bar{\nu}_e}=-V_{\nu_e}$ and $V_{\bar{\nu}_{\mu,\tau}}=-V_{\nu_{\mu,\tau}}$. Note that the relevant quantity for the flavour evolution is the difference of the $\nu_e$ and $\nu_{\mu,\tau}$ potentials, since an overall phase term can always be rotated away.
The matrix $M$ contains the TMMs:
\begin{equation} \label{M-matrix}
M=\left(\begin{array}{ccc}0 & \mu_{e \mu} & \mu_{e \tau} \\ -\mu_{e \mu} & 0 & \mu_{\mu \tau} \\ -\mu_{e \tau} & -\mu_{\mu \tau} & 0\end{array}\right).
\end{equation}
The three flavours of neutrinos and anti-neutrinos are grouped together in a column given by
\begin{equation}
\nu=\left(\begin{array}{c}\nu_{e} \\ \nu_{\mu} \\ \nu_{\tau}\end{array}\right), \quad \bar{\nu}=\left(\begin{array}{c}\bar{\nu}_{e} \\ \bar{\nu}_{\mu} \\ \bar{\nu}_{\tau}\end{array}\right).
\end{equation}
It is the $B_{\perp} M$ term that couples the neutrinos to the antineutrinos, and gives rise to RSFP conversions. Note that in this analysis, we neglect the neutrino self-interaction terms, which are known to cause collective oscillations among different flavours~\cite{Duan:2006an,Hannestad:2006nj}. This is primarily because the negligible proportions of $\bar{\nu}_e$ and $\nu_{\mu,\tau}$ suppresses collective oscillations during the neutronization epoch.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{density-profile.pdf}
\caption{The quantities, $\rho(1-2Y_e)$ and $\rho Y_e$, relevant for RSFP and MSW transitions respectively. These are taken from the static profile of a $15 M_\odot$ star with solar metallicity~\cite{Woosley:1995ip}. Here we consider $R_\odot=696340$ Km.}
\label{fig:profile}
\end{figure}
We rewrite the EoMs in terms of the electron number fraction per nucleon $Y_e=n_e/(n_p+n_n)$, such that the electron number density $n_e=Y_e \rho/m_N$, where $\rho$ is the matter density and $m_N$ is the nucleon mass.
For our discussions with regards to resonant flavour conversions, the relevant potential difference for the MSW flavour conversion is
\begin{equation} \label{MSW-potential}
V_{\nu_e}-V_{\nu_{\mu,\tau}}=\sqrt{2}G_F \frac{\rho}{m_N} Y_e=-(V_{\bar{\nu}_e}-V_{\bar{\nu}_{\mu,\tau}}).
\end{equation}
whereas for the RSFP flavour conversions, it is
\begin{equation} \label{RSFP-potential}
V_{\nu_e}-V_{\bar{\nu}_{\mu,\tau}}=-\sqrt{2}G_F \frac{\rho}{m_N} (1-2Y_e)=-(V_{\bar{\nu}_e}-V_{\nu_{\mu,\tau}}).
\end{equation}
Clearly, in order to proceed with the neutrino flavour evolution, it is important to know the matter profile $\rho$, as well as the
electron number fraction, $Y_e$. We use the static profile model for a progenitor mass $15 M_\odot$ and solar metallicity as shown in Fig.\,\ref{fig:profile}~\cite{Woosley:1995ip}. We focus on the almost isotopically neutral region where $(1-2Y_e)$ is small.
The magnetic field within a SN can be parameterized as a dipole,
\begin{equation} \label{mag-field}
B_{\perp}=B_0\left(\frac{r_0}{r}\right)^3,
\end{equation}
where $r_0$ is the radius of the iron core, which is set to $r_0\approx 0.0024 R_\odot$ for a progenitor mass $15 M_\odot$~\cite{Totani:1996wf}. The magnitude of the magnetic field at the equator of the iron surface is given by $B_0$, and its value can be deduced from the observation of white dwarfs to be in the range $10^7$ G $\lesssim B_0 \lesssim 10^{10}$ G \cite{Totani:1996wf}.
\subsection{Level crossing scheme} \label{level-crossing}
To understand the neutrino flavour evolution, it is useful to work in the basis in which the $\nu_\mu-\nu_\tau$ and $\bar{\nu}_\mu-\bar{\nu}_\tau$ subspaces are simultaneously diagonal. The flavour states in the new basis are primed: $\nu_{\mu}'$, $\nu_{\tau}'$, $\bar{\nu}_{\mu}'$ and $\bar{\nu}_{\tau}'$. In addition, the magnetic moment elements in Eq.\,\ref{M-matrix} alter accordingly to the change of basis and are called $\mu_{e \mu'}$ and $\mu_{e \tau'}$. For simplicity, we set $\mu_{\mu \tau}=0$ as all our results are insensitive to it\footnote{For a detailed discussion of the bounds on $\mu_{\mu \tau}$, see \cite{Guzzo:2012rf}}.
We compute the energy levels of the instantaneous matter eigenstates of the Hamiltonian in Eq.\,\ref{evolution} for both mass orderings and show them in Fig.\,\ref{fig:LevCros}. The values of the mixing parameters were artificially chosen for illustration purpose in Fig.~\ref{fig:LevCros}, but the arrangements of the resonances is in good agreement with calculation using current mixing best fit values. For very high densities, $\rho \gtrsim 10^6$ $\text{g/cm}^3$, the matter eigenstates coincide with the flavour states in the primed basis. Both diagrams display five resonances: the usual MSW-H (converts $\nu_{e} \rightleftharpoons \nu_{\tau}^{\prime}$ for NO and $\bar{\nu}_{e} \rightleftharpoons \bar{\nu}_{\tau}^{\prime}$ for IO), and MSW-L (converts $\nu_{e} \rightleftharpoons \nu_{\mu}^{\prime}$ for both orderings), and three additional spin-flavour resonances, RSFP-H ($\bar{\nu}_{e} \rightleftharpoons \nu_{\tau}^{\prime}$ for NO and $\nu_{e} \rightleftharpoons \bar{\nu}_{\tau}^{\prime}$ for IO), RSFP-L ($\bar{\nu}_{e} \rightleftharpoons \nu_{\mu}^{\prime}$ for NO and IO) and RSFP-E ($\nu_{e} \rightleftharpoons \bar{\nu}_{\mu}^{\prime}$ for NO and $\bar{\nu}_{e} \rightleftharpoons \nu_{\mu}^{\prime}$ for IO). Out of these, the MSW-H, MSW-L, RSFP-H, and RSFP-L can be estimated using the two-flavour approximation. The condition for the MSW-H(L) is given by
\begin{equation}
\sqrt{2}G_F \frac{\rho}{m_N} Y_e \approx \frac{|\Delta m^2_{j1}|}{2E}\cos \theta_{1j}
\end{equation}
where $j=3$ for H and $j=2$ for L. For RSFP-H(L), the corresponding resonance condition is
\begin{equation}
\sqrt{2}G_F \frac{\rho}{m_N} (1-2Y_e) \approx \frac{|\Delta m^2_{j1}|}{2E}\cos \theta_{1j}.
\end{equation}
On the other hand, the RSFP-E only appears in a full three-flavour analysis~\cite{Akhmedov:2003fu}. Furthermore, note that the RSFP-H(L) happen at higher densities than MSW-H(L), because the former is driven by $(1-2Y_e)\lesssim 10^{-3}$, whereas the latter is governed only by $Y_e \approx 0.5$, hence the densities need to be higher in RSFP-H(L).
The corresponding adiabaticity parameter for the MSW and RSFP resonances are
\begin{equation} \label{ad-MSW}
\gamma_{\text{MSW-H(L)}}=\frac{\sin ^{2} 2 \theta_{1 j}}{\cos 2 \theta_{1 j}} \frac{\Delta m_{j 1}^{2}}{2 E} \left|\frac{1}{\rho Y_e} \frac{d\left(\rho Y_e\right)}{d r}\right|^{-1},
\end{equation}
and
\begin{equation} \label{ad-RSFP}
\gamma_{\text{RSFP-H(L)}} \simeq \frac{8 E}{\Delta m_{j1}^{2}}\left(\mu_{e \beta'} B_{\perp }\right)^{2} \left|\frac{1}{\rho (1-2Y_e)} \frac{d\left(\rho (1-2Y_e)\right)}{d r}\right|^{-1},
\end{equation}
where $\beta=\mu$ for L and $\beta=\tau$ for H.
Eqs.\,(\ref{ad-MSW}) and (\ref{ad-RSFP}) must be evaluated at the resonance point and adiabaticity is guaranteed for $\gamma \gg 1$. When neutrinos encounter an adiabatic resonance, the transitions between different energy levels in Fig.~\ref{evolution} are negligible. If adiabaticity is not met, a instantaneous matter eigenstate follows the dashed lines in Fig.~\ref{evolution} and hops from its energy level to a close one with probability,
\begin{equation} \label{hopping}
P_{\rm res}\approx e^{-\frac{\pi}{2} \gamma_{\rm res}},
\end{equation}
where $\gamma_{\rm res}$ assumes the form in Eqs.~(\ref{ad-MSW}) or (\ref{ad-RSFP}) with the subscript ${\rm res}$ referring to the specific resonance. In the non-adiabatic limit, $\gamma_{\rm res} \ll 1$, the hopping probability in (\ref{hopping}) approaches unity and the matter eigenstates corresponding the two energy levels involved in the resonance exchange their previous fluxes.
For lack of precise analytical expression for the adiabaticity of RSFP-E \cite{Akhmedov:2003fu}, we rely on numerical numerical calculations.
In what follows, we discuss these resonances in detail for both mass orderings.
\subsubsection{Normal mass ordering (NO)}
The analysis is simplified because of the adiabaticity of the MSW-H and MSW-L resonances. We find, numerically, for
\begin{equation} \label{muB-small}
\mu_\nu B_0 \lesssim 10^{-2} \mu_B \text{G}\,,
\end{equation}
the RSFP-L and RSFP-E are completely non-adiabatic. On the other hand, the RSFP-H can transit from completely non-adiabatic to adiabatic in the interval
\begin{equation} \label{muB-range}
0.5 \times 10^{-3} \mu_B \text{G} \lesssim \mu_\nu B_0 \lesssim 10^{-2} \mu_B \text{G}\,.
\end{equation}
The closer $\mu_\nu B_0$ is to $10^{-2} \mu_B \text{G}$, the more adiabatic is the RSFP-H, whereas for values $\gtrsim 10^{-2}\mu_B \text{G}$, adiabaticity at RSFP-H saturates to a maximum. The probability that transitions happen between energy levels of neutrinos due to non-adiabaticity of the RSFP-H is given by $p_H$ (i.e., $p_H=P_{\text{RSFP-H}}$, see Eq.~(\ref{hopping})).
\begin{figure} [!t]
\includegraphics[width=1\textwidth]{LevelCrossingDiagram-NH.pdf}\\
\vspace{0.1in}
\includegraphics[width=1\textwidth]{LevelCrossingDiagram-IH.pdf}
\caption{Energy level crossing diagrams for neutrinos and antineutrinos for the Hamiltonian considered in Eq.\,(\ref{evolution}). The top panel shows the crossings for NO, while the bottom panel shows the same for IO. The axes are arbitrarily scaled to highlight the resonances.}
\label{fig:LevCros}
\end{figure}
The energy levels of matter eigenstates in case of NO are given in the top panel of Fig.\,\ref{fig:LevCros}. The coloured lines indicate the evolution of the matter eigenstates (solid for neutrinos, dashed for antineutrinos), whereas the dashed black lines track the flavour content. We find that in the presence of non-zero magnetic moment deep inside the star, the $\nu_e$ state coincides with initial $\bar{\nu}_2$, and $\bar{\nu}_\mu'$ with $\nu_3$, therefore $F_{\bar{\nu}_2}=F_{\nu_e}^0$, whereas $F_{\nu_3}=F_{\bar{\nu}_\mu'}^0$. As the non-adiabatic RSFP-E resonance is hit, the flux associated with the matter eigenstates $\nu_2$ and $\nu_3$ switch, thereby $F_{\nu_e}^0 = F_{\nu_3}$ and $F_{\bar{\nu}_\mu'}^0 = F_{\bar{\nu}_2}$.
The next resonance encountered by $\nu_3$ is the MSW-H, which is adiabatic, while $\bar{\nu}_2$ finds no other resonances, and hence the neutrino states exit the SN as
\begin{eqnarray}
F_{\nu_3}&=&F_{\nu_e}^0,\,\\
F_{\bar{\nu}_2} &=& F_{\bar{\nu}_\mu'}^0.
\end{eqnarray}
The $\bar{\nu}_e$ is produced deep inside as a $\nu_2$, whereas the $\nu_\tau'$ is produced mainly as a $\nu_1$. As these states evolve, they encounter the RSFP-H, where the $\nu_2$ can flip to the $\nu_1$ and vice versa with a probability $p_H$. At this stage, the $\nu_2$ and $\nu_1$ fluxes can be written as an admixture of the initial $\bar{\nu}_e$ flux and the $\nu_\tau'$ flux. As these states propagate, the $\nu_2$ crosses adiabatic MSW-H and MSW-L resonances, therefore $F_{\nu_2}$ remains unchanged, and finally exits the SN as
\begin{equation}
F_{\nu_2} = (1-p_H) F_{\bar{\nu}_e}^0 + p_H F_{\nu_\tau'}^0\,.
\end{equation}
Meanwhile, after the RSFP-H, the $\nu_1$ flux can be written as $F_{\nu_1} = p_H F_{\bar{\nu}_e}^0+(1-p_H) F_{\nu_\tau'}^0
$. As it propagates outwards, it encounters two resonances as well, the non-adiabatic RSFP-L, and the adiabatic MSW-L (check inset plot). At the RSFP-L point, transition happens between $\nu_1$ and $\bar{\nu}_1$. Since initial $\bar{\nu}_1$ flux deep inside the SN is $F_{\bar{\nu}_1}=F_{\nu_\mu'}^0$, after encountering the RSFP-L resonance, one has
\begin{eqnarray}
F_{\nu_1} &=& F_{\nu_\mu'}^0\,,\\
F_{\bar{\nu}_1} &=& p_H F_{\bar{\nu}_e}^0+(1-p_H) F_{\nu_\tau'}^0.
\end{eqnarray}
Finally, as the $\bar{\nu}_3$ is not affected by any of the resonances, we have
\begin{equation}
F_{\bar{\nu}_3}=F_{\bar{\nu}_\tau'}^0.
\end{equation}
Therefore, the $\nu_e$ and the $\bar{\nu}_e$ fluxes at the Earth are
\begin{eqnarray}
F_{\nu_e}&=&|U_{e1}|^2 F_{\nu_\mu'}^0 + |U_{e2}|^2 \left[(1-p_H) F_{\bar{\nu}_e}^0 + p_H F_{\nu_\tau'}^0\right]+ |U_{e3}|^2 F_{\nu_e}^0, \\
F_{\bar{\nu}_e} &=& |U_{e1}|^2 \left[p_H F_{\bar{\nu}_e}^0+(1-p_H) F_{\nu_\tau'}^0 \right] + |U_{e2}|^2 F_{\bar{\nu}_\mu'}^0 + |U_{e3}|^2 F_{\bar{\nu}_\tau'}^0.
\label{eq:NONeutrinoflux}
\end{eqnarray}
Now, if the neutrino magnetic moment were zero, there would be no RSFP, and the flux arriving at the Earth can be obtained by setting $p_H = 1$ in the above equations. Thus, in the limit where all the non-electron neutrino fluxes are set equal $(F_{\nu_\mu'}=F_{\nu_\tau'}=F_{\bar{\nu}_\mu'}=F_{\bar{\nu}_\tau'} = F_{\nu_x})$, one can write Eq.\,\ref{eq:NONeutrinoflux} as
\begin{eqnarray}
F_{\nu_e} &=& F_{\nu_e}^{MSW}-|U_{e2}|^2 (1-p_H)(F_{\nu_x}^0-F_{\bar{\nu}_e}^0) \\
F_{\bar{\nu}_e}&=&F_{\bar{\nu}_e}^{MSW}+|U_{e1}|^2 (1-p_H)(F_{\nu_x}^0-F_{\bar{\nu}_e}^0)
\label{Fnue-NH}
\end{eqnarray}
where $F_{\nu}^{MSW}= F_{\nu}(p_H=1) $.
With this information, we can calculate the changes in the neutronization fluence due to the presence of a non-zero TMM. From Fig.~\ref{fig:Neut_initial_flux}, note that for
$E\lesssim 20$ MeV, $F_{\bar{\nu}_e}^0 \approx F_{\nu_x}^0$, and hence the effect due to a non-zero TMM cancels out. On the other hand, for $E \gtrsim 20$ MeV, we have $F_{\nu_x}^0 > F_{\nu_e}^0$. This ends up reducing the $\nu_e$ flux, and increasing the $\bar{\nu}_e$ flux received at the Earth.
We estimate that in case of complete adiabaticity of the RSFP-H resonance ($p_H=0$), for $E=25$ MeV, one has
\begin{equation}\label{ratios-NH}
\frac{F_{\nu_e}}{F_{\nu_e}^{MSW}} \sim 0.85 \hspace{0.5cm} \text{and} \hspace{0.5cm} \frac{F_{\bar{\nu}_e}}{F_{\bar{\nu}_e}^{MSW}} \sim 1.5 \hspace{0.5cm}
\end{equation}
Therefore, fingerprints of RSFP conversion can be visible in upcoming experiments sensitive to the neutronization burst flux of a galactic SN.
\subsubsection{Inverted mass ordering (IO)}
A similar argument can be applied if the mass ordering is inverted (see bottom panel of Fig.\,\ref{fig:LevCros}).
At very high densities, we have $F_{\nu_1}=F_{\nu_e}^0$ and $F_{\nu_2}=F_{\bar{\nu}_\tau'}^0$.
As the neutrinos propagate, a RSFP-H causes a flip between a $\nu_1$ and $\nu_2$ with probability $p_H$. Since the MSW-L resonance is adiabatic, the $\nu_2$ finally emerges as
\begin{equation}
F_{\nu_2}=p_H F_{\nu_e}^0 + (1-p_H) F_{\bar{\nu}_\tau'}^0.
\end{equation}
On the other hand, the $\nu_1$, with flux $F_{\nu_1}=(1-p_H) F_{\nu_e}^0 + p_H F_{\bar{\nu}_\tau'}^0$ just after RSFP-H, goes through the adiabatic MSW-H and the non-adiabatic RSFP-E, where it flips completely to $\bar{\nu}_1$, and vice-versa. To get the final $\nu_1$ flux, we need to track the flavour evolution of $\bar{\nu}_1$.
The $\bar{\nu}_1$ is produced deep inside as a $\bar{\nu}_e$, while $\bar{\nu}_3$ is produced as a $\nu_\mu'$. At RSFP-L, $\bar{\nu}_1$ and $\bar{\nu}_3$ exchange fluxes so that $F_{\bar{\nu}_1}=F_{\nu_\mu'}^0$ and $F_{\bar{\nu}_3}=F_{\bar{\nu}_e}^0$. The $\bar{\nu}_3$ flux remains unchanged and exits the SN as
\begin{equation}
F_{\bar{\nu}_3}=F_{\bar{\nu}_e}^0.
\end{equation}
On the other hand, the $\bar{\nu}_1$ exchanges flux with $\nu_1$ at RSFP-E and reaches the surface of the star as
\begin{equation}
F_{\bar{\nu}_1}=(1-p_H) F_{\nu_e}^0 + p_H F_{\bar{\nu}_\tau'}^0\,,
\end{equation}
whereas the $\nu_1$ is emitted as
\begin{equation}
F_{\nu_1}=F_{\nu_\mu'}^0\,.
\end{equation}
The $\bar{\nu}_2$ and $\nu_3$ have relatively trivial dynamics, and they are emitted as
\begin{eqnarray}
F_{\nu_3} &=& F_{\nu_\tau'}^0\,,\\
F_{\bar{\nu}_2} &=& F_{\bar{\nu}_\mu'}^0.
\end{eqnarray}
Collecting these fluxes, the $\nu_e$ and $\bar{\nu}_e$ fluxes at the Earth are given by
\begin{eqnarray}\label{eq:NuefluxIO}
F_{\nu_e}&=&|U_{e1}|^2 F_{\nu_\mu'}^0 + |U_{e2}|^2 \left[p_H F_{\nu_e}^0 + (1-p_H) F_{\bar{\nu}_\tau'}^0\right]+ |U_{e3}|^2 F_{\nu_\tau'}^0\,,\\
\label{eq:NuebarfluxIO}
F_{\bar{\nu}_e} &=& |U_{e1}|^2 \left[(1-p_H) F_{\nu_e}^0 + p_H F_{\bar{\nu}_\tau'}^0 \right] + |U_{e2}|^2 F_{\bar{\nu}_\mu'}^0 + |U_{e3}|^2 F_{\bar{\nu}_e}^0\,.
\end{eqnarray}
In terms of the pure MSW fluxes (which is obtained by setting $p_H=1$ in Eqs.\,(\ref{eq:NuefluxIO}-\ref{eq:NuebarfluxIO})), we get
\begin{eqnarray}
F_{\nu_e}&=&F_{\nu_e}^{MSW}-|U_{e2}|^2 (1-p_H)(F_{\nu_e}^0- F_{\nu_x}^0)\,\\
F_{\bar{\nu}_e}&=&F_{\bar{\nu}_e}^{MSW}+|U_{e1}|^2 (1-p_H)(F_{\nu_e}^0- F_{\nu_x}^0)\,,
\end{eqnarray}
where all the non-electron neutrino fluxes are set equal $(F_{\nu_\mu'}=F_{\nu_\tau'}=F_{\bar{\nu}_\mu'}=F_{\bar{\nu}_\tau'} = F_{\nu_x})$.
Thus, the RSFP signal for inverted ordering will be more relevant at $E\lesssim 20$ MeV once $F_{\nu_e}^0 \gg F_{\nu_x}^0$. At $E=10$ MeV we estimate
\begin{equation} \label{ratios-IH}
\frac{F_{\nu_e}}{F_{\nu_e}^{MSW}} \sim 0.3 \hspace{0.5cm} \text{and} \hspace{0.5cm} \frac{F_{\bar{\nu}_e}}{F_{\bar{\nu}_e}^{MSW}} \sim 7.
\end{equation}
From Eqs.~(\ref{ratios-NH}) and (\ref{ratios-IH}), the RSFP signal will be more dominant in the IO compared to NO, due to a large enhancement in the $\bar{\nu}_e$ flux, which is quite uncharacteristic of the neutronization epoch. As a result, experiments sensitive to $\bar{\nu}_e$ from a SN will suddenly see a large flux during the shock breakout phase, which can be a smoking-gun signal of the presence of neutrino TMM.
\section{Sensitivity in upcoming experiments}\label{SEC-05}
In this section, we study the impact of RSFP in the neutronization burst signal at upcoming neutrino detectors.
As discussed in the previous sections, the effect of the RSFP is to reduce the $\nu_e$ flux, and increase the $\bar{\nu}_e$. Depending on the mass ordering, the effect can be quite large. As a result, sensitivity can be maximised using a combination of detectors sensitive to the $\nu_e$ and the $\bar{\nu}_e$ flux. We focus on the Deep Underground Neutrino Experiment (DUNE), which can detect $\nu_e$, and the Hyper-Kamiokande (HK) which is mainly sensitive to the $\bar{\nu}_e$ flux. The number of events of species $\nu_\alpha$ detected per unit energy is
\begin{equation} \label{event-spectrum}
\frac{dN_{\nu_{\alpha}}}{dE_r}=\frac{N_{\rm tar}}{4 \pi R^2} \int d E_t F_{\nu_{\alpha}}(E_t)\sigma_{\alpha}(E_t) W(E_r,E_t),
\end{equation}
with $N_{\rm tar}$ the number of targets in the detector material, $R$ the distance of the supernova from the Earth, $F_{\nu_\alpha}$ is the $\nu_\alpha$ flux that reaches the detector, $\sigma_\alpha$ is the relevant neutrino interaction cross section and $W$ is the Gaussian energy resolution function with width $\sigma_E$ that depends on that experimental setup. $E_t$ is the actual value of neutrino energy, while $E_r$ is the experimentally reconstructed energy.
For the analysis presented in this paper, we choose a fiducial distance of $R=10\,$kpc. For the neutrino oscillation parameters, we assume the best fit values in \cite{deSalas:2020pgw}.
The TMM matrix $M$ in Eq.\,\ref{M-matrix}, with $\mu_{\mu \tau}=0$, has two independent elements $\mu_{e \mu}$, $\mu_{e \tau}$. We consider the following cases:
\begin{enumerate}
\item \label{case 1} both $\mu_{e \mu}$ and $\mu_{e \tau}$ are equal, set to $\mu_\nu$, and considered free parameters.
\item \label{case 2} $\mu_{e \mu}=0$ and $\mu_{e \tau}=\mu_\nu$ is a free parameter.
\item \label{case 3} $\mu_{e \mu}=\mu_\nu$ is considered a free parameter, while $\mu_{e \tau}=0$.
\end{enumerate}
Because the TMMs always appear together with $B_\perp (r)$ in the equations, all our results are sensitive to the quantity $\mu_\nu B_0$, and not independently to $\mu_\nu$ or $B_0$. Hence, to derive constraints on $\mu_\nu$, one needs prior knowledge of $B_0$ from simulations and/or astrophysical observations.
The schemes of level crossing and resonances are the same for all cases. For the specific interval in eq.~(\ref{muB-range}), MSW-H and MSW-L are always adiabatic while RSFP-L and RSFP-E are non-adiabatic. However, the degree of adiabaticity of RSFP-H, given by the hoping probability $p_H$, is strongly dependent on the configuration of the matrix $M$, specifically on the magnitude of $\mu_{e \tau'}$ (see Eq.~(\ref{ad-RSFP})). For the set of mixing parameters and magnetic field assumed in this work, $(\mu_{e \tau'})_1>(\mu_{e \tau'})_3>(\mu_{e \tau'})_2$ for the entire energy range close to RSFP-H, where the subscript points to one of the cases above. Therefore, the strength of conversions at RSFP-H has the following hierarchy: case \ref{case 1}$>$ case \ref{case 3} $>$ case \ref{case 2} for the same value of $\mu_{\nu} B_0$. As a consequence, the fluxes on Earth will change accordingly and that translates into different experimental sensitivities.
The DUNE underground facility in South Dakota uses $40$-kton of liquid argon in time projection chambers. The experiment will be able to detect neutrinos from few MeV up to GeV. The relevant interaction for SN neutrinos that reach the detector is the charged current scattering with argon nuclei, $\nu_{e}+{ }^{40}\mathrm{Ar} \rightarrow { }^{40}\mathrm{K}^{*}+e^{-}$.
The final state consists of electrons, and excited states of the potassium nuclei, which de-excite producing a cascade of photons in the energy range between 5 and 50 MeV. For our analysis, we use the Monte Carlo event simulator MARLEY to generate the $\nu_e-\mathrm{Ar}$ interaction cross section. The energy resolution is taken to be $\sigma_E/\mathrm{MeV}=0.11 \sqrt{E_{r} / \mathrm{MeV}}+0.02 E_{r} / \mathrm{MeV}$, which implies that $\sigma_E/E_r \sim 5 \%$ at $10$ MeV. In addition, we use bins of $5\,$MeV, with the energy reconstruction threshold set at $4\,$MeV.
HK, an upgrade of the Super-Kamiokande (SK) detector, is a next generation water-Cherenkov detector, with a fiducial volume of $187$ kt in each of its two tanks. It is expected to record a huge number of events in case of a galactic SN. HK primarily detects $\bar{\nu}_e$ through the Inverse Beta Decay (IBD) channel, $ \bar{\nu}_{e}+p \rightarrow e^{+}+n$. The IBD cross section for the MeV range is extracted from \cite{Strumia:2003zx} and the energy resolution is the same of SK, namely $\sigma_E/\mathrm{MeV}=0.6 \sqrt{E_r/\mathrm{MeV}}$ with $\sigma_E/E_r \sim 20 \%$ at $10$ MeV. We simulate events at HK detector using bins of $8$ MeV, and a detector threshold of $3$ MeV.
We compute the number of events at each detector for values of the parameter $\mu_\nu B_0$ lying within the range given in (\ref{muB-range}), such that the RSFP-L/E are non-adiabatic, whereas the RSFP-H can transition from non-adiabatic to adiabatic. We focus on the most general Case \ref{case 1}. In Fig.~\ref{spectrum-NO}, we show how the neutrino event spectra is altered at DUNE and HK, assuming normal mass ordering, in the presence of a non-zero neutrino TMM. The figure nicely illustrates the features we discussed in Sec.~\ref{level-crossing} using the level crossing diagram: a spectral distortion, resulting in suppression of events for DUNE, while simultaneously enhancing the events in HK, for bins higher than $10$ MeV. We find that our results are sensitive to values of $\mu_\nu B_0$ as small as $10^{-3} \mu_B\,$G; the spectral distortion becomes more prominent as these values are increased. Case \ref{case 2} and \ref{case 3} also demonstrate similar spectral distortions, albeit smaller, and hence we do not show them here.
In fig.~\ref{spectrum-IO}, we show a similar analysis for the IO, where this effect is expected to be more drastic. We find that for energy bins $E_r \lesssim 25$ MeV, there is a sharp decrease in events for DUNE, while HK records an almost three-fold increase in number of events. This is consistent with our understanding in Sec.~\ref{level-crossing} (see the estimations in eq.~(\ref{ratios-IH})). The decrease in events in DUNE, combined with the sharp increase in HK, illustrates that the combination of RSFP-H with MSW-H is very efficient in converting $\nu_e$ to $\bar{\nu}_e$. This can be considered as a smoking gun signal of non-zero TMMs of neutrinos.
\begin{figure}[!t]
\centering
\includegraphics[width=0.45\textwidth]{DUNE-NH-bins.pdf}~~
\includegraphics[width=0.45\textwidth,height=0.3\textwidth]{HK-NH-bins.pdf}
\caption{Expected event spectrum for a supernova explosion $10$ kpc away in DUNE (left, $\nu_e$ channel) and HK (right, $\bar{\nu}_e$ channel) for distinct values of $\mu_\nu B_0$. We assume NO for neutrino masses. }
\label{spectrum-NO}
\end{figure}
\begin{figure}[!t]
\includegraphics[width=0.45\textwidth]{DUNE-IH-bins.pdf}~~
\includegraphics[width=0.455\textwidth, height=0.295\textwidth]{HK-IH-bins.pdf}
\caption{Expected event spectrum for a supernova explosion $10$ kpc away in DUNE (left, $\nu_e$ channel) and HK (right, $\bar{\nu}_e$ channel) for distinct values of $\mu_\nu B_0$. We assume IO for neutrino masses. }
\label{spectrum-IO}
\end{figure}
We perform a $\chi^2$ analysis to study the sensitivity of DUNE and HK to the presence of neutrino TMM, using the signals from a future galactic SN. We use the Poissonian $\chi^2$ estimator,
\begin{equation} \label{chisq}
\chi^{2}=\sum_{i=1}^{n} 2\left[F_{i}-D_{i}+D_{i} \ln \left(D_{i} / F_{i}\right)\right]\,,
\end{equation}
where $F_i$ is the number of events predicted for a specific non zero value of $\mu_\nu B_0$ and $D_i$ is the expected number of events assuming $\mu_\nu B_0=0$. The subscript $i$ refers to the $i$-th bin and $n$ is the number of bins of the experiment.
Figs.~\ref{chisq-NO} and \ref{chisq-IO} show $\chi^2$ as a function of $\mu_\nu B_0$ for NO and IO, respectively. The null hypothesis is that neutrinos undergo pure MSW flavour conversions, and there is no RSFP involved. As discussed before, the results are sensitive to the magnitude of the product $\mu_\nu B_0$, and it is not possible to disentangle one effect from the other. DUNE shows very little sensitivity to RSFP for NO, as we show in Fig.~\ref{chisq-NO}. On the other hand, for IO (see Fig.\,\ref{chisq-IO}), DUNE can exclude at $95 \%$ C.L. ,
\begin{eqnarray} \label{limits-DUNE-IO}
(\mu_\nu B_0)_1 &\gtrsim& 3 \times 10^{-3} \mu_B \text{G}\,,\\ (\mu_\nu B_0)_2 &\gtrsim& 7 \times 10^{-3} \mu_B \text{G}\,,\\ (\mu_\nu B_0)_3 &\gtrsim& 6 \times 10^{-3}\mu_B \text{G}\,,
\end{eqnarray}
where the subscripts refers to specific cases \ref{case 1}, \ref{case 2} and \ref{case 3} respectively.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{DUNE-NH-chisq.pdf}~
\includegraphics[width=0.5\textwidth,height=0.325\textwidth]{HK-NH-chisq.pdf}
\caption{$\chi^2$ in Eq.\,(\ref{chisq}) as a function of $\mu_\nu B_0$ assuming NO for DUNE (left) and HK (right). The analysis compares the RSFP hypothesis with the pure MSW one for the different cases \ref{case 1}, \ref{case 2} and \ref{case 3} defined in the text. As expected, the strength of the signal and hence the sensitivity is always greater for case \ref{case 1} for both mass orderings.}
\label{chisq-NO}
\end{figure}
\begin{figure}[!t]
\includegraphics[width=0.5\textwidth]{DUNE-IH-chisq.pdf}~
\includegraphics[width=0.5\textwidth]{HK-IH-chisq.pdf}
\caption{$\chi^2$ in Eq.~(\ref{chisq}) as a function of $\mu_\nu B_0$ assuming NO for DUNE (left) and HK (right) that compares the RSFP hypothesis with the pure MSW one for the different cases \ref{case 1}, \ref{case 2} and \ref{case 3} defined in the text. }
\label{chisq-IO}
\end{figure}
Due to its large size, HK will be sensitive to both the mass orderings. For the NO, HK can exclude, at $95 \%$ C.L.,
\begin{eqnarray} \label{limits-HK-NO}
(\mu_\nu B_0)_1 &\gtrsim& 4.5 \times 10^{-3} \mu_B \text{G}\,,\\ (\mu_\nu B_0)_2 &\gtrsim& 9 \times 10^{-3} \mu_B \text{G}\,,\\ (\mu_\nu B_0)_3 &\gtrsim& 8 \times 10^{-3}\mu_B \text{G}\,.
\end{eqnarray}
The best sensitivity to TMM comes from HK in case of IO, and it can exclude, at $95 \%$ C.L.,
\begin{eqnarray} \label{limits-HK-IO}
(\mu_\nu B_0)_1 &\gtrsim& 0.6 \times 10^{-3} \mu_B \text{G}\,,\\
(\mu_\nu B_0)_2 &\gtrsim& 1.4 \times 10^{-3} \mu_B \text{G}\,,\\
(\mu_\nu B_0)_3 &\gtrsim& 1.2 \times 10^{-3}\mu_B \text{G}. \label{limits-HK-IO-3}
\end{eqnarray}
As discussed, for a given $B_0$, these bounds above can be translated into bounds on the TMM of neutrinos. Table~\ref{summary_NMM} summarizes the derived sensitivities to $\mu_\nu$ assuming $B_0=10^{10}\,$G.
\begin{table}[]
\centering
\resizebox{0.8\textwidth}{!}{%
\begin{tabular}{|c|llllll|}
\hline
\hline
\multirow{3}{*}{\textbf{Experiments}} &
\multicolumn{6}{l|}{\textbf{Sensitivities on Neutrino Magnetic Moments (in $\mu_B$)}} \\ \cline{2-7}
&
\multicolumn{2}{c|}{\textbf{CASE 1}} &
\multicolumn{2}{c|}{\textbf{CASE 2}} &
\multicolumn{2}{c|}{\textbf{CASE 3}} \\ \cline{2-7}
&
\multicolumn{1}{c|}{\textbf{NO}} &
\multicolumn{1}{c|}{\textbf{IO}} &
\multicolumn{1}{c|}{\textbf{NO}} &
\multicolumn{1}{c|}{\textbf{IO}} &
\multicolumn{1}{c|}{\textbf{NO}} &
\multicolumn{1}{c|}{\textbf{IO}} \\ \hline
\textbf{HK} &
\multicolumn{1}{l|}{$4.5 \times 10^{-13}$} &
\multicolumn{1}{l|}{$6 \times 10^{-14}$} &
\multicolumn{1}{l|}{$9 \times 10^{-13}$} &
\multicolumn{1}{l|}{$1.4 \times 10^{-13}$} &
\multicolumn{1}{l|}{$8 \times 10^{-13}$} &
\multicolumn{1}{l|}{$1.2 \times 10^{-13}$} \\
\hline
\textbf{DUNE} &
\multicolumn{1}{l|}{$-$} &
\multicolumn{1}{l|}{$3 \times 10^{-13}$} &
\multicolumn{1}{l|}{$-$} &
\multicolumn{1}{l|}{$7 \times 10^{-13}$} &
\multicolumn{1}{l|}{$-$} &
\multicolumn{1}{l|}{$6 \times 10^{-13}$} \\
\hline
\hline
\end{tabular}
}
\caption{Experimental sensitivities on neutrino magnetic moments for different benchmark scenarios for a fixed value of magnetic field strength $B_0 = 10^{10}$ G.}
\label{summary_NMM}
\end{table}
Uncertainties may affect the derived sensitivity. However, such impact need not be drastic as the neutronization burst is a robust feature of all SN simulations with a relatively weak dependence on progenitor mass and/or specific traits of the models \cite{Kachelriess:2004ds, Serpico:2011ir, OConnor:2018sti, Tang:2020pkp}. Indeed, other studies show that different progenitor models and neutrino emission parameters yield changes that are smaller than the statistical uncertainties for current water Cherenkov detectors \cite{Kachelriess:2004ds}. For this reason, we do not perform a dedicated study about the influence of uncertainty in the fluence parameters on the expected sensitivities to $\mu_\nu B_0$.
Systematic uncertainties can also impact the above sensitivities, but owing to energy dependent spectral distortion of the neutrino signal in case of nonzero $\mu_\nu B_0$, only mild changes are expected. To see how systematics affect the $\chi^2$, we modify Eq.~(\ref{chisq}) as
\begin{equation} \label{chisq-sys}
\chi^{2}=\sum_{i=1}^{n} 2\left[(1+\xi)F_{i}-D_{i}+D_{i} \ln \left(\frac{D_{i} }{(1+\xi) F_{i}}\right)\right]+\frac{\xi^2}{\sigma^2} \,,
\end{equation}
where $\xi$ is nuisance parameter and $\sigma$ an overall normalization error that we assume to be $50 \%$. $\xi$ is allowed to vary in the $3 \sigma$ range and we minimize the $\chi^2$ over it. We focus on case \ref{case 1} and summarize new exclusion limits at $95 \%$ C.L. below
\begin{eqnarray} \label{limits-sys}
\text{DUNE(IO):} \hspace{0.2cm} (\mu_\nu B_0)_1 &\gtrsim& 3.5 \times 10^{-3} \mu_B \text{G}\,,\\
\text{HK(NO):} \hspace{0.2cm} (\mu_\nu B_0)_1 &\gtrsim& 6 \times 10^{-3} \mu_B \text{G}\,,\\
\text{HK(IO):} \hspace{0.2cm} (\mu_\nu B_0)_1 &\gtrsim& 0.7 \times 10^{-3}\mu_B \text{G},
\end{eqnarray}
with no sensitivity from DUNE in case of normal ordering.
A final comment on the adiabaticity of the RSFP-E resonance is in order. For $\mu_\nu B_0 > 10^{-2}$ $\mu_B \text{G}$, the RSFP-E starts to be partially adiabatic. This can end up affecting the energy spectra in both the mass orderings. We checked that it is possible for the event spectra in the NO even to mimic that of the IO in certain cases. We do not comment on these cases further and leave the analysis to a future study.
\section{Implications on neutrino properties}
In this section, we mull over the possible impact of our results on the uncharted territories of neutrino physics. To do so, we first briefly summarize the experimental constraints and theoretical predictions on neutrino magnetic moments, along with our results, in Fig.~\ref{sum}. Existing limits on neutrino magnetic moments are shown in blue, while the red lines indicate the future sensitivities on neutrino magnetic moments at DUNE and HK, based on our analysis, by setting magnetic field strength $B_0 =10^{10}$ G. Note that as compared to existing limits, the sensitivity can be upgraded by two or three orders of magnitude (from $\mathcal{O} [10^{-11}]~ \mu_B$ to $\mathcal{O} [10^{-14}]~ \mu_B$) by utilizing the neutrino spectra from the SN neutronization burst phase in forthcoming neutrino experiments like the DUNE, and HK. This is more stringent than the expected sensitivity when considering the scattering at the DUNE near detector \cite{Mathur:2021trm}. This will have consequences for three of the most important unanswered questions in neutrino physics:
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\textwidth]{summary_plot.pdf}
\caption{Summary of experimental constraints and theoretical predictions on neutrino magnetic moments. The red lines indicate the future sensitivities on neutrino magnetic moments at DUNE and HK experiments based on our analysis. Here we set magnetic field strength $B_0 =10^{10}$ G. Theoretical predictions for maximum possible strengths of neutrino magnetic moments are summarized from Ref.~\cite{Babu:2020ivd} and references therein. See text for details.}
\label{sum}
\end{figure}
\begin{itemize}
\item {\textsl{Dirac/Majorana nature of neutrinos:} }
Based on an Effective Field Theory (EFT) analysis, it has been shown that Dirac neutrino magnetic moments over $10^{-14}\mu_B$ would not be natural, since this would generate unacceptably large neutrino masses at higher loops \cite{Bell:2005kz}. Weak interaction corrections to neutrino mass originating from the magnetic moment operator are excessive in the case of a Dirac neutrino magnetic moment. However, such a correction is small for Majorana neutrinos since it is proportional to the mass differences of charged leptons.
Because of EFT naturalness considerations, the transition magnetic moments for Majorana neutrinos are permitted to be significantly greater than the Dirac scenario \cite{Davidson:2005cs,Bell:2006wi}. For instance, if the new physics scale is roughly a TeV, EFT would allow the transition magnetic moment as large as $\mu_{\nu} \sim 10^{-7}\mu_B$. Therefore, if DUNE or Hyper-Kamiokande experiment measure neutrino magnetic moments at the level of $10^{-13}~ \mu_B$, it is more likely that neutrinos are Majorana in nature.
\item {\textsl{Neutrino mass ordering:}} The SN neutronization burst has long been advertised as one of the cleanest means to identify the neutrino mass ordering. The presence (absence) of a distinct peak during the burst phase indicates inverted (normal) mass ordering. However, this is true only in the absence of any new physics in the neutrino sector. If neutrinos have a finite magnetic moment, it can result in the suppression of the peak in the IO due to $\nu_e\rightarrow \bar{\nu}_e$ conversion. As a result, this might make identification of the mass ordering from the burst phase ambiguous.
\item {\textsl{Neutrino mass generation mechanism:}} Due to similar chiral structure of the neutrino magnetic moment and neutrino mass operator, while removing the photon line from the loop diagram that yields the neutrino magnetic moment, a neutrino mass term is generated, and $\mu_\nu$ generally becomes proportional to $m_\nu$. One can estimate $m_\nu$ originating from such diagrams as
\begin{equation}
m_\nu \sim \frac{ \mu_\nu}{\mu_B} \,\frac{\Lambda^2}{2 m_e},
\label{order}
\end{equation}
where $\Lambda$ is the mass scale of the heavy particle inside the loop diagram. Experimental searches generally prefer charged particles with masses heavier than 100 GeV \cite{Babu:2019mfe}. Without extra symmetries (and in the absence of severe fine-tuning), this implies that neutrino mass $m_\nu$ must be $\sim 0.1$ MeV in order to generate neutrino magnetic moments in the order of $10^{-11}~\mu_B$ for a new physics scale $\sim 100$ GeV. This prediction of active neutrino mass contradicts the observed neutrino masses by six orders of magnitude. Therefore, one requires leptonic symmetries \cite{Voloshin:1987qy, Barr:1990um, Babu:2020ivd} in order to get large neutrino magnetic moments while being consistent with tiny neutrino masses and mixings utilizing their different Lorentz structure.
There are several BSM extensions to generate neutrino masses and mixings. However, most of these extensions cannot accommodate large neutrino magnetic moments ($\gtrsim 10^{-14}\mu_B$). For instance, if right-handed neutrinos are introduced in the SM to generate tiny Dirac neutrino mass, neutrino magnetic moment can be expressed as \cite{Fujikawa:1980yx}
\begin{equation}
\mu_\nu = \frac{ e G_F m_\nu}{8 \sqrt{2}\pi^2} = 3 \times 10^{-20} \mu_B\, \left(\frac{m_\nu}{0.1~{\rm eV}}\right)~,
\label{dirac}
\end{equation}
and it can be as large as $\sim 10^{-20}~\mu_B$. In the standard seesaw scenario, where neutrinos are Majorana particles, the transition magnetic moments as a result of Standard Model interactions are given by
\cite{Pal:1981rm}
\begin{equation}
\mu_{ij} = -\frac{3eG_F}{32\sqrt{2}\pi^2}(m_i \pm m_j)\sum_{\ell = e, \mu,\tau}U_{\ell i}^* U_{\ell j} \frac{m_\ell^2}{m_W^2}.
\end{equation}
The strength of transition neutrino magnetic moment is even smaller and of the order of $\sim 10^{-23}~\mu_B$ in this scenario. In the minimal Left-Right symmetric model, neutrino magnetic moment can be comparatively enhanced due to non-standard interactions via $W_R^\pm$ gauge boson \cite{Giunti:2014ixa}. However, the current experimental limit on the mixing angle between $W_R^\pm$ and $W^\pm$ does not permit the strength to be greater than $\sim 10^{-14}~\mu_B$. Similar strength of neutrino magnetic moment can be achievable in $R$-parity-violating supersymmetric extensions \cite{Kim:1976gk}. It was recently systematically analyzed and demonstrated in Ref.~\cite{Babu:2020ivd} that if neutrino magnetic moments are measured at the current experimental sensitivity $\mathcal{O}(10^{-11})~\mu_B$ level, neutrinos are most likely Majorana particles, and the models based on $SU(2)_H$ symmetry\cite{Babu:2020ivd, Babu:1989wn, Babu:1990wv} fit well within this category. Neutrino mass models \cite{Zee:1980ai, Babu:1992vq, Babu:2019mfe} based on spin-symmetry mechanism \cite{Barr:1990um} can accommodate neutrino transition magnetic moments as large as $\mathcal{O}(10^{-12})~\mu_B$ \cite{Babu:2020ivd}. All these predictions on neutrino magnetic moments in different neutrino mass models \cite{Babu:2020ivd, Lindner:2017uvt, Giunti:2014ixa} are shown in green coloured lines in Fig.~\ref{sum}. It is quite interesting to see that the investigation of neutrino magnetic moments can be an excellent tool in searching for the theory underlying the neutrino mass generating mechanism.
\end{itemize}
\begin{figure}[!t]
\centering
\includegraphics[width=0.65\textwidth]{all_su2h.pdf}
\caption{Different theoretical and experimental limits and future sensitivities on $SU(2)_H$ horizontal symmetric model. See text for details.
}
\label{pheno:su2h}
\end{figure}
Any UV complete model of large neutrino magnetic moments is expected to provide a more extensive phenomenology. For completeness of our study, we consider the model based on $SU(2)_H$ horizontal symmetry~\cite{Babu:2020ivd} which generates naturally large neutrino transition magnetic moments consistent with light neutrino masses and other existing experimental constraints. Several phenomenological consequences of this model have been recently studied in a different context~\cite{Babu:2020ivd, Babu:2021jnu}.
Following similar conventions and the same Lagrangian structure from Ref.~\cite{Babu:2020ivd}, we summarize the different phenomenological implications in the $SU(2)_H$ model parameter space in Fig.~\ref{pheno:su2h}. The transition magnetic moments of neutrino arise through charged Higgs induced quantum loop corrections. Due to SU(2)$_H$ symmetric limit, the contributions from two loop diagrams add up for neutrino magnetic moments, whereas it subtracts for neutrino mass contributions. In Fig.~\ref{pheno:su2h}, $h^+$ denotes the responsible charged scalar field, and $f$ is the relevant Yukawa term contributing to the neutrino magnetic moments. The colour shaded regions are excluded from current experimental searches: the light blue shaded region is excluded from charged Higgs searches at LEP experiment; the red shaded region is excluded from Borexino NSI searches \cite{Agarwalla:2019smc}; NSI from the global fit to neutrino oscillation data \cite{Esteban:2018ppq} constrains dark green shaded region, IceCube atmospheric data \cite{Esmaili:2013fva} on NSI imposes bound as indicated by light green shaded region; the yellow shaded region is bounded from $\tau $ lifetime constraints \cite{Babu:2020ivd}; the gray shaded region is excluded from neutrino magnetic moment searches at Borexino experiment \cite{Borexino:2017fbd}. There is an LEP mono-photon search limit \cite{Berezhiani:2001rs} as well; however, we find that it is comparatively less constrained than the above-mentioned scenarios. We also project the future DUNE sensitivity \cite{Chatterjee:2021wac} from NSI searches by the blue dashed line. The brown dashed line represents the future sensitivity from IceCube by looking at Glashow-like resonance features \cite{Babu:2019vff, Babu:2022fje} induced by charged scalars in the ultrahigh-energy neutrino event spectrum. The gray dashed line is the astrophysical constraint originating from red giant and horizontal branch stars $|\mu_\nu| \leq 1.5 \times 10^{-12} \mu_B$ (95\% CL) \cite{Viaux:2013lha, Viaux:2013hca, Capozzi:2020cbu}. However, it has been recently pointed out that this astrophysical limit can be relaxed by considering ``neutrino trapping mechanism" \cite{Babu:2020ivd, Babu:2021jnu}. The red dashed lines indicate the future sensitivities on neutrino magnetic moments at DUNE and HK experiments based on our analysis and setting magnetic field strength $B_0 =10^{10}\,$G. We can see that a large region of the $SU(2)_H$ model parameter space can be explored by looking at imprints of neutrino magnetic moments on supernova neutrino signal at DUNE and HK.
\section{Conclusions }\label{SEC-07}
The neutronization burst from a future galactic SN has the potential to provide a wealth of information on neutrino transition magnetic moments (TMM). The neutrino spectra during this phase is usually dominated by a large fraction of $\nu_e$ with sub-dominant $\bar{\nu}_e$ and $\nu_{\mu,\tau}$. However, the presence of a non-zero neutrino TMM can change this picture completely. The combination of spin-flavour conversions due to a TMM, and resonant flavour conversions due to mass-mixings, can lead to a suppression of the $\nu_e$, while simultaneously enhancing the $\bar{\nu}_e$ spectra in this epoch. This tell-tale signature can be used to put strong bounds on neutrino TMMs using the spectra from a future galactic SN. With immense experimental effort underway in the detection of neutrinos from a future galactic SN, it is timely to analyze the impact of neutrino TMM on the signal.
In this work, we have studied the neutrino flavour evolution inside a SN in the presence of a finite TMM, and considered the effect it has on the spectra from the burst phase. By simulating the event spectra in upcoming neutrino experiments like as DUNE and Hyper-Kamiokande, we have found that the neutrino TMMs that may be probed by these experiments for a SN happening at 10$\,$kpc are two or three orders of magnitude (from $\mathcal{O} [10^{-11}]~ \mu_B$ to $\mathcal{O} [10^{-14}]~ \mu_B$ for the allowed range of magnetic field strengths) better than the current terrestrial and astrophysical bounds. We have discussed the uncertainties present in such an analysis, and the kind of impact it can have on our results. Furthermore, we have analyzed how this realization can shed light on three essential neutrino properties: (a) the Dirac/Majorana character of the neutrino, (b) mass ordering, and (c) the neutrino mass generation mechanisms. Finally, for completeness of our study, we have considered a model based on a $SU(2)_H$ horizontal symmetry, which generates large neutrino TMMs, and extrapolated our bounds onto the model parameter space. We have found that in such models, a large section of the parameter space can be explored using the neutrino signal from a future galactic SN.
\section*{Acknowledgments}
We thank Evgeny Akhmedov, Manfred Lindner, Orlando Peres and Suprabh Prakash for useful discussions. MS would like to thank Ivan Martinez-Soler for help with MARLEY simulations. YPPS thanks the Max-Planck-Institut f{\"u}r Kernphysik Particle and Astroparticle Physics division for warm hospitality during the completion of this work. The work of YPPS is supported in part by the FAPESP funding Grants No. 2014/19164-453 6, No. 2017/05515-0 and No. 2019/22961-9.
{\footnotesize
\bibliographystyle{utphys}
| 6b8ef7b4552a849f67e2725997f6fcff660cde7a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Throughout this paper all rings are commutative and $(R,\mathfrak{m},k)$ is a Noetherian local ring. The goal of this paper is to prove a rigidity theorem for Ext.
\begin{definition} Let $M$ be a nonzero finitely generated $R$ module. We say that $M$ is \emph{Ext rigid} if, for all finitely generated $R$ modules $N$, $\ext_{R}^{n}(M,N)=0$ for some $n\leq\grade{M}$ implies $\ext_{R}^{i}(M,N)=0$ for $i\leq n$. (Recall $\grade{M}=\inf\{i|\ext_{R}^{i}(M,R)\neq 0\}$.)
\end{definition}
\begin{theorem*} (Theorem \ref{maintheorem} below) Let $R$ be a regular local ring or an unramified hypersurface and let $M$ be a finitely generated $R$ module. Then $M$ is Ext rigid.
\end{theorem*}
Recall that $R$ is an \emph{unramified} local ring if $R$ is equicharacteristic or is mixed characteristic and char($k$)$\not\in\mathfrak{m}^{2}$. Also, $R$ is a \emph{hypersurface} if it is not regular and the $\mathfrak{m}$-adic completion $\hat{R}$ of $R$ is the quotient of a regular local ring by a nonzerodivisor: $\hat{R}\cong Q/x$ where $(Q,\mathfrak{n},k)$ is a regular local ring and $x\in\mathfrak{n}^{2}$ is a nonzerodivisor. We give the following as a corollary to the main theorem:
\begin{corollary*} (Corollary \ref{maincorollary} below) Let $R$ be an unramified hypersurface and let $M$ be a nonzero finitely generated $R$ module. Then $\ext_{R}^{i}(M,M)\neq 0$ for $0\leq i\leq\grade{M}$.
\end{corollary*}
For a Noetherian local ring $R$ and a finitely generated CM module $M$ of finite projective dimension, a result of Ischebek \cite[Theorem 17.1]{matsumura} gives us $\ext_{R}^{i}(M,R)=0$ for
$$i<\depth{R}-\dim{M}=\depth{R}-\depth{M}=\pdim_{R}M.$$
Since $\ext_{R}^{\pdim_{R}M}(M,R)\neq 0$, we have $\grade{M}=\pdim_{R}M$, so this corollary answers a question of Jorgensen \cite[Question 2.7]{jorgensen} when the module is CM:
\begin{question}[Jorgensen] \label{jorgensenquestion} Let $M$ be a nonzero finitely generated module of finite projective dimension over a complete intersection $R$ of positive codimension. Does $\ext_{R}^{n}(M,M)=0$ for some $n\geq 1$ imply that $\pdim_{R}M<n$?
\end{question}
A result of Jothilingam says that for a finitely generated module over a regular local ring $R$, $\ext_{R}^{n}(M,M)=0$ implies that $\pdim_{R}M<n$ (this can be found in \cite{jothilingam}). In \cite{jorgensen} Jorgensen gives an extension of this result and in \cite[Proposition 5.4]{dao2} Dao shows:
\begin{proposition}[Dao] Let $R$ be an unramified hypersurface and let $M$ be a finitely generated $R$ module. Then the answer to Jorgensen's Question \ref{jorgensenquestion} is ``yes" provided $[M]=0$ in $\overline{G}(R)_{\mathbb{Q}}$.
\end{proposition}
Recall that $G(R)$ is the Grothendieck group of finitely generated modules over $R$ and $\overline{G}(R)_{\mathbb{Q}}=\overline{G}(R)\otimes_{\mathbb{Z}}\mathbb{Q}$ where $\overline{G}(R)=G(R)/\mathbb{Z}\cdot[R]$ is the reduced Grothendieck group. Note that in our corollary we do not need the module to be zero in $\overline{G}(R)_{\mathbb{Q}}$.\newline
Our main theorem is motivated by the following result of Dao \cite[Proposition 2.8]{dao}:
\begin{theorem}[Dao] \label{daointro} Let $R$ be an unramified hypersurface and let $M$ and $N$ be finitely generated $R$ modules so that $\length(\tor_{i}^{R}(M,N))<\infty$ for $i\gg 0$. If $\theta^{R}(M,N)=0$, then $(M,N)$ is Tor rigid.
\end{theorem}
For a finitely generated module $M$, we say that $M$ is \emph{Tor rigid} if, for all finitely generated $R$ modules $N$, $\tor_{n}^{R}(M,N)=0$ for some $n\geq 0$ implies $\tor_{i}^{R}(M,N)=0$ for all $i\geq n$. For a hypersurface $R$ and finitely generated $R$ modules $M$ and $N$ with $\length(\tor_{i}^{R}(M,N))<\infty$ for $i\gg 0$, define $\theta^{R}(M,N)$ to be
$$\theta^{R}(M,N)=\length(\tor_{2j}^{R}(M,N))-\length(\tor_{2j+1}^{R}(M,N))$$
for $j\gg 0$. Results of Auslander \cite{auslander} and of Lichtenbaum \cite{lichtenbaum} show that every module over a regular local ring is Tor rigid. Auslander originally used this fact to study torsion free modules. In \cite{hochster} Hochster defined the pairing $\theta^{R}$. Note that unlike Dao's theorem, the $\theta^{R}$ pairing is not used in our main theorem and we do not need $\length(\ext_{R}^{i}(M,N))<\infty$ for $i\gg 0$.\newline
I would like to thank my advisor Mark Walker for his continuous support and patience. His comments enormously contributed to the flow of the proofs and the paper. I would also like to thank David Jorgensen for his comments on an earlier draft of this paper.
\section{The Main Theorem and its Corollary}
In this section we prove the main theorem and its corollary. We prove the main theorem in two steps. We first show that it is true when $\ext_{R}^{i}(M,N)$ has finite length for all $i$ and then prove the general case by reducing to the finite length case. We start by giving a definition and a lemma:
\begin{definition} Let $M$ be a finitely generated $R$ module and let $g=\grade{M}$. If
$$\dots\xrightarrow{\partial_{2}}F_{1}\xrightarrow{\partial_{1}}F_{0}\xrightarrow{\partial_{0}}0$$
is a minimal free resolution of $M$ over $R$, define $E_{R}(M)$ to be
$$E_{R}(M)=\text{coker}(F_{g-1}^{*}\xrightarrow{\partial_{g}^{*}}F_{g}^{*}),$$
where $F_{i}^{*}=\Hom_{R}(F_{i},R)$.
\end{definition}
\begin{lemma} \label{exttotor} Let $M$ be a finitely generated $R$ module and let $g=\grade{M}$. If $N$ is an $R$ module, then
$$\ext_{R}^{i}(M,N)\cong\tor_{g-i}^{R}(E_{R}(M),N)$$
for $0\leq i\leq g-1$.
\end{lemma}
\begin{proof}
Let
$$F_{\bullet}=(\dots\rightarrow F_{1}\rightarrow F_{0}\rightarrow 0)$$
be a minimal free resolution of $M$ over $R$ and consider
$$F_{\leq g}=(0\rightarrow F_{g}\rightarrow\dots\rightarrow F_{0}\rightarrow 0).$$
Since $g=\grade{M}$, we have $\ext_{R}^{i}(M,R)=0$ for $0\leq i\leq g-1$, so
$$F_{\leq g}^{*}=(0\rightarrow F_{0}^{*}\rightarrow\dots\rightarrow F_{d}^{*}\rightarrow 0)$$
is a minimal free resolution of $E_{R}(M)$ over $R$. The result follows from the fact that $\Hom_{R}(F_{\leq g},N)$ and $F_{\leq g}^{*}\otimes_{R}N$ are isomorphic.
\end{proof}
Recall the following definition:
\begin{definition} Fix $j$. Let $Q$ be a ring and let $M$ and $N$ be $Q$ modules with $\length(\tor_{i}^{Q}(M,N))<\infty$ for $i\geq j$. If $\tor_{i}^{Q}(M,N)=0$ for $i\gg 0$, then define $\chi_{j}^{Q}(M,N)$ to be:
$$\chi_{j}^{Q}(M,N)=\sum_{i\geq j}(-1)^{i-j}\length(\tor_{i}^{Q}(M,N)).$$
\end{definition}
The following theorem was proven by Hochster and Lichtenbaum. It was proved by Lichtenbaum in \cite{lichtenbaum} in most cases. The rest of the cases were proved by Hochster in \cite{hochster2}.
\begin{theorem}[Hochster and Lichtenbaum] \label{lichtenbaumhochster} Let $(Q,\mathfrak{n},k)$ be an unramified regular local ring and let $M$ and $N$ be finitely generated $Q$ modules. Let $j\geq0$ be an integer and assume that $\length(\tor_{i}^{Q}(M,N))<\infty$ for $i\geq j$. Then the following hold:\newline
1. $\chi_{j}^{Q}(M,N)\geq0$.\newline
2. If $j\geq 1$, then $\chi_{j}^{Q}(M,N)=0$ if and only if $\tor_{i}^{Q}(M,N)=0$ for all $i\geq j$.
\end{theorem}
We make the following definition and prove an analog of the above theorem for Ext.
\begin{definition} Let $Q$ be a ring and let $M$ and $N$ be $Q$ modules. Let $j\geq 0$ be an integer and assume that $\length(\ext_{Q}^{i}(M,N))<\infty$ for $i\leq j$. Define $\overline{\xi}_{j}^{Q}(M,N)$ to be
$$\overline{\xi}_{j}^{Q}(M,N)=\sum_{i=0}^{j}(-1)^{-i}\length(\ext_{Q}^{j-i}(M,N)).$$
\end{definition}
\begin{lemma} \label{analogofLH} Let $(Q,\mathfrak{n},k)$ be an unramified regular local ring and let $M$ and $N$ be finitely generated $Q$ modules. Let $i$ be an integer so that $1\leq i\leq\grade{M}-1$ and assume that $\length(\ext_{Q}^{j}(M,N))<\infty$ for $j\leq i$. Then the following hold:\newline
1. $\overline{\xi}_{i}^{Q}(M,N)\geq 0$.\newline
2. $\overline{\xi}_{i}^{Q}(M,N)=0$ if and only if $\ext_{Q}^{j}(M,N)=0$ for $j\leq i$.
\end{lemma}
\begin{proof}
Let $g=\grade{M}$. Then $\ext_{Q}^{j}(M,N)=\tor_{g-j}^{Q}(E_{Q}(M),N)$ for $0\leq j\leq g-1$ by Lemma \ref{exttotor}, so by Theorem \ref{lichtenbaumhochster} due to Hochster and Lichtenbaum we have
$$\overline{\xi}_{i}^{Q}(M,N)=\chi_{g-i}^{Q}(E_{Q}(M),N)\geq 0.$$
Also, if $\overline{\xi}_{i}^{Q}(M,N)=0$, then $\chi_{g-i}^{Q}(E_{Q}(M),N)=0$. Therefore
$$\ext_{Q}^{j}(M,N)=\tor_{g-j}^{Q}(E_{Q}(M),N)=0$$
for $j\leq i$ again using Theorem \ref{lichtenbaumhochster}.
\end{proof}
\begin{lemma} \label{analogofDAO} Let $R$ be an unramified hypersurface and let $M$ and $N$ be finitely generated $R$ modules. Let $i$ be an integer so that $i\leq\grade{M}$ and assume $\length(\ext_{R}^{j}(M,N))<\infty$ for $j\leq i$. If $\ext_{R}^{i}(M,N)=0$, then $\ext_{R}^{j}(M,N)=0$ for $j\leq i$.
\end{lemma}
\begin{proof}
Note that the completion of an unramified hypersurface is an unramified hypersurface. Also, $\hat{R}$ is a faithfully flat extension of $R$, so by the Cohen structure theorem and \cite[Theorem 21.1]{matsumura} we can assume $R=Q/x$ for some unramified regular local ring $(Q,\mathfrak{n},k)$ and some nonzerodivisor $x\in \mathfrak{n}^{2}$.\newline
We can also assume that $i>0$. Suppose $\ext_{R}^{i}(M,N)=0$ and consider the following long exact sequence:
\begin{tikzcd}[cells={text width={width("$\ext_{R}^{-1}(M,N) \rightarrow\dots$")},align=center},
column sep=2em]
0 \rightarrow \ext_{R}^{0}(M,N) \arrow{r} & \ext_{Q}^{0}(M,N) \arrow{r} \arrow[d, phantom, ""{coordinate, name=Z}] & \ext_{R}^{-1}(M,N) \rightarrow\dots \arrow[dll,rounded corners,to path={ -- ([xshift=2ex]\tikztostart.east)|- (Z) [near end]\tikztonodes-| ([xshift=-2ex]\tikztotarget.west)-- (\tikztotarget)}] \\
\ext_{R}^{i-1}(M,N) \arrow{r} & \ext_{Q}^{i-1}(M,N) \arrow{r} \arrow[d, phantom, ""{coordinate, name=Z}] & \ext_{R}^{i-2}(M,N) \arrow[dll,rounded corners,to path={ -- ([xshift=2ex]\tikztostart.east)|- (Z) [near end]\tikztonodes-| ([xshift=-2ex]\tikztotarget.west)-- (\tikztotarget)}] \\
\ext_{R}^{i}(M,N) \arrow{r} & \ext_{Q}^{i}(M,N) \arrow{r}{\phi} \arrow[d, phantom, ""{coordinate, name=Z}] & \ext_{R}^{i-1}(M,N) \arrow[dll,rounded corners,to path={ -- ([xshift=2ex]\tikztostart.east)|- (Z) [near end]\tikztonodes-| ([xshift=-2ex]\tikztotarget.west)-- (\tikztotarget)}] \\
\text{coker}(\phi) \arrow{r} & 0.
\end{tikzcd}
It follows that
$$\overline{\xi}_{i}^{Q}(M,N)+\length(\text{coker}(\phi))=\length(\ext_{R}^{i}(M,N)).$$
Let $g=\grade_{R}M$. Then Lemma 2 in section 18 of \cite{matsumura} implies $g=\grade_{Q}M-1$, so we have $1\leq i\leq g<\grade_{Q}M$. Therefore by Lemma \ref{analogofLH} we have $\overline{\xi}_{i}^{Q}(M,N)\geq 0$, so
$$0\leq \overline{\xi}_{i}^{Q}(M,N)+\length(\text{coker}(\phi))=\length(\ext_{R}^{i}(M,N))=0.$$
This gives us $\ext_{R}^{j}(M,N)=0$ for $j\leq i$ by Lemma \ref{analogofLH} again.
\end{proof}
We can now prove the main theorem. The proof of this theorem is similar to the proof of Proposition 2.8 in \cite{dao}.
\begin{theorem} \label{maintheorem} Let $R$ be a regular local ring or an unramified hypersurface and let $M$ be a finitely generated $R$ module. Then $M$ is Ext rigid.
\end{theorem}
\begin{proof}
Let $N$ be a finitely generated $R$ module. The theorem follows from a result of Jothilingam in the regular case: when $R$ is a regular local ring, if $\ext_{R}^{n}(M,N)=0$ for some $n\leq\grade{M}$, the Lemma in \cite{jothilingam} tells us that
$$\ext_{R}^{n-1}(M,R)\otimes_{R}N\cong\ext_{R}^{n-1}(M,N).$$
Since $n-1<\grade{M}$, we have $\ext_{R}^{n-1}(M,R)=0$, so $\ext_{R}^{n-1}(M,N)=0$.\newline
Now assume that $R$ is an unramified hypersurface. We can assume $R$ is complete as in the previous proof, and hence by \cite[Lemma 3.4]{auslander} $R_{p}$ is an unramified hypersurface for all primes $p$ of $R$. Also, since
$$\grade{M}=\inf\{i|\ext_{R}^{i}(M,R)\neq 0\},$$
we have $\grade{M}\leq\grade_{R_{p}}{M_{p}}$ for all $p\in\supp{M}$.\newline
We proceed by induction on $\dim{R}$. Since $\grade{M}\leq\depth{R}=\dim{R}$, there is nothing to show when $\dim{R}=0$, so assume $\dim{R}>0$ and suppose $\ext_{R}^{i}(M,N)=0$ for some $i\leq\grade{M}$.\newline
Let $p$ be a nonmaximal prime in $R$. If $M_{p}=0$, then $\ext_{R_{p}}^{j}(M_{p},N_{p})=0$ for all $j$. If $M_{p}\neq 0$, then induction tells us that $\ext_{R_{p}}^{j}(M_{p},N_{p})=0$ for $j\leq i$ because $\ext_{R_{p}}^{i}(M_{p},N_{p})=0$, $i\leq\grade{M}\leq\grade_{R_{p}}{M_{p}}$, and $\dim{R_{p}}<\dim{R}$. This gives us $\ext_{R_{p}}^{j}(M_{p},N_{p})=0$ for $j\leq i$ and all nonmaximal primes $p$, and so $\length(\ext_{R}^{j}(M,N))<\infty$ for $j\leq i$. Since $\ext_{R}^{i}(M,N)=0$, Lemma \ref{analogofDAO} tells us that $\ext_{R}^{j}(M,N)=0$ for $j\leq i$.
\end{proof}
\begin{corollary} \label{maincorollary} Let $R$ be an unramified hypersurface and let $M$ be a nonzero finitely generated $R$ module. Then $\ext_{R}^{i}(M,M)\neq 0$ for $0\leq i\leq\grade{M}$. In particular, the answer to Jorgensen's Question \ref{jorgensenquestion} is ``yes" for such $R$ and all CM $R$ modules $M$.
\end{corollary}
\begin{proof}
Since $\Hom_{R}(M,M)\neq 0$ and since $\pdim_{R}M=\grade{M}$ for CM $R$ modules with finite projective dimension, this follows from Theorem \ref{maintheorem}.
\end{proof}
As noted in the proof of Corollary \ref{maincorollary}, the main theorem gives the following:
\begin{quote}
Assume $R$ is a regular local ring or an unramified hypersurface and let $M$ be a finitely generated CM $R$ module with finite projective dimension. If $N$ is a finitely generated $R$ module and if $\ext_{R}^{n}(M,N)=0$ for some $n\leq\pdim_{R}M$, then $\ext_{R}^{i}(M,N)=0$ for $i\leq n$.
\end{quote}
One might hope that this property would hold after dropping the CM assumption, but the following example shows that this is not the case.
\begin{example} Let $Q=k[[x,y,z]]$ and let $M=Q/(x^{2},xy,xz)$. Then $\pdim_{Q}M=3$, $\ext_{Q}^{2}(M,Q)=0$, and $\ext_{Q}^{1}(M,Q)\neq 0$.
\end{example}
| 089ab1db808c3f06ad20319663238df1b12971fe | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
\add{To date,} WASP \citep[Wide Angle Search for Planets;][]{2007MNRAS.375..951C}, KELT \citep[Kilodegree Extremely Little Telescope;][]{2007PASP..119..923P} and TESS \citep{2015JATIS...1a4003R} surveys \add{have} confirm\add{ed} that there are 17 hot Jupiters around hot stars ($>7000$ K). Despite their small number, their orbital obliquities tend to have a wide range. This tendency indicates that they did not \add{experience} realignment by the tidal torque \citep{2012ApJ...757...18A}. \add{Generally}, hot host stars rotate rapidly, \add{yielding} oblateness larger than those of slowly rotating stars. A misaligned orbit and a fast-rotating star force faster \add{precession}.\add{At present}, there are only two planets whose nodal precessions have been detected: Kepler-13Ab \citep{2012MNRAS.421L.122S} and WASP-33b \citep{2015ApJ...810L..23J}. These planets are hot Jupiters\add{,} revolving in misaligned orbits around rapidly rotating hot stars.
WASP-33b, a hot Jupiter ($R_p=1.5 R_J$) around an A-type ($T_{\mathrm{eff}}=7430 \pm 100$K) and rapidly rotating ($V\sin i_s = 85.6$ km s$^{-1}$) star, is the only planet whose precession has been detected by Doppler tomography based on transit spectral data. This planet was discovered by \citet{2010MNRAS.407..507C}\add{;} its nodal precession has been detected by more than one Doppler tomographic measurement \citep{2015ApJ...810L..23J, 2016MNRAS.455..207I, 2020PASJ...72...19W, 2021A&A...653A.104B}. The latest study on this topic \citep{2021A&A...653A.104B} derived the angle between the stellar spin axis and line of sight $i_s = 90.11 \pm 0.12$ deg as well as the stellar gravitational quadrupole moment $J_2 = (6.73 \pm 0.22) \times 10^{-5}$ from the measurements; however, the value of $i_s$ does not match that of \cite{2016MNRAS.455..207I} within 3 $\sigma$, \add{whereas} the value of $J_2$ is \add{also} not consistent with that of \cite{2020PASJ...72...19W} within 3 $\sigma$. Moreover, \citep{2021A&A...653A.104B} revealed that the change \add{in the transit chord} of WASP-33b is \add{slightly overly} complicated \add{for hereafter estimating} the position of the transit chord. Thus, we performed more additional observations and datasets to \add{obtain} more accurate values of $i_s$ and $J_2$,\add{as well as a} clearer forecasting of the transit chord.
In this paper, we report \add{the nodal precession of} WASP-33b for a longer period of 11 years by considering additional Doppler tomographic measurements obtained through spectral transit and transit photometric observations. Transit photometry is an important method for measuring the impact parameter $b$ and its change, as \add{demonstrated by} \citet{2012MNRAS.421L.122S}\add{, who} applied \add{this} to detect \add{the nodal precession of} Kepler-13Ab using Kepler's photometric datasets. In Section \ref{sec:DT}, we present our spectral datasets for Doppler tomography and the steps for measuring the orbital parameters, \add{i.e.,} $b$ and the projected spin-orbit obliquity $\lambda$, the angle between the stellar spin axis and the planetary orbital axis. We explain how to handle our transit photometric datasets for determining $b$ in Section \ref{sec:PH}. Moreover, we fit the values of these two parameters from the spectral and photometric measurements with the nodal precession model, which is described in Section \ref{FitMod}. We display the behaviour of WASP-33b's nodal precession and the derived parameters of WASP-33b and its host star in Section \ref{result}. In Section \ref{Discuss}, we \add{discuss} the nodal precession \add{results}. Finally, we present our conclusions in Section \ref{Concl}.
\section{Doppler Tomographic Measurement} \label{sec:DT}
We can \add{simultaneously} measure $\lambda$ and $b$ from the transit spectral data via Doppler tomography. When a planet passes in front of the stellar disk, a bump\add{, referred to as} a planetary shadow\add{,} appears in the stellar line profile. The orbital configuration of the planet can be derived from this shadow motion.
\subsection{Observation Datasets} \label{sec:obs}
We used eight spectroscopic datasets \add{for} WASP-33 around planetary transits.
One of them was obtained with the High Dispersion Spectrograph \citep[HDS;][]{10.1093/pasj/54.6.855} at the 8.2 m Subaru telescope on 19 October, 2011 UT. The other two datasets were obtained by the Harlan J. Smith Telescope (HJST) with \add{the} Robert G. Tull Coud\'e Spectrograph \citep[TS23;][]{1995PASP..107..251T} at McDonald Observatory on 12 November 2008 UT and 4 October 2014 UT. \add{These} datasets \add{were also} used in our previous study \cite{2020PASJ...72...19W}. To extract each line profile from each spectrum, we adopted least-squares deconvolution \citep[LSD;][]{1997MNRAS.291..658D}; the observed spectrum \add{was} regarded as a convolution of a line profile and a series of delta functions. Under this method, we obtained a list of \add{the} absorption lines from the Vienna Atomic Line Database \citep[VALD;][]{2000BaltA...9..590K} to create a series of delta functions. \add{We then} derived each line profile and \add{the} error bars with the matrix calculations in \cite{2010A&A...524A...5K}.
\add{We then} have analysed three spectroscopic datasets obtained on 28 September 2016 UT, 12 January 2018 UT\add{,} and 2 January 2019 UT \add{from} the high-resolution HARPS-N spectrograph \citep[][]{2012SPIE.8446E..1VC}\add{,} which is mounted at the Telescopio Nazionale Galileo (TNG). These three datasets have been extracted and published in \cite{2021A&A...653A.104B}, \add{which were} used \add{in} the extracted line profile series \add{in this study}.
We also We then included two aditional spectral datasets. One was \add{an} extracted dataset taken by HJST/TS23 on 11 December 2016 UT. \add{Data from} 2016 also include\add{d} 10 in-transit spectra. The other dataset was obtained from the 188 cm telescope with HIgh Dispersion Echelle Spectrograph \citep[HIDES;][]{1999PYunO....S..77I} at Okayama Astro-Complex (OAC) in Japan on 27 December, 2019 UT. We utilized a wavelength range from $4980$\AA\ to $6220$\AA\, except for the Na D lines and \add{wavelength} regions around bad pixels. We reduced these spectral data by \add{the} subtracting bias and dark features, flat dividing, and performing one-dimensional spectrum and wavelength calibration. We then utilized their continua\add{,} as well as the \add{HDS} process. We selected alp Leo\add{, a} rapidly-rotating star \citep[$V \sin i_s \sim 300$ km s$^{-1}$;][]{2002ApJ...573..359A}\add{,} to erase the Earth's atmospheric absorption lines. \add{We then} shifted the spectra to the barycentric frame using \texttt{PyRAF} \citep{2012ascl.soft07011S}. Finally, we extracted each line profile of each exposure using the LSD.
\begin{table*
\caption{Details \add{on the spectral datasets}}
\label{result_SP}
\centering
\small
\begin{tabular}{l l c c c c c}
\hline
Date (UT) & Instrument & Number of spectra & Exposure time (s)& Resolution &SNR at 5500\AA&Reference\\
\hline
12 Nov 2008& HJST/TS23 & 13 & 900 &60,000& 140$^{\dag}$&\cite{2010MNRAS.407..507C}\\
19 Oct 2011& Subaru/HDS & 35 & 600 (33 spectra), 480 (2 spectra) &110,000& 160$^{\dag}$&\cite{2020PASJ...72...19W}\\
4 Oct 2014& HJST/TS23 & 21 & 900 &60,000& 280$^{\dag}$&\cite{2015ApJ...810L..23J}\\
28 Sep 2016 & TNG/HARPS-N & 40 & 600 &115,000& 110$^{\ddag}$&\add{\cite{2021A&A...653A.104B}} \\
11 Dec 2016 & HJST/TS23 & 21 & 900 &60,000& 250$^{\dag}$&This work \\
12 Jan 2018 & TNG/HARPS-N & 23 & 900 &115,000& 170$^{\ddag}$&\add{\cite{2021A&A...653A.104B}} \\
2 Jan 2019 & TNG/HARPS-N & 33 & 600 &115,000& 120$^{\ddag}$&\add{\cite{2021A&A...653A.104B}} \\
27 Dec 2019& OAC/MuSCAT & 12 & 1200 &65,000& 50$^{\dag}$&This work\\
\hline
\hline
\multicolumn{2}{l}{$^{\dag}$ SNR per pixel}\\
\multicolumn{2}{l}{$^{\ddag}$ SNR per extracted pixel}\\
\end{tabular}
\end{table*}
\subsection{Extracting planetary shadow} \label{PlSh}
\add{We derived a median line profile from all exposure data including in-transit for each epoch. We applied a median line profile, not a mean profile, because it avoided the effects of outliers.}
We subtracted the median line profile from each exposure line profile to calculate the time\add{-}series of the line profile residuals. \add{Both the} planetary shadow due to the transit \add{of WASP-33b} and a striped pattern due to the non-radial pulsations on the surface of WASP-33 \citep{2010MNRAS.407..507C} \add{were present in the line profile residuals}. \citet{2011A&A...526L..10H} showed that the pulsation period was approximately 68 min \add{based on} photometric observation\add{s}. However, determin\add{ing} the period from the Doppler tomographic results \add{was difficult owing to} the irregular patterns.
To extract only the planetary shadow, we applied a Fourier filtering technique \citep{2015ApJ...810L..23J} because the planet is retrograde, \add{while} the pulsations are prograde. First, we performed a two-dimensional Fourier transform. There were components derived from the pulsations in the Fourier space in the first and third quadrants, and \add{the} planetary shadow's components in the second and fourth quadrants. Second, we created a filter in which we set unity in \add{the} two diagonal quadrants, including power from the planetary shadow, and zero in the other quadrants, including power from the pulsation with a Hann function between these quadrants. Finally, we multiplied the Fourier space by the filter and performed an inverse Fourier transform on the filtered Fourier space to extract the planetary shadow. Figure \ref{figbff} \add{illustrates these procedures}.
\begin{figure*
\centering
\includegraphics[width=155mm]{Figure_2/residuals_bff_add.pdf}
\vspace*{-0.8cm}
\caption{Doppler tomographic datasets and Fourier filters. Left column: observed residuals \add{for the} line profile series. The vertical dotted lines show $v=0$ and $\pm v \sin i_{s}$. The bottom, middle, and upper horizontal dotted lines show the beginning, middle, and end of the WASP-33b's transit, respectively. Middle column: Fourier spaces after Fourier transform of the residuals \add{for} the line profile series. These colour scales are shown as square roots. The faint narrow structure from the bottom right to the upper left is a component of WASP-33b's planetary transit. \add{In contrast, t}he bright-wide structure from the bottom left to the upper right is a pulsation component. Right column: filtered Fourier space \add{such} that only the transit component remains.}
\label{figbff}
\end{figure*}
\subsection{Deriving parameters} \label{DB}
To obtain the best-fit values and uncertainties of \add{the} transit parameters, we adopted the Markov chain Monte Carlo (MCMC) method using the code \texttt{EMCEE} \citep{2013PASP..125..306F}.
We modelled a planetary shadow \add{via} a convolution between a rotational broadening profile and a Gaussian line profile owing to intrinsic broadening, thermal broadening, and micro-turbulence. \add{The detailed} equations to derive the model of the planetary shadow are described in the appendix of \cite{2020PASJ...72...19W}\add{.} \add{We then} applied the same filter to the model of the planetary shadow following the procedures described in Section \ref{PlSh}.
We fitted the observed residuals of the five datasets to the models with 21 parameters using MCMC: \add{the} $\lambda$, $b$ and transit mid-time $T_{c}$ of each epoch, $V\sin i_{s}$, $R_{p}/R_{s}$, $a/R_{s}$, two quadratic limb darkening coefficients, and the FWHM of the Gaussian line profile. The limb darkening coefficients \add{were} derived \add{via} the triangular sampling method \add{reported in} \citet{2013MNRAS.435.2152K}, \add{with} $q_{1}$, and $q_{2}$. We estimated that $q_{1}$ and $q_{2}$ of HDS, TS23, and HIDES \add{were} equivalent. They \add{were} calculated from the stellar parameters, \add{i.e.}, the effective temperature $T_{\mathrm{eff}}$, surface gravity log $g$, and metallicity. We set the priors of $\lambda$ and $b$ for all epochs and the FWHM as uniform functions \add{while that of} the priors of the other parameters \add{were set} as Gaussian priors. For \add{the} values and widths of \add{the} Gaussian priors, we set the priors \add{for} $R_{p}/R_{s}$ and $a/R_{s}$ based on the values and uncertainties \add{reported in} \citet{2013AA...553A..44K}\add{;} the priors of each $T_{c}$ of each epoch from $P_{\mathrm{orb}}$ in \citet{2014A&A...561A..48V} and $T_{0}$ in \citet{2019A&A...622A..71V}\add{;} \add{those} of $q_{1}$, and $q_{2}$ calculated \add{via} $\mathrm{PyLDTk}$ \citep{Parviainen2015, Husser2013}\add{;} and \add{those} of $V\sin i_{s}$ from \citet{2015ApJ...810L..23J}.
For the fitting, we maximised the logarithm of the posterior probability\add{,} $\ln L_{\mathrm{post}}$\add{:}
\begin{eqnarray}
\label{eq1}
\ln L_{\mathrm{post}} = -\sum_{i} \frac{(O_{i}-C_{i})^{2}}{\sigma^{2}_{i}} - \sum_{j} \frac{(p_{j}-\mu_{j})^{2}}{s^{2}_{j}},
\end{eqnarray}
where $O_{i}$ is the data, $C_{i}$ is the model, $\sigma_{i}$ is the error for the $i$th data point, $p$ is the parameter value at the gained iteration of the Markov chain, $\mu$ is the value from the literature, and $s$ is the uncertainty from the literature. \add{Index} $j$ denote\add{s} the parameters of the Gaussian priors.
We set the range of the uniform prior of each $\lambda$ to $-180 \mathrm{deg} <\lambda<-90 \mathrm{deg}$ and that of each $b$ as $-1<b<1$.
To converge these parameter values, we ran 4,000 steps, cut off the first 2,000 steps as burn-in, and iterated this set 100 times. \add{Figures \ref{figMCMC} and \ref{figMCMC_2} in Appendix \ref{MCMC_res} plot the} posterior distributions.
\section{Photometric Measurement} \label{sec:PH}
\subsection{Photometric Observations of WASP-33b's Transit}
We observed the transit of WASP-33b by photometry using two instruments \add{with} multicolour simultaneous cameras: Multicolor Simultaneous Camera for studying Atmospheres of Transiting exoplanets \citep[MuSCAT;][]{2015JATIS...1d5001N} on the 188 cm telescope at OAC and MuSCAT2 \citep{2019JATIS...5a5001N} on \add{the} Telescopio Carlos S\'{a}nchez (TCS) 1.52 m telescope at the Teide Observatory (OT). MuSCAT has three channels for \add{the} $g^{'}_{2}$ (400-550 nm), $r^{'}_{2}$ (550-700 nm), and $z_{\mathrm{s},2}$ (820-920 nm) bands. In contract, MuSCAT2 contains four channels for \add{the} $g^{'}_{2}$, $r^{'}_{2}$, $z_{\mathrm{s},2}$, and $i^{'}_{2}$ (700-820 nm) bands. These bands are the Astrodon Photometrics Generation 2 -type Sloan filters. We obtained the dataset \add{from} MuSCAT on 5 November 2017 UT. We set the exposure times of \add{the} $g^{'}_{2}$, $r^{'}_{2}$ and $z_{\mathrm{s},2}$ bands for 4, 4, and 10 s, respectively. We also obtained the dataset \add{from} MuSCAT2 on 11 October 2018 UT. The exposure time of \add{the} $g^{'}_{2}$, $r^{'}_{2}$, $i^{'}_{2}$, and $z_{\mathrm{s},2}$ \add{was} 3, 2, 5, and 12 s, respectively.
To produce the light curves of WASP-33b, aperture photometry was performed using the pipeline proposed by \citet{2011PASJ...63..287F}. This process determines the stellar barycentre in every frame. We used BD+36 488, the second brightest star image in the frame, as a companion star to correct the atmospheric extinction. Next, the pipeline calculated the shift in the stellar position relative to the reference frame. \add{This} approach photometers the target star and a comparison star with a fixed aperture radius. Here\add{,} we set the aperture radii to 36 and 40 pixels for MuSCAT and MuSCAT2, respectively. After the sky background in the torus area centred on the stellar barycentr, the pipeline \add{subtracted} the sky background from WASP-33b's flux and \add{that of the comparison}. Finally, the WASP-33b light curve was obtained by dividing its flux by the comparison star flux.
\subsection{Light Curve Fitting} \label{LCF}
To measure WASP-33b's impact parameter $b$ in 2017 and 2018, we constructed light curve models with a Gaussian process using the Python code \texttt{exoplanet} \citep{exoplanet:exoplanet}. WASP-33b's light curve not only \add{showed} dimming by the transit, but also a short sinusoidal-wave-like feature due to the stellar pulsations. Thus, following a previous study by \citet{2015ApJ...810L..23J}, we applied a Matern 3/2 kernel $\bf{K}_{ker}$, whose element is expressed as \add{follows:}
\begin{eqnarray}
\label{eq2}
k_{i,j}= \alpha^2\left(1+\frac{\sqrt{3}|t_i-t_j|}{l}\right)\exp \left(-\frac{\sqrt{3}|t_i-t_j|}{l}\right)+\sigma_{i} ^{2}\delta_{i,j}
\end{eqnarray}
for the Gaussian process. $i$ and $j$ denote the orders of the photometric observation's data, $t_i$ and $t_j$ are the observation times, $\alpha$ and $l$ are the hyper parameters indicating the amplitude and timescale of the stellar variations, respectively, and $\sigma$ is the error of data point $i$.
We then fitted the light curves on two epochs to the models with the following 30 parameters using MCMC: baseline $B$ for each light curve, $b$ and $T_{c}$ of each epoch, two quadratic limb darkening coefficients $u_1$, $u_2$, and $\alpha$ of each band, $R_{p}/R_{s}$, $P_{\mathrm{orb}}$, $a/R_{s}$, and $l$. We set the priors of $B$ and $b$ for both epochs, and $R_{p}/R_{s}$ as uniform functions. \add{We then set} the priors of $T_{c}$ for each epoch, $u_1$, $u_2$, and $\alpha$ for each band, $P_{\mathrm{orb}}$, $a/R_{s}$, and $l$ \add{as} Gaussian priors. For the values and widths of \add{the} Gaussian priors, we referred to the values and widths of Gaussian priors from \citet{2015ApJ...810L..23J} for $\alpha$ and $l$ and \citet{2014A&A...561A..48V} for $P_{\mathrm{orb}}$; the others \add{were obtained} in the same manner as the spectral analysis. We set the logarithm of the likelihood $\ln P_{\mathrm{like}}$ as
\begin{eqnarray}
\label{LH_GP}
\ln P_{\mathrm{like}} = -\frac{1}{2}(\ln |\bf{K}_{ker}|+\bf{r} ^{T}\bf{K}_{ker} ^{-1}\bf{r})
\end{eqnarray}
because we adopted Gaussian process for this fitting \citep{10.5555/1162254}.
\add{Here,} $\bf{r}$ is a series of residuals obtained by subtracting the model data from the observation data. In the MCMC process \add{in} \texttt{PyMC} \citep{exoplanet:pymc3}, we ran 1,000 steps, cut off the first 500 steps as burn-in, and iterated this set 20 times. \add{Figures \ref{figMCMC_Ph} and \ref{figMCMC_Ph_2} in Appendix \ref{MCMC_res} plot the} posterior distributions.
\section{Fitting with Nodal Precession Model} \label{FitMod}
The angular momentum of WASP-33b's planetary orbit $|\overrightarrow{L_{p}}|$ (= $2 \pi M_{p} a^{2}/P_{\mathrm{orb}}$) is \add{significantly} smaller than the stellar rotational angular momentum of its host star $|\overrightarrow{L_{s}}|$; $|\overrightarrow{L_{p}}|/|\overrightarrow{L_{s}}|$ is $\sim 0.05$ using \add{the values} of $|\overrightarrow{L_{s}}|$ from \citet{2011Ap&SS.331..485I}, \add{those} of $M_{p}$ from \citet{2015AA...578L...4L}, $a$, and $P$ from \citet{2010MNRAS.407..507C}. In this case, we can regard the stellar rotational axis as a stable vector and calculate the changes in $b$ and $\lambda$
\begin{eqnarray}
b(t)&=&\frac{a}{R_s} \left(\cos \psi \cos i_s + \sin \psi \sin i_s \cos \theta(t) \right)
\label{impact_pro} \\
\tan \lambda(t)&=&\frac{\sin \psi \sin \theta(t)}{\sin \psi \cos i_s \cos \theta(t)-\cos \psi \sin i_s}.
\label{lambda_pro}
\end{eqnarray}
Here $\theta$, the nodal angle, can be expressed as \add{follows:}
\begin{eqnarray}
\theta (t)&=&-\frac{3 \pi J_{2} R_s^{2} \cos \psi}{P_{\mathrm{orb}}a^2}t +\theta_0,
\label{theta_pro}
\end{eqnarray}
where the slope of Equation \ref{theta_pro} is the precession speed from \citet{2013ApJ...774...53B}.
\add{We then} fit\add{ted} the model \add{in} Equations \ref{impact_pro} and \ref{lambda_pro} with the values measured by \add{the} MCMC using \texttt{PyMC}. We considered $\psi$, $\theta (t=2008)$, $i_s$, and $J_{2}$ as free parameters and set their priors as uniform functions. Here, we \add{set} $\theta$ in 2008 as $\theta_0$, \add{i.e.,} the initial value of $\theta$ in Equation \ref{theta_pro}. For this fitting, we set the logarithm of the likelihood $\ln P_{\mathrm{like}}$ as:
\begin{eqnarray}
\ln L_{\mathrm{like,pre}}=-\frac{1}{2}\sum_{i} \frac{(O_{\mathrm{mes}, i}-C_{\mathrm{mod},i})^{2}}{\sigma_{\mathrm{mes},i}^{2}}
\label{chi_pre}
\end{eqnarray}
where $O_{\mathrm{mes}, i}$ is the measured value of $\lambda$ and $b$ of each epoch, $C_{\mathrm{mod},i}$ is the model value of $\lambda$ and $b$, and $\sigma_{\mathrm{mes},i}$ is the uncertainty of the measured $\lambda$ and $b$, respectively. We consider that the values $b$ in 2017 and 2018 were zero, as shown in Figure \ref{figMCMC_Ph}.
We ran 20,000 steps, cut off the first 10,000 steps as burn-in, and iterated this set 20 times.
The posteriors from the MCMC are shown in Figure \ref{result_system}, and the values are listed in Table \ref{result_2}. \add{Figure \ref{change}} exhibit\add{s} the changes in $\lambda$ and $b$ \add{for} WASP-33b.
\section{Results} \label{result}
\add{Figure \ref{resulr_ph} shows the} line profile residuals and the best-fitted filtered models. \add{Table \ref{result_1} lists the} best values of $\lambda$ and $b$. Our results for $\lambda$ and $b$ in 2014 \add{were} in excellent agreement with the values of \citet{2015ApJ...810L..23J} within 1$\sigma$, whereas those in 2008 were marginally consistent with \citet{2015ApJ...810L..23J} within $2\sigma$. Moreover, our results for $\lambda$ in 2016 and 2018 from HARPS-N \add{were} consistent with those of \cite{2021A&A...653A.104B}, whereas our $\lambda$ in 2019 from HARPS-N differs from that in \cite{2021A&A...653A.104B} by $\sim 2\sigma$.
\begin{figure*
\centering
\includegraphics[width=155mm]{Figure_2/residuals_MCMC_add.pdf}
\vspace*{-0.8cm}
\caption{Fitting \add{of the} filtered residual data by MCMC. \add{Same} type of colour scale as \add{that in} the left column \add{of} Figure \ref{figbff}. Left column: \add{r}esidual data remain\add{ing} as a planetary shadow. Middle column: \add{f}iltered models of a planetary shadow using \add{the} best-fit values. Right column: the difference between the first row and the second rows.}
\label{resulr_ph}
\end{figure*}
\add{Figures \ref{resulr_LC_1} and \ref{resulr_LC_2} show the} best-fit light curve models. \add{Table \ref{result_1} lists the} ranges of $b$ from MuSCAT and MuSCAT2. The posteriors of both $b$ in Figure \ref{figMCMC_Ph} exhibit a truncated normal distribution with a minimum value of $\sim$ 0. Hence, we set 1$\sigma$ as a 68\% confidence interval \add{based on} the minimum value \add{of} 0 because \add{they were identical,} but \add{with} opposite signs of impact parameters\add{, which yielded} the same transit light curves.
\begin{figure*
\centering
\includegraphics[width=1.\textwidth]{Figure_2/MuSCAT1_lc.pdf}
\caption{Light curves of MuSCAT in 2017 with the original cadences (gray points) and 1-minute bins (coloured points). Top row: light curves from \add{the} photometric observation data. The black solid lines represent the models from the MCMC fitting. Middle row: light curves subtracted using \add{the} Gaussian process. Bottom row: residuals between the observed data and model data.}
\label{resulr_LC_1}
\end{figure*}
\begin{figure*
\centering
\includegraphics[width=1.\textwidth]{Figure_2/MuSCAT2_lc.pdf}
\caption{Similar light curves as Figure \ref{resulr_LC_1}, but for \add{the} MuSCAT2 dataset.}
\label{resulr_LC_2}
\end{figure*}
\renewcommand{\arraystretch}{1.1}
\begin{table*
\caption{Measured \add{p}arameters \add{for} WASP-33b}
\label{result_1}
\centering
\small
\begin{tabular}{l c c c c c}
\hline
Date (UT)& $\lambda$ (deg)& $b$& $T_c (\mathrm{BJD_{TDB}})$&Method & Instrument\\
\hline
12 Nov 2008 & $-111.30^{+0.76}_{-0.77}$ & $0.2398^{+0.0062}_{-0.0058}$ &$2454782.92502\pm 0.00016$& Doppler tomography & HJST/TS23 \rule[-1mm]{0mm}{5mm}\\
19 Oct 2011& $-113.96\pm 0.30$ & $0.1578\pm 0.0027$&$2455853.96863^{+0.00014}_{-0.00015}$& Doppler tomography& Subaru/HDS \rule[-1mm]{0mm}{5mm}\\
4 Oct 2014& $-113.00\pm 0.37$ & $0.0845\pm+0.0031$ &$2456934.77139\pm0.00015$& Doppler tomography & HJST/TS23 \rule[-1mm]{0mm}{5mm}\\
28 Sep 2016& $-111.39\pm0.23$ & $0.0413\pm 0.0019$ &$2457660.59250\pm0.00014$& Doppler tomography & TNG/HARPS-N \rule[-1mm]{0mm}{5mm}\\
11 Dec 2016& $-111.32^{+0.49}_{-0.47}$ & $0.0432\pm 0.0039$ &$2457733.78452\pm0.00015$& Doppler tomography & HJST/TS23 \rule[-1mm]{0mm}{5mm}\\
5 Nov 2017& - & $|b|<0.132$ &$2458063.14903\pm0.00015$& Photometry &OAC/MuSCAT\rule[-1mm]{0mm}{5mm}\\
11 Oct 2018& - & $|b|<0.067$ &$2458403.49234\pm0.00014$& Photometry &TCS/MuSCAT2\rule[-1mm]{0mm}{5mm}\\
12 Jan 2018& $-111.46\pm 0.28$ & $0.0034^{+0.0024}_{-0.0023}$ &$2458131.46135\pm0.00015$& Doppler tomography & TNG/HARPS-N \rule[-1mm]{0mm}{5mm}\\
2 Nov 2019& $-111.64\pm 0.28$ & $-0.0272^{+0.0020}_{-0.0021}$ &$2458486.44296\pm0.00016$& Doppler tomography & TNG/HARPS-N \rule[-1mm]{0mm}{5mm}\\
27 Dec 2019 & $-112.24^{+0.97}_{-1.02}$ & $-0.0592^{+0.0066}_{-0.0065}$ &$2458845.08379^{+0.00015}_{-0.00016}$& Doppler tomography & OAC/HIDES \rule[-1mm]{0mm}{5mm}\\
\hline
\hline
\end{tabular}
\end{table*}
Moreover, \add{Figure \ref{result_system} presents} posteriors from MCMC with the nodal precession model\add{.} \add{Table \ref{result_2}lists} the values of $\psi$, $\theta (t=2008)$, $i_s$, and $J_{2}$ . Here, we note that \citet{2016MNRAS.455..207I} may have consider\add{ed} $\psi$ as a variable value. \add{Figure \ref{change} shows the} change in $\lambda$ and $b$ \add{for} WASP-33b.
\begin{figure
\centering
\includegraphics[width=85mm]{Figure_2/sys_param_allN_add.pdf}
\caption{MCMC corner plots for \add{the} architecture angles, $\psi$, $\theta_{\mathrm{2008}}$ and $i_s$, and stellar quadrupole moment $J_2$ of \add{the} WASP-33b System}
\label{result_system}
\end{figure}
\begin{figure*
\centering
\begin{minipage}{0.45\hsize}
\includegraphics[width=85mm]{Figure_2/Spin-orbit_obliquity_line_add.pdf}
\end{minipage}
\begin{minipage}{0.45\hsize}
\includegraphics[width=85mm]{Figure_2/Spin-orbit_obliquity_line_long.pdf}
\end{minipage}\\
\begin{minipage}{0.45\hsize}
\includegraphics[width=85mm]{Figure_2/impact_line_add.pdf}
\end{minipage}
\begin{minipage}{0.45\hsize}
\includegraphics[width=85mm]{Figure_2/b_line_long_add.pdf}
\end{minipage}
\caption{Changes \add{in} $\lambda$ (upper row) and $b$ (lower row). The left and right columns show the the short\add{-} and long\add{-}term \add{changes}, respectively. The x-axis of the right column is the time in years from the epoch of 2008. The blue circles show values from the HJST/TS23 \add{datasets}, the green triangles are values from Subaru/HDS, the red squares are values from OAC/HIDES, the cyan line is the value range from OAC/MuSCAT, the magenta line is \add{the} value range from TCS/MuSCAT2, and the black solid lines represent the model. In the bottom-right figure, \add{the} two black\add{-}dashed lines show the edges of the stellar disk of WASP-33b.}
\label{change}
\end{figure*}
\begin{table*
\caption{Calculated \add{p}arameters \add{for} WASP-33b}
\label{result_2}
\centering
\small
\begin{tabular}{l c c c c}
\hline
Date& $\psi$ (deg)&$J_2$& $i_s$ (deg)& $\theta_{2008}$ (deg)\\
\hline
Results of this study&$108.19^{+0.95}_{-0.97}$ &$(1.36^{+0.15}_{-0.12})\times 10^{-4}$&$58.3^{+4.6}_{-4.2}$&$73.6^{+1.7}_{-1.5}$ \rule[-2mm]{0mm}{5mm}\\
\citet{2016MNRAS.455..207I} &
$99^{+5}_{-4}$ (in 2008), $103^{+5}_{-4}$ (in 2014)
& $(2.1^{+0.8}_{-0.5})\times 10^{-4}$ &$142^{+10}_{-11}$ & - \rule[-2mm]{0mm}{5mm}\\
\citet{2020PASJ...72...19W}& - &$(9.14\pm 0.51)\times 10^{-5}$& $96^{+10}_{-14}$ & - \rule[-2mm]{0mm}{5mm}\\
\add{\citet{2021A&A...653A.104B}} & $113.99\pm 0.22$ &$(6.73\pm 0.22)\times 10^{-5}$& $90.11\pm 0.12$ & - \rule[-2mm]{0mm}{5mm}\\
\citet{2021arXiv210903250D}& $108.3^{+19.0}_{-15.4}$ & - & $69.8^{+4.0}_{-3.2}$ & - \rule[-2mm]{0mm}{5mm}\\
\hline
\hline
\end{tabular}
\end{table*}
\section{Discussion} \label{Discuss}
We have inspected the nodal precession of WASP-33b with more observations than \add{that used} in previous studies \citep{2015ApJ...810L..23J, 2020PASJ...72...19W, 2021A&A...653A.104B}. This is the first study to verify the nodal precession \add{based on} both Doppler tomographic observatioadd{s} and transit photometry. The errors from the transit photometric observations \add{were} large. However, with the change in the impact parameter (see the left bottom part \add{in} Figure \ref{change}), the results from the transit photometry \add{were} consistent with the predicted values from the decreasing trend \add{for} the Doppler tomographic observations. This indicates that transit photometry can \add{simultaneously} contribute to the measurement\add{s} of the nodal precession using Doppler tomographic data.
In Figure \ref{change}, although the impact parameter of WASP-33b \add{appeared} to change linearly, the change in $\lambda$ may not be along the model of the nodal precession. Thus, \add{we should} observe WASP-33b \add{via} Doppler tomography to clarify whether its $\lambda$ increase\add{s} \add{based on} the model or decrease\add{s} from 2021.
\subsection{Comparison of Stellar Spin Inclination and Quadrupole Moment with Previous Studies}
Our $i_s$ value disagrees with those in previous studies \citep{2015ApJ...810L..23J, 2016MNRAS.455..207I, 2020PASJ...72...19W, 2021A&A...653A.104B}\add{, as} derived from the nodal precession by $\sim$3 $\sigma$ or more. However, our value is \add{similar} to that of \citet{2021arXiv210903250D}, \add{despite values that} are $\sim$2 $\sigma$ \add{different}. Notably, \cite{2021arXiv210903250D} derived WASP-33b's $i_s$ from light curve TESS photometric data considering its oblateness and gravity darkening.
The derived stellar quadrupole moment \add{of WASP-33b was} $J_2 = (1.36^{+0.15}_{-0.12}) \times 10^{-4}$. This value agrees with that of \citet{2016MNRAS.455..207I} within 1.5 $\sigma$, which is larger than those of the other previous studies (add{$>$} 3 $\sigma$) and smaller than the theoretical value ($J_2 = 3.8\times 10^{-4}$) calculated by \cite{2011Ap&SS.331..485I}.
\add{One of the possible reasons for disagreements with the values of $i_s$ and $J_2$ could be the difference in the nodal precession model. \cite{2016MNRAS.455..207I} and \cite{2020PASJ...72...19W} used time variation models for other orbital parameters: the ascending node $\Omega$ and the orbital inclination to the apparent equatorial plane $I$, illustrated in Figure 4 of \cite{2020PASJ...72...19W}, and calculated from $\lambda$ and $b$. Then, \cite{2021A&A...653A.104B} estimated the change in the inclination angle $i_p (=\arccos{(bR_s/a)})$ as a linear function. In this study, we directly used the accurate time variation models of $\lambda$ and $b$, such that our derived values for $i_s$ and $J_2$ would be more accurate than those in previous studies. However, we should clarify the cause of the short-term variation of $\lambda$ to create more detailed nodal precession model.}
\add{In contrast, there is a probability that adding datasets also causes the disagreements. In this study, we found a short-term variation in $\lambda$, which may have been decreasing since 2016, although the reason remains unclear. Therefore,} we \add{have} to obtain more datasets of WASP-33b's transits to \add{more accurately} determine their values.
\subsection{Orbital Evolution of WASP-33b}
We found the real spin-orbit obliquity of WASP-33b $\psi=108.19^{+0.95}_{-0.97}$ deg. This uncertainty is larger than that reported by \citet{2021A&A...653A.104B} \add{owing to} the difference in the precession model, which shows that $d\lambda / dt$ of WASP-33b is always positive, while that of \citet{2021A&A...653A.104B} allows $d\lambda / dt = 0$ at a certain time. The derived value indicates the possibility that WASP-33b has \add{experienced} \add{the planet-planet scattering \citep{2008ApJ...686..580C} or Kozai migration \citep{2007ApJ...669.1298F}, which are the mechanisms that cause the misaligned orbit.} The existence of the WASP-33b's companion star is necessary \add{for clarify} distinguish\add{ing} these two evolution\add{ary} models. Nevertheless, Doppler tomographic observations and transit photometry cannot detect the companion stars.
\citet{2016ApJ...827....8N} found a companion \add{candidate} \add{for} WASP-33\add{,} estimated \add{as} a dwarf star or brown dwarf \add{and} located at 238 AU ($P_{\mathrm{orb}} \sim 3,300$ yrs) from the host star \add{based on} direct imaging. However, \add{this candidate} has not yet been confirmed \add{owing to negligible} proper motion. Thus, additional direct imaging observation\add{s} \add{are} required to confirm whether the companion moves in the same proper motion as the host star or not. If the stellar companion candidate\add{,} with 0.1 $M_{\star}$\add{,} revolves in a circular orbit ($e=0$) and WASP-33b was formed near the snow line, $\sim 5$ AU ($P_{\mathrm{orb}} \sim 12$ yrs), the planet \add{experienced} a Kozai oscillation with the Kozai cycle period\add{,} $P_{\mathrm{Kozai}}$\add{,} \add{of} 13 Myrs \add{based on the} following \add{expression:}
\begin{equation}
\label{KCP}
P_{\mathrm{Kozai}}=\frac{M_s P_c^2}{M_c P_b}(1-e_c^2)^{3/2},
\end{equation}
where $M_s$, $M_c$, $P_c$, $P_b$, and $e_c$ are the stellar mass, companion's mass, companion's period, planetary mass, and companion's eccentricity, respectively \citep{2007ApJ...670..820W}. Although we have to consider that the distance from the host star to the companion star is the sky-projected distance and the eccentricity of the companion star may exist, the Kozai cycle period \add{could} be shorter than the age of WASP-33 ($\sim$ 100 Myrs). Therefore, the stellar companion may have caused the Kozai mechanism \add{for} WASP-33b.
\subsection{Nodal Precession Speed}
We calculated the nodal precession speed $\dot{\theta}=0.507^{+0.025}_{-0.022}$ deg year$^{-1}$ and its period $P_{\mathrm{pre}}=709^{+33}_{-34}$ years, \add{where} the precession \add{was} faster than that of \citet{2016MNRAS.455..207I} ($\dot{\theta}=0.37^{+0.04}_{-0.03}$). \add{we then found} that WASP-33b transits in front of the host star for only $\sim$ 20 \% of the \add{entire} nodal precession period, which indicates that it is rare to discover WASP-33b as a transiting planet. This implies that WASP-33b began transiting in $1977\pm 2$ and will stop transiting in $2055\pm 2$.
\subsection{Nodal Precession Observations of Other Hot Jupiters around Hot Stars}
Doppler tomography has confirmed 17 hot Jupiters around hot stars \add{to date}. The real spin-orbit obliquities of seven of these planets (WASP-33b, Kepler-13Ab, KELT-9b, KELT-17b, MASCARA-1b, MASCARA-4b, and WASP-189b) have been identified thus far, \add{which are} nearly vertical (60 deg < $\psi$ < 120 deg). \add{The remaining hot Jupiters} reveal \add{that} only their projected spin-orbit obliquities were obtained. Three hot Jupiters around hot stars with $\lambda \sim 90$ deg, \add{i.e.,} KELT-26b \citep[$\lambda=91.3^{+6.5}_{-6.3}$ deg;][]{2020AJ....160..111R}, HAT-P-70b \citep[$\lambda=113.1^{+5.1}_{-3.4}$ deg;][]{2019AJ....158..141Z}\add{,} and TOI-1518b \citep[$\lambda=-119.66^{+0.98}_{-0.93}$ deg;][]{2021AJ....162..218C}, are valuable for detecting their observable nodal precessions to derive their $\psi$. Even if $\lambda$ is near 0 deg or 180 deg, we can observe the change in \add{the} transit trajectory and measure $\psi$ \add{near} 90 deg when the star rotates nearly pole-on for the line of sight. However, when $\lambda$ is near 0 or 180 deg and the star rotation axis is almost perpendicular to the line of sight, the transit trajectory \add{barely} moves because $\psi$ should also be near 0 deg or 180 deg. In this case, we can estimate $\psi$ as $\lambda$. Therefore, we should \add{regularly} observe the nodal precessions of these planets around their hot stars to measure their $\psi$ using Doppler tomography and transit photometry.
Kepler-13Ab, orbiting around an A-type star, is the another hot Jupiter whose nodal precession has been detected by only transit photometries \citep{2011ApJ...736L...4S, 2011ApJS..197...10B}. \add{P}revious studies \add{have} measured $\psi$ of this hot Jupiter \add{via} gravity-darkened transit photometry, but the values were different between these two results ($\psi=60\pm 2$ deg in \cite{2015ApJ...805...28M} and $\psi=29\pm 1$ deg in \cite{2018AJ....155...13H}). \add{Al}though Kepler-13Ab is likely to have evolved with Kozai migration \add{owing} to its companion star, Kepler-13B \citep{2012A&A...544L..12S}, the value of its $\psi$ should be verified by adding Doppler tomographic observations and transit photometries to obtain an accurate histogram of the derived $\psi$. \add{As} \citet{2014ApJ...790...30J} measured \add{the} $\lambda$ \add{of Kepler-13Ab} in 2014 \add{via} Doppler tomography, an additional transit spectroscopic observation enabled us to detect the change in $\lambda$ and then to derive its $\psi$ independently from the gravity-darkened transit photometry.
\cite{2020ApJ...888...63A}, \cite{2020A&A...643A..94L}, \cite{2020AJ....160....4A}, and \cite{2021arXiv210905031H} measured \add{the} $\psi$ of MASCARA-4b ($\psi=104^{+7}_{-13}$ deg), WASP-189b ($\psi=85.4\pm 4.3$ deg), KELT-9b ($\psi=87^{+10}_{-11}$deg), and MASCARA-1b ($\psi=72.1^{+2.5}_{-2.4}$ deg), respectively, using gravity-darkened transit photometry.
\cite{2016AJ....152..136Z} derived the $\psi$ of KLET-17b ($\psi=116 \pm 4$ deg) by the technique using \add{the} differential rotation of a host star.
These planets are hot Jupiters around A-type stars\add{:} their nodal precessions are yet to be detected. Observing the nodal precessions of these hot Jupiters is important \add{for verifying} the values of their $\psi$. This observation can also contribute to the investigation\add{s} \add{on} the internal structure of hot stars by deriving the values of their $J_2$.
Although there are 17 hot Jupiters around hot stars whose projected spin-orbit obliquities have been measured, \add{this} number \add{remains} too small to statistically determine the orbital evolution tendency. \citet{2021ApJ...916L...1A} found that planets around solar-like stars with large projected spin-orbit obliquities are likely to revolve on polar orbits ($\psi \sim 90$ deg); however, the tendency remains unclear \add{for} hot Jupiters around hot stars because the real spin-orbit obliquities of only six of them have been revealed.
If normal planet-disk interaction is the main migration, the distribution should gather at $\psi = 0$ deg \citep{2011MNRAS.412.2790L}.
If the orbital evolution by planet-planet scattering is the majority, the orbits are likely to incline \add{at approximately} $\psi = 60$ deg \citep{2011ApJ...742...72N}.
When Kozai migration is the primary evolution, $\psi$ \add{has} a wide range from 10 deg to 140 deg \citep{2015ApJ...799...27P}.
\add{As} \citet{2018ApJS..239....2B} predicted that the TESS mission could find 500 hot Jupiters around A-type stars from the 2-year observation, \add{increasing} the number of hot Jupiters around hot stars \add{will become possible} by validating TESS planet candidates.
Further observations \add{could} lead to uncovering the \add{detailed} $\psi$ distribution around hot stars.
\section{Conclusion} \label{Concl
We analysed \add{the} nodal precession \add{of WASP-33b via} Doppler tomography and transit photometry using various high-dispersion spectrographs including Subaru/HDS, HJST/TS23, OAC/HIDES, and TNG/HARPS-N, and two multicolour simultaneous cameras, OAC/MuSCAT and TCS/MuSCAT2.
Based on the observed change in the projected spin-orbit obliquity $\lambda$ and the impact parameter $b$, we modelled the nodal precession of WASP-33b and derived the real spin-orbit obliquity of WASP-33b as $\psi = 108.19^{+0.95}_{-0.97}$ deg.
Compared \add{with} the results of previous studies, the results of the near-polar orbit did not change. However, our value \add{for} $\psi$ differ\add{ed} from that of \cite{2021A&A...653A.104B} by $>3 \sigma$. \add{This discrepancy may be caused by whether or not $psi$ is a constant value. We assumed $psi$ as a constant while \cite{2021A&A...653A.104B} did not; \cite{2021A&A...653A.104B} adopted $psi$ from its value at the 2011 epoch.}
We also simultaneously derived the stellar spin inclination and the stellar gravitational quadrupole moment of WASP-33 as $i_s = 58.3^{+4.6}_{-4.2}$ deg, and $J_2=(1.36^{+0.15}_{-0.12}) \times 10^{-4}$, respectively. These results \add{differed} by $> 3 \sigma$ from those of previous studies \add{on} nodal precession, except for \add{the} $J_2$ \add{value reported in} \cite{2016MNRAS.455..207I} ($\sim 2.1\sigma$).
\add{A likely reason for these discrepancies is the different nodal precession models. For the first time, we applied the accurate time variation models for $\lambda$ and $b$ to fit the nodal precession. Therefore, our derived values of $\psi$, $i_s$, and $J_2$ are the most accurate to date. Moreover,} additional datasets may have updated these parameter values\add{, thus causing} these differences. Thus, acquiring more datasets will allow \add{for} the derivation of more accurate values of $\psi$, $i_s$, and $J_2$.
We calculated the nodal precession speed $\dot{\theta}=0.507^{+0.025}_{-0.022}$ deg year$^{-1}$ and its period $P_{\mathrm{pre}}=709^{+33}_{-34}$ years, \add{which revealed} that WASP-33b transits in front of the host star for only $\sim$ 20 \% of the \add{entire} nodal precession period. \add{Based on} this result, we speculate that WASP-33b \add{began} transiting in 1977$\pm$2\add{; we} forecast that it \add{will} finish transiting in 2055$\pm$2.
\add{The} TESS survey \add{should} help us increase \add{our ability to count} number of hot Jupiters around hot stars in the future. Applying the proposed methodology to newly discovered hot Jupiters around hot stars is important not only \add{for characterising} each planetary system\add{,} but also \add{for discriminating} the migration mechanisms of such planets and \add{investgating} the internal structure of hot stars.
\section*{Acknowledgements}
This paper is based on data collected at the Subaru Telescope, which is located atop Maunakea and operated by the National Astronomical Observatory of Japan (NAOJ). We wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Maunakea has always had within the indigenous Hawaiian community.
The paper also includes data taken at The McDonald Observatory of The University of Texas at Austin and taken at The Okayama Astrophysical Observatory.
This article is based on observations made with the MuSCAT2 instrument, developed by ABC, at Telescopio Carlos Sánchez operated on the island of Tenerife by the IAC in the Spanish Observatorio del Teide.
Pyraf is a product of the Space Telescope Science Institute, which is operated by AURA for NASA. This work has made use of the VALD database, operated at Uppsala University, the Institute of Astronomy RAS in Moscow, and the University of Vienna. We are grateful to editage (https://www.editage.jp) for English editing.
We acknowledge the GAPS Consortium (Covino et al. 2013) for providing the mean line profiles of their HARPS-N transits, and thank F. Borsa for providing spectral datasets of HARPS-N.
This work is partly supported by JSPS KAKENHI Grant Numbers JP21K20376, JP18H05439, JP20J21872, JP17H04574, JP20K14518, JST CREST Grant Number JPMJCR1761, Astrobiology Center SATELLITE Research project AB022006, and the Astrobiology Center of National Institutes of Natural Sciences (NINS) (Grant Number AB031010).
N.C.B. acknowledges funding from the European Research Council under the European Union’s Horizon 2020 research and innovation program under grant agreement no. 694513.
E. E-B. acknowledges financial support from the European Union and the State Agency of Investigation of the Spanish Ministry of Science and Innovation (MICINN) under the grant PRE2020-093107 of the Pre-Doc Program for the Training of Doctors (FPI-SO) through FSE funds.
\section*{Data Availability}
The raw data from OAC/HIDES, OAC/MuSCAT and TCS/MuSCAT2 will be shared on reasonable request to the corresponding author.
The reduced data of Subaru/HDS were provided by N. Narita by permission, and will be shared on request to the corresponding author with permission of N. Narita.
The line profile data from HJST/TS23 were provided by M. C. Johnson by permission, and will be shared on request to the corresponding author with permission of M. C. Johnson.
The line profile data of TNG/HARPS-N were provided by F. Borsa by permission, and will be shared on request to the corresponding author with permission of F. Borsa.
\bibliographystyle{mnras}
| a0a5dcc7f57b15e244026a20b48b90e08930aca7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
When considering the dynamics of composite quantum systems, there are many regimes where one system can be taken to be classical and the other quantum-mechanical. For example, in quantum thermodynamics we often have a quantum system interacting with a large thermal reservoir that can be treated classically, whilst in atomic physics it is common to consider the behaviour of quantum atoms in the presence of classical electromagnetic fields. Things become more complicated when one considers classical-quantum (CQ) dynamics where the quantum system back-reacts on the classical system. This is particularly relevant in gravity, because we would like to study the back-reaction of thermal radiation being emitted from black holes on space-time, and while the matter fields can be described by quantum field theory, we only know how to treat space-time classically. Likewise in cosmology, vacuum fluctuations are a quantum effect which we believe seeds galaxy formation, while the expanding space-time they live on can only be treated classically. In addition to the need for an effective theory which treats space-time in the classical limit, there has long been a debate about whether one should quantise gravity \cite{cecile2011role,Feynman:1996kb-note,AharonovParadoxes-note,eppley1977necessity,
unruh1984steps,carlip2008quantum,
Mari:2015qva,Baym3035,belenchia2018quantum,kent2018simple,oppenheim_post-quantum_2018,rydving2021gedanken}.
There has even been much discussion on whether quantum-classical coupling can even be consistent. Many proposals for such dynamics~\cite{aleksandrov199536a,quantum_chemistry} are not completely positive (CP)\footnote{A map $\Lambda$ is completely positive, iff $\mathbbm{1}\otimes\Lambda$ is positive. This is the required condition used to derive the GKSL Equation. If it is violated, the dynamics acting on half of an entangled state, give negative probabilities.}, meaning they are at best an approximation and fail outside a regime of validity \cite{boucher1988semiclassical, diosi2000quantum}. The semi-classical Einstein's equation \cite{moller1962theories,rosenfeld1963quantization}, which replaces the quantum operator corresponding to the stress-energy tensor by its expectation value, is another attempt to treat the classical limit from an effective point of view, but it is non-linear in the state, leading to pathological behavior if quantum fluctuations are of comparable magnitude to the stress-energy tensor \cite{page1981indirect}. This is often the precise regime we would like to understand.
However, dynamics introduced in~\cite{blanchard1995event,diosi1995quantum} and studied in \cite{alicki2003completely, poulinKITP2, oppenheim_post-quantum_2018,unrav} do not suffer from such problems, and lead to consistent dynamics. In particular, the master-equation shown in Equation \eqref{eq: cqdyngen}, is linear, preserves the division of classical degrees of freedom and quantum ones, and is completely positive (CP) and preserves normalisation. This ensures that probabilities of measurement outcomes remain positive and always add to $1$. The dynamics is related to the GKSL or Lindblad equation~\cite{GKS76,Lindblad76}, which for bounded generators of the dynamics, is the most general Markovian dynamics for an open quantum system. Likewise, Equation \eqref{eq: cqdyngen} is the most general Markovian classical-quantum dynamics with bounded generators~\cite{oppenheim_post-quantum_2018}. Sub-classes of this master equation along with meeasurement and feedback approaches have been discussed in the context of Newtonian models of gravity~\cite{diosi2011gravity,Kafri_2014,Kafri_2015,tilloy2016sourcing,Tilloy_2017,tilloy2017principle,poulinKITP2},
and further developed into a spatially covariant framework so that Einstein gravity in the ADM formalism \cite{arnowitt2008republication}
emerges as a limiting case \cite{oppenheim_post-quantum_2018,UCLconstraints}.
In this work, we move away from specific realisations of CQ dynamics, in order to discuss their common features and the experimental signatures that follow from this. An early precursor to the discussion here, is the insight of Di\'{o}si \cite{diosi1995quantum} who added classical noise and quantum decoherence to the master equation of \cite{aleksandrov199536a}, and found the noise and decoherence trade-off required for the dynamics to become completely positive. Here we prove that the phenomena found in~\cite{diosi1995quantum,oppenheim_post-quantum_2018,unrav} are generic features of all CQ dynamics; the classical-quantum interaction necessarily induces decoherence on the quantum system, and there is a generic trade-off between the rate of decoherence and the amount of diffusion in the classical phase space. The stronger the interaction between the quantum system and the classical one, the greater the trade-off. One cannot have quantum systems with long-coherence times without inducing a lot of diffusion in the classical system. One can also generalise this result to a trade-off between the rate of diffusion and the strength of more general couplings to Lindblad operators, with decoherence being a special case. This is expressed as Equations \eqref{eq: decdifnonham} and \eqref{eq: generalTradeOffCouplingConstants}, which bounds the product of diffusion coefficients and Lindblad coupling constants in terms of the strength of the CQ-interaction. It is precisely this trade-off which allows the theories considered here, to evade the no-go arguments of Feynmann \cite{cecile2011role,Feynman:1996kb-note}, Aharonov \cite{AharonovParadoxes-note}, Eppley and Hannah \cite{eppley1977necessity} and others \cite{bohr1933on,cecile2011role,dewitt1962definition,boucher1988semiclassical,diosi2000quantum,gisin1989stochastic,Mari:2015qva,Baym3035,belenchia2018quantum,caro1999impediments, salcedo1996absence, sahoo2004mixing, terno2006inconsistency, barcelo2012hybrid,marletto2017we}. The essence of arguments against quantum-classical interactions is that they would prohibit superpositions of quantum systems which source a classical field. Since different classical fields are perfectly distinguishable in principle, if the classical field is in a distinct state for each quantum state in the superposition, the classical field could always be used to determine the state of the quantum system, causing it to decohere instantly. By satisfying the trade-off, the quantum system preserves coherence because diffusion of the classical degrees of freedom mean that the state of the classical field does not determine the state of the quantum system \cite{poulinKITP2,oppenheim_post-quantum_2018}. Equation \eqref{eq: decdifnonham} and other variants we derive, quantify the amount of diffusion required to preserve any amount of coherence.
If space-time curvature is treated classically, then complete positivity of the dynamics means its interaction with quantum fields \textit{necessarily} results in unpredictability and gravitationally induced decoherence.
This trade-off between the decoherence rate and diffusion provides an experimental signature, not only of models of hybrid Newtonian dynamics such as \cite{diosi2011gravity} or post-quantum theories of General Relativity such as \cite{oppenheim_post-quantum_2018} but of any theory which treats gravity as being fundamentally classical. The metric and their conjugate momenta necessarily diffuse away from what Einstein's General Relativity predicts. This experimental signature squeezes classical-quantum theories of gravity from both sides: if one has shorter decoherence times for superpositions of different mass distributions, one necessarily has more diffusion of the metric and conjugate momenta. In Appendix \ref{sec: grav diffusion} we show that the latter effect causes imprecision in measurements of mass such as those undertaken in the Cavendish experiment \cite{cavendish1798xxi,luther1982redetermination,Gundlach2000} or in measurements of Newton's constant ``Big G"\cite{quinn2000measuring,gillies2014attracting,rothleitner2017invited}.
The precision at which a mass can be measured in a short time, thus provides an upper bound on the amount of gravitational diffusion, as quantified by Equation \eqref{eq: variationForceMain}, while decoherence experiments place a lower bound on the diffusion. Our estimates suggest that experimental lower bounds on the coherence time of large molecules \cite{arndt1999wave,Hornberger,Juffman, Juffmandebroglie, Gerlich2011, BassiCollapse}, combined with gravitational experiments measuring the acceleration of small masses \cite{westphal2020measurement,Schm_le_2016,PhysRevLett.124.101101}, already place strong restrictions on theories where space time isn't quantised.
In Section \ref{sec: gravity} we show that several realisations of CQ-gravity are already ruled out, while other realisations produce enough diffusion away from General Relativity to be detectable by future table-top experiments.
Although the absence of such deviations from General Relativity would not be as direct a confirmation of the quantum nature of gravity, such as experiments proposed in~\cite{kafri2013noise,kafri2015bounds,bose2017spin,PhysRevLett.119.240402,Marshman_2020,pedernales2021enhancing,carney2021testing,kent2021non,christodoulou2022locally} to exhibit entanglement generated by gravitons, it would effectively rule out any sensible theory which treats space-time classically. While confirmation of gravitational diffusion would suggest that space-time is fundamentally classical.
The outline of this paper is as follows. In Sec. \ref{sec: cqdnamics} we review the general form of the CQ master equation of classical-quantum systems. The CQ-map can be represented in a manner akin to the Kraus-representation \cite{Kraus} for quantum maps, with conditions for it to be Completely Positive and Trace Preserving (CPTP). We can perform a short time moment expansion of the {\it CQ-map}
taking states at some initial time, to states at a later time.
This gives us the
CQ version of the Kramers-Moyal expansion~\cite{kramers1940brownian,moyal1949stochastic}. The physical meaning of the moments is discussed in Subsection \ref{ss:momentsmeaning}.
In Sec. \ref{sec: tradeoff} we show that there is a general trade-off between decoherence of the quantum system and diffusion in the classical system. We generalize the trade-off to the case of fields in Section \ref{sec: fieldTradeOff} and in Subsection \ref{sec: gravity}, we apply the inequality in the gravitational setting.
The positivity constraints mean that the considerations do not depend on the specifics of the theory, only that it treats gravity classically, and be Markovian. This allows us to discuss some of the observational implications of this result
and we comment on the relevant figures of merit required in interference and precision mass measurements in order to constrain theories of gravity, as they are not always readily available in published reports. In addition to table-top constraints, we consider those due to cosmological observations.
We then conclude with a discussion of our results in Sec. \ref{sec: conclusion}. The Appendix collects or previews a number of technical results.
\section{Classical-Quantum dynamics} \label{sec: cqdnamics}
Let us first review the general map and master equation governing classical-quantum dynamics. The classical degrees of freedom are described by a differential manifold $\mathcal{M}$ and we shall generically denote elements of the classical space by $z$. For example, we could take the classical degrees of freedom to be position and momenta in which case $\mathcal{M}= \mathbb{R}^{2}$ and $z= (q,p)$. The quantum degrees of freedom are described by a Hilbert space $\mathcal{H}$. Given the Hilbert space, we denote the set of positive semi-definite operators with trace at most unity as $S_{\leq 1}(\mathcal{H})$. Then the CQ object defining the state of the CQ system at a given time is a map $\varrho : \mathcal{M} \to S_{\leq 1}(\mathcal{H})$ subject to a normalization constraint $\int_{\mathcal{M}} dz \Tr{\mathcal{H}}{\varrho} =1$. To put it differently, we associate to each classical degree of freedom a sub-normalized density operator, $\varrho(z)$, such that $\Tr{\mathcal{H}}{\varrho} = p(z) \geq 0$ is a normalized probability distribution over the classical degrees of freedom and $\int_{\mathcal{M}} dz \varrho(z) $ is a normalized density operator on $\mathcal{H}$. An example of such a {\it CQ-state} is the CQ qubit depicted as a $2\times 2$ matrix over phase space \cite{unrav}. More generally, we can define any {\it CQ operator} $f(z)$ which lives in the fibre bundle with base space $\mathcal{M}$ and fibre $\mathcal{H}$.
Just as the Lindblad equation is the most general evolution law which maps density matrices to density matrices, we can ask, what is the most general evolution law, which preserves the quantum-classical state-space. Any such dynamics, if it is to preserve probabilities, must be completely positive, norm preserving, and linear in the CQ-state\footnote{That dynamics must be linear can be seen as follows: if someone prepares a system in one of two states $\sigma_0$ or $\sigma_1$ depending on the value of a coin toss ($\proj{0}$ with probability $p$, $\proj{1}$ with probability $1-p$), then the evolution ${\cal L}$ of the system must satisfy $p\proj{0}\otimes {\cal L}\sigma_0 + (1-p)\proj{1}\otimes{\cal L}\sigma_1 = {\cal L}(p\sigma_0 +(1-p)\sigma_1)$ otherwise the system evolves differently depending on whether we are aware of the value of the coin toss. A violation of linearity further implies that when the system is in state $\sigma_0$ it evolves differently depending on what state the system would have been prepared in, had the coin been $\proj{1}$ instead of $\proj{0}$. This motivates our restriction to linear theories.}. We will also require the map to be Markovian on the combined classical-quantum system, which is equivalent to requiring that there is no hidden system which acts as a memory. This is natural if the interaction is taken to be fundamental, but is the assumption which one might want to remove if one thinks of the hybrid theory as an effective description. We thus take these as the minimal requirements that any fundamental classical-quantum theory must satisfy if it is to be consistent.
The most general CQ-dynamics, which maps CQ states onto themselves can be written in the form \cite{oppenheim_post-quantum_2018}
\begin{equation}\label{eq: CPmap}
\varrho(z,t+ \delta t) = \int dz'\sum_{\mu \nu}\Lambda^{\mu\nu}(z|z',\delta t) L_{\mu} \varrho(z',t) L_{\nu}^{\dag}
\end{equation}
where the $L_{\mu}$ are an orthogonal basis of operators and $\Lambda^{\mu\nu}(z|z',\delta t)$ is positive semi-definite for each $z,z'$. Henceforth, we will adopt the Einstein summation convention so that we can drop $\sum_{\mu\nu}$ with the understanding that equal upper and lower indices are presumed to be summed over.
The normalization of probabilities requires
\begin{equation}\label{eq: prob}
\int dz \Lambda^{\mu\nu}(z|z',\delta t) L_{\nu}^{\dag}L_{\mu} =\mathbb{I}.
\end{equation}
The choice of basis $L_\mu$ is arbitrary, although there may be one which allows for unique trajectories \cite{unrav}.
Equation \eqref{eq: CPmap} can be viewed as a generalisation of the Kraus decomposition theorem.
In the case where the classical degrees of freedom are taken to be discrete, Poulin~\cite{PoulinPC2} used the diagonal form of this map to derive the most general form of Markovian master equation for bounded operators, which is the one introduced in~\cite{blanchard1995event}.
When the classical degrees of freedom are taken to live in a continuous configuration space, we need to be a little more careful, since $\varrho(z)$ may only be defined in a distributional sense; for example, $\varrho(z) = \delta(z, \bar{z}) \varrho( \bar{z})$. In this case \eqref{eq: CPmap} is completely positive if the eigenvalues of $\Lambda^{\mu \nu}(z|z',\delta t)$, $\lambda^{\mu}(z|z',\delta t)$, are positive so that $\int dz dz' P_{\mu}(z,z')\lambda^{\mu}(z|z',\delta t) \geq 0$ for any vector with positive components $P_{\mu}(z,z')$ \cite{UCLPawula}.
One can derive the CQ master equation by performing a short time expansion of \eqref{eq: CPmap} in the case when the $L_\mu$ are bounded \cite{oppenheim_post-quantum_2018}. To do so, we first introduce an arbitrary basis of traceless Lindblad operators on the Hilbert space, $L_{\mu} = \{I, L_{\alpha}\}$. Now, at $\delta t=0$ we know \eqref{eq: CPmap} is the identity map, which tells us that $\Lambda^{00} (z|z', \delta t=0) = \delta(z,z')$. Looking at the short time expansion coefficients, by Taylor expanding in $\delta t \ll 1$, we can write
\begin{align}\label{eq: shortime}
&\Lambda^{\mu \nu}(z|z',\delta t) = \delta^{\mu}_{0} \delta^{ \nu}_{0} \delta(z,z') + W^{\mu \nu}(z|z') \delta t + O(\delta t^2).
\end{align}
By substituting the short time expansion coefficients into \eqref{eq: CPmap} and taking the limit $\delta t \to 0$ we can write the master equation in the form
\begin{align}
\frac{\partial \varrho(z,t)}{\partial t} = \int dz' \ W^{\mu \nu}(z|z') L_{\mu} \varrho(z') L_{\nu}^{\dag} - \frac{1}{2}W^{\mu \nu}(z) \{ L_{\nu}^{\dag} L_{\mu}, \varrho \}_+,
\label{eq: cqdyngen}
\end{align}
where $\{, \}_+$ is the anti-commutator, and preservation of normalisation under the trace and $\int dz$ defines
\begin{equation}
W^{\mu \nu}(z) = \int \mathrm{d} z' W^{\mu \nu}(z'| z).
\end{equation}
We see the CQ master equation is a natural generalisation of the Lindblad equation and classical rate equation in the case of classical-quantum coupling. We give a more precise interpretation of the different terms arising when we perform the Kramers-Moyal expansion of the master equation at the end of the section. The positivity conditions from \eqref{eq: CPmap} transfer to positivity conditions on the master equation via \eqref{eq: shortime}. We can write the positivity conditions in an illuminating form by writing the short time expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z',\delta t)$, as defined by equation \eqref{eq: shortime}, in block form
\begin{equation} \label{eq: block1}
\Lambda^{\mu \nu}(z|z',\delta t)
= \begin{bmatrix}
\delta(z,z') + \delta t W^{00}(z|z') & \delta t W^{0\beta}(z|z') \\
\delta t W^{\alpha 0}(z|z') & \delta t W^{\alpha \beta} (z|z') \\
\end{bmatrix} + O(\delta t^2)
\end{equation}
and the dynamics will be positive if and only if $\Lambda^{\mu \nu}(z|z',\delta t)$ is a positive matrix. It is possible to introduce an arbitrary set of Lindblad operators $\bar{L}_{\mu}$ and appropriately redefine the couplings $W^{\mu \nu}(z|z')$ in \eqref{eq: cqdyngen} \cite{oppenheim_post-quantum_2018}. For most purposes, we shall work with a set of Lindblad operators which includes the identity $L_{\mu} = (I, L_{\alpha})$; this is sufficient since any CQ master equation is completely positive if and only if it can be brought to the form in \eqref{eq: cqdyngen}, where the matrix \eqref{eq: block1} is positive.
\subsection{The CQ Kramers-Moyal expansion}\label{sec: kramersMoyal}
In order to study the positivity conditions it is first useful to perform a moment expansion of the dynamics in a classical-quantum version of the Kramers-Moyal expansion~\cite{oppenheim_post-quantum_2018}. In classical Markovian dynamics, the Kramers-Moyal expansion relates the master equation to the moments of the probability transition amplitude and proves to be useful for a multitude of reasons. Firstly, the moments are related to observable quantities; for example, the first and second moments of the probability transition amplitude characterize the amount of drift and diffusion in the system. This is reviewed in Subsection \ref{ss:momentsmeaning}. Secondly, the positivity conditions on the master equation transfer naturally to positivity conditions on the moments, which we can then relate to observable quantities. In the classical-quantum case, we shall perform a short time moment expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z',\delta t) $ and then show that the master equation can be written in terms of these moments. We then relate the moments to observational quantities, such as the decoherence of the quantum system and the diffusion in the classical system.
We work with the form of the dynamics in \eqref{eq: cqdyngen}, using an arbitrary orthogonal basis of Lindblad operators $L_{\mu} = \{\mathbb{I}, L_{\alpha}\}$. We take the classical degrees of freedom $\mathcal{M}$ to be $d$ dimensional, $z= (z_1, \dots z_d)$, and we label the components as $z_i$, $i \in \{1, \dots d\}$. We begin by introducing the moments of the transition amplitude $W^{\mu \nu}(z|z')$ appearing in the CQ master equation \eqref{eq: shortime}
\begin{equation}
D^{\mu \nu}_{n, i_1 \dots i_n}(z'):= \frac{1}{n!}\int dz W^{\mu \nu}(z|z')(z-z')_{i_1} \dots (z-z')_{i_n} .
\end{equation}
The subscripts $i_j \in \{1, \dots d\}$ label the different components of the vectors $(z-z')$. For example, in the case where $d=2$ and the classical degrees of freedom are position and momenta of a particle, $z=(z_1,z_2) = (q,p)$, then we have $(z-z') = (z_1- z'_1, z_2 - z_2') = ( q-q',p-p')$. The components are then given by $(z-z')_1 = (q-q')$ and $(z-z')_2 = (p-p')$. $M^{\mu \nu}_{n, i_1 \dots i_n} (z',\delta t)$ is seen to be an $n$'th rank tensor with $d^n$ components.
In terms of the components $ D^{\mu \nu}_{n, i_1 \dots i_n}$ the short time expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z')$ is given by \cite{UCLPawula}
\begin{equation}\label{eq: shortimeCP}
\Lambda^{\mu \nu}(z|z',\delta t) = \delta_0^{\mu} \delta_0^{\nu} \delta(z,z') + \delta t \sum_{n=0}^{\infty}D^{\mu \nu}_{n, i_1 \dots i_n}(z') \left(\frac{\partial^n }{\partial z_{i_1}' \dots \partial z_{i_n}' }\right) \delta(z,z') + O(\delta t^2),
\end{equation}
and the master equation takes the form \cite{oppenheim_post-quantum_2018}
\begin{align}\label{eq: expansion}\nonumber
\frac{\partial \varrho(z,t)}{\partial t} &= \sum_{n=1}^{\infty}(-1)^n \left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right) \left( D^{00}_{n, i_1 \dots i_n}(z,\delta t) \varrho(z,t) \right)\\ \nonumber
& -i [H(z),\varrho(z) ] + D_0^{\alpha \beta}(z) L_{\alpha} \varrho(z) L_{\beta}^{\dag} - \frac{1}{2} D_0^{\alpha \beta} \{ L_{\beta}^{\dag} L_{\alpha}, \varrho(z) \}_+ \\
& + \sum_{\mu \nu \neq 00} \sum_{n=1}^{\infty}(-1)^n \left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right)\left( D^{\mu \nu}_{n, i_1 \dots i_n}(z) L_{\mu} \varrho(z,t) L_{\nu}^{\dag} \right),
\end{align}
where we define the Hermitian operator $H(z)= \frac{i}{2}(D^{\mu 0}_0 L_{\mu} - D^{0 \mu}_0 L_{\mu}^{\dag}) $ (which is Hermitian since $D^{\mu 0}_0 = D^{0 \mu *}_0$). We see the first line of \eqref{eq: expansion} describes purely classical dynamics, and is fully described by the moments of the identity component of the dynamics $\Lambda^{00}(z|z')$. The second line describes pure quantum Lindbladian evolution described by
the zeroth moments of the components $\Lambda^{\alpha 0}(z|z'), \Lambda^{\alpha \beta}(z|z')$; specifically the (block) off diagonals, $D^{\alpha 0}_0(z)$, describe the pure Hamiltonian evolution, whilst the
components $D^{\alpha \beta}_0(z)$ describe the dissipative part of the pure quantum evolution. Note that the Hamiltonian and Lindblad couplings can depend on the classical degrees of freedom so the second line describes action of the classical system on the quantum one. The third line contains the non-trivial classical-quantum back-reaction, where changes in the distribution over phase space are induced and can be accompanied by changes in the quantum state.
\subsection{Physical interpretation of the moments}
\label{ss:momentsmeaning}
Let us now briefly review the physical interpretation of the moments which will appear in our trade-off relation. In particular, the zeroth moment determines the rate of decoherence (and Lindbladian coupling more generally), the first moment gives the force exerted by the quantum system on the classical system, and the second moment determines the diffusion of the classical degrees of freedom. For this discussion we shall take the classical degrees of freedom to live in a phase space $\Gamma = (\mathcal{M}, \omega)$, where $\omega$ is the symplectic form.
Consider the expectation value of any CQ operator $O(z)$, $\langle O(z)\rangle:=\int dz \tr{O(z)\varrho}$ which doesn't have an explicit time dependence.
Its evolution law can be determined via Equation \eqref{eq: expansion}
\begin{align}
\frac{d\langle O\rangle}{dt}& =\int dz \tr{ O(z)\frac{\partial \varrho}{\partial t}}\nonumber\\
&=
\int dz \mathrm{Tr}\varrho
\left[ -
i [O(z),H(z)] + D_0^{\alpha \beta}(z) L_{\beta}^{\dag}O(z) L_{\alpha} - \frac{1}{2} D_0^{\alpha \beta} \{ L_{\alpha}L_{\beta}^{\dag}, O(z) \}_+
\right.\nonumber \\
& + \left.
\sum_{n=1}^{\infty}
\left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right)\left( D^{\alpha\beta}_{n, i_1 \dots i_n}(z) L_{\beta}^{\dag} L_{\alpha} O(z,t) \right)
\right]
\label{eq Hexpansion2}
\end{align}
where we have used cyclicity of trace and integration by parts, to bring the equation of motion into a form which would enable us to write a CQ version of the {\it Heisenberg representation} \cite{oppenheim_post-quantum_2018} for a CQ operator. If we are interested in the expectation value of phase space variables $O(z)=z_i\mathbbm{I}$ then Equation \eqref{eq Hexpansion2} gives
\begin{align}
\frac{d\langle z_i \rangle}{dt}=
\int dz D^{\mu \nu}_{1,i}\tr{L_\nu^\dagger L_\mu \varrho(z,t)}
\label{eq:semiHamiltonsequation}
\end{align}
with all higher order terms vanishing, and we see that $ \sum_{\mu \nu \neq 00} D^{\mu\nu}_{1,i}\langle L^\dagger_\nu L_\mu\rangle$ governs the average rate at which the quantum system moves the classical system through phase space, and with the back-reaction is quantified by the Hermitian matrix $D_1^{\alpha \mu}:= (D_{1}^{br})^{\alpha \mu}$. The force of this back-reaction is especially apparent if the equations of motion are Hamiltonian in the classical limit as in \cite{oppenheim_post-quantum_2018}. I.e. if we define $H_I(z) := h^{\alpha \beta} L^{\dag}_{\beta} L_{\alpha}$ and take $D^{\alpha\beta}_{1,i}=\omega_i^j d_j h^{\alpha\beta}$ with $\omega$ the symplectic form and $d_j$ the exterior derivative. Then Equation \eqref{eq:semiHamiltonsequation} is analogous to Hamilton's equations,
and the CQ evolution equation after tracing out the quantum system has the form of a Liouville's equation to first order and in the classical limit,
\begin{equation}
\frac{\partial \rho(z,t) }{\partial t} = \{H_c, \rho(z,t) \} + \mathrm{tr}\left(\{H_I(z) , \varrho(z) \}\right) + \dots
\label{eq: limit}
\end{equation}
with $\rho(z):=\tr{\varrho(z)}$.
The significance of the second moment is also seen via Equation \eqref{eq Hexpansion2} to be related to the variance of phase space variables $ \sigma_{z_{i_1}z_{i_2}}:=\langle z_{i_1}z_{i_2}\rangle-\langle z_{i_1}\rangle\langle z_{i_2}\rangle$
\begin{align}
\frac{d\sigma^2_{z_{i_1},z_{i_2}}}{dt}=2\langle D^{\alpha\beta}_{2,i_1,i_2} L^\dagger_\beta L_\alpha \rangle
+\langle z_2 D^{\alpha\beta}_{1,z_{i_1}}L^\dagger_\beta L_\alpha\rangle
-
\langle z_{i_2}\rangle\langle D^{\alpha\beta}_{1,z_{i_1}}L^\dagger_\beta L_\alpha\rangle
+\langle z_{i_1} D^{\alpha\beta}_{1,z_{i_2}} L^\dagger_\beta L_\alpha \rangle
-
\langle z_{i_1}\rangle \langle D^{\alpha\beta}_{1,z_{i_2}}L^\dagger_\beta L_\alpha\rangle
\label{eq:variance}
\end{align}
In the case when $D_{1,z_{i_1}}$ is uncorrelated with $z_{i_2}$ and $D_{1,z_{i_2}}$ uncorrelated with $z_{i_1}$, then the growth of the variance only depends on the diffusion coefficient.
The zeroth moment $D^{\alpha\beta}_0$ is just the pure Lindbladian couplings. The simplest example is the case of a pure decoherence process with a single Hermitian Lindblad operator $L$ and decoherence coupling $D_0$. Then we can define a basis $\{|a \rangle\}$ via the eigenvectors of $L$ and
\begin{align}
\bra{a}\frac{\partial\varrho}{\partial t}\ket{b}
=
-i\bra{a}[H(z),\varrho]\ket{b}-\frac{1}{2}D_0(L(a) - L(b))^2 \langle a| \varrho |b \rangle
\end{align}
and we see that the matrix elements of $\varrho$ which quantify coherence between the states $\ket{a}$,$\ket{b}$ decay exponentially fast with a decay rate of
$D_0(L(a) - L(b))^2$. For a damping/pumping process of a quantum harmonic oscillator with Hamiltonian $H=\omega a^\dagger a$, $L_\downarrow=a$, $L_\uparrow=a^\dagger$, $a$ the creation operator, and $D_0^{\uparrow\uparrow}$, $D_0^{\downarrow\downarrow}$ the non-zero couplings, then standard calculations \cite{breuer2002theory,unrav} show that an initial superposition $\frac{1}{\sqrt{2}}\ket{n+m}$ with $n$,$m$ large and $n\gg m$ will initially decohere at a rate of approximately $(D_0^{\uparrow\uparrow}+D_0^{\downarrow\downarrow})(m+n)/2$, and the state will eventually thermalise to a temperature of $\omega /\log{(D_0^{\downarrow\downarrow}/D_0^{\uparrow\uparrow})}$. So in this case, the Lindblad couplings not only determine the rate of decoherence, but also the rate at which energy is pumped into the harmonic oscillator. In the next section we will derive the trade-off between Lindblad couplings and the diffusion coefficients. Although we will sometimes refer to this as a trade-off between decoherence and diffusion, this terminology is only strictly appropriate for pure decoherence processes, while more generally, it is a trade-off between Lindblad couplings and diffusion coefficients.
\section{A Trade off between decoherence and diffusion} \label{sec: tradeoff}
In this section we use positivity conditions to prove that the trade off between decoherence and diffusion seen in models such as those of \cite{diosi1995quantum,oppenheim_post-quantum_2018,unrav} are in fact a general feature of \textit{all} classical-quantum interactions. We shall also generalise this, and derive a trade-off between diffusion and arbitrary Lindbladian coupling strengths. The trade-off is in relation to the strength of the dynamics and is captured by Equation's \eqref{eq: decdifnonham0}, \eqref{eq: generalTradeOffCouplingConstants} and \eqref{eq: decdifnonham}.
In section \ref{sec: fieldTradeOff} we extend the trade-off to the case where the classical and quantum degrees of freedom can be fields and use this to show that treating the metric as being classical necessarily results in diffusion of the gravitational field.
There are two separate possible sources for the force (or {\it drift}) of the back-reaction of the quantum system on phase space -- it can be sourced by either the $D^{0 \alpha}_{1,i}$ components or the Lindbladian components $D^{\alpha \beta}_{1,i}$. We shall deal with both sources simultaneously by considering a CQ Cauchy-Schwartz inequality which arises from the positivity of
\begin{equation}
\label{eq: observationalpositivitycontinous}
\tr{\int dz dz' \Lambda^{\mu \nu}( z|z') O_{ \mu}(z,z')\rho(z') O_{\nu}^{\dag}(z,z')}\geq 0
\end{equation}
for any vector of CQ operators $O_{\mu}$. One can verify that this must be positive directly from the positivity conditions on $\Lambda^{\mu \nu}(z|z')$ and we go through the details in appendix \ref{sec: positivityCondtionsAppendix}. A common choice for $O_{\mu}$ would be the set of operators $L_{\mu}=\{ \mathbb{I}, L_{\alpha}\}$ appearing in the master equation.
The inequality in Equation \eqref{eq: observationalpositivitycontinous} turns out to be especially useful since it can be used to define a (pseudo) inner product on a vector of operators with components $O_{\mu}$ via
\begin{equation}\label{eq: innerProduct}
\langle \bar{O}_1,\bar{O}_2 \rangle = \int dz dz' \Tr{}{\Lambda^{\mu \nu}( z|z')O_{1 \mu} \varrho(z') O^{\dag}_{2 \nu} }
\end{equation}
where $||\bar{O}|| = \sqrt{\langle \bar{O}, \bar{O} \rangle} \geq 0$ due to \eqref{eq: observationalpositivitycontinous}. Technically this is not positive definite, but this shall not be important for our purpose. Taking the combination $O_{\mu} = ||\bar{O}_2||^2 O_{1 \mu} - \langle \bar{O}_1, \bar{O}_2 \rangle O_{2 \mu}$ for vectors $O_{1\mu}, O_{2\mu}$, positivity of the norm gives
\begin{equation}
||\bar{O}||^2 = \ ||\bar{O}_2||^2 \bar{O}_1 - \langle \bar{O}_1, \bar{O}_2 \rangle \bar{O}_2 ||^2 = ||\bar{O}_2||^2 \left( ||\bar{O}_1||^2 ||\bar{O}_2||^2 - |\langle \bar{O}_1,\bar{O}_2 \rangle |^2 \right) \geq 0,
\end{equation}
and as long as $||\bar{O}_2|| \neq 0$ we have a Cauchy- Schwartz inequality
\begin{equation}\label{eq: continuousCauchySchwarz}
||\bar{O}_1||^2 ||\bar{O}_2||^2 - |\langle \bar{O}_1,\bar{O}_2 \rangle |^2 \geq 0.
\end{equation}
We can use \eqref{eq: continuousCauchySchwarz} to get a trade-off between the observed diffusion and decoherence by picking $O_{2 \mu} = \delta_{\mu}^{\alpha} L_{\alpha}$ and $O_{1 \mu} = b^i(z-z')_i L_{\mu}$, where $L_{\mu}=\{ \mathbb{I}, L_{\alpha}\}$ are the Lindblad operators appearing in the master equation. In this case $||\bar{O}_2|| = \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho L_{\beta}^{\dag}}$ and one can verify using CQ Pawula theorem \cite{UCLPawula}\footnote{In particular, to reach this conclusion one can insert the CQ state into the CQ Cauchy-Schwartz inequality and repeat the proof of the Pawula theorem \cite{UCLPawula}, which must now hold once averaged over the state.} that in order to have non-trivial back-reaction on the quantum system complete positivity demands that $||\bar{O}_2|| > 0$, meaning the Cauchy-Schwartz inequality in Equation \eqref{eq: continuousCauchySchwarz} must hold. By using the short time moment expansion of $\Lambda^{\mu \nu}(z|z')$ defined in Equation \eqref{eq: shortimeCP} and using integration by parts, we then arrive at the observational trade-off between decoherence and diffusion
\begin{equation}\label{eq: expandedTradeOff}
\int dz \Tr{}{2 b^{i *} D^{\mu \nu}_{2, ij} b^j L_{\mu} \varrho(z) L_{\nu }^{\dag} } \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho(z) L_{\beta}^{\dag}} \geq \left| \int dz \Tr{}{b^i D^{\mu \alpha}_{1,i} L_{\mu} \varrho(z) L_{\alpha}^{\dag} } \right|^2,
\end{equation}
which must hold for any positive CQ state $\varrho(z)$. Stripping out the $b^i$ vectors, \eqref{eq: expandedTradeOff} is equivalent to the matrix positivity condition
\begin{equation}\label{eq: decdifnonham0}
0 \preceq 2 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{br} \rangle \langle D_1^{br} \rangle^{\dag}, \ \ \ \forall \varrho(z),
\end{equation}
where we define
\begin{equation} \label{eq: definitionExpectation}
\langle D_{0} \rangle = \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho(z) L_{\beta}^{\dag}} , \ \langle D_1^{br} \rangle_i =\int dz \Tr{}{D^{\mu \alpha}_{1,i} L_{\mu} \varrho(z) L_{\alpha}^{\dag} } , \ \langle D_{2} \rangle_{ij} = \int dz \Tr{}{ D^{\mu \nu}_{2, ij} L_{\mu} \varrho(z) L_{\nu }^{\dag} } .
\end{equation}
Since \eqref{eq: decdifnonham0} holds for all states, the tightest bound is provided by the infimum over all states
\begin{equation}
0 \preceq \inf_{\varrho(z)}\{2 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{br} \rangle \langle D_1^{br} \rangle^{\dag}\} .
\end{equation}
The quantities $\langle D_2 \rangle$ and $\langle D_0\rangle$ appearing in Equation \eqref{eq: decdifnonham0} are related to observational quantities. In particular $\langle D_2 \rangle$ is the expectation value of the amount of classical diffusion which is observed and $\langle D_{0} \rangle$ is related to the amount of decoherence on the quantum system. The expectation value of the back-reaction matrix $\langle D_{1}^{br} \rangle $ quantifies the amount of back-reaction on the classical system.
In the trivial case $D_1^{br}=0$, Equation \eqref{eq: decdifnonham0} places little restriction on the diffusion and Lindbladian rates appearing on the left hand side. We already knew from \cite{GKS76,Lindblad76} that the $D_0^{\alpha\beta}$ must be a positive semi-definite matrix, and we also know that diffusion coefficients must be positive semi-definite.
However, in the non-trivial case, the larger the back-reaction exerted by the quantum system, the stronger the trade-off between the diffusion coefficients and Lindbladian coupling.
Equation \eqref{eq: decdifnonham0} gives a general trade-off between observed diffusion and Lindbladian rates, but we can also find a trade-off in terms of a theory's coupling coefficients alone. We show in appendix \ref{sec: diffusionDecoherenceCouplingTradeOff} that the general matrix trade-off
\begin{equation}\label{eq: generalTradeOffCouplingConstants}
\ D_1^{br} D_0^{-1} D_1^{ br \dag} \preceq 2 D_2
\end{equation}
holds for the matrix whose elements are the couplings $D^{\mu \nu}_{2,ij} , D^{\alpha \mu}_{1,i}, D_0^{\alpha \beta}$ for any CQ dynamics. Moreover, $ (\mathbb{I}- D_0 D_0^{-1})D_1^{br} =0$, which tells us that $D_0$ cannot vanish if there is non-zero back-reaction. Equation \eqref{eq: generalTradeOffCouplingConstants} quantifies the required amount of decoherence and diffusion in order for the dynamics to be completely positive. In Equation \eqref{eq: generalTradeOffCouplingConstants}, and throughout, $D_0^{-1}$ is the generalized inverse of $D_0^{\alpha \beta}$, since $D_0^{\alpha \beta}$ is only required to be positive semi-definite. In the special case of a single Lindblad operator $\alpha=1$ and classical degree of freedom, and when the only non-zero couplings are $D_0^{11}:=D_0$, $D_{2,pp}^{00}:=2D_2$ and $D_{1,q}^0=1$ this trade-off reduces to the condition $D_2D_0\geq 1$ used in \cite{diosi1995quantum}.
It is also useful to try to obtain an observational trade-off in terms of the total drift due to back-reaction as calculated in Equation \eqref{eq:semiHamiltonsequation}
\begin{equation}
\langle D_1^{T} \rangle_i =\sum_{\mu \nu \neq 00}\int dz \Tr{}{D^{\mu \nu}_{1,i} L_{\mu} \varrho(z) L_{\nu}^{\dag} }.
\end{equation}
It follows directly from Equation \eqref{eq: decdifnonham0} that when the back-reaction is sourced by either $D_{1,i}^{0 \mu}$ or $D_{1,i}^{\alpha \beta}$ we can arrive at the observational trade-off in terms of the total drift\footnote{We believe that \eqref{eq: decdifnonham} should hold more generally, though we don't have a general proof. }
\begin{equation}\label{eq: decdifnonham}
0 \preceq 8 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{T} \rangle \langle D_1^{T} \rangle^{\dag}, \ \ \ \forall \varrho(z),
\end{equation}
where the quantities appearing in Equation \eqref{eq: decdifnonham} are now all observational quantities, related to drift, decoherence and diffusion as outlined in the previous subsection \ref{ss:momentsmeaning}.
In the case where the back-reaction is Hamiltonian at first order in the sense of Equation \eqref{eq: limit}, then \eqref{eq: decdifnonham} can be written as
\begin{equation}
\label{eq: decodiff}
\langle \omega \cdot \frac{\partial H_I}{ \partial \vec{z}} \rangle \langle \omega \cdot \frac{\partial H_I}{ \partial \vec{z}} \rangle^{\dag} \preceq 8 \langle D_{2} \rangle \langle D_0 \rangle, \ \ \ \forall \varrho(z) .
\end{equation}
As a result, we can derive a trade-off between diffusion and decoherence for any theory which reproduces this classical limit and treats one of the systems classically.
To summarize, whenever back-reaction of the quantum system on the classical system induces a force on the phase space, then we have a trade-off between the amount of diffusion on the classical system and the strength of decoherence on the quantum system (or more precisely the strength of the Lindbladian couplings $D_0^{\alpha \beta}$). This can be expressed both as a condition on the matrix of coupling co-efficients in the master equation, via Equation \eqref{eq: generalTradeOffCouplingConstants} or in terms of observable quantities using
Equation's \eqref{eq: decdifnonham0} and \eqref{eq: decdifnonham}. In the case when the back-reaction is Hamiltonian, we further have Equation \eqref{eq: decodiff}. We would like to apply this trade-off to the case of gravity in the non-relativistic, Newtonian limit. In order to do so, we will need to generalise the trade-off to the case of quantum fields interacting with classical one, which we do in Section \ref{sec: fieldTradeOff}. The goal will be to understand the implications of treating the metric (or Newtonian potential) as being classical by using the trade-off when the quantum back-reaction induces a force on the gravitational field which, on expectation, is the same as the weak field limit of general relativity.
\section{Trade off in the presence of fields}\label{sec: fieldTradeOff}
We would like to explore the trade-off in the gravitational setting and explore the consequences of treating the gravitational field as being classical and matter quantum. Since gravity is a field theory, we must first discuss classical-quantum master equations in the presence of fields. In the field theoretic case, both the Lindblad operators and the phase space degrees of freedom can have spatial dependence, $z(x), L_{\mu}(x)$ and a general bounded CP map which preserves the classicality of the two systems can be written \cite{oppenheim_post-quantum_2018}
\begin{equation}\label{eq: mapFields}
\rho(z, t) = \int dz' dx dy\Lambda^{\mu \nu}(z|z',t; x,y) L_{\mu}(x, z, z') \varrho(z',0) L_{\nu}^{\dag}(y, z, z'),
\end{equation}
where, as is usually the case with fields, in Equation \eqref{eq: mapFields} it should be implicitly understood that a smearing procedure has been implemented. We elaborate on some of details when fields are introduced in appendix \ref{sec: cqfieldsAppendix}. The condition for \eqref{eq: mapFields} to be completely positive on \textit{all} CQ states is
\begin{equation}\label{eq: fieldPositivityAll}
\int dz dx dy A_{\mu}^*(x, z, z') \Lambda^{\mu \nu}(z|z'; x,y) A_{\nu}(y, z, z')\geq 0
\end{equation}
meaning that $\Lambda^{\mu \nu}(x,y) $ can be viewed as a positive matrix in $\mu \nu$ and a positive kernel in $x,y$. In the field theoretic case one can still perform a Kramers-Moyal expansion and find a trade-off between the coefficients $D_0(x,y), D_1(x,y), D_2(x,y)$ appearing in the master equation. The coefficients now have an $x,y$ dependence, due to the spatial dependence of the Lindblad operators. The coefficients $D_1(x,y),D_2(x,y)$ still have a natural interpretation as measuring the amount of force ({\it drift}) and diffusion, whilst $D_0(x,y)$ describes the purely quantum evolution on the system and can be related to decoherence.
Using the positivity condition in \eqref{eq: fieldPositivityAll} we find the same trade of between coupling constants in Equation \eqref{eq: generalTradeOffCouplingConstants} but where now $D_2(x,y)$ is the $(p+1)n \times (p+1)n$ matrix-kernel with elements $D_{2,ij}^{\mu \nu}(x,y)$, $D_{1}^{br}(x,y)$ is the $( p+1)n \times p$ matrix-kernel with rows labeled by $\mu i$, columns labelled by $\beta$, and elements $D_{1, i}^{ \mu \beta}(x,y)$, and $D_0(x,y)$ is the $p\times p $ decoherence matrix-kernel with elements $D_0^{\alpha \beta}(x,y)$.\footnote{Here $i \in \{1, \dots, n \}$ $\alpha \in \{1, \dots, p\} $ and $\mu \in \{ 1, \dots, p+1\}$,} In the field theoretic trade off we are treating the objects in Equation \eqref{eq: generalTradeOffCouplingConstants} as matrix-kernels, so that for any position dependent vector $b^i_{\mu}(x)$, $( D_2 b )_{i}^{\mu} (x) = \int dy D^{\mu \nu}_{2,i j}(x,y) b^{j}_{\nu}(y) $, whilst for any position dependent vector $a_{\alpha}(x)$, $( D_0 \alpha )^{\alpha} (x) = \int dy D_{0}^{\alpha \beta}(x,y) \alpha_{\beta}(y) $. Explicitly, we find that positivity of the dynamics is equivalent to the matrix condition
\begin{align} \label{eq: generalFieldMatrixInequalityMain}
\int dx dy [ b^*(x), \alpha^*(x)]
\begin{bmatrix}
2 D_{2}(x,y) & D_1^{br}(x,y)\\ D^{br}_1(x,y) & D_0(x,y)
\end{bmatrix}
\begin{bmatrix}
b(y) \\ \alpha(y)
\end{bmatrix} \geq 0
\end{align}
which should be positive for any position dependent vectors $b^{i}_{\mu}(x)$ and $a_{\alpha}(x)$. This is equivalent to trade-off between coupling constants in Equation \eqref{eq: generalTradeOffCouplingConstants} if we view \eqref{eq: generalTradeOffCouplingConstants} as a matrix-kernel equation.
Though we make no assumption on the locality of the Lindbladian and diffusion couplings, we shall hereby assume that the drift back-reaction is local, so that $D_1^{br}(x,y) = \delta(x,y) D_1^{br}(x)$. As we shall see in the next section, this is a natural assumption if we want to have back-reaction which is given by a local Hamiltonian. However, one might not want to assume that the form of the Hamiltonian remains unchanged to arbitrarily small distances. With this locality assumption, Equation \eqref{eq: generalFieldMatrixInequalityMain} gives rise to the same trade-off of Equation \eqref{eq: generalTradeOffCouplingConstants}, where the trade-off is to be interpreted as a matrix kernel inequality. Writing this out explicitly we have
\begin{align}
\label{eq: generalTradeOffCouplingConstants_Fields_local}
\int dx dy \alpha_{\nu}^{i*}(x)D_{1,i}^{\mu \alpha}(x) (D_0^{-1})_{\alpha \beta}(x,y) D_{1,j}^{\beta \nu}(x')\alpha_{\nu}^i(x') \leq \int dx dy 2\alpha_{\mu}^{i *}(x) D_{2,ij}^{\mu \nu}(x,y) \alpha_{\nu}^j(y),
\end{align}
where asking that this inequality holds for all vectors $\alpha_{\mu}^i(x)$ is equivalent to the matrix-kernel trade-off condition of Equation \eqref{eq: generalTradeOffCouplingConstants}. We give two examples of master Equations satisfying the coupling constant trade-off in appendix \ref{app: examplesOfKernals}. The decoherence-diffusion trade-off tells us how much diffusion and stochasticity is required to maintain coherence when the quantum system back-reacts on the classical one. If the interaction between the classical and quantum degrees of freedom is dictated by unbounded operators, such as the mass density, then there can exist states for which the back-reaction can be made arbitrarily large. This is the case for a quantum particle interacting with its Newtonian potential through its mass density at arbitrarily short distances. Hence, if one considers a particle in a superposition of two peaked mass densities, then there can be an arbitrarily large response in the Newtonian potential around those points, and either there must be an arbitrary amount of diffusion, or the decoherence must occur arbitrarily fast. The former is unphysical, while the latter turns out to be the case in simple examples of theories such as those discussed in Appendix \ref{sec: decoherenceRatesAppendix}.
Since our goal is to experimentally constrain classical-quantum theories of gravity, we shall hereby ask that the map \eqref{eq: mapFields} is CP when acting on all physical states $\rho$. If one allows for arbitrarily peaked mass distributions then the coupling constant trade-off of Equation \eqref{eq: generalTradeOffCouplingConstants_Fields_local} should be satisfied. In the field theoretic case, we can similarly find an observational trade-off, relating the expected value of the diffusion matrix $\langle D_2(x,y) \rangle$ to the expected value of the drift in a physical state $\varrho$ as we did in Section \ref{sec: tradeoff}.
This is done explicitly in Appendix \ref{sec: cqfieldsAppendix}, using a field theoretic version of the Cauchy-Schwartz inequality given by Equation \eqref{eq: continuousCauchySchwarzField}, we find
\begin{equation}\label{eq: observationalTradeOffFields0}
2 \langle D_2(x,x) \rangle \int dx' dy' \langle D_0(x',y') \rangle \succeq \langle D_1^{br}(x) \rangle \langle D_1^{br}(x) \rangle^{\dag},
\end{equation}
where equation \eqref{eq: observationalTradeOffFields} is to be understood as a matrix inequality with entries
\begin{equation}\label{eq: fieldCoeffDefinitions}
\begin{split}
& \langle D_0(x,y)\rangle = \int dz \Tr{}{D_0^{\alpha \beta} L_{\alpha}(x) \varrho L_{\beta}^{\dag}(y)}, \\
& \langle D_1^{br}(x,y)\rangle_i = \int dz \Tr{}{D_{1,i}^{\mu \alpha} L_{\mu}(x) \varrho L_{\alpha}^{\dag}(x)}, \\
& \langle D_2(x,y)\rangle_{ij} = \int dz \Tr{}{D_{2,ij}^{\mu \nu} L_{\alpha}(x) \varrho L_{\beta}^{\dag}(y)}.
\end{split}
\end{equation}
Similarly, when the back-reaction is sourced by either $D_{1,i}^{0 \mu}$ or $D_{1,i}^{\alpha \beta}$ it follows from Equation \eqref{eq: observationalTradeOffFields0} we can arrive at the observational trade-off in terms of the total drift due to back-reaction
\begin{equation}\label{eq: observationalTradeOffFields}
8 \langle D_2(x,x) \rangle \int dx' dy' \langle D_0(x',y') \rangle \succeq \langle D_1^{T}(x) \rangle \langle D_1^{T}(x) \rangle^{\dag},
\end{equation}
where
\begin{equation}\label{eq: totdriftfieldCoeffDefinition}
\langle D_1^T(x)\rangle_i = \int dz \Tr{}{D_{1,i}^{0 \alpha}(x) \varrho L_{\alpha}^{\dag}(x)+ D_{1,i}^{ \alpha 0}(x) L_{\alpha} \varrho (x) + D_{1,i}^{\alpha \beta}(x) L_{\alpha}(x) \varrho L_{\beta}^{\dag}(x)}.
\end{equation}
We shall now use the trade-off to study the consequences of treating the gravitational field classically. We will consider the back-reaction of the mass on the gravitational field to be governed by the Newtonian interaction (or more accurately, a weak field limit of General Relativity). We shall then find that experimental bounds on coherence lifetimes for particles in superposition require large diffusion in the gravitational field in order to be maintained and this can be upper bounded by gravitational experiments.
To summarise this section, we have derived the trade-off between decoherence and diffusion for classical-quantum field theories, both in terms of coupling constants of the theory and in terms of observational quantities. This trade-off puts tight observational constraints on classical theories of gravity which we now discuss.
\section{Physical constraints on the classicality of gravity}
\label{sec: gravity}
In this section we apply the trade-off of Equation \eqref{eq: generalFieldMatrixInequalityMain} to the case of gravity.
Since the trade-offs derived in the previous section depend only on the back-reaction, or drift term, they are insensitive to the particulars of the theory. We shall consider the Newtonian, non-relativistic limit of a classical gravitational field which we reproduce in Appendix \ref{sec: cqlimit}. It is in taking this limit where some care should be taken, since one is gauge fixing the full general relativistic theory. We denote $\Phi$ to be the Newtonian potential and in the weak field limit of General Relativity, it has a conjugate momenta we denote by $\pi_\Phi$. We assume
\begin{enumerate}[label=(\roman*)]
\item The theory satisfies the assumptions used to derive the master equation as in Sections \ref{sec: fieldTradeOff}; in particular that the theory be a completely positive norm-preserving Markovian map, and that we can perform a short-time Kramers-Moyal expansion as in Appendix \ref{sec: cqfieldsAppendix}.
\item We apply the theory to the weak field limit of General Relativity, where as recalled in Appendix \ref{sec: cqlimit} the Newtonian potential interacts with matter through its mass density $m(x)$,
\begin{equation}
H_I(\Phi) = \int d^3 x \Phi(x) m(x).
\label{eq:interaction_ham}
\end{equation}
and the conjugate momentum to $\Phi$ satisfies
\begin{align}
\dot{\pi}_{\Phi} = \frac{\nabla^2 \Phi}{4 \pi G} - m(x)
\label{eq:dotpi}
\end{align}
where in the $c \to \infty$ limit we recover Poisson's equation for the Newtonian potential.
We assume this limit of General Relativity is satisfied on expectation, at least to leading order.
\item In relating $D_0$ to the decoherence rate of a particle in superposition, we shall assume that the state of interest is well approximated by a state living in a Hilbert space of fixed particle number. We believe this is a mild assumption: ordinary non-relativistic quantum mechanics is described via a single particle Hilbert space, and we frequently place composite massive particles in superposition and they do not typically decay into multiple particles.
\item We will assume that the diffusion kernel $ D_2(\Phi,x,x')$ does not depend on $\pi_\Phi$ i.e. it is {\it minimally coupled}. This is reasonable,
since in the purely classical case matter couples to the Newtonian potential and not its conjugate momenta.
\label{ass:minimal}
\end{enumerate}
With these assumptions, and
treating the matter density as a quantum operator $\hat{m}(x)$, this tells us that in order for the back-reaction term to reproduce the Newtonian interaction on average
\begin{align}
\tr{\{H_I,\varrho\}}=
\tr{\int d^3 x \ \hat{m}(x) \frac{\delta \varrho}{\delta \pi_\Phi}} = -\sum_{\mu \nu \neq 00} \tr{\int d^3 x D_{1,\pi_{\Phi}}^{\mu \nu}(\Phi, \pi_{\Phi}, x) L_{\mu}(x) \frac{\delta \varrho}{\delta \pi_\Phi} L_{\nu}^{\dag}(x)},
\label{eq:Nsemi}
\end{align}
then we must pick
\begin{equation}\label{eq: firstMomentGravity}
\langle D_{1,\pi_{\phi}}^{T}(\Phi, \pi_{\Phi}, x) \rangle = - \langle \hat{m}(x) \rangle,
\end{equation}
meaning that the back-reaction matrix $D_{1,\pi_{\Phi}}^{\mu \alpha}$ is non vanishing. In Appendix \ref{sec: cqlimit} we give examples of master equations for which \eqref{eq: firstMomentGravity} is satisfied, but their details are irrelevant since we only require the expectation of the back-reaction force to be the expectation value of the mass -- a necessary condition for the theory to reproduce Newtonian gravity.
As a consequence of the coupling constant and observational trade-offs derived in Equations \eqref{eq: generalTradeOffCouplingConstants_Fields_local} and \eqref{eq: observationalTradeOffFields0}, a non-zero $D_{1,\pi_{\Phi}}$ implies that there must be diffusion in the momenta conjugate to $\pi_{\Phi}$.
This diffusion is equivalent to adding a stochastic random process $J(x,t)$ (the Langevin picture), to the equation of motion \eqref{eq:dotpi} to give
\begin{equation}\label{eq: diffusivePi}
\dot{\pi}_{\Phi} = \frac{\nabla^2 \Phi}{4 \pi G} - m(x) + u(\Phi,\hat{m})
J(t,x),
\end{equation}
where we allow some {\it colouring} to the noise via a function $u(\Phi,\hat{m})$ which can depend on $\Phi$, and the matter distribution $\hat{m}$ (assumption \ref{ass:minimal}). The noise process satisfies
\begin{equation}\label{eq: statsJ}
\mathbb{E}_{m, \Phi}[u J(x,t)] =0, \ \ \ \mathbb{E}_{m, \Phi}[uJ(x,t) u J(y,t') ]= 2\langle D_2(x,y,\Phi) \rangle \delta(t,t'),
\end{equation}
where we have defined $\langle D_2(x,y,\Phi) \rangle = \Tr{}{ D_{2}^{\mu \nu} (x,y, \Phi) L_{\mu}(x) \rho L^{\dag}_{\nu}(y)}$, and $\rho$ is the quantum state for the decohered mass density.
Here the $m,\Phi$ subscripts of $\mathbb{E}_{m,\Phi}$ allow for the possibility that the statistics of the noise process can be dependent on the Newtonian potential and mass distribution of the particle. The restriction on $\mathbb{E}_{m, \Phi}[uJ(x,t)]$ follows from assumption (ii). If $uJ(x,t)$ is Gaussian, Equation \eqref{eq: statsJ}
completely determines the noise process, but in general, higher order correlations are possible, although they need not concern us here, since we are only interested in bounding the effects due to $D_2(x,y,\Phi)$.
In the non-relativistic limit, where $c\rightarrow\infty$, we can take $\dot{\pi}_{\Phi}$ to be small in comparison to the other terms, and we recover Poisson's equation for gravity, but with a stochastic contribution to the mass. This is precisely as expected on purely physical grounds: in order to maintain coherence of any mass in superposition, there must be noise in the Newtonian potential and this must be such that we cannot tell which element of the superposition the particle will be in, meaning the Newtonian potential should look like it is being sourced in part by a random mass distribution. In other words, the trade-off requires that the stochastic component of the coupling obscures the amount of mass $m$ at any point.
The solution to Equation \eqref{eq: diffusivePi} is given by
\begin{equation}\label{eq: poissonnoiseMain}
\Phi(t,x) \approx -G \int d^3 x' \frac{[m(x',t)-u(\Phi, \hat{m})J(x',t)]}{|x-x'|},
\end{equation}
and a formal treatment of solutions to non-linear stochastic integrals of the form of Equation \eqref{eq: diffusivePi} can be found in \cite{DalangHigh}. A higher precision calculation would involve a full simulation of CQ dynamics and in Appendix \ref{sec: cqlimit} we show in full detail the evolution the Newtonian potential looks like for general continuous CQ theories using the continuous unraveling of CQ dynamics introduced in \cite{UCLHealing}. We find the effects are qualitatively and quantitatively the same as Equation \eqref{eq: poissonnoiseMain}.
In \cite{UCLPawula} it was shown that there are two classes of CQ dynamics, at least in the sense that there are those with continuous trajectories in phase space and those which contain discrete jumps. For the class of continuous CQ models (see \cite{diosi2011gravity} and appendix \ref{ss:continuous}), we know that $J(x,t)$ should be described by a white noise process in time, and its statistics should be independent of the mass density of the particle.
We go through the full CQ calculation for the continuous models in Appendix \ref{eq: newtUnrav}.
For the discrete class (see \cite{oppenheim_post-quantum_2018, oppenheim2021constraints} and Appendix \ref{ss: discrete}), $J(x,t)$ can involve higher order moments, and will generally be described by a jump process \cite{UCLPawula, unrav}. It's statistics can also depend on the mass density, since in general the diffusion matrix $D^{\mu \nu}_{2,ij}$ couples to Lindblad operators. It is worth noting that the discrete CQ theories considered in \cite{oppenheim_post-quantum_2018, unrav, UCLconstraints} generically suppress higher order moments, and often we expect that we can approximate the dynamics by a Gaussian process, but this need not be the case in general.
This variation in Newtonian potential leads to observational consequences which can be used to experimentally test and constrain CQ theories of gravity for various choices of kernels appearing in the CQ master equation. One immediate consequence is that the variation in Newtonian potential leads to a variation of force experienced by a particle or composite mass via $\vec{F}_{tot} = -\int d^3 x m(x) \nabla \Phi(x)$. We can also estimate the time averaged force via $\frac{1}{T}\int_0^T\vec{F}_{tot}$ where $T$ is the time over which the force is measured and is the useful quantity when comparing with experiments. Using Equation \eqref{eq: poissonnoiseMain}, in Appendix \ref{sec: tabletop} we find that the variance of the magnitude of the time averaged force experienced by a particle in a Newtonian potential is given by
\begin{equation}\label{eq: variationForceMain}
\sigma_F^2 = \frac{2G^2}{T} \int d^3x d^3y d^3x' d^3y' m(x) m(y) \frac{(\vec{x}- \vec{x}') \cdot (\vec{y}- \vec{y}')}{|x-x'|^3 |y-y'|^3} \langle D_2(x',y',\Phi)\rangle,
\end{equation}
where the variation is averaged over a time period $T$. We will use this to estimate the variation in precision measurements of mass, such as modern versions of the Cavendish experiement for various choices of $\langle D_2(x',y',\Phi)\rangle$.
On the other hand, experimentally measured decoherence rates can be related to $D_0$. We explore the calculation of decoherence rates in gravity in detail in~\cite{gravitydecoherence}. The important point is that the decoherence rate is dominated by the background Newtonian potential $\Phi_b$ due to the Earth. In Appendix \ref{sec: decoherenceRatesAppendix}, we show that for a mass whose quantum state is a superposition of two states $|L\rangle$ and $|R\rangle$ of approximately orthogonal mass densities $m_L(x),m_R(x)$, and whose separation we take to be larger than the correlation range of $D_0(x,y)$, the decoherence rate is given by
\begin{equation}\label{eq: decoherencerateMain}
\lambda = \frac{1}{2}\int dx dy D_0^{\alpha \beta}( x,y) (\langle L | L_{\beta}^{\dag}(y) L_{\alpha}(x) |L \rangle + \langle R | L_{\beta}^{\dag}(y) L_{\alpha}(x) |R \rangle ).
\end{equation}
Via the coupling constant trade-off, Equations \eqref{eq: variationForceMain} and \eqref{eq: decoherencerateMain} then give rise to a double sided squeeze on the coupling $D_2$. Equation \eqref{eq: variationForceMain} upper bounds $D_2$ in terms of the uncertainty of
acceleration measurements seen in gravitational torsion experiments, whilst the coupling constant trade-off Equation \eqref{eq: decoherencerateMain} lower bounds $D_2$ in terms of experimentally measured decoherence rates arising from interferometry experiments.
We now show this for various choices of diffusion kernel, with the details given in Appendix \ref{sec: tabletop}. The diffusion coupling strength will be characterized by the coupling constant $D_2$, which we take to be a dimension-full quantity with units $kg^2 s m^{-3}$, and is related to the rate of diffusion for the conjugate momenta of the Newtonian potential. We upper bound $D_2$ by considering the variation of the time averaged acceleration $\sigma_a = \frac{\sigma_F}{M}$ for a composite mass $M$ which contains $N$ atoms which we treat as spheres of constant density $\rho$ with radius $r_N$ and mass $m_N$. We lower bound $D_2$ via the coupling constant trade-off of Equation \eqref{eq: generalFieldMatrixInequalityMain} and then by considering bounds on the coherence time for particles with total mass $M_{\lambda}$, which have typical length scale when in superposition $R_{\lambda}$ and volume $V_{\lambda}$.
For continuous dynamics $\langle D_{2}(x,y, \Phi) \rangle$ = $D_2(x,y,\Phi)$ since the diffusion is not associated to any Lindblad operators. Let us now consider a very natural kernel, namely $D_2(x,y;\Phi) = D_2(\Phi) \delta(x,y)$ which is both translation invariant, and does not create any correlations over space-like separated regions. In general, the squeeze will depend on the functional choice of $D_2(\Phi)$ on the Newtonian potential. However, in the presence of a large background potential $\Phi_b$, such as that of the Earth's, we will often be able to approximate $D_2(\Phi) = D_2(\Phi_b)$. This is true for kernels which depend on $\Phi$ and $\nabla \Phi$, though the approximation does not hold for all kernels, for example $D_2 \sim -\nabla^2 \Phi$ of Equation \eqref{eq: lapbeltweak} which creates diffusion only where there is mass density. For diffusion kernels $D_2(\Phi_b)$ where the background potential is dominant we find the promised squeeze on $D_2(\Phi_b)$
\begin{equation}\label{eq: contLocal}
\frac{\sigma_a^2 N r_N^4 T}{V_b G^2 } \geq D_2 \geq \frac{M_{\lambda}^2}{V_{\lambda} \lambda},
\end{equation}
where $V_b$ is the volume of space over which the background Newtonian potential is significant. $V_b$ enters since the variation in acceleration is found to be
\begin{equation}\label{eq: intD2Main}
\sigma_a^2 \sim \frac{D_2 G^2 }{r_N^4 N T} \int d^3x' D_2(\Phi_b) ,
\end{equation}
where the $d^3 x'$ integral is over all space. This immediately rules out continuous theories with noise everywhere, i.e, with a diffusion coefficient independent of the Newtonian potential since the integral will diverge.
Standard Cavendish type classical torsion balance experiments measure accelerations of the order $10^{-7}ms^{-2}$, so a very conservative bound is $\sigma_a \sim 10^{-7}ms^{-2}$, whilst for a kg mass $N\sim 10^{26}$ and $r_N\sim 10^{-15} m$. Conservatively taking $V_b \sim r_E^2 h \ m^3$ where $r_E$ is the radius of the Earth and $h$ is the atmospheric height gives $D_2 \leq 10^{-41}kg^2 s m^{-3}$. The decoherence rate $\lambda$ is bounded by various experiments~\cite{bassi}. Typically, the goal of such experiments is to witness interference patterns of molecules which are as massive as possible. Taking a conservative bound on $\lambda$, for example that arising from the interferometry experiment of~\cite{Gerlich2011} which saw coherence in large organic fullerene molecules with total mass $M_{\lambda} = 10^{-24}kg$ over a timescale of $0.1s$, gives an upper bound on the decoherence rate $\lambda < 10^{1} s^{-1}$. Fullerene molecules have typical size $ 10^{-9}m$. After passing through the slits the molecule becomes delocalized in the transverse direction on the order of $10^{-7}m$ before being detected. Since the interference effects are due to the superposition in the transverse $x$ direction, which is the direction of alignment of the gratings, it seems like a reasonable assumption to take the size of the wavepacket in the remaining $y,z$ direction to be the size of the fullerene, since we could imagine measuring the $y,z$ directions without effecting the coherence. We therefore take the volume $V_{\lambda} \sim 10^{-9} 10^{-9} 10^{-7}m^3 = 10^{-25}m^3$, which gives $D_2 \geq 10^{-24} kg^2 s m^{-3}$, and suggests that classical-quantum theories of gravity with local continuous noise need to have a dependence on the Newtonian potential which will suppress the diffusion by $20$ orders of magnitude. This happens to be the case for the kernel from Section \ref{ss:lapkernel} and whose motivation comes from the constraint algebra \cite{oppenheim2021constraints}.
On the other hand, the discrete models appear less constrained due to the suppression of the noise away from the mass density. For example consider the local discrete jumping models, such as the one given in Section \ref{ss: discrete} which have $\langle D_2(x,y,\Phi_b)\rangle = \frac{ l_P^3 D_2(\Phi_b)}{m_P } m(x)$, where $m_P = \sqrt{\frac{\hbar c}{G}}$ is the Planck mass and $l_{P} = \sqrt{\frac{\hbar G}{c^3}}$ is the Planck length, required to ensure $D_2$ has the units of $kg^2 s m^{-3}$. We find the squeeze on $D_2$
\begin{equation}\label{eq: disLocal}
\frac{ \sigma_a^2 N r_N^4 T}{ m_N G^2} \geq \frac{l_P^3}{m_P} D_2 \geq \frac{ M_{\lambda}}{\lambda},
\end{equation}
and plugging in the numbers tells us that discrete theories of classical gravity are not ruled out by experiment and we find $10^{-1}kg s \geq \frac{l_P^3}{m_P} D_2 \geq 10^{-25}kg s $.
We can also consider other noise kernels, with examples and some discussion given in Section \ref{app: examplesOfKernals}. A natural kernel is $D_2(x,y,\Phi_b) = - l_P^2 D(\Phi_b) \nabla^2 \delta(x,y)$. The inverse Lindbladian kernel satisfying the coupling constants trade-off is to zeroeth order in $\Phi(x)$, the Diosi-Penrose kernel $D_0(x,y,\Phi_b) = \frac{D_0(\Phi_b)}{|x-y|}$. For this choice of dynamics, we find the squeeze for $D_2$ in terms of the variation in acceleration
\begin{equation}\label{eq: contDP}
\frac{\sigma_a^2 N r_N^3 T}{G^2} \geq l_P^2 D_2 \geq \frac{M^2_{\lambda}}{R_{\lambda} \lambda} .
\end{equation}
Using the same numbers as for the local continuous model, with $R_{\lambda} \sim V_{\lambda}^{1/3}\sim10^{-9}m$ we find that classical torsion experiments upper bound $D_2$ by $10^{-9}kg^2 s m^{-1} \geq l_P^2 D_2$, whilst interferometry experiments bound $D_2$ from below via $l_P^2 D_2 \geq 10^{-40} kg^2 s m^{-1} $.
Equations \eqref{eq: contLocal}, \eqref{eq: disLocal} and \eqref{eq: contDP} show that classical theories of gravity are squeezed by experiments from both ways. We have here been extremely conservative, and we anticipate that further analysis, as well as near term experiments, can tighten the bounds by orders of magnitude. There are several proposals for table-top experiments to precisely measure gravity, some of which have recently been performed, and which could give rise to tighter upper bounds on $D_2$. Some of these experiments involve millimeter-sized masses whose gravitational coupling is measured via torsional pendula~\cite{westphal2020measurement,Schm_le_2016}, or rotating attractors~\cite{PhysRevLett.124.101101}. With such devices, the gravitational coupling between small masses can be measured while limiting the amount of other sources of noise. There are proposals for further mitigating the noise due to the environment, including the inertial noise, gas particles collisions, photon scattering on the masses, and curvature fluctuations due to other sources~\cite{Chevalier_2020,vandekamp2020quantum,toros2020relative}. Other experiments are based on interference between masses; for example, atomic interferometers allow for the measurement of the curvature of space-time over a macroscopic superposition~\cite{Chou1630,PhysRevLett.118.183602}.
We can get stronger lower bounds via improved coherence experiments. Typically, the goal of such experiments is to witness interference patterns of molecules which are as massive as possible, while here, we see that the experimental bound on CQ theories is generically obtained by maximizing the coherence time for massive particles with as small wave-packet size $V_{\lambda}$.
Thus far in this section we have considered local effects on particles due to the diffusion. While this enables us to rule out some types of theories, the bounds are generally weak if one wanted to rule out all of them. However, it may be possible to do so via cosmological considerations. In attempting to place experimental constraints on this diffusion, it is also worth considering other regimes, such as longer range effects which might be detected by gravitational wave detectors such as LIGO.
In Appendix \ref{sec: energyProduction}, we begin a study of the cosmological effects of the diffusion by studying the observational trade-off in Equation \eqref{eq: observationalTradeOffFields}. For the class of CQ theories sourced by either the $D_{1,i}^{0 \mu}$ term, or the $D_{1,i}^{\alpha \beta}$ this gives a
lower bound for the diffusion of the conjugate momentum in terms of the mass density of the particle and its decoherence rate $\lambda$
\begin{equation}\label{eq: piDispersionmain}
\frac{d\sigma^2_{\pi_\Phi}(x)}{dt} \geq \frac{ |\langle m(x) \rangle|^2}{ 8 \lambda},
\end{equation}
which leads to an estimate of the production rate of stochastic waves in terms of their gravitational kinetic energy. This can be lower bounded in terms of experimentally decoherence rates
\begin{equation}
\frac{d \Delta \bar{E}}{dt} \geq \int d^3 x \frac{ c^2 G \pi |\langle m(x) \rangle|^2}{12 \lambda } .
\label{eq: diffusion rate trade-offmain}
\end{equation}
The diffusion is akin to the stochastic production of gravitational waves, but these waves need not be transverse (see Appendix \ref{sec: grav diffusion}).
One advantage of studying this regime, is that Equation \eqref{eq: diffusion rate trade-offmain} is a bound which holds very generally, and which is independent of the choice of kernel, since the kinetic energy rate is lower bounded in terms of a experimentally measured decoherence rate.
Since fullerene interference experiments require $\frac{d \Delta \bar{E}}{dt} \sim 10^{-19} Js^{-1}$ over the wave-packet size of nucleons, this implies that the energy density of waves must be produced at a rate of at least $ \sim 10^{5} Js^{-1}m^{-3}$. If this was produced over all space, as required by the diffusion kernel $D_2(x,x')=D_2\delta(x,x')$ (already ruled out by precision Cavendish experiments), this gives an apparent energy density in the ball park of $10^{22}J/m^3$ accumulated over the age of the universe. Since the observed expansion rate of the universe puts its energy density at $10^{-9}J/m^3$, this discrepancy would appear to rule out continuous realisations of classical gravity with the diffusion kernel $\delta(x,x')$. This can be seen as a specific instance of diffusion kernels $D_2(x,x')$ which diverge as $x\rightarrow x'$, something we comment on in Section \ref{app: examplesOfKernals}. However, even here, care should be taken, partly because our estimate is non-relativistic\footnote{We find that extrapolating too far into the past runs afoul of the gauge fixing condition used to derive the Newtonian limit.}, and partly because our understanding of cosmology requires some degree of modesty -- after all, quantum field theory predicts an energy density of $10^{113} J/m^3$, and yet we do not see this reflected in the acceleration of the universe.
We leave a detailed study of the effect of gravitational diffusion on LIGO to future work. It suffices to mention that the effect will again depend on the form of the kernel $D_2(x,x')$. Our estimates \cite{UCLLigo} suggest that local effects from table-top experiments currently place stronger bound on gravitational theories than LIGO currently does. In particular, unlike for gravitational wave measurements, which are reasonably high frequency events requiring extraordinary high precision in relative displacement of the arm length from its average, it is preferential to have a lower precision measurement, but which occurs over a longer time period to allow for the diffusion in path length to build up, and with a smaller uncertainty in the average length of the arm itself. Furthermore, since the LIGO arm is kept in a vacuum, we do not expect strong bounds on discrete models where the diffusion is associated to an energy density.
\section{Discussion}\label{sec: conclusion}
A number of direct proposals to test the quantum nature of gravity are expected to come online in the next decade or two. These are based on the detection of entanglement between mesoscopic masses inside matter-wave interferometers~\cite{kafri2013noise,kafri2015bounds,bose2017spin,PhysRevLett.119.240402,Marshman_2020,pedernales2021enhancing,carney2021testing,christodoulou2022locally}. For these experiments, some theoretical assumptions are needed: one requires that it is only gravitons which travel between the two masses and mediate the creation of entanglement. If this is the case, then the onset of entanglement implies that gravity is not a classical field. These can be thought of as experiments which if successful, would confirm the quantum nature of gravity (although other alternatives to quantum theory are possible \cite{galley2021nogo}).
Here, we come from the other direction, by supposing that gravity is instead classical, and then exploring the consequences.
Theories in which gravity is fundamentally classical were thought to have been ruled out by various no-go theorems and conceptual difficulties. However, these no-go theorems are avoided if one allows for non-deterministic coupling as in \cite{blanchard1995event,diosi1995quantum,diosi2011gravity,alicki2003completely,poulinKITP2,oppenheim_post-quantum_2018,unrav,UCLconstraints,UCLPawula}. We have here proven that this feature is indeed necessary, and made it quantitative by exploring the consequences of complete positivity on any dynamics which couples quantum and classical degrees of freedom. Complete positivity is required to ensure the probabilities of measurement outcomes remain positive throughout the dynamics. We have shown that any theory which preserves probabilities and treats one system classically, is required to have fundamental decoherence of the quantum system, and diffusion in phase space, both of which are signatures of information loss. Using a CQ version of the Kramers-Moyal expansion, we have derived a trade-off between decoherence on the quantum system, and the system's diffusion in phase space. The trade-off is expressed in terms of the strength of the back-reaction of the quantum system on the classical one. We have derived the trade-off both in terms of coupling constants of the theory, and in terms of observational quantities that can be measured experimentally.
In the case of gravity, the observational trade-off places a lower bound on the rate of diffusion of the gravitational degrees of freedom as expressed by Equation \eqref{eq: piDispersionmain} in terms of the decoherence rate of particles in superposition. We find that theories which treat gravity as fundamentally classical, are not ruled out by current experiments, however we have been able to rule out a broad parameter space of such theories. This is done partly through table-top observations via Equations \eqref{eq: contLocal}, \eqref{eq: disLocal} and \eqref{eq: contDP}. Given any diffusion kernel, we can compute the inaccuracy of mass measurements due to fluctuations in the gravitational field, and using the trade-off, we can derive a bound on the associated decoherence rate. This allows us to rule out broad classes of theories in terms of their diffusion kernel. For example, we are able to rule out a number of theories which are continuous in phase space. Then, using the trade-off of Equation \eqref{eq: diffusion rate trade-offmain}, we saw that there was some tension with cosmological observations and kernels such as that of Equation \eqref{eq: Xlocal} and \eqref{eq: lapbelt}, which produce diffusion over all space. However, we are not confident enough in our understanding of cosmology in CQ theories to rule these out.
Any theory which treats gravity classically has fairly limited freedom to evade the effects of the trade-off. There is freedom to choose the diffusion or decoherence kernels $D_2(x,x')$ and $D_0(x,x')$, but the trade-off restricts one in terms of the other. Then, because of the results proven in \cite{UCLPawula}, one can consider two classes of theory, those which are continuous realisations and whose diffusion can only depend on the gravitational degrees of freedom, and discrete theories whose diffusion can also depend directly on the matter fields. Examples of both classes of theory are given in Appendix \ref{sec: cqlimit}.
Finally, one could consider theories which do not reproduce the weak field limit of General Relativity to all distances, namely we could imagine that the interaction Hamiltonian of Equation \eqref{eq:interaction_ham} does not hold to arbitrarily short distances, or arbitrarily high mass densities. This would correspond to modifying $D_1(x,x')$ in some way, either by making it slightly non-local, or by disallowing arbitrarily high mass densities, or by including an additional contribution such as the friction term discussed in \ref{ss:continuous}. All of these modifications would seem to violate Lorentz invariance in some way\footnote{This may only be a concern if it results in any inconsistency with low energy observations, since a theory of quantum gravity would also likely have an anomaly at the Planck scale.}.
Here, we have only given an order of magnitude estimate of when gravitational diffusion will lead to appreciable deviations from Newtonian gravity or General Relativity. We have done so in a number of regimes. The most promising being table-top experiments which precisely measure the mass of an object. This is an area which is important from the perspective of weight standards, for example those undertaken by NIST on the 1kg mass standard K20 and K4 \cite{abbott2019mass}. The increased precision and measuring time of Kibble Balances \cite{Chao2019Design} and atomic interferometers \cite{Chou1630,PhysRevLett.118.183602,peters2001high,menoret2018gravity} would make such measurements an ideal testing ground, both to further constrain the diffusion kernel, and to look for diffusion effects, whose dependence on the test mass is outlined in Appendix \ref{sec: grav diffusion}. Here, we have found that the time $T$ over which results of the measurement are made, affects the strength of the bound, and it would be helpful if future experiments reported this value. Since we have found that CQ theories predict an uncertainty in mass measurements it is perhaps intriguing that different experiments to measure Newton's constant $G$ yield results whose relative uncertainty differ by as much as $5 \cdot10^{-4}$ m$^3$kg$^{-1}$s$^{-2}$, which is more than an order of magnitude larger than the average reported uncertainty \cite{quinn2000measuring,gillies2014attracting,rothleitner2017invited}. If one were to try and explain the discrepancy in $G$ measurements via gravitational diffusion, then for all the kernels we studied in Section \ref{sec: gravity} we find that the variation in acceleration depends on $\frac{1}{\sqrt{N}}$ the number of nucleons in the test mass, so that masses with smaller volume should yield larger uncertainty and this would be the effect to look for in measurement discrepancies. The relatively large uncertainty in such measurements, also makes it challenging for table-top experiments to place strong upper bounds on gravitational diffusion.
We have also estimated the effect that this diffusion would have on the energy density of the universe, and in the production of stochastic waves in terms of gravitational kinetic energy in the weak field limit. We have found that spatially uncorrelated and continuous realisations of classical gravity which reproduce General Relativity at short distances, appear to be ruled out by cosmological considerations as well, since the energy density of the stochastic wave contribution is high enough that it should effect the expansion rate of the universe. However, this is a regime where we do not understand the theory well, and so we are cautious about making too strong a claim. We have also found that the stochastic production of gravitational kinetic energy waves is in a regime which could be detectable by LIGO, an effect which constrains the form of $D_2(x,x')$. However, initial estimates suggest that this is less constraining than table top experiments. For this to be definitive, a more precise understanding of gauge artifacts and of the dynamics that the diffusion induces on the Newtonian potential is required, especially over longer time scales.
How this diffusion might effect dynamics over galactic scales and longer times, requires a fuller General Relativistic treatment, a study which we undertake in \cite{UCLDMDNE} where we find that it causes deviations from what general relativity predicts.
Turning to the other side of the trade-off, improved decoherence times would further squeeze theories in which gravity remains classical. While a current experimental challenge is to demonstrate interference patterns using larger and larger mass particles, we here find that some of our bounds depend on the expectation of the particle's mass density, either in terms of $\langle m^2(x)\rangle/\lambda$, or in ways which depend on the particular kernel. Thus interference experiments with particles of high {\it mass density} rather than mass can be preferable. There are also kernels, for which the relevant quantity is the expectation of the mass density, which will depend on the size of the wave-packet used in the interference experiment, a quantity which is rarely obtainable from most papers which report on such experiments. While this dependence might initially appear counter-intuitive, it follows from the fact that in order to relate the trade-off in terms of coupling constants to observational quantities, and in particular, the decoherence rate, we took expectation values of the relevant quantities to get a trade-off in terms of only averages. And indeed the decoherence rate, which is an expectation value, can easily depend on the wave-packet density, as we see from examples is Section \ref{sec: decoherenceRatesAppendix}.
Since we here show that all theories which treat gravity classically necessarily decohere the quantum system, another constraint on theories which treat gravity classically is given by constraints on fundamental decoherence. These are usually constrained by bounds on anomalous heating of the quantum system \cite{bps}. However, these constraints are not in themselves very strong, since fundamental decoherence effects can be made arbitrarily weak.
In the simplified model in Appendix \ref{sec: cqlimit}, the strength of the decoherence depends on the strength of the gravitational field, thus, constraints due to heating \cite{bps,ghirardi1986unified,ballentine1991failure,pearle1999csl,bassi2005energy,adler2007lower,lochan2012constraining,nimmrichter2014optomechanical,bahrami2014testing,laloe2014heating,bahrami2014proposal,goldwater2016testing,tilloy2019neutron,donadi2020underground} can be suppressed, either by scaling the Lindbladian coupling constants, or by having strong decoherence effects more pronounced near stronger gravitational fields such as near black holes where one expects information loss to occur. The necessity for decoherence to heat the quantum system is further weaked by the fact that the dynamics are not Markovian on the quantum fields, if one integrates out the classical degrees of freedom, space-time acts as a memory. This potentially captures some of the non-Markovian features advocated in \cite{unruh-wald-onbps}, who recognised that Markovianity is a key assumption in attempts to rule out fundamental decoherence or information loss. Here however, we see that there is less freedom than one might imagine. If the Lindbladian coupling constants are made small to reduce heating, the gravitational diffusion must be large. Thus, heating constraints which place bounds on $D_0(x,x')$ place additional constraints on $D_2(x,x')$.
While the absence of diffusion could rule out theories where gravity is fundamentally classical, the presence of such deviations, at least on short time scales, might not by itself be a confirmation of the classical nature of gravity. Such effects could instead be caused by quantum theories of gravity whose classical limit is effectively described by \cite{oppenheim_post-quantum_2018} or perhaps \cite{hu2008stochastic}. In other words, one might expect some gravitational diffusion, because from an effective theory point of view, one is in a regime where space-time is behaving classically. However, the
trade-off we have derived is a direct consequence of treating the background space-time as fundamentally classical. In a fully quantum theory of gravity, the interaction of the gravitational field with particles in a superposition of two trajectories will cause decoherence, but coherence can then be restored when the two trajectories converge. This is what happens when electrons interact with the electromagnetic field while passing through a diffraction grating, yet still form an interference pattern at the screen. This is a non-Markovian effect,
and the trade-off we derived is a direct consequence of the positivity condition, which is a direct consequence of the Markovian assumption. In the non-Markovian theory where General Relativity is treated classically, one still expects the master equation to take the form found in \cite{oppenheim_post-quantum_2018}, but without the matrix whose elements are $D_n^{\mu\nu}$ needing to be positive semi-definite \cite{Hall2014, Breuer2016}.
\section*{Acknowledgements}
We would like thank Sougato Bose, Joan Camps, Matt Headrick, Isaac Layton, Juan Maldacena, Andrea Russo, Andy Svesko and Bill Unruh for valuable discussions and Lajos Di\'{o}si and Antoine Tilloy for their very helpful comments on an earlier draft of this manuscript
JO is supported by an EPSRC Established Career Fellowship, and a Royal Society Wolfson Merit Award, C.S. and Z.W.D.~acknowledges financial support from EPSRC. This research was supported by the National Science Foundation under Grant No. NSF PHY11-25915 and by the Simons Foundation {\it It from Qubit} Network. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Economic Development, Job Creation and Trade.
\section{Introduction}
When considering the dynamics of composite quantum systems, there are many regimes where one system can be taken to be classical and the other quantum-mechanical. For example, in quantum thermodynamics we often have a quantum system interacting with a large thermal reservoir that can be treated classically, whilst in atomic physics it is common to consider the behaviour of quantum atoms in the presence of classical electromagnetic fields. Things become more complicated when one considers classical-quantum (CQ) dynamics where the quantum system back-reacts on the classical system. This is particularly relevant in gravity, because we would like to study the back-reaction of thermal radiation being emitted from black holes on space-time, and while the matter fields can be described by quantum field theory, we only know how to treat space-time classically. Likewise in cosmology, vacuum fluctuations are a quantum effect which we believe seeds galaxy formation, while the expanding space-time they live on can only be treated classically. In addition to the need for an effective theory which treats space-time in the classical limit, there has long been a debate about whether one should quantise gravity \cite{cecile2011role,Feynman:1996kb-note,AharonovParadoxes-note,eppley1977necessity,
unruh1984steps,carlip2008quantum,
Mari:2015qva,Baym3035,belenchia2018quantum,kent2018simple,oppenheim_post-quantum_2018,rydving2021gedanken}.
There has even been much discussion on whether quantum-classical coupling can even be consistent. Many proposals for such dynamics~\cite{aleksandrov199536a,quantum_chemistry} are not completely positive (CP)\footnote{A map $\Lambda$ is completely positive, iff $\mathbbm{1}\otimes\Lambda$ is positive. This is the required condition used to derive the GKSL Equation. If it is violated, the dynamics acting on half of an entangled state, give negative probabilities.}, meaning they are at best an approximation and fail outside a regime of validity \cite{boucher1988semiclassical, diosi2000quantum}. The semi-classical Einstein's equation \cite{moller1962theories,rosenfeld1963quantization}, which replaces the quantum operator corresponding to the stress-energy tensor by its expectation value, is another attempt to treat the classical limit from an effective point of view, but it is non-linear in the state, leading to pathological behavior if quantum fluctuations are of comparable magnitude to the stress-energy tensor \cite{page1981indirect}. This is often the precise regime we would like to understand.
However, dynamics introduced in~\cite{blanchard1995event,diosi1995quantum} and studied in \cite{alicki2003completely, poulinKITP2, oppenheim_post-quantum_2018,unrav} do not suffer from such problems, and lead to consistent dynamics. In particular, the master-equation shown in Equation \eqref{eq: cqdyngen}, is linear, preserves the division of classical degrees of freedom and quantum ones, and is completely positive (CP) and preserves normalisation. This ensures that probabilities of measurement outcomes remain positive and always add to $1$. The dynamics is related to the GKSL or Lindblad equation~\cite{GKS76,Lindblad76}, which for bounded generators of the dynamics, is the most general Markovian dynamics for an open quantum system. Likewise, Equation \eqref{eq: cqdyngen} is the most general Markovian classical-quantum dynamics with bounded generators~\cite{oppenheim_post-quantum_2018}. Sub-classes of this master equation along with meeasurement and feedback approaches have been discussed in the context of Newtonian models of gravity~\cite{diosi2011gravity,Kafri_2014,Kafri_2015,tilloy2016sourcing,Tilloy_2017,tilloy2017principle,poulinKITP2},
and further developed into a spatially covariant framework so that Einstein gravity in the ADM formalism \cite{arnowitt2008republication}
emerges as a limiting case \cite{oppenheim_post-quantum_2018,UCLconstraints}.
In this work, we move away from specific realisations of CQ dynamics, in order to discuss their common features and the experimental signatures that follow from this. An early precursor to the discussion here, is the insight of Di\'{o}si \cite{diosi1995quantum} who added classical noise and quantum decoherence to the master equation of \cite{aleksandrov199536a}, and found the noise and decoherence trade-off required for the dynamics to become completely positive. Here we prove that the phenomena found in~\cite{diosi1995quantum,oppenheim_post-quantum_2018,unrav} are generic features of all CQ dynamics; the classical-quantum interaction necessarily induces decoherence on the quantum system, and there is a generic trade-off between the rate of decoherence and the amount of diffusion in the classical phase space. The stronger the interaction between the quantum system and the classical one, the greater the trade-off. One cannot have quantum systems with long-coherence times without inducing a lot of diffusion in the classical system. One can also generalise this result to a trade-off between the rate of diffusion and the strength of more general couplings to Lindblad operators, with decoherence being a special case. This is expressed as Equations \eqref{eq: decdifnonham} and \eqref{eq: generalTradeOffCouplingConstants}, which bounds the product of diffusion coefficients and Lindblad coupling constants in terms of the strength of the CQ-interaction. It is precisely this trade-off which allows the theories considered here, to evade the no-go arguments of Feynmann \cite{cecile2011role,Feynman:1996kb-note}, Aharonov \cite{AharonovParadoxes-note}, Eppley and Hannah \cite{eppley1977necessity} and others \cite{bohr1933on,cecile2011role,dewitt1962definition,boucher1988semiclassical,diosi2000quantum,gisin1989stochastic,Mari:2015qva,Baym3035,belenchia2018quantum,caro1999impediments, salcedo1996absence, sahoo2004mixing, terno2006inconsistency, barcelo2012hybrid,marletto2017we}. The essence of arguments against quantum-classical interactions is that they would prohibit superpositions of quantum systems which source a classical field. Since different classical fields are perfectly distinguishable in principle, if the classical field is in a distinct state for each quantum state in the superposition, the classical field could always be used to determine the state of the quantum system, causing it to decohere instantly. By satisfying the trade-off, the quantum system preserves coherence because diffusion of the classical degrees of freedom mean that the state of the classical field does not determine the state of the quantum system \cite{poulinKITP2,oppenheim_post-quantum_2018}. Equation \eqref{eq: decdifnonham} and other variants we derive, quantify the amount of diffusion required to preserve any amount of coherence.
If space-time curvature is treated classically, then complete positivity of the dynamics means its interaction with quantum fields \textit{necessarily} results in unpredictability and gravitationally induced decoherence.
This trade-off between the decoherence rate and diffusion provides an experimental signature, not only of models of hybrid Newtonian dynamics such as \cite{diosi2011gravity} or post-quantum theories of General Relativity such as \cite{oppenheim_post-quantum_2018} but of any theory which treats gravity as being fundamentally classical. The metric and their conjugate momenta necessarily diffuse away from what Einstein's General Relativity predicts. This experimental signature squeezes classical-quantum theories of gravity from both sides: if one has shorter decoherence times for superpositions of different mass distributions, one necessarily has more diffusion of the metric and conjugate momenta. In Appendix \ref{sec: grav diffusion} we show that the latter effect causes imprecision in measurements of mass such as those undertaken in the Cavendish experiment \cite{cavendish1798xxi,luther1982redetermination,Gundlach2000} or in measurements of Newton's constant ``Big G"\cite{quinn2000measuring,gillies2014attracting,rothleitner2017invited}.
The precision at which a mass can be measured in a short time, thus provides an upper bound on the amount of gravitational diffusion, as quantified by Equation \eqref{eq: variationForceMain}, while decoherence experiments place a lower bound on the diffusion. Our estimates suggest that experimental lower bounds on the coherence time of large molecules \cite{arndt1999wave,Hornberger,Juffman, Juffmandebroglie, Gerlich2011, BassiCollapse}, combined with gravitational experiments measuring the acceleration of small masses \cite{westphal2020measurement,Schm_le_2016,PhysRevLett.124.101101}, already place strong restrictions on theories where space time isn't quantised.
In Section \ref{sec: gravity} we show that several realisations of CQ-gravity are already ruled out, while other realisations produce enough diffusion away from General Relativity to be detectable by future table-top experiments.
Although the absence of such deviations from General Relativity would not be as direct a confirmation of the quantum nature of gravity, such as experiments proposed in~\cite{kafri2013noise,kafri2015bounds,bose2017spin,PhysRevLett.119.240402,Marshman_2020,pedernales2021enhancing,carney2021testing,kent2021non,christodoulou2022locally} to exhibit entanglement generated by gravitons, it would effectively rule out any sensible theory which treats space-time classically. While confirmation of gravitational diffusion would suggest that space-time is fundamentally classical.
The outline of this paper is as follows. In Sec. \ref{sec: cqdnamics} we review the general form of the CQ master equation of classical-quantum systems. The CQ-map can be represented in a manner akin to the Kraus-representation \cite{Kraus} for quantum maps, with conditions for it to be Completely Positive and Trace Preserving (CPTP). We can perform a short time moment expansion of the {\it CQ-map}
taking states at some initial time, to states at a later time.
This gives us the
CQ version of the Kramers-Moyal expansion~\cite{kramers1940brownian,moyal1949stochastic}. The physical meaning of the moments is discussed in Subsection \ref{ss:momentsmeaning}.
In Sec. \ref{sec: tradeoff} we show that there is a general trade-off between decoherence of the quantum system and diffusion in the classical system. We generalize the trade-off to the case of fields in Section \ref{sec: fieldTradeOff} and in Subsection \ref{sec: gravity}, we apply the inequality in the gravitational setting.
The positivity constraints mean that the considerations do not depend on the specifics of the theory, only that it treats gravity classically, and be Markovian. This allows us to discuss some of the observational implications of this result
and we comment on the relevant figures of merit required in interference and precision mass measurements in order to constrain theories of gravity, as they are not always readily available in published reports. In addition to table-top constraints, we consider those due to cosmological observations.
We then conclude with a discussion of our results in Sec. \ref{sec: conclusion}. The Appendix collects or previews a number of technical results.
\section{Classical-Quantum dynamics} \label{sec: cqdnamics}
Let us first review the general map and master equation governing classical-quantum dynamics. The classical degrees of freedom are described by a differential manifold $\mathcal{M}$ and we shall generically denote elements of the classical space by $z$. For example, we could take the classical degrees of freedom to be position and momenta in which case $\mathcal{M}= \mathbb{R}^{2}$ and $z= (q,p)$. The quantum degrees of freedom are described by a Hilbert space $\mathcal{H}$. Given the Hilbert space, we denote the set of positive semi-definite operators with trace at most unity as $S_{\leq 1}(\mathcal{H})$. Then the CQ object defining the state of the CQ system at a given time is a map $\varrho : \mathcal{M} \to S_{\leq 1}(\mathcal{H})$ subject to a normalization constraint $\int_{\mathcal{M}} dz \Tr{\mathcal{H}}{\varrho} =1$. To put it differently, we associate to each classical degree of freedom a sub-normalized density operator, $\varrho(z)$, such that $\Tr{\mathcal{H}}{\varrho} = p(z) \geq 0$ is a normalized probability distribution over the classical degrees of freedom and $\int_{\mathcal{M}} dz \varrho(z) $ is a normalized density operator on $\mathcal{H}$. An example of such a {\it CQ-state} is the CQ qubit depicted as a $2\times 2$ matrix over phase space \cite{unrav}. More generally, we can define any {\it CQ operator} $f(z)$ which lives in the fibre bundle with base space $\mathcal{M}$ and fibre $\mathcal{H}$.
Just as the Lindblad equation is the most general evolution law which maps density matrices to density matrices, we can ask, what is the most general evolution law, which preserves the quantum-classical state-space. Any such dynamics, if it is to preserve probabilities, must be completely positive, norm preserving, and linear in the CQ-state\footnote{That dynamics must be linear can be seen as follows: if someone prepares a system in one of two states $\sigma_0$ or $\sigma_1$ depending on the value of a coin toss ($\proj{0}$ with probability $p$, $\proj{1}$ with probability $1-p$), then the evolution ${\cal L}$ of the system must satisfy $p\proj{0}\otimes {\cal L}\sigma_0 + (1-p)\proj{1}\otimes{\cal L}\sigma_1 = {\cal L}(p\sigma_0 +(1-p)\sigma_1)$ otherwise the system evolves differently depending on whether we are aware of the value of the coin toss. A violation of linearity further implies that when the system is in state $\sigma_0$ it evolves differently depending on what state the system would have been prepared in, had the coin been $\proj{1}$ instead of $\proj{0}$. This motivates our restriction to linear theories.}. We will also require the map to be Markovian on the combined classical-quantum system, which is equivalent to requiring that there is no hidden system which acts as a memory. This is natural if the interaction is taken to be fundamental, but is the assumption which one might want to remove if one thinks of the hybrid theory as an effective description. We thus take these as the minimal requirements that any fundamental classical-quantum theory must satisfy if it is to be consistent.
The most general CQ-dynamics, which maps CQ states onto themselves can be written in the form \cite{oppenheim_post-quantum_2018}
\begin{equation}\label{eq: CPmap}
\varrho(z,t+ \delta t) = \int dz'\sum_{\mu \nu}\Lambda^{\mu\nu}(z|z',\delta t) L_{\mu} \varrho(z',t) L_{\nu}^{\dag}
\end{equation}
where the $L_{\mu}$ are an orthogonal basis of operators and $\Lambda^{\mu\nu}(z|z',\delta t)$ is positive semi-definite for each $z,z'$. Henceforth, we will adopt the Einstein summation convention so that we can drop $\sum_{\mu\nu}$ with the understanding that equal upper and lower indices are presumed to be summed over.
The normalization of probabilities requires
\begin{equation}\label{eq: prob}
\int dz \Lambda^{\mu\nu}(z|z',\delta t) L_{\nu}^{\dag}L_{\mu} =\mathbb{I}.
\end{equation}
The choice of basis $L_\mu$ is arbitrary, although there may be one which allows for unique trajectories \cite{unrav}.
Equation \eqref{eq: CPmap} can be viewed as a generalisation of the Kraus decomposition theorem.
In the case where the classical degrees of freedom are taken to be discrete, Poulin~\cite{PoulinPC2} used the diagonal form of this map to derive the most general form of Markovian master equation for bounded operators, which is the one introduced in~\cite{blanchard1995event}.
When the classical degrees of freedom are taken to live in a continuous configuration space, we need to be a little more careful, since $\varrho(z)$ may only be defined in a distributional sense; for example, $\varrho(z) = \delta(z, \bar{z}) \varrho( \bar{z})$. In this case \eqref{eq: CPmap} is completely positive if the eigenvalues of $\Lambda^{\mu \nu}(z|z',\delta t)$, $\lambda^{\mu}(z|z',\delta t)$, are positive so that $\int dz dz' P_{\mu}(z,z')\lambda^{\mu}(z|z',\delta t) \geq 0$ for any vector with positive components $P_{\mu}(z,z')$ \cite{UCLPawula}.
One can derive the CQ master equation by performing a short time expansion of \eqref{eq: CPmap} in the case when the $L_\mu$ are bounded \cite{oppenheim_post-quantum_2018}. To do so, we first introduce an arbitrary basis of traceless Lindblad operators on the Hilbert space, $L_{\mu} = \{I, L_{\alpha}\}$. Now, at $\delta t=0$ we know \eqref{eq: CPmap} is the identity map, which tells us that $\Lambda^{00} (z|z', \delta t=0) = \delta(z,z')$. Looking at the short time expansion coefficients, by Taylor expanding in $\delta t \ll 1$, we can write
\begin{align}\label{eq: shortime}
&\Lambda^{\mu \nu}(z|z',\delta t) = \delta^{\mu}_{0} \delta^{ \nu}_{0} \delta(z,z') + W^{\mu \nu}(z|z') \delta t + O(\delta t^2).
\end{align}
By substituting the short time expansion coefficients into \eqref{eq: CPmap} and taking the limit $\delta t \to 0$ we can write the master equation in the form
\begin{align}
\frac{\partial \varrho(z,t)}{\partial t} = \int dz' \ W^{\mu \nu}(z|z') L_{\mu} \varrho(z') L_{\nu}^{\dag} - \frac{1}{2}W^{\mu \nu}(z) \{ L_{\nu}^{\dag} L_{\mu}, \varrho \}_+,
\label{eq: cqdyngen}
\end{align}
where $\{, \}_+$ is the anti-commutator, and preservation of normalisation under the trace and $\int dz$ defines
\begin{equation}
W^{\mu \nu}(z) = \int \mathrm{d} z' W^{\mu \nu}(z'| z).
\end{equation}
We see the CQ master equation is a natural generalisation of the Lindblad equation and classical rate equation in the case of classical-quantum coupling. We give a more precise interpretation of the different terms arising when we perform the Kramers-Moyal expansion of the master equation at the end of the section. The positivity conditions from \eqref{eq: CPmap} transfer to positivity conditions on the master equation via \eqref{eq: shortime}. We can write the positivity conditions in an illuminating form by writing the short time expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z',\delta t)$, as defined by equation \eqref{eq: shortime}, in block form
\begin{equation} \label{eq: block1}
\Lambda^{\mu \nu}(z|z',\delta t)
= \begin{bmatrix}
\delta(z,z') + \delta t W^{00}(z|z') & \delta t W^{0\beta}(z|z') \\
\delta t W^{\alpha 0}(z|z') & \delta t W^{\alpha \beta} (z|z') \\
\end{bmatrix} + O(\delta t^2)
\end{equation}
and the dynamics will be positive if and only if $\Lambda^{\mu \nu}(z|z',\delta t)$ is a positive matrix. It is possible to introduce an arbitrary set of Lindblad operators $\bar{L}_{\mu}$ and appropriately redefine the couplings $W^{\mu \nu}(z|z')$ in \eqref{eq: cqdyngen} \cite{oppenheim_post-quantum_2018}. For most purposes, we shall work with a set of Lindblad operators which includes the identity $L_{\mu} = (I, L_{\alpha})$; this is sufficient since any CQ master equation is completely positive if and only if it can be brought to the form in \eqref{eq: cqdyngen}, where the matrix \eqref{eq: block1} is positive.
\subsection{The CQ Kramers-Moyal expansion}\label{sec: kramersMoyal}
In order to study the positivity conditions it is first useful to perform a moment expansion of the dynamics in a classical-quantum version of the Kramers-Moyal expansion~\cite{oppenheim_post-quantum_2018}. In classical Markovian dynamics, the Kramers-Moyal expansion relates the master equation to the moments of the probability transition amplitude and proves to be useful for a multitude of reasons. Firstly, the moments are related to observable quantities; for example, the first and second moments of the probability transition amplitude characterize the amount of drift and diffusion in the system. This is reviewed in Subsection \ref{ss:momentsmeaning}. Secondly, the positivity conditions on the master equation transfer naturally to positivity conditions on the moments, which we can then relate to observable quantities. In the classical-quantum case, we shall perform a short time moment expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z',\delta t) $ and then show that the master equation can be written in terms of these moments. We then relate the moments to observational quantities, such as the decoherence of the quantum system and the diffusion in the classical system.
We work with the form of the dynamics in \eqref{eq: cqdyngen}, using an arbitrary orthogonal basis of Lindblad operators $L_{\mu} = \{\mathbb{I}, L_{\alpha}\}$. We take the classical degrees of freedom $\mathcal{M}$ to be $d$ dimensional, $z= (z_1, \dots z_d)$, and we label the components as $z_i$, $i \in \{1, \dots d\}$. We begin by introducing the moments of the transition amplitude $W^{\mu \nu}(z|z')$ appearing in the CQ master equation \eqref{eq: shortime}
\begin{equation}
D^{\mu \nu}_{n, i_1 \dots i_n}(z'):= \frac{1}{n!}\int dz W^{\mu \nu}(z|z')(z-z')_{i_1} \dots (z-z')_{i_n} .
\end{equation}
The subscripts $i_j \in \{1, \dots d\}$ label the different components of the vectors $(z-z')$. For example, in the case where $d=2$ and the classical degrees of freedom are position and momenta of a particle, $z=(z_1,z_2) = (q,p)$, then we have $(z-z') = (z_1- z'_1, z_2 - z_2') = ( q-q',p-p')$. The components are then given by $(z-z')_1 = (q-q')$ and $(z-z')_2 = (p-p')$. $M^{\mu \nu}_{n, i_1 \dots i_n} (z',\delta t)$ is seen to be an $n$'th rank tensor with $d^n$ components.
In terms of the components $ D^{\mu \nu}_{n, i_1 \dots i_n}$ the short time expansion of the transition amplitude $\Lambda^{\mu \nu}(z|z')$ is given by \cite{UCLPawula}
\begin{equation}\label{eq: shortimeCP}
\Lambda^{\mu \nu}(z|z',\delta t) = \delta_0^{\mu} \delta_0^{\nu} \delta(z,z') + \delta t \sum_{n=0}^{\infty}D^{\mu \nu}_{n, i_1 \dots i_n}(z') \left(\frac{\partial^n }{\partial z_{i_1}' \dots \partial z_{i_n}' }\right) \delta(z,z') + O(\delta t^2),
\end{equation}
and the master equation takes the form \cite{oppenheim_post-quantum_2018}
\begin{align}\label{eq: expansion}\nonumber
\frac{\partial \varrho(z,t)}{\partial t} &= \sum_{n=1}^{\infty}(-1)^n \left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right) \left( D^{00}_{n, i_1 \dots i_n}(z,\delta t) \varrho(z,t) \right)\\ \nonumber
& -i [H(z),\varrho(z) ] + D_0^{\alpha \beta}(z) L_{\alpha} \varrho(z) L_{\beta}^{\dag} - \frac{1}{2} D_0^{\alpha \beta} \{ L_{\beta}^{\dag} L_{\alpha}, \varrho(z) \}_+ \\
& + \sum_{\mu \nu \neq 00} \sum_{n=1}^{\infty}(-1)^n \left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right)\left( D^{\mu \nu}_{n, i_1 \dots i_n}(z) L_{\mu} \varrho(z,t) L_{\nu}^{\dag} \right),
\end{align}
where we define the Hermitian operator $H(z)= \frac{i}{2}(D^{\mu 0}_0 L_{\mu} - D^{0 \mu}_0 L_{\mu}^{\dag}) $ (which is Hermitian since $D^{\mu 0}_0 = D^{0 \mu *}_0$). We see the first line of \eqref{eq: expansion} describes purely classical dynamics, and is fully described by the moments of the identity component of the dynamics $\Lambda^{00}(z|z')$. The second line describes pure quantum Lindbladian evolution described by
the zeroth moments of the components $\Lambda^{\alpha 0}(z|z'), \Lambda^{\alpha \beta}(z|z')$; specifically the (block) off diagonals, $D^{\alpha 0}_0(z)$, describe the pure Hamiltonian evolution, whilst the
components $D^{\alpha \beta}_0(z)$ describe the dissipative part of the pure quantum evolution. Note that the Hamiltonian and Lindblad couplings can depend on the classical degrees of freedom so the second line describes action of the classical system on the quantum one. The third line contains the non-trivial classical-quantum back-reaction, where changes in the distribution over phase space are induced and can be accompanied by changes in the quantum state.
\subsection{Physical interpretation of the moments}
\label{ss:momentsmeaning}
Let us now briefly review the physical interpretation of the moments which will appear in our trade-off relation. In particular, the zeroth moment determines the rate of decoherence (and Lindbladian coupling more generally), the first moment gives the force exerted by the quantum system on the classical system, and the second moment determines the diffusion of the classical degrees of freedom. For this discussion we shall take the classical degrees of freedom to live in a phase space $\Gamma = (\mathcal{M}, \omega)$, where $\omega$ is the symplectic form.
Consider the expectation value of any CQ operator $O(z)$, $\langle O(z)\rangle:=\int dz \tr{O(z)\varrho}$ which doesn't have an explicit time dependence.
Its evolution law can be determined via Equation \eqref{eq: expansion}
\begin{align}
\frac{d\langle O\rangle}{dt}& =\int dz \tr{ O(z)\frac{\partial \varrho}{\partial t}}\nonumber\\
&=
\int dz \mathrm{Tr}\varrho
\left[ -
i [O(z),H(z)] + D_0^{\alpha \beta}(z) L_{\beta}^{\dag}O(z) L_{\alpha} - \frac{1}{2} D_0^{\alpha \beta} \{ L_{\alpha}L_{\beta}^{\dag}, O(z) \}_+
\right.\nonumber \\
& + \left.
\sum_{n=1}^{\infty}
\left(\frac{\partial^n }{\partial z_{i_1} \dots \partial z_{i_n} }\right)\left( D^{\alpha\beta}_{n, i_1 \dots i_n}(z) L_{\beta}^{\dag} L_{\alpha} O(z,t) \right)
\right]
\label{eq Hexpansion2}
\end{align}
where we have used cyclicity of trace and integration by parts, to bring the equation of motion into a form which would enable us to write a CQ version of the {\it Heisenberg representation} \cite{oppenheim_post-quantum_2018} for a CQ operator. If we are interested in the expectation value of phase space variables $O(z)=z_i\mathbbm{I}$ then Equation \eqref{eq Hexpansion2} gives
\begin{align}
\frac{d\langle z_i \rangle}{dt}=
\int dz D^{\mu \nu}_{1,i}\tr{L_\nu^\dagger L_\mu \varrho(z,t)}
\label{eq:semiHamiltonsequation}
\end{align}
with all higher order terms vanishing, and we see that $ \sum_{\mu \nu \neq 00} D^{\mu\nu}_{1,i}\langle L^\dagger_\nu L_\mu\rangle$ governs the average rate at which the quantum system moves the classical system through phase space, and with the back-reaction is quantified by the Hermitian matrix $D_1^{\alpha \mu}:= (D_{1}^{br})^{\alpha \mu}$. The force of this back-reaction is especially apparent if the equations of motion are Hamiltonian in the classical limit as in \cite{oppenheim_post-quantum_2018}. I.e. if we define $H_I(z) := h^{\alpha \beta} L^{\dag}_{\beta} L_{\alpha}$ and take $D^{\alpha\beta}_{1,i}=\omega_i^j d_j h^{\alpha\beta}$ with $\omega$ the symplectic form and $d_j$ the exterior derivative. Then Equation \eqref{eq:semiHamiltonsequation} is analogous to Hamilton's equations,
and the CQ evolution equation after tracing out the quantum system has the form of a Liouville's equation to first order and in the classical limit,
\begin{equation}
\frac{\partial \rho(z,t) }{\partial t} = \{H_c, \rho(z,t) \} + \mathrm{tr}\left(\{H_I(z) , \varrho(z) \}\right) + \dots
\label{eq: limit}
\end{equation}
with $\rho(z):=\tr{\varrho(z)}$.
The significance of the second moment is also seen via Equation \eqref{eq Hexpansion2} to be related to the variance of phase space variables $ \sigma_{z_{i_1}z_{i_2}}:=\langle z_{i_1}z_{i_2}\rangle-\langle z_{i_1}\rangle\langle z_{i_2}\rangle$
\begin{align}
\frac{d\sigma^2_{z_{i_1},z_{i_2}}}{dt}=2\langle D^{\alpha\beta}_{2,i_1,i_2} L^\dagger_\beta L_\alpha \rangle
+\langle z_2 D^{\alpha\beta}_{1,z_{i_1}}L^\dagger_\beta L_\alpha\rangle
-
\langle z_{i_2}\rangle\langle D^{\alpha\beta}_{1,z_{i_1}}L^\dagger_\beta L_\alpha\rangle
+\langle z_{i_1} D^{\alpha\beta}_{1,z_{i_2}} L^\dagger_\beta L_\alpha \rangle
-
\langle z_{i_1}\rangle \langle D^{\alpha\beta}_{1,z_{i_2}}L^\dagger_\beta L_\alpha\rangle
\label{eq:variance}
\end{align}
In the case when $D_{1,z_{i_1}}$ is uncorrelated with $z_{i_2}$ and $D_{1,z_{i_2}}$ uncorrelated with $z_{i_1}$, then the growth of the variance only depends on the diffusion coefficient.
The zeroth moment $D^{\alpha\beta}_0$ is just the pure Lindbladian couplings. The simplest example is the case of a pure decoherence process with a single Hermitian Lindblad operator $L$ and decoherence coupling $D_0$. Then we can define a basis $\{|a \rangle\}$ via the eigenvectors of $L$ and
\begin{align}
\bra{a}\frac{\partial\varrho}{\partial t}\ket{b}
=
-i\bra{a}[H(z),\varrho]\ket{b}-\frac{1}{2}D_0(L(a) - L(b))^2 \langle a| \varrho |b \rangle
\end{align}
and we see that the matrix elements of $\varrho$ which quantify coherence between the states $\ket{a}$,$\ket{b}$ decay exponentially fast with a decay rate of
$D_0(L(a) - L(b))^2$. For a damping/pumping process of a quantum harmonic oscillator with Hamiltonian $H=\omega a^\dagger a$, $L_\downarrow=a$, $L_\uparrow=a^\dagger$, $a$ the creation operator, and $D_0^{\uparrow\uparrow}$, $D_0^{\downarrow\downarrow}$ the non-zero couplings, then standard calculations \cite{breuer2002theory,unrav} show that an initial superposition $\frac{1}{\sqrt{2}}\ket{n+m}$ with $n$,$m$ large and $n\gg m$ will initially decohere at a rate of approximately $(D_0^{\uparrow\uparrow}+D_0^{\downarrow\downarrow})(m+n)/2$, and the state will eventually thermalise to a temperature of $\omega /\log{(D_0^{\downarrow\downarrow}/D_0^{\uparrow\uparrow})}$. So in this case, the Lindblad couplings not only determine the rate of decoherence, but also the rate at which energy is pumped into the harmonic oscillator. In the next section we will derive the trade-off between Lindblad couplings and the diffusion coefficients. Although we will sometimes refer to this as a trade-off between decoherence and diffusion, this terminology is only strictly appropriate for pure decoherence processes, while more generally, it is a trade-off between Lindblad couplings and diffusion coefficients.
\section{A Trade off between decoherence and diffusion} \label{sec: tradeoff}
In this section we use positivity conditions to prove that the trade off between decoherence and diffusion seen in models such as those of \cite{diosi1995quantum,oppenheim_post-quantum_2018,unrav} are in fact a general feature of \textit{all} classical-quantum interactions. We shall also generalise this, and derive a trade-off between diffusion and arbitrary Lindbladian coupling strengths. The trade-off is in relation to the strength of the dynamics and is captured by Equation's \eqref{eq: decdifnonham0}, \eqref{eq: generalTradeOffCouplingConstants} and \eqref{eq: decdifnonham}.
In section \ref{sec: fieldTradeOff} we extend the trade-off to the case where the classical and quantum degrees of freedom can be fields and use this to show that treating the metric as being classical necessarily results in diffusion of the gravitational field.
There are two separate possible sources for the force (or {\it drift}) of the back-reaction of the quantum system on phase space -- it can be sourced by either the $D^{0 \alpha}_{1,i}$ components or the Lindbladian components $D^{\alpha \beta}_{1,i}$. We shall deal with both sources simultaneously by considering a CQ Cauchy-Schwartz inequality which arises from the positivity of
\begin{equation}
\label{eq: observationalpositivitycontinous}
\tr{\int dz dz' \Lambda^{\mu \nu}( z|z') O_{ \mu}(z,z')\rho(z') O_{\nu}^{\dag}(z,z')}\geq 0
\end{equation}
for any vector of CQ operators $O_{\mu}$. One can verify that this must be positive directly from the positivity conditions on $\Lambda^{\mu \nu}(z|z')$ and we go through the details in appendix \ref{sec: positivityCondtionsAppendix}. A common choice for $O_{\mu}$ would be the set of operators $L_{\mu}=\{ \mathbb{I}, L_{\alpha}\}$ appearing in the master equation.
The inequality in Equation \eqref{eq: observationalpositivitycontinous} turns out to be especially useful since it can be used to define a (pseudo) inner product on a vector of operators with components $O_{\mu}$ via
\begin{equation}\label{eq: innerProduct}
\langle \bar{O}_1,\bar{O}_2 \rangle = \int dz dz' \Tr{}{\Lambda^{\mu \nu}( z|z')O_{1 \mu} \varrho(z') O^{\dag}_{2 \nu} }
\end{equation}
where $||\bar{O}|| = \sqrt{\langle \bar{O}, \bar{O} \rangle} \geq 0$ due to \eqref{eq: observationalpositivitycontinous}. Technically this is not positive definite, but this shall not be important for our purpose. Taking the combination $O_{\mu} = ||\bar{O}_2||^2 O_{1 \mu} - \langle \bar{O}_1, \bar{O}_2 \rangle O_{2 \mu}$ for vectors $O_{1\mu}, O_{2\mu}$, positivity of the norm gives
\begin{equation}
||\bar{O}||^2 = \ ||\bar{O}_2||^2 \bar{O}_1 - \langle \bar{O}_1, \bar{O}_2 \rangle \bar{O}_2 ||^2 = ||\bar{O}_2||^2 \left( ||\bar{O}_1||^2 ||\bar{O}_2||^2 - |\langle \bar{O}_1,\bar{O}_2 \rangle |^2 \right) \geq 0,
\end{equation}
and as long as $||\bar{O}_2|| \neq 0$ we have a Cauchy- Schwartz inequality
\begin{equation}\label{eq: continuousCauchySchwarz}
||\bar{O}_1||^2 ||\bar{O}_2||^2 - |\langle \bar{O}_1,\bar{O}_2 \rangle |^2 \geq 0.
\end{equation}
We can use \eqref{eq: continuousCauchySchwarz} to get a trade-off between the observed diffusion and decoherence by picking $O_{2 \mu} = \delta_{\mu}^{\alpha} L_{\alpha}$ and $O_{1 \mu} = b^i(z-z')_i L_{\mu}$, where $L_{\mu}=\{ \mathbb{I}, L_{\alpha}\}$ are the Lindblad operators appearing in the master equation. In this case $||\bar{O}_2|| = \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho L_{\beta}^{\dag}}$ and one can verify using CQ Pawula theorem \cite{UCLPawula}\footnote{In particular, to reach this conclusion one can insert the CQ state into the CQ Cauchy-Schwartz inequality and repeat the proof of the Pawula theorem \cite{UCLPawula}, which must now hold once averaged over the state.} that in order to have non-trivial back-reaction on the quantum system complete positivity demands that $||\bar{O}_2|| > 0$, meaning the Cauchy-Schwartz inequality in Equation \eqref{eq: continuousCauchySchwarz} must hold. By using the short time moment expansion of $\Lambda^{\mu \nu}(z|z')$ defined in Equation \eqref{eq: shortimeCP} and using integration by parts, we then arrive at the observational trade-off between decoherence and diffusion
\begin{equation}\label{eq: expandedTradeOff}
\int dz \Tr{}{2 b^{i *} D^{\mu \nu}_{2, ij} b^j L_{\mu} \varrho(z) L_{\nu }^{\dag} } \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho(z) L_{\beta}^{\dag}} \geq \left| \int dz \Tr{}{b^i D^{\mu \alpha}_{1,i} L_{\mu} \varrho(z) L_{\alpha}^{\dag} } \right|^2,
\end{equation}
which must hold for any positive CQ state $\varrho(z)$. Stripping out the $b^i$ vectors, \eqref{eq: expandedTradeOff} is equivalent to the matrix positivity condition
\begin{equation}\label{eq: decdifnonham0}
0 \preceq 2 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{br} \rangle \langle D_1^{br} \rangle^{\dag}, \ \ \ \forall \varrho(z),
\end{equation}
where we define
\begin{equation} \label{eq: definitionExpectation}
\langle D_{0} \rangle = \int dz \Tr{}{D^{\alpha \beta}_0 L_{\alpha} \varrho(z) L_{\beta}^{\dag}} , \ \langle D_1^{br} \rangle_i =\int dz \Tr{}{D^{\mu \alpha}_{1,i} L_{\mu} \varrho(z) L_{\alpha}^{\dag} } , \ \langle D_{2} \rangle_{ij} = \int dz \Tr{}{ D^{\mu \nu}_{2, ij} L_{\mu} \varrho(z) L_{\nu }^{\dag} } .
\end{equation}
Since \eqref{eq: decdifnonham0} holds for all states, the tightest bound is provided by the infimum over all states
\begin{equation}
0 \preceq \inf_{\varrho(z)}\{2 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{br} \rangle \langle D_1^{br} \rangle^{\dag}\} .
\end{equation}
The quantities $\langle D_2 \rangle$ and $\langle D_0\rangle$ appearing in Equation \eqref{eq: decdifnonham0} are related to observational quantities. In particular $\langle D_2 \rangle$ is the expectation value of the amount of classical diffusion which is observed and $\langle D_{0} \rangle$ is related to the amount of decoherence on the quantum system. The expectation value of the back-reaction matrix $\langle D_{1}^{br} \rangle $ quantifies the amount of back-reaction on the classical system.
In the trivial case $D_1^{br}=0$, Equation \eqref{eq: decdifnonham0} places little restriction on the diffusion and Lindbladian rates appearing on the left hand side. We already knew from \cite{GKS76,Lindblad76} that the $D_0^{\alpha\beta}$ must be a positive semi-definite matrix, and we also know that diffusion coefficients must be positive semi-definite.
However, in the non-trivial case, the larger the back-reaction exerted by the quantum system, the stronger the trade-off between the diffusion coefficients and Lindbladian coupling.
Equation \eqref{eq: decdifnonham0} gives a general trade-off between observed diffusion and Lindbladian rates, but we can also find a trade-off in terms of a theory's coupling coefficients alone. We show in appendix \ref{sec: diffusionDecoherenceCouplingTradeOff} that the general matrix trade-off
\begin{equation}\label{eq: generalTradeOffCouplingConstants}
\ D_1^{br} D_0^{-1} D_1^{ br \dag} \preceq 2 D_2
\end{equation}
holds for the matrix whose elements are the couplings $D^{\mu \nu}_{2,ij} , D^{\alpha \mu}_{1,i}, D_0^{\alpha \beta}$ for any CQ dynamics. Moreover, $ (\mathbb{I}- D_0 D_0^{-1})D_1^{br} =0$, which tells us that $D_0$ cannot vanish if there is non-zero back-reaction. Equation \eqref{eq: generalTradeOffCouplingConstants} quantifies the required amount of decoherence and diffusion in order for the dynamics to be completely positive. In Equation \eqref{eq: generalTradeOffCouplingConstants}, and throughout, $D_0^{-1}$ is the generalized inverse of $D_0^{\alpha \beta}$, since $D_0^{\alpha \beta}$ is only required to be positive semi-definite. In the special case of a single Lindblad operator $\alpha=1$ and classical degree of freedom, and when the only non-zero couplings are $D_0^{11}:=D_0$, $D_{2,pp}^{00}:=2D_2$ and $D_{1,q}^0=1$ this trade-off reduces to the condition $D_2D_0\geq 1$ used in \cite{diosi1995quantum}.
It is also useful to try to obtain an observational trade-off in terms of the total drift due to back-reaction as calculated in Equation \eqref{eq:semiHamiltonsequation}
\begin{equation}
\langle D_1^{T} \rangle_i =\sum_{\mu \nu \neq 00}\int dz \Tr{}{D^{\mu \nu}_{1,i} L_{\mu} \varrho(z) L_{\nu}^{\dag} }.
\end{equation}
It follows directly from Equation \eqref{eq: decdifnonham0} that when the back-reaction is sourced by either $D_{1,i}^{0 \mu}$ or $D_{1,i}^{\alpha \beta}$ we can arrive at the observational trade-off in terms of the total drift\footnote{We believe that \eqref{eq: decdifnonham} should hold more generally, though we don't have a general proof. }
\begin{equation}\label{eq: decdifnonham}
0 \preceq 8 \langle D_2 \rangle \langle D_0 \rangle -\langle D_1^{T} \rangle \langle D_1^{T} \rangle^{\dag}, \ \ \ \forall \varrho(z),
\end{equation}
where the quantities appearing in Equation \eqref{eq: decdifnonham} are now all observational quantities, related to drift, decoherence and diffusion as outlined in the previous subsection \ref{ss:momentsmeaning}.
In the case where the back-reaction is Hamiltonian at first order in the sense of Equation \eqref{eq: limit}, then \eqref{eq: decdifnonham} can be written as
\begin{equation}
\label{eq: decodiff}
\langle \omega \cdot \frac{\partial H_I}{ \partial \vec{z}} \rangle \langle \omega \cdot \frac{\partial H_I}{ \partial \vec{z}} \rangle^{\dag} \preceq 8 \langle D_{2} \rangle \langle D_0 \rangle, \ \ \ \forall \varrho(z) .
\end{equation}
As a result, we can derive a trade-off between diffusion and decoherence for any theory which reproduces this classical limit and treats one of the systems classically.
To summarize, whenever back-reaction of the quantum system on the classical system induces a force on the phase space, then we have a trade-off between the amount of diffusion on the classical system and the strength of decoherence on the quantum system (or more precisely the strength of the Lindbladian couplings $D_0^{\alpha \beta}$). This can be expressed both as a condition on the matrix of coupling co-efficients in the master equation, via Equation \eqref{eq: generalTradeOffCouplingConstants} or in terms of observable quantities using
Equation's \eqref{eq: decdifnonham0} and \eqref{eq: decdifnonham}. In the case when the back-reaction is Hamiltonian, we further have Equation \eqref{eq: decodiff}. We would like to apply this trade-off to the case of gravity in the non-relativistic, Newtonian limit. In order to do so, we will need to generalise the trade-off to the case of quantum fields interacting with classical one, which we do in Section \ref{sec: fieldTradeOff}. The goal will be to understand the implications of treating the metric (or Newtonian potential) as being classical by using the trade-off when the quantum back-reaction induces a force on the gravitational field which, on expectation, is the same as the weak field limit of general relativity.
\section{Trade off in the presence of fields}\label{sec: fieldTradeOff}
We would like to explore the trade-off in the gravitational setting and explore the consequences of treating the gravitational field as being classical and matter quantum. Since gravity is a field theory, we must first discuss classical-quantum master equations in the presence of fields. In the field theoretic case, both the Lindblad operators and the phase space degrees of freedom can have spatial dependence, $z(x), L_{\mu}(x)$ and a general bounded CP map which preserves the classicality of the two systems can be written \cite{oppenheim_post-quantum_2018}
\begin{equation}\label{eq: mapFields}
\rho(z, t) = \int dz' dx dy\Lambda^{\mu \nu}(z|z',t; x,y) L_{\mu}(x, z, z') \varrho(z',0) L_{\nu}^{\dag}(y, z, z'),
\end{equation}
where, as is usually the case with fields, in Equation \eqref{eq: mapFields} it should be implicitly understood that a smearing procedure has been implemented. We elaborate on some of details when fields are introduced in appendix \ref{sec: cqfieldsAppendix}. The condition for \eqref{eq: mapFields} to be completely positive on \textit{all} CQ states is
\begin{equation}\label{eq: fieldPositivityAll}
\int dz dx dy A_{\mu}^*(x, z, z') \Lambda^{\mu \nu}(z|z'; x,y) A_{\nu}(y, z, z')\geq 0
\end{equation}
meaning that $\Lambda^{\mu \nu}(x,y) $ can be viewed as a positive matrix in $\mu \nu$ and a positive kernel in $x,y$. In the field theoretic case one can still perform a Kramers-Moyal expansion and find a trade-off between the coefficients $D_0(x,y), D_1(x,y), D_2(x,y)$ appearing in the master equation. The coefficients now have an $x,y$ dependence, due to the spatial dependence of the Lindblad operators. The coefficients $D_1(x,y),D_2(x,y)$ still have a natural interpretation as measuring the amount of force ({\it drift}) and diffusion, whilst $D_0(x,y)$ describes the purely quantum evolution on the system and can be related to decoherence.
Using the positivity condition in \eqref{eq: fieldPositivityAll} we find the same trade of between coupling constants in Equation \eqref{eq: generalTradeOffCouplingConstants} but where now $D_2(x,y)$ is the $(p+1)n \times (p+1)n$ matrix-kernel with elements $D_{2,ij}^{\mu \nu}(x,y)$, $D_{1}^{br}(x,y)$ is the $( p+1)n \times p$ matrix-kernel with rows labeled by $\mu i$, columns labelled by $\beta$, and elements $D_{1, i}^{ \mu \beta}(x,y)$, and $D_0(x,y)$ is the $p\times p $ decoherence matrix-kernel with elements $D_0^{\alpha \beta}(x,y)$.\footnote{Here $i \in \{1, \dots, n \}$ $\alpha \in \{1, \dots, p\} $ and $\mu \in \{ 1, \dots, p+1\}$,} In the field theoretic trade off we are treating the objects in Equation \eqref{eq: generalTradeOffCouplingConstants} as matrix-kernels, so that for any position dependent vector $b^i_{\mu}(x)$, $( D_2 b )_{i}^{\mu} (x) = \int dy D^{\mu \nu}_{2,i j}(x,y) b^{j}_{\nu}(y) $, whilst for any position dependent vector $a_{\alpha}(x)$, $( D_0 \alpha )^{\alpha} (x) = \int dy D_{0}^{\alpha \beta}(x,y) \alpha_{\beta}(y) $. Explicitly, we find that positivity of the dynamics is equivalent to the matrix condition
\begin{align} \label{eq: generalFieldMatrixInequalityMain}
\int dx dy [ b^*(x), \alpha^*(x)]
\begin{bmatrix}
2 D_{2}(x,y) & D_1^{br}(x,y)\\ D^{br}_1(x,y) & D_0(x,y)
\end{bmatrix}
\begin{bmatrix}
b(y) \\ \alpha(y)
\end{bmatrix} \geq 0
\end{align}
which should be positive for any position dependent vectors $b^{i}_{\mu}(x)$ and $a_{\alpha}(x)$. This is equivalent to trade-off between coupling constants in Equation \eqref{eq: generalTradeOffCouplingConstants} if we view \eqref{eq: generalTradeOffCouplingConstants} as a matrix-kernel equation.
Though we make no assumption on the locality of the Lindbladian and diffusion couplings, we shall hereby assume that the drift back-reaction is local, so that $D_1^{br}(x,y) = \delta(x,y) D_1^{br}(x)$. As we shall see in the next section, this is a natural assumption if we want to have back-reaction which is given by a local Hamiltonian. However, one might not want to assume that the form of the Hamiltonian remains unchanged to arbitrarily small distances. With this locality assumption, Equation \eqref{eq: generalFieldMatrixInequalityMain} gives rise to the same trade-off of Equation \eqref{eq: generalTradeOffCouplingConstants}, where the trade-off is to be interpreted as a matrix kernel inequality. Writing this out explicitly we have
\begin{align}
\label{eq: generalTradeOffCouplingConstants_Fields_local}
\int dx dy \alpha_{\nu}^{i*}(x)D_{1,i}^{\mu \alpha}(x) (D_0^{-1})_{\alpha \beta}(x,y) D_{1,j}^{\beta \nu}(x')\alpha_{\nu}^i(x') \leq \int dx dy 2\alpha_{\mu}^{i *}(x) D_{2,ij}^{\mu \nu}(x,y) \alpha_{\nu}^j(y),
\end{align}
where asking that this inequality holds for all vectors $\alpha_{\mu}^i(x)$ is equivalent to the matrix-kernel trade-off condition of Equation \eqref{eq: generalTradeOffCouplingConstants}. We give two examples of master Equations satisfying the coupling constant trade-off in appendix \ref{app: examplesOfKernals}. The decoherence-diffusion trade-off tells us how much diffusion and stochasticity is required to maintain coherence when the quantum system back-reacts on the classical one. If the interaction between the classical and quantum degrees of freedom is dictated by unbounded operators, such as the mass density, then there can exist states for which the back-reaction can be made arbitrarily large. This is the case for a quantum particle interacting with its Newtonian potential through its mass density at arbitrarily short distances. Hence, if one considers a particle in a superposition of two peaked mass densities, then there can be an arbitrarily large response in the Newtonian potential around those points, and either there must be an arbitrary amount of diffusion, or the decoherence must occur arbitrarily fast. The former is unphysical, while the latter turns out to be the case in simple examples of theories such as those discussed in Appendix \ref{sec: decoherenceRatesAppendix}.
Since our goal is to experimentally constrain classical-quantum theories of gravity, we shall hereby ask that the map \eqref{eq: mapFields} is CP when acting on all physical states $\rho$. If one allows for arbitrarily peaked mass distributions then the coupling constant trade-off of Equation \eqref{eq: generalTradeOffCouplingConstants_Fields_local} should be satisfied. In the field theoretic case, we can similarly find an observational trade-off, relating the expected value of the diffusion matrix $\langle D_2(x,y) \rangle$ to the expected value of the drift in a physical state $\varrho$ as we did in Section \ref{sec: tradeoff}.
This is done explicitly in Appendix \ref{sec: cqfieldsAppendix}, using a field theoretic version of the Cauchy-Schwartz inequality given by Equation \eqref{eq: continuousCauchySchwarzField}, we find
\begin{equation}\label{eq: observationalTradeOffFields0}
2 \langle D_2(x,x) \rangle \int dx' dy' \langle D_0(x',y') \rangle \succeq \langle D_1^{br}(x) \rangle \langle D_1^{br}(x) \rangle^{\dag},
\end{equation}
where equation \eqref{eq: observationalTradeOffFields} is to be understood as a matrix inequality with entries
\begin{equation}\label{eq: fieldCoeffDefinitions}
\begin{split}
& \langle D_0(x,y)\rangle = \int dz \Tr{}{D_0^{\alpha \beta} L_{\alpha}(x) \varrho L_{\beta}^{\dag}(y)}, \\
& \langle D_1^{br}(x,y)\rangle_i = \int dz \Tr{}{D_{1,i}^{\mu \alpha} L_{\mu}(x) \varrho L_{\alpha}^{\dag}(x)}, \\
& \langle D_2(x,y)\rangle_{ij} = \int dz \Tr{}{D_{2,ij}^{\mu \nu} L_{\alpha}(x) \varrho L_{\beta}^{\dag}(y)}.
\end{split}
\end{equation}
Similarly, when the back-reaction is sourced by either $D_{1,i}^{0 \mu}$ or $D_{1,i}^{\alpha \beta}$ it follows from Equation \eqref{eq: observationalTradeOffFields0} we can arrive at the observational trade-off in terms of the total drift due to back-reaction
\begin{equation}\label{eq: observationalTradeOffFields}
8 \langle D_2(x,x) \rangle \int dx' dy' \langle D_0(x',y') \rangle \succeq \langle D_1^{T}(x) \rangle \langle D_1^{T}(x) \rangle^{\dag},
\end{equation}
where
\begin{equation}\label{eq: totdriftfieldCoeffDefinition}
\langle D_1^T(x)\rangle_i = \int dz \Tr{}{D_{1,i}^{0 \alpha}(x) \varrho L_{\alpha}^{\dag}(x)+ D_{1,i}^{ \alpha 0}(x) L_{\alpha} \varrho (x) + D_{1,i}^{\alpha \beta}(x) L_{\alpha}(x) \varrho L_{\beta}^{\dag}(x)}.
\end{equation}
We shall now use the trade-off to study the consequences of treating the gravitational field classically. We will consider the back-reaction of the mass on the gravitational field to be governed by the Newtonian interaction (or more accurately, a weak field limit of General Relativity). We shall then find that experimental bounds on coherence lifetimes for particles in superposition require large diffusion in the gravitational field in order to be maintained and this can be upper bounded by gravitational experiments.
To summarise this section, we have derived the trade-off between decoherence and diffusion for classical-quantum field theories, both in terms of coupling constants of the theory and in terms of observational quantities. This trade-off puts tight observational constraints on classical theories of gravity which we now discuss.
\section{Physical constraints on the classicality of gravity}
\label{sec: gravity}
In this section we apply the trade-off of Equation \eqref{eq: generalFieldMatrixInequalityMain} to the case of gravity.
Since the trade-offs derived in the previous section depend only on the back-reaction, or drift term, they are insensitive to the particulars of the theory. We shall consider the Newtonian, non-relativistic limit of a classical gravitational field which we reproduce in Appendix \ref{sec: cqlimit}. It is in taking this limit where some care should be taken, since one is gauge fixing the full general relativistic theory. We denote $\Phi$ to be the Newtonian potential and in the weak field limit of General Relativity, it has a conjugate momenta we denote by $\pi_\Phi$. We assume
\begin{enumerate}[label=(\roman*)]
\item The theory satisfies the assumptions used to derive the master equation as in Sections \ref{sec: fieldTradeOff}; in particular that the theory be a completely positive norm-preserving Markovian map, and that we can perform a short-time Kramers-Moyal expansion as in Appendix \ref{sec: cqfieldsAppendix}.
\item We apply the theory to the weak field limit of General Relativity, where as recalled in Appendix \ref{sec: cqlimit} the Newtonian potential interacts with matter through its mass density $m(x)$,
\begin{equation}
H_I(\Phi) = \int d^3 x \Phi(x) m(x).
\label{eq:interaction_ham}
\end{equation}
and the conjugate momentum to $\Phi$ satisfies
\begin{align}
\dot{\pi}_{\Phi} = \frac{\nabla^2 \Phi}{4 \pi G} - m(x)
\label{eq:dotpi}
\end{align}
where in the $c \to \infty$ limit we recover Poisson's equation for the Newtonian potential.
We assume this limit of General Relativity is satisfied on expectation, at least to leading order.
\item In relating $D_0$ to the decoherence rate of a particle in superposition, we shall assume that the state of interest is well approximated by a state living in a Hilbert space of fixed particle number. We believe this is a mild assumption: ordinary non-relativistic quantum mechanics is described via a single particle Hilbert space, and we frequently place composite massive particles in superposition and they do not typically decay into multiple particles.
\item We will assume that the diffusion kernel $ D_2(\Phi,x,x')$ does not depend on $\pi_\Phi$ i.e. it is {\it minimally coupled}. This is reasonable,
since in the purely classical case matter couples to the Newtonian potential and not its conjugate momenta.
\label{ass:minimal}
\end{enumerate}
With these assumptions, and
treating the matter density as a quantum operator $\hat{m}(x)$, this tells us that in order for the back-reaction term to reproduce the Newtonian interaction on average
\begin{align}
\tr{\{H_I,\varrho\}}=
\tr{\int d^3 x \ \hat{m}(x) \frac{\delta \varrho}{\delta \pi_\Phi}} = -\sum_{\mu \nu \neq 00} \tr{\int d^3 x D_{1,\pi_{\Phi}}^{\mu \nu}(\Phi, \pi_{\Phi}, x) L_{\mu}(x) \frac{\delta \varrho}{\delta \pi_\Phi} L_{\nu}^{\dag}(x)},
\label{eq:Nsemi}
\end{align}
then we must pick
\begin{equation}\label{eq: firstMomentGravity}
\langle D_{1,\pi_{\phi}}^{T}(\Phi, \pi_{\Phi}, x) \rangle = - \langle \hat{m}(x) \rangle,
\end{equation}
meaning that the back-reaction matrix $D_{1,\pi_{\Phi}}^{\mu \alpha}$ is non vanishing. In Appendix \ref{sec: cqlimit} we give examples of master equations for which \eqref{eq: firstMomentGravity} is satisfied, but their details are irrelevant since we only require the expectation of the back-reaction force to be the expectation value of the mass -- a necessary condition for the theory to reproduce Newtonian gravity.
As a consequence of the coupling constant and observational trade-offs derived in Equations \eqref{eq: generalTradeOffCouplingConstants_Fields_local} and \eqref{eq: observationalTradeOffFields0}, a non-zero $D_{1,\pi_{\Phi}}$ implies that there must be diffusion in the momenta conjugate to $\pi_{\Phi}$.
This diffusion is equivalent to adding a stochastic random process $J(x,t)$ (the Langevin picture), to the equation of motion \eqref{eq:dotpi} to give
\begin{equation}\label{eq: diffusivePi}
\dot{\pi}_{\Phi} = \frac{\nabla^2 \Phi}{4 \pi G} - m(x) + u(\Phi,\hat{m})
J(t,x),
\end{equation}
where we allow some {\it colouring} to the noise via a function $u(\Phi,\hat{m})$ which can depend on $\Phi$, and the matter distribution $\hat{m}$ (assumption \ref{ass:minimal}). The noise process satisfies
\begin{equation}\label{eq: statsJ}
\mathbb{E}_{m, \Phi}[u J(x,t)] =0, \ \ \ \mathbb{E}_{m, \Phi}[uJ(x,t) u J(y,t') ]= 2\langle D_2(x,y,\Phi) \rangle \delta(t,t'),
\end{equation}
where we have defined $\langle D_2(x,y,\Phi) \rangle = \Tr{}{ D_{2}^{\mu \nu} (x,y, \Phi) L_{\mu}(x) \rho L^{\dag}_{\nu}(y)}$, and $\rho$ is the quantum state for the decohered mass density.
Here the $m,\Phi$ subscripts of $\mathbb{E}_{m,\Phi}$ allow for the possibility that the statistics of the noise process can be dependent on the Newtonian potential and mass distribution of the particle. The restriction on $\mathbb{E}_{m, \Phi}[uJ(x,t)]$ follows from assumption (ii). If $uJ(x,t)$ is Gaussian, Equation \eqref{eq: statsJ}
completely determines the noise process, but in general, higher order correlations are possible, although they need not concern us here, since we are only interested in bounding the effects due to $D_2(x,y,\Phi)$.
In the non-relativistic limit, where $c\rightarrow\infty$, we can take $\dot{\pi}_{\Phi}$ to be small in comparison to the other terms, and we recover Poisson's equation for gravity, but with a stochastic contribution to the mass. This is precisely as expected on purely physical grounds: in order to maintain coherence of any mass in superposition, there must be noise in the Newtonian potential and this must be such that we cannot tell which element of the superposition the particle will be in, meaning the Newtonian potential should look like it is being sourced in part by a random mass distribution. In other words, the trade-off requires that the stochastic component of the coupling obscures the amount of mass $m$ at any point.
The solution to Equation \eqref{eq: diffusivePi} is given by
\begin{equation}\label{eq: poissonnoiseMain}
\Phi(t,x) \approx -G \int d^3 x' \frac{[m(x',t)-u(\Phi, \hat{m})J(x',t)]}{|x-x'|},
\end{equation}
and a formal treatment of solutions to non-linear stochastic integrals of the form of Equation \eqref{eq: diffusivePi} can be found in \cite{DalangHigh}. A higher precision calculation would involve a full simulation of CQ dynamics and in Appendix \ref{sec: cqlimit} we show in full detail the evolution the Newtonian potential looks like for general continuous CQ theories using the continuous unraveling of CQ dynamics introduced in \cite{UCLHealing}. We find the effects are qualitatively and quantitatively the same as Equation \eqref{eq: poissonnoiseMain}.
In \cite{UCLPawula} it was shown that there are two classes of CQ dynamics, at least in the sense that there are those with continuous trajectories in phase space and those which contain discrete jumps. For the class of continuous CQ models (see \cite{diosi2011gravity} and appendix \ref{ss:continuous}), we know that $J(x,t)$ should be described by a white noise process in time, and its statistics should be independent of the mass density of the particle.
We go through the full CQ calculation for the continuous models in Appendix \ref{eq: newtUnrav}.
For the discrete class (see \cite{oppenheim_post-quantum_2018, oppenheim2021constraints} and Appendix \ref{ss: discrete}), $J(x,t)$ can involve higher order moments, and will generally be described by a jump process \cite{UCLPawula, unrav}. It's statistics can also depend on the mass density, since in general the diffusion matrix $D^{\mu \nu}_{2,ij}$ couples to Lindblad operators. It is worth noting that the discrete CQ theories considered in \cite{oppenheim_post-quantum_2018, unrav, UCLconstraints} generically suppress higher order moments, and often we expect that we can approximate the dynamics by a Gaussian process, but this need not be the case in general.
This variation in Newtonian potential leads to observational consequences which can be used to experimentally test and constrain CQ theories of gravity for various choices of kernels appearing in the CQ master equation. One immediate consequence is that the variation in Newtonian potential leads to a variation of force experienced by a particle or composite mass via $\vec{F}_{tot} = -\int d^3 x m(x) \nabla \Phi(x)$. We can also estimate the time averaged force via $\frac{1}{T}\int_0^T\vec{F}_{tot}$ where $T$ is the time over which the force is measured and is the useful quantity when comparing with experiments. Using Equation \eqref{eq: poissonnoiseMain}, in Appendix \ref{sec: tabletop} we find that the variance of the magnitude of the time averaged force experienced by a particle in a Newtonian potential is given by
\begin{equation}\label{eq: variationForceMain}
\sigma_F^2 = \frac{2G^2}{T} \int d^3x d^3y d^3x' d^3y' m(x) m(y) \frac{(\vec{x}- \vec{x}') \cdot (\vec{y}- \vec{y}')}{|x-x'|^3 |y-y'|^3} \langle D_2(x',y',\Phi)\rangle,
\end{equation}
where the variation is averaged over a time period $T$. We will use this to estimate the variation in precision measurements of mass, such as modern versions of the Cavendish experiement for various choices of $\langle D_2(x',y',\Phi)\rangle$.
On the other hand, experimentally measured decoherence rates can be related to $D_0$. We explore the calculation of decoherence rates in gravity in detail in~\cite{gravitydecoherence}. The important point is that the decoherence rate is dominated by the background Newtonian potential $\Phi_b$ due to the Earth. In Appendix \ref{sec: decoherenceRatesAppendix}, we show that for a mass whose quantum state is a superposition of two states $|L\rangle$ and $|R\rangle$ of approximately orthogonal mass densities $m_L(x),m_R(x)$, and whose separation we take to be larger than the correlation range of $D_0(x,y)$, the decoherence rate is given by
\begin{equation}\label{eq: decoherencerateMain}
\lambda = \frac{1}{2}\int dx dy D_0^{\alpha \beta}( x,y) (\langle L | L_{\beta}^{\dag}(y) L_{\alpha}(x) |L \rangle + \langle R | L_{\beta}^{\dag}(y) L_{\alpha}(x) |R \rangle ).
\end{equation}
Via the coupling constant trade-off, Equations \eqref{eq: variationForceMain} and \eqref{eq: decoherencerateMain} then give rise to a double sided squeeze on the coupling $D_2$. Equation \eqref{eq: variationForceMain} upper bounds $D_2$ in terms of the uncertainty of
acceleration measurements seen in gravitational torsion experiments, whilst the coupling constant trade-off Equation \eqref{eq: decoherencerateMain} lower bounds $D_2$ in terms of experimentally measured decoherence rates arising from interferometry experiments.
We now show this for various choices of diffusion kernel, with the details given in Appendix \ref{sec: tabletop}. The diffusion coupling strength will be characterized by the coupling constant $D_2$, which we take to be a dimension-full quantity with units $kg^2 s m^{-3}$, and is related to the rate of diffusion for the conjugate momenta of the Newtonian potential. We upper bound $D_2$ by considering the variation of the time averaged acceleration $\sigma_a = \frac{\sigma_F}{M}$ for a composite mass $M$ which contains $N$ atoms which we treat as spheres of constant density $\rho$ with radius $r_N$ and mass $m_N$. We lower bound $D_2$ via the coupling constant trade-off of Equation \eqref{eq: generalFieldMatrixInequalityMain} and then by considering bounds on the coherence time for particles with total mass $M_{\lambda}$, which have typical length scale when in superposition $R_{\lambda}$ and volume $V_{\lambda}$.
For continuous dynamics $\langle D_{2}(x,y, \Phi) \rangle$ = $D_2(x,y,\Phi)$ since the diffusion is not associated to any Lindblad operators. Let us now consider a very natural kernel, namely $D_2(x,y;\Phi) = D_2(\Phi) \delta(x,y)$ which is both translation invariant, and does not create any correlations over space-like separated regions. In general, the squeeze will depend on the functional choice of $D_2(\Phi)$ on the Newtonian potential. However, in the presence of a large background potential $\Phi_b$, such as that of the Earth's, we will often be able to approximate $D_2(\Phi) = D_2(\Phi_b)$. This is true for kernels which depend on $\Phi$ and $\nabla \Phi$, though the approximation does not hold for all kernels, for example $D_2 \sim -\nabla^2 \Phi$ of Equation \eqref{eq: lapbeltweak} which creates diffusion only where there is mass density. For diffusion kernels $D_2(\Phi_b)$ where the background potential is dominant we find the promised squeeze on $D_2(\Phi_b)$
\begin{equation}\label{eq: contLocal}
\frac{\sigma_a^2 N r_N^4 T}{V_b G^2 } \geq D_2 \geq \frac{M_{\lambda}^2}{V_{\lambda} \lambda},
\end{equation}
where $V_b$ is the volume of space over which the background Newtonian potential is significant. $V_b$ enters since the variation in acceleration is found to be
\begin{equation}\label{eq: intD2Main}
\sigma_a^2 \sim \frac{D_2 G^2 }{r_N^4 N T} \int d^3x' D_2(\Phi_b) ,
\end{equation}
where the $d^3 x'$ integral is over all space. This immediately rules out continuous theories with noise everywhere, i.e, with a diffusion coefficient independent of the Newtonian potential since the integral will diverge.
Standard Cavendish type classical torsion balance experiments measure accelerations of the order $10^{-7}ms^{-2}$, so a very conservative bound is $\sigma_a \sim 10^{-7}ms^{-2}$, whilst for a kg mass $N\sim 10^{26}$ and $r_N\sim 10^{-15} m$. Conservatively taking $V_b \sim r_E^2 h \ m^3$ where $r_E$ is the radius of the Earth and $h$ is the atmospheric height gives $D_2 \leq 10^{-41}kg^2 s m^{-3}$. The decoherence rate $\lambda$ is bounded by various experiments~\cite{bassi}. Typically, the goal of such experiments is to witness interference patterns of molecules which are as massive as possible. Taking a conservative bound on $\lambda$, for example that arising from the interferometry experiment of~\cite{Gerlich2011} which saw coherence in large organic fullerene molecules with total mass $M_{\lambda} = 10^{-24}kg$ over a timescale of $0.1s$, gives an upper bound on the decoherence rate $\lambda < 10^{1} s^{-1}$. Fullerene molecules have typical size $ 10^{-9}m$. After passing through the slits the molecule becomes delocalized in the transverse direction on the order of $10^{-7}m$ before being detected. Since the interference effects are due to the superposition in the transverse $x$ direction, which is the direction of alignment of the gratings, it seems like a reasonable assumption to take the size of the wavepacket in the remaining $y,z$ direction to be the size of the fullerene, since we could imagine measuring the $y,z$ directions without effecting the coherence. We therefore take the volume $V_{\lambda} \sim 10^{-9} 10^{-9} 10^{-7}m^3 = 10^{-25}m^3$, which gives $D_2 \geq 10^{-24} kg^2 s m^{-3}$, and suggests that classical-quantum theories of gravity with local continuous noise need to have a dependence on the Newtonian potential which will suppress the diffusion by $20$ orders of magnitude. This happens to be the case for the kernel from Section \ref{ss:lapkernel} and whose motivation comes from the constraint algebra \cite{oppenheim2021constraints}.
On the other hand, the discrete models appear less constrained due to the suppression of the noise away from the mass density. For example consider the local discrete jumping models, such as the one given in Section \ref{ss: discrete} which have $\langle D_2(x,y,\Phi_b)\rangle = \frac{ l_P^3 D_2(\Phi_b)}{m_P } m(x)$, where $m_P = \sqrt{\frac{\hbar c}{G}}$ is the Planck mass and $l_{P} = \sqrt{\frac{\hbar G}{c^3}}$ is the Planck length, required to ensure $D_2$ has the units of $kg^2 s m^{-3}$. We find the squeeze on $D_2$
\begin{equation}\label{eq: disLocal}
\frac{ \sigma_a^2 N r_N^4 T}{ m_N G^2} \geq \frac{l_P^3}{m_P} D_2 \geq \frac{ M_{\lambda}}{\lambda},
\end{equation}
and plugging in the numbers tells us that discrete theories of classical gravity are not ruled out by experiment and we find $10^{-1}kg s \geq \frac{l_P^3}{m_P} D_2 \geq 10^{-25}kg s $.
We can also consider other noise kernels, with examples and some discussion given in Section \ref{app: examplesOfKernals}. A natural kernel is $D_2(x,y,\Phi_b) = - l_P^2 D(\Phi_b) \nabla^2 \delta(x,y)$. The inverse Lindbladian kernel satisfying the coupling constants trade-off is to zeroeth order in $\Phi(x)$, the Diosi-Penrose kernel $D_0(x,y,\Phi_b) = \frac{D_0(\Phi_b)}{|x-y|}$. For this choice of dynamics, we find the squeeze for $D_2$ in terms of the variation in acceleration
\begin{equation}\label{eq: contDP}
\frac{\sigma_a^2 N r_N^3 T}{G^2} \geq l_P^2 D_2 \geq \frac{M^2_{\lambda}}{R_{\lambda} \lambda} .
\end{equation}
Using the same numbers as for the local continuous model, with $R_{\lambda} \sim V_{\lambda}^{1/3}\sim10^{-9}m$ we find that classical torsion experiments upper bound $D_2$ by $10^{-9}kg^2 s m^{-1} \geq l_P^2 D_2$, whilst interferometry experiments bound $D_2$ from below via $l_P^2 D_2 \geq 10^{-40} kg^2 s m^{-1} $.
Equations \eqref{eq: contLocal}, \eqref{eq: disLocal} and \eqref{eq: contDP} show that classical theories of gravity are squeezed by experiments from both ways. We have here been extremely conservative, and we anticipate that further analysis, as well as near term experiments, can tighten the bounds by orders of magnitude. There are several proposals for table-top experiments to precisely measure gravity, some of which have recently been performed, and which could give rise to tighter upper bounds on $D_2$. Some of these experiments involve millimeter-sized masses whose gravitational coupling is measured via torsional pendula~\cite{westphal2020measurement,Schm_le_2016}, or rotating attractors~\cite{PhysRevLett.124.101101}. With such devices, the gravitational coupling between small masses can be measured while limiting the amount of other sources of noise. There are proposals for further mitigating the noise due to the environment, including the inertial noise, gas particles collisions, photon scattering on the masses, and curvature fluctuations due to other sources~\cite{Chevalier_2020,vandekamp2020quantum,toros2020relative}. Other experiments are based on interference between masses; for example, atomic interferometers allow for the measurement of the curvature of space-time over a macroscopic superposition~\cite{Chou1630,PhysRevLett.118.183602}.
We can get stronger lower bounds via improved coherence experiments. Typically, the goal of such experiments is to witness interference patterns of molecules which are as massive as possible, while here, we see that the experimental bound on CQ theories is generically obtained by maximizing the coherence time for massive particles with as small wave-packet size $V_{\lambda}$.
Thus far in this section we have considered local effects on particles due to the diffusion. While this enables us to rule out some types of theories, the bounds are generally weak if one wanted to rule out all of them. However, it may be possible to do so via cosmological considerations. In attempting to place experimental constraints on this diffusion, it is also worth considering other regimes, such as longer range effects which might be detected by gravitational wave detectors such as LIGO.
In Appendix \ref{sec: energyProduction}, we begin a study of the cosmological effects of the diffusion by studying the observational trade-off in Equation \eqref{eq: observationalTradeOffFields}. For the class of CQ theories sourced by either the $D_{1,i}^{0 \mu}$ term, or the $D_{1,i}^{\alpha \beta}$ this gives a
lower bound for the diffusion of the conjugate momentum in terms of the mass density of the particle and its decoherence rate $\lambda$
\begin{equation}\label{eq: piDispersionmain}
\frac{d\sigma^2_{\pi_\Phi}(x)}{dt} \geq \frac{ |\langle m(x) \rangle|^2}{ 8 \lambda},
\end{equation}
which leads to an estimate of the production rate of stochastic waves in terms of their gravitational kinetic energy. This can be lower bounded in terms of experimentally decoherence rates
\begin{equation}
\frac{d \Delta \bar{E}}{dt} \geq \int d^3 x \frac{ c^2 G \pi |\langle m(x) \rangle|^2}{12 \lambda } .
\label{eq: diffusion rate trade-offmain}
\end{equation}
The diffusion is akin to the stochastic production of gravitational waves, but these waves need not be transverse (see Appendix \ref{sec: grav diffusion}).
One advantage of studying this regime, is that Equation \eqref{eq: diffusion rate trade-offmain} is a bound which holds very generally, and which is independent of the choice of kernel, since the kinetic energy rate is lower bounded in terms of a experimentally measured decoherence rate.
Since fullerene interference experiments require $\frac{d \Delta \bar{E}}{dt} \sim 10^{-19} Js^{-1}$ over the wave-packet size of nucleons, this implies that the energy density of waves must be produced at a rate of at least $ \sim 10^{5} Js^{-1}m^{-3}$. If this was produced over all space, as required by the diffusion kernel $D_2(x,x')=D_2\delta(x,x')$ (already ruled out by precision Cavendish experiments), this gives an apparent energy density in the ball park of $10^{22}J/m^3$ accumulated over the age of the universe. Since the observed expansion rate of the universe puts its energy density at $10^{-9}J/m^3$, this discrepancy would appear to rule out continuous realisations of classical gravity with the diffusion kernel $\delta(x,x')$. This can be seen as a specific instance of diffusion kernels $D_2(x,x')$ which diverge as $x\rightarrow x'$, something we comment on in Section \ref{app: examplesOfKernals}. However, even here, care should be taken, partly because our estimate is non-relativistic\footnote{We find that extrapolating too far into the past runs afoul of the gauge fixing condition used to derive the Newtonian limit.}, and partly because our understanding of cosmology requires some degree of modesty -- after all, quantum field theory predicts an energy density of $10^{113} J/m^3$, and yet we do not see this reflected in the acceleration of the universe.
We leave a detailed study of the effect of gravitational diffusion on LIGO to future work. It suffices to mention that the effect will again depend on the form of the kernel $D_2(x,x')$. Our estimates \cite{UCLLigo} suggest that local effects from table-top experiments currently place stronger bound on gravitational theories than LIGO currently does. In particular, unlike for gravitational wave measurements, which are reasonably high frequency events requiring extraordinary high precision in relative displacement of the arm length from its average, it is preferential to have a lower precision measurement, but which occurs over a longer time period to allow for the diffusion in path length to build up, and with a smaller uncertainty in the average length of the arm itself. Furthermore, since the LIGO arm is kept in a vacuum, we do not expect strong bounds on discrete models where the diffusion is associated to an energy density.
\section{Discussion}\label{sec: conclusion}
A number of direct proposals to test the quantum nature of gravity are expected to come online in the next decade or two. These are based on the detection of entanglement between mesoscopic masses inside matter-wave interferometers~\cite{kafri2013noise,kafri2015bounds,bose2017spin,PhysRevLett.119.240402,Marshman_2020,pedernales2021enhancing,carney2021testing,christodoulou2022locally}. For these experiments, some theoretical assumptions are needed: one requires that it is only gravitons which travel between the two masses and mediate the creation of entanglement. If this is the case, then the onset of entanglement implies that gravity is not a classical field. These can be thought of as experiments which if successful, would confirm the quantum nature of gravity (although other alternatives to quantum theory are possible \cite{galley2021nogo}).
Here, we come from the other direction, by supposing that gravity is instead classical, and then exploring the consequences.
Theories in which gravity is fundamentally classical were thought to have been ruled out by various no-go theorems and conceptual difficulties. However, these no-go theorems are avoided if one allows for non-deterministic coupling as in \cite{blanchard1995event,diosi1995quantum,diosi2011gravity,alicki2003completely,poulinKITP2,oppenheim_post-quantum_2018,unrav,UCLconstraints,UCLPawula}. We have here proven that this feature is indeed necessary, and made it quantitative by exploring the consequences of complete positivity on any dynamics which couples quantum and classical degrees of freedom. Complete positivity is required to ensure the probabilities of measurement outcomes remain positive throughout the dynamics. We have shown that any theory which preserves probabilities and treats one system classically, is required to have fundamental decoherence of the quantum system, and diffusion in phase space, both of which are signatures of information loss. Using a CQ version of the Kramers-Moyal expansion, we have derived a trade-off between decoherence on the quantum system, and the system's diffusion in phase space. The trade-off is expressed in terms of the strength of the back-reaction of the quantum system on the classical one. We have derived the trade-off both in terms of coupling constants of the theory, and in terms of observational quantities that can be measured experimentally.
In the case of gravity, the observational trade-off places a lower bound on the rate of diffusion of the gravitational degrees of freedom as expressed by Equation \eqref{eq: piDispersionmain} in terms of the decoherence rate of particles in superposition. We find that theories which treat gravity as fundamentally classical, are not ruled out by current experiments, however we have been able to rule out a broad parameter space of such theories. This is done partly through table-top observations via Equations \eqref{eq: contLocal}, \eqref{eq: disLocal} and \eqref{eq: contDP}. Given any diffusion kernel, we can compute the inaccuracy of mass measurements due to fluctuations in the gravitational field, and using the trade-off, we can derive a bound on the associated decoherence rate. This allows us to rule out broad classes of theories in terms of their diffusion kernel. For example, we are able to rule out a number of theories which are continuous in phase space. Then, using the trade-off of Equation \eqref{eq: diffusion rate trade-offmain}, we saw that there was some tension with cosmological observations and kernels such as that of Equation \eqref{eq: Xlocal} and \eqref{eq: lapbelt}, which produce diffusion over all space. However, we are not confident enough in our understanding of cosmology in CQ theories to rule these out.
Any theory which treats gravity classically has fairly limited freedom to evade the effects of the trade-off. There is freedom to choose the diffusion or decoherence kernels $D_2(x,x')$ and $D_0(x,x')$, but the trade-off restricts one in terms of the other. Then, because of the results proven in \cite{UCLPawula}, one can consider two classes of theory, those which are continuous realisations and whose diffusion can only depend on the gravitational degrees of freedom, and discrete theories whose diffusion can also depend directly on the matter fields. Examples of both classes of theory are given in Appendix \ref{sec: cqlimit}.
Finally, one could consider theories which do not reproduce the weak field limit of General Relativity to all distances, namely we could imagine that the interaction Hamiltonian of Equation \eqref{eq:interaction_ham} does not hold to arbitrarily short distances, or arbitrarily high mass densities. This would correspond to modifying $D_1(x,x')$ in some way, either by making it slightly non-local, or by disallowing arbitrarily high mass densities, or by including an additional contribution such as the friction term discussed in \ref{ss:continuous}. All of these modifications would seem to violate Lorentz invariance in some way\footnote{This may only be a concern if it results in any inconsistency with low energy observations, since a theory of quantum gravity would also likely have an anomaly at the Planck scale.}.
Here, we have only given an order of magnitude estimate of when gravitational diffusion will lead to appreciable deviations from Newtonian gravity or General Relativity. We have done so in a number of regimes. The most promising being table-top experiments which precisely measure the mass of an object. This is an area which is important from the perspective of weight standards, for example those undertaken by NIST on the 1kg mass standard K20 and K4 \cite{abbott2019mass}. The increased precision and measuring time of Kibble Balances \cite{Chao2019Design} and atomic interferometers \cite{Chou1630,PhysRevLett.118.183602,peters2001high,menoret2018gravity} would make such measurements an ideal testing ground, both to further constrain the diffusion kernel, and to look for diffusion effects, whose dependence on the test mass is outlined in Appendix \ref{sec: grav diffusion}. Here, we have found that the time $T$ over which results of the measurement are made, affects the strength of the bound, and it would be helpful if future experiments reported this value. Since we have found that CQ theories predict an uncertainty in mass measurements it is perhaps intriguing that different experiments to measure Newton's constant $G$ yield results whose relative uncertainty differ by as much as $5 \cdot10^{-4}$ m$^3$kg$^{-1}$s$^{-2}$, which is more than an order of magnitude larger than the average reported uncertainty \cite{quinn2000measuring,gillies2014attracting,rothleitner2017invited}. If one were to try and explain the discrepancy in $G$ measurements via gravitational diffusion, then for all the kernels we studied in Section \ref{sec: gravity} we find that the variation in acceleration depends on $\frac{1}{\sqrt{N}}$ the number of nucleons in the test mass, so that masses with smaller volume should yield larger uncertainty and this would be the effect to look for in measurement discrepancies. The relatively large uncertainty in such measurements, also makes it challenging for table-top experiments to place strong upper bounds on gravitational diffusion.
We have also estimated the effect that this diffusion would have on the energy density of the universe, and in the production of stochastic waves in terms of gravitational kinetic energy in the weak field limit. We have found that spatially uncorrelated and continuous realisations of classical gravity which reproduce General Relativity at short distances, appear to be ruled out by cosmological considerations as well, since the energy density of the stochastic wave contribution is high enough that it should effect the expansion rate of the universe. However, this is a regime where we do not understand the theory well, and so we are cautious about making too strong a claim. We have also found that the stochastic production of gravitational kinetic energy waves is in a regime which could be detectable by LIGO, an effect which constrains the form of $D_2(x,x')$. However, initial estimates suggest that this is less constraining than table top experiments. For this to be definitive, a more precise understanding of gauge artifacts and of the dynamics that the diffusion induces on the Newtonian potential is required, especially over longer time scales.
How this diffusion might effect dynamics over galactic scales and longer times, requires a fuller General Relativistic treatment, a study which we undertake in \cite{UCLDMDNE} where we find that it causes deviations from what general relativity predicts.
Turning to the other side of the trade-off, improved decoherence times would further squeeze theories in which gravity remains classical. While a current experimental challenge is to demonstrate interference patterns using larger and larger mass particles, we here find that some of our bounds depend on the expectation of the particle's mass density, either in terms of $\langle m^2(x)\rangle/\lambda$, or in ways which depend on the particular kernel. Thus interference experiments with particles of high {\it mass density} rather than mass can be preferable. There are also kernels, for which the relevant quantity is the expectation of the mass density, which will depend on the size of the wave-packet used in the interference experiment, a quantity which is rarely obtainable from most papers which report on such experiments. While this dependence might initially appear counter-intuitive, it follows from the fact that in order to relate the trade-off in terms of coupling constants to observational quantities, and in particular, the decoherence rate, we took expectation values of the relevant quantities to get a trade-off in terms of only averages. And indeed the decoherence rate, which is an expectation value, can easily depend on the wave-packet density, as we see from examples is Section \ref{sec: decoherenceRatesAppendix}.
Since we here show that all theories which treat gravity classically necessarily decohere the quantum system, another constraint on theories which treat gravity classically is given by constraints on fundamental decoherence. These are usually constrained by bounds on anomalous heating of the quantum system \cite{bps}. However, these constraints are not in themselves very strong, since fundamental decoherence effects can be made arbitrarily weak.
In the simplified model in Appendix \ref{sec: cqlimit}, the strength of the decoherence depends on the strength of the gravitational field, thus, constraints due to heating \cite{bps,ghirardi1986unified,ballentine1991failure,pearle1999csl,bassi2005energy,adler2007lower,lochan2012constraining,nimmrichter2014optomechanical,bahrami2014testing,laloe2014heating,bahrami2014proposal,goldwater2016testing,tilloy2019neutron,donadi2020underground} can be suppressed, either by scaling the Lindbladian coupling constants, or by having strong decoherence effects more pronounced near stronger gravitational fields such as near black holes where one expects information loss to occur. The necessity for decoherence to heat the quantum system is further weaked by the fact that the dynamics are not Markovian on the quantum fields, if one integrates out the classical degrees of freedom, space-time acts as a memory. This potentially captures some of the non-Markovian features advocated in \cite{unruh-wald-onbps}, who recognised that Markovianity is a key assumption in attempts to rule out fundamental decoherence or information loss. Here however, we see that there is less freedom than one might imagine. If the Lindbladian coupling constants are made small to reduce heating, the gravitational diffusion must be large. Thus, heating constraints which place bounds on $D_0(x,x')$ place additional constraints on $D_2(x,x')$.
While the absence of diffusion could rule out theories where gravity is fundamentally classical, the presence of such deviations, at least on short time scales, might not by itself be a confirmation of the classical nature of gravity. Such effects could instead be caused by quantum theories of gravity whose classical limit is effectively described by \cite{oppenheim_post-quantum_2018} or perhaps \cite{hu2008stochastic}. In other words, one might expect some gravitational diffusion, because from an effective theory point of view, one is in a regime where space-time is behaving classically. However, the
trade-off we have derived is a direct consequence of treating the background space-time as fundamentally classical. In a fully quantum theory of gravity, the interaction of the gravitational field with particles in a superposition of two trajectories will cause decoherence, but coherence can then be restored when the two trajectories converge. This is what happens when electrons interact with the electromagnetic field while passing through a diffraction grating, yet still form an interference pattern at the screen. This is a non-Markovian effect,
and the trade-off we derived is a direct consequence of the positivity condition, which is a direct consequence of the Markovian assumption. In the non-Markovian theory where General Relativity is treated classically, one still expects the master equation to take the form found in \cite{oppenheim_post-quantum_2018}, but without the matrix whose elements are $D_n^{\mu\nu}$ needing to be positive semi-definite \cite{Hall2014, Breuer2016}.
\section*{Acknowledgements}
We would like thank Sougato Bose, Joan Camps, Matt Headrick, Isaac Layton, Juan Maldacena, Andrea Russo, Andy Svesko and Bill Unruh for valuable discussions and Lajos Di\'{o}si and Antoine Tilloy for their very helpful comments on an earlier draft of this manuscript
JO is supported by an EPSRC Established Career Fellowship, and a Royal Society Wolfson Merit Award, C.S. and Z.W.D.~acknowledges financial support from EPSRC. This research was supported by the National Science Foundation under Grant No. NSF PHY11-25915 and by the Simons Foundation {\it It from Qubit} Network. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Economic Development, Job Creation and Trade.
| 7d495ec03e67357cd737613a10e27c9a9d502d63 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
According to the predictions of the Einstein's Equivalence Principle (EEP) the light emitted in the stellar photospheres suffers a redshift when escaping its gravitational field, which is directly correlated to the mass-to-radius ratio of the observed star. In numbers, the measured effect is of the order of several
hundred m\,s$^{-1}$ for main-sequence stars and tens of times smaller for giants. For white dwarfs the gravitational effect is
several tens of km\,s$^{-1}$ owing to their comparatively small radius and extremely massive cores. That explains why the
first attempts at detecting this relativistic effect was achieved in these types of stars
by \citet{1967ApJ...149..283G} and \citet{1972ApJ...177..441T}.
The proximity of the Sun allows high-quality observations that are
unattainable for other stars. However, since the Sun is a main-sequence star, the
gravitational effect is relatively small and the observations need to consider the shift
associated with the convective nature of the photosphere. In the solar spectrum, for example,
the absorption line wavelength is blueshifted by about 400 m\,s$^{-1}$ after correction for
the known gravitational redshift \citep{2003A&A...401.1185L}. In general, the precise amount
of convective shift depends on the strength of the absorption line, i.e. stellar metallicity,
surface gravity, and spectral type, or effective temperature \citep{2002ApJ...566L..93A, 2013A&A...550A.103A}.
Several measurements have been made of the Sun \citep{2014MNRAS.443.1837R}, the most precise estimate being
performed by \citet{2020A&A...643A.146G}, who take into account all these effects to
determine the Sun's gravitational redshift at 633.1 m\,s$^{-1}$ with an uncertainty of $\sim $ 1\%.
Extending the test to other non-degenerate stars is of interest in order to test the EEP and the accuracy of stellar models. Nevertheless, as outlined above, such experiments
are difficult: on the one hand, the hydrodynamic effect of the stellar photosphere produces a
counterbalancing effect that tends to cancel out the relatively small gravitational redshift effect of
the typical main-sequence star; on the other hand, obtaining data with high enough quality requires
considerable effort in terms of observing time and instrumentation of high spectral resolution.
These considerations explain the relatively low number of such attempts \citep{1997A&A...322..460N, 2019MNRAS.483.5026L, 2002A&A...381..446M}. Among these, it is worth mentioning the
work by \citet{2011A&A...526A.127P}, who measured the shift of spectral lines for a sample of
144 member stars in the M67 open cluster. The aim was to measure a radial velocity shift
difference between the main-sequence and giant stars. The detection was unsuccessful, so the
authors conclude that the convective shift probably mimicked the gravitational effect.
More recently, \citet{2019ApJ...871..119D} used \textit{Gaia} radial velocity data to measure
the redshift in the main-sequence stars and found values that are much smaller than the
theoretical predictions. Again, the authors attributed this deficiency to the convective
blueshift effect. A more successful attempt to detect the gravitational effect was carried out
by \citet{2021arXiv210201079M} in non-degenerate stars. These authors measured the differences
in velocities between the components of pairs of co-moving stars formed by dwarf--giant binaries.
Instead of assuming the stellar hydrodynamical effect, they used a parametric model
\cite{2013A&A...550A.103A} to estimate the convective shift. The final results provide
compelling evidence for the compatibility between observations and theoretical predictions. Actually, measuring the gravitational redshift in galaxy clusters (e.g., \citealt{1995A&A...301....6C, 2011Natur.477..567W, 2013MNRAS.435.1278K}) and quasars \citep{Mediavilla_2021} could also be used to test General Relativity versus alternative theories of gravitation.
Nowadays, the availability of huge databases with astrometric, photometric and spectroscopic information, such as \href{https://www.sdss.org}{SDSS} and \textit{Gaia} \citep{2016A&A...595A...1G}, has provided sufficient objects to detect the gravitational redshift in a statistical way.
By using a sample of member stars of open clusters catalogued by \textit{Gaia}, we aim to
detect the effect of stellar gravitational redshift in non-degenerate stars. The
data analyzed contain several thousand non-degenerate stars covering a wide range of mass and
radius that potentially allows us to detect relative variations in their gravitational redshift. Unfortunately, the relatively weak gravitational fields existing in non-degenerated stars and the accuracy of the data used in this work does not allow to discriminate between different gravitational theories.
The paper is structured in a standard way. After this introduction, Section 2 describes
the sample and methodology used; in Section 3 are presented the analysis, main results, and
their interpretation. Finally, conclusions and planned future work are given in Section 4.
\section{Methodology and Sample} \label{sec:method}
For this study, we use a sample of stars in open clusters (OCs) of our galaxy.
OCs are young stellar clusters which have been formed from the same molecular cloud,
so their members share roughly the same age and metallicity \citep{2020SSRv..216...69A}.
As co-moving systems bounded by mutual gravitational attraction, their
the stellar peculiar velocity in the radial direction may be assumed to be random and
to cancel out on average for each OC.
\begin{figure}[tb!]
\plotone{figure1.pdf}
\caption{\textit{Left panel}: Distribution of the stellar radial velocities
($RV_i$) relative to its host cluster ($RV_{\rm OC}$). The solid lines are the
Gaussian distribution fits for each sample, with corresponding standard deviations
($\sigma$) depicted in the label. $N$ denotes the number of stars. \textit{Right panel}:
Histogram distribution of the member stars ($N$) population in open cluster ($N_{\rm OC}$).
In both panels, the step-type histogram corresponds to the full data set of 86 clusters
and 8899 stars, labeled GR; the filled histogram corresponds to a subset of 51
clusters and 4464 stars for which it is possible to estimate the convective shift,
labeled GR+CS. See main text for details. \label{fig:histRV}}
\end{figure}
We took advantage of the largest radial velocity ($RV$) catalogue for stars in OCs by
\citet{2021A&A...647A..19T}. It gathers nearly 30\,000 $RV$ measurements of 1382 OC
members based on \textit{Gaia}-RVS and ground-based surveys (see references within).
The complete OCs catalogue and the initial membership probability assessment came mostly
from \citet{2020A&A...640A...1C} and \citet{2020A&A...633A..99C}, both based on \textit{Gaia} DR2 astrometry.
The initial 1382 OCs data were completed with stellar parameters from the \textit{Gaia}
DR2 based StarHorse catalogue \citep{2019A&A...628A..94A}. This provides estimations of
photo-astrometric distances, extinctions, and other astrophysical parameters such as
effective temperature ($T_{\rm eff}$), surface gravity ($\log g$) and metallicity ([Fe/H]).
The final dataset contains 1342 OCs and 21\,403 stars classified as OC members with a
minimum membership probability of 0.4, a threshold estimated by the authors. That
constitutes the initial data set used in this paper.
In order to select an appropriate subsample for this work, only stars with uncertainties
up to 2 km\,s$^{-1}$ in $RV$ measurements were considered. This first restriction removes
24\% of the initial sample objects from the analysis, and was motivated by the need to discard not
only objects with poor measurements but also those that are part of binary systems, whose velocities could
be dominated by the binary orbit itself.
To constrain the stellar membership of the host cluster provided by \citet{2018A&A...619A.155S},
radial velocities were used as an additional criterion \citep{1997A&A...322..460N}. It was assumed that the radial
velocities of the stars in a given OC correspond to a sample of a Gaussian distribution
centred on $RV_{\rm OC}$ with a standard deviation $\sigma_{\rm OC}$. $RV_{\rm OC}$ was estimated as
the mean value of the radial velocities of the member stars through an iterative method
that rejects as outliers those stars that were originally classified as OC members, but
that have velocities outside the $\pm 3 \sigma_{\rm OC}$-clipping. It was assumed that the
spectral shift corresponds entirely to the motion of the stars, i.e.\ any other possible
contributor, such as the gravitational redshift or convective shift, was ignored at this stage.
This approach is likely to produce a small bias (of a few hundred m\,s$^{-1}$)
in the estimates of $RV_{\rm OC}$ and $\sigma_{\rm OC}$. The relevance of this minor effect is
considered in the results and conclusions sections of this paper.
After estimating the radial velocity for each cluster, two additional restriction
were applied. First, we consider only those clusters with a minimum number of stars
in order to gain confidence in our estimation of the kinematic parameters of the cluster.
Several thresholds were considered and $N_{\rm min} \geq 20$ was chosen as a good option to
keep a relatively large sample. To ensure a tight gravitational binding, those clusters
with $\sigma_{\rm OC}> 3$ km\,s$^{-1}$ were also discarded. With these restrictions in place, we were left with $\sim$42\% of the original dataset.
Additional properties of the clusters such as morphology, age, position within our galaxy,
or the stellar type population, etc., could be also considered as efficient criteria to optimize the sample. However, we decided
to ignore them to minimize the number of assumptions and to give a more general validity
to the results. Neither has it been considered the information of tangential velocities
that should provide a better estimation of the kinematic and dynamical state of the OC.
That might constitute an additional criteria to determine the membership probability of the
stars\footnote{A further analysis is possible by comparing tangential and radial velocities of
the stars, which might provide an alternative way to estimate gravitational redshift.}.
Finally, a total sample of 86 OCs and 8899 stars was selected for further analysis (Table \ref{tab:basesample}). A
subsample of 51 OCs and 4464 stars are obtained after applying the restrictions based on
the hydrodynamical model to estimate the convective shift effect of the stars (see below).
Figure \ref{fig:histRV} depicts the kinematic and descriptive information for both the GR (step
histogram) and GR+CS (filled histogram) data samples.
It is expected that the velocity distribution for each cluster follow a Gaussian form, centered
on $RV_{\rm OC}$ with a dispersion of $\sigma_{\rm OC}$. In order to consider all the clusters, the relative
radial velocity $\Delta RV$ distribution was depicted in Figure \ref{fig:histRV} (\textit{left
panel}), subtracting the corresponding $RV_{\rm OC}$. After combining all the clusters, the relative
velocity histogram fits a single Gaussian distribution centered at 0 with
$\sigma_{\rm GR} = 1.6$ km \, s$^{-1}$ and $\sigma_{\rm GR+CS} = 1.7$ km \, s$^{-1}$, respectively for GR and GR+CS data.
The velocity dispersion ranges from $-$8.41 to 8.08 km\,s$^{-1}$ that roughly corresponds to
the $\pm 3$-sigma level of the dispersion.
The \textit{right panel} shows the histogram distribution of member stars $N$, illustrating
how populated are the clusters ($N_{\rm OC}$). It serves to compare quantitatively the numbers
of both data samples.
\subsection{Gravitational Redshift} \label{subsec: GR}
To calculate a star's predicted gravitational redshift (GR) it is necessary to know its mass and
radius, as $GR \sim M/R$. As a reference value, we estimate the gravitational redshift in the vicinity of the Sun
to be ${GR}_\odot = GM_\odot/cR_\odot = 636.31$\footnote{\label{ref_units} Calculated
using \href{https://docs.astropy.org/en/stable/constants/}{AstroPy} astrophysical constants.} m\,s$^{-1}$.
Therefore, for a given star of mass $M_*$ and radius $R_*$, the predicted gravitational effect is
given by $GR_*=G(M_*/M_\odot)/c (R_*/R_\odot)$.
The stellar masses are provided by StarHorse catalogue \citep{2019A&A...628A..94A} and their radii
can be easily estimated through their surface gravity ($\log g \sim M/R^2$), $\log g$
also being available for each star. The reference value for $\log g_\odot$ is $4.438^\mathrm{\ref{ref_units}}$.
Figure \ref{fig:CMD}, \textit{left panel}, illustrates the Hertzprung--Russell (HR) diagram for our total
data sample, color-coded with the GR predictions (km\,s$^{-1}$). The GR values span a range from a
minimum of 4.45 m\,s$^{-1}$ for red-giant stars, gradually increasing towards the main-sequence stars
with a typical value of 0.60 km\,s$^{-1}$ and reaching a maximum of 1.30 km\,s$^{-1}$ for stars above
the main-sequence turn-off, some of these being known as blue stragglers \citep{2019A&A...627A.119C}.
\begin{figure}[tb!]
\makebox[\textwidth][c]{\includegraphics[width=1.05\textwidth]{figure2.pdf}}
\caption{Comprehensive Hertzprung--Russell (HR) diagram with dereddened colors and magnitudes.
\textit{Left panel}: 86 clusters and 8899 stars, color-coded by the stars' predicted
gravitational redshift (GR) results (km\,s$^{-1}$). \textit{Middle panel}: 51 clusters
and 4464 stars, color-coded by the stars' convective shift (CS) results (km\,s$^{-1}$).
\textit{Right panel}: these 51 clusters and 4464 stars, color-coded by the sum of the
GR and CS results (km\,s$^{-1}$). \label{fig:CMD}}
\end{figure}
\subsection{Convective Shift} \label{subsec: CS}
Convective shift (CS) is a hydrodynamic effect in the observed $RV$ due to the convective
motions of the stellar photosphere \citep{2003A&A...401.1185L}. It produces a net kinematic
displacement that is reflected in the line-of-sight velocity shift measured for each star.
To estimate the effective contribution of the CS we used the 3D hydrodynamical model developed
by \citet{2013A&A...550A.103A}, which is described as a function of three stellar parameters;
namely, the stellar effective temperature ($T_{\rm eff}$), surface gravity ($\log g$), and metallicity
([Fe/H]). The model is constrained mainly to the main-sequence (MS) and red-giant-branch (RGB)
stellar parameters. The allowed parameter ranges constitute the additional restrictions on the
final data set, as mentioned before. The ranges for each parameter are: $T_{\rm eff}[K] \in
[3790.0, 6730.0]$; $\log g [dex] \in [9.184 \cdot 10^{-4}*T_{\rm eff}-2.482, 4.5]$; [Fe/H][dex] $ \in [-3.0, 0.0]$.
After the restrictions were imposed, we were left with a total of 51 clusters and 4464 member stars
(see Figure \ref{fig:histRV} for comparison with the initial data set). The estimated CS
contribution per star is color-coded in the \textit{middle panel} of the HR diagram (Figure \ref{fig:CMD}).
The CS results span a range from a minimum of $-$322.7 m\,s$^{-1}$ (i.e. blueshifted) for a typical sun-like
star, to a maximum of 804.9 m\,s$^{-1}$ (i.e. redshifted) for the early F-type dwarfs. It is worth
highlighting that less than 2\% of the stars have a CS greater than 0.3 km\,s$^{-1}$, in agreement
with the predictions.
The total \emph{velocity shift} contribution to the observed $RV$, estimated from the sum of the GR and CS
effects, is illustrated in the \textit{right panel} of Figure \ref{fig:CMD}. The results are clearly
differentiated between the red-giant branch and the main-sequence stars, with a minimum blueshift of
$-$50.84 m\,s$^{-1}$ and a maximum redshift (for early F-type dwarfs) of 1.16 km\,s$^{-1}$, respectively.
The majority of the sample, namely 74\% of stars, have a total predicted shift within the range of 0.2--0.7 km\,s$^{-1}$.
\section{Analysis and Results} \label{sec:analysis}
The basic hypothesis taken into account
is that the observed spectral shift, in a given star, is a combination of (i) the average radial
velocity of its OC; (ii) the radial component of the peculiar velocity of the star within its OC;
(iii) the gravitational redshift ($z_{\rm GR}$) which is equivalent from an observational
point of view to a radial velocity $GR = c z_{\rm GR}$, and (iv) the effect of the convective shifts (CS)
due to the stellar turbulence at the photosphere. The large size and the broad range of stellar
parameters in our sample allow us to consider other second-order effects (e.g., the kinematics of the outer
layer, orbital motion in binary systems, etc.) as sources of random noise. The radial velocity of each
OC was estimated as the average of the radial velocities of it member stars, while the peculiar velocity
of each star within its own OC was also considered a source of random noise.
In summary, the observed redshift (after subtracting the mean velocity of the OC) is $\Delta RV = GR +
CS + \sigma $. To test this hypothesis, the data was parameterized as $\Delta RV = a \cdot (GR + CS) + b$,
where $a$ is the slope correlating the expected results with the observations. Values $a \sim 1$ indicate
agreement between theory and observations.
Two different approaches were taken to archive our goal. The first one is based on analyzing a
fictitious cluster that has been built through the combination of the stellar members of all the OCs, and
analyze it as a single system (hereafter Method A). The second approach considers only the
\textit{best}-clusters to analyze them independently (hereafter Method B). In Method A, after subtracting
the radial velocity of each cluster, we jointly analyze all the stars of host selected clusters with at
least 20 members, i.e. those stars in the step-histograms shown in Figure 1. Method B analyzes
OCs with at least 50 members independently. The individual outcome parameters $a$ and $b$ are
later combined into a \textit{final} result. The analysis and results for each method are described below.
\subsection{Method A}
\begin{figure}[t!]
\plotone{figure3.pdf}
\caption{\textit{Upper panel}: Bimodal Histogram of GR + CS (km\,s$^{-1}$) for stellar members of 51 OCs
($N \geq 20$), where $N$ is the number of the stars (right Y-axis). The 2-peak distribution shows the
presence of the (two) main stellar groups, i.e. red-giants (left) and main-sequence stars (right).
\textit{Lower panel}: Linear Regression for $f(x) = \Delta RV$ as a function of $x = $ GR + CS, both in
(km\,s$^{-1}$). The blue dots represent stellar members of the 51 OCs. In both figures, the red line is
the result of a least-squares fit to the 4464 stars; the red diamonds are the binned GR + CS data
corresponding to the average value of $\Delta RV$, shown for illustrative purpose only.\label{fig:OLS1}}
\end{figure}
First, we ignore the CS contribution and consider our primary data set of 86 OCs and 8899 stars.
The aim is to prove, even if without the CS effect, that there is a not-null result when correlating
$\Delta RV$ with the theoretical GR. Moreover, the CS algorithm was developed specifically for the
\textit{Gaia} Radial Velocity Spectrometer \citep{2013A&A...550A.103A}, while the $RV$ data used have
multiple sources (see ref.\ in \citealt{2021A&A...647A..19T}). To avoid any possible bias from the
CS algorithm, it is important first to analyze the GR contribution alone.
By least-squares fitting, we perform a linear regression for the $\Delta RV$ as a function of GR
predictions. The results obtained are: $a = 0.425 \pm 0.077$ (slope) and $b = - 0.187 \pm 0.038$
(intercept, or the bias term). This results may be interpreted as a significant correlation between
the observed redshift and the $M/R$ ratio of the stars, although the amplitude of this correlation is
about half the value predicted by GRT. As in previous studies, we can associate this difference to
the neglected effect of convective shift. The value of $b$ is interpreted as the mean value of the bias
associated with the estimation of the radial velocity of each cluster.
In the following, we include in the analysis the CS contribution of each star using the hydrodynamic
model proposed by \citeauthor{2013A&A...550A.103A}, as explained in the previous section.
Figure \ref{fig:OLS1} (\textit{upper panel}) shows the histogram of the GR + CS distribution for these
stars. It exhibits a bimodal shape where the peaks correspond to red-giants (left) and main-sequence stars
(right). A linear regression was performed for the $\Delta RV$ as a function of the sum of GR and CS
predictions. Figure \ref{fig:OLS1} (\textit{lower panel}) shows the member stars combined into a single,
fictitious cluster and the regression line with a slope $a = 0.899 \pm 0.137$, and intercept $b = -0.359
\pm 0.058$. The value obtained for $a$ is compatible with unity and is therefore consistent with the
joint predictions of the GTR and convective shift model. The value of the $b$ parameter may be interpreted
as the shift due to the average bias in the estimates of the radial velocities of the clusters
(calculated ignoring the GR + CS term of each star). In fact, the mean $\overline{GR + CS}$ for the whole
data set (4,464 objects) is $0.383$ km\,s$^{-1}$, compatible with the absolute value of $|b| = 0.359$ km\,s$^{-1}$.
One of the biggest limitations of this model is the systematic bias that could arise when different
types of clusters are combined; for example, combining different clusters having only one-type
stellar populations (i.e., either red giants or main-sequence stars). In that case, when subtracting the
$RV_{\rm OC}$ (defined as the mean of $RV_{i}$, for all \textit{i}-member stars) and afterwards combining
the clusters, any substantial difference between the two stellar types is ruled out. Increasing the
number of stars per cluster ($N$) should avoid this systematic bias. Actually, performing a quick analysis
varying the number of stars per combined cluster, we realized that the slope $a$ converges to unity with
increasing $N$. A more conservative analysis would be to perform a linear fit per cluster with a higher $N$
threshold, and then combine the slope results. This analysis has been done in the following section.
\subsection{Method B}
The systematic bias introduced in the estimation of the radial velocity of each OC is a major drawback of Method A.
This has motivated the following analysis in which each OC is analyzed individually by estimating the
parameters $a_j$ and $ b_j$ for each \textit{j}-OC and then combining their results. In this analysis,
the bias in the estimation of the radial velocity of each OC is absorbed in the parameter $b_j$ and does not affect the slopes $a_j$.
\begin{figure}[t!]
\makebox[\textwidth][c]{\includegraphics[width=\textwidth]{figure4.pdf}}
\caption{Linear-fit slope ($a$)
as a function of $N$ stars per cluster. The dashed line
is the weighted mean slope ($a$) from the 28 OC ($N\geq 50$) results, with a value of
$0.977 \pm 0.218$. The blue shadow region is the $ \pm \sigma $ cumulative weighted average of each OC
slope result. Note how the the results converge smoothly to the expected value ($a \sim 1$) with
increasing $N$. The gray shadow region is the the $ \pm \sigma $ cumulative weighted average with
decreasing $N$. The error bars are standard deviation errors per cluster. Note that the X-axis is in
log-scale for visual clarity. \label{fig:OLS2}}
\end{figure}
Numerous tests were carried out, changing the threshold for the minimum number of stars, particularly
the $N\geq 20$ assumed in Method A. The results obtained are robust against the specific
threshold adopted, although the parameters obtained for OCs with relatively low numbers of stars are
very noisy and contribute little to the average value. Therefore, a higher limit was applied for
this analysis. The following results correspond to a subsample of OCs with at least 50 stars
($N \geq 50$). This restriction reduced the original GR-sample to the \textit{best} 47 OCs with 7713 stars.
When the CS contribution is included in the analysis, the complete GR + CS sample is reduced to 28 OCs with 3779 member stars.
The $a_j$ and $b_j$ parameters obtained in the analysis of each OC were weighted according to their
respective errors $\varepsilon_{a_j}$ and $\varepsilon_{b_j}$. The averaged results, ignoring the CS
contribution, are $a = 0.643 \pm 0.164$ and $b = -0.257 \pm 0.067$. The $a$ value obtained is consistent
with the non-null hypothesis at $\sim$$4 \sigma$ level. Modeling the CS effect, the weighted average
results are $a = 0.977 \pm 0.218$ and $b = -0.395 \pm 0.082$.
Figure \ref{fig:OLS2} and Table \ref{tab:results}
show the $a_j$ results and their standard errors for each OC when considering the CS contribution (the \textit{best}-cluster subsample).
The blue- and gray-shaded regions enclose the $a \pm \varepsilon_a$ values obtained from the
cumulative mean along the increasing (blue) and decreasing (gray) $N$ values.
In both cases it is worth noting the progressive and smooth convergence to the predicted value $a=1$.
Moreover, splitting the data between the most and least populated clusters, should reproduce compatible
results within the expected outcome. Combining only the first 18 OCs ($N \leq 150$), we get
$a = 1.14 \pm 0.32$, while the last 10 OCs ($N > 150$) give $a = 1.01 \pm 0.32$. Therefore, the
analysis of two independent data sets successfully reproduces the predictions.
Although the results for all the OCs show compatibility within the $\pm 3\sigma$ with the $a=1$ value,
the uncertainties in their respective estimations are relatively large and render it impossible to
reject the null hypothesis in the analysis of single OCs.
Interestingly, the result found in the widely analyzed M67 (NGC 2682) cluster ($ a = 1.499 \pm 0.587$, or $1.178 \pm 0.388$ without modelling the CS contribution) is in tension with previous studies \citep{2011A&A...526A.127P}.
In general, the error bars show a strong
anticorrelation with the number of stellar members of each OC. However, some of the cases show
comparatively higher than expected slopes with low error bars (i.e., Trumpler 5). This could be a consequence
of a low ratio of dwarf or giants, kinematic segregation between types, or irregularities in their spatial
or kinematic distributions. These hypothesis will require further in-depth study that is beyond the scope of this paper.
\begin{deluxetable*}{lCLCLlRLC}[htb!]
\tablecolumns{9}
\tablewidth{0pt}
\tablecaption{Fit results and parameters$^i$ for OCs with at least 50 members. \label{tab:results}}
\tablehead{
\colhead{\vspace{-0.2cm} Cluster } & \colhead{a} & \colhead{$\varepsilon_{a}$} & \colhead{b} & \colhead{$\varepsilon_{b}$} & \colhead{N} & \colhead{$RV_{OC}$} & \colhead{$\varepsilon_{RV_{OC}}$} & \colhead{$\sigma_{OC}$}
\vspace{-0.2cm}
\\
}
\startdata
Trumpler\_5 & 2.409 & 0.587 & -1.014 & 0.268 & 478 & 51.889 & 0.095 & 2.293 \\
NGC\_2141 & 1.164 & 0.533 & -0.479 & 0.229 & 277 & 26.802 & 0.078 & 1.462 \\
NGC\_2420 & 1.360 & 0.393 & -0.671 & 0.187 & 240 & 74.939 & 0.049 & 0.839 \\
NGC\_2682 & 1.499 & 0.587 & -0.640 & 0.251 & 238 & 34.451 & 0.063 & 1.093 \\
NGC\_2243 & 0.220 & 0.452 & -0.117 & 0.200 & 217 & 59.929 & 0.064 & 1.048 \\
Berkeley\_39 & -0.575 & 0.532 & 0.128 & 0.215 & 194 & 59.053 & 0.079 & 1.287 \\
Trumpler\_20 & 0.084 & 0.660 & 0.043 & 0.284 & 185 & -39.553 & 0.110 & 1.973 \\
NGC\_2158 & 3.261 & 0.959 & -1.167 & 0.383 & 171 & 28.678 & 0.186 & 2.694 \\
NGC\_6253 & 1.171 & 1.036 & -0.395 & 0.433 & 152 & -28.375 & 0.102 & 1.549 \\
NGC\_6705 & 1.820 & 1.138 & -1.083 & 0.536 & 151 & 36.081 & 0.149 & 2.597 \\
NGC\_2477 & 2.338 & 0.547 & -0.715 & 0.192 & 149 & 8.155 & 0.099 & 1.330 \\
Berkeley\_32 & -0.119 & 0.472 & 0.017 & 0.192 & 132 & 106.216 & 0.069 & 0.863 \\
NGC\_3532 & 2.477 & 1.154 & -1.265 & 0.587 & 131 & 5.803 & 0.067 & 1.479 \\
Collinder\_261 & 1.041 & 0.781 & -0.243 & 0.244 & 126 & -24.423 & 0.094 & 1.309 \\
NGC\_7789 & 2.640 & 4.879 & -0.192 & 0.437 & 109 & -54.134 & 0.124 & 1.499 \\
Berkeley\_36 & 1.404 & 1.237 & -0.408 & 0.390 & 80 & 63.235 & 0.163 & 1.704 \\
NGC\_2516 & -1.340 & 1.900 & 1.002 & 0.998 & 80 & 24.200 & 0.054 & 1.067 \\
Melotte\_71 & 2.680 & 1.083 & -1.133 & 0.482 & 76 & 51.811 & 0.173 & 1.745 \\
NGC\_2425 & -0.289 & 0.737 & 0.033 & 0.344 & 72 & 104.085 & 0.113 & 1.155 \\
Melotte\_22 & -2.000 & 2.258 & 1.283 & 1.201 & 70 & 6.077 & 0.051 & 0.998 \\
NGC\_6475 & 0.632 & 2.986 & -0.516 & 1.483 & 67 & -14.717 & 0.231 & 2.180 \\
NGC\_2632 & 1.581 & 0.936 & -1.102 & 0.482 & 63 & 35.459 & 0.071 & 0.963 \\
Collinder\_69 & -1.249 & 2.518 & 0.942 & 0.965 & 56 & 28.388 & 0.170 & 1.744 \\
Haffner\_10 & -0.153 & 0.695 & -0.002 & 0.285 & 56 & 87.999 & 0.114 & 0.946 \\
NGC\_6005 & 1.126 & 1.942 & -0.203 & 0.810 & 54 & -24.567 & 0.279 & 2.348 \\
Ruprecht\_147 & 0.800 & 1.042 & -0.323 & 0.534 & 52 & 42.397 & 0.127 & 1.093 \\
Ruprecht\_134 & 5.138 & 1.404 & -1.141 & 0.418 & 52 & -39.909 & 0.221 & 1.917 \\
NGC\_6819 & 3.346 & 5.965 & -0.309 & 0.588 & 51 & 2.811 & 0.213 & 1.592 \\
\enddata
\tablenotetext{i}{Values for $b$, $\varepsilon_b$, $RV$, $\varepsilon_{RV}$ and $\sigma_{OC}$ are in km\,s$^{-1}$.}
\tablecomments{The table shows the results obtained for the 28 OCs with at least 50 member stars. OCs are sorted in descending order with N.
Table 1 is published in its entirety in the electronic
edition of the {\it Astrophysical Journal}. A portion is shown here
for guidance regarding its form and content. The full version includes the results without modelling the CS.}
\end{deluxetable*}
\section{Conclusions}
The aim of this paper has been to estimate the gravitational redshift in non-degenerate stars, i.e., the light
shift when photons escape the gravitational potential of the star, as predicted by the Einstein's Equivalence Principle. This was done by analyzing astrometric, photometric, and spectroscopic information
provided mainly by the \textit{Gaia} survey of more than 70\,000 non-degenerate stars previously classified
as members of galactic open clusters. The study also takes into account information on the
spectral shift of the stars (i.e., radial velocities) provided by the ground-based surveys. Radial velocities
were assumed to be a combined result of kinematic factors (motion of the OC, peculiar velocity of the stars
within its own OC, and convective shift) and the gravitational redshift. When considering
each OC, the membership of the stars was re-estimated on the basis of their kinematic information. To ensure
the general validity of the results, a minimum number of restrictions on the properties of the OCs and their
members was assumed. That gives a baseline sample that contains 86 clusters and 8899 member stars. The
main conclusions of the study are the following:
\begin{itemize}
\item Ignoring the effect of convective shift, the estimated redshift, which depends on the M/R ratio, is lower in amplitude by roughly a factor of 1.5 than the theoretical expectations from EEP. Similar results were obtained regardless the applied method (A, B).
\item By estimating the contribution of convective shift, the original sample is reduced to 51
OCs with more than 20 members each, and a total of 4464 stars. The results show a correlation between
predictions and observations as quantified by a significant correlation $a = 0.899 \pm 0.137$.
This value is compatible to within 1$\sigma$ with the predictions (i.e. $a = 1$).
\item An improved method applied to the \textit{best}-cluster subsample containing 28 OCs ($N \geq 50$
each) and 3779 stars gives a correlation in accordance with the predictions; namely, $a = 0.977 \pm 0.218$.
\item When splitting the complete GR + CS sample of 28 OCs into two independent sets -- (i) OCs with lower
number of members ($N \leq 150$ per cluster) and (ii) OCs with higher number of members ($N > 150$ per cluster) -- the results are $a = 1.14 \pm 0.32$ and $a = 1.01 \pm 0.32$, respectively. This latter approach gives further confidence in the overall results.
\end{itemize}
In summary, the result is robust since it was extensively tested on a variety of constraints and restrictions
used to form the sample. This constitutes one of the widest and more conclusive estimations of gravitational redshift in non-degenerate
stars. The result shows the
potential of large astronomical observations in testing physical theories and is a step forward toward proving the universal
character of physical laws. Potentially, incorporating information on tangential velocities jointly
with the radial velocities provided by \textit{Gaia}-DR3, and other ground-based surveys, will significantly
enhance the sample. A better and wider spatial sampling will add further significance to the result of this work.
\begin{acknowledgments}
We thank Y. Tarricq for kindly providing the $RV$ data prior to the public release. We also thank
M. L\'opez-Corredoira and E. Mediavilla for useful comments and insights into several aspects such as methodology and interpretation of this study. We would like also to acknowledge an anonymous referee for his/her useful comments, and in particular those clarifying the scope of the results. This work has made use of data from the European Space Agency (ESA) mission
{\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia}Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC has been provided
by national institutions, in particular the institutions participating in the {\it Gaia} Multilateral
Agreement. This work was funded by the Agencia Estatal de Investigación (AEI), of the spanish Ministerio de Ciencia e Innovación (MICINN) under the project PID2019-110614GB-C21/AEI/10.13039/501100011033. N.R. also acknowledges support from the AEI and the European Social Fund (ESF) under grant \textit{Ayuda para
contratos predoctorales para la formación de doctores} with reference PRE2020-095880.
\end{acknowledgments}
\facilities{\textit{Gaia} DR2 \citep{2016A&A...595A...1G, 2018A&A...616A...1G}.}
\software{Astropy \citep{astropy:2013, astropy:2018}, Pandas \citep{reback2020pandas},
Matplotlib \citep{Hunter:2007}, Statsmodels \citep{seabold2010statsmodels}, AdjustText (\url{https://github.com/Phlya/adjustText}).
}
\newpage
| 8b3361ef2cb14e44ceda7a04d6bccb357d39f9d4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Polar auxin transport}
The phytohormone auxin is a central player in practically all aspects of the development and growth of plants, for example in phyllotaxis, root development and the initiation of lateral roots, the formation of vascular tissues in stems, the patterning of leaf veins, and flower development \cite{Paque2016}. The pattern formation principles underlying these developmental mechanisms have been uncovered to a large part through an intensive cross-talk between experimental approaches and mathematical modeling \cite{Shi2018,Autran2021,Cieslak2021}. Auxin is transported between cells and between cells and the cell walls both through diffusion and through transport proteins that are localized at the cell membrane of the cell. These transport proteins are distributed in a polarized manner inside the cells, and the polarization of adjacent cells is coordinated in plant tissue, leading to a directed transport of auxin through plant tissues in a mechanism called polar auxin transport (PAT) \cite{Adamowski2015}. For example, in fully developed seed plants, auxin is synthesized in leaves, then is transported through the central tissues of the stem and the root towards the root tips, where it redirected along the superficial tissues of the root back to towards the stem and recycled towards the internal tissues of the root \cite{Adamowski2015}.
Despite new details being uncovered incessantly (see e.g.~\cite{Hajny2020}), it is still incompletely understood what mechanisms drive the polarization of auxin transporters inside cells and the coordinated polarization among adjacent cells. In a series of classical experiments, Sachs applied artificial auxin to bean plants, and observed that these become the source new vascular tissue that then joins the existing vasculature; see e.g.~\cite{Sachs1975} and the review~\cite{Hajny2022}. These initial observations, together with the discovery of auxin transporters including PIN1 suggested that auxin drives the polarization of its own transporters, and hence the direction of its own transport (reviewed in~\cite{merks2007canalization,Hajny2022}). Initial models aimed to explain the formation of transport channels as observed in Sachs' experiments. These models therefore assumed that the rate of auxin flux from cell to cell further polarised auxin transport. This positive feedback led to the self-organised formation of auxin transport channels in a process called auxin canalisation. When it was realised that auxin accumulations mark the formation of a new leaves at the shoot apex, an alternative model was proposed, in which cells polarised towards the locally increased concentrations of auxin, thus forming self-organised accumulation of auxin \cite{Reinhardt:2003ww}. Mathematical models of the self-organisation of polar auxin transport therefore follow these two broad
categories. `With-the-gradient' models formalise the canalisation hypothesis and assume that the rate of cell polarisation depends
on the auxin \textit{flux} towards the relevant neighbour~\cite{Mitchison1980,mitchison1981polar,RollandLagan:2005ej,RollandLagan:2008db}. 'Up-the-gradient' models assume that PIN polarizes in the direction of neighbouring cells
at a rate that positively depends on the auxin \textit{concentration} in that neighbour~\cite{Jonsson.pnas06,Smith:2006du}. Attempts to reconcile these two seemingly contradicting ideas have followed two broad approaches. The first approach proposed that with-the-gradient and up-the-gradient models act at different positions of the plant or at different stages during development. For example Bayer et al.~\cite{Bayer:2009iq} proposed that the up-the-gradient model act at superficial tissue layers of the shoot apical meristem where it forms auxin accumulation points leading to the initial of new leaves. The deeper tissue layers could follow the with-the-gradient model channeling auxin away from the auxin accumulation point towards the vascular tissues \cite{Bayer:2009iq}. A similar approach was recently taken to explain the leaf venation patterning in combination with auxin convergence at the edge of the leaf primordium~\cite{Holloway2021}. The second approach looked for variants of the with-the-gradient or up-the-gradient models that could explain both auxin canalisation and auxin canalisation depending on the parameter settings. In this line of reasoning Walker et al. have proposed a with-the-gradient hypothesis for phyllotaxis \cite{Walker:2013wl}, whereas one of us has proposed an up-the-gradient hypothesis for canalisation~\cite{merks2007canalization}.
\subsection{Mathematical motivation}
In order to distinguish between the available phenomenological models of auxin-driven pattern formation and the general developmental principles that they represent, mathematical insight into the models' structure and the models' solutions will be crucial. This will help pinpoint key differences between the model structures and may uncover potential structural instabilities in the models upon which evolution may have acted, so as to produce new developmental patterning modules~\cite{Benitez2018}.
From the mathematical side,
almost all previous studies have focused on the types of patterns that can be generated by different models once the transitory dynamics have died out. An important example is the study
by Van Berkel and coworkers \cite{van2013polar},
where a number of models for polar auxin transport are recast into a common mathematical framework that allows them to be compared.
A steady state analysis for a general class of
active transport models can be found in \cite{draelants2015localized},
using advanced tools such as snaking from the field of bifurcation theory. Both periodic and stationary patterns are examined in \cite{allen2020mathematical}, where the authors consider an extended
`with-the-flux' model. Haskovec and his coworkers derive local and global existence results together with an appropriate continuum limit
for their graph-based diffusion model in \cite{haskovec2019auxin}.
Important qualitative examples of the with-the-gradient model are the formation of regularly spaced auxin maximums that lead to the growth of new leaves, as well as the formation of auxin channels that precede the formation of veins. Our goal here is to move beyond the well-studied equilibrium settings above and focus instead on understanding the dynamical behavior that leads to these patterns. In particular, we provide a rigorous framework to study a class of wave solutions that underpin the dynamical behaviour associated to with-the-gradient model. Ultimately, we hope that this analytic approach will provide an additional lens through which models of PAT can be examined and compared.
\subsection{The model}
\begin{figure}[t]
\centering\includegraphics[width=\textwidth]{figs/figTransport}
\caption[]{Schematic representation of the model \eqref{eqn: main system}.
Black arrows represent transport, red arrows describe polarization
and the green dashed arrows indication promotion. In particular, the PIN1
polarization rate correlates positively with the neighbouring auxin concentration,
making this a model of `up-the-gradient' type.
}
\label{fig:int:model}
\end{figure}
Inspired by \cite{Heisler2006,merks2007canalization}, the system we will study is given by
\begin{equation}\label{eqn: main system}
\begin{cases}
\dot{A}_j = T_{\act}\left(R_{j-1}\frac{A_{j-1}}{k_a+A_{j-1}} - R_j\frac{A_j}{k_a+A_j}\right) + T_{\diff}(A_{j+1}-2A_j+A_{j-1}), \\
\\
\dot{P}_j = -k_1 \frac{A_{j+1}}{k_r + A_{j+1}}\left(\frac{P_j}{k_m+P_j}\right) + \alpha{A}_j , \\
\\
\dot{R}_j = k_1\frac{A_{j+1}}{k_r+A_{j+1}}\left(\frac{P_j}{k_m+P_j}\right),
\end{cases}
\end{equation}
posed on the one-dimensional lattice $j \in \mathbb{Z}$;
see Fig. \ref{fig:int:model}. The variable $A_j(t)$ denotes the auxin concentration in cell $j \in \mathbb{Z}$, while $P_j(t)$ and $R_j(t)$ represent the unpolarized respectively right-polarized PIN1 in this cell.
The PIN1 hormone is the PIN-variant that is believed to play the most important role in auxin transport \cite{heisler2010alignment}.
The parameters appearing in the problem are all strictly positive and labelled in the same manner as in \cite{merks2007canalization}\footnote{For presentation purposes, the parameters $L$ and $r$ appearing in \cite{merks2007canalization} have been set to unity.}. In particular, $T_{\mathrm{act}}$ and $T_{\mathrm{diff}}$ denote the strengths of the active PIN1-induced rightward auxin transport and its diffusive counterpart, respectively. Unpolarized PIN1 is formed in the presence of auxin at a rate $\alpha$, while $k_1$ denotes the polarization rate. Finally, $k_a$, $k_r$, and $k_m$ are the Michaelis constants associated to the active transport of auxin and the polarization of PIN1, which depends on the auxin-concentration in the right-hand neighbouring cell. In particular, this model is of `up-the-gradient' type.
The main difference compared to \cite{merks2007canalization}
is that we are neglecting the presence of left-polarized PIN1
and have set the decay and depolarization rates of PIN1 to zero. Although this step of course imposes a pre-existing polarity on the system, we need to do this for technical reasons that we explain in the sequel. For now we simply point out that we wish to focus our attention on the dynamics of rightward auxin propagation, which takes place on timescales that are much faster than these decay and depolarization processes, and that the results will give novel insight into the full problem.
\begin{figure}[t]
\centering\includegraphics[width=\textwidth]{figs/figProfile}
\caption[]{Left: cartoon of the waveprofiles $(\phi_A, \phi_P, \phi_R)$, illustrating the definition of the width $w$ of the auxin-pulse
and the limits \eqref{eqn: profile limits}.
Right: numerical simulation of an auxin pulse passing through cell 25, leaving
a residue of (polarized) PIN1. We used the procedure described in {\S}\ref{sec:sub:int:mr}, with $A_1(0) = A_{\diamond} = 0.15$. The remaining parameters were fixed as $T_{\mathrm{act}} = 800$, $T_{\mathrm{diff}} = 0.15$,
$k_a = 1$, $k_m = k_r = 100$, $k_1 = 200$ and $\alpha = 0.1$.
}
\label{fig:int:profile}
\end{figure}
We will look for solutions of the special type
\begin{equation}
\label{eq:int:ansatz:a:p:r}
(A_j, P_j, R_j)(t) = (\phi_A, \phi_P, \phi_R)(j - c t),
\end{equation}
with $c > 0$, in which we impose the limits
\begin{equation}\label{eqn: profile limits}
\lim_{\xi \to - \infty} \phi_A(\xi) = 0,
\qquad \qquad
\lim_{\xi \to \infty} (\phi_A, \phi_P, \phi_R)(\xi) = 0.
\end{equation}
From a modelling perspective, such solutions represent a pulse of auxin that moves to the right through a one-dimensional row of cells. Ahead of the wave the cells are clear of both polarized and unpolarized PIN, but behind the wavefront a residual amount of PIN is left in the cells, representing the coordinated polarisation of the tissue.
In reality these residues start to depolarize and decay, which can be included by
adding linear decay terms to \eqref{eqn: main system}.
This leads to the expanded system
\begin{equation}\label{eqn:int:main:sys:expanded}
\begin{cases}
\dot{A}_j = T_{\act}\left(R_{j-1}\frac{A_{j-1}}{k_a+A_{j-1}} - R_j\frac{A_j}{k_a+A_j}\right) + T_{\diff}(A_{j+1}-2A_j+A_{j-1}), \\
\\
\dot{P}_j = -k_1 \frac{A_{j+1}}{k_r + A_{j+1}}\left(\frac{P_j}{k_m+P_j}\right) +\alpha{A}_j
+ k_2 R_j
- \delta P_j , \\
\\
\dot{R}_j = k_1\frac{A_{j+1}}{k_r+A_{j+1}}\left(\frac{P_j}{k_m+P_j}\right) - k_2 R_j,
\end{cases}
\end{equation}
in which the positive parameters $\delta$ and $k_2$ represent the decay and depolarization rate of PIN1, respectively.
Mathematically, these terms can be included into our framework
provided that the parameters $\delta$ and $k_2$ are small compared to the amplitude of the pulses, but we do not pursue this level of generality in the current paper for presentational clarity. Note in any case that in
\cite{merks2007canalization} these parameters were chosen to be orders of magnitude smaller than $\alpha$ and $k_1$.
\begin{figure}[t]
\centering\includegraphics[width=\textwidth]{figs/figScalings}
\caption[]{Scaling behaviour of the wavespeed $c$ (left) and the auxin width $w$ (right) against the height $h_A$ of the auxin pulse. The dashed lines represent the explicit predictions \eqref{eq:int:scaling:relations}.
The circles arise from numerical simulations, following the procedure
described in {\S}\ref{sec:sub:int:mr} with several
different values for $A_\diamond$. The other parameters were chosen as in Fig. \ref{fig:int:profile}.
}
\label{fig:int:scalings}
\end{figure}
Travelling waves have played a fundamental role in the analysis of
many spatially discrete systems \cite{kev,MPB,CGW08,HJHFZHNGM,VL28}.
They can be seen as a
lossless mechanism to transport matter or energy over arbitrary
distances. As such, they are interesting in their own right,
but they can also be viewed as building blocks to describe
more complicated behaviour of nonlinear systems \cite{AW75,AW78}.
In the present case for example, one can construct
wavetrain solutions to \eqref{eqn:int:main:sys:expanded}
by adding a persistent auxin source;
see Fig.~\ref{fig:int:wave:trains} and Supplementary Video S1.
Initially, these solutions
can be seen in an approximate sense as a concatenation of the
individual auxin pulses that we consider here \cite{moserBSc}. As a
consequence of the amplitude variations, small speed differences occur between these pulses which leads to highly interesting collision processes.
Due to this type of versatility, travelling waves play an
important role in many applications and have been extensively studied
in a variety of settings
\cite{sandstede2002stability,kev,
hochstrasser1989energy,jones1991construction}.
\subsection{Main results}
\label{sec:sub:int:mr}
Our goal will be to obtain quantitative scaling information concerning
the speed and shape of these waves. In particular, we will show rigorously
that \eqref{eqn: main system} admits a family of travelling wave solutions
that are parameterized by the amplitude of the auxin-pulse. In addition, we show that the speed and width of these waves scale with this amplitude
via a fractional power law.
We state our results in full technical detail in Corollary \ref{cor: main corollary} below.
\begin{figure}[t]
\centering\includegraphics[width=\textwidth]{figs/figLimits}
\caption[]{Convergence of the (scaled) profiles $\phi_A$ (left), $\phi_P$ (center) and $\phi_R$ (right) to their limits $(\phi_A^*, \phi_P^*, \phi_R^*)$. To perform the scalings, we wrote $h_A = \norm{\phi_A}_{L^\infty}$, compressed space by a factor of $h_A^{2/5}$ and divided the three profiles by the respective factors $(h_A, h_A^{1/5}, h_A^{2/5})$,
in line with the relations \eqref{eq:int:scaling:relations}.
}
\label{fig:int:limits}
\end{figure}
More precisely, we provide an explicit triplet of functions
$(\phi_A^*, \phi_P^*, \phi_R^*)$ that satisfy the limits
\eqref{eqn: profile limits}
and construct solutions to
\eqref{eqn: main system} of the form
\begin{equation}
\label{eq:int:mr:a:p:r:profiles:scales}
\begin{array}{lcl}
\big( A_j, P_j, R_j \big)(t)
&=& \Big( \epsilon \phi_A^*, \epsilon^{1/5} \phi_P^*, \epsilon^{2/5} \phi_R^* \Big)
\Big( \epsilon^{2/5}( j - c_* \epsilon^{2/5} t ) \Big)
\\[0.2cm]
& & \qquad \qquad \qquad
+ \Big( \mathcal{O}( \epsilon^{17/15}),
\mathcal{O}( \epsilon^{1/3} ),
\mathcal{O}( \epsilon^{3/5} )
\Big),
\end{array}
\end{equation}
for a constant $c_*$, which we state exactly in \eqref{eq:int:def:expl:constants}. Here the limiting profile $\phi_A^*$
is scaled in such a way that $\norm{ \phi_A^* }_{L^{\infty}} = 1$.
Upon introducing the heights\footnote{
Here we use the abbreviation
$\norm{A}_{\infty} = \textstyle{\sup_{j,t}} |A_j(t)|$ and its analogues for $P$ and $R$.
}
\begin{equation}
(h_A, h_P, h_R) = \big( \norm{A}_\infty, \norm{P}_\infty, \norm{R}_\infty \big)
\end{equation}
associated to the three components of our waves, this choice ensures that
the auxin-height $h_A$ is equal to the parameter $\epsilon>0$ at leading order.
In particular, comparing
this to \eqref{eq:int:ansatz:a:p:r} we uncover the
leading order scaling relations
\begin{equation}
\label{eq:int:scaling:relations}
c \sim c_* h_A^{2/5},
\qquad \qquad w \sim w_* h_A^{-2/5},
\qquad \qquad h_P \sim h_P^* h_A^{1/5},
\qquad \qquad h_R \sim h_R^* h_A^{2/5}
\end{equation}
for the speed $c$, width\footnote{We define the width of the auxin pulse as the distance between the two points where the pulse attains 5\%
of its maximum value.} $w$ and heights of the wave.
Here
the constant $w_*$ denotes the width of the limiting profile $\phi_A^*$, while the other constants are given explicitly by
\begin{equation}
\label{eq:int:def:expl:constants}
\begin{array}{lcl}
c_* & = & \left(\frac{9\alpha{k}_1T_{\act}T_{\diff}^2}{8k_ak_mk_r}\right)^{1/5},
\\[0.4cm]
h_P^* & = & \sqrt{6} \left(\frac{9\alpha^6 {k}_a^4 k_m^4 k_r^4 T_{\diff}^2}{8 k_1^4T_{\act}^4}\right)^{1/10},
\\[0.4cm]
h_R^* &= & 3\left(\frac{9\alpha k_a^4 {k}_1T_{\diff}^2}{8k_r k_m T_{\act}^4}\right)^{1/5}.
\end{array}
\end{equation}
In particular, for a fixed height of the auxin-pulse our results state that the speed and residual PIN1 will increase as the PIN1-production parameter $\alpha > 0$ is increased.
\begin{figure}[t]
\centering\includegraphics[width=\textwidth]{figs/figWaveTrains}
\caption[]{Six snapshots of a wavetrain simulation for the expanded system
\eqref{eqn:int:main:sys:expanded}. Higher pulses travel faster than lower pulses, in correspondence with the scaling relations \eqref{eq:int:scaling:relations}. These speed differences lead to merge events where even higher pulses are formed, which detach from the bulk.
We used the procedure described in {\S}\ref{sec:sub:int:mr}, taking $A_1(0) = A_{\diamond} = 0.0$ but adding $0.025$ to $\dot{A}_1(t)$ to simulate a constant auxin influx at the left boundary.
We picked $\delta = 0.1$ and $k_2 = 0.2$,
leaving the remaining parameters
from Fig. \ref{fig:int:profile} unchanged. The full simulation can be found in supplementary video S1.
}
\label{fig:int:wave:trains}
\end{figure}
Although our proof requires the parameter $\epsilon > 0$ and hence the amplitude of the auxin-pulses to be small,
this branch of solutions continues to exist well beyond this asymptotic regime.
Indeed, we numerically confirmed the existence (and stability) of these waves by a direct simulation of \eqref{eqn: main system}
on a row of cells $j \in \{1 , \ldots 500 \}$,
initialized with $A_j(0) = P_j(0) = R_j(0) =0$ for $2 \le j\le 500$,
together with $P_1(0) = R_1(0)=0$ and $A_1(0) = A_{\diamond}$ for some $A_{\diamond} > 0$ that we varied between simulations.
In order to close the system, we used the Neumann-type condition $A_0(t) = A_1(t)$ on the left-boundary, together with $R_0(t) = 0$ and a sink condition $A_{501}(t) = 0$
on the right. An example of such a simulation can be found
in Fig.~\ref{fig:int:profile} (right). By varying the initial auxin concentration $A_{\diamond}$, we were able to generate waves with a range of amplitudes. We subsequently numerically computed the speed and width of these waves, which allowed us to confirm the leading order behaviour \eqref{eq:int:scaling:relations}; see Fig.~\ref{fig:int:scalings}. In addition, we verified the
convergence to the limiting profiles $(\phi_A^*,\phi_P^*, \phi_R^*)$ by comparing the appropriately rescaled numerical waveprofiles; see
Fig.~\ref{fig:int:limits}.
\subsection{Cross-diffusion}
From a mathematical perspective, the problem
\eqref{eqn: main system} is interesting due to its interpretation
as a so-called cross-diffusion problem, where the transport coefficient
of one component is influenced by one of the other components. Work in this
area was stimulated by developments in the modeling of bacterial cell membranes \cite{SHIH20191469} and biofilms \cite{EmereniniBlessingO2015AMMo},
where self-organization of biological molecules plays an important role.
In the continuum regime, such problems are tough to analyze on account
of potential degeneracies in the coefficients. The well-posedness of the underlying problem was analyzed in
\cite{SonnerStefanie2011Otwo}, while a numerical method for such problems
was developed in \cite{ghasemi_sonner_eberl_2018}.
The key phenomenological assumption behind such models is that particles
behave differently when they are isolated compared to when they are
part of a cluster. A simplified agent-based approach to capture this
mechanism can be found in
\cite{johnston2017co}, which reduces naturally to a scalar
PDE with nonlinear diffusion in the continuum limit. After adding
a small regularization term, it is possible to
use geometric singular perturbation theory
to show that this PDE
admits travelling wave solutions \cite{LI2021132916}.
In this setting, the steepness of the wavefronts provides
the necessary scale-separation required for rigorous results.
Our approach in this paper proceeds along entirely different lines,
using the amplitude of the auxin pulse as a small continuation parameter
to construct a family of travelling wave solutions
to \eqref{eqn: main system}. The key insight is that one can extract an effective limiting system by scaling the width and speed of the wave
in an appropriate fashion and sending the amplitude to zero.
By means of a fixed-point analysis one
can show in a rigorous fashion that solutions to this limiting system
can be continued to form a family of solutions to the full system.
\subsection{Relation to FPUT pulses}
Our technique is a generalization of the approach developed
by Friesecke and Pego \cite{friesecke-pego1} to construct
small-amplitude travelling pulse solutions to the
Fermi-Pasta-Ulam-Tsingou (FPUT) problem
\cite{fput-original,dauxois}
\begin{equation}\label{newton}
\ddot{x}_j
= F(x_{j+1} - x_j) -F(x_{j} - x_{j-1}),
\qquad \qquad j \in \mathbb{Z}.
\end{equation}
This models an infinite, one-dimensional chain of particles that
can only move horizontally and are connected to their nearest neighbours by springs. These springs transmit a force
\begin{equation}
F(r) = r + r^2
\end{equation}
that hence depends nonlinearly on the relative distance $r$
between neighbouring particles; see \cite{friesecke-pego1, herrmann-matthies-asymptotic, pankov} for the impact of other choices.
The FPUT system is well-established as a fundamental model
to study the propagation of disturbances
through spatially discrete systems,
such as granular media, artificial metamaterials, DNA strands, and electrical transmission lines \cite{brill, kev}.
Looking for a travelling wave in the relative displacement coordinates, one introduces an Ansatz of the form
\begin{equation}
x_{j+1}(t) - x_j(t) = \phi(j - \sigma t ),
\end{equation}
which leads to the scalar functional differential equation of mixed type (MFDE)
\begin{equation}\label{mono:fput:tw:eqn}
\sigma^2 \phi''(\xi)
= F\big(\phi(\xi+1)\big) - 2 F\big(\phi(\xi)\big) + F\big(\phi(\xi - 1) \big) .
\end{equation}
Following the classic papers by Friesecke in combination with Wattis \cite{friesecke-wattis}
and Pego \cite{friesecke-pego1, friesecke-pego2, friesecke-pego3, friesecke-pego4},
we introduce the scaling
\begin{equation}
\phi(\xi) = \ep^2 \varphi_{\ep}(\ep \xi )
\end{equation}
and write $\sigma = \sigma_{\epsilon}$,
which transforms \eqref{mono:fput:tw:eqn} into the MFDE
\begin{equation}
\label{eq:int:fput:trv:wave:scaled}
\sigma_{\epsilon}^2 \epsilon^2 \varphi_{\epsilon}''
= \big( S^{\epsilon} + S^{-\epsilon} - 2 \big) \big[ \varphi_{\epsilon} + \epsilon^2 \varphi_{\epsilon}^2 \big] .
\end{equation}
Here the shift operator $S^d$ acts as
\begin{equation}\label{eqn: shift intro}
(S^d f)(\xi) = f(\xi + d)
\end{equation}
for any $d \in \mathbb{R}$. Since the symbol $S^{\epsilon} + S^{-\epsilon}- 2$ represents a discrete Laplacian, we can interpret \eqref{eq:int:fput:trv:wave:scaled} as a wave equation with a nonlinear diffusion term. To some extent, this clarifies the link with our original problem \eqref{eqn: main system} and the discussion above.
Applying the Fourier transform to \eqref{eq:int:fput:trv:wave:scaled}
with $k$ as the frequency variable, we arrive at
\begin{equation}
-\sigma_{\epsilon}^2 \epsilon^2 k^2 \hat{\varphi}_{\epsilon}(k) = 2(\cos (\epsilon k) -1) \big[ \hat{\varphi_{\epsilon}} + \epsilon^2\hat{\varphi_{\epsilon}^2} \big](k)
= - 4 \sin^2(\epsilon k/2) \big[ \hat{\varphi_{\epsilon}} + \epsilon^2 \hat{\varphi_{\epsilon}^2} \big](k).
\end{equation}
Upon introducing the symbol
\begin{equation}
\widetilde{\mathcal{M}}_{\mathrm{FPUT}}^{(\epsilon)}(k) =
\frac{4 \epsilon^2 \sin^2(\epsilon k/2)}{ \sigma_{\epsilon}^2 \epsilon^2 k^2 - 4 \sin^2(\epsilon k/2)},
\end{equation}
this can be recast into the compact form
\begin{equation}
\hat{\varphi_{\epsilon}}(k) = \widetilde{\mathcal{M}}_{\mathrm{FPUT}}^{(\epsilon)}(k) \hat{\varphi_{\epsilon}^2}(k).
\end{equation}
Upon choosing the speed
\begin{equation}\label{eqn: FP wave speed}
\sigma_{\ep} = 1 + \frac{\ep^2}{3},
\end{equation}
we can exploit the expansion
$\sin^2(z/2) = \frac{1}{4} z^2 - \frac{1}{48} z^4 + O(z^6)$
to obtain the pointwise limit
\begin{equation}
\widetilde{\mathcal{M}}_{\mathrm{FPUT}}^{(\epsilon)}(k) \to \frac{12}{8 + k^2}, \qquad \qquad \epsilon \to 0 .
\end{equation}
Using the fact that $(8 + k^2)$ is the Fourier symbol for $8 - \partial^2_{\xi}$, this suggests that the
relevant
system for $\varphi_{\epsilon}$ in the formal $\epsilon \to 0$ limit is given by
\begin{equation}
8 \varphi_* - \varphi_*'' = 12 \varphi_*^2,
\end{equation}
which has the nontrivial even solution
\begin{equation}\label{eqn: FP sech}
\varphi_*(\xi) =
\sech^2(\sqrt{2} \xi).
\end{equation}
By casting the problem in an appropriate
functional analytic framework, one can show
that this explicit solution $\varphi_*$ can be continued
to yield solutions $\varphi_{\epsilon}$
to \eqref{eq:int:fput:trv:wave:scaled} for small $\epsilon > 0$.
In this fashion, one establishes the existence of a family of pulse solutions \cite{friesecke-pego1}
\begin{equation}
x_{j+1}(t) - x_j(t)
= \epsilon^2 \sech^2\Big(\sqrt{2} \epsilon (j - \sigma_{\epsilon} t) \Big)
+ \mathcal{O} (\epsilon^4).
\end{equation}
Roughly speaking, the main mathematical contribution in this paper is that
we show how this analysis can be generalized to the setting of
\eqref{eqn: main system}. The first main obstacle is that this
is a multi-component system, which requires us to
explicitly reduce the order before a tractable limit can be obtained.
The second main obstacle is that the analysis of our Fourier symbol is
considerably more delicate, since in our setting the wavespeed $c$
converges to zero instead of one as $\ep \to 0$.
Indeed, the denominator
of $\widetilde{\mathcal{M}}_{\mathrm{FPUT}}^{(\epsilon)}$ above depends
only on the product $\epsilon k$, while in our case
there is a separate dependence on $\epsilon^2 k$.
This introduces a quasi-periodicity into the problem that requires
our convergence analysis to carefully distinguish between `small' values of $k$ and several separate regions of `large' $k$.
The third main difference is that we cannot use formal spectral arguments to analyze the limiting linear operator, which in our case is related to the Bernoulli equation. Instead, we apply a direct solution technique using variation-of-constants formulas. On the one hand this is much more explicit, but on the other hand the resulting estimates are rather delicate
on account of the custom function spaces involved.
\subsection{Discussion}
Due to the important organizing role that wave solutions often play in complex systems, scaling information such as \eqref{eq:int:scaling:relations} can be used as the starting point to uncover more general dynamical information concerning models
such as \eqref{eqn: main system} and related models of polar auxin trasnport. As such, we hope that the ideas
we present here will provide a robust analytical tool to analyze
different types of models as well. The resulting insights and predictions could
help to prioritize competing models on the basis of
dynamical experimental observations. Indeed, scaling laws
appear to play a role in
many aspects of biological systems,
such as the structural properties of vascular systems \cite{razavi2018scaling},
the mass dependence of metabolic rates \cite{west2004life}
and the functional constraints imposed by size
\cite{schmidt1984scaling}.
Although we have included only right-polarizing PIN in our system,
we believe that our techniques can be adapted to cover the full
case where also left-polarizing PIN is included. However, the
computations rapidly become unwieldy and the limiting system
is expected to differ qualitatively. For this reason, we have
not chosen to pursue this level of generality in the present paper,
as it would only obscure the main ideas behind our framework.
One of the main generalizations that we intend to pursue
in the future is to study the model in two spatial dimensions.
This is motivated by recent
numerical observations concerning the formation of auxin channels
and their associated PIN walls
under the influence of travelling patterns that are localized in
both spatial dimensions \cite{althuisBSc}.
\subsection{Notation}
We summarize a few aspects of our (mostly standard) notation.
\begin{enumerate}[label=$\bullet$]
\item
If $f = f(X)$ is a differentiable function on $\R$, then we sometimes write $f' = \partial_X[f]$.
\item
If $\X$ and $\mathcal{Y}$ are normed spaces, then we denote the space of bounded linear operators from $\X$ to $\Y$ by $\b(\X,\Y)$.
We put $\b(\X) := \b(\X,\X)$.
\end{enumerate}
\subsection{Acknowledgments}
HJH and TEF acknowledge support from the Netherlands Organization for Scientific Research (NWO) (grant 639.032.612).
\section{The Travelling Wave Problem}
\label{sec:travelingwave}
\subsection{Rewriting the original problem \eqref{eqn: main system}}
We will reduce the problem \eqref{eqn: main system} to a system of equations involving only $A_j$ and $P_j$, and it will be this resulting system on which we make the long wave-scaled travelling wave Ansatz.
\subsubsection{Changes of notation}
We begin by rewriting \eqref{eqn: main system} in a slightly more compressed manner that also exposes more transparently the leading order terms in the nonlinearities.
Let $\delta^{\pm}$ be the left and right difference operators that act on sequences $(x_j)$ in $\R$ via
\[
\delta^+x_j := x_{j+1}-x_j
\quadword{and}
\delta^-x_j := x_j-x_{j-1}.
\]
Next, for $k$, $x \in \R$ with $k+x \ne 0$ we have
\[
\frac{x}{k+x}
= \frac{x}{k} - \frac{x^2}{k(k+x)}.
\]
We put
\begin{equation}\label{eqn: Qsf1}
\Qsf_1(x,y)
:= \frac{x^2y}{k_a+x}
\end{equation}
and compress
\begin{equation}
\label{eq:twv:def:tau:1:2}
\tau_1 := \frac{T_{\act}}{k_a}
\quadword{and}
\tau_2 := T_{\diff}
\end{equation}
to see that our equation for $A_j$ now reads
\[
\dot{A}_j
= \tau_2\delta^+\delta^-A_j
- \tau_1\delta^-(R_jA_j)
+ \tau_1\delta^-\Qsf_1(A_j,R_j).
\]
Next, we abbreviate
\begin{equation}
\label{eq:twv:def:kappa}
\kappa
:= \frac{k_1}{k_rk_m}
\end{equation}
and put
\begin{equation}\label{eqn: Qsf2}
\Qsf_2(x,y)
:= \kappa\left(\frac{k_ry + k_mx + xy}{(k_r+x)(k_m+y)}\right)xy
\end{equation}
to see that, the equation for $P_j$ is
\[
\dot{P}_j
= -\kappa{A}_{j+1}P_j
+ \alpha{A}_j
+ \Qsf_2(A_{j+1},P_j).
\]
The equation for $R_j$ is updated similarly, and so we have rewritten \eqref{eqn: main system} as
\begin{equation}\label{eqn: main system2}
\begin{cases}
\dot{A}_j = \tau_2\delta^+\delta^-A_j - \tau_1\delta^-(R_jA_j) + \tau_1\delta^-\Qsf_1(A_j,R_j), \\
\\
\dot{P}_j = -\kappa{A}_{j+1}P_j + \alpha{A}_j + \Qsf_2(A_{j+1},P_j), \\
\\
\dot{R}_j = \kappa{A}_{j+1}P_j - \Qsf_2(A_{j+1},P_j).
\end{cases}
\end{equation}
We observe that the equation for $R_j$ depends only on $A_{j+1}$ and $P_j$ and therefore can be solved by direct integration.
Before we do that, however, we rewrite the new equation for $P_j$ using Duhamel's formula.
\subsubsection{Rewriting the $P_j$ equation}\label{sec: Pj analysis}
We can view the equation for $P_j$ in \eqref{eqn: main system2} as a first-order linear differential equation forced by $\alpha{A}_j + \Qsf_2(A_{j+1},P_j)$, and so we can solve it via the integrating factor method.
For $f$, $g \in L^1$ and $h \in L^{\infty}$ we introduce the operators
\begin{equation}\label{eqn: Esf}
\Esf(f)(s,t)
:= \exp\left(-\kappa\int_s^t f(\xi) \dxi\right), \ s, t \in \R,
\end{equation}
\begin{equation}\label{eqn: Psf1}
\Psf_1(f,g)(t)
:= \alpha\int_{-\infty}^t \Esf(f)(s,t)g(s) \ds,
\end{equation}
and
\begin{equation}\label{eqn: Psf2}
\Psf_2(f,h)(t)
:= \int_{-\infty}^t \Esf(f)(s,t)\Qsf_2(f(s),h(s)) \ds.
\end{equation}
Recall from \eqref{eqn: profile limits} that we want $P_j$ to vanish at $-\infty$.
The unique solution for $P_j$ in \eqref{eqn: main system2} that does vanish at $-\infty$ must satisfy
\[
P_j(t)
= \Psf_1(A_{j+1},A_j)(t) + \Psf_2(A_{j+1},P_j)(t).
\]
\subsubsection{Solving the $R_j$ equation}\label{sec: Rj analysis}
Since, per \eqref{eqn: profile limits}, we want $R_j$ to vanish at $-\infty$, and since we are assuming that each $A_j$ vanishes sufficiently fast at both $\pm\infty$ and $P_j$ vanishes at $-\infty$ and remains bounded at $+\infty$, we may solve for $R_j$ by integrating the third equation in \eqref{eqn: main system2} from $-\infty$ to $t$.
For $f$, $g \in L^1$ and $h \in L^{\infty}$, we define more integral operators:
\begin{equation}\label{eqn: Rsf1}
\Rsf_1(f,g)(t)
:= \kappa\tau_1\int_{-\infty}^t f(s)\Psf_1(f,g)(s) \ds, \ t \in \R,
\end{equation}
\begin{equation}\label{eqn: Rsf2}
\Rsf_2(f,g,h)(t)
:= \int_{-\infty}^t \big(\kappa{f}(s)\Psf_2(f,g)(s)-\Qsf_2(f(s),\Psf_1(f,g)(s)+\Psf_2(f,h)(s))\big) \ds,
\end{equation}
and
\begin{equation}\label{eqn: Rsf}
\Rsf(f,g,h)(t)
:= \Rsf_1(f,g)(t) + \Rsf_2(f,g,h)(t).
\end{equation}
We have defined $\Psf_1$ and $\Psf_2$ just above, respectively, in \eqref{eqn: Psf1} and \eqref{eqn: Psf2} and $\Qsf_2$ earlier in \eqref{eqn: Qsf2}.
Then the solution to the third equation in \eqref{eqn: main system2} that vanishes at $-\infty$ is
\begin{equation}\label{eqn: Rj}
R_j(t)
= \Rsf(A_{j+1},A_j,P_j)(t)
= \Rsf_1(A_{j+1},A_j)(t) + \Rsf_2(A_{j+1},A_j,P_j)(t).
\end{equation}
\subsubsection{The final system for $A_j$ and $P_j$}
We rewrite (part of) the $A_j$ equation once more to incorporate the new expression for $R_j$.
For $f$, $g \in L^1$ and $h \in L^{\infty}$ and $t \in \R$ put
\begin{equation}\label{eqn: Nsf}
\Nsf(f,g,h)(t)
:= \tau_1\Qsf_1(g(t),\Rsf(f,g,h)(t)) - \tau_1\Rsf_2(f,g,h)(t)g(t),
\end{equation}
where we defined $\Qsf_1$ in \eqref{eqn: Qsf1}.
Then $A_j$ must satisfy
\[
\dot{A}_j
= \tau_2\delta^+\delta^-A_j - \delta^-\big(\Rsf_1(A_{j+1},A_j)A_j\big) + \delta^-\Nsf(A_{j+1},A_j,P_j),
\]
and so our system for $A_j$ and $P_j$ is now
\begin{equation}\label{eqn: main system3}
\begin{cases}
\dot{A}_j = \tau_2\delta^+\delta^-A_j - \delta^-\big(\Rsf_1(A_{j+1},A_j)A_j\big) + \delta^-\Nsf(A_{j+1},A_j,P_j), \\
\\
P_j = \Psf_1(A_{j+1},A_j) + \Psf_2(A_{j+1},P_j).
\end{cases}
\end{equation}
That is, using the formula \eqref{eqn: Rj} for $R_j$ in terms of $A_j$ and $P_j$, we can solve \eqref{eqn: main system2} if we can solve \eqref{eqn: main system3}.
We will make two changes of variables on \eqref{eqn: main system3}.
First, in Section \ref{sec: tw}, we make a travelling wave Ansatz for $A_j$ and $P_j$.
We reformulate \eqref{eqn: main system3} for the travelling wave profiles as the system \eqref{eqn: tw syst fm} below.
Then, in Section \ref{sec: lw}, we introduce our long wave scaling on these travelling wave profiles.
After numerous adjustments, we arrive at the final system \eqref{eqn: lw syst-ep} for the scaled travelling wave profiles, which we solve in Section \ref{sec: lw syst-nu soln}.
The reader uninterested in these intermediate stages may wish to proceed directly to Theorem \ref{thm: lw tw syst final}, which discusses the equivalence of the problem \eqref{eqn: main system3} for $A_j$ and $P_j$ and the ultimate long wave system \eqref{eqn: lw syst-nu}.
Of course, our notation must keep up with these changes of variables, and we summarize in Table \ref{table: notation} the evolution of a typical operator's typesetting across these different problems.
\begin{table}
\[
\begin{tabular}{c|l}
\hline
Symbol &Use \\
\hline\\[-10pt]
$\Rsf$ &The original problem \eqref{eqn: main system3} \\
\hline\\[-10pt]
$\tRsf^c$ &The travelling wave problem \eqref{eqn: tw syst fm} \\[3pt]
\hline\\[-10pt]
$\bRsf^{\ep}$ &The preliminary long wave problem \eqref{eqn: lw syst-ep} \\[3pt]
\hline\\[-10pt]
$\Rcal^{\nu}$ &The final long wave problem \eqref{eqn: lw syst-nu} \\[3pt]
\hline
\end{tabular}
\]
\caption{Summary of notational evolution.}
\label{table: notation}
\end{table}
\begin{remark}\label{rem: speed of sound}
The linearization of \eqref{eqn: main system3} at 0 yields
\[
\dot{A}_j
= \tau_2\delta^+\delta^-A_j,
\qquad
P_j = R_j = 0.
\]
If we follow the discussion after \cite[Thm.\@ 1.1]{friesecke-pego1}, as well as \cite[Rem.\@ 2.2]{faver-wright}, and look for plane wave solutions $A_j(t) = e^{ikj-i\omega{t}}$ with $\omega$, $k \in \R$, we find the dispersion relation
\begin{equation}\label{eqn: disp rel}
-i\omega
= 2\tau_2(\cos(k)-1).
\end{equation}
The only real solutions are $\omega=0$ and $k \in 2\pi\Z$.
Previously, in \cite{friesecke-pego1, faver-wright} a nontrivial dispersion relation $\omega = \omega(k)$ was found by making the same kind of plane wave Ansatz, and the result `phase speed' $k \mapsto \omega(k)/k$ had a nonzero maximum $c_s$, which was called the `speed of sound.'
These articles then proceeded to look for travelling waves with speed slightly above their respective values of $c_s$; these were `supersonic' waves.
For us, $\omega(k)$ is identically zero, which suggests that the speed of sound for our auxin problem is 0.
Our long wave scaling in Section \ref{sec: lw} analytically justifies this intuition.
\end{remark}
\subsection{The travelling wave Ansatz}\label{sec: tw}
We now look for solutions $A_j$ and $P_j$ to \eqref{eqn: main system3} of the form
\begin{equation}\label{eqn: tw ansatz}
A_j = \phi_1(j-ct)
\quadword{and}
P_j = \phi_2(j-ct).
\end{equation}
The profiles $\phi_1$ and $\phi_2$ are real-valued functions of a single real variable and $c \in \R$.
The following manipulations will be justified if we assume $\phi_1 \in H_q^1$ and $\phi_2 \in W^{1,\infty}$; we discuss the exponentially localized Sobolev space $H_q^1$ in Appendix \ref{app: expn loc Sob space}.
Furthermore, since we want $P_j$ to vanish at $-\infty$ and be asymptotically constant at $+\infty$, per the limits \eqref{eqn: profile limits} and the numerical predictions of Fig.~\ref{fig:int:profile}, we expect that $\phi_2$ should vanish at $+\infty$ and be asymptotically constant at $-\infty$.
We will convert the problem \eqref{eqn: main system3} for $A_j$ and $P_j$ into a nonlocal system for $\phi_1$ and $\phi_2$, with $c$ as a parameter.
Doing so amounts to little more than changing variables {\it{many}} times in the integral operators defined in Sections \ref{sec: Pj analysis} and \ref{sec: Rj analysis} and gives us a host of new integral operators that will constitute the problem for $\phi_1$ and $\phi_2$.
In what follows we assume $f \in L^1$ and $g \in L^{\infty}$, so that the operators below are defined in the special cases of $f = \phi_1 \in H_q^1$ and $g = \phi_2 \in W^{1,\infty}$.
First, for $x$, $v \in \R$, put
\begin{equation}\label{eqn: tEsf-c}
\tEsf^c(f)(v,x)
:= \exp\left(\frac{\kappa}{c}\int_v^x f(u+1) \du\right)
\end{equation}
and
\begin{equation}\label{eqn: tPsf1-c}
\tPsf_1^c(f)(x)
:= \frac{\alpha}{c}\int_x^{\infty} \tEsf^c(f)(v,x)f(v) \dv.
\end{equation}
Then we use the Ansatz \eqref{eqn: tw ansatz} and the definition of $\Psf_1$ in \eqref{eqn: Psf1} to find
\[
\Psf_1(A_{j+1},A_j)(t)
= \alpha\int_{-\infty}^t \exp\left(-\kappa\int_s^t \phi_1(j-c\xi+1)\dxi\right)\phi_1(j-cs) \ds
= \tPsf_1^c(\phi_1)(j-ct).
\]
Here we have substituted $u = j-c\xi$ in the exponential's integral and then $v = j-cs$ throughout.
Similar substitutions, which we do not discuss, yield the following identities.
Put
\begin{equation}\label{eqn: tPsf2-c}
\tPsf_2^c(f,g)(x)
:= \frac{1}{c}\int_x^{\infty} \tEsf^c(f)(v,x)\Qsf_2(f(v+1),g(v)) \dv,
\end{equation}
so that with $\Psf_2$ defined in \eqref{eqn: Psf2} we have
\[
\Psf_2(A_{j+1},P_j)(t)
= \tPsf_2^c(\phi_1,\phi_2)(j-ct).
\]
Thus $\phi_2$ must satisfy
\begin{equation}\label{eqn: tw eqn for phi2}
\phi_2
= \tPsf_1^c(\phi_1) + \tPsf_2^c(\phi_1,\phi_2),
\end{equation}
which indicates that, as expected, $\phi_2$ should vanish at $+\infty$ and be asymptotically constant at $-\infty$.
Now we reformulate the equation for $A_j$, equivalently, for $\phi_1$.
Put
\begin{equation}\label{eqn: tRsf1-c}
\tRsf_1^c(f)(x)
:= \frac{\kappa\tau_1}{c}\int_x^{\infty} f(u+1)\tPsf_1^c(f)(u) \du,
\end{equation}
so that with $\Rsf_1$ defined in \eqref{eqn: Rsf1} we have
\[
\Rsf_1(A_{j+1},A_j)(t)
= \tRsf_1^c(\phi_1)(j-ct).
\]
Put
\begin{equation}\label{eqn: Rsf2-c}
\tRsf_2^c(f,g)(x)
:= \frac{1}{c}\int_x^{\infty} \big(\kappa{f}(u+1)\tPsf_2^c(f,g)(u) - \Qsf_2(f(u+1),g(u))\big) \du
\end{equation}
and
\begin{equation}\label{eqn: Rsf-c}
\tRsf^c(f,g)
:= \tRsf_1^c(f) + \tRsf_2^c(f,g),
\end{equation}
so that with $\Rsf_2$ defined in \eqref{eqn: Rsf2} and $\Rsf$ in \eqref{eqn: Rsf} we have
\[
\Rsf_2(A_{j+1},A_j,P_j)(t) = \tRsf_2^c(\phi_1,\phi_2)(j-ct)
\quadword{and}
\Rsf(A_{j+1},A_j,P_j)(t) = \tRsf^c(\phi_1,\phi_2)(j-ct).
\]
Last, put
\begin{equation}\label{eqn: Nsf-c}
\tNsf^c(f,g)(x)
:= \tau_1\tRsf_2^c(f,g)(x)f(x)-\tau_1\Qsf_1(f(x),\tRsf^c(f,g)(x)),
\end{equation}
so that with $\Nsf$ defined in \eqref{eqn: Nsf} we have
\[
\Nsf(A_{j+1},A_j,P_j)(t)
= \tNsf^c(\phi_1,\phi_2)(j-ct).
\]
For a function $f \colon \R \to \R$ and $d \in \R$, define, as in \eqref{eqn: shift intro}, the shift operator $S^d$ by
\begin{equation}\label{eqn: shift operator}
(S^df)(x)
:= f(x+d).
\end{equation}
This final piece of notation, along with the equation \eqref{eqn: tw eqn for phi2}, allows us to convert the problem \eqref{eqn: main system3} for $A_j$ and $P_j$ into the following nonlocal system for $\phi_1$ and $\phi_2$:
\begin{equation}\label{eqn: tw syst}
\begin{cases}
-c\phi_1'
= \tau_2(S^1-2+S^{-1})\phi_1
+ (S^{-1}-1)\big(\tRsf_1^c(\phi_1)\phi_1+ \tNsf^c(\phi_1,\phi_2)\big), \\
\\
\phi_2
= \tPsf_1^c(\phi_1) + \tPsf_2^c(\phi_1,\phi_2).
\end{cases}
\end{equation}
\subsection{The Fourier multiplier structure}
We summarize our conventions and definitions for Fourier transforms and Fourier multipliers in Appendix \ref{app: fourier}.
If we take the Fourier transform of the equation for $\phi_1$ in \eqref{eqn: tw syst}, we find
\[
\big(ick + 2\tau_2(\cos(k)-1)\big)\hat{\phi}_1(k)
= (1-e^{-ik})\ft\big[\tRsf_1^c(\phi_1)\phi_1+ \tNsf^c(\phi_1,\phi_2)\big](k).
\]
For $k \in \R$, we have $ick+2\tau_2(\cos(k)-1) = 0$ if and only if $k=0$.
Consequently, the function
\begin{equation}\label{eqn: tMc}
\tMsf_c(k)
:= \frac{1-e^{-ik}}{ick+2\tau_2(\cos(k)-1)}
\end{equation}
has a removable singularity at 0 and is in fact analytic on $\R$.
We therefore define $\Msf_c$ to be the Fourier multiplier with symbol $\tMsf_c$, i.e., $\Msf_c$ satisfies
\[
\hat{\Msf_cf}(k)
= \tMsf_c(k)\hat{f}(k).
\]
We discuss some further properties of Fourier multipliers in Appendix \ref{app: fm}.
Now the problem \eqref{eqn: tw syst} is equivalent to
\begin{equation}\label{eqn: tw syst fm}
\begin{cases}
\phi_1 = \Msf_c\big(\tRsf_1^c(\phi_1)\phi_1 + \tNsf^c(\phi_1,\phi_2)\big) \\
\phi_2 = \tPsf_1^c(\phi_1) + \tPsf_2^c(\phi_1,\phi_2).
\end{cases}
\end{equation}
\section{The Long Wave Problem}
\subsection{The long wave scaling}\label{sec: lw}
We now make the long wave Ansatz
\begin{equation}\label{eqn: lw ansatz}
\phi_1(x) = \ep\psi_1(\ep^{\mu}x),
\qquad
\phi_2(x) = \ep^{\beta}\psi_2(\ep^{\mu}x),
\quadword{and}
c = \ep^{\gamma}c_0.
\end{equation}
We assume, as with $\phi_1$ and $\phi_2$, that the scaled profiles satisfy $\psi_1 \in H_q^1$ and $\psi_2 \in W^{1,\infty}$.
We think of $\ep > 0$ as small and keep the exponents $\beta$, $\gamma$, $\mu > 0$ arbitrary for now; eventually we will pick
\[
\gamma = \mu = \frac{2}{5}
\quadword{and}
\beta = \frac{1}{5}.
\]
The reasoning behind this choice is by no means obvious at this point and will not be for some time; leaving $\mu$, $\beta$, and $\gamma$ arbitrary will allow this choice to appear more naturally (at the cost of temporarily more cumbersome notation).
As we intuited in Remark \ref{rem: speed of sound}, our wave speed is now close to 0, which is the auxin problem's natural `speed of sound.'
The parameter $c_0$ affords us some additional flexibility in choosing the wave speed.
A properly chosen value of $c_0$ will cause the maximum of the leading-order term of $\phi_1$ to be $\ep$, which will fulfill our promise in Section \ref{sec:sub:int:mr} that the auxin-height is, to leading order, $\ep$.
Friesecke and Pego introduce a similar auxiliary parameter into their $\ep$-dependent wave speed, see \cite[Eq.\@ (2.5), (2.13)]{friesecke-pego1}.
This parameter allows them to prove that the dependence of their travelling wave profile on wave speed is sufficiently regular in different function spaces, a result needed for their subsequent stability arguments in \cite{friesecke-pego2, friesecke-pego3, friesecke-pego4}.
We did not provide this extra parameter in our version \eqref{eqn: FP wave speed} of the Friesecke-Pego wave speed, but rather we selected it so that the amplitude of the leading order $\sech^2$-profile term in \eqref{eqn: FP sech} is 1.
Similarly, we will not pursue their depth of wave-speed analysis on our profiles' dependence on $c_0$.
We convert \eqref{eqn: tw syst fm} to another nonlocal system for $\psi_1$ and $\psi_2$, which now depends heavily on the parameter $\ep$.
As before, this process mostly amounts to changing variables in many integrals.
For example, we use the definition of $\tPsf_1^c$ in \eqref{eqn: tPsf1-c} and the Ansatz \eqref{eqn: lw ansatz} to find
\begin{equation}\label{eqn: tPsf1-c lw intermed}
\tPsf_1^c(\phi_1)(x)
= \frac{\alpha}{\ep^{\gamma}c_0}\int_x^{\infty} \tEsf^c(\phi_1)(v,x)\ep\psi_1(\ep^{\mu}v) \dv,
\end{equation}
where, using the definition of $\tEsf^c$ in \eqref{eqn: tEsf-c}, we have
\[
\tEsf^c(\phi_1)(v,x)
= \exp\left(\frac{\kappa}{\ep^{\gamma}c_0}\int_v^x \ep\psi_1(\ep^{\mu}u + \ep^{\mu}) \du\right) \\
= \exp\left(\frac{\kappa}{c_0}\ep^{1-(\gamma+\mu)}\int_{\ep^{\mu}v}^{\ep^{\mu}x} \psi_1(U + \ep^{\mu}) \dU\right).
\]
Here we have substituted $U = \ep^{\mu}u$.
Now for $f \in L^1$ we put
\begin{equation}\label{eqn: E}
\E(f)(V,X)
:= \exp\left(\frac{\kappa}{c_0}\int_V^X f(U) \dU\right), \ V, X \in \R,
\end{equation}
so that \eqref{eqn: tPsf1-c lw intermed} becomes
\[
\tPsf_1^c(\phi_1)(x)
= \frac{\alpha}{c_0}\ep^{1-\gamma}\int_x^{\infty} \E(\ep^{1-(\gamma+\mu)}S^{\ep^{\mu}}\psi_1)(\ep^{\mu}v,\ep^{\mu}x)\psi_1(\ep^{\mu}v) \dv.
\]
Here $S^{\ep^{\mu}}$ is the shift operator defined in \eqref{eqn: shift operator} with $d = \ep^{\mu}$.
We substitute again with $V = \ep^{\mu}v$ and define
\begin{equation}\label{eqn: bPsf1-ep}
\bPsf_1^{\ep}(f)(X)
:= \frac{\alpha}{c_0}\int_X^{\infty} \E(\ep^{1-(\gamma+\mu)}S^{\ep^{\mu}}f)(V,X)f(V)\dV
\end{equation}
to conclude that
\[
\tPsf_1^c(\phi_1)(x)
= \ep^{1-(\gamma+\mu)}\bPsf_1^{\ep}(\psi_1)(\ep^{\mu}x).
\]
Similar careful substitutions will allow us to reformulate the integral operators from Section \ref{sec: tw} in terms of the long wave Ansatz.
First, however, we define
\begin{equation}\label{eqn: bQsf-ep}
\bQsf_1^{\ep}(X,Y) := \frac{X^2Y}{k_a+\ep{X}}
\quadword{and}
\bQsf_2^{\ep}(X,Y) := \kappa\frac{k_rY+k_m\ep^{1-\beta}X+\ep{XY}}{(k_r+\ep{X})(k_m+\ep^{\beta}Y)}XY .
\end{equation}
When $\ep \ne 0$, this definition permits the very convenient factorizations
\[
\Qsf_1(\ep{X},\ep^{1-(\gamma+\mu)}Y)
= \ep^{3-(\gamma+\mu)}\bQsf_1^{\ep}(X,Y)
\quadword{and}
\Qsf_2(\ep{X},\ep^{\beta}Y)
= \ep^{1+2\beta}\bQsf_2^{\ep}(X,Y),
\]
where $\Qsf_1$ was defined in \eqref{eqn: Qsf1} and $\Qsf_2$ in \eqref{eqn: Qsf2}.
Now we work on the travelling wave integral operators.
Below we will assume $f \in L^1$ and $g \in L^{\infty}$.
Put
\begin{equation}\label{eqn: bPsf2-ep}
\bPsf_2^{\ep}(f,g)(X)
:= \frac{1}{c_0}\int_X^{\infty} \E(\ep^{1-(\gamma+\mu)}S^{\ep^{\mu}}f)(V,X)\bQsf_2^{\ep}(f(V+\ep^{\mu}),g(V)) \dV,
\end{equation}
so that with $\tPsf_2^c$ defined in \eqref{eqn: tPsf2-c} we have
\[
\tPsf_2^c(\phi_1,\phi_2)(x)
= \ep^{1-(\gamma+\mu)+2\beta}\bPsf_2^{\ep}(\psi_1,\psi_2)(\ep^{\mu}x).
\]
This converts the second equation in \eqref{eqn: tw syst fm} for $\phi_2$ to
\[
\ep^{\beta}\psi_2(\ep^{\mu}x)
= \ep^{1-(\gamma+\mu)}\bPsf_1^{\ep}(\psi_1)(\ep^{\mu}x) + \ep^{1-(\gamma+\mu)+2\beta}\bPsf_2^{\ep}(\psi_1,\psi_2)(\ep^{\mu}x).
\]
Passing to $X = \ep^{\mu}x$, we find that $\psi_2$ must satisfy
\begin{equation}\label{eqn: psi2 lw eqn intermed}
\psi_2(X)
= \ep^{1-(\gamma+\mu)-\beta}\bPsf_1^{\ep}(\psi_1)(X) + \ep^{1-(\gamma+\mu)+\beta}\bPsf_2^{\ep}(\psi_1,\psi_2)(X).
\end{equation}
Now put
\begin{equation}\label{eqn: bRsf1-ep}
\bRsf_1^{\ep}(f)(X)
:= \frac{\kappa\tau_1}{c_0}\int_X^{\infty} \bPsf_1^{\ep}(f)(V)f(V + \ep^{\mu}) \dV,
\end{equation}
so that with $\tRsf_1^c$ defined in \eqref{eqn: tRsf1-c} we have
\[
\tRsf_1^c(\phi_1)(x)
= \ep^{2(1-(\gamma+\mu))}\bRsf_1^{\ep}(\psi_1)(\ep^{\mu}x).
\]
Put
\begin{equation}\label{eqn: bRsf2-ep}
\bRsf_2^{\ep}(f,g)(X)
:= \frac{1}{c_0}\int_X^{\infty} \big(\ep^{1-(\gamma+\mu)}\kappa{f}(V+\ep^{\mu})\bPsf_2^{\ep}(f,g)(V)-\bQsf_2^{\ep}(f(V+\ep^{\mu}),g(V))\big) \dV
\end{equation}
and
\begin{equation}\label{eqn: bRsf-ep}
\bRsf^{\ep}(f,g)(X)
:= \ep^{1-(\gamma+\mu)}\bRsf_1^{\ep}(f)(X) + \ep^{2\beta}\Rsf_2^{\ep}(f,g)(X),
\end{equation}
so that with $\tRsf_2^c$ defined in \eqref{eqn: Rsf2-c} and $\tRsf^c$ defined in \eqref{eqn: Rsf-c} we have
\[
\tRsf_2^c(\phi_1,\phi_2)(x)
= \ep^{1-(\gamma+\mu)+2\beta}\bRsf_2^{\ep}(\psi_1,\psi_2)(\ep^{\mu}x)
\quadword{and}
\tRsf^c(\phi_1,\phi_2)(x)
= \ep^{1-(\gamma+\mu)}\bRsf^{\ep}(\psi_1,\psi_2)(\ep^{\mu}x).
\]
Finally, put
\begin{equation}\label{eqn: bNsf-ep}
\bNsf^{\ep}(f,g)(X)
:= \tau_1\bRsf_2^{\ep}(f,g)(X)f(X)
- \ep^{1-2\beta}\tau_1\bQsf_1^{\ep}(f(X),\bRsf^{\ep}(f,g)(X)),
\end{equation}
so that with $\tNsf^c$ defined in \eqref{eqn: Nsf-c} we have
\[
\tNsf^c(\phi_1,\phi_2)(x)
= \ep^{2-(\gamma+\mu)+2\beta}\bNsf^{\ep}(\psi_1,\psi_2)(\ep^{\mu}x).
\]
The definition of scaled Fourier multipliers from \eqref{eqn: scaled fm} tells us that, for $\ep > 0$, $\Msf_{\ep^{\gamma}c_0}^{(\ep^{\mu})}$ is the Fourier multiplier satisfying
\[
\hat{\Msf_{\ep^{\gamma}c_0}^{(\ep^{\mu})}f}(k)
= \tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)\hat{f}(k),
\]
where $\tMsf_{\ep^{\gamma}c_0}$ is defined by taking $c = \ep^{\gamma}c_0$ in \eqref{eqn: tMc}.
This converts the first equation in \eqref{eqn: tw syst fm} for $\phi_1$ to
\[
\ep\psi_1(\ep^{\mu}x)
= \Msf_{\ep^{\gamma}c_0}^{(\ep^{\mu})}[\ep^{2(1-(\gamma+\mu))}\bRsf_1^{\ep}(\psi_1)\ep\psi_1 + \ep^{2-(\gamma+\mu)+2\beta}\bNsf^{\ep}(\psi_1,\psi_2)](\ep^{\mu}x).
\]
We factor this to reveal
\begin{equation}\label{eqn: psi1 lw eqn intermed}
\psi_1(X)
= \ep^{2(1-(\gamma+\mu))}\Msf_{\ep^{\gamma}c_0}^{(\ep^{\mu})}\big[\bRsf_1^{\ep}(\psi_1)\psi_1 + \ep^{-1+\gamma+\mu+2\beta}\Nsf^{\ep}(\psi_1,\psi_2)\big](X).
\end{equation}
We abbreviate
\begin{equation}\label{eqn: bMu-ep}
\bMsf_{\ep}
:= \ep^{2(1-(\gamma+\mu))}\Msf_{\ep^{\gamma}c_0}^{(\ep^{\mu})}
\end{equation}
to conclude from \eqref{eqn: psi1 lw eqn intermed} and the prior equation \eqref{eqn: psi2 lw eqn intermed} for $\psi_2$ that the long wave profiles must satisfy
\begin{equation}\label{eqn: lw syst-ep}
\begin{cases}
\psi_1 = \bMsf_{\ep}\big[\bRsf_1^{\ep}(\psi_1)\psi_1 + \ep^{-1+\gamma+\mu+2\beta}\bNsf^{\ep}(\psi_1,\psi_2)\big] \\
\\
\psi_2
= \ep^{1-(\gamma+\mu+\beta)}\bPsf_1^{\ep}(\psi_1) + \ep^{1-(\gamma+\mu)+\beta}\bPsf_2^{\ep}(\psi_1,\psi_2).
\end{cases}
\end{equation}
We have been tacitly assuming that all of the exponents on powers of $\ep$ above are nonnegative so that the various $\ep$-dependent operators and prefactors are actually defined at $\ep=0$.
In particular, this demands
\begin{equation}\label{eqn: expn bare minimum}
1-2\beta \ge 0,
\qquad
-1+\gamma+\mu+2\beta \ge 0,
\quadword{and}
1-(\gamma+\mu+\beta) \ge 0.
\end{equation}
\subsection{The formal long wave limit and exponent selection}\label{sec: formal lw limit}
Our intention is now to take the limit $\ep \to 0$ in the equations \eqref{eqn: lw syst-ep} for $\psi_1$ and $\psi_2$.
Doing so in a way that the limit is both meaningful (i.e., defined and nontrivial) and reflective of what the numerics predict at $\ep=0$ will teach us what the exponents $\mu$, $\gamma$, and $\beta$ should be, beyond the requirements of \eqref{eqn: expn bare minimum}.
\subsubsection{The formal limit on $\bMsf_{\ep}$ and the selection of the exponents $\gamma$ and $\mu$}\label{sec: bMu-ep formal}
We want to assign a `natural' definition to $\bMsf_0$, where $\bMsf_{\ep}$ was defined, for $\ep > 0$, in \eqref{eqn: bMu-ep}.
However, we relied above on having $\ep > 0$ to invoke the scaled Fourier multiplier identity \eqref{eqn: scaled fm} that gave us $\bMsf_{\ep}$, and naively setting $\ep = 0$ in that identity is meaningless.
Additionally, we should be careful that the prefactor $\ep^{2(1-(\gamma+\mu))}$ in \eqref{eqn: bMu-ep} does not lead us to define $\bMsf_0 = 0$; otherwise, we would have $\psi_1 = 0$ when $\ep = 0$, and that is not what the numerics in Fig.~\ref{fig:int:profile} predict.
A natural starting point, then, is to study $\bMsf_{\ep}$ in the limit $\ep \to 0^+$, and this amounts to considering the limit of its symbol, whose definition we extract from the definition of $\bMsf_{\ep}$ in \eqref{eqn: bMu-ep} and the definition of the scaled Fourier multiplier in \eqref{eqn: scaled fm}.
Thus, for each $k \in \R$, we want the limit
\begin{equation}\label{eqn: Mc delta limit}
\lim_{\ep \to 0^+} \ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
\end{equation}
to exist without being identically zero.
The function $\tMsf_{\ep^{\gamma}c_0}$ was defined in \eqref{eqn: tMc}.
To calculate this limit, we first state the Taylor expansions
\begin{equation}\label{eqn: N}
1-e^{-iz} = iz+iz^2N_1(z)
\quadword{and}
\cos(z)-1 = -\frac{z^2}{2}+ \frac{iz^4N_2(z)}{2\tau_2}
\end{equation}
for $z \in \C$.
The functions $N_1$ and $N_2$ are analytic and uniformly bounded on strips in the sense that
\begin{equation}\label{eqn: Cq}
C_q
:= \sup_{x \in \R} |N_1(x\pm{iq})| + |N_2(x\pm{iq})|
< \infty
\end{equation}
for any $q > 0$.
The choice of constants on $N_1$ and $N_2$ will permit some useful cancellations later.
Then
\[
\tMsf_c(k)
= \frac{ik+ik^2N_1(k)}{ick -\tau_2k^2 + ik^4N_2(k)}
= \frac{1+kN_1(k)}{c+i\tau_2k + k^3N_2(k)},
\]
and so
\begin{equation}\label{eqn: Mc-ep}
\ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
= \ep^{2(1-(\gamma+\mu))}\frac{1+\ep^{\mu}kN_1(\ep^{\mu}k)}{\ep^{\gamma}c_0 + i\tau_2\ep^{\mu}k + \ep^{3\mu}k^3N_2(\ep^{\mu}k)}.
\end{equation}
At this point it does not make sense to set $\ep = 0$, as then the denominator would be identically zero.
So, we would like to factor some power of $\ep$ out of the denominator.
Since the first term in the denominator has a factor of $\ep^{\gamma}$ and the second a factor of $\ep^{\mu}$, we assume $\gamma = \mu$ and remove the power of $\ep$ from both the first and the second terms.
We discuss the choice of $\gamma=\mu$ further in Remark \ref{rem: why gamma=nu}.
Then
\begin{equation}\label{eqn: Mc-ep2}
\ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
= \ep^{2(1-2\gamma)}\tMsf_{\ep^{\gamma}c_0}(\ep^{\gamma}k)
= \ep^{2(1-2\gamma)-\gamma}\frac{1+\ep^{\gamma}kN_1(\ep^{\gamma}k)}{c_0+i\tau_2k+\ep^{2\gamma}k^3N_2(\ep^{\gamma}k)}.
\end{equation}
Pointwise in $k$ we have
\[
\lim_{\ep \to 0^+} \frac{1+\ep^{\gamma}kN_1(\ep^{\gamma}k)}{c_0+i\tau_2k+\ep^{2\gamma}k^3N_2(\ep^{\gamma}k)}
= \frac{1}{c_0+i\tau_2k},
\]
and so we want
\[
2(1-2\gamma)-\gamma
= 0
\]
so that the prefactor of $\ep^{2(1-2\gamma)-\gamma}$ in \eqref{eqn: Mc-ep2} does not induce a trivial or undefined limit.
Thus we take
\[
\gamma
= \mu
= \frac{2}{5}.
\]
Certainly doing so does not contradict any of the inequalities in \eqref{eqn: expn bare minimum}, provided that $\beta$ is chosen appropriately.
Moreover, the power of $2/5$ agrees with the height-speed-width relations suggested in Fig.\@ \ref{fig:int:scalings}.
And so
\[
\lim_{\ep \to 0^+} \ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
= \lim_{\ep \to 0^+} \ep^{4/5}\tMsf_{\ep^{2/5}c_0}(\ep^{2/5}k)
= \frac{1}{c_0+\tau_2ik}.
\]
Put
\begin{equation}\label{eqn: tvarpi-0}
\tM^{(0)}(z)
:= \frac{1}{c_0+i\tau_2z},
\end{equation}
so $\tM^{(0)}$ is analytic on any strip $\set{z \in \C}{|\im(z)| < q}$ for $q \in (0,\tau_2/c_0)$.
Let $\M^{(0)}$ be the Fourier multiplier with symbol $\tM^{(0)}$.
Lemma \ref{lem: Beale fm lemma} then gives the following properties of $\M^{(0)}$; the identities \eqref{eqn: varpi-0 ids} are direct calculations with the Fourier transform.
\begin{lemma}\label{lem: varpi0}
Fix $q \in (0,\tau_2/c_0)$.
Then $\M^{(0)} \in \b(H_q^r,H_q^{r+1})$ for all $r$.
More generally, if $f \in H^1$ and $g \in L^2$, then
\begin{equation}\label{eqn: varpi-0 ids}
\M^{(0)}(c_0+ \tau_2\partial_X)f = f
\quadword{and}
(c_0+\tau_2\partial_X)\M^{(0)}g = g.
\end{equation}
\end{lemma}
Because of the identities \eqref{eqn: varpi-0 ids}, we write $\M^{(0)} = (c_0+\tau_2\partial_X)^{-1}$.
The formal analysis above then leads us to expect
\begin{equation}\label{eqn: bMu-ep to 0}
\lim_{\ep \to 0^+} \bMsf_{\ep}
= \M^{(0)}
= (c_0+\tau_2\partial_X)^{-1}.
\end{equation}
However, we have not yet proved this rigorously by any means.
\begin{remark}\label{rem: why gamma=nu}
Here is why we take $\gamma=\mu$ when factoring the power of $\ep$ out of the denominator in \eqref{eqn: Mc-ep}.
First, taking $\gamma > \mu$ produces
\[
\ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
= \ep^{2(1-(\gamma+\mu))-\mu}\frac{1+\ep^{\mu}kN_1(\ep^{\mu}k)}{\ep^{\gamma-\mu}c_0 + i\tau_2k + \ep^{2\mu}k^3N_2(\ep^{\mu}k)}
\]
instead of \eqref{eqn: Mc-ep2}.
If $2(1-(\gamma+\mu))-\mu > 0$, then the right side above is identically zero at $\ep = 0$, and so we demand $2(1-(\gamma+\mu))-\mu = 0$; there are many pairs of $\gamma$ and $\mu$ that work here.
But then
\[
\lim_{\ep \to 0^+} \frac{1+\ep^{\mu}kN_1(\ep^{\mu}k)}{\ep^{\gamma-\mu}c_0 + i\tau_2k + \ep^{2\mu}k^3N_2(\ep^{\mu}k)}
= \frac{1}{i\tau_2k}.
\]
This suggests that instead of \eqref{eqn: bMu-ep to 0}, we have
\[
\lim_{\ep \to 0^+} \bMsf_{\ep}
= (\tau_2\partial_X)^{-1}.
\]
However, this is meaningless: differentiation is not invertible from $H_q^r$ to $H_q^{r+1}$.
Taking $\gamma < \mu$ also does not work.
In that case, instead of \eqref{eqn: Mc-ep2} we would have found
\[
\ep^{2(1-(\gamma+\mu))}\tMsf_{\ep^{\gamma}c_0}(\ep^{\mu}k)
= \ep^{2(1-(\gamma+\mu))-\gamma}\frac{1+\ep^{\mu}kN_1(\ep^{\mu}k)}{c_0 + i\tau_2\ep^{\mu-\gamma}k + \ep^{3\mu-\gamma}k^3N_2(\ep^{\mu}k)}.
\]
Since $\gamma < \mu$ we find
\[
\lim_{\ep \to 0^+} \frac{1+\ep^{\mu}kN_1(\ep^{\mu}k)}{c_0 + i\tau_2\ep^{\mu-\gamma}k + \ep^{3\mu-\gamma}k^3N_2(\ep^{\mu}k)}
= \frac{1}{c_0}.
\]
We would then want $2-3\gamma-2\mu = 0$ to prevent a nontrivial limit.
Choosing $\gamma$ and $\mu$ appropriately, we conclude that at $\ep = 0$ the equation for $\psi_1$ from \eqref{eqn: lw syst-ep} formally reduces to
\[
\psi_1
= \frac{1}{c_0}\bRsf_1^0(\psi_1)\psi_1.
\]
Numerically we expect $\psi_1(X) > 0$ for all $X$ when $\ep = 0$, and so, using the definition of $\bRsf_1^0$ from \eqref{eqn: bRsf1-ep}, we have
\[
c_0
= \bRsf_1^0(\psi_1)(X)
= \frac{\alpha\kappa\tau_1}{c_0^2}\int_X^{\infty} \left(\int_V^{\infty} \psi_1(W)\dW\right)\psi_1(V) \dV.
\]
Differentiating, we find
\[
\left(\int_X^{\infty}\psi_1(W) \dW\right)\psi_1(X)
= 0.
\]
But since $\psi_1(W) > 0$ for all $W$, we cancel the integral factor to find $\psi_1(X) = 0$, a contradiction to our numerical predictions.
\end{remark}
\subsubsection{The formal leading order equation for $\psi_1$}\label{sec: formal leading order1}
At $\ep = 0$ the equation for $\psi_1$ in \eqref{eqn: lw syst-ep} becomes (again, formally)
\[
\psi_1
= \M^{(0)}\big(\bRsf^0(\psi_1)\psi_1\big)
= (c_0+\tau_2\partial_X)^{-1}\big(\bRsf^0(\psi_1)\psi_1\big).
\]
This is equivalent to
\begin{equation}\label{eqn: psi1 nu=0}
c_0\psi_1+\tau_2\psi_1'
= \bRsf^0(\psi_1)\psi_1.
\end{equation}
We will rewrite this equation so that each term is a perfect derivative.
The definition of $\bRsf_1^{\ep}$ in \eqref{eqn: bRsf1-ep}, valid for all $\ep$, gives
\begin{equation}\label{eqn: bRsf0}
\bRsf^0(\psi_1)(X)
= \frac{\alpha\kappa\tau_1}{c_0^2}\int_X^{\infty} \left(\int_V^{\infty} \psi_1(W) \dW\right)\psi_1(V) \dV.
\end{equation}
Write
\[
\Psi_1(X)
:= \int_X^{\infty} \psi_1(W) \dW,
\]
so that $\Psi_1' = -\psi_1$.
The double integral from \eqref{eqn: bRsf0} is
\begin{multline*}
\int_X^{\infty} \left(\int_V^{\infty} \psi_1(W) \dW\right)\psi_1(V) \dV
= -\int_X^{\infty} \Psi_1(V)\Psi_1'(V) \dV
= -\int_X^{\infty} \partial_V\left[\frac{\Psi_1^2(V)}{2}\right] \dV \\
= \frac{\Psi_1(X)^2}{2}.
\end{multline*}
Here we are using the requirement that $\psi_1 \in H_q^1$, which implies $\Psi_1(X) \to 0$ as $X \to \infty$.
Thus
\[
\bRsf^0(\psi_1)\psi_1
= -\left(\frac{\alpha\kappa\tau_1}{2c_0^2}\right)\Psi_1^2\Psi_1'
= -\left(\frac{\alpha\kappa\tau_1}{6c_0^2}\right)\partial_X[\Psi_1^3] .
\]
Then \eqref{eqn: psi1 nu=0} is equivalent to
\[
\tau_2\Psi_1'' + c_0\Psi_1' - \left(\frac{\alpha\kappa\tau_1}{6c_0^2}\right)\partial_X[\Psi_1^3]
= 0.
\]
We integrate both sides from 0 to $\infty$ and use the aforementioned fact that $\Psi_1$ and its derivatives are required to vanish at $\infty$ to find
\begin{equation}\label{eqn: Bernoulli}
\tau_2\Psi_1' + c_0\Psi_1 - \left(\frac{\alpha\kappa\tau_1}{6c_0^2}\right)\Psi_1^3
= 0.
\end{equation}
This is a Bernoulli equation, and it has the solution
\begin{equation}\label{eqn: Sigma}
\Psi_1(X)
= \Sigma(X)
:= \left(\frac{6c_0^3}{\alpha\kappa\tau_1+6c_0^2\exp\big(2c_0X/\tau_2+\theta\big)}\right)^{1/2}.
\end{equation}
Here $\theta \in \R$ is an arbitrary phase shift.
It follows that putting
\begin{equation}\label{eqn: sigma}
\psi_1(X)
= \sigma(X)
:= -\Sigma'(X)
= \frac{(6c_0^3)^{3/2}\exp\big(2c_0X/\tau_2+\theta\big)}{\tau_2\big[\alpha\kappa\tau_1+6c_0^2\exp\big(2c_0X/\tau_2+\theta\big)\big]^{3/2}}
\end{equation}
solves \eqref{eqn: psi1 nu=0}.
Friesecke and Pego \cite{friesecke-pego1} do not incorporate a phase shift like $\theta$ into their leading order $\sech^2$-type KdV solution, since their broader existence result relies on working in spaces of even functions.
We will not need such symmetry in our subsequent arguments (nor could we achieve it, since no translation of $\sigma$ is even or odd), and so we will leave $\theta$ as an arbitrary free parameter and not specify its value.
\subsubsection{The formal leading order equation for $\psi_2$ and the selection of the exponent $\beta$}\label{sec: formal leading order2}
From our choice of $\gamma=\mu=2/5$ and the inequalities in \eqref{eqn: expn bare minimum}, we need, at the very least,
\[
\frac{1}{10}
\le \beta
\le \frac{1}{5}.
\]
If the strict inequality $\beta < 1/5$ holds, then at $\ep = 0$ the equation for $\psi_2$ in \eqref{eqn: lw syst-ep} reduces to the trivial result $\psi_2 = 0$.
This is not at all what we expect numerically from Figure \ref{fig:int:profile}; rather, we anticipate that $\psi_2$ will asymptote to some nonzero constant at $\infty$.
However, if we instead take $\beta$ so that
\[
0
= 1-(\gamma+\mu+\beta)
= \frac{1}{5}-\beta,
\]
which is to say,
\[
\beta
= \frac{1}{5},
\]
then the equation for $\psi_2$ in \eqref{eqn: lw syst-ep} at $\ep = 0$ becomes
\[
\psi_2
= \bPsf_1^0(\psi_1).
\]
Putting
\begin{equation}\label{eqn: zeta}
\psi_2(X)
= \zeta(X)
:= \bPsf_1^0(\sigma)(X)
= \frac{\alpha}{c_0}\int_X^{\infty} \sigma(V) \dV
\end{equation}
therefore solves the leading order equation for $\psi_2$.
We really have
\[
\zeta(X)
= \frac{\alpha}{c_0}\Sigma(X)
= \frac{\alpha}{c_0}\left(\frac{6c_0^3}{\alpha\kappa\tau_1+6c_0^2e^{2c_0X/\tau_2+\theta}}\right)^{1/2},
\]
where $\Sigma$ was defined in \eqref{eqn: Sigma}.
\subsection{The final long wave system}
With the choices of exponents $\gamma=\mu=2/5$ and $\beta = 1/5$, it becomes convenient to introduce the new small parameter
\begin{equation}\label{eqn: nu-ep}
\nu
:= \ep^{2/5}
\end{equation}
into the problem \eqref{eqn: lw syst-ep} and then recast that problem more cleanly in terms of $\nu$.
First, the long wave Ansatz \eqref{eqn: lw ansatz} becomes
\begin{equation}\label{eqn: nu-ep lw}
\phi_1(x) = \nu^{5/2}\psi_1(\nu{x}),
\qquad
\phi_2(x) = \nu^{1/2}\psi_2(\nu{x}),
\quadword{and}
c = \nu{c}_0.
\end{equation}
Proceeding very much as in Section \ref{sec: lw}, we then define
\begin{equation}\label{eqn: nl-nu}
\nl_1^{\nu}(X,Y ) := \frac{X^2Y}{k_a(k_a+\nu^{5/2}{X})}
\quadword{and}
\nl_2^{\nu}(X,Y) := \kappa\frac{k_rY+k_m\nu^2X+\nu^{5/2}{XY}}{(k_r+\nu^{5/2}{X})(k_m+\nu^{1/2}Y)}XY
\end{equation}
for $X$, $Y \in \R$, while for $f \in L^1$ and $g \in L^{\infty}$, we put
\begin{equation}\label{eqn: P1-nu}
\P_1^{\nu}(f)(X)
:= \frac{\alpha}{c_0}\int_X^{\infty} \E(\nu^{1/2}S^{\nu}f)(V,X)f(V) \dV,
\end{equation}
where $\E$ was defined in \eqref{eqn: E}, and
\begin{equation}\label{eqn: P2-nu}
\P_2^{\nu}(f,g)(X)
:= \frac{1}{c_0}\int_X^{\infty} \E(\nu^{1/2}S^{\nu}f)(V,X)\nl_2^{\nu}(f(V+\nu),g(V)) \dV,
\end{equation}
\begin{equation}\label{eqn: Rcal1-nu}
\Rcal_1^{\nu}(f)(X)
:= \frac{\kappa\tau_1}{c_0}\int_X^{\infty} \P_1^{\nu}(f)(V)f(V + \nu) \dV,
\end{equation}
\begin{equation}\label{eqn: Rcal2-nu}
\Rcal_2^{\nu}(f,g)(X)
:= \frac{1}{c_0}\int_X^{\infty} \big(\nu^{1/2}\kappa{f}(V+\nu)\P_2^{\nu}(f,g)(V)-\nl_2^{\nu}(f(V+\nu),g(V))\big) \dV,
\end{equation}
\begin{equation}\label{eqn: Rcal-nu}
\Rcal^{\nu}(f,g)(X)
:= \Rcal_1^{\nu}(f)(X) + \nu^{1/2}\Rcal_2^{\nu}(f,g)(X),
\end{equation}
and
\begin{equation}\label{eqn: Ncal-nu}
\Ncal^{\nu}(f,g)(X)
:= \tau_1\Rcal_2^{\nu}(f,g)(X)f(X)
- \nu^{3/2}\tau_1\nl_1^{\nu}(f(X),\Rcal^{\nu}(f,g)(X)).
\end{equation}
\begin{remark}\label{rem: ops rem}
The operators $\P_1^{\nu}$ and $\Rcal_1^{\nu}$ map $L^1$ into $L^{\infty}$, while $\P_2^{\nu}$ and $\Rcal_2^{\nu}$ map $L^1 \times L^{\infty}$ into $L^{\infty}$, and $\Ncal^{\nu}$ maps $L^1 \times L^{\infty}$ into $L^1$.
More precisely, we could replace $L^1$ with $H_q^1$ and $L^{\infty}$ with $W^{1,\infty}$ and the preceding statement would still be true; see the estimates in Appendix \ref{app: aux ests}.
The operator $\Rcal_1^0$ has the especially simple form
\begin{equation}\label{eqn: Rcal1-0}
\Rcal_1^0(f)(X)
= \left(\frac{\alpha\kappa\tau_1}{6c_0^2}\right)\int_X^{\infty} \left(\int_V^{\infty} f(W) \dW\right) f(V) \dV
\end{equation}
and therefore is differentiable from $L^1$ to $L^{\infty}$.
\end{remark}
Last, for $\nu > 0$, let $\M^{(\nu)}$ be the Fourier multiplier with symbol
\begin{equation}\label{eqn: M-nu}
\tM^{(\nu)}(z)
:= \nu\frac{1-e^{-i\nu{z}}}{ic_0\nu^2z+2\tau_2(\cos(\nu{z})-1)}.
\end{equation}
When $\nu = 0$ we have already defined $\M^{(0)}$ as the Fourier multiplier whose symbol $\tM^{(0)}$ is given in \eqref{eqn: tvarpi-0}.
Now we can state precisely the convergence result that we formally anticipated in Section \ref{sec: bMu-ep formal}, specifically in the limit \eqref{eqn: bMu-ep to 0}.
\begin{proposition}\label{prop: varpi-nu conv}
Fix $q \in (0,c_0/\tau_2)$.
There exist $\nu_{\M}$, $C_{\M} > 0$ such that if $0 < \nu < \nu_{\M}$, then
\[
\norm{\M^{(\nu)}-\M^{(0)}}_{\b(H_q^1)}
\le C_{\M}\nu^{1/3}.
\]
\end{proposition}
We prove this proposition in Appendix \ref{app: proof of prop varpi-nu conv}.
More broadly, we can summarize the work above on the travelling wave Ansatz and subsequent long wave scaling and exponent selection for the system \eqref{eqn: main system3} in the following theorem.
\begin{theorem}\label{thm: lw tw syst final}
Suppose
\[
\begin{cases}
A_j(t) = \nu^{5/2}\psi_1(\nu(j-\nu{c}_0t)), \\
P_j(t) = \nu^{1/2}\psi_2(\nu(j-\nu{c}_0t))
\end{cases}
\]
for some $\psi_1 \in H_q^1$ and $\psi_2 \in L^{\infty}$, where $c_0$, $\nu > 0$ and $q \in (0,c_0/\tau_2)$.
Then $A_j$ and $P_j$ satisfy \eqref{eqn: main system3} if and only if $\psi_1$ and $\psi_2$ satisfy
\begin{equation}\label{eqn: lw syst-nu}
\begin{cases}
\psi_1 = \M^{(\nu)}\big(\Rcal_1^{\nu}(\psi_1)\psi_1 + \nu^{1/2}\Ncal^{\nu}(\psi_1,\psi_2)\big), \\
\psi_2 = \P_1^{\nu}(\psi_1) + \nu\P_2^{\nu}(\psi_1,\psi_2).
\end{cases}
\end{equation}
Moreover, taking
\[
\psi_1 = \sigma
\quadword{and}
\psi_2 = \zeta = \P_1^0(\sigma),
\]
where $\sigma$ is defined in \eqref{eqn: sigma} and $\zeta$ is given explicitly in \eqref{eqn: zeta}, solves \eqref{eqn: lw syst-nu} when $\nu = 0$.
\end{theorem}
We proceed to analyze the system \eqref{eqn: lw syst-nu} with a quantitative contraction mapping argument that tracks its dependence on $\nu$.
\section{Analysis of the Long Wave Problem}\label{sec: lw syst-nu soln}
\subsection{The perturbation Ansatz for the long wave problem \eqref{eqn: lw syst-nu}}
Throughout this section we keep $q \in (0,c_0/\tau_2)$ fixed.
We make the perturbation Ansatz
\begin{equation}\label{eqn: perturb ansatz}
\psi_1 = \sigma + \eta_1
\quadword{and}
\psi_2 = \zeta + \eta_2
\end{equation}
for the long wave problem \eqref{eqn: lw syst-nu}.
Here $\eta_1 \in H_q^1$ and $\eta_2 \in W^{1,\infty}$ are unknown.
We abbreviate
\[
\etab
= (\eta_1,\eta_2)
\in \X := H_q^1 \times W^{1,\infty},
\]
where $\X$ has the norm
\[
\norm{\etab}_{\X}
:= \norm{\eta_1}_{H_q^1} + \norm{\eta_2}_{W^{1,\infty}}.
\]
The Ansatz \eqref{eqn: perturb ansatz} solves the system \eqref{eqn: lw syst-nu} if and only if $\eta_1$ and $\eta_2$ solve
\begin{equation}\label{eqn: eta1 eta2}
\begin{cases}
\T\eta_1 = \sum_{k=1}^5 \V_{1k}^{\nu}(\etab), \\
\eta_2 = \sum_{k=1}^3 \V_{2k}^{\nu}(\etab),
\end{cases}
\end{equation}
where
\begin{equation}\label{eqn: T}
\T\eta_1
:= \eta_1-\M^{(0)}\big[\Rcal_1^0(\sigma)\eta_1+\big(D\Rcal_1^0(\sigma)\eta_1\big)\sigma\big],
\end{equation}
\begin{equation}\label{eqn: V1k}
\begin{aligned}
\V_{11}^{\nu}(\etab) &:= \big(\M^{(\nu)}-\M^{(0)}\big)\big[\Rcal_1^{\nu}(\sigma+\eta_1)(\sigma+\eta_1)\big], \\
\V_{12}^{\nu}(\etab) &:= \M^{(0)}\big[\big(\Rcal_1^{\nu}(\sigma+\eta_1)-\Rcal_1^0(\sigma+\eta_1)\big)(\sigma+\eta_1)\big], \\
\V_{13}^{\nu}(\etab) &:= \M^{(0)}\big[\big(\Rcal_1^0(\sigma+\eta_1)-\Rcal_1^0(\sigma)-D\Rcal_1^0(\sigma)\eta_1\big)\sigma\big], \\
\V_{14}^{\nu}(\etab) &:= \M^{(0)}\big[\big(\Rcal_1^0(\sigma+\eta_1)-\Rcal_1^0(\sigma)\big)\eta_1\big], \\
\V_{15}^{\nu}(\etab) &:= \nu^{1/2}\M^{(0)}\Ncal^{\nu}(\sigma+\eta_1,\zeta+\eta_2)
\end{aligned}
\end{equation}
and
\begin{equation}\label{eqn: V2k}
\begin{aligned}
\V_{21}^{\nu}(\etab) &:= \P_1^{\nu}(\sigma+\eta_1)-\P_1^0(\sigma+\eta_1), \\
\V_{22}^{\nu}(\etab) &:= \P_1^0(\sigma+\eta_1)-\zeta, \\
\V_{23}^{\nu}(\etab) &:= \nu\P_2^{\nu}(\sigma+\eta_1,\zeta+\eta_2).
\end{aligned}
\end{equation}
We claim that $\T$ is `right-invertible' in the following sense, which we make rigorous in Appendix \ref{app: proof of prop T invert}.
\begin{proposition}\label{prop: T invert}
Let $q \in (0,c_0/\tau_2)$.
There exists $\Scal \in \b(H_q^1)$ such that $\T\Scal{g} = g$ for all $g \in H_q^1$.
\end{proposition}
The operator $\T$ is really the linearization at $\psi_1 = \sigma$ and $\nu=0$ of the first equation in \eqref{eqn: lw syst-nu}.
Such a linearization at the limiting localized solution appears as a key operator in numerous FPUT problems, including \cite{friesecke-pego1, faver-wright, hoffman-wright}, and the invertibility of this operator is a property essential to the development of the right fixed point formula for the given problem.
Our treatment of the invertibility of $\T$ in Appendix \ref{app: proof of prop T invert} is rather different from the analogous inversions in those papers, as here $\T{f}=g$ is really a linearized Bernoulli equation in disguise, rather than the linearized KdV travelling wave profile equation.
In particular, solving $\T{f} = g$ turns into a first-order linear problem, which we can solve explicitly with an integrating factor.
While this requires a fair amount of calculus, we do avoid the more abstract spectral theory that manages the second-order KdV linearizations (see, e.g., \cite[Lem.\@ 4.2]{friesecke-pego1}).
Due to Proposition \ref{prop: T invert}, for $\eta_1 \in H_q^1$ and $\eta_2 \in W^{1,\infty}$ to solve \eqref{eqn: eta1 eta2}, it suffices to take
\begin{equation}\label{eqn: Nfrak1-nu}
\eta_1
= \Scal\sum_{k=1}^5 \V_{1k}^{\nu}(\etab)
=: \Nfrak_1^{\nu}(\etab).
\end{equation}
Subsequently, $\eta_1$ and $\eta_2$ solve \eqref{eqn: eta1 eta2} if and only if
\begin{equation}\label{eqn: Nfrak2-nu}
\eta_2
= \V_{21}^{\nu}(\etab)
+ \V_{22}^{\nu}(\Nfrak_1^{\nu}(\etab))
+ \V_{23}^{\nu}(\etab)
= : \Nfrak_2^{\nu}(\etab).
\end{equation}
We have replaced $\eta_1$ with its fixed point expression \eqref{eqn: Nfrak1-nu} in $\V_{22}^{\nu}$ for the sake of better estimates later; see Appendix \ref{app: V22-nu Lip} for a more precise discussion.
Finally, set
\begin{equation}\label{eqn: Nfrakb-nu}
\Nfrakb^{\nu}(\etab)
:= (\Nfrak_1^{\nu}(\etab),\Nfrak_2^{\nu}(\etab)),
\end{equation}
so $\Nfrakb^{\nu}$ maps $\X$ to $\X$.
More precisely, this follows from the mapping estimates in Appendix \ref{app: map}.
We conclude that the problem \eqref{eqn: eta1 eta2} is equivalent to the fixed point problem
\begin{equation}\label{eqn: final fixed point}
\etab
= \Nfrakb^{\nu}(\etab),
\end{equation}
which we now solve.
\subsection{The solution of the fixed point problem \eqref{eqn: final fixed point}}
For $\rho > 0$, we define the ball
\[
\Bfrak(\rho)
:= \set{\etab \in \X}{\norm{\etab}_{\X} \le \rho}.
\]
We prove the following estimates in Appendix \ref{app: proof of main contraction ests}.
\begin{proposition}\label{prop: main contraction ests}
There exist $C_{\star}$, $\nu_{\star} > 0$ such that if $0 < \nu < \nu_{\star}$ then the following hold.
\begin{enumerate}[label={\bf(\roman*)}, ref={(\roman*)}]
\item\label{part: main mapping}
If $\etab \in \Bfrak(C_{\star}\nu^{1/3})$, then $\Nfrakb^{\nu}(\etab) \in \Bfrak(C_{\star}\nu^{1/3})$.
\item\label{part: main Lipschitz}
If $\etab$, $\grave{\etab} \in \Bfrak(C_{\star}\nu^{1/3})$, then
\[
\norm{\Nfrakb^{\nu}(\etab)-\Nfrakb^{\nu}(\grave{\etab})}_{\X}
\le \frac{1}{2}\norm{\etab-\grave{\etab}}_{\X}.
\]
\end{enumerate}
\end{proposition}
Proposition \ref{prop: main contraction ests} guarantees that $\Nfrakb^{\nu}$ is a contraction on $\Bfrak(C_{\star}\nu^{1/3})$ for each $0 < \nu < \nu_{\star}$, and so Banach's fixed point theorem gives the following solution to \eqref{eqn: eta1 eta2}.
\begin{theorem}\label{thm: eta1 eta2}
Let $C_{\star}$, $\nu_{\star} > 0$ be as in Proposition \ref{prop: main contraction ests}.
For each $0 < \nu < \nu_{\star}$, there exists $\etab^{\nu} \in \Bfrak(C_{\star}\nu^{1/3})$ such that $\etab^{\nu} = \Nfrakb^{\nu}(\etab^{\nu})$.
\end{theorem}
Theorems \ref{thm: lw tw syst final} and \ref{thm: eta1 eta2}, along with the integral formulations of Section \ref{sec: Rj analysis} and the relation $\nu = \epsilon^{2/5}$, per \eqref{eqn: nu-ep}, together yield the following solutions to our original problem \eqref{eqn: main system} for $A_j$, $P_j$, and $R_j$.
These results are paraphrased nontechnically in \eqref{eq:int:mr:a:p:r:profiles:scales}.
\begin{corollary}\label{cor: main corollary}
Let $\alpha$, $\kappa$, $\tau_1$, $\tau_2$, $c_0 > 0$, $q \in (0,c_0/\tau_2)$, and $\theta \in \R$.
Define the leading-order profile terms
\[
\phi_A^*(X)
:= \left(\frac{6\sqrt{6}c_0^{9/2}}{\tau_2}\right)\left(\frac{\exp(2c_0X/\tau_2+\theta)}{\big[\alpha\kappa\tau_1+6c_0^2\exp(2c_0X/\tau_2+\theta)\big]^{3/2}}\right),
\]
\[
\phi_P^*(X)
:= \big((6c_0)^{1/2}\alpha\big)\left(\frac{1}{\big[\alpha\kappa\tau_1+6c_0^2\exp\big(2c_0X/\tau_2+\theta)/\tau_2\big)\big]^{1/2}}\right),
\]
and
\[
\phi_R^*(X)
:= (3\alpha\kappa{c}_0)\left(\frac{1}{\alpha\kappa\tau_1+6c_0^2\exp(2c_0X/\tau_2+\theta)/\tau_2)}\right).
\]
There exists $\ep_{\star} > 0$ such that for each $0 < \ep < \ep_{\star}$, there are $\phi_A^{\ep} \in H_q^1 \cap \Cal^{\infty}$ and $\phi_P^{\ep}$, $\phi_R^{\ep} \in W^{1,\infty} \cap \Cal^{\infty}$ with the following properties.
\begin{enumerate}[label={\bf(\roman*)}]
\item
Let
\[
A_j(t)
= \ep\phi_A^*(\ep^{2/5}(j-\ep^{2/5}{c}_0t)) + \ep^{17/15}\phi_A^{\ep}(\ep^{2/5}(j-\ep^{2/5}c_0t)),
\]
\[
P_j(t)
= \ep^{1/5}\phi_P^*(\ep^{2/5}(j-\ep^{2/5}c_0t))+\ep^{1/3}\phi_P^{\ep}(\ep^{2/5}(j-\ep^{2/5}c_0t)),
\]
and
\[
R_j(t)
= \ep^{2/5}\phi_R^*(\ep^{2/5}(j-\ep^{2/5}c_0t)) + \ep^{3/5}\phi_R^{\ep}(\ep^{2/5}(j-\ep^{2/5}c_0t)).
\]
Then the triple $(A_j,P_j,R_j)$ solves \eqref{eqn: main system}.
\item
The remainder terms $\phi_A^{\ep}$, $\phi_P^{\ep}$, and $\phi_R^{\ep}$ satisfy
\[
\sup_{0 < \ep < \ep_{\star}}
\norm{\phi_A^{\ep}}_{H_q^1}
+ \norm{\phi_P^{\ep}}_{W^{1,\infty}}
+ \norm{\phi_R^{\ep}}_{W^{1,\infty}}
< \infty.
\]
\item
The functions $\phi_P^{\ep}$ and $\phi_R^{\ep}$ vanish exponentially fast at $+\infty$ and are asymptotically constant at $-\infty$ in the following sense: there exist $\ell_P^{\ep}$, $\ell_R^{\ep} \in \R$ such that
\[
\sup_{0 < \ep < \ep_{\star}} \left(|\ell_P^{\ep}| + \sup_{X \ge 0} e^{qX}|\phi_P^{\ep}(X)|
+ \sup_{X \le 0} e^{-qX}|\phi_P^{\ep}(X)-\ell_P^{\ep}|\right)
< \infty
\]
and
\[
\sup_{0 < \ep < \ep_{\star}} \left(|\ell_R^{\ep}| + \sup_{X \ge 0} e^{qX}|\phi_R^{\ep}(X)|
+ \sup_{X \le 0} e^{-qX}|\phi_R^{\ep}(X)-\ell_R^{\ep}|\right)
< \infty.
\]
\end{enumerate}
\end{corollary}
In order to achieve the normalization $\norm{\phi_A^*}_{L^{\infty}} = 1$, as discussed in Section \ref{sec:sub:int:mr}, we need to use the explicit choice
\begin{equation}\label{eqn: cstar}
c_0
= \left(\frac{9\alpha\kappa\tau_1\tau_2^2}{8}\right)^{1/5}
=: c_*,
\end{equation}
as used in \eqref{eq:int:mr:a:p:r:profiles:scales}.
Furthemore, we then get
\begin{equation}\label{eqn: P R Linfty}
\norm{\phi_P^*}_{L^{\infty}}
= \left(\frac{6\alpha}{\kappa\tau_1}\right)^{1/2}\left(\frac{9\alpha\kappa\tau_1\tau_2^2}{8}\right)^{1/10}
\quadword{and}
\norm{\phi_R^*}_{L^{\infty}}
= \frac{3}{\tau_1}\left(\frac{9\alpha\kappa\tau_1\tau_2^2}{8}\right)^{1/5}.
\end{equation}
Substituting the abbreviations \eqref{eq:twv:def:tau:1:2} and \eqref{eq:twv:def:kappa} into the quantities in \eqref{eqn: cstar} and \eqref{eqn: P R Linfty} then leads to the identities \eqref{eq:int:def:expl:constants}.
| bb2ca0c31eb74d5b991c6ea60cee1a5c67e4df46 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Overview}
\label{sec:Overview}
Essentially everything of astronomical interest is either part of a galaxy, or from a galaxy, or otherwise relevant to the origin or evolution of galaxies. Diverse examples are that the isotropic composition of meteorites provides clues to the history of star formation billions of years ago, and cosmological tests for the deceleration of the Universe are strongly affected by changes in the luminosities of galaxies during the lookback time sampled. The aim of this article is to review some of the vital connections that galaxy evolution makes among many astronomical phenomena.
The evolution of galaxies can be broadly divided into three areas: dynamical evolution, chemical evolution, and the evolution of photometric properties. Chemical and photometric evolution are the main topics of this review, and dynamical processes will be considered only where they directly affect the others. The article is intended to be self-contained for readers with a general knowledge of astronomy and astrophysics but with no special expertise on galaxies. Because the background and literature relevant to the evolution of stars and gas in galaxies are so extensive, complete coverage cannot be attempted. Instead, comprehensive reviews and recent papers are emphasized in the references; these include detailed works in closely related fields, such as galaxy formation and nucleosynthesis, that are treated very briefly here. The history of the subject has been discussed in several recent reviews \citep[e.g.][]{Sandage1975GalaxiesUniverse, Spinrad1975TheSpectra, Trimble1975TheElements, vandenBergh1975StellarGalaxies}, so here the emphasis will be on current ideas.
\medskip
This first Section gives an overview of relevant properties of galaxies and theoretical ideas on how they formed and evolved, followed by an outline of the rest of the article.
\subsection{Stellar Populations and Chemical Composition of Galaxies}
\label{subsec:Stellar_Populations}
Galaxies are made of stars and interstellar matter (ISM), with a variety of properties that correlate remarkably with the forms of galaxies as they appear on the sky. Introductory references to supplement the following brief outline include the classic lectures of \citet{Baade1963EvolutionGalaxies} and subsequent reviews by \citet{Morgan1969OnGalaxies}, \citet{King1971StellarGalaxies, King1977Introduction:Day}, \citet{vandenBergh1975StellarGalaxies}, and \citet{Sandage1975GalaxiesUniverse}. The \emph{Hubble Atlas of Galaxies} \citep{Sandage1961c} is an essential source book for photographs and descriptions of typical galaxies, while the \emph{Atlas of Peculiar Galaxies} \citep{Arp1966AtlasGalaxies} illustrates a great variety of unusual forms.
\medskip
Most galaxies can be arranged in a natural sequence of morphological types, the Hubble sequence. The ``earliest'' are elliptical (E) galaxies, with apparently smooth distributions of yellow--red stars and seldom any signs of internal structure (except that individual stars and globular clusters can be resolved in the closest ones); and the ``latest'' are irregulars (Irr I) with no symmetry and with disorganized patches of blue stars, hot gas, and dust throughout\footnote{By a historical misfortune, the early-type galaxies are dominated by late-type (red) stars, and the late-type galaxies are dominated by early-type (blue) stars!}. In between are the spirals, in order from Sa types with large nuclear bulges and tightly wound arms to Sc types with small bulges and open arms; the nuclear bulges closely resemble elliptical galaxies, while spiral arms, or more often bits of spiral arms, are traced by patchy blue stars and gas and by dark dust lanes, superposed on an underlying smoother disk of redder stars. The galaxies classified as S0 have disks but lack spiral structure and its associated young stars and gas; Hubble put them between types E and Sa, but \citet{vandenBergh1976a} pictures them as a sequence parallel to the spirals, ordered by bulge-to-disk ratio. Another ``dimension'' of the Hubble sequence is based on the presence or absence of a bar in the center, from the ends of which the spiral arms seem to emerge; there are barred spirals of types SBa with tight arms through SBc with loosely wound arms. The Hubble sequence has been elaborated by \citet{deVaucouleurs1959ClassificationGalaxies} into a more detailed system indicating intermediate forms and additional structural features.
Although the sequence E--S0--Sa--Sb--Sc--Irr I was defined by the shapes of galaxies, it is very much a sequence of stellar populations. The earliest galaxies show no signs of hot young stars and (usually) have undetectably little gas or dust, while the latest galaxies are undergoing active star formation and have significant amounts of ISM. The Yerkes classification system of galaxies \citep{Morgan1957AGalaxies, Morgan1969OnGalaxies} emphasizes the correlation between forms and stellar populations of galaxies. For spirals, this system is based simply on the central concentration of light, and it is found that the highly concentrated galaxies have late-type nuclear spectra (indicating a mixture of G, K, and M stars), while those with little central concentration have early-type spectra (dominated by B stars) even in the nucleus. The Yerkes types of galaxies are correlated loosely with Hubble types, which is not surprising since the Hubble classification is based partly on the amount of light concentrated into a nuclear bulge and partly on the tightness of spiral structure. The increasing prevalence of blue, young stars as one goes from elliptical through irregular galaxies is reflected in a general progression from redder to bluer integrated colors.
From studies of the integrated spectra and colors of galaxies, it is concluded that nearly all galaxies have some very old stars, and that their different proportions of red and blue stars in the light are due mainly to different current rates of star formation. Most (if not all) galaxies appear to be many billions of years old, but whereas significant star formation ceased billions of years ago in ellipticals, it has occurred at a roughly constant rate in the latest types of galaxies. Theories of galaxy formation, and of star formation within galaxies, are challenged to explain this close correlation between the forms of galaxies and their histories of star formation.
Some galaxies fit nowhere into a tidy sequence of types, and they are classified as simply ``peculiar'' or as one of the special types of peculiars such as ring galaxies or Irr~II (which are early-type galaxies in form but strewn with dust, young stars etc.). Many of the peculiar systems are interacting galaxies, distorted by tidal effects \citep{Toomre1972GalacticTails}. Sometimes the colors of peculiar galaxies are so blue that any underlying old population is completely outshone by newly-formed stars, and sometimes they are so red that most of the activity must be hidden in dust clouds. Although these very peculiar galaxies are rare, they are of special interest since some of them may be undergoing the kinds of violent changes and rapid star formation that characterized normal galaxies in their youth \citep{Larson1976c}.
\medskip
The chemical compositions of galaxies in general show much less diversity than their stellar populations. Few instances are known of interstellar gas that is deficient or overabundant in heavy elements by more than a factor of four relative to the Sun, and although stellar metallicities range over more than two orders of magnitude, they are usually within a factor of three of Solar\footnote{The term ``metals'' in this field generally includes all elements heavier than helium. True metals, especially iron, are most easily detected in the spectra of stars, while the common non-metallic elements, especially oxygen, are most accessible in the ISM. Another loose usage in this field is that the word ``element'' frequently means a particular nuclide (e.g., the ``element'' $^{13}$C).}. Many of the differences among chemical compositions in galaxies are systematic: metallicities of both stars and gas tend to decrease outward from the centers of galaxies, and average metallicities tend to increase with galaxy luminosity.
Models for chemical evolution of galaxies aim to account for their compositions in terms of the production of elements by stars (mainly) and the mixing of stellar ejecta with interstellar gas. Gas flows often play important roles in chemical evolution, diluting the products of nucleosynthesis with unenriched material from outside the galaxy, and carrying metals from one part of the galaxy to another.
Many more details are known about the stellar population in the Solar vicinity of our own galaxy than anywhere else, and some of these details contain clues to large-scale processes of galaxy formation. For example, stars with high space velocities that make them members of the halo\footnote{The word ``halo'' is used in the astronomical literature in two ways: (i) its classical meaning, a spheroidal population of ordinary stars, typified by globular clusters; and (ii) an invisible spheroidal component that provides much more galactic mass than can be ascribed to known stars and gas. The first meaning is implied throughout this article, unless otherwise stated.} population are metal-poor by factors of ten or more relative to the Sun, while stars with disk motions have metallicities almost entirely within a factor of three of Solar. Since the classic paper of \citet{Eggen1962EvidenceCollapsed.}, this difference has been interpreted as evidence that the halo stars formed first, before much chemical enrichment by deaths of massive stars had taken place. Those authors also used the kinematics of halo stars, and a model for the collapse of the Galaxy from particular initial conditions, to infer a free-fall timescale, ${\sim} 2 \times 10^{8} \: \rm yr$, for the collapse and formation of halo stars. Ages that have recently been determined for halo and disk stars now point toward a slower collapse, however; globular clusters in the Galaxy appear to have a range of ages from ${\sim} 12 - 16 \: \rm Gyr$, while very few disk stars near the Sun are older than $10 \: \rm Gyr$ \citep{Demarque1977StellarGalaxy, Saio1977AgesII, Saio1977OnStars, Twarog1980AnNeighborhood}. The picture presented by these values is of a rather slow collapse, and a very long timescale for star formation to get underway in the outer disk of the Galaxy. A very young star cluster at least $10 \: \rm kpc$ from the Sun in the Galactic anticenter direction has recently been assigned a metallicity as low as those of halo stars \citep{Christian1979UBV21}, so it seems that the outermost disk is still in a very early stage of chemical evolution.
\medskip
In general, composition differences among the old stars of galaxies serve as frozen-in tracers of their early chemical evolution, while abundances in young stars and the ISM indicate the progress of continuing enrichment. These effects are closely tied to the history of star formation in various regions of galaxies, which relates them to factors determining photometric properties. It will be seen later in this review that the aspects of galactic evolution of greatest importance for photometric evolution and for chemical evolution, respectively, are to a large extent complementary, so the two fields of study tend to yield different types of information about the evolution of galaxies.
\subsection{Galaxy Formation and Evolution}
\label{subsec:Galaxy_Formation}
The striking correlations between the forms and contents of galaxies, and systematic changes of content with position in galaxies, point to the importance of dynamical processes in chemical and photometric evolution. Many theoretical ideas on the formation and dynamical evolution of galaxies have been motivated by these correlations, and dynamical models in turn have led to further understanding of such properties. A brief sketch of some current theoretical views of galaxy formation and dynamical evolution will therefore be given here. A fuller non-technical introduction to the field is given by \citet{Larson1977a}, and more technical reviews are by \citet{Doroshkevich1978SpatialGalaxies}, \citet{Gott1977RecentFormation}, \citet{Jones1976TheObservation}, \citet{Larson1976b}, and \citet{Rees1977CosmologyFormation}.
\medskip
In the conventional cosmological picture, the primeval gas emerging from the Big Bang some $10 - 20 \: \rm Gyr$ ago consisted almost entirely of hydrogen and helium. There must have been large enough density perturbations in this gas for some regions to be locally bound and to collapse despite the expansion of the Universe, and the existence today of galaxies and bound clusters with a $10^{10}$-fold range of masses shows that there was a very wide spread of density perturbations. A favored mass scale is ${\sim} 10^{6} \: \rm M_{\odot}$, which is the Jeans mass in the Universe after (re-)combination of the primeval plasma about $10^{6} \: \rm yr$ after the Big Bang. Maybe star formation first began in lumps of this mass, which could themselves have been part of larger-scale perturbations destined to collapse eventually into a massive galaxy. Although models for galaxy formation often start from an idealized smooth density distribution, gas in this state would be Jeans unstable on much smaller scales than the protogalaxy, and one would expect star formation to occur in regions of much higher density than average.
Independently of cosmological details, a protogalaxy can be pictured as a lumpy gas cloud governed by self-gravity (and perhaps by perturbations due to neighbors) rather than by the expansion of the Universe. If no stars formed, the gaseous protogalaxy would collapse, collisions between gas clouds would dissipate the energy parallel to the net angular momentum vector, and the system would become a flat disk. (The process of ``dissipation'' in this context usually implies that some of the kinetic energy of the gas is lost by collisionally induced radiation). If, on the other hand, stars form in much less than the collapse time, there would be no dissipation and a spheroidal system of stars would be produced; subclustering due to the initially lumpy gas distribution would be wiped out except in regions with about the mass and density of globular clusters. A spheroidal system can also be formed in much longer than the free-fall time if dissipation takes much longer, perhaps because the gas is in clouds that collide infrequently.
In general, the formation of spheroidal stellar systems -- elliptical galaxies and the bulges of spirals -- requires that stars form on a timescale less than that of gaseous dissipation, while the formation of a disk requires the opposite situation. The amount of disk possessed by a galaxy therefore depends on the amount of gas that remained to collapse dissipatively, after efficient star formation in the spheroidal component. This dynamical picture is consistent with the great ages, ${\sim} 10^{10} \: \rm yr$, inferred for stars in spheroidal systems, and the contrasting range of ages (including ongoing star formation) in disks.
In detail, the differences between spheroidal and disk systems are not quite so clear-cut, because the formation of condensed nuclei in elliptical galaxies and the central bulges of spirals probably involved some dissipation. The enhanced metallicities of nuclear stars are thought to be due to an inward concentration of metals, as gas enriched by massive stars further out dissipated its energy and condensed toward the center.
S0 galaxies appear to be intermediate systems, having disks but no young stars. One class of theories of their origin suggests that these properties are intrinsic, in that star formation efficiently used up the gas after it had collapsed to a disk. In another class of theories, S0 galaxies are former spirals that are devoid of young stars because their ISM was swept away, either in a collision with another galaxy or by an intergalactic wind due to motion of the galaxy through intergalactic gas. The high frequency of S0 galaxies in dense clusters, where intergalactic gas has been detected, suggests that at least some were formed in the latter way, but other S0 galaxies are quite isolated and may be of the ``intrinsic'' type.
Sweeping by collisions and intracluster gas are examples of interactions with the environment that probably affect the evolution of many galaxies, including apparently normal ones and very peculiar objects. Other effects include mergers of galaxies in the centers of clusters and in close binaries, tidal disruption in close passages, and bursts of star formation due to such violent interactions among gas-rich galaxies. A spiral galaxy may also accrete gas from its surroundings during most of its life, gradually forming a more massive and more extended disk and acquiring new gas to fuel continued star formation. Various aspects of interactions between galaxies and their environment are reviewed by \citet{Ostriker1977a}, \citet{Saar1978InteractionGalaxies}, and \citet{Toomre1977MergersConsequences}.
\subsection{Plan of this Review}
\label{subsec:Review_plan}
From the preceding outline, it should be clear that attempts to understand the evolution of stars and gas in galaxies inevitably get involved in very diverse aspects of astronomical theory and observation. This is not a field in which one can hope to develop a complete theory from a simple set of assumptions, because many relevant data are unavailable or ambiguous, and because galactic evolution depends on many complicated dynamical, atomic, and nuclear processes which themselves are incompletely understood. Because a logical development is inappropriate, this article treats different aspects of galactic evolution as pieces of a jigsaw puzzle that may some day be put together.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig1.png}
\caption{A schematic outline of the links between various constituents of galaxies and evolutionary processes. \emph{Solid lines} indicate the main factors in chemical evolution, and \emph{dashed lines} are for photometric evolution.}
\label{fig:fig1}
\end{figure}
Figure~\ref{fig:fig1} is a very schematic diagram of relations among some of the most important factors in galactic evolution. It serves to summarize parts of the preceding outline, to suggest how a final jigsaw might be linked, and to show how the topics discussed below are related to each other. Two main sets of processes and constituents are indicated.
\begin{enumerate}
\item Chemical evolution depends mainly on the areas connected by solid lines, which indicate that stars form from the ISM and return some of their mass, including new products of nucleosynthesis, to the ISM at death; the composition of the ISM also depends on gas flows to and from the region under consideration.
\vspace{1mm}
\item Photometric properties depend mainly on the areas on the areas connected by dashed lines, indicating that the history of star formation and the evolution of individual stars determine the population at any time.
\end{enumerate}
\medskip
Sections~\ref{sec:Aims and Methods 1},~\ref{sec:Evolution solar neighborhood},~and~\ref{sec:Chemical Evolution of Galaxies} below are on chemical evolution and Sections~\ref{sec:Approaches to Photometric Evolution}~and~\ref{sec:Colors and Star Formation Rates} are on the evolution of photometric properties. Since the formation and evolution of stars are vital to all aspects of galactic evolution, these processes are reviewed first in Section~\ref{sec:Star_formation}.
\section{The Formation and Evolution of Stars}
\label{sec:Star_formation}
The evolution of galaxies depends critically on properties of stars as a function of mass: their formation rates, evolution in the Hertzsprung--Russell (HR) diagram, lifetimes, and the mass and composition of material returned to the ISM. This Section reviews the properties of stars that are most relevant to later discussion of galaxy evolution.
\subsection{Basic Physical Properties of Stars}
\label{subsec:Physical_Properties}
Table~\ref{tab:table1} lists for reference some properties of main sequence (MS) stars with chemical compositions typical of the nearby disk population, i.e. with Solar or slightly lower metallicities. Because stars evolve in the HR diagram while still on the MS, this Table only applies to a particular stage of evolution for each mass: stars with total MS lifetimes less than $10 \: \rm Gyr$ (masses $ \gtrsim 1 \: \rm M_{\odot}$) are listed with the properties they have in the middle of this lifetime; less massive stars are listed at age $5 \: \rm Gyr$ or with mean empirical properties. The theoretical properties of MS stars of a given mass and composition differ slightly in different series of calculations, and they are sensitive to the helium and heavy-element abundances, even within the range of values applicable to disk stars, so the tabulated values should not be accepted too literally. This Table is intended only to provide approximate relations between such quantities as spectral type and MS lifetime, for quick reference.
\begin{table}
\centering
\caption{Representative data for main sequence stars. Quantities tabulated are typical values for stars of approximately Solar composition, from a variety of sources including mainly the following. Masses $m \leq 0.6 \: \rm M_{\odot}$: data from \citet{Veeder1974LuminositiesPhotometry}; $0.8 \leq m \leq 10 \: \rm M_{\odot}$: theoretical tracks of \citet{Mengel1979StellarSequence} and \citet{Alcock1978TheStars}; $m \geq 20 \: \rm M_{\odot}$: theoretical tracks of \citet{Chiosi1978MassiveMass-loss.} with maximum mass loss rates. For $m \leq 0.6 \: \rm M_{\odot}$, the mean empirical main sequence is given; for $0.8 \leq m \leq 1.0 \: \rm M_{\odot}$, points are for an age of $5 \: \rm Gyr$; and for more massive stars, points are for the middle of the main sequence lifetime ($\tau_{\rm ms} / 2$).
}
\label{tab:table1}
\begin{tabular}{ccccccc}
\hline
$m \: \rm \left( M_{\odot} \right)$ & $\tau_{\rm ms}$ & $\log \left( L / \rm L_{\odot} \right)$ & $M_{\rm V}$ & $\log \ T_{\rm eff}$ & $B - V$ & Sp\\
\hline
0.15 & -- & -2.5 & 14.2 & 3.48 & 1.80 & M7\\
0.25 & -- & -2.0 & 12.0 & 3.52 & 1.60 & M5\\
0.4 & -- & -1.4 & 10.0 & 3.57 & 1.48 & M1\\
0.6 & -- & -0.9 & 7.6 & 3.64 & 1.18 & K5\\
0.8 & 25 & -0.4 & 6.0 & 3.70 & 0.88 & K1\\
0.9 & 15 & -0.2 & 5.4 & 3.73 & 0.76 & G8\\
1.0 & 10 & 0.0 & 4.9 & 3.76 & 0.64 & G2\\
1.1 & 6.4 & 0.2 & 4.3 & 3.79 & 0.56 & F8\\
1.2 & 4.5 & 0.4 & 3.7 & 3.82 & 0.47 & F6\\
1.3 & 3.2 & 0.5 & 3.5 & 3.84 & 0.42 & F5\\
1.4 & 2.5 & 0.7 & 3.0 & 3.86 & 0.36 & F2\\
1.5 & 2.0 & 0.8 & 2.8 & 3.88 & 0.30 & F0\\
2 & 0.75 & 1.3 & 1.4 & 3.98 & 0.00 & A0\\
3 & 0.25 & 2.1 & -0.2 & 4.10 & -0.12 & B7\\
4 & 0.12 & 2.6 & -0.6 & 4.18 & -0.17 & B5\\
6 & 0.05 & 3.2 & -1.5 & 4.30 & -0.22 & B3\\
8 & 0.03 & 3.6 & -2.2 & 4.35 & -0.25 & B1\\
10 & 0.02 & 3.9 & -2.7 & 4.40 & -0.27 & B0.5\\
15 & 0.01 & 4.4 & -3.7 & 4.45 & -0.29 & B0.5\\
20 & 0.008 & 4.7 & -4.3 & 4.48 & -0.30 & B0\\
30 & 0.006 & 5.1 & -5.1 & 4.51 & -0.31 & O9.5\\
40 & 0.004 & 5.4 & -5.7 & 4.53 & -0.31 & O9\\
60 & 0.003 & 5.7 & -6.2 & 4.58 & -0.32 & O5\\
\hline
\end{tabular}
\end{table}
An overview of stellar evolution in the HR diagram is provided by Figures~\ref{fig:fig2}~and~\ref{fig:fig3}. Figure~\ref{fig:fig2} is a theoretical HR diagram with the zero-age MS and evolutionary tracks for stars of a few masses; the tracks are drawn lightly in regions where stellar evolution is rapid (relative to neighboring points on the same track) since few stars are observed at those stages. A given star spends only about $10\%$ of its MS lifetime in stages of evolution beyond the MS, so the supergiant and giant regions of the HR diagram contain a relatively ephemeral population of stars. Much the same information is given empirically in Figure~\ref{fig:fig3}, which is a reproduction of the composite color--magnitude diagram for open clusters of \citet{Sandage1969IsochronesClusters}. The clusters have too few stars for the faster stages of evolution to be well represented; not only are there gaps between the MS and the giant branch, but the latest giants are absent from all but one or two clusters. Despite this problem, the diagram of \citet{Sandage1969IsochronesClusters} gives a vivid picture of stellar evolution.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig2.png}
\caption{Theoretical HR diagram indicating main sequence lifetimes and evolutionary tracks of stars with approximately Solar composition. The zero-age main sequence (ZAMS) is shown, with a few stellar masses ($\rm M_{\odot}$) and lifetimes (yr). Also shown are representative evolutionary tracks, shaded heavily where a star evolves slowly relative to adjacent points. \emph{Sources}: ZAMS and tracks for $m \leq 3 \: \rm M_{\odot}$: \citet{Mengel1979StellarSequence}, $Z = 0.01$, $Y = 0.20$; ZAMS for $3 < m \leq 10 \: \rm M_{\odot}$: \citet{Alcock1978TheStars}, $Z = 0.01$, $Y = 0.29$; track for $m = 4.5 \rm \: M_{\odot}$: \citet{Harris1976EffectsClouds}, $Z = 0.01$, $Y = 0.28$; ZAMS and tracks for $m \geq 20 \rm \: M_{\odot}$: \citet{Chiosi1978MassiveMass-loss.}, case with maximum mass loss rate, $Z = 0.02$, $Y = 0.28$.}
\label{fig:fig2}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig3.png}
\caption{Composite color--magnitude diagram for open clusters \citep{Sandage1969IsochronesClusters}. Ages at the right refer to the blue point for a given cluster. Modern distances and age scales correspond to younger ages, e.g., $5 \times 10^{9} \: \rm yr$ for NGC~188 \citep{Demarque1977StellarGalaxy}. This diagram is an empirical counterpart of Figure~\ref{fig:fig2}.}
\label{fig:fig3}
\end{figure}
\medskip
Using these Figures, one can envisage the evolution of a single generation of stars in a galaxy, starting with their populations on the zero-age MS. As time passes, the top of the MS peels away and short-lived supergiants appear; the MS turnoff evolves down to mid-B spectral type in $10^{8} \: \rm yr$, and by $10^{9} \: \rm yr$ late A and later stars remain on the MS; between $10^{9}$ and $10^{10} \: \rm yr$, the turnoff evolves from late A to early G, and there is a population of evolving K and M giants. An elliptical galaxy, for example, that had no star formation for the last many billions of years, would be expected to consist only of K and M giants and the MS up to a late F or early G turnoff. In a galaxy like our own, where star formation has presumably been continuous for billions of years, the whole array of MS stars, turnoff, and giant branches can be found.
\subsection{The Initial Mass Function}
\label{subsec:IMF}
Because the properties of stars depend strongly on their masses, the distribution of stellar masses at birth is a function vital to the photometric and chemical evolution of galaxies. Various definitions and notations for this function are used in the literature, so care is needed to avoid confusion in comparing different papers. The following notation is used here: the number of stars formed in the mass interval ($m$, $m + dm$) and in the time interval ($t$, $t + dt$) is
\begin{equation}
\phi \left( m \right) \psi \left( t \right) dm \ dt,
\label{eq:eq2.1}
\end{equation}
where $\psi \left( t \right)$ is the total mass of stars formed per unit time, and $\phi \left( m \right)$, which may itself be a time-dependent function, is therefore normalized so that
\begin{equation}
\int_{0}^{\infty} m \phi \left( m \right) dm = 1.
\label{eq:eq2.2}
\end{equation}
The function $\psi \left( t \right)$ is called the star formation rate (SFR) and $\phi \left( m \right)$ is called the initial mass function (IMF). It is often useful to approximate $\phi \left( m \right)$ by a power law, at least over some range of masses, and in these cases the notation used here is
\begin{equation}
\phi \left( m \right) \propto m^{-\left( 1 + x \right)},
\label{eq:eq2.3}
\end{equation}
where $x$ is called the slope of the IMF\footnote{Quantities equivalent to $x$, $-x$, $\left( 1 + x \right)$, and $-\left( 1 + x \right)$ have been called ``the slope of the IMF'' by different authors.}.
A comprehensive review of the IMF in the Solar neighborhood and elsewhere has recently been written by \citet{Scalo1978TheSpectrum}, so the following account includes only those aspects that are most relevant to the rest of this article.
\subsubsection{The local IMF}
\label{subsubsec:Local IMF}
The IMF in the Solar neighborhood can be derived from counts of field stars, using principles originally derived by \citet{Salpeter1955TheEvolution.}, and generalized to allow for a time-dependent SFR by \citet{Schmidt1959TheFormation., Schmidt1963TheMass.}, among others. A recent thorough rediscussion of the data and methods is given by \citet{Miller1979TheNeighborhood}.
Stars are counted as a function of absolute magnitude, and the counts must first be reduced to the present mass distribution of stars, $n(m)$, where $n(m) \ dm$ will denote the present number of MS stars in the mass interval ($m$, $m + dm$). Several non-trivial problems arise at this stage, including the following:
\begin{enumerate}
\item The number of stars in each absolute magnitude interval must be corrected for post-MS stars, which requires a knowledge of color or spectral type; these corrections can be large -- e.g., near $M_{\rm V} = +1$ there are about equal numbers per unit volume of A dwarfs and K giants.
\vspace{1mm}
\item Field stars of different masses have very different distributions perpendicular to the Galactic plane, the scaleheight decreasing with increasing mass. The function $n(m)$ obtained for stars in a unit volume therefore differs systematically from the function for a unit column perpendicular to the plane. Because one is normally interested in the whole population of stars formed at all heights (or formed in the plane and accelerated to more extended orbits), near the Sun's galactocentric distance, the star counts should be reduced to the \emph{average number per square parsec} at this distance from the Galactic center.
\vspace{1mm}
\item The transformation between magnitudes and masses of MS stars depends on their chemical composition.
\vspace{1mm}
\item Stars evolve in the HR diagram while on the MS, so there is not a unique mass--luminosity relation even for a given composition.
\end{enumerate}
The papers cited above discuss how these problems can be handled, and the resulting uncertainties in $n(m)$.
Let us assume that we know the present mass distribution of MS stars per square parsec in the Solar neighborhood, $n(m)$. This function can be used to estimate the IMF and SFR as follows. Stars with lifetimes $\left( \tau_{\rm m} \right)$ greater than the age of the Galaxy $\left( t_{1} \right)$ have accumulated since star formation began ($t \equiv 0$), so Equation~(\ref{eq:eq2.1}) gives directly
\begin{equation}
n(m) = \int_{0}^{t_{1}} \phi \left( m \right) \psi \left( t \right) dt, \: \: \tau_{\rm m} \geq t_{1}.
\label{eq:eq2.4}
\end{equation}
If the IMF is assumed constant, this expression simplifies to
\begin{equation}
n(m) = \phi \left( m \right) \overline{\psi}_{1} t_{1}, \: \: \tau_{\rm m} \geq t_{1},
\label{eq:eq2.5}
\end{equation}
where $\overline{\psi}_{1}$ is the average past SFR; Equation~(\ref{eq:eq2.5}) can be used in any case if $\phi \left( m \right)$ is interpreted as the average past IMF. The mass with $\tau_{\rm m} = t_{1}$ is called the present turnoff mass ($m_{1}$) since it defines the lowest MS turnoff point in the HR diagram for local field stars. In practice, one must consider how the mass--lifetime relation depends on metallicity, but to illustrate the principles here these effects will be ignored; approximate round numbers are $m_{1} \simeq 1 \: \rm M_{\odot}$, $t_{1} \simeq 10 \: \rm Gyr$.
The present population of stars with $\tau_{\rm m} < t_{1}$ includes only those that were formed at times less than $\tau_{\rm m}$ ago, so from Equation~(\ref{eq:eq2.1}) their mass distribution is
\begin{equation}
n(m) = \int_{t_{1} - \tau_{\rm m}}^{t_{1}} \phi \left( m \right) \psi \left( t \right) dt, \: \: \tau_{\rm m} < t_{1}.
\label{eq:eq2.6}
\end{equation}
If $\phi(m)$ is constant, it can be taken out of the integral, but $n(m)$ still depends on details of the past SFR. A simpler equation can be written for stars with lifetimes shorter than the present timescale for changes in $\psi \left( t \right)$:
\begin{equation}
n(m) = \phi \left( m \right) \psi_{1} \tau_{\rm m}, \: \: \tau_{\rm m} \ll t_{1},
\label{eq:eq2.7}
\end{equation}
where $\psi_{1}$ is the present SFR. The function $n(m)$ is supposed to be averaged over the patchy distribution of the youngest stars in the Solar neighborhood, so Equation~(\ref{eq:eq2.7}) is probably a good approximation for $m \gtrsim 2 \: \rm M_{\odot}$, i.e., for stars with lifetimes $\lesssim 1 \: \rm Gyr$ and MS spectral types earlier than about A0; it would hold for all $m > m_{1}$ if the SFR were constant.
From the empirical function $n(m)$, therefore, one can use Equation~(\ref{eq:eq2.5}) to derive the shape of the (average) IMF for stars below about $1 \: \rm M_{\odot}$ and Equation~(\ref{eq:eq2.7}) to derive the shape of the IMF for stars above about $2 \: \rm M_{\odot}$. These two pieces of the IMF are not determined with the same multiplicative factors, since the quantities actually determined are $\phi \left( m \right) \overline{\psi}_{1} t_{1}$ and $\phi \left( m \right) \psi_{1}$, respectively. In particular, the relative values of the IMF for $1 \: \rm M_{\odot}$ and $2 \: \rm M_{\odot}$ depend on the ratio
\begin{equation}
T_{1} \equiv \frac{\overline{\psi}_{1} t_{1}}{\psi_{1}},
\label{eq:eq2.8}
\end{equation}
which is evidently a timescale for star formation in the Solar neighborhood.
There is an intermediate mass interval, $1 \lesssim m / \rm M_{\odot} \lesssim 2$, for which Equation~(\ref{eq:eq2.6}) cannot be simplified, so the shape of the IMF cannot be derived independently of details of $\psi(t)$. This gap is usually filled by assuming that the functions for higher and lower masses can be smoothly interpolated, as illustrated, for example, by \citet[][Figure~1]{Schmidt1963TheMass.}. In this way, $\phi(m)$ for the intermediate mass range is determined without too much ambiguity and constraints are set on the quantity $T_{1}$.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig4.png}
\caption{The initial mass function in the Solar neighborhood, shown in the form $\log \left[ \psi_{1} m \phi(m) \right]$ versus $\log \ m$, where $\phi(m)$ is the IMF itself, normalized as in Equation~(\ref{eq:eq2.2}), and $\psi_{1}$ is the present star formation rate in $\rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$. Masses are in solar units. \emph{Solid line and shaded area}: range of functions derived by \citet{Miller1979TheNeighborhood}, allowing for uncertainties in the basic data and in the history of the star formation rate. \emph{Dashed line}: power-law segments, given by Equation~(\ref{eq:eq2.9}); the slopes of these lines are $-x$ in the notation of Equation~(\ref{eq:eq2.3}). (The dashed line and the shaded area are discrepant for $-0.4 < \log \ m < 0.0$ because different stellar luminosity functions were used).}
\label{fig:fig4}
\end{figure}
\medskip
The IMF derived along the preceding lines by \citet{Miller1979TheNeighborhood} is illustrated in Figure~\ref{fig:fig4}, where the quantity plotted (logarithmically) is $m \phi(m) \psi_{1}$. The solid line is their analytical relation for the case of a constant SFR and a Galactic age $t_{1} = 12 \: \rm Gyr$: $m \phi(m) \psi_{1} = 3.83 \exp [ -1.09 ( \log m + 1.02 )^{2} ] \: \rm pc^{-2} \ Gyr^{-1}$ (with $m$ in solar units). The shaded area shows the range of uncertainty allowed by their analytic fits to limiting cases, based on uncertainties in the basic data and the limits for smoothly joining the low- and high-mass ends, as explained above. \citet{Miller1979TheNeighborhood} stress that, if one is making models for the Solar neighborhood, the choice of an IMF within the shaded area should be made consistently with the adopted SFR function, for otherwise the two functions together would not reproduce the observed star counts.
Power laws are very tractable analytically, so it is tempting to fit straight lines to the IMF derived from the star counts. \citet{Salpeter1955TheEvolution.} originally derived a famous slope $x = 1.35$ for the local IMF, but subsequent star counts and stellar lifetimes show that a single power law for the whole mass spectrum is a poor approximation; serious systematic errors can be introduced in some applications by ignoring the increase of slope with mass in the local IMF. The four dashed straight lines in Figure~\ref{fig:fig4} are intended as a compromise between the data and analytical convenience. (The discrepancy with the shaded area between $0.4$ and $1 \: \rm M_{\odot}$ arises because the dashed line for those masses is based on the luminosity function of \citet{Wielen1974TheCatalogue.}, which is flatter than the function adopted by \citealp[][Figure~1]{Miller1979TheNeighborhood}). The equations of these four lines, which have slopes $-x$ in the logarithmic plot (see Equation~\ref{eq:eq2.3}), are:
\begin{equation}
\begin{aligned}
m \phi(m) \psi_{1} = & \ 1.00 m^{-0.25}, \: \: 0.4 < m \leq 1.0, \\
& \ 1.00 m^{-1.0}, \: \: 1.0 < m \leq 2.0, \\
& \ 1.23 m^{-1.3}, \: \: 2.0 < m \leq 10.0, \\
& \ 12.3 m^{-2.3}, \: \: 10.0 < m \leq 50.0,
\label{eq:eq2.9}
\end{aligned}
\end{equation}
in units of $\rm pc^{-2} \ Gyr^{-1}$, with masses in solar units. (It is fortuitous that the coefficient 1.00 occurs in these equations). A slope is not given for $m < 0.4 \: \rm M_{\odot}$ because M dwarfs contribute little to the light of any galaxy and nothing directly to chemical evolution; their contribution to the total mass can be included with the invisible objects below $0.1 \: \rm M_{\odot}$.
Integrating Equation~(\ref{eq:eq2.9}) with respect to mass, we find that the present SFR for masses $0.4 - 50 \: \rm M_{\odot}$ is $3.0 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$. The total SFR, $\psi_{1}$, could be derived if we knew the contribution from objects below $0.4 \: \rm M_{\odot}$, and $\psi_{1}$ in turn would allow the IMF to be normalized as in Equation~(\ref{eq:eq2.2}). As described in Section~\ref{subsubsec:Local SFR}, \citet{Miller1979TheNeighborhood} find $\psi_{1}$ in this way by defining as ``stars'' only objects above $0.1 \: \rm M_{\odot}$, for which there are star counts giving the shape of the IMF. Alternatively, $\psi_{1}$ can be estimated indirectly (see Section~\ref{subsubsec:Local SFR}). The values found are within a factor of ${\sim} 2$ of $10 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$, which is several times greater than the rate for objects above $0.4 \: \rm M_{\odot}$.
\subsubsection{The IMF at other times and places}
\label{subsubsec:Other IMF}
Several questions can be asked about the constancy of the IMF. Is the IMF the same in regions other than the Solar neighborhood? Is it constant in time, locally and elsewhere? If variations occur, do they involve the shape of the IMF for stars above say $0.4$ or $1 \: \rm M_{\odot}$, or do they involve only the mass fraction in ``inert'' low-mass objects?
\medskip
Star clusters in the Galaxy and the Magellanic Clouds have a variety of MS luminosity functions indicating significant variations in the IMF on the scale of clusters (\citealp{DaCosta1977TheClusters, Freeman1977StarGalaxies}; others reviewed by \citealp{Scalo1978TheSpectrum}); also, there are regions of the Milky Way where the only newborn stars seem to be T Tauri stars, with masses $\lesssim 2 \: \rm M_{\odot}$, and other regions abounding in young OB stars. These observations, however, do not necessarily imply variations on galaxy-wide scales in averages over many sites of star formation. In fact, counts of field stars in nearby galaxies have not revealed any significant deviations from the local IMF \citep{Butcher1977ACloud, Hardy1977TheRegions, Lequeux1979ComparisonGroup}. The clearest evidence for large-scale variations is the mass function for nearby halo stars of \citet{Schmidt1975TheStars}; this function has $x \simeq 2$ for $0.25 - 0.75 \: \rm M_{\odot}$, whereas local disk stars have $x < 1$ in the same mass range. There are also some early-type spiral galaxies with widespread patches of young blue stars but a deficiency of \textsc{H~ii} regions, suggesting that the IMF is deficient in O stars; examples are the Sombrero galaxy, M104 \citep{vandenBergh1976b, Schweizer1978GalaxiesTails}, and NGC~2841 \citep{Kormendy1977}.
The upper part of the IMF $\left( m > 0.4 \: \rm M_{\odot} \right)$ can be studied indirectly by comparing colors of galaxies with those of models based on various IMFs. As discussed in Section~\ref{sec:Colors and Star Formation Rates}, the results are consistent with the local IMF holding (on large scales) everywhere, but in some respects the test is rather insensitive. For example, a galaxy like the Sombrero is so dominated by old yellow--red stars that the \emph{integrated} colors at optical wavelengths would be changed negligibly by the presence or absence of stars above $10 \: \rm M_{\odot}$ in the IMF. The red--infrared spectra of galaxies show that giant stars, rather than late dwarfs, provide most of the light at these wavelengths, and this result sets some constraints on the IMF for stars of ${\sim} 0.4 - 1 \: \rm M_{\odot}$: if the IMF were much steeper than the local function (say if $x > 2$), low-mass dwarfs would contribute much of the infrared light, contrary to the spectroscopic information.
Because of this giant dominance in the infrared light, photometric observations are very insensitive to the mass fraction in stars below $0.4 \: \rm M_{\odot}$, i.e. M dwarfs; these stars could be numerous enough to dominate the mass of the system while contributing very little light! A handle on the mass fraction in invisible stars is given by the mass-to-luminosity ($M / L$) ratio, which is crudely a ratio of very low-mass to more massive stars. In the Solar neighborhood, the integrated luminosity of known stars (per $\rm pc^{3}$), including giants, divided by the dynamically-estimated density of mass, yields a ratio $M / L_{\rm B} \simeq 3 \: \rm M_{\odot} / L_{B \odot}$ \citep{Faber1979MassesGalaxies}. If the IMF and proportion of invisible matter were invariant, all galaxies should have this $M / L$ ratio, apart from a systematic increase by a factor of ${\sim} 10$ from the bluest galaxies to the reddest (because of the decreasing fraction of young blue stars), and apart from a scatter ${\sim} 10\%$ due to different mass fractions of ISM. In fact, observed $M / L$ values show a \emph{scatter} by a factor of 10 or more at a given color, and a tendency to increase from values of a few in the central parts of galaxies to ${\sim} 100$ in the outer regions \citep{Faber1979MassesGalaxies}. It appears that different proportions of hidden matter exist in different galaxies, and that the invisible mass fraction increases with radius. If this matter condensed before ordinary star formation began in galaxies, as in some pictures of the formation of heavy invisible halos \citep[e.g.][]{White1978CoreClustering}, then it would have no direct effect on chemical or photometric evolution. (It would have indirect effects, because the motions of gas and stars in the galaxy would be influenced by the hidden mass). However, if there are variations in the low-mass part of the stellar IMF, chemical evolution would be affected significantly, because variable proportions of stellar matter would be returned to the ISM by evolving stars.
Certain chemical properties of galaxies can be explained by invoking variations in the shape of the IMF for massive stars. For example, the extreme paucity of metal-poor dwarfs in the Solar neighborhood could be due to an initial burst of massive metal-producing stars; and variations in relative abundances of heavy elements could be due to variable proportions of the stars that synthesize such elements. But in all cases, other explanations not involving variations in the IMF are available, as discussed in Sections~\ref{sec:Evolution solar neighborhood}~and~\ref{sec:Chemical Evolution of Galaxies}.
\medskip
In summary, significant variations in the IMF may be rare enough for the assumption of a universal function (the local IMF) to be useful for many contexts. However, it is equally relevant to ask how galactic evolution would be affected by changes in both the form of the IMF for visible stars and the relative weight of very low-mass objects.
\subsection{Rates of Star Formation}
\label{subsec:SFRs}
The rate of star formation is one of the main factors in galactic evolution. The shape of a galaxy depends on the timescale for star formation relative to collapse and dissipation timescales, the colors and luminosity depend on the age distribution of stars, and chemical evolution depends on the SFR relative to the gas flow rates and the gas mass. Various ways of estimating SFRs in galaxies are reviewed here.
\subsubsection{The local SFR}
\label{subsubsec:Local SFR}
There are several ways of estimating the present SFR in the Solar neighborhood, which is the quantity in $\rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$ denoted $\psi_{1}$ above. Stars above ${\sim} 2 \: \rm M_{\odot}$ are no problem, because they are so short-lived that their numbers scale directly with $\psi_{1}$ (Equation~\ref{eq:eq2.7}), but the ages of less massive stars are unknown, as is the mass fraction contained in objects below $0.1 \: \rm M_{\odot}$. Most of the methods that have been suggested for overcoming these problems reduce in principle to estimating the average past SFR ($\overline{\psi}_{1}$) and then the ratio $\overline{\psi}_{1} / \psi_{1}$ \footnote{Another approach in the literature is to find the formation rate of relatively massive stars (above $2 \: \rm M_{\odot}$ or a greater limit), and to scale to the total SFR via an \emph{assumed} IMF. Obviously, the problems mentioned above are sidestepped rather than solved, so the conclusions should be regarded as correspondingly uncertain.}. The following two methods are typical.
\begin{enumerate}
\item A dynamical estimate of the total surface density in the Solar neighborhood, called the ``Oort limit'', can be derived from motions of stars perpendicular to the Galactic plane; its value is probably ${\sim} 80 - 100 \: \rm M_{\odot} \ pc^{-2}$ \citep{Oort1965StellarDynamics}. Of the total surface density, ISM in various forms accounts for $5 - 10 \: \rm M_{\odot} \ pc^{-2}$ \citep{Gordon1976CarbonNucleons}, so stars are presumed to contribute ${\sim} 70 - 95 \: \rm M_{\odot} \ pc^{-2}$ (neglecting any possible non-stellar condensed objects). This is not the total mass of stars ever formed, because some fraction $R$ has been ejected back to the ISM by evolving and dying stars; $R$ depends on the IMF, whose normalization is not known without $\psi_{1}$ itself, but all reasonable normalizations give $R$ in the range ${\sim} 0.1 - 0.3$ (as shown in Section~\ref{subsec:Assumptions}). The mass of stars ever formed is thus $10\% - 30\%$ greater than the mass now present; the mean past SFR is thus given by $75 < \overline{\psi}_{1} t_{1} < 125 \: \rm M_{\odot} \ pc^{-2}$.
The next step is to use the ratio $T_{1}$ (Equation~\ref{eq:eq2.8}), derived by joining smoothly the IMFs with different multiplicative factors for stars above $2 \: \rm M_{\odot}$ and below $1 \: \rm M_{\odot}$, respectively. Plausible values of $T_{1}$ lie between 5 and $20 \: \rm Gyr$ \citep{Tinsley1976ChemicalInhomogeneities, Miller1979TheNeighborhood}, so we are left with the following range of values for $\psi_{1}$ itself: $1.5 < \psi_{1} < 25 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$. The uncertainties at various stages of the derivation are mutually independent, and the $20\%$ uncertainty in $R$ is relatively unimportant. A more stringent upper limit could be obtained if one could assume that the SFR has not increased since the time the oldest disk stars formed, ${\sim} 12 \: \rm Gyr$ ago; then $T_{1} > 12 \: \rm Gyr$ and $\psi_{1} < 10 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$. However, although the assumption of a monotonically decreasing SFR is common, recent work on the age distribution of disk stars \citep{Twarog1980AnNeighborhood} shows that the SFR more likely \emph{increased} between 12 and $5 \: \rm Gyr$ ago, and was approximately constant in the last ${\sim} 5 \: \rm Gyr$. In this case, $T_{1}$ is probably between about 5 and $10 \: \rm Gyr$, and the above limits on $\overline{\psi}_{1} t_{1}$ lead to $8 < \psi_{1} < 25 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$. Values only ${\sim} 10 - 20\%$ of these have usually been derived in the literature, on the assumption that $\psi (t)$ has been a decreasing function for at least $10 \: \rm Gyr$.
\vspace{1mm}
\item Instead of using the Oort limit to estimate the total mass of stars ever formed, one can use star counts to obtain the existing population (and to allow for the total mass of stars that have evolved away in the time since they formed) down to ${\sim} 0.1 \: \rm M_{\odot}$, and neglect any less massive objects. In this way, \citet{Miller1979TheNeighborhood} derive a range of values $43 < \overline{\psi}_{1} t_{1} < 144 \: \rm M_{\odot} \ pc^{-2}$. Combining these values with appropriate limiting values of $T_{1}$, they find self-consistent values in the range $3 < \psi_{1} < 7 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$.
\end{enumerate}
Method (i) above gave larger values than those of \citet{Miller1979TheNeighborhood} for two reasons: the estimated mass of stars now present above $0.1 \: \rm M_{\odot}$ falls short of estimates based on the Oort limit (although the discrepancy may not be significant), and \citet{Miller1979TheNeighborhood} considered both decreasing and increasing SFRs. Altogether, it is plausible to suggest that $\psi_{1}$ lies within a factor of about two of $10 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$.
\subsubsection{The SFR elsewhere}
\label{subsubsec:Elsewhere SFR}
Estimates of the SFR in other places use various indicators of the numbers of fairly massive stars, and convert these to a total SFR by means of an assumed IMF, which is usually the local function (or an assumed power-law approximation). Methods that have been used include the following.
\begin{enumerate}
\item The brightest individual stars, supergiants, and occasionally upper-MS stars, can be counted in nearby galaxies. Apart from complications due to differences in chemical composition, the numbers of massive stars in given stages of evolution should be directly proportional to the SFR, if the IMF is the same everywhere; their numbers relative to counts of similar stars in the Solar neighborhood thus scale fairly directly with the SFR. \citet{Lequeux1979ComparisonGroup} has applied this method to galaxies of the Local Group, using various types of stars, and several authors \citep[e.g.][]{Searle1973TheGalaxies, Larson1974PhotometricGalaxies} have used such rare objects as supernovae in the same way for more distant galaxies.
\vspace{1mm}
\item If hot young stars are present in a region, they generally emit most of the Lyman continuum photons that are available for ionizing interstellar hydrogen. The flux of $\rm H \alpha$, $\rm H \beta$, or free--free radio emission is therefore approximately proportional to the SFR. This method has been applied to external galaxies by \citet{Cohen1976H-alphaGalaxies} using integrated $\rm H \alpha$ equivalent widths, and by \citet{Huchra1977StarGalaxies} using $\rm H \beta$; \citet{Smith1978StarGalaxy.} have studied SFRs in the Galaxy using radio emission from \textsc{H~ii} regions.
\vspace{1mm}
\item The integrated colors of galaxies depend strongly on their relative proportions of young and old stars; as a result, UBV colors in general provide a rough estimate of the ratio (SFR) / (mass of old stars). This approach will be discussed in detail in Section~\ref{sec:Colors and Star Formation Rates}.
\vspace{1mm}
\item Regions of star formation are often heavily obscured by dust, but even so the infrared luminosity (starlight re-radiated by dust grains) gives a measure of the SFR, as discussed in Section~\ref{subsubsec:Highly reddened galaxies}.
\end{enumerate}
Most of these methods rely strongly on the assumption of a universal IMF, since they sample only the very massive stars; integrated colors can reveal large departures from an assumed IMF (Section~\ref{subsubsec:Other IMF}), but even these are completely blind to stars below ${\sim} 0.5 \: \rm M_{\odot}$. Even granted the assumption of universality, important uncertainties arise in scaling from the observations to a total SFR, because the statistics of very massive stars are poor. Large systematic errors can arise from the use of a single power law for the whole IMF (a common practice), as can be clearly seen from Figure~\ref{fig:fig4}.
Despite the uncertainties, the results of various studies have led to a coherent picture. In the Galaxy, some $10\%$ of the current star formation is occurring in the innermost $1 \: \rm kpc$, and most of the remainder is concentrated in a ring between 5 and $8 \: \rm kpc$ from the center, which is the site of most of the Galaxy's giant molecular clouds, infrared emission, and other signs of intense star formation. (There are several relevant reviews in the symposium volume edited by \citealp{Burton1979TheGalaxy}). Morphologically normal galaxies show a systematic trend of SFR with type, which can be described as a progression from a very short timescale for star formation in the earliest types to very long in the latest. Some peculiar galaxies appear to be undergoing intense bursts of star formation. These properties will be discussed later in Section~\ref{sec:Colors and Star Formation Rates}.
\subsubsection{Factors affecting the SFR}
\label{subsubsec:SFR factors}
An obvious question to ask is \emph{why} stars form at various rates in various places. Factors affecting SFRs have been reviewed by \citet{Larson1977b}, and some salient points will be mentioned here.
\medskip
A popular assumption, following \citet{Schmidt1959TheFormation.}, is that the SFR varies as a power $n$ of the gas density. However, attempts to determine $n$ empirically run into two problems: $n$ appears to vary within and among galaxies, indicating that the gas density is not the only relevant quantity; and the very definition of ``gas density'' is ambiguous since it depends strongly on the spatial resolution of the observations. Consequently, a power-law formula for the SFR cannot be given the status of a physical law, but it can be regarded as a useful parametrization in some circumstances (as in the later study of \citealp{Schmidt1963TheMass.}).
The physical environment needed for star formation is evidently cool, gravitationally bound gas clouds, so the main question is: under what conditions does interstellar gas clump into clouds? \citet{Larson1977b} argues that gas compression is the key requirement for star formation, and that both large- and small-scale dynamical processes are important. On small scales, interstellar shock and ionization fronts have been widely discussed as causes of star formation. These processes can be self-sustaining, since supernovae and \textsc{H~ii} regions arising from recently-formed stars may induce further star formation nearby \citep[e.g.][]{Elmegreen1977SequentialAssociations}. Large-scale compression mechanisms that may lead to star formation include gravitational settling of gas into a thin layer, density waves of bar or spiral form, high-velocity collisions between gas streams in young or interacting galaxies, and accretion of intergalactic clouds. If star formation can be self-propagating over long distances, differential rotation can spread the active regions out to produce spiral structure \citep{Gerola1978StochasticGalaxies}. Large-scale processes are especially relevant to the long-term evolution of galaxies, and since some of them involve interactions with external matter, galaxies do not necessarily evolve as independent systems. This point has been emphasized in a review by \citet{Saar1978InteractionGalaxies}. Further discussion of star formation can be found in the reviews cited, many references therein, and in a symposium volume edited by \citet{DeJong1977StarFormation}.
\medskip
Although some understanding has developed of the conditions under which stars form, no formula for the SFR has emerged that can make a useful \emph{prediction} for any galaxy. \citet{Lynden-Bell1977OnFormation} has pointed out that a proper expression for the SFR would contain so many unknown parameters as to be useless. He wrote, ``This rate $R$ (the SFR) probably depends on gas density $\rho_{\rm g}$, gas sound speed $c_{\rm s}$, shock frequency $\omega_{\rm s}$, shock strength $V_{\rm T} / c_{\rm s}$, gas rotation $\Omega$ and shearing rate $A$, the magnetic field strength $\lvert \vec{B} \rvert$, the gas metal abundance $Z$, and possibly the background star density $\rho_{*}$. Thus
\begin{equation*}
R = R \left( \rho_{\rm g}, \ c_{\rm s}, \ \omega_{\rm s}, \ V_{\rm T} / c_{\rm s}, \ \Omega, \ A, \ \lvert \vec{B} \rvert, \ Z, \ \rho_{*} \right).
\label{eq:eq2.Lynden-Bell}
\end{equation*}
However, if we knew the true functional form of $R$ and offered it to a galaxy builder he would probably tell us `Oh, go and jump in the lake, that's far too complicated'.''
The only feasible approach to galaxy building is to be less ambitious than to want a physical formula for the SFR. Instead, schematic expressions, guided by the above ideas on relevant factors, can be tested to see how galactic evolution is affected by various parameters. Some examples already mentioned just above and in Section~\ref{sec:Overview} are that self-propagating star formation can lead naturally to spiral structure, and that the final shape of a galaxy depends on the ratios of the timescales for star formation, collapse, and gaseous dissipation in the protogalaxy. Many examples of schematic expressions for the SFR, in models for chemical and photometric evolution, will be considered later.
\subsection{Stellar Evolution Beyond the Main Sequence}
\label{subsec:Beyond MS}
Late stages of stellar evolution are important to galaxies because stars burn much of their nuclear fuel after leaving the main sequence. The products of this late fuel guzzling provide most of the chemical enrichment of galaxies, and the energy liberated provides most of their integrated light. The following outline of stellar evolution and nucleosynthesis emphasizes aspects that will be relevant later in the discussion of galaxies, and necessarily only skims the surface of these topics. Further details and references can be found in the comprehensive review by \citet{Trimble1975TheElements}, and in other papers cited below.
\subsubsection{Stars near Solar mass}
\label{subsubsec:Solar mass}
Giant stars of approximately Solar metallicity and mass are especially interesting because they provide most of the light of elliptical and S0 galaxies and the nuclear bulges of spirals. The dominant old-disk population of the Solar neighborhood provides a convenient sample of such stars for detailed study; these may be somewhat younger than most stars in spheroidal systems, but evolution on the giant branch is thought to be rather insensitive to the initial stellar mass in the small range whose lifetimes are ${\sim} 5 - 15 \: \rm Gyr$.
\medskip
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig5.png}
\caption{Color--magnitude diagrams for the old open cluster M67 (\citealp{Racine1971Photometry+12}; \emph{filled circles}) and for red giants in old-disk groups (\citealp{Eggen1962EvidenceCollapsed., TinsleyGunn1976b}; \emph{open circles}); \emph{crosses} denote variable stars. The M67 data have been reduce to $M_{\rm bol}$ versus $R-I$, as in \citet{TinsleyGunn1976b}. This diagram illustrates the types of stars that apparently contribute most of the light in old, metal-rich systems, such as giant elliptical galaxies and the nuclear bulges of spirals. Because of the small total cluster population, there are no late giants in M67, and because of selection effects the visually brightest stars are over-represented in the groups.}
\label{fig:fig5}
\end{figure}
Figure~\ref{fig:fig5} gives color--magnitude diagrams of the old open cluster M67 and a sample of field giants, which are necessary for completeness because the cluster is so poor that it contains none of the short-lived M giants. Selection effects in the field sample mean that earlier giants are badly under-represented relative to later types, and an unbiased luminosity function would show roughly a factor of two decrease in the number of stars per bolometric magnitude interval, for each magnitude increase in luminosity up the giant branch. Some relevant aspects of the stellar evolution responsible for this distribution of stars are as follows \citep[e.g.][]{Iben1974PostStars, TinsleyGunn1976b}:
\begin{enumerate}
\item When the central $10\%$ of hydrogen has been converted to helium, the star leaves the MS and its site of hydrogen burning moves from the center to a shell around the helium core; as the shell is established, the star moves along the subgiant branch (at $M_{\rm bol} \simeq 2.5$ in Figure~\ref{fig:fig5}) to the base of the giant branch, at which point the helium core mass is ${\sim} 0.2 \: \rm M_{\odot}$.
\vspace{1mm}
\item During the first ascent of the giant branch to the so-called ``tip'' at $M_{\rm bol} \simeq -3$, the helium core grows to ${\sim} 0.45 \: \rm M_{\odot}$. Although this ascent occupies only about $10\%$ of the preceding lifetime, the energy emitted, which is nearly proportional to the mass of hydrogen burned, is greater.
\vspace{1mm}
\item Helium ignition in the degenerate core -- the ``helium flash'' -- halts the ascent of the giant branch, and the star very quickly drops to a lower luminosity, $M_{\rm bol} \simeq 0.5$, where it remains for about $10^{8} \: \rm yr$ while the core helium is burned to carbon (and a little oxygen). This stage appears as the horizontal branch to the left of the giant branch in HR diagrams of globular clusters, but most metal-rich stars form a clump of early K giants superposed on the ascending giant branch; some metal-rich blue horizontal-branch stars are found in the old-disk population, and these may represent a minority that lost a large fraction of their envelopes at the helium flash \citep[e.g.][]{Butler1976MetalWindow}. Although the total energy emitted by core-helium-burning stars is relatively small, blue horizontal-branch stars can make an important contribution to the light of galaxies whose bluest stars otherwise are at the turnoff.
\vspace{1mm}
\item The star then ascends the giant branch for a second time (called the asymptotic giant branch stage because it appears as an asymptote to the left of the first giant branch in the HR diagrams of globular clusters), with two shells burning hydrogen and helium respectively. The star can become brighter and cooler than the first giant branch ``tip'', as seen in Figure~\ref{fig:fig5} but not in all HR diagrams for globular clusters.
\vspace{1mm}
\item The end of this stage of evolution is determined by mass loss, without which the whole star would become a carbon-oxygen white dwarf. In fact, there is much evidence that several tenths of a solar mass of envelope are lost before the burning shells reach out to consume the whole initial mass of hydrogen. For example, the average mass of white dwarfs is ${\sim} 0.7 \: \rm M_{\odot}$, well below the turnoff mass (${\sim} 1 \: \rm M_{\odot}$) for field stars; planetary nebulae, most of which belong to the old-disk population, represent ${\sim} 0.2 \: \rm M_{\odot}$ of ejected matter; and spectroscopic observations of M giants show that they are losing mass in a wind. The existence of both very late M giants and blue horizontal-branch stars in the old-disk population suggests that mass loss occurs at a variety of rates among stars of a given initial mass and composition. Since the average final core mass of solar-mass stars must be the white dwarf mass, ${\sim} 0.7 \: \rm M_{\odot}$, the energy liberated during the second ascent of the giant branch is comparable to that liberated on the first ascent, with the difference that most of it now appears in the near infrared from very cool stars.
\end{enumerate}
Current stellar models do not predict post-MS evolution accurately enough to be used directly in synthetic galaxies. The main uncertainties are the depth of the surface convection zone, which determines the effective temperature at a given luminosity, and mass loss; a related problem is the outward mixing of products of nucleosynthesis, which are observed on the surfaces of many giants but cannot be predicted in detail. Consequently, galaxy models must rely heavily on empirical studies of giant populations. The implications of these problems for understanding photometric properties of elliptical galaxies are reviewed by \citet{Faber1977ThePopulations}. Mass loss and mixing also mean that even these low-mass stars contribute significant amounts of newly synthesized elements to the ISM, including for example carbon and nitrogen, which are often overabundant at the surfaces of giants and in planetary nebulae. The isotopes $^{13}$C and $^{14}$N are by-products of hydrogen burning by the CNO cycle, while $^{12}$C must be a product of core helium burning \citep[e.g.][]{Trimble1975TheElements, IbenI.1978OnMedium}.
\subsubsection{Stars of $1 - 4 \: \rm M_\odot$}
\label{subsubsec:1 - 4 Msun}
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{Figures/Harris_1976_fig2.png}
\includegraphics[width=0.49\textwidth]{Figures/Harris_1976_fig3.png}
\includegraphics[width=0.49\textwidth]{Figures/Harris_1976_fig4.png}
\includegraphics[width=0.49\textwidth]{Figures/Harris_1976_fig5.png}
\caption{Composite HR diagrams for (mainly) post-main-sequence stars in galactic clusters grouped by age \citep{Harris1976EvolvedClusters.}. Ages ($\tau$ in yr) are indicated. These diagrams illustrate the distribution of supergiants expected in galaxies containing stars of the appropriate ages.}
\label{fig:fig6}
\end{figure*}
White dwarfs appear in open clusters where their progenitors must have been stars with initial masses up to ${\sim} 4 \: \rm M_{\odot}$ \citep[e.g.][]{Tinsley1975b}. The substantial mass loss required is in agreement with stellar models that allow for winds \citep{Fusi-Pecci1976OnSun., Mengel1976SteadyIgnition.}; these models suggest that stars up to some limit ${\sim} 3 - 6 \: \rm M_{\odot}$ lose so much mass that they fizzle before the point of carbon ignition, which would otherwise occur (possibly explosively) when the degenerate core reached $1.4 \: \rm M_{\odot}$. Therefore, stars with MS lifetimes in the whole range from ${\sim} 10^{8}$ to ${\sim} 10^{10} \: \rm yr$, from mid-B to early G on the MS, evolve internally similarly to the solar-mass stars discussed above, and they make similar contributions to nucleosynthesis.
\subsubsection{Stars above $8 \: \rm M_\odot$}
\label{subsubsec:8 Msun}
Stars initially more massive than ${\sim} 8 \: \rm M_{\odot}$, which are early B and O stars on the MS, are believed to ignite carbon (quietly) in a non-degenerate core and to proceed through many stages of nuclear burning. Models acquire an onion-skin structure of successive elements, with unaltered envelope material (mostly hydrogen) on the outside, then layers of helium, carbon, oxygen, neon, magnesium, etc., with the ashes of each nuclear reaction providing fuel for the next. These stars and their contributions to nucleosynthesis are discussed by \citet{Arnett1978OnStars}, \citet{Weaver1978PresupernovaStars}, and \citet{Chiosi1979TheGalaxy}, the last stressing the very important effects of mass loss. Theoretical models predict that the stars undergo core collapse after the last exoergic reaction, formation of iron-peak elements, and the event is expected to send a shock wave through the star leading to explosive nucleosynthesis of most of the less abundant elements. Because Type~II supernovae are associated with the most massive stars in galaxies, they are believed to represent these explosive stellar deaths. The lower initial mass limit for stars with this fate is somewhere in the range $6 - 10 \: \rm M_{\odot}$, depending on the structure of stellar cores at carbon ignition; it is sometimes suggested that there is also an upper limit for such deaths, stars above the higher limit collapsing to black holes that swallow most of the new elements \citep[e.g.][]{Wheeler1978a}.
\medskip
The evolution of massive stars in the HR diagram (as well as internally) must be strongly affected by mass loss, since winds of ${\sim} 10^{6} \: \rm M_{\odot} \ yr^{-1}$ are inferred from the spectra of MS O stars and much higher mass-loss rates are common in supergiants \citep{Conti1978MassStars}. These rates mean that the stellar mass is changing significantly on an evolutionary timescale. The evolutionary tracks of such stars in the HR diagram cannot be predicted reliably \citep[e.g.][]{Chiosi1978MassiveMass-loss., Stothers1978StellarStars}, so models for galaxies with young stars are subject to some uncertainty. Empirical illustrations of the evolutionary tracks of massive stars are provided by compilations of HR diagrams of open clusters in age groups \citep{Harris1976EvolvedClusters.}, of which the youngest few are reproduced in Figure~\ref{fig:fig6}. The general trends with age are clear, but there are obviously many uncertainties due to the small numbers of supergiants, and a scatter that may be due partly to the presence of clusters with different metallicities.
\subsubsection{Stars of intermediate mass}
\label{subsubsec:Intermediate mass}
There may be some stars of mid-B type on the MS, with masses in the interval ${\sim} 4 - 8 \: \rm M_{\odot}$, that neither die quietly as white dwarfs nor follow the sequence of nuclear burning stages through core collapse and (Type~II) supernova explosion. Such stars would be characterized by carbon ignition in a degenerate core, and it is possible that they do not exist if the upper limit for white dwarf formation exceeds the lower limit for non-degeneracy. If they do exist, their final evolution is very uncertain, since it is not known whether degenerate carbon ignition would blow a star apart or lead quietly to further burning; this problem is reviewed by \citet{Wheeler1978a}.
Regardless of their final fate, stars of about $4 - 8 \: \rm M_{\odot}$\footnote{Values of stellar masses will from now on be understood to be initial values, unless otherwise stated.} make important contributions to galactic enrichment in elements that are produced during hydrogen and helium burning and are mixed up to the surface. Overabundances of carbon, nitrogen, barium and other elements are seen in the spectra of many giants in this mass range, and their theoretical origin is discussed by \citet{IbenI.1978OnMedium}.
The evolution of stars between 4 and $8 \: \rm M_{\odot}$ in the HR diagram is subject to the same uncertainties that beset more and less massive stars. The groups of clusters in \citet{Harris1976EvolvedClusters.} again give a useful guide, but they are rather sparse, as seen in Figure~\ref{fig:fig6}.
\medskip
About half of all supernovae in spiral galaxies are Type~I supernovae (SNI), which, because of their sites in galaxies are associated with stars below ${\sim} 6 \: \rm M_{\odot}$. Their origin has been controversial for some time, as reviewed by \citet{Tammann1974StatisticsSupernovae, Tammann1977AStatistics}, \citet{Tinsley1975b, Tinsley1977a}, and \citet{Wheeler1978a}. This leaves a serious gap in our understanding of stellar deaths and nucleosynthesis, especially since SNI appear to eject significant quantities of iron \citep{Kirshner1975Supernova5253}. Alternative theories include the following.
\begin{enumerate}
\item SNI are exploding white dwarfs. This has long been the standard theory, because SNI occur in elliptical galaxies where only stars ${\sim} 1 \: \rm M_{\odot}$ are conventionally believed to be dying. However, the white dwarf theory conflicts with the large envelopes (of supergiant size and $\gtrsim 1 \: \rm M_{\odot}$) that are invoked to explain the early spectra and light curves of SNI, and with the clear association of SNI with a young stellar population in spiral and irregular galaxies \citep{Oemler1979TypeStars}.
\vspace{1mm}
\item SNI are stars ${\sim} 2 \: \rm M_{\odot}$ that live for ten times longer than normal by mixing completely on the MS. \citet{Wheeler1978b} proposed this model to be consistent with the envelope masses, populations in elliptical galaxies, the extreme deficiency of hydrogen in SNI spectra, and some properties of the Crab nebula. Its problems are that occasional complete mixing of MS stars must be postulated ad hoc, and that often SNI are associated with a much younger stellar population.
\vspace{1mm}
\item SNI are stars of intermediate mass, ${\sim} 4 - 6 \: \rm M_{\odot}$. This was suggested by \citet{Renzini1976OnSupergiants} on the grounds that these stars could blow away their hydrogen-rich envelopes before dying, and by \citet{Oemler1979TypeStars} to account for the association of SNI with ongoing star formation in spiral and irregular galaxies. This hypothesis implies a rate of star formation in elliptical galaxies that should be marginally detectable.
\vspace{1mm}
\item There are two kinds of SNI, those from low-mass stars that appear in elliptical galaxies, and those from more massive stars that account for the frequency in spiral and irregular galaxies \citep{Dallaporta1973OnSupernovae}. While this hypothesis avoids the problems of the preceding ones, the properties of SNI (while not exactly homogeneous) do not divide naturally into two classes suggestive of very different stellar origins.
\end{enumerate}
In any case, the production of iron by SNI and the red-giant nucleosynthesis mentioned above show that stars with a very wide range of masses contribute to galactic enrichment in various elements.
\subsubsection{Effects of initial composition}
\label{subsubsec:Initial composition}
Stars now forming and evolving in different regions of galaxies have at least a ten-fold range of metallicities, and the first stars everywhere were presumably very metal-poor. It is therefore important to ask how stellar evolution is affected by initial composition.
\medskip
The evolution of stars from ${\sim} 0.7 - 1 \: \rm M_{\odot}$ has been widely studied as a function of metallicity ($Z$) and helium abundance ($Y$), for the interpretation of HR diagrams of globular clusters \citep[e.g.][]{Iben1974PostStars, Mengel1979StellarSequence, Sweigart1978EvolutionaryStars}. Much less attention has been paid to more massive stars. In the largest systematic survey to date, stars up to $7 \: \rm M_{\odot}$ with a wide range of $Z$ and $Y$ have been followed to the base of the giant branch \citep{Mengel1979StellarSequence}; \citet{Sweigart1978EvolutionaryStars} have followed some of these up the giant branch. A few more massive models have been studied by \citet{Trimble1973EvolutionStars} and \citet{Harris1976EffectsClouds}, and \citet{Alcock1978TheStars} have followed the evolution of stars up to $10 \: \rm M_{\odot}$, with various $Z$, through core helium exhaustion. The results of these calculations suggest that important systematic differences in nucleosynthesis may occur as a result of initial $Z$ differences, but they do not go far enough to provide quantitative estimates. Models for chemical evolution usually assume that the production of primary elements -- those that are made from hydrogen and helium that were initially in the star -- is a function of stellar mass alone, but this assumption may be leading to systematic errors. (An exception is the study undertaken by \citealp{Arnett1971GalacticNucleosynthesis} of explosive nucleosynthesis of certain neutron-rich primary elements which may depend systematically on the star's initial $Z$; \citealp{Arnett1971GalacticNucleosynthesis} showed how such effects could be tested using relative abundances of metals in metal-poor stars).
\medskip
Some elements are believed to be synthesized in amounts that depend \emph{directly} on the initial composition of the star, because they are made not only from hydrogen and helium but also from heavier elements that were present initially. These are called secondary elements, and some examples are:
\begin{enumerate}
\item $^{13}$C and $^{14}$N are made in stellar envelopes as by-products of hydrogen burning in the CNO cycle, mainly from $^{12}$C that was present initially\footnote{If any fresh $^{12}$C, a product of helium-burning in the star's own core, gets mixed to the hydrogen-burning region and processed in this way, the resulting $^{13}$C and $^{14}$N would be called primary products.}.
\vspace{1mm}
\item Barium is made from the addition of neutrons to iron nuclei, probably in the intershell region of asymptotic-branch giants \citep[e.g.][]{IbenI.1978OnMedium}.
\end{enumerate}
Because the structure of stars at different stages of evolution depends on their initial composition, these and other elements may also be affected \emph{indirectly} by initial composition.
Extensive calculations of stellar models lie ahead before the dependence of both primary and secondary nucleosynthesis on composition is adequately understood. Almost all calculations to date have used Solar relative abundances of the heavy elements (all scaled with the single initial metallicity parameter $Z$), and it will be important also to find out how stellar evolution and nucleosynthesis depend on variations among these relative abundances.
Another set of questions concerns changes in the HR diagram that would affect the colors of galaxies with stars of different compositions. A population of low $Z$ is in general bluer than normal for two reasons: the effective temperature of a star of a given mass and evolutionary stage is (in most cases of interest) higher at lower $Z$, because metal-poor stars tend to have smaller radii; and most colors at a given effective temperature are bluer at lower $Z$, because the reduced line-blanketing restores more flux at short wavelengths than at long wavelengths. These effects are partially offset when one considers populations of a given age, because the turnoff mass is lower at lower $Z$, but this factor is relatively unimportant. Some implications of these changes will be discussed in Section~\ref{subsec:Data2} and Section~\ref{subsubsec:Variations in metallicity}.
\section{Aims and Methods of Chemical Evolution}
\label{sec:Aims and Methods 1}
Studies of chemical evolution aim to account for abundance distributions of the elements, including the variation of stellar metallicities with age and position in the Galaxy, abundance gradients in galaxies, variations in relative abundances of elements heavier than helium, and related observations. The main processes governing chemical evolution are indicated in Figure~\ref{fig:fig1}: star formation, nucleosynthesis, mass loss from evolving and dying stars, and gas flows. Because of the importance of dynamical processes, another motivation for studying chemical evolution is that it leads to clues to the dynamical evolution of galaxies, as examples in Sections~\ref{sec:Evolution solar neighborhood}~and~\ref{sec:Chemical Evolution of Galaxies} will show. Details of the basic processes have been extensively reviewed elsewhere, so the emphasis here will be on methods of modeling chemical evolution, and selected applications. Reviews for further background material include \citet{Audouze1976ChemicalGalaxies.}, \citet{Faber1977ThePopulations}, \citet{Pagel1978a, Pagel1978b}, \citet{Trimble1975TheElements}, \citet{vandenBergh1975StellarGalaxies}, and others cited in Section~\ref{sec:Star_formation}. There is no clearer introduction to chemical evolution than the seminal paper by \citet{Schmidt1963TheMass.} on the subject.
\subsection{Basic Assumptions and Equations}
\label{subsec:Assumptions}
Models for chemical evolution follow, analytically or numerically, abundance changes in the ISM of a region and the resulting abundance distributions in stars. For most purposes, the region under study in a particular model can be assumed to have ISM of uniform composition and to gain or lose mass through gas flows only; complicated models for whole galaxies are generally divided up, for computation, into zones with these properties. The basic equations are then easy to derive.
The total mass $M$ of the system changes according to the net inflow rate $f$ of gas:
\begin{equation}
\frac{dM}{dt} = f.
\label{eq:eq3.1}
\end{equation}
($f$ is often called the infall or accretion rate). The mass of stars $M_{\rm s}$ changes via star formation and mass loss from evolving stars:
\begin{equation}
\frac{dM_{\rm s}}{dt} = \psi - E,
\label{eq:eq3.2}
\end{equation}
where $\psi$ is the SFR and $E$ is the total ejection rate from stars of all masses and ages. Similarly, the mass of ISM, called ``gas'' ($M_{\rm g}$), changes through star formation, ejection, and net inflow into the region:
\begin{equation}
\frac{dM_{\rm g}}{dt} = -\psi + E + f,
\label{eq:eq3.3}
\end{equation}
Since $M = M_{\rm s} + M_{\rm g}$, the sum of Equations~(\ref{eq:eq3.2})~and~(\ref{eq:eq3.3}) is simply (\ref{eq:eq3.1}).
The fraction of gas in the system is written
\begin{equation}
\mu \equiv \frac{M_{\rm g}}{M},
\label{eq:eq3.4}
\end{equation}
so the mass of stars is obviously
\begin{equation}
M_{\rm s} = \left( 1 - \mu \right) M.
\label{eq:eq3.5}
\end{equation}
The ejection rate can be written in terms of the IMF and SFR, Equation~(\ref{eq:eq2.1}), if one makes the usual approximation that each star undergoes its entire mass loss after a well-defined lifetime. Let $w_{\rm m}$ be the remnant mass and $\tau_{\rm m}$ the lifetime of a star of (initial) mass $m$; the star was therefore formed at time ($t - \tau_{\rm m}$) if it dies at time $t$. Then the ejection rate at time $t$ is
\begin{equation}
E(t) = \int_{m_{\rm t}}^{\infty} \left( m - w_{\rm m} \right) \psi \left( t - \tau_{\rm m} \right) \phi(m) \ dm,
\label{eq:eq3.6}
\end{equation}
where $m_{\rm t}$ is the mass with $\tau_{\rm m} = t$ (the turnoff mass) and $\phi(m)$ is the IMF at time ($t - \tau_{\rm m}$). The approximation of sudden mass loss at the end of each star's lifetime is usually a reasonable one, since for most stars nearly all of the mass loss is thought to occur in a final small fraction of their lives; mass loss on the MS is important for O stars, however, so the approximation fails if timescales $< 10^{7} \: \rm yr$ are of interest.
\medskip
For illustration, let us consider the evolution of a single ``metal'' abundance parameter $Z$, where ``metals'' are one or all of the common primary elements (e.g., C, O, Fe), and let us assume that their production is a function only of stellar mass, not of initial composition. (The methods for calculating this schematic $Z$ are readily adapted to more realistic cases). Let us also assume for simplicity that material ejected from stars is mixed instantly throughout the ISM in the region under study. Then the mass of metals in the gas, which is $ZM_{\rm g}$, evolves via star formation (putting metals from the ISM into stars), ejection, and gas flows, according to the equation
\begin{equation}
\frac{dZM_{\rm g}}{dt} = -Z \psi + E_{Z} + Z_{\rm f} f,
\label{eq:eq3.7}
\end{equation}
where $E_{Z}$ is the total ejection rate of metals from stars and $Z_{\rm f}$ is the mean metal abundance in infalling gas. $E_{Z}$ includes both newly synthesized metals and those that were in the star from birth and re-ejected. Let $p_{\rm zm}$ be the mass fraction of a star of mass $m$ that is converted to metals and ejected; then the rate of ejection of new metals from stars at time $t$ is
\begin{equation}
\int_{m_{\rm t}}^{\infty} m p_{\rm zm} \psi \left( t - \tau_{\rm m} \right) \phi(m) \ dm.
\label{eq:eq3.8}
\end{equation}
Unprocessed material occupies a mass ($m - w_{\rm m} - m p_{\rm zm}$) of the ejected part of a star of mass $m$, and the metal abundance in this region is $Z(t - \tau_{\rm m})$. Thus the total ejection rate of old and new metals is obtained by adding the ejection rate from all masses of these unprocessed metals to the rate (Equation~\ref{eq:eq3.8}):
\begin{equation}
\begin{medsize}
E_{Z}(t) = \int\limits_{m_{\rm t}}^{\infty} \left[ \left( m - w_{\rm m} - m p_{\rm zm} \right) Z \left( t - \tau_{\rm m} \right) + m p_{\rm zm} \right] \psi \left( t - \tau_{\rm m} \right) \phi(m) \ dm.
\label{eq:eq3.9}
\end{medsize}
\end{equation}
The mean metallicity of stars ever formed, denoted $Z_{\rm s}$, can be obtained by writing an equation for the conservation of metals: a mass $Z_{\rm s} M_{\rm s}$ of metals is stored in stars, a mass $Z M_{\rm g}$ is in the gas, and their sum is the mass of new metals ever ejected, which is obtained by integrating Equation~(\ref{eq:eq3.8}) over time. Thus $Z_{\rm s}$ is given by
\begin{equation}
Z_{\rm s} M_{\rm s} + Z M_{\rm g} = \int_{0}^{t} \int_{m_{\rm t^{\prime}}}^{\infty} m p_{\rm zm} \psi \left( t^{\prime} - \tau_{\rm m} \right) \phi(m) \ dt^{\prime} dm.
\label{eq:eq3.10}
\end{equation}
\medskip
It is convenient to define some integrals that depend on the IMF and parameters of stellar evolution but not on $\psi(t)$. The \emph{returned fraction} $R$ is defined by the integral
\begin{equation}
R \equiv \int_{m_{1}}^{\infty} \left( m - w_{\rm m} \right) \phi(m) \ dm,
\label{eq:eq3.11}
\end{equation}
where $m_{1}$ is the present turnoff mass, and is usually taken to be $1 \: \rm M_{\odot}$ in evaluating $R$. Because the IMF is normalized as in Equation~(\ref{eq:eq2.2}), $R$ is the fraction of mass put into stars at a given time that is thereafter returned to the ISM in the lifetime of a solar-mass star; for example, if the stars in a galaxy were all formed in a single initial burst with a total mass $M$, then ${\sim} 10^{10} \: \rm yr$ later the mass in stellar form would be $M(1 - R)$. An estimate of $R$ can be obtained using the local IMF as approximated in Equation~(\ref{eq:eq2.9}), with a white dwarf mass of $0.7 \: \rm M_{\odot}$ as the remnant mass $m_{\rm w}$ for $m \leq 4 \: \rm M_{\odot}$ and a neutron star mass of $1.4 \: \rm M_{\odot}$ for $m > 4 \: \rm M_{\odot}$; the result is, $R = 0.17 \times (10 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1} / \psi_{1})$.
A second useful integral is the \emph{yield} $y$, which is the mass of new metals ejected (eventually) when unit mass of matter is locked into stars. The yield can be defined for any element of interest, and for the ``metals'' considered above it is given by
\begin{equation}
y \equiv \frac{1}{1 - R} \int_{m_{1}}^{\infty} m p_{\rm zm} \phi(m) \ dm.
\label{eq:eq3.12}
\end{equation}
Again, if a mass $M$ of stars were formed in a single initial burst, then ${\sim} 10^{10} \: \rm yr$ later, the mass of new metals ejected would be $y(1 - R)$. Sometimes the quantities $p_{\rm zm}$ are called stellar yields, and $y$ is called the net yield.
Although $R$ and $y$ are defined with a specific lower mass limit, their values do not depend strongly on $m_{1}$, so they turn out to be useful parameters in many situations. For example, $R$ has been used in Section~\ref{subsubsec:Local SFR} to relate the mass of stars now in the Solar neighborhood to the mass ever formed, giving a reasonable estimate even though star formation did not occur in a single burst; and the yield will prove to be a first approximation to the abundance of metals resulting from a given IMF and stellar production parameters (e.g., $p_{\rm zm}$), independently of the detailed history of the system.
Many insights into chemical evolution can be obtained using analytical approximations to the above equations, while for detailed modeling they can be cast into forms suitable for numerical computation. These techniques will be outlined in turn.
\subsection{Analytical Approximations}
\label{subsec:Approximations}
The equations of chemical evolution are especially easy to handle if one makes an approximation known as \emph{instantaneous recycling}: stars are divided into two classes, those that live forever (masses less than some value $m_{1}$) and those that die as soon as they are born ($m > m_{1}$). Thus the lower mass limits in the integrals in Equations~(\ref{eq:eq3.6})~--~(\ref{eq:eq3.10}) are taken to be the time-independent $m_{1}$, and the assumed instant deaths allow the arguments ($t - \tau_{\rm m}$) to be replaced by simply ($t$). (Note that the assumption now made of death immediately after birth, so $\tau_{\rm m} \equiv 0$, is a much stronger and less realistic assumption than the earlier one of sudden mass loss, which simply allowed $\tau_{\rm m}$ to be well-defined for each star). Equations~(\ref{eq:eq3.6})~and~(\ref{eq:eq3.9}) reduce at once to the much simpler relations,
\begin{equation}
E(t) = R \psi(t),
\label{eq:eq3.13}
\end{equation}
\begin{equation}
E_{Z}(t) = R Z(t) \psi(t) + y(1 - R) \left[ 1 - Z(t) \right] \psi(t),
\label{eq:eq3.14}
\end{equation}
in which all time-dependent quantities are evaluated at the current time $t$ only. Since $Z \ll 1$ in all situations of interest, Equation~(\ref{eq:eq3.14}) can be written more simply
\begin{equation}
E_{Z}(t) = R Z(t) \psi(t) + y(1 - R) \psi(t).
\label{eq:eq3.15}
\end{equation}
Substituting in Equations~(\ref{eq:eq3.2}),~(\ref{eq:eq3.3}),~and~(\ref{eq:eq3.7}), we have now
\begin{equation}
\frac{dM_{\rm s}}{dt} = (1 - R) \psi,
\label{eq:eq3.16}
\end{equation}
\begin{equation}
\frac{dM_{\rm g}}{dt} = -(1 - R) \psi + f,
\label{eq:eq3.17}
\end{equation}
\begin{equation}
\frac{dZM_{\rm g}}{dt} = -Z(1 - R) \psi + y(1 - R) \psi + Z_{\rm f} f.
\label{eq:eq3.18}
\end{equation}
An equation for the metal abundance $Z$ of the gas, rather than the mass of metals it contains, is obtained by substituting Equation~(\ref{eq:eq3.17}) into (\ref{eq:eq3.18}), using the identity $M_{\rm g} dZ / dt = d(ZM_{\rm g}) / dt - Z dM_{\rm g} / dt$; the result is
\begin{equation}
M_{\rm g} \frac{dZ}{dt} = y(1 - R) \psi + \left( Z_{\rm f} - Z \right) f.
\label{eq:eq3.19}
\end{equation}
For simplicity, let us assume that the IMF is constant. (The methods developed for a constant IMF can be adapted when necessary to more complicated cases). The quantities $R$ and $y$ are therefore constants.
Equation~(\ref{eq:eq3.16}) now gives an expression for the mass of stars at time $t$,
\begin{equation}
M_{\rm s} (t) = (1 - R) \int_{0}^{t} \psi \left( t^{\prime} \right) \ dt^{\prime} \equiv (1 - R) \overline{\psi} t.
\label{eq:eq3.20}
\end{equation}
The mean metallicity of stars can be obtained from Equation~(\ref{eq:eq3.10}), using the instantaneous recycling approximation and the assumption of a constant IMF: the integrals with respect to time and mass separate, and are simply $M_{\rm s} / (1 - R)$ and $y(1 - R)$ respectively, so the result is $Z_{\rm s} M_{\rm s} + Z M_{\rm g} = y M_{\rm s}$, from which
\begin{equation}
Z_{\rm s} = y - \frac{\mu}{1 - \mu} Z.
\label{eq:eq3.21}
\end{equation}
This result shows that $Z \rightarrow y$ as $\mu \rightarrow 0$, which simply states that, when there is no gas, all the metals ever made and ejected ($y M_{\rm s}$) are incorporated into later generations of stars. An estimate of the yield for the local IMF is therefore the mean metallicity of stars in the Solar neighborhood (where the gas fraction is small, $\mu \sim 0.05$), i.e., $y \simeq 0.8 \: \rm Z_{\odot}$. The approximation $y \sim \: \rm Z_{\odot}$ will be useful below.
Further solutions to the above equations depend on the assumptions made about gas flows, so two extreme cases will be considered for illustration.
\subsubsection{A closed system, initially unenriched gas}
\label{subsubsec:closed system}
In this case, we set $f = 0$, and the initial values are $M_{\rm g0} = M$, $M_{\rm s0} = 0$, $Z_{0} = Z_{\rm s0} = 0$. Obviously, $M = \rm constant$. Time can be eliminated as an explicit variable if Equation~(\ref{eq:eq3.19}) is divided by (\ref{eq:eq3.17}), with the result
\begin{equation}
M_{\rm g} \frac{dZ}{dM_{\rm g}} = -y,
\label{eq:eq3.22}
\end{equation}
i.e.,
\begin{equation}
Z = y \ln \left( \frac{M_{\rm g0}}{M_{\rm g}} \right) = y \ln \left( \frac{M}{M_{\rm g}} \right) = y \ln \left( \frac{1}{\mu} \right).
\label{eq:eq3.23}
\end{equation}
This solution is valid as long as $Z \ll 1$. If not, Equation~(\ref{eq:eq3.14}) must be used instead of (\ref{eq:eq3.15}); the result is then $Z = 1 - \mu^{y}$, which reduces to (\ref{eq:eq3.23}) if $Z \ll 1$, and has the limit $Z \rightarrow 1$ as $\mu \rightarrow 0$. (This formally correct limit is of course absurd, and is a consequence of neglecting the effects of metallicity on stellar nucleosynthesis). Original derivations of these results were given by \citet{Talbot1971TheModel}, and Equation~(\ref{eq:eq3.23}) was derived independently by \citet{Searle1972InferencesGalaxies}. Through Equations~(\ref{eq:eq3.21})~and~(\ref{eq:eq3.23}), $Z_{\rm s}$ can also be expressed as a function of $\mu$.
\subsubsection{A system with infall balanced by star formation}
\label{subsubsec:balanced system}
Here it is assumed that star formation just keeps up with the rate of infall plus stellar gas loss:
\begin{equation}
\psi = f + R \psi,
\label{eq:eq3.24}
\end{equation}
so that, from Equation~(\ref{eq:eq3.17}),
\begin{equation}
M_{\rm g} = {\rm constant} = M_{0}.
\label{eq:eq3.25}
\end{equation}
The picture is that the total mass grows by infall, while star formation maintains a constant gas mass. Other initial conditions are $M_{\rm s0} = 0$ and $Z_{0} = Z_{\rm s0} = 0$, and the infalling gas will be assumed to be unenriched so that $Z_{\rm f} = 0$. To eliminate time in this case, we divide Equation~(\ref{eq:eq3.17}) by (\ref{eq:eq3.1}), with the result
\begin{equation}
M_{\rm g} \frac{dZ}{dM} = y - Z.
\label{eq:eq3.26}
\end{equation}
It is useful to solve this Equation in terms of a parameter $\nu$ that gives the ratio of mass accreted to the initial mass:
\begin{equation}
\nu \equiv \frac{M - M_{0}}{M_{0}} = \frac{M - M_{\rm g}}{M_{\rm g}} = \frac{1}{\mu} - 1.
\label{eq:eq3.27}
\end{equation}
Then Equation~(\ref{eq:eq3.26}) has the solution
\begin{equation}
Z = y \left(1 - e^{-\nu} \right),
\label{eq:eq3.28}
\end{equation}
a result due originally to \citet{Larson1972b}. Equation~(\ref{eq:eq3.21}) again gives $Z_{\rm s}$. An important difference between this infall model and the preceding closed model is that here $Z \rightarrow y$ as $\mu \rightarrow 0$; an equilibrium is set up between the rate of infall of metal-free gas and the rate of enrichment by evolving stars.
\subsubsection{Generalities}
\label{subsubsec:generalities}
Although the above two models are extremely schematic, their Equations~(\ref{eq:eq3.23})~and~(\ref{eq:eq3.28}) for the ``metal'' abundance of the gas have some common properties that are found in a wide variety of models for chemical evolution. The following generalizations are approximately true even when instantaneous recycling is a poor approximation, but they break down if the IMF is time-dependent.
\begin{enumerate}
\item \emph{$Z$ is proportional to the net yield, $y$.} Thus any primary elements whose stellar production parameters are independent of composition will have abundances in proportion to their respective yields. The yields in turn can be calculated from Equation~(\ref{eq:eq3.12}), for a given IMF, independently of any model for star formation or gas flows in the system. The utility of this result is that theories of nucleosynthesis can be tested by comparing predicted relative yields of various elements with their observed relative abundances.
\vspace{1mm}
\item \emph{$Z / y$ depends chiefly on current properties of the system}, and is insensitive to its past history. This statement is exactly true in the above models because of the assumption of instantaneous recycling (which led to the cancellation of time as an explicit variable). In numerical models that allow for finite stellar lifetimes, the same result holds approximately, and relevant current properties include the gas fraction ($\mu$) and the ratio of SFR to gas flow rate ($\psi / f$).
\vspace{1mm}
\item \emph{$Z / y$ depends rather weakly on model-dependent quantities} such as $\mu$ and $\psi / f$. Consequently, regions of galaxies with gas fractions differing by orders of magnitude may have interstellar abundances differing by only small factors. This prediction allows one to test whether theories of nucleosynthesis give the right absolute amounts of elements, as well as the right relative amounts; in particular, the yield of an element, calculated for the local IMF, should be in order of magnitude equal to its abundance in the Solar System or nearby stars.
\end{enumerate}
\subsection{Numerical Models}
\label{subsec:Numerical Models}
Chemical evolution is best studied numerically in cases where approximations that make the analytical approach transparent (or possible) break down. Situations in which instantaneous recycling is no longer a useful approximation include the following:
\begin{enumerate}
\item If the SFR is a strongly decreasing function of time, systematic errors result from setting $\psi (t - \tau_{\rm m}) = \psi(t)$; in particular, low-mass stars formed early die later in much greater numbers (relative to massive stars) than would be predicted on the basis of the current SFR.
\vspace{1mm}
\item If one is interested in time-dependent abundance ratios arising from nucleosynthesis in stars of different lifetimes, the effects would be entirely lost by neglecting finite stellar lifetimes. Even if instantaneous recycling can be assumed, the analytical approach becomes intractable in models with time-dependent IMFs, or in considerations of radioactive elements with lifetimes of billions of years, except in very special cases.
\end{enumerate}
For numerical computation, equations such as (\ref{eq:eq3.1})~--~(\ref{eq:eq3.10}) are expressed as differences in finite steps of time and sums over a grid of stellar masses. The programming required is little more than careful book-keeping, so the effort in models for chemical evolution goes not into techniques but into the astrophysical input. Many examples of numerical models will be found in the references cited in Sections~\ref{sec:Evolution solar neighborhood}~and~\ref{sec:Chemical Evolution of Galaxies}, in the discussion of particular problems.
\section{Chemical Evolution in the Solar Neighborhood}
\label{sec:Evolution solar neighborhood}
The extensive data available on chemical compositions of nearby stars and \textsc{H~ii} regions, and details of the isotopic composition of Solar System material, have led to many efforts to build consistent models for chemical evolution in the Solar neighborhood. In trying to explain the abundance distributions, these models also shed light on dynamical processes that have affected the local region of the Galaxy, and they set some constraints on theories of nucleosynthesis.
The ``neighborhood'' must be carefully defined to include the range of environments occupied by nearby stars and their predecessors in their motions around the Galaxy, and to include stars that were formed near the plane but now occupy a thicker disk. A small ``local swimming hole'' around the Sun (in Baade's words) will not do. Instead, the Sun's ``nucleogenetic pool'' is usefully defined as a region about a kpc wide in the Galactic plane (to include most stellar epicyclic motions and periodic gas motions), extended around the Galaxy at the Sun's galactocentric distance (to include passages of local material through spiral arms and related small-scale patchiness), and extending about a kpc out of the plane (to include stars whose periodic motions carry them far from the plane). This region is shaped like a cylindrical shell, and quantities like the mass of stars, SFR, etc. are expressed in units per square parsec, referring to \emph{an average column at the Solar galactocentric distance}. Spatial averages over this pool can be interpreted as short-term time averages during the history of a given small amount of material, and models generally consider explicitly only timescales longer than passages between spiral arms, etc. (Some special studies, such as the interpretation of elements with radioactive lifetimes ${\sim} 10^{6} - 10^{9} \: \rm yr$, do need to consider short-term changes). Although the region has been defined to include most relevant periodic motions, it cannot be treated as a closed system, because it is probably affected by systematic radial gas flows in the disk and by infall from outside, especially at early times.
Halo stars are sometimes treated as part of the ``neighborhood'' (in which case its extent out of the plane must be many kpc), but more often the halo is treated as a separate subsystem of the Galaxy, whose effects on the evolution of the disk are perturbations from outside. Properties of halo stars are in any case relevant to the Solar neighborhood, so they will be discussed along with the properties of disk stars.
\subsection{Outline of Relevant Data}
\label{subsec:Data1}
Chemical abundances in the Solar neighborhood and in the Solar System have been reviewed in detail by \citet{Trimble1975TheElements} and several authors at a recent conference at Yale University. An outline of the points most relevant to this article will be given here. Because of systematic radial gradients in chemical abundances in the Galaxy (Section~\ref{subsec:Data2}), the following remarks apply only to stars within a few kpc of the Sun's galactocentric distance.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig7.png}
\caption{The relation between age and metallicity for disk stars in the Solar neighborhood \citep{Twarog1980AnNeighborhood}. Boxes denote the range of ages referred to by each point, and the $\pm 1 \sigma$ spread in $\rm [Fe / H]$ for the age group. Error bars attached to the points denote the standard error in the mean. Note that the mean metallicity increases by a factor of less than four during the lifetime of the disk.}
\label{fig:fig7}
\end{figure}
Stars with halo kinematics are all metal-poor, with $\rm [Fe / H] \lesssim -1$, whereas stars with disk motions nearly all have $\rm [Fe / H] \gtrsim -0.5$. (As usual, abundance ratios in square brackets are logarithms of the ratios relative to their Solar values). Since most halo stars are probably older than most disk stars (Section~\ref{subsec:Stellar_Populations}), their composition differences indicate a systematic increase of metallicity with time of formation. The disk stars alone show a systematic increase of metallicity with time of formation, illustrated in Figure~\ref{fig:fig7}, but the trend is only about a factor of three over the whole lifetime of the disk. Open clusters show a very weak dependence of metallicity on age, if any, because they are found at such a wide range of galactocentric distances that the radial gradient in metallicity obscures the time-dependence.
A powerful tool for chemical evolution is the metallicity distribution of all stars ever formed, which is given by samples of G--M dwarfs since these stars have remained on the MS since they formed. The distribution can be roughly approximated by a log-normal function, with a mean $\langle \left[ \rm Fe / H \right] \rangle = -0.1$ and standard deviation $\sigma \left( \left[ \rm Fe / H \right] \right) = 0.2$, allowing for observational errors (\citealp{Pagel1975MetalNeighbourhood, Pagel1978a}; and references therein). Its striking characteristic is the extreme paucity of metal-poor stars: only a few percent of stars in the local cylinder (including halo stars) have $\rm [Fe / H] < -1$. The lack of metal-poor disk stars proves to be one of the most powerful constraints on models for the Solar neighborhood, as discussed below.
Stellar ``metallicities'' are commonly denoted $\rm [Fe / H]$ because most photometric indices of the abundance of heavy elements are sensitive to blanketing by numerous iron lines. Detailed abundance analyses show that other elements heavier than helium usually occur in fixed proportions, so a single metallicity index is often adequate. However, the exceptions to this uniformity are often of great interest. Certain abundance variations in \emph{evolved} stars are believed to show products of the stars' own nucleosynthesis that are mixed to the surface and lost in a stellar wind (Section~\ref{subsec:Beyond MS}) -- direct observational evidence for enrichment of the ISM. Some elements also show relative abundance variations in \emph{unevolved} stars, and in these cases the anomalies must have been present in the ISM from which the star formed.
\medskip
Besides these interesting abundance variations, there are others that should be ignored or ``corrected'' for in the context of chemical evolution. Extraneous effects include depletion of metals in the diffuse interstellar gas due to their incorporation into grains; anomalous interstellar abundance ratios, especially between isotopes of the same element, caused by chemical fractionation\footnote{This is a case of real chemistry, not in the special sense of ``chemical'' evolution of galaxies. Other terminology has been suggested to avoid the term ``chemical evolution'' now that real chemistry of the ISM is becoming an important topic in astronomy, but the alternative suggestions have their own problems. For example, ``nuclear evolution'' can be confused with the evolution of the nuclei (centers) of galaxies, and it obscures the great importance of gas flows, in addition to nucleosynthesis, in determining the evolution of chemical abundances in galaxies.}; and large abundance variations of various elements on the surfaces of Ap and related stars, produced by differential diffusion of elements in the stellar atmospheres \citep{Trimble1975TheElements}. Unfortunately, it is not always clear whether a particular composition anomaly is due to some such extraneous effect, to nucleosynthesis in the star observed, or to significant composition changes in the ISM.
\subsection{The ``G-Dwarf Problem''}
\label{subsec:G-Dwarf}
The metallicity distribution of long-lived disk dwarfs leads to a problem originally discovered by \citet{vandenBergh1962TheAbundances.} and \citet{Schmidt1963TheMass.}: there are far fewer metal-poor stars than one predicts using the most straightforward assumptions. This contradiction is called the ``G-dwarf problem'', although a stellar metallicity distribution has now been confirmed for M dwarfs \citep{Mould1978InfraredDwarfs}. The straightforward assumptions involved are the following:
\begin{enumerate}
\item the Solar neighborhood can be modeled as a closed system;
\vspace{1mm}
\item it started as $100 \%$ metal-free gas;
\vspace{1mm}
\item the IMF is constant; and
\vspace{1mm}
\item the gas is chemically homogenous at any time.
\end{enumerate}
Any model based on these assumptions is called a \emph{``simple'' model}. It is now recognized that assumptions (i) and (ii) are almost certainly false and that more realistic alternatives lead to a natural solution of the G-dwarf problem; assumptions (iii) and (iv) are probably also unrealistic, but there is less evidence for believing that they are violated sufficiently to solve the problem entirely. Although the G-dwarf problem is no longer regarded as a difficulty in chemical evolution, it will be reviewed here because of its earlier importance in motivating many models. Moreover, a review of this problem serves to emphasize that simple models cannot be used consistently in any arguments relating to chemical evolution in the Solar neighborhood -- the G-dwarf metallicities are so much at variance with such models that they must ignore very important effects, and could lead to highly misleading results in other contexts. This conclusion also leads to suspicion of simple models for other regions of galaxies.
\medskip
The problem itself can be derived analytically, following \citet{Pagel1975MetalNeighbourhood}, using the instantaneous recycling approximation and the results for a simple model in Section~\ref{subsubsec:closed system}. Let us consider the cumulative metallicity distribution of stars ever formed (which is represented observationally by dwarfs of mid-G type and later). From Equation~(\ref{eq:eq3.5}), the fraction of all stars that had been made while the gas fraction was $\geq \mu$ is
\begin{equation}
\frac{M_{\rm s}}{M_{\rm s1}} = \frac{1 - \mu}{1 - \mu_{1}},
\label{eq:eq4.1}
\end{equation}
where subscripts 1 denote present values. From Equation~(\ref{eq:eq3.23}), these stars were formed while the metallicity of the gas was $\leq y \ln \ \mu^{-1}$. The fraction of stars with metallicities $\leq Z$, denoted $S(Z)$, is therefore
\begin{equation}
S(Z) = \frac{M_{\rm s}}{M_{\rm s1}} = \frac{1 - \exp(-Z / y)}{1 - \mu_{1}}.
\label{eq:eq4.2}
\end{equation}
It is useful to eliminate the yield using the relation $y = Z_{1} / (\ln \ \mu_{1}^{-1})$ (again Equation~\ref{eq:eq3.23}), so Equation~(\ref{eq:eq4.2}) can finally be written in terms of the metallicity ratio $Z / Z_{1}$ and the present gas fraction $\mu_{1}$:
\begin{equation}
S(Z) = \frac{1 - \mu_{1}^{Z / Z_{1}}}{1 - \mu_{1}}.
\label{eq:eq4.3}
\end{equation}
This distribution is compared in Figure~\ref{fig:fig8} with a log-normal approximation to the data. (The model illustrated has $\mu_{1} = 0.05$, and the shape of the curve is insensitive to $\mu_{1}$ within the range of likely values). Obviously, there are far fewer metal-poor stars in the Solar neighborhood than predicted by this formula. The assumptions defining the simple model allow only two significant quantities to define a specific model: the IMF and the SFR. However, the IMF affects only the yield, which does not appear in Equation~(\ref{eq:eq4.3}) because metallicities have been scaled to the present value; the SFR was eliminated in the derivation of Equation~(\ref{eq:eq3.23}), and numerical models show that very nearly the same result is obtained if the approximation of instantaneous recycling is dropped. Thus the failure of the ``simple'' class of models is independent of the particular IMF or SFR used. It may seem at first sight that one could eliminate the problem by making the metals rapidly using a large SFR at early times; however, since there is initially unenriched gas and a constant IMF, a fixed quota of metal-poor G dwarfs must form along with the massive stars that produce metals -- no matter how rapidly it all happens.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig8.png}
\caption{Cumulative stellar metallicity distribution functions. $S(Z)$ is the fraction of stars that have metallicities $\leq Z$, with a maximum value $Z_{1}$. \emph{Solid line}: log-normal representation of the data for stars in the Solar neighborhood; $Z_{1}$ is taken as $2 \: \rm Z_{\odot}$ since more metal-rich stars are very rare; $\langle \log \ Z / \rm Z_{\odot} \rangle = -0.1$, and $\sigma( \log \ Z / \rm Z_{\odot}) = 0.15$, allowing for observational errors \citep[e.g.][]{Pagel1978a}. \emph{Long dashes}: the ``simple'' model for chemical evolution (Section~\ref{subsec:G-Dwarf}). \emph{Short dashes}: an extreme infall model (Section~\ref{subsubsec:infall}). \emph{Dots}: a model with a finite initial metallicity (Section~\ref{subsubsec:pre-enrichment}).}
\label{fig:fig8}
\end{figure}
Attempts to solve the G-dwarf problem have motivated numerous alternative models for the Solar neighborhood, in which one or more of the four straightforward assumptions are dropped. The following outline considers each assumption in turn, and some types of model will be developed from a different viewpoint in Section~\ref{subsec:Effects of Galaxy Formation}. Further details and references, on these and other models, can be found in reviews by \citet{Audouze1976ChemicalGalaxies.}, \citet{Pagel1975MetalNeighbourhood}, \citet{Tinsley1974ConstraintsNeighborhood} and \citet{Trimble1975TheElements}.
\subsubsection{Infall}
\label{subsubsec:infall}
It is unlikely that the region called the ``Solar neighborhood'' has evolved as a closed system. In particular, the disk could have grown gradually by accretion, on a timescale similar to that for star formation, as proposed by \citet{Larson1972a} and \citet{Sciama1972OnChromosphere}. The effect on the stellar metallicity distribution is then as follows \citep{Larson1972b}. Suppose the disk consists initially of unenriched gas, as in the simple case, but with a very small mass (surface density). The first stars formed are metal-poor, but their numbers are proportional only to the initial disk mass; star formation substantially enriches the gas while the disk is still far below its final mass. Then further star formation continually enriches the new gas that is accreted, so ultimately the only very metal-poor stars are the few initial ones.
The model in Section~\ref{subsubsec:balanced system} illustrates this behavior analytically. According to Equation~(\ref{eq:eq3.28}), $Z$ remains essentially constant at the yield value after the disk mass has increased to several times its initial value. The stellar metallicity distribution, $S(Z)$, can be derived analytically for this model \citep{Tinsley1975a}, and the result is a function strongly peaked at $Z = y$, as shown in Figure~\ref{fig:fig8}. Here we have a problem opposite to that of the simple model: there are too few stars in the model with $Z < Z_{1}$! The problem is alleviated if the width of the observed distribution is due more to errors than has been thought, but in any case this particular model is very schematic, showing effects of infall in a most extreme way, and more realistic models with infall (Section~\ref{subsec:Effects of Galaxy Formation}) predict metallicity distributions intermediate between the extreme and the simple cases.
\subsubsection{Pre-enrichment of the disk gas}
\label{subsubsec:pre-enrichment}
If the gas that formed the disk had a significant initial metal abundance, the number of more metal-poor disk stars would obviously be zero. This idea was introduced by \citet{Truran1971EvolutionaryGalaxy}, who postulated a pregalactic burst of massive stars. \citet{Ostriker1975GalacticHalos} considered the effects of massive stars in the young halo, whose ejecta enriched the gas that was condensing into a disk; the likely importance of this effect will be reviewed below (Section~\ref{subsubsec:Metals from the young halo}).
The effects of pre-enrichment can be seen analytically by simply adding a finite initial metal abundance $Z_{0}$ to the solution of Equation~(\ref{eq:eq3.22}). Then (\ref{eq:eq3.23}) becomes $Z = Z_{0} + y \ln \ \mu^{-1}$, and the exponent $Z / Z_{1}$ in (\ref{eq:eq4.3}) is replaced by $(Z - Z_{0}) / (Z_{1} - Z_{0})$. Figure~\ref{fig:fig8} shows the resulting stellar metallicity distribution in the case $Z_{0} = 0.2 Z_{1}$, and there is clearly no excess of metal-poor stars. A more realistic picture will be discussed in Section~\ref{subsec:Effects of Galaxy Formation}.
\subsubsection{Variable IMF}
\label{subsubsec:Variable IMF}
The original solution to the G-dwarf problem, proposed by \citet{Schmidt1963TheMass.}, was to postulate an IMF richer in massive stars at early times; the number of long-lived stars formed before the gas reaches a given metallicity is therefore reduced. (An analytical solution illustrating this effect is given by \citealp{Pagel1975MetalNeighbourhood}). Many alternative models with variable IMFs have been considered subsequently, but they all suffer from being ad hoc: although arguments can be offered for massive stars forming more easily at low metallicity, they are not compelling, and there is no independent way of verifying suggested changes in the IMF in the distant past.
\subsubsection{Metal-enhanced star formation}
\label{subsubsec:Metal-enhanced star formation}
\citet{TalbotRaymondJ.1973TheFormation} suggested that there are always large chemical inhomogeneities in the ISM, and that stars form preferentially in regions with above-average metallicity. \citet{TalbotRaymondJ.1974SensitivityElements} has developed this idea into a physical model, in which the enhanced cooling by heavy elements favors star formation in metal-rich regions. For this process alone to solve the G-dwarf problem, it would require both very large chemical inhomogeneities and very inefficient star formation in metal-poor regions; more plausibly, metal-enhanced star formation leads only to some reduction in the proportion of metal-poor stars.
\medskip
In summary, none of the four assumptions underlying the simple model is likely to be strictly true, and departures from any or all of them could alter the predicted metallicity distribution in the required sense of including fewer metal-poor stars.
\subsection{Effects of Galaxy Formation}
\label{subsec:Effects of Galaxy Formation}
In retrospect, the failure of the simple model is not surprising, because the Solar neighborhood was treated as a closed box, whereas gas flows during the formation of the Galaxy would affect chemical evolution strongly in two ways:
\begin{enumerate}
\item some pre-enrichment of the initial disk gas by massive stars in the young halo, and
\vspace{1mm}
\item later accretion of metal-poor gas.
\end{enumerate}
The ``formation'' of the Galaxy could in the latter sense be continuing today.
\subsubsection{Metals from the young halo}
\label{subsubsec:Metals from the young halo}
The amount of initial enrichment by massive halo stars can be estimated from the present mass of long-lived stars in the halo. Let $y_{\rm h}$ be the yield during the time of halo formation. Then, by definition of the yield, a mass $M_{\rm h}$ of surviving halo stars implies that a mass ${\sim} y_{\rm h} M_{\rm h}$ of metals was made and ejected at an early time; of these metals, a mass $Z_{\rm h} M_{\rm h}$ remains locked in halo stars (where $Z_{\rm h}$ is their mean metallicity), and the remaining $(y_{\rm h} - Z_{\rm h}) M_{\rm h}$ either fell to the disk or was lost from the Galaxy. \citet{Ostriker1975GalacticHalos} suggest on geometrical grounds that about half of those metals were retained by the disk. Now the halo yield, $y_{\rm h}$, may differ from the disk yield because the IMF for low-mass stars appears to differ between the two regions (Section~\ref{subsubsec:Other IMF}), but for a very rough estimate we can consider $y_{\rm h} \sim {\rm Z_{\odot}} \gg Z_{\rm h}$. The mass of metals falling on the disk is therefore ${\sim} 0.5 \ {\rm Z_{\odot}} M_{\rm h}$, and the initial metallicity of the disk gas is this quantity divided by the initial disk mass. The appropriate value of $M_{\rm h}$ is not the mass of halo stars in a local cylinder, but the mass (mostly further out) from which ejected gas would fall to this part of the disk; since the surface density of halo stars is locally only a few percent of the disk value, and it decreases with increasing radius \citep{Schmidt1975TheStars}, $M_{\rm h}$ is probably only a small fraction of the present mass of the disk in the local cylinder. However, the initial disk mass may also have been small. Altogether, it is possible that metals captured from the young halo could have been enough to have enriched the initial disk gas significantly. A value of $Z_{0} / \rm Z_{\odot} \sim 0.2$ would completely solve the G-dwarf problem, as illustrated above.
Early enrichment by halo stars appears as an important effect in the dynamical models for the formation of disk galaxies of \citet{Larson1976a}, whose chemical evolution has been studied further by \citet{Tinsley1978ChemicalDisks}. At model radii corresponding to the Sun's position in the Galaxy, $Z$ reaches about a third of its final value before stars with disk kinematics begin to form, and the initial disk mass in these models is about $10\%$ of the final value. As a result, the fraction of local stars with $Z < \rm Z_{\odot} / 3$ is no greater than observed.
It may be concluded that the disk in the Solar neighborhood plausibly started with a small fraction of its present mass, in the form of gas that had been significantly enriched by massive stars in the young halo.
\subsubsection{Later metal-poor infall}
\label{subsubsec:Later metal-poor infall}
In the dynamical models just mentioned, the outer disk grows on a timescale of billions of years by accretion of metal-poor gas; star formation occurs on a similar timescale, so the metallicity tends to reach a roughly constant value near the yield, as in the schematic infall model of Section~\ref{subsubsec:balanced system}.
A significant additional factor, first suggested by \citet{Lynden-Bell1975TheGalaxies}, is that the ratio of SFR to infall rate ($\psi / f$) is unlikely to remain strictly constant in a realistic picture; its value is more likely to increase as the surrounding gas is used up, as indeed occurs in the dynamical models. An asymptotic relation between $\psi / f$ and $Z$ can be derived using the Equations of Section~\ref{subsec:Approximations} with the instantaneous recycling approximation. Let
\begin{equation}
(1 - R) \psi = kf,
\label{eq:eq4.4}
\end{equation}
where $k$ is a constant, so Equation~(\ref{eq:eq3.19}) becomes
\begin{equation}
M_{\rm g} = \frac{dZ}{dt} = \left( ky + Z_{\rm f} - Z \right) f.
\label{eq:eq4.5}
\end{equation}
According to this Equation, there is an asymptotic value of $Z$,
\begin{equation}
Z \rightarrow ky + Z_{\rm f},
\label{eq:eq4.6}
\end{equation}
so the metal abundance of the gas tends toward a value $ky$ if $Z_{\rm f} = 0$. Thus if $\psi / f$ increases, so does $Z$, and there is steady enrichment rather than the constant $Z$ obtained in Section~\ref{subsubsec:balanced system}. Examples of this behavior are seen in the numerical models of \citet{Tinsley1978ChemicalDisks}, in particular in the regions between ${\sim} 7$ and $15 \: \rm kpc$ which have general properties resembling the Solar neighborhood. The slow increase of $Z$ leads to a less peaked stellar metallicity distribution than that in the ``extreme'' infall model shown in Figure~\ref{fig:fig8}.
\subsubsection{Timescales for chemical evolution}
\label{subsubsec:Timescales for chemical evolution}
The original evidence for infall of extragalactic gas was the explanation of the high-velocity clouds of neutral hydrogen \citep{Oort1970TheHydrogen.}, which required an infall rate comparable to the SFR. Further evidence is given by the timescale for star formation to consume the ISM in the Solar neighborhood, as pointed out by \citet{Larson1972a}: the local surface density of ISM is ${\sim} 8 \: \rm M_{\odot} \ pc^{-2}$ while the SFR is ${\sim} 10 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$ (Section~\ref{subsubsec:Local SFR}), so the ISM is being used up on a timescale
\begin{equation}
\tau_{*} \equiv \frac{M_{\rm g}}{(1 - R) \psi} \sim 1 \: \rm Gyr.
\label{eq:eq4.7}
\end{equation}
If there is not a source of new gas, flowing in at about the SFR, we must be living very near the end of active star formation in this part of the disk. Various constraints related to interstellar pressure balance and Galactic X-ray emission allow an infall rate up to ${\sim} 1 \: \rm M_{\odot} \ pc^{-2} \ Gyr^{-1}$ in the Solar neighborhood \citep{Cox1976AccretionRate}. This is enough to extend the lifetime of the gas against total consumption by only ${\sim} 10\%$, but all of the above estimates are rather uncertain, and it is possible that $f \sim \psi$ in the Solar neighborhood, so the gas is continually replenished.
A related timescale is that for chemical enrichment, which can be written
\begin{equation}
\tau_{\rm Z} \equiv \frac{Z}{\left| dZ / dt \right|} = \frac{Z M_{\rm g}}{\left| y(1 - R) \psi + \left( Z_{\rm f} - Z \right) f \right|},
\label{eq:eq4.8}
\end{equation}
using Equation~(\ref{eq:eq3.19}). If infall were negligible, this Equation would give $\tau_{\rm Z} = \tau_{*} (Z / y) \simeq \tau_{*} \sim 1 \: \rm Gyr$; we would then expect to see $Z$ increasing on a timescale of $1 \: \rm Gyr$ now, which is not consistent with the data in Figure~\ref{fig:fig7}. However, if gaseous inflow is occurring at a rate comparable to the SFR, $\tau_{\rm Z}$ could be much longer.
Other processes than infall have been suggested to account for the slow enrichment rate in the Solar neighborhood. For example, various authors beginning with \citet{Truran1971EvolutionaryGalaxy} have suggested a time-dependent IMF with a greater yield in the past; thus at present $Z \gg y$, so $\tau_{\rm Z} \gg \tau_{*}$.
Whatever such additional processes may affect the evolution and distribution of metal abundance, it seems almost inevitable that gas flows related to galaxy formation have a strong influence on chemical evolution. The dynamical models discussed above were designed to reproduce the structural properties of galaxies \citep{Larson1976a}, so it is significant that they predict both pre-enrichment of the disk gas and later metal-poor infall, which thus appear to be the most natural solutions to the G-dwarf and timescale problems of the Solar neighborhood. From these results, it is to be expected that the chemical evolution of other regions of galaxies is affected by gas flows; examples will be reviewed in Section~\ref{sec:Chemical Evolution of Galaxies}.
\subsection{Relative Abundances of the Elements}
\label{subsec:Relative Abundances of the Elements}
Relative abundances of elements heavier than helium provide information on both nucleosynthesis and galactic evolution, in a variety of ways that are discussed extensively in the literature (see reviews by \citealp{Audouze1976ChemicalGalaxies.}, \citealp{Pagel1978a}, and \citealp{Trimble1975TheElements}). Here the emphasis will be on points relevant to galactic evolution.
\subsubsection{Primary elements from different stars}
\label{subsubsec:Primary elements from different stars}
Stars of different masses eject new primary elements in different proportions, so if the ISM is not well mixed we expect to see some variations in relative abundances on the surfaces of unevolved stars. (Variations would arise similarly from any dependence of primary nucleosynthesis on the initial compositions of stars, Section~\ref{subsubsec:Initial composition}). Until recently, most reported variations were within the uncertainties of measurement and analysis, but some trends are now emerging.
The relative abundances of carbon, oxygen, and iron (for which the data are reviewed by \citealp{Pagel1978a}) are an interesting example. All unevolved stars with reliable carbon abundances have $\rm [C / Fe] = 0$ within the errors, i.e. a Solar abundance ratio\footnote{This statement is tentative, since very few metal-poor stars have quantitative carbon abundances and some globular clusters show preliminary evidence for carbon excesses and/or deficiencies \citep{McClure1979ObservationalPopulations}.}, but oxygen is systematically overabundant in metal-poor stars, a typical value being $\rm [O / Fe] = 0.5$ in stars with $\rm [Fe / H] < -1$. Since these stars have halo kinematics, it appears that the yield of oxygen was relatively high during the early life of the Galaxy. Oxygen is the most abundant element after H and He, so it would be more appropriate to say that the yield of iron (and perhaps carbon) was relatively low.
One possible explanation is that the IMF was different, favoring massive stars, which have a higher oxygen/carbon ratio in their ejecta according to current stellar models \citep[e.g.][]{Arnett1978OnStars}. Another possibility is that significant fractions of iron and carbon are ejected by stars with such long lives that they die too late to enrich the halo stars. This idea agrees qualitatively with the mass dependence of stellar nucleosynthesis outlined in Section~\ref{subsec:Beyond MS}: carbon stars (and planetary nebulae) with initial masses from ${\sim} 1 - 8 \: \rm M_{\odot}$ indicate that carbon production takes place partly in very long-lived stars; Type~I supernovae, with iron-enriched envelopes, show that iron is ejected at least partly by stars of $\lesssim 6 \: \rm M_{\odot}$; but there is no evidence for oxygen production outside the massive stars ($\gtrsim 10 \: \rm M_{\odot}$) that are theoretically predicted to be its main source. Models showing the effects of nucleosynthesis in stars with such a wide range of lifetimes have been studied by \citet{Tinsley1979StellarEvolution}. It is found that systematic variations in relative abundances arise if star formation occurs on timescales less than the lifetime of the least massive star producing significant amounts of an element of interest. For example, iron produced by stars of intermediate mass (${\sim} 4 - 6 \: \rm M_{\odot}$) is underabundant relative to oxygen if stars form on a timescale ${\sim} 10^{8} \: \rm yr$. Even if the whole time for halo formation is several billion years, as indicated by the age spread of globular clusters (Section~\ref{subsec:Stellar_Populations}), such variations will appear in halo stars if they were formed in a series of short bursts. A very inhomogeneous protogalaxy, with stars forming in dense lumps, has been suggested on dynamical grounds (Section~\ref{subsec:Galaxy_Formation}), so it is interesting to note how chemical abundances could be affected.
\subsubsection{Secondary elements}
\label{subsubsec:Secondary elements}
Secondary elements are those synthesized from ``seed'' elements heavier than helium that were already present in the star at birth; some examples have been mentioned in Section~\ref{subsubsec:Initial composition}. The stellar yields of these elements increase with metallicity, so their abundances are expected to increase more rapidly (and to be relatively lower in metal-poor stars) than those of primary elements. As reviewed by \citet{Pagel1978a}, this prediction is borne out qualitatively by relative underabundances of nitrogen and barium in metal-poor stars, but the data are in quantitative disagreement with theoretical models.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig9.png}
\caption{Ratios of primary to secondary abundances versus metallicity. \emph{Data points} (and shaded area with many points) are from sources cited by \citet{Tinsley1979StellarEvolution}, and \emph{dashed lines} correspond to a simple scenario for secondary nucleosynthesis (Section~\ref{subsubsec:Secondary elements}). The nitrogen data are all for dwarfs; barium data include giants, but these do not give a different picture from the dwarfs alone, so there is probably no significant contamination by the stars' own nucleosynthesis.}
\label{fig:fig9}
\end{figure}
The canonical prediction for secondary abundances is that they increase as the square of primary abundances, e.g. $\rm Ba / H \propto (Fe / H)^{2}$ or $\rm Ba / Fe \propto Fe / H$. This prediction arises, as follows, from the conventional assumption that a star of given mass produces and ejects an amount of secondary element in direct proportion to the star's initial primary seed abundance. Let us consider a secondary element ``$s$'', with primary seed $Z$, and let $p_{\rm sm}$ be the mass of new $s$ ejected by a star of mass $m$ with initial metallicity $\rm Z_{\odot}$; with another initial metallicity $Z$, the amount of new $s$ ejected is $p_{\rm sm} (Z / \rm Z_{\odot})$. Thus the expression like (\ref{eq:eq3.8}) for the rate of ejection of new $s$ is
\begin{equation}
\int_{m_{\rm t}}^{\infty} mp_{\rm sm} \frac{Z \left( t - \tau_{\rm m} \right)}{\rm Z_{\odot}} \psi \left( t - \tau_{\rm m} \right) \phi(m) \ dm.
\label{eq:eq4.9}
\end{equation}
The instantaneous recycling approximation to this expression can be written $y_{\rm s} (Z / {\rm Z_{\odot}}) (1 - R) \psi$, where $Z$ and $\psi$ are current values. The equation analogous to (\ref{eq:eq3.19}) for the abundance $X_{\rm s}$ of $s$ is therefore
\begin{equation}
M_{\rm g} \frac{dX_{\rm s}}{dt} = y_{\rm s} \left( \frac{Z}{\rm Z_{\odot}} \right) (1 - R) \psi + \left( X_{\rm sf} - X_{\rm s} \right) f.
\label{eq:eq4.10}
\end{equation}
In the simple model of Section~\ref{subsubsec:closed system}, the equation analogous to (\ref{eq:eq3.22}) is
\begin{equation}
M_{\rm g} \frac{dX_{\rm s}}{dM_{\rm g}} = -y_{\rm s} \left( \frac{Z}{\rm Z_{\odot}} \right).
\label{eq:eq4.11}
\end{equation}
Now $Z$ is given by Equation~(\ref{eq:eq3.23}), so (\ref{eq:eq4.11}) can be solved for $X_{\rm s}$, with the result
\begin{equation}
X_{\rm s} = \frac{1}{2} \left( \frac{y_{\rm s}}{y \rm Z_{\odot}} \right) Z^{2},
\label{eq:eq4.12}
\end{equation}
which is the canonical prediction $X_{\rm s} \propto Z^{2}$ mentioned above. Since Solar abundances are near the mean disk values, they must also be approximately the yield values, so the prediction is equivalent to $X_{\rm s} / Z \propto Z$ with a zero-point near the Solar values.
Stellar nitrogen and barium abundances are plotted in Figure~\ref{fig:fig9}, where the canonical relation would correspond in each case to the straight line at $45^{\circ}$ through about the Solar values (zeros in the logarithmic notation). It can be seen that the nitrogen abundances scatter greatly about the predicted relation, mainly exceeding the predicted value in metal-poor stars; and barium abundances are almost all greater than predicted, no matter where one places this zero-point among the data for stars near Solar composition.
\medskip
What is wrong? There is a logical flaw in making predictions with the simple model because it is known to be untenable on account of the stellar metallicity distributions (Section~\ref{subsec:G-Dwarf}); furthermore, instantaneous recycling may be a bad approximation for secondary elements that are made in red giants down to initial masses ${\sim} 1 \: \rm M_{\odot}$. Plausible modifications to the above theory lead to no better agreement, however, as the following examples show \citep{Tinsley1979StellarEvolution}.
\begin{enumerate}
\item Numerical calculations allowing for delayed ejection by long-lived stars make the situation worse, because they predict even smaller secondary abundances in the most metal-poor stars.
\vspace{1mm}
\item Infall makes the situation even worse, because it predicts a steeper rise than $X_{\rm s} / Z \propto Z^{2}$ at the metal-rich end.
\end{enumerate}
The latter result can be derived simply in the extreme infall model of Section~\ref{subsubsec:balanced system}. Use of Equation~(\ref{eq:eq4.10}), with $X_{\rm sf} = 0$ and (\ref{eq:eq3.28}) for $Z$, leads to a result analogous to (\ref{eq:eq3.28}),
\begin{equation}
X_{\rm s} = \frac{y_{\rm s} y}{\rm Z_{\odot}} \left( 1 - e^{-\nu} -\nu e^{-\nu} \right).
\label{eq:eq4.13}
\end{equation}
This formula reduces to Equation~(\ref{eq:eq4.12}) in the limit of low abundances, but its asymptotic limit, $X_{\rm s} \rightarrow y_{\rm s} y / \rm Z_{\odot}$, is twice the value obtained by extrapolating (\ref{eq:eq4.12}) to the asymptotic value $Z \rightarrow y$. Thus a plot of $X_{\rm s} / Z$ versus $Z$ for this model would have a slope of $45^{\circ}$ at first, but would become steeper at the top, obviously in disagreement with the data in Figure~\ref{fig:fig9}.
From these results, it appears that the problem in Figure~\ref{fig:fig9} is not due to a simplistic choice of model for chemical evolution, but to wrong assumptions about nucleosynthesis. For nitrogen, it has often been suggested that a primary source of nucleosynthesis could contribute significantly \citep[e.g.][]{Truran1971EvolutionaryGalaxy}, and this component of the nitrogen abundance would make the ratio $\rm N / Fe$ more nearly constant. A possible scenario for primary nitrogen production is that some carbon produced by helium burning in stars of intermediate mass gets mixed into the hydrogen-burning shell region and there converted to nitrogen; the models in \citet{IbenI.1976FurtherStar} illustrate this possibility. Nitrogen abundances in \textsc{H~ii} regions, in our own and other galaxies, show a scatter similar to the stellar data, and primary production has again been proposed as an explanation \citep[e.g.][]{Smith1975SpectrophotometricGalaxies, Edmunds1978NitrogenGalaxies}.
For barium there is no obvious solution to the discrepancy in Figure~\ref{fig:fig9}, since a primary source is unlikely in terms of nuclear physics. The calculations of stellar structure and nucleosynthesis used to predict barium production are so complicated that results for Solar initial metallicities cannot be extrapolated to others \citep{IbenI.1978OnMedium}; it is therefore likely that the conventional assumption, that production scales directly with the initial $\rm Fe / H$ of a star of given mass, is wrong.
\medskip
Other aspects of secondary nucleosynthesis relevant to galactic evolution are discussed in the references cited at the beginning of Section~\ref{subsec:Relative Abundances of the Elements} and in the proceedings of a meeting on CNO isotopes in astrophysics \citep{Audouze1977CNOAstrophysics}.
\subsubsection{Radioactive elements}
\label{subsubsec:Radioactive elements}
The field of nucleochronology (or cosmochronology or nucleo-cosmochronology) uses meteoritic abundances of the decay products of certain radioactive elements to derive information on the age distribution of elements in the early Solar System. This information is relevant to the rates of star formation and nucleosynthesis before the Sun formed, and to the age of the Galaxy. Nucleochronology has been reviewed recently by \citet{Fowler1979NuclearReport}, \citet{Schramm1974Nucleo-Cosmochronology}, and \citet{Trimble1975TheElements}, and its relevance to galactic evolution is discussed further by \citet{Reeves1972SpatialNucleosynthesis}, \citet{Reeves1976ThePhenomena}, and in other papers referenced below. Some general principles will be outlined here.
\medskip
Let us consider the abundance $X_{\rm i}$ of a radioactive nuclide with decay constant $\lambda_{\rm i}$ and net yield $y_{\rm i}$. The equations of chemical evolution must be modified to allow for decay in the ISM and in stellar envelopes; the equation analogous to (\ref{eq:eq3.7}) is
\begin{equation}
\frac{d X_{\rm i} M_{\rm g}}{dt} = -X_{\rm i} \psi + E_{\rm i} + X_{\rm if} f - \lambda_{\rm i} X_{\rm i} M_{\rm g},
\label{eq:eq4.14}
\end{equation}
where $E_{\rm i}$ is given, analogously to Equation~(\ref{eq:eq3.9}), by
\begin{equation}
\begin{medsize}
E_{\rm i}(t) = \int\limits_{m_{\rm t}}^{\infty} \left[ \left( m - w_{\rm m} - m p_{\rm im} \right) X_{\rm i} \left( t - \tau_{\rm m} \right) e^{-\lambda_{\rm i} \tau_{\rm m}} + m p_{\rm im} \right] \psi \left( t - \tau_{\rm m} \right) \phi(m) \ dm.
\label{eq:eq4.15}
\end{medsize}
\end{equation}
The decay timescale, $\lambda_{\rm i}^{-1}$, is of course independent of the system's timescales for star formation and infall, so straightforward analytical solutions for $X_{\rm i}$ are possible only in special cases, even if instantaneous recycling is used. In fact, the long-lived radioactive elements are useful just \emph{because} their abundances are sensitive to evolutionary details (such as the past SFR and the age) that cancel in the instantaneous recycling approximation and hardly affect other abundance parameters in detailed models.
The simplest situation would be if all the elements formed in an initial burst. Then their decay in the ISM and stellar envelopes would lead to abundances at time $t$ given by $X_{\rm i}(t) = X_{\rm i0} \exp( -\lambda_{\rm i} t)$, where $X_{\rm i0}$ is the initial abundance and is proportional to the amount of element $i$ made, i.e. to its yield $y_{\rm i}$. In particular, at the time $T$ when the Solar System formed, two such elements would have an abundance ratio
\begin{equation}
\frac{X_{\rm i}(T)}{X_{\rm j}(T)} = \frac{X_{\rm i0}}{X_{\rm j0}} \exp \left[ - \left( \lambda_{\rm i} - \lambda_{\rm j} \right) T \right],
\label{eq:eq4.16}
\end{equation}
where $X_{\rm i0} / X_{\rm j0} = y_{\rm i} / y_{\rm j}$. The value of $T$ could in this case be found by substituting into Equation~(\ref{eq:eq4.16}) the yield ratio (given by nucleosynthesis theory) and the abundance ratio in the early Solar System (obtained from the abundances of decay products in meteorites). The age of the Galaxy, which is simply $T$ plus the age of the Sun, could then be determined.
Of course, the elements were probably synthesized continually during the period $T$ from first star formation in the Galaxy until the Solar System condensed, so Equation~(\ref{eq:eq4.16}) is unrealistic. As first shown by \citet{Schramm1970NucleochronologiesElements}, one can define a useful age parameter, here denoted $T_{\rm ij}$, analogously to the solution of (\ref{eq:eq4.16}) for $T$:
\begin{equation}
T_{\rm ij} \equiv \frac{1}{\lambda_{\rm i} - \lambda_{\rm j}} \ln \left[ \frac{y_{\rm i} / y_{\rm j}}{X_{\rm i}(T) / X_{\rm j} (T)} \right],
\label{eq:eq4.17}
\end{equation}
which can, in principle, be evaluated from meteoritic and nuclear data independently of Galactic evolution. In the limit of long-lived elements, ($\lambda_{\rm i} T \ll 1, \: \lambda_{\rm j} T \ll 1$), $T_{\rm ij}$ is just the mean age of elements in the gas, $T_{Z}$, at the time when the Solar System formed \citep{Tinsley1975c}.
The relation between the mean age ($T_{Z}$) and the elapsed time ($T$) is model-dependent, so estimates of $T_{\rm ij}$ for long-lived pairs of elements do not give $T$ directly. Different possibilities include the following.
\begin{enumerate}
\item If essentially all nucleosynthesis of the relevant elements took place in an initial burst, then $T = T_{Z}$.
\vspace{1mm}
\item The simple model for chemical evolution gives the intuitive result that the mean age is half the elapsed time, $T_{Z} = T / 2$; however, because this model is discrepant with stellar metallicities, one cannot assume that $T$ is given simply by $2 T_{Z}$ in reality.
\vspace{1mm}
\item In extreme infall models, the ISM and heavy elements in it have a mean age ${\sim} M_{\rm g} / f$ at all times greater than $M_{\rm g} / f$; so if chemical evolution in the disk was strongly affected by infall before the Solar System formed, the value of $T_{Z}$ obtained from meteoritic abundances may reflect only the timescale for infall, independently of the age $T$.
\vspace{1mm}
\item There are consistent models for the Solar neighborhood, with some infall and/or some early enrichment, that have values of $T_{Z} \simeq T / 2$ (as emphasized by \citealp{Hainebach1977CommentsNucleocosmochronology}), but since not all plausible models have this relation it cannot be used confidently (as emphasized by \citealp{Tinsley1977b}).
\end{enumerate}
In summary, there is a large uncertainty in any age estimate of the Galaxy derived from nucleochronology, except of course that the age of the Solar System is a reliable lower limit!
\medskip
The initial Solar System abundances of short-lived radioactive elements are sensitive to rates of nucleosynthesis at times immediately preceding the solidification of the meteorites. Their abundances suggest that the nucleosynthesis of most elements ceased ${\sim} 10^{8} \: \rm yr$ before the solidification time, but some material was produced only ${\sim} 10^{6} \: \rm yr$ earlier. Interpretations of these timescales include passage of the pre-Solar material through spiral arms at $10^{8}$-yr intervals; enrichment by fresh supernova ejecta each $10^{6} \: \rm yr$, which could result from the average supernova rate in the Solar neighborhood; a last-minute supernova that triggered the formation of the Solar System; and locking of radioactive elements, with their decay products, into grains long before the Solar System formed. These possibilities are reviewed briefly by \citet{Podosek1978IsotopicMaterials}, and discussed in detail by several authors in a conference proceedings edited by \citet{Gehrels1978ProtostarsPlanets.}. As yet there is no consensus on the interpretation of short-lived radioactivities in the early Solar System, but ultimately they should provide valuable information on star formation and interstellar chemistry.
\section{Chemical Evolution of Galaxies}
\label{sec:Chemical Evolution of Galaxies}
For other galaxies, and for regions of our own outside the Solar neighborhood, there is much less detailed information on abundance distributions, but there are some striking trends that call for explanations involving the formation and later dynamical evolution of galaxies. A few relevant properties have been mentioned in Section~\ref{subsec:Stellar_Populations}, and now further details will be described and some of the theoretical models reviewed. Other general reviews of this subject include those by \citet{Audouze1976ChemicalGalaxies.}, \citet{Pagel1978b}, and \citet{Trimble1975TheElements}.
\subsection{Outline of Relevant Data}
\label{subsec:Data2}
Abundances are very often found to decrease outward in galaxies: gradients have been observed in the \textsc{H~ii} regions of disks, in disk stars, and in the stars of spheroidal systems including elliptical galaxies and the bulge-halo components of spirals.
\medskip
In the Galactic disk, \textsc{H~ii} regions within a few kpc of the Sun have an average gradient $d {\rm [O / H]} / dr \simeq -0.1 \: \rm kpc^{-1}$ (where $r$ is the galactocentric distance), while stars of intermediate age have a gradient $d {\rm [Fe / H]} / dr \simeq -0.05 \: \rm kpc^{-1}$; an open cluster about $10 \: \rm kpc$ from the Sun in the anticenter direction, of age only $\lesssim 10^{8} \: \rm yr$, is apparently as metal-poor as $\rm [Fe / H] < -1$ \citep{Christian1979UBV21}. (Oxygen and iron abundances are quoted for the ISM and stars, respectively, because these are the best observed elements). The uncertainties are such that the apparent age dependence of the gradient may or may not be real. These data are reviewed by \citet{Janes1977ADisk}, \citet{Mayor1977ChemicalGalaxy}, and \citet{Peimbert1977ChemicalGalaxy}. \textsc{H~ii} regions in external galaxies generally show gradients of a similar magnitude (e.g. \citealp{Smith1975SpectrophotometricGalaxies, Shields1978TheM101}; and references therein). However, only a marginal gradient appears in the Large Magellanic Cloud and none in the Small Cloud \citep{Pagel1978AClouds}.
The most obvious explanation for these gradients would be that the outer regions of disks are less chemically evolved than the inner regions, in the sense of having converted a smaller fraction of their gas to stars. The simple model, for example, would predict a $Z$ gradient given by $Z = y \ln \ \mu^{-1}$ arising from a gradient in the gas fraction $\mu$ (Section~\ref{subsubsec:closed system}). However, the best studied galaxies (the Milky Way and M101) probably do not have a sufficient gradient in $\mu$ for this explanation to suffice. \citet{Gordon1976CarbonNucleons} found that the combined surface densities of atomic and molecular hydrogen lead to a nearly constant gas fraction ($\mu \sim 0.05$) at $R > 4 \: \rm kpc$ in the Galaxy, and \citet{Shields1978TheM101} noted a similar problem in M101. The amount of ISM interior to the Sun could be overestimated, since the $\rm H_{2}$ density is derived from observations of CO molecules on the assumption of a constant abundance ratio $\rm CO / H_{2}$; if in fact $\rm CO / H_{2}$ increases inward because of a $\rm C / H$ abundance gradient, then there is less $\rm H_{2}$ than had been thought at small radii \citep{Peimbert1977ChemicalGalaxy}. With this correction, the Galaxy probably has some gradient in $\mu$, which could account in part for the interstellar composition gradient. Of course, the simple model is known to be invalid in the Solar neighborhood, so we do not expect the formula $Z = y \ln \ \mu^{-1}$ to explain gradients in detail. Other ways of generating gradients will be mentioned in Section~\ref{subsec:Abundance Gradients in Disks}.
The Galactic halo stars also have a metallicity gradient. Studies of individual stars in globular clusters show that a spread of metallicities of $\rm [Fe / H] \sim -2$ to $0$ occurs in the innermost $5 \: \rm kpc$ (measured from the Galactic center, at any angle to the disk), while the upper limit drops to ${\sim} {-1}$ at greater radii \citep{Cowley1978SpectraGalaxy, Searle1978CompositionsHalo}. It is not clear whether a systematic decline in iron and/or CNO abundances persists further out \citep{McClure1979ObservationalPopulations, Kraft1978EvidenceClusters}.
\medskip
Many elliptical and S0 galaxies have gradients of integrated colors and line strengths in their spectra that are best interpreted as composition gradients. The same quantities also vary systematically with the absolute magnitudes of E and S0 galaxies, as illustrated in Figure~\ref{fig:fig10}, indicating that the brighter galaxies are more metal-rich. A thorough review of this subject is given by \citet{Faber1977ThePopulations}. The analysis and calibration of abundance effects in the integrated light of a galaxy are much more complicated than for individual stars, because line strengths are strongly affected by the mixture of stellar temperatures and luminosities, and because the whole population in the HR diagram is shifted by effects of metallicity on the interiors and atmospheres of stars (Section~\ref{subsubsec:Initial composition}). Until recently, it was not clear whether all elements heavier than helium enter the composition trends in E and S0 galaxies, or whether a few with strong spectral features (N, Mg, Na), are mainly responsible. \citet{Cohen1978AbundancesM13} has now made a detailed observational study of some lines of Ca, Na, Mg, and Fe, together with approximate theoretical predictions of how their strengths should vary with the composition of a population of stars; she finds no evidence against the abundances of all of these elements varying in the same proportions.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig10.png}
\caption{A color--magnitude diagram for elliptical and S0 galaxies in several clusters of galaxies \citep{Visvanathan1977TheClusters}. The color $(u - V)_{\rm c}$ is corrected for reddening, redshift, and aperture effects; magnitudes are adjusted to the distance of the Virgo cluster using redshift ratios. (\emph{Crosses} denote Virgo galaxies not used by \citealp{Visvanathan1977TheClusters} in their linear fit to the data). The \emph{straight lines} are a linear fit, with $\pm 2 \sigma$ boundaries. Despite the excellent linear fit over a large range of magnitudes, the brightest few points for the Virgo cluster alone (\emph{filled circles}) and for other clusters (\emph{open circles}) show a tendency to level off in color.}
\label{fig:fig10}
\end{figure}
The color--magnitude relation for elliptical galaxies is linear over a wide range of metallicities (Figure~\ref{fig:fig10}), which suggests a power-law relation between metallicity and luminosity. A tentative calibration, subject to revision when both the data and the theoretical color--metallicity relation are more secure, is
\begin{equation}
Z_{\rm s} \propto L_{\rm B}^{0.3}
\label{eq:eq5.1}
\end{equation}
\citep{Tinsley1978b}, where $Z_{\rm s}$ is the average metallicity of stars in an elliptical galaxy of blue luminosity $L_{\rm B}$. This relation was derived by from population models differing only in metallicity, so that the stars in all cases had the same IMF and age (as outlined in Section~\ref{subsubsec:Evolutionary models}). Such models also predict that the mass-to-luminosity ratio should depend on metallicity, yielding a relation $M_{\rm s} / L_{\rm B} \propto L_{\rm B}^{0.13}$. A relation very similar to this has been obtained observationally by \citet{Schechter1979ObservationsGalaxies} for the cores of elliptical galaxies. Equation~(\ref{eq:eq5.1}) therefore corresponds to a tentative metallicity--mass relation,
\begin{equation}
Z_{\rm s} \propto M_{\rm s}^{0.25},
\label{eq:eq5.2}
\end{equation}
where $M_{\rm s}$ is the mass of stars (not any extended hidden halo material), and the main uncertainties in the exponent are due to the color--magnitude data and to the color--metallicity calibration. There is some evidence that the color--magnitude relation levels off at the magnitudes of the brightest cluster galaxies, as can be seen, for example, from the brightest few points in Figure~\ref{fig:fig10}.
\medskip
\begin{figure*}
\includegraphics[width=0.97\textwidth]{Figures/Fig11.png}
\caption{(a) Star formation rates and (b) metallicities of newly formed stars (i.e., $Z$ of the gas), at several radii in a collapse model for the formation of a spherical galaxy \citep{Larson1974a}. The radius in pc is marked on each curve, and the three ticks indicate the times at which star formation is $10\%$, $50\%$, and $90\%$ complete (relative to the final mass of stars) at that radius.}
\label{fig:fig11}
\end{figure*}
Heavy elements have been detected in intergalactic gas, and since these elements almost certainly come from galactic stars they are relevant to the chemical evolution of galaxies. A feature due to iron has been observed in the diffuse X-ray emission spectra of several rich clusters of galaxies; the interpretation is that these clusters contain hot gas (${\sim} 10^{8} \: \rm K$), emitting by thermal Bremsstrahlung, with approximately Solar abundances of iron. The mass of intergalactic gas inferred from the data is model-dependent \citep[e.g.][]{Bahcall1977ParametersVirgo., Perrenod1978TheSources}, and is roughly between 1 and 30 times the total luminosity of cluster galaxies, in solar units. Now the mass of stars in galaxies is ${\sim} 10$ times their total luminosity, in solar units\footnote{This mass is not to be confused with the virial masses of clusters, which are ${\sim} 100$ times the total luminosity and which provide the main evidence for hidden non-stellar matter in association with galaxies.}, and their average metallicity is approximately Solar, so the intergalactic medium (IGM) in the rich clusters apparently contains about the same total mass of iron as do the stars themselves. These observations suggest that galaxies sometimes lose enough metal-rich gas to affect their own chemical evolution substantially.
Another striking observation of metal-rich IGM is absorption due to Ca and Mg in the spectrum of a quasar (3C~232) lying $1.9^{\prime}$ on the sky from a foreground spiral galaxy (NGC~3067); the absorption redshift matches that of the galaxy, and neutral hydrogen absorption has been detected at the same redshift in the radio spectrum of the quasar \citep{Haschick1975Neutral3067}. The line strengths and positions of the objects imply that there is gas with roughly Solar abundances at least $17 \: \rm kpc$ from the center of the galaxy \citep{Boksenberg1978The3067}.
\subsection{Abundance Gradients in Spheroidal Systems}
\label{subsec:Abundance Gradients in Spheroidal Systems}
Most stars in elliptical galaxies and in the bulge-halo components of spirals were probably formed within a few times $10^{9} \: \rm yr$ at a time ${\sim} 10^{10} \: \rm yr$ ago. The abundance gradients in these systems therefore reflect processes occurring during the time of star formation in a protogalaxy; several such processes have been suggested as possible causes of gradients.
\subsubsection{Dissipative collapse}
\label{subsubsec:Dissipative collapse}
The most extensive models exploring the effects of star formation during the collapse of an initially gaseous protogalaxy are those reviewed by \citet{Larson1976b}. In the spheroidal components of model disk galaxies, and in model ellipticals, star formation becomes increasingly concentrated toward the center as the density builds up. This effect is illustrated in Figure~\ref{fig:fig11}~(a), which shows the SFR as a function of time at several radii in a spherical collapse model. Stars formed at a given radius remain in orbits with little net inward motion, but the gas sinks further in because it is dissipative (i.e., its kinetic energy of radial motion is partly lost via collisionally induced radiation). Thus the metals ejected by evolving stars are carried inward by the gas, and an abundance gradient develops in the gas. As stars continue to form, their compositions reflect this gaseous abundance gradient. Figure~\ref{fig:fig11}~(b) shows the evolution of metallicity of newly formed stars (i.e., $Z$ of the gas) at several radii in a spherical model, and the rapid development of a gradient is clear. The same process of dissipation produces a central concentration in the gas density, which leads to a condensed nucleus of stars.
If there were no dissipation, the stars and gas would collapse together and the metals would not be concentrated inward. Thus in the outer parts of some of these models, where the protogalactic density is too low for dissipation to be effective, no stellar abundance gradient appears. The possible lack of a gradient in metallicities of globular clusters beyond ${\sim} 10 \: \rm kpc$ from the Galactic center has therefore been interpreted as showing that the collapse of the Galaxy began with the stars and gas in free-fall; conversely, the gradient at smaller radii is interpreted as showing the effects of dissipation at a later stage of the collapse \citep{Larson1977c}.
\subsubsection{A gradient in the IMF}
\label{subsubsec:A gradient in the IMF}
Aside from any effects of gas flows, negative metallicity gradients could be produced by gradients in the IMF that led to a yield decreasing outward. Since the yield (Equation~\ref{eq:eq3.12}) depends on the relative numbers of metal-producing stars, possibilities would be a steeper slope for massive stars or more low-mass ``inert'' stars, at larger radii. In the latter case, the stars that survive to the present would still have a radial gradient in their mass function, with an interesting consequence: most of the luminosity of an old population of stars comes from dwarfs near the MS turnoff and evolving giants, while most of the mass is in less massive objects that contribute little light; thus the $M / L$ ratio increases with the proportion of very low-mass stars, and the postulated gradient in the IMF would lead to an outward increase of $M / L$. Such a trend is indeed observed, although it is sometimes ascribed to an extended ``halo'' of non-stellar condensed objects that formed too soon to affect chemical evolution (Section~\ref{subsubsec:Other IMF}). \citet{vandenBergh1975StellarGalaxies} has suggested that the IMF tends to have more massive stars in regions of high density, and this view of the origin of metallicity gradients is part of his evidence. The hypothesis of a gradient in the IMF in spheroidal systems has no convincing theoretical basis, and the trends it would explain can arise in other ways, but nevertheless systematic variations in the IMF could be as important as they are hard to verify.
\subsubsection{Finite stellar lifetimes}
\label{subsubsec:Finite stellar lifetimes}
The timescale for star formation in a protogalaxy could be comparable to the lifetimes of some metal-producing stars, in which case stars formed early would be relatively unenriched. Thus if the outermost stars formed before most of the central ones, there would be a negative metallicity gradient. In the models in Section~\ref{subsubsec:Dissipative collapse} \citep{Larson1976b} it is assumed that all metals are produced by stars with lifetimes $< 3 \times 10^{7} \: \rm yr$, so this effect is negligible; but iron production by Type~I supernovae could in fact be significant on longer timescales (Section~\ref{subsubsec:Intermediate mass}). What timescales are relevant? The minimal collapse time for a protogalaxy is the free-fall time,
\begin{equation}
t_{\rm ff} = 1.7 \times 10^{6} \left( \frac{M}{10^{11} \: \rm M_{\odot}} \right)^{-\frac{1}{2}} \left( \frac{R}{1 \: \rm kpc} \right)^{\frac{3}{2}} \: \rm yr,
\label{eq:eq5.3}
\end{equation}
where $M$ and $R$ are the mass and radius. For example, a galaxy with $M = 2 \times 10^{11} \: \rm M_{\odot}$ and $R = 15 \: \rm kpc$ has $t_{\rm ff} = 7 \times 10^{7} \: \rm yr$; and a protogalaxy of the same mass collapsing from $R = 50 \: \rm kpc$ has $t_{\rm ff} = 4 \times 10^{8} \: \rm yr$. Much longer timescales for star formation are possible if the dissipation is slow, and the collapse time of the system can be much longer if its boundary initially expands with the Universe \citep{Gunn1972OnEvolution}. At least the outer parts of large galaxies could therefore be metal-poor partly because of the finite lifetimes of metal-producing stars.
A potential test is to look for variations in relative abundances. For example, if oxygen comes only from very massive stars but iron comes partly from stars of intermediate mass (Section~\ref{subsec:Beyond MS}; Section~\ref{subsubsec:Primary elements from different stars}), then iron should be more deficient than oxygen in the outermost stars. The hypothesis of a gradient in the IMF of massive stars would predict the opposite trend in relative abundances. Current data do not detect any gradients in relative abundances, but oxygen itself has not been studied and nor have the faint outer regions of elliptical galaxies.
It is quite possible that all of the processes discussed above were effective in producing abundance gradients in spheroidal systems, so clear choices among the theories are not to be expected.
\subsection{The Metallicity--Mass Relation for Elliptical Galaxies}
\label{subsec:The Metallicity--Mass Relation for Elliptical Galaxies}
The correlation between metallicity and mass (color and luminosity) of elliptical galaxies has been explained in several ways, of which two will be reviewed here. These each involve dynamical effects during galaxy formation, resulting in less complete conversion of the protogalactic gas to stars, and so to a smaller final mean stellar metallicity, in smaller systems. One could, of course, invoke differences in the IMFs of galaxies as a function of their mass, but there is no independent evidence for a trend of the required form.
\subsubsection{Supernova-driven winds}
\label{subsubsec:Supernova-driven winds}
Star formation and chemical enrichment are cut off in a protogalaxy if the residual gas is lost, and a possible loss mechanism is a galactic wind energized by supernova explosions. Galactic winds were first analyzed for a steady-state case by \citet{Johnson1971GalacticWinds} and \citet{Mathews1971GalacticWinds}; similar analyses have been made for nuclear bulges of spirals by \citet{Faber1976HWinds} and for bulge-disk systems by \citet{Bregman1978GalacticSequence.}. A galaxy sustains a steady-state wind if the supernova rate divided by the rate of supply of gas (from evolving stars) gives the gas enough energy to escape from the galactic potential well. For protogalaxies, we are interested not in the steady state, but in conditions for the initiation of a wind that can remove essentially all of the residual gas. \citet{Larson1974b} discussed possible effects of supernovae in heating the gas, and adopted a simple approximation as the condition for its removal: the residual gas is assumed to be lost suddenly when the total heat input from supernovae has provided the gas with the escape energy, assuming uniform conditions throughout the protogalaxy. This approximation is plausible enough to suggest how the loss condition scales with the relevant parameters, but there are unavoidably large uncertainties in the astrophysics involved so the results are not very secure. The scaling can be derived as follows.
\medskip
Let $E$ be the thermal energy imparted to the ISM by supernovae when a unit mass of stars is formed; $E$ is proportional to the fraction of stars that become supernovae, to the mean kinetic energy of material ejected in a supernova explosion, and to the efficiency with which this energy is transferred to the ISM as heat. (The last factor is the most uncertain). As an approximation, let $E$ be treated as a constant, despite finite stellar lifetimes and complicated effects of the clumpiness of the ISM, its chemical composition, etc. Let us consider a spherical protogalaxy of mass $M$ that has formed a mass $M_{\rm s}$ of stars and has residual gas mass $M_{\rm g} = M - M_{\rm s}$. The condition for gas to escape can be written
\begin{equation*}
\rm Potential \: energy \: of \: gas = Energy \: from \: supernovae,
\label{eq:eq5.energy}
\end{equation*}
i.e,
\begin{equation}
K \frac{GMM_{\rm g}}{R} = EM_{\rm s},
\label{eq:eq5.4}
\end{equation}
where $K$ depends on the density distribution in the galaxy and will be assumed constant as another simplification. Large elliptical galaxies are observed to be more tightly bound than small ones, so a greater fraction of their gas must be converted to stars before the condition (\ref{eq:eq5.4}) is satisfied; therefore, their surviving stars have a greater mean metallicity. Other consequences of this scenario are that the more massive galaxies collapse more extensively before star formation is cut off, so they are predicted to have more condensed nuclei and steeper metallicity gradients than smaller galaxies (Section~\ref{subsubsec:Dissipative collapse}). These trends are observed, lending support to this type of origin for the increase of metallicity with mass.
The form of the metallicity--mass relation can be accounted for using the same approximate model. Let the initial mass--radius relation for protogalaxies have the form
\begin{equation}
M \propto R^{\alpha},
\label{eq:eq5.5}
\end{equation}
so Equation~(\ref{eq:eq5.4}) can be written
\begin{equation}
M^{1 - \frac{1}{\alpha}} \left( M - M_{\rm s} \right) \propto M_{\rm s}.
\label{eq:eq5.6}
\end{equation}
Asymptotic equations for the mean metallicity of stars can be derived from very general considerations: the mass of metals synthesized and ejected is $yM_{\rm s}$, so at early stages of evolution when $M \simeq M_{\rm g}$, we have approximately
\begin{equation}
Z_{\rm s} \propto \frac{yM_{\rm s}}{M_{\rm g}} \simeq \frac{yM_{\rm s}}{M}, \: \: \left( Z_{\rm s} \ll y \right).
\label{eq:eq5.7}
\end{equation}
At late stages, Equation~(\ref{eq:eq3.21}) predicts that in all cases where mass is conserved,
\begin{equation}
Z_{\rm s} \rightarrow y, \: \: {\rm as} \: \: M_{\rm s} \rightarrow M.
\label{eq:eq5.8}
\end{equation}
The results from numerical collapse models verify these relations in cases of interest here. Substituting Equation~(\ref{eq:eq5.7}) into (\ref{eq:eq5.6}), we find the stellar metallicity--mass relation,
\begin{equation}
Z_{\rm s} \propto M_{\rm s}^{\frac{\alpha - 1}{2 \alpha -1}}, \: \: \left( Z_{\rm s} \ll y \right).
\label{eq:eq5.9}
\end{equation}
The tentative empirical relation (\ref{eq:eq5.2}) is obtained if $\alpha = 1.5$, which agrees fairly well with the observed mass--radius relation for elliptical galaxies if one considers how the stellar system must swell (to conserve energy) when the gas is lost \citep{Tinsley1978b}. According to Equation~(\ref{eq:eq5.8}), the power-law relation between $Z_{\rm s}$ and $M_{\rm s}$ must level off at large masses, with $Z_{\rm s} \rightarrow y$ in the limit when essentially all the original material is converted to stars; this behavior agrees with the levelling of the color--magnitude relation at the magnitudes of the brightest cluster galaxies.
The critical parameter $E$ can plausibly have a value that would give the right scale for the $Z_{\rm s}$--$M_{\rm s}$ relation \citep{Larson1974b}, but its value is very uncertain so the success of this theory must be considered tentative. The interaction between supernovae and the ISM could, in fact, be so weak as to drive a wind in only the very smallest protogalaxies.
\subsubsection{Bursts of star formation in merging subsystems}
\label{subsubsec:Bursts of star formation in merging subsystems}
Since the largest elliptical galaxies are the most metal-rich, a natural hypothesis is that chemical enrichment accompanied the \emph{growth} of galaxies by successive mergers among small subsystems. As noted in Section~\ref{subsec:Galaxy_Formation}, gaseous protogalaxies probably consisted of many dense lumps, so it is only a change of viewpoint to consider these as merging subsystems rather than as a collapsing unit. Moreover, extrapolations backward in time from the incidence of strongly interacting galaxies in sight today suggest that collisions and coalescence were common processes in the early lives of galaxies \citep{Toomre1977MergersConsequences, Vorontsov-Velyaminov1977NewGalaxies}. A property of colliding galaxies most relevant to chemical evolution is that they often appear to be undergoing intense bursts of star formation induced by the dynamical disturbance (Section~\ref{subsec:Colors of Peculiar Galaxies}), so it is reasonable to assume that star formation was caused in the past by coalescence of subsystems in a protogalaxy. A qualitative model of chemical enrichment by this process has been proposed by \citet{Tinsley1979StellarGalaxies}: elliptical galaxies form by a hierarchical sequence of mergers among subsystems, starting from small unenriched gas clouds; a burst of star formation occurs at each merger, so at each stage of growth the fraction of the total mass in stars increases and the mean metallicities of stars and gas increase. In this picture, the final mass of an elliptical galaxy is determined by the total mass of the small pieces initially present in its surroundings. When these have all been mopped up, efficient star formation stops. Any residual gas may get swept away if the system is moving through an ambient IGM, or possibly blown out in a wind; if it remains bound to the system, it could settle to a disk and leave the ``elliptical galaxy'' as the central bulge of a spiral.
The resulting $Z_{\rm s}$--$M_{\rm s}$ relation depends on the \emph{efficiency} of star formation as a function of the mass of the system (i.e., the system that has been built after a given number of mergers), where efficiency is defined as the mass of stars formed (in a given burst) per unit mass of gas. An approximately power-law relation between $Z_{\rm s}$ and $M_{\rm s}$ can be obtained only if the efficiency increases with the total mass of the system, i.e., with successive mergers. For example, a relation
\begin{equation}
{\rm Efficiency \: of \: star \: formation \propto (Total \: mass)}^{p},
\label{eq:eq5.10}
\end{equation}
where $p$ is a constant, leads to
\begin{equation}
Z_{\rm s} \propto M_{\rm s}^{\frac{p}{1 + p}}, \: \: \left(Z_{\rm s} \ll y \right),
\label{eq:eq5.11}
\end{equation}
with the usual limit $Z_{\rm s} \rightarrow y$ when all the gas is consumed. The relation (\ref{eq:eq5.10}) can be justified qualitatively by considerations of gas compression during collisions and mergers of subsystems. To reproduce the tentative empirical relation (\ref{eq:eq5.2}), Equation~(\ref{eq:eq5.11}) needs $p = 1/3$, which is consistent with the compression arguments. Equation~(\ref{eq:eq5.11}) results from (\ref{eq:eq5.10}) independently of such details as the mass distribution of merging pieces, and it can be understood as follows: Equation~(\ref{eq:eq5.7}) is true in any models with mass conservation (including here conservation of the total mass of merging pieces), while Equation~(\ref{eq:eq5.10}) gives, dimensionally,
\begin{equation*}
\frac{M_{\rm s}}{M_{\rm g}} \propto M^{p},
\label{eq:eq5.mass_ratio}
\end{equation*}
so that
\begin{equation}
M_{\rm s} \propto M^{1 + p} \: \: {\rm when} \: M_{\rm g} \simeq M \: \: \left( M_{\rm s} \ll M \right);
\label{eq:eq5.12}
\end{equation}
Equation~(\ref{eq:eq5.11}) then follows from (\ref{eq:eq5.7}) and (\ref{eq:eq5.12}). The power law is again predicted to level off, with $Z_{\rm s} \rightarrow y$ at high masses, according to Equation~(\ref{eq:eq5.8}).
As a theory for the origin of the metallicity--mass relation, this model has the advantage of invoking processes that can be studied in nearby interacting galaxies, but it remains to be seen whether the structural properties of elliptical galaxies are fully consistent with its dynamical implications.
\subsubsection{Mergers of stellar systems}
\label{subsubsec:Mergers of stellar systems}
The color--magnitude (metallicity--mass) relation for elliptical galaxies is apparently affected in a way that has nothing to do with chemical evolution: central cluster galaxies accrete their neighbors, by the process of dynamical friction. There is no star formation during these mergers, because the galaxies involved are ellipticals or S0s with almost no gas. Thus the growth in luminosity is not accompanied by chemical enrichment, and it can make the growing system bluer because the surrounding galaxies that it accretes are generally smaller than the central giant. Galactic cannibalism by dynamical friction was first proposed by \citet{Ostriker1975AnotherGalaxies}, and later papers (e.g. \citealp{Hausman1978GalacticClusters}, and references therein) have developed its implications for cosmological tests, the origin of core--halo structure of cD galaxies, the luminosity function of galaxies in clusters, and the color--magnitude relation itself. The process obviously tends to make the color--magnitude relation turn over toward bluer colors at the bright end. This effect has been proposed as a test for the occurrence of cannibalism in clusters, but the results are not unambiguous because there is an intrinsic upper limit, $Z_{\rm s} \rightarrow y$, to the average stellar metallicity in the models discussed above, that leads to a flattening of the relation anyway. Strong evidence that galaxies in the centers of clusters \emph{do} merge with each other is given by the lumpy appearance of the central supergiant (cD) members of some clusters; the lumps are interpreted as recently swallowed galaxies, and the timescale for them to merge into a smooth system is generally $< 10^{9} \: \rm yr$ \citep{Gunn1977ConcludingRemarks}.
\subsection{The Intergalactic Medium and Gas Lost from Galaxies}
\label{subsec:The Intergalactic Medium and Gas Lost from Galaxies}
Loss of interstellar gas from galaxies can both affect their own evolution, as discussed for example in Section~\ref{subsec:The Metallicity--Mass Relation for Elliptical Galaxies} above, and be a significant source of metal-enriched IGM.
\subsubsection{Loss of metals from galaxies}
\label{subsubsec:Loss of metals from galaxies}
The mass of metals lost from an elliptical galaxy can be estimated by the following argument, which is independent of the method of gas loss. The mass of metals ever made by stars in the galaxy is ${\sim} yM_{\rm s}$ (by the definition of the yield, Equation~\ref{eq:eq3.12}), and the mass of metals presently in its stars is $Z_{\rm s} M_{\rm s}$, so the mass lost to the IGM at some stage is ${\sim} (y - Z_{\rm s}) M_{\rm s}$. This reasoning was used by \citet{Larson1975GasGalaxies} to predict a substantial metal-rich IGM in clusters of galaxies, and a number of models with similar results have been advanced since the iron X-ray line was discovered. An essentially model-independent estimate can be made as follows.
Let $\phi(M_{\rm s})$ be the mass function of elliptical galaxies in a cluster. Then the total mass of metals they have supplied to the ISM is
\begin{equation}
M_{Z1} = \int \left[ y - Z_{\rm s} \left( M_{\rm s} \right) \right] M_{\rm s} \phi \left( M_{\rm s} \right) \ dM_{\rm s},
\label{eq:eq5.13}
\end{equation}
where $Z_{\rm s} (M_{\rm s})$ is a function derivable from the color--magnitude relation. In practice, $M_{\rm s}$ is expressed in terms of luminosity, and $\phi(M_{\rm s})$ is obtained from the luminosity function. The value of $y$ should be taken as the maximum $Z_{\rm s}$ of an elliptical galaxy, which is hard to obtain since the extensive outer regions that are probably metal-poor are seldom observed; setting $y$ equal to the mean metallicity of local stars (a little under $\rm Z_{\odot}$) is equivalent if elliptical galaxies have the local IMF. In a calculation equivalent to the one just outlined, \citet{Tinsley1979StellarGalaxies} found that a cluster of elliptical galaxies would contain a mass ${\sim} (2 - 5) \: \rm M_{\odot} Z_{\odot}$ of intergalactic metals per solar luminosity.
This is a very significant quantity of metals. For example, about $1/3$ of the luminosity of the Coma cluster is due to its elliptical galaxies, so they would provide a mass ${\sim} 1 \: \rm M_{\odot} Z_{\odot}$ of metals per solar luminosity of the cluster, corresponding to ${\sim} 0.1 \: \rm Z_{\odot}$ of metals per unit mass of galaxies (the ordinary stellar mass). If the bulges of spiral and S0 galaxies also lost their metals due to the IGM, this estimate could be doubled, but some of those metals may be in the disks (cf. Section~\ref{subsubsec:Metals from the young halo}). Iron can be considered as a representative metal in this calculation, so we predict ${\sim} 1 \times \rm (Fe / H)_{\odot} \times M_{\odot}$ of iron in the IGM per solar luminosity of the cluster. The actual mass of iron is quite uncertain, and could be equal to the predicted amount.
\subsubsection{Overall gas loss from galaxies}
\label{subsubsec:Overall gas loss from galaxies}
The total mass of gas lost from elliptical galaxies is a less predictable quantity, depending on gas flows within the galaxies and gain or loss of gas during the time of star formation. Nevertheless, some estimates are interesting.
\medskip
In order of magnitude, almost any model will predict a mean metallicity of the lost gas that exceeds the mean metallicity of stars, since the gas has the composition of the last and most metal-rich stars formed; i.e., $Z_{\rm i} \gtrsim Z_{\rm s}$. The mass of gas lost is therefore $M_{\rm gi} = M_{\rm Zi} / Z_{\rm i} \lesssim M_{\rm Zi} / Z_{\rm s}$. With $Z_{\rm s} \lesssim y \sim \rm Z_{\odot}$ for the mean of a typical cluster of galaxies, we therefore expect that $M_{\rm gi} \sim M_{\rm Zi} / \rm Z_{\odot}$, very roughly. This implies that the elliptical galaxies in the Coma cluster have ejected ${\sim} 1 \: \rm M_{\odot}$ of gas per solar luminosity of the cluster, which is at the lower end of the range of estimates of the cluster gas content, from X-ray data (Section~\ref{subsec:Data2}). Various specific calculations using the models discussed in Sections~\ref{subsubsec:Supernova-driven winds}~and~\ref{subsubsec:Bursts of star formation in merging subsystems} lead to similar results within a factor ${\sim} 3$. It is therefore uncertain whether gas loss from ellipticals could be the entire source of gas in a cluster like Coma, or whether some of the IGM is simply primordial material that was never in a galaxy. (In the latter case, the intergalactic metals could nevertheless have been supplied by galaxies).
An interesting comment on the origin of the cluster IGM has been made by \citet{Ostriker1977a}: the distribution of morphological types of galaxies in clusters like Coma differs from the field distribution in having a much smaller fraction of spirals, many more S0s, and somewhat more ellipticals \citep{Oemler1974TheClusters}. If one ``corrects'' the cluster galaxies by adding disk matter until the overall ratio (disks)/(elliptical galaxies + bulges) equals that of the field, the mass of extra disk matter required is ${\sim} 50\%$ of the (ordinary) mass of galaxies in the cluster. This in turn is a significant fraction of the mass of cluster IGM. \citet{Ostriker1977a} therefore proposes that some of the IGM is material that would have been made into disks in a less dense environment, but instead was swept up into a hot ambient IGM. This idea ties in with several scenarios for the formation of disk galaxies, in which the disk is made from diffuse gas that is accreted after the formation of a spheroidal component. For example, in the model of \citet{Ostriker1975GalacticHalos}, a significant fraction of the disk is shed by halo stars; and in the picture of \citet{Tinsley1979StellarGalaxies}, the disk forms from diffuse gas after denser pieces have merged to make the bulge.
\subsubsection{Ejection from evolving stars in elliptical galaxies}
\label{subsubsec:Ejection from evolving stars in elliptical galaxies}
The stellar population in elliptical galaxies (and S0 galaxies and the bulge--halo components of spirals) is predominantly very old, and the light of these galaxies is dominated by red giants. It is almost certain that such stars lose a few tenths of a solar mass, between the MS turnoff and the end of their lives, to die as white dwarfs of typically $0.7 \: \rm M_{\odot}$ (Section~\ref{subsubsec:Solar mass}). This mass has been included in the total mass loss considered above, but it is interesting to calculate also the present \emph{rate} of mass loss by stars in elliptical galaxies.
\medskip
For an analytical estimate, let us assume that all the stars in the system formed at the same time, $t = 0$. Let $M_{0}$ be the mass of stars formed, and let $\phi(m)$ be the IMF; the mass of stars formed in the mass interval $(m, \: m + dm)$ is therefore
\begin{equation}
n(m) \ dm = M_{0} \phi(m) \ dm,
\label{eq:eq5.14}
\end{equation}
by Equation~(\ref{eq:eq2.1}). Now imagine these stars peeling off the MS and dying soon afterward as they reach their lifetimes $\tau_{\rm m}$. The number of stars dying per unit time is clearly
\begin{equation}
D(t) = n \left( m_{\rm t} \right) \left| \frac{dm}{d \tau_{\rm m}} \right|_{\tau_{\rm m} = t},
\label{eq:eq5.15}
\end{equation}
where $m_{\rm t}$ is the turnoff mass ($\tau_{\rm m} = t$). The stellar mass--lifetime relation can be approximated by a power law,
\begin{equation}
\frac{m}{m_{1}} = \left( \frac{\tau_{\rm m}}{\tau_{1}} \right)^{-\theta},
\label{eq:eq5.16}
\end{equation}
where $\tau_{1}$ is the lifetime of a fiducial mass $m_{1}$ and $\theta \simeq 0.25$ in the mass range of interest ($m_{\rm t} \sim 1 \: \rm M_{\odot}$). It is convenient to use a power-law IMF, Equation~(\ref{eq:eq2.3}), normalized to $\phi(m_{1}) \equiv \phi_{1}$; masses in only the small range ${\sim} 0.5 - 1 \: \rm M_{\odot}$ are relevant to the following calculation, so this IMF may be a reasonable approximation even if a single power law would not apply to all masses. The ejection rate can be obtained by multiplying $D(t)$ by $(m_{\rm t} - w_{\rm m})$, the mass lost per star with remnant mass $w_{\rm m}$, with the result
\begin{equation}
E(t) = M_{0} \phi_{1} \theta \frac{m_{1}}{\tau_{1}} \left( m_{\rm t} - w_{\rm m} \right) \left( \frac{t}{\tau_{1}} \right)^{-1 + \theta x}.
\label{eq:eq5.17}
\end{equation}
Since $(m_{\rm t} - w_{\rm m})$ changes slowly with time and $\theta x$ is probably only a few tenths, this expression shows that the ejection rate $E(t)$ varies approximately as $t^{-1}$. In Section~\ref{subsubsec:The stellar mass loss rate relative to luminosity}, an analytical expression is derived for the luminosity of stars in this model, and it is shown that Equation~(\ref{eq:eq5.17}) leads to a ratio of ejection rate to integrated blue luminosity of the population,
\begin{equation}
\frac{E}{L_{\rm B}} \sim 0.02 \: \rm M_{\odot} \ L_{B \odot}^{-1} \ Gyr^{-1}
\label{eq:eq5.18}
\end{equation}
at a present time ${\sim} 10^{10} \: \rm yr$.
\medskip
Most elliptical galaxies have a mass of neutral hydrogen that is less than 0.1 times their luminosity, in solar units, and many better studied ellipticals have less neutral hydrogen than 0.01 times their luminosity \citep{Knapp1978HGalaxies}. \citet{Faber1976HWinds} argue that significant amounts of ISM cannot be hiding in elliptical galaxies in ionized or molecular form. Thus the ejection rate given by Equation~(\ref{eq:eq5.18}) would provide more than the observed amount of gas in a few Gyr, or even in less than $1 \: \rm Gyr$. Possible fates for this gas have been thoroughly discussed by \citet{Faber1976HWinds}; they note that star formation at the rate in Equation~(\ref{eq:eq5.18}) would be detectable (unless only low-mass stars form), so they conclude that the gas ejected from stars is being continually lost from the galaxies. On the other hand, \citet{Oemler1979TypeStars} argue that star formation at the rate required to use up this gas could have escaped detection in most ellipticals, and could account for their supernova rate.
\subsection{Abundance Gradients in Disks}
\label{subsec:Abundance Gradients in Disks}
Abundance gradients in disk stars and gas cannot be fully accounted for by gradients in the gas fraction (Section~\ref{subsec:Data2}), so it is of interest to see whether dynamical processes analogous to those discussed in Section~\ref{subsec:Abundance Gradients in Spheroidal Systems}, for spheroidal systems, could be responsible. A gradient in the IMF could again be invoked, but this mechanism will not be discussed further.
\begin{figure*}
\includegraphics[width=0.97\textwidth]{Figures/Fig12.png}
\caption{(a) Star formation rates at several radii in the equatorial plane of a collapse model for the formation of a disk galaxy \citep{Larson1976a}. The radius in pc is marked on each curve, and the three ticks indicate the times at which star formation is $10\%$, $50\%$, and $90\%$ complete (relative to the final mass of stars) at that radius. (b) Metal abundances in the gas (relative to the yield) in the equatorial plane of the same model \citep{Tinsley1978ChemicalDisks}. In this Figure, the radii are given in kpc, ticks have the same meaning as before, and open circles denote the time of maximum gas density at each radius.}
\label{fig:fig12}
\end{figure*}
\subsubsection{Effects of infall}
\label{subsubsec:Effects of infall}
The idea that disks of galaxies form by accretion, incorporating metals from the young halo, has been suggested by dynamical models and supported by the metallicities of stars in the Solar neighborhood (Section~\ref{subsec:Effects of Galaxy Formation}). The properties of the accretion process that affect chemical evolution are the metallicity of infalling gas ($Z_{\rm f}$) and the ratio of SFR to infall rate ($\psi / f$). If there is a radial gradient in these quantities, there must be a corresponding metallicity gradient in the disk. In particular, the metallicity of the gas at any time tends to a value given by Equation~(\ref{eq:eq4.6}), $Z \rightarrow ky + Z_{\rm f}$, where $k = (1 - R) \psi / f$. Thus the disk gas has about this metallicity at the present time, in regions where infall is at all effective, i.e., where $k$ is not so large that $Z$ takes too long to approach its asymptotic value; the stars in turn reflect the metallicity of the gas at the time when they formed.
In the dynamical models studied by \citet{Tinsley1978ChemicalDisks}, the value of $\psi / f$ decreases outward in the disk at all times, because star formation is less efficient at low densities; $Z_{\rm f}$ is negligible at late times at all radii, but it has a significant negative gradient at early stages when metals from the young halo (central bulge) were most important near the center. Figure~\ref{fig:fig12}~(a) illustrates the SFR versus time at several radii in the equatorial plane of one of these models, and Figure~\ref{fig:fig12}~(b) gives the corresponding metallicities in the gas. At small radii, most stars form early from metal-rich infalling gas, while the outer regions experience star formation on a much longer timescale and are still relatively metal-poor. The gas at the present time thus has a negative metallicity gradient due mainly to the gradient in $\psi / f$, while the gradient in stellar abundances is due partly to the early gradient in $Z_{\rm f}$. The sizes of the model gradients are comparable to observed values, so this very schematic model may possibly be showing some effects that occur in real disk galaxies.
Generalizing these results, we can conclude as in Section~\ref{subsec:Effects of Galaxy Formation} that the chemical properties of disks could plausibly be strongly affected by gas flows that constitute the formation of the disk itself.
\subsubsection{Effects of radial gas flows}
\label{subsubsec:Effects of radial gas flows}
Radial inflow of gas in disks possibly occurs as a result of transfer of angular momentum by viscosity, loss of angular momentum from the gas to spiral or bar-like density waves, and other mechanisms \citep[e.g.][]{Kalnajs1978AObservations, Ostriker1977b}. These processes are rather speculative, since the inflow in many models could be a numerical artifact, but it is interesting to see how chemical evolution could be affected by flow velocities of a plausible magnitude.
The metals are concentrated inward by this process, as by gaseous dissipation in a collapsing spheroidal system (Section~\ref{subsubsec:Dissipative collapse}). Let us consider an annulus of a galaxy between radii $r$ and $r + \delta r$, measured in the disk. The chemical evolution of this annulus can be studied in the instantaneous recycling approximation, using Equations~(\ref{eq:eq3.17})~and~(\ref{eq:eq3.19}). Let $M_{\rm g}$ and $\psi$ in those equations be replaced by $2 \pi r M_{\rm g} \delta r$ and $2 \pi r \psi \delta r$, respectively, where $M_{\rm g}$ and $\psi$ now denote the corresponding surface densities; let $f$ be replaced by the net rate of inflow into the annulus, i.e., $F(r) - F(r + \delta r) = -(\partial F / \partial r) \delta r$, where $F$ is a flow rate (in $\rm M_{\odot} \ yr^{-1}$) with a positive sign for outward motion; and let $Z_{\rm f} f$ be replaced by the net rate of inflow of metals, which is $Z(r) F(r) - Z(r + \delta r) F(r + \delta r) = -Z (\partial F / \partial r) \delta r - (\partial Z / \partial r) F \delta r$. The equations then reduce to
\begin{equation}
\frac{\partial M_{\rm g}}{\partial t} = -(1 - R) \psi - \frac{1}{2 \pi r} \frac{\partial F}{\partial r},
\label{eq:eq5.19}
\end{equation}
and
\begin{equation}
M_{\rm g} \frac{\partial Z}{\partial t} = y(1 - R) \psi - \frac{1}{2 \pi r} \frac{\partial Z}{\partial r} F.
\label{eq:eq5.20}
\end{equation}
Equation~(\ref{eq:eq5.20}) shows that the radial flow is consistent with a steady-state abundance gradient,
\begin{equation}
\frac{\partial (Z / y)}{\partial r} \sim 2 \pi r \frac{(1 - R) \psi}{F},
\label{eq:eq5.21}
\end{equation}
which is negative if the flow is inward. The flow causes $Z$ to change on a timescale
\begin{equation}
\tau_{\rm F} \sim \frac{2 \pi r^{2} M_{\rm g}}{\lvert F \rvert}.
\label{eq:eq5.22}
\end{equation}
$F$ can be expressed in terms of the flow velocity $v$, where $\lvert F \rvert = $ (mass of gas in the annulus)/(time for gas to flow across $\delta r$) $= 2 \pi r M_{\rm g} \lvert v \rvert$. The timescale for radial flow to be effective is thus
\begin{equation}
\tau_{\rm F} \sim \frac{r}{\lvert v \rvert},
\label{eq:eq5.23}
\end{equation}
and the corresponding gradient can be written
\begin{equation}
\frac{\partial (Z / y)}{\partial \ \ln(r)} \sim \frac{\tau_{\rm F}}{\tau_{*}},
\label{eq:eq5.24}
\end{equation}
where $\tau_{*} \equiv M_{\rm g} / (1 - R) \psi$ is the timescale for star formation to use up the gas. These relations show that rapid inflow, with a timescale less than that for star formation, quickly obliterates any radial metallicity gradient, while slow inflow can lead to a significant one.
Substituting values of $\psi$, $M_{\rm g}$, and $r$ for the Solar neighborhood, it is found that the interstellar abundance gradient (Section~\ref{subsec:Data2}) is consistent with inflow at a few $\rm km \ s^{-1}$, carrying a flux ${\sim} 1 \: \rm M_{\odot} \ yr^{-1}$; the timescale for the gradient to change is a few Gyr. There is no strong evidence for the occurrence of systematic gas flows of this magnitude in the Galaxy, but nor can they be ruled out.
\citet{Sanders1977ChemicalGalaxy} has suggested that the deep minimum in the surface density of gas in the Galaxy, in an annulus between $0.6$ and $4 \: \rm kpc$, could be due to inflow into the central $600 \: \rm pc$, where the total quantity of ISM is enough to fill the depleted annulus. If so, inflow could perhaps be fueling the strong star formation at the Galactic center.
\section{Approaches to Photometric Evolution}
\label{sec:Approaches to Photometric Evolution}
Evolution of stars in galaxies affects not only their chemical compositions, but also their integrated luminosities, colors, and spectra. Photometric and chemical evolution can be studied separately, because they depend largely on complementary properties of a galaxy: the colors at a given time are governed strongly by the current rate of star formation relative to its past average value, whereas the chemical composition depends mainly on the integrated past star formation relative to the gas content and on the ratio of the SFR to gas flow rates. In studying photometric evolution, we can ignore the effects of ISM, except in correcting colors for any reddening or gaseous emission, and we can avoid assumptions relating the SFR to the gas supply. Of course, a complete understanding of the properties of a galaxy would include the relations among its history of star formation, gas content and gas flows, and chemical composition (cf. Figure~\ref{fig:fig1}), but more can be learned by tackling pieces of the puzzle separately first.
\subsection{Aims and Methods}
\label{subsec:Aims and Methods}
Models for the stellar population of a galaxy address three related questions: what types of stars are present, what history of star formation produced them, and what was the population and its photometric properties in the past? The answers to these questions have many applications, such as interpreting in terms of star formation the correlations between photometric and morphological properties of galaxies, and predicting changes on cosmological timescales.
Methods of constructing population models can be divided into three categories: ``population synthesis'' with no explicit evolution, evolutionary models, and analytical approximations.
\subsubsection{Population synthesis}
\label{subsubsec:Population synthesis}
This approach is to find the ``best'' mixture of stars to match the colors of the galaxy under study. The inferred distribution of stars in the HR diagram then contains information on their past formation rate and the IMF, and it is often possible to judge the mean chemical composition and even to detect minor components of high or low metallicity. The procedure is to observe the colors of the galaxy and a variety of nearby stars, generally including narrow-band photoelectric colors and indices giving the strengths of spectral features that are sensitive to stellar temperature, luminosity, or composition. Then synthetic colors are computed for various mixtures of stars and compared with the galaxy colors. The search for an optimal mixture can be made in many ways, ranging from trial-and-error to elaborate computer algorithms; the method generally used, quadratic programming, was introduced to the field by \citet{Faber1972QuadraticSynthesis.}. Because of observational errors, and because the available nearby stars do not include all types in the galaxy under study, a perfect fit is seldom found, and the solution that (formally) minimizes the errors is not necessarily the most plausible. The choice of a ``best'' synthetic population must therefore be based on imposed astrophysical constraints; these include such requirements as a smooth MS luminosity function, and a distribution of subgiants and giants that could plausibly arise from evolution off the MS. The lack of an objectively defined best fit means that the final solution depends strongly on the imposed constraints, as emphasized by \citet{Williams1976PopulationGalaxies}.
There are often several astrophysically acceptable synthetic populations that match the galaxy colors equally well but correspond to significantly different histories of star formation. An example of such ambiguity appears in models for elliptical galaxies, reviewed by \citet{Faber1977ThePopulations}. All studies agree that \emph{most} of the light of these galaxies, from blue through infrared wavelengths, comes from an old stellar population with a distribution in the HR diagram like an old open cluster plus an extended giant branch (cf. Figure~\ref{fig:fig5}). However, such models almost always fail by a few percent to account for the light around $3500~\si{\angstrom}$ (e.g. $U - B$ is predicted to be too red by ${\sim} 0.1 \: \rm mag$), so they are lacking some hot stellar component that is present in the real galaxies. To date it has been impossible to determine whether the hot stars are a few upper-main-sequence stars, implying ongoing star formation, or a minor population of old objects such as horizontal-branch stars or blue stragglers (which can be seen in the color--magnitude diagram for the old cluster M67, Figure~\ref{fig:fig5}). Obviously, it would be very interesting to know if typical elliptical galaxies really are still making stars at a slow residual rate! For the central bulge of M31, which is optically indistinguishable from an elliptical galaxy, there are broadband colors down to $1550~\si{\angstrom}$, but even these data have not resolved the ambiguity \citep{Wu1980TheM81}. A new avenue has been opened by a demonstration that the integrated light of stars in a nuclear bulge region of our own galaxy matches exactly the integrated light of comparable regions of spirals and ellipticals \citep{Whitford1978SpectralGalaxies}. The brighter stars are individually observable in the Galactic bulge, so a star-by-star synthesis of their contribution to the light is possible. Perhaps in the end this approach will tell whether an old population alone can account for the ultraviolet light.
\medskip
Another general problem is that even the best defined regions in the HR diagram cannot be interpreted uniquely in terms of a past history of star formation. The models are insensitive to many details of the IMF and SFR, for two basic reasons:
\begin{enumerate}
\item the integrated light of galaxies is dominated by regions of the HR diagram that depend theoretically on rather few parameters of star formation; and
\vspace{1mm}
\item some types of stars, such as red giants, may have evolved from a wide range of MS masses (Section~\ref{subsubsec:1 - 4 Msun}), so they cannot be traced uniquely back to an initial mass and time of star formation.
\end{enumerate}
The second of these problems is avoided in the evolutionary method described next, but the first remains and will be discussed below.
\subsubsection{Evolutionary models}
\label{subsubsec:Evolutionary models}
This approach relies primarily on stellar evolution theory to suggest allowable populations, as follows. Theoretical tracks (or isochrones) of stars in the HR diagram are used to compute the stellar population that would arise, at a given age, from a given SFR and IMF, with a given chemical composition; the integrated colors are then calculated, using observed colors of stars in appropriate parts of the HR diagram, and the results are compared with the colors of the galaxy under study. The aim is to derive from a series of models the SFR, IMF, age, and composition(s) that best match the galaxy, and thereby to learn not only about its present stellar population but also about its past history and past photometric properties.
In practice, stellar evolution is not well enough understood for fully theoretical models to be reliable. The main problems are related to late stages of evolution, including particularly the giant branches in old stellar populations, whose effects on models for elliptical galaxies are reviewed by \citet{Faber1977ThePopulations}. These problems are alleviated by using statistical studies of nearby giants to provide semi-empirical evolutionary tracks (Section~\ref{subsubsec:Solar mass}), and by allowing the most uncertain types of stars to be present in numbers that are treated as adjustable parameters. This method thus closely resembles some non-evolutionary population syntheses in which the constraints are chosen to represent stellar evolutionary tracks \citep{OConnell1976GalaxyEllipticals}.
The evolutionary approach has several advantages. The best established aspects of stellar evolution theory are incorporated, so the resulting population is a \emph{possible} one as far as can be determined. Uncertainties cannot be formally calculated, but from trials with a variety of assumptions one can estimate subjectively the allowable range of parameters. Often this range is small enough to lead to useful conclusions about the past history of star formation, and to predictions of photometric changes of cosmological interest. For example, it is possible to determine the slope of the IMF in elliptical galaxies closely enough to be sure that their integrated luminosity declines with increasing age (Section~\ref{subsec:Evolution of a Single Generation of Stars}).
Uncertainties in the conclusions from this method arise partly from uncertainties in stellar evolution, and partly from the intrinsic insensitivity of integrated colors to many parameters of interest -- a problem found earlier with population syntheses. Two parts of the HR diagram tend to dominate the integrated light, as illustrated spectroscopically by the work of, e.g., \citet{Morgan1957AGalaxies} and \citet{Morgan1969OnGalaxies}: B stars on the upper main sequence, and late G through early M giants. These dominant regions are extended out to O stars in ultraviolet light and to late M giants in the infrared. If young stars are absent, low-mass giants dominate at visual and longer wavelengths, so the colors depend much more on stellar evolution than on the IMF or past SFR; at shorter wavelengths, however, turnoff stars are seen so the colors give some information on the age of the system. If young stars are present, the light at short wavelengths is dominated by OB stars, whose relative numbers depend on the IMF and whose total numbers (relative to red stars) depend on the ratio of the present SFR to its integrated past value. Stars with lifetimes from a few times $10^{8} \: \rm yr$ to just below the age of the system (usually A and F stars) contribute relatively little light, so there is little information on either their part of the IMF or the detailed time-dependence of the SFR. In Section~\ref{subsec:UBV Colors of Normal Galaxies}, models will be discussed that illustrate the dominance of the upper main sequence and/or low-mass giants, depending on the SFR.
\medskip
Programs for constructing evolutionary models have been described by \citet{Tinsley1968EvolutionGalaxies, Tinsley1972a, Tinsley1978b}, \citet{Searle1973TheGalaxies}, \citet{TinsleyGunn1976a}, and \citet{Larson1978StarGalaxies}. The mechanical details are far less troublesome than the input ``data'' representing stellar tracks, and it is easy to obtain numerical accuracy far exceeding the astrophysical certainty of the calculations. There are two types of this technique.
\begin{enumerate}
\item The first method is to supply the computer with evolutionary tracks in the HR diagram for stars with a series of discrete masses, or with isochrones for a series of discrete ages; separate stellar data are used for each chemical composition of interest. Then, for a given IMF and SFR, the calculation yields the numbers of stars on a large grid of points in the HR diagram, as a function of the age of the system.
\vspace{1mm}
\item The second method uses the first type of program once only for each IMF and composition, to give the integrated colors at a series of ages of a model whose SFR consists of a single initial burst. These are then regarded as the colors of ``generations'' of stars with a given age (and IMF and composition).
\end{enumerate}
A model with any prescribed SFR can then be treated, at each age, as the sum of such generations in proportions given by the SFR. The number of generations whose properties must be combined to obtain the integrated colors of any model is much smaller than the number of points in the HR diagram that are referred to directly in the first method, so the second approach is more economical. In either method, it is clearly possible to add arbitrary numbers of stars of undetermined evolutionary status, in the spirit of population synthesis.
While models with Solar metallicity can rely on nearby stars to provide colors and semi-empirical evolutionary tracks, there is no such convenient sample for other compositions. In making models for non-Solar metallicities, it is often most convenient to change the ``standard'' models differentially, rather than starting from scratch with tracks and colors for each set of stars. \citet{Faber1973VariationsGalaxies} first used the metallicity effects discussed in Section~\ref{subsubsec:Initial composition} to estimate differential changes in the integrated colors of elliptical galaxies, as a function of metallicity, and her methods have been adapted by others subsequently. Recent results for elliptical galaxies have been cited in Section~\ref{subsec:Data2}. The calculations of metallicity effects in integrated light are still much less secure than one would like, and there is a need for more basic work on stellar evolution and atmospheres at non-Solar compositions, including non-Solar abundance ratios among elements heavier than helium \citep{Faber1977ThePopulations}.
\subsubsection{Analytical approximations}
\label{subsubsec:Analytical approximations}
Some of the results from evolutionary models can be understood qualitatively using analytical approximations. These have proved particularly tractable for models in which all the stars form in a single initial burst, which is a first approximation to the population in elliptical galaxies. Such models will be considered next.
\subsection{Evolution of a Single Generation of Stars}
\label{subsec:Evolution of a Single Generation of Stars}
Many numerical models designed to match detailed photometry of elliptical galaxies have shown that nearly all the light at visual and longer wavelengths can be accounted for by a very old population, with a turnoff near the Sun's position on the MS. The metallicities of the dominant stars appear to be within a factor of two of Solar in wide-aperture observations of giant ellipticals, although their centers may be more metal-rich and small ellipticals are metal-poor (Section~\ref{subsec:Data2}). Reviews by \citet{vandenBergh1975StellarGalaxies} and \citet{Faber1977ThePopulations} cover the history and recent status of this subject, and a few subsequent developments have been referred to in Section~\ref{subsec:Aims and Methods}. The implications of a predominantly very old population for the evolution of elliptical galaxies are best understood using analytical approximations.
\subsubsection{Content and luminosity}
\label{subsubsec:Content and luminosity}
Let us consider a single generation of stars, formed with total mass $M_{0}$ in a short burst (as in Section~\ref{subsubsec:Ejection from evolving stars in elliptical galaxies}), with a fixed chemical composition near Solar. The population evolves by peeling off the MS, as can be visualized from Figures~\ref{fig:fig2}~and~\ref{fig:fig3}.
The IMF will be taken to be a power law, normalized to $\phi(m_{1}) \equiv \phi_{1}$, where $m_{1}$ is the turnoff mass at a fiducial time $\tau_{1}$. The power-law approximation need only hold over a small mass interval, since the light at present comes almost entirely from stars between $0.4 \: \rm M_{\odot}$ and turnoff, and the turnoff mass at ages of interest, ${\sim} 5 - 20 \: \rm Gyr$, lies in the small range ${\sim} 0.9 - 1.2 \: \rm M_{\odot}$.
At a time $t$ after star formation, the MS stars present have masses from the lower limit at formation, $m_{\rm L}$, up to the turnoff mass $m_{\rm t}$, which is given by substituting $\tau_{\rm m} = t$ in Equation~(\ref{eq:eq5.16}). Thus the number of dwarfs with masses in the interval $(m, \ m + dm)$ is, by Equation~(\ref{eq:eq2.3}),
\begin{equation}
\begin{medsize}
n_{\rm d} (m) \ dm = M_{0} \phi(m) \ dm = M_{0} \phi_{1} \left( \frac{m}{m_{1}} \right)^{-(1 + x)} \ dm, \: \: m_{\rm L} \leq m \leq m_{\rm t}.
\label{eq:eq6.1}
\end{medsize}
\end{equation}
Stars slightly more massive than $m_{\rm t}$ are present as giants, and their total number is the number of stars that were on the MS with lifetimes between $t$ and $t - \tau_{\rm g}$, where $\tau_{\rm g}$ is the duration of post-MS evolution for masses ${\sim} m_{\rm t}$. (The term ``giants'' is used loosely here to mean all post-MS stars; the analysis can easily be modified to refer to any portion of post-MS evolution). The number of giants is therefore
\begin{equation}
n_{\rm g} (t) = M_{0} \phi(m_{\rm t}) \left| \frac{dm}{d \tau_{\rm m}} \right|_{\tau_{\rm m} = t}; \: \: \tau_{\rm g} = M_{0} \phi_{1} \theta \frac{m_{1}}{\tau_{1}} \tau_{\rm g} \left( \frac{t}{\tau_{1}} \right)^{-1 + \theta x}.
\label{eq:eq6.2}
\end{equation}
The luminosity of individual dwarfs in the mass range of interest can be approximated by a power law,
\begin{equation}
\ell_{\rm d} (m) = \ell_{1} \left( \frac{m}{m_{1}} \right)^{\alpha},
\label{eq:eq6.3}
\end{equation}
where $\alpha \simeq 5$. For giants, an average luminosity $\ell_{\rm g}$ is defined so that the product $\ell_{\rm g} \tau_{\rm g}$ gives correctly the integrated light output during post-MS evolution. The values of $\ell_{1}$, $\alpha$, and $\ell_{\rm g}$ of course depend on the wavelength interval of interest, and so do the results below relating to luminosities. (For bolometric light, the product $\ell_{\rm g} \tau_{\rm g}$ is proportional to the amount of nuclear energy used, but it has no such interpretation in restricted wavelength bands).
The integrated luminosities and masses of dwarfs and giants can now be derived from Equations~(\ref{eq:eq6.1})~--~(\ref{eq:eq6.3}) and Equation~(\ref{eq:eq5.16}). It will be assumed in the integrals that $m_{\rm L} \ll m_{1}$. The total mass of dwarfs at time $t$ depends critically on whether the slope of the IMF $(x)$ is less than or greater than 1:
\begin{subequations}
\begin{empheq}[left={ M_{\rm d} (t) = \int\limits_{m_{\rm L}}^{m_{\rm t}} m n_{\rm d} (m) \ dm = \empheqlbrace\,}]{align}
& \frac{M_{0} \phi_{1} m_{1}^{2}}{x - 1} \left( \frac{m_{\rm L}}{m_{1}} \right)^{-x + 1}, \: \: \: \, x > 1, \label{eq:eq6.4a} \\
& M_{0} \phi_{1} m_{1}^{2} \ \ln \left( \frac{m_{\rm t}}{m_{\rm L}} \right), \: \: \: \: \: \; x = 1, \label{eq:eq6.4b} \\
& \frac{M_{0} \phi_{1} m_{1}^{2}}{1 - x} \left( \frac{t}{\tau_{1}} \right)^{-\theta (1 - x)}, \: \: x < 1. \label{eq:eq6.4c}
\end{empheq}
\end{subequations}
Giants have a total mass ${\sim} m_{\rm t} n_{\rm g} (t)$, and one can quickly verify that the mass ratio of giants to dwarfs is greatest in the case $x < 1$, and is at most ${\sim} \tau_{\rm g} / t \sim 0.1$; the contribution of giants to the total mass will therefore be neglected. The integrated luminosity of dwarfs is
\begin{equation}
L_{\rm d} (t) = \int_{m_{\rm L}}^{m_{\rm t}} \ell_{\rm d} (m) n_{\rm d} (m) \ dm = \frac{M_{0} \phi_{1} m_{1} \ell_{1}}{\alpha - x} \left( \frac{t}{\tau_{1}} \right)^{-\theta (\alpha - x)},
\label{eq:eq6.5}
\end{equation}
on the assumption $x < \alpha$, which is justified below. Finally, the integrated luminosity of giants is
\begin{equation}
L_{\rm g} (t) = \ell_{\rm g} n_{\rm g} (t) = M_{0} \phi_{1} \theta \frac{m_{1}}{\tau_{1}} \ell_{\rm g} \tau_{\rm g} \left( \frac{t}{\tau_{1}} \right)^{-1 + \theta x}.
\label{eq:eq6.6}
\end{equation}
The above relations will be used to derive some interesting properties of this single generation of stars.
\subsubsection{Remnants of dead stars}
\label{subsubsec:Remnants of dead stars}
There may be a significant dark mass in the form of remnants of stars initially above $m_{\rm t}$, especially if the IMF has a fairly shallow slope so these stars were relatively numerous. Although it is probably a very poor approximation to extrapolate the IMF to high masses with the slope $x$ used near $1 \: \rm M_{\odot}$, the equations will be written to show how the contributions of remnants can be estimated in the simplest cases. (These results can easily be modified to allow for a variable slope). In this approximation, it will be assumed that all remnants have the same mass $w$, and that all stars above $m_{\rm t}$ are dead. Then the total mass of remnants is $w$ times the number of stars formed with masses between $m_{\rm t}$ and the upper limit $m_{\rm U}$:
\begin{equation}
M_{\rm w} (t) = w \int_{m_{\rm t}}^{m_{\rm U}} M_{0} \phi(m) \ dm = \frac{M_{0} \phi_{1} m_{1} w}{x} \left( \frac{t}{\tau_{1}} \right)^{\theta x},
\label{eq:eq6.7}
\end{equation}
assuming $m_{\rm U} \ll m_{\rm t}$ and $x > 0$. The relative mass of remnants is potentially greatest if $x < 1$, and then Equation~(\ref{eq:eq6.4c}) shows that $M_{\rm w} / M_{\rm d} \sim w / m_{\rm t}$, which could be close to unity. This result is obviously strongly dependent on the assumption of a single power law for the whole IMF, which would exaggerate the mass of remnants if, for example, elliptical galaxies have a curved IMF like the function in the Solar neighborhood (Figure~\ref{fig:fig4}). It may be concluded that dead remnants could possibly affect the total mass by a factor ${\sim} 2$, which cannot be predicted with any confidence from constraints on the slope of the IMF at turnoff.
\subsubsection{The ratio of giants to dwarfs in the light}
\label{subsubsec:The ratio of giants to dwarfs in the light}
Some spectral features in the integrated light of elliptical galaxies depend sensitively on the relative amounts of light contributed by giant and dwarf stars at the feature wavelength. Examples are an iron hydride band at $0.99 \: \rm \upmu m$, known as the Wing-Ford band, which \citet{Whitford1977ThePopulations.} has found to be extremely strong in late dwarfs but weak in late giants; and a carbon monoxide band at $2.2 \: \rm \upmu m$, studied especially by \citet[][and earlier papers cited therein]{Frogel1978PhotometricGalaxies}, which has the opposite behavior, being much stronger in late giants than in late dwarfs. Since the light of elliptical galaxies at those wavelengths must be dominated by late-type stars, the galaxies should show a weak FeH band and a strong CO band if giants outshine dwarfs, and vice versa. As the following analysis shows, the relative luminosities of giants and dwarfs give important information on the slope of the IMF, which in turn affects many other properties of elliptical galaxies including the rate of evolution of total luminosity; it is the significance of this effect for cosmological tests (Section~\ref{subsubsec:Evolution of luminosity and the Hubble diagram}) that has motivated much of the analysis of spectral features.
Equations~(\ref{eq:eq6.5})~and~(\ref{eq:eq6.6}) together give an approximate expression for the relative luminosities of giants and dwarfs:
\begin{equation}
G(t) \equiv \frac{L_{\rm g} (t)}{L_{\rm d} (t)} = \theta (\alpha - x) \frac{\ell_{\rm g} \tau_{\rm g}}{\ell_{1} \tau_{1}} \left( \frac{t}{\tau_{1}} \right)^{\theta \alpha - 1}.
\label{eq:eq6.8}
\end{equation}
An alternative expression is obtained by substituting Equations~(\ref{eq:eq5.16})~and~(\ref{eq:eq6.3}):
\begin{equation}
G(t) = \theta (\alpha - x) \frac{\ell_{\rm g} \tau_{\rm g}}{\ell_{\rm d} \left( m_{\rm t} \right) t}.
\label{eq:eq6.9}
\end{equation}
The term $\ell_{\rm g} \tau_{\rm g}$ is the amount of energy radiated (at a given wavelength) by a star of approximately turnoff mass after it leaves the MS, while $\ell_{\rm d} \left( m_{\rm t} \right) t$ is the energy radiated during MS evolution. Thus Equation~(\ref{eq:eq6.9}) says that the value of $G$ in bolometric light is, in order of magnitude, equal to the ratio of nuclear fuel consumed after leaving the MS to that consumed on the MS; since stars near $1 \: \rm M_{\odot}$ burn the hydrogen in only $10\%$ of their mass while on the MS but in $70\%$ before they die (Section~\ref{subsubsec:Solar mass}), this fuel ratio is ${\sim} 6$. This high value is the underlying reason why giants can outshine dwarfs in the integrated light of a galaxy, despite their very short lifetimes. Giants tend to be especially dominant at long wavelengths, because most of the energy from the giant branch as a whole comes from red giants.
The fuel burning ratio is not the only factor affecting $G$, however. The term $(\alpha - x)$ in Equation~(\ref{eq:eq6.9}) introduces a dependence on $x$, the slope of the IMF. A larger value of $x$ reduces the contribution of giants simply by reducing the number of stars in the mass range of giants (just above turnoff) relative to those still on the MS. The dependence of $G$ on $x$ is of great practical importance, since it allows spectroscopic criteria to set constraints on $x$. The work of \citet{Whitford1977ThePopulations.} and \citet{Frogel1978PhotometricGalaxies} shows that the red--infrared light of elliptical galaxies is strongly dominated by giants, to an extent that $x$ must be less than 2, and possibly less than 1. These constraints are consistent with the IMF in the Solar neighborhood, which has $x < 1$ in the relevant mass range (Figure~\ref{fig:fig4} and Equation~\ref{eq:eq2.9}).
The infrared spectra of elliptical galaxies set constraints not only on the IMF but also on the relative numbers of M giants of different spectral types that populate the giant branch. As noted in Section~\ref{subsubsec:Evolutionary models}, these numbers are not firmly predicted by stellar evolution theory, so studies of galaxy spectra can add to an understanding of late stages in the lives of low-mass stars. This application of galaxy models is discussed by \citet{Faber1977ThePopulations}, \citet{Tinsley1978b}, and references therein.
\subsubsection{The stellar mass loss rate relative to luminosity}
\label{subsubsec:The stellar mass loss rate relative to luminosity}
An expression for the rate of mass loss from stars has been derived in Section~\ref{subsubsec:Ejection from evolving stars in elliptical galaxies}, but Equation~(\ref{eq:eq5.17}) is not in a useful form for comparing with observable quantities. It is possible to obtain a useful equation for the ejection rate per unit integrated luminosity, because both quantities scale with the populations of stars near turnoff.
From Equations~(\ref{eq:eq6.5})~and~(\ref{eq:eq6.8}), the total luminosity can be written
\begin{equation}
L(t) = \left[ 1 + G(t) \right] L_{\rm d} (t) = \frac{M_{0} \phi_{1} m_{1} \ell_{1}}{\alpha - x} (1 + G) \left( \frac{t}{\tau_{1}} \right)^{-\theta (\alpha - x)}.
\label{eq:eq6.10}
\end{equation}
Then, with Equation~(\ref{eq:eq5.17}), the ratio of ejection rate to luminosity is
\begin{equation}
\frac{E(t)}{L(t)} = \frac{\theta (\alpha - x)}{\ell_{1} \tau_{1}} \frac{m_{\rm t} - w_{\rm m}}{1 + G} \left( \frac{t}{\tau_{1}} \right)^{\theta \alpha - 1},
\label{eq:eq6.11}
\end{equation}
which shows that the ratio depends only slowly on time. A more useful relation for finding the present ratio is given by substituting Equation~(\ref{eq:eq5.16})~and~(\ref{eq:eq6.3}) to eliminate $\ell_{1}$ and $\tau_{1}$, with the result
\begin{equation}
\frac{E(t)}{L(t)} = \theta (\alpha - x) \frac{m_{\rm t} - w_{\rm m}}{1 + G} \frac{1}{\ell_{\rm d} \left(m_{\rm t} \right) t}.
\label{eq:eq6.12}
\end{equation}
This ratio can be estimated for present-day ellipticals as follows. From spectroscopic studies in \emph{blue} light, $G \simeq 1$ (the value of $G$ is greater in red or bolometric light); and approximate values of the other quantities are $\alpha \simeq 5$, $\theta \simeq 0.25$, $m_{\rm t} \simeq 1 \: \rm M_{\odot}$, $w_{\rm m} \simeq 0.7 \: \rm M_{\odot}$, $\ell_{\rm d} \simeq 1 \: \rm L_{\rm B \odot}$, $t \simeq 10 \: \rm Gyr$, $x \simeq 1$. The result from Equation~(\ref{eq:eq6.12}) is then $E / L_{\rm B} \simeq 0.015 \: \rm M_{\odot} \ L_{\rm B \odot} Gyr^{-1}$, of which the significance was discussed in Section~\ref{subsubsec:Ejection from evolving stars in elliptical galaxies}.
\subsubsection{The mass-to-luminosity ratio}
\label{subsubsec:The mass-to-luminosity ratio}
An analytical estimate of $M_{\rm s} / L$ can be made using the mass of stars $M_{\rm s} \simeq M_{\rm d} (t)$ (neglecting the small contribution of giants and the very uncertain contribution of dead remnants), and the total luminosity $L(t)$. From Equations~(\ref{eq:eq6.4a})~--~(\ref{eq:eq6.4c}), it is clear that the result depends strongly on whether $x \lessgtr 1$. Moreover, it depends critically on the assumption that $x$ is constant down to $m_{\rm L}$, since the least massive stars (or sub-stellar objects) can be numerous enough to dominate the mass while contributing negligibly to the light. If $x < 1$, the result from Equations~(\ref{eq:eq6.4c})~and~(\ref{eq:eq6.10}) is
\begin{equation}
\frac{M_{\rm s}}{L} = \frac{\alpha - x}{1 - x} \frac{1}{1 + G} \frac{m_{\rm t}}{\ell_{\rm d} \left( m_{\rm t} \right)}, \: \: x < 1,
\label{eq:eq6.13}
\end{equation}
which is proportional to the mass-to-luminosity ratio of turnoff stars. If $x > 1$ (or if $x$ increases from a value below 1 at turnoff to above 1 at smaller masses) Equation~(\ref{eq:eq6.4a}) shows that $M_{\rm s} / L$ increases in proportion to $m_{\rm L}^{-(x - 1)}$, so it is sensitive to a quantity that cannot be determined photometrically.
In all cases, photometric data (star counts, population syntheses, spectroscopic estimates of $x$) yield only a \emph{lower limit} to the true mass-to-luminosity ratio $(M / L)$ of a galaxy, since any amount of mass could be present in hidden form. When the masses of galaxies are determined dynamically, the empirical $M / L$ values often increase to such large values in the outer regions that a large amount of hidden mass must indeed be present \citep[e.g.][]{Spinrad1978HalosRatios}. For this reason, values of $M / L$ determined from population syntheses and equivalent methods are sometimes called ``photometric $M / L$ ratios'' to distinguish them from the ratios of actual mass (defined dynamically) to luminosity.
\subsubsection{Evolution of luminosity and the Hubble diagram}
\label{subsubsec:Evolution of luminosity and the Hubble diagram}
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig13.png}
\caption{Schematic Hubble diagram showing how both the deceleration parameter ($q_{0}$) and evolution of galaxies affect the departure from linearity. \emph{Lines} are schematic ``theoretical'' curves for two values of $q_{0}$, \emph{dots} are hypothetical data points, and \emph{arrows} indicate qualitatively how they should be corrected if the net effect of evolution is to make distant galaxies intrinsically brighter than nearby ones. More precisely, evolution (in this sense) at the rate of a few percent of a galaxy's luminosity per Gyr makes the true value of $q_{0}$ smaller by about unity than the value inferred from the uncorrected data points.}
\label{fig:fig13}
\end{figure}
In one of the classic cosmological tests, the Hubble diagram, logarithmic redshifts of galaxies are plotted against against their apparent magnitudes, as illustrated schematically in Figure~\ref{fig:fig13}. For a sample with a well-defined mean absolute magnitude, this diagram can be regarded heuristically as a plot of ``recession velocity'' versus ``distance''. At small redshifts, the regression line is linear with a slope corresponding to Hubble's Law, $\rm redshift \propto distance$. At large redshifts, the deviation from linearity measures the change in the ratio ``velocity'' / ``distance'' with distance itself; since the lookback time (the light-travel time) increases with distance, the curvature of the Hubble diagram thus gives a measure of the past expansion rate of the Universe, and in particular of its deceleration. The deceleration parameter $q_{0}$ can take only positive values in the simplest cosmological models of General Relativity, the Friedmann models, and $1/2$ is a critical value: if $q_{0} > 1/2$, the deceleration is large enough for the expansion eventually to be reversed, but if $0 < q_{0} \leq 1/2$, the Universe will expand forever; if in fact $q_{0}$ is negative, indicating that the expansion is accelerating, more complicated cosmological models are required. Evolution of galaxies enters the picture because the lookback times sampled must be many Gyr for the deceleration to be detectable; the galaxies then had significantly different luminosities, so the ``distance'' parameter, apparent magnitude, cannot be estimated on the assumption of a constant absolute magnitude. The departure of the Hubble diagram from linearity is very sensitive to evolution: if the luminosities of elliptical galaxies grow fainter at a few percent per Gyr, for example, the apparent value of $q_{0}$ (inferred from the shape of the Hubble diagram) exceeds its true value by several tenths. This problem has been discussed by \citet{Humason1956RedshiftsNebulae.}, \citet{Sandage1961a, Sandage1961b}, \citet{Gunn1975SpectrophotometryCosmology}, and \citet{Tinsley1972b, Tinsley1977c}. For an approximate estimate of the evolutionary correction to $q_{0}$, the above analytical equations can be used.
From Equation~(\ref{eq:eq6.10}), we have
\begin{equation}
\frac{d(\ln \ L)}{d(\ln \ t)} = -\theta (\alpha - x) + \frac{t}{1 + G} \frac{dG}{dt},
\label{eq:eq6.14}
\end{equation}
and Equation~(\ref{eq:eq6.8}) can be used to evaluate $dG / dt$. The term $\ell_{\rm g} \tau_{\rm g}$ in the expression for $G(t)$ depends only slowly on time, because giant branch evolution depends only weakly on mass in the relevant range, so only the explicit time-dependence need be considered and that term gives $(t / G) (dG / dt) = \theta \alpha - 1$. Substituting in Equation~(\ref{eq:eq6.14}), we have
\begin{equation}
\frac{d(\ln \ L)}{d(\ln \ t)} = -\theta (\alpha - x) + \frac{G}{1 + G} (\theta \alpha - 1).
\label{eq:eq6.15}
\end{equation}
The second term in Equation~(\ref{eq:eq6.15}) is not very important, since $(\theta \alpha - 1)$ is a few tenths and $G / (1 + G)$ lies between 0 and 1. The main term is therefore simply $(-\theta \alpha + \theta x)$, which can be written
\begin{equation}
\frac{d(\ln \ L)}{d(\ln \ t)} \simeq -1.3 + 0.3 x.
\label{eq:eq6.16}
\end{equation}
Essentially the same result is obtained for the evolution of luminosity in numerical population models. The examples in Figure~\ref{fig:fig14} show the predicted dependence on the IMF: the rate at which $M_{V}$ gets dimmer is slower in models with a larger value of $x$. Since giants supply most of the light, this behavior is mainly because, when $x$ is large, the giant branch is fed by a more richly populated main sequence as time goes on.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig14.png}
\caption{Evolution of colors and magnitudes of single-generation models for the stellar population in elliptical galaxies \citep{TinsleyGunn1976a}. Curves are for three values of the slope of the IMF: \emph{solid lines}, $x = 2$; \emph{dashes}, $x = 1$; \emph{dots}, $x = 0$. Note that if $x$ is small, colors evolve slowly but magnitudes evolve quickly.}
\label{fig:fig14}
\end{figure}
\medskip
In the Hubble diagram, evolution means that departures from linearity are due not only to $q_{0}$ but also to systematic changes in the absolute magnitudes of galaxies (Figure~\ref{fig:fig13}). If the curvature is interpreted without regard to evolution, the result is an apparent value of $q_{0}$ that differs from the true value by
\begin{equation}
\Delta q_{0} \equiv {\rm apparent \ value - true \ value} \simeq -1.5 \frac{d(\ln \ L)}{d(\ln \ t)}
\label{eq:eq6.17}
\end{equation}
\citep[e.g.][]{Tinsley1977c}. A first-order estimate, from Equation~(\ref{eq:eq6.16}), is therefore
\begin{equation}
\Delta q_{0} \simeq 2.0 - 0.4 x.
\label{eq:eq6.18}
\end{equation}
The slope of the IMF, $x$, emerges as the critical parameter. As discussed in Section~\ref{subsubsec:The ratio of giants to dwarfs in the light}, spectroscopic studies indicate that $x < 2$, and possibly $x < 1$. In the first case, $|\Delta q_{0}| \gtrsim 1$, and in the second case, $|\Delta q_{0}| \gtrsim 1.5$. In either case, the correction for evolution is big enough to make a qualitative difference to the type of cosmology inferred.
Current estimates of the apparent value of $q_{0}$ range from ${\sim} 0$ \citep{Gunn1975SpectrophotometryCosmology} to ${\sim} 1.5$ \citep{Kristian1978TheLine}, the differences being due to unknown sampling and observational effects. Downward corrections of order unity are clearly important in determining whether the true value of $q_{0}$ is greater than $1/2$, less than $1/2$, or even negative.
A value of $x \gtrsim 5$ would be needed to make stellar evolution negligible in the Hubble diagram, but such a steep IMF would make the late dwarfs dominate the infrared light to an extent that is precluded by the giant-dominated spectra of elliptical galaxies. A possible loophole is the following. Equation~(\ref{eq:eq6.16}) depends on the value of $x$ for stars near turnoff, while the infrared spectra depend on the ratio of giants to dwarfs of types K and M; if the IMF were to turn over sharply between ${\sim} 0.5$ and $1 \: \rm M_{\odot}$ (i.e., having far fewer less massive stars that the turnoff slope would predict), one could have both a steep slope at turnoff and a very small contribution from dwarfs to the infrared light. This idea is, of course, completely ad hoc, since the IMF in the Solar neighborhood has $x < 2$ for all masses $< 10 \: \rm M_{\odot}$, and does not cut off above ${\sim} 0.2 \: \rm M_{\odot}$. It is therefore most reasonable to conclude that elliptical galaxies have giant-dominated spectra because the IMF has a fairly shallow slope at turnoff; if so, their luminosity evolves fast enough to make the apparent value of $q_{0}$ exceed its true value by 1 or more.
However, this is not all we need to know to unravel the Hubble diagram. The galaxies used for this test are the central cluster giants, which are believed to grow secularly by cannibalizing their neighbors (Section~\ref{subsubsec:Mergers of stellar systems}). This process could plausibly lead to a growth rate in the total stellar population of several percent per Gyr, with a corresponding increase of luminosity in opposition to the effect just discussed. The dynamical effects cannot yet be calculated accurately enough for a correction to be applied to the Hubble diagram, so this test does not yet give a usefully accurate value of $q_{0}$. The situation is reviewed by \citet{Tinsley1977c}.
\subsubsection{Evolution of colors}
\label{subsubsec:Evolution of colors}
Predictions of color evolution are of interest because they can be tested by observations of distant elliptical galaxies whose ages are several Gyr younger than nearby galaxies.
The colors of a single-generation population become redder with age, if the main course of stellar evolution is peeling off the MS at turnoff and following the red giant branch. The main contribution to the color change is the redward evolution of the turnoff, since giant evolution is insensitive to turnoff mass in the range of interest. Consequently, colors evolve faster if the light is less giant-dominated, i.e. if $x$ is larger, in contrast to the integrated luminosities just discussed. This behavior is illustrated in Figure~\ref{fig:fig14}.
Qualitatively different behavior is predicted if the stars can lose enough mass to become blue horizontal-branch stars, instead of the red ``clump'' giants that normally represent the core helium burning stage of metal-rich low-mass stars (Section~\ref{subsubsec:Solar mass}). It has been suggested that such stars lose mass at a variety of rates, some becoming late red giants and others becoming blue. Numerical models for galaxy populations in which mass loss occurs stochastically on the red giant branch have been studied by \citet{Ciardullo1978TheGalaxies}. Because evolution to a blue position in the HR diagram occurs only if the star has a small mass of envelope left, the fraction of giants becoming blue increases as the turnoff mass decreases. The upshot is that the integrated colors of the model galaxies evolve blueward after ${\sim} 8 \: \rm Gyr$.
Observations of distant elliptical galaxies are ambiguous on this point, as reviewed by \citet{Spinrad1977TheRedshifts}. Some of the most distant central cluster galaxies known, with redshifts ${\sim} 0.6$, have intrinsic colors that are bluer than those of nearby ellipticals, but the distant galaxies were selected on the basis of strong radio emission so they may be atypical. If they are typical, the color change is about that expected according to the type of models that evolve monotonically toward redder colors (e.g. Figure~\ref{fig:fig14}); the lookback time sampled is ${\sim} 4 - 7 \: \rm Gyr$, depending on the cosmological model. Another sample of central cluster galaxies with redshifts up to nearly $0.5$ has no systematic dependence of color on redshift that can be disentangled from the intrinsic scatter \citep{Wilkinson1978SpectralGalaxies}.
Dramatic color differences between nearby and distant galaxy populations in clusters have been discovered by \citet{Butcher1978The295, Butcher1978TheClusters}. Nearby clusters, i.e. those with lookback times $< 1 \: \rm Gyr$, have galaxy populations that are strongly correlated with the cluster morphology: loose, irregular clusters have a large fraction of spiral galaxies, and centrally concentrated, regular clusters have very few spirals and mainly S0 and elliptical galaxies; the brighter galaxies in regular clusters are correspondingly all red. However, in two regular clusters with lookback times ${\sim} 5 \: \rm Gyr$, the bright galaxies are found to have a wide range of colors, including many that are as blue as late-type spiral galaxies. On the assumption that the distant regular clusters represent younger versions of the nearby ones, these very blue galaxies must evolve (in a few Gyr) into red S0s or ellipticals. The color change observed is many times greater than any predictions based on the evolution of single-generation populations, so it is concluded that those galaxies were actively forming stars just a few Gyr ago. Presumably, they are mainly the precursors of S0 galaxies seen in nearby clusters, in which star formation is undetectable.
\section{Colors and Star Formation Rates}
\label{sec:Colors and Star Formation Rates}
The stellar populations in most galaxies are far more complicated than those in ellipticals, because young stars are important contributors to the light. The time-dependence of the SFR is therefore an important parameter in addition to the three quantities (age, IMF, and metallicity) used to characterize old populations, and the latter quantities could also be changing in time. Moreover, the colors of spiral and irregular galaxies are often affected by internal reddening and gaseous emission lines. Despite the complications presented by these galaxies, it is especially interesting to try to understand their photometric properties in terms of histories of star formation. Applications of such studies include explaining correlations between form and photometric properties, finding what physical conditions are conducive to star formation, and searching for young galaxies.
Models for galaxies with ongoing star formation are usually numerical; analytical approximations are cumbersome except in the simplest case of a constant SFR \citep{Tinsley1973AnalyticalGalaxies}. This Section will consider models that study only a few simple properties, mainly just UBV colors. Although more can be learned from spectroscopic details, the UBV system has the advantage of an extensive and homogeneous compilation of galaxy colors in the Second Reference Catalogue of Bright Galaxies (\citealp{deVaucouleurs1976SecondGalaxies.}; to be referred to as \citetalias{deVaucouleurs1976SecondGalaxies.}).
\subsection{UBV Colors of Normal Galaxies}
\label{subsec:UBV Colors of Normal Galaxies}
The UBV colors of a sample of morphologically normal galaxies are shown in Figure~\ref{fig:fig15}; the crosses are all elliptical and S0 galaxies, and the dots are a variety of morphological types, which we consider first. The colors of these galaxies form such a narrow distribution in the two-color diagram that it is tempting to look for one dominant parameter that could vary among galaxies and lead to a one-dimensional set of colors. Because the appearance and spectra of galaxies suggest a progression of star-forming activity, ranging from very active in late-type irregulars to negligible in ellipticals, it is natural to suggest that the color sequence is due to different proportions of young and old stars. Population syntheses and evolutionary models have confirmed this view, and their conclusions can be summarized (with some oversimplification) in a ``standard scenario'' for galaxy evolution: normal galaxies have the same IMF and mean stellar metallicities, and they are of the same age, but they differ in the time-dependence of their SFRs; in particular, the latest (bluest) types of galaxies form stars on a long timescale, while the earliest (reddest) ceased star formation long ago. This hypothesis is obviously inaccurate in detail, but it provides a useful starting point. It will be used to construct a series of ``standard'' galaxy models, whose colors will be compared with observations, and then the effects of factors other than the SFR will be considered in turn.
\subsubsection{``Standard'' models}
\label{subsubsec:Standard models}
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig15.png}
\caption{Two-color diagram for morphologically normal galaxies and globular clusters. \emph{Filled circles}: galaxies from the \emph{Hubble Atlas} \citep{Sandage1961c}, excluding peculiars and those with galactic latitudes $|b| < 20^{\circ}$, with corrected colors from the \citetalias{deVaucouleurs1976SecondGalaxies.}; the \emph{error cross} is for this sample, and the \emph{solid line} is its mean locus estimated by eye \citep{Larson1978StarGalaxies}. \emph{Crosses}: E and S0 galaxies in the Virgo cluster, with colors from \citet{Sandage1972AbsoluteColor}, corrected for reddening according to the \citetalias{deVaucouleurs1976SecondGalaxies.} formulae. \emph{Open circles}: galactic globular clusters (excluding those with $E_{B - V} > 0.05$), with colors from \citet{Harris1979GlobularGalaxies}, corrected for reddening according to the \citetalias{deVaucouleurs1976SecondGalaxies.} formulae.}
\label{fig:fig15}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig16.png}
\caption{Theoretical two-color diagram for galaxies with monotonic SFRs, Solar metallicity, and the local IMF (Section~\ref{subsec:UBV Colors of Normal Galaxies}). \emph{Heavy line}: ``standard'' models, i.e. those of age $10 \: \rm Gyr$, with SFRs ranging from constant at the top to a single initial burst at the bottom. \emph{Light solid lines}: models differing from the standard set only in age, as indicated. \emph{Dashes}: models differing from the standard set only in having an IMF with a constant slope, as marked. \emph{Dash-dot line}: models differing from the preceding set with $x = 1$ only in having an upper stellar mass limit $m_{\rm U} = 10 \: \rm M_{\odot}$, whereas all other models shown have $m_{\rm U} = 30 \: \rm M_{\odot}$. \emph{Arrows}: approximate estimates of the effect on colors of blue and red galaxies, respectively, of altering the metallicity by the factor indicated.}
\label{fig:fig16}
\end{figure}
The methods discussed in Section~\ref{subsubsec:Evolutionary models} have been used by various authors to construct models corresponding to the standard scenario; the (typical) results shown here are from \citet{Larson1978StarGalaxies}. Let us consider models with the IMF of the Solar neighborhood, Solar (or old-disk) metallicity, and an age of $10 \: \rm Gyr$; the exact choice of these standard parameters is not critical, as shown below. The models have monotonically decreasing SFRs ($\psi$), ranging from constant to a single initial burst lasting $10^{7} \: \rm yr$. Different series of models have different monotonic functions $\psi(t)$ between these extremes, such as exponential functions, negative powers of time, and combinations of the constant and single-burst models as two components in different proportions. The colors of these series in the UBV diagram all lie very near the locus indicated by a single heavy line in Figure~\ref{fig:fig16}: the model with a constant SFR is at the top of this line, that with an initial burst is at the bottom, and the form of the curve in between is essentially the same for all series with various functional forms for $\psi(t)$. The theoretical locus for these standard models is very close to the observed mean locus for galaxies of different types (line in Figure~\ref{fig:fig15}), so the standard scenario is at least superficially consistent. Two conclusions can be stated.
\begin{enumerate}
\item \emph{The UBV colors of normal galaxies can in general be accounted for by models with the same age, metallicity, and IMF}; most of the observed colors lie in the range predicted for monotonically decreasing SFRs, within the observational errors. A further conclusion is that late-type galaxies are not necessarily young, even though their appearance and blue-light spectra are dominated by short-lived OB stars; the integrated colors are instead consistent with an underlying population of stars with ages up to many billions of years. These conclusions have been stressed in the context of evolutionary models by \citet{Tinsley1968EvolutionGalaxies}, \citet{Searle1973TheGalaxies}, \citet{Larson1974PhotometricGalaxies, Larson1978StarGalaxies}, and \citet{Huchra1977StarGalaxies}. Caveats and deviations from the norm are discussed below.
\vspace{1mm}
\item \emph{Models with monotonically declining SFRs (and with the same age, metallicity, and IMF) define a one-parameter sequence in the ($U - B$, $B - V$) plane}. Inspection of the models shows that the parameter is the ratio of the present SFR to its average value over the lifetime of the system; or equivalently the SFR per unit mass of stars ($\psi_{1} / M_{\rm s}$ or $\psi_{1} / \overline{\psi}_{1} t_{1}$); or equivalently the inverse of these quantities, a \emph{timescale for star formation} $T_{1} \equiv \overline{\psi}_{1} t_{1} / \psi_{1}$, in the notation of Section~\ref{subsec:IMF}. Galaxies of the latest morphological types are the bluest objects in Figure~\ref{fig:fig15}, and these evidently have the longest timescales for star formation, while the earliest types, which are the reddest, have the shortest timescales; this point will be discussed further in Section~\ref{subsubsec:Relevance to the formation and structure of normal galaxies}.
\end{enumerate}
The one-parameter sequence shows that UBV colors for a given value of $T_{1}$ are almost independent of the functional form of $\psi(t)$, as long as it is monotonic. An unfortunate consequence of this result is that the UBV colors of a galaxy (if near the mean locus in Figure~\ref{fig:fig15}) cannot give any more information about the SFR than the quantity $T_{1}$. In practice, they give less information, because of ambiguities due to possible variations of metallicity, etc., as shown below.
The sensitivity of colors to the single parameter $T_{1}$ is due to the dominance of low-mass giants and/or young OB stars in the light of galaxies, as discussed in Section~\ref{subsubsec:Evolutionary models}. In effect, the contribution of low-mass giants is proportional to the number of long-lived stars ever formed, and the contribution of upper-main-sequence stars is proportional to the present SFR, so their ratio is proportional to $T_{1}$. The integrated colors are insensitive to details of the past SFR because A -- F dwarfs with intermediate lifetimes contribute relatively little light, and because the nature of the giant branch changes little over a wide range of MS lifetimes for the precursor stars. For the same reasons, it is difficult to extract significantly more information about the history of star formation in a galaxy from more detailed photometry than from UBV colors.
\medskip
We next consider some possible problems with the simple one-parameter scenario.
\subsubsection{Possible effects of errors}
\label{subsubsec:Possible effects of errors}
Three systematic discrepancies between the models and data can be seen on comparing Figures~\ref{fig:fig16}~and~\ref{fig:fig15}: the heavy theoretical line lies about $0.05 \: \rm mag$ about the empirical mean locus, some galaxies are bluer than the bluest model, and some are redder than the reddest model. The systematic offset is no more than could be due to errors in the stellar evolution tracks, judged from series of models based on alternative tracks. If this offset is corrected ad hoc by moving the heavy theoretical line downward, there are still some bluer and redder galaxies than predicted. The differences seem to be too big to ascribe to uncertainties in the stellar evolution used, and they cannot be corrected by redefining the ``standard'' age or metallicity, since an improvement at the red end would leave more discrepant galaxies at the blue end, and vice versa. Nor can the ``standard'' IMF be changed, since if the IMF is universal it must be the same as the local function. Therefore, not all of the discrepancies between the heavy line and the data are due to theoretical errors within the framework of the standard scenario.
Although the mean error bar shown for the data in Figure~\ref{fig:fig15} is small, some of the colors may have significantly larger errors due to uncertainties in the reduction. The colors plotted were corrected in the \citetalias{deVaucouleurs1976SecondGalaxies.} on a statistical basis for Galactic and internal reddening, so excessively red and blue galaxies could result from inappropriate corrections in a few cases. A reddening vector (from the \citetalias{deVaucouleurs1976SecondGalaxies.}) is shown in Figure~\ref{fig:fig15}, and it indicates that galaxies away from the ends of the distribution could not be moved far from the mean locus except by extremely large over- or underestimates of their reddening, because the vector happens to lie almost parallel to the mean locus itself.
Emission lines can affect the colors of late-type galaxies, but the estimates made by \citet{Huchra1977StarGalaxies} indicate that morphologically normal galaxies are unlikely to have strong enough gaseous emission for this to be important.
\medskip
In summary, it seems likely that some normal galaxies have colors that are too red or too blue to be accounted for by the standard scenario. Two questions arise: How can the discrepant galaxies be accounted for? And could normal galaxies have significant variations in age, IMF, or metallicity that do not show up on the UBV plot?
\subsubsection{Variations in age}
\label{subsubsec:Variations in age}
Light lines in Figure~\ref{fig:fig16} indicate the effects of allowing ages between 5 and $20 \: \rm Gyr$. The loci for different ages overlap, so most of the galaxies in Figure~\ref{fig:fig15} could have any ages in this range. The extreme colors, however, do depend on age, and the bluest and reddest data points could be accounted for if the ages of galaxies vary by a factor ${\sim} 4$. We shall see that this is not the only possible explanation of those data points, since metallicity effects are probably important.
\subsubsection{Variations in metallicity}
\label{subsubsec:Variations in metallicity}
Effects of different stellar metallicities ($Z_{\rm s}$) can be estimated as outlined in Section~\ref{subsubsec:Evolutionary models}, and some approximate results are indicated in Figure~\ref{fig:fig16}; the slope of the vector for red galaxies is empirical, but the slope for blue galaxies and the length of each vector are uncertain by factors ${\sim} 2$.
As discussed in Section~\ref{subsec:Data2}, the sequence of colors for E--S0 galaxies (crosses in Figure~\ref{fig:fig15}) is regarded as one of metallicity. This sequence closely overlaps the locus of galaxies with different SFRs and ages, so UBV colors alone cannot unambiguously give the SFR parameter ($T_{1}$), age, and $Z_{\rm s}$ for a population of stars. The reddest points in Figure~\ref{fig:fig15} are giant elliptical galaxies, which almost certainly have a mean $Z_{\rm s}$ greater than Solar (in the aperture used for the colors); if the galaxies have some residual star formation, it is undetected to date, but it could affect the colors enough to change the estimated $Z_{\rm s}$ and/or age somewhat.
The bluest points in Figure~\ref{fig:fig15} are small late-type galaxies. It is known from studies of the gaseous emission lines in some such galaxies that they can be significantly metal-poor (e.g. a factor of 4 in the Small Magellanic Cloud; \citealp{Pagel1978AClouds}), so a low $Z_{\rm s}$ could help to make these points very blue. The effects of abundance changes on the colors of blue galaxies are too uncertain to say whether another effect, such as a somewhat younger age, is also required to account for their being bluer than the standard sequence.
\subsubsection{Variations in the IMF}
\label{subsubsec:Variations in the IMF}
To show possible effects of variations in the IMF, Figure~\ref{fig:fig16} includes the loci of UBV colors of models differing from the standard sequence only in their IMF. Two of the variants have constant slopes, $x = 1$ and $x = 2$, while the third has $x = 1$ and an upper limit $m_{\rm U} = 10 \: \rm M_{\odot}$ compared to $30 \: \rm M_{\odot}$ in all other cases\footnote{The models in Figures~\ref{fig:fig16}~and~\ref{fig:fig18} use an IMF slightly different from Equation~(\ref{eq:eq2.9}): the upper limit is $30 \: \rm M_{\odot}$ (except as stated), and the slope is $x = 1.3$ ($\phi \propto m^{-2.3}$) for all $m > 2 \: \rm M_{\odot}$. The UBV colors would be little affected if Equation~(\ref{eq:eq2.9}) itself, with $m_{\rm U}$ taking any value $\geq 30 \: \rm M_{\odot}$, were used \citep[cf.][]{Huchra1977StarGalaxies}.}. For blue galaxies, the local IMF gives colors between those for $x = 1$ and $x = 2$. In general, the colors are redder with a larger value of $x$ or a smaller value of $m_{\rm U}$, since there are relatively few upper-MS stars. Comparisons with Figure~\ref{fig:fig15} show that the variants illustrated are about the largest deviations from the local IMF that one could have without predicting a greater color spread than is observed.
This conclusion applies only to the bluer galaxies, and only to stars $\gtrsim 1 \: \rm M_{\odot}$ that contribute significantly to their light. It is clear from Figure~\ref{fig:fig16} that the UBV colors of redder galaxies ($B - V \gtrsim 0.8$) are very insensitive to the variations of IMF considered. Additional information on the IMF, derived from spectroscopic studies and $M / L$, was discussed in Section~\ref{subsubsec:Other IMF}. Possible departures from the local IMF discussed there are a lack of stars above $10 \: \rm M_{\odot}$ in some early-type spirals, and ubiquitous variations in the fraction of very low-mass objects. In elliptical galaxies, the IMF between ${\sim} 0.4$ and $1 \: \rm M_{\odot}$ cannot be very much steeper than the local function (Section~\ref{subsubsec:The ratio of giants to dwarfs in the light}).
\subsubsection{Relevance to the formation and structure of normal galaxies}
\label{subsubsec:Relevance to the formation and structure of normal galaxies}
To summarize the preceding discussion, every aspect of the standard scenario has been shown to have its weaknesses: the metallicity, IMF, and age are known to vary from one galaxy to another and/or could vary significantly without affecting the locus of normal UBV colors. Nevertheless, it is true that the main parameter causing the progression of colors of morphologically normal galaxies is the timescale for star formation. The average UBV colors of galaxies of different Hubble types lie along the middle of the distribution in Figure~\ref{fig:fig15}, with the latest types at the top and the earliest at the bottom \citep{deVaucouleurs1977QualitativeGalaxies.}. Thus \emph{there is a strong correlation between the structure of galaxies and their timescales for star formation}. Star formation seems to be most efficient in the galaxies with the highest bulge-to-disk ratios, and, among the spirals, it is most efficient in those with the tightly wound spiral arms.
\medskip
An obvious question is whether the shape of a galaxy is a consequence of its efficiency of star formation, or whether, conversely, the timescale for star formation is determined by the structure. Both effects are believed to be present. On one hand, more efficient early star formation leads to a galaxy with a greater bulge-to-disk ratio (Section~\ref{subsec:Galaxy_Formation}): the formation of a spheroidal component requires that the stars in this component formed on a timescale less than that for gaseous dissipation in the protogalaxy. On the other hand, the present structure of a galaxy governs its large-scale dynamics, which in turn has important effects on star formation (Section~\ref{subsubsec:SFR factors}); for example, a more prominent bulge implies a greater surface density, which can lead to stronger gas compression and so to more efficient star formation in the disk. The papers cited in Section~\ref{subsec:Galaxy_Formation} and Section~\ref{subsubsec:SFR factors} give many more detailed discussions, and further references, on the origin of the Hubble sequence of galaxy types and the numerous properties of galaxies that correlate with their forms.
S0 galaxies, which are disk galaxies without spiral structure, are an inscrutable class. Like elliptical galaxies, they have colors consistent with no ongoing star formation (or possibly a little, showing at short wavelengths; Section~\ref{subsubsec:Population synthesis}). There are currently two types of theory on the origin of S0 galaxies: in the first, they are former spirals that have no more star formation because their ISM was lost, either in a collision with another galaxy or by ram-pressure sweeping due to motion through an ambient IGM; in the second type of theory, S0 galaxies had intrinsically very efficient star formation in their disks at early stages. The second type of theory is preferred by some authors because S0 galaxies in isolation and in dense clusters have essentially the same low contents of neutral hydrogen \citep{Faber1976HWinds} and the same distributions of colors \citep{Sandage1978Color-absoluteGalaxies}, and because there are some structural differences between S0s and spirals \citep{Burstein1978TheGalaxies.}. Nevertheless, the first picture is supported circumstantially by the high proportion of S0 galaxies in clusters, especially in their central regions, and especially in clusters with hot IGM \citep{Oemler1974TheClusters, Melnick1977TheClusters}. Arguments based on colors are not decisive, because models in which star formation stopped long ago have similar present colors to those in which it stopped only a few Gyr ago \citep{Biermann1975OnGalaxies}. Whatever mechanism cuts off star formation, the very blue galaxy content of distant, regular clusters strongly suggests that many S0 galaxies were actively making stars only ${\sim} 4 \: \rm Gyr$ ago (Section~\ref{subsubsec:Evolution of colors}). It is especially interesting that some nearby clusters contain ``anemic'' spirals, with weak spiral structure and a subnormal neutral hydrogen content, that have been interpreted as disk systems at a stage of evolution between normal spirals and stripped S0s \citep{vandenBergh1976a}.
\subsection{Colors of Peculiar Galaxies}
\label{subsec:Colors of Peculiar Galaxies}
Galaxies with morphological peculiarities have peculiar colors too, as illustrated in Figure~\ref{fig:fig17}, which is a UBV diagram for systems in the \emph{Atlas of Peculiar Galaxies} \citep{Arp1966AtlasGalaxies}. The width of their color distribution is in striking contrast to the narrow locus of normal galaxies (Figure~\ref{fig:fig15}), and on closer inspection the width turns out to be due almost entirely to interacting galaxies, which are shown as crosses in Figure~\ref{fig:fig17} \citep{Larson1978StarGalaxies}. The implication is that dynamical disturbances have led to an unusual star formation history, so a study of these galaxies and their colors might shed some light on the process of star formation in general.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig17.png}
\caption{Two-color diagram for galaxies in the \emph{Atlas of Peculiar Galaxies} \citep{Arp1966AtlasGalaxies}, excluding those with galactic latitudes $|b| < 20^{\circ}$, with corrected colors from the \citetalias{deVaucouleurs1976SecondGalaxies.} and other sources cited by \citet{Larson1978StarGalaxies}. \emph{Crosses} denote interacting systems. \emph{Open circles} are two Type~I Seyfert galaxies, whose colors may be affected by non-thermal emission.}
\label{fig:fig17}
\end{figure}
\subsubsection{Bursts of star formation and blue colors}
\label{subsubsec:Bursts of star formation and blue colors}
If the SFR in a galaxy does not decrease monotonically, colors very different from those of the standard models in Section~\ref{subsec:UBV Colors of Normal Galaxies} can be obtained. The idea of star formation in ``flashes'' or ``bursts'' was introduced by \citet{Searle1972InferencesGalaxies} (see also \citealp{Searle1973TheGalaxies}) to explain the very blue colors of some dwarf irregular galaxies, and it has appeared in other contexts including elliptical galaxies with patches of star formation \citep{vandenBergh1975StellarGalaxies} and the peculiar galaxies discussed here.
\medskip
The effects of a burst of star formation on a formerly red galaxy are illustrated in Figure~\ref{fig:fig18}, where the heavy curve is the locus of standard models aged $10 \: \rm Gyr$, from Figure~\ref{fig:fig16}. The colors of younger galaxies are shown in two extreme cases: the dotted line is a model with a constant SFR, evolving through the ages shown (in Gyr), and the heavy dashed line (on the left) is a model whose star formation stopped at $10^{7} \: \rm yr$. The latter curve can be regarded as the evolution of a cluster of stars formed in a period of $10^{7} \: \rm yr$, or equivalently as the colors resulting from stars formed in a burst lasting $10^{7} \: \rm yr$. The light solid lines are the loci of models made of two components:
\begin{enumerate}
\item a red galaxy aged $10 \: \rm Gyr$, with no ongoing star formation; and
\vspace{1mm}
\item stars formed in a burst of duration $10^{7} \: \rm yr$, seen at ages $10^{7} \: \rm yr$ (upper line) and $10^{8} \: \rm yr$ (lower light solid line).
\end{enumerate}
\begin{figure}
\includegraphics[width=0.47\textwidth]{Figures/Fig18.png}
\caption{Theoretical two-color diagram showing the colors of young galaxies (\emph{dotted and heavy dashed lines}) and an old red galaxy with bursts of star formation of various strengths and ages (\emph{light solid and dashed lines}). Details of these lines and their labels are explained in Section~\ref{subsubsec:Bursts of star formation and blue colors}. The \emph{heavy solid line} is the locus of standard old models, from Figure~\ref{fig:fig16}.}
\label{fig:fig18}
\end{figure}
The numbers along the upper curve are a burst strength parameter, defined as the mass ratio of stars formed in the burst to stars in the old red galaxy. Finally, the light dashed lines represent the evolution of a composite system, from age $10^{7} \: \rm yr$ when the star formation in the burst stops; these are lines of constant burst strength, and they cross the lower light solid line when the age is $10^{8} \: \rm yr$. (The heavy dashed line is the limiting case of a burst of infinite strength, i.e., without any underlying old stars). It can be seen that a burst of star formation in a red galaxy gives colors initially above the normal UBV locus, since the young stars cause an ``ultraviolet excess''; as the burst ages, the colors evolve across the normal locus after nearly $10^{8} \: \rm yr$; then they fall below the normal line and eventually become imperceptibly different from the colors of an undisturbed old galaxy.
Evidently the colors of peculiar galaxies in Figure~\ref{fig:fig17} can be explained by bursts of star formation of various strength and ages in galaxies with various initial colors. Using a series of models like those in Figure~\ref{fig:fig18}, \citet{Larson1978StarGalaxies} found that most of the \citet{Arp1966AtlasGalaxies} galaxies can be accounted for with bursts of strength less than $5\%$ and duration $< 2 \times 10^{7} \: \rm yr$; a few more deviant colors are probably due to observational scatter, internal reddening, non-thermal continuum emission (in two Type~I Seyfert galaxies in the sample), and strong gaseous emission lines, whose effects on colors are discussed by \citet{Huchra1977StarGalaxies}. The strongest bursts of star formation are inferred for galaxies in distorted close pairs, often with bridges or tails, and in apparently single systems with tails and filamentary streamers. Dynamical models for colliding galaxies predict features like these in cases of strong tidal deformation and recent mergers \citep{Toomre1977MergersConsequences}, so it appears that violent dynamical interactions lead to star formation. For this reason, it has been suggested that the stars in elliptical galaxies could have formed in bursts due to collisions and mergers among protogalactic subsystems (Section~\ref{subsubsec:Bursts of star formation in merging subsystems}).
Figure~\ref{fig:fig18} shows that the colors of models with bursts of strength $10\%$ and infinity are very similar. The differences are less than the observational uncertainties for many faint peculiar galaxies, and the theoretical uncertainties in stellar evolution, metallicity effects, etc. In other words, it is not possible to tell from UBV colors alone whether a galaxy is really young or has $90\%$ of its mass in old stars! Some of the galaxies near the upper left in Figure~\ref{fig:fig17} are very chaotic in appearance, and are tempting candidates for truly young galaxies, but in all cases the colors are inconclusive and the appearance could be due to a violent collision or irregularly distributed star formation in an old galaxy.
Colors at longer wavelengths, such as $V - K$, are much more sensitive than $B - V$ to the presence of some old stars that would distinguish between a truly young system and one with a burst strength ${\sim} 10\%$ \citep{Struck-Marcell1978StarRadiation}. Galaxies with very active star formation are often dusty, so red colors could be due to reddening rather than to age; on the other hand, very blue values of $V - K$ would indicate a lack of red giants, and much stronger limits could be put on the mass of any underlying old component. As yet, there are no known nearby galaxies in which the dominant young stellar component could not be masking a significant mass of old stars.
\subsubsection{Highly reddened galaxies}
\label{subsubsec:Highly reddened galaxies}
Some regions of galaxies that are suspected of having intense star formation are extremely dusty, and they show thermal infrared (IR) emission that is interpreted as re-radiation of starlight by the dust. Example of such regions include the centers of M82 and NGC~253, and the dust band around NGC~5128 \citep[e.g.][]{Kleinmann1977InfraredSources, Telesco1978ExtendedA/}. Star formation is indicated by early-type spectra and blue colors in unobscured patches \citep[e.g.][]{vandenBergh1971The82, vandenBergh1978Multi-ColourA}, emission from interstellar molecules \citep[e.g.][]{Whiteoak1978RadioGalaxies}, and the lack of more plausible explanations for the IR emission \citep[e.g.][]{Kleinmann1977InfraredSources}. In NGC~5128, the dust band has an IR luminosity of a few times $10^{10} \: \rm L_{\odot}$ \citep{Telesco1978ExtendedA/}, which rivals the visual luminosity of the entire elliptical galaxy.
If the IR luminosity is assumed to represent the bolometric luminosity of buried stars, an SFR can be estimated \citep{Struck-Marcell1978StarRadiation}: models like those of Section~\ref{subsec:UBV Colors of Normal Galaxies} show that any system with a mass-to-luminosity ratio $M_{\rm s} / L_{\rm bol} < 0.5$ is so dominated by young stars that $L_{\rm bol}$ is almost directly proportional to the SFR; with the local IMF, the relation is
\begin{equation}
\psi \simeq (0.1 - 0.4) \frac{L_{\rm bol}}{\rm L_{\odot}} \: \rm M_{\odot} \ Gyr^{-1}.
\label{eq:eq7.1}
\end{equation}
An upper limit to the time for which star formation could have continued at this rate is approximately $M_{\rm s} / \psi$, so the limiting timescale $\tau_{\rm s}$ depends only on $M_{\rm s} / L_{\rm bol}$, according to the relation
\begin{equation}
\tau_{\rm s} \equiv \frac{M_{\rm s}}{\psi} \sim (3 - 10) \frac{M_{\rm s} / L_{\rm bol}}{\rm M_{\odot} / L_{\odot}} \: \rm Gyr.
\label{eq:eq7.2}
\end{equation}
Some galactic nuclei have such strong IR emission that $M_{\rm s} / L_{\rm bol}$ is only a few hundredths, so the timescale is only a few times $10^{8} \: \rm yr$. The dust-band region of NGC~5128 is also making stars at a prodigious rate: given a luminosity ${\sim} 10^{10} \: \rm L_{\odot}$, Equation~(\ref{eq:eq7.1}) leads to an SFR of about $2 \: \rm M_{\odot} \ yr^{-1}$, so a respectable disk of stars could be built in just a few times the dynamical timescale of the system. \citet{vandenBergh1975StellarGalaxies} has suggested that NGC~5128 could evolve into an early-type spiral seen edge-on, like the Sombrero galaxy M104.
Since most of the bolometric light comes from massive stars, the above ratios of SFR to $L_{\rm bol}$ could be overestimates if low-mass stars are not forming. For example, in the models discussed, stars above $10 \: \rm M_{\odot}$ contribute $90\%$ of $L_{\rm bol}$, but stars below $1 \: \rm M_{\odot}$ account for $80\%$ of the mass formed into stars. There is no reason to suspect a lack of low-mass stars, however, especially since low-mass protostars (T Tauri stars) are associated with dark clouds in the Milky Way.
These very dusty galaxies with intense star formation lead again to the question of what truly young galaxies would look like. In the absence of dust they would be very blue, like the young models in Figure~\ref{fig:fig18}, but it seems that the regions of galaxies with the most intense bursts are the very reddened ones just described. This question is important in the context of primeval galaxies at large redshifts, i.e. the early stages of most normal galaxies that now have ages ${\sim} 10 \: \rm Gyr$. Starting with the ideas of \citet{Partridge1967AreVisible}, most models for primeval galaxies have assumed that hot young stars would be visible and lead to detectable luminosities (despite the great distances) during a brilliant early burst of star formation \citep[e.g.][]{Meier1976TheGalaxies, Kaufman1977YoungRedshifts, Sunyaev1978ObservableRedshift.}. A model for a very dusty primeval galaxy has been studied by \citet{Kaufman1976PrimevalLuminosities}, and \citet{Sunyaev1978ObservableRedshift.} considered the possibility of substantial radiation from dust. If the very dusty nearby galaxies are the best analogs of truly primeval systems, as suggested by \citet{Larson1976c}, their prospects for detection at optical wavelengths are dim. Observational searches have so far produced null results.
Another factor making primeval galaxies hard to detect could be that most galaxies have such a long timescale for star formation that there is not a very bright phase at an early peak SFR \citep{Tinsley1978a, Tinsley1979StellarGalaxies}. One argument is that most spiral galaxies have colors that correspond to rather long timescales for star formation (Section~\ref{subsec:UBV Colors of Normal Galaxies}), precluding a significant early burst with a corresponding peak in luminosity. Moreover, even elliptical galaxies could form their stars over rather long time intervals if star formation occurs during a series of mergers among subsystems (Section~\ref{subsubsec:Bursts of star formation in merging subsystems}). According to these ideas, late-type galaxies could be fainter in the past than they are now, and early-type galaxies could experience their brightest evolutionary stages only a few Gyr ago. Galaxies in interestingly early evolutionary stages have indeed already been found at moderate redshifts: distant clusters have excess numbers of blue galaxies (Section~\ref{subsubsec:Evolution of colors}; \citealp{Butcher1978The295}), and counts in the field show a large excess of blue galaxies at faint apparent magnitudes \citep{Kron1978PhotometryGalaxies.}. Further studies of these phenomena will surely shed light on the ways in which stars and galaxies have formed during cosmological time.
\section{Conclusion}
\label{sec:Conclusion}
Returning to the outline of galactic evolution in Figure~\ref{fig:fig1}, one can see how much remains to be learned before the jigsaw puzzle will be complete enough for a clear picture to emerge. Essentially every aspect of the subject needs further observational and theoretical study, so galactic evolution will long be a fertile field for research.
\section*{Acknowledgements}
I am grateful to Pierre Demarque, Richard B. Larson, Curtis Struck-Marcell, and Bruce A. Twarog for their suggestions and help in the preparation of this review. The work was supported in part by the National Science Foundation (Grant AST77-23566) and the Alfred P. Sloan Foundation.
\bibliographystyle{mnras}
| f070be937239913639caae680fc40bf73c9cd39e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Deep visuomotor policies that map from pixels to actions end-to-end can represent complex manipulation skills \cite{levine_finn_2016}, but have shown to be sensitive to the choice of action space \cite{martin2019iros} -- e.g., Cartesian end effector actions (i.e. task space \cite{berenson2011task}) perform favorably when learning policies for tabletop manipulation, while joint actions have shown to fare better for whole-body motion control \cite{peng2017learning, Tan-RSS-18}.
In particular, policies modeled by deep networks are subject to spectral biases \cite{ronen2019convergence,battaglia2018relational, bietti2019inductive} that make them more likely to learn and generalize the low-frequency patterns that exist in the control trajectories.
Hence, choosing the right action space in which to define these trajectories remains a widely recognized problem in both reinforcement learning \cite{peng2017learning} as well as imitation learning, where demonstrations can be provided in a wide range of formats -- e.g., continuous teleoperation in Cartesian or joint space, kinesthetic teaching, etc. -- each changing the underlying characteristics of the training data.
\begin{figure}[t!]
\centering
\vspace{0.5em}
\noindent\includegraphics[width=1.0\linewidth]{images/splash-v4.png}
\caption{Enabled by an implicit policy formulation, Implicit Kinematic Policies (IKP) provide both Joint and Cartesian action representations (linked via forward kinematics) to a model that can pick up on inductive patterns in both spaces. Subsequently, this allows our end-to-end policies to extract the best action space for a wide variety of manipulation tasks ranging from tabletop block sorting to whole-arm sweeping.}
\label{fig:teaser}
\vspace{-1.5em}
\end{figure}
Although considerable research has been devoted to finding the right action space for a given application \cite{peng2017learning}, much less attention has been paid to figuring out how our models could instead discover for themselves which optimal combination of action spaces to use.
This goal remains particularly unclear for most \textit{explicit} policies $f_\theta(\textbf{o})=\hat{\textbf{a}}$ with feed-forward models that take as input observations $\textbf{o}$ and output actions $\textbf{a}\in\mathcal{A}$. Explicit formulations must either specify a single action space to output from the model and potentially convert predictions into another desired space (e.g. with inverse kinematics), or use multiple action spaces with multiple outputs and losses which can be subject to conflicting gradients \cite{yu2020gradient} from inconsistent predictions \cite{martin2019iros}.
In this work, we demonstrate that we can train deep policies to learn which action space to use for imitation, made possible by implicit behavior cloning (BC) \cite{florence2021implicit}. In contrast to its explicit counterpart, \textit{implicit} BC formulates imitation as the minimization of an energy-based model (EBM) $\hat{\textbf{a}}=\arg\min_{\textbf{a} \in \mathcal{A}} E_{\theta}(\textbf{o},\textbf{a})$ \cite{lecun2006tutorial}, which regresses the optimal action via sampling or gradient descent \cite{welling2011bayesian, du2019implicit}. Since actions are now inputs to the model, we propose presenting the model with the same action represented in multiple spaces while remaining consistent between each other, allowing the model to pick up on inductive patterns \cite{battaglia2018relational, bietti2019inductive} from all action representations.
We study the benefits of this multi-action-space formulation with Cartesian task and joint action spaces in the context of learning manipulation skills. Since both spaces are linked by the kinematic chain, we can integrate a differentiable kinematic module within the deep network -- a formulation which we refer to as Implicit Kinematic Policies (IKP). IKP can be weaved into an implicit autoregressive control strategy introduced in \cite{florence2021implicit}, where each action dimension is successively and uniformly sampled at a time and passed as input to the model. This exposes the model to not only the joint configuration and end-effector pose, but also the Cartesian action representations of {\em{every rigid body link}} of the arm as input. This can provide downstream benefits when learning whole-body manipulation tasks that may have emergent patterns in the trajectory of any given link of the robot. Furthermore, a key aspect of Implicit Kinematic Policies is that in addition to exposing the implicit policy to both Cartesian and joint action spaces, it also enables incorporating learnable layers throughout the kinematic chain, which can be used to optimize for residuals in either space
Our main contribution is Implicit Kinematic Policies, a new formulation that integrates forward kinematics as a differentiable module within the deep network of an autoregressive implicit policy, exposing both joint and Cartesian action spaces in a kinematically consistent manner. Behavior cloning experiments across several complex vision-based manipulation tasks suggest that IKP, without any modifications, is broadly capable of efficiently learning both prehensile and non-prehensile manipulation tasks, even in the presence of miscalibrated joint encoders -- results that may pave the way for more data efficient learning on low-cost or cable-driven robots, where low-level joint encoder errors (due to drift, miscalibration, or gear backlash) can propagate to large non-linear artifacts in the Cartesian end effector trajectories. IKP achieves 85.9\%, 97.5\% and 92.4\% on a sweeping, non-prehensile box lifting and precise insertion task respectively -- performance that is on par with or exceeds that of standard implicit or explicit baselines (using the best empirically chosen action space per individual task). We also provide qualitative experiments on a real UR5e robot suggesting IKP's ability to learn from noisy human demonstrations in a data-efficient manner.
Our experiments provide an initial case study towards more general action-space-agnostic architectures for end-to-end robot learning.
\section{Related Work}
\subsection{Learning Task-Specific Action Representations}
While much of the early work in end-to-end robot learning has focused on learning explicit policies that map pixels to low-level joint torques \cite{ddpg2016, levine_finn_2016}, more recent work has found that the choice of action space can have a significant impact on downstream performance for a wide range of robotic tasks ranging from manipulation to locomotion \cite{martin2019variable, duan2021learning, peng2017learning, Viereck_2018, varin2019comparison}. In the context of deep reinforcement learning for locomotion, \cite{peng2017learning} find that using PD joint controllers achieves higher sample efficiency and reward compared to torque controllers on bipedal and quadrupedal walking tasks in simulation, whereas \cite{Viereck_2018} interestingly find that torque controllers outperform PD joint controllers for a real robot hopping task. In contrast, \cite{duan2021learning} shows that using task space actions are more effective for learning higher level goals of legged locomotion compared to joint space controllers and demonstrate results on a real bipedal robot.
Careful design of the action space is equally important when learning manipulation skills with high DoF robots \cite{martin2019variable, varin2019comparison}. Prior approaches successfully use direct torque control or PD control with joint targets to learn a variety of tasks such as pick-and-place, hammering, door opening \cite{rajeswaran2018learning} and stacking \cite{popov2017dataefficient}, but these methods often require lots of data and interaction with the environment to generalize well and are also sensitive to feedback gains in the case of PD control.
Other works default to low-level residual cartesian PD controllers in order to learn prehensile tasks such as block stacking \cite{johannink2018residual} and insertion \cite{silver2019residual}, but \cite{martin2019variable, varin2019comparison} instead vouch for a task-space impedance controller to support more compliant robot control. However, this approach is agnostic to the full joint configuration of the robot and focuses on tabletop manipulation which is generally better defined in task space. Fundamentally, these works share the common theme of choosing the single best action space for a particular task. In this work, we are more concerned with the problem of extracting the best combination of action spaces.
\subsection{Difficulties in Learning High-Frequency Functions}
The proposed benefits of implicit kinematic policies rely on the hypothesis that simpler patterns are easier for deep networks to learn. Although theoretical analyses of two-layer networks and empirical studies of random training labels have shown that deep networks are prone to memorization \cite{Zhang2017UnderstandingDL}, further studies on network convergence behavior and realistic dataset noise have found that this result does not necessarily explain performance in practice. Instead, deep networks first learn patterns across samples before memorizing data points \cite{Arpit2017ACL}. \cite{Rahaman2019Spectral} formalize this finding and show that deep networks have a learning bias towards low-frequency functions by using tools from Fourier analysis and \cite{Basri2019TheCR} extend this result by showing that networks fit increasingly higher frequency functions over the course of training. Within learning-based robotics, multiple works have found the related result that higher-level action spaces lead to improved learnability. Specifically, the frequency of the action space can have an outsized effect on the robustness and quality of learned policies \cite{peng2017learning}. This could explain why learned, latent actions improve performance and generalizability for policies trained either with reinforcement learning \cite{Haarnoja2018LatentSP,Hausman2018LearningAE,Nair2020HierarchicalFS,Pflueger2020PlanSpaceSE} or imitation learning \cite{fox2017multi}.
\begin{figure}[t]
\centering
\noindent\includegraphics[width=\linewidth]{images/implicit_vs_explicit.png}
\caption{Integrating forward kinematics (FK), in particular, synergizes with implicit policies (a, b), where actions $\textbf{a}$ are sampled in joint space, and both joints and corresponding Cartesian actions along with observations $\textbf{o}$ can be fed as inputs to the EBM $f_\theta$. For explicit policies (c), it becomes less clear how to expose both action spaces in a kinematically consistent way to the deep network, except via predicting joints from observations, then using forward kinematics to backpropagate gradients from losses imposed on the final output Cartesian actions (in red). In this work, we investigate both formulations and find that implicit (b) yields better performance.
}
\label{fig:key-idea}
\vspace{-2em}
\end{figure}
\section{Background}
\subsection{Problem Formulation}
We consider the imitation learning setting with access to a fixed dataset of observation-action pairs. We frame this problem as supervised learning of a policy $\pi: \mathcal{O} \rightarrow \mathcal{A}$ where $\mathbf{o} \in \mathcal{O} = \mathbb{R}^m$ represents an observational input to the policy and $\mathbf{a} \in \mathcal{A} = \mathbb{R}^d$ represents the action output. The policy $\pi$, with parameters $\theta$, can either be formulated as an {\em{explicit}} function $f_{\theta}(\mathbf{o}) \rightarrow \hat{\mathbf{a}}$ or as an {\em{implicit}} function $\arg\min_{\mathbf{a}\in\mathcal{A}} E(\mathbf{o}, \mathbf{a}) \rightarrow \hat{\mathbf{a}}$. In this work, we adopt the implicit formulation and build upon recent work which we describe in more detail in Section \ref{implicitbcsec}. We operate in the continuous control setting, where our policy infers a desired target configuration at 10 Hz, and a low-level controller asynchronously reaches desired configurations with a joint-level PD controller. Our $\mathcal{O}$ contains both image observations and proprioceptive robot state (i.e. from joint encoders).
\subsection{Implicit BC}
\label{implicitbcsec}
We use an energy-based, contrastive loss to train our implicit policy, modeled with the same late-fusion deep network architecture as in \cite{florence2021implicit} with 26 convolutional ResNet layers \cite{resnets_2016} for the image encoder and 20 dense ResNet layers for the EBM. Specifically, to optimize $E_{\theta}(\cdot)$, we train an InfoNCE style loss \cite{Oord2018RepresentationLW} on observation-action samples \cite{florence2021implicit}. Our dataset consists of $\{\mathbf{o}_i, \mathbf{a}^*_i\}$ for $\mathbf{o}_i \in \mathbb{R}^m$ and $\mathbf{a}_i \in \mathbb{R}^d$ and from regression bounds $\mathbf{a}_{\min} , \mathbf{a}_{\max} \in \mathbb{R}^d$ we generate negative samples $\{\tilde{\mathbf{a}}_i^j\}_{j=1}^{N_{\text{neg.}}}$. The loss equates to the negative log likelihood of $p(\mathbf{a}|\mathbf{o})$ where we use negative samples to approximate the normalizing constant:
\begin{align*}
\mathcal{L}_{\text{InfoNCE}} &= \sum_{i=1}^N -\log(\tilde{p}_\theta(\mathbf{a}^*_i|\mathbf{o}_i,\{\tilde{\mathbf{a}}_i^j\}_{j=1}^{N_{\text{neg.}}})) \\
\tilde{p}_\theta(\mathbf{a}^*_i|\mathbf{o}_i,\{\tilde{\mathbf{a}}_i^j\}_{j=1}^{N_{\text{neg.}}}) &= \frac{e^{-E_\theta(\mathbf{o}_i, \mathbf{a}^*_i)}}{e^{-E_\theta(\mathbf{o}_i, \mathbf{a}^*_i)} + \sum_{j=1}^{N_{\text{neg.}}}e^{-E_\theta(\mathbf{o}_i, \tilde{\mathbf{a}}_i^j)}}
\end{align*}
To perform inference, given some observation $\mathbf{o}$, we minimize our learned energy function, $E_{\theta}(\mathbf{o}, \mathbf{a})$, over actions in $\mathcal{A}$ using a sampling-based optimization procedure.
\label{implicitbc}
\label{background}
\section{Method}
Here we present the details of our proposed Implicit Kinematic Policies (IKP) method. First, we will present our extension to the implicit policy formulation which provides multiple action space representations as input. (Sec.~\ref{subsec:multi-action-space}). Sec.~\ref{subsec:multi-action-space-explicit} also compares the implicit multi-action-space formulation to its explicit counterpart and discusses the relevant tradeoffs. We then describe how to perform autoregressive training and inference with implicit policies in a way that exposes the action trajectories of every joint and link in the robot to the model (Sec.~\ref{subsec:autoregressive-formulation}). Finally, we discuss a motivating application of controlling miscalibrated robots for precise tasks, and how IKP is uniquely suited to automatically compensate for this noise through the use of strategically placed residual blocks in the model (Sec.~\ref{subsec:residual-formulation}).
\subsection{Multi-Action-Spaces With Implicit Policies}\label{subsec:multi-action-space}
Our formulation enables the implicit policy $\hat{\mathbf{a}} = \arg\min_{\mathbf{a}} E_{\theta}(\mathbf{o}, \mathbf{a})$ to have access to multiple action spaces, which are constrained to be consistent. Specifically, our implicit multi-action-space formulation is of the form:
\vspace{-.85em}
\begin{equation}
\begin{aligned}
\underset{\mathbf{a}, \mathbf{a}', ...}{\arg\min} \quad & E_{\theta}(\mathbf{o}, \mathbf{a}, \mathbf{a}', ...)\\
\textrm{s.t.} \quad & \mathbf{a} = \mathcal{T}(\mathbf{a}'), ...
\end{aligned}
\label{eq:general-multi-action-space}
\end{equation}
where $\mathbf{a} \in \mathcal{A}$ is one parameterization of the action space, $\mathbf{a}' \in \mathcal{A}'$ is a different parameterization of the action space, and $\mathcal{T}(\cdot): \mathcal{A}' \rightarrow \mathcal{A}$ is a transformation between the two action spaces. Of course, $N$ different consistent parameterizations of the action space could be represented, which is depicted in Eq.~\ref{eq:general-multi-action-space} by the ellipses (...).
In particular for robots, we are interested in representing both {\em{joint-space}} and {\em{cartesian-space}} actions, which is a case of Eq.~\ref{eq:general-multi-action-space} in which the transformation between the two action spaces is forward kinematics (FK):
\vspace{-0.75em}
\begin{equation}
\begin{aligned}
\underset{\mathbf{a}_{\text{joints}}, \ \mathbf{a}_{\text{cartesian}}}{\arg\min} \quad & E_{\theta}(\mathbf{o}, \mathbf{a}_{\text{joints}}, \mathbf{a}_{\text{cartesian}})\\
\textrm{s.t.} \quad & \mathbf{a}_{\text{cartesian}} = FK(\mathbf{a}_{\text{joints}})
\end{aligned}
\label{eq:joints-cartesian-implicit}
\end{equation}
Specifically, we accomplish this by first sampling input actions in joint space $\mathbf{a}_{\text{joints}} \in \mathcal{A}_{\text{joints}}$, then computing the corresponding Cartesian (task) space actions via forward kinematics ($FK$), $\mathbf{a}_{\text{cartesian}} = FK(\mathbf{a}_{\text{joints}})$, and finally concatenating and passing the combined representation into the model $E_{\theta}(\mathbf{o}, \mathbf{a}_{\text{joints}}, \mathbf{a}_{\text{cartesian}})$. A visualization of this model is shown in the middle portion of Fig.~\ref{fig:key-idea}. Our hypothesis is that the model can use this redundant action representation to exploit patterns in both spaces, akin to {\em{automatically}} discovering the best combination of action spaces -- this hypothesis will be tested in our Experiments section.
As we will show in Sec.~\ref{subsec:autoregressive-formulation}, we can perform training and inference for $E_{\theta}$ through using autoregressive derivative-free optimization.
\begin{figure*}[t!]
\label{fkdiagram}
\centering
\noindent\includegraphics[width=\textwidth]{images/overview.png}
\caption{ \textbf{Method overview.} Given an image captured from an RGB camera overlooking the robot workspace (a), we feed it as input to a deep convolutional network (b) to get a latent state representation. We then predict desired robot joint actions by leveraging implicit autoregression \cite{florence2021implicit, nash2019autoregressive} (c) with state-conditioned EBMs to progressively sample each action dimension (joint angle) at a time: i.e., we uniform sample $\text{j}_n$, feed it to $\text{FK}_n$ (forward kinematics to link $n$, prepended with deep layers) to get the Cartesian pose $\text{C}_{n}$, which is then concatenated with the latent state and all previously sampled argmin joint dimensions $[\text{j}_n, \text{j}_{n-1}, ..., \text{j}_0]$ and Cartesian representations $[\text{C}_{n-1}, ..., \text{C}_0]$ and fed to an 8-layer EBM $\text{E}_{n}$ to compute the argmin over $\text{j}_n$.
}
\label{fig:method-overview}
\vspace{-1.5em}
\end{figure*}
\subsubsection{Multi-Action-Spaces With Explicit Policies}
\label{subsec:multi-action-space-explicit}
Consider the above formulation in contrast to an explicit policy, $\hat{\mathbf{a}}=f_{\theta}(\mathbf{o})$, where the mapping $f_{\theta}(\cdot)$ maps to one specific action space. It is possible to provide different losses on different transformations of the explicit policy's action space, for example $\mathcal{L}_{\text{joints}}(\hat{\mathbf{a}}, \mathbf{a}^*)$ and $\mathcal{L}_{\text{cartesian}}(FK(\hat{\mathbf{a}}), FK(\mathbf{a}^*))$, where $FK$ represents differentiable forward kinematics. However, this can lead to conflicting gradients and requires choosing relative weightings, $\lambda$, between these losses: $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{joints}} + \lambda \mathcal{L}_{\text{cartesian}}$. This issue of loss balancing is exacerbated if the explicit policy were to regress the Cartesian representation of every link in the robot in addition to the joint configuration and end-effector pose. Additionally, the explicit policy predicts a single joint configuration even though there may be multiple correct configurations when using kinematically redundant robots. It is possible to first regress Cartesian actions and subsequently recover the corresponding joint configuration via inverse kinematics, however, IK solvers are only approximately differentiable which poses challenges for end-to-end training. A visualization of this explicit policy is shown in the rightmost column of Fig.~\ref{fig:key-idea}.
\subsection{Autoregressive Implicit Training and Inference}
\label{subsec:autoregressive-formulation}
As in \cite{florence2021implicit}, we employ an autoregressive derivative-free optimization method for both training and inference. Sampling-based, derivative-free optimization synergizes with the multi-action-space constrained energy model (Eq.~\ref{eq:general-multi-action-space}), since gradient-based optimization (i.e. through Langevin dynamics \cite{florence2021implicit}), while possible, would require a solution to synchronize between the different action spaces.\footnote{Specifically, when using the implicit model gradients, with learning rate $\lambda$, the actions are updated as $\mathbf{a}_{\text{joints}}^{k+1} = \mathbf{a}_{\text{joints}}^{k} - \lambda \nabla_{\mathbf{a}_{\text{joints}}} E_{\theta}(\cdot)$ and $\mathbf{a}_{\text{cartesian}}^{k+1} = \mathbf{a}_{\text{cartesian}}^{k} - \lambda \nabla_{\mathbf{a}_{\text{cartesian}}} E_{\theta}(\cdot)$. The updated actions
actions may not be consistent with each other -- even for an analytical model, there will be differences in the first-order derivatives of the different spaces.}
The autoregressive procedure uses $m$ models, one model $E_{\theta}^{j}(\mathbf{o}, \mathbf{a}^{0:j})$ for each dimension $j = 1, 2, ...,m$. In contrast with prior work \cite{florence2021implicit}, instead of $\mathbf{a}$ only representing a single action space, our proposed method represents, for each $j$th model, both the 6DoF Cartesian pose of the $j$th link {\em{and}} the joint of that link into $\mathbf{a}$. Similarly in $\mathbf{o}$ we also represent the dual Cartesian-and-joint state of each link. Since our method strictly samples actions in joint space and the state is represented by the joint configuration, the Cartesian representations for both the state $\mathbf{o}$ and actions $\mathbf{a}$ are acquired through $FK(\mathbf{o})$ and $FK(\mathbf{a})$ for every dimension $j = 1, 2, ...,m$. We visualize this procedure in Fig. \ref{fig:method-overview}
\subsection{Residual Forward Kinematics}
\label{subsec:residual-formulation}
In robotics, we often assume access to accurate kinematic descriptions and joint encoders, but this assumption can be a significant source of error in the context of low-cost or cable-driven robots. In these settings, the errors often take the form of fixed, unknown linear offsets in each joint motor encoder or link representation. If calibration isn't performed, both cases can cause heavy non-linear offsets in end-effector space which can drastically affect the performance on high-precision tasks. By adding dense residual blocks both before and after joint actions are sampled in the autoregressive procedure, we can encourage our model to learn these linear offsets at each joint and link rather than solving the more difficult problem of learning a highly non-linear offset in task space. This extension is only possible due to the full differentiability of the forward kinematics layer, which allows gradients to flow through the residual blocks. Concretely, we redefine our multi-action-space formulation with residuals as follows:
\vspace{-1.5em}
\begin{equation}
\begin{aligned}
\underset{\mathbf{a}_{\text{joints}}, \ \mathbf{a}_{\text{cartesian}}}{\arg\min} \quad & E_{\theta}(\mathbf{o}, \mathbf{a}_{\text{joints}}, \mathbf{a}_{\text{cartesian}})\\
\textrm{s.t.} \quad & \mathbf{a}_{\text{cartesian}} = FK\big(\mathbf{a}_{\text{joints}} + \Delta_{\theta}(\mathbf{a}_{\text{joints}})\big)
\end{aligned}
\label{eq:joints-cartesian-residuals-implicit}
\end{equation}
where $\Delta_{\theta}(\mathbf{a}_{\text{joints}}) \in \mathcal{A}_{\text{joints}}$ is a new learnable module in the implicit model which gives the model the inductive bias that there may be imperfect calibration of the joint measurements (i.e., biased joint encoder measurements). The module $\Delta_{\theta}(\cdot)$ can be learned during training and we hypothesize that the EBM may be able to automatically learn the $\Delta_{\theta}(\cdot)$ which provides the lowest-energy fit of the data. This can be interpreted as automatically calibrating the biased joint encoders. Experiments testing this hypothesis are in Section~\ref{subsec:residual-experiments}.
\begin{figure}[t]
\centering
\noindent\includegraphics[width=\linewidth]{images/Sweeping+Flipping.png}
\caption{Expert trajectories for bimanual sweeping are characterized by distinct patterns in Cartesian space (b), e.g., y-values experience a mode-change in the latter half of the episode when the policy switches bowls in which to drop particles. Such patterns are less salient in joint space (a). The opposite is true for bimanual flipping, where linear patterns emerge in joint space (c), whereas they are less salient in Cartesian space (d) especially with randomized end effector poses (semi-transparent plots).}
\label{fig:tasks}
\vspace{-1.5em}
\end{figure}
\section{Experiments}
We evaluate IKP across several vision-based continuous control manipulation tasks with quantitative experiments in simulation, as well as qualitative results on a real robot. All tasks require generalization to unseen object configurations at test time. The goals of our experiments are two-fold: (i) across tasks where one action space substantially outperforms the other, we investigate whether IKP can achieve the best of both and perform consistently well across all tasks, and (ii) given a miscalibrated robot with unknown offsets in the joint encoders (representing miscalibration or low-cost encoders), we study whether IKP can autonomously learn to compensate for these offsets while still succeeding at the task. Across all experiments, a dataset of expert demonstrations is provided by a scripted oracle policy (in simulation), or by human teleoperation (in real).
\subsection{\textbf{Simulated Bimanual Sweeping and Flipping}}
In a simulated environment, we evaluate on two bimanual tasks -- sweeping and flipping -- both of which involve two 7DoF KUKA IIWA robot arms equipped with spatula-like end-effectors positioned over a $0.4m^2$ workspace. The setup for sweeping is identical to that presented in Florence et al. \cite{florence2021implicit}, where given a pile of 50-60 particles on the workspace, the task is to scoop and evenly distribute all particles into two bowls located next to the workspace. For flipping, given a bowl of 20-30 particles attached to the top of a $0.2m^3$ box, the task is to flip the box to pour the particles into a larger bowl positioned near the base of the arms. Both tasks are designed to involve significant coordination between the two arms. During sweeping, for example, scooping up particles and transporting them to the bowls requires carefully maintaining alignment between the tips of the spatulas to avoid dropping particles. Flipping, on the other hand, requires aligning the elbow surfaces against the sides of the box, and using friction to carry the box as it pours the particles into the bowl -- any subtle misalignment of the elbow against the sides can lead to the box slipping away or tipping over. For both tasks, we generate fixed datasets of 1,000 demonstration episodes using a scripted oracle with access to privileged state information, including object poses and contact points, which are not accessible to policies.
We train two Implicit BC \cite{florence2021implicit} baselines \textit{without} Residual Kinematics: one using a 12DoF Cartesian action space (6DoF end-effector pose for each arm), and another using a 14DoF joint action space (7DoF for each arm). From the quantitative results presented in Tab. \ref{table:bimanual-experiments}, we observe that the performance of Implicit BC on bimanual sweeping degrades substantially when using a joint action space. This is likely that learning the task (\i.e. distribution of oracle demonstrations) involves recognizing a number of Cartesian space constraints that are less salient in joint space: e.g., keeping the Cartesian poses of the spatulas aligned with each other while scooping and transporting particles, and maintaining z-height with the tabletop.
For bimanual flipping, on the other hand, we observe the opposite: where the performance of Implicit BC performs well with joint actions, but poorly with Cartesian actions. We conjecture that the task involves more whole-body manipulation, where fitting policies in joint space are more likely to result in motion trajectories that accurately conform to the desired elbow contact with the box.
\vspace{-.7em}
\begin{table}[h]
\centering
\caption{\scriptsize Performance Measured in Task Success (Avg $\pm$ Std \% Over 3 Seeds)}
\begin{tabular}[b]{@{}lcccccccc@{}}
\toprule
Task & Sweeping & Flipping \\
{\em{Oracle's action space}} & \em{cartesian} & \em{joints} \\
Policy Action Space & & \\
\midrule
Cartesian & 79.4 $\pm$ 2.1 & 38.6 $\pm$ 5.2\\
Joints & 44.3 $\pm$ 3.2 & \textbf{98.4 $\pm$ 1.4} \\
Joints + Cartesian (Ours) & \textbf{85.9 $\pm$ 1.5} & \textbf{97.5 $\pm$ 1.2} \\
\bottomrule
\end{tabular}
\vspace{-1.0em}
\label{table:bimanual-experiments}
\end{table}
Our proposed method, Implicit Kinematic Policies (labeled as Joints + Cartesian in Tab. \ref{table:bimanual-experiments}) leverages both action spaces. Results suggest its performance is not only on par with the best bespoke action space for the task, but also surprisingly exceeds the performance of Cartesian actions for bimanual sweeping. Upon further inspection of the action trajectories for bimanual sweeping, it is evident that the Cartesian space action trajectories (Fig. \ref{fig:tasks}b) are lower frequency and contain more piecewise linear structure in comparison to the joint space action trajectories (Fig.~\ref{fig:tasks}a). Implicit policies have shown to thrive in conditions where this structure is present \cite{florence2021implicit}. In contrast, for bimanual flipping, the joint space action trajectories (Fig.~\ref{fig:tasks}c) appear much more linear and low frequency than the corresponding arc-like Cartesian trajectories (Fig.~\ref{fig:tasks}d). We also add small perturbations to the end-effector pose highlighted by the semi-transparent lines in plot (d) of Fig.~\ref{fig:tasks}. Note that the Implicit (Cartesian) policy uses a generic IK solver which can return multiple distinct joint target commands, causing less stable trajectories overall.
\begin{figure}[t]
\centering
\noindent\includegraphics[scale=0.2]{images/Insertion+Sweeping.png}
\vspace{-0.5em}
\caption{Simple constant joint encoder errors (which may appear from drift in low-cost or cable-driven robots) can propagate to non-linear offsets in Cartesian space. By learning joint space residuals to compensate for these offsets, IKP is better equipped to generalize over such errors than standard end-to-end policies trained in Cartesian space for tabletop manipulation.}
\label{fig:residual-deltas}
\vspace{-1.5em}
\end{figure}
\subsection{\textbf{Simulated Miscalibrated Sweeping and Insertion}}
\label{subsec:residual-experiments}
We design two additional tasks to evaluate IKP's ability to learn high precision tasks in the presence of inaccurate joint encoders, described in Sec.~\ref{subsec:residual-formulation}. To replicate such a scenario, we simulate these errors by adding constant $<$ 2 degree offsets to each of the six revolute joints on a UR5e robot. These offsets are visualized and described in more detail in Fig.~\ref{fig:residual-deltas}. We test the effect of these offsets on the bimanual sweeping task and a new insertion task where the goal is to insert an L-shaped block into a tight fixture. Demonstrations for both tasks are provided in the form of a Cartesian scripted oracle with privileged access to the underlying joint offsets and can compensate for them. This is akin to a human using visual servoing to compensate for inaccurate encoders when teleoperating a real robot to collect expert demonstrations. Although both tasks are generally well-defined in Cartesian space as a series of linear step functions, the joint offsets induce high-frequency non-linear artifacts in the end-effector trajectory (Fig.~\ref{fig:residual-deltas}b and \ref{fig:residual-deltas}d) causing poor performance with the Implicit (Cartesian) policy (results in Tab.~\ref{table:noisy-experiments}). Poor performance persists for miscalibrated insertion despite using 10x the data. Alternatively, even though the encoder offsets only cause linear shifts in the joint trajectories (Fig.~\ref{fig:residual-deltas}a and \ref{fig:residual-deltas}c), both tasks are less structured in joint space and as a result Implicit (Joints) performs significantly worse on miscalibrated bimanual sweeping and comparatively worse on miscalibrated insertion. Implicit (Joints) performs relatively well on the miscalibrated insertion task due to the simplicity of the pick and place motion, but struggles to generalize when the block nears the edge of workspace as the joint trajectories become increasingly non-linear in those regions. The performance only slightly improves when using 1000 demonstrations. IKP provides a best-of-all-worlds solution by learning the linear offsets in joint space through the residual blocks while simultaneously exploiting the unperturbed Cartesian trajectories through forward kinematics on the shifted joint actions to generalize. IKP achieves the highest performance on miscalibrated insertion with both 100 and 1000 demonstrations, and significantly higher performance on miscalibrated bimanual sweeping. Interestingly, not only do Explicit and ExplicitFK perform significantly worse than both Implicit (Joints) and IKP for both tasks, but ExplicitFK also provides little to no additional benefit over Explicit, even in the presence of more data.
\begin{table}[h]
\centering
\caption{\scriptsize Miscalibrated Joint Encoder Experiments, Performance Measured in Task Success (Avg $\pm$ Std. \% Over 3 Seeds). (``J+C'') is short for Joints+Cartesian.}
\begin{tabular}[b]{@{}lcccccccc@{}}
\toprule
Task & Sweeping & \multicolumn{2}{c}{Block Insertion}\\
{\em{Oracle's action space}} & \em{cartesian} & \multicolumn{2}{c}{\em{cartesian}}\\
Method (Action Space) & 1000 & 100 & 1000\\
\midrule
Explicit (Joints) & 38.2 $\pm$ 3.4 & 73.8 $\pm$ 4.3 & 74.2 $\pm$ 2.5\\
ExplicitFK (J + C) (Ours) & 40.3 $\pm$ 4.6 & 72.4 $\pm$ 3.8 & 75.4 $\pm$ 2.8\\
Implicit (Cartesian) & 3.1 $\pm$ 2.2 & 0.0 $\pm$ 0.0 & 0.0 $\pm$ 0.0\\
Implicit (Joints) & 46.3 $\pm$ 1.8 & 82.3 $\pm$ 2.9 & 85.2 $\pm$ 3.1 \\
IKP (J + C) (Ours) & \textbf{84.5 $\pm$ 1.2} & \textbf{88.8 $\pm$ 3.4} & \textbf{92.4 $\pm$ 2.6}\\
\bottomrule
\end{tabular}
\vspace{-1.0em}
\label{table:noisy-experiments}
\end{table}
\subsection{\textbf{Real Robot Sorting, Sweeping, and Alignment}}
We conduct qualitative experiments with a real UR5e robot on two tasks: 1) sorting two blocks into bowls and subsequently sweeping the bowls, and 2) aligning a red block with a blue block, where both tasks have initial locations randomized). Our goal with these experiments is two-fold: i) to demonstrate that IKP can run on real robots with noisy human continuous teleop demonstrations with as few as 100 demonstrations, and ii) to show that Implicit Kinematic Policies can perform both whole-body and prehensile tasks. Demonstrations for the block sorting portion of the first task are provided using continuous Cartesian position control based teleoperation at 100 Hz with a 3D mouse. Once the blocks are sorted, the teleoperator switches to a joint space PD controller with the same 3D mouse to guide the arm into an extended pose in order to sweep the bowls. For the alignment task, demonstrations are also provided using x, y and z Cartesian PD control at 100 Hz. Both tasks take 480${\times}$640 RGB images (downsampled to 96${\times}$96) from an Intel RealSense D435 camera at a semi-overhead view as input to the policy, and no extrinsic camera calibration is used. We show successful rollouts for both tasks in Fig.~\ref{fig:real-tasks}.
\begin{figure}[t]
\centering
\noindent\includegraphics[scale=0.23]{images/Real_Robot_Figure.png}
\caption{Real robot IKP rollouts on a sweeping and sorting task (top two rows) and a block alignment task (bottom row).}
\label{fig:real-tasks}
\vspace{-2em}
\end{figure}
\section{Discussion}
In future work, we will investigate using the multi-action-space formulation to extend beyond joint and cartesian PD control by incorporating the forward and/or inverse robot dynamics into the network which will allow us to expose joint torques and velocities to our model in a fully differentiable way. We hypothesize that this additional information will be particularly helpful in dynamic manipulation tasks and visual locomotion where torque and velocity action trajectories may appear more structured. We would also like to further test our residual framework within IKP on robots that have non-linear drift in joint space which may represent a larger set of low-cost robots. Finally, we would like to utilize our fully differentiable residual forward kinematics module to learn the link parameters themselves, which can be a promising direction for controlling soft and/or continuum robots.
\section*{Acknowledgments}
\small The authors would like to thank Vincent Vanhoucke, Vikas Sindhwani, Johnny Lee, Adrian Wong, Daniel Seita and Ryan Hoque for helpful discussions and valuable feedback on the manuscript.
\bibliographystyle{IEEEtranS}
| 54712c99fcf7ca56b9c80d1756371dada4c84fd4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction \label{sec:intro}}
Chemical systems with many close-lying electronic states or, more generally, strongly correlated electrons pose a significant challenge for modern electronic structure theories in computational quantum chemistry~\cite{lowdin1958correlation,sherrill2007bond,krylov2007breaking,stein2014seniority,gaggioli2019beyond}. When transition metals or heavier elements are involved, degenerate and
nearly degenerate electronic states are common, and single-reference electronic structure methods such as Kohn--Sham density functional theory often fail~\cite{neese2009prediction,jacob2012spin,yu2016perspective}. In these situations one has to use multireference methods to generate multiconfigurational wave functions and accurately describe these near degeneracies~\cite{szalay2012multiconfiguration,park2020multireference,cramer2006theoretical}.
Scientists also want to compute properties of large chemical systems or solids with accurate quantum chemistry methods, in spite of steep computational requirements. One way to achieve such computations is to use fragmentation methods. Many variations of fragmentation methods exist~\cite{gordon2012fragmentation, collins2015energy, raghavachari2015accurate, fedorov2011geometry}, but the common feature is that a large molecular system is divided into fragments and quantum-mechanical calculations are performed on the fragments. An especially important case is the application of fragmentation methods to multireference wave functions because of the exponential explosion of the computational cost with respect to the size of the active space of electronic configurations.
In the complete active space self-consistent field (CASSCF) method~\cite{Roos1980}, all the electronic configurations that can be formed for a given number of active electrons distributed in a given number of
active orbitals are included in the wave function. Thus, the wave function scales exponentially with the number of active electrons and orbitals, and the method has only limited application to chemically relevant systems. If one wants to study systems containing, for example, several transition metals~\cite{LiManni2021,hallmen2019toward,sharma2019multiple,hogue2018spin,malrieu2014magnetic}, the active site of a protein~\cite{Levine2020}, or extended organic chains in their ground and excited states~\cite{Levine2020,Sharma2019b}, more affordable multireference methods have to be developed. This is one of the major challenges of modern electronic structure theory.
Reducing the computational cost of CASSCF or other multiconfiguration self-consistent field calculations is pursued both in the development of new well-motivated theoretical approximations and in the application of new developments in computational hardware~\cite{Hohenstein2015,Snyder2016}. On the theoretical side, one strategy is to identify subspaces of the CAS that can be treated on different footings~\cite{olsen1988determinant,malmqvist1990restricted} or interact with one another only weakly~\cite{ma2011generalized,Ivanic2003,Parker2014,Nishio2019,Kathir2020}.
\change{The localized active-space self-consistent field (LASSCF) method~\cite{Hermes2019,Pandharkar2019a,Hermes2020,pandharkar2021localized}, also known as the cluster mean-field (cMF) method,~\cite{Jimenez-Hoyos2015} is an example of such a strategy}. LASSCF is designed for applications in which electrons are strongly correlated in different weakly interacting physical regions of a molecule and approximates the strongly correlated part of the wave function as a single antisymmetrized product of subspace wave functions. The computational cost of LASSCF is a linear function of the number of such unentangled subspaces.
Some of the authors have recently shown that LASSCF accurately reproduces the CASSCF spin-state energy gaps of bimetallic compounds and the simultaneous dissociation of two double bonds in bisdiazene at a significantly reduced cost~\cite{Pandharkar2019a,Hermes2020}. However, LASSCF fails to recover any electron correlation between fragments, for example in the \textit{cis-trans} isomerization of stilbene and similar systems~\cite{pandharkar2021localized}.
\change{Moreover, methods to restore the missing correlation variationally~\cite{Abraham2020c}, perturbatively~\cite{Jimenez-Hoyos2015,Papastathopoulos-Katsaros2021}, or \emph{via} the coupled-cluster (CC) approach~\cite{Wang2020b} on classical computers must usually enumerate a general many-body basis for each fragment. That is, they inherit the complications of multireference perturbation and CC theory~\cite{Lyakh2012,park2020multireference} over traditional single-reference perturbative or truncated coupled-cluster (CC) corrections based on second quantization~\cite{Moller1934,ShavittBartlett_textbook}.}
Recently, the development of quantum computers has led to an increased interest in novel quantum algorithms, especially for computational quantum chemistry, which is widely seen as a potential ``killer app'' of quantum computers~\cite{cao2019quantum,head2020quantum,alan_2020_review}. \change{The quantum phase estimation (QPE) quantum algorithm~\cite{lloyd1996universal} can potentially offer exponential speedups when large fault-tolerant quantum computers are available~\cite{kitaev1995quantum,abrams1999quantum},
under the assumption that an initial state with non-negligible overlap can be prepared~\cite{kitaev2002classical,o2021electronic}.
Additionally, the variational unitary coupled cluster (UCC) requires
only a polynomial number of gates to represent on a quantum computer, whereas representing the same ansatz classically has no known polynomial solution~\cite{romero2018strategies,peruzzo2014variational}.}
For the noisy, intermediate-scale quantum (NISQ)~\cite{preskill2018quantum} devices that we have today, these algorithms are not tenable, since they require coherence times far beyond what is available. Variational algorithms, such as the variational quantum eigensolver (VQE)~\cite{peruzzo2014variational}, have been used to perform calculations of the ground state energy of small molecules, with limited accuracy, on NISQ devices~\cite{wecker2015progress,mcclean2017hybrid,kandala2017hardware}. Quantum algorithms that have less stringent requirements compared with full QPE, and at the same time accuracy beyond that demonstrated by variational algorithms such as VQE, will be required to productively use the progressively larger and higher-quality quantum devices as they become available in the next few years.
In this paper we describe a framework for such quantum algorithms, inspired by classical LASSCF. The wave function within a fragment is solved by using one method (e.g., QPE), and correlation between fragments is encoded variationally by using an ansatz that entangles the fragments. This approach goes beyond what can be achieved with classical fragment methods, such as LASSCF, by providing additional correlation between fragments, while significantly reducing the total computational time (estimated via the number of gates) compared with full
QPE.
\section{Theory \label{sec:theory}}
\subsection{Multireference Methods with Exponential Scaling}
We seek to find the ground state of the second-quantized molecular Hamiltonian for a given number of $M$ electrons,
\begin{equation}\label{qchem_ham}
\hat{H} = h^p_q \crop{p}\anop{q} + \frac{1}{4} h^{pr}_{qs} \crop{p}\crop{r}\anop{s}\anop{q},
\end{equation}
where $\crop{p}$ ($\anop{p}$) creates (annihilates) an electron in spin orbital $p$; $h^p_q$ and $h^{pr}_{qs}$ are the one- and antisymmetrized two-electron Hamiltonian matrix elements, respectively; and repeated internal indices are summed. Generally, for $N$ spin orbitals, $\hat{H}$ has a
sparse-matrix representation in a space of size $O\binom{N}{M}$
and has $O(N^4)$ elements. Full-configuration interaction (FCI) determines the exact energy within a given one-electron basis set
(the FCI energy) at exponential cost. Methods such as CASSCF (and its restricted \cite{olsen1988determinant,malmqvist1990restricted} and generalized \cite{fleig2001generalized,ma2011generalized} active space approximations) or
selected configuration interaction~\cite{li2020accurate,li2018fast}, can go beyond FCI in system size, maintaining comparable accuracy, but still scale exponentially. The density matrix renormalization group~\cite{olivares2015ab,knecht2015new,kurashige2011second,marti2011density} and coupled cluster methods~\cite{ShavittBartlett_textbook} can scale polynomially
but introduce (sometimes
uncontrollable) approximation errors. Here we briefly
describe the LASSCF algorithm~\cite{Hermes2019,Hermes2020}, which
will serve as the basis for our fragment-based quantum algorithms.
\subsection{LASSCF \label{sec:theory_lasscf}}
In LASSCF, the wave function of a molecule is approximated as
\begin{equation}
\ket{{\rm LAS}} = \bigwedge_K \ket{\Psi_K} \wedge \ket{\Phi}, \label{eq:las_wfn}
\end{equation}
where $\ket{\Psi_K}$ is a general many-body wave function describing $M_K$
electrons occupying $N_K$ active orbitals of the $K$th ``fragment'' or ``active
subspace,'' $\ket{\Phi}$ is a single determinant spanning the complement of the
complete active space, and the wedge operator (``$\wedge$'') implies an
antisymmetrized product.
In the variational\cite{Hermes2020} implementation of LASSCF, this wave function
is obtained by minimizing the LAS energy,
\begin{equation}
E_{\rm{LAS}} = \braket{{\rm LAS}|\hat{H}|{\rm LAS}}, \label{eq:elas}
\end{equation}
with respect to all orbital rotations and configuration interaction (CI) vectors defining $\ket{\rm{LAS}}$. This is accomplished by introducing a unitary operator (see the Supporting Information of Ref.\ \onlinecite{Hermes2020}) that is parameterized in terms of all nonredundant transformations of the orbitals and CI vectors,
\begin{equation}
\ket{{\rm LAS}} \to \hat{U}_{\mathrm{orb}}\prod_K \hat{U}_{\mathrm{CI},K}\ket{{\rm LAS}}, \label{eq:ulas}
\end{equation}
where
\change{
\begin{eqnarray}
\hat{U}_{\mathrm{orb}} &=& \exp{x^k_l\left(\crop{k}\anop{l}-\crop{l}\anop{k}\right)}, \label{eq:uorb} \\
\hat{U}_{\mathrm{CI},K} &=& \exp{x_{\vec{k}}\left(\ket{\vec{k}}\bra{\Psi_K}-\ket{\Psi_K}\bra{\vec{k}}\right)}, \label{eq:ucik}
\end{eqnarray}
}
where $k,l$ index individual spin orbitals in two different subspaces (including the
inactive and virtual subspaces outside of the CAS) and where $\ket{\vec{k}}$ is a determinant or configuration state function. First and second derivatives of Eq.\ (\ref{eq:elas}) with respect to the generator amplitudes ($x^{k}_{l}$ and $x_{\vec{k}}$) are obtained by using the Baker--Campbell--Hausdorff (BCH) expansion, and the energy is minimized by repeated applications of the preconditioned conjugate gradient (PCG) method~\cite{Bernhardsson1999,Stalring2001}.
The orbital unitary operator, $\hat{U}_{\mathrm{orb}}$, corresponds to the
UCC correlator truncated after the first (``singles'')
term:
\begin{eqnarray}
\hat{U}_{\rm UCC} &\equiv& \exp{\hat{T}_{\rm UCC}}, \label{eq:ucc} \\
\hat{T}_{\rm UCC} &\equiv& x^k_l \left(\crop{k}\anop{l}-\mathrm{h.c.}\right)
+ \frac{1}{4}x^{km}_{ln} \left(\crop{k}\crop{m}\anop{n}\anop{l} - \mathrm{h.c.}\right) + \ldots. \label{eq:tcc}
\end{eqnarray}
\change{The use of the more general cluster operator, Eq.\ (\ref{eq:tcc}), in place of the orbital rotation unitary operator, Eq.\ (\ref{eq:uorb}), corresponds to a multireference unitary coupled cluster method~\cite{Hoffmann1988} built on top of a $\ket{\rm LAS}$ reference wave function. Such a method is expected to more flexible than LASSCF itself, in that doubles and higher-order cluster amplitudes could encode electron correlation and entanglement between active subspaces. This would require the reference
wave function, $\ket{\rm LAS}$, to be updated by explicit exponentiation of
the general cluster operator, Eq.\
(\ref{eq:tcc}), after each execution of the PCG algorithm. On classical
computer hardware, however, this is not an efficient way to extend LASSCF.}
\subsection{LAS Methods on Quantum Computers \label{sec:theory_qlas}}
Here we describe an algorithm for molecular calculations that goes beyond the limited accuracy of standard VQE~\cite{kandala2017hardware,mcclean2017hybrid},
while having dramatically reduced computational complexity compared with QPE [see Methods section]. The
algorithm exploits the structure of the molecule by separating it into
coupled fragments, as is done in the classical algorithm, LASSCF. The quantum
algorithm, however, goes beyond classical LASSCF by providing some degree of
entanglement between the fragments.
The algorithm begins by segmenting the orbital active space of a given molecule into distinct fragments defined by non-overlapping orbital subspaces, as in classical LASSCF. For instance, orthogonalized atomic orbitals generated by using the meta-L\"{o}wdin method\cite{Sun2014} can be sorted into localized fragments and then projected onto a guess for the CAS of a given molecule to produce localized active orbitals. We construct an effective Hamiltonian that omits non-mean-field interfragment interactions, resulting in a sum of local fragment Hamiltonians,
\begin{equation}
\hat{H}_{{\rm eff}} = \sum_K^{n_f} \Bigg( \tilde{h}^{k_1}_{k_2} \crop{k_1}\anop{k_2} + \frac{1}{4} h^{k_1k_3}_{k_2k_4} \crop{k_1}\crop{k_3}\anop{k_4}\anop{k_2}\Bigg),
\end{equation}
where $k_1,k_2,\ldots$ index distinct active orbitals of the $K$th fragment and where
\begin{eqnarray}
\tilde{h}^{k_1}_{k_2} &=& h^{k_1}_{k_2} + h^{k_1i}_{k_2i}
+ \sum_{L\neq K} h^{k_1l_1}_{k_2l_2}
\gamma^{l_1}_{l_2},
\label{eq:h1eff}
\end{eqnarray}
where $i$ and $l_n$ index respectively inactive orbitals [i.e., those defining $\ket{\Phi}$ in Eq.\ (\ref{eq:las_wfn})] and active orbitals of the $L$th fragment and where $\gamma^{l_1}_{l_2}$ is a \change{one-electron reduced} density matrix element for spin orbitals $l_1$ and $l_2$\change{,
\begin{equation}
\gamma^{l_1}_{l_2} \equiv \braket{{\rm LAS}|\crop{l_1}\anop{l_2}|{\rm LAS}} = \braket{\Psi_L|\crop{l_1}\anop{l_2}|\Psi_L}.
\end{equation}}
Given a set of localized active orbitals that minimize the LASSCF energy, if the density matrices in Eq.\ (\ref{eq:h1eff}) are obtained from a classical LASSCF calculation on the same system, then the QPE algorithm applied to $\hat{H}_{{\rm eff}}$ generates the active-space part of the LASSCF wave function, $\ket{{\rm QLAS}} = \bigwedge_K \ket{\Psi_K}$, on the quantum computer. The same result is achieved if density matrices are obtained self-consistently from the QPE evaluation. If the density matrices are obtained in some other way, for instance from $\ket{{\rm HF}}$, then an approximation to the LASSCF wave function is obtained.
\change{The QPE step provides the initial $\ket{{\rm QLAS}}$ for each fragment step by repeating the measurement of the phase until it is consistent with the phase representing the ground state energy,
which collapses the system into the ground state wavefunction. This introduces some overhead, as each fragment will need to be in the ground state to continue to the next step. Furthermore,
a full QPE solve, estimating the ground state energy, must be performed initially to provide a comparison value.}
A sequence of UCC with singles and doubles (UCCSD) circuits, with variable parameters, is then
applied across $m$ fragments each (which we term $m$-local),
leading to the LAS-UCC wave function,
\change{
\begin{equation}
\ket{{\rm QLAS}(\bm{x})} \rightarrow \prod_\zeta \hat{U}_{{\rm UCCSD},\zeta}(\bm{x}) \ket{{\rm QLAS}}, \label{eq:mlocal}
\end{equation}
where $\hat{U}_{{\rm UCCSD},\zeta}(\bm{x})$ is the UCCSD ansatz including
only creation/annihilation operators within the $m$ fragments
that it spans, $\zeta$ is a list of fragment indices of size $m$, and $\bm{x}$ are the associated singles and doubles
cluster amplitudes.}
\change{The factorization of Eq.\ (\ref{eq:tcc}) implied by Eq.\ (\ref{eq:mlocal}) is based on the intuition that physically adjacent active subspaces are likely to be more strongly entangled to one another than subspaces on opposite ends of a large molecule.}
The parameters of the UCCSD circuit are varied to minimize the total energy of the full system, as in VQE \change{[see also Methods section]:
\begin{equation}\label{eq:vqe}
E = \min_{\bm{x}} \braket{ {\rm QLAS}(\bm{x}) | \hat{H} |{\rm QLAS}(\bm{x}) }.
\end{equation}
}
A schematic representation of the described circuit is shown in Fig.~\ref{fig:circuit}.
This provides electron correlation between the fragments, in a way
that scales exponentially on classical computers, but only polynomially on
quantum computers. Moreover, this procedure provides a better
estimate of the ground state energy than the product wave function or the UCCSD would
provide alone. Note that, unlike LASSCF, this method is not strictly variational (despite the use of VQE) because the initial product-state wave function, $\bigwedge_K \ket{\Psi_K}$, is not variationally reoptimized in the presence of the UCCSD correlators.
The QPE circuits could also be replaced with a local variational ansatz, leading to a fully variational algorithm, which we term LAS-VQE and describe in the Supplementary Information.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figs/circuit.png}
\caption{Diagram of example circuit using LAS-UCC. The system of interest is first separated into distinct fragments. QPE is used on each fragment to solve for the approximate unentangled ground state. Correlation between fragments is then added in, variationally, through a unitary coupled cluster ansatz.}
\label{fig:circuit}
\end{figure}
To understand the large improvement in computational complexity of our approach,
we focus on a system of $n_f$
fragments, with the number of orbitals per fragment, $N_K$, constant as the number of fragments
grows. The total system size is defined by
$N=N_K n_f$ orbitals. We also
assume that each fragment interacts with only the $m$ geometrically nearest fragments and
that $m$ does not grow with $n_f$. These are reasonable assumptions for many
interesting molecules and mirror the assumptions made in classical LASSCF.
Under these assumptions, the QPE solver for the unentangled fragments does not
grow with $N$, since $N_K$ is
assumed to be fixed while $n_f$ grows.
The number of small QPE sections grows linearly with the number of fragments, of course.
Typically, the Jordan--Wigner transformation would introduce
an $O(N)$ term
to enforce the anticommutation relations among the orbital creation
and annhilation operators.
However, in the case of
linear chains, as we study here, ordering the orbitals such that all up and down occupied and virtual orbitals in a given fragment
are close, the high-weight $Z$ part of the Jordan--Wigner transformation effectively cancels out, causing no scaling with total number of orbitals. See Supplementary Information for more details. Together, this
leads to an overall
$O( n_f N_K^4) \approx O(N)$ (linear) number of gates to solve for the
$n_f$ unentangled product wave functions.
The UCCSD correlator, which is then applied, has $O\big(m^4 N_K^4)$ terms in the cluster
operator for each correlator, because the UCCSD circuit spans only $m$ fragments.
Neither $m$ nor $N_K$ grows with the total size (number of spin orbitals) of the system, $N$. The number of $m$-local correlators grows as $O(n_f)$.
Again, by careful ordering of the orbitals, the
Jordan--Wigner transformation does not introduce any scaling overhead.
The complexity of the $m$-local UCCSD correlator is then
$O(n_f m^4 N_K^4) \approx O(N)$ (linear).
This creates an overall linear scaling in the number of gates
for linear chain geometries, with respect to only the total size of
the system, $N$, and is polynomially ($O(N^4)$) better than performing QPE alone, while
providing accuracy above VQE using the UCCSD ansatz and classical LASSCF. Many of the
gates can be done in parallel,
such as the local QPE circuits and the
different $m$-local UCCSD correlators, leading to an
expected overall sub-linear depth. If the fragments are
coupled in a geometry more complicated than a linear chain, the UCCSD
correlator will potentially incur the $O(N)$ Jordan-Wigner overhead,
leading to an overall $O(N^2)$ scaling for arbitrary geometries with
an expected $O(N)$ depth.
\subsection{Illustrative Molecular Systems}
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figs/molecules.png}
\caption{Two model systems used for testing. (a) The asymmetric hydrogen dimer, $(\rm{H}_2)_2$. Each H$_2$
molecule is a fragment described by a 2-electron, 2-spatial orbital or (2,2) active subspace in the dimer's LAS wave function. The potential
energy surface is scanned along the distance between the two H$_2$ bond
midpoints, indicated by the black double line.
(b) The \emph{trans}-butadiene molecule at its CASSCF(8,8)/6-31G
ground-state equilibrium geometry. Dashed boxes depict the two
notional fragments containing the two (4,4) active subspaces in the LAS
wave function. Black double lines indicate the internal coordinate along which
the potential energy surface is scanned; the two terminal methylene units
are simultaneously removed from the
central acetylene unit.}
\label{fig:molecules}
\end{figure}
In the calculations discussed below, we consider two systems, depicted in
Fig.~\ref{fig:molecules}. The first, shown in Fig.~\ref{fig:molecules}(a),
is a simplistic model of weakly interacting fragments, consisting of two H$_2$
molecules at various distances between their two midpoints using a minimal STO-3G
atomic orbital (AO) basis set, and the two active subspaces in the LAS wave
function correspond to the active spaces of the two H$_2$ molecules. We use this small basis set because of the size limitations of today's quantum computers and simulations. The bond lengths and internal
angles of this system are set arbitrarily to remove point group symmetry so
that differences between various methods are not obscured by the simplicity of a
symmetrized electronic wave function. The interaction between the two fragments
in this model system are weak, and the LAS wave function is therefore expected to
provide an excellent model of the FCI wave function except when the distance
between the two molecules is very small. We additionally extend this system
up to 20 H$_2$ in a linear chain, where we estimate only the
total number of quantum resources necessary.
The second system, depicted in Fig.~\ref{fig:molecules}(b), is the
\emph{trans}-butadiene molecule. The
potential energy surface of this molecule is scanned along the internal
coordinate corresponding to the simultaneous stretching of both the C=C
double bonds, leading to the removal of two methylene units from a
central C$_2$H$_2$ (distorted acetylene-like) unit. In the LAS wave function, the molecule is divided
into two fragments split across the central C--C bond, and each fragment
is described by a (4,4) active subspace. Several molecular orbitals are therefore
left inactive, described by an unfragmented single determinant.
We employed the 6-31G AO basis set in this case.
The \emph{trans}-butadiene system is a chemical model of the case of two strongly interacting units in a system, where
the value of the stretching internal coordinate is a proxy for the strength of
electron correlation. Near the equilibrium geometry, dividing the
active space into two fragments is chemically reasonable: each fragment encloses
one $\pi$-bond, and inasmuch as electron correlation affects the system at all,
it is a reasonable approximation to consider it only locally. However, as the
C=C double bonds are elongated, electrons from the two broken $\pi$ bonds
recouple across the central C$_2$H$_2$ unit, which spans the fissure between
the two LAS fragments. The LAS wave function cannot model a $\pi$ bond in this
position, and the LASSCF method breaks down.
\section{Results and Discussion \label{sec:results}}
\begin{figure}
\centering
\includegraphics[width=0.96\linewidth]{figs/h4_results.png}
\caption{Energies for (H$_2$)$_2$ calculated by CASCI, LASSCF, and LAS-UCC. The inset shows the error, with respect to CASCI, of LASSCF and LAS-UCC. The black dashed line represents chemical accuracy.
LAS-UCC is able to obtain chemical accuracy, with respect to CASCI, at all distances. LASSCF cannot obtain chemical accuracy at sufficiently short distances. }
\label{fig:h4}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.96\linewidth]{figs/c4h6_results.png}
\caption{Energies for C$_4$H$_6$ calculated by CASCI, LASSCF, and LAS-UCC. The inset shows the error, with respect to CASCI, of LASSCF and LAS-UCC. The black dashed line represents chemical accuracy. LAS-UCC obtains chemical accuracy across the potential energy surface, whereas LASSCF, which cannot accurately represent the correlation between the fragments, fails to obtain chemical accuracy for most points.}
\label{fig:c4h6}
\end{figure}
\subsection{LAS-UCC}
We demonstrate the efficacy of our framework by simulating the two benchmark
molecules, (H$_2$)$_2$ and \emph{trans}-butadiene, described above.
We compare three methods: LASSCF, CAS configuration interaction in the basis of LASSCF orbitals (CASCI),
and our new algorithm, LAS-UCC. LASSCF
represents the best unentangled set of wave functions and is equivalent to the
solution after the QPE circuits but before the use of the UCCSD ansatz. Note that
CASCI is slightly different from CASSCF since the orbitals are not variationally
reoptimized. CASCI solves for the FCI wave function within the active space; in
this case, it is equivalent to using QPE across the whole molecule
and represents the reference result in these studies.
Figure~\ref{fig:h4} shows the results of applying the methods to the hydrogen dimer as the two H$_2$ molecules are
pulled apart. We see that LASSCF, CASCI, and
LAS-UCC agree except for very small distances where LASSCF no longer provides accurate
energies.
Figure~\ref{fig:c4h6} shows the results for \emph{trans}-butadiene,
a model of strongly correlated fragments. Here, as the terminal methylene units
are removed, the interfragment correlation grows as a double bond is formed
between the fragments. The UCCSD ansatz can accurately represent this level of
entanglement, allowing LAS-UCC to achieve nearly CASCI accuracy, whereas LASSCF
fails to account for this entanglement. \change{With a standard Hartree-Fock initial state, as is typically done in VQE, the UCCSD
ansatz is unable to obtain chemical accuracy for the large distances.
We also attempted to use the so-called `hardware-efficient' ansatz~\cite{kandala2017hardware},
but were unable to obtain results significantly better than Hartree-Fock using depths up to 10 (which corresponds to a similar
number of parameters as the UCCSD ansatz) at equilibrium.}
\subsection{Resource Estimates}
To demonstrate the scaling advantage of our method, we perform resource estimation for the number
of logical quantum gates necessary for several different quantum algorithms: the QPE algorithm over the
full unfragmented molecule; the UCCSD ansatz over the full unfragmented molecule; and the two
steps of our proposed LAS-UCC method, the fragmented QPE
and the 2-local UCCSD (which corresponds to
the circuit depicted in Fig.~\ref{fig:circuit}). We estimate the number of resources
needed for the QPE algorithm if only a single Trotter time step were needed; $O(1000)$
time steps will be needed for typical systems to get to chemical accuracy~\cite{alan_2020_review,
garnet_2020_review}.
Note that these estimates represent only the number of two-qubit CNOT gates, which we use as a
primary gauge of the number of total resources.
Single-qubit gates are also necessary; the
estimates for these resources can be found in the Supplementary Information and scale similarly to the number of CNOT gates.
\change{We also note here that we are only comparing the scaling number of gates;
QPE, with a sufficiently good initial state and enough Trotter states, will of course be the most accurate
of all compared algorithms.}
We use a model system of
an increasing number of H$_2$ molecules and look at how the number of CNOT gates increases
as the number of molecules increases, as shown in Fig.~\ref{fig:gate_counts}.
As the number
of H$_2$ molecules increases, the number of gates needed for all methods also increases. As
predicted in the complexity analysis of QPE [see Methods section], the
total number of gates for a single Trotter step in the QPE algorithm grows as $O(N^5)$.
Similarly, the number of gates needed for a global UCCSD ansatz also grows as $O(N^5)$,
as expected~\cite{mcclean2017hybrid}. This result is compared with the much smaller number of gates necessary
to implement the two steps of our LAS-UCC algorithm. As expected, both the QPE and UCCSD
parts of LAS-UCC provide dramatic scaling advantages, with the 2-local UCCSD ansatz and the QPE of the reduced Hamiltonian both scaling as only $O(N)$.
We note that, in addition to evaluating the quantum circuits here, an
additional optimization loop is needed when using the UCCSD ansatz, whether it is
global or 2-local. Using a 2-local UCCSD ansatz also greatly reduces the number
of parameters that need to be optimized compared with a global UCCSD ansatz.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figs/gate_counts.png}
\caption{Estimated two-qubit gate counts using various algorithms. The QPE estimates assume only a single Trotter step; \change{$O(1000)$ will need to be taken to obtain chemical accuracy}.
Polynomials of various orders have been plotted to demonstrate the scaling. Our algorithm, LAS-UCC, requires both the LAS-QPE and 2-UCC circuits and thus has an overall $O(N)$ scaling, compared with the $O(N^5)$ scaling of UCC and QPE.}
\label{fig:gate_counts}
\end{figure}
\subsection{Discussion}
Here we compare LAS-UCC with the two quantum algorithms that it
is composed of: QPE and variational UCCSD.
Compared with global QPE, LAS-UCC reduces the total quantum resource cost by
approximating the system with noninteracting fragments and adding in some interaction between fragments (those described by a UCCSD ansatz spanning
the fragments). This in general reduces the accuracy; but as shown in the preceding sections,
LAS-UCC provides accuracy comparable to CASCI (and therefore global QPE) for the
systems considered here. The \emph{trans}-butadiene molecule is a model for larger, more complicated
systems of strongly interacting units. Many single molecular magnets have such pockets of strong correlation localized on the metal centers, which moderately interact with each other~\cite{doi:10.1021/ja0316824,baniodeh2018high}. With LAS-UCC we not only can obtain the wave function efficiently but also can selectively couple the fragments with the UCC correlator, offering further insight into the nature of these interactions. Affordable and accurate modeling of phenomena such as singlet fission \cite{smith2013recent,casanova2018theoretical} in molecular crystals of conjugated organic compounds can be performed with LAS-UCC, as fault-tolerant quantum computers become available.
This approach will also be used to study chemical processes involving interfragment bond formation and breaking while still treating all points on a potential energy surface at comparable footing.
Compared with standard UCCSD, LAS-UCC can be seen as augmenting UCCSD with a
multireference initial state. Instead of using single-determinant Hartree--Fock,
as is standard in VQE demonstrations of UCCSD~\cite{doi:10.1063/1.5133059,doi:10.1126/science.abb9811,mccaskey2019quantum,PhysRevX.6.031007,kandala2017hardware}, LAS-UCC uses the unentangled
product state of the ground state wave functions of each fragment (which is
also the LASSCF wavefunction). This provides additional accuracy,
above standard single-reference UCCSD, at a negligible increase in cost.
When using a global UCCSD
ansatz, the increase in the number of gates is negligible, even when taking
into account the $O(1000)$ time steps that would be needed to implement the
QPE step. Using the $m$-local ansatz provides further reduction.
\change{There have been other proposals for preparing interesting,
multireference initial state in context of efficiently finding
states with large overlap with the true ground state~\cite{tubman2018postponing,sugisaki2018quantum}. These
algorithms could be used in-place of the QPE part of LAS-UCC to provide
the initial state and potentially
adapted to give similar LASSCF-like states with similar overhead reductions as shown for QPE.}
Moreover, recent advances in VQE algorithms have developed various ways to reduce the cost associated with the UCC correlator~\cite{grimsley2019adaptive,alan_2020_review,cao2019quantum,cerezo2021variational,bharti2021noisy,fedorov2021unitary}.
As presented in the Theory section, LAS-UCC can also be seen as a post-LASSCF method that recouples select fragments at a level of theory beyond the mean field.
The addition of the doubles or higher terms in the cluster operator provides a way to systematically improve the accuracy beyond the LASSCF reference.
\change{On classical computers, such an approach requires truncating~\cite{Hoffmann1988} or approximating~\cite{Neuscamman2010} the non-terminating BCH expansion in a more or less arbitrary way.}
Not every system will be accurately described by LAS-UCC, of course, but one can systematically increase the accuracy in several ways, while
increasing the total resource cost. Increasing the
size of each fragment (which in turn decreases the number of fragments)
gradually increases the accuracy, until the limit of a single fragment,
where the UCCSD ansatz becomes redundant and the algorithm
becomes simply global QPE. On the UCC side, the order of the ansatz can be
increased. Triples, quadruples, and so on can be included at increasing cost. If
using an $m$-local ansatz, the scaling is unaffected, but the total
number of gates increases. The locality of the ansatz, $m$, can also be increased,
providing explicit correlation between more geometrically distant fragments.
\section{Conclusions \label{sec:conclusions}}
We introduced LAS-UCC, a quantum algorithm that combines a fragmentation of the wave function of a chemical system with QPE and variational UCCSD
to compute the ground state
energy of such a system. LAS-UCC can describe compounds containing strongly interacting
fragments, and it provides a polynomial scaling advantage in
the number of quantum gates compared with
other quantum algorithms such as QPE and UCCSD. Since the fragments' reduced Hamiltonians have fewer terms and by ensuring the locality of the Jordan-Wigner transform, the overall gate count will be $O(N)$
with respect to the total size of
the system $N$ for for linear geometries and $O(N^2)$ more generally,
compared with $O(N^5)$ requirements for QPE. We also demonstrated the
accuracy of LAS-UCC on (H$_2$)$_2$ and \emph{trans}-butadiene molecules and performed resource
estimations of larger systems to \change{provide evidence for potential scaling advantages.}
As larger fault-tolerant quantum computers are developed, we expect
that our algorithm will be able to provide accurate calculations of
large and useful chemical systems, such as molecular magnets and qubits, photovoltaic materials, and large biomolecules that are out of reach of classical computing algorithms but
for which QPE would be too expensive.
\section{Methods}
\subsection{Quantum Algorithms}
Here we describe two quantum algorithms that serve as the primary
components for our fragment-based quantum algorithm.
\subsubsection{Quantum Phase Estimation}
The quantum phase estimation algorithm solves for the eigenvalue, $\lambda_k$, for
an eigenvector $|v_k\rangle$ of some unitary matrix, $U$. In addition to its
use in quantum chemistry, it forms the basis
for many important quantum algorithms, such as Shor's prime number factoring
algorithm~\cite{shor1999polynomial} and the Hassidim--Harrow--Lloyd algorithm for inverting
matrices~\cite{PhysRevLett.103.150502}. For quantum chemistry problems, the unitary matrix $U$ is
generated by the Hamiltonian, $H$ (eq.~\eqref{qchem_ham}), over time steps $\tau$:
\begin{equation}\label{qpe_phase}
U |v_k\rangle = e^{-i\hat{H}\tau} |v_k\rangle = e^{i 2\pi \phi} |v_k \rangle,
\end{equation}
and the desired energy is mapped to the phase acquired,~$E=-2\pi \phi/\tau$,
where units have been chosen such that $\hbar=1$. By combining real-time evolution
of the Hamiltonian, $\hat{H}$, with
application of the quantum Fourier transform
(QFT)~\cite{365700}, the
value of the energy can be obtained in polynomial time using a quantum computer.
The computational complexity of the QPE is directly related to the complexity of
implementing the unitary propagator $U=e^{-i \hat{H} \tau}$. Many strategies for implementing $U$ exist,
including Trotterization~\cite{PhysRevA.64.022319,PhysRevA.91.022311}, Taylorization~\cite{PhysRevLett.114.090502}, and
qubitization~\cite{Low2019hamiltonian}. The Hamiltonian, Eq.~\eqref{qchem_ham},
has $O(N^4)$ terms, where $N$ is the number of spin orbitals. Each term in the
Hamiltonian can be transformed into a Pauli string (that is,
a product of Pauli operators $X$, $Y$, $Z$, or $I$) via one of the many
fermion-to-spin transformations, such as the Jordan--Wigner~\cite{jordan1993paulische},
parity~\cite{bravyi2017tapering}, and Bravyi--Kitaev~\cite{BRAVYI2002210} transformations. In this
work we focus on QPE using Trotterization with
the Jordan--Wigner transformation since they serve as standard
reference points for the other variations. The complexity
of QPE for the Hamiltonian, Eq.~\eqref{qchem_ham},
using Trotterization with the Jordan--Wigner transformation is $O(N^5)$:
$N^4$ arising from the number of terms in the Hamiltonian and
an additional $N$ from the Jordan--Wigner transform.
Although QPE can obtain estimates of the ground state energy with
only a polynomial number of quantum gates, the overheads are still too large for
near-term quantum computers. The success of the QPE algorithm directly depends on the
overlap of the initial state (which is often taken to be the
Hartree-Fock state) and the true ground state. Realistic estimates, taking into account overheads
such as quantum error correction, put the needed number of qubits to perform QPE
on interesting molecules in the millions~\cite{elfving2020will,liu2021prospects,kim2021fault}.
QPE is analogous to a Fourier analysis of a correlation
function; and, for a given energy accuracy, $\epsilon$, it requires propagation
efforts (maximum times) on the order of $O(1/\epsilon)$~\cite{alan_2020_review,
garnet_2020_review}. Since the circuit depth for evaluating the propagator
for individual fragments will naturally be lower than for the full system,
the QPEs involved in our LAS approach will be significantly cheaper
than full QPE.
\subsubsection{Variational Quantum Eigensolver}
The variational quantum eigensolver is a hybrid quantum-classical
algorithm that relies on the variational principle to find an estimate of the ground state
energy of a given molecule. A circuit with variable parameters, $\theta$,
serves as an ansatz, whose energy is evaluated on a quantum computer and whose
parameters are iteratively optimized by a classical computer. For a circuit ansatz
$|\psi(\theta)\rangle$, VQE estimates the energy as
\begin{equation}\label{eq:vqe2}
E = \min_{\theta} \braket{ \psi(\theta) | \hat{H} |\psi(\theta) }.
\end{equation}
The Hamiltonian, $\hat{H}$, is transformed into a sum of Pauli strings via a
fermion-to-spin transformation, and
the expectation value of each term is measured from the quantum computer
separately and summed on the classical computer. VQE has
much less stringent quantum resource requirements than QPE has, since it
offloads much of the work (such as optimization) to the classical computer. Hence, VQE has been used in proof-of-principle calculations for small
molecules~\cite{doi:10.1063/1.5133059,mccaskey2019quantum,PhysRevX.6.031007}.
The accuracy of VQE is determined by the quality of the ansatz,
$|\psi(\theta)\rangle$. The UCCSD
ansatz is an interesting choice as wave function for VQE since there is no known way to efficiently implement UCCSD on classical computers~\cite{BARTLETT1989133,taube2006new,kutzelnigg1991error}, but it can be implemented with $O(N^5)$
gates on quantum computers~\cite{PhysRevA.95.020501,doi:10.1063/1.5011033,mcclean2017hybrid}. The UCCSD ansatz is
\begin{equation}
\ket{\psi_{{\rm UCCSD}}} = \hat{U}_{{\rm UCCSD}} \ket{{\rm HF}} = \exp{\hat{T}_{\rm UCCSD}} \ket{{\rm HF}},
\end{equation}
where $\hat{T}_{\rm UCCSD}$ is defined by truncating the more general
cluster operator of Eq.\ (\ref{eq:tcc}) at the second term. While the UCCSD ansatz can be implemented
on NISQ devices for small molecules~\cite{doi:10.1126/science.abb9811,kandala2017hardware}, it is limited
in its accuracy because of only including up to doubles excitations.
\subsection{Computational Methods}
To calculate the accuracy of the proposed method for small molecules,
we use the following strategy. We first use a classical LASSCF solver,
as implemented in the \textit{mrh} package~\cite{mrh_software}, to find the best product wave function.
This effectively provides an equivalent solution to that of the QPE step
of our proposed algorithm. We then represent this product wave function as a CI vector in the complete active Fock space and apply a UCCSD correlator, as well as its derivatives with respect to all amplitudes, to this reference CI vector. We employ the factorization reported by Chen et al.~\cite{Chen2020b} to avoid the BCH expansion and its inevitable approximate truncation. The resulting $\ket{{\rm QLAS}}$ CI vector and its derivatives ($\ket{\delta {\rm QLAS}}$) with respect to the unitary coupled cluster amplitudes are used to compute the energy, $\braket{{\rm QLAS}|\hat{H}|{\rm QLAS}}$, and its derivatives, $\braket{\delta{\rm QLAS}|\hat{H}|{\rm QLAS}}$. We then minimize the former using the latter and the Broyden–-Fletcher-–Goldfarb-–Shanno algorithm.
We find that this approach is more
efficient than directly simulating the quantum circuits. We note that this method
scales exponentially on classical computers.
To provide gate count estimates, we use the Q\# package~\cite{svore2018q}, generally
following the framework of Ref.~\onlinecite{low2019q}. The
full and reduced Hamiltonians are produced by using the \textit{mrh} package~\cite{mrh_software}, and
both Hamiltonians are then passed to the Q\# package to estimate the number of CNOT gates using
the QPE algorithm with a single Trotter time step for each.
Additionally, we estimate the number of CNOT gates
necessary to calculate various UCCSD ansatzes, including a global
UCCSD ansatz over the whole unfragmented molecule and multiple 2-local ansatzes
that span only two fragments. We count only the number of logical quantum
gates needed. Real quantum computers will require additional overheads,
owing to limited connectivity and the need to use expensive quantum error
correction protocols to deal with inevitable errors~\cite{liu2021prospects,kim2021fault}. Furthermore,
we provide gate counts only; no attempt was made to count gate depth, which
is typically smaller, because many gates can be implemented in parallel.
\section{Contributions}
L.G., S.G. M.O. and M.R.H. designed the project. M.O. wrote the quantum algorithm and tested it. M.R.H. wrote the LASSCF classical code. R.P. tested the codes and performed some of the calculations. Y.A. helped with the theory and suggested testing calculations. M.O. and M.R.H. wrote the initial draft of the manuscript.
All authors contributed to the scientific discussions and manuscript revisions.
\begin{acknowledgments}
This research is based on work supported by Laboratory Directed Research and Development (LDRD) funding from Argonne National Laboratory, provided by the Director, Office of Science, of the U.S. DOE under Contract No. DE-AC02-06CH11357.
This work was performed, in part, at the Center for Nanoscale Materials, a U.S. Department of Energy Office of Science User Facility, and supported by the U.S. Department of Energy, Office of Science, under Contract No. DE-AC02-06CH11357.
MRH and LG are partially supported by the U.S.
Department of Energy (DOE), Office of Basic Energy
Sciences, Division of Chemical Sciences, Geosciences, and
Biosciences under grant no. USDOE/DE-SC002183.
This material is based upon work supported by the U.S. Department of Energy, Office of Science, National Quantum Information Science Research Centers.
We gratefully acknowledge the computing resources provided on Bebop, a high-performance computing cluster operated by the Laboratory Computing Resource Center at Argonne National Laboratory and University of Chicago Research Computing Center.
\end{acknowledgments}
\section{LAS-VQE}
Here we describe a more approximate approach than LAS-UCC wherein the QPE circuits for the fragments are replaced by UCC ansatzes, leading to a fully variational method.
The chemical knowledge that guides us in defining subspaces of the LAS wave function can also be used to reduce the size of the unitary operator for the UCC ansatzes. It suggests that operators corresponding to the higher excitation from one fragment to another do not affect the wave function significantly. Thus, we introduce a modified
ansatz for truncated UCC circuits where we consider the “locality” of the excitation. An excitation involving only orbitals localized on a particular fragment---that is defined by the user---is classified as a “local excitation.” In
this modified UCC ansatz the user not only can truncate the UCC excitation to a
certain maximum number (doubles, triples, etc.) but also can impose a constraint of
locality on the higher excitations.
We develop here the theory and the algorithm of this method and name it LAS-VQE. In LAS-VQE, all the singles
excitations are included while only the local ones are included for the higher
excitations. This corresponds to having a mean-field interfragment interaction
and allowing all possible orbital rotations.
This does not reduce the number of qubits required to represent the system, but
it does lower the complexity of the circuit as discussed below.
In principle, one could use the locality argument to include only local
excitations, allowing us to reduce the number of qubits required. This separate-fragments VQE approach, however, fails to account for any interaction between the fragments.
We choose to always include the singles excitations (orbital rotations) across
the entire system to account for some interfragment correlation and to make the
method less sensitive to the localization schemes used.
We also choose to perform in the singles excitation among spin-restricted
orbitals. We achieve this by using the same parameter to control the alpha and
beta excitations corresponding to a given pair of spatial orbitals. This is
necessary to avoid artificial lowering of energies due to spontaneous symmetry
breaking (the spin unrestricted solution). An example of the effect of this is
seen for the hydrogen dimer shown in Figure~2 of the main text. The nonspin adapted
UCCS energy is about 0.2 mHartree lower than the Hartree--Fock and has an spin
contamination of 0.09, showing the symmetry breaking. Working with
spin-restricted orbitals not only allows us to prepare symmetry preserving
states for the subsequent excitations, but also lowers the number of parameters
at the singles level by a factor of 2. This becomes a significant advantage
for the classical optimizer in VQE as the size of the system gets larger.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figs/las_vqe_log_scale.png}
\caption{Energy difference (in Hartree) between FCI and the approximate methods as a function of the distance between the midpoints of the two H$_2$ molecules}
\label{fig:las_vqe_log_scale}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figs/percent_correlation.png}
\caption{Percentage of correlation energy accounted for by the various methods for different intermolecular distances}
\label{fig:percent_corr}
\end{figure}
We study the ground state electronic structure of the hydrogen dimer in various
conformations with the STO-3G basis set. We use CASSCF with a (4,4) active
space and LASSCF with a ((2,2),(2,2)) active space, UCCSD and LAS-VQE. Each H$_2$
molecule is considered as one fragment on which the molecular orbitals are
localized. As one would expect, the performance of LASSCF with respect to CASSCF
(which is also FCI in this case) worsens with decreasing distance between the
two H$_2$ molecules. Figure~\ref{fig:las_vqe_log_scale} shows the energy difference between the various
methods and the FCI/CAS(4,4) absolute energies. UCCSD in this case is not
exactly FCI but does give accurate results. LAS-VQE is close to
(but slightly lower than) the LASSCF energy. The LAS-VQE energy is not
variationally bound by the LASSCF that it is based on.
In contrast to the LAS-UCC method discussed in this paper, this method uses a
single determinant
Hartree--Fock reference state. The UCCSD energy is a variational lower bound to
the LAS-VQE method. There is, however, a significant advantage when it comes to
the computational cost: gate count and the depth of the circuit. The number of
doubles operators scales linearly with the number of fragments.
The Table~\ref{tab:lasvqe_cost} shows the number of gates and parameters required for the
various calculations for a system with increasing number of H$_2$ molecules with
the same fragmentation scheme. LAS-VQE in considerably
lower in both circuit depth and number of parameters,
compared with full UCCSD.
\begin{table}[ht]
\centering
\caption{Circuit depth and number of parameters for the various methods for an increasing number of H$_2$ fragments }
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
&\multicolumn{3}{|c|}{Circuit Depth} & \multicolumn{3}{|c|}{Parameters} \\ \cline{2-7}
& UCCS & LAS-VQE & UCCSD & UCCS & LAS-VQE & UCCSD \\ \hline
1 $\times$ H$_2$ & 14 & 97 & 97 & 1 & 2 & 2 \\ \hline
2 $\times$ H$_2$ & 58 & 141 & 1791 & 4 & 6 & 22 \\ \hline
3 $\times$ H$_2$ & 147 & 230 & 11206 & 9 & 12 & 108 \\ \hline
4 $\times$ H$_2$ & 296 & 379 & 42789 & 16 & 20 & 344 \\ \hline
\end{tabular}
\label{tab:lasvqe_cost}
\end{table}
\change{
\section{Justification of Eq. (9) of the main text}
Substituting Eqs. (4) and (6) into the Eq. (3) of the main text and differentiating with respect to generator amplitude $x_{\vec{k}}$ yields
\begin{equation}
\pdv{E_{\textrm{LAS}}}{x_{\vec{k}}} = \bra{\Phi}\bigwedge_{L\neq K}\bra{\Psi_L}\wedge\braket{\vec{k}|\hat{H}-E_{\textrm{LAS}}|\Psi_K}\bigwedge_{M\neq K}\ket{\Psi_M}\wedge\ket{\Phi},
\end{equation}
the vanishing of which corresponds to the minimization of the LAS energy given by Eq.\ (3). Within the Hilbert space of the $K$th fragment, the energy minimization conditions for the $x_{\vec{k}}$ amplitudes corresponds to an eigenproblem,
\begin{equation}
\hat{H}_K\ket{\Psi_K}=E_{\textrm{LAS}}\ket{\Psi_K},
\end{equation}
where
\begin{eqnarray}
\hat{H}_K &\equiv& \bra{\Phi}\bigwedge_{L\neq K}\bra{\Psi_L}\hat{H}\bigwedge_{M\neq K}\ket{\Psi_M}\wedge\ket{\Phi}
\nonumber\\ &=& \tilde{h}^{k_1}_{k_2} \crop{k_1}\anop{k_2} + \frac{1}{4} h^{k_1k_3}_{k_2k_4} \crop{k_1}\crop{k_3}\anop{k_4}\anop{k_2},
\end{eqnarray}
where $\tilde{h}^{k_1}_{k_2}$ is given by Eq.\ (10) of the main text. This effective Hamiltonian describes the $K$th fragment without interacting with any other fragment; $\hat{H}_K$ for various $K$ mutually commute.
For a system composed of noninteracting fragments $A$ and $B$ with Hamiltonians $\hat{H}_A$ and $\hat{H}_B$, energies $E_A$ and $E_B$, and wave functions $\ket{\Psi_A}$ and $\ket{\Psi_B}$, it is generally true that
\begin{equation}
\left(\hat{H}_A+\hat{H}_B\right)\ket{\Psi_A}\wedge\ket{\Psi_B} = \left(E_A+E_B\right)\ket{\Psi_A}\wedge\ket{\Psi_B}.
\end{equation}
Noting that $\hat{H}_{\textrm{eff}}$ given by Eq.\ (9) of the main text is obviously
\begin{equation}
\hat{H}_{\textrm{eff}}=\sum_K \hat{H}_K,
\end{equation}
we conclude that
\begin{equation}
\hat{H}_{\textrm{eff}}\bigwedge_K\ket{\Psi_K} = n_f\times E_{\textrm{LAS}}\bigwedge_K\ket{\Psi_K}.
\end{equation}
In other words, $\hat{H}_{\textrm{eff}}$ models interacting fragments as non-interacting subsystems described by unrelated, mutually commuting effective Hamiltonian terms. Thus, the QPE algorithm applied to $\hat{H}_{\textrm{eff}}$ generates $\ket{\textrm{QLAS}}\equiv \bigwedge_K\ket{\Psi_K}$ on the quantum circuit.
}
\section{Orbital Ordering and the Jordan-Wigner Transformation}
The Jordan-Wigner transformation~\cite{jordan1993paulische} is one of the many ways of transforming fermion operators into spin operators, and is a necessary
step for performing quantum chemistry calculations on quantum computers.
Here, we very briefly discuss the overheads associated with the Jordan-Wigner
transformation and how it can be mitigated in certain fragmented geometries.
The Jordan-Wigner transformation transforms a fermion creation operator for
orbital $j$, $a_j^\dagger$, out of a total of $N$ spin orbitals as follows
\begin{equation}
\tilde{a}_j^\dagger = Z^{\otimes j - 1} \otimes \frac{X-iY}{2} \otimes I^{\otimes N-1},
\end{equation}
where $\tilde{a}_j^\dagger$ is the transformed operator;
$X$, $Y$, and $Z$ are the Pauli operators; and the notation $Z^{\otimes N}$
denotes applying the tensor operation, $\otimes$, $N$ times for Z (i.e., $Z^{\otimes 4} = Z \otimes Z \otimes Z \otimes Z$. The term $Z^{\otimes j - 1}$ ensures that
the transformed operators obey the correct fermionic anti-commutation relations.
In standard quantum algorithms for quantum chemistry, these $Z$ strings introduce
very high-weight (that is, having many terms that are not $I$) terms into the Hamiltonian
and cluster operators, leading to an $O(N)$ overhead. To see this,
we focus on just the one-body terms.
Under the Jordan-Wigner transformation, and assuming $j>k$,
\begin{equation}\label{eq:one-body}
\tilde{a}_j^\dagger \tilde{a}_k = I^{\otimes k -1} \otimes \frac{Z(X-iY)}{2} \otimes Z^{\otimes j-k-1} \otimes\frac{X-iY}{2} \otimes I^{\otimes N-j}.
\end{equation}
The weight (number of non-identity terms) of this operator is $j-k+1$.
In the most general case, there will exist terms where spin orbital $1$ and $N$ will have a nonzero component, and the resulting weight is $N$, which introduces
the $O(N)$ overhead normally associated with the Jordan-Wigner transformation.
There are many techniques for generally reducing this overhead~\cite{tranter2018comparison,PhysRevA.95.032332}. In our simple
chain of H$_2$ molecules, we can remove this scaling by choosing a particular
ordering of the orbitals. One typical ordering of the orbitals
is the following: all occupied spin up orbitals, all virtual spin up orbitals,
all occupied spin down orbital, all virtual spin down orbitals. If this
ordering is followed when using the local actives spaces of each fragment,
the $O(N)$ overhead is still applicable. Take, for example, just the one-body term
from the first occupied spin up orbital of the first fragment
to the first virtual spin up orbital
for the first fragment. In this case, $j=1$ and $k=O(N)$, leading to the
typical $O(N)$ Jordan-Wigner scaling. If, instead, the orbitals are ordered
with all occupied spin up orbitals of fragment one, followed by all virtual
spin up orbitals of fragment one, followed the same for the second fragment,
and so on, the overhead is removed. Now, $j=1$ and $k=O(N_k)$, where $N_k$
is the number of spin orbitals in each fragment. Since $N_k$ is assumed to be
fixed as the number of fragments grows, the $O(N)$ scaling is reduced to a constant.
To remove the Jordan-Wigner overhead from mixed spin two-body terms, the ordering
can be further changed so that each spin up orbital is immediately followed
by its analogous spin down orbital. This shows that the Jordan-Wigner $O(N)$
overhead can be removed from all terms in the reduced Hamiltonian, removing the
$O(N)$ scaling from the QPE part of LAS-UCC. For UCCSD correlators, the Jordan-Wigner overhead can be removed for certain geometries, such as the linear H$_2$ chain
studied in the main text. More complex geometries, such as higher dimensional
lattices, will re-introduce scaling terms, as the orbitals cannot be simultaneously
ordered to give low-weight Pauli strings for terms which couple both in the
multiple dimensions, such as $x$ and $y$ in a 2D lattice.
\section{Single Qubit Rotation Gates}
Figure~\ref{fig:rot_gates} shows the estimated number of arbitrary
single qubit rotation gates needed to implement full QPE across the
whole molecule, the fragmented LAS-QPE, full UCCSD, and the 2-local
UCCSD correlator with increasing number of H$_2$ molecules. Just
like in the main text, LAS-UCC has polynomially fewer gates,
compared with QPE or UCCSD, scaling only linearly for the
linear chain geometry. QPE and UCCSD only need $O(N^4)$ single
qubit rotations because the Jordan-Wigner overhead only shows up in the
CNOT gates.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figs/rot_gate_counts.png}
\caption{Estimated single qubit arbitrary rotation gate counts using various algorithms. The QPE estimates assume only a single Trotter step. Polynomials of various orders have been plotted to demonstrate the scaling. Our algorithm, LAS-UCC, requires both the LAS-QPE and 2-UCC circuits and thus has an overall $O(N)$ scaling, compared with the $O(N^4)$ scaling of UCC and QPE.}
\label{fig:rot_gates}
\end{figure}
\change{
\section{Strong correlation of \emph{trans}-butadiene}
}
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figs/pes_si-eps-converted-to.pdf}
\caption{\change{Potential energy curve of the \textit{trans}-butadiene molecule along the symmetric double C=C bond dissociation coordinate, computed with CASCI, LASSCF, and CCSD(T). Total energies for each method are shifted to zero at the equilibrium geometry.}}
\label{fig:c4h6_si}
\end{figure}
\change{
The symmetric double-bond dissociation curve of the \textit{trans}-butadiene molecule (cf.\ Figs.\ 2b and 4 of the main text) calculated using the CCSD(T) method is plotted in Fig.\ \ref{fig:c4h6_si} and compared to CASCI and LASSCF curves. In the dissociation limit, in which the molecule objectively can be modeled as three non-interacting diradical fragments, the CCSD(T) method recovers a greater amount of correlation energy than LASSCF, \textit{albeit} not as much as LAS-UCC or CASCI. Near the equilibrium geometry, the three methods are indistinguishable.
However, unlike LASSCF, CASCI, or LAS-UCC, all of which are multireference methods, the single-reference CCSD(T) method cannot smoothly connect the region of the equilibrium geometry to the bond dissociation limit. The discontinuous and unphysical behavior of CCSD(T) in the region of $R=1.8 \AA$ is consistent with the tendency of single-reference coupled cluster at low truncation orders to struggle to dissociate multiple or high-order bonds.\cite{Li2001} This \textit{trans}-butadiene potential energy curve is an example of a system in which both static and dynamical correlation must simultaneously be accounted for in order to obtain quantitatively accurate results.
\section{UCCSD operator ordering in LAS-UCC}
The following listings from the \textit{mrh} package describe the operator order for the first-order Trotterized UCCSD correlator used in our classical LAS-UCC calculations. Individual spatial orbitals are indexed first by fragment, and then in decreasing order of their natural orbital occupancy (before application of the UCCSD correlator). Listing \ref{lst:spatexc} contains the code for generating the combinations of spatial-orbital excitation patterns, as well as a docstring summarizing the effects of the code presented in Listings \ref{lst:spinexc} and \ref{lst:spincases} for the case of UCCSD. Listings \ref{lst:spinexc} and \ref{lst:spincases} contain the code for generating valid spin-orbital excitation patterns for any general set of spatial-orbital excitation patterns of any order in a UCC operator. Note that ``np'' is numpy and ``combinations\_with\_replacement'' is from the itertools module. This code is from the commit identified by the SHA-1 hash 8bb3c2d46ab5f43c8787a365b22713b5410855f2, which differs from the commit which performed the calculations reported in the main text (fe564cd1c54c404edf8bce1dd7118b2d53acdd7a) only in docstrings, test scripts, and line wrapping.
}
\begin{lstlisting}[caption={\change{Code for generating the spatial-orbital list of the UCCSD correlator in LAS-UCC within \textit{mrh}}}, language=Python, float=*, label=lst:spatexc]
def get_uccsd_op (norb, t1=None, t2=None):
''' Construct and optionally initialize semi-spin-adapted unitary CC
correlator with singles and doubles spanning a single undifferentiated
orbital range. Excitations from spatial orbital(s) i(, j) to spatial
orbital(s) a(, b) are applied to the ket in the order
U|ket> = u^n(n-1)_nn u^n(n-1)_n(n-1) u^n(n-2)_nn ... u^11_22 u^11_21
... u^n_(n-1) u^n_(n-2) ... u^3_2 u^3_1 u^2_1 |ket>
where ^ indicates creation operators (a, b; rows) and _ indicates
annihilation operators (i, j; columns). The doubles amplitudes are
arbitrarily chosen in the upper-triangular space (a,b <= i,j), but the
lower-triangular space is used for the individual double pairs
(a > b, i > j) and for the singles amplitudes (a > i). In all cases,
row-major ordering is employed.
The spin cases of a given set of orbitals a, b, i, j are grouped
together. For singles, spin-up (a) and spin-down (b) amplitudes are
constrained to be equal and the spin-up operator is on the right (i.e.,
is applied first). For doubles, the spin case order is
u|ket> -> ^bb_bb ^ab_ab ^ab_ba ^ba_ab ^ba_ba ^aa_aa |ket>
For spatial orbital cases in which the same index appears more than
once, spin cases that correspond to nilpotent (eg., ^pp_qr ^aa_aa),
undefined (eg., ^pq_pq ^ab_ab), or redundant (eg., ^pq_pq ^ab_ba)
operators are omitted.
Args:
norb : integer
Total number of spatial orbitals. (0.5 * #spinorbitals)
Kwargs:
t1 : ndarray of shape (norb,norb)
Amplitudes at which to initialize the singles operators
t2 : None
NOT IMPLEMENTED. Amplitudes at which to initialize the doubles
operators
Returns:
uop : object of class FSUCCOperator
The callable UCCSD operator
'''
t1_idx = np.tril_indices (norb, k=-1)
ab_idxs, ij_idxs = list (t1_idx[0]), list (t1_idx[1])
pq = [(p, q) for p, q in zip (*np.tril_indices (norb))]
for ab, ij in combinations_with_replacement (pq, 2):
ab_idxs.append (ab)
ij_idxs.append (ij)
uop = FSUCCOperator (norb, ab_idxs, ij_idxs)
x0 = uop.get_uniq_amps ()
if t1 is not None: x0[:len (t1_idx[0])] = t1[t1_idx]
if t2 is not None: raise NotImplementedError ("t2 initialization")
uop.set_uniq_amps_(x0)
return uop
\end{lstlisting}
\begin{lstlisting}[caption={\change{Code for generating spin-orbital excitation operators of UCC(SD) in \textit{mrh}. The function ``spincases'' is presented in Listing \ref{lst:spincases}}}, language=Python, float=*, label=lst:spinexc]
class FSUCCOperator (uccsd_sym0.FSUCCOperator):
''' A callable spin-adapted (Sz only) unrestricted coupled cluster
operator. For single-excitation operators, spin-up and spin-down
amplitudes are constrained to be equal. All spin cases for a given
spatial-orbital excitation pattern (from i_idxs to a_idxs) are grouped
together and applied to the ket in ascending order of the index
(a_spin) * nelec + i_spin
where 'a_spin' and 'i_spin' are the ordinal indices of the spin
cases returned by the function 'spincases' for a_idxs (creation
operators) and i_idxs (annihilation operators) respectively, and nelec
is the order of the generator (1=singles, 2=doubles, etc.) Nilpotent
or undefined spin cases (i.e., because of spatial-orbital index
collisions) are omitted.
'''
def __init__(self, norb, a_idxs, i_idxs):
# Up to two equal indices in one generator are allowed
# However, we still can't have any equal generators
self.a_idxs = []
self.i_idxs = []
self.symtab = []
for ix, (a, i) in enumerate (zip (a_idxs, i_idxs)):
a = np.ascontiguousarray (a, dtype=np.uint8)
i = np.ascontiguousarray (i, dtype=np.uint8)
errstr = 'a,i={},{} invalid for number-sym op'.format (a,i)
assert (len (a) == len (i)), errstr
#errstr = 'a,i={},{} degree of freedom undefined'.format (a,i)
#assert (not (np.all (a == i))), errstr
if len (a) == 1: # Only case where I know the proper symmetry
# relation between amps to ensure S**2
symrow = [len (self.a_idxs), len (self.i_idxs)+1]
self.a_idxs.extend ([a, a+norb])
self.i_idxs.extend ([i, i+norb])
self.symtab.append (symrow)
else:
for ix_ab, (ab, ma) in enumerate (zip (*spincases (a, norb))):
if np.amax (np.unique (ab, # nilpotent escape
return_counts=True)[1]) > 1: continue
for ix_ij, (ij, mi) in enumerate (zip (*spincases (
i, norb))):
if mi != ma: continue # sz-break escape
if np.all (ab==ij): continue # undefined escape
if np.all (a==i) and ix_ab>ix_ij:
continue # redundant escape
if np.amax (np.unique (ij, return_counts=True)[1]) > 1:
continue # nilpotent escape
self.symtab.append ([len (self.a_idxs)])
self.a_idxs.append (ab)
self.i_idxs.append (ij)
self.norb = 2*norb
self.ngen = len (self.a_idxs)
assert (len (self.i_idxs) == self.ngen)
self.uniq_gen_idx = np.array ([x[0] for x in self.symtab])
self.amps = np.zeros (self.ngen)
self.assert_sanity ()
\end{lstlisting}
\begin{lstlisting}[caption={\change{Code for generating a valid list of spin cases corresponding to a set of field operators for particular spatial orbitals in \textit{mrh}}}, language=Python, float=*, label=lst:spincases]
def spincases (p_idxs, norb):
''' Compute the spinorbital indices corresponding to all spin cases of a
set of field operators acting on a specified list of spatial orbitals
The different spin cases are returned 'column-major order':
aaa...
baa...
aba...
bba...
aab...
The index of a given spincase string ('aba...', etc.) can be computed
as
p_spin = int (spincase[::-1].replace ('a','0').replace ('b','1'), 2)
Args:
p_idxs : ndarray of shape (nelec,)
Spatial orbital indices
norb : integer
Total number of spatial orbitals
Returns:
p_idxs : ndarray of shape (2^nelec, nelec)
Rows contain different spinorbital cases of the input spatial
orbitals
m : ndarray of shape (2^nelec,)
Number of beta (spin-down) orbitals in each spin case
'''
nelec = len (p_idxs)
p_idxs = p_idxs[None,:]
m = np.array ([0])
for ielec in range (nelec):
q_idxs = p_idxs.copy ()
q_idxs[:,ielec] += norb
p_idxs = np.append (p_idxs, q_idxs, axis=0)
m = np.append (m, m+1)
p_sorted = np.stack ([np.sort (prow) for prow in p_idxs], axis=0)
idx_uniq = np.unique (p_sorted, return_index=True, axis=0)[1]
p_idxs = p_idxs[idx_uniq]
m = m[idx_uniq]
return p_idxs, m
\end{lstlisting}
\change{
For the UCCSD runs of the main text, Fig.~(4), we use ordering
consistent with Qiskit's UCCSD generation.
We have included a short code listing (Listing~\ref{lst:hfuccsdexc}) which will reproduce the excitation ordering.
}
\begin{lstlisting}[caption={\change{Code for generating the excitation operators of the UCCSD ansatz for the UCCSD calculations of Fig.~(4) of the main text. }}, language=Python, float=*, label=lst:hfuccsdexc]
def generate_fermionic_excitations_occ(num_excitations,num_spin_orbitals,num_particles,
alpha_occ,alpha_unocc,beta_occ,beta_unocc):
#Assumes ordering [occ_alpha,unocc_alpha,occ_beta,unocc_beta]
alpha_excitations = []
# generate alpha-spin orbital indices for occupied and unoccupied ones
# the Cartesian product of these lists gives all possible single alpha-spin excitations
alpha_excitations = list(itertools.product(alpha_occ, alpha_unocc))
# the Cartesian product of these lists gives all possible single beta-spin excitations
beta_excitations = list(itertools.product(beta_occ, beta_unocc))
# we can find the actual list of excitations by doing the following:
# 1. combine the single alpha- and beta-spin excitations
# 2. find all possible combinations of length `num_excitations`
pool = itertools.combinations(
alpha_excitations + beta_excitations, num_excitations
)
excitations = list()
visited_excitations = set()
for exc in pool:
# validate an excitation by asserting that all indices are unique:
# 1. get the frozen set of indices in the excitation
exc_set = frozenset(itertools.chain.from_iterable(exc))
# 2. all indices must be unique (size of set equals 2 * num_excitations)
# 3. and we palso don't want to include permuted variants of identical excitations
if len(exc_set) == num_excitations * 2 and exc_set not in visited_excitations:
visited_excitations.add(exc_set)
occ, unocc = zip(*exc)
exc_tuple = (occ, unocc)
excitations.append(exc_tuple)
return excitations
\end{lstlisting}
| 1e22e6feff2b008cd337221e5d84d102dc251c10 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $P$ be a point on an elliptic curve $E$ defined over a discrete valuation field $K$. In this paper, we study some properties of the sequence $\{nP\}_{n\in\mathbb{N}}$ of the multiples of the point $P$. It is known that this sequence has some remarkable properties, in particular when a Weierstrass model is fixed
\[y^2+a_1xy+a_3y=x^3+a_2x^2+a_4x+a_6\]
the elements $nP=\left(\frac{\phi_n(x,y)}{\psi_n^2(x,y)},\frac{\omega_n(x,y)}{\psi_n^3(x,y)}\right)$ for $P=(x,y)$ satisfy certain recurrent formulas (see properties (A), (B), (C) below or \cite[Exercise 3.7]{aoc}).
When the curve $E$ is defined over a number field the sequence $(\psi_{n}(x,y))_{n\in\mathbb{N}}$ has been studied as an example of a divisibility sequence, a notion which was studied classically for Lucas sequences and by Ward for higher degree recurrences. Many more results are known, including \cite{ward} and \cite{silvermanprimitive}.
Over function fields, in particular over the function field of a smooth projective curve $C$ over an algebraically closed field $k$ one can study the set of effective divisors $\{D_{nP}\}$ where $D_{nP}$ is the divisor of poles of the element $x(nP)$ for a fixed $n$, cf. \cite{ingrammahesilvstreng}, \cite{Naskrecki_NYJM}, \cite{Naskrecki_Streng}, \cite{Cornelissen}.
The question which we investigate in this paper is to what extent, for a fixed discrete valuation $v$ in the field $K$, the numbers $v(\psi_n)$ and $v(\phi_{n})$ can be both positive and what is the maximum exponent $r$ for the power $\pi^{r}$ of the local at $v$ uniformizing element $\pi\in K$ which divides both $\psi_n$ and $\phi_n$. Such a situation only happens when the point $P$ has bad reduction at the place $v$ and we produce a complete formula which deals with all Kodaira reduction types.
Our main result has the same formulation as the main theorem proved by Yabuta and Voutier \cite{yv21} for finite extensions of $\mathbb{Q}_{p}$ but our proof is much shorter and depends exclusively on the elementary properties of the N\'{e}ron local heights. Up to some standard facts from \cite{langellcur},\cite{Silverman_Advanced}, \cite{tate}, and \cite{shioda} our proof is essentially self-contained.
Let $n$ be a positive integer and define
\begin{equation}
k_{v,n}=\min\left\{v\left(\psi_n^2(x(P))\right),v\left(\phi_n(x(P))\right)\right\}.
\end{equation}
\begin{theorem}\label{main}
Let $R$ be a discrete valuation ring with quotient field $K$. Let $v$ be the valuation of $K$. Let $E/K$ be an elliptic curve defined by a Weierstrass model in minimal form with respect to $v$, let $P=(x,y)\in E(K)$ and let $n$ be a positive integer.
If $P\neq O$ is non-singular modulo $v$, then
\[
k_{v,n}=\min\left(0,n^2v(x(P))\right).
\]
If $P\neq O$ is singular modulo $v$, then
\[
k_{v,n}=n^2c_v(P)-c_v(nP).
\]
The function $c_v(nP)$ will be defined in Definition \ref{def:cv} and it is periodic in $n$.
For the case of standard reduction types at $v$ we compute the value of $n^2c_v(P)-c_v(nP)$ in Table \ref{tab:common_vals}. The constants $m_P$ and $a_P$ will be defined in the next section.
\end{theorem}
{\small\renewcommand{\arraystretch}{1.5}
\begin{table}[!htb]
\begin{tabular}{c|c|cl}
Kodaira symbol & $m_{P}$ & $n^2c_v(P)-c_v(nP)$ & \\
\hline
$III^{*}$ & $2$ & $3n^2/2$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $3(n^2-1)/2$ & \text{{\rm if} $n \not\equiv 0 \bmod{m_{P}}$} \\ \hline
$IV^{*}$ & $3$ & $4n^2/3$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $4(n^2-1)/3$ & \text{{\rm if} $n \not\equiv 0 \bmod{m_{P}}$} \\ \hline
$III$ & $2$ & $n^2/2$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $(n^2-1)/2$ & \text{{\rm if} $n \not\equiv 0 \bmod{m_{P}}$} \\ \hline
$IV$ & $3$ & $2n^2/3$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $2(n^2-1)/3$ & \text{{\rm if} $n \not\equiv 0 \bmod{m_{P}}$} \\ \hline
$I_{m}^{*}$ & $2$ & $n^2$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $n^2-1$ & \text{{\rm if} $n \not\equiv 0 \bmod{m_{P}}$} \\ \hline
$I_{m}^{*}$
& $4$ & $n^2(m+4)/4$ & \text{{\rm if} $n \equiv 0 \bmod{m_{P}}$} \\
& & $(n^2-1)(m+4)/4$ & \text{{\rm if} $n \equiv 1,3 \bmod{m_{P}}$} \\
& & $(n^2(m+4)/4)-1$ & \text{{\rm if} $n \equiv 2 \bmod{m_{P}}$} \\ \hline
$I_{m}$ & $\dfrac{m}{\gcd \left( a_{P}, m \right)}$
& $n^2\dfrac{a_{P}\left( m-a_{P} \right)}{m}-\dfrac{n'\left( m-n' \right)}{m}$
& \text{{\rm if} $a_{P}n \equiv n' \bmod{m}$} \\
\end{tabular}
\caption{Tables of the common valuation $k_{v,n}$ for standard Kodaira types. In the last line we took $n'$ as the smallest positive integer such that $a_{P}n \equiv n' \bmod{m}$.}\label{tab:common_vals}
\end{table}
}
If the residue field $\kappa$ is perfect, then the Kodaira symbol of the curve is one of those of Table \ref{tab:common_vals} and it can be computed using \cite[Table C.15.1]{aoc}. If $\kappa$ is not perfect, then one can compute the Kodaira symbol of the curve using \cite{szydlo}.
We also emphasize that the formula for the cancellation $k_{v,n}$ is a quadratic polynomial in $n$ with rational coefficients which depend only on the correction terms $c_v(P)$ defined by Cox-Zucker \cite{coxzucker} and used extensively by \cite{shioda} in his theory of Mordell-Weil lattices. This dependence is deduced naturally from a translation of the local N\'{e}ron height into the language of elliptic surfaces. Although the difference between these notions is purely psychological, we find it more natural to formulate our argument using correction terms and intersection pairing of curves.
There are two key facts which make the proof ultimately very short.
\begin{itemize}
\item[1.] The N\'{e}ron local height is a sum of two ingredients: local intersection pairing with the zero section and the ``correction part'' which depends only on the component which the given point hits at the place $v$, cf. \cite{langellcur}, \cite{J_S_Muller},\cite{Busch_Muller} \cite{Holmes_PhD}.
\item[2.] The valuation $v(\psi_n(P))$ is a quadratic polynomial in $n$ whose coefficients depend on the local N\'{e}ron height at arguments $nP$ and $P$, respectively, and on the valuation of the minimal discriminant $\Delta$ of the elliptic curve $E$, cf. Lemmas \ref{lem:valuation_lem_1},\ref{lemma:psiinter}
\end{itemize}
On the way we reprove in our setting some results of \cite{ayad} and obtain a much simpler analysis of the ''troublemaker sequences'' used by Stange \cite{Stange} and \cite{yv21} in their proofs. In fact, we dare to say that our current formulation is optimal in its exposition, emphasizing the role of the component group at places $v$ of bad reduction and the mapping of point $P$ to its elements $comp_v(P)$ at places $v$ of bad reduction.
In the number field case, the sequence $k_{v,n}$ received many attentions during the last 30 years, cf. \cite{ayad}, \cite{chhadiv}, \cite{ingram}, \cite{yv21}. There were many attempt to understand the shape of $k_{v,n}$ and the goal of this paper is to give a very precise formula in the most general setting.
The study of the sequence $k_{v,n}$ has some interesting applications. For example, knowing the behaviour of $k_{v,n}$ was necessary to show that every elliptic divisibility sequence satisfies a recurrence relation in the rational case, cf. \cite{recurrence}. Another application can be found in \cite{Stange} and \cite{ingram}, where the study of $k_{v,n}$ was necessary to study the problem of understanding when a multiple of $P$ is integral.
\section*{Acknowledgements}
The second author has been supported by MIUR (Italy) through PRIN 2017 ``Geometric, algebraic and analytic methods in arithmetic''.
\section{First definitions}
Let $R$ be a discrete valuation ring with quotient field $K$ and residue field $\kappa$.
Let $v$ be the valuation of $K$ and assume that $v(K^*)=\Z$. Let $E$ be an elliptic curve over $K$ which is a generic fibre of the regular proper model $\pi:\mathcal{E}\rightarrow\Spec R=\{o,s\}$. The fibre $E_s=\pi^{-1}(s)$ is the special fibre (singular or not). We denote by $\widetilde{\mathcal{E}}\rightarrow\Spec R=\{o,s\}$ the associated N\'{e}ron model which is the subscheme of smooth points over $R$.
Assume that Weierstrass model of $E$ is in minimal form with respect to $v$ and let $P=(x,y)\in E(K)$. We denote by $\Delta$ the discriminant of the minimal model $E$.
The point $P\in E(K)$ extends to a section $\sigma_{P}:\Spec R\rightarrow\overline{\mathcal{E}}$ by the N\'{e}ron model property. Let $\Phi_{v}$ denote the group of components of the special fibre $E_s$. We have a natural map $comp_{v}:E(K)\rightarrow\Phi_{v}$ which sends the point $P$ to the element of the component group. We say that the point $P$ is non-singular modulo $v$ if the image $comp_v(P)$ is the identity element, otherwise the point $P$ is singular modulo $v$.
Let $E_0(K)$ be the subgroup of $E(K)$ of points that reduce to non-singular points modulo $v$ and it is well known that $comp_v(P)$ belongs to a cyclic component of $\Phi_v$. Let $a_P$ be the order of this component and let $m_P$ be the order of $comp_v(P)$ in $\Phi_v$.
\begin{definition}
Let $(P.O)_v$ denote the local intersection number of the point $P\in E(K)\setminus\{O\}$ with the zero point $O$ at $v$ defined by the formula
\[(P.O)_v=\max\{0,-v(x(P))\}\]
for the minimal model $E$.
\end{definition}
\begin{remark}
For the behaviour of the sequence of integers $\{(nP.O)_v\}_n$, see \cite[Section 7 and Lemma 8.2]{Naskrecki_NYJM} in the function field case and \cite[Lemma 5.1]{Stange} in the number field case.
\end{remark}
\begin{definition}\label{def:cv}
With a pair $(P,v)$ we associate a rational number $c_v(P)$ which is called \textit{the correction term}. Let $\widehat{\lambda}_{v}(P)$ be the local N\'{e}ron height as defined in \cite[Theorem III.4.1]{langellcur}. Define
\[
c_v(P)=(P.O)_v+v(\Delta)/6-2\widehat{\lambda_v}(P).
\]
\end{definition}
\begin{remark}
Observe that $c_v(P)$ depends only on $comp_v(P)$. This follows from \cite[Theorem 11.5.1]{Lang_Fundamentals} when $P$ is singular modulo $v$ and from \cite[Theorem VI.4.1]{Silverman_Advanced} when $P$ is non-singular modulo $v$. In particular, \cite[Theorem VI.4.1]{Silverman_Advanced} shows that $c_v(P)=0$ when $P$ is non-singular modulo $v$.
In Lemma \ref{tab:corr_terms} we will show how to explicitly compute $c_v(P)$ in the case when the residue field $\kappa$ is perfect.
Finally, note that the definition of local N\'{e}ron height in \cite[Theorem III.4.1]{langellcur} is given for discrete valuation fields, as is noted by Lang in \cite[Page 66]{langellcur}.
\end{remark}
The division polynomials are defined in \cite[Exercise 3.7]{aoc}. We recall the properties that we will use:
\begin{itemize}
\item[(A)] If $nP$ is not equal to the identity of the curve $O$, then \[x(nP)=\frac{\phi_n(x(P))}{\psi_n^2(x(P))};\]
\item[(B)] The polynomials have integer coefficients. Moreover, $\psi_n^2(x)$ has degree $n^2-1$ and $\phi_n(x)$ is monic and has degree $n^2$;
\item[(C)] It holds \[\phi_n(x(P))=x(P)\psi_n^2(x(P))-\psi_{n-1}(P)\psi_{n+1}(P).\]
\end{itemize}
\begin{remark}
Our proof clearly works also for number fields. In that setting the result was known \cite{yv21}, but the two proofs are completely different. Indeed, the proof of Yabuta and Voutier is almost completely arithmetical meanwhile our is more geometrical.
\end{remark}
\begin{remark}
We will work assuming that $v(0)=\infty$. Under this assumption, our theorem works also in the case when $nP=O$. In this case $\psi_n^2(x(P))=0$ and then $k_{v,n}(P)=v(\phi_n(x(P)))$.
\end{remark}
\section{Proof of the main theorem}
\begin{lemma}\label{lem:valuation_lem_1}
If $nP\neq O$, then
\begin{equation}\label{eq:psi_lambda_formula}
\widehat{\lambda}(nP) = n^2 \widehat{\lambda}(P)+v(\psi_{n}(P))-\frac{n^2-1}{12}v(\Delta),
\end{equation}
where $\Delta$ is the discriminant of the curve.
\end{lemma}
\begin{proof}
This lemma is stated as an exercise in \cite[Exercise 6.4 (e)]{Silverman_Advanced}. First, assume that $P$ is a non-torsion point. Thanks to \cite[Theorem III.4.1]{langellcur}, we have
\begin{equation}\label{eq:lambda}
\widehat{\lambda}(R+Q)+\widehat{\lambda}(R-Q)=2\widehat{\lambda}(R)+2\widehat{\lambda}(Q)+v(x(R)-x(Q))-v(\Delta)/6
\end{equation}
for all $R,Q\in E(K)$ with $R,Q,R\pm Q\neq O$.
We prove the lemma by induction. If $n=1$ it is obvious since $\psi_1=1$. If $n=2$, then it follows from \cite[Theorem VI.1.1]{Silverman_Advanced}. So, we assume that the lemma is true for $i\leq n$ and we show that it holds also for $n+1\geq 3$. Put $R=nP$ and $Q=P$. So, $R,Q,R\pm Q\neq O$. Then,
\begin{equation}\label{eq:n+1}
\widehat{\lambda}((n+1)P)=-\widehat{\lambda}((n-1)P)+2\widehat{\lambda}(nP)+2\widehat{\lambda}(P)+v(x(nP)-x(P))-v(\Delta)/6.
\end{equation}
For the definition of division polynomials we have
\begin{equation}\label{eq:lam}
x(nP)-x(P)=\frac{\phi_n(x(P))}{\psi_n^2(x(P))}-x(P)=x(P)+\frac{\psi_{n+1}(P)\psi_{n-1}(P)}{\psi_n^2(x(P))}-x(P)=\frac{\psi_{n+1}(P)\psi_{n-1}(P)}{\psi_n^2(x(P))}.
\end{equation}
Put $d_n=\widehat{\lambda}(nP)-v(\psi_{n}(P))$ and then combining (\ref{eq:n+1}) and (\ref{eq:lam}) we have
\[
d_{n+1}=2d_n-d_{n-1}+2d_1-\frac{v(\Delta)}6.
\]
By induction, we have $d_i=i^2d_1-(i^2-1)v(\Delta)/12$ for $i\leq n$ and then
\begin{align*}
d_{n+1}&=(2n^2-(n-1)^2+2)d_1-v(\Delta)\left(\frac{2(n^2-1)-((n-1)^2-1)+2}{12}\right)\\&=(n+1)^2d_1-v(\Delta)\left(\frac{(n+1)^2-1}{12}\right).
\end{align*}
Therefore,
\[
\widehat{\lambda}((n+1)P)=v(\psi_{n+1}(P))+d_{n+1}=v(\psi_{n+1}(P))+(n+1)^2\widehat{\lambda}(P)-v(\Delta)\left(\frac{(n+1)^2-1}{12}\right).
\]
Now, we need to prove the lemma in the case when $P$ is a torsion point. Fix $n\in \mathbb{N}$. First, we can assume that $K$ is complete since clearly the statement does not change. In a finite extension of $K$ we can find a non-torsion point $P'$ such that $(P-P'.O)_v$ is very large. The choice of $P'$ will depend on $n$. Thanks to \cite[Theorem VI.1.1.c]{Silverman_Advanced}, the N\'{e}ron local height does not change if we substitute $K$ with a finite extension, so we can assume $P'\in E(K)$. Since $(P-P'.O)_v$ is very large, $(nP.O)_v=(nP'.O)_v$ and $(P.O)_v=(P'.O)_v$ (this is certainly true if $(P-P'.O)_v>(nP.O)_v$). Moreover, $comp_v(P)=comp_v(P')$ and then, from \cite[Theorem 11.5.1]{Lang_Fundamentals}, $\widehat{\lambda}_v(P)=\widehat{\lambda}_v(P')$ and $\widehat{\lambda}_v(nP)=\widehat{\lambda}_v(nP')$. Furthermore, $v(\psi_n(P))=v(\psi_n(P'))$ since $\psi_n$ is a continuous function and $v$ is discrete. We know that the lemma holds for $P'$ since it is a non-torsion point and then
\[
\widehat{\lambda}(nP)=\widehat{\lambda}(nP') = n^2 \widehat{\lambda}(P')+v(\psi_{n}(P'))-\frac{n^2-1}{12}v(\Delta)=n^2 \widehat{\lambda}(P)+v(\psi_{n}(P))-\frac{n^2-1}{12}v(\Delta).
\]
\end{proof}
Recall that, in the case $\kappa$ perfect, the Kodaira symbol of an elliptic curve can be computed using \cite[Table C.15.1]{aoc}.
\begin{lemma}\label{tab:corr_terms}
Assume that the residue field $\kappa$ is perfect. Then, the value of $c_v(P)$ is computed in the following table.
\begin{center}
\begin{tabular}{c|c|c}
Kodaira symbol & $m_{P}$& $c_v(P)$ \\ \hline
Any & $1$ & $0$ \\ \hline
$III^{*}$ & $2$ & $3/2$ \\
\hline
$IV^{*}$ & $3$ & $4/3$ \\
\hline
$III$ & $2$ & $1/2$ \\
\hline
$IV$ & $3$ & $2/3$ \\
\hline
$I_{m}^{*}$ & $2$ & $1$ \\
\hline
$I_{m}^{*}$
& $4$ & $(m+4)/4$ \\
\hline
$I_{m}$ & $\dfrac{m}{\gcd \left( a_{P}, m \right)}$
& $\dfrac{a_{P}\left( m-a_{P} \right)}{m}$\\
\end{tabular}
\end{center}
In the last line we are assuming that $0<a_P< m$.
\end{lemma}
\begin{proof}
If $P$ has non-singular reduction, then $\widehat{\lambda}(P)=v(\Delta)/12+(P.O)_v/2$ thanks to \cite[Theorem VI.4.1]{Silverman_Advanced}. Observe that $\widehat{\lambda}(P)$ is denote by $\lambda(P)$ in \cite{Silverman_Advanced}. Hence, $c_v(P)=0$. Observe that this is true even without the hypothesis $\kappa$ perfect.
If $P$ has singular reduction, then $(P.O)_v=0$ since the identity is a non-singular point (see \cite[Page 46]{aoc}). Define $\lambda_v(P)=2\widehat{\lambda}_v(P)+\frac{v(\Delta)}{6}$. As is pointed out in \cite[Equation 11]{cremonaprickett}, our $\lambda_v(P)$ is equal to the $\lambda(P)$ of that paper. So, by definition,
\[
c_v(P)=(P.O)_v+v(\Delta)/6-2\widehat{\lambda_v}(P)=v(\Delta)/3-\lambda_v(P).
\]
Hence, we need to prove that $v(\Delta)/3-\lambda(P)$ is equal to the value in the table. The values of $v(\Delta)$ are computed in \cite[Table C.15.1]{aoc}. The values of $\lambda(P)$ are computed in \cite[Proposition 6]{cremonaprickett} and one can easily check that the values are equal case by case. For example, if the Kodaira type is III, then $v(\Delta)=3$ and $v(\lambda_v(P))=1/2$. So, $c_v(P)=1/2$. Observe that \cite[Proposition 6]{cremonaprickett} is stated for number fields, but the proof works in the exact same way for discrete valuation fields with perfect residue field. The hypothesis perfect residue field is necessary in order to use Tate's algorithm \cite{tate}.
In order to make the proof clearer, we just sketch the proof in a particular case. Take $P$ a singular point at $v$. We will prove the Lemma in the case when the Kodaira symbol is IV. The other cases are similar.
First of all, observe that $\widehat{\lambda_v}(P)=\widehat{\lambda_v}(-P)$. Indeed, if we put $\widehat{\lambda_v'}(P)=\widehat{\lambda_v}(-P)$, then $\widehat{\lambda_v'}$ satisfies (\ref{eq:lambda}). Thanks to \cite[Theorem III.4.1]{langellcur}, we have $\widehat{\lambda_v'}=\widehat{\lambda_v}$ since $\widehat{\lambda_v}$ is unique.
Moreover, if $P\notin E_0(K)$, then $\widehat{\lambda}_v(P)$ just depends on the image of $P$ in $E(K)/E_0(K)$. This follows from \cite[Theorem 11.5.1]{Lang_Fundamentals}. Furthermore, $\widehat{\lambda_v}(P)$ does not depend on the choice of the minimal Weierstrass model defining the curve (see \cite[Page 63]{langellcur}). Since the Kodaira symbol is IV, we have that $P$ has order $3$ in $E(K)/E_0(K)$ and then $\widehat{\lambda_v}(2P)=\widehat{\lambda_v}(-P)=\widehat{\lambda_v}(P)$. Using Lemma \ref{lem:valuation_lem_1} with $n=2$, we have
\[
3\widehat{\lambda_v}(P)=\frac{v(\Delta)}{4}-v(\psi_2(P))
\]
and
\[
c_v(P)=(P.O)_v+\frac{v(\Delta)}6-2\widehat{\lambda_v}(P)=\frac{v(\psi_2^2(x(P)))}{3}.
\]
So, in order to prove that $c_v(P)=2/3$ we just need to show that
\[
v(\psi_2^2(x(P)))=2.
\]
From the definition of $\psi_2^2$ we have $\psi_2^2(x(P))=4x(P)^3+b_2x(P)^2+2b_4x(P)+b_6$. Since $\widehat{\lambda_v}(P)$ does not depend on the choice of the minimal Weierstrass equation defining the curve, we can assume that the coefficients satisfy the Tate's algorithm (see \cite{tate}). In this case, $v(b_2)>0$, $v(b_4)>1$, and $v(b_6)=2$. Moreover, $v(x(P))>0$ since $P$ is non-singular if and only if $v(x(P))>0$. This is proved in the number fields case in \cite[Lemma 2.8 (i)]{yv21}, but the proof holds in the same way also in our case.
Hence,
\[
2=v(b_6)=v(4x(P)^3+b_2x(P)^2+2b_4x(P)+b_6)=v(\psi_2^2(P))
\]
and we are done.
\end{proof}
In order to compute the value of $c_v(P)$ in the case when $\kappa$ is not perfect, one can use the analogue of Tate's algorithm in the non-perfect case, cf. \cite{szydlo}.
\begin{lemma}\label{lemma:psiinter}
Let $P$ be a point on the elliptic curve $E$ as above and let $\psi_{n}(P)$ be the value of the $n$-division polynomial at the point $P$. For every $n$ such that $nP\neq O$, we have
\[v(\psi_n^2(P)) = (nP.O)_v-n^2(P.O)_v+n^2 c_v(P)-c_v(nP).\]
\end{lemma}
\begin{proof}
From (\ref{eq:psi_lambda_formula}) we have
\[
v(\psi_{n}(P))=\widehat{\lambda}(nP) +\frac{n^2-1}{12}v(\Delta)- n^2 \widehat{\lambda}(P).
\]
From Definition \ref{def:cv} we have
\[
\widehat{\lambda}(P)=\frac{v(\Delta)}{12}-\frac{c_v(P)}{2}+\frac{(P.O)_v}2.
\]
Hence,
\begin{align*}
v(\psi_{n}(P))&=\widehat{\lambda}(nP) +\frac{n^2-1}{12}v(\Delta)- n^2 \widehat{\lambda}(P)\\&
=\frac{v(\Delta)}{12}-\frac{c_v(nP)}2+\frac{(nP.O)_v}2 +\frac{n^2-1}{12}v(\Delta)- n^2\frac{v(\Delta)}{12}+n^2\frac{c_v(P)}2-\frac{n^2(P.O)_v}2\\&
=\frac{(nP.O)_v}2-\frac{n^2(P.O)_v}2+\frac{1}{2}n^2 c_v(P)-\frac{1}{2}c_v(nP).
\end{align*}
\end{proof}
\begin{remark}
Thanks to the previous lemma, one can easily show that the valuation $v(\psi_{n}(P))$ is independent of the choice of the minimal Weierstrass model.
\end{remark}
Put $n_P$ as the smallest positive integer such that $n_P P$ reduces to the identity modulo $v$. Observe that $n_P$ does not necessarily exist.
\begin{lemma}\label{lemma:div}
If $(kP.O)_v>0$, then $(mkP.O)_v\geq (kP.O)_v$ for all $m\geq 1$. Moreover, if $n_P$ exists, then $(kP.O)_v>0$ if and only if $n_P\mid k$.
\end{lemma}
\begin{proof}
Put $P'=kP$ and then $v(x(P'))<0$. Hence, $v(\phi_m(x(P')))=m^2v(x(P'))$ (recall that $\phi_m$ is monic) and $v(\psi_m^2(x(P')))\geq (m^2-1)v(x(P'))$. So,
\[
v(x(mkP))=v(x(mP'))=v(\phi_m(x(P')))-v(\psi_m^2(x(P')))\leq v(x(P'))=v(x(kP))<0.
\]
If $n_P\mid k$, then $(kP.O)_v>0$ for the first part of the lemma. If $n_P\nmid k$, then $k=qn_P+r$ with $0<r<n_P$. Recall that the identity is a non-singular point of the curve reduced modulo $v$. Hence, if $(kP.O)_v>0$, then $kP$ reduces to the identity modulo $v$ and then also the point $rP=kP-q(n_PP)$ reduces to the identity. This is absurd from the definition of $n_P$.
\end{proof}
\begin{proposition}
Assume that $P$ has non-singular reduction and $(P.O)_v=0$. Then, $k_{v,n}(P)=0$ for all $n\geq 1$.
\end{proposition}
\begin{remark}
This proposition was proved for number fields by Ayad \cite{ayad}. Their proof works also in our case, but it is very different from our.
\end{remark}
\begin{proof}
Since $P$ has non-singular reduction, then $c_v(kP)=0$ for any $k$ (see Lemma \ref{tab:corr_terms}). Hence, from Lemma \ref{lemma:psiinter}, $v(\psi_k(P)) = (kP.O)_v$ for all $k$ such that $kP\neq O$.
Assume $nP\neq O$.
If $(nP.O)_v>0$, then due to Lemma \ref{lemma:div} the numbers $((n-1)P.O)_v, ((n+1)P.O)_v$ are zero. Thus
$v(\psi_n(P)^2) =(nP.O)_v>0$, $v(\psi_{n-1}(P)\psi_{n+1}(P))=0$ and $v(\phi_n(P)) = 0$ from the properties (A), (B), (C), proving $k_{v,n}(P)=0$.
If $(nP.O)_v = 0$, then $v(\psi_n(P)^2)=0$ and $v(\phi_{n}(P))\geq 0$, hence again $k_{v,n}(P)=0$. Here we are using that $\psi_n^2$ and $\phi_n$ have integer coefficients.
Assume now that $nP=O$. Then $\psi_n^2(x(P))=0$ and so $\phi_n(x(P))=\psi_{n-1}(P)\psi_{n+1}(P)$. Observe that $((n\pm 1)P.O)_v=0$ since otherwise $(P.O)_v$ would be strictly positive. Hence, $k_{v,n\pm1}(P)=v(\psi_{n\pm1}^2(P))$. For the first part of the lemma, $k_{v,n\pm1}(P)=0$ and then
\[
v(\phi_n(x(P)))=v(\psi_{n-1}(P)\psi_{n+1}(P))=\frac{k_{v,n-1}(P)+k_{v,n+1}(P)}2=0.
\]
\end{proof}
Now, we are ready to prove our main result.
\begin{proof}[Proof of Theorem \ref{main}]
Assume that $P$ has non-singular reduction.
If $v(x(P))< 0$, then, using that $\psi_n^2$ (resp. $\phi_n$) has integer coefficients and degree $n^2-1$ (resp. $n^2$), we have $v(\psi_n^2(x(P)))\geq (n^2-1)v(x(P))$ and $v(\phi_n(x(P)))=n^2v(x(P))$ (recall that $\phi_n$ is monic). So, $k_{v,n}(P)=n^2v(x(P))$. If $v(x(P))\geq 0$, then $(P.O)_v=0$ and we conclude thanks to the previous proposition.
Assume now that $P$ has singular reduction, that $n_P$ exists, and that $n_P\mid n$. Then, $v(x(nP))<0$. Therefore, $v(\phi_n(x(P)))<v(\psi_n^2(x(P)))$ and so $k_{v,n}=v(\phi_n(x(P)))$. Since $P$ is singular we have $v(x(P))\geq 0$. Indeed, if $v(x(P))< 0$ then $P$ is the identity modulo $v$ and the identity is a non-singular point. Therefore, $v(\phi_n(x(P)))< v(x(P)\psi_n^2(x(P)))$. Recall that $\phi_n(x(P))=x(P)\psi_n^2(x(P))+\psi_{n-1}(P)\psi_{n+1}(P)$. So, \[v(\phi_n(x(P)))=v(\psi_{n+ 1}(P))+v(\psi_{n- 1}(P)).\] Observe that $((n\pm1)P.O)_v=0$ and then in particular $(n\pm 1)P\neq O$. Recall that $c_v(P)$ depends only on $comp_v(P)$ and observe that $c_v(P)=c_v(-P)$. Thus, $c_v((n\pm 1)P)=c_v(P)$ since $m_P\mid n_P\mid n$, . Hence, we can apply Lemma \ref{lemma:psiinter} and we have \[v(\psi_{n\pm 1}(P)) =\frac{(n\pm 1)^2 c_v(P)-c_v((n\pm 1)P)}2= \frac{((n\pm 1)^2-1)c_v(P)}2.\]
Thus,
\[k_{v,n}=v(\phi_n(x(P)))=v(\psi_{n- 1}(P)\psi_{n+1}(P))= n^2 c_v(P)=n^2 c_v(P)-c_v(nP).\]
It remains to prove the case when $P$ has singular reduction and that $n_P$ does not exist or it exist but $n_P\nmid n$. In this case $v(x(nP))\geq 0$ and so $k_{v,n}=v(\psi_n^2(x(P)))$. Observe that $(P.O)_v=(nP.O)_v=0$ and then from Lemma \ref{lemma:psiinter},
\[
k_{v,n}=v(\psi_n^2(x(P)))=n^2c_v(P)-c_v(nP).
\]
To conclude the proof, we want to show how to compute the values $n^2c_v(P)-c_v(nP)$, as we did in Table \ref{tab:common_vals}. If the Kodaira symbol is not $I_m$, then the value can be computed directly using Lemma \ref{tab:corr_terms}. If the Kodaira symbol is $I_m$, then one can easily prove that $a_{nP}\equiv a_Pn\mod m$. Therefore, \[n^2c_v(P)-c_v(nP)=n^2\dfrac{a_{P}\left( m-a_{P} \right)}{m}-\dfrac{n'\left( m-n' \right)}{m}\]
with $n'$ the smallest positive integer such that $a_{P}n \equiv n' \bmod{m}$ and this agrees with our computation of $n^2c_v(P)-c_v(nP)$ in Table \ref{tab:common_vals}.
\end{proof}
Now, we show an example of how to compute $k_{v,n}$ in some particular cases. For more details on the following examples, see \cite{Naskrecki_Acta}.
\begin{example}
Let $\kappa=\mathbb{C}$, $C=\P_\kappa^1$, and $K=\kappa(C)$. Then, $K=\mathbb{C}(t)$. Consider the elliptic curve $y^2=x(x-f^2)(x-g^2)$ defined over $K$ with $f^2+g^2=h^2$ for three polynomials $(f,g,h)=(t^2-1,2t,t^2+1)$. Let $P=((f-h)(g-h),(f+g)(f-h)(g-h))\in E(K)$. We denote by $v_1$ the place such that $v_1(t-1)=1$. One can easily show that $E$ is singular modulo $v_1$ since $v_1(f)=1$. Moreover $v_1(g-h)=2$ and then $v_1(x(P))>0$ and $v_1(y(P))>0$. So, $P$ is non-singular modulo $v$. By direct computation, $v_1(\Delta_E)=4$ and $v_1(j(E))=-4$. Then, $E$ is in minimal form over $K_{v_1}$ and $E/K_{v_1}$ has Kodaira symbol $I_4$.
By direct computation $2P=(t^4 + 2t^2 + 1,-2t^5 + 2t )$ and then $2P$ is non-singular modulo $v_1$, observing that $v_1(x(2P))=0$. So, $a_{P,v_1} =2$ and $m_{P,v_1}=2$. Using Lemma \ref{tab:corr_terms}, we have
\[
c_{v_1}(P)=\frac{2(4-2)}{4}=1.
\]
Therefore, thanks to Theorem \ref{main},
\[
k_{v_1,n}(P)=\begin{cases}
n^2-1 \text{ if $n$ is odd,}\\
n^2 \text{ if $n$ is even.}
\end{cases}
\]
This agrees with the direct computation of some cases for $n$ small. If $n=2$, then
\[
\psi_2^2(x(P))=4(x(P)(x(P)-f^2)(x(P)-g^2))=16(t^2+2t-1)^2(t^2-2t+1)^2
\]
and $v_1(\psi_2^2(x(P)))=4$.
If $n=3$, then
\[
\psi_3^2(x(P))=256(t^4+4t-1)^2(t^2+2t-1)^4(t-1)^8
\]
and so $v_1(\psi_3^2(x(P)))=8$. Moreover, $v_1(\psi_4^2(x(P)))=18$ and $v_1(x(P))=2$. Therefore, using $\phi_n(x(P))=x(P)\psi_n^2(x(P))-\psi_{n-1}(P)\psi_{n+1}(P)$ we obtain
\[
\phi_2(x(P))=\left(8(t^2+2t-1)(t^2-2t+1)\right)^2(2(t-1)^2)-16(t^4+4t-1)(t^2+2t-1)^2(t-1)^4
\]
and then $v(\phi_2(x(P)))= 4$. In the same way $v_1(\phi_3(x(P)))= 10$. Hence, we conclude that $k_{v_1,2}=4$ and
$k_{v_1,3}=8$.
\end{example}
\begin{example}
Consider $E/K$ and $P$ as in the previous example. Let $\alpha$ be a root of $t^2+2t-1=0$ and $v_\alpha$ be the place such that $v_\alpha(t-\alpha)=1$. In this case $E(K_{v_\alpha})$ has Kodaira type $I_2$ and $P$ is singular modulo $v_\alpha$. Therefore, $a_{P,v_\alpha} =1$ and $m_{P,v_\alpha}=2$. So,
\[
c_{v_\alpha}(P)=\frac{1(2-1)}{2}=\frac 12.
\]
Hence,
\[
k_{v_\alpha,n}(P)=\begin{cases}
\frac{n^2-1}2 \text{ if $n$ is odd,}\\
\frac{n^2}2 \text{ if $n$ is even.}
\end{cases}
\]
With the computation of the previous example one can check that this is true in the cases with $n$ small. For example, we know that $\psi_3^2(x(P))=256(t^4+4t-1)^2(t^2+2t-1)^4(t-1)^8$ and then $v_\alpha(\psi_3^2(x(P)))=4$. Moreover, $3P$ is singular and then $v_\alpha(\phi_3(x(P)))\geq v_\alpha(\psi_3^2(x(P)))$. So, $k_{v_\alpha,3}(P)=4$.
\end{example}
\bibliographystyle{alpha}
| 06e3f018719013ef623f9330e6f9cc838c11aeb4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Anomaly detection (AD) has numerous real-world applications, especially for tabular data, including detection of fraudulent transactions, intrusions for cybersecurity, and adverse outcomes in healthcare.
We propose a novel method that targets the following challenges in tabular AD:
\begin{itemize}[noitemsep,nolistsep,leftmargin=*]
\item \textbf{Noisy and irrelevant features}:
Tabular data (such as Electronic Health Records) often contain noisy or irrelevant features caused by measurement noise, outliers and inconsistent units.
Yet, even a change in a small subset of features may be considered anomalous.
\item \textbf{Heterogeneous features}:
Unlike image or text data, tabular data features can have values with significantly different types (numerical, boolean, categorical, and ordinal), ranges and distributions.
\item \textbf{Some labeled data}:
In many applications, often a small set of labeled data is available.
AD accuracy can be substantially boosted by using labeled information to identify some representative anomalies and ignore irrelevant ones.
\item \textbf{Interpretability}:
Without interpretable outputs, humans cannot understand the rationale behind anomalies, take actions to improve the AD model performance and build trust.
Verification of model accuracy is particularly challenging for tabular data as they are not easy to visualize for humans.
An interpretable AD model should identify important features used to classify anomalies to enable verification and build trust.
\end{itemize}
Most AD methods for tabular data fail to address the above challenges -- their performance often deteriorates with noisy features (see Sec.~\ref{sec:experiments}), they cannot incorporate labeled data, and they cannot provide interpretability (see Sec.~\ref{sec:related_work}).
In this paper, we aim to address these challenges by proposing a \textbf{D}ata-efficient \textbf{I}nterpretable \textbf{AD} approach, which we call \textbf{DIAD}. Our model architecture is inspired by Generalized Additive Models (GAMs) that have been shown to obtain high accuracy and interpretability for tabular data~\citep{caruana2015intelligible,chang2021interpretable,liu2021controlburn} and are widely used in many applications such as finding outlier patterns and auditing fairness~\citep{tan2018distill}.
We propose to employ intuitive notions of Partial Identification (PID) as AD objective and propose to learn them with differentiable GAMs.
PID scales to high-dimensional features and handles heterogeneous features, while the differentiable GAM architecture allows fine-tuning with labeled data and retain interpretability.
Furthermore, by fine-tuning using a differentiable AUC loss with a small amount of labeled samples, DIAD outperforms other semi-supervised learning AD methods.
For example, DIAD improves from 86.2\% to 89.4\% AUC with 5 labeled anomalies by unsupervised AD.
DIAD provides rationale on why an example is classified as anomalous using the GAM graphs and also gives insights on the impact of labeled data on decision boundary, which can be used to build global understanding about the task and to help improve the AD performance.
\section{Related Work}
\label{sec:related_work}
AD methods only with normal data are widely studied~\citep{kddtutorial}.
The methods closest to the proposed are the tree-based AD.
Isolation Forest (IF)~\citep{liu2008isolation} grows decision trees randomly and the shallower the tree depth for the data is, the more anomalous the data is.
However, this approach shows performance degradation when feature dimensionality increases.
Robust Random Cut Forest (RRCF, \citet{guha2016robust}) further improves IF by choosing features to split based on the size of its range, but it is more sensitive to scale.
PIDForest~\citep{gopalan2019pidforest} zooms on the features which have large variances which makes it robust to noisy or irrelevant features.
Another family of AD methods are based on generative approaches that learn to reconstruct input features, and use the error of reconstructions or density to identify anomalies.
\citet{bergmann2018improving} employs auto-encoders for image data.
DAGMM~\citep{zong2018deep} first learns an auto-encoder and uses a Gaussian Mixture Model to estimate density in the low-dimensional latent space.
Since these generative approaches are based on reconstructing input features, they may not easily fit in high-dimensional tabular data with noisy and heterogeneous features and small amount of labeled examples.
\setlength\tabcolsep{1.5pt}
\begin{table}[tbp]
\centering
\caption{Comparison of related works.}
\label{table:related_work}
\resizebox{\columnwidth}{!}{
\begin{tabular}{c|ccccc}
\toprule
& \small{\textbf{\makecell{Utilizing\\unlabeled\\data}}} & \small{\textbf{\makecell{Handling\\noisy\\features}}} & \small{\textbf{\makecell{Handling\\heterogen-\\ous features}}} & \small{\textbf{\makecell{Utilizing\\labeled\\data}}} & \small{\textbf{\makecell{Interpret-\\ability}}} \\
\midrule
PIDForest & \checkmark & \checkmark & \checkmark & \ding{55} & \ding{55} \\
DAGMM & \checkmark & \ding{55} & \ding{55} & \ding{55} & \checkmark \\
GOAD & \checkmark & \checkmark & \checkmark & \ding{55} & \ding{55} \\
Deep SAD & \checkmark & \checkmark & \checkmark & \checkmark & \ding{55} \\
ICL & \checkmark & \ding{55} & \checkmark & \ding{55} & \checkmark \\
DevNet & \ding{55} & \checkmark & \checkmark & \checkmark & \ding{55} \\ \midrule
\makecell{DIAD\\(Ours)} & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark \\
\bottomrule
\end{tabular}
}
\end{table}
Recently, methods with pseudo-tasks have been proposed for AD.
One major line of work is to predict geometric transformations~\citep{golan2018deep, bergman2020classification}, such as rotation or random transformations, and use the prediction errors to detect anomalies.
\citet{qiu2021neural} instead learns a set of diverse transformations and show improvements for tabular and text data.
CutPaste~\citep{li2021cutpaste} learns to classify the images with patches replaced by another image, combined with density estimation in the latent space for image data.
Several recent works focus on contrastive learning.
\citet{tack2020csi} learns to distinguish synthetic samples from the original distribution and achieves success on image data.
\citet{sohn2021learning} first learns a distribution-augmented contrastive representation and then uses a one-class classifier to identify anomalies.
Internal Contrastive Learning (ICL, \citet{shenkar2022anomaly}) tries to distinguish between in-window and out-of-window features by a sliding window over features and utilizes the error to identify anomalies, achieving state-of-the-art performance for tabular data.
A few AD works focus on explainability.
\citet{vinh2016discovering,liu2020lp} explain anomalies from other off-the-shelf detectors; thus, their explanations may not reflect the rationales of the detectors.
\citet{liznerski2021explainable} proposes to identify anomalies with a one-class classifier; it uses a CNN without fully connected layers so each output unit corresponds to a receptive field in the input image.
This allows visualizations of which part of the input images leads to high error in the output with accurate localization; however, this approach is limited to image data.
Several works have been proposed for semi-supervised AD.
Deep SAD~\citep{ruff2019deep} extends the deep one-class classification method DSVDD~\citep{ruff2018deep} into the semi-supervised setting.
However, their approach is non-interpretable and performs even worse than the unsupervised One-Class SVM for tabular data, while the DIAD outperforms it.
DevNet~\citep{pang2019deep} formulates the AD problem into a regression problem and achieves better sample complexity under limited labeled data.
\citet{yoon2020vime} trains an embedding similar to BERT~\citep{devlin2018bert} combined with consistency loss~\citep{sohn2020fixmatch} and achieves the state-of-the-art semi-supervised performance in tabular data.
See Table~\ref{table:related_work} for a comparison.
\section{Methods}
\paragraph{Model Architecture}
To render AD interpretable and allow back-propagation to learn from labeled data, our work is inspired from the NodeGAM~\citep{chang2021node}, an interpretable and differentiable tree-based GAM model.
\paragraph{What are GAM and GA$^2$M?}
GAMs and GA$^2$Ms are considered as white-box models since they allow humans to visualize their functional forms in 1-D or 2-D plots by not allowing any 3-way or more feature interactions.
Specifically, given an input $x\in\mathbb{R}^{D}$, a label $y$, a link function $g$ (e.g. $g$ is $\log ({p}/{1 - p})$ in binary classification), main effects $f_j$ for each feature $j$, and 2-way feature interactions $f_{jj'}$, GAM and GA$^2$M are expressed as:
\begin{align*}
&\text{\textbf{GAM}:\ \ \ \ }
g(y) = f_0 + \sum_{j=1}^D f_j(x_j), \ \ \ \\
&\text{\textbf{GA$^2$M}:\ \ \ \ }
g(y) = f_0 + \sum_{j=1}^D f_j(x_j) + \sum_{j=1}^D\sum_{j' > j} f_{jj'}(x_j, x_{j'}).
\end{align*}
Unlike commonly-used deep learning architectures that use all the feature interactions, GAMs and GA$^2$M are restricted to lower-order interaction (1 or 2-way), so the impact of each $f_j$ or interaction $f_{jj'}(x_j, x_{j'})$ can be visualized independently as a graph.
That means, for $f_j$, we may plot $x_j$ on the x-axis and $f_j(x_j)$ on the y-axis. To plot $f_{jj'}$, we show a scatter plot with $x_j$ on the x-axis and $x_j'$ on the y-axis, where color indicates the value $f_{jj'}$.
Note that a linear model is a special case of GAM.
Humans can easily simulate the decision making of a GAM by reading $f_j$ and $f_{jj'}$ from the graph for each feature $j$ and $j'$ and taking the sum.
Here we review NodeGA$^2$M relevant to our changes.
NodeGA$^2$M is based on GA$^2$M but uses the neural trees to learn feature functions $f_j$ and $f_{jj'}$.
Specifically, NodeGA$^2$M consists of $L$ layers where each layer has $m$ differentiable soft oblivious decision trees (ODT) of equal depth $C$ where each tree only interacts with at most 2 features.
Next, we describe ODTs.
\begin{algorithm}[tb]
\caption{A soft decision tree}
\label{alg:pid_tree}
\begin{algorithmic}
\STATE {\bfseries Input:} Mini-batch $\bm{X} \in \mathbb{R}^{B \times D}$, Temperature $T$ ($T$ $\xrightarrow{}$ 0)
\STATE {\bfseries Symbols:} Tree Depth $C$, Entmoid $\sigma$
\STATE {\bfseries Trainable Parameters}: Feature selection logits $\bm{F}^1, \bm{F}^2 \in \mathbb{R}^{D}$, split Thresholds $\bm{b} \in \mathbb{R}^{C}$, split slope $\bm{S} \in \mathbb{R}^{C}$,
\vspace{-5pt}
\\\hrulefill
\STATE $\bm{F} = [\bm{F}^1, \bm{F}^2, \bm{F}^1,...]^T \in \mathbb{R}^{D \times C}$ \COMMENT{Alternating $\bm{F}^1$, $\bm{F}^2$}
\STATE $G = \bm{X} \cdot \text{EntMax}(\bm{F} / T, \text{dim=0}) \in \mathbb{R}^{B \times C}$
\FOR{$c=1$ {\bfseries to} $C$}
\STATE $H^c = \sigma(\frac{(\bm{G^c} - b^c)}{S^c \cdot T})$ \COMMENT{Soft binary split}
\ENDFOR
\STATE
$
\bm{e} = \left( {\begin{bmatrix}
H^1 \\
(1 - H^1)
\end{bmatrix} }
\otimes \dots \otimes
{\begin{bmatrix}
(H^C) \\
(1 - H^C)
\end{bmatrix} }
\right) \in \mathbb{R}^{B \times 2^C}
$
\STATE
$E = \text{sum}(\bm{e}, \text{dim=0}) \in \mathbb{R}^{2^C}$ \COMMENT{Sum across batch}
\STATE {\bfseries Return:} $E$ count
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[tbp]
\caption{DIAD Update}
\label{alg:pid_update}
\begin{algorithmic}
\STATE {\bfseries Input:} Mini-batch $\bm{X}$, Tree model $\mathcal{M}$, Smoothing $\delta$, Leaf weights $w_{tl}$ for each tree $t$ and leaf $l$ in $\mathcal{M}$
\vspace{-5pt}
\\\hrulefill
\STATE $\bm{X}$ = MinMaxTransform($X$, min=$-1$, max=$1$)
\STATE $\bm{X_U} \sim U[-1, 1]$ \COMMENT{Data uniformly drawn from [-1, 1]}
\STATE $E^{tl} = \mathcal{M}(X)$ \COMMENT{counts -- See Alg.~\ref{alg:pid_tree}}
\STATE $E_U^{tl} = \mathcal{M}(X_U)$
\STATE $E^{tl} = E^{tl} + \delta$, $E_U^{tl} = E_U^{tl} + \delta$ \COMMENT{Smooth the counts}
\STATE $V_{tl} = \frac{E^{tl}}{\sum_{n'} E^{tl'}}$ \COMMENT{Volume ratio}
\STATE $D_{tl} = \frac{E_U^{tl}}{\sum_{n'} E_U^{tl'}}$ \COMMENT{Data ratio}
\STATE $M_{tl} = \frac{V_{tl}^2}{P_{tl}}$ \COMMENT{Second moments}
\STATE $m_t \sim \text{Bernoulli}(p)$ \COMMENT{Sample masks per tree $t$}
\STATE $M_{tl} = \frac{m_t}{p} * M_{tl} + (1 - m_t) * M_{tl}$ \COMMENT{Per-tree dropout}
\STATE $L_M = -\sum_{t,l} M_{tl}$ \COMMENT{Maximize the second moments}
\STATE $\hat{s}_{tl} = {V_{tl}} / {P_{tl}}$ \COMMENT{Sparsity}
\STATE $s_{tl} = (\frac{2\hat{s}_{tl}}{(\max \hat{s}_{tl} - \min \hat{s}_{tl})} - 1)$ \COMMENT{Normalize to [-1, 1] -- Eq.~\ref{eq:norm_sp}}
\STATE $w_{tl} = (1 - \gamma)w_{tl} + \gamma s_{tl}$ \COMMENT{Update weights -- Eq.~\ref{eq:sp_update}}
\STATE Optimize $L_M$ by Adam optimizer
\end{algorithmic}
\end{algorithm}
\paragraph{Differentiable ODTs:}
An ODT works like a traditional decision tree except for all nodes in the same depth share the same input features and thresholds, which allows parallel computation and makes it suitable for deep learning.
Specifically, an ODT of depth $C$ compares chosen $C$ input features to $C$ thresholds, and returns one of the $2^C$ possible options.
Mathematically, for feature functions $F^c$ which choose what features to split, splitting thresholds $b^c$, and a response vector $\bm{R} \in \mathbb{R}^{2^C}$,
the tree output $h(\bm{x})$ is given as:
\begin{equation}
h(\bm{x}) = \bm{R} \cdot \left( {\begin{bmatrix}
\mathbb{I}(F^1(\bm{x}) - b^1) \\
\mathbb{I}(b^1 - F^1(\bm{x}))
\end{bmatrix} }
{\otimes}
\dots {\otimes}
{\begin{bmatrix}
\mathbb{I}(F^C(\bm{x}) - b^C) \\
\mathbb{I}(b^C - F^C(\bm{x}))
\end{bmatrix} },
\right).
\end{equation}
where $\mathbb{I}$ is the step function, $\otimes$ is the outer product and $\cdot$ is the inner product.
Both feature functions $F^c$ and $\mathbb{I}$ would prevent differentiability.
To address this, $F^c(\bm{x})$ is replaced with a weighted sum of features with a temperature annealing that makes it gradually one-hot:
\begin{equation}
\label{eq:split}
F^c(\bm{x}) = \sum_{j=1}^D x_j \text{entmax}_{\alpha}(F^c / T)_{j}, \ \ \ \ T \rightarrow 0.
\end{equation}
where $\bm{F}^c \in \mathbb{R}^{D}$ are the logits for which features to choose, and entmax$_{\alpha}$~\citep{entmax} is the entmax normalization function as the sparse version of softmax such that the sum equals to $1$.
As $T \rightarrow 0$, the output of entmax will gradually become one-hot.
Also, we replace $\mathbb{I}$ with entmoid which works like a sparse sigmoid that has output values between $0$ and $1$.
Unlike NodeGAM, we also introduce a temperature annealing in the entmoid to make it closer to $\mathbb{I}$ since it performs better under our AD objective. That is:
\begin{equation}
\text{entmoid}(\frac{\bm{x}}{T}) \rightarrow \mathbb{I}, \ \ \ \ \text{as\ \ } T \rightarrow 0.
\end{equation}
Differentiability of all operations (entmax, entmoid, outer and inner products), render ODT differentiable.
\paragraph{Stacking trees into deep layers:}
Similar to residual connections, all tree outputs $\bm{h}(\bm{x})$ from previous layers become the inputs to the next layer.
For input features $\bm{x}$, the inputs $\bm{x}^l$ to each layer $l$ become:
\begin{equation}
\bm{x}^1 = \bm{x}, \ \ \ \bm{x}^l = [\bm{x}, \bm{h}^{1}(\bm{x}^1), ... , \bm{h}^{(l-1)}(\bm{x}^{(l-1)})] \text{\ \ for\ \ } l > 1.
\end{equation}
The final output of the model $\hat{y}(\bm{x})$ is the average of all the tree outputs ${\bm{h}_1,...,\bm{h}_L}$ of all $L$ layers:
\begin{equation}
\label{eq:final_output}
\hat{y}(x) = \sum\nolimits_{l=1}^L \sum\nolimits_{i=1}^{m} h_{li}(\bm{x^l}) / ({L \cdot m}).
\end{equation}
\setlength\tabcolsep{2pt}
\paragraph{GA$^2$M design}
To allow only maximum two-way interactions, for each tree, at most two logits $\bm{F}^1$ and $\bm{F}^2$ are allowed, and let the rest of the depth the same as either $\bm{F}^1$ or $\bm{F}^2$: $\bm{F}^c = \bm{F}^{\left \lfloor{c / 2}\right \rfloor}$ for $c > 2$ -- this allows at most $2$ features to interact within each tree.
Also, we avoid the connection between two trees that focus on different feature combinations, since it may create higher feature interactions.
See Alg.~\ref{alg:pid_tree} for pseudo code.
\begin{table*}[tbp]
\centering
\caption{Unsupervised AD performance (\% of AUC) on 18 tabular datasets for DIAD and 9 baselines. Metrics with standard error overlapped with the best number are bolded. Methods not involving randomness do not have standard error. We show the number of samples (N) and the number of features (P). Datasets are ordered by N.}
\label{table:unsup_perf}
\begin{tabular}{c|cccccccccc|cc}
\toprule
& \textbf{DIAD} & \textbf{PIDForest} & \textbf{IF} & \textbf{COPOD} & \textbf{PCA} & \textbf{ICL} & \textbf{kNN} & \textbf{RRCF} & \textbf{LOF} & \textbf{OC-SVM} & N & P \\
\midrule
Vowels & 78.3\tiny{ $\pm$ 0.9 } & 74.0\tiny{ $\pm$ 1.0 } & 74.9\tiny{ $\pm$ 2.5 } & 49.6 & 60.6 & 90.8\tiny{ $\pm$ 2.1 } & \textbf{97.5} & 80.8\tiny{ $\pm$ 0.3 } & 5.7 & 77.8 & 1K & 12 \\
Siesmic & 72.2\tiny{ $\pm$ 0.4 } & 73.0\tiny{ $\pm$ 0.3 } & 70.7\tiny{ $\pm$ 0.2 } & 72.7 & 68.2 & 65.3\tiny{ $\pm$ 1.6 } & \textbf{74.0} & 69.7\tiny{ $\pm$ 1.0 } & 44.7 & 60.1 & 3K & 15 \\
Musk & 90.8\tiny{ $\pm$ 0.9 } & \textbf{100.0}\tiny{ $\pm$ 0.0 } & \textbf{100.0}\tiny{ $\pm$ 0.0 } & 94.6 & \textbf{100.0} & 93.3\tiny{ $\pm$ 0.7 } & 37.3 & 99.8\tiny{ $\pm$ 0.1 } & 58.4 & 57.3 & 3K & 166 \\
Satimage & \textbf{99.7}\tiny{ $\pm$ 0.0 } & 98.2\tiny{ $\pm$ 0.3 } & 99.3\tiny{ $\pm$ 0.1 } & 97.4 & 97.7 & 98.0\tiny{ $\pm$ 1.3 } & 93.6 & 99.2\tiny{ $\pm$ 0.2 } & 46.0 & 42.1 & 6K & 36 \\
Thyroid & 76.1\tiny{ $\pm$ 2.5 } & \textbf{88.2}\tiny{ $\pm$ 0.8 } & 81.4\tiny{ $\pm$ 0.9 } & 77.6 & 67.3 & 75.9\tiny{ $\pm$ 2.2 } & 75.1 & 74.0\tiny{ $\pm$ 0.5 } & 26.3 & 54.7 & 7K & 6 \\
A. T. & 78.3\tiny{ $\pm$ 0.6 } & \textbf{81.4}\tiny{ $\pm$ 0.6 } & 78.6\tiny{ $\pm$ 0.6 } & 78.0 & 79.2 & 79.3\tiny{ $\pm$ 0.7 } & 63.4 & 69.9\tiny{ $\pm$ 0.4 } & 43.7 & 67.0 & 7K & 10 \\
NYC & 57.3\tiny{ $\pm$ 0.9 } & 57.2\tiny{ $\pm$ 0.6 } & 55.3\tiny{ $\pm$ 1.0 } & 56.4 & 51.1 & 64.5\tiny{ $\pm$ 0.9 } & \textbf{69.7} & 54.4\tiny{ $\pm$ 0.5 } & 32.9 & 50.0 & 10K & 10 \\
Mammography & 85.0\tiny{ $\pm$ 0.3 } & 84.8\tiny{ $\pm$ 0.4 } & 85.7\tiny{ $\pm$ 0.5 } & \textbf{90.5} & 88.6 & 69.8\tiny{ $\pm$ 2.7 } & 83.9 & 83.2\tiny{ $\pm$ 0.2 } & 28.0 & 87.2 & 11K & 6 \\
CPU & 91.9\tiny{ $\pm$ 0.2 } & 93.2\tiny{ $\pm$ 0.1 } & 91.6\tiny{ $\pm$ 0.2 } & \textbf{93.9} & 85.8 & 87.5\tiny{ $\pm$ 0.3 } & 72.4 & 78.6\tiny{ $\pm$ 0.3 } & 44.0 & 79.4 & 18K & 10 \\
M. T. & 81.2\tiny{ $\pm$ 0.2 } & 81.6\tiny{ $\pm$ 0.3 } & 82.7\tiny{ $\pm$ 0.5 } & 80.9 & \textbf{83.4} & 81.8\tiny{ $\pm$ 0.4 } & 75.9 & 74.7\tiny{ $\pm$ 0.4 } & 49.9 & 79.6 & 23K & 10 \\
Campaign & 71.0\tiny{ $\pm$ 0.8 } & \textbf{78.6}\tiny{ $\pm$ 0.8 } & 70.4\tiny{ $\pm$ 1.9 } & \textbf{78.3} & 73.4 & 72.0\tiny{ $\pm$ 0.5 } & 72.0 & 65.5\tiny{ $\pm$ 0.3 } & 46.3 & 66.7 & 41K & 62 \\
smtp & 86.8\tiny{ $\pm$ 0.5 } & \textbf{91.9}\tiny{ $\pm$ 0.2 } & 90.5\tiny{ $\pm$ 0.7 } & 91.2 & 82.3 & 82.2\tiny{ $\pm$ 2.0 } & 89.5 & 88.9\tiny{ $\pm$ 2.3 } & 9.5 & 84.1 & 95K & 3 \\
Backdoor & \textbf{91.1}\tiny{ $\pm$ 2.5 } & 74.2\tiny{ $\pm$ 2.6 } & 74.8\tiny{ $\pm$ 4.1 } & 78.9 & 88.7 & \textbf{91.8}\tiny{ $\pm$ 0.6 } & 66.8 & 75.4\tiny{ $\pm$ 0.7 } & 28.6 & 86.1 & 95K & 196 \\
Celeba & \textbf{77.2}\tiny{ $\pm$ 1.9 } & 67.1\tiny{ $\pm$ 4.8 } & 70.3\tiny{ $\pm$ 0.8 } & 75.1 & \textbf{78.6} & 75.4\tiny{ $\pm$ 2.6 } & 56.7 & 61.7\tiny{ $\pm$ 0.3 } & 56.3 & 68.5 & 203K & 39 \\
Fraud & \textbf{95.7}\tiny{ $\pm$ 0.2 } & 94.7\tiny{ $\pm$ 0.3 } & 94.8\tiny{ $\pm$ 0.1 } & 94.7 & 95.2 & \textbf{95.5}\tiny{ $\pm$ 0.2 } & 93.4 & 87.5\tiny{ $\pm$ 0.4 } & 52.5 & 94.8 & 285K & 29 \\
Census & \textbf{65.6}\tiny{ $\pm$ 2.1 } & 53.4\tiny{ $\pm$ 8.1 } & 61.9\tiny{ $\pm$ 1.9 } & \textbf{67.4} & 66.1 & 58.4\tiny{ $\pm$ 0.9 } & 64.6 & 55.7\tiny{ $\pm$ 0.1 } & 45.0 & 53.4 & 299K & 500 \\
http & 99.3\tiny{ $\pm$ 0.1 } & 99.2\tiny{ $\pm$ 0.2 } & \textbf{100.0}\tiny{ $\pm$ 0.0 } & 99.2 & 99.6 & 99.3\tiny{ $\pm$ 0.1 } & 23.1 & 98.4\tiny{ $\pm$ 0.2 } & 64.7 & 99.4 & 567K & 3 \\
Donors & \textbf{87.7}\tiny{ $\pm$ 6.2 } & 61.1\tiny{ $\pm$ 1.3 } & 78.3\tiny{ $\pm$ 0.7 } & 81.5 & 82.9 & 65.5\tiny{ $\pm$ 11.8 } & 61.2 & 64.1\tiny{ $\pm$ 0.0 } & 40.2 & 70.2 & 619K & 10 \\
\midrule
Average & \textbf{82.5} & 80.7 & 81.2 & 81.0 & 80.5 & 80.3 & 70.6 & 76.8 & 40.2 & 71.0 & - & - \\
Rank & \textbf{3.6} & 4.4 & 4.0 & 4.2 & 4.2 & 4.7 & 6.6 & 6.7 & 9.8 & 6.8 & - & - \\
\bottomrule
\end{tabular}
\vspace{-5pt}
\end{table*}
\paragraph{AD Objectives}
Here we introduce the AD objective for our tree-based model: Partial Identification (PID)~\citep{gopalan2019pidforest}.
Consider all patients admitted into an ICU.
We might consider patients with blood pressure (BP) as 300 as anomalous, since it deviates from most others.
In this example, BP of 300 is in a "sparse" feature space since very few people are more than 300.
To formalize this intuition, we need to introduce the concept of volume.
We first define the max and min value of each feature value.
Then, we define the volume of a tree leaf as the product of the proportion of the splits within the min and max value.
For example, assuming the max value of BP is 400 and min value is 0, the tree split of "BP $\ge$ 300" has a volume $0.25$.
We define the sparsity $s_l$ of a tree leaf $l$ as the ratio between the volume of the leaf $V_l$ and the \% of data in the leaf $D_l$:
$$
s_l = {V_l} / {D_l},
$$
and we treat the higher sparsity as more anomalous.
Let's assume only less than 0.1\% have values more than 300 and the volume of "BP $\ge$ 300" is 0.25, this patient is quite anomalous in the data by having a large sparsity $\frac{0.25}{0.1\%}$.
To learn the effective splitting of regions with high vs. low sparsity, we optimize the tree structures to maximize the variance of sparsity across leafs, as it splits the space into a high (anomalous) and a low (normal) sparsity region.
Note that the expected sparsity weighted by the number of data in each leaf is a constant $1$.
Given each tree leaf $l$, the \% of data in the leaf is $D_l$, sparsity $s_l$:
\begin{equation}
\mathbb{E}[s] = \sum_l [D_l s_l] = \sum_l [D_l \frac{V_l}{D_l}] = \sum_l [V_l] = 1.
\end{equation}
Therefore, maximizing variance equals to maximizing the second moment of sparsity since the first moment is $1$:
\begin{equation}
\max \text{Var}[s] = \max \sum\nolimits_l D_l s_l^2 = \max \sum\nolimits_l {V_l^2}/{D_l}.
\end{equation}
\paragraph{Estimating volume and the data ratio}
The above objectives require estimating \% of volume $V_l$ and \% of data $D_l$ for each leaf $l$.
However, calculating volume exactly is not trivial in an oblivious decision tree since it involves complicated rules extractions.
Instead, we sample random points uniformly in the input space, and count the number of the random points that end up in each tree leaf.
And more points in a leaf indicate higher volume.
To avoid the zero count in the denominator, we use Laplacian smoothing, which adds a constant $\delta$ to each count.
We find it's crucial to set a large $\delta$, around 50-100, to encourage models to ignore the tree leaves with fewer counts.
Similarly, we estimate $D_l$ by counting the data ratio in each mini-batch.
We add $\delta$ for both $V_l$ and $D_l$.
\paragraph{Regularization}
To encourage diverse trees, we introduce a per-tree dropout noise on the estimated momentum to make each tree operate on a different subset of samples in a mini-batch.
We also restrict each tree to only split on $\rho\%$ of features randomly to promote diverse trees (Supp.~\ref{appx:hparams}).
\paragraph{Updating leafs' weight}
We set the leafs' response as the sparsity to reflect the degree of the anomaly.
However, since sparsity estimation involves randomness, we set the response as the damped value of sparsity to stabilize the performance.
Specifically, given the training step $i$, sparsity $s_l^i$ for each leaf $l$, and the update rate $\gamma$:
\begin{equation}
w_l^i = (1 - \gamma)w_l^{(i-1)} + \gamma s_l^i.
\label{eq:sp_update}
\end{equation}
\paragraph{Normalizing sparsity}
Because of the residual connections, the output of each tree adds with the input features and the summation becomes the input to the next tree.
But this creates a very large magnitude difference -- as the output of trees could have sparsity values up to $10^4$ but the input feature is normalized to -1 and 1, naive optimization tends to ignore the input features.
Also, the large outputs of trees make fine-tuning hard and lead to inferior performance in the semi-supervised setting (Sec.~\ref{sec:ablation}).
To circumvent this magnitude difference, we linearly scale the min and max value of the estimated sparsity to -1 and 1:
\begin{equation}
\hat{s}_l = {V_l} / {D_l},\ \ \ \ \ s_l = {2\hat{s}_l} / {(\max_l \hat{s}_l - \min_l \hat{s}_l)} - 1.
\label{eq:norm_sp}
\end{equation}
Algorithm ~\ref{alg:pid_update} overviews all training update steps.
\paragraph{Incorporating labeled data}
To optimize the labeled data in the imbalanced setting, we optimize the differentiable AUC loss~\citep{yan2003optimizing} which has been shown effective in the imbalanced setting.
Specifically, given a mini-batch of labeled positive data $X_P$ and labeled negative data $X_N$, model $\mathcal{M}$, it minimizes
\begin{equation}
L_{PN} = \frac{1}{|X_P| |X_N|} \sum_{x_p \in X_P, x_n \in X_N} \max(\mathcal{M}(x_n) - \mathcal{M}(x_p), 0).
\end{equation}
We compare this AUC loss to commonly-used Binary Cross Entropy (BCE) loss (Sec.~\ref{sec:ablation}).
\paragraph{Data Loader} Similar to Devnet~\citep{devnet}, we upsample the positive samples to make each mini-batch have the same number of positive and negative samples.
We find it improves over uniform sampling (Sec.~\ref{sec:ablation}).
\setlength\tabcolsep{2pt}
\begin{figure*}[tbp]
\begin{center}
\begin{tabular}{ccccc}
& (a) Vowels & (b) Satimage & (c) Thyroid & (d) NYC \\
\raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small AUC}}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/vowels.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/satimage-2_nolegend.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/thyroid_nolegend.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/nyc_taxi_nolegend.pdf} \vspace{-5pt} \\
& \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies \vspace{5pt} \\
& (e) CPU & (f) Campaign & (g) Backdoor & (h) Donors \\
\raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small AUC}}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/cpu_utilization_asg_misconfiguration.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/campaign_nolegend.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/backdoor_nolegend.pdf}
& \includegraphics[width=0.24\linewidth]{figures/ss_summary2/donors_nolegend.pdf} \vspace{-5pt} \\
& \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies & \ \ \ \ \ No. Anomalies \vspace{5pt} \\
\end{tabular}
\end{center}
\caption{
Semi-supervised AD performance on 8 tabular datasets (out of 15) with varying number of anomalies.
Our method `DIAD' (blue) outperforms other semi-supervised baselines. Summarized results can be found in Table.~\ref{table:ss_summary}.
See the remaining plots with 7 tabular datasets in Supp.~\ref{appx:ss_figures_appx}.
}
\label{fig:ss}
\end{figure*}
\section{Experiments}
\label{sec:experiments}
We evaluate DIAD on various tabular datasets, in both unsupervised and semi-supervised settings. Detailed experimental settings and additional results are in Supplementary.
\subsection{Unsupervised Anomaly Detection}
We compare methods on $20$ tabular datasets, including $14$ datasets used in \citet{gopalan2019pidforest} and $6$ larger datasets from \citet{devnet}.
Since it's hard to tune hyperparameters in the unsupervised setting, for fair comparisons we compare all baselines using default hyperparameters. We run experiments 8 times with different random seeds and take average if methods involve randomness.
\begin{table*}[t]
\caption{Unsupervised AD performance (\% of AUC) with additional 50 noisy features for DIAD and 9 baselines. We find both DIAD and OC-SVM deteriorate around 2-3\% while other methods deteriorate 7-17\% on average.}
\label{table:unsup_noisy_perf}
\centering
\begin{tabular}{c|cccccccccc}
\toprule
& \textbf{DIAD} & \textbf{PIDForest} & \textbf{IF} & \textbf{COPOD} & \textbf{PCA} & \textbf{ICL} & \textbf{kNN} & \textbf{RRCF} & \textbf{LOF} & \textbf{OC-SVM} \\
\midrule
Thyroid & 76.1\tiny{ $\pm$ 2.5 } & 88.2\tiny{ $\pm$ 0.8 } & 81.4\tiny{ $\pm$ 0.9 } & 77.6 & 67.3 & 75.9\tiny{ $\pm$ 2.2 } & 75.1 & 74.0\tiny{ $\pm$ 0.5 } & 26.3 & 54.7 \\
Thyroid (noise) & 71.1\tiny{ $\pm$ 1.2 } & 76.0\tiny{ $\pm$ 2.9 } & 64.4\tiny{ $\pm$ 1.6 } & 60.5 & 61.4 & 49.5\tiny{ $\pm$ 1.6 } & 49.5 & 53.6\tiny{ $\pm$ 1.1 } & 50.8 & 49.4 \\
Mammography & 85.0\tiny{ $\pm$ 0.3 } & 84.8\tiny{ $\pm$ 0.4 } & 85.7\tiny{ $\pm$ 0.5 } & 90.5 & 88.6 & 69.8\tiny{ $\pm$ 2.7 } & 83.9 & 83.2\tiny{ $\pm$ 0.2 } & 28.0 & 87.2 \\
Mammography (noise) & 83.1\tiny{ $\pm$ 0.4 } & 82.0\tiny{ $\pm$ 2.2 } & 71.4\tiny{ $\pm$ 2.0 } & 72.4 & 76.8 & 69.4\tiny{ $\pm$ 2.4 } & 81.7 & 79.1\tiny{ $\pm$ 0.7 } & 37.2 & 87.2 \\ \midrule
Average $\downarrow$ & $\bm{3.5}$ & 7.5 & 15.6 & 17.6 & 8.9 & 13.4 & 13.9 & 12.2 & -16.8 & $\bm{2.7}$ \\
\bottomrule
\end{tabular}
\end{table*}
\paragraph{Baselines}
We compare with ICL~\citep{shenkar2022anomaly}, a recent deep-learning AD method, and non-deep learning methods including PIDForest~\citep{gopalan2019pidforest}, COPOD~\citep{copod}, PCA, k-nearest neighbors (kNN), RRCF~\citep{guha2016robust}, LOF~\citep{breunig2000lof} and OC-SVM~\citep{scholkopf2001estimating}.
We use 2 aggregate metrics to summarize the performances across datasets: (1) \textbf{Average}: we take the average of AUC across datasets, (2) \textbf{Rank}: to avoid dominant impact of a few datasets, we calculate the rank of each method in each dataset and average across datasets (lower rank is better).
We demonstrate overall AUC performances in Table~\ref{table:unsup_perf}.
On average, our method performs the best in both Average and Rank.
DIAD, using up to 2nd order interactions, performs better or on par with other models for most datasets.
Compared to PIDForest, DIAD often underperforms on smaller datasets such as Musk and Thyroid, but outperforms on larger datasets like Backdoor, Celeba, Census and Donors.
In Table.~\ref{table:unsup_noisy_perf}, we evaluate the robustness of AD methods with the additional noisy features.
More specifically, we follow the experiment settings in \citet{gopalan2019pidforest} to include 50 additional noisy features which are randomly sampled from $[-1, 1]$ to Thyroid and Mammography datasets, and create Thyroid (noise) and Mammography (noise) respectively.
In Table.~\ref{table:unsup_noisy_perf}, we show that the performance of DIAD is robust with additional noisy features (76.1$\rightarrow$71.1, 85.0$\rightarrow$83.1), while others show significant performance degradation. For example, on Thyroid (noise), ICL decreases from 75.9$\rightarrow$49.5, KNN from 75.1$\rightarrow$49.5, and COPOD from 77.6$\rightarrow$60.5.
\setlength\tabcolsep{2pt}
\begin{figure*}[tbp]
\centering
\begin{tabular}{cccccccc}
& (a) Contrast (Sp=$0.38$) & & (b) Noise (Sp=$0.21$) & & (c) Area (Sp=$0.18$) & & \makecell{(d) Area x Gray Level\\(Sp=$0.05$)} \\
\raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Sparsity}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/mammography/0_imp0.382_contrast.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Sparsity}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/mammography/1_imp0.206_rms_noise_flucutation.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Sparsity}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/mammography/2_imp0.181_area.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Gray Level}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/mammography/3_imp0.053_area_gray_level.pdf}\vspace{-5pt} \\
& \ \ \ \ \ Contrast & & \ \ \ \ \ Noise & & \ \ \ \ \ Area & & Area \vspace{5pt} \\
\end{tabular}
\caption{
Explanations of the most anomalous sample in the Mammography dataset.
We show the top 4 contributing GAM plots with 3 features (a-c) and 1 two-way interaction (d).
(a-c) x-axis is the feature value, and y-axis is the model's predicted sparsity (higher sparsity represents more anomalous). Model's predicted sparsity is shown in the blue line.
The red backgrounds indicate data density and the green line indicates the value of the most anomalous sample with Sp as its sparsity.
The model finds it anomalous because it has high Contrast, Noise and Area differing from values where majority of other samples have.
(d) x-axis is the Area and y-axis is the Gray Level with color indicating the sparsity (blue/red indicates anomalous/normal). The green dot is the value of the data that has 0.05 sparsity.
}
\label{fig:mammography_gam_plots}
\end{figure*}
\subsection{Semi-supervised Anomaly Detection}
\label{sec:ss_result}
Next, we focus on the semi-supervised setting.
We show how much DIAD can improve the performance with small labeled data in comparison to alternatives.
We first divide the data into 64\%-16\%-20\% train-val-test splits, and within the training set only a small part of data is labeled.
Specifically, we assume the existence of labels for a small subset of the training set (5, 15, 30, 60 or 120 positives and the corresponding negatives to have the same anomaly ratio).
The validation set is used for model selection and we report the average performances evaluated on the disjoint 10 data splits.
We compare with 3 baselines: (1) \textbf{DIAD w/o PT}: we directly optimize our model under the small labeled data without the first AD pre-training stage. (2) \textbf{CST}: we compare with the Consistency Loss proposed in~\citet{vime} which regularizes the model to make similar predictions between unlabeled data under dropout noise injection. (3) \textbf{DevNet}~\citep{devnet}: the state-of-the-art semi-supervised AD methods.
Hyperparameters are in Supp.~\ref{appx:ss_hparams}.
\setlength\tabcolsep{4pt}
\begin{table}[tbp]
\centering
\caption{Summary of semi-supervised AD performances. We show the average \% of AUC across 15 datasets with varying number of anomalies.}
\label{table:ss_summary}
\begin{tabular}{c|cccccc}
\toprule
\textbf{No. Anomalies} & \textbf{0} & \textbf{5} & \textbf{15} & \textbf{30} & \textbf{60} & \textbf{120} \\
\midrule
DIAD & \textbf{87.1} & \textbf{89.4} & \textbf{90.0} & \textbf{90.4} & \textbf{89.4} & \textbf{91.0} \\
DIAD w/o PT & - & 86.2 & 87.6 & 88.3 & 87.2 & 88.8 \\
CST & - & 85.3 & 86.5 & 87.1 & 86.6 & 88.8 \\
DevNet & - & 83.0 & 84.8 & 85.4 & 83.9 & 85.4 \\
\bottomrule
\end{tabular}
\end{table}
Fig.~\ref{fig:ss} shows the AUC across 8 of 15 datasets (the rest can be found in Supp.~\ref{appx:ss_figures_appx}). The proposed version of DIAD (blue) outperforms DIAD without pre-training (orange) consistently in 14 of 15 datasets (except Census dataset), which demonstrates that learning the PID objectives from unlabeled data help improve the performance.
Second, both the consistency loss (green) and DevNet (red) do not always improve the performance in comparison to the supervised setting.
To conclude, DIAD outperforms all baselines and improve from the unlabeled setting.
\setlength\tabcolsep{2pt}
\begin{figure*}[tbp]
\begin{center}
\begin{tabular}{cccccccc}
& (a) Great Chat & & \makecell{(b) Great Messages\\Proportion} & & (c) Fully Funded & & (d) \makecell{Referred Count} \\
\raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Output}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/ft_donors/great_chat=t.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Output}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/ft_donors/great_messages_proportion.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Output}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/ft_donors/fully_funded=t.pdf}
& \raisebox{4\normalbaselineskip}[0pt][0pt]{\rotatebox[origin=c]{90}{\small Output}}\hspace{-5pt}
& \includegraphics[width=0.23\linewidth]{figures/ft_donors/teacher_referred_count.pdf}\vspace{-5pt} \\
\end{tabular}
\end{center}
\caption{
4 GAM plots on the Donors dataset before (orange) and after (blue) fine-tuning on the labeled samples. In (a, b) we show two features that the labeled information agrees with the notion of sparsity; thus, after fine-tuning the magnitude increases. In (c, d) the label information disagrees with the notion of sparsity; thus, the magnitude changes or decreases after the fine-tuning.
}
\label{fig:donors_ft}
\end{figure*}
\subsection{Qualitative analyses on GAM explanations}
\paragraph{Explaining anomalous data}
To let domain experts understand and debug why a sample is considered anomalous, we demonstrate explaining the most anomalous sample considered by DIAD on Mammography dataset.
The task is to detect breast cancer from radiological scans, specifically the presence of clusters of microcalcifications that appear bright on a mammogram.
The 11k images are segmented and preprocessed by vision pipelines and extracted 6 image-related features including the area of the cell, constrast, and noise etc.
In Fig.~\ref{fig:mammography_gam_plots}, we show the most anomalous data and see which feature contributes the most for sparsity (i.e. anomalous).
We illustrate why the model predicts this sample as anomalous; the unusually-high `Contrast' (Fig.~\ref{fig:mammography_gam_plots}(a)) of the image differs from other samples.
Also, the unusually high noise (Fig.~\ref{fig:mammography_gam_plots}(b)) and `Large area' (Fig.~\ref{fig:mammography_gam_plots}(c)) also makes it anomalous.
Finally, it is also considered quite anomalous in a 2-way interaction (Fig.~\ref{fig:mammography_gam_plots}(d)).
Since the sample has `middle area' and `middle gray level' which constitute a rare combination for the dataset.
\paragraph{Qualitative analyses on the impact of fine-tuning with labeled data}
In Fig.~\ref{fig:donors_ft}, we visualize how predictions change before and after fine-tuning with labeled samples on Donors dataset.
Donors dataset consists of 620k educational proposals for K12 level with 10 features, and the anomalies are defined as the top 5\% ranked proposals as outstanding.
Here, we show 4 GAM plots before and after fine-tuning.
Figs.~\ref{fig:donors_ft} a \& b show that both `Great Chat' and `Great Messages Proportion' increase its magnitude after fine-tuning, showing that the sparsity of these two features is consistent with the labels.
On the other hand, Figs.~\ref{fig:donors_ft} c \& d show that after fine-tuning, the model learns the opposite trend.
The sparsity definition treats values with less density as more anomalous -- in this case \textit{`Fully Funded'=0} is treated as more anomalous.
But in fact `Fully Funded' is a good indicator of outstanding proposals, so after fine-tuning model learns that \textit{`Fully Funded'=1} is in fact more anomalous.
This shows the importance of incorporating labeled data into the model to let the model correct its anomaly objective and learn what the intended anomalies are.
\section{Ablation and sensitivity analysis}
\label{sec:ablation}
To analyze the source of gains, we perform ablation studies with some variants of DIAD. The results are presented in Table~\ref{table:ablation}. First, we find fine-tuning with AUC is better than BCE.
Sparsity normalization plays an important role in fine-tuning, since sparsity could have values up to $10^4$ which negatively affect fine-tuning.
Upsampling the positive samples also contributes to performance improvements.
\setlength\tabcolsep{6pt}
\begin{table}[h!]
\centering
\vspace{-10pt}
\caption{Ablation study for semi-supervised AD. We test our method with fine-tuning only AUC vs. BCE loss. The performance benefits from more labels. Removing sparsity normalization substantially decreases the performance.}
\label{table:ablation}
\begin{tabular}{c|ccccc}
\toprule
\textbf{No. Anomalies} & \textbf{5} & \textbf{15} & \textbf{30} & \textbf{60} & \textbf{120} \\
\midrule
DIAD & \textbf{89.4} & \textbf{90.0} & \textbf{90.4} & \textbf{89.4} & \textbf{91.0} \\
Only AUC & 88.9 & 89.4 & 90.0 & 89.1 & 90.7 \\
Only BCE & 88.8 & 89.3 & 89.4 & 88.3 & 89.2 \\
\makecell{Unnormalized\\sparsity} & 84.1 & 85.6 & 85.7 & 84.2 & 85.6 \\
No upsampling & 88.6 & 89.1 & 89.4 & 88.5 & 90.1 \\
\bottomrule
\end{tabular}
\end{table}
\setlength\tabcolsep{5pt}
\begin{table}[h!]
\centering
\caption{Semi-supervised AD performance with 25\% of the original validation data (4\% of total data).
}
\label{table:ss_val_ratio_main}
\begin{tabular}{c|ccccc}
\toprule
& \multicolumn{5}{c}{25\% val data (4\% of total data)} \\
\midrule
\textbf{No. Anomalies} & \textbf{5} & \textbf{15} & \textbf{30} & \textbf{60} & \textbf{120} \\
\midrule
DIAD & \textbf{89.0} & \textbf{89.3} & \textbf{89.7} & \textbf{89.1} & \textbf{90.4} \\
\makecell{DIAD w/o PT}& 85.4 & 87.1 & 86.9 & 86.4 & 87.9 \\
CST & 83.9 & 84.9 & 85.7 & 85.6 & 88.2 \\
DevNet & 82.0 & 83.4 & 84.4 & 82.0 & 84.6 \\
\bottomrule
\end{tabular}
\end{table}
In practice we might not have a large (e.g. 16\% of the labeled dat) validation dataset, as in Sec.~\ref{sec:ss_result}, thus, it would be valuable to evaluate the performances of DIAD with a smaller validation dataset. In Table~\ref{table:ss_val_ratio_main}, we reduce the validation dataset size to only 4\% of the labeled data and find DIAD still consistently outperforms others. Additional results can be found in Supp.~\ref{appx:ss_less_val}.
We also perform a sensitivity analysis in Supp.~\ref{appx:sensitivity} that varies hyperparameters in the unsupervised AD benchmarks.
Our method performs quite stable in less than $2\%$ differences across a variety of hyperparameters.
\section{Discussions and Conclusions}
As all unsupervised AD methods rely on approximate objectives to discover anomalies such as reconstruction loss, predicting geometric transformations, or contrastive learning. The objectives inevitably would not align with labels on some datasets, as inferred from the performance ranking fluctuations across datasets.
This motivates for abilities to incorporate labeled data to boost performances and incorporate interpretability to find out whether the model could be trusted and whether the approximate objective aligns with the human-defined anomalies.
Beyond the inspirations from NodeGAM for the model architecture and PID loss for the objective, we introduce novel contributions that are key for highly accurate and interpretable AD: we modify the architecture by temperature annealing, introduce a novel way to estimate and normalize sparsity, propose a novel regularization for improved generalization, and introduce semi-supervised AD via supervised fine-tuning of the unsupervised learnt representations.
Our contributions play a crucial role to push both unsupervised and semi-supervised AD benchmarks.
Furthermore, our method provides interpretability which is crucial in high-stakes applications with needs of explanations such as finance or healthcare.
| 6272d3cc40693b9ce2ee6b71ab80a72650778ddb | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Colourings of simple $3$--dimensional hyperbolic polytopes with elements of $\mathbb{Z}^3_2$ were first used by Vesnin \cite{Ves87} in order to produce closed hyperbolic $3$--manifolds from the so--called L\"obell polyhedra.
Later on, Davis and Januszkiewicz \cite{DJ91} used a similar idea in order to produce complex $n$--manifolds with torus actions. The fixed point set of the conjugation involution of any such manifold produces a real $n$--manifold endowed with $\mathbb{Z}_2$--action \cite[Corollary 1.9]{DJ91}. The latter gives rise to a $\mathbb{Z}^n_2$--colouring that corresponds to a manifold known as a small cover. The technique of colourings of right--angled polyhedra was then generalised in the works \cite{KMT, KS?}, and the recent work of Choi and Park \cite{CP2} provides an efficient method to compute cohomology rings of manifolds obtained via colourings. Right--angled polyhedra attracted a lot of interest per se and as constituent pieces of hypebolic manifolds, cf. \cite{Atkinson, Dufour, Kolpakov, KM, PV}.
Colourings of polytopes provide an accessible way to build hyperbolic manifolds with prescribed topological characteristics. Cf. the recent surveys \cite{Martelli, Vesnin} for more information about the recent results in this area.
One of the classes of manifolds with relatively simple topological structure are homology spheres. We are mostly interested in \textit{rational homology spheres} (termed $QHS$ for short) and, moreover, in those that are hyperbolic manifolds (which we call hyperbolic $QHS$).
Having a hyperbolic structure, however, implies that their fundamental groups are highly non--trivial, and this leaves only a narrow escape for hyperbolic $QHS$ to exist. For example, there are no hyperbolic QHS in dimension $2$. The first hyperbolic $QHS$ in dimension $3$ was constructed by Seifert and Weber \cite{SW}. A $QHS$ produced by $\mathbb{Z}^3_2$--colouring of the right--angled hyperbolic dodecahedron can be found in the work of Garrison and Scott \cite{GS}. There is still no known example of a hyperbolic $QHS$ in dimensions $\geq 4$. Moreover, it is known that even such a large class as arithmetic hyperbolic manifolds cannot produce a QHS of even dimension $n > 4$ \cite{emery-qhs}.
Recently, Reid \cite{Reid-lecture} posed the question on whether there exists an arithmetic hyperbolic $QHS$ in dimension $n\geq 3$ that bounds geometrically. We can answer it in the affirmative for $3$--dimensional QHS.
\begin{thm}\label{theorem:qhs-bounds}
There is an arithmetic hyperbolic rational homology $3$--sphere that bounds geometrically.
\end{thm}
It appears that we cannot use colourings
to produce hyperbolic rational homology spheres in higher dimensions. Indeed, compact right--angled polytopes exist in $\mathbb{H}^n$ only for $n \leq 4$ \cite{PV}. The only compact $4$--dimensional right--angled polytopes known so far appear to be the $120$--cell and the associated ``garlands'' produced by ``gluing'' several copies of the $120$--cell along isometric facets. However,
none of such polytopes can produce a manifold with Euler characteristic $2$ via colouring due to the fact that the orbifold Euler characteristic of the right-angled $120$-cell is $\frac{17}{2}$, and that the orbifold Euler characteristic is multiplicative under covers.
We might be able, nonetheless, to use colourings of right-angled $n$-cubes to produce Euclidean rational homology spheres, which exist only in odd dimension. In dimension $3$, there is one colouring up to equivalence that produces the Hantzsche--Wendt manifold \cite[p.~8]{FKS}. In higher dimensions, Proposition \ref{prop:Euclidean QHS} gives a combinatiorial condition to determine whether a colouring of an $n$-cube is a rational homology sphere. In even dimension, rational homology spheres have Euler characteristic $2$ and cannot be Euclidean, since the latter implies Euler characteristic $0$ via the Gauss--Bonnet formula.
As for spherical rational homology spheres, the canonical colouring of the right-angled $n$-simplex that associates the $i$-th standard basis vector to its $i$-th facet, $1 \leq i \leq n+1$, produces the $n$-sphere in any dimension $n\geq 1$.
\subsection*{Outline of the paper}
In Section~\ref{sec:preliminaries} we provide all the necessary definition and basic facts about hyperbolic manifolds obtained via colourings of right--angled polyhedra, including results about their cohomology and coloured isometries. In Section~\ref{sec:extensions} we introduce the concept of colouring extensions and prove some general results related to it. In Section~\ref{sec:3-qhs} we provide an easy algorithmic criterion for a coloured compact right--angled hyperbolic $3$--polytope to produce a QHS. In Section~\ref{sec:fpf-isometries} we show how to promote geodesical embeddings to geometrical boundaries, and prove Theorem~\ref{theorem:qhs-bounds}.
\section{Preliminaries}\label{sec:preliminaries}
A finite-volume polytope $\mathcal{P} \subset \mathbb{X}^n$ (for $\mathbb{X}^n = \mathbb{S}^n, \mathbb{E}^n, \mathbb{H}^n$ being spherical, Euclidean and hyperbolic $n$-dimensional space, respectively, cf. \cite[Chapters 1--3]{Ratcliffe}) is called {\itshape right-angled} if any two codimension $1$ faces (or facets, for short) are either intersecting at a right angle or disjoint. It is known that compact hyperbolic right-angled polytopes cannot exist if $n > 4$ \cite{PV}. The only compact right-angled spherical and Euclidean polytopes are the $n$-simplex and the $n$-parallelotope, respectively. A sufficient condition for an abstract $3$-polytope to be realizable as right-angled hyperbolic one is given in \cite[Theorem 2.4]{Vesnin}. There is no such classification for right-angled $n$-polytopes with $n\geq 4$.
\subsection{Colourings of right-angled polytopes}
One of the important properties of hyperbolic right-angled polytopes is that their so-called colourings provide a rich class of hyperbolic manifolds. By inspecting the combinatorics of a colouring, one may obtain important topological and geometric information about the associated manifold.
\begin{defn}
Let $\mathcal{S}$ be an $n$-dimensional simplex with the set of vertices $\mathcal{V}=\{v_1,\dots,v_{n+1}\}$ and $W$ an $\mathbb{Z}_2$-vector space. A map $\lambda: \mathcal{V} \rightarrow W$ is called {\itshape a colouring} of $\mathcal{S}$. Such a colouring $\lambda$ of $\mathcal{S}$ is {\itshape proper} if the vectors $\lambda(v_i), \ i=1,\dots, n+1$ are linearly independent.
\end{defn}
\begin{defn}
Let $K$ be a simplicial complex with the set of vertices $\mathcal{V}$. Then {\itshape a colouring} of $K$ is a map $\lambda: \mathcal{V} \rightarrow W$. A colouring $\lambda$ of $K$ is {\itshape proper} if $\lambda$ is proper on each simplex of $K$.
\end{defn}
Notice that if $\mathcal{P} \subset \mathbb{X}^n$ is a compact right-angled polytope then $\mathcal{P}$ is necessarily simple and the boundary complex $K_\mathcal{P}$ of its dual $\mathcal{P}^*$ is a simplicial complex. Below we shall provide a more general definition that works also for the case of non-compact polytopes of finite volume.
\begin{defn}
Let $\mathcal{P} \subset \mathbb{X}^n$ be a polytope with the set of facets $\mathcal{F}$ and $K_\mathcal{P}$ be the maximal simplicial subcomplex of the boundary of $\mathcal{P}^*$. A \textit{colouring} of $\mathcal{P}$ is a map $\lambda: \mathcal{F} \rightarrow W$, where $W$ an $\mathbb{Z}_2$-vector space. This map naturally defines a colouring of $K_\mathcal{P}$. Then $\lambda$ is called \textit{proper} if the induced colouring on $K_\mathcal{P}$ is proper.
\end{defn}
If the polytope $\mathcal{P}$ or the vector space $W$ are clear from the context, then we will omit them and simply refer to $\lambda$ as a colouring. The \textit{rank} of $\lambda$ is the $\mathbb{Z}_2$-dimension of $\mathrm{im}\, \lambda$. We will always assume that colourings are surjective, in the sense that the image of the map $\lambda$ is a generating set of vectors for $W$.
A colouring of a right-angled $n$-polytope $\mathcal{P}$ naturally defines a homomorphism, which we still denote by $\lambda$ without much ambiguity, from the associated right-angled Coxeter group~$\Gamma(\mathcal{P})$, that is generated by reflections in all the facets of $\mathcal{P}$, into $W$ with its natural group structure. Being a Coxeter polytope, $\mathcal{P}$ has a natural orbifold structure as the quotient $\mathbb{X}^n /_{\Gamma(\mathcal{P)}}$.
\begin{prop}[\cite{DJ91}, Proposition 1.7]\label{DJ}
If the colouring $\lambda$ is proper, then $\ker \lambda < \Gamma(\mathcal{P})$ is torsion-free, and $\mathcal{M}_\lambda = \mathbb{X}^n /_{\ker \lambda}$ is a manifold without boundary. Moreover, if $\mathcal{P}$ is compact, $\mathcal{M}_\lambda$ is closed.
\end{prop}
If the dimension of the vector space $W$ is minimal, i.e. equal to the maximum number of vertices in the simplices of $\mathcal{P}^*$, $\mathcal{M}_\lambda$ is called \textit{a small cover} of $\mathcal{P}$. This is equivalent to $\text{rank} \, \lambda = n$.
\begin{rem}\label{rem:manifold cover tesselation}
If $\lambda:\Gamma \to \mathbb{Z}_2 ^m$ is surjective, $M_\lambda$ is homeomorphic to the topological manifold $(\mathcal{P} \times \mathbb{Z}_2 ^m)/_{\sim}$, where
\begin{equation}\label{colouring 2}
(p,g) \sim (q,h) \Leftrightarrow \left\{
\begin{array}{ll}
p=q \text{ and } g=h & \text{if } p\in \text{int}(\mathcal{P}) \\
p=q \text{ and } g - h \in W_f & \text{if } p \in \partial \mathcal{P},
\end{array}
\right.
\end{equation}
with $f= F_{i_1} \cap \ldots \cap F_{i_l}$ being the unique face that contains $p$ as its interior point and $W_f = \lambda(\mathrm{Stab}_\Gamma(f))$ is the subspace of $W = \mathbb{Z}_2 ^k$ generated by $\lambda(F_{i_1})$, \ldots, $\lambda(F_{i_l})$.
In other words, this means that $(\mathcal{P} \times \mathbb{Z}_2 ^k)/_{\sim}$ can be obtained from $\mathcal{P} \times \mathbb{Z}_2 ^k$ by identifying distinct copies $F\times \{g\}$ and $F\times \{h\}$ through the reflection in $F \in \mathcal{F}$ (which acts on the orbifold $\mathcal{P}$ as the identity) whenever $g-h =\lambda(F)$. Then it is clear that $\mathcal{M}_\lambda$ can be tessellated by $2^k$ copies of $\mathcal{P}$.
\end{rem}
\begin{defn}
Let $\mathrm{Sym}(\mathcal{P})$ be the group of symmetries of a polytope $\mathcal{P}\subset \mathbb{X}^n$, which acts on the set of facets $\mathcal{F}$ by permutations. Two $W$-colourings $\lambda,\mu:\mathcal{F}\to W$ of $\mathcal{P}$ are called \textit{equivalent in the sense of Davis and Januszkiewicz} or, simply, $DJ$-\textit{equivalent}, if there exists a combinatorial symmetry $s \in \mathrm{Sym}(\mathcal{P})$ and an invertible linear transformation $m \in \mathrm{GL}(W)$ such that $\lambda = m \circ \mu \circ s$.
\end{defn}
It is easy to see that $DJ$-equivalent proper colourings of a polytope $\mathcal{P} \subset \mathbb{X}^n$ define isometric manifolds. The converse is also known to be true for orientable manifold covers of the $3$-cube \cite[Proposition 3.2]{FKS} and orientable small covers of compact, hyperbolic $3$-polytopes \cite[Theorem 3.13]{Vesnin}. Therefore, unless explicitly said, we shall always consider colourings up to equivalence.
\begin{defn}
We say that a $\mathbb{Z}_2^k$-colouring $\lambda$ is \textit{orientable} if the orbifold $M_\lambda$ is orientable. Here, $\lambda$ is not necessarily proper.
\end{defn}
Equivalence gives us the following criterion for orientability.
\begin{prop}[\cite{KMT}, Lemma 2.4]\label{orientability}
The orbifold $M_\lambda$ is orientable if and only if $\lambda$ is equivalent to a colouring that assigns to each facet a colour in $W \cong \mathbb{Z}^k_2$ with an odd number of entries $1$.
\end{prop}
\subsection{Computing the homology of colourings.} Given a right-angled polytope $\mathcal{P} \subset \mathbb{X}^n$ with a $\mathbb{Z}^k_2$-colouring $\lambda$, let us enumerate the facets $\mathcal{F}$ of $\mathcal{P}$ in some order. Then we can assume that $\mathcal{F} = \{ 1, 2, \ldots, m \}$. Let $\Lambda$ be \textit{the defining matrix} of $\lambda$ that consists of the column vectors $\lambda(1), \ldots, \lambda(m)$ exactly in this order. Hence $\Lambda$ is a matrix with $k$ rows and $m$ columns. More precisely, $\Lambda$ represents the abelianization of $\lambda$, i.e. the former is a map such that $\Lambda \circ \mathrm{ab} = \lambda$, where $\mathrm{ab}:\Gamma \to \mathbb{Z}_2 ^m$ is the abelianization map that takes $r_i$, the reflection of the facet $i$, to $e_i$.
Let $\mathrm{Row}(\Lambda)$ denote the row space of $\Lambda$, while for a vector $\omega \in \mathrm{Row}(\Lambda)$ let $K_\omega$ be the simplicial subcomplex of the complex $K=K_\mathcal{P}$ spanned by the vertices $i$, also labelled by the elements of $\{1, 2, \ldots, m\}$, such that the $i$-th entry of $\omega$ equals $1$.
Then the rational cohomology of $\mathcal{M}_{\lambda}$ can be computed via the following formula, c.f. \cite[Theorem 1.1]{CP2}.
\begin{equation}\label{eq:cohomology}
H^p(\mathcal{M}_{\lambda},\mathbb{Q})\cong \underset{\omega \in \mathrm{Row}(\Lambda)}{\bigoplus}\widetilde{H}^{p-1}(K_{\omega},\mathbb{Q}).
\end{equation}
Moreover, the cup product structure is given by the maps \cite[Main Theorem]{CP2}:
\begin{equation}\label{eq:cup products}
\widetilde{H}^{p-1}(K_{\omega_1},\mathbb{Q}) \otimes \widetilde{H}^{q-1}(K_{\omega_2},\mathbb{Q}) \mapsto \widetilde{H}^{p+q-1}(K_{\omega_1 + \omega_2},\mathbb{Q}).
\end{equation}
One practical application of Equation \ref{eq:cohomology} is the computation of Betti numbers of colourings of Euclidean $n$-cubes, which turns out to be very simple.
\begin{cor}[\cite{Ferrari}, Proposition 5.1.11]\label{cor:Ferrari}
Let $\mathcal{C}$ be an $n$-cube with opposite facets in pairs labelled with the numbers $\{2i-1,2i\}$, for each $i \in [n]$, and $\lambda$ be any colouring of $\mathcal{C}$. Let $T_j\subset \mathbb{Z}_2^{2n}$, $0\le j \le n$, be the sets such that
\begin{align*}
T_1 =&\{e_1+e_2, \ldots, e_{2i-1}+e_{2i}, \ldots, e_{2n-1}+e_{2n}\} \\
T_j =&\big\{x_1+\ldots+x_j \,\big|\, \{x_1, \ldots, x_j\} \subset T_1\big\}.
\end{align*}
For any $j \in \{1, \ldots, n\}$, we have the following:
$$\beta_j ^\mathbb{Q} (\mathcal{M}_\lambda)=\big|\mathrm{Row} (\Lambda) \cap T_j\big|. $$
\end{cor}
We conclude this section by proving the following useful proposition.
\begin{cor}\label{prop:orientability_criterion}
Let $\mathcal{P} \subset \mathbb{X}^n$ be a right-angled polytope with a proper colouring $\lambda$. Then the manifold $\mathcal{M}_{\lambda}$ is orientable if and only if the vector $\varepsilon=(1,\dots,1)$ belongs to $\mathrm{Row}(\Lambda)$.
\end{cor}
\begin{proof}
By Proposition \ref{orientability}, if $\lambda$ is orientable then the sum of all lines of $\Lambda$ equals exactly the vector $\varepsilon$.
Conversely, if $\varepsilon \in \mathrm{Row}(\Lambda)$, then $K_{\varepsilon}=K\sim \mathbb{S}^{n-1}$ and $\widetilde{H}^{n-1}(K_{\varepsilon},\mathbb{Q})\cong \mathbb{Q}$ then we immediately have $H^n(\mathcal{M}_{\lambda},\mathbb{Q})\cong \mathbb{Q}$ by Equation \ref{eq:cohomology}. Since $\mathcal{M}_\lambda$ is a manifold without boundary, then $\mathcal{M}_\lambda$ is orientable.
\end{proof}
\section{Colouring Extensions}\label{sec:extensions}
\begin{defn}
Let $\lambda:\mathcal{F}\to \mathbb{Z}_2 ^k$ be any colouring. A colouring $\mu:\mathcal{F}\to \mathbb{Z}_2 ^{k+1}$ is called \textit{an extension of} $\lambda$ if there is a linear projection $p:\mathbb{Z}_2 ^{k+1}\to \mathbb{Z}_2 ^k$
such that $\lambda=p\circ\mu$.
\end{defn}
\begin{prop}\label{prop:double-cover colouring}
Let $\lambda:\mathcal{F}\to \mathbb{Z}_2 ^k$ be any colouring and $\mu:\mathcal{F}\to \mathbb{Z}_2 ^{k+1}$ its extension. Then $\mathcal{M}_\mu$ double-covers $\mathcal{M}_\lambda$. Moreover, if $\lambda$ is proper, so is $\mu$.
\end{prop}
\begin{proof}
Let $\Gamma=\Gamma(\mathcal{P})$ be the reflection group associated with $\mathcal{P}$. Let $\lambda:\Gamma \to \mathbb{Z}_2^k$ and $\mu:\Gamma \to \mathbb{Z}_2^{k+1}$ be the homomorphisms induced by $\lambda$ and $\mu$, respectively. By definition, we have that $\lambda = p \circ \mu$, and it follows that $\mathrm{ker} \, \lambda=\mathrm{ker} \, (p \circ \mu)=\mu^{-1}(\mathrm{ker}\, p)$. Moreover, $\mathrm{Im}\,p\cong \mathbb{Z}_2 ^k$ and $|\mathrm{ker} \, p|=[\mathbb{Z}_2 ^{k+1}:\mathrm{Im}\,p]=2$. Thus $\mathrm{ker} \, p=\{0,v_0\}$ for some $v_0\in \mathbb{Z}_2^{k+1}$, $v_0 \neq 0$.
Since $\mu$ is surjective, there exists $u_0\in \mu^{-1}(v_0)\neq \emptyset$. Since $\mu$ is a homomorphism, $\mu^{-1}(v_0)=u_0+ \mathrm{ker}\, \mu $. Then $\mathrm{ker} \, \lambda = \mathrm{ker} \, \mu \sqcup (g_0+\mathrm{ker}\,\mu)$, and thus $\mathrm{ker}\,\mu \triangleleft_{2} \mathrm{ker}\, \lambda$. Hence $\mathcal{M}_\mu$ is a double cover of $\mathcal{M}_\lambda$.
Finally, assume that $\{\lambda(F_1), \ldots, \lambda(F_s)\} \subset \mathbb{Z}_2 ^k$ is a set of linearly independent colours. By using the fact that $\lambda = p \circ \mu$, we easily obtain that $\mu(F_1), \ldots, \mu(F_s)$ are linearly independent. Hence, if $\lambda$ is proper then $\mu$ is proper too.
\end{proof}
One direct application of Equation \eqref{eq:cohomology} to extensions of colourings is the following.
\begin{prop}\label{prop:double-cover properties}
Let $\lambda:\Gamma \to \mathbb{Z}_2 ^k$ be a colouring and $\mu$ its extension. Let also $\Lambda$ and $\mathrm{M}$ be their respective defining matrices. Then, up to equivalence, $\mathrm{M}$ is the matrix obtained from $\Lambda$ by adding an extra row vector $v\in \mathbb{Z}_2 ^m =\mathrm{ab}(\Gamma)$, such that $v\notin \mathrm{Row}(\Lambda)$. Moreover, if $\lambda$ is orientable, so is $\mu$. Finally, for all $p\ge 0$,
\begin{equation*}
H^p(\mathcal{M}_\mu,\mathbb{Q})=H^p(\mathcal{M}_\lambda,\mathbb{Q})\oplus \underset{\omega \in \mathrm{Row}(\Lambda)}{\bigoplus}\widetilde{H}^{p-1}(K_{\omega+v},\mathbb{Q}).
\end{equation*}
\end{prop}
\begin{proof}
Up to equivalence, the projection $p:\mathbb{Z}_2 ^{k+1} \to \mathbb{Z}_2 ^k$ is just the canonical projection onto the first $k$ coordinates, given that, otherwise, we can just pre- and post-compose it with a suitable maps in $GL_{k+1}(\mathbb{Z}_2)$ and $GL_k(\mathbb{Z}_2)$, respectively. Then, since $p\circ \mathrm{M} = \Lambda$, it is clear that $\mathrm{M}$ is the matrix $\Lambda$ with another row $v\in\mathbb{Z}_2^m$ added. Moreover, $\mu$ is surjective if and only if $\mathrm{M}$ is surjective, and the latter holds if and only if $v\notin \mathrm{Row}(\Lambda)$. The colouring extensions can be seen in red in the diagram below:
\vspace{0.25in}
\[
\begin{tikzcd}
\Gamma \arrow[overlay, r, red, "\mu"] \arrow[overlay, d, "ab"] \arrow[overlay, dr, out=90, in=0, looseness=2.5, "\lambda"]
& \mathbb{Z}_2 ^{k+1} \arrow[overlay, d, "p"] \\
\mathbb{Z}_2 ^m \arrow[overlay, r, "\Lambda"] \arrow[overlay, ur, red, crossing over, "M"]
& \mathbb{Z}_2 ^k
\end{tikzcd}
\]
Clearly, $\mathrm{Row}(\mathrm{M})=\mathrm{Row}(\Lambda)\sqcup\big(v+\mathrm{Row}(\Lambda)\big)$. If $\lambda$ is orientable, by Corollary \ref{prop:orientability_criterion}, $\epsilon \in \mathrm{Row}(\Lambda) \subset \mathrm{Row}(\mathrm{M})$, that is, that $\mu$ is orientable. We conclude by applying Equation \eqref{eq:cohomology}.
\end{proof}
In the case of non-orientable colourings, we can define a specific type of extension.
\begin{defn}
Let $\lambda:\mathcal{F} \to \mathbb{Z}_2 ^k$ be a colouring. The \textit{orientable extension} $\delta:\mathcal{F}\to\mathbb{Z}_2^{k+1}$ of $\lambda$ is given by $$\delta(F)=\lambda(F)+\Big(1+\sum_{i=1}^k \lambda(F)_i\Big)e_{k+1}.$$
\end{defn}
\begin{prop}\label{prop:orientable double covers}
Let $\lambda:\mathcal{F} \to \mathbb{Z}_2 ^k$ be a non-orientable colouring and $\delta$ its orientable extension. Then $\mathcal{M}_\delta$ is the orientable double cover of $\mathcal{M}_\lambda$.
\end{prop}
\begin{proof}
Up to equivalence, any basis of colours $\{\lambda(F_1),...,\lambda(F_k)\}$ in $\mathbb{Z}_2 ^k$ can be chosen to be the canonical vectors, and we have that $\delta(F_i)=\lambda(F_i)=e_i$. Then it is enough to take any colour $\lambda(F_0)$ with an even number of entries $1$, which exists since $\lambda$ is non-orientable, and we have that $\delta(F_0)_{k+1} = 1$. Thus $\{\delta(F_0),...,\delta(F_k)\}$ is a basis and $\delta$ is surjective.
Now, let $p:\mathbb{Z}_2 ^{k+1}\to \mathbb{Z}_2 ^k$ be the canonical projection onto the first $k$ coordinates: we have that $\lambda=p\circ \delta$. Then Proposition \ref{prop:double-cover colouring} implies that $\mathcal{M}_\delta$ is a double cover of $\mathcal{M}_\lambda$. Moreover, $\mathcal{M}_\delta$ is orientable by Proposition \ref{orientability}.
\end{proof}
Conversely, there is a criterion to tell whether a given colouring $\mu$ is an extension of some other colouring $\lambda$.
\begin{prop}\label{prop:stabilizer condition}
Let $\mu:\Gamma(\mathcal{P})\to W$ be a colouring, and let $W_p = \mu\big(\mathrm{Stab}_\Gamma(p)\big)$ for any vertex $p$ of $P$. Then $\mu$ is an extension of some proper colouring if and only if $\bigcup_{p} W_p \subsetneq W$.
\end{prop}
\begin{proof}
Assume that there is a projection $p:W\cong\mathbb{Z} _2^k\to\mathbb{Z}_2 ^{k-1}$ such that $p\circ \mu$ is a proper colouring. Then, for any codimension $s$ face $f=F_1\cap...\cap F_s$ of $\mathcal{P}$ we have $(p\circ\mu)(F_1)+...+(p\circ\mu)(F_s)\neq 0$. This means that $\mu(F_1)+...+\mu(F_s)\notin \mathrm{ker}\, p$. As in the proof of Proposition~\ref{prop:double-cover colouring}, we have that $\mathrm{ker}\, p=\{0,v_0\}$ for some $v_0\in W$ and, in particular, $\mu(F_1)+...+\mu(F_s)\neq v_0$ for any face $f=F_1\cap...\cap F_s$. It follows that $v_0\notin W_f$ for any such face $f$ and, in particular, for any vertex $q\in \mathcal{P}$.
Conversely, assume there is a vector $v_0\in W\setminus \bigcup_{q} W_q$. Then $\mu(F_1)+ \ldots +\mu(F_s) \neq v_0$ for any codimension $s$ face $f = F_1\cap \ldots \cap F_s$ of $\mathcal{P}$. Let $W\cong \mathbb{Z}_2 ^k$ and $p:\mathbb{Z}_2 ^k \to \mathbb{Z}_2 ^{k-1}$ be the
projection along $v_0$. Let us then set $\lambda = p\circ \mu$. This is a proper colouring since $\mu(F_1)+ \ldots +\mu(F_s)\neq v_0$ implies that $\lambda(F_1)+ \ldots +\lambda(F_s)\neq p(v_0) = 0$ for all faces $f = F_1\cap \ldots \cap F_s$. By definition, $\mu$ is an extension of $\lambda$.
\end{proof}
\begin{example}[The Hantzsche--Wendt colouring]
Let $\lambda$ be the colouring of the $3$-cube defined in \cite[p.~8]{FKS} such that $\mathcal{M}_\lambda$ is the Hantzsche--Wendt manifold \cite{HW}. In particular, $\mathrm{rank}\,\lambda = 4$. However, we have that $\bigcup_{p} W_q = W$, and it follows from Proposition \ref{prop:stabilizer condition} that $\lambda$ is not an extension of any colouring.
\end{example}
\section{Rational Homology 3-Spheres}\label{sec:3-qhs}
A manifold $M$ is called a \textit{homology $n$-sphere},or $HS^n$, if $H_i(M, \mathbb{Z}) = H_i(\mathbb{S}^n, \mathbb{Z})$, for all $i = 1, \ldots, n$. Similarly, $M$ is called a \textit{rational homology $n$-sphere}, or $QHS^n$, if $H_i(M, \mathbb{Q}) = H_i(\mathbb{S}^n, \mathbb{Q})$, for all $i = 1, \ldots, n$. Clearly, any $HS$-- or $QHS$--manifold is orientable. Additionally, we say a $CW$-complex is a \textit{(rational) homology point}, or $(Q)HP$, if all its reduced $\mathbb{Z}$-- ($\mathbb{Q}$--) homology groups are trivial.
By applying Equation \ref{eq:cohomology}, we have the following.
\begin{lemma}\label{lemma:subcomplexes are HP}
An orientable $\mathcal{M}_{\lambda}$ is a $QHS^n$ if and only if for all $\omega \in \mathrm{Row}(\Lambda) \setminus \{0, \varepsilon\}$, $K_\omega$ is an $QHP$.
\end{lemma}
\begin{proof}
The only non-trivial cohomology groups of $\mathcal{M}_{\lambda}$ are $H_n(\mathcal{M}_{\lambda},\mathbb{Q})\cong H^0(\mathcal{M}_{\lambda},\mathbb{Q})\cong \widetilde{H}^{-1}(K_0,\mathbb{Q})\cong \mathbb{Q}$ and $H_0(\mathcal{M}_{\lambda},\mathbb{Q})\cong H^n(\mathcal{M}_{\lambda},\mathbb{Q})\cong \widetilde{H}^{n-1}(K_\varepsilon,\mathbb{Q})\cong \mathbb{Q}$. Therefore, every other simplicial subcomplex $K_\omega$ must have trivial reduced homology groups.
\end{proof}
In the case of colourings of $n$-cubes, by applying Corollary \ref{cor:Ferrari}, we have the following.
\begin{prop}\label{prop:Euclidean QHS}
Let $\mathcal{C}$ be an $n$-cube with its sides labelled as in Corollary \ref{cor:Ferrari}, and $T_i$ be its corresponding $T$-sets. Let $\mathcal{T}=\bigcup_{i<n} T_i$. Then we have that $\mathcal{M}_\lambda$ is a $QHS$ if and only if $\mathrm{Row}\Lambda \cap \mathcal{T}=\emptyset$.
\end{prop}
By applying Equation \ref{eq:cup products}, we get a useful consequence.
\begin{lemma}\label{lemma:HP condition}
Let $\mathcal{M}_{\lambda}$ be an orientable colouring and $\omega \in \mathrm{Row}(\Lambda)\setminus\{0,\epsilon\}$. Then $K_\omega$ is an $QHP$ if and only if $K_{\epsilon-\omega}$ is an $QHP$.
\end{lemma}
\begin{proof}
Assume that $K_\omega$ is not an $QHP$. Then $\tilde{H}^*(K_\omega,\mathbb{Q})$ has non-trivial $0$-- or $1$--cohomology. First, let $0\neq \alpha \in \widetilde{H}^0(K_{\omega},\mathbb{Q})$. By Equation \eqref{eq:cohomology}, $\alpha \in H^1(\mathcal{M}_{\lambda},\mathbb{Q})$. By \cite[Corollary 3.39]{Hatcher}, there exists $\beta \in H^2(\mathcal{M}_{\lambda},\mathbb{Q})$ such that $\alpha\smile \beta$ is the generator of $H^3(\mathcal{M}_{\lambda},\mathbb{Q})$. Then, by Equation \eqref{eq:cohomology}, $\alpha \smile \beta$ is the generator of $\widetilde{H}^2(K_\epsilon, \mathbb{Q})$, since $K_\epsilon$ is homotopically $\mathbb{S}^2$. Finally, by Equations \eqref{eq:cohomology} and \eqref{eq:cup products}, we obtain that $0\neq \beta \in \widetilde{H}^1(K_{\epsilon-\omega},\mathbb{Q})$, since otherwise the product $\alpha \smile \beta$ would not belong to $\widetilde{H}^2(K_\epsilon, \mathbb{Q})$. The case of $\widetilde{H}^1(K_{\omega},\mathbb{Q})\neq 0$ is completely analogous.
\end{proof}
Thus, we can improve Lemma \ref{lemma:subcomplexes are HP} algorithmically by checking only the connectivity of some graphs.
\begin{cor}\label{cor:algorithm}
An orientable $\mathcal{M}_{\lambda}$ is a $QHS^n$ if and only if for all $\omega \in \mathrm{Row}(\Lambda) \setminus \{0, \varepsilon\}$ the $1$-skeleton of $K_\omega$ is connected.
\end{cor}
\begin{proof}
If one proper subcomplex $K_\omega$ has a non-trivial cycle then, by Lemma \ref{lemma:HP condition}, the complementary complex $K_{\varepsilon-\omega}$ will be disconnected. It suffices therefore to check if all proper, non-empty subcomplexes $K_\omega$ are connected. Clearly the connectivity of $K_\omega$ depends only on the connectivity of its $1$-skeleton.
\end{proof}
In the case of double covers, the transfer homomorphisms \cite[Section 3.G]{Hatcher} can be used in order to obtain the following statement.
\begin{lemma}\label{lemma:QHS double-covers}
Let $Y$ be a closed manifold that is a double cover of another manifold $X$. If $Y$ is a $QHS$, then $X$ is either a $QHS$ or homologically trivial.
\end{lemma}
\begin{proof}
Let $\varphi$ be the involution acting on $Y$ such that $X\cong \faktor{Y}{G}$, with $G=\langle \varphi \rangle\cong \mathbb{Z}_2$. By \cite[Section 3.G]{Hatcher}, we have that $H^*(X,R)\cong H^*(Y,R)^G$ for any ring $R\ni |G|^{-1}$, where $H^*(Y,R)^G$ is the subalgebra of the algebra $H^*(Y,R)$ fixed by the induced action of $G$.
Then, we take $R=\mathbb{Q}\ni \frac{1}{2}$. If $Y$ is a $QHS$, we have that all its rational cohomology classes are trivial, except for $H^0(Y,\mathbb{Q})\cong H^n(Y,\mathbb{Q})\cong \mathbb{Q}$, where $n$ is the dimension of $Y$. Since $H^*(X,\mathbb{Q})$ is isomorphic to a
subalgebra of $H^*(Y,\mathbb{Q})$, the non-trivial cohomology classes of the former are at most those of the latter. Moreover, $H^0(X,\mathbb{Q})\cong\mathbb{Q}$, since $X$ is connected. If $X$ is orientable then $H^n(X,\mathbb{Q})\cong\mathbb{Q}$, and it follows that $X$ is a $QHS$. If $X$ is non-orientable then $H^n(X,\mathbb{Q})=0$, and it follows that $X$ is homologically trivial.
\end{proof}
\begin{rem}
The action of $G$ on the algebra can be seen directly by the action of $\varphi_*$ on the homology classes. Since $\varphi$ takes the connected $Y$ to its only connected component, $\varphi_*$ fixes its $0$-homology class. The action of $\varphi_*$ on the fundamental class $[Y]$, instead, is multiplication either by $+1$ or by $-1$, depending on whether $\varphi$ preserves or changes the orientation. Given that $\varphi^* (\alpha)=\alpha\circ\varphi_*$ for any cohomology element $\alpha$, we conclude that the action of $\varphi^*$ on the ring of cohomology is $\varphi^*(1+[Y]^*)=1\pm [Y]^*$, where the sign depends on whether $\varphi$ preserves orientation. If $\varphi$ is orientation-preserving then $X$ is orientable and, indeed, the whole ring is fixed under the action of $G$.
\end{rem}
Thus we can use the following algorithm in order to determine whether an extension of a colouring is a $QHS$.
\begin{lemma}\label{lemma:QHS extension algorithm}
Let $\lambda$ be a proper colouring such that $\mathcal{M}_\lambda$ is a $QHS^3$. Let $\mu$ be any extension of $\lambda$, obtained by adding a row vector $v$ to $\Lambda$. Then $\mathcal{M}_\mu$ is a $QHS^3$ if and only if for every pair $\{\omega,\epsilon-\omega\} \subset \mathrm{Row}(\Lambda)$, we have that $K_{\omega+v}$ is connected and has only trivial homology $1$--cycles.
\end{lemma}
\begin{proof}
By Proposition \ref{prop:double-cover properties}, we have that $\mathcal{M}_\mu$ is a $QHS$ if and only if $K_{\omega+v}$ is homologically trivial for every $\omega \in \mathrm{Row}(\Lambda)$. By Lemma \ref{lemma:HP condition}, $K_{\omega+v}$ is homologically trivial if and only if $K_{\epsilon-(\omega+v)}$ is so. Since $\epsilon-(\omega+v)=(\epsilon-\omega)+v$ and, by Proposition~\ref{prop:orientability_criterion}, $\epsilon \in \mathrm{Row}(\Lambda)$, we have that also $\epsilon-\omega \in \mathrm{Row}(\Lambda)$. Therefore, it is enough to check whether $K_{\omega+v}$ is homologically trivial for each pair $\{\omega,\epsilon-\omega\} \subset \mathrm{Row}(\Lambda)$. Since $K$ is homeomorphic to $\mathbb{S}^2$ and $K_{\omega + v}$ is a proper subcomplex of $K$, then $K_{\omega + v}$ is homologically trivial if and only if it is connected and has only trivial homology $1$--cycles.
\end{proof}
\section{Geodesic Embeddings and Geometric Boundaries}\label{sec:fpf-isometries}
First, we describe a way of promoting geodesic embeddings to geometric boundaries.
\begin{lemma}\label{lemma-embed-bound}
Let $M$ be an orientable hyperbolic $n$-manifold that has a fixed point free involution $\varphi \in \mathrm{Isom}(M)$. If $M$ embeds geodesically then it also bounds geometrically.
\end{lemma}
\begin{proof}
Let $M$ embed into an orientable manifold $N'$ as a totally geodesic submanifold of codimension $1$. Let us consider $N$ obtained from $N'$ by cutting it along $M$ and taking a connected component. Then either $\partial N = M$ and we are done, or $\partial N = M \sqcup M$, and we can quotient out one copy of $M$ in $\partial N$ by self-identifying it via $\varphi$. Given that $\varphi$ is a fixed point free involution, the resulting metric space $N_\varphi$ will be a hyperbolic manifold with a single boundary component isometric to $M$. Moreover, $N_\varphi$ is orientable or not depending on whether $\varphi$ is orientation-reversing or not.
\end{proof}
\begin{prop}\label{prop:dodecahedron colourings that bound}
Let $\lambda:\Gamma(\mathcal{D})\to W$ be a proper colouring of the right-angled dodecahedron. If $\bigcup_q W_q \subsetneq W$, then $M_\lambda$ bounds geometrically. Equivalently, any extension of a proper colouring of $\mathcal{D}$ bounds geometrically.
\end{prop}
\begin{proof}
By \cite[Theorem 2]{Martelli 2}, $\mathcal{M}_\lambda$ embeds (c.f. also \cite{KRS}). If $\bigcup_q W_q \subsetneq W$ then, by Proposition \ref{prop:stabilizer condition}, $\lambda$ is an extension of some colouring $\mu$ and, by Proposition~\ref{prop:double-cover colouring}, we have that $\mathcal{M}_\lambda$ double-covers $\mathcal{M}_\mu$. Therefore, there exists a fixed point free involution $\varphi \in \mathrm{Isom}(\mathcal{M}_\lambda)$ such that $\faktor{\mathcal{M}_\lambda}{\langle \varphi \rangle}\cong \mathcal{M}_\mu$. Then $\mathcal{M}_\lambda$ bounds geometrically by Lemma \ref{lemma-embed-bound}.
\end{proof}
\begin{thm}\label{thm:arithmetic-QHS-that-bounds}
There is an arithmetic hyperbolic $QHS^3$ that bounds geometrically.
\end{thm}
\begin{proof}
Take the only orientable small cover $\lambda$ of $\mathcal{D}$ given in \cite[p. 6]{GS}. Thus
\begin{equation*}
\Lambda=\begin{pmatrix}
1 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 0 & 0 \\
0 & 1 & 0 & 0 & 1 & 1 & 0 & 1 & 1 & 0 & 0 & 1 \\
0 & 0 & 1 & 1 & 0 & 1 & 0 & 1 & 1 & 0 & 1 & 0 \\
\end{pmatrix},
\end{equation*}
where the labeling of the faces of $\mathcal{D}$ is given in Figure \ref{im:dodecahedron}. By Equation \eqref{eq:cohomology}, $\mathcal{M}_\lambda$ is a $QHS$. If we find an extension $\mu$ of $\lambda$ such that $\mathcal{M}_\mu$ is a $QHS$, then we are done by Proposition \ref{prop:dodecahedron colourings that bound}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.42]{dodecahedron_extension.pdf}
\caption{The dodecahedron used in the proof of Theorem~\ref{thm:arithmetic-QHS-that-bounds} with its face labelling. The red, green, blue and yellow subcomplexes are $K_{13}$, $K_{14}$, $K_{34}$ and $K_v$, respectively.}
\label{im:dodecahedron}
\end{figure}
By Lemma \ref{lemma:QHS extension algorithm}, it is enough to find a row vector $v\in \mathbb{Z}_2 ^{12}$ such that the complexes $K_{\omega + v}$ are connected and without non-trivial cycles for $\omega \in \{0,(1,0,0)^T \Lambda,(1,0,1)^T \Lambda,(0,0,1)^T \Lambda\}$.
Recall that $\mathrm{Row}(\Lambda)=\{x^T \Lambda\mid x \in \mathbb{Z}_2 ^3\}$ and $\epsilon=(1,1,1)^T\Lambda$. Let $\omega= x^T \Lambda$. Then for a face $F^*$ of $\mathcal{D}$ we have that $F^* \in K_{\omega}$ if and only if $x \cdot \lambda(F) = 1$. Hence the complexes $K_{ij}=K_{(e_i+e_j)^T\Lambda}$ for $\{i,j\}\in\{1,2,3\}$ are precisely the subcomplexes of $K$ with vertices coloured by $e_i$ and $e_j$, while the subcomplexes $K_{i4}=K_{(e_i)^T\Lambda}$ are the subcomplexes coloured by $e_i$ and $e_1+e_2+e_3$.
Due to the constraint that $K_v$ be homologically trivial, the choice of vertices $(F_i)^*$ in $K$ such that $\mu(F_i)_4 = 1$ (or, equivalently, the choice of $v_i \neq 0$) should define a homologically trivial subcomplex. By choosing $K_v$ as the simplex $\{3,7,9\}$ around which the three complexes $K_{13}$, $K_{14}$, $K_{13}$ are ``wrapped'', we have precisely that all four subcomplexes $K_{v+\omega}$ are homologically trivial as shown in Figure~\ref{im:dodecahedron}.
Explicitly, the colouring $\mu$ with defining matrix
\begin{equation*}
\mathrm{M} = \begin{pmatrix}
1 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 0 & 0 \\
0 & 1 & 0 & 0 & 1 & 1 & 0 & 1 & 1 & 0 & 0 & 1 \\
0 & 0 & 1 & 1 & 0 & 1 & 0 & 1 & 1 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 0 & 0 \\
\end{pmatrix}
\end{equation*}
is an extension of $\Lambda$ by Proposition \ref{prop:double-cover properties}, and $\mathcal{M}_\mu$ is a $QHS$ by Equation \eqref{eq:cohomology}.
Hence, by Proposition \ref{prop:dodecahedron colourings that bound}, $\mathcal{M}_\mu$ bounds geometrically. Finally, since $\mathcal{D}$ is arithmetic \cite[Lemma 3.8]{Vesnin}, it follows that $\mathcal{M}_\mu$ is also arithmetic.
\end{proof}
\begin{rem}
A computer search among all possible extensions of $\Lambda$ returned that there are, up to $DJ$--equivalence, $7$ extensions which are $QHS^3$. All of them have trivial coloured symmetry group \cite[Section 2.3]{KS?}. However, the number of equivalence classes up to isometry might be smaller, given that the exact equivalence between isometry classes and colouring classes of compact hyperbolic $3$-polytopes holds only for small covers \cite[Theorem 3.13]{Vesnin}.
\end{rem}
\vspace{-0.1in}
| 903c34034c4ed485c83fc7ab4e9b81b2bbec37dc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
One of the central objects studied in general relativity are isolated gravitational systems such as stars, black holes and galaxies. Mathematically, they are modeled by asymptotically flat initial data sets (IDS) which are triples $(M,g,k)$ consisting of an asymptotically flat, complete, smooth Riemannian 3-manifold $(M,g)$ together with a smooth, symmetric two-tensor $k$.
\
More precisely, $(M,g)$ contains a compact set $\mathcal{C}\subset M$ such that we can write $M\setminus \mathcal{C}=\cup_{\ell=1}^{\ell_0}M_{end}^{\ell}$ where the ends $M_{end}^\ell$ are pairwise disjoint and diffeomorphic to the complement of a ball $\mathbb{R}^3 \setminus B_1$.
Furthermore, there exists a coordinate system in each end satisfying
\begin{equation}\label{asymflat}
|\partial^l (g_{ij}-\delta_{ij})(x)|=O(|x|^{-\tau-l}),\quad l=0,1,2,\quad\quad
|\partial^l k_{ij}(x)|=O(|x|^{-\tau-1-l}),\quad l=0,1,
\end{equation}
for some $\tau>\tfrac{1}{2}$.
To each initial data set $(M,g,k)$ we associate the energy density $\mu$ and the momentum density $J$ defined by
\begin{align}
\mu=\frac12(R+(\text{Tr}_g k)^2-|k|^2),\quad\quad J=\div_g(k-\text{Tr}_gkg)
\end{align}
where $R$ is the scalar curvature of $g$.
Moreover, we define the ADM energy $E$ and linear momentum $P$ by
\begin{equation}\label{EP definition}
E=\lim_{r\rightarrow\infty}\frac{1}{16\pi}\int_{S_{r}}\sum_i \left(g_{ij,i}-g_{ii,j}\right)\upsilon^j dA,\quad\quad
P_i=\lim_{r\rightarrow\infty}\frac{1}{8\pi}\int_{S_{r}} \left(k_{ij}-(\mathrm{Tr}_g k)g_{ij}\right)\upsilon^j dA
\end{equation}
where $\upsilon$ is the outer unit normal to the sphere $S_r$ and $dA$ is its area element.
In order to ensure that $E$ and $P$ are well-defined in equation \eqref{EP definition}, we impose additionally $\mu,J\in L^1(M)$, and throughout this paper we assume that $g\in C^{2,\alpha}(M)$ and $k\in C^{1,\alpha}(M)$.
\
A fundamental results about initial data sets is the positive mass theorem (PMT):
\begin{theorem}
Suppose $(M,g,k)$ is a complete asymptotically flat initial data set satisfying the dominant energy condition (DEC) $\mu\ge|J|$.
Then $E\ge|P|$.
\end{theorem}
This result has been first established by Schoen-Yau in \cite{SY2} using the Jang equation and by Witten in \cite{Witten} using spinors.
Further proofs have been given by \cite{Eichmair, EHLS, HKK}, and the important special case $k=0$ has been treated in \cite{SY1, HuiskenIlmanen, Li, BKKS, AMO}. We refer to \cite{HKK} for a more detailed historical overview.
\
It has been conjectured that if $E=|P|$, the IDS embeds isometrically in Minkowski spacetime with second fundamental form $k$.
This has been already confirmed under additional decay assumptions on $g$ and $k$ by Beig-Chrusciel and Huang-Lee in \cite{BeigChrusciel, HuangLee}.
More precisely, Beig-Chrusciel assume additionally $g_{ij}-\delta_{ij}\in C^{3,\alpha}_{-\tau}(M)$, $k_{ij}\in C^{2,\alpha}_{-\tau-1}(M)$ and $\mu,J\in C^{1,\alpha}_{-3-\epsilon}(M)$ for some constants $\tau>\frac12$, $\epsilon>0$ and $0<\alpha <1$.
Huang-Lee assume additionally $\mu,J\in C^{0,\alpha}_{-3-\epsilon}(M)$ and $\text{Tr}_gk\in C^0_{-2-\epsilon}(M)$.
As observed in \cite{HKK}, the decay condition $\text{Tr}_gk\in C^0_{-2-\epsilon}(M)$ can be omitted by combining \cite{HuangLee} with \cite{HKK}.
However, the general case is still an open question and is for instance listed as conjecture in \cite{Lee}, page 226.
\
Furthermore, we would like to point on that \cite{HuangLee} and \cite{CM} addressed the rigidity conjecture in higher dimension under certain additional assumptions.
However, the situation becomes more subtle, see for instance the counter example constructed in \cite{HuangLee}.
Finally, we would also like to point out the paper \cite{HuangJangMartin} on the rigidity of asymptotically hyperbolic manifolds.
\
In this manuscript we establish the following result which removes the additional decay assumptions required in previous papers:
\begin{theorem}\label{Theorem: main}
Let $(M,g,k)$ be a complete asymptotically flat initial data set satisfying the dominant energy condition $\mu\ge|J|$.
Moreover, suppose that $E=|P|$.
Then $E=|P|=0$ and $(M,g,k)$ arises as spacelike slice of Minkowski spacetime $\R^{3,1}$.
\end{theorem}
Our theorem is optimal in the sense that we merely need to assume $\mu,J\in L^1(M)$ which is required to ensure that $E$ and $P$ are finite and independent of the coordinate system used.
\
Our proof is short, elementary and relies on two ingredients:
First, we use the integral formula for spacetime harmonic functions $u$ established in \cite{HKK}.
Using the integral formula, we deduce that $E=|P|$ implies that all level-sets of $u$ have vanishing Gaussian curvature.
Second, we employ the fundamental theorem of surfaces which states that if $(M,g,k)$ satisfies the Gauss and Codazzi equations, $(M,g,k)$ embeds isometrically into Minkowski spacetime.
Combining the flatness of the level-sets with Liouville's theorem, we verify that the Gauss and Codazzi equations are indeed satisfied.
We expect that this method can also be applied in other settings such as asymptotically hyperbolic manifolds.
\
\textbf{Acknowledgements.}
The authors would like to thank Hubert Bray, Demetre Kazaras, Marcus Khuri and Dan Lee for stimulating discussions.
\section{Preliminaries}\label{S:Preliminaries}
There are several tools available to study IDS such as the Jang equation \cite{SY2}, spinors \cite{Witten} and marginally outer trapped surfaces \cite{EHLS}.
In \cite{HKK} a new method to study IDS has been introduced: \emph{spacetime harmonic functions}.
The main result of \cite{HKK} states the following:
\begin{theorem}
There exists an asymptotically linear, spacetime harmonic function $u\in C^{2,\alpha}(M_{ext})$, i.e., a function $u$ solving the differential equation $\Delta u=-\text{Tr}_gk|\nabla u|$, such that
\begin{align}\label{integral formula}
E-|P|\ge\frac1{16\pi}\int_{M_{ext}}\left( \frac{|\nabla^2u+k|\nabla u||^2}{|\nabla u|}+2\mu|\nabla u|+2\langle J,\nabla u\rangle\right).
\end{align}
\end{theorem}
We refer to \cite{HKK} for a discussion of the exterior region $M_{ext}$ and to \cite{BHKKZ} for a detailed motivation of spacetime harmonic functions.
The above theorem yields directly:
\begin{corollary}\label{cor:main}
Let $(M,g,k)$ be an asymptotically flat initial data set satisfying the dominant energy condition $\mu\ge|J|$ and suppose $E=|P|$. Then $M=M_{ext}=\R^3$ and there exists an asymptotically linear spacetime harmonic function $u\in C^{3,\alpha}(M)$ satisfying
\begin{align}
\nabla^2u=&-k|\nabla u|,\\
\mu|\nabla u|=&-\langle J,\nabla u\rangle.
\end{align}
Moreover, $|\nabla u|\ne0$ and the level sets $\Sigma_t=\{u=t\}$ are flat with second fundamental form $h=-k|_{T\Sigma_t}$.
\end{corollary}
\begin{proof}
The identities for $\nabla^2u=-k|\nabla u|$ and $\mu|\nabla u|=-\langle J,\nabla u\rangle$ follow immediately from the integral formula \eqref{integral formula}.
This also implies $h=-k|_{T\Sigma_t}$.
Lemma 7.1 and Proposition 7.2 in \cite{HKK} established that $|\nabla u|\ne0$ and $M=M_{ext}=\R^3$.
The claim $u\in C^{3,\alpha}(M)$ follows immediately from Schauder estimates in combination with the non-vanishing of $|\nabla u|$.
Finally, the claim that the level sets have vanishing Gaussian curvature is implied from the following computation, also see \cite{BHKKZ2}.
Since $\mu|\nabla u|=-\langle J,\nabla u\rangle$, the Gaussian equations yield
\begin{align}
\Delta|\nabla u|=\frac1{|\nabla u|}(-K|\nabla u|+|k|^2|\nabla u|^2-\langle \div k,\nabla u\rangle |\nabla u|)
\end{align}
where $K$ is the Gaussian curvature of $\Sigma_t$.
On the other side, we have by the equation $\nabla^2u=-k|\nabla u|$
\begin{align}
\Delta |\nabla u|=|k|^2|\nabla u|-\langle \div k,\nabla u\rangle
\end{align}
which finishes the proof.
\end{proof}
To prove rigidity of the spacetime PMT we will need to use every piece of information given by this corollary.
\section{Proof of Theorem \ref{Theorem: main}}\label{S:main}
Throughout this section we assume $E=|P|$, and let $u$ be the asymptotically linear spacetime harmonic function from Corollary \ref{cor:main}.
Let $e_3=\frac{\nabla u}{|\nabla u|}$.
For a fixed level set $\Sigma$, we can express the level set metric by $dx^2_1+dx^2_2$ which is possible since $\Sigma$ is flat.
Let $e_1=\partial_{x_1}$, $e_2=\partial_{x_2}$, then we extend $e_1$, $e_2$ to the entire manifold such that $\{e_1,e_2,e_3\}$ forms an orthonormal frame. We use Greek letter $\alpha$, $\beta$ to denote $e_1,e_2$, and Roman letters $i,j,k,l$ to denote $e_1,e_2,e_3$.
\
We define $\bar R_{ijkl}=R_{ijkl}+k_{il}k_{jk}-k_{ik}k_{jl}$ and say that $(M,g,k)$ satisfies the Gauss- and Codazzi equations if $\bar R_{ijkl}=0$ and $\nabla_ik_{jk}-\nabla_jk_{ik}=0$ for all $i,j,k,l$.
\begin{proposition}\label{Fundamental theorem}
Suppose $(M,g,k)$ is asymptotically flat, and satisfies the Gauss and Codazzi equations.
Then $(M,g,k)$ arises as a subset of Minkowski spacetime.
\end{proposition}
This is the Lorentzian version of the well-known fundamental theorem for hypersurfaces.
For the convenience of the reader we provide a proof in Appendix \ref{A:fundamental}.
In the next two lemma we demonstrate that the majority of the Gauss- and Codazzi equations are already satisfied.
\begin{lemma}\label{codazzi 1}
We have
\begin{align}
0=&\nabla_1 k_{23}-\nabla_2 k_{13},\\
0=& \nabla_\alpha k_{\beta\beta}-\nabla_\beta k_{\alpha\beta},\\
0=&\nabla_\alpha k_{33}-\nabla_3k_{\alpha 3}.
\end{align}
\end{lemma}
\begin{proof}
The first identity follows from
\begin{align}
\nabla_1k_{23}-\nabla_2k_{13}=&-\nabla_1\frac{\nabla^2_{23}u}{|\nabla u|}+\nabla_2\frac{\nabla^2_{13}u}{|\nabla u|}=R_{2133}=0.
\end{align}
Observe that $J_\alpha=0$ implies
\begin{align}
\nabla_\beta k_{\alpha\beta}-\nabla_\alpha k_{\beta\beta}+\nabla_3k_{\alpha3}-\nabla_\alpha k_{33}=0.
\end{align}
Thus, the last two identities follow from
\begin{align}
\nabla_3k_{\alpha 3}-\nabla_\alpha k_{33}=-\nabla_3\frac{\nabla^2_{\alpha 3}u}{|\nabla u|}+\nabla_\alpha\frac{\nabla^2_{33}u}{|\nabla u|}=R_{\alpha 333}=0.
\end{align}
\end{proof}
\begin{lemma}\label{Lemma:spacetime curvature}
We have
\begin{align}
\bar R_{1212}=&0,\\
\bar R_{\alpha\beta3\alpha}=&0,\\
\bar R_{\alpha33\beta}=&A_{\alpha\beta}.
\end{align}
where $A_{\alpha\beta}:=\nabla_3k_{\alpha\beta}-\nabla_\alpha k_{\beta3}$.
\end{lemma}
\begin{proof}
Using the Gauss equations we obtain
\begin{align}
R_{1212}=2K-h_{11}h_{22}+h_{12}^2.
\end{align}
Thus, the first identity follows from $K=0$ and $h=-k|_\Sigma$.
Next, we compute
\begin{align}
R_{\alpha\beta3\alpha}=(\nabla_\alpha\nabla_\beta-\nabla_\beta\nabla_\alpha)\frac{\nabla_\alpha u}{|\nabla u|}.
\end{align}
Using the spacetime Hessian equation $\nabla^2u=-k|\nabla u|$, we obtain
\begin{align}
\bar R_{\alpha\beta\alpha3}=\nabla_\alpha k_{\alpha\beta}-\nabla_\beta k_{\alpha\alpha}=0,
\end{align}
where the last equality follows from the previous lemma.
Finally, the third identity follows in the same spirit as the second one.
\end{proof}
Next, we show that $A_{\alpha\beta}$ is vanishing.
This will be achieved by PDE methods in combination with the asymptotics of $g,k$.
\begin{lemma}\label{L:F existence} On each level set,
there exists a twice differentiable function $F$ such that
\begin{align}
\nabla^\Sigma_{\alpha\beta}F=|\nabla u|^{-2}A_{\alpha\beta}.
\end{align}
\end{lemma}
For the proof of this lemma we need to additionally assume that $g\in C^3(M)$ and $k\in C^2(M)$.
However, we provide an alternative approach to the spacetime PMT rigidity in Appendix \ref{B:Killing development}.
This approach does not require such additional regularity of $g$ and $k$ and therefore establishes Theorem \ref{Theorem: main} in full generality.
\begin{proof}
We first show that $\partial_2 (|\nabla u|^{-2} A_{11})=\partial_1 (|\nabla u|^{-2}A_{12})$ and $\partial_1 (|\nabla u|^{-2}A_{22})=\partial_2 (|\nabla u|^{-2}A_{12})$.
Since $\langle \nabla_{e_\alpha}e_3,e_\beta\rangle=-k_{\alpha \beta} $ and applying Lemma \ref{codazzi 1}, we obtain
\begin{align}
\partial_2 A_{11}
=& \partial_2 (\nabla_3k_{11}-\nabla_1k_{13})
\\=&\nabla_2(\nabla_3k_{11}-\nabla_1k_{13})-k_{2}^\alpha\nabla_\alpha k_{11}+2k_{21}\nabla_3 k_{31}
\\&-k_{21}\nabla_3k_{13}-k_{21}\nabla_1 k_{33}+k_{2}^\alpha\nabla_1 k_{1\alpha}
\\=&\nabla_2(\nabla_3k_{11}-\nabla_1k_{13}).
\end{align}
Therefore, we have
\begin{align}
&\partial_2 A_{11}-\partial_1 A_{12}
\\=&\nabla_2(\nabla_3k_{11}-\nabla_1k_{13})-\nabla_1(\nabla_3k_{12}-\nabla_2k_{13})
\\=& \nabla_3\nabla_2k_{11}-2R_{231}^ik_{1i}-\nabla_2\nabla_1k_{13}
-(\nabla_3\nabla_1k_{12}-R_{131}^ik_{i2}-R_{132}^ik_{1i})
\\&+(\nabla_2\nabla_1k_{13}-R_{121}^ik_{i3}-R_{123}^ik_{1i})
\\=& \nabla_3\nabla_2k_{11}-\nabla_3\nabla_1 k_{12}
-R_{2312}k_{12}-R_{2313}k_{13}+R_{1312}k_{22}
\\&+R_{1313}k_{23}-R_{1212}k_{23}-R_{1213}k_{33}
\\=& \nabla_3\nabla_2k_{11}-\nabla_3\nabla_1 k_{12}
-(k_{12}k_{23}-k_{22}k_{13})k_{12}-(k_{12}k_{33}-k_{23}k_{13}-\nabla_3k_{12}+\nabla_1k_{23})k_{13}
\\&+(k_{11}k_{23}-k_{12}k_{13})k_{22}+(k_{11}k_{33}-k_{13}^2-\nabla_3k_{11}+\nabla_1k_{13})k_{23}
\\&-(k_{11}k_{22}-k_{12}^2)k_{23}-(k_{11}k_{23}-k_{13}k_{12})k_{33}
\\=&\nabla_3\nabla_2k_{11}-\nabla_3\nabla_1 k_{12}-(-\nabla_3k_{12}+\nabla_1k_{23})k_{13}+(-\nabla_3k_{11}+\nabla_1k_{13})k_{23}.
\label{a2-c1,1}
\end{align}
Because $\langle\nabla_3e_\alpha,e_3\rangle=-\langle\nabla_3 e_3,e_\alpha\rangle=k_{\alpha3}$, we deduce
\begin{align}
&\nabla_3\nabla_2k_{11}-\nabla_3\nabla_1 k_{12}
\\=&\partial_3(\nabla_2 k_{11})-\nabla_{\nabla_{3}e_2}k_{11}-2\nabla_2k(\nabla_3 e_1,e_1)
\\&-\partial_3(\nabla_1k_{12})+\nabla_{\nabla_3e_1}k_{12}+\nabla_1k(\nabla_3e_1,e_2)+\nabla_1k(e_1,\nabla_3e_2)
\\=&-\langle e_1,\nabla_3e_2\rangle\nabla_1k_{11}-k_{23}\nabla_3k_{11}-2\langle\nabla_3e_1,e_2\rangle\nabla_2k_{21}-2k_{13}\nabla_2k_{31}
+\langle e_2,\nabla_3 e_1\rangle \nabla_2k_{12}
\\&+k_{13}\nabla_3k_{12}+\langle e_2,\nabla_3e_1\rangle\nabla_1k_{22}+k_{13}\nabla_1k_{32}+\langle e_1,\nabla_3e_2\rangle\nabla_1k_{11}+k_{23}\nabla_1k_{13}
\\=&k_{23}(\nabla_1k_{13}-\nabla_3k_{11})-k_{13}(\nabla_1k_{32}-\nabla_{3}k_{12}).
\label{a2-c1,2}
\end{align}
Combing Equation \eqref{a2-c1,1} and \eqref{a2-c1,2} yields
\begin{equation}
\partial_2A_{11}-\partial_1A_{12}=2A_{12}k_{13}-2A_{11}k_{23}.
\end{equation}
Moreover, we have $\partial_\alpha |\nabla u|=-k_{\alpha 3}$ which implies
\begin{align}
&\partial_2(|\nabla u|^{-2} A_{11})-\partial_1(|\nabla u|^{-2} A_{12})
\\=&|\nabla u|^{-2}(\partial_2 A_{11}-\partial_1 A_{12})+A_{11}\partial_2|\nabla u|^{-2}-A_{12}\partial_1 |\nabla u|^{-2}
\\=&|\nabla u|^{-2}(2A_{12}k_{13}-2A_{11}k_{23})+2A_{11}|\nabla u|^{-2}k_{23}-2A_{12}|\nabla u|^{-2}k_{13}
\\=&0.
\end{align}
Therefore, $|\nabla u|^{-2}A_{11} dx_1+|\nabla u|^{-2}A_{12} dx_2$ is closed, where $dx_1$ and $dx_2$ are the dual 1-forms of $e_1$ and $e_2$.
Since the topology of a level set is trivial, there exists on each level set a function which we suggestively denote by $F_1$ such that $dF_1=|\nabla u|^{-2}A_{11} dx_1+|\nabla u|^{-2}A_{12} dx_2$.
Replacing the roles of $e_1$ and $e_2$, there exists another function $F_2$ such that $dF_2=|\nabla u|^{-2}A_{12} dx_1+|\nabla u|^{-2}A_{22} dx_2$.
Next, we compute
\begin{align}
d(F_1dx_1+F_2dx_2)=&\frac{\partial F_1}{\partial x_2}dx_2\wedge dx_1+\frac{\partial F_2}{\partial x_1}dx_1\wedge d x_2
\\=&(|\nabla u|^{-2}A_{12}-|\nabla u|^{-2}A_{12})dx_2\wedge dx_1=0.
\end{align}
Thus there exists an $F$ with $dF=F_1dx_1+F_2dx_2$.
\end{proof}
\begin{lemma}\label{F:linear}
On each level set, $F$ is a linear function with respect to $x_1$ and $x_2$, i.e. $\nabla^2_\Sigma F=0$.
\end{lemma}
\begin{proof}
First observe that $F$ is superharmonic on each level set, i.e.
\begin{align}
\Delta^\Sigma F\ge 0
\end{align}
which follows immediately from
\begin{equation}
\Delta^\Sigma F=
|\nabla u|^{-2}(A_{11}+A_{22})=-|\nabla u|^{-2}J_3
=|\nabla u|^{-2}\mu\ge0.
\end{equation}
Since $\partial^l k_{ij}=O(|x|^{-\tau-l-1})$, $\tau>\frac{1}{2}$, $l=0,1$, we obtain
\begin{equation}
F_{11}=\nabla^\Sigma_{11}F=\nabla_3k_{11}-\nabla_1 k_{13}=O(|x|^{-\tau-2}).
\end{equation}
Therefore, integrating $F_{11}$ twice on the level set, we see that $F=L+B$, where $L$ is a linear function with respect to $\{x_1,x_2\}$, and $B$ is a bounded function.
Combining this with our previous observation yields $\Delta^\Sigma B=\Delta^\Sigma F\ge0$. Thus, $B$ is constant in view of Liouville's theorem.
\end{proof}
\begin{proof}[Proof of Theorem \ref{Theorem: main}]
Since $\nabla^2_\Sigma F=0$, $(M,g,k)$ satisfies the Gauss and Codazzi equations which completes the proof in view of the Proposition \ref{Fundamental theorem}.
\end{proof}
| 3e0cf209474d871b56cc1d9b359cb0c08f6bc034 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Level design is a core feature of what defines a video game. When constructed correctly, it is a main determinant of player experience. From the designer's perspective, level design can be either a tedious, but necessary step in the game's development or a creatively freeing process - sometimes it is both. Most levels are designed with the intent to teach the game's interactable space - the mechanics - to the player in a way that is (ideally) engaging, fun, visually pleasing, intuitive, and informative \cite{rogers2014level,koster2013theory,green2017press}.
Levels designed for tutorial sections of the game create simplistic and low-risk environments. These levels are direct, and sometimes oblique in their intention so the player can grasp the core mechanics of the game as quickly as possible. As the player becomes more familiar with the mechanics and how they work together in the game's system, the levels should also increase in complexity and challenge. The general design of these levels in turn needs to be as complex and engaging both visually and functionally~\cite{khalifa2019intentional,anthropy2014game}.
Most games demonstrate each mechanic at least once throughout the entire level space; combining and ordering them in a way that builds itself based on the player's current skill as they get more familiar with the game \cite{totten_2016,anthropy2014game}. However, designing levels to explore multiple combinations of mechanics is an arduous task to undertake for a level designer. While it is unlikely that a player would play all of these levels, creating this possibility space of levels would allow the level designer to hand-pick and order them in a way that the mechanics feel coherent. Furthermore, an adaptive game with a diverse set of levels would allow the player to explore different combinations of mechanics according to their own pace and preference. For example, if a player is having difficulties with a certain necessary mechanic, such as long jumps in most Super Mario games or the spin attack in most Legend of Zelda games, having a specific subset of levels with a focus on these more challenging mechanics would allow the player to develop their skill with the mechanic better than a single tutorial level~\cite{anthropy2014game}. In contrast, if a player has fully mastered a mechanic, the game could select a level that uses the mechanic in a more challenging situation or a level with an entirely different mechanic space for them to master next. Levels could be selected automatically from the generated level space and dynamically ordered in a way that adapts to the player's current skill level as opposed to a hard-coded level ordering~\cite{yannakakis2011experience}. Generating a diversity of levels that explore multiple mechanic combinations would save time in the design process and allow for more creative flexibility in a way that manually designed levels could not provide.
This paper presents a system that seeks to combine human design and AI-driven design to enable mixed-initiative collaborative game level creation. Users can choose to start from a blank slate with their work while adding their own edits then have an AI back-end evolve their work towards a pre-defined objective. This objective function can be defined by minimalism in design, maximization of game mechanic coverage, overall quality, or any other feature that could contribute to the quality of the level. Alternatively, users may select from a variety of AI suggestions and pre-generated samples to begin their work and then make changes as necessary. This design process is not limited to the initializing step of the level; the user and AI system can switch their roles as designers at any point in the creation process. Concurrently, the AI system will look at what its previous users have created and submitted, and ask new users to design levels that complement what's already there. With this design process, the mechanic space of a game can be fully explored and every combination of mechanics can be represented by a level. With a human-based rating system, the automated system can learn to design levels with better quality and the human users can design levels that are missing from this mechanic combination space.
This project demonstrates the mixed-initiative collaborative process through level design for the independent, Sokoban-like game `Baba is You' - a game whose mechanics are defined and modified by the level design itself and the player's interaction with it. Levels can be made either by users, AI, or a mixed combination of both and uploaded to the level database to be used for future creations and to improve the quality of the AI's objective function.
\subsection{Baba is Y'all v1 (prototype)}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{imgs/level_matrix.png}
\caption{Baba is Y'all Version 1 Main Screen (from April 2020)}
\label{fig:levelMat1}
\end{figure}
The first version of Baba is Y'all (BiY v1) was released officially in March 29th, 2020, and promoted chiefly on Twitter. This version served as a prototype and proof-of-concept system for mixed-initiative AI-assisted game content collaboration specifically for designing levels in the game `Baba is You' (Arvi 'Hempuli' Teikari, 2017).
This system was built on concepts from three different areas of content creation:
\begin{itemize}
\item \textbf{Crowdsourcing:} a model used by different systems that allows a large set of users to contribute toward a common goal provided by the system~\cite{brabham2013crowdsourcing}. For example, Wikipedia users participate to fill in missing information for particular content.
\item \textbf{User content creation:}
allows players to create levels for a game/system and upload them online to the level database for other players to play and enjoy - i.e. Super Mario Maker (Nintendo, 2015), Line Rider (inXile Entertainment, 2006), and LittleBigPlanet (Media Molecule, 2008).
\item \textbf{Quality diversity:} the underlying technique behind our system. It ensures that the levels made from combining the first 2 concepts are of both good quality and diverse in terms of the feature space they are established in~\cite{pugh2016quality}. For this system, the feature space is defined as the potential game mechanics implemented in each level.
\end{itemize}
The Baba is Y'all website (as shown in figure~\ref{fig:levelMat1}) was a prototype example of a mixed-initiative collaborative level designing system. However, the site was limited by the steep learning curve required to interact with the system~\cite{charity2020baba}. Features of the site were overwhelming to use and lack cohesion in navigating the site.
\subsection{Baba is Y'all v2 (updated release)}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{imgs/dark_main.png}
\caption{Baba is Y'all Version 2 Main Screen (as of September 2021)}
\label{fig:levelMat2}
\end{figure}
The second version of Baba is Y'all\footnote{http://equius.gil.engineering.nyu.edu/} (BiY v2) was released on May 27th, 2021 and designed to have a more user-friendly setup. It was similarly promoted via Twitter and on mailing lists. This version includes a cleaner, more compact, and more fluid user interface for the entire website and consolidated many of the separate features from the BiY v1 site onto fewer pages for easier access. Three main webpages were created for this updated system.
Unlike the previous version, which showed all of the mechanic combination levels (both from the database and unmade) in random order, the updated level selection page adds level tabs that separates levels by recently added (New), highest rated (Top), and levels with rules that had not been made yet (Unmade.) A carousel scrolling feature shows 9 levels at a time to not overwhelm the player with choices (as shown in figure~\ref{fig:levelMat2}). The level rating system is also included on the main page as a tab, as well as the search feature. The personal level selection tab allows users to see their previously submitted levels and login to their account to submit levels with their username as the author or co-author.
The updated level editing page consolidates both the user editing with the PCG level evolution onto one page. Users can easily switch between manually editing the level themselves and allowing the PCG back-end system to edit the level while pausing in between. Users can also select rule objectives for the system to evolve towards implementing. To fight the problem of blank canvas paralysis, users can start from a set of different types of levels (both PCG and user-made)~\cite{krall2012artist}. Once a level is successfully solved, users may name the level upon submission - further personalizing the levels and assigning authorship.
A slideshow tutorial is provided for the users and describes every feature and function of the site instead of the walkthrough video that was featured on BiY v1. Users can also play a demo version of the `Baba is You' (Arvi 'Hempuli' Teikari, 2017) game to familiarize themselves with the game mechanics/rule space and how they interact with each other (game dynamics). For quick assistance, a helper tool is provided on the level editing page as a refresher on how to use the editing tool.
In addition to updating the features and collecting more data about the levels created, we conducted a formal user study with 76 participants to gather information about which features they chose to use for their level creation process and their subjective opinion on using the site overall. This user study, as well as the general level statistics collected from the site's database, showed that our new interface better facilitated the user-AI collaborative experience to create more diverse levels.
\section{Background and Related Work}
The Baba is Y'all system uses the following methods in the collaborative level design process: procedural content generation to create new levels from the AI backend, quality diversity to maintain the different kinds of levels produced from the system and show the coverage of game mechanics across each level, crowdsourcing so the AI may learn to create new levels from previously submitted "valid" levels - either those made exclusively by users, the system itself, or a combination of both, and finally mixed-initiative AI so that the user and evolutionary algorithm can develop the level together. Each method is described as the following:
\subsection{Procedural Content Generation}
Procedural content generation (PCG) is defined as the process of using a computer program to create content that with limited or indirect user input \cite{shaker2016procedural}. Such methods can make an automated, quicker, and more efficient content creation process, and also enable aesthetics based on generation. PCG has been used in games from the 1980's Rogue to its descendent genre of the Rogue-likes used in games such as Spelunky (Mossmouth, LLC, 2008) and Hades (Supergiant Games, 2020), as well as games that revolve around level and world generation such as Minecraft (Mojang, 2011) and No Man's Sky (Hello Games, 2016). PCG can be used to build levels such as The Binding of Isaac (Edmund McMillen, 2011), enemy encounters such as Phoenix HD (Firi Games, 2011), or item or weapon generation such as Borderlands (Gearbox Software, 2009). In academia, PCG has been explored in many different game facets for generating assets \cite{ruela2017procedural, gonzalez2020generating}, mechanics \cite{khalifa2019general, togelius2008experiment,browne2010evolutionary}, levels \cite{snodgrass2016learning,charity2020mech}, boss fights~\cite{siu2016programming}, tutorials~\cite{khalifa2019intentional,green2018atdelfi}, or even other generators \cite{kerssemakers2012procedural,earle2021learning,earle2021illuminating,khalifa2020multi}.
A plethora of AI methods underpin successful PCG approaches, including evolutionary search \cite{togelius2010search}, supervised and unsupervised learning \cite{summerville2018procedural,liu2021deep}, and reinforcement learning \cite{khalifa2020pcgrl}. The results of these implementations have led to PCG processes being able to generate higher quality, more generalizable, and more diverse content. PCG is used in the Baba is Y'all system to allow the mutator module to create new `Baba is You' levels.
\subsection{Quality Diversity}
Quality-diversity (QD) search based methods are increasing in usage for both game researchers and AI researchers \cite{pugh2016quality,gravina2019procedural}. Quality-diversity techniques are search based techniques that try to generate a set of diverse solutions while maintaining high level of quality for each solution. A well-known and popular example is MAP-Elites, an evolutionary algorithm that uses a multi-dimensional map instead of a population to store its solutions~\cite{mouret2015illuminating}. This map is constructed by dividing the solution space into a group of cells based on a pre-defined behavior characteristics. Any new solution found will not only be evaluated for fitness but also for its defined characteristics then placed in the correct cell in the MAP-Elites map. If the cell is not empty, both solutions compete and only the fitter solution survives. Because of the map maintenance and the cell competition, MAP-Elites can guarantee a map of diverse and high quality solutions, after a finite number of iterations through the generated population.
The MAP-Elites algorithm has also been extended into Constrained MAP-Elites \cite{khalifa2018talakat, khalifa2019intentional, alvarez2019empowering}, Covariance Matrix Adaptation using MAP-Elites (CMA-ME) \cite{fontaine2020covariance}, Monte Carlo Elites~\cite{sfikas2021monte}, MAP-Elites via Gradient Arborescence~\cite{fontaine2021differentiable}, and etc. For this project, we use the Constrained MAP-Elites algorithm to maintain a diverse population of `Baba is You' levels where the behavior characteristic space of the matrix is defined by the starting and ending rules of a level when it is submitted.
\subsection{Crowdsourcing data and content}
Some, but relatively few, games allow users to submit their own custom creations using the game's engine as most games do not have their source code available or even partially accessible for modifications to add more content in the context of the game. Whether through a built-in level editing system seen in games like Super Mario Maker (Nintendo, 2015), LittleBigPlanet (MediaMolecule, 2008), or LineRider (inXile Entertainment, 2006) or through a modding community that alter the source code for notable games such as Skyrim (Bethesda, 2011) Minecraft (Mojang, 2011,) or Friday Night Funkin' (Ninjamuffin99, 2020), players can create their own content to enhance their experience and/or share with others.
In crowdsourcing, many users contribute data that can be used for a common goal. Some systems like Wikipedia rely entirely on content submitted by their user base in order to provide information to others on a given subject. Other systems like Amazon's MechanicalTurk crowdsource data collection, such as research experiments \cite{buhrmester2016amazon}, by outsourcing small tasks to multiple users for a small wage. An example of a game generator based on crowdsourced data is Barros et al.’s DATA Agent \cite{barros2018killed,green2018data}, which uses crowd-sourced data such as Wikipedia to create a point-click adventure game sourced from a large corpus of open data to generate interesting adventure games.
What differentiates the Baba is Y'all system from other level editing systems or interactive PCG systems is that the Baba is Y'all site has a central goal: populate the MAP-Elites matrix with levels that cover all possible rule combinations. With this system, users may freely create the levels they want, but they may also work towards completing the global goal of making levels with a behavior characteristic that has not been made before. Participation in this task is encouraged by the AI back-end system that keeps track of missing cells in the MAP-Elites matrix.
\subsection{Mixed-Initiative AI}
Mixed-initiative AI systems involve a co-creation of content between a human user and an artificially intelligent system~\cite{yannakakis2014mixed}. Previous mixed-initiative systems include selecting from and evolving a population of generated images \cite{secretan2008picbreeder,bontrager2018deep}, composing music \cite{mann2016ai,tokui2000music}, and creating game levels through suggestive feedback \cite{machado2019pitako}. Mixed-initiative and collaborative AI level editors for game systems have thoroughly been explored in the field as well through direct and indirect interaction with the AI backend system \cite{shaker2013ropossum,liapis2013sentient,butler2013mixed,guzdial2018co,zhou2021toward,bhaumik2021lode,alvarez2019empowering,smith2010tanagra,delarosa2021mixed}.
Since the release of the first Baba is Y'all prototype and paper~\cite{charity2020baba}, the implementation of mixed-initiative systems have grown in the game and AI research field. Bhaumik implemented an AI constrained system with their Lode Encoder level editing tool that only allowed users to edit a level from a set of levels generated by a variational autoencoder - forcing users to only edit from a palette provided by the AI back-end tool~\cite{bhaumik2021lode}. Delarosa used a reinforcement learning agent in a mixed-initiative web app to collaboratively suggest edits to Sokoban levels \cite{delarosa2021mixed}. Zhou used levels generated with the AI-assisted level editor Morai Maker (a Super Mario level editor) to apply transfer learning for level editing to Zelda \cite{zhou2021toward}. These recent developments look more into how the human users are affected through their relationship with collaborating with these AI systems and how it can be improved through examining the dimensionality of the QD algorithm, the evolutionary process, or the human-system interaction itself \cite{alvarez2020exploring}. We look to incorporate these new perspectives into this updated iteration of Baba is Y'all and evaluate the effects through a user study.
\section{System Description}
The updated Baba is Y'all site's features were condensed into 2 main pages to make navigation and level editing much easier and intuitive:
\begin{itemize}
\item \textbf{The Home Screen:} contains the level matrix \textit{Map Module}, the search page, the \textit{Rating Module} page, and the \textit{User Profile} page. From here, users can also change the visuals of the site from light to dark mode, view the tutorial section or the site stats page by clicking on the Baba and Keke sprites respectively at the top of the page, and create a new level from scratch by clicking on various 'Create New Level' buttons placed on various subpages. Figure~\ref{fig:levelMat2} shows the starting page of the home screen.
\item \textbf{The Level Editor Screen:} contains both the \textit{Editor Module} and the \textit{Mutator Module}. Users can also test their levels with themselves or with the Keke solver by clicking on the Baba and Keke icons at the bottom of the canvas. Figure~\ref{fig:editor_screen} shows the starting page of the level editor screen.
\end{itemize}
In the following subsections, we are going to explain the different modules that constitutes these two main screens. Each of the following modules are either being used in the home screen, the level editor screen, or both.
\subsection{Baba is You}
`Baba is You' (Arvi ``Hempuli'' Teikari, 2019) is a puzzle game where players can manipulate the rules of a level and properties of the game objects through Sokoban-like movements of pushing word blocks found on the map. These dynamically changing rules create interesting exploration spaces for both procedurally generating the levels and solving them. The different combinations of rules can also lead to a large diversity of level types that can be made in this space.
The general rules for the `Baba is You' game can be referred to from our previous paper~\cite{charity2020baba}. To reiterate, there are three types of rule formats in the game:
\begin{itemize}
\item \textbf{X-IS-(KEYWORD)} a property rule stating that the game object class `X' has a certain property such as `WIN', `YOU', `MOVE', etc.
\item \textbf{X-IS-X} a reflexive rule stating that the game object class `X' cannot be changed to another game object class.
\item \textbf{X-IS-Y} a transformative rule changing all game objects of class `X' into game objects of class `Y'.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.4\linewidth]{imgs/simple_level.png}
\caption{An example of a simple `Baba is You' level.}
\label{fig:simple_map}
\end{figure}
The game sprites are divided into two main different classes: the object class and the keyword class. Sprites in the object class represent the interactable objects in the map as well as the literal word representation for the object. Sprites in the keyword class represent the rules of the level that manipulate the properties of the objects. For example, figure~\ref{fig:simple_map} shows four different object class sprites [BABA (object and corresponding word) and FLAG (object and corresponding word)] and three different keyword class sprites [IS (x2), YOU, and WIN]. The keyword class sprites are arranged in two rules: `BABA-IS-YOU' allowing the player to control all the Baba objects and `FLAG-IS-WIN' indicating that reaching any flag object will make the player win the level. The system has a total of 32 different sprites: 11 object class sprites and 21 keyword class sprites. Because the game allows rule manipulation, object classes are arbitrary in the game as they serve only to provide a variety of objects for rules to affect and for aesthetic pleasure.
\subsection{Game Module}
The game module is responsible for simulating a `Baba is You' level. It also allows users to test the playability of levels either by directly playing through the level themselves or by allowing a solver agent to attempt to solve it. This component is used on the home screen when a user selects a level to play and the editor screen for a user to test their created level.
Because the game rules are dynamic and can be altered by the player at any stage in the solution, the system keeps track of all the active rules at every state. Once the win condition has been met, the game module records the current solution, the active rules at the start of the level, and the active rules when the solution has been reached. These properties are saved to be used and interpreted by the Map module (section~\ref{sec:map_module}).
The activated rules are used as the level's characteristic feature representation and saved as a chromosome to the MAP-Elites matrix.
The game module provides an AI solver called 'KEKE' (based on one of the characters traditionally used as an autonomous 'NPC' in the game). KEKE uses a greedy best-first tree search algorithm that tries to solve the input level. The branching space is based on the five possible inputs a player can do within the game: move left, move right, move up, move down, and do nothing. The algorithm uses a heuristic function based on a weighted average of the Manhattan distance to the centroid distance for 3 different groups: keyword objects, objects associated with the `WIN' rule, and objects associated with the `PUSH' rule. These were chosen based on their critical importance for the user solving the level - as winning objects are required to complete the level, keyword objects allow for manipulation of active rules, and pushable objects can directly and indirectly affect the layout of a level map and therefore the accessibility of player objects to reach winning objects. The heuristic function is represented by the following equation:
\begin{equation}
h = (n + w + p) / 3
\end{equation}
where $h$ is the final heuristic value for placement in the priority queue, $n$ is the minimum Manhatttan distance from any player object to the nearest winnable object, $w$ is the minimum Manhatttan distance from any player object to the nearest word sprite, and $p$ is the minimum Manhatttan distance from any player object to the nearest pushable object.
As an update for this version of the system, the agent can run for a maximum of 10000 iterations and can be stopped at any time. A user may also attempt to solve part of the level themselves and the KEKE solver can pick up where the user left off to attempt to solve the remainder of the level. This creates a mixed-initiative approach to solving the levels in addition to editing the levels. However, even with this collaborative approach, the system still has limitations and difficulty solving levels with complex solutions - specifically solutions that require back-tracking across the level after a rule has been changed. The solver runs on the client side of the site and is limited by the capacity of the user's computational resources. Future work will look into improving the solver system to reduce computational resource. We will also look for better solving algorithms to improve the utility of the solver such as Monte Carlo Tree Search (MCTS) with reversibility compression~\cite{cook2021monte}.
\subsection{Editor Module}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\linewidth]{imgs/editor_screen.png}
\caption{A screenshot of the level editor screen}
\label{fig:editor_screen}
\end{figure}
The editor module of the system allows human users to create their own `Baba is You' levels in the same vain of Super Mario Maker (Nintendo, 2015). Figure~\ref{fig:editor_screen} shows the editor window that is available for the user. The user can place and erase any game sprite or keyword at any location on the map using the provided tools. As a basis, the user can start modifying either a blank map, a basic map (a map with X-IS-YOU and Y-IS-WIN rules already placed with X and Y objects), a randomly generated map, or an elite level provided by the Map Module. Similar to Super Mario Maker (Nintendo, 2015), the created levels can only be submitted after they are tested by the human player or the AI agent to check for solvability. For testing the level, the editor module sends the level information to the game module to allow the user to test it.
This updated version of the site also includes an undo and redo feature so that users may erase any changes they make. A selection and lasso feature is also available so users can select specific areas of the level and move them to another location. Unlike the previous version, all tiles are available to the user on the same screen and the user may seamlessly transition from the editor module to the mutator module and vice versa for ease of access and better interactivity and collaboration between the AI system and the user.
\subsection{Mutator Module}\label{sec:mutator_module}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{imgs/evolver_screen.png}
\caption{A screenshot of the level evolver page}
\label{fig:evolver_screen}
\end{figure}
The Mutator module is a procedural content level generator. More specifically, the Baba is Y'all system uses an evolutionary level generator that defines a fitness function based on a version of tile-pattern Kullback-Liebler Divergence (ETPKLDiv\footnote{https://github.com/amidos2006/ETPKLDiv}) algorithm~\cite{lucas2019tile}. Figure~\ref{fig:evolver_screen} shows the updated interface used by the evolver. As mentioned before in the previous subsection, this version of the mutator module can interface seamlessly with the other modules to allow the user more ease of access between manual editing and evolutionary editing. The user can easily transfer the level from the editor module to the mutator module and vice versa.
When switching between the editor module and the mutator module, the level loses its pure procedurally generated or pure human-designed quality and becomes a hybrid of the two - thus mixed-initiative interaction between the algorithm and the user.
The evolver interface provides the user with multiple customizations such as the initialization method, stopping criteria, evolution pausing, and an application of a mutation function allowing manual user control. With these features, the user is not directly changing the evolution process itself, but instead guiding and limiting the algorithm towards generating the level they want.
The ETPKLDiv algorithm uses a 1+1 evolution strategy, also known as a hillclimber, to improve the similarity between the current evolved levels and a reference level. The algorithm uses a sliding window of a fixed size to calculate the probability of each tile configuration (called tile patterns) in both the reference level and the evolved level and tries to minimize the Kullback-Liebler Divergence between both probability distributions.
Like Lucas and Volz, we use a window size of 3x3 for the tile selection. This was to maximize the probability of generating initial rules for a level, since rules in `Baba is You' are made up of 3 tiles. However, in our project, we used 2+2 evolution strategy instead of 1+1 used to allow slightly more diversity in the population~\cite{lucas2019tile}. We also modified the fitness function to allow it to compare with more than one level. The fitness value also includes the potential solvability of the level ($p$), the ratio of empty tiles ($s$), and the ratio of useless sprites ($u$).
The final fitness equation for a level is as follows:
\begin{equation}\label{eq:fitness}
fitness_{new} = min(fitness_{old}) + u + p + 0.1 \cdot s
\end{equation}
where $fitness_{old}$ is the Kullback-Lievler Divergence fitness function from the Lucas and Volz work~\cite{lucas2019tile} compared to a reference level. The minimum operator is added as we are using multiple reference levels instead of one and we want to pick the fitness of the most similar reference level.
In the updated version of Baba is Y'all, we recalculate the ratio of useless objects ($u$) used in the original version's equation. The value $u$ is defined as the combined percentage of unnecessary object and word sprites in the level. This is broken up into 2 variables $o$ and $w$ for the objects and words respectively. The $o$ value corresponds to the objects that are not required or predicted to act as a constraint or solution for the level. The value for $o$ can be calculated as follows:
\begin{equation}
o = \frac{i}{j}
\end{equation}
where $i$ is the number of objects sprites initialized in the level without a related object-word sprite and $j$ is the total number of object sprites initialized in the level. While the $w$ value corresponds to the words that have no associated object in the map (this does not apply to keyword class words such as ``KILL'' or ``MOVE''). The value for $w$ can be calculated as follows:
\begin{equation}
w = \frac{k}{l}
\end{equation}
where $k$ is the number of word sprites initialized in the level without a related object-word sprite and $l$ is the total number of word sprites initialized in the level. To combine both variables $o$ and $w$ into the one variable $u$ a constant ratio is applied. In the system, 0.85 is applied to the $o$ variable and 0.15 to $w$. This is to more weight on reducing the number of useless object sprites as opposed to useless word sprites, as word sprites can be used to modify the properties of objects or transform other object sprites.
The $u$ value is implemented in order to prevent noise within the level due to having object tiles that cannot be manipulated in any way or have relevancy to the level. A human-made level may include these ``useless'' tiles for aesthetic purposes or to give the level a theme - similar to the original `Baba is You' levels. However, the PCG algorithm optimizes towards efficiency and minimalist levels, therefore ignoring the subjective aspect of a level's quality (which can be added later by the user).
The playability of the level ($p$) is a binary constraint value that determines whether a level is potentially winnable or not. The value can be calculated as follows:
\begin{equation}
p =
\begin{cases}
1, & \text{has [`X-IS-YOU' rule, `WIN' keyword]} \\
0, & otherwise
\end{cases}
\end{equation}
This is to ensure any levels that are absolutely impossible to play or win are penalized in the population and less likely to be mutated and evolved from in future generations. We used a simple playability constraint check instead of checking for playability using the solver because the solver take time to check for playability. Also, all playable levels by the solver usually end up being easy levels due to the limited search space we are given for the best first algorithm.
The ratio of empty tiles ($s$) is the ratio of empty space tiles to all of the tiles in the level. The equation can be calculated as follows:
\begin{equation}
s = \frac{e}{t}
\end{equation}
where $e$ is the number of empty spaces in the level and $t$ is the total number of tiles found in the level. The value $s$ is multiplied with a value of $0.1$ in equation~\ref{eq:fitness} to avoid heavy penalization for having any empty spaces in a level and to prevent encouragement for levels to mutate towards populating the level with an overabundance of similar tiles in order to eliminate any empty space.
The Mutator module is not run as a back-end process to find more levels, instead it has to be done manually by the user. This is done due to the fact that some generated levels cannot be solved without human input. One might wonder why not generate a huge corpus of levels and ask the users later to test them for the system. This could result in the system generating a multitude of levels that are either impossible to solve or are solvable but not subjectively ``good'' levels - levels the user would not find pleasing or enjoyable. This overabundance of ``garbage'' levels could lead to a waste of memory and a waste human resources. By allowing the user direct control over which levels are submitted from the generation algorithm, it still guarantees that the levels are solvable and with sufficient quality and promote using the tool in a mixed-initiative approach. Future work will explore implementing a fully autonomous generator and associated solver to expand the archive of levels without human input.
\subsection{Objective Module}\label{sec:objective_module}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\linewidth]{imgs/obj_screen.png}
\caption{A screenshot of the rule objective screen}
\label{fig:objective_screen}
\end{figure}
In conjunction with the Mutator module (section~\ref{sec:mutator_module}), an Objective Module has been implemented to help guide the evolver towards generating levels that match selected objectives - or rules - set by either the Map Module or the user. Like before this will nudge both the user and the evolver back-end towards creating levels with mechanic combinations that have not been made in the site database.
Users can select from the table of mechanics which sets of rules to include in the level - whether initially at the start of the level, at the solution, or either. Initial rules can be found automatically when the user or evolver edits the level, final rules can only be determined at the end of the level - when the solution has been found. Active rules are highlighted with a green backlight in the table and change accordingly when a rule is created or removed.
The evolver also prioritizes levels that match as many of the selected rules as possible. A cascading function is used to rank the generated levels from the chromosome population. The evolver first evaluates how well a generated level corresponds to the selected objectives then looks at the fitness function. With this, the evolver becomes more involved with expanding the level database for the site and actively tries to help the user fill these missing levels.
\subsection{Rating Module}\label{sec:rating_module}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\linewidth]{imgs/rating_screen.png}
\caption{A screenshot of the rating screen with 2 levels shown}
\label{fig:rating_screen}
\end{figure}
Like the original system, a rating for a single level is determined by comparison to another level within the site database. The user must determine the better level based on two qualities: level of challenge and quality of aesthetic design. A level that is considered `more challenging' could indicate that the solution search space for the level takes longer to arrive at or is not as intuitive or straightforward. A level that is considered to have `better design' represents that the level is more visually pleasing and elegant with its map representation - a quality that is hard to generate automatically with AI. Users can select between the two levels for each feature by shifting a slider towards one level or the other.
\subsection{Map Module}\label{sec:map_module}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\linewidth]{imgs/map_screen.png}
\caption{A screenshot of the map selection screen}
\label{fig:select_screen}
\end{figure}
The Map module functions as both storing all of the levels in the site database as well as recommending specific levels to the user to use for their own level creation process. The Map module is the core module of the system. To maintain distinguish-ability between quality and diverse levels, we implemented the MAP-Elites algorithm for this module.
\begin{table}[t]
\caption{Chromosome Rule Representation}
\centering
\begin{tabular}{|p{0.2\linewidth}|p{0.7\linewidth}|}
\hline
Rule Type & Definition \\
\hline
\hline
X-IS-X & objects of class X cannot be changed to another class \\
X-IS-Y & objects of class X will transform to class Y \\
X-IS-PUSH & X can be pushed \\
X-IS-MOVE & X will autonomously move \\
X-IS-STOP & X will prevent the player from passing through it\\
X-IS-KILL & X will kill the player on contact\\
X-IS-SINK & X will destroy any object on contact\\
X-IS-[PAIR] & both rules 'X-IS-HOT' and 'X-IS-MELT' are present \\
X,Y-IS-YOU & two distinct objects classes are controlled by the player \\
\hline
\end{tabular}
\label{tab:rrp}
\end{table}
When a level is submitted to be archived, the system uses the list of active rules at the start and the end of the level as behavior characteristic for the input level to determine its location in the map. There are 9 different rules checked for in each level - based on the possible rule mechanics that can be made in the Game module system. Table \ref{tab:rrp} shows the full list of possible rules. Since these rules can be active at the beginning or at the end, it makes the number of behavior characteristics equal to 18 instead of 9 which provide us with a map of $2^{18}$ cells.
The Map Module can recommend levels to start from when designing a new level. Like the Mutator Module (section~\ref{sec:mutator_module}), it also takes the Objective Module (section~\ref{sec:objective_module}) into consideration when selecting its recommendations. The Map Module can provide levels that most similarly match the objectives chosen and provide either other levels the user has previously made or high rated (and intuitively high quality) ``elite'' levels.
In this project we are using a multi population per each cell of the Map-Elites similar to the constrained Map-Elites~\cite{khalifa2018talakat}. The quality of the level is determined by user ratings - performed by the Rating Module.
\subsection{User Profiles}
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\linewidth]{imgs/user_screen.png}
\caption{A screenshot of the user profile screen for the user 'Milk'}
\label{fig:profile_screen}
\end{figure}
The user profiles feature is the newest addition to the Baba is Y'all site. Like the original system, if a user creates a profile through the site's login system and submits a level, they get authorship attributed to the submitted level. Users can also find their previously made levels on the profile page - called ``My Levels'' - and replay them, edit them, or view the level's mechanic combination. A user's personal stats for their level submissions can also be viewed on the page including the number of levels submitted, number of rule combinations contributed, and their top rated level. This feature was implemented to provide more user agency and personalization on the site and give users better access to their own submitted levels.
Through the search page, players can search for specific levels by username or by level name. This creates a sense of authorship over each of the levels, even if the level wasn't designed with any human input (i.e. a level with PCG.js as the author) and encourages the collaborative nature of the site between AI and human. Users may also share links to site levels via the game page.
\section{Results}
The following results were extracted from the entire Baba is Y'all v2 site and includes data from levels made from participants not involved with the study.
\subsection{User and Author-based Data}
All users on the Baba is Y'all site had the option of registering for a new account to easily find their saved work as well as attribute personal authorship to any levels they submitted. Those who participated in the user study were given pre-made usernames in order to verify the levels they submitted from their responses and to protect their identities. These users only had to provide an email address to register for both the site and the survey.
The site had a total of 727 unique users registered - only 78 (10\%) came from outside of the user study while the rest of the users participated in the survey.
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\linewidth]{imgs/Level-types.png}
\caption{Sample levels generated for the system. The left column is user generated levels, the middle column is evolver module levels, and the right column is mixed-initiative user and evolver levels}
\label{fig:level_types}
\end{figure}
We looked into all the levels created by the users and we divided them based on how the mixed-initiative tool was used to create them. We divided them into three main categories (as shown in figure~\ref{fig:level_types}):
\begin{itemize}
\item \textbf{User-Only levels:} were created from a blank map exclusively by the human user without any AI assistance.
\item \textbf{PCG-only levels:} were created solely by the AI tool without any human input aside from choosing which tool to use and when.
\item \textbf{Mixed-author levels:} involved both the human user as well as the AI tool in the creation process of the level.
\end{itemize}
\begin{table}[ht]
\begin{center}
\begin{tabular}{|c c c|}
\hline
Author Type & Number & \%\\
\hline\hline
User-only & 103 & 66.45 \\
\hline
PCG-only & 16 & 10.32\\
\hline
Mixed-author & 36 & 23.23\\
\hline
\hline
Total & 155 & 100\\
\hline
\end{tabular}
\end{center}
\caption{Authorship for levels submitted}
\label{tab:level_author}
\end{table}
The majority of the levels submitted were user only (66.45\%), however almost a quarter (23.23\%) of the levels submitted had mixed-authorship. Table \ref{tab:level_author} shows the full data for this area. Looking at this table, we notice that the amount of submitted levels are a lot less than total number of users ($155$ levels and $727$ users). This big difference in the numbers is due to releasing the system online with no security measures. This attracted a lot of bots that created multiple accounts so they could fill out the user survey via the link provided, but did not submit any levels.
\subsection{Level-based Data}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\linewidth]{site_graphs/rule_perc.png}
\caption{Site results for the rule distribution across levels submitted}
\label{fig:level_rule_dist}
\end{figure}
Looking into all the $155$ submitted levels, we found only $74$ different cells in the MAP-Elites matrix were covered. This is less than 1\% of the whole number of possible rule combinations ($2^{18}$ possible combinations). Figure~\ref{fig:level_rule_dist} shows the rule distributions over all of the levels submitted. The X-is-KILL rule was used the most in over half of the levels submitted and the X-is-STOP rule was used the second-most at 44.52\%. This may be because these rules create hazards for the player and add more depth to the level and solution. Meanwhile, the X-is-[PAIR] rule was used the least in only 12.9\% of the levels submitted. This is likely due to the lock-and-key nature of the rule combinations that require more intentionally placed word blocks that can also be accomplished with the X-is-SINK or X-is-KILL rule.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|c c c c|}
\hline
User Type & \# Rules & Sol. Length & Map Size (\# tiles) \\
\hline\hline
User-only & 2.563 $\pm$ 2.19 & 25.834 $\pm$ 26.11 & 117.883 $\pm$ 50.84\\
\hline
PCG-only & 1.00 $\pm$ 1.17 & 19.062 $\pm$ 14.68 & 95.437 $\pm$ 25.60\\
\hline
Mixed-author & \textbf{2.833 $\pm$ 2.56} & \textbf{26.027 $\pm$ 20.36} & \textbf{127.722 $\pm$ 49.91}\\
\hline
\end{tabular}
\end{center}
\caption{Averaged attributes for different types of created levels}
\label{tab:avg_author}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{site_graphs/rule_dist.png}
\caption{Rule distributions across the different authored levels}
\label{fig:rule_dist}
\end{figure}
The relation between rules and the different type of authors can be shown in table~\ref{tab:avg_author}. Some levels may use no rules at all (only containing the required X-is-YOU and X-is-WIN rules.) The mixed-author levels has the highest number of average rules per level ($2.833$), while PCG-only levels have the lowest average ($1$). The rule distributions for each author type are shown in Figure \ref{fig:rule_dist}. The PCG-authored levels had the least variability between rules while the Mixed-authored levels had the most variability. Mixed-author levels also had the highest average solution length and highest average level size, with PCG levels having the lowest for both attributes.
\section{User Study}
The following results were extracted from a Google Form survey given to the experiment participants. Users were instructed to play a level already made on the site, create a new level using the level editor, test it, and finally submit it to the site. They were also given the option to go through the tutorial of the site if they were unfamiliar with the `Baba is You' game or needed assistance with interacting with the level editor tool.
Of the $727$ users registered on the site, only a total of $170$ responses were received, however, only $76$ of these responses were valid. These responses were evaluated based on cross-validation and verification between the saved level on the website and the level ID they submitted via the survey that they claimed they authored. Many of these invalid responses contained levels that either did not exist in the database or were claimed to be authored by another user already. The following results are taken from the self-reported subjective survey given to the valid $76$ users.
\subsection{Demographic Data}\label{sec:demographics}
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{hor_survey_graphs/freq_v2.png}
\caption{A. Frequency for playing games; B. Frequency for designing levels for games}
\label{fig:freq_des_play}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{hor_survey_graphs/pref_v2.png}
\caption{Preference for solving or making puzzles}
\label{fig:design_pref}
\end{figure}
Half of the users who completed the survey answered that they frequently played video games (more than 10 hours a week) with around 80\% of the users stating they play for at least 2 hours a week (figure~\ref{fig:freq_des_play}). Conversely, only 28.9\% of users responded that they spend 2 or more hours a week designing levels for games with 40.8\% of users stating they never design levels at all (figure~\ref{fig:design_pref}). When asked if they prefer to solve or make puzzles, 50\% of participants responded that they prefer to solve puzzles, while only 6.6\% preferred the latter. 40.8\% of users were split on the preference for designing and solving puzzles.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{hor_survey_graphs/experience_v2.png}
\caption{A. Experience playing Sokoban; B. Experience with 'Baba is You'; C. Experience with AI-assisted level editing tools}
\label{fig:exp_graph}
\end{figure}
We asked participants if they had ever played the original game `Baba is You' by Hempuli (either the jam version or the Steam release as both contain the rules used in the Baba is Y'all site), played a Sokoban-like game (puzzle games with pushing block mechanics), and have experience with AI-assisted level editing tools. Figure~\ref{fig:exp_graph} shows the distribution of the users' answers for these questions. Only 30\% of participants had played the game before, meanwhile 22\% had heard of it but had never played it. For the rest, this study would be their first experience with the game. Interestingly enough, 96\% of the participants stated they had played a Sokoban-like game so we can infer that the learning curve would not be too harsh for the new players. Concerning AI-Assisted level editing tools, 75\% of users had never used them before, with 5.3\% stating they were unsure if they had ever used one - thus the learning curve for AI-collaboration would be much higher and new to participants.
\subsection{Self-Reported Site Interactions}
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{survey_graphs/feat.png}
\caption{Survey results for users' reports on the features they used}
\label{fig:feat_report}
\end{figure}
Figure \ref{fig:feat_report} shows the full list of features that participants interacted with on the site. Users were given the optional task to go through the tutorial section of the Baba is Y'all site to familiarize themselves with both the mechanics of the original `Baba is You' game, the AI assisted tools available to them through the level editor, and the site layout and navigation itself. 81.6\% of users went through this tutorial (whether fully or partially was not recorded.) The second task for users was to play a level that was previously submitted to the website database. 100\% of users were able to solve a level by themselves, however 72.4\% of users reported choosing to watch the Keke AI solver complete the submitted level as well. The third and final task for the participants was to submit their own `Baba is You' level using the level editor. Here, users were asked the most about their involvement with the AI system. Some users chose to create more than one level, so they may have multiple experiences and their design choices may not be mutually exclusive (i.e. using a blank level and also using an AI-suggested level.)
For the initial creation of the level, 88.2\% of users chose to start with a blank map. 9.2\% of users started with a level that had already been submitted to the level database - either a level that had been ranked as an elite level or a level created by the user themselves (in the case that they submitted more than one level during this study.) 6.6\% of users started with a level that was suggested from the 'Unmade' page - ideally with the intent to make a level with a rule combination that had not been made yet - thus expanding the MAP-Elites rule combination matrix in the database. Unfortunately, we forgot to ask users in the survey if they started with the random level option that was also provided by the AI assistance tool - so we lack data to report on this statistic.
For editing the level, 81.6\% of users reported editing a level completely by hand without any AI assistance. 27.6\% of users edited the level with help from either the evolver algorithm or the mutator functions provided by the AI assistance back-end. 19.7\% of users reported using the objective table to aid the evolver tool in creating the level. We think this low percentage is attributed the fact that a large population of users were unfamiliar with the system or `Baba is You' game overall. This - as well as the lack of selection for level comparison from the previously submitted levels in the database - made using the evolver tool towards certain goals too steep of a task to accomplish and learn. Finally, when testing the level, 59.2\% of users reported using the Keke solver AI when testing their levels and 72.4\% of users named their levels.
While not required in the tasks given, we also asked participants about any extra site features they chose to explore. 23.7\% of users reported submitting a level rating from the 'Rate' page. 51.3\% of users reported using the 'Search' tool to search for specific levels (what their search criteria was we did not ask.) Finally, 19.7\% of users reported using the 'Share Level' to share a submitted level link with others online.
The least used interactions - 'Started with a database-saved map in the level editor', 'Started with a level suggestion from the Unmade page', and 'Used the objectives table to evolve levels' - were also all related to the AI mixed-initiation of the system. The first could be attributed to a lack of overall levels in the database (at the start of the experiment there were only around 40 available levels) therefore leading to a lack of viable options for the user to choose from. However, the lack of usage for the other two features could be attributed to the opposite problem of having too many options to choose from - again due to lack of levels available to choose from in the database. Trying to make a level with constrained parameters may have also been too steep of a task to accomplish for someone who was totally unfamiliar with the system or even the `Baba is You' game overall. There was also no incentive for a player to create a level suggested by the system as opposed to making a level from scratch. We also didn't explicitly instruct users to make a level from the suggested set, and instead allowed them to make whatever level they wanted with the editor - whether with the prompted ruleset or from their own ideas.
\section{Discussion}
\subsection{Data Analysis}
It is clear from both the submitted level statistics of the site and the self-reported user survey that mixed-authorship is not the preference for users when designing levels. Many users would still prefer to have total control over their level design process from start to finish. For future work, we can look to limit user control and encourage more AI-assistance with the design process similar to the work done by Bhaumik et al.~\cite{bhaumik2021lode}.
The limitations of the AI back-end (both the evolver and solver) may be at fault for the lack of AI interaction. The mutator and evolver system are dependent on previously submitted levels and level ratings in order to ``learn'' how to effectively evolve levels towards high quality design. As a result, the assistant tool is always learning what makes a ``good'' level from human input. If there is a lack of available data for the tool to learn from, the AI will be unable to create quality levels - causing the user to less likely submit mixed-initiative co-created levels, and causing a negative feedback loop.
The fitness function defined for the evolver and mutator tool may be inadequate for level designing. It could produce a level that is deemed ``optimal'' in quality by its internal definition, but may actually be sub-par in quality for a human user. Another flaw in the AI-collaboration system, could be that the users lacked direct control on the evolver and mutator and attempting to use them in middle of creation might have been more problematic as it could destroy some of the level structures that the users were working on. Future work could remedy this problem by giving users various mutation "options" similar to the AI selections in RLBrush \cite{delarosa2021mixed} and Pitako. \cite{machado2019pitako} Finally, the `Keke' AI solver was also lacking in performance as a few participants mentioned that the solver was unable to solve their prototype levels that they themselves could end up solving in just a couple of moves. An improved AI solver would help with the level creation efficiency.
\subsection{User Comments and Feedback}
We gave the participants opportunities to provide open feedback about their experience using the site in order to gather more subjective data about their experience as well as collect suggestions for potential new features.
Almost no users experienced any technical difficulties or bugs that prevented them from using the site. The few that did mentioned formatting issues with site caused by their browser (i.e. icons too close together, loading the helper gifs, font colors.) However, one user mentioned that this issue may have been because they were using the site from their phone (we unfortunately did not provide users with instructions to complete the study on a desktop or laptop.) In the future, we will be sure to exhaustively test the site on as many browsers as possible - both desktop-based and mobile - to be more accessible.
Some users were confused by the tutorial and the amount of information it conveyed for the entire site citing it as ``intimidating'', ``overwhelming'', and ``a bit complex''. However, other users reported the lack of information saying it was ``not detailed'', or had ``sufficient information [...] but could have been delivered in a more comprehensible way.'' To make the game more accessible, we will most likely try to make the tutorial section less intimidating to new users by limiting the amount of information shown (possibly through a ``table of contents'' as suggested by one participant) while still being comprehensible enough to understand the level editor and tools.
For feature suggestions, many users wished for larger maps and vocabulary - like those found in the Steam-release `Baba is You' game. Users also wished for a save feature that would allow them to make ``drafts'' of their level to come back later to edit. Many users also suggested a co-operative multiplayer feature for level editing and level solving - we can assume with another human and not an AI agent.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{hor_survey_graphs/browser_v2.png}
\caption{User feedback for likelihood to return using the site after the experiment}
\label{fig:reuse_likelihood}
\end{figure}
While the results of the statistics on the levels submitted were disappointing for involvement of the AI assisting tool, we also asked users how likely they would continue using the site after the experiment. 38.2\% of users said they would continue to use the site, while 55.3\% said they would maybe use the site (figure~\ref{fig:reuse_likelihood}). Many users were optimistic and encouraging with the concept of incorporating AI and PCG technologies with level design - citing the project as a ``cool project'', ``a very unique experience'', a ``lovely game and experiment'', and ``very fun.'' At the time of writing, a few users did return, as their 'Keke' assigned usernames were shown as authors on the New page, long after the study was completed. Most notably, the Keke subject user Keke978 who took up the username 'Jme7' and contributed 28 more levels to the site after the study was concluded and currently holds the title for most levels submitted and most rule combinations on the site.
Many users also provided us with constructive feedback for feature implementation, site usability, and suggestions for improvement with how to further incorporate the AI back-end interactivity. As shown in figure~\ref{fig:exp_graph}, 70\% of users who played with the system had never played the game `Baba is You' and 75\% of people had never used an AI-assisted level editor tool before this experiment. Based on this information and retainability of users to complete the survey and provide the constructive feedback, we can extrapolate 2 conclusions: 1. the game stands alone, independent of `Baba is You', as an entertainment system; and 2. for people with even limited AI-gaming experience, as long as they are not completely foreign to gaming, this project has the ability to grasp their attention long enough to understand it, tinker around, and then give constructive feedback.
\section{Conclusion and Future Work}
The results from the user study have demonstrated both the benefits and limitations of a crowd-sourced mixed-initiative collaborative AI system. Currently, users still prefer to edit most of the content themselves, with minimal AI input - due to the lack of submitted content and ratings for the AI to learn from. Pretraining the AI system before incorporating it into the full system would be recommended to create more intelligent systems that can effectively collaborate with their human partners for designing and editing content. This would lead to more helpful suggestions on the evolver's end as well as better designed levels overall. This project is the start of a much longer and bigger investigation into the concept of crowd-sourced mixed initiative systems that can use quality diversity methods to produce content and we have many more ideas to improve upon the Baba is Y’all system.
As suggested by many participants in the user study, we would like to incorporate level design collaborations between multiple users and multiple types of evolutionary algorithms all at once to create levels. Our system would take inspiration from collaboration tools such as LodeEncoder \cite{bhaumik2021lode}, RLBrush \cite{delarosa2021mixed}, and Roblox (Roblox Corporation, 2006). This would broaden the scope and possibilities of level design and development even further to allow more creativity and evolutionary progress within the system. This collaboration setting will open multitude of interesting problems to investigate such as authorship.
Outside of the `Baba is You' game, we would like to propose the development of an open-source framework to allow mixed-initiative crowd-sourcing level design for any game or game clone. Such games could include Zelda, Pacman, Final Fantasy, Kirby, or any other game as long as we have a way to differentiate between levels mechanically and we can measure minimum viable quality of levels. Adding more games to the mixed-initiative framework would allow an easier barrier of entry to players who may have been unfamiliar with the independent game `Baba is You' but is very familiar with triple-A games produced by companies such as Nintendo.
We would like to also propose a competition for the online `Keke' solver algorithm for the challenging levels. In this competition, users would submit their own agent that can solve the user-made and artificially created `Baba is You' levels. Ideally, this improve the solver of the `Baba is Y'all' system but also introduce a novel agent capable of solving levels with dynamically changing content and rules - an area that has not been previously explored in the field. Development for this framework for this competition has already begun at the time of writing this paper.
Finally, we would like to propose the creation of a fully autonomous level generator and solver that can act as a user to our system.
This generator-solver pair would work parallel to the current system's mixed-initiative approach, but with a focus on coverage to exhaustively find and create levels for every combination of mechanics. With a redefined fitness function and updated solver (possibly from the Keke Solver Competition,) this could be more efficient than having users manually submit the levels, while still using content created by human users to maintain the mixed-initiative approach.
There are many new directions we can take the Baba is Y'all system and the concept of crowd-sourced collaborative mixed-initiative level design as a whole and this project will hopefully serve as a stepping stone into the area and provide insight on how AI and users can work together in a crowd-sourced website to generate new and creative content.
\section*{Acknowledgment}
The authors would like to thank the Game Innovation Lab, Rodrigo Canaan, Mike Cook, and Jack Buckley for their feedback on the site in its beta version as well as the numerous users who participated in the study and left feedback.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
| 61decaa1f993790b6b05ece2d83435830a9a74c1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Brazil reported COVID-19 first official case on February 25, 2020\cite{AgenciaBrasil2020_1}, in the city of São Paulo, which is the largest municipality in the country. Similar notifications in other Brazilian cities quickly followed, being accompanied by a devastating scenario of contagion and death \cite{Lovisolo2020,OF-COVID19-Relatorio30}, as shown in Fig.~\ref{fig:fig1}, which presents the prevalence of cases and deaths in the country until October 2021. The first official notification in Rio de Janeiro, the second largest and most visited city by tourists, occurred on March 6, 2020\cite{Cavalcante2020}, with the epidemiological situation quickly evolving to one of the worst scenarios at the national level, among the highest number of confirmed cases and deaths by COVID-19, as can be seen in the prevalence maps of Fig.~\ref{fig:fig2}.
\begin{figure}
\centering
\includegraphics[scale=0.2]{fig1_a.jpeg}
\vspace{2mm}
\includegraphics[scale=0.2]{fig1_b.jpeg}
\caption{Prevalence of COVID-19 cases and deaths confirmed in Brazil between March 2020 and October 2021 \cite{OF-COVID19-Relatorio30}.}
\label{fig:fig1}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.2]{fig2_a.jpeg}
\vspace{2mm}
\includegraphics[scale=0.2]{fig2_b.jpeg}
\caption{Prevalence of COVID-19 cases and deaths confirmed in Rio de Janeiro city between March 2020 and October 2021 \cite{OF-COVID19-Relatorio30}.}
\label{fig:fig2}
\end{figure}
Furthermore, Fig.~\ref{fig:fig3} shows the incidence of confirmed cases (top left) and deaths (top right) in Rio de Janeiro from January 2020 to December 2021, organized by first symptoms date and event date, respectively. It also shows a comparison between new notifications per week and total notifications for cases (bottom left) and deaths (bottom right), where periods of incidence proportional to prevalence (exponential growth) are noticed. In this figure, the raw data are represented in magenta dots and the 7 days moving average by the green curve. Multiple waves of contagion/deaths can be seen, the four most pronounced in April 2020, December 2020, April 2021, and August 2021, respectively. Small oscillations over time, which resembles an endemic period, can also be seen. In addition, the surveillance data for new confirmed cases (top left of Fig.~\ref{fig:fig3}) also highlights the existence of case notifications in January and February 2020, well before March 1st, the current date where it is already confirmed (through surveillance data) that community transmission had started. These cases, registered by the date reported for the first symptoms, were not counted at the beginning of March 2020 because their registration in the system only occurred a posteriori. For this reason, March 6th was announced that time as the day of the first confirmed case. Although these early records may be due to errors in the registration, due to the provision of false or misleading information by patients, the fact that they are not isolated notifications may suggest that community transmission of SARS-CoV-2 in Rio de Janeiro city may have started well before the period close to March 1st, where community transmission is known to have already occurred.
\begin{figure*}
\centering
\includegraphics[scale=0.45]{fig3_a.pdf}
\includegraphics[scale=0.45]{fig3_b.pdf}\\
\vspace{5mm}
\includegraphics[scale=0.45]{fig3_c.pdf}~~~~~~~~~
\includegraphics[scale=0.45]{fig3_d.pdf}~~~~~
\caption{Evolution of the incidence of confirmed cases (top left) and deaths (top right) of COVID-19 in Rio de Janeiro, from January 2020 to December 2021, organized by first symptoms date and event date, respectively. Also available is the comparison between new notifications per day and total notifications for cases (bottom left) and deaths (bottom right). The raw data are represented in magenta dots, and the 7 days moving average by the green curve. It can be noted the presence of multiple waves of contagion and deaths, where the four most pronounced peaks were in April 2020, December 2020, April 2021, and August 2021, respectively.}
\label{fig:fig3}
\end{figure*}
As at the beginning of the pandemic the Brazilian epidemiological surveillance system was not prepared to massively register COVID-19 cases, it is not only possible but probable that the introduction of the virus in the community took place in a period earlier than of the current estimates for the starting date of community transmission (March 1st). The early cases mentioned above provide direct, albeit weak (due to possible recording errors or biases), evidence in this sense.
The SARS-CoV-2 fluctuation in Rio de Janeiro, as well as the significant number of notifications before the first official confirmation, are peculiar epidemiological phenomena, which differ from the typical situation observed in the majority of Brazilian cities. Thus, this situation deserves to be further investigated, since understanding this multiple wave behavior can provide insights into different stages of the epidemic spread, information that can be useful to guide decision-makers in future outbreaks.
In this context, one piece of information that is particularly interesting is the likely start date of each community outbreak, i.e.,
when the transmission inside a community is due to the locals. These dates are important because they can indicate key events, such as the introduction period of a new viral strain in the population\cite{Voloch2021, resendeetal}, or a drastic change in social behavior that is capable of inciting a new phase of epidemic expansion in a given population.
Several mathematical approaches can be used to access the dynamics of an epidemic wave, including the estimation of their initial date. Compartmental models\cite{Brauer2008,Brauer2017,Martcheva2015} based on differential equations are very natural for this purpose, with the underling parametrization obtained with aid of data-assimilation techniques, such as Kalman filter\cite{Rajnesh2021p1}, nonlinear regression\cite{Kucharski2020,Magal2020p3040,Subhas2020p071101,Lobato2021nody}, Bayesian statistics\cite{Cotta2020p220,Lyra2020}, neural networks and other machine learning tools\cite{Magri2020,He2020,ALBANI2021p6088,Kuhl2021}, etc. Such compartmental models are also fundamental in approaches that employ the concept of complex networks to describe the epidemic dynamics in a large population with heterogeneous spatial distribution\cite{Chen2020,Costa2020p043306,Aleta2020p964,Aleta2020p157,Ventura2021}. It is also quite common to describe the evolution of epidemic waves with aid of purely phenomenological models, supported by frequentist\cite{Liu2020p1527,Vasconcelos2020,Pelinovsky2020p110241,Vasconcelos2021,Vasconcelos2021SR,Chol-jun2022,Pelinovsky2022p111699,Contoyiannis2022p043109,Contoyiannis2022p011103} or Bayesian\cite{Romadhon2021,Calatayud2022} statistical approaches.
There are already some works in the open literature concerned with estimating the start date of a local epidemic outbreak of COVID-19, such as Delatorre et al.\cite{DELATORRE}, which tries to infer the starting date of the SARS-CoV-2 spread in Western Europe and the Americas; Batista and Cunha~Jr\cite{Fernando_Batista}, that did the same exercise for the early stages of COVID-19 in Portugal; and Zhai et al.\cite{Zhai2021p013155}, which investigated the initial COVID-19 dynamics of 10 states in the U.S.A., New York City, United Kingdom, Italy, and Spain.
Despite the consistency of the results presented by the first two studies, by comparison with evidence (other epidemic observations) that support the estimated dates, they are more focused on reporting dates, having relatively few details about the underlying mathematical methodologies. The third one, on the other side, is rich in methodological detail, presenting a mechanistic framework based on a non-Markovian delayed compartmental model, which provides a relatively general tool for inferring the start date of an outbreak. However, in terms of validating the results, the authors' arguments implicitly rely heavily on the assumptions that the available data are sufficiently accurate and informative to calibrate the dynamic model, and that the latter provides a sufficiently realistic representation of the epidemic outbreak of interest. The scenario where at least one of these assumptions is not valid is not rare, so there is space in the literature for contributions that attack the same problem on other fronts.
Seeking to contribute to expanding the arsenal to estimate the initial date of an epidemic community outbreak, this paper discusses a generic parametric statistical approach for estimating the start date of an epidemic outbreak based on epidemiological surveillance data. The fundamental idea is to present a procedure that incorporates elements of generality, to be applied in typical epidemic outbreaks, but that is simple enough to be used by researchers who do not have strong training in epidemiology or more advanced statistical methods (e.g. Bayesian inference). In this way, we combine the surveillance data with algebraic multiple waves models, nonlinear regression, and information criteria to obtain a simple but representative mathematical model of the underlying outbreak, which provides an interval estimate for the start date of the referred contagion wave. The proposed methodology is illustrated with aid of COVID-19 data from Rio de Janeiro city, which has complex dynamics, with multiple contagion waves and (the typical) very irregular data.
The paper is organized as follows. Section~\ref{sec:methodology} describes the statistical framework. Section~\ref{sec:Results and discussion} presents the results of a case study in Rio de Janeiro. Finally, conclusions are shown in Section~\ref{sec:Conclusions}.
\section{Method}
\label{sec:methodology}
The methodology employed here combines epidemiological surveillance data, algebraic statistical models, nonlinear regression, and a model selection procedure to infer the systematic behavior of COVID-19 outbreaks, focusing on estimating a possible start date for each contagion wave. A schematic version of this framework can be seen in Fig.~\ref{fig:fig4}, and each of its steps is described below.
\begin{figure*}
\centering
\includegraphics[scale = 0.33]{fig4.pdf}
\caption{Statistical framework adopted to describe the COVID-19 outbreaks and infer the start period of each contagion wave. From epidemiological data, several statistical models are built with the aid of a linear regression process via Monte Carlo simulation. Information criteria are used to choose the most representative model(s), which is (are) used to provide an interval estimate for the start date of the COVID-19 contagion wave.}
\label{fig:fig4}
\end{figure*}
\subsection{Surveillance data}
To monitor the evolution of COVID-19 spread, time series that quantify the new reported cases and deaths may be used. Very often such time series are obtained by a process of selection and agglutination of raw data from a large spreadsheet, where each line corresponds to a patient (duly anonymized), containing many dozens of information, such as zip code, date of first symptoms, and date of outcome (recovery or death), etc. These data are registered by the municipality health authorities. In the case of Rio de Janeiro city, they come from two databases: (i) e-SUS VE; and (ii) SIVEP-Gripe. The first records the flu syndrome in ordinary cases, while the second is responsible for the severe acute respiratory syndromes records. Both are compiled (without duplication) by the Municipal Health Department, which makes them available on a website\cite{PainelRioCovid} widely accessible to the general public. Data for new cases are organized by the date of first symptoms, while new deaths are recorded by the date of the event. Other locations have their registration systems, but the vast majority of them have most of the features in common with the database described above.
Epidemiological data are notoriously problematic for statistical analysis purposes, as delay and underreporting effects are unavoidable, due to the physical impossibility of knowing all the cases that occur, neither in real-time nor with a few days delay\cite{Gamerman2022}. In practice, as the city of Rio de Janeiro has not adopted a massive random testing policy, only part of the cases of infection by COVID-19 were confirmed with the laboratory tests, essentially those that seek medical attention. Thus, the available data for cases evolution provide a distorted (biased) picture of the epidemiological reality, with far fewer cases reported than the actual number of infected, and with an estimated date for infection beginning shifted in time from the true onset, since the infection onset is estimated by the date of first symptoms (which occurs on average 5 days after the infection event). Death records also suffer from underreporting, but on a much smaller scale, as in this case registration is compulsory. Typically, deaths from COVID-19 that are not counted are those with no formal diagnosis of the disease by a laboratory test. In this setting, the time series for the number of new deaths per day can be considered a proxy to assess the evolution of contagion by the disease, since there is a clear correlation between it and the temporal evolution of the total number of cases (part of the cases evolves to death, more cases occurring, the expectation is more deaths, fewer cases, fewer deaths)\cite{Gamerman2022}.
In the analysis reported in this paper, only death data are considered. Since these data are recorded by the actual date of the event (in this case, the death), no delay is inserted into data, except that one related to the insertion of data into the system, which can take up to 30 days, but which only compromises the accuracy of the end of the time series, and can be corrected with the use of nowcasting techniques\cite{Gamerman2022}.
As the time series of cases and deaths are subject to much fluctuation, due to the natural variability of the disease, but also due to imperfections in the surveillance system, treating the data in advance is necessary to make them suitable for visual analysis, since extracting patterns from raw data is quite difficult. In this sense, these data are smoothed with the aid of 7 days moving average filter, to reduce excess fluctuation. As can be seen in Fig.~\ref{fig:fig3}, this treatment helps to display the data evolution pattern. But it should be mentioned that due to the use of only past data in the smoothing (for reasons of causality), the moving average inserts a delay of a few days in the data series. To avoid introducing this bias into statistical estimates, only raw data are considered in the statistical estimation processes presented below. Moving averages are used in this paper only to visually illustrate the trend of raw data.
\subsection{Ensemble of statistical models}
The three typical phases of an epidemic outbreak (expansion, transition, and exhaustion)\cite{brauer2001,Murray2002,Kuhl2021} usually are well described by logistic curves when the underlying population is relatively homogeneous, so that contamination comes out of a close interaction between two people. This scenario is assumed as a working hypothesis in this paper, in a way that several statistical models, based on the algebraic solution of a logistic differential equation, are used here to represent the multiple outbreaks of COVID-19 that are characterized by a dataset that collects deaths records for a certain locality into a time series. This is inspired by the approach adopted by Batista and Cunha~Jr\cite{Fernando_Batista} to study the early stages of COVID-19 in Portugal.
It means the we have the statistical
\subsubsection*{Single wave model}
When just a single wave of contagion is of interest, the Verhulst logistic model \cite{Martcheva2015,Brauer2008,Murray2002,brauer2001} is employed. This model assumes a growth rate proportional to the disease prevalence in the population at time $t$, denoted by $C(t)$, but with a constant of proportionality that changes in time, so that it reduces when $C(t)$ grows, until reaching a maximum sustainable population in the limit when $t \to \infty$. This logistic model is defined by the differential equation
\begin{equation}\label{log02}
\dfrac{dC}{dt} = r \, C \, \left(1 - \dfrac{C}{K}\right) \, ,
\end{equation}
which has as solution the classical logistic curve
\begin{equation}\label{Ct}
C(t) = \displaystyle \frac{K}{1+e^{\displaystyle -r \, (t-\tau)}} \, ,
\end{equation}
where $r$ represents the infection growth rate, $K$ is the final number of notifications at the end of the outbreak, and $\tau$ describes the (initial condition dependent) instant of inflection associated to this curve, when the exponential growth ends, and a sudden deceleration begins. The derivative of $C(t)$, dubbed the incidence curve $I(t)$, represents the number of new death notifications per day, and is given by
\begin{equation}
I(t) = \dfrac{dC}{dt} = \frac{r \, K \, e^{\displaystyle -r\, (t-\tau)}}{\left(1 + e^{\displaystyle -r\,(t-\tau)} \right)^2} \, .
\label{It}
\end{equation}
Other authors use generalized logistic curves to deal with this kind of epidemic data \cite{Vasconcelos2020,Fernando_Batista,Zou2020p1}. Despite this being a possibility, the present work opted for the classic logistic curve for the sake of parsimony, as it is capable of providing a reasonable representation of Rio de Janeiro outbreaks.
\subsubsection*{Multiple waves model}
When more than one wave of contagion matters, a multimodal epidemic curve with $N$ peaks is considered, which demands the estimation of $3N$ parameters
\begin{equation}
\theta = (K_1,\, r_1,\, \tau_1,\, \cdots , \, K_N,\, r_N,\, \tau_N) \, ,
\label{theta_eq}
\end{equation}
resulting in a global prevalence curve
\begin{equation}
C(t) = \displaystyle\sum_{i=1}^{N} \frac{K_i}{1+e^{\displaystyle -r_i \, (t-\tau_i)}} \, ,
\label{sumCN}
\end{equation}
which results, by differentiation, in the global incidence curve
\begin{equation}
I(t) = \displaystyle\sum_{i=1}^{N} \displaystyle \frac{r_i \, K_i \, e^{\displaystyle -r_i \, (t-\tau_i)}}{ \left(\displaystyle 1+e^{\displaystyle -r_i \, (t-\tau_i)} \right)^2} \, .
\label{INt}
\end{equation}
\subsubsection*{Calibration procedure}
The calibration process of each statistical model consists of identifying the underlying parameters with the aid of new deaths per day time series $(I_1, \cdots, I_n)$, that correspond to the discrete-time instants $(t_1, \cdots, t_n)$. This task requires the minimization of the \textit{Roots Mean Squared Error} (RMSE)
\begin{equation}
RMSE = \sqrt{\dfrac{1}{n}\sum_{j = 1}^{n} \left(I_{j} - I(t_{j} \,) \right)^2} \,,
\label{eq:RMSE}
\end{equation}
that measures the ``discrepancy'' between data values and the corresponding predictions given by the model that calculates the incidence $I(t)$. In practice, RMSE represents the weighted average of the square of the residuals generated by the calibrated model\cite{Wasserman2004,Hastie2009}.
The RMSE numerical minimization procedure employs a Trust Region algorithm\cite{Nocedal2006,Bonnans2009}, where bounds for the parameter value, as well as an initial guess, are specified. Such bounds are determined based on numerical experimentation followed by visual inspection of the fitted curves. To minimize the dependence of the fitted curve with the prescribed value for the initial guess, Monte Carlo simulation\cite{kroese2011,cunhajr2014p1355} is used, where random values for the initial guess (within the admissible region) are drawn and used to generate a fitting curve. Among all the obtained fittings, the one with the smallest RMSE is chosen.
To assess the quality of the statistical fit, this work also considers the coefficient of determination
\begin{equation}
R^2 = 1 - \dfrac{\sum_{j = 1}^{n}\left( I_{j} - I(t_{j}) \,\right )^2 }{\sum_{j = 1}^{n} \left( I_{j} - \bar{I} \, \right)^2} \,,
\label{Rsq}
\end{equation}
where the time series $(I_1, \cdots, I_n)$ average is given by
\begin{equation*}
\bar{I} = \dfrac{1}{n}\sum_{j = 1}^{n} I_{j} \, .
\end{equation*}
The last metric describes how much of the total variance generated by the observed data is explained by the calibrated model. The value of $R^2$ varies between 0 and 1, so that the closer to 1 the more significant the model is, as the value of $R^2$ is the proportion of the original variability of the data that is explained\cite{Bishop2006,Hastie2009}.
\subsection{Model selection}
Different models can be fitted to the same dataset, choosing which is the most suitable requires the use of rational criteria to avoid potential bias. Thus, two information criteria are used here to choose the most suitable model following a rational criterion, which seeks to balance simplicity and predictability between the models. They are: (i) Akaike information criterion (AIC); (ii) Bayesian information criterion (BIC)\cite{Bishop2006,Hastie2009,Brunton2019}.
The AIC metric is given by
\begin{equation}
AIC = -2 \, \log{ L(\hat{\theta})} + 2 \, p \,,
\label{AIC}
\end{equation}
where $p$ is the number of model parameters, $\hat{\theta}$ the estimated parameters vector which maximizes the likelihood function $L(\theta)$. On the other hand, the BIC metric is defined as
\begin{equation}
BIC = - 2 \ \log{ L(\hat{\theta})} + p \, \log{n} \,,
\label{BIC}
\end{equation}
where $n$ is the number of observations. The model with the lowest AIC and BIC values is considered the best fitting model. Eventually these metrics can be combined with facts (documented in data) to assist in choosing the model.
\subsection{Outbreak start dates}
The estimate of the epidemic outbreak start date is not done in a punctual way, because an exact date in a context full of uncertainties such as an epidemic outbreak is a fragile estimate, and meaningless from any modeling point of view that makes sense. Thus, an interval estimate is considered in this work, where the upper limit of an admissible interval for the beginning of the outbreak is estimated with the help of the confidence band that encompasses the uncertainty of the statistical model. In this way, after the construction of a statistical model that fits the data, a 95\% confidence interval is obtained around the model's response curve. Such a prediction interval\cite{Wasserman2004}, with $(1-\alpha) \times 100\%$ confidence, is defined by
\begin{equation}
I(t) \pm z_{\alpha/2} \, \xi_{n} \,,
\label{ICzj2}
\end{equation}
where $z_{\alpha/2}$ is the quantile of the Student-$\text{t}$ distribution with $n$ degrees of freedom and statistical significance $\alpha = 0.05$, while
\begin{equation}
\xi_{n} = S \, \sqrt{1 + \frac{1}{n} \, \frac{\sum_{j = 1}^{n}(I_{j} - I(t)\,)^2}{\sum_{j = 1}^{n}(I_{j} - \bar{I} \,)^2}} \,,
\label{ICzj2}
\end{equation}
with
\begin{equation}
S = \frac{\sum_{j=1}^n (I(t_j)-I_j)^2}{n-2} \, .
\end{equation}
The intersection between the boundaries of this envelope and the time axis provides a range of possible dates for the start of the outbreak under investigation so that the rightmost point of this intersection is assumed to be an upper bound for the outbreak start date.
\section{Results and discussion}
\label{sec:Results and discussion}
In this section, our objective is to obtain an interval estimation for the starting dates of each COVID-19 outbreak in Rio de Janeiro city. The start date in this context is understood to be the day on which a certain level of prevalence is achieved so that new infection events occur every day from that point forward. This starting date may be related to the introduction of a novel viral strain in the community or to a key event that initiated a new chain of contagion.
\subsection*{The starting of SARS-CoV-2 community transmission}
The data shown in Fig.~\ref{fig:fig3} allows us to observe the existence of several waves of contagion in Rio de Janeiro city, which resulted in 6 waves of deaths, four big explosions, and two small boosts. In this first analysis, only the first of these waves is considered.
Once the exact date where one wave ends and another begins is extremely uncertain, makes more sense to speak about this event in an interval sense, a plausible period where that date is contained. In this way, by visual inspection of Fig.~\ref{fig:fig3}, it is verified that the first wave of deaths starts between the middle of March and the beginning of April, and ends by the end of June or the beginning of July 2020. Between March and April 2020, the epidemic surveillance system in Rio de Janeiro was still adapting to the pandemic, testing to determine if death from respiratory disease was due to COVID-19 was not yet mandatory, so that underreporting in this period (certainly above average) can bias the estimate of the outbreak start date. In May 2020, on the other side, the COVID-19 death tracking system was better developed, so the first wave phase records should not be so biased. Thus, for the sake of minimizing bias during statistical model calibration, the training data considered here uses the time series of new deaths per day between May 1st and July 1st.
Table~\ref{tab1} shows the values of the parameters found in the calibration process for a single wave model, the estimated date for the epidemic outbreak start, as well as the quality metrics of the fitting and information criteria, for several scenarios of the parameter $\tau$, which represents the instant of inflection of the logistic model (the peak of the incidence curve). These multiple scenarios are considered because the evolution of a patient who dies, between the time of infection and death, has a variable duration. Therefore, considering a range of possible dates around the peak of the data (April 28, day 127 of the time series) is a strategy to reduce the influence of this uncertainty on the inference process. By generating a family of possible curves (ensemble of models), the adopted information criteria (AIC and BIC) and the known facts about the outbreak will indicate which of those curves are a plausible representation for the wave of deaths.
\begin{table*}
\centering
\caption{{\bf Estimated parameters, the respective confidence intervals, fitting metrics, and information criteria measures for the statistical models used to represent the first wave of deaths by COVID-19 in Rio de Janeiro (for several scenarios of peak day).
}}
\label{tab1}
\vspace{2mm}
\begin{tabular}{c|c|c|c|c|cc|cc}
\toprule
$\tau$ & $K \times 10^3$ & $r \times 10^{-3} $ & deaths start date & cases start date & RMSE & R$^2$ & AIC & BIC\\
(day) & (people) & (day$^{-1}$) & upper bound & upper bound & & & & \\
\midrule
119 & 9.70 (9.14, 10.26) & 53 (49, 57) & Mar 13, 2020 & Mar 02, 2020 & 13.8 & 0.86 & 3.99 & 8.24\\
120 & 9.70 (9.18, 10.21) & 53 (49, 57) & Mar 12, 2020 & Mar 01, 2020 & 12.7 & 0.89 & 3.99 & 8.24\\
121 & 9.70 (9.21, 10.18) & 53 (49, 56) & Mar 11, 2020 & Feb 29, 2020 & 11.8 & 0.90 & 3.99 & 8.24\\
122 & 9.70 (9.23, 10.16) & 53 (50, 56) & Mar 11, 2020 & Feb 29, 2020 & 11.1 & 0.91 & 4.19 & 8.44\\
123 & 9.70 (9.25, 10.15) & 52 (49, 55) & Mar 11, 2020 & Feb 29, 2020 & 10.7 & 0.92 & 4.04 & 8.30\\
125 & 9.28 (8.84, ~9.73) & 55 (51, 57) & Mar 15, 2020 & Mar 04, 2020 & 10.8 & 0.92 & 3.98 & 8.24\\
127 & 8.72 (8.28, ~9.15) & 57 (53, 60) & Mar 20, 2020 & Mar 09, 2020 & 11.0 & 0.91 & 4.01 & 8.27\\
129 & 8.18 (7.74, ~8.62) & 60 (56, 64) & Mar 26, 2020 & Mar 15, 2020 & 11.4 & 0.91 & 4.00 & 8.25\\
131 & 7.68 (7.22, ~8.13) & 63 (59, 68) & Mar 31, 2020 & Mar 20, 2020 & 12.2 & 0.89 & 4.00 & 8.26\\
\bottomrule
\end{tabular}
\end{table*}
Note that for each model corresponds to a date, which is the estimate of the upper limit for the beginning of the wave of deaths. Due to the correlation between cases and deaths, there is a temporal shift (with a certain probability distribution) between this limit date on the curve of deaths and its counterpart on the curve of cases. Hawryluk et al.\cite{Hawryluk2020} points out that the COVID-19 average time from symptom onset to death in Brazil, during the first wave, was around 15.2 (11.2, 17.8) days, so that a conservative estimate for the upper bound for the date of onset of cases (fifth column of Table~\ref{tab1}) can be obtained by subtracting 11 days from the date estimated by the curve of deaths (fourth column of Table~\ref{tab1}).
\begin{figure*}[ht!]
\centering
\includegraphics[scale = 0.3]{fig5_a.pdf}
\includegraphics[scale = 0.3]{fig5_b.pdf}
\includegraphics[scale = 0.3]{fig5_c.pdf}\\
\vspace{5mm}
\includegraphics[scale = 0.3]{fig5_d.pdf}
\includegraphics[scale = 0.3]{fig5_e.pdf}
\includegraphics[scale = 0.3]{fig5_f.pdf}\\
\vspace{5mm}
\includegraphics[scale = 0.3]{fig5_g.pdf}
\includegraphics[scale = 0.3]{fig5_h.pdf}
\includegraphics[scale = 0.3]{fig5_i.pdf}
\caption{Logistic curves fitted to the time series associated with the incidence of deaths during the first wave of COVID-19 in Rio de Janeiro city, and the corresponding prediction band, for several scenarios of peak day: 119 (top left), 120 (top center), 121 (top right), 122 (middle left), 123 (middle center), 125 (middle right), 127 (bottom left), 129 (bottom center), 131 (bottom right). The dashed vertical lines indicate the upper bound estimation for the starting date of the first epidemic wave. Data used in the model calibration: May 1st to July 1st, 2020.}
\label{fig:fig5}
\end{figure*}
\begin{figure*}[ht!]
\centering
\includegraphics[scale = 0.3]{fig6_a.pdf}
\includegraphics[scale = 0.3]{fig6_b.pdf}
\includegraphics[scale = 0.3]{fig6_c.pdf}\\
\vspace{5mm}
\includegraphics[scale = 0.3]{fig6_d.pdf}
\includegraphics[scale = 0.3]{fig6_e.pdf}
\includegraphics[scale = 0.3]{fig6_f.pdf}\\
\vspace{5mm}
\includegraphics[scale = 0.3]{fig6_g.pdf}
\includegraphics[scale = 0.3]{fig6_h.pdf}
\includegraphics[scale = 0.3]{fig6_i.pdf}\\
\caption{Logistic curves fitted to the time series associated with the prevalence of deaths during the first wave of COVID-19 in Rio de Janeiro city, and the corresponding prediction band, for several scenarios of peak day: 119 (top left), 120 (top center), 121 (top right), 122 (middle left), 123 (middle center), 125 (middle right), 127 (bottom left), 129 (bottom center), 131 (bottom right). Data used in the model calibration: May 1st to July 1st, 2020.}
\label{fig:fig6}
\end{figure*}
Furthermore, the fitting metrics of Table~\ref{tab1} suggest that all the calibrated statistical models have good adherence to the data, which can be visually confirmed in Fig.~\ref{fig:fig5}, which shows 9 candidate models for the single wave incidence curve, each one corresponding to a different value for $\tau$. The corresponding prevalence curves can be seen in the sequence, in Fig.~\ref{fig:fig6}. In these figures, the raw epidemic data is represented by magenta dots, the 7 days moving average by a thin green line, the fitted epidemic curve is given by a thick blue line, and the 95\% confidence band is shown in gray.
In qualitative terms, all the candidate models provide a good description of the descending part of the wave, and they all more or less agree on the inference of the "hidden part" of the wave, i.e., that part of the outbreak which was not captured by the epidemic surveillance system on the early stages of the epidemic process. It is also noted that models with a lower value of $\tau$ (inflection point furthest from the peak of data) tend to estimate underreporting in death records more conservatively, as their prevalence curves are more detached from the observations (known to be underreported). However, despite the qualitative similarity between all candidate models, from a quantitative point of view, differences exist and materialize themselves in the estimation of the upper bound for the starting dates. These quantitative differences in date predictions are compared with known information about the onset of community transmission, to eliminate models that provide predictions inconsistent with observations. This can be seen as a complementary step in the model selection procedure, to delineate the most plausible model(s) for the available data.
The surveillance data used to construct the Fig.~\ref{fig:fig3} presents sparse notifications for cases in January and February, and no null records from March 1st. This is a direct indication that at the beginning of March there was already community transmission. Thus, by inconsistency with direct evidence, we will eliminate all models that point to a date greater than March 1st as the upper bound for the wave of cases start date, which leaves us with four candidate models, corresponding to $\tau \in \{120, 121, 122, 123 \}$.
Among these four models, $\tau = 120$ and $121$ have the lowest values for AIC and BIC, being in principle the natural choices for the most representative model. And between these two candidate models, $\tau =121$ has the lowest RMSE and highest $R^2$, so it is the optimal choice in light of fitting metrics and model selection criteria. This model indicates \underline{\emph{February 29, 2020}} as an upper bound for the first wave of cases starting date.
\subsection*{Consistency on the inference of first wave starting date}
It is very clear that the comparison with the evidence of the beginning of community transmission provides a good filter to discard models that overestimate the desired date, but does not provide information about the other models. This role is in charge of the model selection criteria, which show a marginally small difference in their numerical values for all the remnants candidate models, which may indicate that all the models under evaluation are more or less equally capable of describing the first epidemic wave. The fitting metrics RMSE and $R^2$ can be used as a supplement, in this case, the drawback is that, in the case of overfitting, they may provide very poor indicators of the model quality. Thus, in a scenario of metrics, as indicated in Table~\ref{tab1}, additional considerations to indicate the consistency of the statistical fit are desirable.
Following this idea, a combination (arithmetic mean) of the forecast obtained by the four candidate models is used as an estimator for the starting date of cases, providing again \underline{\emph{February 29, 2020}} as a possible upper limit for the beginning of community transmission of COVID-19 in the city of Rio de Janeiro. Once this is an interval estimate, it is possible (and probable) that the threshold of dozens of cases in the prevalence of the disease has occurred sometime in February, instead of March as current data suggest.
To strengthen this conclusion we use a hypothesis test, where the null and alternative hypotheses adopted are $$H_0: \mbox{cases start date} \geq \mbox{March 1, 2020} \, ,$$ and $$H_1: \mbox{cases start date} < \mbox{March 1, 2020} \, ,$$ respectively. Only the estimated upper bound dates that were not discarded in the comparison with the community transmission data are considered in the test, providing a piece of strong evidence against the null hypothesis, with a p-value 0.0288.
This result is indirect evidence, coming from a prediction with a mathematical model, in favor of the thesis that the virus was already circulating in the city before March 2020, which adds to the direct (but weak) evidence that appears in the record of cases by first symptoms.
\subsection*{The appearance of other waves of contagion}
The exact periods when the other waves of contagion begin are so uncertain as that one of the first wave. Thus, periods that contain the other waves are determined by visual inspection as follows: Jun-Nov 2020 (second wave); Nov 2020-Mar 2021 (third wave); Mar-May 2021 (fourth wave); May-Jul 2021 (fifth wave); Jul-Dec 2021 (sixth wave). Logistic curves that fit training data, determined within each of these time windows, are fitted following a procedure similar to that adopted in the analysis of the first wave.
The results of these model calibrations can be seen in Table~\ref{tab2}, which shows for the 2nd to the 6th epidemic waves, and several peak date scenarios, information related to identified parameters, wave starting dates, fitting metric, and model selection criteria. Selecting the models as described in the first wave, we obtain the following models as the most representative: $\tau = 268$ (second wave); $\tau = 355$ (third wave); $\tau = 465$ (fourth wave); $\tau = 505$ (fifth wave); $\tau = 608$ (sixth wave).
The incidence curves that correspond to the best models of each wave can be seen in Fig.~\ref{fig:fig7}. In all these cases, the incidence curves show good adherence to the data, so that the respective models provide a reasonable description of the epidemic behavior underlying each analyzed period. Using these models (in the same way as was done in the first wave), we have as upper bounds for the start dates of the other epidemic waves: July 17, 2020 (second wave); November 6, 2020 (third wave); March 2, 2021 (fourth wave); March 16, 2021 (fifth wave); July 9, 2021 (sixth wave).
Using the parameters identified in each of the six waves above as the initial guess of a regression process that seeks to calibrate a model with 6 waves, Eq.(\ref{INt}), we obtain the incidence curve and prediction band that are shown in Fig.~\ref{fig:fig8}, which also shows the start dates that were estimated above. This figure clarifies a point that may have left doubts in the reader, about the temporal proximity between the fourth and fifth waves. The sum of the two can explain the asymmetry observed in the fall of the fourth wave, which is interpreted in the light of this model as being the superposition of two contagion waves that emerged temporally close. This is a plausible scenario, for example, if there are two key events in the city that are triggers for scaling up contagion events. Or if two regions of the city, geographically not close, experience an increase in contagion almost simultaneously.
\begin{table*}
\centering
\caption{{\bf Estimated parameters, the respective confidence intervals, fitting metrics, and information criteria measures for the statistical models used to represent the 2nd to 6th waves of deaths by COVID-19 in Rio de Janeiro (for several scenarios of peak day).
}}
\label{tab2}
\vspace{2mm}
\begin{tabular}{c|c|c|c|c|c|cc|cc}
\toprule
& $\tau$ & $K \times 10^3$ & $r \times 10^{-3} $ & deaths start date & cases start date & RMSE & R$^2$ & AIC & BIC\\
& (day) & (people) & (day$^{-1}$) & upper bound & upper bound & & & & \\
\midrule
\multirow{4}{*}{2nd wave}
& 264 & 4.30 (3.63, 4.70) & 28 (23, 34) & Jul 16, 2020 & Jul 05, 2020 & 5.37 & 0.27 & 124.7 & 129.73\\
& 266 & 4.36 (3.69, 5.03) & 28 (23, 33) & Jul 17, 2020 & Jul 06, 2020 & 5.34 & 0.27 & 88.2 & 93.30\\
& 268 & 4.48 (3.78, 5.18) & 27 (22, 32) & Jul 17, 2020 & Jul 06, 2020 & 5.35 & 0.27 & 4.3 & 9.36\\
& 270 & 4.67 (3.90, 5.43) & 26 (21, 31) & Jul 16, 2020 & Jul 05, 2020 & 5.40 & 0.26 & 22.8 & 27.84\\
\midrule
\multirow{4}{*}{3rd wave}
& 349 & 6.85 (6.22, 7.48) & 47 (41, 53) & Nov 06, 2020 & Oct 26, 2020 & 15.5 & 0.52 & 4.02 & 9.09\\
& 351 & 6.66 (6.15, 7.17) & 50 (45, 55) & Nov 06, 2020 & Oct 26, 2020 & 13.7 & 0.62 & 4.00 & 9.07\\
& 353 & 6.57 (6.14, 7.00) & 47 (47, 56) & Nov 06, 2020 & Oct 26, 2020 & 12.2 & 0.70 & 4.00 & 9.07\\
& 355 & 6.54 (6.16, 6.92) & 48 (48, 56) & Nov 06, 2020 & Oct 26, 2020 & 11.0 & 0.76 & 4.02 & 9.11\\
\midrule
\multirow{4}{*}{4th wave}
& 461 & 6.01 (5.56, 6.47) & 70 (63, 77) & Mar 03, 2021 & Feb 20, 2021 & 13.2 & 0.75 & 4.00 & 8.25\\
& 463 & 6.12 (5.73, 6.52) & 69 (63, 74) & Mar 02, 2021 & Feb 19, 2021 & 11.3 & 0.82 & 4.00 & 8.25\\
& 465 & 6.36 (5.95, 6.77) & 65 (60, 71) & Mar 02, 2021 & Feb 19, 2021 & 10.9 & 0.83 & 4.00 & 8.25\\
& 467 & 6.73 (5.25, 7.21) & 61 (55, 66) & Mar 02, 2021 & Feb 19, 2021 & 11.6 & 0.81 & 4.00 & 8.26\\
\midrule
\multirow{4}{*}{5th wave}
& 504 & 8.11 (6.78, 9.44) & 32 (26, 39) & Mar 14, 2021 & Mar 03, 2021 & 09.4 & 0.55 & 68.88 & 73.13\\
& 505 & 8.11 (6.65, 9.57) & 32 (25, 38) & Mar 16, 2021 & Mar 05, 2021 & 09.7 & 0.52 & 11.52 & 15.78\\
& 506 & 8.11 (6.51, 9.72) & 31 (24, 39) & Mar 19, 2021 & Mar 08, 2021 & 10.0 & 0.48 & 16.91 & 21.16\\
& 507 & 8.11 (6.35, 8.87) & 31 (23, 39) & Mar 21, 2021 & Mar 10, 2021 & 10.3 & 0.45 & 37.77 & 42.02\\
\midrule
\multirow{4}{*}{6th wave}
& 600 & 4.86 (4.62, 5.10) & 49 (46, 52) & Jul 05, 2021 & Jun 24, 2021 & 6.3 & 0.92 & 4.27 & 9.89\\
& 602 & 5.06 (4.81, 5.31) & 47 (44, 50) & Jul 02, 2021 & Jun 21, 2021 & 6.5 & 0.91 & 4.33 & 9.95\\
& 604 & 4.86 (4.63, 5.10) & 49 (46, 52) & Jul 05, 2021 & Jun 24, 2021 & 6.3 & 0.92 & 4.27 & 9.89\\
& 608 & 4.70 (4.48, 4.92) & 51 (48, 54) & Jul 09, 2021 & Jun 28, 2021 & 6.3 & 0.92 & 4.07 & 9.69\\
\bottomrule
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale = 0.3]{fig7_a.pdf}
\includegraphics[scale = 0.3]{fig7_b.pdf}
\vspace{5mm}
\includegraphics[scale = 0.3]{fig7_c.pdf}
\includegraphics[scale = 0.3]{fig7_d.pdf}
\includegraphics[scale = 0.3]{fig7_e.pdf}
\caption{Logistic curves fitted to the time series associated with COVID-19 incidence of deaths in Rio de Janeiro city, and the corresponding prediction bands, for several contagion waves. The dashed vertical lines indicate the upper bound estimation for the starting dates of the epidemic waves. Second wave (top) data: August 1st to November 1st, 2020; Third wave (middle top) data: November 1st, 2020 to February 1st, 2021; Fourth wave (middle middle) data: March 1st to May 1st, 2021; Fifth wave (middle bottom) data: May 1st to July 1st, 2021; and Sixth wave (bottom) data: August 1st to December 1st, 2021.}
\label{fig:fig7}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale = 0.5]{fig8.pdf}
\caption{Logistic curve with 6-waves fitted to the time series associated with COVID-19 incidence of deaths in Rio de Janeiro city, and the corresponding prediction band. The dashed vertical lines indicate the upper bound estimations for the starting dates of the epidemic waves. Training data: April 1st, 2020 to December 31st, 2022.}
\label{fig:fig8}
\end{figure*}
\subsection*{Correlation between starting dates and key events}
It is worth remembering at this point that the Carnival festivities of that year took place between the 21st and 26th of February (Friday to Wednesday) 2020. This temporal proximity between the plausible period for the beginning of community transmission (February 29 or before) and the Carnival is a shred of evidence in favor of the hypothesis widely circulated among epidemiologists\cite{Ashktorab2021} that the Carnival celebrations in the streets were a critical event for the spread of COVID-19 in Rio de Janeiro. Correlations with other key events, such as business reopenings, extended holidays, festive dates, etc., can be similarly investigated. Due to space limitations, we do not present other investigations of this type in this paper.
\section{Conclusions}
\label{sec:Conclusions}
The dynamics of COVID-19 may have multiple waves of contagion in the same geographic location as a side effect. Due to the complex and not fully understood nature of the disease transmission process, specifying the exact moment of emergence of new community contagion outbreaks is a hard task. This work presented a statistical approach capable of describing the multiple waves of contagion observed, as well as providing an interval estimation of the possible start date for each outbreak.
This methodology was illustrated with the COVID-19 data of Rio de Janeiro city and determined possible starting dates for the multiple outbreaks during the years 2020 and 2021. The results obtained show that logistic models with one or more waves can be used to provide mathematical descriptions, with good adherence to the data, in epidemiological scenarios with complex transmission dynamics. In addition, they support that SARS-CoV-2 probably started to be disseminated locally in Rio de Janeiro as early as February 2020.
\section*{Supplementary Material}
The supplementary material is composed of the data and codes used to calibrate the statistical models, which are also are available in the repository \url{https://github.com/americocunhajr/COVID19Waves}
\begin{acknowledgments}
The authors thank Mrs. Gabrielle Pereira and \emph{\mbox{COVID-19:} Observatório Fluminense} initiative for preparing Figures~\ref{fig:fig1} and \ref{fig:fig2}. This work was supported by Funda\c c\~ao de Amparo \`a Pesquisa do Estado de S\~ao Paulo, FAPESP (process: 2015/50122-0); Funda\c c\~ao Carlos Chagas Fillho de Amparo \`a Pesquisa do Estado do Rio de Janeiro, FAPERJ (processes: 211.037/2019 and 201.294/2021); Coordenação de Aperfei\c coamento de Pessoal de N\'ivel Superior, CAPES (process: 88887.506931/2020-00), and Conselho Nacional de Desenvolvimento Cient\'ifico e Tecnol\'ogico , CNPq (process: 441016/2020-0).
\end{acknowledgments}
\section*{Data Availability Statement}
\label{sec:DataStatement}
The data that support the findings of this study are available within the article and its supplementary material.
\section*{References}
| 216c49d3c3c01d1cb048d02f4147f02ec75e920c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}
Following the pioneering work of Shannon on information theory, Kullback and Leibler defined a divergence of two probability measures $P$ and $Q$ \cite{kullback-leibler-1951} \[D_{KL}(P||Q)= \sum_{x} P(x) \log \frac{P(x)}{Q(x)} \text{ with the convention } 0 \cdot \infty =0.\] This would later come to be known as the Kullback-Leibler divergence of probability measures. Later, R\'enyi generalised the notion of Kullback-Leibler divergence \cite{renyi1961}. For $0< \alpha \not = 1$, R\'enyi defined
\[
I_\alpha (P||Q)= \frac{1}{\alpha -1} \log \sum_x P(x)^\alpha Q(x)^{1-\alpha}.
\]
Quantizing these developments Umegaki in \cite{umegaki-1962} extended these notions to the setting of von Neumann algebras. If
$\mathcal{M}$ be a von Neumann algebra with a faithful normal trace $\tau$, and $\rho, \sigma$ are elements of the noncommutative $L^1(\mathcal{M},\tau)$ space, Umegaki defined
\[
I(\rho||\sigma)= \left\{
\begin{array}{ll}
\tau (\rho \log \rho - \rho \log \sigma), &
\text{ if } \text{supp}\, \rho \subseteq \text{supp}\, \sigma;\\
\infty, & \text{ otherwise}.
\end{array}\right.
\] Araki defined the relative entropy between two states of a von Neumann algebra using the theory of relative modular operators in
\cite{araki1976relative} and \cite{araki1977relative} (See Appendix \ref{appendix:relative-modular} and Section \ref{sec:counter-eg}). Similar to the generalisation R\'enyi obtained for the Kullback-Leibler divergence, Petz and Ohya \cite{petz1986quasientropy, petz-ohya-1993} generalised the notion of Araki's relative entropy to what is known today as Petz-R\'enyi $\alpha$-relative entropy. The Petz-R\'enyi $\alpha$-relative entropy of two density operators (positive trace class operators with unit trace, also known as states) $\rho$ and $\sigma$ is defined as \[D_{\alpha} (\rho \lvert\lvert \sigma) = \frac{1}{\alpha-1}\log \tr \rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}, \alpha \in (0,1)\cup (1,\infty). \] Datta in \cite{Datta2009} studied an entropic quantity called $D_{\operatorname{min}}(\rho||\sigma)$ in the finite dimensional setting, and verified that $D_{\operatorname{min}}(\rho||\sigma) = \lim_{\alpha\rightarrow0}D_{\alpha}(\rho||\sigma)$ \cite[equation (10)]{Datta2009}. Recently, Seshadreeshan, Lami and Wilde in \cite{Ses-Lam-Wil-2018} and Parthasarathy in \cite{Par2021a} studied the Petz-R\'enyi relative entropy in the infinite dimesnional setting, for the particular case of gaussian states. A related entropic quantity called sandwiched R\'enyi relative entropy gained prominence in the recent years due to several nice properties satisfied by them, \cite{muller-tomamichel-etal-2013,frank-leib-2013, wilde-winter-yang-2014, Datta_Leditzky_2014, Par2021b}. It should be noted that all of these entropic quantities are important in quantum information theory in general and quantum hypothesis testing in particular. One may refer to \cite{petz-hiai-1991} and the other articles cited above to know more about different uses of relative entropy.
Our aim in this article is to prove that the Petz-R\'enyi relative entropy enjoys several important properties even in the infinite dimensional setting which is not well understood in the literature.
Some important results obtained in this article are: \begin{enumerate}
\item \textit{A new formula:} For $\alpha\in (0,1)\cup (1,\infty)$, a formula to compute the Petz-R\'enyi $\alpha$-relative entropy between states $\rho$ and $\sigma$, in terms of certain classical probability measures on the positive integer lattice of the plane $\mathbb{R}^2$ (Theorems \ref{thm:the-formula-general}, and \ref{thm:quantum-reduces-to-classical}).
\item \textit{Positivity and Monotonicity:} $D_{\alpha}(\rho||\sigma)\geq 0$ and is nondecreasing in $\alpha$ (Theorems \ref{thm:3-veh-quantum} and \ref{thm:8-veh-quantum}).
\item \textit{Limiting Cases:} Formulae for limits of $D_{\alpha}(\rho||\sigma)$ as $\alpha\rightarrow 0, 1,$ and $\infty$. The von Neumann relative entropy $D(\rho||\sigma)$ is recovered as \[\lim_{\alpha \uparrow 1} D_{\alpha}(\rho||\sigma) = D(\rho||\sigma)\](Theorems \ref{thm:4-veh-quantum}, \ref{thm:6-veh-quantum} anf \ref{thm:von-Neumann-entropy}).
\item Necessary and sufficient conditions for $\lim_{\alpha\downarrow0}D_{\alpha}(\rho||\sigma) =0$ (Theorem \ref{thm:23-veh-quantum}).
\item Necessary and sufficient conditions for $D_{\alpha}(\rho||\sigma) =\infty$ for some $\alpha\in [0,1)$ (Theorem \ref{thm:24-veh-quantum}).
\item \textit{A counter example:} Information theoretic definition of the von Neumann relative entropy is different from Araki's definition of relative entropy. This disproves a recent attempt to prove such a result in \cite[Theorem 20]{luczak2019relative} (Example \ref{eg:counter-example}, Theorem \ref{thm:counter-example}).
\item \textit{Continuity:} $D_\alpha (\rho||\sigma)$ is continuous in $\alpha$ on \[\mathcal{A} = \{\alpha\in [0,\infty]\,:\, 0\leq \alpha\leq 1\textnormal{ or } D_{\alpha}(\rho||\sigma)<\infty\}\]
(Theorem \ref{thm:7-veh-quantum}).
\item \textit{Quantum Pinsker Type inequality:} For $\alpha\in (0,1]$ \begin{align*}
\frac{\alpha}{2}\norm{\rho-\sigma}_2^4\leq D_{\alpha}(\rho||\sigma),
\end{align*}
where $\norm{\cdot}_2$ denotes the Hilbert-Schmidt norm (Theorem \ref{thm:quantum-pinsker}).
\item \textit{Skew symmetry:} for $0<\alpha<1$, \[D_\alpha(\rho||\sigma) = \frac{\alpha}{1-\alpha}D_{1-\alpha}(\sigma||\rho)\]
(Proposition \ref{prop:8-veh-quantum}).
\item \textit{Concavity:} The function $(1-\alpha)D_{\alpha}(\rho||\sigma)$ is concave in $\alpha$ (Proposition \ref{cor:2-veh-quantum}).
\end{enumerate}
\section{Preliminaries}\label{sec:preliminaries}
\subsection{A Criteria for Finite Trace}Let us recall the definition of the trace of a positive operator, as this definition have an important role in our analysis. We refer to Section 11 in Chapter 2 of \cite{vN2018} for this definition of the trace. Let $\mathcal{K}$ denote an arbitrary but fixed complex separable Hilbert space for this section. In what follows, a selfadjoint operator $T$ (not necessarily bounded) is said to be positive if \begin{equation}
\mel{x}{T}{x}\geq 0,\quad \forall x\in D(T),
\end{equation} (See Definition 9 of Chapter 2 (page 62) of \cite{vN2018}). A positive operator defined in this way has spectrum on the positive half of the real line.
\begin{defn}\label{defn:trace}
Let $T$ be a positive operator defined on a dense subspace $D(T)\subseteq \mathcal{K}$. Let $\{u_j\} \subseteq D(T)$ be any orthonormal basis (onb) of $\mathcal{K}$ (such an onb always exists by applying Gram-Schmidt process to a countable dense subset of $D(T)$). Define the trace of $T$, denoted by $\tr T$, by \begin{align}
\tr T = \sum\limits_j \mel{u_j}{T}{u_j} \in [0, \infty].
\end{align}
Trace of $T$ thus defined is independent of the choice of the onb $\{u_j\}$ \cite{vN2018}. A positive operator T is said to be a trace class operator if $\tr T <\infty$.
\end{defn}
The following lemma shows that a positive trace class operator (as in the previous definition) extends as a bounded operator on $\mathcal{K}$.
\begin{lem}\label{sec:gaussian-states-lem}
Let $D(T)\subseteq \mathcal{K}$ be a dense subspace
and let $T:D(T)\rightarrow \mathcal{K}$ be a positive operator. Assume that there exists an orthonormal basis $\{u_k\}\subseteq D(T)$ such that \begin{align*}
\sum\limits_k \mel{u_k}{T}{u_k} = M<\infty,
\end{align*} i.e., $T$ is a trace class operator in the sense of Definition \ref{defn:trace},
then $T$ extends to a bounded operator on $\mathcal{K}$.
\end{lem}
\begin{proof}
It is enough to prove that $T$ is bounded. Let $h\in D(T)$ be a finite linear combination of $u_k$'s, i.e., $h = \sum_k \braket{u_k}{h}\ket{u_k}$, where $\braket{u_k}{h} = 0$ for all but finitely many $k$'s. Then \begin{align*}
Th = \sum_k \braket{u_k}{h}T\ket{u_k}.
\end{align*}
Let $\xi^T$ denote the spectral measure associated with $T$, $\sqrt{T}$ exists as a positive operator and then we have \begin{align*}
D(\sqrt{T}) &= \{u\in \mathcal{K}: \int_{\mathbb{R}}\abs{x}\mel{u}{\xi^T(\dd x)}{u}<\infty\}\\&
\supseteq \{u\in \mathcal{K}: \int_{\mathbb{R}}\abs{x}^2\mel{u}{\xi^T(\dd x)}{u}<\infty\} = D(T).
\end{align*}
Therefore, by applying Cauchy-Schwartz inequality twice we have
\begin{align*}
0\leq \mel{h}{T}{h} &= \braket{\sum \limits_j \braket{u_j}{h}\ket{u_j}}{\sum \limits_k \braket{u_k}{h}T\ket{u_k}}\\
& = \sum\limits_{j,k}\overline{\braket{u_j}{h}}\braket{u_k}{h}\mel{u_j}{T}{u_k}\\
& \leq \sum\limits_{j,k}\abs{\overline{\braket{u_j}{h}}\braket{u_k}{h}\mel{u_j}{T}{u_k}}\\
&\leq \sum\limits_{j,k}\abs{\overline{\braket{u_j}{h}}\braket{u_k}{h}\norm{\sqrt{T}u_j}\norm{\sqrt{T}u_k}}\\
& = \left(\sum\limits_{k}\abs{\braket{u_k}{h}}\norm{\sqrt{T}u_k}\right)^2\\
& \leq \left(\sum\limits_{k}\abs{\braket{u_k}{h}}^2\right)\left(\sum\limits_{k}\norm{\sqrt{T}u_k}^2\right)\\
&= M \norm{h}^2, \quad \forall h\in D(T).
\end{align*}
Hence $T$ defined on $D(T)$ extends to a bounded operator on $\mathcal{K}$.
\end{proof}
In order to study the Petz-R\'enyi $\alpha$-relative entropy (see Definition \ref{defn:alpha-relative-entropy}) of two states, $\rho$ and $\sigma$ we need to understand whether an operator of the form $\rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}$ is trace class or not. Theorem \ref{sec:gaussian-states-prop} in this section provides an equivalent criteria to deal with this situation. We do not expect Theorem \ref{sec:gaussian-states-prop} to be new but we could not find a suitable reference to cite the same. First we recall the definition of the support of a positive operator.
\begin{defn}\label{defn:support}
Define the support of a positive operator $\tau$, denoted by $\supp\tau$ as \[\supp\tau = (\ker\tau)^{\perp} = \overline{\ran}\tau.\]
\end{defn}
\begin{defn}\label{sec:states-observables-1}\label{defn:moments}Let $\mathcal{K}$ be a complex separable Hilbert space and $\mathcal{K}_1\subseteq \mathcal{K}$ be a closed subspace. Let
$X$ be a (possibly unbounded) selfadjoint operator defined on $D(X)\subseteq\mathcal{K}_1\subseteq \mathcal{K}$ with spectral measure $\xi^{X}$ and let $\tau$ be a positive compact operator on $\mathcal{K}$ with $\supp \tau \subseteq \mathcal{K}_1$. Define a (possibly infinite) positive measure $\mu^{\tau,X}$ on $(\mathbb{R}, \mathcal{B}_{\mathbb{R}})$, where $\mathcal{B}_{\mathbb{R}}$ is the Borel $\sigma$-algebra of $\mathbb{R}$, by \[\mu^{\tau,X}(E):= \tr \xi^{X}(E)\tau,\quad \forall E \in \mathcal{B}_{\mathbb{R}}.\] Then $\mu^{\tau,X}$ is called the \emph{distribution of $X$ with respect to $\tau$}.
\end{defn}
\begin{rmk} \label{rmk:distribution}
If $\tau$ is a positive compact operator on $\mathcal{K}$ with $\supp \tau \subseteq \mathcal{K}_1$, then by the spectral theorem,
\begin{equation}
\label{eq:15}
\tau = \sum_{i}p_{i}\ketbra{u_i}{u_{i}},
\end{equation}
where $p_{i}\geq 0$ and $\{u_i\}$ is an orthonormal basis in $\mathcal{K}$, $u_i\in \mathcal{K}_1$ when $p_i\neq 0$. In this case, \begin{equation}
\label{eq:9}
\mu^{\tau,X}(E) = \sum\limits_{\{i\,:\, p_i\neq 0\}}^{}p_i \left\langle u_i|\xi^X(E)|u_i \right\rangle,
\end{equation}
for any Borel set $E \subseteq \mathbb{R}$. Thus $\mu^{\tau,X}$ is supported inside the spectrum of $X$. Furthermore, $\mu^{\tau,X}$ is a $\sigma$-finite measure.
\end{rmk}
\begin{thm}\label{sec:gaussian-states-prop}
Let $\tau$ be a
positive compact operator on $\mathcal{K}$ with spectral decomposition
\begin{align*}
\tau = \sum\limits_{i\in\mathcal{I}}^{}p_i\ketbra{u_i},
\end{align*} where $p_i\geq 0$ and $\{u_i\}_{i\in\mathcal{I}}$ an orthonormal basis of
$\mathcal{K}$. Let $X$ be a positive operator defined on a dense subspace $D(X)$ of a closed subspace $\mathcal{K}_1$ of $\mathcal{K}$.
Assume that \begin{align}\label{item:8}
u_i\in D(X), \text{ for all } i\in \mathcal{I} \text{ such that } p_i\neq 0.
\end{align}Then the operator $\tau^{1/2}X\tau^{1/2}$ initially defined on the dense subspace $\spn\{u_i\}_{i\in \mathcal{I}}$ by \begin{align*}
\tau^{1/2}X\tau^{1/2}u_i= \begin{cases}
\sqrt{p_i} \tau^{1/2}Xu_i, & \text{ when } p_i\neq 0\\
0, &\text{ when } p_i = 0
\end{cases}
\end{align*} extends as a positive selfadjoint operator on a dense domain in $\mathcal{K}$. Furthermore,
\begin{align}
\label{item:7} \tr \tau^{1/2}X\tau^{1/2} = \int_{0}^{\infty}x \mu^{\tau,X}(\dd x),
\end{align}
where $\mu^{\tau,X}$ is the distribution of $X$ with respect to $\tau$ (see Definition \ref{sec:states-observables-1}).
\end{thm}
\begin{proof}
Note that $\spn\{u_i\}_{i\in \mathcal{I}}$ is a dense subspace of $\mathcal{K}$. Clearly, $\tau^{1/2}X\tau^{1/2}$ defined as in the statement of the Theorem is a symmetric operator. For any $x\in \spn\{u_i\}_{i\in \mathcal{I}}$, we have \[\mel{x}{\tau^{1/2}X\tau^{1/2}}{x} = \mel{\tau^{1/2}x}{X}{\tau^{1/2}x} \geq 0.\]
Now the fact that $\tau^{1/2}X\tau^{1/2}$ defined on $\spn\{u_i\}_{i\in \mathcal{I}}$ extends as a positive selfadjoint operator is a consequence of the Friedrichs extension (\cite[Theorem 5.1.13]{pedersen2012analysis}, also \cite[Theorem X.23]{ReS75ii}). We fix the Friedrichs extension of $\tau^{1/2}X\tau^{1/2}$ and use same notation for the extended operator.
Now we prove (\ref{item:7}). Since $\tau^{1/2}X\tau^{1/2}$ is a positive operator, we have
\begin{align*} \tr \tau^{1/2}X\tau^{1/2}&=
\sum\limits_{i\in \mathcal{I}}^{}\mel{u_i}{\tau^{1/2}X\tau^{1/2}}{u_i}\\ &=
\sum\limits_{\{i|p_i\neq 0\}}^{}p_i\mel{u_i}{X}{u_i}\\
& =\sum\limits_{\{i|p_i\neq 0\}}^{}p_i \int\limits_{0}^{\infty}x\mel{u_i}{\xi^X(\dd
x)}{u_i}\numberthis \label{eq:20}\\
& = \int\limits_{0}^{\infty}x \sum\limits_{\{i|p_i\neq 0\}}^{}p_i\mel{u_i}{\xi^X(\dd
x)}{u_i}\numberthis \label{eq:21.1}\\
& = \int\limits_{0}^{\infty}x \sum\limits_{\{i|p_i\neq 0\}}^{}p_i\mel{u_i}{\xi^X(\dd
x)}{u_i}\\
&=\int_{0}^{\infty}x \mu^{\tau,X}(\dd x),
\end{align*} where (\ref{eq:20}) follows from the spectral theorem, (\ref{eq:21.1}) follows because all the terms are positive and we used (\ref{eq:9}) in the last line above.
\end{proof}
\begin{rmk}
Since
$T=\tau^{1/2}X\tau^{1/2}$ is a positive operator, Lemma
\ref{sec:gaussian-states-lem} shows that $\tau^{1/2}X\tau^{1/2}$ is a
trace class operator if and only if $\int_{0}^{\infty}x \mu^{\tau,X}(\dd x)<\infty$.
\end{rmk}
\section{Petz-R\'enyi Relative Entropy of States}
\begin{defn}\label{defn:alpha-relative-entropy}
Let $\rho$ and $\sigma$ be any two states on a Hilbert space $\mathcal{K}$. For $\alpha \in (0,1)\cup (1,\infty)$, the Petz-R\'enyi $\alpha$-relative entropy of $\rho$ given $\sigma$ is
\begin{equation}\label{eq:1.1}
D_{\alpha} (\rho \lvert\lvert \sigma) = \frac{1}{\alpha-1}\log \tr \rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2},
\end{equation}
whenever $ \rho^{\alpha/2}\sigma^{(1-\alpha)}\rho^{\alpha/2} $ is a positive selfadjoint operator defined on a dense subspace operator on $\mathcal{K}$. Otherwise, $D_{\alpha} (\rho \lvert\lvert \sigma) =\infty$.
\end{defn}
\begin{rmks}\begin{enumerate}
\item
When $\alpha>1$, we note that $\sigma^{(1-\alpha)}$ is defined as the pseudoinverse of $\sigma$ raised to the power $(\alpha-1)$. If the spectral decomposition of $\sigma$ is $\sum_{j}\lambda_j\ketbra{f_j}$ with $\lambda_j>0$, then \begin{align}\label{eq:pseudo-inverse}
\sigma^{(1-\alpha)} = \sum_{j}\lambda_j^{(1-\alpha)}\ketbra{f_j}.
\end{align}
It may be noted from the spectral theorem (Theorem 12.4 in \cite{Par12}) that the pseudo inverse as defined above is a selfadjoint operator (not necessarily bounded) because its spectral measure is supported on the real line. Furthermore, by (\ref{eq:pseudo-inverse}), we also have $\sigma^{(1-\alpha)}$ is a positive operator.
\item Let $\norm{A}_p$ denote the Schatten $p$-norm of an operator $A$. When $0<\alpha<1$, we know that $\rho^{\alpha}$ and $\sigma^{(1-\alpha)}$ are bounded operators. In this case we have,
\[ \tr \rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2} = \tr \rho^{\alpha}\sigma^{1-\alpha}\\
\leq \norm{\rho^{\alpha}}_{\frac{1}{\alpha}}\norm{\sigma^{1-\alpha}}_{\frac{1}{1-\alpha}}=1.\]
Hence $0\leq D_{\alpha}(\rho||\sigma)$ for $0<\alpha<1$.
\end{enumerate}
\end{rmks}
\subsection{Main Results}
\begin{thm}\label{thm:the-formula-general}
Let $\mathcal{K}$ be a complex Hilbert space with $\dim \mathcal{K} = \abs{\mathcal{I}}$, where $\mathcal{I}\subseteq \mathbb{N}$ may be a finite or infinite set. Let $\rho$ and $\sigma$ be states on $\mathcal{K}$ with spectral decomposition \begin{align}\label{eq:spectral-rho-sigma}\begin{split}
\rho &= \sum_{i\in \mathcal{I}}r_i \ketbra{u_i}, \quad r_i\geq 0,\quad \sum_i r_i = 1,\quad \{u_i\}_{i} \text{ is an orthonormal basis};\\
\sigma &= \sum_{j\in \mathcal{I}}s_j \ketbra{v_j}, \quad s_j\geq0, \quad \sum_j s_j = 1, \quad \{v_j\}_j\text{ is an orthonormal basis}.
\end{split}
\end{align}
Let $\alpha\in (0,1)\cup(1,\infty)$. Then the Petz-R\'enyi $\alpha$-relative entropy of $\rho$ and $\sigma$ is finite if and only if the following two conditions are satisfied: \begin{enumerate}
\item\label{item:1-general} $u_i\in D(\sigma^{(1-\alpha)})$, $\forall i $ such that $r_i\neq 0$, which is same as \begin{equation}\label{eq:condition-1}
\sum\limits_j s_j^{2(1-\alpha)}\abs{\braket{u_i}{v_j}}^2<\infty, \quad \forall i \text{ such that } r_i\neq 0
\end{equation}
\item \label{item:2-general}$\sum\limits_{i,j} r_i^{\alpha}s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2 <\infty$.
\end{enumerate}
Furthermore, \begin{align}
\label{eq:alpha-relative-entropy}
D_{\alpha} (\rho \lvert\lvert \sigma) = \frac{1}{\alpha-1}\log \sum\limits_{i,j} r_i^{\alpha}s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2,
\end{align}
whenever the Condition \ref{item:1-general} above is satisfied.
\end{thm}
\begin{proof}
By Lemma \ref{sec:gaussian-states-lem} and Theorem \ref{sec:gaussian-states-prop}, the operator $\rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}$ is trace class if and only if the following two conditions are satisfied:
\begin{enumerate}[label=(\roman*)]
\item \label{item:1-general-proof} $u_i\in D(\sigma^{(1-\alpha)})$, $\forall i $ such that $r_i\neq 0$;
\item \label{item:2-general-proof} $\int_0^\infty x\mu^{\rho^\alpha,\sigma^{(1-\alpha)}}(\dd x)<\infty$,
\end{enumerate}
where \ref{item:1-general-proof} ensures that the operator $\rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}$ is well defined and \ref{item:2-general-proof} provides the value of a sum determining the trace. We will show that Condition \ref{item:1-general-proof} above is equivalent to (\ref{eq:condition-1}) and Condition \ref{item:2-general-proof} above is equivalent to Condition \ref{item:2-general} in the statement of the present theorem. To this end, let $\xi^{\sigma^{1-\alpha}}$ denote the spectral measure corresponding to the operator $\sigma^{1-\alpha}$. Then \begin{align}
\label{eq:spectral-measure-sigma}
\xi^{\sigma^{1-\alpha}}\{s_j^{1-\alpha}\} = \sum\limits_{\{r\in \mathbb{Z}|s_j=s_r\}}\ketbra{v_r}.
\end{align} Now by spectral theory, for $i$ such that $r_i\neq 0$ in Condition \ref{item:1-general-proof} \begin{align*}
& u_i\in D(\sigma^{(1-\alpha)})\\
& \Leftrightarrow \int\limits_0^\infty \abs{x}^2\mel{u_i}{\xi^{\sigma^{1-\alpha}}(\dd x)}{u_i}<\infty\quad \\
& \Leftrightarrow \sum\limits_{j} \abs{s_j}^{2(1-\alpha)}\mel{u_i}{\xi^{\sigma^{1-\alpha}}\{s_j\}}{u_i}<\infty \\
& \Leftrightarrow \sum\limits_{j} \abs{s_j}^{2(1-\alpha)}\mel{u_i}{\left(\sum\limits_{\{r\in \mathbb{Z}|s_j=s_r\}}\ketbra{v_r}\right)}{u_i}<\infty\\
& \Leftrightarrow \sum\limits_{j}\sum\limits_{\{r\in \mathbb{Z}|s_j=s_r\}} \abs{s_j}^{2(1-\alpha)}\mel{u_i}{\left(\ketbra{v_r}\right)}{u_i}<\infty\\
& \Leftrightarrow \sum\limits_{j}\sum\limits_{\{r\in \mathbb{Z}|s_j=s_r\}} \abs{s_j}^{2(1-\alpha)}\abs{\braket{u_i}{v_r}}^2<\infty\\
& \Leftrightarrow \sum\limits_{j} \abs{s_j}^{2(1-\alpha)}\abs{\braket{u_i}{v_j}}^2<\infty
\end{align*}
which proves (\ref{eq:condition-1}).
Now a similar computation as above using (\ref{item:7}), (\ref{eq:9}) and the fact that $\mu^{\rho^\alpha,\sigma^{(1-\alpha)}}$ is supported on the spectrum of $\sigma^{1-\alpha}$ shows that \begin{align*}
\tr \rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2} &= \int_0^\infty x\mu^{\rho^\alpha,\sigma^{(1-\alpha)}}(\dd x)\\
&= \sum\limits_j s_j^{1-\alpha}\sum\limits_ir_i^{\alpha}\mel{u_i}{\xi^{\sigma^{1-\alpha}}\{s_j^{1-\alpha}\}}{u_i}\\
& = \sum\limits_{i,j}r_i^{\alpha}s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2.
\end{align*}
Thus Condition \ref{item:2-general-proof}
is same as Condition \ref{item:2-general} and also (\ref{eq:alpha-relative-entropy}) is proved.
\end{proof}
\begin{rmks}\begin{enumerate}\item It may be noted that the Condition \ref{item:1-general} merely states that the operator $\rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}$ is well defined. Hence the Condition \ref{item:2-general} is the only condition necessary and sufficient for the finiteness of the Petz-R\'enyi $\alpha$-relative entropy between two states $\rho$ and $\sigma$, if we already know that the operator $\rho^{\alpha/2}\sigma^{1-\alpha}\rho^{\alpha/2}$ is well defined.
\item
In our definition of Petz-R\'enyi relative entropy between $\rho$
and $\sigma$ the finiteness of $D_\alpha(\rho||\sigma)$ implies (by Lemma \ref{sec:gaussian-states-lem}) that the operator $\rho^{\alpha/2}\sigma^{(1-\alpha)}\rho^{\alpha/2}$ is a bounded operator. In Chapter 2 of \emph{Mathematical Foundations of Quantum Mechanics}
\cite{vN2018}, John von Neumann defines the trace of the product of two unbounded positive operators $A,B$, which we denote as $\operatorname{vN-Tr} \{AB\}$, by \begin{align}\label{eq:vN-trace}
\operatorname{vN-Tr} \{AB\} = \sum\limits_j \norm{\sqrt{A}\sqrt{B}u_j}^2,
\end{align}
where $\{u_j\}$ is an orthonormal basis of the Hilbert space contained
in the domain of $B$. Then the sum on the right side of (\ref{eq:vN-trace}) is independent of the choice of basis. We can define \begin{align*}
\operatorname{vN-D}_{\alpha}(\rho||\sigma) := \frac{1}{\alpha-1}\log \left(\operatorname{vN-Tr} \{\rho^{\alpha}\sigma^{(1-\alpha)}\}\right).
\end{align*}
Now we compute the $\operatorname{vN-Tr} \{\rho^{\alpha}\sigma^{(1-\alpha)}\}$ using an orthonormal basis consisting of eigenvectors of $\sigma$ as in Theorem \ref{thm:the-formula-general},
\begin{align*}
\operatorname{vN-Tr} \{\rho^{\alpha}\sigma^{(1-\alpha)}\} &=\sum\limits_{j} \norm{\rho^{\alpha/2}\sigma^{(1-\alpha)/2}v_j}^2\\
& = \sum\limits_{j}
s_j^{1-\alpha}\norm{\rho^{\alpha/2}v_j}^2\numberthis
\label{eq:alternate-formula}\\
& = \sum\limits_{j}
s_j^{1-\alpha}\mel{v_j}{\rho^{\alpha}}{v_j}\\
& = \sum\limits_{j}
s_j^{1-\alpha}\mel{v_j}{\left(\sum\limits_ir_i^{\alpha}\ketbra{u_i}\right)}{v_j}\\
& = \sum\limits_{i,j}r_i^{\alpha}s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2.
\end{align*}
Thus the
formula (\ref{eq:alpha-relative-entropy}) for the relative
entropy remains the same in this case also. Since $\rho^{\alpha}\sigma^{(1-\alpha)}$ is a densely defined operator, we do not require Condition~1 of Theorem \ref{thm:the-formula-general} to hold in this case. The definition of trace according to von Neumann has the property, $\operatorname{vN-Tr} AB = \operatorname{vN-Tr} BA$, whenever $AB$ and $BA$ are well-defined. It is not clear to us (from the book \cite{vN2018}) whether von Neumann wants to define the trace only when $AB$ and $BA$ both are well defined or not. Nevertheless, we have $ \operatorname{vN-Tr} \rho^{\alpha}\sigma^{1-\alpha} = \operatorname{vN-Tr} \sigma^{1-\alpha}\rho^{\alpha}$ whenever the latter is well-defined, but in this case Condition~1 in Theorem \ref{thm:the-formula-general} is needed. For example, if $\alpha>1$ and $\sigma$ is any state, then
$\sigma^{1-\alpha}$ is an unbounded operator. Let $u\in \mathcal{K}$ be such that $u\notin D(\sigma^{1-\alpha})$ and choose $\rho = \ketbra{u}$. In this case $\sigma^{1-\alpha}\rho^\alpha$ is not well defined.
\end{enumerate}
\end{rmks}
Now we prove one of the major results in this article. The reader is referred to Appendix \ref{sec:classical-renyi-devergence} for the basic definitions and properties related to the classical Petz-R\'enyi divergence.
\begin{thm}[Petz-R\'enyi relative entropy reduces to R\'enyi relative divergence]\label{thm:quantum-reduces-to-classical}
Let $\mathcal{K}$ be a complex Hilbert space with $\dim \mathcal{K} = \abs{\mathcal{I}}$, where $\mathcal{I} =\{1,2,\dots,n\}$ or $\mathcal{I}=\mathbb{N}$.
Let $\rho$ and $\sigma$ be as in \eqref{eq:spectral-rho-sigma}.
Define $P$ and $Q$ on $\mathcal{I}\times \mathcal{I}$ such that,
\begin{align}
\label{eq:P-and-Q}\begin{split}
P(i,j)= p_{ij} &= r_i\abs{\braket{u_i}{v_j}}^2,\\
Q(i,j)= q_{ij}& = s_j\abs{\braket{u_i}{v_j}}^2, \quad \forall (i,j)\in \mathcal{I}\times\mathcal{I}.
\end{split}
\end{align} Then $P$ and $Q$ are probability measures on $\mathcal{I}\times \mathcal{I}$ such that \begin{equation}\label{eq:main-result}
D_{\alpha}(\rho||\sigma) = D_{\alpha}(P||Q) ,\quad \forall \alpha\in (0,1)\cup(1,\infty),
\end{equation}
where the $D_\alpha$ on the left side of the equation above denotes the Petz-R\'enyi relative entropy of quantum states $\rho$ and $\sigma$, and that on the right side is the classical R\'enyi relative divergence of probability distributions $P$ and $Q$.
\end{thm}
\begin{proof}
Since $\{u_i\}_{i}$ and $\{v_j\}_{j}$ are orthonormal bases, \[\sum_{i,j}p_{ij} = \sum_{i}r_i\sum_j\abs{\braket{u_i}{v_j}}^2 = \sum_{i}r_i = 1.\]
Similarly, $\sum_{i,j}q_{ij} = 1$. Hence $P$ and $Q$ are probability distributions on $\mathcal{I}\times \mathcal{I}$. Let $\mu$ be the counting measure on $\mathcal{I}\times \mathcal{I}$, then clearly $P\ll\mu$, $Q\ll\mu$, \begin{align*}
\dv{P}{\mu}\,(i,j)=p_{ij}, \quad \textnormal{and} \quad \dv{Q}{\mu}\,(i,j)=q_{ij},
\end{align*} where $\dv{P}{\mu}$ and $\dv{Q}{\mu}$ are the respective Radon-Nikodym derivatives. Now by Definition \ref{defn:renyi-divergence-Kullback}, and equation (\ref{eq:alpha-relative-entropy}) in Theorem \ref{thm:the-formula-general}, for $\alpha\in (0,1)\cup (1,\infty)$ the R\'enyi divergence of $P$ from $Q$ is given by
\begin{align*}
D_{\alpha}(P||Q) &= \frac{1}{\alpha-1} \log \int_{\mathcal{I}\times \mathcal{I}} \left(\dv{P}{\mu}\right)^\alpha \left(\dv{Q}{\mu}\right)^{1-\alpha} \dd \mu \\
&= \frac{1}{\alpha-1} \log \sum\limits_{i,j} \left(r_i\abs{\braket{u_i}{v_j}}^2\right)^\alpha\left(s_j\abs{\braket{u_i}{v_j}}^2\right)^{1-\alpha}\\
&=\frac{1}{\alpha-1} \log \sum\limits_{i,j} r_i^\alpha s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2\\
&= D_{\alpha}(\rho||\sigma),
\end{align*}
where for $\alpha>1$, we adopt the conventions $0^{1-\alpha} = \infty$ and $0\cdot \infty = 0$ as in Definition \ref{defn:renyi-divergence-Kullback}
\end{proof}
\begin{rmk}\label{rmk:quamtum-classical-measurement}
In the setting of Theorem \ref{thm:quantum-reduces-to-classical} and its proof, for $i,j\in \mathcal{I}$ define $A_{ij} = \braket{u_i}{v_j}\ketbra{u_i}{v_j}$. Then $\sum_{i,j}A_{ij}A_{ij}^\dagger = I = \sum_{i,j}A_{ij}^\dagger A_{ij}$. Hence both $\{A_{ij}A_{ij}^\dagger\}$ and $\{A_{ij}^\dagger A_{ij}\}$ are POVM's. Furthermore, it may be noted that $p_{ij} = \tr \rho A_{ij}A_{ij}^\dagger$ and $q_{ij} = \tr \sigma A_{ij}^\dagger A_{ij}$. Thus the probability measures $P$ and $Q$ in the previous theorem are precisely those measures that are obtained by measuring $\rho$ and $\sigma$, respectively in $\{A_{ij}A_{ij}^\dagger\}$ and $\{A_{ij}^\dagger A_{ij}\}$.
\end{rmk}
Now we prove a lemma and two propositions which describe some relationships between the pairs $(\rho, \sigma)$ and $(P,Q)$ described in (\ref{eq:spectral-rho-sigma}) and (\ref{eq:P-and-Q}).
\begin{lem}\label{lem:support-condition}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}). Then $\supp \rho \subseteq \supp \sigma$ if and only if $s_j=0$ for some $j$ implies that for every $i$ at least one of the two quantities $\{\braket{u_i}{v_j}, r_i \}$ is equal to zero.
\end{lem}
\begin{proof} ($\Rightarrow$) Assume $\supp \rho \subseteq \supp \sigma$.
The assumption that $s_j=0$ implies that $v_j\in \ker \sigma$. Now the condition $\braket{u_i}{v_j}\neq 0$ implies that $u_i\notin (\ker{\sigma})^{\perp} =\supp \sigma$. Since $\supp \sigma\supseteq\supp \rho$ and $u_i$ is an eigenvector of $\rho$ we see that \[u_i\notin\supp\rho \Rightarrow u_i\in \ker \rho \Rightarrow r_i=0.\]
($\Leftarrow$) We will prove that $\ker \sigma \subseteq \ker \sigma$. It is enough to prove that every $v_j$ for which $s_j = 0$ belong to $\ker \rho$. Fix such $j$ such that $s_j=0$, by our assumption, $\braket{u_i}{v_j}=0$ for all $i$ such that $r_i\neq 0$ and hence $\sum_{\{i:r_i\neq 0\}}\braket{u_i}{v_j}\ket{u_i} = 0$. This means that the projection of $v_j$ to $\supp \rho$ is the vector is the zero vector which means that $v_j\in \ker \rho$.
\end{proof}
\begin{prop}\label{prop:rho-equals-sigma-P-equals-Q}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}) and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}) then
\[P=Q \Leftrightarrow \rho=\sigma.\]
\end{prop}
\begin{proof}Clearly, if $\rho=\sigma$ then $P=Q$. Now assume $P=Q$.
By definition \begin{align}\label{eq:P-equals-Q-implies}
P=Q \Rightarrow r_i\abs{\braket{u_i}{v_j}}^2 = s_j\abs{\braket{u_i}{v_j}}^2,\quad \forall i,j.
\end{align}
Therefore, $s_j=0$ for some $j$ implies that for every $i$ at least one of the two quantities $\{\braket{u_i}{v_j}, r_i \}$ is equal to zero. Now by Lemma \ref{lem:support-condition}, $\supp\rho \subseteq \supp \sigma$. Also $\supp \sigma \subseteq \supp \rho$ by symmetry of the situation. Hence we have \[\supp\rho = \supp \sigma.\]
Therefore, $\ker\rho = \ker\sigma$. To prove that $\rho = \sigma$, we will prove that the nonzero eigenvalues and corresponding eigenspaces of $\rho$ and $\sigma$ are same.
Now fix $i_0\in \mathcal{I}$ such that $r_{i_0}\neq 0$. If $r_{i_0}\neq s_j$ for all $j$, the second equality in (\ref{eq:P-equals-Q-implies}) shows that $\braket{u_{i_0}}{v_j} =0$ for all $j$ such that $s_j\neq 0$. But this is impossible because $\{v_j|s_j\neq0\}$ is an orthonormal basis for $\supp \sigma$ and $0\neq u_{i_0}\in \supp \sigma$. Therefore, for each $i_0\in \mathcal{I}$ such that $r_{i_0}\neq 0$, there exists $j_0\in \mathcal{I}$ such that $r_{i_0}=s_{j_0}.$ Hence $\operatorname{sp}(\rho) \subseteq\operatorname{sp}(\sigma)$. A similar argument shows that $\operatorname{sp}(\sigma) \subseteq \operatorname{sp}(\rho)$. Thus we have \[\operatorname{sp}(\rho) =\operatorname{sp}(\sigma).\]
Fix ${i_0}$ and ${j_0}$ such that $r_{i_0}=s_{j_0}$.
Let $R_{i_0}$ denote the eigenspace of $\rho$ corresponding to the eigenvalue $r_{i_0}$ and $S_{j_0}$ denote the eigenspace of $\sigma$ corresponding to the eigenvalue $s_{j_0}$. We will show that $R_{i_0}=S_{j_0}$, which will complete the proof since $r_{i_{0}}$ is an arbitrary non zero eigenvalue. It is enough to show the following two claims: \begin{enumerate}
\item \label{claim:1} If $r_i=r_{i_0}$ for some $i$ then $u_i\perp v_j$ for all $j\in \mathcal{I}$ with $s_j\neq s_{j_0}$;
\item If $s_j=s_{j_0}$ for some $j$ then $v_j\perp u_i$ for all $i\in \mathcal{I}$ with $r_i\neq r_{i_0}$.
\end{enumerate}
Proof of both the claims above are similar so we will prove \ref{claim:1} only. Fix $i$ such that $r_i=r_{i_0}$ and $j$ such that $s_j\neq s_{j_0}$. Since $r_{i_0}=s_{j_0}$, we have $r_i\neq s_j$. Thus by our assumption $\braket{u_i}{v_j}=0$, which proves \ref{claim:1}.
\end{proof}
\begin{prop}\label{prop:support-condition-iff-absolute-continuity}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}) and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}), then \[\supp \rho \subseteq \supp \sigma \Leftrightarrow P\ll Q.\]
\end{prop}
\begin{proof}Assume $ P\centernot{\ll} Q$, we have
\begin{align*}
P\centernot{\ll} Q &\Leftrightarrow \exists (i,j) \in \mathcal{I}\times \mathcal{I} \textnormal{ such that } \braket{u_i}{v_j} \neq 0, s_j=0,r_i\neq 0\\
&\Leftrightarrow \exists (i,j) \in \mathcal{I}\times \mathcal{I} \textnormal{ such that } \braket{u_i}{v_j} \neq 0, v_j\in \ker \sigma, u_i \in \supp \rho\\
&\Leftrightarrow \exists i\in \mathcal{I} \textnormal{ such that } u_i \in \supp \rho, u_i \notin (\ker \sigma )^\perp\\
&\Leftrightarrow \exists i\in \mathcal{I} \textnormal{ such that } u_i \in \supp \rho, u_i \notin \supp \sigma\\
&\Leftrightarrow \supp \rho \nsubseteq \supp \sigma
\end{align*}
\end{proof}
\subsubsection{The Limiting Cases}
The definition of Petz-R\'enyi $\alpha$-relative entropy excludes the values $0$, $1$ and $\infty$ of $\alpha$. Nevertheless, we can give meaning to the entropic quantities corresponding to these values of $\alpha$ and they are important in applications too \cite{Datta2009, muller-tomamichel-etal-2013, Datta_Leditzky_2014}. In this section, we study the limits at $0$ and $\infty$ of $D_{\alpha}(\cdot||\cdot)$. The case $\alpha= 1$ is requires more analysis and we discuss that in detail in Section \ref{sec:vN-entropy}.
First we prove that the Petz-R\'enyi $\alpha$-relative entropy is nondecreasing in $\alpha$, which will help us to extend the definition of $D_{\alpha}$ to the values $0, 1$ and $\infty$.
\begin{thm} \label{thm:3-veh-quantum}
For $\alpha\in (0,1)\cup(1,\infty)$ the Petz-R\'enyi entropy, $D_{\alpha}(\rho||\sigma)$ is nondecreasing in $\alpha$.
\end{thm}
\begin{proof}
This is an easy consequence of Theorem \ref{thm:quantum-reduces-to-classical} and Theorem \ref{thm:3-veh}.
\end{proof}
Theorem \ref{thm:3-veh-quantum} enables us to extend the definition of $D_{\alpha}(\rho||\sigma)$ to the values $\alpha = 0,1$ and $\infty$ as in the following definition.
\begin{defn}\label{defn:extended-orders-quantum}
The Petz-R\'enyi relative entropies of orders $0, 1$ and $\infty$ are defined as \begin{align*}
D_0(\rho||\sigma) &= \lim\limits_{\alpha \downarrow 0} D_{\alpha}(\rho||\sigma),\\
D_1(\rho||\sigma) &= \lim\limits_{\alpha \uparrow 1} D_{\alpha}(\rho||\sigma),\\
D_{\infty}(\rho||\sigma) &= \lim\limits_{\alpha \uparrow \infty} D_{\alpha}(\rho||\sigma).
\end{align*}
\end{defn}
With the definition above, we have the following corollary.
\begin{cor}
For $\alpha\in [0,\infty]$, the function $\alpha\mapsto D_{\alpha}(\rho||\sigma)$ is nondecreasing and thus \[D_0(\rho||\sigma)\leq D_{\alpha}(\rho||\sigma)\leq D_{\infty}(\rho||\sigma), \quad \forall \alpha\geq0.\]
\end{cor}
\begin{rmk} Following the notations used in \cite{Datta2009},
the quantity $D_0(\rho||\sigma)$ may also be written as $D_{\operatorname{min}}(\rho||\sigma)$.
\end{rmk}
If $\rho = \sum_ir_i\ketbra{u_i}{u_i}$ is a spectral decomposition of $\rho$, where $\{u_i\}$ is an orthonormal basis of $\mathcal{K}$, then
\begin{equation}\label{eq:support-formula}
\supp\rho = \overline{\spn}\{u_i|r_i\neq 0 \}.
\end{equation}
In \cite{Datta2009}, Datta observes in the finite dimensional setting that, \begin{align}\label{eq:datta-leditzky}
D_0(\rho||\sigma) = -\log \tr \Pi_{\rho}\sigma,
\end{align}
where $\Pi_{\rho}$ is the projection onto the support of $\rho$, i.e., by keeping the notations of Theorem \ref{thm:the-formula-general}, \[\Pi_{\rho} = \textnormal{Projection onto the }\spn\{u_i\,|\,r_i\neq 0\}.\] In the finite dimensions, this result follows from our formula (\ref{eq:alpha-relative-entropy}) as well, because we have \begin{align}\label{eq:datta-leditzky-computation}
\begin{split}
D_0(\rho||\sigma)&= \lim_{\alpha\rightarrow0}\frac{1}{\alpha-1}\log \sum\limits_{i,j} r_i^{\alpha}s_j^{1-\alpha}\abs{\braket{u_i}{v_j}}^2\\
& = -\log \sum\limits_{\{i,j|r_i\neq 0\}}s_j^{}\abs{\braket{u_i}{v_j}}^2\\
&=-\log \tr \sum_{\{i|r_i\neq 0\}}\ketbra{u_i}{u_i}\sum_{j}s_j\ketbra{v_j}{v_j}\\
&= -\log \tr \Pi_{\rho}\sigma.
\end{split}
\end{align}
Now we prove (\ref{eq:datta-leditzky}) in the infinite dimensional situation. A priori, the computation in (\ref{eq:datta-leditzky-computation}) cannot go through in infinite dimensions because we need a limit theorem to pass the limit through the infinite sum. Nevertheless, Theorem \ref{thm:4-veh} helps us to prove the desired result and the following proof works in both finite and infinite dimensional setting.
\begin{thm}\label{thm:4-veh-quantum}
The Petz-R\'enyi relative entropy satisfies, \begin{align}\label{eq:datta-leditzky-infinite}
D_0(\rho||\sigma) = -\log \tr \Pi_{\rho}\sigma,
\end{align}
where $\Pi_{\rho}$ is the projection onto $\supp \rho$.
\end{thm}
\begin{proof}
Keeping the notations of Theorem \ref{thm:the-formula-general}, we have by Theorems \ref{thm:quantum-reduces-to-classical} and \ref{thm:4-veh},
\begin{align*}
D_0(\rho||\sigma) &= -\log (Q ({\{p_{ij}>0\}}))\\
& = \sum_{\{i,j|r_i>0, \braket{u_i}{v_j}\neq0\}}s_j\abs{\braket{u_i}{v_j}}^2\\
& = -\log \sum\limits_{\{i,j|r_i\neq 0\}}s_j^{}\abs{\braket{u_i}{v_j}}^2\\
&=-\log \tr \sum_{\{i|r_i\neq 0\}}\ketbra{u_i}{u_i}\sum_{j}s_j\ketbra{v_j}{v_j}\\
&= -\log \tr \Pi_{\rho}\sigma.
\end{align*}
\end{proof}
\begin{rmk}
On a related note, it may be recalled that a sandwiched R\'enyi relative entropy, $\tilde{D}_{\alpha}$ was introduced by M\"{u}ller-Lennert et al. in \cite{muller-tomamichel-etal-2013}. Furthermore, Datta and Leditzky in their Theorem 1 of \cite{Datta_Leditzky_2014} proved that \[\lim_{\alpha\rightarrow 0 }\tilde{D}_{\alpha} = D_0(\rho||\sigma),\]
whenever $\supp \rho =\supp \sigma$.
\end{rmk}
\begin{thm}\label{thm:6-veh-quantum}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}). Then
\begin{equation}
\label{eq:limit-infty-quantum}
D_{\infty}(\rho||\sigma) = \log \underset{}{\sup}\left\{\frac{r_i}{s_j}\,:\,\braket{u_i}{v_j}\neq 0\right\},
\end{equation}
with the conventions that $0/0 = 0$ and $x/0 = \infty$ if $x>0$.
\end{thm}
\begin{proof} Let $P$ and $Q$ be as in (\ref{eq:P-and-Q}).
By the definition of $D_{\infty}(\rho||\sigma)$ and Theorem \ref{thm:quantum-reduces-to-classical}, we have
\begin{align*}
D_{\infty}(\rho||\sigma)&= \lim_{\alpha\rightarrow \infty}D_{\alpha}(\rho||\sigma)= \lim_{\alpha\rightarrow \infty}D_{\alpha}(P||Q) = D_{\infty}(P||Q)
\end{align*}
But by equation (\ref{eq:limit-infty-countable-sample-space}), \begin{align*}
D_{\infty}(P||Q) &= \log \underset{(i,j)\in \mathcal{I}\times \mathcal{I}}{\sup}\,\frac{r_i \abs{\braket{u_i}{v_j}^2}}{s_j\abs{\braket{u_i}{v_j}^2}}\\
&=\log \underset{}{\sup}\left\{\frac{r_i}{s_j}\,:\,\braket{u_i}{v_j}\neq 0\right\}
\end{align*}
with the conventions that $0/0 = 0$ and $x/0 = \infty$ if $x>0$.
\end{proof}
In the following Corollary, we remove the condition $\braket{u_i}{v_j}\neq 0$ from the formula for $D_{\infty}(\rho||\sigma)$ obtained in the previous theorem. But this will result in imposing an extra condition on the $\rho$ and $\sigma$.
\begin{cor}\label{cor:d-infty}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}). Assume further that \begin{align}\label{eq:D-infty-hypothesis}
\braket{u_i}{v_j}=0 \text{ for some } (i,j) \Rightarrow r_i = 0,
\end{align}
then
\begin{align}
D_{\infty}(\rho||\sigma) =\begin{cases}
\log \norm{\rho}\norm{\sigma^{-1}}, \quad &\text{ when } \dim \mathcal{K}<\infty;\\
\infty,\quad &\text{ when } \dim \mathcal{K}=\infty,
\end{cases}
\end{align}
where $\norm{\cdot}$ denotes the operator norm, $\norm{\sigma^{-1}}$ is defined to be infinity if $\sigma$ is not invertible, and $\log \infty:=\infty$.
\end{cor}
\begin{proof}
Notice that in Equation (\ref{eq:limit-infty-quantum}) we can remove the restriction
$\braket{u_i}{v_j} \not = 0$, since when $\braket{u_i}{v_j} = 0$, then by the hypothesis (\ref{eq:D-infty-hypothesis}), the fraction $\frac{r_i}{s_j}$ is equal to $0$ and it does
not contribute to the $\sup \{ \frac{r_i}{s_j}: (i,j) \in \mathcal{I}\times \mathcal{I}\}$. Now
\begin{align*}
D_{\infty}(\rho||\sigma)&= \log \underset{\underset{}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup}\,\frac{r_i }{s_j}.
\end{align*}
Note that if $\dim \mathcal{K} = \infty$ then $s_j=0$ for some $j$ or there is a subsequence of $s_j$'s which converges to zero. In both cases we get $\log \underset{\underset{}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup}\,\frac{r_i }{s_j} = \infty$. On the other hand, if $\dim \mathcal{K} <\infty$ we have
\begin{align*}
D_{\infty}(\rho||\sigma)&= \log \underset{\underset{r_i\neq0}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup}\,\frac{r_i }{s_j}\\
&= \underset{\underset{r_i\neq0}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup} \log (\frac{r_i}{s_j})\\
&= \underset{\underset{r_i\neq0}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup} \left(\log r_i+\log{{s_j^{-1}}}\right) \\
&= \underset{r_i\neq0}{\sup} \{\log r_i\} +\underset{}{\sup} \{\log {s_j^{-1}}\}\\
&=\log \left(\underset{r_i\neq0}{\sup} \{r_i\}\right)+\log\left( \underset{}{\sup} \left\{{s_j^{-1}}\right\}\right)\\
&= \log \norm{\rho} +\log \norm{\sigma^{-1}}\\
&= \log \norm{\rho}\norm{\sigma^{-1}}.
\end{align*}
\end{proof}
\begin{rmk}\label{rmk:camparison-datta}Recall the definition of $D_{\operatorname{max}}(\rho||\sigma)$ from equation (7) in \cite{Datta2009}, \begin{align*}
D_{\operatorname{max}}(\rho||\sigma):= \log \norm{\sigma^{-1/2}\rho\sigma^{-1/2}},
\end{align*}
whenever $\supp\rho \subseteq\supp \sigma$. When $\supp \rho \subseteq \supp \sigma$, the equation (31) of \cite{Datta2009} notes that \begin{align}
\label{eq:datta-d-max}
D_{\operatorname{max}}(\rho||\sigma) \leq \log \sup\left\{\frac{1}{s_j}\,:\, s_j\neq 0\right\}.
\end{align}
Let $\operatorname{sp}(A)$ denote the spectrum of an operator $A$. Note that \[\norm{\sigma^{-1/2}\rho\sigma^{-1/2}}=\sup \operatorname{sp}(\sigma^{-1/2}\rho\sigma^{-1/2})=\sup \operatorname{sp}(\rho\sigma^{-1})\leq\norm{\rho\sigma^{-1}}\leq \norm{\rho} \norm{\sigma^{-1}}.\]
Therefore, when (\ref{eq:D-infty-hypothesis}) is satisfied and $\sigma$ is invertible, Corollary \ref{cor:d-infty} shows that \begin{align}\label{eq:comparison-datta}
D_{\operatorname{max}}(\rho||\sigma):= \log \norm{\sigma^{-1/2}\rho\sigma^{-1/2}}&\leq \log \norm{\rho} \norm{\sigma^{-1}} = \log \norm{\rho}+\log \norm{\sigma^{-1}}\nonumber\\&\leq \log \norm{\sigma^{-1}} = \log\sup\left\{\frac{1}{s_j}\,:\, s_j\neq 0\right\}.
\end{align} Thus equation (\ref{eq:comparison-datta}) above improves (\ref{eq:datta-d-max}) under the extra assumption (\ref{eq:D-infty-hypothesis}) and $\sigma$ is invertible.
\end{rmk}
Next example shows that the expressions in Theorem \ref{thm:6-veh-quantum} and Corollary \ref{cor:d-infty} for $D_{\infty}(\rho||\sigma)$ are different in general.
\begin{eg}
Let $\dim \mathcal{K} = 2$ and let $\{u_1,u_2\}$ be any orthonormal basis of $\mathcal{K} $. Let \begin{align*}
\rho = \sigma = \frac{2}{3}\ketbra{u_1}+\frac{1}{3}\ketbra{u_2}.
\end{align*}
Then $r_1=s_1=2/3$, $r_2=s_2=1/3$ and \[\left\{\frac{r_i}{s_j}\,:\,\braket{u_i}{v_j}\neq 0\right\} =\{1\}, \quad \left\{\frac{r_i}{s_j}\,:\, r_i\neq0\right\}= \{2,1\}. \]
Hence \[\log \underset{\underset{r_i\neq0,}{(i,j)\in \mathcal{I}\times \mathcal{I}}}{\sup}\,\frac{r_i }{s_j}=\log 2=1 > 0 = \log \underset{}{\sup}\left\{\frac{r_i}{s_j}\,:\,\braket{u_i}{v_j}\neq 0\right\}. \]
\end{eg}
\subsubsection{von Neumann Relative Entropy} \label{sec:vN-entropy} In this section, we discuss the limiting case $D_1(\rho||\sigma)$ for the Petz-R\'enyi $\alpha$-relative entropy. Theorem \ref{thm:von-Neumann-entropy} shows that this limit is the von Neumann entropy. The result is straight forward in the finite dimensional setting. But the infinite dimensional situation is a bit tricky because of reasons we shall see below. One definition of the von Neumann entropy in the infinite dimensional setting is Araki's definition \cite{araki1976relative, araki1977relative} for general semifinite von Neumann Algebra, using the heavy machinery of relative modular operator. To avoid this, we will show that it makes sense to define the von Neumann entropy as $D_1(\rho||\sigma)$ itself when $\mathcal{K} $ is infinite dimensional. We begin with the definition of von Neumann relative entropy in the finite dimensional setting.
In the finite dimensional situation, the von Neumann relative entropy $D(\rho||\sigma)$ of two states $\rho$ and $\sigma$ are defined as \begin{align*}
D(\rho||\sigma) = \begin{cases}
\tr\rho(\log\rho-\log\sigma), &\textnormal{ when } \supp \rho \subseteq \supp \sigma;\\
\infty, & \textnormal{otherwise.}
\end{cases}
\end{align*}
Let $\dim \mathcal{K} = n<\infty$, and let $\rho = \sum_{i=1}^{n}r_i \ketbra{u_i}, \sigma = \sum_{j=1}^{n}s_j \ketbra{v_j} $, where $\{u_i\}_{i=1}^{n}$, $\{v_j\}_{j=1}^{n}$ are orthonormal bases of $\mathcal{K} $. The operators $\log\rho$ and $\log \sigma$ have to be appropriately understood when $\rho$ and $\sigma$ have a non trivial kernel. We may define
\begin{align}\begin{split}\label{eq:log-finite-dim}
\log \rho &= \sum_{r_i\neq0}\log r_i \ketbra{u_i} \textnormal{ on } \supp \rho\\
\textnormal{and }\log \sigma &= \sum_{s_j\neq0}\log s_j \ketbra{v_j} \textnormal{ on } \supp \sigma.
\end{split}
\end{align}
Under the assumption $\supp \rho \subseteq \supp \sigma$, the operator $(\log \rho -\log \sigma)$ is well defined on $\supp \rho$. Now we describe another way to make sense of the definition of the von Neumann entropy. With the convention $0\cdot\infty=0$, the operator $\rho\log\rho$ can be defined as \[\rho\log\rho =\sum_{j=1}^{n}r_i\log r_i \ketbra{u_i}. \]
Also, $\rho\log\sigma$ can be defined as \begin{align}\label{eq:rholog-sigma-finite-dim}
\rho\log \sigma = \sum_{i,j=1}^n r_i\log s_j\braket{u_i}{v_j}\ketbra{u_i}{v_j},\quad\textnormal{whenever } \supp \rho \subseteq \supp \sigma,
\end{align}
because of Lemma \ref{lem:support-condition}.
We wish to set up a definition of $D(\rho||\sigma)$ which works well both in the finite and infinite dimensional setting. To this end, we first define the operator $-\log \sigma$ on a dense subspace of $\supp \sigma$ by the spectral decomposition
\begin{align}\label{eq:log-sigma}
-\log \sigma = \sum\limits_{\{j\in \mathcal{I}\,|\, s_j\neq 0\}}-\log s_j\ketbra{v_j},
\end{align}
where $v_j$ is considered as an element of the Hilbert space $\supp \sigma$.
Note that, by the spectral theorem the domain of $-\log \sigma$ is given by \begin{align}\label{eq:domain-logsigma}
D(-\log \sigma) = \{v\in \supp \sigma\,|\, \sum_{\{j\in \mathcal{I}\,|\,s_j\neq 0\}} \abs{\log s_j}^2\abs{\braket{v}{v_j}}^2<\infty\}\subseteq \supp \sigma.
\end{align}
It is easy to see that $\spn\{v_j\,:\, s_j\neq 0\}\subseteq D(-\log \sigma)$.
Now $(-\log \sigma)$ is a positive (selfadjoint) operator on $D(-\log \sigma)\subseteq \supp\sigma$. A similar definition applies to $\log \rho$ also, and $\log \rho$ defined on $D(\log \rho)\subseteq \supp \rho$ is a selfadjoint operator with spectrum on the negative half of the real line. It may be observed that $\log \rho $ (or $-\log \sigma)$ as defined above is bounded if and only if $\rho$ (or $\sigma$) has only a finite number of nonzero eigenvalues.
\begin{eg}[$\supp \rho \subseteq \supp \sigma$, but the operator $(\log\rho -\log\sigma)$ does not make sense and $\rho^{1/2}(-\log\sigma)\rho^{1/2}$ is not densely defined]\label{eg:counter-example} \label{eg:support-condition-not-enough}
Let $\{v_j\}_{j=1}^{\infty}$ be an orthonormal basis of $\mathcal{K} $. Take
\begin{align*}
\sigma&= \sum_{j=1}^{\infty} 2^{-j}\ketbra{v_j}\\
\rho &= \ketbra{u_1}, \textnormal{ where } u_1 =\left(\sum_{j=1}^{\infty}\frac{1}{j^2}\right)^{-1/2}\sum_{j=1}^{\infty}\frac{1}{j}\ket{v_j}.
\end{align*} Then $\supp \sigma = \mathcal{K} $ and
the condition $\supp \rho \subseteq \supp \sigma$ is trivially satisfied. Note that
\[\sum_{j=1}^{\infty}\abs{\log 2^{-j}}^2\abs{\braket{u_1}{v_j}}^2 =\left(\sum_{j=1}^{\infty}\frac{1}{j^2}\right)^{-1/2} \sum_{j=1}^{\infty}\frac{j^2}{j^2}=\infty, \] so by (\ref{eq:domain-logsigma}),
$u_1\notin D(-\log \sigma)$. But for any $ v\in \mathcal{K} $ with $v\neq 0$, $\rho^{1/2}v = \braket{u_1}{v}\ket{u_1}$ and hence $\rho^{1/2}v\in D(-\log \sigma)$ if and only if $v\in \ker \rho$. Therefore, \begin{enumerate}
\item $ D(\rho^{1/2}(-\log\sigma) \rho^{1/2}) = \ker \rho$ and hence it is not densely defined;
\item $\log \rho$ is defined only on $\spn \{u_1\}$ and $\log \rho =0$ on $\spn \{u_1\}$;
\item $D(\log \rho -\log \sigma) =\{0\}$;
\item $\rho^{1/2}(-\log\sigma) \rho^{1/2}=0$ on $\ker \rho$ and $\rho^{1/2}(-\log\rho)\rho^{1/2} =0 $.
\end{enumerate}
Thus we have \begin{align}
\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2} = 0.
\end{align}
\end{eg}
Let $\mathcal{K} $ be a finite or infinite dimensional Hilbert space. Let $\rho $ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}).
Assume that \begin{align}
\label{eq:support-condition-infinite-dim}
\spn \{u_i\,|\,r_i\neq 0 \}\subseteq \spn \{v_j\,|\,s_j\neq 0\}.
\end{align}
Now by the first half of Theorem \ref{sec:gaussian-states-prop}, $\rho^{1/2}(-\log \sigma)\rho^{1/2}$ first defined on the dense domain $\spn\{u_i\}_{i\in \mathcal{I}} $ extends as a positive (selfadjoint) operator on a dense domain in $\mathcal{K} $. Next theorem shows that our definition of the von-Neumann entropy coincides with the standard definition of the same in the finite dimensional setting, and also generalizes nicely to the infinite dimensional setting, whenever the quantities involved do make sense.
\begin{thm}\label{thm:von-Neumann-entropy}Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}).
Then
\begin{align}\label{eq:vn-entropy}
\lim_{\alpha\uparrow 1} D_{\alpha}(\rho||\sigma)= \begin{cases}
\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}, &\text{when } (\ref{eq:support-condition-infinite-dim}) \text{ is satisfied} \\
&\text{and at least one of the }\\
&\text{quantities }\\
&\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}\\
&\textnormal{or } \\&\tr \rho^{1/2}(-\log\rho)\rho^{1/2} \text{is finite;}\\
&\\
\infty, & \text{when } \supp \rho \nsubseteq \supp \sigma.\\
\end{cases}
\end{align}
In particular, if $\dim \mathcal{K} <\infty $, then \begin{align}
\lim_{\alpha\uparrow 1} D_{\alpha}(\rho||\sigma) =\begin{cases}
\tr \rho(\log \rho -\log \sigma),& \text{ when } \supp \rho \subseteq \supp \sigma;\\
\infty, & \text{otherwise}.
\end{cases}
\end{align}
Moreover, if $D_1(\rho||\sigma) = \infty$ or there exists $\beta>1$ such that $D_{\beta}(\rho||\sigma)<\infty$, then also \begin{align}\label{eq:vn-entropy-equals-right-limit}
D_1(\rho||\sigma) = \lim\limits_{\alpha\downarrow 1}D_{\alpha}(\rho||\sigma).
\end{align}
\end{thm}
\begin{proof}
Assume first that (\ref{eq:support-condition-infinite-dim}) is satisfied. By the remarks after (\ref{eq:support-condition-infinite-dim}) the operators $\rho^{1/2}(-\log\sigma)\rho^{1/2}$ and $\rho^{1/2}(-\log \rho)\rho^{1/2}$ are densely defined positive operators. We have
\begin{align*}
\tr \rho^{1/2}(-\log\sigma)\rho^{1/2} &= \sum\limits_{i}\mel{u_i}{\rho^{1/2}(-\log\sigma)\rho^{1/2}}{u_i}\\
&=\sum\limits_{i}r_i\mel{u_i}{(-\log\sigma)}{u_i}\\
&=\sum\limits_{i}\sum\limits_{\{j\,|\, s_j\neq 0\}}r_i(-\log s_j)\mel{u_i}{(\ketbra{v_j})}{u_i}\\
&=\sum\limits_{i}\sum\limits_{\{j\,|\, s_j\neq 0\}}-r_i(\log s_j)\abs{\braket{u_i}{v_j}}^2.
\end{align*}
By Remark \ref{rmk:support-condition}, Lemma \ref{lem:support-condition} is satisfied in our case. Now by using the convention $0\cdot\infty = 0$, we can sum over all possible $j$'s in the last sum above and write \begin{align}\label{eq:first-term-vn-entropy}
\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}= \sum\limits_{i,j}-r_i(\log s_j)\abs{\braket{u_i}{v_j}}^2.
\end{align}
Similarly, we also have $\tr \rho^{1/2}(-\log\rho)\rho^{1/2}= \sum\limits_{i}-r_i(\log r_i)\abs{\braket{u_i}{v_j}}^2.$ Since $\sum_j\abs{\braket{u_i}{v_j}}^2 = 1$ for all $i$, we can write
\begin{align}\label{eq:second-term-vn-entropy}
\tr \rho^{1/2}(-\log\rho)\rho^{1/2}= \sum\limits_{i,j}-r_i(\log r_i)\abs{\braket{u_i}{v_j}}^2.
\end{align}
By (\ref{eq:first-term-vn-entropy}) and (\ref{eq:second-term-vn-entropy}), we have \begin{align}\label{eq:vn-entropy-formula-before-clubbing the terms}
& \tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}\nonumber\\
&\phantom{...........}= \sum\limits_{i,j}-r_i(\log s_j)\abs{\braket{u_i}{v_j}}^2 -\sum\limits_{i,j}-r_i(\log r_i)\abs{\braket{u_i}{v_j}}^2\nonumber\\
&\phantom{...........}=\sum\limits_{i,j} r_i(\log r_i)\abs{\braket{u_i}{v_j}}^2-\sum\limits_{i,j}r_i(\log s_j)\abs{\braket{u_i}{v_j}}^2.\end{align}
If at least one of the quantities $\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}$ or $\tr\rho^{1/2}(-\log\rho)\rho^{1/2}$ are finite,we can combine the two summations above to write,\[\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2} = \sum\limits_{i,j} r_i(\log r_i)\abs{\braket{u_i}{v_j}}^2-r_i(\log s_j)\abs{\braket{u_i}{v_j}}^2.\] By using Lemma \ref{lem:support-condition}, and the conventions $\frac{0}{0}=0$, $\frac{x}{0}=\infty$ if $x>0$, $0\cdot\infty = 0$, to get
\begin{align*}
\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}
& =\sum_{i,j}r_i\abs{\braket{u_i}{v_j}}^2\log\frac{r_i}{s_j}\\
&= \sum_{i,j}r_i\abs{\braket{u_i}{v_j}}^2\log\frac{r_i\abs{\braket{u_i}{v_j}}^2}{s_j\abs{\braket{u_i}{v_j}}^2}\\
&= D(P||Q),
\end{align*}where $D(P||Q)$ is the Kullback-Leibler divergence of $P$ and $Q$,
where $P$ and $Q$ are as in (\ref{eq:P-and-Q}). Now by
Theorem \ref{thm:5-veh} we get
\begin{align*}
D_1(\rho||\sigma) = D(P||Q) = \tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}
\end{align*}
whenever (\ref{eq:support-condition-infinite-dim}) is satisfied and at least one of the quantities $\tr \rho^{1/2}(-\log\sigma)\rho^{1/2}$ or $\tr\rho^{1/2}(-\log\rho)\rho^{1/2}$ is finite.
To prove the second equality in (\ref{eq:vn-entropy}), assume that $\supp \rho \nsubseteq \supp \sigma$. Then Proposition \ref{prop:support-condition-iff-absolute-continuity} shows that $P\centernot{\ll} Q$, and hence by equation (\ref{eq:kullback-leibler-equals-infty}), and Theorem \ref{thm:5-veh} the Kullback-Leibler divergence $D(P||Q) = \infty = \lim_{\alpha\uparrow} D_\alpha(P||Q) = D_1(\rho||\sigma)$ in this case.
Finally, equation (\ref{eq:vn-entropy-equals-right-limit}) follows from Theorem \ref{thm:5-veh} and Remark \ref{rmk:5-veh-quantum}
\end{proof}
\begin{rmk}
\label{rmk:support-condition}
Note that (\ref{eq:support-condition-infinite-dim}) is same as the condition $\supp\rho\subseteq\supp\sigma$ when $\dim \mathcal{K} $ is finite. Also, the condition (\ref{eq:support-condition-infinite-dim}) implies $\supp\rho\subseteq\supp\sigma$. Hence Lemma \ref{lem:support-condition} holds in this case as well.
\end{rmk}
\begin{eg}\label{eg:infty-infty} If $\rho$ is any state such that $\tr \rho^{1/2}(-\log\rho)\rho^{1/2} = -\sum_{\{i\,:\, r_i\neq 0\}} r_i \log r_i= \infty$, then $ \tr \rho^{1/2}(-\log\rho)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}= \infty -\infty$. A specific example of such a state is obtained by taking $\rho = \sum_{i}r_i\ketbra{u_i}{u_i}$ where $\{u_i\}$ is an orthonormal basis and $r_i=\left(\sum_j \frac{1}{j(\log j)^2} \right)^{-1} \frac{1}{i(\log i)^2}$.
\end{eg}
Example \ref{eg:infty-infty} suggests that it is impossible to define the von-Neumann entropy in the infinite dimensional setting by $D(\rho||\sigma) = \tr \rho^{1/2}(-\log\sigma)\rho^{1/2}-\tr\rho^{1/2}(-\log\rho)\rho^{1/2}$, even when $(\ref{eq:support-condition-infinite-dim})$ is satisfied because this can lead to $D(\rho||\sigma)= \infty-\infty$, in certain situations.
Finally, in the light of Theorem \ref{thm:von-Neumann-entropy} we propose the following definition for von Neumann relative entropy.
\begin{defn}\label{defn:vn-entropy}
Define the von-Neumann relative entropy $D(\rho||\sigma)$ of two states $\rho$ and $\sigma$ by
\begin{align}
D(\rho||\sigma) = D_1(\rho||\sigma):= \lim_{\alpha \uparrow 1}D_{\alpha}(\rho||\sigma).
\end{align}
\end{defn}
\begin{rmk}
\label{rmk:5-veh-quantum} Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}) and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}). By Theorem \ref{thm:5-veh} and Theorem \ref{thm:quantum-reduces-to-classical}, the von-Neumann entropy of $\rho$ and $\sigma$ is the Kullback-Leibler divergence of $P$ and $Q$, i.e., we have \begin{align}\label{eq:5-veh-quantum}
D(\rho||\sigma) = D(P||Q).
\end{align}
\end{rmk}
\subsubsection{Example: Araki's Relative Entropy is different from Information Theoretic Relative Entropy when \texorpdfstring{\ensuremath{\dim \mathcal{K} =\infty}}{}}\label{sec:counter-eg}
In this section we discuss a counter example to Theorem 20 of \cite{luczak2019relative} thus disproving that result.
Consider $\rho$ and $\sigma$ as in Example \ref{eg:counter-example}. It was shown that $\supp \rho \subseteq \supp \sigma$ and \[\tr \rho^{1/2}(\log \rho) \rho^{1/2}-\tr\rho^{1/2}(\log \sigma)\rho^{1/2}=0.\] We take the semifinite von Neumann algebra $\B{\mathcal{K} }$. If $\eta\in \B{\mathcal{K} }$ is a density operator i.e., $\eta\geq 0$, and $\tr \eta =1$, then $\eta$ can be considered as a normal state $\varphi_{\eta}$ on the von Neumann algebra $\B{\mathcal{K} }$ via its action \[X\overset{\varphi_{\eta}}{\longmapsto} \tr \eta X, \quad \forall X\in \B{\mathcal{K} }.\] We will show that Araki's relative entropy $S(\varphi_{\rho}||\varphi_\sigma)$, of $\rho$ and $\sigma$ as in Example \ref{eg:counter-example} is\[ S(\varphi_{\rho}||\varphi_\sigma) = \infty\]
in this case.
The definition of Araki's relative entropy uses the notion of relative modular operator. A quick introduction to the description of relative modular operator in our setting is provided in Appendix \ref{appendix:relative-modular}. Let $\Bt{\mathcal{K} }$ be the Hilbert space of Hilbert-Schmidt operators on $\mathcal{K} $ with scalar product given by $\braket{a}{b}_2= \tr a^{\dagger}b$ for $a,b\in \Bt{\mathcal{K} }$. If $\rho$ and $\sigma$ are density operators on $\mathcal{K} $ with $\ker \sigma =\{0\}$ then the relative modular operator $\Delta_{\rho,\sigma}$ can be seen as the closure of densely defined operator $X\sigma \mapsto \rho X$, $\forall X\in \B{\mathcal{K} }$ and $\Delta_{\rho,\sigma}$ is a unbounded positive selfadjoint operator on the Hilbert space $\left(\Bt{\mathcal{K} }, \braket{\cdot}{\cdot}_2\right)$ satisfying \begin{align}\label{eq:relative-modular-action}
\Delta_{\rho,\sigma} (X\sigma) = \rho X, \quad \forall X\in \B{\mathcal{K} }.
\end{align}
Consider two density operators $\rho$ and $\sigma$ on $\mathcal{K} $. When considered as normal states $\varphi_{\rho}$ and $\varphi_{\sigma}$ on the von Neumann algebra $\B{\mathcal{K} }$, let $E_{\rho,\sigma}$ denote the spectral measure associated with the relative modular operator $\Delta_{\rho,\sigma}$ of $\rho$ and $\sigma$. Araki defined the relative entropy \cite{araki1976relative,araki1977relative} of $\varphi_\rho$ and $\varphi_\sigma$ as
\begin{align}
\label{eq:araki-entropy}
S(\varphi_{\rho}||\varphi_\sigma):= \begin{cases}
\int\limits_{0+}^{\infty} \log \lambda \mel{\sqrt{\rho}}{E_{\rho,\sigma}(\dd \lambda)}{\sqrt{\rho}}_2,& \text{ when } \supp \rho \subseteq \supp \sigma;\\
\infty, &\text{ otherwise.}
\end{cases}
\end{align}
When $\mathcal{K} $ is finite dimensional, it is well known \cite{araki1976relative,araki1977relative,petz1986quasientropy} that \begin{align}\label{eq:araki-equals-it=finite-dim}
S(\varphi_{\rho}||\varphi_\sigma)= \begin{cases}
\tr \rho^{1/2}(\log \rho) \rho^{1/2}-\tr\rho^{1/2}(\log \sigma)\rho^{1/2},&\text{ when } \supp\rho\subseteq\supp \sigma;\\
\infty, &\text{ otherwise}.
\end{cases}
\end{align}
Recently, there have been attempts \cite{luczak2019relative} to prove a similar formula as in (\ref{eq:araki-equals-it=finite-dim}) when $\dim \mathcal{K} =\infty$. In this section, we provide a counter example to show that (\ref{eq:araki-equals-it=finite-dim}) will not hold in general when $\mathcal{K} $ is infinite dimensional.
\begin{lem}\label{lem:relative-modular-spectral}
Let $\rho$ and $\sigma$ be as in Example \ref{eg:counter-example}. Let $X_{1,j}=\ketbra{u_1}{v_j}\in \Bt{\mathcal{K} }$, $j=1,2,\dots$. Then $\{X_{1,j}\}_{j=1}^{\infty}$ is an orthonormal set in $\Bt{\mathcal{K} }$ and the spectral decomposition of $\Delta_{\rho,\sigma}$ is given by
\begin{align}\label{eq:relative-modular-spectral}
\Delta_{\rho,\sigma}= \sum_{j = 1}^{\infty} 2^j \ketbra{X_{1,j}},
\end{align}
where $\ketbra{X_{1,j}}$ is considered as a rank one projection acting on $\Bt{\mathcal{K} }$.
\end{lem}
\begin{proof}
Extend $\{u_1\}$ to $\{u_1, u_2,\dots\}$ as an orthonormal basis of $\mathcal{K} $. Define \begin{align}\label{eq:X-ell-j-defn}
X_{\ell,j} = \ketbra{u_\ell}{v_j}, \quad \forall \ell, j = 1,2,\cdots.
\end{align}
Since $\{u_\ell\}$ and $\{v_j\}$ are orthonormal bases for $\mathcal{K} $, it is easy to see that the double sequence \[\{X_{\ell,j}\}_{\ell, j=1}^{\infty} \text{ is an orthonormal basis of } \Bt{\mathcal{K} }. \]
To complete the proof, we will show the following: \begin{enumerate}
\item \label{item:counter-eg-1} $X_{\ell,j}\in D(\Delta_{\rho,\sigma})$ for all $\ell,j=1,2,\cdots$;
\item \label{item:counter-eg-2} for $\ell\neq 1$, $X_{\ell,j}\in \ker \Delta_{\rho,\sigma}$;
\item\label{item:counter-eg-3} $\Delta_{\rho,\sigma} (X_{1,j}) = 2^j X_{1,j}$ for all $j=1,2,\dots$.
\end{enumerate}
To prove \ref{item:counter-eg-1}, by (\ref{eq:relative-modular-action}) it is enough to prove that there exists $Y_{\ell,j}\in \B{\mathcal{K} }$ such that $X_{\ell,j} = Y_{\ell,j}\sigma$. Note that \begin{align}\label{eq:X-ell-j}
2^jX_{\ell,j}\sigma =2^j\ketbra{u_{\ell}}{v_j}\left(\sum_{k=1}^{\infty} 2^{-k}\ketbra{v_k}\right) =\ketbra{u_{\ell}}{v_j}= X_{\ell,j},\quad \forall \ell,j=1,2,\dots.
\end{align}
Now to prove \ref{item:counter-eg-2} and \ref{item:counter-eg-3}, note that by (\ref{eq:X-ell-j}) and (\ref{eq:relative-modular-action})
\begin{align*}
\Delta_{\rho,\sigma} (X_{\ell,j}) &= \Delta_{\rho,\sigma} (2^jX_{\ell,j}\sigma)=2^j\rho X_{\ell,j}\\
&=2^j\ketbra{u_1}\ketbra{u_{\ell}}{v_j}\\
&=\begin{cases}
2^j\ketbra{u_1}{v_j}, & \text{ if } \ell=1\\
0, &\text{ if } \ell\neq 1.
\end{cases}
\end{align*}
Thus $\Delta_{\rho,\sigma} (X_{1,j}) = 2^j X_{1,j}$ for all $j=1,2,\dots$ and $\Delta_{\rho,\sigma} (X_{\ell,j}) =0$ when $\ell\neq 1$.
\end{proof}
\begin{cor} Let $\rho$ and $\sigma$ be as in Example \ref{eg:counter-example} and $X_{\ell,j}$ be as in (\ref{eq:X-ell-j-defn}).
The spectral measure $E_{\rho,\sigma}$ of $\Delta_{\rho,\sigma}$ is supported on the set $\{0,2,2^2,\dots\}$ and \begin{align}
\label{eq:spectral-measure-relative-modular}
&E_{\rho,\sigma}\{0\} \phantom{.}=\sum_{\ell=2}^\infty\sum_{j=1}^{\infty}\ketbra{X_{\ell,j}},\\
&E_{\rho,\sigma}\{2^j\} =\ketbra{X_{1,j}} \quad j=1,2,\dots.
\end{align}
\end{cor}
\begin{proof}
Follows easily from Lemma \ref{lem:relative-modular-spectral} and the fact that $\{X_{\ell,j}\}_{\ell, j=1}^{\infty}$ forms an orthonormal basis of $\Bt{\mathcal{K} }$.
\end{proof}
\begin{thm}\label{thm:counter-example}
Let $\rho$ and $\sigma$ be as in Example \ref{eg:counter-example}, then Araki's relative entropy of $\rho$ and $\sigma$ is \[S(\varphi_{\rho}||\varphi_\sigma) = \infty,\] but
\[\tr \rho^{1/2}(\log \rho) \rho^{1/2}-\tr\rho^{1/2}(\log \sigma)\rho^{1/2}=0.\]
\end{thm}
\begin{proof} The second equality in the statement of the theorem is already shown in Example \ref{eg:counter-example}. To prove the first equality we will evaluate the integral in (\ref{eq:araki-entropy}). To this end, fix $j\in \mathbb{N}$ and note that \begin{align*}
\mel{\sqrt{\rho}}{E_{\rho,\sigma}\{2^j\}}{\sqrt{\rho}}_{2}&= \mel{\sqrt{\rho}}{\left(\ketbra{X_{1,j}}\right)}{\sqrt{\rho}}_{2}\\
&=\braket{\sqrt{\rho}}{X_{1,j}}_2\braket{X_{1,j}}{\sqrt{\rho}}_2\\
&=\abs{\braket{\sqrt{\rho}}{X_{1,j}}_2}^2\\
&=\abs{\tr {\sqrt{\rho}}{X_{1,j}}}^2.
\end{align*}
Recall that $\rho=\ketbra{u_1}=\sqrt{\rho}$ and $X_{1,j}=\ketbra{u_1}{v_j}$. Therefore, \begin{align*}
\mel{\sqrt{\rho}}{E_{\rho,\sigma}\{2^j\}}{\sqrt{\rho}}_{2} = \abs{\braket{u_1}{v_j}}^2.
\end{align*}
Finally, since $u_1=\left(\sum_{k=1}^{\infty}\frac{1}{k^2}\right)^{-1/2}\sum_{k=1}^{\infty}\frac{1}{k}\ket{v_k}$, we have \begin{align}\label{eq:measure-araki-entropy}
\mel{\sqrt{\rho}}{E_{\rho,\sigma}\{2^j\}}{\sqrt{\rho}}_{2} = \left(\sum_{k=1}^{\infty}\frac{1}{k^2}\right) \frac{1}{j^2}.
\end{align}
Now \begin{align*}
S(\varphi_{\rho}||\varphi_\sigma)&=
\int\limits_{0+}^{\infty} \log \lambda \mel{\sqrt{\rho}}{E_{\rho,\sigma}(\dd \lambda)}{\sqrt{\rho}}_2\\
&=\sum_{j=1}^{\infty} \log 2^j\left(\sum_{k=1}^{\infty}\frac{1}{k^2}\right) \frac{1}{j^2}\\
&=\left(\sum_{k=1}^{\infty}\frac{1}{k^2}\right)\sum_{j=1}^{\infty} \frac{j}{j^2}=\infty,
\end{align*}which completes the proof.
\end{proof}
\subsubsection{Quantum Pinsker Type Inequality}
\begin{lem}\label{lem:total-variation}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}) and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}), then the total variation distance $V(P,Q)$ between $P$ and $Q$ is given by \begin{align}
V(P,Q) = \sum_{i,j}\abs{r_i-s_j}\abs{\braket{u_i}{v_j}}^2
\end{align}
\end{lem}
\begin{proof}
By the definition of the total variation distance as in (\ref{eq:total-variation}), we have \begin{align*}
V(P,Q) &= \sum_{x\in \mathcal{I}\times \mathcal{I}}\abs{\frac{P(x)}{\mu(x)}-\frac{Q(x)}{\mu(x)}}\mu(x) = \sum_{x\in \mathcal{I}\times \mathcal{I}}\abs{P(x)-Q(x)}\\
&=\sum_{i,j}\abs{r_i-s_j}\abs{\braket{u_i}{v_j}}^2,
\end{align*}
which completes the proof.
\end{proof}
\begin{lem}\label{lem:tr-rho-sigma-square}
Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}), and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}), then \begin{align}
\norm{\rho-\sigma}_2^2 = \sum_{i,j}(r_i-s_j)^2\abs{\braket{u_i}{v_j}}^2\leq V(P,Q),
\end{align}
where $\norm{\cdot}_2$ denotes the Hilbert-Schmidt norm.
\end{lem}
\begin{proof}
Note that \begin{align}\label{eq:tr-rho-sigma-square}
\norm{\rho-\sigma}_2^2 = \tr (\rho-\sigma)^2= \tr \rho^2+\tr\sigma^2 -2\tr \rho\sigma.
\end{align}
By (\ref{eq:spectral-rho-sigma}), \[\rho\sigma = \sum_{i,j}r_is_j\braket{u_i}{v_j}\ketbra{u_i}{v_j}\]
is a trace class operator. Hence \[\tr \rho\sigma = \sum_{i,j}r_is_j\abs{\braket{u_i}{v_j}}^2. \]
Along with the fact that $\tr\rho^2 = \sum_{i}r_i^2$ and $\tr \sigma^2 = \sum_j s_j^2$, we get from (\ref{eq:tr-rho-sigma-square}),\begin{align*}
\norm{\rho-\sigma}_2^2 &= \sum_{i}r_i^2+ \sum_j s_j^2-2\sum_{i,j}r_is_j\abs{\braket{u_i}{v_j}}^2\\
&=\sum_{i,j}r_i^2\abs{\braket{u_i}{v_j}}^2+ \sum_{i,j} s_j^2\abs{\braket{u_i}{v_j}}^2-2\sum_{i,j}r_is_j\abs{\braket{u_i}{v_j}}^2\\
&=\sum_{i,j} (r_i-s_j)^2\abs{\braket{u_i}{v_j}}^2.
\end{align*}
To prove the last inequality in (\ref{eq:tr-rho-sigma-square}) note that $r_i,s_j\in [0,1]$ for all $i,j$ and thus we have \[(r_i-s_j)^2\leq\abs{r_i-s_j}, \quad \forall i,j.\]
Lemma \ref{lem:total-variation} completes the proof.
\end{proof}
\begin{thm}[Quantum Pinsker Type Inequality]\label{thm:quantum-pinsker}
For $\alpha\in (0,1]$, the Petz-R\'enyi relative entropy satisfies the quantum Pinsker's inequality, \begin{align}
\frac{\alpha}{2}\norm{\rho-\sigma}_2^4\leq D_{\alpha}(\rho||\sigma),
\end{align}
where $\norm{\cdot}_2$ denotes the Hilbert-Schmidt norm.
\end{thm}
\begin{proof} Let $\rho$ and $\sigma$ be as in (\ref{eq:spectral-rho-sigma}) and let $P$ and $Q$ be as in (\ref{eq:P-and-Q}).
By Theorem \ref{thm:31-veh} and Theorem \ref{thm:quantum-reduces-to-classical} we have\[\frac{\alpha}{2}V^2(P,Q)\leq D_{\alpha}(P||Q) =D_{\alpha}(\rho||\sigma) ,\]
where $V(P,Q)$ denote the total variation distance between $P$ and $Q$. Lemma \ref{lem:tr-rho-sigma-square} completes the proof.
\end{proof}
\begin{rmk}
The quantum Pinsker's inequality presented above bounds the relative Petz-R\'{e}nyi entropy from
below by a function of the Hilbert-Schmidt norm of the difference of two quantum states. Usually, in the Pinsker's type inequalities the relative Petz-Renyi entropy between two states on a finite dimensional Hilbert space
is bounded from below by an increasing function of their von Neumann relative entropy, which in turn, is bounded from below by a
function of the trace class norm of their difference (for example one may refer to equation (43) in \cite{rastegin-2013}). Our inequality is weaker since we only use the Hilbert-Schmidt norm, but on the
other hand, our inequality is obtained using only the classical Pinsker's inequality on the R\'{e}nyi relative entropy.
\end{rmk}
\subsubsection{Continuity, Positivity, Symmetry and Concavity}
We begin this section with three examples which illustrate behaviour of $D_{\alpha}(\rho||\sigma)$ when $\alpha\geq 1.$ These examples will help us to understand continuity points of $D_{\alpha}(\rho||\alpha)$
\begin{eg}[$D_1(\rho||\sigma)<\infty$ but $D_{\alpha}(\rho||\sigma) = \infty, \forall \alpha >1$] Let $\{u_i\}_{i=1}^{\infty}$ be any orthonormal basis on $\mathcal{K} $. Take \begin{align*}
\rho &= \sum_{i=1}^{\infty}2^{-i}\ketbra{u_i}\\
\sigma & =s^{-1} \sum_{j=1}^{\infty}2^{-j^2}\ketbra{u_j},
\end{align*}
where $s=\left(\sum_j 2^{-j^2}\right)^{1/2}$. In this case, keeping the notations in (\ref{eq:spectral-rho-sigma}), and (\ref{eq:P-and-Q}), $r_i=2^{-i}$, $s_j=s^{-1}2^{-j^2}$, $\braket{u_i}{v_j}= \delta_{i,j}$, we have
\begin{align*}
D_1(\rho||\sigma) = D(P||Q)&= \sum_{i}r_i\log \left( \frac{r_i}{s_i}\right) = \sum_{i}2^{-i}\log \left( \frac{s2^{-i}}{2^{-i^2}}\right)\\
&= \sum_{i}2^{-i} \log \left(s2^{-i+i^2}\right)= \sum_{i}2^{-i} \log s+ \sum_{i}2^{-i} \log \left(2^{-i+i^2}\right) \\
&= \sum_{i}2^{-i} \log s + \sum_{i}2^{-i}(-i+i^2) <\infty.
\end{align*}
On the other hand, for $\alpha>1$, \begin{align*}
\sum_{i}r_i^{\alpha}s_i^{(1-\alpha)} &=s^{-1(1-\alpha)} \sum_{i}2^{-\alpha i}2^{-(1-\alpha)i^2} = s^{-1(1-\alpha)}\sum_{i}2^{(\alpha-1)i^2-\alpha i}=\infty
\end{align*}
because $(\alpha-1)>0$. Therefore,
\begin{align*}
D_{\alpha}(\rho||\sigma) = D_{\alpha}(P||Q)&= \frac{1}{\alpha-1} \log \sum_{i}r_i^{\alpha}s_i^{1-\alpha} = \infty.
\end{align*}
\end{eg}
\begin{eg}[$D_{\alpha}(\rho||\sigma)<\infty$ for $1<\alpha< 2$ but $D_{2}(\rho||\sigma) = \infty$] Let $\{u_i\}_{i=1}^{\infty}$ be any orthonormal basis on $\mathcal{K} $. Take \begin{align*}
\rho &= \sum_{i=1}^{\infty}2^{-i}\ketbra{u_i}\\
\sigma & =s^{-1} \sum_{j=1}^{\infty}2^{-2i}\ketbra{u_j},
\end{align*}
where $s=\left(\sum_j 2^{-2i}\right)^{1/2}$. In this case, keeping the notations in (\ref{eq:spectral-rho-sigma}), and (\ref{eq:P-and-Q}), $r_i=2^{-i}$, $s_j=s^{-1}2^{-2i}$, $\braket{u_i}{v_j}= \delta_{i,j}$. We have
for $\alpha>1$, \begin{align*}
\sum_{i}r_i^{\alpha}s_i^{(1-\alpha)} &=s^{-1(1-\alpha)} \sum_{i}2^{-\alpha i}2^{-(1-\alpha)2i} = s^{-1(1-\alpha)}\sum_{i}2^{(\alpha-1)2i-\alpha i}= s^{-1(1-\alpha)}\sum_{i}2^{(\alpha-2)i}.
\end{align*}
The above series converges for $1<\alpha<2$ and diverges for $\alpha=2$. Therefore,
\begin{align*}
D_{\alpha}(\rho||\sigma) = D_{\alpha}(P||Q)&= \frac{1}{\alpha-1} \log \sum_{i}r_i^{\alpha}s_i^{1-\alpha}
\end{align*}
is finite for $1<\alpha<2$ and diverges for $\alpha=2$.
\end{eg}
\begin{eg}[$D_{2}(\rho||\sigma)<\infty$ but $D_{\alpha}(\rho||\sigma) = \infty$ for $\alpha>2$] Let $\{u_i\}_{i=1}^{\infty}$ be any orthonormal basis on $\mathcal{K} $. Take \begin{align*}
\rho &= \sum_{i=1}^{\infty}2^{-i}\ketbra{u_i}\\
\sigma & =s^{-1} \sum_{j=1}^{\infty}i^{2}2^{-2i}\ketbra{u_j},
\end{align*}
where $s=\left(\sum_j i^{2}2^{-2i}\right)^{1/2}$. In this case, keeping the notations in (\ref{eq:spectral-rho-sigma}), and (\ref{eq:P-and-Q}), $r_i=2^{-i}$, $s_j=s^{-1}i^{2}2^{-2i}$, $\braket{u_i}{v_j}= \delta_{i,j}$. We have
for $\alpha\geq 2$, \begin{align*}
\sum_{i}r_i^{\alpha}s_i^{(1-\alpha)} &=s^{-1(1-\alpha)} \sum_{i}2^{-\alpha i}i^{2(1-\alpha)}2^{-(1-\alpha)2i}\\
&= s^{-1(1-\alpha)}\sum_{i}i^{2(1-\alpha)}2^{(\alpha-1)2i-\alpha i}\\
&= s^{-1(1-\alpha)}\sum_{i}i^{2(1-\alpha)}2^{(\alpha-2)i}.
\end{align*}
The above series converges for $\alpha=2$ and diverges for $\alpha>2$. Therefore,
\begin{align*}
D_{\alpha}(\rho||\sigma) = D_{\alpha}(P||Q)&= \frac{1}{\alpha-1} \log \sum_{i}r_i^{\alpha}s_i^{1-\alpha}
\end{align*}
is finite for $\alpha=2$ and diverges for $\alpha>2$.
\end{eg}
\begin{thm}
\label{thm:7-veh-quantum}
The Petz-R\'enyi relative entropy $D_\alpha (\rho||\sigma)$ is continuous in $\alpha$ on $\mathcal{A} = \{\alpha\in [0,\infty]\,|\, 0\leq \alpha\leq 1\textnormal{ or } D_{\alpha}(\rho||\sigma)<\infty\}$.
\end{thm}
\begin{proof} By Theorem \ref{thm:quantum-reduces-to-classical} we know $D_\alpha (\rho||\sigma) = D_\alpha (P||Q)$ where $\rho$ and $\sigma$ are as in (\ref{eq:spectral-rho-sigma}) and $P$ and $Q$ are as in (\ref{eq:P-and-Q}). The result follows because the same result is true for the classical R\'enyi relative divergence (refer Theorem \ref{thm:5-veh}).
\end{proof}
\begin{thm}
\label{thm:8-veh-quantum} For any order $\alpha\in [0,\infty]$,
\[D_{\alpha}(\rho||\sigma)\geq 0.\] For $\alpha>0$, $D_{\alpha}(\rho||\sigma) = 0$ if and only if $\rho=\sigma$. For $\alpha = 0$, $D_{\alpha}(\rho||\sigma) = 0$ if and only if $\supp \sigma \subseteq \supp \rho$.
\end{thm}
\begin{proof}
Follows easily from Theorem \ref{thm:8-veh} and Proposition \ref{prop:rho-equals-sigma-P-equals-Q} because of Theorem \ref{thm:quantum-reduces-to-classical}.
\end{proof}
\begin{prop}
\label{prop:8-veh-quantum}
For any $0<\alpha<1$, the Petz-R\'enyi relative entropy shows the following skew-symmetry property \[D_\alpha(\rho||\sigma) = \frac{\alpha}{1-\alpha}D_{1-\alpha}(\sigma||\rho).\]
\end{prop}
\begin{proof}
Follows from Proposition \ref{prop:8-veh}.
\end{proof}
Note that in particular, Petz-R\'enyi relative entropy is symmetric for $\alpha=1/2$, and that skew-symmetry does not hold for $\alpha = 0$ and $\alpha = 1$.
\begin{thm}
\label{thm:16-veh-quantum}
For any $0<\alpha\leq \beta <1$,\[\frac{\alpha}{\beta}\frac{1-\beta}{1-\alpha}D_{\beta}(\rho||\sigma)\leq D_\alpha (\rho||\sigma)\leq D_{\beta}(\rho||\sigma).\]
\end{thm}
\begin{proof}
Follows from Theorems \ref{thm:16-veh} and \ref{thm:quantum-reduces-to-classical}.
\end{proof}
\begin{rmk}In the light of the Theorem~\ref{thm:16-veh-quantum} we can discuss about a topology on the set of states arising from $D_{\alpha}$.
For a fixed $\alpha \in (0,1)$, one can define $\alpha$-left open ball with center $\rho$ and radius $r>0$ to be the set
$\{ \sigma\,|\, D_\alpha(\rho||\sigma) <r\}$, and subsequently define $\alpha$-left open sets to be the union of $\alpha$-left open balls.
Notice that Theorem~\ref{thm:16-veh-quantum} yields that for $\alpha, \beta \in (0,1)$, the $\alpha$-left topology is equivalent to the
$\beta$-left topology. Similarly, one can define $\alpha$-right open balls and $\alpha$-right topologies by reversing the order
of $\rho$ and $\sigma$ in the definition of $\alpha$-left topology. Proposition~\ref{prop:8-veh-quantum}, combined with the fact that the
$\alpha$-left topologies are all equivalent for $0<\alpha<1$, gives that the $\alpha$-left topologies are equivalent with the
$\beta$-right topologies for all $\alpha, \beta \in (0,1)$.
\end{rmk}
\begin{thm}
\label{thm:23-veh-quantum}The following conditions are equivalent:
\begin{enumerate}
\item \label{item:vEH-1-quantum}$\supp \sigma \subseteq \supp \rho,$
\item \label{item:vEH-2-quantum} $\tr \Pi_\rho\sigma = 1$, where $\Pi_\rho$ is the orthogonal projection onto $\supp \rho$.
\item\label{item:vEH-3-quantum} $D_0(\rho||\sigma) = 0$,
\item\label{item:vEH-4-quantum} $\lim_{\alpha \downarrow 0} D_{\alpha}(\rho||\sigma) = 0.$
\end{enumerate}
\end{thm}
\begin{proof}
\ref{item:vEH-1-quantum} $\Leftrightarrow$ \ref{item:vEH-3-quantum} follows from Theorem \ref{thm:8-veh-quantum}.
\ref{item:vEH-3-quantum} $\Leftrightarrow$ \ref{item:vEH-2-quantum} follows from Theorem \ref{thm:4-veh-quantum}.
\ref{item:vEH-3-quantum} $\Leftrightarrow$ \ref{item:vEH-4-quantum} follows from Theorem \ref{thm:7-veh-quantum}.
\end{proof}
\begin{thm}
\label{thm:24-veh-quantum}The following conditions are equivalent:
\begin{enumerate}
\item \label{item:vEH-5-quantum}$\supp \rho\perp \supp \sigma,$
\item \label{item:vEH-6-quantum} $\tr \Pi_\rho\sigma = 0$, where $\Pi_\rho$ is the orthogonal projection onto $\supp \rho$,
\item\label{item:vEH-7-quantum} $D_{\alpha}(\rho||\sigma) = \infty$ for some $\alpha\in [0,1)$,
\item\label{item:vEH-8-quantum} $D_{\alpha}(\rho||\sigma) = \infty$ for all $\alpha\in [0,\infty].$
\end{enumerate}
\end{thm}
\begin{proof}
\ref{item:vEH-5-quantum} $\Leftrightarrow$ $\supp \sigma \subseteq \ran (I-\Pi_{\rho})$ $\Leftrightarrow$ $\tr (I-\Pi_{\rho})\sigma = \tr \sigma =1$ $\Leftrightarrow$ the statement \ref{item:vEH-6-quantum}.
By Theorem \ref{thm:4-veh-quantum}, the statement \ref{item:vEH-6-quantum} $\Leftrightarrow$ $D_0(\rho||\sigma) = \infty$ $\Leftrightarrow$ statement \ref{item:vEH-8-quantum} by Theorem \ref{thm:3-veh-quantum}.
Finally, \ref{item:vEH-7-quantum} $\Leftrightarrow$ \ref{item:vEH-8-quantum} because of Theorem \ref{thm:16-veh-quantum}
\end{proof}
\begin{prop}\label{cor:2-veh-quantum}
The function $[0, \infty] \ni \alpha \mapsto (1-\alpha)D_{\alpha}(\rho||\sigma)$ is concave, with the conventions that it is $0$ at $\alpha = 1$ even if $D(\rho||\sigma) = \infty$ and that it is $0$ at $\alpha = \infty$ if $\rho=\sigma$.
\end{prop}
\begin{proof}
Follows from Corollary \ref{cor:2-veh} and Theorem \ref{thm:quantum-reduces-to-classical}.
\end{proof}
\section{Conclusion and Discussion}
The major contributions of this article are: (1) we provide a procedure for obtaining results related to Petz-R\'enyi relative entropy using the classical results available for R\'enyi divergence of probability measures; and (2) we give an example to show that Araki's definition of von Neumann relative entropy in infinite dimensions does not agree with its usual information theoretic definition.
We illustrated (1) by proving several results about Petz-R\'enyi relative entropy from the classical results available in the survey article \cite{Erven-Harremos-2014}. All our results work both in finite and infinite dimensional setting which is a strength of this method. Hence these results are particularly useful in continuous variable quantum information theory as well. For instance, we use the results of the current work in order to study the Petz-R\'enyi relative entropy of gaussian states in a follow-up article \cite{androulakis-john-2022}. There have been several recent new results in the classical theory of R\'enyi divergences after the publication of \cite{Erven-Harremos-2014}. This shows that more results about Petz-R\'enyi relative entropy can be proved using the Theorem \ref{thm:quantum-reduces-to-classical} of this article. This method directly works for proving a quantum counterpart of any result about the classical R\'enyi divergence involving only two probability distributions. An interesting open problem is to develop methods in order to overcome the technical difficulties one encounters while trying to obtain a quantum version of a result starting from a classical result in which more than two probability distributions are involved. One example of such result, that may be explored is the data processing inequality which is satisfied by many other entropic quantities.
\section*{Acknowledgement}
The second author thanks The Fulbright Scholar Program and United States-India Educational Foundation for providing funding and other support to conduct this research through a Fulbright-Nehru Postdoctoral Fellowship.
| 00c8532b8dc652c041c5c94c26f0a30f1b07a30c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
The Fornax dwarf spheroidal galaxy (dSph) is the second brightest dSph after Sagittarius in the Milky Way (MW) halo \citep{McConnachie2012}. It has been discovered using photographic plates \citep{Shapley1938a,Shapley1938b}. Fornax, together with Leo I, Sculptor, Leo II, Sextans, Carina, Ursa Minor (UMi), Sagittarius (Sgr), and Draco, are known as the classical dSphs, since they are relatively bright \citep{Walker2012,Simon2019}. They are small and faint with very low stellar density, 0.0079 $L_\odot {\rm pc}^{-3}$ \cite[about 20 times less dense than that of solar neighborhood, see][]{Hammer2019}, while their member stars show hot kinematics, i.e., large line-of-sight (\textsc{los}) velocity dispersion ($\sigma_{\textsc{los}}$), at an order of 5 to 10 km/s, \citep[e.g.,][]{Walker2009}. This may indicate large dynamical mass (i.e., dark-matter, DM) compare to their stellar mass, if they are dispersion-supported systems as speculated from their spheroidal appearance.
To understand their nature, a large number of studies have focused on the stellar structures of these dwarf galaxies to seek evidence for MW tides, tidal tails or extra-tidal debris outside the King tidal radius that is obtained by fitting the stellar density profile of a dSph. Except for Sagittarius which is an obvious case that is undergoing tidal stripping by the MW \citep{Majewski2003}, for three of other classical dSphs evidence or hints of tidal-debris were found: Leo I, Carina, and UMi \citep{Sohn2007,Munoz2006,Munoz2005,Palma2003}. Using deep photometry, \citet{Battaglia2012,Battaglia2013} argued that Carina could be the best candidate with tidal tail detection, but this is not confirmed by \citet{McMonigal2014}. The main reason for different conclusions could be the uncertainty of background determination.
In more than two-third (7/9) of the classical dSphs, a "break" has been detected in their surface density profiles.
These are
Sculptor \citep{Westfall2006},
Carina \citep{Kuhn1996,Majewski2000,Majewski2005,Munoz2006},
Leo~I \citep{Sohn2007},
Draco \citep{Wilkinson2004},
Sextans \citep{Gould1992},
UMi \citep{Kocevski2000,Matinez2001,Palma2003,Munoz2005}
and, of cause, Sgr \citep{Ibata2001,Majewski2003}.
This break corresponds to an excess of density that extends outside the King tidal radius. It has been argued \citep[][and references therein]{Westfall2006} to be evidence that MW tides have significantly affected the structure of dSphs. Note that the presence of a "break" is not always acknowledged. For example, \citet{Wilkinson2004} found an obvious break in the density profile of Draco, while \citet{Segall2007} pictured Draco as a featureless dSph, without a break in its density profile.
In fact, with a careful inspection of the results by \citet[][see their Fig. 10]{Segall2007}, one could notice a break-like feature at the same radius (25 arcmin) as that found by \citet{Wilkinson2004}. The break feature at 25 arcmin is unlikely due to an artifact of background subtraction, since at this radius the density ($\approx 1-2$~stars/arcmin$^2$) is almost an order of magnitude higher than the background ($\approx 0.1$~stars/arcmin$^2$) in both studies. One should note, however, that an over-subtraction of background may lead to an artificial cut-off of extended features.
The strong contradiction between the conclusions of \citet{Wilkinson2004} and \citet{Segall2007} could be due to the fact that outskirt density profiles of dSphs are very sensitive to background determination, as stressed by \citet{Irwin1995}.
Simulations were designed to test the effects of MW tides on dSphs. They reveal that MW tides may create significant morphological perturbations and breaks in the density profiles \citep[e.g.,][]{Read2006,Munoz2008,Penarrubia2008,Klimentowski2009,Kazantzidis2011}, but cannot inflate the central velocity dispersions to the observed values \citep{Piatek1995,Oh1995}. In order to reproduce the large $\sigma_{\textsc{los}}$, one has to assume large amounts of dark matter to dominate the gravity of dSphs, providing sufficient large velocity dispersions \citep[e.g.,][]{Munoz2008}. These simulations also support the idea that if MW tides affect the outskirts of these classical dSphs, their cores can be safely approximated by the assumption of dynamical equilibrium \citep{Simon2019,Battaglia2013,Walker2012}.
Based on this assumption, dynamical analysis using the Jeans equation have been applied to infer the total mass of dSphs via their $\sigma_{\textsc{los}}$ \citep[e.g.,][]{Walker2012,Wolf2010} or via the radial profile of $\sigma_{\textsc{los}}$ for deriving the DM distribution \citep{Walker2011}. If the large $\sigma_{\textsc{los}}$ values are due to stars in dSphs at dynamical equilibrium with DM \citep{Walker2012}, the assumed DM pushes the theoretical tidal radius well beyond the stellar extent of the dSphs.
Consequently, a DM halo "shields" its stars, including those apparent extra-tidal stars, from any external tides. In such a situation, the "break" in density profiles of dSphs could be interpreted as a second stellar component of the dwarfs, though its origin is still unclear.
Nevertheless, a consensus has been reached that a huge amount of dark-matter (DM) is required to consistently explain the large $\sigma_{\textsc{los}}$, as well as the complex star formation histories of dSphs \citep{Battaglia2013}. A direct confirmation is expected via the annihilation of DM particles \citep{Walker2012}, but no such confirmation has been reported yet.
MW dSphs have been widely recognized as the most DM dominated objects in the universe. This emphasizes their important role in the cosmological context, since it supports a prediction by Lambda cold dark matter ($\Lambda$CDM) cosmology at the smallest galactic scales, despite some debates on $\Lambda$CDM predictions, such as The Plane of Satellites Problem \citep{Pawlowski2018}, the Missing Satellites Problem, the Core-Cusp Problem (specifically on Fornax), and the Too-Big-to-Fail problem \citep{Bullock2017}.
\begin{table*}
\small
\caption{Observations and methods for studying Fornax.}
\label{tab:method}
{\centering \small
\begin{tabular}{lccccl}
\hline
\hline
& FOV (sq. deg.) & seeing (") & Bands(limiting mag) & $\Delta$mag$^{\rm a}$ & Method \\ \hline
\citet{Munoz2018} & $\approx$0.6 & $<1$ & $g$(25.6); $r$(25.3) & 6.2 & CMD \\
\citet{delPino2015} & $\approx$2 & --- & B(23);V & $\approx$5 & CMD \\
\citet{Coleman2005} & $\approx$2 & 1.8 & V(20.7); I & 2.5 & CMD \\
\citet{Battaglia2006} & $\approx$2.6 & $<1$ & V(23); I(22) & 4.5 & CMD \\
\citet{Wang2019DES} & 25 &$ \approx$1 & $g$(23.5); $r$(23.5)& 4.3 & Matched filter on CMDs\\
\citet{Bate2015} & 25 & $<1$& $g$(23.1); $r$(22.4); $i$(21.4)& 5 & Matched filter on CMDs\\
\citet{Irwin1995} & 25 & 2-3 & B(22); R(20-21)& --- & direct counting (mixing star and galaxy, and color)\\
This work & 400 & 0.7$^{\rm b}$ & G (20.8), RP, BP & 3 & CMDs + PM + Parallax\\
\hline
\end{tabular}
}\\
$^{\rm a}$ Estimate of the magnitude range of RGB population covered by observation.
$^{\rm b}$ Note that the 0.7" is the effective angular resolution of Gaia EDR3 \citep{Fabricius2021}.
\end{table*}
Previously, the success of the DM-dominated scenario for dSph was also strengthened by the 'failure' of DM-free scenarios \citep{Klessen2003,Piatek1995}, known as 'tidal scenario' that were proposed at relatively early time, such as \citet{Kuhn1989,Kuhn1993,Kroupa1997,Fleck2003,Metz2007}.
All these early propositions of DM-free scenarios assumed that dSphs are long-lived MW satellites since up to $\sim$8 to 10 Gyr ago. Such investigations, especially simulations, were done assuming a single old stellar component. With better resolved stellar populations, it has been revealed that Carina, Fornax, Leo I, UMi and LeoII have their star formation extend until recent 1 or 2 Gyrs \citep{Weisz2014,Pace2020}. In Fornax, a 100-Myr-old stellar populations has been discovered \citep{deBoer2013}, together with an HI cloud that is suspected to be associated with Fornax (Bouchard et al. 2006).
These facts suggest that their gas may have played an important role at the last stage evolution of dSph dynamics.
The impact of the gas component has been investigated by \citet{Yang2014} for the DM-free case, and by \citet{Mayer2001} and \cite{Mayer2010} for the DM-dominated case.
A recent work \citep{Hammer2019,Hammer2018} identified a strong anti-correlation between the internal acceleration ($\sigma_{\textsc{los}}^2$/$r_{\rm half}$) of dSphs and their distances to the Galactic Centre. Such an anti-correlation is unexpected in the DM scenario, since in such a case the internal acceleration is the DM gravity, for which no specific relation is expected with the Galactocentric distance. Such an anti-correlation is, however, predicted if dwarfs are experiencing MW tidal shocks in the framework of the impulse approximation for fast encounters \citep{Binney2008}. This scenario requires dSph progenitors to be gas-rich dwarf galaxies approaching the MW in the past few billion years. In such a case, after loosing their gas due to ram pressure induced by the MW halo gas, dSph stars spherically expand because of the loss of a major part of the dSph potential. Due to the spherical expansion, part of the residual stellar body (about 25\%, see \citealt{Hammer2019}) exchanges kinetic energy through MW tidal shocks, which may explain the large velocity dispersions of dSphs without the need for DM. This mechanism has been previously investigated with numerical simulations by \citet{Yang2014}. It predicts anisotropic velocity dispersions, with large values along the line of sight, which is nearly the Galactic Center direction, and small values on the axis perpendicular to the dSph orbital motion \citep{Hammer2018}.
Possibly, this is another mechanism that could explain the nature of dSphs, without invoking a dominant dark matter component.
Gaia\footnote{https://www.cosmos.esa.int/web/gaia/home} DR2 (and more recently, EDR3) reveal dSph orbital motions that appear not consistent with that of DM dominated MW satellites (or sub-haloes). This is because they lie too close to their pericenters \citep{Hammer2020,Li2021}, they are preferentially distributed into a Vast Polar Structure \citep{Pawlowski2014} within which they orbit, and their tangential velocities and angular momenta are excessively large, suggesting a recent infall \citep{Hammer2021}.
Gaia will bring a complete proper motion (PM) coverage of the whole sky, which provides an independent method (comparing to the widely used color-magnitude diagram, i.e., CMD method) to distinguish stars belonging to dwarf galaxies from that of MW because they move differently. This will be helpful to reach a deeper and more accurate background determination, allowing to further identify possible tidal debris around dSph and weak structures in dSph.
In any scenario, MW tides may play a role in shaping dSphs, but how and by how much?
Answering this fundamental question requires to reach the true edges of dSphs by eliminating most contamination from non-member stars. Besides observational depth, the requirement of a pure member sample is preponderant, i.e., removing the contamination from MW stars and background Quasi-Stellar Objects (QSOs). Member stars of dSphs in the RGB branch are bright enough to be observed by Gaia, and so one should be able to robustly retrieve the dSphs' true extent using a highly pure member sample.
In this paper, we present a case study on Fornax, using Gaia EDR3 published in December 2020 \citep{Brown2020}.
Thanks to its homogeneous coverage and data quality, we can explore the data over a very large area.
Both coverage and calibrations across large field are difficulties for ground-based and mosaic-type observations on dSphs.
\section{Methodology}
\label{sec:method}
Our goal is to search possible faint structures in the outskirts of Fornax using Gaia data.
This is a challenging task because Gaia data is limited to a relative shallow depth, i.e., $G<21$~mag, covering about a 3 magnitude range of the RGB branch of Fornax\citep{Helmi2018}.
In Table~\ref{tab:method}, we compare major photometric works that have been dedicated to the study of Fornax' structure. \citet{Munoz2018} carried out the deepest observation, but only focused on the central region. \citet{delPino2015}, \citet{Coleman2005} and \citet{Battaglia2006} carried out observations that barely cover the nominal tidal radius of Fornax. Thus, the estimation of background in these studies may have more uncertainties, as acknowledged in Battaglia et al.. \citet{Wang2019DES}, \citet{Bate2015} and \citet{Irwin1995} presented deep photometry studies covering the wider field around Fornax. All these studies, except \citet{Irwin1995}, used CMDs to pre-select member candidates of different stellar population such as RGB, red clump, etc.. In doing so, many Galactic foreground stars and compact background galaxies that differ in color from Fornax stellar populations can be excluded from the candidate sample. All studies performed star/galaxy separation and exclude extended sources before applying the CMD selection, except \citet{Irwin1995} (due to poor seeing conditions).
When using the CMD method to select member candidates, there are always non-member sources selected into the candidate sample due to their colors blending with Fornax's stellar population. The fraction of such blending is partly intrinsic and partly dependent on photometric accuracy: the better the accuracy, the lower the contamination. These contaminating sources will appear as a background in the spatial distribution which can be removed statistically. Theoretically, such a background due to contaminating sources should be smoothly distributed compared to the structure of Fornax. Even if the background is not flat, it could be modelled by fitting a surface if necessary, and the fluctuation of the background will be recognized as an uncertainty. Importantly, the uncertainty of such a background directly defines the limit to which we are able to probe faint structures of, and around, a dSph such as Fornax.
Gaia data reach more shallow photometry and less accurate color measurement compared to the ground-based observations (see Sect.~\ref{sec:data}). Yet, the Fornax RGB branch can still be recognized easily. We may select only RGB candidates to study its morphology and possible debris under the assumption that the RGB is representative of the stellar population of the galaxy, knowing that Fornax is dominated by an old stellar population even though recent star formation has been discovered \citep{Weisz2014,deBoer2013,delPino2013,Rusakov2021}.
If there were tidal tails in Fornax, detecting them requires that we search as far as possible from the object, and that we reach the real background where there are no member stars by definition. Thus, choosing a secure reference as our background region is mandatory. A symmetric region, and as far as possible around the object in question, would be ideal because possible tidal debris are expected to distribute along a specific direction.
If member stars of Fornax were to fall into the background region, for example due to the presence of a long tidal tail, the background will be overestimated in principal, which may lead to artificial truncation of the density profile and will reduce or even remove possible signatures of faint structures. Gaia's full-sky coverage easily provides us the opportunity to perform a study over a large area, which will provide us with sufficient counts in the background region to accurately evaluate the uncertainty of the background level.
Table~\ref{tab:method} compares the seeings from all studies. In term of spatial resolution, Gaia EDR3 is the best amongst all studies in the table. A good effective spatial resolution helps to obtain a better completeness of the source catalog, and hence reduces the effect of crowding in dense region. The crowding effect could reduce the estimated spatial density in the central region, so that the measurement of the shape of an object, such as its ellipticity and radius, will be affected. As shown by \citet{Bate2015}, the crowding effect may cause an overestimation of the Sersic radius by about 5\%. This could be one of the reasons\footnote{On the other hand, an over-subtraction of background could also cause an overestimation of the Sersic radius.} why other studies e.g., \citet{Battaglia2006} or \citet{Munoz2018}, found relatively larger Sersic radii without considering the effect of crowding.
\citet{Irwin1995} used observations with worse seeing conditions,
which should lead to a more severe crowding effect.
To test this, we fit a Sersic profile to their density profile of Fornax (the data in their Table 3). We found a Sersic radius of 21.9 arcmin (see Appendix~\ref{sec:ih95}), which is the largest amongst all studies of Fornax, and consistent with our analysis of the impact of seeing on the study of the dwarf's structure. The other aspect, shallow photometry, also helps to reduce the crowding effect. Bright or faint RGB stars are the same stellar population, thus, they should trace the same structure of Fornax. However faint RGB stars are numerous, hence could result in a distribution affected by crowding, especially towards the center of the galaxy. Under such consideration, a shallower depth of photometry may be a better way to probe the structure in the central region.
A drawback with shallow photometric data will be less significance in statistics. Nevertheless, our major interest is to search faint structure in the outskirts where the crowding effect becomes negligible if using Gaia data.
The main mission of Gaia is to provide PM measurement of all sources, which is an independent parameter space from the CMD.
As Fornax is moving around the MW, in principle we should be able to distinguish Fornax member stars from non-member sources in a certain region in PM space. Thus, we could select member stars by PM.
Similar to the problem with the CMD-selection method,
there will be contaminating sources selected as candidates due to their intrinsic PM overlapping with that of Fornax, and the error bars in PM will introduce more contamination.
By combining the selections from both CMD and PM, i.e., two independent parameter spaces, principally we are able to eliminate more contaminating sources, and hence reducing the level of the background in the spatial distribution. In addition, parallaxes could be used for eliminating foreground stars close the Sun in case they fall in the RGB and PM selections.
Fornax is located at high Galactic latitude ($b=-65.6$ deg) where the foreground contamination from the MW is relatively small when compared to many other dSphs at lower galactic latitude. We may thus expect to reach deeper background depth.
Finally, another tricky problem is that we have to keep contaminating sources in sufficient numbers to accurately measure the background level, and especially its uncertainty, in order to characterise the depth and the significance of faint structures that we are seeking. In other words, it is not necessary to completely remove contamination, but we need to optimize the selection with all the above considerations to obtain the least contaminated candidate sample and to keep sufficient counts to evaluate the spatial background level.
Note that we do not consider the method of likelihood selection, e.g., \citet{Pace2019} where a prior spatial distribution of a Plummer sphere is assumed for the target. Such an assumption may lead to biases if we want to search faint and unknown structures. In brief, in term of methodology of member candidate selection, we will use Gaia EDR3 data to select member candidates having the color of Fornax' RGB, having similar PM as Fornax' mean motion, and being reasonable distant from the Sun. All selection conditions will be scaled by the error bar of each relevant quantity.
\section{Data}
\label{sec:data}
We have chosen a field of 20 by 20 degrees centered on Fornax. This very large area, more than 30 times larger than Fornax in diameter (half light diameter of 0.6 degree, Munoz et al. 2018), should be enough to reach and robustly determine the outskirts of Fornax. First, we selected a raw sample ($S_{\rm raw}$) of sources from EDR3\footnote{Gaia Archive at https://gea.esac.esa.int/archive/} with the following conditions to control the quality of the data:
Not \texttt{duplicated\_source};
Not QSO;
with color \texttt{bp\_rp} measured;
with astrometry solutions (either 5-parameters or 6-parameters);
\texttt{G~<~20.8};
\texttt{ruwe < 1.4};
and \texttt{$C^{*} < 1.0$}. The $C^{*}$ is the corrected \texttt{phot\_bp\_rp\_excess\_factor} introduced by \citet{Riello2020}, which is very helpful to remove background galaxies (see their Figure~$\!$21); bright QSOs can be identified by cross-matching with the confirmed QSO sample provided in the Gaia EDR3 database, i.e., the table \texttt{gaiaedr3.agn\_cross\_id}.
For this raw sample, we have median uncertainties in photometry: $0.008$ in \texttt{G}-band, $0.15$ in \texttt{BP} and $0.09$ in \texttt{RP}, resulting in a less accurate in color (\texttt{BP}$-$\texttt{RP}) than the ground-based observation, e.g., \citet{Battaglia2006}. Nevertheless, the RGB branch of Fornax stars can still be recognized unambiguously.
We checked Galactic extinction over the field, which is $E(B-V)= 0.04$ magnitude on average, and then we do not apply any correction in the following analysis. For a better interpretation of internal structures in both morphology and kinematics, we have adopted the projection introduced by \citet{Helmi2018,Luri2020}, including the corresponding transformation of proper motions:
\begin{equation} \label{eq:1}
\small
\begin{split}
x & = \cos{\delta} \sin{(\alpha - \alpha_{\textsc c})} \\
y & = \sin{\delta} \cos{\delta_{\textsc c}} - \cos{\delta} \sin{\delta_{\textsc c}}\cos{(\alpha - \alpha_{\textsc c})} \\
\begin{bmatrix}
\mu_{x} \\
\mu_{y}
\end{bmatrix} & = \mathsf{M}
\begin{bmatrix}
\mu_{\alpha*} \\
\mu_{\delta}
\end{bmatrix} \\
\mathsf{C}_{\mu_{x,y}} & = \mathsf{M}\, \mathsf{C}_{\mu_{\alpha*,\delta}} \mathsf{M}^{T}, \\{\rm where}\\
\mathsf{M} & = \!\!
\begin{bmatrix}
\cos{(\alpha - \alpha_{\textsc c})} & -\sin{\delta} \sin{(\alpha - \alpha_{\textsc c})} \\
\sin{\delta_{\textsc c}} \sin{(\alpha - \alpha_{\textsc c})} & \!\!\! \cos{\delta}\cos{\delta_{\textsc c}
\!+\!\sin{\delta} \sin{\delta_{\textsc c}} \cos{(\alpha - \alpha_{\textsc c})}
}
\end{bmatrix},
\end{split}
\end{equation}
$\mathsf{C}_{\mu_{x,y}}$ and $\mathsf{C}_{\mu_{\alpha*,\delta}}$ are covariance matrix.
At the morphological center we have $({\mu}_x,{\mu}_y)_\textsc{c} = ({\mu}_{\alpha*},{\mu}_\delta)_\textsc{c}$. All the following analysis steps are carried out in this projection frame, unless specified otherwise.
\begin{figure}
\includegraphics[width=8.5cm]{Sample0_inspec1.jpg}
\caption{Panel~($a$) spatial distribution of sources showing the spatial definition of the object sample, $S_{\rm obj}$ (red, a circle of 0.4-degree radius), the reference sample, $S_{\rm ref}$ (blue, an annulus of 6-10 degree), and 10\% of the raw sample, $S_{\rm raw}$ in gray. Panel~($b$), PM distribution of $S_{\rm obj}$ showing that Fornax member stars concentrate in an elliptical region. The average PM of Fornax is marked by the black dot. The ellipse, also in panel ($c,d$), has a major axis of 2.0 mas/yr, a minor axis $1.06$~mas/yr and an orientation 62.9 degree (measured from $x$-axis, anti-clockwisely). Panel~($c$), PM distribution for $S_{\rm ref}$, where we can see a concentration of sources at zero proper motion, which could be mostly due to background compact galaxies, such as QSOs; and that MW stars distribute very broadly peaking around $(\mu_x,\mu_y) = (2.5,-2)$ and overlapping with Fornax member stars.
Panel~($d$) illustrates the distribution of a comparison sample, i.e., possible contaminating sources, selected from $S_{\rm ref}$ in an area that corresponds to that of $S_{\rm obj}$, which means both of the comparison sample and $S_{\rm obj}$ have the same sky coverage.
Hence, the source distribution in panel $b$ and $d$ are directly comparable. From this comparison, one could infer the level of contamination. This comparison sample of contamination is taken arbitrarily from $S_{\rm ref}$ as a ring from 7.0 to 7.0114 degree.
}
\label{fig:check}
\end{figure}
\begin{figure}
\includegraphics[width=8.5cm]{cmds_v2.jpg}
\caption{CMD for different samples, i.e., the title of each panel where "PM selection" refers to the selection method in Sect~\ref{sec:selPM}. In all panels, solid black lines set limits for the selection of RGB stars in each CMD; vertical orange and red lines illustrate color cuts at \texttt{BP-RP}=0.8 and 1.1, respectively. }
\label{fig:cmdselections}
\end{figure}
\begin{figure*}
\includegraphics[width=18cm]{FornaxSample0__cmdofpm_v2.png}
\caption{Top row: CMDs for the sample $S_{\textsc{\,pm}}$, as a function of elliptical annuli. The range of each annulus and the corresponding area are indicated at the top of each panel. The vertical orange and purple lines indicate color cut at \texttt{BP-RP} larger than 0.8 and 1.1, respectively. The red line marks \texttt{BP-RP}=1.1 and \texttt{G} = 19.2, the conditions for defining $S_2$. Bottom row: CMDs of possible contaminating sources that corresponds to the same sky coverage for each elliptical annulus in the top row. Each sample in the bottom row is taken randomly from $S_{\rm ref}$ in a ring, which radius interval is indicated at the top of the panels. }
\label{fig:cmdofpm}
\end{figure*}
\section{Member star candidates}
\label{sample}
In order to detect possible faint structures far from the center of Fornax, we use the combination of several available parameters, such as the color-magnitude diagram (CMD), proper motions (PMs), and parallaxes to remove at best contaminations from QSOs or Galactic stars. For each parameters we carefully define specific procedures for selecting stars associated to Fornax. Final samples are optimized and carried out iteratively in all parameter spaces in order to obtain the best signal-to-noise ratio (S/N) for the surface density profile.
\subsection{Selection by PM}
\label{sec:selPM}
As indicated in Figure~\ref{fig:check}-$a$, we defined an object sample ($S_{\rm obj}$, red) of sources that are located within the central 0.4-degree radius and a reference sample ($S_{\rm ref}$, blue) based on a 6-10 degrees annulus. The latter is also defined as the reference or background region throughout of the paper.
Fornax member stars (see panel $b$) concentrate in an elliptical region in PM space. At Fornax' distance 147 kpc, 10 km/s on sky is equivalent to 0.014 mas/yr, which is far smaller than the median uncertainty (0.391~mas/yr) for an individual source in $S_{\rm raw}$. Thus, this distribution in PM space reflects only the uncertainties of the astrometry determination and a global anti-correlation between $\mu_x$ and $\mu_y$ for member stars. We measured the average shape of this distribution, indicated in panel ($b$) by an ellipse. The details of this calculation is explained below.
MW stars, i.e., $S_{\rm ref}$ (see panels $c$) show quite a scattered distribution with a peak shifted from that of Fornax stars.
In panel~($d$), we illustrate the level of contaminating sources for $S_{\rm obj}$ by showing a comparison sample, see more details in the figure caption.
If a star has its PM consistent with the mean PM of Fornax within its error bar, it can be considered as a member candidate. As we mentioned above, the dominant Fornax population shown in Figure~\ref{fig:check}-($b$) actually reflects the random error distribution of PM for member stars. If we define a selection region following this distribution, we will obtain a sample of member candidates. The distribution can be characterized by an ellipse centered on the mean PM of Fornax denoted by $({\mu}_x,{\mu}_y)_\textsc{c}$.
On the other hand, the ellipse should not extend too far from the mean PM of Fornax otherwise the peak distribution of MW star will be included, increasing the contamination. In Figure~\ref{fig:check}-($b$), the size of the ellipse is set at $2.0$~mas/yr in major axis, which seems a best choice according to our following analysis.
Our goal is to search Fornax member stars in the full field of $S_{\rm raw}$, thus, principally we should not apply any spatial condition for the selection.
With all these considerations, we implemented the following procedure to perform the selection with PM.
In order to characterise the PM distribution of Fornax' stars as seen in Figure~\ref{fig:check}-($b$), we first prepare a core-sample which is defined within an elliptical radius of 1 degree according to the morphological parameters of Fornax, i.e., the on-sky center ($\alpha_\textsc{c}, \delta_\textsc{c}$), ellipticity and position angle (PA). These parameters can be initialized with values from the literature, e.g., \citep{Battaglia2006}, and revised later using our own selected sample.
With the core sample, we can calculate the corresponding error-weighted average PM, i.e., $({\mu}_x,{\mu}_y)_\textsc{c}$ and the axis-ratio and orientation of the ellipse representing the PM distribution.
Note that the properties of an elliptical distribution are calculated by averaging the source counts, following the method in \citet{Bertin1996}.
To define the member candidate sample selected by PM, namely "$S_{\textsc{\,pm}}$", we required the following two conditions:
\begin{enumerate}
\item The PM of a candidate must follow the correlation between $\mu_x$ and $\mu_y$, i.e., fall inside the ellipse that characterises the PM distribution, e.g., Figure~\ref{fig:check}-($b$). The size of the ellipse is limited to $2.0$~mas/yr as a global condition.
This cut-off is 5 times larger than the median uncertainty of individual sources, i.e, 0.391 mas/yr, so it is reasonably large to fully over the error distribution of member stars in PM space, while avoiding the increasing contamination beyond the ellipse in Figure~\ref{fig:check}-($c$) .
\item As uncertainties of measurements are linked to the magnitude of sources, we applied a more restrict selection by requiring that each source has its PM ($\mu$) to be consistent with $({\mu}_x,{\mu}_y)_\textsc{c}$ within 3 times of its own uncertainty. The uncertainties of $\mu$ for each star is calculated via error propagation by considering the correlation coefficient between $\mu_x$ and $\mu_y$.
\end{enumerate}
The PM selection procedure is done by successive iterations, because each change of the morphological parameters of Fornax and the characterizations of the PM distribution will affect the candidate sample. In doing so, we kept a full consistency between selection conditions and the corresponding output. This selection by PM is applied, respectively, to the three Fornax-member samples that will be introduced in the next section. All calculated parameters are listed in Table~\ref{tab:fornax}.
\subsection{ Selection by CMDs}
\label{sec:selCMD}
Figure~\ref{fig:cmdselections}-$a$ shows the CMD of $S_{\textsc{pm}}$ where the RGB branch is seen clearly. We define sources enclosed by two black lines in the CMD as sample $S_{\textsc{\,rgb}}$ that is selected by only colors and magnitude.
By comparing to panel-$b$ and $c$, we notice that many bright MW stars have been excluded, leaving fainter source around $\texttt{BP-RP}=0.6$ that may contaminate faint Fornax RGB stars when $\texttt{BP-RP}$ less than around 1. Further checking (see panel-$d$) we notice that these contaminating sources may be faint QSOs that are not identified yet. Note that the confirmed QSOs have been excluded when we selected $S_{\rm raw}$.
To avoid contamination, we may consider more or less restrictive conditions, such as using further limits in \texttt{BP-RP} (see vertical lines in Figure~\ref{fig:cmdselections}) and combine them with $S_{\textsc{\,rgb}}$ to obtain CMD selected sample.
To maximize the total number of selected Fornax members, we limit $S_{\textsc{\,rgb}}$ in the magnitude range $17.3 < \texttt{G} < 20.8$, which covers the tip of RGB and guarantees the data quality. Fainter than $\texttt{G} > 20.8$, the uncertainties in astrometry and photometry increase very quickly, which reduces the efficiency in separating stars of Fornax from those of the MW.
After testing many possible selections by limiting magnitude range and cut-offs in \texttt{BP-RP}, we finally choose three representative selections by CMD to establish the Fornax-member samples (see Sect~\ref{sec:final}).
In the top row of Figure~\ref{fig:cmdofpm}, we plot CMDs as a function of elliptical annuli, which is an efficient way to explore how far from the center Fornax stars could distribute.
In the bottom row of Figure~\ref{fig:cmdofpm}, we plot, illustratively, comparison samples of possible contamination that correspond to each elliptical annulus. Each comparison sample in the bottom row is chosen arbitrarily from $S_{\rm ref}$ by requiring the same sky coverage as in the corresponding top panel. In doing so, we can compare the CMD distribution directly in number between member candidates (top) and background (bottom).
For example, in panel-$c$, if we count the number of points inside the RGB region for top and bottom panel respectively, and subtract them, we find that there is an excess of counts of around 30 in the top panel. This is a strong signal that Fornax member stars exist in this annulus. In panel-$d$, we found an excess count of 4, which is a much weaker signal of about two sigma. This suggests that Fornax' member stars may extend out to 2-degree elliptical radius.
Properly subtracting the contamination can only be done statistically in the following density analysis. The illustration in Figure~\ref{fig:cmdofpm} is only to visualize the significance of Fornax' RGB stellar populations at different radii.
\subsection{Selection by parallax}
\label{sec:selparallax}
Fornax is located at a distance $147\pm 4$ kpc from the Sun \citep{deBoer2016}, corresponding to a parallax $0.00680$~mas, which is far smaller than the typical error of individual sources in Gaia data. Hence, in principle the parallax distribution for Fornax stars in EDR3 reflects only the uncertainties in parallax determination. Thus we applied a gentle filter by rejecting some sources with \texttt{|parallax\_over\_error|}~$ > 3.5$, see Figure~\ref{fig:verify}. Correspondingly, we defined $S_{\textsc{plx}}$ for selection \texttt{|parallax\_over\_error|}~$ <= 3.5$ to be combined with other conditions (see Table~\ref{tab:stat}).
\begin{table}
\small
\caption{Fornax parameters derived from this work.}
\label{tab:fornax}
{\centering \small
\begin{tabular}{lrrrr}
\hline
\hline
Parameters & S0 & S1 & S2 \\ \hline
$\alpha_\textsc{c}$(J2000.0) & $2^\mathrm{h}39^\mathrm{m}50.9^\mathrm{s}$ & $2^\mathrm{h}39^\mathrm{m}50.9^\mathrm{s}$ & $2^\mathrm{h}39^\mathrm{m}51.0^\mathrm{s}$ \\
$\delta_\textsc{c}$(J2000.0) & $-34^\circ30{}^\prime53{}^{\prime\prime}$ & $-34^\circ30{}^\prime54{}^{\prime\prime}$ & $-34^\circ30{}^\prime49{}^{\prime\prime}$ \\
Ellipticiy & 0.325 & 0.317 & 0.321 \\
Position angle (deg)$\!\!\!\!\!\!$ & 47.7 & 47.3 & 46.8 \\
\hline
${\mu}_{\alpha*,\textsc{c}}$(mas/yr) $\,\,\,^\dagger$$\!\!\!\!\!\!$& $0.381\pm 0.001 $ &$0.381\pm 0.001 $& $0.381\pm 0.002 $ \\
${\mu}_{\delta,\textsc{c}}$(mas/yr) $\!\!\!\!\!\!$$\!\!\!\!\!\!$ & $-0.367\pm 0.002$ &$\!\!\!\!\!\!$$-0.367\pm 0.002$& $\!\!\!\!\!\!$$-0.366\pm 0.002$ \\
Axis-ratio & 0.531 & 0.533 & 0.610 \\
Orientation & 62.9 & 64.4 & 65.7 \\
\hline
\end{tabular}
}\\
$^\dagger$For PMs, only the random error is quoted here, while for the systematic errors one may refer to \citet{Li2021} : 0.018 and 0.019 mas/yr, respectively.
\end{table}
\subsection{Three final samples of Fornax member candidates}
\label{sec:final}
After applying all the selection conditions we obtain our final samples of Fornax member candidates $S_{\textsc{0}}$ , $S_{\textsc{1}}$, and $S_{\textsc{2}}$ with different conditions, as summarized in Table~\ref{tab:stat}.
$S_0$ selects the largest number of member candidates as possible, and it provides the best S/N for the following data analysis. $S_1$ is defined by limiting the color to be redder, in order to better remove QSOs and MW stars that overlap in color with Fornax RGB, as seen in Figure~\ref{fig:cmdselections}. Thus $S_1$ provides a less contaminated background that could still be determined accurately given its large number of stars. $S_2$ is a very strict sample of bright stars having the best measurements in PM and parallax (see Figure~\ref{fig:verify}). With $S_2$, we may probe the extent of Fornax based on bright stars, also to verify the structure analysis from the other two samples. In the following we analyse the three samples in parallel, in order to consolidate stable and variable structures, as well as the robustness in our structure analysis. The contamination fraction (CC) from background counts within 2.1 degree radius is estimated and given in Table.~\ref{tab:stat}.
We verified the member sample selection in PM space and parallax space, see Figure~\ref{fig:verify}. The global background is obtained through the 6-10 degree annulus of Figure~\ref{fig:check} (see panel(a)), for which the mean density and Poisson noise can be robustly determined for each of the three Fornax samples (Table~\ref{tab:stat}).
Figure~\ref{fig:hist} compares the spatial distribution of $S_{\rm raw}$ and one of the Fornax-member sample, $S_{0}$. It shows that the gradient observed in the background counts (left panel) disappears in the right panel, meaning that most of the MW stars have been excluded. Moreover, it illustrates that there is no obvious large scale structure, such as tails in the large field of view.
\begin{table*}
\small
\caption{Statistic properties of the three Fornax candidate samples.}
\label{tab:stat}
\begin{tabular}{lccccccccc}
\hline
\hline
\multicolumn{2}{c}{{Samples and Conditions }} & $N$ & CC$^{\dagger}$ & $\Sigma^{\rm 2D}_{\rm bg}$ & ${\rm S.B.}_{\rm lim}$ $^{\dagger\dagger}$ & $\Sigma^{\rm 1D}_{\rm bg}$ & ${\rm S.B.}_{\rm lim}$ $^{\dagger\dagger}$& (S/N)$_*$ & Cnt.$_*^{\rm net}$ \\
&&& & ($10^{-3}$arcmin$^{-2}$) & mag/arcsec$^2$ & ($10^{-3}$arcmin$^{-2}$) & mag/arcsec$^2$ & \\ \hline
&&& $\!\!\!\!\!\!$($r_{\rm ell}<$2.1deg)$\!\!\!\!\!\!$ & \multicolumn{2}{c}{ 2D surface density} & \multicolumn{2}{c}{ 1D surface density} & \multicolumn{2}{c}{($1.3<r_{\rm ell}<2.1$)} \\
$S_{\textsc{0}}$:$\!\!\!\!\!\!$ & $S_{\textsc{base}}$$^{\dagger\dagger\dagger}$ {\scriptsize \&} ($\texttt{G} < 20.8$ {\scriptsize \&} $\texttt{BP-RP} > 0.8$) & 21359 &53 / 17140 & 1.63$\pm$1.77 & 33.55 & 1.625$\pm$0.047 & 36.00 & 9.1 & 117 \\
$S_{\textsc{1}}$:$\!\!\!\!\!\!$ & $S_{\textsc{base}}$ {\scriptsize \&} ($\texttt{G} < 20.8$ {\scriptsize \&} $\texttt{BP-RP} > 1.1$) & 12818 &14 / 11669& 0.41$\pm$0.91 & 33.86 & 0.421$\pm$0.024 & 36.59 & 6.8 & 55 \\
$S_{\textsc{2}}$:$\!\!\!\!\!\!$ & $S_{\textsc{base}}$ {\scriptsize \&} ($\texttt{G} < 19.2$ {\scriptsize \&} $\texttt{BP-RP} > 1.1$) & 3883 &1 / 3769& 0.03$\pm$0.25 & 33.99 & 0.030$\pm$0.006 & 37.94 & 4.4 & 20 \\
\hline
\end{tabular}\\
$^{\dagger}$ Estimated contamination counts from background density over the total counts inside $r_{\rm ell}<2.1$ degree.\\
$^{\dagger\dagger}$ V-band surface brightness limits are defined at the 3-sigma level of above the mean background of each sample; all values have the same error $\pm0.16$ mag.\\
Note that the last two columns are the signal-to-noise ratio and net counts of member stars (after subtracting the corresponding background counts) in between 1.3 and 2.1 elliptical radius.\\
$^{\dagger\dagger\dagger}$ $S_{\textsc{base}}$ = ($S_{\textsc{pm}}$ $\cap$ $S_{\textsc{rgb}}$ $\cap$ $S_{\textsc{plx}}$), which are defined in Sect~\ref{sample}. \\
\end{table*}
\begin{figure}
\includegraphics[width=8.5cm]{FornaxSample1__verify.png}
\includegraphics[width=8.5cm]{FornaxSample2__verify.png}
\caption{Verification of the final samples (green dots) in PM space (left panels) and parallax versus parallax error (right panels). Only $S_1$ and $S_2$ are shown in the top and bottom row, respectively. $S_0$ has almost identical distribution as $S_1$ in these two parameter spaces, except with a larger number of sources. Blue dots are the corresponding reference sample in the magnitude ranges of $S_1$ and $S_2$, respectively. In the right panels, violet dots are those rejected by the parallax condition indicated by the two black lines (see more in Sect~\ref{sec:selparallax}), and larger dots are those within the central 1 degree radius.
}
\label{fig:verify}
\end{figure}
\begin{figure}
\includegraphics[height=4.4cm]{spacehist2.png}\includegraphics[height=4.4cm]{Sample0_spacehist1.png}
\caption{
{\it Left panel: } Inspection of background for the raw sample $S_{\rm raw}$. Histograms of star counts are shown respectively for $x$ and $y$ projection. Gradient of counts can be clearly seen in both projections, which is due to the MW stars.
{\it Right panel :} The same figure as the {\it Left}, but for $S_{0}$, where the gradient of $S_{\rm raw}$ has clearly disappeared, indicating our selection method is very efficient in removing MW stars.
}
\label{fig:hist}
\end{figure}
\section{results}
\subsection{Fornax morphology}
The spatial distribution of the three samples are shown in the left panels of Figure~\ref{fig:morph}. A smoothed contour map of the spatial distribution is shown in the right panels. We applied a fixed Gaussian kernel of FWHM 0.25 degree for smoothing. Only the central 6x6 sq degrees is shown; outside this scale no substructures can be identified. We defined an annulus covering 6 to 10 degrees as the background reference, and calculated the mean densities and fluctuations for each sample (see Table~\ref{tab:stat}).
Morphologies from different samples show that the main body (those black contours above 10-sigma level) of Fornax appears quite regular and substantially symmetric. Overall, distributions of the three samples between 1.3 and 2.1 degrees are in good agreement, despite some variations.
$S_0$ shows a possible extended feature on the minor axis in the north-west direction, at position $(x,y)=(-0.8,0.5)$, i.e., the last black contour. This feature is not confirmed in $S_1$ and $S_2$. Sample $S_2$ shows less significant extended features, though there are still few stars at 2.1 degrees, see also Figure~\ref{fig:cmdofpm}-$d$ and the relevant discussion in Sect.~\ref{sec:selCMD} about the significance.\\
\begin{figure}
\includegraphics[width=8cm]{FornaxSample0_points.pdf}
\includegraphics[width=8cm]{FornaxSample1_points.pdf}
\includegraphics[width=8cm]{FornaxSample2_points.pdf}
\caption{Morphology of Fornax by $S_0$, $S_1$ and $S_2$, from top to bottom, respectively. Direct dot plots are shown in the left panels, and the smoothed contour maps are shown on the right. In all panels, two ellipses (gray dashed lines) of 1.3 and 2.1 degree in radius, describing the morphological shape (Table.~\ref{tab:fornax}), are superposed. The red arrow in left-middel panel indicates the mean PM of Fornax (see Table~\ref{tab:fornax}). For $S_0$, contour levels range from $5.30\times10^{-3}$ to 14.1 counts/arcmin$^2$ with logarithmic interval factor 1.483. For $S_1$, $2.75\times10^{-3}$ to 9.87 counts/arcmin$^2$, the interval factor 1.490. For $S_2$, $0.740\times10^{-3}$ to 3.10 counts/arcmin$^2$, the interval factor 1.510. Note that the peak density is lower than that of 1D density profile, because the contours has been smoothed by a constant Gaussian kernel of 0.25 degree, indicated by a red circle at the top-left corner. In all contour maps, thin and dark-gray contours are between 3 to 10 sigma significance above the 2D background fluctuations (see Table~\ref{tab:stat}). }
\label{fig:morph}
\end{figure}
\subsection{Surface Density Profile of Fornax}
\label{sec:radialprofile}
We derive and analyze the radial surface-density profile for each member candidate sample by applying the same algorithm as described below. For simplicity, we choose to present $S_1$ in this section, while $S_0$ and $S_2$ are presented in Appendix~\ref{sec:dens02}.
Following the method of \citet[see also \citet{Battaglia2006}]{Irwin1995}, the radial profile is built directly from the candidate catalog, by evaluating the mean density in each elliptical annulus indexed by the semimajor axis. We assumed a mean geometry, i.e an P.A. and ellipticity as listed in Table~\ref{tab:fornax}, for each candidate sample, respectively. Such an assumption is helpful to explore the extent of a dSph in the low-density outskirts and can provide an easy way to understand and interpret the results, \citep[see][for a detailed discussion]{Irwin1995}.
Figure~\ref{fig:denprof} shows the radial surface-density profile for $S_1$.
Error bars are estimated from Poisson noise based on the counts in each annulus. In order to see details, we used high-resolution and logarithmic spacing in radius. We also request a minimal count of 9 stars in a single radius bin (resulting in a 33\% statistic error at most), otherwise, it will be combined into its neighboring annulus. Then the profile is obtained with adaptive resolution especially at center and outskirts.
We have examined the background uncertainty with two different methods. Method-1 calculates a mean number density within the background reference, i.e. the annulus from 6 to 10 degree, and the Poisson noise of counts as its uncertainty. Method-2 calculates the background density using the error weighted mean of all stars lying between 6 to 10 degree from the density profile \citep{Battaglia2006}. The latter include background fluctuations. Both methods provide the same mean value, and we have chosen the larger error bar as a conservative estimation of the background uncertainty. Under such considerations, for $S_2$, we have chosen the Poisson noise as the estimate of background uncertainty, while for $S_0$ and $S_1$, we have adopted the uncertainties using Method-2. As indicated in the figure, after subtracting the background, we are able to trace the density profile by almost 5 decades from the center to the outskirts above the 3-sigma level (defined as detection limit) of the background fluctuations (Table~\ref{tab:stat}).
Our results reach 2 decades, i.e., 5 magnitudes, deeper than previous studies mentioned in Table~\ref{tab:method}.
Here we evaluate the depth or the equivalent surface brightness limit of our results. To do so, we first assume that our density profiles in the central region trace the same structure, i.e. density profile, as found by \citet{Munoz2018}. This underlying assumption is based on the fact that the RGB dominates the stellar light in Fornax, and are considered to be representative of the whole Fornax surface brightness made by \citet{Munoz2018}, who included other stellar population besides RGB stars.
Then, we re-scale the density profile by matching the mean density inside the 0.3-degree elliptical radius to the effective surface brightness $24.77\pm0.16$ mag arcsec$^{-2}$
in the V-band (defined inside the same elliptical radius) found by \citet{Munoz2018}. The mean density inside the 0.3-degree elliptical radius are 17.318, 11.901, 3.612 star/arcmin$^2$ for $S_0$, $S_1$ and $S_2$ respectively. These values are applicable to the calibration of the corresponding 2D density maps.
The reached surface brightness limits are listed in Table~\ref{tab:stat}. The statistical error of evaluating the mean density is very small because of the large number of counts, so the uncertainty of the equivalent surface brightness is mostly coming from the physical calibration by \citet{Munoz2018}, i.e., 0.16 mag.
\begin{table}
\caption{Best-fitting parameters of different theoretical density profiles :
King model (core radius, $r_{\rm c}$, tidal radius, $r_{\rm t}$),
Sersic model (Sersic radius, $R_{\rm S}$, shape index, $m$),
exponential model (scale radius, $r_{\rm s}$)
and Plummer model (scale radius, $b$). Peak density of each profile $I_0$ are all in unit stars/arcmin$^2$. The last columns give the reduced and the mean $\chi^2$, respectively. The latter is evaluated for the points between 1.3 to 2.1 degree. The goal is to compare between Sersic and double-Sersic model. }
\small
\label{tab:fit}
\centering
\begin{tabular}{llccccc}
\hline
\hline
Model$\!\!\!\!$ & \multicolumn{2}{l}{{ Sample$\!\!\!\!$$\!\!\!\!$}} & \multicolumn{2}{l}{{parameters}} & $\!\!\!\!$ $\chi^2/\nu$ & $\!\!\!\!$$\overline{\chi^2_*}$ \\ \hline
\multicolumn{2}{l}{King} & $I_{\rm 0,K}$ & $r_{\rm c}$ [$^\prime$] & $r_{\rm t}$ [$^\prime$] & \\
& $S_0$ & 33.9$\pm$3.6 & 15.3$\pm$1.7 & 76.0$\pm$5.2 & 1.16 & - \\
& $S_1$ & 24.4$\pm$2.8 & 15.4$\pm$1.9 & 71.6$\pm$5.2 & 1.28 & - \\
& $S_2$ & ~7.5$\pm$1.3 & 14.7$\pm$2.5 & 76.3$\pm$8.7 & 0.88 & - \\ \\
\multicolumn{2}{l}{Sersic} & $I_{\rm 0,S}$ & $R_{\rm S}$ [$^\prime$] & m & \\
& $S_0$ & 25.4$\pm$2.7 & 14.5$\pm$0.6 & 0.83$\pm$0.04 & 1.41 & 5.15\\
& $S_1$ & 17.4$\pm$2.0 & 14.7$\pm$0.7 & 0.80$\pm$0.04 & 1.26 & 4.64 \\
& $S_2$ & ~5.8$\pm$1.0 & 14.1$\pm$1.0 & 0.84$\pm$0.07 & 0.92 & 3.83\\ \\
\multicolumn{2}{l}{double-Sersic} & & & & \\
\multicolumn{1}{r}{ c.1} & \multirow{2}*{$S_0$} & 22.3$\pm$2.6 & 15.8$\pm$0.8 & 0.74$\pm$0.05 & \multirow{2}*{ 1.07 } & \multirow{2}*{0.94} \\
\multicolumn{1}{r}{ c.2} & & ~1.0$\pm$0.8 & 18.7$^{+3.3}_{-6.5}$ & 1.00$^{+0.12}_{-0.29}$ & \\ \\
\multicolumn{1}{r}{ c.1} & \multirow{2}*{$S_1$} & 15.0$\pm$1.9 & 15.9$\pm$0.9 & 0.71$\pm$0.05 & \multirow{2}*{0.97} & \multirow{2}*{0.60}\\
\multicolumn{1}{r}{ c.2} & & ~0.83$\pm$0.65 & 19.7$^{+3.3}_{-6.6}$ & 0.91$^{+0.11}_{-0.29}$ & \\ \\
\multicolumn{1}{r}{ c.1} & \multirow{2}*{$S_2$} & ~4.8$\pm$0.9 & 15.5$\pm$1.2 & 0.73$\pm$0.08 & \multirow{2}*{0.78} & \multirow{2}*{0.37}\\
\multicolumn{1}{r}{ c.2} & & ~0.33$\pm$0.3 & 19.6$^{+4.0}_{-12.}$ & 0.92$^{+0.14}_{-0.90}$ & \\ \\
\multicolumn{2}{l}{Exponetial} & $I_{\rm 0,E}$ & $r_{\rm e}$ [$^\prime$] & & \\
& $S_0$ & 34.5$\pm$3.6 & 10.7$\pm$0.4 & & 2.48 & - \\
& $S_1$ & 24.6$\pm$2.8 & 10.3$\pm$0.5 & & 2.47 & - \\
& $S_2$ & ~7.65$\pm$1.3 & 10.6$\pm$0.7 & & 1.25 & - \\ \\
\multicolumn{2}{l}{Plummer} & $I_{\rm 0,P}$ & $b$ [$^\prime$] & & \\
& $S_0$ & 27.0$\pm$2.8 & 16.4$\pm$0.7 & & 10.85 & - \\
& $S_1$ & 19.6$\pm$2.2 & 15.4$\pm$0.7 & & 12.01 & - \\
& $S_2$ & 6.1$\pm$1.1 & 16.0$\pm$1.1 & & 4.56 & - \\
\hline
\end{tabular}
\end{table}
Following \citet{Battaglia2006}, we fit different theoretical density profiles to the data, and the results are listed in Table~\ref{tab:fit}. For all theoretical profiles we have adopted the same notation as \citet{Battaglia2006}. First, let us compare the results between the three samples. For each theoretical profile, the best fitting parameters are similar from one sample to another, within the error bars. We also note that the reduced-$\chi^2$ generally decreases from $S_0$ to $S_2$, which could reflect the increase of the fraction of true Fornax members from $S_0$ to $S_2$, but could also possibly be due to the decrease of the sample size which leads to less statistical significance.
Second, let us focus on $S_1$. Fornax extends much farther than the King tidal radius (see Table~\ref{tab:fit}), outside which the observed density profile shows a change of slope, in a similar way as the observed breaks found in other classical dSphs \citep[e.g.,][]{Westfall2006}.
Furthermore, we can safely reject the Plummer profile, according its huge reduced-$\chi^2$ value.
Although the exponential profile gives a reasonable description of the data, it shows some systematic biases at different radii. This can be seen more clearly in the middle panel of Figure~\ref{fig:denprof} where the best-fit of exponential profile deviates from the data points.
The Sersic profile provides a very good fit to the data out to 1~degree (see, the lower panel in Figure~\ref{fig:denprof}). The departure (i.e., the break) beyond this radius reveals that an additional component exists in Fornax. To characterize this additional component we introduce a double-Sersic model to decompose the density profile. It gives an obvious improvement of the fitting, as shown by the significant changes of the mean $\chi^2$ indicated in the last columns of Table.~\ref{tab:fit}, suggesting the need of this second component. The second Sersic component , i.e., "c.2" in Figure~\ref{fig:denprof}, starts to dominate the density at 0.9 degree, where the two components reach the same density. It includes 10, 9, and 12.4\% of the total net counts by integrating the theoretical profiles for $S_1$, $S_0$, and $S_2$ respectively. Accounting for the 2D morphology that shows a nearly symmetric morphology, it suggests that the second component behaves like a surrounding halo, though there could be projection effects.
We find that within 1 degree, the density profiles derived for Fornax in this paper are almost the same as the result obtained by \citet{Coleman2005} (see Figure~\ref{fig:denprof} open triangles). Interestingly, \citet{Coleman2005} used only bright RGB star selection too, covering an even shallower 2.5 mag of RGB stars.
We have re-scaled their profile by normalizing the average density within 0.1 degree.
The case "All" in \citet{Battaglia2006} has also been re-scaled and plotted in Figure~\ref{fig:denprof}. There is a visible discrepancy between 0.3 to 1 degree, which agrees with fitting results. For example, with a single Sersic profile and $S_1$, we found a smaller characteristic radius than \citet{Battaglia2006}, i.e.,14$^\prime$.7 versus 17$^\prime$.6. We note, however, that our result is consistent with the result by \citet{Bate2015} after their correction of the crowding effect in the center of the galaxy. These comparisons are in line with what we discussed and expected in Sect.~\ref{sec:method}, the relatively more shallow photometric data of Gaia and its good image quality could lead to results much less affected by the crowding effect at the center of Fornax.
\begin{figure*}
\includegraphics[width=15.cm]{FornaxSample1__density_v2.pdf}
\caption{Surface density profile indexed by elliptical radius for $S_1$. In the main panel (top), the corresponding background densities have been subtracted. The horizontal dashed lines indicate the detection limits (Table~\ref{tab:stat}). Different best fitting theoretical density profiles and observed profiles are shown (see the legend). {The two components of the double-Sersic profile are denoted by "c.1" and "c.2".} The vertical dashed lines indicate 1.3 and 2.1 degree, respectively.
The green arrow indicate the location of the best-fitting King tidal radius $r_{\rm t}$.
Two diagnostic panels are shown below: 1) Normalised data (dimmed into gray) and different models with respect to the double-Sersic model; 2) the density profile zoomed to the background level (gray solid and dashed lines for the mean and 3-sigma uncertainty, respectively), for inspecting how density profiles vary at large radii before subtracting the mean background.
}
\label{fig:denprof}
\end{figure*}
\section{Discussion and Conclusion}
Thanks to Gaia's homogeneous data quality and full coverage of PM measurements, we have studied the stellar structure of Fornax dSphs over a huge area (400 sq deg), and reached an extreme depth of photometry, down to 12 magnitudes fainter than the central density of Fornax, which is equivalent to a surface brightness $37.94\pm0.16$~mag/arcsec$^2$ in the V-band. Our study demonstrates that using Gaia's multiple and independent parameter spaces, i.e. CMD, PM, and parallax, allows to perform a member candidate selection that provides a breakthrough for studying dSph outer structures compared to former studies with ground-based deep observations.
Above the detection limit, we have discovered a significant second component in the Fornax dSph, which resembles a stellar halo due to its nearly symmetric morphology. This component represents about 10\% of the total mass of Fornax, and extends out to 2.1 degree (i.e. 5.4 kpc or almost 7 times the half-light radius). This additional component in the Fornax outskirts is well described by a Sersic profile with index of $0.92^{+0.11}_{-0.29}$, which is consistent with an exponential profile. This might suggest that Fornax could extend even further. What is the origin of this halo-like structure?
\citet{Battaglia2015} presented a set of dedicated simulations to fit Fornax, in a frame for which it is DM dominated and a long-lived satellite of the MW. All their models but one (assuming that mass follows light), can reproduce quantitatively both the $\sigma_{\textsc{los}}$ radial profile and the observed morphology out to 1.3 degree \citep{Battaglia2006}. In these models the stellar component of the simulated object is very stable against tidal stripping that removes only up to 10\% of the initial stellar mass after 12 Gyr of orbital integration. However, none of these models predict a second component similar to that we found from Gaia data.
Using cosmological simulations, \citet{Genina2020} study analogs of Fornax according to its star formation history. They argue that the Fornax dSph may show tidal tails at around 35-36 mag/arcsec$^2$, which is only 6 magnitudes in contrast to the central density of the simulated object (see their Fig.~6). With our Gaia results, we have been able to trace the density profile of Fornax within a 12 magnitude range, and we do not find any tidal tail over a large field surrounding Fornax (up to 400 sq degrees).
Conversely to clues of tidal stripping in Fornax, we found that the outskirt of Fornax, i.e. $r_{\rm ell} > 1.3$ degree, resembles a halo-like structure because it appears rather symmetric. It suggests that other mechanisms than tidal stripping are at work in Fornax. Among them, a former dwarf-dwarf merger was supported by the discovery of a shell structure near Fornax \citep{Coleman2004}, but has been disproved later by \citet{Bate2015}.\\ \\
Alternatively, the MW dSphs could be affected by tidal shocks \citep{Hammer2018,Hammer2019}, and in the following we investigate whether it could naturally explain the secondary component of stars in the Fornax outskirts. The mechanism includes the transformation of a Fornax progenitor that was a gas-rich dwarf galaxy. Fornax shows very recent star formation (50-100 Myr ago, \citealt{Coleman2008}), and if we were observing this galaxy at that epoch it would have resembled a gas-rich dwarf irregular (dIrr, \citealt{Battaglia2013}). The case of a DM-free progenitor was first investigated by \citet{Yang2014} with numerical simulations. If gas was dominating the former Fornax dwarf, after the last gas removal, the residual stars would have expanded following a spherical geometry due the loss of gravity. As shown by \citet[and references therein]{Hammer2019}, the tidal stripping term becomes negligible and energy exchanges with the MW are dominated by tidal shocks, which can dominate the \textsc{los}-kinematics. Not all stars are in resonance with the MW or tidally shocked, an average fraction of only 25\% is sufficient to explain the observed hot kinematics. We suggest that the second stellar component revealed by Gaia data may be due to the recent expansion of stars in Fornax. Detailed numerical modeling will be required to reproduce the overall properties of Fornax in the tidal shock scenario.
Such a scenario assumes that the MW halo is filled by diffuse and ionized gas (up to few million Kelvin), know as the Circumgalactic Medium \citep[CGM,][]{Tumlinson2017}. There is direct evidence of such diffuse gas surrounding the MW \citep{Miller2013,Anderson2010}. There are also many indirect indications for the existence of this diffuse gas, such as the strong dichotomy of the distribution of dSphs and dwarf irregular galaxies within and beyond 300 kpc, respectively \citep{Grcevich2009}. The presence of head-tail High-Velocity Clouds (HVCs) \citep{Kalberla2006} at $>$ 50 kpc from the Magellanic Stream implies a density of ionized gas of about $10^{-4} {\rm cm}^{-3}$ at 50-70 kpc and $10^{-5} {\rm cm}^{-3}$ out to a few hundred kpc. When gas-rich dwarf galaxies fall into the MW, their cold gas will feel the ram pressure generated by the hot gas in the MW halo, and later on they could be fully stripped. The best example for such an on-going process is The Magellanic Stream, since the ram pressure tails lagging behind the Magellanic Clouds are indeed observed and reproduced by a 'ram-pressure+collision' model \citep{Hammer2015,Wang2019}. It is reasonable that such a strong ram pressure force could remove the gas from the in-falling dwarf galaxies \citep{Grcevich2009,Yang2014}. Fornax can be interpreted as an archetype for this scenario, because of its very recent star formation implying a recent removal of its gas. \citet{Rusakov2021}, after examining possible scenarios and simulations, found that the observed star formation history (SFH) of Fornax is better explained by periodic passages about the MW. But their proposition cannot explain the two star formation peaks within the last Gyr (see their Fig.~12), because the last orbital period of Fornax is at least 2 Gyr \citep{Rusakov2021}.
Our scenario could explain naturally the subsequent star formation history of Fornax. After Fornax's progenitor (which in this scenario is supposed to be a gas-rich and dark matter free dwarf galaxy) entered the MW halo, around 2 Gyr ago, the gas in the dwarf galaxy will be compressed due to ram pressure, igniting star formation. Because there is no dark-matter, which means the potential of the galaxy is very shallow, the feedback of star formation could push the gas to larger radii of the galaxy, hence reducing or even halting the star formation process. As gas is cooling down and falls back to the center of the galaxy, with the help of ram pressure a new cycle (i.e. epoch) of star formation will be activated, and so on. This process could explain the multiple epochs of star formation in the last 1 Gyr. The oscillation of such a star formation process depends on the mass of the dwarf galaxy and the strength of ram pressure, and thus could put some constraints on the mass of Fornax.
As Fornax approached its current position, it experienced the last ram-pressure process, and then lost its last gas in the recent few hundred Myr. The last event of very recent gas removal is consistent with the observed 100-300 Myr stellar population in Fornax.
If confirmed, the HI cloud superposed on Fornax and discovered by \citet{Bouchard2006} could be the last gas cloud that is leaving the Fornax dSph now.
Our discovery of this second component, i.e. a halo-like stellar component in Fornax, motivates us to revisit the understanding of the "breaks" in dSphs density profiles.
With Fornax, almost all dSphs, except Leo~II,
are reported to display such a break in their density profiles. If this second component of Fornax can be explained as the result of expanding stars, how about the other dSphs? It could be interesting to identify the morphologies of the second component responsible of the break in their density profile, and whether they are consistent with tidal stripping or alternatively with tidal shocking. The latter could be verified with internal kinematics via PMs by Gaia in the future, see our Introduction or \citet[][]{Hammer2018,Hammer2019}.
\section*{Acknowledgements}
The authors would like to thank Fr\'ed\'eric Arenou, Carine Babusiaux, Piercarlo Bonifacio, Jianling Wang, and Hefan Li for their helpful disucssions.
MSP acknowledges funding of a Leibniz-Junior Research Group (project number J94/2020) via the Leibniz Competition, and thanks the Klaus Tschira Stiftung and German Scholars Organization for support via a KT Boost Fund.
This work has been supported by the National Natural Foundation of China (NSFC No. 12041302 and No. 11973042). We also thank for its support the International Research Program Tianguan, which is an agreement between the CNRS, NAOC and the Yunnan University.
This work presents results from the European Space Agency (ESA) space mission Gaia. Gaia data are being processed by the Gaia Data Processing and Analysis Consortium (DPAC). Funding for the DPAC is provided by national institutions, in particular the institutions participating in the Gaia MultiLateral Agreement (MLA). The Gaia mission website is https://www.cosmos.esa.int/gaia. The Gaia archive website is https://archives.esac.esa.int/gaia.
\section*{Data Availability}
The data underlying this article are available at https://doi.org/10.5281/zenodo.5021235.
We provide the catalog of Fornax member candidate, including the three final samples, as well as the data of surface density profiles.
\section*{}{\bf Note added in proof:}
During the publication of the paper, we noticed
that stellar halos of MW dSphs were also discovered recently in the
Fornax \citep{Stringer2021}, and Tucana II \citep{Chiti2021} sourroundings. These findings confirm our result and motivate us to
further study other dSphs.
| af1eb02db91efa2973554dc0ee9820609b2eb395 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{Sec:intro}
Over the past decades there have been great strides in the classification and characterization of topological phases of matter \cite{Kane-Mele-2005, Kane-Mele-2007, bernevig_zhang, RMP_ChingKai,Schnyder_2008, Kitaev_2009, cohomology1, DW_ashvin, cohomology2, wanggu, kapustin2015fermionic, freed2021reflection, maissam_new_5, bhardwaj2017state, cheng_bi_you_gu, maissam_new_3, Aasen:2021vva}.
Distinctions between these phases are encoded in the entanglement structure of their quantum wave functions rather than in patterns of broken symmetry \cite{Oshikawa_2010, senthil_2013, zeng2019quantum}.
While some gapless systems such as topological semimetals \cite{{Savrasov_2011, Zyuzin_2012}} and quantum spin liquids \cite{Savary_2016, sl-review-2} exhibit topological characteristics, in this work, we focus exclusively on topological phases with a bulk energy gap, which often coexists with gapless modes localized on system boundaries.
Such gapped topological phases fall into distinct equivalence classes that cannot be adiabatically deformed into each other without encountering a phase transition at which the bulk gap closes.
In some cases, these distinctions rely on the presence of protecting global symmetry(ies); if such symmetries are broken --- either spontaneously or explicitly --- then phases may be deformed into each other without encountering a phase boundary.
Such phases are said to be `symmetry protected topological' (SPT) phases \cite{cohomology1, cohomology2}.
Another class of phases does {\it not} rely on such symmetry protection, and are said to have `intrinsic' topological order \cite{Kitaev_toric_code, wen2004quantum, Levin_Wen}.
A separate distinction can be usefully drawn between invertible topological phases --- those with no non-trivial topological excitations --- and non-invertible phases which host such excitations.
Invertible topological order can thus either be intrinsic, as in the chiral $p_x + ip_y$ superconductor \cite{Read_Green}, or symmetry-protected, as exemplified by three dimensional (3D) time-reversal invariant topological band insulators (TIs), which rely on a combination of $\ms{U}(1)$ particle number conservation and time-reversal symmetry (TRS) $\mathcal T$ \cite{Kane-Mele-2007}.
Non-invertible topological orders have to be intrinsic and are then distinguished from each other and from trivial orders by the quantum statistics or braiding properties of their topological excitations.
However such features can be `enriched' by global symmetries, which allow finer distinctions to be made between distinct patterns of quantum number fractionalization \cite{Wang_2019}.
Examples of gapped non-invertible topological orders include Kitaev's toric code \cite{Kitaev_toric_code}, Abelian and non-Abelian fractional quantum Hall states \cite{Wen_1995, xiao-gang_zee, Read_Rezayi, Mooreread2}, and quantum spin liquids \cite{Savary_2016, sl-review-2, PUTROV2017254, Xiao_2016, Apoorv_2017}.
Invertible and non-invertible topological orders are linked through the notion of `anomalous' fractionalization, a focus of this paper. An invertible SPT with purely on-site symmetries in $d$ dimensions generically has gapless modes on $d-1$ dimensional boundaries that respect the protecting symmetries. These gapless modes are anomalous in that they cannot be realized in a strictly $d-1$ dimensional system equipped with the same symmetries, and therefore require the bulk in order to exist --- a feature often termed the `bulk-boundary correspondence' \cite{Ryu_2012, Chang_Tse_2014, kapustin2014symmetry, Chang_Tse_2016, Han_2017, Tiwari_2018, Witten_2016, witten_path}.
One route to gapping these modes in the absence of a bulk phase transition involves symmetry breaking, but an additional possibility emerges on the two-dimensional (2D) surface of a 3D SPT: namely, the formation of a non-invertible 2D topological order, enriched by the same symmetries \cite{Senthil_2013b}.
This necessarily involves interactions, since non-invertible orders are intrinsically interacting. The bulk-boundary correspondence is now encoded in the fact that the resulting symmetry-enriched topological (SET) order is {\it also} anomalous: its fractionalized quasiparticles transform under symmetry in a manner that is impossible in a strictly two-dimensional system, but is admissible on the 2D surface of a 3D SPT \cite{Senthil_2013b, Chen_2014, Bonderson_2013}. A specific example of this is furnished by the 3D TRS invariant topological insulator \cite{Kane-Mele-2007}. The gapless `surface termination' that preserves symmetry is a single 2D Dirac fermion, which would violate theorems on `fermion doubling' were it to appear in a purely 2D TRS lattice system \cite{NIELSEN1981219, Witten_2016, witten_path}. The second, gapped, possibility is the non-invertible $\mathcal T$-Pfaffian topological order which contains non-Abelian anyonic excitations \cite{Chen_2014, Bonderson_2013}. Notably, despite respecting TRS $\mathcal T$, the anyon content of the $\mathcal T$-Pfaffian requires a non-zero chiral central charge; this is incompatible with $\mathcal T$-symmetry in a strictly 2D system, but can be realized in a $\mathcal T$-preserving manner on the 2D surface of the 3DTI. Similar gapped anomalous surface topological orders (STOs) have been proposed for many bosonic and fermionic SPT phases \cite{Chen_2015, Fidkowski_2013, barkeshli2017, Barkeshli:2016mew, Meng_2018, kobayashi2019gapped, juven_wen_witten, Tachikawa_TRS, Seiberg:2016rsg, juven_wen_witten, maissam_new_1, maissam_new_2} and have been used as the basis of a classification of interacting electronic topological insulators.
The introduction of spatial symmetries adds richness to the topological classification, allowing the identification of crystalline \cite{fu_crystalline, isobe_fu, fu_review, titus_lecture, gauging_crystalline} and `higher order' symmetry-protected topological phases (HOTPs) of matter~\cite{benalcazar-bernevig-hughes-prb-2017,Schindlereaat-2018,Yuxuan-2018, Piet-2017, Chen-Fang-2019, Peterson_2018, Serra_Garcia_2018, imhof2018topolectrical, luka_2019, yizhi_hoint_2018, shiozaki_2019, bultinck_2019, Ingham_2020, rasmussen_2020}. A topological crystalline insulator/superconductor, as its name implies, requires one or more crystalline symmetries, and exhibits gapless states only on surfaces that preserve those symmetries. However, generic surfaces may preserve symmetries only within certain high-symmetry subsystems, e.g., reflection symmetric lines on a 2D surface, or rotation symmetric points on a 1D or 2D edge; different patches of surfaces can be mapped into each other under symmetry \cite{Hermele_1, Hermele_2, Hermele_3}. These facts can complicate a straightforward definition of a signature on a $d-1$ dimensional surface as it may not fully realize all the symmetries of the bulk. Higher-order topology resolves this complication, by identifying robust signatures on `boundaries of boundaries' and generalizations thereof. An $n^\text{th}$ order topological phase of matter in $d$ dimensions hosts gapless excitations on $d-n$ dimensional boundary subsystems. In the $d=3$ case of interest to us, a first-order topological phase has gapless 2D surface states, a second-order phase is gapless along 1D high-symmetry lines of its 2D surface, and a third-order phase has gapless modes localized to 0D points on its surface. (The latter two cases are often termed `hinges' or `corners', reflecting their spatial locations when the protecting symmetry is a point-group). A large class of HOSPTs have been identified in fermionic and bosonic systems, and candidate solid-state materials have been proposed to host gapless modes protected by higher-order topology.
In 3D, the concept of anomalous surface topological order also generalizes to HOSPTs, but in a distinct fashion from the $n=1$ case. This was demonstrated in Ref.~\onlinecite{apoorv_2019} in the specific setting of HOSPTs protected by a combination of ${\ms C}_{2n}$ rotations and ${\mathcal T}$, which host gapless chiral modes on the hinges of a ${\ms C}_{2n}$-symmetric sample. It was demonstrated that a consistent STO for ${\ms C}_{2n}{\mathcal T}$ HOSPTs could be generated by placing a cousin of the $\mathcal T$-Pfaffian topological order, with the same anyons and symmetry transformation properties, but with $\mathcal T$-symmetry broken in two opposite senses, on adjacent patches of the surface that get mapped into each other under the action of ${\ms C}_{2n}$. In a purely 2D setting --- imagine these phases `painted' on a hollow ${\ms C}_{2n}$-symmetric shell --- this pattern would necessarily involve chiral boundary modes between topological orders with distinct senses of $\mathcal{T}$-breaking. Their absence ---and hence the presence of an {\it anomalous gapped boundary} --- is because the gapless modes can be gapped while preserving symmetry when combined with those contributed by the HOSPT bulk, which counter-propagate and have the same symmetry properties. Thus, the manifestation of higher order topology in this strongly-interacting setting is through the anomalous gapped boundaries of a certain symmetry-enforced patterning of topological orders.
[We note there have been further subtle notions introduced which distinguish bulk-boundary phenomena in the context of {\it weakly} interacting electronic systems, such as boundary-obstructed topological phases \cite{Khalaf_Benalcazar_Hughes_Queiroz_2019, tiwari2020chiral,Ezawa_2020, Ronny_2020, Jahin_2022}, phases with obstructed atomic limits, and fragile topology \cite{Bradlyn_2017,Po_2018, Cano_2018, Po_2019, Khalaf_PRX, Bradlyn_2019, iraola2021topological, rjslager_1, rjslager_2}; we focus on strongly interacting, non-fragile phases and do not consider these below.]
\begin{table*}[ht]\label{Table: Main}
\begin{tabular}{ c|c|c||c|c }
\toprule
\hline
& \multicolumn{2}{c||}{Second Order Topology} & \multicolumn{2}{c}{Third Order Topology}\\
\hline
AZ Class & Hinge Modes& STOs&Zero Modes& STOs \\
\hline
A & Chiral Dirac & 2D $\mathcal{T}$-Pfaffian $^\dagger$ & & \\
\hline
AIII & & & Dirac & $(\mathsf{SO}(3)_3)^4$ $^\ddagger
\\
\hline
AI & & & & \\
\hline
BDI & & & Majorana & $(\mathsf{SO}(3)_3)^2 $ \\
\hline
D & Chiral Majorana & $\mathsf{SO}(3)_3 $ & Majorana & $(\mathsf{SO}(3)_3)^2 $ \\
\hline
DIII & Helical Majorana & $\mathsf{SO}(3)_3 \times \overline{\mathsf{SO}(3)_3}$ & Majorana Kramers Pair& $(\mathsf{SO}(3)_3 \times \overline{\mathsf{SO}(3)_3})^2$ \\
\hline
AII & Helical Dirac & 2D $\mathcal{T}$-Pfaffian$\times \overline{\text{2D $\mathcal{T}$-Pfaffian}}$ & & \\
\hline
CII & & & Majorana Kramers Pair & $(\mathsf{SO}(3)_3 \times \overline{\mathsf{SO}(3)_3})^2$ \\
\hline
C& Chiral Majorana & $^\S$
& & \\
\hline
CI & & & & \\
\hline
\bottomrule
\multicolumn{5}{l}{\footnotesize $^\dagger$ The 2D $\mathcal{T}$-Pfaffian has the same anyon content of $\mathcal{T}$-Pfaffian, albeit without TRS.} \\
\multicolumn{5}{l}{\footnotesize $^\ddagger$ Also enriched by $\mathsf{U}(1)$ charge conservation symmetry.} \\
\multicolumn{5}{l}{\footnotesize $^\S$ Note that there is no STO for 2nd order class C. The 2nd order class C can be obtained by breaking the TRS in the first order class CI, } \\
\multicolumn{5}{l}{\footnotesize similar to the case of 2nd order class A which is obtained by breaking the TRS in the first order class AII. The STO for the 2nd order class C}\\
\multicolumn{5}{l}{\footnotesize should have the same anyon content as the STO for the first order class CI, which was excluded in Ref.~\onlinecite{ci-noSTO}. In the above paper, the authors}\\
\multicolumn{5}{l}{\footnotesize argue that, due to disorder, interaction is always relevant in the first order class CI,and will cause spontaneous symmetry breaking of TRS, } \\
\multicolumn{5}{l}{\footnotesize thus ruling out STOs that preserve TRS.} \\
\end{tabular}
\caption{\label{tab}Summary of surface topological order for all inversion symmetric higher order topological phases in the AZ classes. The superscripts that appear in the column of third order STO denote the number of copies of the topological order e.g. $(\mathsf{SO}(3)_3)^2$ means two copies of $\mathsf{SO}(3)_3$. The STO we put on the surface is always inversion symmetric since we place a topological order and its inversion symmetric partner on the surface so that the original gapless line/point modes can be gapped.}
\end{table*}
In this work, we construct topologically ordered surface terminations for three dimensional electronic topological insulators and superconductors both with and without TRS (classes A, AII, AIII, D, DIII, CII and BDI within the Altland-Zirnbauer classification scheme) whose higher-order topology is enabled by the additional presence of three-dimensional spatial inversion symmetry, denoted $\mathcal{I}$. A band-theoretic classification indicates that the surfaces of such phases host one-dimensional chiral or helical Dirac hinge modes along an inversion-invariant line, or degenerate zero-dimensional corner modes at antipodal points, which cannot be gapped by any symmetric free-fermion perturbation \cite{khalaf_inversion}. We show that these hinge and corner modes may be gapped out upon introducing an inversion-symmetric configuration of fractionalized phases with non-Abelian anyons on the surface. The surface now realizes a fully gapped and symmetric topologically ordered state. Crucially, this surface fractionalization pattern is anomalous as it would be impossible to assemble a configuration of topological orders with the relevant symmetry properties for a system in purely two dimensions, i.e. without invoking the mode contributed by the three dimensional bulk. Compared with Ref.~\onlinecite{apoorv_2019}, this work discusses the action of crystalline symmetry on the STO beyond merely arranging the STOs in a symmetry-respecting configuration.
The rest of the paper is organized as follows. In Section \ref{Sec:2ndOrder} and \ref{Sec:3rdOrder}, we present the construction of surface topological order for second and third order inversion symmetric topological phases respectively. The main results of the paper are summarised in Table \ref{Table: Main}. Technical details are collected in several appendices.
\section{Surface Topological Order for Second Order Topological Phases} \label{Sec:2ndOrder}
\subsection{Class A $+$ Inversion: HOTI with Chiral Dirac Hinge Mode} \label{Sec:HOTI}
\begin{figure}[tbh!]
\centering
\subfloat[Before pasting STO]{%
\includegraphics[width=.44\linewidth]{drawing_1_b.png}%
}\hfill
\subfloat[After pasting STO]{%
\includegraphics[width=.49\linewidth]{drawing_1.png}%
}
\caption{Surface topological order for second order topological phases protected by inversion symmetry in A class.}
\label{figure_for_sto}
\end{figure}
A second-order topological insulator (HOTI) protected solely by inversion symmetry can be obtained by perturbing a 3D topological insulator with a surface mass term that breaks TRS ($\mathcal{T}$) but preserves inversion symmetry ($\mathcal{I}$) \cite{khalaf_inversion}. The bulk of the 3D topological insulator can be captured by the Bloch Hamiltonian
\begin{equation}
H(\vec{k})=\sum_{i}\sin(k_i)\sigma_i\otimes \tau_x-(2-\sum_{i}\cos(k_i))\sigma_0\otimes \tau_z,
\end{equation}
where $\vec{\sigma}, \vec{\tau}$ are Pauli matrices that act on spin and orbital degrees of freedom respectively. Inversion symmetry and TRS are given by $\mathcal{I}=\sigma_0\otimes \tau_z$ and $\mathcal{T}=i\sigma_y \otimes \tau_0\mathcal{K}$ respectively, where $\mathcal{K}$ denotes complex conjugation. In order to construct the hinge state we first inspect the linearized surface Hamiltonian which takes the form \cite{khalaf_inversion}
\begin{equation}
h(\vec{k},\vec{r})=-(\vec{k}\times \hat{n}_{\vec{r}})\cdot \vec{\tilde{\sigma}},
\end{equation}
where $\hat{n}_{\vec{r}}$ is the unit vector normal to the surface, which is assumed to be spherical. The projected surface Hamiltonian is expressed in a rotated $\tilde{\sigma}$ basis in which the symmetry operators act as $\mathcal{I}=-\tilde{\sigma}_0$ and $\mathcal{T}=i\tilde{\sigma}_y \mathcal{K}$. The surface hinge mode is obtained by perturbing the 3D TI with a TRS breaking mass term of the form $\delta h(\vec{r})=m_{\vec{r}}(\hat{n}_{\vec{r}}\cdot \vec{\tilde{\sigma}})$.
Inversion symmetry imposes the constraint $m_{\vec{r}}=-m_{-\vec{r}}$ on the mass profile, signalling the vanishing of the mass term along some inversion symmetric curve. For concreteness, we consider the setup illustrated in Fig.~\ref{figure_for_sto} wherein the mass changes sign across the equator. It is known that such a domain wall hosts a chiral fermionic mode \citep{Jansen:1994ym}.
In order to gap out the chiral hinge mode, we induce topological orders denoted as $\mathcal A_{\mathrm{N}}$ and $\mathcal A_{\mathrm{S}}$ on the top/bottom halves of the surface. It is worth noting that the calculations below will be similar to the calculations in Ref.~\onlinecite{apoorv_2019}, albeit with an explicit treatment of crystalline symmetry.
3D inversion symmetry imposes $\mathcal{A}_\mathrm{S}=\bar{\mathcal{A}}_\mathrm{N}$, where $\bar{\mathcal{A}}$ denotes the orientation-reversed version of $\mathcal{A}$.
The topological order $\left[\text{Ising}\times \ms{U}(1)_{-8}\right]/\mathbb Z_{2}$ is a suitable choice for $\mathcal A_{\mathrm{N}}$.
This is the same topological order as the $\mathcal T$-Pfaffian, if we ignore TRS, therefore we refer to it as 2D $\mathcal T$-Pfaffian.
The edge theory of $\mathcal A_{\mathrm{N}}$ contains a chiral Majorana mode and an anti-chiral compact boson mode.
These are the edge fields corresponding to the bulk $\text{Ising}$ and $\ms{U}(1)_{-8}$ topological orders respectively.
Concretely, the edge of $\mathcal{A}_\mathrm{N}$ and $\mathcal A_{\mathrm{S}}$ are described by the Lagrangians \cite{Bonderson_2013, Chen_2014}
\begin{align}
\mathcal{L}_{\partial \mathcal{A}_\mathrm{N}}=&\;-\frac{2}{4\pi}\partial_x \phi_\mathrm{N}(\partial_t-\partial_x)\phi_\mathrm{N}+i \gamma_\mathrm{N}(\partial_t+\partial_x)\gamma_\mathrm{N}, \nonumber \\
\mathcal{L}_{\partial \mathcal{A}_\mathrm{S}}=&\;-\frac{2}{4\pi}\partial_x \phi_\mathrm{S}(\partial_t-\partial_x)\phi_\mathrm{S}+i \gamma_\mathrm{S}(\partial_t+\partial_x)\gamma_\mathrm{S}.
\end{align}
where $\gamma_{\mathrm{N}}$ and $\gamma_{\mathrm{S}}$ are Majornana-Weyl modes while $\phi_{\mathrm{N}}$ and $\phi_{\mathrm{S}}$ are compact bosonic modes. The chiral hinge additionally contains a single Dirac mode contributed by the HOTI bulk which can be described by the bosonized Lagrangian \cite{Wen_1990, Wen_1995, Senechal_2004}
\begin{equation}
\mathcal{L}_{0}=\frac{1}{4\pi}\partial_x \phi(\partial_t+\partial_x)\phi,
\label{eq:chiralmode_from_bulk}
\end{equation}
where $\phi$ is a compact boson. The combined Lagrangian describing the equatorial hinge is therefore given by
\begin{align}
\mathcal{L}_{\text{Hinge}}=\mathcal{L}_{\partial \mathcal{A}_\mathrm{N}}+\mathcal{L}_{\partial \mathcal{A}_\mathrm{S}}+\mathcal{L}_{0}.
\end{align}
The two Majorana modes can be combined into a Dirac mode which can be subsequently bosonized and written in terms of the compact boson $\phi_{\mathrm{M}}$ using
\begin{equation}
\psi_{\mathrm{M}}\sim e^{i \phi_\mathrm{M}(x)} \sim e^{-i \frac{\pi}{4}}\gamma_\mathrm{N}+e^{i \frac{\pi}{4}}\gamma_\mathrm{S}, \label{bosonized:1}
\end{equation}
where we have suppressed the Klein factors for brevity \cite{von_Delft_1998}. The benefit of bosonizing the Majorana pair is that it allows for the description of the hinge in terms of the $K$-matrix Luttinger liquid formalism which is easier to work with. The hinge is then described by the Lagrangian
\begin{align}
\mathcal L_{\text{Hinge}}=\frac{1}{4\pi}\partial_{x}\Phi^{\ms{T}} K \partial_t \Phi -\frac{1}{4\pi}\partial_{x}\Phi^{\ms{T}}\partial_{x}\Phi,
\label{eq:HOTI_hinge}
\end{align}
where $\Phi^{\ms{T}}=(\phi_{\mathrm{M}}, \phi_{\mathrm{N}},\phi_{\mathrm{S}},\phi)$, the $K=\text{diag}(1,-2,-2,1)$ and the charge vector $t^{\ms{T}}=(0,1,1,1)$. Our intention is to add to Eq.~\eqref{eq:HOTI_hinge} generic interactions represented by cosine terms that gap out all the hinge modes when driven to strong coupling
\begin{align}
\delta \mathcal{L}&=\sum_{I=1,2}\delta\mathcal L_{I}=\sum_{I=1,2}\lambda_I(x)\cos[\Lambda_I^{\mathsf{T}}K\Phi-\alpha_{I}]
\label{eq:gap_terms}
\end{align}
Apart from being simultaneously gappable (see App.~\ref{appendix:k-matrix} for details), the gapping vectors $\Lambda_{I}$ need to satisfy a number of symmetry criteria related to inversion, charge conservation and gauge symmetry derived from a $\mathbb Z_{2}$ redundancy in our description of the hinge. First, the inversion symmetry acts as $\mathcal I: \Phi(x)\mapsto \mathrm{I}\Phi(-x)$ where $\mathrm{I}=(-1)\oplus \sigma^{x}\oplus (+1)$. Second, in order to respect $\mathsf{U}(1)$ symmetry, we impose charge neutrality condition, namely we require that $\Lambda^{\mathsf{T}}t=0$. Finally, due to the fermionic nature of $\mathcal{A}_\mathrm{N}$ and $\mathcal{A}_\mathrm{S}$, an additional gauge symmetry $\mathbb{Z}^\mathrm{N}_2\times \mathbb{Z}^\mathrm{S}_2$ is imposed. The generators $\ms g_{\alpha}$ of $\mathbb Z_{2}^{\alpha}$ (where $\alpha=\mathrm{N},\mathrm{S}$) implement the transformation
\begin{align}
\ms g_{\mathrm{N},\mathrm{S}}&: \phi_{\mathrm{N},\mathrm{S}}\mapsto \phi_{\mathrm{N},\mathrm{S}}\pm\frac{\pi}{2},\quad \gamma_{\mathrm{N},\mathrm{S}}\mapsto -\gamma_{\mathrm{N},\mathrm{S}}.
\end{align}
Since the fermionic operator $\Psi_{\alpha}\simeq\gamma_{\alpha} e^{2i\phi_{\alpha}}$ is invariant under $\mathbb Z_{2}^{\alpha}$, the gauge symmetry imposes that any admissible cosine term tunnels only local operators, that is, fermions or combinations thereof. Additionally, as a consequence of the $\mathbb Z_{2}^{\mathrm{N}}\times \mathbb Z_{2}^{\mathrm{S}}$ symmetry we need to fix the compactification $\phi_{\mathrm{N},\mathrm{S}}\sim \phi_{\mathrm{N},\mathrm{S}}+\pi$.
Two cosine terms are required to open a gap for the combined hinge theory. The first gapping vector can be chosen to be $\Lambda_1^{\mathsf{T}}=(0,-2,-2,4)$. Such a term is inversion-symmetric if $\lambda_1(x)=\lambda_1(-x)$ [see Eq.~\eqref{eq:gap_terms}] therefore $\lambda_{1}$ can be chosen to be constant and $\alpha=0$, i.e.,
\begin{align}
\delta \mathcal L_{1}=\lambda_{1}\cos\left(4\phi_{\mathrm{N}}+4\phi_{\mathrm{S}}+4\phi\right).
\label{eq:gap_term_1}
\end{align}
This gapping term also respects $\mathsf{U}(1)$ and $\mathbb{Z}^\mathrm{N}_2\times \mathbb{Z}^\mathrm{S}_2$ symmetry as can be checked explicitly. Upon adding Eq.~\eqref{eq:gap_term_1} to the original gapless hinge described by Eq.~\eqref{eq:HOTI_hinge}, the combination of fields $\langle \phi_{\mathrm{N}}+\phi_{\mathrm{S}}+\phi\rangle $ acquire a vacuum/groundstate expectation value, thereby breaking the $\mathbb{Z}^\mathrm{N}_2\times \mathbb{Z}^\mathrm{S}_2$ symmetry into a diagonal $\mathbb Z_{2}$ subgroup denoted as $\mathbb Z_{2}^{\text{diag}}$, generated by $\ms{g}_{\text{diag}}:=\ms g_{\mathrm{N}}\ms g_{\mathrm{S}}$ with the action
\begin{equation}
\ms{g}_{\text{diag}}: \phi_\mathrm{N}\mapsto \phi_\mathrm{N}+\frac{\pi}{2}, \phi_\mathrm{S}\mapsto \phi_\mathrm{S}-\frac{\pi}{2},\phi_{\mathrm{M}}\mapsto \phi_{\mathrm{M}}+\pi.
\end{equation}
The second gapping vector can be chosen as $\Lambda_2^{\mathsf{T}}=(2,1,-1,0)$. Since the two gapping vectors $\Lambda_{1,2}$ satisfy the Haldane criterion $\Lambda_i^{\mathsf{T}}K \Lambda_j=0$, the bosonic fields $\Lambda^{\ms{T}}K\Phi$ can simultaneously acquire a vacuum expectation value. The second gapping term $\delta \mathcal L_2$ is also charge neutral since $\Lambda_2^{\ms{T}}t=0$ as well as invariant under the residual $\mathbb Z_{2}^{\text{diag}}$ symmetry. Finally, the term is $\delta \mathcal L_{2}$ is inversion symmetric if $\alpha_{2}=n\pi$ and $\lambda_{2}(-x)=(-1)^{n}\lambda_{2}(x)$. By choosing $n\in \mathbb Z_{\text{even}}$, we can fix $\lambda_{2}$ to be constant everywhere.
To summarize, we have shown that the two cosine terms corresponding to $\Lambda_{1,2}$ satisfy the symmetry requirements as well as the Haldane criteria. Therefore, they can be simultaneously driven to strong coupling thereby completely gapping out the hinge without breaking any symmetry. We note that an inversion symmetric configuration of $\mathcal A_{\mathrm{N}}$ and $\mathcal A_{\mathrm{S}}$ illustrated in Fig.~\ref{figure_for_sto} without the mode contributed from the bulk is clearly ingappable on the hinge as the modes on the hinge carry a non-vanishing chiral central charge. Therefore as a pattern of 2D inversion symmetric topological order, $\mathcal A_{\mathrm{N}}\oplus \mathcal A_{\mathrm{S}}$ is anomalous and cancels the higher-order anomaly coming from the bulk.
\subsection{Class AII $+$ Inversion: HOTI with Helical Dirac Hinge Mode} \label{Sec:AII_HOTI}
\begin{figure}[tbh!]
\centering
\subfloat[Before pasting STO]{%
\includegraphics[width=.475\linewidth]{drawing_2_b.png}%
}\hfill
\subfloat[After pasting STO]{%
\includegraphics[width=.49\linewidth]{drawing_2.png}%
}
\caption{Surface topological order configuration for inversion and TRS protected higher-order topological insulator. }
\label{figure_for_sto:2}
\end{figure}
It was shown in Refs.~\onlinecite{khalaf_inversion, Khalaf_PRX} that TRS-invariant insulators (with $\mathcal T^{2}=-1$) enriched by additional inversion symmetry can support non-trivial second-order topology. On inversion symmetric open geometries, models within the non-trivial second-order phase host robust helical Dirac modes along an inversion symmetric hinge on the surface. The helical hinge modes are similar to those obtained on the edge of a quantum spin Hall insulator and form a Kramer's pair which is stable against interactions \cite{Kane-Mele-2005}. Here we show that these modes can be gapped out by inducing topological order on the surface. Before getting into the details of the surface topological order, we briefly review the free fermion model for the helical HOTI. The strategy is to start with a doubled model and subsequently add a perturbation that gaps out the surface leaving behind a robust hinge. In this case, the parent theory consists of two 3D topological insulators. We consider the same geometrical settings as in Fig.~\ref{figure_for_sto:2}. The surface Hamiltonian is given by
\begin{equation}
h_1(\vec{k},\vec{r})=\tau_0\otimes (\vec{k}\times \hat{n}_{\vec{r}})\cdot\vec{\sigma}.
\end{equation}
The inversion and TRS are represented by $\mathcal{I}=\tau_0\otimes (-\sigma_0)$ and $\mathcal{T}=\tau_0\otimes (i\sigma_y\mathcal{K})$. A surface mass term that respects TRS can be added to the Hamiltonian
\begin{equation}
\delta h_1(\vec{r})=\tau_y\otimes m_{\vec{r}}(\hat{n}_{\vec{r}}\cdot \vec{\sigma}).
\end{equation}
Inversion symmetry demands that $\mathcal{I}\delta h_1(\vec{r})\mathcal{I}^{-1}=\delta h_1(-\vec{r})$ which further imposes the condition $m_{\vec{r}}=-m_{-\vec{r}}$, signalling the vanishing of the mass term along some inversion symmetric curve. For reasons identical to the chiral HOTI case, this indicates the existence of gapless helical modes along the equator. TRS in the above construction acts within each flavor of the above model, i.e diagonally in the $\vec{\tau} $ space. We find it convenient to work with an equivalent description of the helical HOTI in which TRS acts by switching fermion flavors, such that the model can be thought of as a stacking of a chiral HOTI with its time-reversed copy. We consider the following Hamiltonian which is related to $h_1+\delta h_{1}$ by a unitary transformation
\begin{align}
h_2(\vec{k}, \vec{r})&=\tilde{\tau}_0 \otimes (\lambda_y\tilde{\sigma}_y+m_{\vec{r}}n_x\tilde{\sigma}_x+m_{\vec{r}}n_z\tilde{\sigma}_z)\nonumber \\
&+\tilde{\tau}_z \otimes (\lambda_x\tilde{\sigma}_x+\lambda_z\tilde{\sigma}_z+m_{\vec{r}}n_y\tilde{\sigma}_y), \label{hamiltonian:2}
\end{align}
where $\vec{\lambda}\equiv \vec{k}\times \hat{n}_{\vec{r}}$. TRS is represented by
\begin{equation}
\mathcal{T}=i\tilde{\tau}_y\otimes \tilde{\sigma}_0 \mathcal{K}, \quad \ \mathcal T^{2}=-1. \label{trs:2}
\end{equation}
Upon performing an analysis similar to the one described above one is left with a surface which is gapped everywhere except an inversion symmetric hinge which hosts a pair of gapless helical modes $\left\{\psi, \bar{\psi}\right\}$ that form a Kramers doublet.
\noindent A natural candidate for an STO that can gap out the helical hinge mode is given by stacking the STO from the previous section and its time-reversed copy. It remains to be shown that this construction furnishes modes on the hinge that are robust by themselves, but when considered along with the helical modes contributed by the bulk lead to a completely gapped hinge. As illustrated in Fig.~\ref{figure_for_sto:2}, we set up the STO configuration by placing $(\mathcal{A}_{\mathrm{N}},\mathcal{A}_{\mathrm{S}})$ on the northern/southern hemispheres of a spherical surface geometry. Here $\mathcal A_{\mathrm{N},\mathrm{S}}$ stand for the product topological orders consisting of 2D $\mathcal{T}$-Pfaffian topological orders and their time-reversed copies. The edge theory contains 8 modes. We divide these into bosonic modes $\Phi=(\phi_\mathrm{N}, \phi_\mathrm{S},\bar{\phi}_{\mathrm{N}},\bar{\phi}_{\mathrm{S}})^{\ms{T}}$ and fermionic modes $\Gamma=(\gamma_{\mathrm{N}},\gamma_{\mathrm{S}},\bar{\gamma}_\mathrm{N},\bar{\gamma}_\mathrm{S})^{\ms{T}}$. The action of TRS is encoded in the matrices $T_{\Phi}:=\sigma^{x}\otimes \mathbb 1_{2}$ and $T_{\Gamma}=i\sigma^{y}\otimes\mathbb 1_{2}$ such that under TRS
\begin{align}
\mathcal T:&\;
\begin{bmatrix}
\Phi \\
\Gamma
\end{bmatrix}
\longmapsto
\begin{bmatrix}
T_{\Phi}\Phi \\
T_{\Gamma}\Gamma
\end{bmatrix}, \quad
i \ \longmapsto -i.
\end{align}
As before we need to impose a gauge symmetry that ensures that the cosine terms only tunnel combinations of fields that are built from local fermionic operators. The full fermionic gauge symmetry group is $\mathbb Z_{2}^{4}=\prod_{\alpha}\mathbb{Z}^{\alpha}$ where $\alpha=\mathrm{N,S,\bar{N},\bar{S}}$. The generators of this group denoted as $\ms{g}_{\alpha}$ act as
\begin{align}
\ms{g}_{\alpha}:&\; \begin{bmatrix}
\gamma_{\alpha} \\
\phi_{\alpha}
\end{bmatrix}
\longmapsto
\begin{bmatrix}
-\gamma_{\alpha} \\
\phi_{\alpha}
\end{bmatrix}
+
s_{\alpha} \begin{bmatrix}
0 \\
\frac{\pi}{2}
\end{bmatrix},
\label{eq:fourgauge}
\end{align}
where $s_{\alpha}=-1$ for $\alpha=\mathrm{S},\mathrm{S}_{T}$ and $+1$ otherwise. Inversion squares to +1 and simply maps the fields on the northern hemisphere to their counterparts on the southern hemisphere and vice versa. We now proceed to gap out the edge modes. Firstly we combine the Majorana fermions into Dirac fermions
\begin{align}
\psi_\mathrm{M}&\sim e^{i\phi_\mathrm{M}}\sim e^{-i\frac{\pi}{4}}\gamma_\mathrm{N}+e^{i\frac{\pi}{4}}\gamma_\mathrm{S}, \nonumber \\ \bar{\psi}_\mathrm{M}&\sim e^{-i\bar{\phi}_\mathrm{M}}\sim e^{i\frac{\pi}{4}}\bar{\gamma}_\mathrm{N}+e^{-i\frac{\pi}{4}}\bar{\gamma}_\mathrm{S}.
\end{align}
Since the Majorana fermions are themselves Kramers pairs, the action of TRS can be deduced as
\begin{align}
\mathcal T:&\;
\begin{bmatrix}
\phi_\mathrm{M} \\
\bar{\phi}_\mathrm{M}
\end{bmatrix}
\longmapsto
\begin{bmatrix}
\bar{\phi}_\mathrm{M} \\
\phi_\mathrm{M}+\pi
\end{bmatrix}.
\end{align}
The edge is now effectively described by the following $K$-matrix and charge vector $t$
\begin{align}
K&=\mathrm{diag}(-1,2,2,-1,1,-2,-2,1),\nonumber \\
t&=(0,1,1,1,0,1,1,1)^{\mathsf{T}},
\end{align}
in the basis $(\phi_\mathrm{M},\phi_\mathrm{N}, \phi_\mathrm{S},\phi,\bar{\phi}_\mathrm{M},\bar{\phi}_{\mathrm{N}},\bar{\phi}_{\mathrm{S}},\bar{\phi})^{\ms{T}}$. Consider the gapping terms
\begin{align}
\hspace{-7pt}\delta \mathcal{L}&=\cos[4\phi_\mathrm{N}+4\phi_\mathrm{S}+4\phi]+\cos[2\phi_\mathrm{N}-2\phi_\mathrm{S}-2\phi_\mathrm{M}]\nonumber \\
&+\cos[4\bar{\phi}_\mathrm{N}+4\bar{\phi}_\mathrm{S}+4\bar{\phi}]+\cos[2\bar{\phi}_\mathrm{N}-2\bar{\phi}_\mathrm{S}-2\bar{\phi}_\mathrm{M}].
\end{align}
Note that this expression is basically the gapping term for an inversion symmetric HOTI plus its time-reversed copy. Thus, we only need to check whether the above expression breaks TRS. Clearly, it does not break TRS explicitly; however, since both $\langle \bar{\phi}_\mathrm{N}+\bar{\phi}_\mathrm{S}+\bar{\phi}\rangle $ and $\langle\bar{\phi}_\mathrm{N}-\bar{\phi}_\mathrm{S}-\bar{\phi}_\mathrm{M} \rangle$ transform to their TRS copies with extra $\pi$ phase, naively it seems like TRS is broken spontaneously. We note that the gauge group is now broken to $\mathbb{Z}^{\mathrm{diag}}_2\times \bar{\mathbb{Z}}^{\mathrm{diag}}_2$, where
\begin{align}
\ms{g}_{\text{diag}}&: \phi_\mathrm{N}\mapsto \phi_\mathrm{N}+\frac{\pi}{2}, \phi_\mathrm{S}\mapsto \phi_\mathrm{S}+\frac{\pi}{2},\phi_{\mathrm{M}}\mapsto \phi_{\mathrm{M}}+\pi,\nonumber \\
\bar{\ms{g}}_{\text{diag}}&: \bar{\phi}_\mathrm{N}\mapsto \bar{\phi}_\mathrm{N}+\frac{\pi}{2}, \bar{\phi}_\mathrm{S}\mapsto \bar{\phi}_\mathrm{S}+\frac{\pi}{2},\bar{\phi}_{\mathrm{M}}\mapsto \bar{\phi}_{\mathrm{M}}+\pi.
\end{align}
We can see that $\langle \phi_\mathrm{N}+\phi_\mathrm{S}+\phi \rangle\sim \langle \phi_\mathrm{N}+\phi_\mathrm{S}+\phi \rangle+\pi$, $\langle \phi_\mathrm{N}-\phi_\mathrm{S}-\phi_\mathrm{M} \rangle\sim \langle \phi_\mathrm{N}-\phi_\mathrm{S}-\phi_\mathrm{M} \rangle+\pi$ as they are related to each other by a gauge transformation. Therefore TRS is not broken spontaneously either. Finally, we emphasize that without inversion symmetry, the above STO is non-anomalous as we can paste a copy of a quantum spin Hall liquid on, e.g., the southern hemisphere, with its edge modes residing on the equator and gap out the helical modes contributed by the STO without invoking the bulk.
\subsection{Class D $+$ Inversion: HOTSC with Chiral Majorana Hinge Mode} \label{Sec:ClassD_HOTSC}
We briefly review the free fermion model for the chiral HOTSC in class D. The strategy to construct a second-order phase is to start with a class DIII topological superconductor and add a TRS-breaking perturbation that gaps out the surface leaving behind a robust hinge protected by inversion symmetry. The surface Hamiltonian is given by \cite{khalaf_inversion}
\begin{equation}
h(\vec{k},\vec{r})=-(\vec{k}\times \hat{n}_{\vec{r}})\cdot\vec{\sigma},
\end{equation}
The inversion, time reversal and particle-hole symmetries are generated by $\mathcal{I}=-\sigma_0$, $\mathcal{T}=i\sigma_y \mathcal K$ and $\mathcal P= -(\hat{n}_{\vec{r}}\cdot\vec{\sigma})\sigma_{y}\mathcal{K}$. The surface can be deformed by the mass term
\begin{equation}
\delta h(\vec{r})=m_{\vec{r}}(\hat{n}_{\vec{r}}\cdot \vec{\sigma}),
\end{equation}
that breaks TRS. Inversion symmetry demands that $\mathcal{I}\delta h(\vec{r})\mathcal{I}^{-1}=\delta h(-\vec{r})$ which further imposes the condition $m_{\vec{r}}=-m_{-\vec{r}}$, signalling the vanishing of the mass term along some inversion symmetric curve which hosts gapless chiral modes. Due to the additional particle-hole symmetry as compared with the class A chiral HOTI, these are Majorana as opposed to Dirac modes.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{drawing_3.png}
\caption{An illustration of the procedure used to gap out the $(3,0)$ configuration of chiral Majorana modes $\chi_{i}$ with $i=1,2,3$. We introduce $\gamma^{N/S}_i$ Majoranas by pasting $p\pm ip$ superconductors on the surface and end up with a total of nine Majorana modes on the hinge. These are subsequently gapped out by introducing the STO $\mathcal A_{N,S}$ on the northern and southern hemispheres respectively. The nine Majorana modes are described by the $\ms{SO}(9)_1$ WZW which splits into two copies of $\ms{SO}(3)_3$ theories that gap out upon coupling to the edge modes provided by the STOs.}
\label{STO_Chiral_HOTSC}
\end{figure}
Before turning to the surface topological order, we first inspect the stability of the chiral hinge modes to inversion symmetric surface pasting of $p\pm ip$ superconductors.
%
Let us consider the situation where there are $N_{\pm}$ chiral co-propagating Majorana hinge modes denoted as $\chi^{\pm}_{i}$ with $i=1,\dots,N_{\pm}$ that transform under inversion as
\begin{align}
\mathcal I:
\begin{bmatrix}
\chi_{i}^{\pm}(\theta)\\
\chi_{i}^{\pm}(\theta+\pi)
\end{bmatrix}\mapsto
\begin{bmatrix}
\pm i\chi_{i}^{\pm}(\theta+\pi)\\
\mp i\chi_{i}^{\pm}(\theta)
\end{bmatrix},
\end{align}
where $\theta$ is introduced to parameterize the equator on which the Majoranas are propagating. The above symmetry action can be derived from the bulk symmetry using a recursive Jackiw-Rebbi procedure (see App.~\ref{App:Jackiw_Rebbi}) \citep{jackiw_rebbi, minghao_notes} and satisfies the basis-invariant relations $\{\mathcal{I}, \mathcal{P}\}=0$, $\left[\mathcal I,\mathcal T\right]=0$ and $\mathcal I^{2}=1$. we can always paste a $p+ ip$ and $p- ip$ topological superconductor on the northern and southern hemispheres respectively, which contribute a pair of chiral hinge modes denoted as $\gamma^{\mathrm{N},\mathrm{S}}$. The inversion action on these modes is represented as
\begin{align}
\mathcal I:
\begin{bmatrix}
\gamma^{\mathrm{N}}(\theta)\\
\gamma^{\mathrm{N}}(\theta+\pi)\\
\gamma^{\mathrm{S}}(\theta)\\
\gamma^{\mathrm{S}}(\theta+\pi)
\end{bmatrix}\mapsto
\begin{bmatrix}
i\gamma^{\mathrm{S}}(\theta+\pi)\\
-i\gamma^{\mathrm{S}}(\theta)\\
i\gamma^{\mathrm{N}}(\theta+\pi)\\
-i\gamma^{\mathrm{N}}(\theta)
\end{bmatrix}.
\end{align}
Consider the linear combinations $\gamma^{\pm}=(\gamma^{\mathrm{N}}\pm \gamma^{\mathrm{S}})/\sqrt{2}$ that transform under inversion as
\begin{align}
\mathcal I:
\begin{bmatrix}
\gamma^{\pm}(\theta)\\
\gamma^{\pm}(\theta+\pi)
\end{bmatrix}\mapsto
\begin{bmatrix}
\pm i\gamma^{\pm}(\theta+\pi)\\
\mp i\gamma^{\pm}(\theta)
\end{bmatrix}.
\end{align}
Henceforth we denote left/right-moving modes with/without an overbar. The configuration $(N_{+},N_{-})$ with net $N_{+}+N_{-}$ right movers can always be transformed to $(N_{+}-1,N_{-}-1)$ by surface pasting. Therefore we have the equivalence relation
\begin{align}
(N_{+},N_{-})\sim (N_{+}-n,N_{-}-n),
\label{eq:hinge_config}
\end{align}
where $n\in \mathbb Z$. Consequently, we can always transform a configuration $(N_{+},N_{-})$ into a configuration with all positive parity modes $(N_{+}-N
_{-},0)$.
For this reason we will only need to consider the stability of such modes under surface pasting of topologically ordered phases.
The classification of inversion symmetry-protected higher-order phases in class D is given by the group $\mathbb Z_{4}$ which can be indexed by $(N_+,0)$ (See Sec.~\ref{Sec:3rdOrder} for details).
For the present discussion it will suffice to construct the STO for the generator of $\mathbb Z_{4}$ which may be treated as $(3,0)$.
Since an odd number of Majorana fermions cannot be tamed by Abelian bosonization, the $K$-matrix approach we previously employed must be abandoned.
Instead, we use non-Abelian bosonization to approach the problem.
We remark here that the method in this section is similar in spirit to Ref.~\onlinecite{Meng_2018,teo_majorana}, however, adapted to inversion symmetry.
We consider the HOTSC to have a spherical geometry as illustrated in Fig.~\ref{STO_Chiral_HOTSC} with three chiral majorana modes $\chi_{i}^{+}$ on the inversion-symmetric equator. In order to show that the chiral hinge can be gapped, we find it convenient to proceed in two steps. First we add additional degrees of freedom on the hinge by a purely surface pasting of $p\pm ip$ superconductors that preserves the inversion symmetry.
In the second step, we induce inversion-symmetric topological order on the surface to gap out the combined hinge modes contributed by the $p\pm ip$ superconductors and the bulk higher-order superconductor.
We begin by adding three copies of a $p+ip$ superconductor on the northern hemisphere and three copies of a $p-ip$ superconductor on the southern hemisphere.
As a result, we end up with $6$ additional Majorana modes on the equator which we label as $\{\gamma^\mathrm{N}_i,\gamma^\mathrm{S}_i \}$, with $i=1,\dots, 3$.
The hinge is described by the $\mathfrak{so}(9)_1$ WZW theory \citep{cft_fran}
\begin{align}
S=\int\mathrm{d}t\mathrm{d}\theta i{\Psi}^{\ms{T}}(
\partial_{t}-\partial_{\theta}){\Psi},
\label{eq:so(9)_1_wzw}
\end{align}
where we have introduced a 9 component Majorana spinor field $\Psi$. The operator product expansion (OPE) of the Majorana operators satisfies the standard relations
\begin{align}
\chi_{i}^{+}(z)\chi_{j}^{\mathrm{N}}(w)\sim&\; \frac{\delta_{ij}}{z-w}+\dots, \nonumber \\
\gamma_{i}^{\mathrm{N}}(z)\gamma_{j}^{\mathrm{N}}(w)\sim&\; \frac{\delta_{ij}}{z-w}+\dots, \nonumber \\
\gamma_{i}^{\mathrm{S}}(z)\gamma_{j}^{\mathrm{S}}(w)\sim&\; \frac{\delta_{ij}}{z-w}+\dots.
\end{align}
We introduce $\mathfrak{so}(9)_{1}$ currents which can be expressed as fermion biliears,
\begin{align}
\mathcal J^{\ms{A}}(z)=&\; \frac{i}{2}{\Psi^{\dagger}}(z)\Sigma^{\ms{A}}{\Psi}(z),
\label{eq:current_defn}
\end{align}
where $\ms{A}$ is a Lie-algebra index, $\Sigma^{\ms{A}}$ are the generators of the $\mathfrak{so}(9)$ lie algebra and $z$ are holomorphic coordinates defined as $z=\theta+i t$ on the hinge. The currents $\left\{\mathcal J^{\ms{A}} \right\}_{\ms{A}=1,\dots \text{dim}(\mathfrak{so}(9))}$ satisfy the OPE
\begin{align}
\mathcal{J}^{\ms{A}}(z) \mathcal{J}^{\ms{B}}(w)\sim \frac{\delta^{\ms{AB}}}{(z-w)^{2}}+ \frac{i f^{AB}_{C}\mathcal J^{C}(w)}{z-w} + \dots,
\end{align}
where $f^{\ms{AB}}_{\ms{C}}$ are the structure constants for $\mathfrak{so}(9)$. The action of inversion on the different Majorana operators is as follows
\begin{align}
\mathcal I:&\; \chi^{+}_i(\theta) \longrightarrow i\chi^{+}_{i}(\theta+\pi), \nonumber \\
\mathcal I :&\; \gamma_{i}^{\mathrm{N}}(\theta)\longrightarrow i\gamma_{i}^{\mathrm{S}}(\theta+\pi), \nonumber \\
\mathcal I:&\; \gamma_{i}^{\mathrm{S}}(\theta)\longrightarrow i\gamma_{i}^{\mathrm{N}}(\theta+\pi).
\end{align}
In order to construct the surface topological order that can absorb the Majorana hinge modes, it is useful to work with an embedding of $ \mathfrak{so}(3)^{(1)}_{3}\times \mathfrak{so}(3)^{(2)}_3\subset \mathfrak{so}(9)_{1} $. Since inversion symmetry is an essential part of our setup, we need to be careful about its action on the various embedded components. We work with a choice of embedding such that the two copies of $\mathfrak{so}(3)_3$ are swapped under the action of inversion. Let us index the components of the spinor $\Psi$ by a tuple $(i,j)$ where $i,j=1,2,3$. We define the different components such that they have the following simple transformation rule under inversion
\begin{align}
\mathcal I:&\Psi_{(i,j)}(\theta) \longmapsto i\Psi_{(j,i)}(\theta+\pi)
\end{align}
with
\begin{align}
\Psi_{(i,i)}=&\; \chi^{+}_{i}, \nonumber \\
\Psi_{(2,3)}=&\; \gamma_{1}^{\mathrm{N}}, \quad \Psi_{(3,2)}= \gamma_{1}^{\mathrm{S}}, \nonumber \\
\Psi_{(3,1)}=&\; \gamma_{2}^{\mathrm{N}}, \quad \Psi_{(1,3)}= \gamma_{2}^{\mathrm{S}}, \nonumber \\
\Psi_{(1,2)}=&\; \gamma_{3}^{\mathrm{N}}, \quad \Psi_{(2,1)}= \gamma_{3}^{\mathrm{S}}.
\label{tupleassignchiral}
\end{align}
In order to construct the $\mathfrak{so}(3)_3$ current operators, consider the matrices defined as $\sigma^{\ms{a},1}:=\mathrm{L}^{\ms{a}}\otimes \mathrm{Id}_{3}$ and $\sigma^{\ms{a},2}:=\mathrm{Id}_{3}\otimes \mathrm{L}^{\ms{a}}$, $\ms{a}=1,2,3$, where $\mathrm{L}^{\ms{a}}$ are the generators of $\mathfrak{so}(3)$ in the fundamental representation. These matrices generate two decoupled $\mathfrak{so}(3)$ algebras
\begin{align}
\left[\sigma^{\ms{a},\kappa},\sigma^{\ms{b},\kappa'}\right]=\delta^{\kappa\kappa'}f^{\ms{ab}}_{\ms{c}}\sigma^{\ms{c},\kappa}.
\label{Eq:embedded_matrices}
\end{align}
Using this decomposition we define the following $\mathfrak{so}(3)_{3}\times \mathfrak{so}(3)_{3}$ currents
\begin{equation}
\mathcal J^{\ms{a},\kappa}=\frac{i}{2}\Psi^{\dag}\sigma^{\ms{a},\kappa}\Psi,
\end{equation}
which explicitly take the form
\begin{align}
\mathcal J^{1,1}=&\; \frac{i}{2}\left[\left(\gamma^{\mathrm{S}}_{3}\right)^{(\dag)}\gamma^{\mathrm{N}}_{2}+\left(\chi_2^{+} \right)^{(\dag)}\gamma_{1}^{\mathrm{S}}+\left(\gamma_{1}^{\mathrm{N}}\right)^{(\dag)}\chi_{3}^{+}\right]+\text{h.c.}, \nonumber \\
\mathcal J^{1,2}=&\; \frac{i}{2}\left[\left(\gamma^{\mathrm{N}}_{3}\right)^{(\dag)}\gamma^{\mathrm{S}}_{2}+\left(\chi_{2}^{+}\right)^{(\dag)}\gamma_{1}^{\mathrm{N}}+\left(\gamma_{1}^{\mathrm{S}}\right)^{(\dag)}\chi_{3}^{+}\right]+\text{h.c.}, \nonumber \\
\mathcal J^{2,1}=&\; \frac{i}{2}\left[\left(\chi_{1}^{+}\right)^{(\dag)}\gamma^{\mathrm{N}}_{2}+\left(\gamma_{3}^{\mathrm{N}}\right)^{(\dag)}\gamma_{1}^{\mathrm{S}}+\left(\gamma_{2}^{\mathrm{S}}\right)^{(\dag)}\gamma_{3}^{+}\right]+\text{h.c.}, \nonumber \\
\mathcal J^{2,2}=&\; \frac{i}{2}\left[\left(\chi_{1}^{+}\right)^{(\dag)}\gamma^{\mathrm{S}}_{2}+\left(\gamma_{3}^{\mathrm{S}}\right)^{(\dag)}\gamma_{1}^{\mathrm{N}}+\left(\gamma_{2}^{\mathrm{N}}\right)^{(\dag)}\chi_{3}^{+}\right]+\text{h.c.}, \nonumber \\
\mathcal J^{3,1}=&\; \frac{i}{2}\left[\left(\chi_{1}^{+}\right)^{(\dag)}\gamma^{\mathrm{S}}_{3}+\left(\gamma_{3}^{\mathrm{N}}\right)^{(\dag)}\chi_{2}^{+}+\left(\gamma_{2}^{\mathrm{S}}\right)^{(\dag)}\gamma_{1}^{\mathrm{N}}\right]+\text{h.c.}, \nonumber \\
\mathcal J^{3,2}=&\; \frac{i}{2}\left[\left(\chi_{1}^{+}\right)^{(\dag)}\gamma^{\mathrm{N}}_{3}+\left(\gamma_{3}^{\mathrm{S}}\right)^{(\dag)}\chi_{2}^{+}+\left(\gamma_{2}^{\mathrm{N}}\right)^{(\dag)}\gamma_{1}^{\mathrm{S}}\right]+\text{h.c}.. \label{currentschiralhinge}
\end{align}
The reason why we write $(\dag)$ on the Majorana operator is to remind ourselves of the subtlety related to the imaginary action of inversion, e.g., if $\mathcal I: \chi^{+}_i(\theta) \mapsto i\chi^{+}_{i}(\theta+\pi)$, then $\mathcal I: (\chi^{+}_i(\theta))^{(\dag)} \mapsto (\chi^{+}_{i}(\theta+\pi))^{(\dag)}(-i)$. We can verify that inversion acts as $\mathcal{I}: \mathcal{J}^{\ms{a},1}\leftrightarrow\mathcal{J}^{\ms{a},2}$ on the $\mathfrak{so}(3)$ currents. From the standard OPE for Majorana operators, we can extract the OPE for the $\mathfrak{so}(3)_{3}$ currents, and verify that the level is indeed 3 (see App.~\ref{appendix:wzw}). The stress tensor decomposes as
\begin{equation}
T_{\mathfrak{so}(9)_1}=T_{\mathfrak{so}(3)^{(1)}_3}+T_{\mathfrak{so}(3)^{(2)}_3},
\end{equation}
which means that the chiral central charges of the embedded sectors add up to give the chiral central charge of the $\mathfrak{so}(9)_1$ WZW theory. Having formulated the hinge modes as two copies of $\mathfrak{so}(3)_3$, it is a straightforward task to gap them out by adding surface topological order. We introduce $\mathcal A_{\mathrm{N}}=\overline{\mathcal A_{\mathrm{S}}}=\overline{\ms{SO}(3)}_{3}$ whose edge conformal field theories and corresponding current operators we denote as $\overline{\mathfrak{so}(3)}_{3,\mathrm{N/S}}$ and $\bar{\mathcal J}^{\ms{a}}_{\mathrm{N/S}}$, respectively \cite{Meng_2018}. Under inversion the currents transform as
\begin{equation}
\mathcal{I}: \bar{\mathcal J}^{\ms{a}}_{\mathrm{N}}(\theta)\longmapsto \bar{\mathcal J}^{\ms{a}}_{\mathrm{S}}(\theta+\pi).
\end{equation}
The hinge modes $\mathcal{J}^{\ms{a},\kappa}$ and the edge modes of the surface topological order $\bar{\mathcal{J}}^{\ms{a}}_{\mathrm{N/S}}$ can together be gapped out upon adding the gapping term
\begin{equation}
\delta \mathcal{L}=\lambda(\theta)\sum_{\ms{a}=1}^{3}\Big[\bar{\mathcal J}^{\ms{a}}_\mathrm{N}(\theta)\mathcal J^{\ms{a},1}(\theta)+\bar{\mathcal J}^{\ms{a}}_\mathrm{S}(\theta)\mathcal J^{\ms{a},2}(\theta)\Big],
\end{equation}
which is inversion symmetric if $\lambda(\theta)=\lambda(\theta+\pi)$. Therefore, we can choose $\lambda$ to be constant. To summarize, we have shown that the hinge modes $(3,0)$ can first be mapped to $(3,6)$ by purely surface pasting of $p\pm ip$ superconductors. Thereafter, two copies of $\mathfrak{so}(3)_3$ can be embedded in the $(3,6)$ configuration which can be gapped out by a surface pasting of $\mathsf{SO}(3)_3$ topological order.
\subsection{Class DIII $+$ Inversion: HOTSC with Helical Majorana Hinge Modes}
\label{sec:STO_DIII}
\noindent Class DIII superconductors enriched by inversion symmetry support non-trivial second and third-order topological phases which host robust helical modes and Majorana Kramers zero modes on inversion symmetric loci on the surface \cite{khalaf_inversion}. The helical hinge modes are similar to those obtained on the edge of a 2D TRS invariant topological superconductor, i.e they contain a Majorana Kramers' pair of counter-propagating modes. Here we show that these modes can be gapped out if we allow for the possibility of surface topological order.
First we briefly review the free-fermion model for the helical HOTSC. We can start with two class DIII topological superconductors with opposite topological index and add symmetry-respecting perturbations that gap out the surface, leaving behind a robust hinge. The surface Hamiltonian prior to adding such a mass term is given by \begin{equation}
h(\vec{k},\vec{r})=-\rho_z\otimes (\vec{k}\times \hat{n}_{\vec{r}})\cdot\vec{\sigma}, \label{eq:2ndorderdiii}
\end{equation}
where $\rho_{\mu}$ and $\sigma_{\mu}$ are the Pauli matrices in the orbital and spin space respectively. The inversion, time reversal and particle-hole symmetries are generated by $\mathcal{I}=-\rho_z\otimes \sigma_0$, $\mathcal{T}=\rho_0\otimes i\sigma_y \mathcal {K}$ and $\mathcal P= -\rho_{z}\otimes (\hat{n}_{\vec{r}}\cdot \vec{\sigma})\sigma_{y}\mathcal{K}$. To perturb the Hamiltonian, we add the following mass term that respects the symmetries
\begin{equation}
\delta h(\vec{r})=m_{\vec{r}}\rho_x\otimes \sigma_0.\label{eq:mass2ndorderdiii}
\end{equation}
Inversion symmetry demands that $\mathcal{I}\delta h(\vec{r})\mathcal{I}^{-1}=\delta h(-\vec{r})$ which further imposes the condition $m_{\vec{r}}=-m_{-\vec{r}}$, signalling the vanishing of the mass term along some inversion symmetric curve which we choose to be at the equator. After Jackiw-Rebbi projection, we may verify that a pair of helical Majorana modes resides at the hinge. \par
The classification of inversion symmetry enriched higher-order phases in class DIII is given by the group $\mathbb Z_{4}$ which can be indexed by the number of Majorana helical hinge modes modulo 4 (See Sec.~\ref{Sec:3rdOrder} for details).
For the present discussion it will suffice to construct the STO for the generator of $\mathbb Z_{4}$ which may be treated as $(3,0)$.
We first show that an odd number of helical modes is stable to weak interaction, and they are only unstable to adding topological order on the surface.
\subsubsection{Stability of odd number of helical hinge modes}
For concreteness, let us begin with the setup with three pairs of helical Majorana modes on the hinge. The Lagrangian density for the Majorana modes can be written as
\begin{align}
\mathcal L=\sum_{j=1}^{2n+1}\Big[i\chi_j(
\partial_{t}-\partial_{x})\chi_{j}+i\bar{\chi}_j(
\partial_{t}+\partial_{x})\bar{\chi}_{j}\Big],
\label{eq:action_so(2n+1)WZW}
\end{align}
with $n=1$ which is the non-chiral $\mathfrak{so}(3)_{1}$ WZW theory. For simplicity, we drop the explicit hermitian conjugation from the equations in this subsection as they do not play a role unless we are dealing with a imaginary symmetry representation such as inversion in Eq.~\eqref{eq:inv_repn_maj}. Here we are only interested in the stability of pairs of helical-modes under TRS which has a real representation. The holomorphic and anti-holomorphic currents that generate the $\mathfrak{so}(3)$ current algebra are
\begin{align}
\mathcal J^{\ms{a}}=&\; \frac{i}{2}\chi_j \mathrm{L}^{\ms{a}}_{jk} \chi_k=\frac{i}{2}\epsilon^{\ms{a}jk}\chi_{j}\chi_{k} \nonumber \\
\bar{\mathcal J}^{\ms{a}}=&\; \frac{i}{2}\bar{\chi}_j \mathrm{L}^{\ms{a}}_{jk} \bar{\chi}_k=\frac{i}{2}\epsilon^{\ms{a}jk}\bar{\chi}_{j}\bar{\chi}_{k} ,
\end{align}
where $\ms{a}=1,\dots,\text{dim}(\mathfrak{so}(3))$, and as before $\mathrm{L}^{\ms{a}}$ are the generators for the $\mathfrak{so}(3)$ Lie algebra. The model is TRS invariant with the TRS action given by Eq.~\eqref{eq:TRS_Majorana} for each pair $(\chi_{j},\bar{\chi}_{j})$. We are interested in the stability of this model to TRS invariant perturbations. More precisely, whether the theory can be completely gapped out without breaking TRS. At the quadratic level we can add the following terms to the Hamiltonian
\begin{align}
\delta H =&\; \sum_{j,k,l}im_{j}\epsilon^{jkl}(\chi_k\bar{\chi}_l+\bar{\chi}_k\chi_l)+\sum_{j}i\widetilde{m}_j\chi_j\bar{\chi}_j \nonumber \\
=&\; \sum_{j}\left[ m_{j}\mathcal O^{j} + i\widetilde{m}_{j}\chi_{j}\bar{\chi}_{j}\right],
\end{align}
where, in the second second line we have defined the fermion bilinear $\mathcal O^{j}=i\epsilon^{jkl}(\chi_{k}\bar{\chi}_{l}
+\bar{\chi}_{k }\chi_{l})$. TRS imposes that $\widetilde{m}_{j}=-\widetilde{m}_{j}=0$, while there are no such constraints on $m_{j}$. The operators $\mathcal O^{j}$ satisfy the algebra
\begin{align}
\left[\mathcal O^{j},\mathcal O^{k}\right]=4(\chi_{j}{\chi}_{k}+\bar{\chi}_{j}\bar{\chi}_{k}),
\label{eq:non_abelian_haldane}
\end{align}
which suggests that these operators cannot condense/acquire a ground state expectation value simultaneously. This may pose an obstruction to symmetrically gapping out the theory. Since the model is quadratic, we simply diagonalize the Hamiltonian and check whether this is the case. The full Hamiltonian reads
\begin{align}
H=&\;\sum_j\int\mathrm{d}x\Big\{ iv \left(\chi_j\partial_x \chi_j-\bar{\chi}_j\partial_x \bar{\chi}_j\right)+\sum_{j}im_{j}\mathcal O^{j}\Big\} \nonumber \\
=&\; \int_{k}dk \Psi^{\ms{T}}_kH(k) \Psi_{k},
\end{align}
where in the second line, we have introduced the spinor $\Psi^{\ms{T}}=(\chi_1, \bar{\chi}_1, \chi_2, \bar{\chi}_2,\chi_3, \bar{\chi}_3)$ and transformed to momentum space. The explicit form of $H(k)$ is
\begin{align}
H(k)= vk \mathrm{Id}_{3}\otimes \sigma^{z}+ \sum_{j}i m_{j}\mathrm{L}^{j}\otimes \sigma^{y}.
\end{align}
The spectrum of $H(k)$ is gapless with the eigenvectors $|\psi_{1}\rangle =(m_1,m_2,m_3)^{\ms{T}}\otimes (1,0)^{\ms{T}}$ and $|\psi_{2}\rangle =(m_1,m_2,m_3)^{\ms{T}}\otimes (0,1)^{\ms{T}}$ having eigenvalues $\pm vk$. We are able to find the above two vectors due to the simple condition that
\begin{align}
\text{Ker}\Big[\sum_{j} m_{j}\mathrm{L}^{j}\Big]\neq \varnothing,
\end{align}
which follows from the fact that $\mathrm{L\equiv} \sum_{j}m_{j}\mathrm{L}^{j}$ is a generic $3\times 3$ anti-symmetric matrix, and is therefore singular, since
\begin{equation}
\mathrm{det}(\mathrm{L}^{\ms T})=\mathrm{det}(-\mathrm{L})=(-1)^3\mathrm{det}(\mathrm{L}).
\end{equation}
The above argument can be directly generalized to any odd number of helical modes, since the corresponding $\mathrm{L}$ will always be singular, and result in the existence of gapless eigenvectors. More generally, we may use a mathematical theorem \citep{yangrmp1962} that states any anti-symmetric matrix $\mathrm{L}$ can be block diagonalized by conjugating with an orthogonal matrix. We can then verify that for an even number of helical modes, we can block diagonalize the Hamiltonian into $2\times 2$ blocks with a gapped spectrum.
Having established the stability of an odd number of helical modes at the non-interacting level, we proceed to examine the effect of four fermion or current-current interaction terms for the $\mathfrak{so}(3)_1\times \overline{\mathfrak{so}(3)}_1$ theory. The action of TRS on the currents is
\begin{align}
\mathcal T: \mathcal J^{\ms{a}} \longleftrightarrow -\bar{\mathcal J}^{\ms{a}}.
\end{align}
Therefore the general form of TRS invariant current-current interaction terms is
\begin{equation}
\delta {H}_{\ms{int}}=\sum_{\ms{a}}\lambda_{\ms{a}}\mathcal J^{\ms{a}}\bar{\mathcal J}^{\ms{a}}+\sum_{\ms{a,b}}\lambda_{\ms{ab}}\left(\mathcal J^{\ms{a}}\bar{\mathcal J}^{\ms{b}}+\bar{\mathcal J}^{\ms{a}}{\mathcal J}^{\ms{b}}\right).
\end{equation}
We examine the $\lambda_{\ms{a}}$ terms first. The term $\mathcal J^{\ms{a}}\bar{\mathcal J}^{\ms{a}}$ can be decomposed into two kinds of bilinears: those of the form $\chi^{j}\bar{\chi}^{j}$, $j\neq \ms{a}$, and those of the form $\mathcal O^{\ms{a}}$. In other words if $\lambda_{\ms{a}}$ were to flow to strong coupling, at least one of these bilinears would be expected to acquire a groundstate expectation value. Since the former kind breaks TRS, this would lead to a groundstate that spontaneously breaks TRS. Alternatively, we could consider the scenario where $\mathcal O^{\ms{a}}$ acquires an expectation value. An important observation is that
\begin{equation}
\left(\mathcal O^j\right)^2\propto \mathcal J^{j}\bar{\mathcal J}^{j},
\end{equation}
up to a constant term. Therefore, by ramping up $\lambda^{1}$ for example, we can gap out the modes $\chi_{2,3}$ and $\bar{\chi}_{2,3}$ by condensing $\mathcal O^{\ms{1}}\propto (\chi_{2}\bar{\chi_{3}}+\bar{\chi}_{2}\chi_{3})$. Crucially though, we cannot gap out the entire theory by simultaneously condensing $\mathcal O^{1,2,3}$, as these operators satisfy the non-trivial algebra in Eq.~\eqref{eq:non_abelian_haldane}. This can be understood as a generalization of the Haldane criterion to non-Abelian current algebras.
Next, we turn to the terms of the form $\lambda_{\ms{ab}}(\mathcal J^{\ms{a}}\bar{\mathcal J}^{\ms{b}}+\bar{\mathcal J}^{\ms{a}}\mathcal J^{\ms{b}})$. We are interested in how the groundstate at $\lambda_{\ms{ab}}\to \infty$ transforms under TRS. To this end, we decouple the interaction term into possible products of fermion bilinears and ask whether we can find a decoupling where each bilinear is invariant under TRS. Let us illustrate this procedure with an explicit example. Consider the term
\begin{align}
\mathcal J^{1}\bar{\mathcal J}^2+\bar{\mathcal J}^1{\mathcal {J}}^2=&\;(i\chi_2\chi_3)(i\bar{\chi}_1\bar{\chi}_3)+(i\bar{\chi}_2\bar{\chi}_3)(i\chi_1\chi_3)\nonumber \\
=&\; (i\chi_{3}\bar{\chi}_{3})(i\bar{\chi}_{1}\chi_{2}-i\chi_{1}\bar{\chi}_2).
\end{align}
This is the only possible decoupling for the term proportional to $\lambda_{12}$; the terms proportional to the other $\lambda_{\ms{ab}}$'s all have similarly unique decouplings. Crucially, both the bilinears in the decoupling transform non-trivially under TRS and such an interaction cannot have a TRS invariant groundstate.
The above considerations generalize to any odd number of helical modes. For an even number of helical modes, say $2n$, we can construct an interaction term that gaps out all the modes while preserving the TRS. Consider the matrices $\widetilde{\mathrm{L}}_{2n}^{\ms{a}}$ with $\ms{a}=1,\dots,n$ which generate a $\mathfrak{so}(2)^{n}$ subgroup $\mathfrak{so}(2n)$. The matrix $\mathrm{L}_{2n}^{\ms{a}}$ basically generates rotations in ${x}_{2\ms{a}-1}\text{-}x_{2\ms{a}}$ plane in $\mathbb R^{2n}$. Then we can construct the currents
\begin{align}
\mathcal J^{\ms{a}}:=\frac{i}{2}\chi_{j}\widetilde{\ms{L}}_{2n,jk}^{\ms{a}}\chi_{k}=i\chi_{2\ms{a}-1}\chi_{2\ms{a}},
\end{align}
and analogously we define the antiholomorphic currents $\bar{\mathcal {J}}^{\ms{a}}$. Then we may write down the interaction term
\begin{align}
\delta H=&\; \lambda\sum_{\ms{a}=1}^{n}\mathcal J^{\ms{a}}\bar{\mathcal J}^{\ms{a}} \nonumber \\
= &\; \lambda\sum_{\ms{a}=1}^{n} (i\chi_{2\ms{a}-1}\bar{\chi}_{2\ms{a}}+i\bar{\chi}_{2\ms{a}-1}\chi_{2\ms{a}})^{2} + \text{const.}.
\end{align}
Since the terms $(i\chi_{2\ms{a}-1}\bar{\chi}_{2\ms{a}}+i\bar{\chi}_{2\ms{a}-1}\chi_{2\ms{a}})$ are TRS invariant and commute mutually for all $\ms{a}$, adding such a term gaps out all $2n$ helical modes simultaneously.
\subsubsection{Gapping out the surface with topological order} In this subsection, we describe the STO for second-order 3D class DIII HOTSC protected by inversion symmetry. Here we start with the system that originally carries three pairs of helical Majorana modes along the hinge. Since the class DIII hinge modes can be regarded as a stack of class D hinge modes with their time-reversed partners, a natural candidate of the STO for class DIII is given by stacking the STO for class D with its time-reversed partner which is $\mathsf{SO}(3)_3\times \overline{\mathsf{SO}(3)_3}$.
We consider the HOTSC to have a spherical geometry with three pairs of helical Majorana modes $\chi_i,\bar{\chi}_i$ propagating along the equator on the surface. First we add three copies of class DIII superconductors on the northern and southern hemisphere, ending up with $6$ additional pairs of helical Majorana modes on the equator which we label as $\gamma^{\mathrm{N/S}}_i,\bar{\gamma}^{\mathrm{N/S}}_i$. The hinge is described by the $\mathfrak{so}(9)_1\times \overline{\mathfrak{so}(9)}_1$ WZW theory. The action of inversion on the different Majorana operators is as follows
\begin{align}
\mathcal I:&\;
\begin{bmatrix}
\chi_i \\
\gamma_{i}^{\mathrm{N}} \\
\gamma_i^{\mathrm{S}} \\
\bar{\chi}_{i} \\
\bar{\gamma}^{\mathrm{N}}_{i} \\
\bar{\gamma}^{\mathrm{S}}_{i}
\end{bmatrix}(\theta)
\longrightarrow
\begin{bmatrix}
i\chi_i \\
i\gamma_{i}^{\mathrm{S}} \\
i\gamma_i^{\mathrm{N}} \\
-i\bar{\chi}_{i} \\
-i\bar{\gamma}^{\mathrm{S}}_{i} \\
-i\bar{\gamma}^{\mathrm{N}}_{i}
\end{bmatrix}(\theta + \pi).
\label{eq:inv_action_conformal_embedding}
\end{align}
Next, we carry out the conformal embedding procedure i.e. we embed $\mathfrak{so}(3)_3\times \mathfrak{so}(3)_3 \subset \mathfrak{so}(9)_1$ and $\overline{\mathfrak{so}}(3)_3\times \overline{\mathfrak{so}}(3)_3 \subset \overline{\mathfrak{so}}(9)_1$. Since the recipe is identical to that described for the holomorphic CFT in Sec.~\ref{Sec:ClassD_HOTSC}, we do not repeat the procedure here. Eventually, we end up with chiral and anti-chiral current operators that transform under the inversion-symmetry action as
\begin{align}
\mathcal I:&\;
\begin{bmatrix}
\mathcal J^{\ms{a},1} \\
\bar{\mathcal J}^{\ms{a},1}
\end{bmatrix}(\theta)
\longmapsto
\begin{bmatrix}
\mathcal J^{\ms{a},2} \\
\bar{\mathcal J}^{\ms{a},2}
\end{bmatrix}(\theta+\pi),
\end{align}
where `$\ms{a}$' labels the generators of $\mathfrak{so}(3)$. Similarly, under TRS, the currents transform as
\begin{align}
\mathcal T: \mathcal J^{\ms{a},\kappa} \longleftrightarrow -\bar{\mathcal J}^{\ms{a},\kappa},
\end{align}
where $\kappa\in \left\{1,2\right\}$. In order to gap out these current operators, we introduce topological order $\mathcal A_{\mathrm{N}}$ and $\mathcal A_{\mathrm{S}}$ on the northern and southern hemispheres, respectively, with $\mathcal A_{\mathrm{N}}=\ms{SO}(3)_3\times \overline{\ms{SO}(3)}_3$. Conveniently, the symmetry transformation properties of the edge modes induced on the hinge from the topological order are identical to those of the aforementioned modes obtained from the conformal embedding procedure. We denote the modes provided by the topological orders on the northern and southern hemispheres with subscripts $\mathrm{N}$ and $\mathrm{S}$. Under inversion and TRS,
\begin{align}
\mathcal I:&\;\mathcal J^{\ms{a}}_{\mathrm{N,S}}(\theta)\longmapsto \mathcal J^{\ms{a}}_{\mathrm{S,N}}(\theta+\pi), \nonumber \\
\mathcal I:&\; \bar{\mathcal J}^{\ms{a}}_{\mathrm{N,S}}(\theta)\longmapsto \bar{\mathcal J}^{\ms{a}}_{\mathrm{S,N}}(\theta+\pi), \nonumber \\
\mathcal T:&\; \mathcal J^{\ms{a}}_{\mathrm{N,S}}(\theta)\longleftrightarrow -\bar{\mathcal J}^{\ms{a}}_{\mathrm{N,S}}(\theta).
\end{align}
The hinge modes $\mathcal{J}^{\ms{a},\kappa}, \bar{\mathcal{J}}^{a,\kappa}$ and the edge modes of the surface topological order $\mathcal{J}^{\mathsf{a}}_{\mathrm{N/S}}, \bar{\mathcal{J}}^{\ms{a}}_{\mathrm{N/S}}$ can together be gapped out upon adding the gapping term
\begin{align}
\delta \mathcal{L}&=\lambda(\theta)\sum_{\ms{a}=1}^{3}\Big[\bar{\mathcal J}^{\ms{a}}_\mathrm{N}(\theta)\mathcal J^{\ms{a},1}(\theta)+\bar{\mathcal J}^{\ms{a}}_\mathrm{S}(\theta)\mathcal J^{\ms{a},2}(\theta)\nonumber \\
&+\mathcal J^{\ms{a}}_\mathrm{N}(\theta)\bar{\mathcal J}^{\ms{a},1}(\theta)+\mathcal J^{\ms{a}}_\mathrm{S}(\theta)\bar{\mathcal J}^{\ms{a},2}(\theta)\Big],
\end{align}
which is TRS invariant, and inversion symmetric if $\lambda(\theta)=\lambda(\theta+\pi)$. Therefore, we can choose $\lambda$ to be constant.
\section{Surface Topological Order for Third Order Topological Phases}\label{Sec:3rdOrder}
In this subsection, we discuss the surface topological order for third order topological phases protected by inversion symmetry in addition to possible Altland-Zirnbauer symmetries. In total there are five Altland-Zirnbauer classes that support non-trivial third-order topology upon imposing inversion symmetry. These are D, BDI, AIII, DIII and CII. In what follows we present the STO for classes D, BDI and AIII together, as these classes and consequently their STOs are closely related.
\begin{figure}[tbh!]
\includegraphics[width=0.3\textwidth]{drawing_4.png}
\caption{Surface topological order for third order topological phases protected by inversion symmetry. In the picture, the red stars denote the zero modes located at antipodal points of the surface.}
\label{STO_3rdOrder_HOTSC}
\end{figure}
\subsection{Class D, BDI and AIII}
We begin with the discussion of class D. Note that the third-order inversion-symmetric class D superconductor can be obtained by stacking two copies of second-order inversion-symmetric class D superconductors whose surface contains the configuration $(1,0)\oplus (1,0)=(2,0)\sim (1,-1)$ in the notation used in Sec.~\ref{Sec:ClassD_HOTSC}. The configuration $(1,-1)$ contains a pair of counter-propagating chiral modes $\chi^{+}$ and $\bar{\chi}^{-}$ which are unstable to a mass term $im(\theta)(\chi^{+}(\theta))^{(\dag)}\bar{\chi}^{-}(\theta)$. The inversion symmetry imposes that $m(\theta+\pi)=-m(\theta)$ and consequently, the mass vanishes at two anti-podal points which contain Majorana zero-modes. Furthermore since we can always gap out Majorana modes in pairs, $(2,0)\oplus (2,0)=(4,0)\sim (0,0)$. This agrees with the result \citep{luka,khalaf_inversion} that 3D class D higher-order topological superconductors enriched by inversion symmetry are classified by $\mathbb Z_{4}$ which is an extension of $\mathbb Z_{2}$ (second-order phases) by $\mathbb Z_{2}$ (third-order phases).\par
Here we describe the procedure to gap out two surface Majorana zero modes (or equivalently the (2,0) configuration) by pasting inversion symmetric surface topological order. First we start with the configuration $(-6,-6)$ which is obtained by pasting 6 inversion symmetric copies of $p\pm ip$ on the surface. Next we paste $\mathfrak{so}(3)_3\times \mathfrak{so}(3)_3$ on the northern and southern hemispheres which effectively provides additional modes corresponding to $(12,6)$. Upon such a surface pasting, we end up with $(-6,-6)\oplus (12,6) = (6,0)\sim (2,0)$. Since we can create two Majorana zero modes on antipodal points on the surface without manipulating the bulk, we can always absorb the surface modes contributed by a third order class D superconductor.\par
Having obtained the STO for class D, we now proceed to discuss the surface topological order for 3D inversion TSC in BDI class. According to Ref.~\onlinecite{khalaf_inversion}, 3D inversion-symmetric TSC in class BDI has a third order phase but no non-trivial first or second order phases. Conveniently, we make use of the notion of \emph{block state} introduced in Ref.~\onlinecite{Hermele_1, Hermele_2}. Generically, a block state $\state{\Psi}$ has the form $\state{\Psi}=\bigotimes_{b\in B}\state{\psi_{b}}$, where $B$ is a collection of blocks, and block $b$ is a $d_b$-dimensional system embedded in a $d$-dimensional space. In our case, the wavefunction of a non-trivial third order phase is physically equivalent
(up to a inversion-symmetric finite depth unitary circuit) to a block state $\state{\Psi}=\state{\Psi_1}\otimes \state{\Psi_a}$ where $\state{\Psi_1}$ denotes a state in the non-trivial phase of the 1D BDI Majorana chain embedded in the 3D space with inversion symmetry, and $\state{\Psi_a}$ denotes a state, describing the rest of the 3D space, in the trivial phase. The 1D BDI Majorana chain can be described by the following Hamiltonian
\begin{equation}
\hat{H}=\frac{1}{2}\int dx \hat{\psi}^\dag(x) H(x)\hat{\psi}(x),
\end{equation}
where
\begin{equation}
H(x)=-i\tau_y\partial_x+m(x)\tau_z, \quad \hat{\psi}^\dag=(a^\dag_x,a_x),
\end{equation}
with $a^\dag_{x}$ being a complex fermion creation operator. The Altland-Zirnbauer symmetries are represented as
\begin{equation}
\mathcal{T}=\mathcal{K},\quad \mathcal{P}=\tau_x\mathcal{K},\quad \mathcal{I}=\tau_z.
\end{equation}
The mass profile $m(x)$ is positive (negative) inside (outside) the chain. If we denote the two edges of this chain as $x^+$ and $x^-$ separately, then Jackiw-Rebbi procedure shows that these two edges host Majorana zero modes $\gamma_{x^+}$ and $\gamma_{x^-}$, which have the form \cite{ruben_2017}
\begin{align}
\gamma_{x^+}&=\int dx \frac{1}{\sqrt{2}}(a_x+a^\dag_x)e^{\int^x_{x_a}dx'm(x')},\nonumber \\
\gamma_{x^-}&=\int dx \frac{i}{\sqrt{2}}(-a_x+a^\dag_x)e^{-\int^x_{x_b}dx'm(x')},
\end{align}
where $x_{a/b}$ are parameters that ensure the normalization condition $\int dx e^{2\int^x_{x_a}dx'm(x')}=\int dx e^{-2\int^x_{x_b}dx'm(x')}=1$.
The actions of TRS and inversion are as follows
\begin{equation}
\mathcal{T}\gamma_{x^+}\mathcal{T}^{-1}=\gamma_{x^+}, \quad \mathcal{T}\gamma_{x^-}\mathcal{T}^{-1}=-\gamma_{x^-}.\label{eq:trsbdi}
\end{equation}
\begin{equation}
\mathcal{I}\gamma_{x^+}\mathcal{I}^{-1}=i\gamma_{x^-},\quad \mathcal{I}\gamma_{x^-}\mathcal{I}^{-1}=-i\gamma_{x^+}.\label{eq:invbdi}
\end{equation}
The above symmetry actions satisfy the basis invariant relations $\{\mathcal{I}, \mathcal{P}\}=0$, $\left[\mathcal I,\mathcal T\right]=0$, $\mathcal T^{2}=1$ and $\mathcal I^{2}=1$, independently agreeing with Ref.~\onlinecite{khalaf_inversion}. We thus propose that the STO for 3D third order inversion TSC in BDI class to be $\mathsf{SO}(3)_3\times\overline{\mathsf{SO}(3)_3}$, which is the same as we found for the 3D third order inversion TSC in D class. Recall that in Section \ref{Sec:ClassD_HOTSC}, we showed that an inversion symmetric surface realization of $\mathsf{SO}(3)_3$ contains a gapless modes (denoted as $\mathcal{J}^{\ms{a}}_{\mathrm{N/S}}$ currents) appearing on an inversion-symmetric line (chosen as the equator for our convenience). We then used such a current mode to gap out the 9 chiral Majorana modes on the equator contributed by the bulk as well as the surface pasting of $p\pm ip$ superconductors.
Similarly, $\mathsf{SO}(3)_3\times\overline{\mathsf{SO}(3)_3}$, which is the proposed STO for class BDI, has 9 pairs of counter-propagating Majorana modes, without the contribution from the bulk. Let us denote these modes as $\alpha_i,\bar{\alpha}_i$ with $i=1,\dots,9$. The TRS action on the Majorana modes is
\begin{equation}
\mathcal{T}: i\mapsto -i, \begin{bmatrix}
\alpha_i\\
\bar{\alpha}_i
\end{bmatrix}\mapsto \begin{bmatrix}
\bar{\alpha}_i\\
\alpha_i
\end{bmatrix}.
\end{equation}
While the inversion acts naturally on the current operators contributed by the STO on the equatorial hinge as
\begin{align}
\mathcal{I}:
\begin{bmatrix}
\mathcal{J}^{\ms{a}}_\mathrm{N} \\
\bar{\mathcal{J}}^{\ms{a}}_\mathrm{N}
\end{bmatrix} (\theta)
\longleftrightarrow
\begin{bmatrix}
\mathcal{J}^{\ms{a}}_\mathrm{S} \\
\bar{\mathcal{J}}^{\ms{a}}_\mathrm{S}
\end{bmatrix}(\theta+\pi).
\end{align}
After the conformal embedding, we regroup the Majoranas as $\chi_i,\gamma^{(1)}_i,\gamma^{(2)}_i,\bar{\chi}_i,\bar{\gamma}^{(1)}_i,\bar{\gamma}^{(2)}_i$, where the definition is taken as
\begin{align}
\chi_{1,2,3}\equiv&
\alpha_{1,2,3}, \nonumber \\
\gamma^{(1)}_{1,2,3}\equiv&\;
\alpha_{4,5,6}, \nonumber \\
\gamma^{(2)}_{1,2,3}\equiv&\; \alpha_{7,8,9},
\end{align}
and similarly for the $\bar{\chi}_i,\bar{\gamma}^{(1)}_i,\bar{\gamma}^{(2)}_i$. Inversion action on the Majoranas is
\begin{align}
\mathcal I:&\;
\begin{bmatrix}
\chi_i \\
\gamma_{i}^{(1)} \\
\gamma_i^{(2)} \\
\bar{\chi}_{i} \\
\bar{\gamma}^{(1)}_{i} \\
\bar{\gamma}^{(2)}_{i}
\end{bmatrix}(\theta)
\longrightarrow
\begin{bmatrix}
i\chi_i \\
i\gamma_{i}^{(2)} \\
i\gamma_i^{(1)} \\
-i\bar{\chi}_{i} \\
-i\bar{\gamma}^{(2)}_{i} \\
-i\bar{\gamma}^{(1)}_{i}
\end{bmatrix}(\theta + \pi).
\end{align}
We now proceed to gap these modes in groups. For the $\gamma$'s and $\bar{\gamma}$'s, we can write the following gapping term
\begin{equation}
\delta \hat{H}_1=\sum_i m_1(\theta)(i(\gamma_i^{(1)})^{(\dag)}\bar{\gamma}_i^{(1)}-i(\gamma_i^{(2)})^{(\dag)}\bar{\gamma}_i^{(2)}),
\end{equation}
which gaps all the $\gamma$ and $\bar{\gamma}$. For the $\chi$'s, $\bar{\chi}$'s, we can write the following gapping term:
\begin{equation}
\delta \hat{H}_2=\sum_im_{2,i}(\theta)(i\chi^{(\dag)}_i\bar{\chi}_i).
\end{equation}
Note that inversion forces $m_{2,i}(\theta)=-m_{2,i}(\theta+\pi)$, and each counter propagating $\chi,\bar{\chi}$ contributes a pair of Majorana zero modes. Since we have three pairs of counter-propagating $\chi,\bar{\chi}$, two pairs of Majorana zero modes will be gapped out, and we are therefore left with one protected pair of Majorana zero modes on the equator located at inversion-symmetric positions. Furthermore, the Jackiw-Rebbi procedure shows that these Majorana zero modes transform in the exact same way as in Eq.~\eqref{eq:trsbdi} and Eq.~\eqref{eq:invbdi}. Thus the zero modes from the STO can gap out the zero modes from the BDI bulk.
What about the third order inversion symmetric topological phase in AIII? According to Ref.~\onlinecite{khalaf_inversion}, a 3D inversion symmetric TI in class AIII has only a third-order implementation. Physically, this phase can be thought of as a 1D AIII Su-Schrieffer-Heeger (SSH) chain which is in non-trivial phase inserted into a 3D manifold. There is a close connection between the 1D SSH chain and the 1D BDI chain, as pointed out in Ref.~\onlinecite{ruben_2017}. we can establish an exact mapping from two copies of the BDI Kitaev chain to one copy of SSH chain. The two dangling Majorana zero modes form a Dirac zero mode, which is the dangling zero mode of the SSH chain. Since two copies of Kitaev chain have an emergent $\ms{O}(2)$ symmetry, its subgroup $\ms{SO}(2)$ corresponds to the $\ms{U}(1)$ symmetry for the AIII chain. Crucially, the TRS in the BDI chain corresponds to the sub-lattice/chiral symmetry of the AIII chain (We comment that the TRS is still anti-unitary in the Fock space after the mapping, but it is unitary on the single-particle Hamiltonian \cite{ruben_2017}). Therefore, we naturally conclude that the STO for the 3D third order AIII phase is equivalent to the STO for two copies of 3D third order BDI phase, with chiral symmetry implemented in the same way as the TRS in the STO for BDI phase.
\subsection{Class DIII}
In this subsection, we discuss the STO for third order class DIII HOTSC. To that end, we first demonstrate the fact that, for 3D class DIII HOTSC protected by inversion, the classification is $\mathbb{Z}_4$, which is an extension of $\mathbb{Z}_2$ (second-order phases) by $\mathbb{Z}_2$ (third-order phases). Consider the second-order case, where there is a pair of helical hinge modes. We denote the Majorana hinge modes as $(\chi,\bar{\chi})$. The TRS action on these modes is
\begin{align}
\mathcal T:
\begin{bmatrix}
\chi(\theta)\\
\bar{\chi}(\theta)
\end{bmatrix}\mapsto
\begin{bmatrix}
\bar{\chi}(\theta)\\
-\chi(\theta)
\end{bmatrix},
\label{eq:TRS_Majorana}
\end{align}
while the inversion action on these modes is
\begin{align}
\mathcal I:
\begin{bmatrix}
\chi(\theta)\\
\bar{\chi}(\theta)\\
\chi(\theta+\pi)\\
\bar{\chi}(\theta+\pi)
\end{bmatrix}\mapsto
\begin{bmatrix}
i\chi(\theta+\pi)\\
-i\bar{\chi}(\theta+\pi)\\
-i\chi(\theta)\\
i\bar{\chi}(\theta)
\end{bmatrix}.
\label{eq:inv_repn_maj}
\end{align}
As in the case of Class D, the above action can be derived using recursive Jackiw-Rebbi procedures. Now suppose we have two copies of such helical modes $(\chi_1,\bar{\chi}_1,\chi_2,\bar{\chi}_2)$, with symmetry action exactly the same as the above. A gapping term can be written down in the 1D model
\begin{equation}
\delta \hat{h}=\int \mathrm{d}\theta \left[im(\theta)\left(\chi^{(\dag)}_1(\theta)\bar{\chi}_2(\theta)+\bar{\chi}^{(\dag)}_1(\theta)\chi_2(\theta)\right)\right].
\end{equation}
The above term is TRS invariant, and inversion symmetry imposes $m(\theta)=-m(\theta+\pi)$. Thus the 1D system is gapped out except at two inversion symmetric points. Furthermore, these two point modes can be gapped out if we take a double stacking of this model. We therefore conclude that the class DIII HOTSC protected by inversion has classification $\mathbb{Z}_4$, which is an extension of $\mathbb{Z}_2$ (second-order phases) by $\mathbb{Z}_2$ (third-order phases).\par
To gap out the third order topology, we make use of the fact that the third order phase is obtained by two copies of second order phase. Now since the STO for the second order phase is $\mathsf{SO}(3)_3\times\overline{\mathsf{SO}(3)_3}$, we conclude that the STO for third order topological phase is two copies of $\mathsf{SO}(3)_3\times\overline{\mathsf{SO}(3)_3}$ topological order.
\subsection{Class CII}
We now briefly discuss the STO for the third order 3D HOTSC in class CII. Similarly to the case AIII, we can view the 3D inversion-symmetric third order CII phase as a 1D inversion-symmetric CII chain embedded in a 3D manifold. The 1D CII chain always has even number of Majorana zero modes at its edge, instead of single Majorana zero mode at the edge of Kitaev chain. we can view the edge zero modes for CII chain as a Kramers' pair of Majorana zero modes, similar to the edge mode in the case of 1D DIII chain. However 1D CII chain has a $2\mathbb{Z}$ classification whereas the 1D DIII chain has a $\mathbb{Z}_2$ classification \cite{RMP_ChingKai}. Here we briefly look at the following linearised 1D model for CII chain taken from Ref.~\onlinecite{cii_wang_zhao}
\begin{equation}
H(k)=-k\tau_z\sigma_y+m\tau_x, \quad \mathcal{T}=\sigma_y \mathcal{K}, \quad \mathcal{P}=\tau_y \mathcal{K}.
\end{equation}
The chain is in the topological phase when $m>0$, and the zero modes are trapped at the $m=0$ domain wall. We perform the Jackiw-Rebbi projection to track the symmetry action on the zero modes, and found that $\mathcal{T}_{\mathrm{edge}}\sim \mathcal{P}_{\mathrm{edge}}\sim i\sigma_y \mathcal{K}$. Because of this, stacking any number of copies of CII chains cannot enable us to gap these zero modes out, as TRS must commute with the mass term, whereas the PHS must anti-commute with the mass term. The fact that CII can be viewed as a stacking of a Kitaev chain and its TRS copy with $\mathcal{P}^2=-1$ implementation of PHS leads us to conjecture that the STO for the third order 3D inversion CII phase to be the same as the STO for the third order 3D DIII phase, with the $\mathcal{P}^2=-1$ implementation of PHS on the STO level.
\section{Summary and outlook}
In this work, we have established via an explicit construction that all inversion-symmetric higher-order topological insulators and superconductors (except for AZ class C) admit gapped surfaces with anomalous topological order.
While we have done so for the case of inversion symmetric electronic phases, we expect it to hold more generally for both Bosonic and electronic three-dimensional higher-order phases with various spatial symmetries.
This consequently extends the list of symmetric surface terminations of 3D second and third-order topological phases to include `anomalous gapped surfaces'.
A technical consequence of this work is the study of spatial symmetries and their anomalies in 2D topological orders.
In this regard, there is a recent systematic algebraic framework to study the spatial symmetry enrichment of modular tensor categories \cite{maissam_new_4}.
In particular, obstructions to such symmetry enrichment must precisely encode anomalies that in turn can be compensated by crystalline topological phases in one higher dimension.
It would be interesting to pursue this direction in future work.
\section*{Acknowledgements}
We would like to thank Eslam Khalaf for useful comments.
We acknowledge support from the European Research Council under the European Union Horizon 2020 Research and Innovation Programme, through Starting Grant [Agreement No. 804213-TMCS, ML and SP] and [ERC-StG-Neupert757867-PARATOP, TN], the Marie Sklodowska
Curie Action [Grant Agreement No 701647, AT], the Swedish Research Council (VR) through grants number 2019-04736 and 2020-00214
and from a Buckee Scholarship at Merton College [ML].
Finally, we acknowledge Alexander, Serena, and Amaia, without whose arrival into our lives this paper would have been finished in half the time but with a tenth the joy.
| 3eb98df565505de16221850411cda31fbf72eda3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Methods}
\subsection*{Electron Microscopy}
Simultaneously acquired EELS and HAADF datasets were collected on a 5-th order aberration-correction Nion UltraSTEM microscope operated at 100 keV with a probe semi-angle of roughly 30 mrad and collection semi-angle of 80-240 mrad and 0-60 mrad for HAADF and EELS, respectively. Both specimens were imaged at 30 pA, for a dwell time of 10 ms (Fig. \ref{fig::ZnSCu}) and 15 ms (Fig. \ref{fig::eels_overview}) receiving a total dose of 3.25 $\times 10^4$ and 7.39 $\times 10^4$ e/Å$^2$. The EELS signals were obtained by integration over the core loss edges, all of which were done after background subtraction. The background EELS spectra were modeled using a linear combination of power laws implemented using the open-source Cornell Spectrum Imager software \cite{cueva2012csi}.
Simultaneously acquired EDX and HAADF datasets were collected on a Thermo Fisher Scientific Titan Themis G2 at 200 keV with a probe semi-angle of roughly 25 mrad, HAADF collection semi-angle of 73-200 mrad, and 0.7 sr EDX solid angle. The CoS specimen was imaged at 100 pA and 40 $\mu$s dwell time for 50 frames receiving a total dose of approximately $2 \times 10^{5}$ e/Å$^2$. The initial chemical distributions were generated from EDX maps using commercial Velox softwarethat produced inital net count estimates (however atomic percent estimates are also suitable).
\subsection*{Fused Multi-Modal Recovery}
Here, fused multi-modal electron microscopy is framed as an inverse problem expressed in the following form:
$\hat{\bm{x}} = \argminop_{\bm{x} \geq 0} \Psi_1(\bm{x}) + \lambda_1 \Psi_2(\bm{x}) + \lambda_2 \mathrm{TV}(\bm{x})$
where $\hat{\bm{x}}$ is the final reconstruction, and the three terms are described in the main manuscript (Eq.~\ref{eq:costFunc}). When implementing an algorithm to solve this problem, we concatenate the multi-element spectral variables ($\bm{x}_i, \bm{b}_i)$ as a single vector:
$\bm{x},~\bm{b}~\in~\mathbb{R}^{n_x n_y n_{i}}$
where $n_{i}$ denotes the total number of reconstructed elements.
The optimization problem is solved by a combination of gradient descent with total variation regularization. We solve this cost function by descending along the negative gradient directions for the first two terms and subsequently evaluate the isotropic TV proximal operator to denoise the chemical maps~\cite{beck2009tv}. The gradients of the first two terms are:
\begin{align}
\nabla_{\bm{x}} \Psi_1(\bm{x}) &= - \gamma \text{diag} \big(\bm{x}^{\gamma-1}\big) \bm{A}^{T} \big(\bm{b}_{H} - \bm{A} \bm{x}^{\gamma} \big) \\
\nabla_{\bm{x}} \Psi_2(\bm{x}) &= \textbf{1} - \bm{b} \oslash (\bm{x} + \varepsilon),
\end{align}
where $\oslash$ denotes point-wise division.
Here, the first term in the cost function,
relating the elastic and inelastic modalities,
has been equivalently re-written as
$\Psi_1 = \frac{1}{2} \big\| \bm{b}_{H} - \bm{A} \bm{x}^{\gamma} \big \|_2^2$,
where $\bm{A}~\in~\mathbb{R}^{n_x n_y \times n_x n_y n_i}$ expresses the summation of all elements as matrix-vector multiplication. Evaluation for the TV proximal operator is in itself another iterative algorithm. In addition, we impose a non-negativity constraint since negative concentrations are unrealistic. We initialize the first iterate with the measured data ($\bm{x}^0_i = \bm{b}_i$), an ideal starting point as it is a local minima for $\Psi_2$.
The inverse of the Lipschitz constant (1/L) is an upper bound of the step-size that can theoretically guarantee convergence. From Lipschitz continuity, we estimated the step size for the model term's gradient ($\nabla \Psi_1$) as: 1/$L_{\nabla\Psi_1} \leq 1/\big(\|\bm{A}\|_1 \|\bm{A}\|_{\infty}\big) = 1/n_i$. The gradient of the Poisson negative log-likelihood ($\Psi_2$) is not Lipschitz continuous, so its descent parameter cannot be pre-computed \cite{dupe2009poissonLips}. We heuristically determined the regularization parameters starting with values with a similar order of magnitude to $1/L_{\nabla \Psi_1}$, then iteratively reduce until the cost function exhibits stable convergence. The regularization parameters were manually selected, however future work may allow automated optimization by the L-curve method or cross-validation \cite{regParamSelect}.
\subsection*{Estimating Standard Error of Recovered Chemical Maps}
Using estimation theory,
we can approximate the uncertainty in a recovered chemical image for unbiased estimators with the model's (Eq.~\ref{eq:costFunc}) Hessian expressed as: $\bm{H}(\bm{x}) = \nabla_{\bm{x}}^2 \Psi_1(\bm{x}) + \nabla_{\bm{x}}^2 \Psi_2(\bm{x}) $, where \begin{align}
\nabla_{\bm{x}}^2 \Psi_1(\bm{x}) &= \text{diag} \Big( \gamma(\gamma - 1) \text{diag} \big(\bm{x}^{\gamma-2}\big) \bm{A}^{T} \big(\bm{b}_{H} - \bm{A} \bm{x}^{\gamma} \big) \Big) \\ & \nonumber \quad + \gamma^2 \text{diag} \big( \bm{x}^{\gamma-2} \big) \bm{A}^T \bm{A} \text{diag}(\bm{x}^{\gamma-1}) \\
\nabla_{\bm{x}}^2 \Psi_2(\bm{x}) &= \text{diag} \big( \bm{b} \oslash (\bm{x} + \varepsilon)^2 \big)
\end{align}
Calculation of standard error follows the Cramer-Rao inequality, which provides a lower bound given by:
$\big(\mathrm{\textbf{Var}}(\hat{x}_j) \geq \big[ \bm{H}^{-1}(\hat{\bm{x}}) \big]_{jj} \big)$ \cite{wei2020crlbPtycho}, where $\mathrm{\textbf{Var}}(\hat{\bm{x}})$ are variance maps for the recovered chemical distributions ($\hat{\bm{x}}$) and subscript $jj$ denotes indices along the diagonal elements. We determined this lower bound from an empirical derivation of the Fisher Information Matrix. From the variance we thus extract standard error maps: $\mathrm{\textbf{Standard~Error}} ~= \sqrt{\mathrm{\textbf{Var}}(\hat{\bm{x}})}$ as demonstrated in Supplemental Figure 10. The average standard error denotes the mean value of all pixels in $\mathrm{\textbf{Standard~Error}}$. Note, the TV regularizer reduces noise and may introduce bias due to smoothing, so the standard error measurements could potentially be lower; our Fisher information derivation provides an upper bound on uncertainty.
\subsection*{Inelastic Scattering Simulations for Atomic Imaging}
The inelastic scattering simulations for the FePt nanoparticle structure (Fig.~\ref{fig::FePtSim}) were performed using the abTEM simulation code \cite{madsen2021abTEM}, using the algorithm described in \cite{brown2019prismEELS}. In this algorithm the initial STEM probe is propagated and transmitted to some depth into the specimen using the scattering matrix method described in the PRISM algorithm \cite{ophus2017PRISM}. Next, the inelastic transition potentials of interest (in this case the L$_{2,3}$ Fe and M$_{4,5}$ Pt edges) were calculated and applied using the methods given in \cite{saldin1987theoryEELS, dwyer2008elasticScatter}, using the GPAW density functional theory code \cite{enkovaara2010gpaw}. Finally, a second scattering matrix is used to propagate the inelastically scattered electrons through the sample and to the plane of the EELS entrance aperture. The elastic signal channels were calculated with the conventional PRISM method using the same parameters.
The atomic structure used in the simulations was a portion of the FePt nanoparticle structure determined from atomic electron tomography \cite{yang2017feptExp}. After cropping out 1/4 of nanoparticle coordinates, the boundaries were padded by 5 Å total vacuum. The STEM probe's convergence semiangle was set to 20 mrad and the voltage to 200 kV. The multislice steps used a slice thicknesses of 2 Å, the wavefunction sampling size was 0.15 Å, and the projected potentials were computed using the infinite Kirkland parameterization \cite{kirkland2020compuTEM}. The EELS detector had a semiangle of 30 mrad, and the STEM probe positions were Nyquist sampled at a step size of 0.31 Å. After completion, we convolved the simulated images with a 0.2 Å Gaussian to account for source size. These simulation parameters required approximately 4 days of calculation time using the CPU mode of abTEM on a workstation with a 40 core Xeon processor clocked at 2.0 GHz.
\section*{Data Availability}
The datasets and codes that support the finding of this study are available from the corresponding author upon reasonable request.
\section*{Introduction}
Modern scanning transmission electron microscopes (STEM) can focus sub-angstrom electron beams on and between atoms to quantify structure and chemistry in real space from elastic and inelastic scattering processes. The chemical composition of specimens is revealed by spectroscopic techniques produced from inelastic interactions in the form of energy dispersive X-rays (EDX)~\cite{alfonso2010atomicEDX,kothleitner2014quantEDX} or electron energy loss (EELS)~\cite{spence1982eels, muller2008atomicEELS}. Unfortunately, high-resolution chemical imaging requires high doses (e.g., >10$^6$ e/Å$^2$) that often exceed the specimen limits---resulting in chemical maps that are noisy or missing entirely~\cite{hart2017direct,cueva2012csi}. Substantial effort and cost to improve detector hardware has brought the field closer to the measurement limits set by inelastic processes~\cite{mcmullan2014deDetectors, kotula2012chemiSTEM}. Direct interpretation of atomic structure at higher-SNR is provided by elastically scattered electrons collected in a high-angle annular dark field detector (HAADF); however, this signal under-describes the chemistry~\cite{lebeau2008quantitative}. Reaching the lowest doses at the highest SNR ultimately requires fusing both elastic and inelastic scattering modalities.
Currently, detector signals---such as HAADF and EDX/EELS---are analyzed separately for insight into structural, chemical, or electronic properties~\cite{su2010correlativeImg}. Correlative imaging disregards shared information between structure and chemistry and misses opportunities to recover useful information. Data fusion, popularized in satellite imaging, goes further than correlation by linking the separate signals to reconstruct new information and improve measurement accuracy~\cite{hall1997introduction,lahat2015datafusionreview,wendy2016jointXray}. Successful data fusion designs an analytical model that faithfully represents the relationship between modalities, and yields a meaningful combination without imposing any artificial connections~\cite{calhoun2016mriFusion}.
\begin{figure*}[ht!]
\centering
\includegraphics[width=\linewidth]{Figs/Fig1_EDX_Intro.png}
\caption{\textbf{Nanoscale multi-modal chemical recovery of CoS catalysts using EDX + HAADF.} \textbf{a)} Schematic highlighting the linked HAADF and EDX modalities collected in the microscope for every probe position. The algorithm links and correlates information between the two signals through an optimization process that produces chemical maps with higher SNRs. \textbf{b)} The raw EDX chemical maps for the Co, S, and O elemental distributions. \textbf{c)} The simultaneous HAADF micrograph of the CoS nanoparticle. \textbf{d)} The multi-modal reconstructions for the elemental distributions. \textbf{e)} EDX RGB overlay of the Co, S, and O maps. Scale bar, 30 nm.}
\label{fig::edx_overview}
\end{figure*}
Here we introduce fused multi-modal electron microscopy, a technique offering high SNR recovery of nanomaterial chemistry by linking correlated information encoded within both HAADF and EDX / EELS. We recover chemical maps by reformulating the inverse problem as a nonlinear optimization which seeks solutions that accurately match the actual chemical distribution in a material. Our approach substantially improves SNRs for chemical maps, often around 300-500$\%$, and can reduce doses over one order of magnitude while remaining consistent with original measurements. We demonstrate on EDX/EELS datasets at sub-nanometer and atomic resolution. Moreover, fused multi-modal electron microscopy recovers a specimen’s relative concentration, allowing researchers to measure local stoichiometry with less-than $15\%$ error without any knowledge of the inelastic cross sections. Convergence and uncertainty estimates are identified along with simulations that provide ground-truth assessment of when and how this approach can fail.
\section*{Results}
\subsection*{Principles of Multi-Modal Electron Microscopy}
Fused multi-modal electron microscopy recovers chemical maps by solving an optimization problem seeking a solution that strongly correlates with (1) the HAADF modality containing high SNR, (2) the chemically sensitive spectroscopic modality (EELS and / or EDX), and (3) encourages sparsity in the gradient domain producing solutions with reduced spatial variation. The overall optimization function results as following:
\begin{align}
\label{eq:costFunc}
\argminop_{\bm{x}_i \geq 0} \quad &\frac{1}{2} \Big\| \bm{b}_{H} - \sum_{i} (Z_i\bm{x}_{i})^{\gamma} \Big \|_2^2 +
\nonumber\\
\lambda_1 \sum_{i} & \Big(\bm{1}^T \bm{x}_i - \bm{b}_{i}^T \log(\bm{x}_i + \varepsilon) \Big) + \lambda_2 \sum_{i} \|\bm{x}_i\|_{\mathrm{TV}},
\end{align}
where $\lambda$ are regularization parameters, $\bm{b}_H$ is the measured HAADF,
$\bm{b}_i$ and $\bm{x}_i$ are the measured and reconstructed chemical maps for element $i$,
$\varepsilon$ herein prevents log(0) issues but can also account for background,
the $\log$ is applied element-wise to its arguments,
superscript $T$ denotes vector transpose,
and
$\bm{1}$ denotes the vector of $n_x n_y$ ones,
where $n_x \times n_y$ is the image size.
The three terms in (\ref{eq:costFunc}) define our multi-modal approach to surpass traditional dose limits for chemical imaging. First, we assume a forward model where the simultaneous HAADF is a linear combination of elemental distributions ($\bm{x}_i^\gamma$ where $\gamma \in$ [1.4, 2]). The incoherent linear imaging approximation for elastic scattering scales with atomic number as $Z_i^\gamma$
where $\gamma$ is typically around 1.7~\cite{hartel1996conditions,krivanek2010atom,hovden2012efficient}.
This $\gamma$ is bounded between 2 for Rutherford scattering from bare nuclear potentials to 4/3
as described by Lenz-Wentzel expressions for electrons experiencing a screened coulombic potential~\cite{crewe1970coloumb, langmore1973eScatter}.
Second, we ensure the recovered signals maintain a high-degree of data fidelity with the initial measurements by using maximum negative log-likelihood for spectroscopic measurements dominated by low-count Poisson statistics~\cite{wendy2017jointXray,odstrcil2018iterativeLS}. In a higher count regime, this term can be substituted with a simple least-squares error. Lastly, we utilize channel-wise total variation (TV) regularization to enforce a sparse gradient magnitude, which reduces noise by promoting image smoothness while preserving sharp features~\cite{osher1992tv}. This sparsity constraint, popularized by the field of compressed sensing (CS), is a powerful yet minimal prior toward recovering structured data~\cite{donoho2006CS,candes2006CS}. When implementing, each of these three terms can and should be weighted by an appropriately selected coefficients that balances their contributions. All three terms are necessary for accurate recovery (Supplementary Figure 1).
\subsection*{High-SNR Recovery of Nanomaterial Chemistry}
Figure~\ref{fig::edx_overview} demonstrates high-SNR recovery for EDX signals of commercial cobalt sulfide (CoS) nano-catalysts for oxygen-reduction applications---a unique class with the highest activity among non-precious metals~\cite{steve2020CoSx}.
Figure \ref{fig::edx_overview}a illustrates the model that links the two modalities (EDX and HAADF) simultaneously collected in the electron microscope.
The low detection rate for characteristic X-rays is due to minimal emission
(e.g., over 50\% for $Z>32$ and below 2\% for $Z<11$) and collection yield ($<9\%$)~\cite{scholssmacher2010edxYield}.
For high-resolution EDX, the low count rate yields a sparse chemical image dominated by shot noise (Fig.~\ref{fig::edx_overview}b). However, noise in the fused multi-modal chemical map is virtually eliminated (Fig.~\ref{fig::edx_overview}d) and recovers chemical structure without a loss of resolution---including the nanoparticle core and oxide shell interface. The chemical maps produced by fused multi-modal EM quantitatively agree with the expected stoichiometry---the specimen core contains a relative concentration of 39$\pm$1.6$\%$, 42$\pm$2.5$\%$ and 13$\pm$2.4$\%$ and exterior shell composition of 26$\pm$2.8$\%$, 11$\pm$2.0$\%$, 54$\pm$1.3$\%$ for Co, S, O respectively.
The dose for this dataset was approximately $\sim$10$^5$~e Å$^{-2}$ and a 0.7 sr EDX detector was used; however, these quantitative estimates remained consistent when the dose was reduced to $\sim$10$^4$~e Å$^{-2}$.
\begin{figure*}[ht!]
\centering
\includegraphics[width=\linewidth]{Figs/Fig2_EELS_Intro.png}
\caption{\textbf{Atomic-scale multi-modal chemical recovery of Co$_{3-x}$Mn$_x$O$_4$ supercapacitors using EELS + HAADF.} \textbf{a)} Schematic highlighting the linked HAADF and EELS modalities collected in the microscope at every probe position. \textbf{b)} Raw EELS maps for the elemental distributions of Co, Mn - L$_{2,3}$ and O - K edges. \textbf{c)} The simultaneous HAADF micrograph of the Co$_{3-x}$Mn$_x$O$_4$ nanoparticle. \textbf{d)} The multi-modal reconstructions for the elemental distributions. \textbf{e)} EELS RGB overlay of the Co, S, and O maps. Scale bar, 2 nm.}
\label{fig::eels_overview}
\end{figure*}
Fused multi-modal electron microscopy accurately recovers chemical structure down to atomic length scales---demonstrated here for EELS spectroscopic signals. EELS derived chemical maps for Co$_{3-x}$Mn$_x$O$_4$ ($x=1.49$) high-performing super-capacitor nanoparticles~\cite{perera2015comno} are substantially improved by fused multi-modal electron microscopy in Figure~\ref{fig::eels_overview}. This composite Co-Mn oxide was designed to achieve a synergy between cobalt oxide's high specific capacitance and manganese oxide's long life cycle~\cite{perera2015comno,bhargava2019comnoSpinel}. While the Co$_{3-x}$Mn$_x$O$_4$ nanoparticle appears chemically homogeneous in the HAADF projection image along the [100] direction (Fig.~\ref{fig::eels_overview}c), core-shell distinctions are hinted at in the raw EELS maps (Fig.~\ref{fig::eels_overview}b). Specifically, these nanoparticles contain a Mn-rich center with a Co shell and homogeneous distribution of O. However the raw EELS maps are excessively degraded by noise, preventing analysis beyond rough assessment of specimen morphology. The multi-modal reconstructions (Fig.~\ref{fig::eels_overview}d) confirm the crystalline Co-rich shell and map the Co/Mn interface in greater detail (Fig.~\ref{fig::eels_overview}e). In the presence of cobalt and manganese, the HAADF image lacks noticeable contrast from oxygen; the resulting oxygen map lacks detail and benefits mostly from regularization.
Figure \ref{fig::ZnSCu} exhibits fused multi-modal electron microscopy at atomic resolution on copper sulphur heterostructured nanocrystals with zinc sulfide caps with potential applications in photovoltaic devices or battery electrodes~\cite{ha2014ZnSCu}. The copper sulfide properties are sensitive to the Cu-S stoichiometry and crystal structure at the interface between ZnS and Cu$_{0.64}$S$_{0.36}$. Figure \ref{fig::ZnSCu} shows high-resolution HAADF and EELS characterization of a heterostructure Cu$_{0.64}$S$_{0.36}$-ZnS interface. Fused multi-modal electron microscopy maps out the atomically sharp Cu$_{0.64}$S$_{0.36}$-ZnS interface and reveals step edges between the two layers. The labeled points on the RGB chemical overlay (Fig.~\ref{fig::ZnSCu}d) shows the chemical ratios produced by multi-modal EM for the Cu$_{0.64}$S$_{0.36}$ and ZnS regions---values which are consistent with the reported growth conditions. Figure \ref{fig::ZnSCu}e shows the algorithm convergence for each of the three terms in the optimization function (Eq.~\ref{eq:costFunc})---smooth and asymptotic decay is an indicator of reliable reconstruction. Refer to Supplementary Figure 2 for an additional demonstration at the atomic-scale on an ordered manganite system.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\linewidth]{Figs/Fig3_ZnSCu_v2.png}
\caption{\textbf{Recovering chemistry in an atomically sharp ZnS-Cu$_{0.64}$S$_{0.36}$ heterointerface interface.} \textbf{a)} The raw EELS maps for the Cu, S, and Zn L$_{2,3}$ edges. \textbf{b)} The multi-modal reconstructions for the elemental compositions. \textbf{c)} The simultaneous HAADF micrograph of the ZnS-Cu$_{0.64}$S$_{0.34}$ interface. \textbf{d)} Color overlay of the Zn, S, and Zn maps. The relative concentration for the constituent elements consist of 48$\pm$5.9$\%$ for Zn, 59.9$\pm$3.2$\%$ for Cu and 38$\pm$2.6$\%$ for S in the Cu$_{0.64}$S$_{0.36}$ layer and 48.9$\pm$6$\%$ in ZnS. \textbf{e)} Convergence plots for the three individual components in the cost function. Scale bar, 1 nm.}
\label{fig::ZnSCu}
\end{figure}
Fused multi-modal imaging of Fe and Pt distributions from inelastic multislice simulations (Fig.~\ref{fig::FePtSim}) provide ground truth solutions to validate recovery at atomic resolution under multiple scattering conditions of an on-axis $\sim$8 nm nanoparticle. Here, we applied Poisson noise (Fig.~\ref{fig::FePtSim}b) containing electron doses of $\sim$10$^9$~e~Å$^{-2}$, to produce chemical maps with noise levels resembling experimental atomic-resolution EELS datasets (SNR $\simeq 5$). We estimated SNR improvements by measuring peak-SNR for the noisy and recovered chemical maps \cite{hore2010psnr}. Qualitatively, the recovered chemical distributions (Fig.~\ref{fig::FePtSim}c) match the original images. Fig.~\ref{fig::FePtSim}d illustrates agreement of the line profiles as the atom column positions and relative peak intensities between the ground truth and multi-modal reconstruction are almost identical.
Simulating EELS chemical maps is computationally demanding as every inelastic scattering event requires propagation of an additional wavefunction~\cite{dwyer2015role,allen2015modelling}---scaling faster than the cube of the number of beams, $O(N^3 \log N)$. Inelastic transition potentials of interest (in this case the L$_{2,3}$ Fe and M$_{4,5}$ Pt edges) were calculated from density function theory (see Methods). Long computation times (nearly 4,000 core-hours) result from a large number of outgoing scattering channels corresponding to the many possible excitations in a sample. For this reason, there is little precedence for inelastic image simulations. We relaxed the runtime by utilizing the PRISM STEM-EELS approximation, achieving over a ten-fold speedup (see Methods)~\cite{brown2019prismEELS}. Future work may explore the effects of smaller ADF collection angles with increased coherence lengths and crystallographic contrast~\cite{hartel1996conditions,zhang2021combining}, or thicker specimens where electron channeling becomes more concerning~\cite{anstis2003limitations,hovden2012channeling}.
\subsection*{Quantifying Chemical Concentration}
\begin{figure}[ht!]
\centering
\includegraphics[width=\linewidth]{Figs/Fig4_FePt_Sim.png}
\caption{\textbf{Inelastic FePt nanoparticle simulation.} \textbf{a)} Ground truth EELS images generated from inelastic simulations. \textbf{b)} EELS maps degraded with Poisson shot noise. SNR shown on top right. \textbf{c)} Recovered atomic-resolution EELS maps for the Fe and Pt distributions. Estimated SNR shown on top right. \textbf{d)} Line profiles of the marked yellow bars (10 pixels in width) in (c) compares the Multi-Modal reconstruction and ground truth. Scale bar, 1 nm. }
\label{fig::FePtSim}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=\linewidth]{Figs/Fig5_Concentration.png}
\caption{\textbf{Measuring relative concentration for experimental and synthetic datasets.} Pixel intensity histograms for an experimental Zr (green), Ni (blue) and synthetic Ga (red) concentration maps. The standard deviation ($\sigma$) for each element is reported. The raw and reconstructed EDX maps are illustrated inside of the plot. Ground truth concentrations are highlighted by the respective colored triangles above the top axis. Stable convergence for the three components in the cost function: model term (orange), data fidelity (magenta), and regularization (turquoise) are illustrated in the inset. Qualitatively the convergence is identical for all three example datasets. Zr and Ni scale bars: 5, 10 nm, respectively.}
\label{fig::stoichiometry}
\end{figure}
Fused multi-modal electron microscopy can produce stoichiometricly meaningful chemical maps without specific knowledge of inelastic cross sections. Here, the ratio of pixel values in the reconstructed maps quantify elemental concentration. We demonstrate quantifiable chemistry on experimental metal oxide thin films with known stoichiometry: NiO~\cite{egerton1994NiO} and ZrO$_{2}$. A histogram of intensities from the recovered chemical maps are fit with Gaussian distributions to determine the average concentration. The recovered pixel values highlighted in Figure \ref{fig::stoichiometry} followed a single Gaussian distribution where the Zr and Ni concentrations are centered about 35$\pm$5.8$\%$ and 50$\pm$2.9$\%$. In both cases, the average Ni and Zr relative concentration is approximately equivalent to the expected ratio from the crystal stoichiometry: 33$\%$ and 50$\%$. The CoS nanoparticle in Fig. \ref{fig::edx_overview} follows a bi-modal distribution for the core and shell phases (Supplementary Figure 5). We found measuring stoichiometry is robust across a range of $\gamma$ values close to 1.7. In cases where $\gamma$ is far off (e.g., $\gamma = 1.0$), the quantification is systematically incorrect (Supplementary Figure 6).
We further validate stoichiometric recovery on a synthetic gallium oxide crystal (Fig~\ref{fig::stoichiometry}) where two overlapping Ga and O thin films of equal thickness have a stoichiometery of Ga$_2$O$_3$. The simulated HAADF signal is proportional to $\sum_{i} (\bm{x}_i Z_{i})^{\gamma}$ where $\bm{x}_{i}$ is the concentration for element ${i}$ and $Z_{i}$ is the atomic number. As shown by the histogram, the simulated results agree strongly with the prior knowledge and successfully recovers the relative Ga concentration. The Gaussian distribution is centered about 40$\pm$0.4$\%$ when the ground truth is 40$\%$. The inset shows convergence plots.
We estimate a stoichiometric error of less-than 15\% for most materials based on the relative concentration's standard deviation ($\pm$7\%) added in quadrature with the variation of solutions ($\pm$6\%).
Although the algorithm shows stable convergence,
the overall quantitative conclusions are slightly sensitive to the selection of hyperparameters.
We estimate incorrect selection of hyperparameters could result in variation of roughly $\pm$6$\%$ from the correct prediction in stoichiometery even when the algorithm converges (convergence shown in Supplementary Figures 8-9). This error is comparable to estimating chemical concentrations directly from EELS / EDX spectral maps from the ratio of scattering cross section against core-loss intensity~\cite{rez1982crossSections}. However, traditional approaches require accurate knowledge of all experimental parameters (e.g., beam energy, specimen-thickness, collection angles) and accurate calculation of the inelastic cross-section typically to provide errors roughly between 5-10$\%$~\cite{egerton1978eelsQuant}.
\subsection*{Influence of Electron Dose}
To better understand the accuracy of fused multi-modal electron microscopy at low doses, we performed a quantitative study of normalized root-mean-square error (RMSE) concentrations for a simulated 3D core-shell nanoparticle (CoS core, CoO shell).
Figure~\ref{fig::3Dsimulation} shows the fused multi-modal reconstruction accuracy across a wide range of HAADF and chemical SNR. The simulated projection images were generated by simple linear incoherent imaging model of the 3D chemical compositions highlighted in Fig.~\ref{fig::3Dsimulation}d--here the probe's depth of focus is much larger than the object. Random Poisson noise corresponding to different electron dose levels was applied to vary the SNR across each pixel.
\begin{figure}[ht!]
\centering
\includegraphics[width=\linewidth]{Figs/Fig6_3DSimulation.png}
\caption{\textbf{Estimating dose requirements for accurate chemical recovery.} \textbf{a)} A RMSE map representing the reconstruction error as a function of multiple spectroscopic and HAADF SNR. Brighter pixels denote results containing the incorrect concentrations from the ground truth. \textbf{b)} Visualization of three points on the phase diagram corresponding to increasing ADF / chemical electron dose. \textbf{c)} A plot of average standard error vs. RMSE demonstrating the two metrics are linearly correlated. \textbf{d)} The 3D model for generating synthetic chemical and ADF projections. }
\label{fig::3Dsimulation}
\end{figure}
Overall, the RMSE simulation map (Fig.~\ref{fig::3Dsimulation}a) shows the core-shell nanoparticle chemical maps are accurately recovered at low-doses (HAADF SNR $\gtrsim 4$ and chemical SNR $\gtrsim 2$); however, they become less accurate at extremely low doses. The RMSE map for multimodal reconstruction shows a predictably continuous degradation in recovery as signals diminish. The degraded and reconstructed chemical maps for various noise levels are highlighted in Figure~\ref{fig::3Dsimulation}b. The Co map closely mirrors the Z-contrast observed in HAADF (not shown) simply because it is the heaviest element present. Usually researchers will perform spectroscopic experiments in the top right corner of Fig.~\ref{fig::3Dsimulation}a
(e.g., HAADF SNR $>20$, chemical SNR $>3$), which for this simulation, provides accurate recovery.
In actual experiments, the ground truth is unknown and RMSE cannot be calculated to assess fused multi-modal electron microscopy. However we can estimate accuracy by calculating an average standard error of our recovered image from the Hessian of our model (see methods). The standard error reflects uncertainty at each pixel in a recovered chemical map by quantifying the neighborhood size for similar solutions (Supplementary Figure 10). The average standard error across all pixels in a fused multi-modal image provides a single value metric of the reconstruction accuracy (see Methods). Figure \ref{fig::3Dsimulation}c shows that RMSE and average standard error correlate, especially at higher doses (SNR > 10).
\section*{Discussion}
While this paper highlights the advantages of multi-modal electron microscopy, the technique is not a black-box solution.
Step sizes for convergence and weights on the terms in the cost function (Eq.~\ref{eq:costFunc}) must be reasonably selected. This manuscript illustrates approaches to assess the validity of concentration measurements using confidence estimation demonstrated across several simulated and experimental material classes. Standard spectroscopic pre-processing methods become ever more critical in combination with multi-modal fusion. Improper background subtraction of EELS spectra or overlapping characteristic X-ray peaks that normally causes inaccurate stoichiometric quantification also reduces the accuracy of fused multi-modal imaging.
Fused multi-modal electron microscopy offers little advantage in recovering chemical maps for elements with insignificant contrast in the HAADF modality.
This property is limiting for analyzing specimens with low-Z elements in the presence of heavy elements (e.g., oxygen and lutetium). Future efforts could resolve this challenge by incorporating an additional complementary elastic imaging mode where light elements are visible, such as annular bright field (ABF)~\cite{findlay2010abf}. However in some instances, fused multi-modal electron microscopy may recover useful information for under-determined chemical signals. For example, in a Bi$_{0.35}$Sr$_{0.18}$Ca$_{0.47}$MnO$_3$ (BSCMO) system~\cite{savitzky2017bscmo}, only the Ca, Mn, and O EELS maps were obtained, yet multimodality remarkably improves the SNR of measured maps despite missing two elements (Supplementary Figure 2).
Although fused multi-modal chemical mapping appears quite robust at nanometer or sub-nanometer resolution, we found atomic-resolution reconstructions can be challenged by spurious atom artifacts which require attention. However, this is easily remedied by down-sampling to frequencies below the first Bragg peaks and analysing a lower resolution chemical map. Alternatively, recovery with minimal spurious atom artifacts is achieved when lower resolution reconstructions are used as an initial guess (Supplementary Figure 11).
In summary, we present a model-driven data fusion algorithm that substantially improves the quality of electron microscopy spectroscopic maps at nanometer to atomic resolutions by using both elastic and inelastic signals.
From these signals, or modalities, each atom's chemical identity and coordination provides essential information about the performance of nanomaterials across a wide range of applications from clean energy, batteries, and opto-electronics, among many others. In both synthetic and experimental datasets, multi-modal electron microscopy shows quantitatively accurate chemical maps with values that reflect stoichiometry. This approach not only improves SNR but opens a pathway for low-dose chemical imaging of radiation sensitive materials. Although demonstrated herein for common STEM detectors (HAADF, EDX, and EELS), this approach can be extended to many other modalities---including pixel array detectors, annular bright field, ptychography, low-loss EELS, etc. One can imagine a future where all scattered and emitted signals in an electron microscope are collected and fused for maximally efficiently atomic characterization of matter.
\input{appendix.tex}
\section*{References}
{\printbibliography[heading=none]}
\section*{Acknowledgements}
R.H. and J.S. acknowledge support from the Army Research Office, Computing Sciences (W911NF-17-S-0002) and Dow Chemical Company.
\section*{Author Contributions}
J.S., R.H., Z.W.D. and Y.J. conceived the idea.
J.S. and R.H. implemented the multi-modal reconstruction algorithms and performed the analysis.
J.F. and Z.W.D. assisted with the algorithm formulation.
C.O. designed and ran the inelastic multi-slice simulations.
S.R. and A.J.F. conducted the EDX experiments.
R.H. and I.E.B. conducted the EELS experiments.
R.R., D. H., S.D.P., synthesized the Co$_{3-x}$Mn$_x$O and ZnSCu nanoparticles.
J.S. and R.H. wrote the manuscript.
All authors reviewed and commented on the manuscript.
\section*{Ethics Declarations}
\subsection*{Competing Interests}
The authors declare no competing interests.
\end{document}
| 3b35871df935cd50eab816266e50f800013ec56b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{section:introduction}
Theoretical performance guarantees and adaptability to real-world constraints determine the deployability, and hence the practical utility, of statistical methods and machine learning algorithms. For example, spectral clustering \citep{NgEtAl:2001:OnSpectralClustering,Luxburg:2007:ATutorialOnSpectralClustering}, one of the most sought after algorithms for clustering and community detection, has been studied under various constraints such as \textit{must-link} and \textit{cannot-link} constraints \citep{KamvarEtAl:2003:SpectralLearning, WangDavidson:2010:FlexibleConstrainedSpectralClustering}, size-balanced clusters \citep{BanerjeeGhosh:2006:ScalableClusteringAlgorithmsWithBalancingConstraints}, and statistical fairness \citep{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}. Commonly used constraints in practice can be categorized as \textit{population-level} (also known as statistical-level) or \textit{individual-level} constraints. However, the only known consistency guarantees for constrained spectral clustering were established in \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} for the first case, where the goal is to find balanced clusters with respect to an auxiliary categorical attribute. In this paper, we study a problem setting where the auxiliary information is encoded in a graph $\mathcal{R}$, which we refer to as a \textit{representation graph}. We study spectral clustering in a given \textit{similarity graph} $\mathcal{G}$ under an individual-level balancing constraint that is specified using the representation graph $\mathcal{R}$. There are two-fold advantages of this setting: \textbf{(i)} our constraint selects clusters that are balanced from each individual's perspective, and \textbf{(ii)} it enables us to define a new variant of the stochastic block model (SBM) \citep{HollandEtAl:1983:StochasticBlockmodelsFirstSteps} that plants the properties of $\mathcal{R}$ into the sampled graphs, making this variant of SBM \textit{representation-aware} (aware of the representation graph $\mathcal{R}$).
\subsection{Problem setting and applications}
\label{section:problem_setting_and_applications}
Let $\mathcal{G}$ denote a similarity graph based on which clusters have to be discovered. We consider a setting where each node in $\mathcal{G}$ specifies a list of its representatives (other nodes in $\mathcal{G}$) using an auxiliary representation graph $\mathcal{R}$. The graph $\mathcal{R}$ is defined on the same set of nodes as $\mathcal{G}$ and its edges specify the ``is representative of'' relationship. For example, $\mathcal{R}$ may be a result of the interactions between individuals that result from values of certain latent node attributes like age and gender. This motivates a \textit{representation constraint} that requires the clusters in $\mathcal{G}$ to be such that each node has a sufficient number of representatives (as per $\mathcal{R}$) in all the clusters. Our goal is to develop and analyze variants of the spectral clustering algorithm with respect to this constraint.
We begin by briefly describing two applications that motivate the above discussed problem. The first application is concerned with the ``fairness'' of clusters. Informally, a node or individual finds the clusters fair if it has a sufficient representation in all the clusters.
The work of
\citet{ChierichettiEtAl:2017:FairClusteringThroughFairlets} requires the clusters to be balanced with respect to various \textit{protected groups} (like gender or race). For example, if $50\%$ of the population is female then the same proportion should be respected in all clusters. This idea of proportional representation has been extended in various forms \citep{RosnerSchmidt:2018:PrivacyPreservingClusteringWithConstraints,BerceaEtAl:2019:OnTheCostOfEssentiallyFairClusterings,BeraEtAl:2019:FairAlgorithmsForClustering} and several efficient algorithms for discovering fair clusters under this notion have been proposed \citep{SchmidtEtAl:2018:FairCoresetsAndStreamingAlgorithmsForFairKMeansClustering,AhmadianEtAl:2019:ClusteringWithoutOverRepresentation,KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}. While the fairness notion mentioned above is a \textit{statistical} fairness notion (i.e., constraints are applied on protected groups as a whole), \citet{ChenEtAl:2019:ProportionallyFairClustering} and \citet{MahabadiEtAl:2020:IndividualFairnessForKClustering} develop \textit{individual} fairness notions that require examples to be “sufficiently close" to their cluster centroids. \citet{AndersonEtAl:2020:DistributionalIndividualFairnessInClustering} pursue a different direction and adapt the fairness notion proposed by \citet{DworkEtAl:2012:FairnessThroughAwareness} to the problem of clustering. Only \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} study spectral clustering in the context of (statistical) fairness. In contrast, we show in Section \ref{section:constraint} that our proposed constraint interpolates between statistical and individual fairness based on the structure of $\mathcal{R}$.
\begin{figure}[t]
\centering
\subfloat[][Protected groups]{\includegraphics[width=0.3\textwidth]{Images/toy_example_protected}\label{fig:toy_example:protected_groups}}%
\hspace{5mm}\subfloat[][Statistically fair clusters]{\includegraphics[width=0.3\textwidth]{Images/toy_example_fairsc}\label{fig:toy_example:fairsc}}%
\hspace{5mm}\subfloat[][Individually fair clusters]{\includegraphics[width=0.3\textwidth]{Images/toy_example_repfairsc}\label{fig:toy_example:repfairsc}}
\caption{An example representation graph $\mathcal{R}$. Panel (a) shows the protected groups recovered from $\mathcal{R}$. Panel (b) shows the clusters recovered by a statistically fair clustering algorithm. Panel (c) shows the ideal individually fair clusters. (Best viewed in color)}
\label{fig:fairness:toy_example}
\end{figure}
\begin{example}
\label{example:statistical_vs_individual_fairness}
To understand the need for individual fairness notions, consider the representation graph $\mathcal{R}$ specified in Figure \ref{fig:toy_example:protected_groups}. All the nodes have a self-loop associated with them that has not been shown for clarity. In this example, $N = 24$, $K = 2$, and every node is connected to $d=6$ nodes (including the self-loop). To use a statistical fairness notion \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets}, one would begin by clustering the nodes in $\mathcal{R}$ to approximate the protected groups, as the members of these protected groups will be each other's representatives to the first order of approximation. A natural choice is to have two protected groups, as shown in Figure~\ref{fig:toy_example:protected_groups} using different colors. However, clustering nodes based on these protected groups can produce the green and yellow clusters shown in Figure~\ref{fig:toy_example:fairsc}. It is easy to verify that these clusters satisfy the statistical fairness criterion as they have an equal number of members from both protected groups. However, these clusters are very "unfair" from the perspective of each individual. For example, node $v_1$ does not have enough representation in the yellow cluster as only one of its six representatives are in this cluster, despite the equal size of both the clusters. A similar argument can be made for every other node in this graph. This example highlights an extreme case where a statistically fair clustering is highly unfair from the perspective of each individual. Figure~\ref{fig:toy_example:repfairsc} shows another clustering assignment, and it is easy to verify that each node in this assignment has the same representation in both red and blue clusters, making it individually fair with respect to $\mathcal{R}$. Our goal is to develop algorithms that prefer the clusters in Figure~\ref{fig:toy_example:repfairsc} over the clusters in Figure~\ref{fig:toy_example:fairsc}.
\end{example}
Another possible application could be in balancing the load on computing resources in a cloud platform. Here, nodes in $\mathcal{G}$ correspond to processes and edges encode similarity among these processes in terms of shareable resources such as a read-only file. The edges in $\mathcal{R}$ on the other hand connect processes that share resources that can only be accessed by one process at a time (say a network channel). The goal is to cluster similar processes in $\mathcal{G}$ while ensuring that neighbors in $\mathcal{R}$ are spread out across clusters to avoid a collision.
\subsection{Contributions and results}
\label{section:contributions_and_results}
Our primary goal is to establish the statistical consistency of constrained spectral clustering, where the constraints are specified from the perspective of each individual node in the graph. Towards this end, we make four contributions.
First, in Section \ref{section:constraint}, we introduce the notion of balance from the perspective of each individual node in the graph and formally specify our representation constraint as a simple linear expression. While we focus on spectral clustering in this paper, the proposed constraint may be of independent interest for other clustering techniques as well.
Second, in Sections~\ref{section:unnormalized_repsc} and \ref{section:normalized_repsc}, we develop \textit{representation-aware} variants of the unnormalized and normalized spectral clustering. The proposed algorithms incorporate our representation constraint as a linear constraint in spectral clustering's optimization objective. The resulting problem can be easily solved using eigen-decomposition and the returned clusters approximately satisfy the constraint.
Third, in Section~\ref{section:rsbm}, we propose a variant of SBM called \textit{representation-aware} stochastic block model ($\mathcal{R}$-SBM). $\mathcal{R}$-SBM encodes a probability distribution over similarity graphs $\mathcal{G}$ conditioned on a given representation graph $\mathcal{R}$. It can be viewed as a model that plants the properties of $\mathcal{R}$ into $\mathcal{G}$. We show that $\mathcal{R}$-SBM generates similarity graphs that present a ``hard'' problem instance to the spectral algorithms in a constrained setting. In Section~\ref{section:consistency_results}, we consider the class of $d$-regular representation graphs and establish the weak-consistency of our algorithms (Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized}) for graphs sampled from $\mathcal{R}$-SBM. To the best of our knowledge, these are the first consistency results for constrained spectral clustering under individual-level constraints.
Fourth, in Section~\ref{section:numerical_results}, we present empirical studies on both simulated and real-world data to verify our theoretical guarantees. A comparison between the performance of the proposed algorithms and their closest counterparts in the literature demonstrates their practical utility. In particular, our experiments show that the $d$-regularity assumption on the representation graph is not necessary in practice.
We conclude the paper in Section \ref{section:conclusion} with a few remarks on promising directions for future work. The proofs of all technical lemmas are presented in the supplementary material \citep{ThisPaperSupp}.
\subsection{Related results}
\label{section:related_results}
Several algorithms for unconstrained clustering such as $k$-means \citep{HofmannBuhmann:1997:PairwiseDataClusteringByDeterministicAnnealing, WagstaffEtAl:2001:ConstrainedKMeansClusteringWithBackgroundKnowledge}, expectation-maximization based clustering \citep{ShentalEtAl:2003:ComputingGaussianMixtureModelsWithEMUsingEquivalenceConstraints}, and spectral clustering \citep{KamvarEtAl:2003:SpectralLearning} have been modified to satisfy the given \textit{must-link} (ML) and \textit{cannot-link} (CL) constraints \citep{BasuEtAl:2008:ConstrainedClustering} that specify pairs of nodes that should or should not belong to the same cluster. In this paper, we restrict our focus to spectral clustering as it provides a deterministic solution to the clustering problem in polynomial time and can detect arbitrarily shaped clusters \citep{Luxburg:2007:ATutorialOnSpectralClustering}. Existing approaches modify spectral clustering by preprocessing the input similarity graph \citep{KamvarEtAl:2003:SpectralLearning, LuCarreiraPerpinan:2008:ConstrainedSpectralClusteringThroughAffinityPropagation}, post-processing the eigenvectors of the Laplacian matrix \citep{LiEtAl:2009:ConstrainedClusteringViaSpectralRegularization}, or modifying the optimization problem solved by spectral clustering \citep{YuShi:2001:GroupingWithBias,YuShi:2004:SegmentationGivenPartialGroupingConstraints, BieEtAl:2004:LearningFromGeneralLabelConstraints, WangDavidson:2010:FlexibleConstrainedSpectralClustering,ErikssonEtAl:2011:NormalizedCutsRevisited, KawaleBoley:2013:ConstrainedSpectralClusteringUsingL1Regularization, WangEtAl:2014:OnConstrainedSpectralClusteringAndItsApplications}. Researchers have also studied spectral approaches that, for example, handle inconsistent \citep{ColemanEtAl:2008:SpectralClusteringWithInconsistentAdvice} or sparse \citep{ZhuEtAl:2013:ConstrainedClustering} constraints, actively solicit constraints \citep{WangDavidson:2010:ActiveSpectralClustering}, or modify variants of spectral clustering \citep{RangapuramHein:2012:Constrained1SpectralClustering, WangEtAl:2009:IntegratedKLClustering}, to name but a few. Other types of constraints such as those on cluster sizes \citep{BanerjeeGhosh:2006:ScalableClusteringAlgorithmsWithBalancingConstraints, DemirizEtAl:2008:UsingAssignmentConstraintsToAvoidEmptyClustersInKMeansClustering} and those that can be expressed as linear expressions \citep{XuEtAl:2009:FastNormalizedCutWithLinearConstraints} have also been explored. While this has been an active area of research, theoretical consistency guarantees on the performance of these constrained spectral clustering algorithms are largely missing from the literature.
(Unconstrained) spectral clustering is backed by strong statistical guarantees that usually take the form ``the algorithm makes $o(N)$ mistakes with probability $1 - o(1)$.'' Here, $N$ is the number of nodes in the similarity graph. These results consider a random model that generates problem instances for the algorithm (similarity graph in this case) with known ground-truth clusters. The high probability bound is with respect to this random model and the mistakes are computed against the ground-truth clusters. An algorithm that satisfies the condition above is called \textit{weakly consistent} \citep{Abbe:2018:CommunityDetectionAndStochasticBlockModels}. A common choice for the random model is the Stochastic Block Model (SBM) \citep{HollandEtAl:1983:StochasticBlockmodelsFirstSteps}. In this model, nodes have predefined community memberships that are used to sample edges with different probabilities. \citet{RoheEtAl:2011:SpectralClusteringAndTheHighDimensionalSBM} established the weak consistency of spectral clustering under the SBM. \citet{LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM} instead used a variant of SBM to sample networks with a more realistic degree distribution \citep{KarrerNewman:2011:StochasticBlockmodelsAndCommunityStructureInNetworks}. Several other variants of SBM have also been used to provide appropriate problem instances such as graphs with overlapping clusters \citep{ZhangEtAl:2014:DetectingOverlappingCommunitiesInNetworksUsingSpectralMethods}, observable node covariates \citep{BinkiewiczEtAl:2017:CovariateAssistedSpectralClustering}, or two alternative set of clusters, one \textit{unfair} and one \textit{fair} \citep{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}. \citet{LuxburgEtAl:2008:ConsistencyOfSpectralClustering} use a different random model where the similarity graph encodes pairwise cosine similarity between input feature vectors that follow a particular probability distribution. \citet{TremblayEtAl:2016:CompressiveSpectralClustering} study a variant of spectral clustering that is faster than the traditional algorithm. \textit{Strong consistency} results are also known in some cases \citep{GaoEtAl:2017:AchievingOptimalMisclassificationProportionInStochasticBlockModels,LeiZhu:2017:AGenericSampleSplittingApproachForRefinedCommunityRecoveryInSBMs, VuEtAl:2018:ASimpleSVDAlgorithmForFindingHiddenPartitions}. Finally, \citet{GhoshdastidarDukkipati:2017:UniformHypergraphPartitioning, GhoshdastidarDukkipati:2017:ConsistencyOfSpectralHypergraphPartitioningUnderPlantedPartitionModel} propose a variant of SBM for hypergraphs and establish the weak consistency of spectral algorithms in this setting.
Theoretical results for constrained clustering are primarily concerned with the computational complexity of the problem. It is known that the problem is NP-hard if one hopes for an exact solution that satisfies all the CL constraints \citep{DavidsonRavi:2005:ClusteringWithConstraints} or the statistical fairness constraint \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets}. Hence, most existing methods only satisfy the constraints approximately \citep{CucuringuEtAl:2016:SimpleAndScalableConstrainedClustering}. Results related to the cluster quality in the constrained setting only consider the algorithm's convergence to the global optima of a relaxed optimization problem \citep{XuEtAl:2009:FastNormalizedCutWithLinearConstraints}. \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} is a notable exception, however, even they consider constraints that apply at the level of protected groups as explained above, and not at the level of individuals. They follow a similar strategy as ours and modify spectral clustering to add a statistical fairness constraint. In Section~\ref{section:constraint}, we argue that a particular configuration of the representation graph $\mathcal{R}$ reduces our constraint to the statistical fairness criterion. Thus, the algorithms proposed in \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} are strictly special cases of the algorithms presented in this paper. To the best of our knowledge, we are the first to establish statistical consistency results for constrained spectral clustering for individual-level constraints. A subset of results presented in this paper are available online as a preliminary draft \citep{ThisPaperArXiv}.
\section{Notation and preliminaries}
\label{section:notation_and_preliminaries}
Let $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ denote a similarity graph, where $\mathcal{V} = \{v_1, v_2, \dots, v_N\}$ is the set of $N$ nodes, and $\mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$ is the set of edges. The aim of clustering is to partition the nodes into $K \geq 2$ clusters $\mathcal{C}_1, \dots, \mathcal{C}_K \subseteq \mathcal{V}$ such that each node in $\mathcal{V}$ belongs to exactly one cluster, i.e., $\mathcal{C}_i \cap \mathcal{C}_j = \phi$ if $i \neq j$ and $\cup_{k=1}^K \mathcal{C}_k = \mathcal{V}$. We further assume the availability of a \textit{representation graph}, denoted by $\mathcal{R} = (\mathcal{V}, \hat{\mathcal{E}})$, that encodes auxiliary information. Notice that $\mathcal{R}$ is defined on the same set of vertices as the similarity graph $\mathcal{G}$, but has a different set of edges $\hat{\mathcal{E}} \subseteq \mathcal{V} \times \mathcal{V}$. The discovered clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ are required to satisfy the constraint encoded by $\mathcal{R}$, as described in Section \ref{section:constraint}. $\mathbf{A} \in \{0, 1\}^{N \times N}$ and $\mathbf{R} \in \{0, 1\}^{N \times N}$ denote the adjacency matrices of graphs $\mathcal{G}$ and $\mathcal{R}$, respectively. We assume that $\mathcal{G}$ and $\mathcal{R}$ are undirected. Further, $\mathcal{G}$ has no self-loops. Thus, $\mathbf{A}$ and $\mathbf{R}$ are symmetric and $A_{ii} = 0$ for all $i \in [N]$, where $[n] \coloneqq \{1, 2, \dots, n\}$ for any integer $n$.
We propose modified variants of spectral clustering in Section \ref{section:algorithms}. Before describing the proposed algorithms, we begin with a brief review of the standard spectral clustering algorithm.
\subsection{Unnormalized spectral clustering}
\label{section:unnormalized_spectral_clustering}
Given a similarity graph $\mathcal{G}$, unnormalized spectral clustering finds clusters by approximately optimizing a quality metric known as the ratio-cut defined as \citep{Luxburg:2007:ATutorialOnSpectralClustering}
\begin{equation*}
\mathrm{RCut}(\mathcal{C}_1, \dots, \mathcal{C}_K) = \sum_{i = 1}^K \frac{\mathrm{Cut}(\mathcal{C}_i, \mathcal{V} \backslash \mathcal{C}_i)}{\abs{\mathcal{C}_i}}.
\end{equation*}
Here, $\mathcal{V} \backslash \mathcal{C}_i$ denotes the set difference between sets $\mathcal{V}$ and $\mathcal{C}_i$. For any two subsets $\mathcal{X}, \mathcal{Y} \subseteq \mathcal{V}$, $\mathrm{Cut}(\mathcal{X}, \mathcal{Y})$ is defined as
$\mathrm{Cut}(\mathcal{X}, \mathcal{Y}) = \frac{1}{2} \sum_{v_i \in \mathcal{X}, v_j \in \mathcal{Y}} A_{ij}$. That is, $\mathrm{Cut}(\mathcal{X}, \mathcal{Y})$ counts the number of edges that have one endpoint in $\mathcal{X}$ and another endpoint in $\mathcal{Y}$.
The Laplacian matrix $\mathbf{L}$ of the similarity graph $\mathcal{G}$ is defined as
\begin{equation}
\label{eq:L_def}
\mathbf{L} = \mathbf{D} - \mathbf{A}.
\end{equation}
Here, $\mathbf{D} \in \mathbb{R}^{N \times N}$ is the degree matrix, which is a diagonal matrix such that $D_{ii} = \sum_{j = 1}^N A_{ij}$, for all $i \in [N]$.
Further, define $\mathbf{H} \in \mathbb{R}^{N \times K}$ as
\begin{equation}
\label{eq:H_def}
H_{ij} = \begin{cases}
\frac{1}{\sqrt{\abs{\mathcal{C}_j}}} & \text{ if }v_i \in \mathcal{C}_j \\
0 & \text{ otherwise.}
\end{cases}
\end{equation}
One can easily verify that $\mathrm{RCut}(\mathcal{C}_1, \dots, \mathcal{C}_K) = \trace{\mathbf{H}^\intercal \mathbf{L} \mathbf{H}}$, where $\mathbf{H}$ corresponds to clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$. Thus, to find good clusters, one can solve:
\begin{equation*}
\min_{\mathbf{H} \in \mathbb{R}^{N \times K}} \,\,\,\, \trace{\mathbf{H}^\intercal \mathbf{L} \mathbf{H}} \,\,\,\, \text{s.t.} \,\,\,\, \mathbf{H} \text{ is of the form \eqref{eq:H_def}.}
\end{equation*}
It is computationally hard to solve this optimization problem due to the combinatorial nature of the constraint \citep{WagnerWagner:1993:BetweenMinCutAndGraphBisection}. Unnormalized spectral clustering instead solves the following relaxed optimization problem:
\begin{equation}
\label{eq:opt_problem_normal}
\min_{\mathbf{H} \in \mathbb{R}^{N \times K}} \,\,\,\, \trace{\mathbf{H}^\intercal \mathbf{L} \mathbf{H}} \,\,\,\, \text{s.t.} \,\,\,\, \mathbf{H}^\intercal \mathbf{H} = \mathbf{I}.
\end{equation}
The above relaxation is often referred to as the spectral relaxation.
By Rayleigh-Ritz theorem \citep[Section 5.2.2]{Lutkepohl:1996:HandbookOfMatrices},
the optimal matrix $\mathbf{H}^*$ is such that it has $\mathbf{u}_1, \mathbf{u}_2, \dots, \mathbf{u}_K \in \mathbb{R}^N$ as its columns, where $\mathbf{u}_i$ is the eigenvector corresponding to the $i^{th}$ smallest eigenvalue of $\mathbf{L}$ for all $i \in [K]$.
The algorithm clusters the rows of $\mathbf{H}^*$ into $K$ clusters using $k$-means clustering \citep{Lloyd:1982:LeastSquaresQuantisationInPCM} to return $\hat{\mathcal{C}}_1, \dots, \hat{\mathcal{C}}_K$. Algorithm \ref{alg:unnormalized_spectral_clustering} summarizes this procedure.
The Laplacian given in \eqref{eq:L_def} is more specifically known as unnormalized Laplacian. The next subsection describes a variant of spectral clustering, known as normalized spectral clustering \citep{ShiMalik:2000:NormalizedCutsAndImageSegmentation, NgEtAl:2001:OnSpectralClustering}, that uses the normalized Laplacian. Unless stated otherwise, we will use spectral clustering (without any qualification) to refer to unnormalized spectral clustering.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\State \textbf{Input:} Adjacency matrix $\mathbf{A}$, number of clusters $K \geq 2$
\State Compute the Laplacian matrix $\mathbf{L} = \mathbf{D} - \mathbf{A}$.
\State Compute the first $K$ eigenvectors $\mathbf{u}_1, \dots, \mathbf{u}_K$ of $\mathbf{L}$. Let $\mathbf{H}^* \in \mathbb{R}^{N \times K}$ be a matrix that has $\mathbf{u}_1, \dots, \mathbf{u}_K$ as its columns.
\State Let $\mathbf{h}^*_i$ denote the $i^{th}$ row of $\mathbf{H}^*$. Cluster $\mathbf{h}^*_1, \dots, \mathbf{h}^*_N$ into $K$ clusters using $k$-means clustering.
\State \textbf{Output:} Clusters $\hat{\mathcal{C}}_1, \dots, \hat{\mathcal{C}}_K$, \textrm{s.t.} $\hat{\mathcal{C}}_i = \{v_j \in \mathcal{V} : \mathbf{h}^*_j \text{ was assigned to the }i^{th} \text{ cluster}\}$.
\end{algorithmic}
\caption{Unnormalized spectral clustering}
\label{alg:unnormalized_spectral_clustering}
\end{algorithm}
\subsection{Normalized spectral clustering}
\label{section:normalized_spectral_clustering}
The ratio-cut objective divides $\mathrm{Cut}(\mathcal{C}_i, \mathcal{V} \backslash \mathcal{C}_i)$ by the number of nodes in $\mathcal{C}_i$ to balance the size of the clusters. The volume of a cluster $\mathcal{C} \subseteq \mathcal{V}$, defined as $\mathrm{Vol}(\mathcal{C}) = \sum_{v_i \in \mathcal{C}} D_{ii}$, is another popular notion of its size. The normalized cut or $\mathrm{NCut}$ objective divides $\mathrm{Cut}(\mathcal{C}_i, \mathcal{V} \backslash \mathcal{C}_i)$ by $\mathrm{Vol}(\mathcal{C}_i)$, and is defined as,
\begin{equation*}
\mathrm{NCut}(\mathcal{C}_1, \dots, \mathcal{C}_K) = \sum_{i = 1}^K \frac{\mathrm{Cut}(\mathcal{C}_i, \mathcal{V} \backslash \mathcal{C}_i)}{\mathrm{Vol}(\mathcal{C}_i)}.
\end{equation*}
As before, one can show that $\mathrm{NCut}(\mathcal{C}_1, \dots, \mathcal{C}_K) = \trace{\mathbf{T}^\intercal \mathbf{L} \mathbf{T}}$ \citep{Luxburg:2007:ATutorialOnSpectralClustering}, where $\mathbf{T} \in \mathbb{R}^{N \times K}$ is specified below.
\begin{equation}
\label{eq:T_def}
T_{ij} = \begin{cases}
\frac{1}{\sqrt{\mathrm{Vol}(\mathcal{C}_j)}} & \text{ if }v_i \in \mathcal{C}_j \\
0 & \text{ otherwise.}
\end{cases}
\end{equation}
Note that $\mathbf{T}^\intercal \mathbf{D} \mathbf{T} = \mathbf{I}$. Thus, the optimization problem for minimizing the NCut objective is
\begin{equation}
\label{eq:normalized_ideal_opt_problem}
\min_{\mathbf{T} \in \mathbb{R}^{N \times K}} \,\,\,\, \trace{\mathbf{T}^\intercal \mathbf{L} \mathbf{T}} \,\,\,\, \text{s.t.} \,\,\,\, \mathbf{T}^\intercal \mathbf{D} \mathbf{T} = \mathbf{I} \text{ and } \mathbf{T} \text{ is of the form \eqref{eq:T_def}.}
\end{equation}
As before, this optimization problem is hard to solve, and normalized spectral clustering solves a relaxed variant of this problem. Let $\mathbf{H} = \mathbf{D}^{1/2} \mathbf{T}$ and define the normalized graph Laplacian as $\mathbf{L}_{\mathrm{norm}} = \mathbf{I} - \mathbf{D}^{-1/2} \mathbf{A} \mathbf{D}^{-1/2}$. Normalized spectral clustering solves the following relaxed problem:
\begin{equation}
\label{eq:opt_problem_normal_normalized}
\min_{\mathbf{H} \in \mathbb{R}^{N \times K}} \,\,\,\, \trace{\mathbf{H}^\intercal \mathbf{L}_{\mathrm{norm}} \mathbf{H}} \,\,\,\, \text{s.t.} \,\,\,\, \mathbf{H}^\intercal \mathbf{H} = \mathbf{I}.
\end{equation}
Note that $\mathbf{H}^\intercal \mathbf{H} = \mathbf{I} \Leftrightarrow \mathbf{T}^\intercal \mathbf{D} \mathbf{T} = \mathbf{I}$. This is again the standard form of the trace minimization problem that can be solved using the Rayleigh-Ritz theorem. Algorithm \ref{alg:normalized_spectral_clustering} summarizes the normalized spectral clustering algorithm.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\State \textbf{Input:} Adjacency matrix $\mathbf{A}$, number of clusters $K \geq 2$
\State Compute the normalized Laplacian matrix $\mathbf{L}_{\mathrm{norm}} = \mathbf{I} - \mathbf{D}^{-1/2}\mathbf{A} \mathbf{D}^{-1/2}$.
\State Compute the first $K$ eigenvectors $\mathbf{u}_1, \dots, \mathbf{u}_K$ of $\mathbf{L}_{\mathrm{norm}}$. Let $\mathbf{H}^* \in \mathbb{R}^{N \times K}$ be a matrix that has $\mathbf{u}_1, \dots, \mathbf{u}_K$ as its columns.
\State Let $\mathbf{h}^*_i$ denote the $i^{th}$ row of $\mathbf{H}^*$. Compute $\tilde{\mathbf{h}}^*_i = \frac{\mathbf{h}^*_i}{\norm{\mathbf{h}^*_i}[2]}$ for all $i = 1, 2, \dots, N$.
\State Cluster $\tilde{\mathbf{h}}^*_1, \dots, \tilde{\mathbf{h}}^*_N$ into $K$ clusters using $k$-means clustering.
\State \textbf{Output:} Clusters $\hat{\mathcal{C}}_1, \dots, \hat{\mathcal{C}}_K$, \textrm{s.t.} $\hat{\mathcal{C}}_i = \{v_j \in \mathcal{V} : \tilde{\mathbf{h}}^*_j \text{ was assigned to the }i^{th} \text{ cluster}\}$.
\end{algorithmic}
\caption{Normalized spectral clustering}
\label{alg:normalized_spectral_clustering}
\end{algorithm}
\section{Representation constraint and representation-aware spectral clustering}
\label{section:algorithms}
\subsection{Representation constraint}
\label{section:constraint}
Here, an individual-level constraint for clustering the graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ is specified using a representation graph $\mathcal{R} = (\mathcal{V}, \hat{\mathcal{E}})$.
One intuitive explanation for $\mathcal{R}$ is that nodes connected in this graph represent each other in some form (say based on opinions if nodes correspond to people).
Let $\mathcal{N}_{\mathcal{R}}(i) = \{v_j \; : \; R_{ij} = 1\}$ denote the set of neighbors of node $v_i$ in $\mathcal{R}$. The size of $\mathcal{N}_{\mathcal{R}}(i) \cap \mathcal{C}_k$ denotes node $v_i$'s representation in cluster $\mathcal{C}_k$.
The goal of this constraint is to ensure that each node has an adequate amount of representation in all clusters.
\begin{definition}
\label{def:balance}
The balance of given clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ with respect to a node $v_i \in \mathcal{V}$ is defined as
\begin{equation}
\label{eq:balance}
\rho_i = \min_{k, \ell \in [K]} \;\; \frac{\abs{\mathcal{C}_k \cap \mathcal{N}_{\mathcal{R}}(i)}}{\abs{\mathcal{C}_\ell \cap \mathcal{N}_{\mathcal{R}}(i)}}.
\end{equation}
\end{definition}
It is easy to see that $0 \leq \rho_i \leq 1$ and higher values of $\rho_i$ indicate that node $v_i$ has an adequate representation in all clusters. Thus, one objective could be to find clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ that solve the following constrained optimization problem.
\begin{equation}
\label{eq:general_optimization_problem}
\min_{\mathcal{C}_1, \dots, \mathcal{C}_K} \;\; f(\mathcal{C}_1, \dots, \mathcal{C}_K) \;\;\;\; \text{s.t.} \;\;\;\; \rho_i \geq \alpha, \; \forall \; i \in [N].
\end{equation}
Here, $f(\cdot)$ is a function that is inversely proportional to the quality of clusters (such as $\mathrm{RCut}$ or $\mathrm{NCut}$), and $\alpha \in [0, 1]$ is a user specified threshold. However, it not clear as to how this approach can be combined with spectral clustering to develop a consistent algorithm. Therefore, we take a slightly different approach, as described below.
First, note that $\rho_i \leq \min_{k, \ell \in [K]} \; \frac{\abs{\mathcal{C}_k}}{\abs{\mathcal{C}_\ell}}$ for all $i \in [N]$. Therefore, the balance $\rho_i$ is maximized when the representatives $\mathcal{N}_{\mathcal{R}}(i)$ of node $v_i$ are split across clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ in proportion of the sizes of these clusters. Our constraint, formally defined below, requires this proportionality condition to be satisfied for each node $v_i \in \mathcal{V}$.
\begin{definition}[Representation constraint]
\label{def:representation_constraint}
Given a representation graph $\mathcal{R}$, clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ in $\mathcal{G}$ satisfy the representation constraint if $\abs{\mathcal{C}_k \cap \mathcal{N}_{\mathcal{R}}(i)} \propto \abs{\mathcal{C}_k}$, for all $i \in [N]$ and $k \in [K]$, or equivalently,
\begin{equation}
\label{eq:representation_constraint}
\frac{\abs{\mathcal{C}_k \cap \mathcal{N}_{\mathcal{R}}(i)}}{\abs{\mathcal{C}_k}} = \frac{\abs{\mathcal{N}_{\mathcal{R}}(i)}}{N}, \;\; \forall k \in [K], \; \forall i \in [N].
\end{equation}
\end{definition}
In other words, the representation constraint requires the representatives of any given node $v_i$ to have a proportional membership in all clusters. For example, if a node $v_i$ is connected to $30\%$ of all the nodes in $\mathcal{R}$, then the clusters discovered in $\mathcal{G}$ must be such that this node has $30\%$ representation in all clusters. We wish to re-emphasize that the representation constraint applies at the level of individual nodes unlike the constraint in \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets} that applies at the level of protected groups. In Sections \ref{section:unnormalized_repsc} and \ref{section:normalized_repsc}, we show that \eqref{eq:representation_constraint} can be integrated with the optimization problem solved by spectral clustering.
Appendix \ref{appendix:constraint} presents additional remarks about the properties of this constraint, in particular, its relation to the statistical-level constraint for categorical attributes \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets, KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}. It shows that \eqref{eq:representation_constraint} recovers the statistical-level constraint for a particular configuration of the representation graph, and generalizes it to individual-level constraint for other configurations. Next, we turn to the feasibility of the constraint.
\paragraph*{Feasibility} The optimization problem in \eqref{eq:general_optimization_problem} can always be solved for a small enough value of $\alpha$ (with the convention that $0/0 = 1$). On the other hand, the constraint in Definition \ref{def:representation_constraint} may not always be feasible. For example, if a node has only two representatives, i.e., $\abs{\mathcal{N}_{\mathcal{R}}(i)} = 2$, and there are $K > 2$ clusters, then \ref{eq:representation_constraint} can never be satisfied as there will always be at least one cluster $\mathcal{C}_k$ for which $\abs{\mathcal{C}_k \cap \mathcal{N}_{\mathcal{R}}(i)} = 0$. We argue in the subsequent subsections that spectral relaxation ensures the approximate satisfiability of the constraint when it is added to spectral clustering's optimization problem. We also describe a necessary assumption that is needed to ensure feasibility in practice and two additional assumptions that are required for our theoretical analysis. Thus, the way we use the proposed constraint makes it widely applicable, even though it is very strong. The general form of the constraint in \eqref{eq:general_optimization_problem} may be of independent interest in the context of other clustering algorithms, but we do not pursue this direction here.
\subsection{Unnormalized representation-aware spectral clustering (\textsc{URepSC})}
\label{section:unnormalized_repsc}
Recall from Section \ref{section:unnormalized_spectral_clustering} that unnormalized spectral clustering approximately minimizes the ratio-cut objective by relaxing an NP-hard optimization problem to solve \eqref{eq:opt_problem_normal}. The lemma below specifies a sufficient condition that implies the constraint in \eqref{eq:representation_constraint}.
\begin{lemma}
\label{lemma:constraint_matrix_unnorm}
Let $\mathbf{H} \in \mathbb{R}^{N \times K}$ have the form specified in \eqref{eq:H_def}. The condition
\begin{equation}
\label{eq:matrix_fairness_criteria}
\mathbf{R} \left( \mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal \right) \mathbf{H} = \mathbf{0}
\end{equation}
implies that the corresponding clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ satisfy the constraint in \eqref{eq:representation_constraint}. Here, $\mathbf{I}$ is the $N \times N$ identity matrix and $\bm{1}$ is a $N$-dimensional all-ones vector.
\end{lemma}
Ideally, we would like to solve the following optimization problem to get the clusters that satisfy the representation constraint,
\begin{equation}
\label{eq:optimization_problem_ideal}
\min_{\mathbf{H}} \;\;\;\; \trace{\mathbf{H}^\intercal \mathbf{L} \mathbf{H}} \;\;\;\; \text{s.t.} \;\;\;\;\mathbf{H} \text{ is of the form \eqref{eq:H_def}} ; \;\;\;\;
\mathbf{R} \left( \mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal \right) \mathbf{H} = \mathbf{0},
\end{equation}
where $\mathbf{L}$ is the unnormalized graph Laplacian defined in \eqref{eq:L_def}. However, as noted in Section \ref{section:unnormalized_spectral_clustering}, the first constraint on $\mathbf{H}$ makes this problem NP-hard. Thus, we solve the following relaxed problem,
\begin{equation}
\label{eq:opt_problem_with_eq_constraint}
\min_{\mathbf{H}} \;\;\;\; \trace{\mathbf{H}^\intercal \mathbf{L} \mathbf{H}} \;\;\;\; \text{s.t.} \;\;\;\; \mathbf{H}^\intercal \mathbf{H} = \mathbf{I}; \;\;\;\; \mathbf{R} \left( \mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal \right)\mathbf{H} = \mathbf{0}.
\end{equation}
Clearly, the columns of any feasible $\mathbf{H}$ must belong to the null space of $\mathbf{R}(\mathbf{I} - \bm{1}\bmone^\intercal / N)$. Thus, any feasible $\mathbf{H}$ can be expressed as $\mathbf{H} = \mathbf{Y} \mathbf{Z}$ for some matrix $\mathbf{Z} \in \mathbb{R}^{N - r \times K}$, where $\mathbf{Y} \in \mathbb{R}^{N \times N - r}$ is an orthonormal matrix containing the basis vectors of the null space of $\mathbf{R}(\mathbf{I} - \bm{1}\bmone^\intercal / N)$ as its columns. Here, $r$ is the rank of $\mathbf{R}(\mathbf{I} - \bm{1}\bmone^\intercal / N)$. Because $\mathbf{Y}^\intercal \mathbf{Y} = \mathbf{I}$, $\mathbf{H}^\intercal \mathbf{H} = \mathbf{Z}^\intercal \mathbf{Y}^\intercal \mathbf{Y} \mathbf{Z} = \mathbf{Z}^\intercal \mathbf{Z}$. Thus, $\mathbf{H}^\intercal \mathbf{H} = \mathbf{I} \Leftrightarrow \mathbf{Z}^\intercal \mathbf{Z} = \mathbf{I}$. The following optimization problem is equivalent to \eqref{eq:opt_problem_with_eq_constraint} by setting $\mathbf{H} = \mathbf{Y} \mathbf{Z}$.
\begin{equation}
\label{eq:optimization_problem}
\min_{\mathbf{Z}} \;\;\;\; \trace{\mathbf{Z}^\intercal \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Z}} \;\;\;\; \text{s.t.} \;\;\;\; \mathbf{Z}^\intercal \mathbf{Z} = \mathbf{I}.
\end{equation}
As in standard spectral clustering, the solution to \eqref{eq:optimization_problem} is given by the $K$ leading eigenvectors of $\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}$. Of course, for $K$ eigenvectors to exist, $N - r$ must be at least $K$, as $\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}$ has dimensions $N - r \times N - r$. The clusters can then be recovered by using $k$-means clustering to cluster the rows of $\mathbf{H} = \mathbf{Y} \mathbf{Z}$, as in Algorithm \ref{alg:unnormalized_spectral_clustering}. Algorithm \ref{alg:urepsc} summarizes this procedure. We refer to this algorithm as unnormalized representation-aware spectral clustering (\textsc{URepSC}).
\begin{algorithm}[t]
\begin{algorithmic}[1]
\State \textbf{Input: }Adjacency matrix $\mathbf{A}$, representation graph $\mathbf{R}$, number of clusters $K \geq 2$
\State Compute $\mathbf{Y}$ containing orthonormal basis vectors of $\nullspace{\mathbf{R}(\mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal)}$
\State Compute Laplacian $\mathbf{L} = \mathbf{D} - \mathbf{A}$
\State Compute leading $K$ eigenvectors of $\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}$. Let $\mathbf{Z}$ contain these vectors as its columns.
\State Apply $k$-means clustering to rows of $\mathbf{H} = \mathbf{Y} \mathbf{Z}$ to get clusters $\hat{\mathcal{C}}_1, \hat{\mathcal{C}}_2, \dots, \hat{\mathcal{C}}_K$
\State \textbf{Return:} Clusters $\hat{\mathcal{C}}_1, \hat{\mathcal{C}}_2, \dots, \hat{\mathcal{C}}_K$
\end{algorithmic}
\caption{\textsc{URepSC}}
\label{alg:urepsc}
\end{algorithm}
\subsection{Normalized representation-aware spectral clustering (\textsc{NRepSC})}
\label{section:normalized_repsc}
We use a similar strategy as the previous section to develop the normalized variant of \textsc{RepSC}. Recall from Section \ref{section:normalized_spectral_clustering} that normalized spectral clustering approximately minimizes the $\mathrm{NCut}$ objective. The lemma below is a counterpart of Lemma \ref{lemma:constraint_matrix_unnorm}. It formulates a sufficient condition that implies our constraint in \eqref{eq:representation_constraint}, but this time in terms of the matrix $\mathbf{T}$ defined in \eqref{eq:T_def}.
\begin{lemma}
\label{lemma:constraint_matrix_norm}
Let $\mathbf{T} \in \mathbb{R}^{N \times K}$ have the form specified in \eqref{eq:T_def}. The condition
\begin{equation}
\label{eq:normalized_matrix_fairness_criteria}
\mathbf{R} \left( \mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal \right) \mathbf{T} = \mathbf{0}
\end{equation}
implies that the corresponding clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ satisfy \eqref{eq:representation_constraint}. Here, $\mathbf{I}$ is the $N \times N$ identity matrix and $\bm{1}$ is a $N$ dimensional all-ones vector.
\end{lemma}
For \textsc{NRepSC}, we assume that the similarity graph $\mathcal{G}$ is connected so that the diagonal entries of $\mathbf{D}$ are strictly positive. We proceed as before to incorporate constraint \eqref{eq:normalized_matrix_fairness_criteria} in optimization problem \eqref{eq:normalized_ideal_opt_problem}. After applying the spectral relaxation, we get
\begin{equation}
\label{eq:optimization_problem_normalized}
\min_{\mathbf{T}} \;\;\;\; \trace{\mathbf{T}^\intercal \mathbf{L} \mathbf{T}} \;\;\;\; \text{s.t.} \;\;\;\; \mathbf{T}^\intercal \mathbf{D} \mathbf{T} = \mathbf{I}; \;\;\;\; \mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal / N) \mathbf{T} = \mathbf{0}.
\end{equation}
As before, $\mathbf{T} = \mathbf{Y} \mathbf{Z}$ for some $\mathbf{Z} \in \mathbb{R}^{N - r \times K}$, where recall that columns of $\mathbf{Y}$ contain orthonormal basis for $\nullspace{\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal / N)}$. This reparameterization yields
\begin{equation*}
\min_{\mathbf{Z}} \;\;\;\; \trace{\mathbf{Z}^\intercal \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Z}} \;\;\;\; \text{s.t.} \;\;\;\; \mathbf{Z}^\intercal \mathbf{Y}^\intercal \mathbf{D} \mathbf{Y} \mathbf{Z} = \mathbf{I}.
\end{equation*}
Define $\mathbf{Q} \in \mathbb{R}^{N - r \times N - r}$ such that $\mathbf{Q}^2 = \mathbf{Y}^\intercal \mathbf{D} \mathbf{Y}$. Note that $\mathbf{Q}$ exists as the entries of $\mathbf{D}$ are non-negative
Let $\mathbf{V} = \mathbf{Q} \mathbf{Z}$. Then, $\mathbf{Z} = \mathbf{Q}^{-1} \mathbf{V}$ and $\mathbf{Z}^\intercal \mathbf{Q}^2 \mathbf{Z} = \mathbf{V}^\intercal \mathbf{V}$ as $\mathbf{Q}$ is symmetric. Reparameterizing again, we get:
\begin{equation*}
\min_{\mathbf{V}} \;\;\;\; \trace{\mathbf{V}^\intercal \mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1} \mathbf{V}} \;\;\;\; \text{s.t.} \;\;\;\; \mathbf{V}^\intercal \mathbf{V} = \mathbf{I}.
\end{equation*}
This again is the standard form of the trace minimization problem and the optimal solution is given by the leading $K$ eigenvectors of $\mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}$. Algorithm \ref{alg:nrepsc} summarizes the normalized representation-aware spectral clustering algorithm, which we denote by \textsc{NRepSC}. Note that the algorithm assumes that $\mathbf{Q}$ is invertible, which requires the absence of isolated nodes in the similarity graph $\mathcal{G}$.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\State \textbf{Input: }Adjacency matrix $\mathbf{A}$, representation graph $\mathbf{R}$, number of clusters $K \geq 2$
\State Compute $\mathbf{Y}$ containing orthonormal basis vectors of $\nullspace{\mathbf{R}(\mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal)}$
\State Compute Laplacian $\mathbf{L} = \mathbf{D} - \mathbf{A}$
\State Compute $\mathbf{Q} = \sqrt{\mathbf{Y}^\intercal \mathbf{D} \mathbf{Y}}$ using the matrix square root
\State Compute leading $K$ eigenvectors of $\mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}$. Set them as columns of $\mathbf{V} \in \mathbb{R}^{N-r \times K}$
\State Apply $k$-means clustering to the rows of $\mathbf{T} = \mathbf{Y} \mathbf{Q}^{-1} \mathbf{V}$ to get clusters $\hat{\mathcal{C}}_1, \hat{\mathcal{C}}_2, \dots, \hat{\mathcal{C}}_K$
\State \textbf{Return:} Clusters $\hat{\mathcal{C}}_1, \hat{\mathcal{C}}_2, \dots, \hat{\mathcal{C}}_K$
\end{algorithmic}
\caption{\textsc{NRepSC}}
\label{alg:nrepsc}
\end{algorithm}
\subsection{Comments on the proposed algorithms}
\label{section:comments_on_the_proposed_algorithms}
Before proceeding with the theoretical analysis in Section \ref{section:analysis}, we first make two remarks about the proposed algorithms.
\paragraph*{Spectral relaxation} Note that the constraints $\mathbf{R}(\mathbf{I} - \bm{1}\bmone^\intercal / N) \mathbf{H} = \mathbf{0}$ and $\mathbf{R}(\mathbf{I} - \bm{1}\bmone^\intercal / N) \mathbf{T} = \mathbf{0}$ imply the satisfaction of our representation constraint only when $\mathbf{H}$ and $\mathbf{T}$ have the form given in \eqref{eq:H_def} and \eqref{eq:T_def}, respectively. Thus, a feasible solution to the relaxed optimization problem in~\eqref{eq:opt_problem_with_eq_constraint} or \eqref{eq:optimization_problem_normalized} may not necessarily result in \textit{representation-aware} clusters. In fact, even in the unconstrained case, there are no general guarantees that bound the difference between the optimal solution of~\eqref{eq:opt_problem_normal} or \eqref{eq:opt_problem_normal_normalized} and the respective optimal solutions of the original NP-hard ratio-cut/normalized-cut problems \citep{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}. Thus, the representation-aware nature of the clusters discovered by solving \eqref{eq:optimization_problem} or \eqref{eq:optimization_problem_normalized} cannot be guaranteed in the general case. Nonetheless, we show in Section~\ref{section:analysis} that the discovered clusters indeed satisfy the representation constraint under certain additional assumptions.
\paragraph*{Computational complexity} Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} have a time complexity of $O(N^3)$ and space complexity of $O(N^2)$. Finding the null space of $\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal / N)$ to calculate $\mathbf{Y}$ and computing the eigenvectors of appropriate matrices are the computationally dominant steps in both cases. This matches the worst-case complexity of the standard spectral clustering algorithm. For small $K$, several approximations can reduce this complexity, but most such techniques require $K = 2$ \citep{YuShi:2004:SegmentationGivenPartialGroupingConstraints,XuEtAl:2009:FastNormalizedCutWithLinearConstraints}.
\section{Analysis}
\label{section:analysis}
In this section, we show that Algorithms~\ref{alg:urepsc} and \ref{alg:nrepsc} recover the ground truth clusters with a high probability under certain assumptions on the representation graph. As we will see in Section \ref{section:consistency_results}, the ground truth clusters satisfy \eqref{eq:representation_constraint} by construction for similarity graphs $\mathcal{G}$ sampled from a modified variant of the Stochastic Block Model (SBM) \citep{HollandEtAl:1983:StochasticBlockmodelsFirstSteps}, described in Section \ref{section:rsbm}. Section \ref{section:consistency_results} presents our main results that establish a high probability upper bound on the number of mistakes made by the proposed algorithms. Corollaries \ref{corollary:weak_consistency_urepsc} and \ref{corollary:weak_consistency_nrepsc} then establish their weak-consistency.
\subsection{$\mathcal{R}$-SBM}
\label{section:rsbm}
The well known Stochastic Block Model (SBM) \citep{HollandEtAl:1983:StochasticBlockmodelsFirstSteps} allows one to sample random graphs with known clusters. It takes a function $\pi: \mathcal{V} \rightarrow [K]$ as input. This function assigns each node $v_i \in \mathcal{V}$ to one of the $K$ clusters. Then, independently for all node pairs $(v_i, v_j)$ such that $i > j$, $\mathrm{P}(A_{ij} = 1) = B_{\pi(v_i) \pi(v_j)},$ where $\mathbf{B} \in [0, 1]^{K \times K}$ is a symmetric matrix. The $(k, \ell)^{th}$ entry of $\mathbf{B}$ specifies the probability of a connection between two nodes that belong to clusters $\mathcal{C}_k$ and $\mathcal{C}_\ell$, respectively. A commonly used variant of SBM assumes $B_{kk} = \alpha$ and $B_{k\ell} = \beta$ for all $k, \ell \in [K]$ such that $k \neq \ell$. We define a variant of SBM with respect to a representation graph $\mathcal{R}$ below and refer to it as Representation-Aware SBM or $\mathcal{R}$-SBM.
\begin{definition}[$\mathcal{R}$-SBM]
\label{def:conditioned_sbm}
A $\mathcal{R}$-SBM is defined by the tuple $(\pi, \mathcal{R}, p, q, r, s)$, where $\pi: \mathcal{V} \rightarrow [K]$ maps nodes in $\mathcal{V}$ to clusters, $\mathcal{R}$ is a representation graph, and $1 \geq p \geq q \geq r \geq s \geq 0$ are probabilities used for sampling edges. Under this model, for all $i > j$,
\begin{equation}
\label{eq:sbm_specification}
\mathrm{P}(A_{ij} = 1) = \begin{cases}
p & \text{if } \pi(v_i) = \pi(v_j) \text{ and } R_{ij} = 1, \\
q & \text{if } \pi(v_i) \neq \pi(v_j) \text{ and } R_{ij} = 1, \\
r & \text{if } \pi(v_i) = \pi(v_j) \text{ and } R_{ij} = 0, \\
s & \text{if } \pi(v_i) \neq \pi(v_j) \text{ and } R_{ij} = 0.
\end{cases}
\end{equation}
\end{definition}
The similarity graphs sampled from a $\mathcal{R}$-SBM have two interesting properties. First, everything else being equal, nodes have a higher tendency to connect with other nodes in the same cluster as $p \geq q$ and $r \geq s$. Thus, $\mathcal{R}$-SBM plants the clusters specified by $\pi$ in the sampled graph $\mathcal{G}$. Second, and more importantly, $\mathcal{R}$-SBM also plants the properties of the given representation graph $\mathcal{R}$ in the sampled graphs $\mathcal{G}$. To see this, note that nodes that are connected in $\mathcal{R}$ have a higher probability of being connected in $\mathcal{G}$ as well ($p \geq r$ and $q \geq s$).
Recall that our algorithms must discover clusters in $\mathcal{G}$ in which the connected nodes in $\mathcal{R}$ are proportionally distributed. However, $\mathcal{R}$-SBM makes two nodes connected in $\mathcal{R}$ more likely to connect in $\mathcal{G}$, even if they do not belong to the same cluster ($q \geq r$). In this sense, graphs sampled from $\mathcal{R}$-SBM are ``hard'' instances from the perspective of our algorithms. When $\mathcal{R}$ itself has a community structure, there are two natural ways to cluster the nodes: \textbf{(i)} based on the ground-truth clusters $\mathcal{C}_1$, $\mathcal{C}_2$, \dots, $\mathcal{C}_K$ specified by $\pi$; and \textbf{(ii)} based on the communities in $\mathcal{R}$. The clusters based on communities in $\mathcal{R}$ are likely to not satisfy the representation constraint in Definition \ref{def:representation_constraint} as tightly connected nodes in $\mathcal{R}$ will be assigned to the same cluster in this case rather than being distributed across clusters.
We show in Section \ref{section:consistency_results} that, under certain assumptions on $\mathcal{R}$, the ground-truth clusters can be constructed so that they satisfy the representation constraint \eqref{eq:representation_constraint}. Assuming that the ground-truth clusters indeed satisfy \eqref{eq:representation_constraint}, the goal is to show that Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} recover the ground-truth clusters with high probability rather than returning any other natural but ``representation-unaware'' clusters.
\subsection{Consistency results}
\label{section:consistency_results}
As noted in Section \ref{section:constraint}, some representation graphs lead to constraints that cannot be satisfied. For our theoretical analysis, we restrict our focus to cases where the constraint in \eqref{eq:representation_constraint} is feasible. Towards this end, an additional assumption on $\mathcal{R}$ is required.
\begin{assumption}
\label{assumption:R_is_d_regular}
$\mathcal{R}$ is a $d$-regular graph for some $K \leq d \leq N$. Moreover, $R_{ii} = 1$ for all $i \in [N]$, and each node in $\mathcal{R}$ is connected to $d / K$ nodes from cluster $\mathcal{C}_i$, for all $i \in [K]$ (including the self-loop).
\end{assumption}
Assumption~\ref{assumption:R_is_d_regular} ensures the existence of a $\pi$ for which the corresponding ground-truth clusters satisfy the representation constraint in \eqref{eq:representation_constraint}. Namely, assuming equal-sized clusters, let $\pi(v_i) = k$, if $(k - 1) \frac{N}{K} \leq i \leq k \frac{N}{K}$ for all $i \in [N]$, and $k \in [K]$. It can be easily verified that the resulting clusters $\mathcal{C}_k = \{v_i : \pi(v_i) = k \}$, $k \in [K]$ satisfy \eqref{eq:representation_constraint}.
Before presenting our main results, we need to set up additional notation. Let $\bm{\Theta} \in \{0, 1\}^{N \times K}$ indicate the ground-truth cluster memberships, i.e., $\Theta_{ij} = 1 \Leftrightarrow v_i \in \mathcal{C}_j$. Similarly, $\hat{\bm{\Theta}} \in \{0, 1\}^{N \times K}$ indicates the clusters returned by the algorithm, i.e., $\hat{\Theta}_{ij} = 1 \Leftrightarrow v_i \in \hat{\mathcal{C}}_j$. Further, let $\mathcal{J}$ be the set of all $K \times K$ permutation matrices. The fraction of misclustered nodes \citep{LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM} is defined as
\begin{equation*}
M(\bm{\Theta}, \hat{\bm{\Theta}}) = \min_{\mathbf{J} \in \mathcal{J}} \frac{1}{N} \norm{\bm{\Theta} - \hat{\bm{\Theta}} \mathbf{J}}[0].
\end{equation*}
As the ground truth clusters $\mathcal{C}_1, \dots, \mathcal{C}_K$ satisfy \eqref{eq:representation_constraint} by construction, a low $M(\bm{\Theta}, \hat{\bm{\Theta}})$ indicates that the clusters returned by the algorithm approximately satisfy \eqref{eq:representation_constraint}. Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized} also use the eigenvalues of the Laplacian matrix in the expected case. We use $\mathcal{L}$ to denote this matrix, and define it as $\mathcal{L} = \mathcal{D} - \mathcal{A}$, where $\mathcal{A} = \mathrm{E}[\mathbf{A}]$ is the expected adjacency matrix of a graph sampled from $\mathcal{R}$-SBM and $\mathcal{D} \in \mathbb{R}^{N \times N}$ is a diagonal matrix such that $\mathcal{D}_{ii} = \sum_{j = 1}^N \mathcal{A}_{ij}$, for all $i \in [N]$. The next two results establish high-probability upper bounds on the fraction of misclustered nodes for \textsc{URepSC} and \textsc{NRepSC} for similarity graphs $\mathcal{G}$ sampled from $\mathcal{R}$-SBM.
\begin{theorem}[Error bound for \textsc{URepSC}]
\label{theorem:consistency_result_unnormalized}
Let $\rank{\mathbf{R}} \leq N - K$ and assume that all clusters have equal sizes. Let $\mu_1 \leq \mu_2 \leq \dots \leq \mu_{N - r}$ denote the eigenvalues of $\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}$, where $\mathbf{Y}$ was defined in Section \ref{section:unnormalized_repsc}. Define $\gamma = \mu_{K + 1} - \mu_{K}$. Under Assumption \ref{assumption:R_is_d_regular}, there exists a universal constant $\mathrm{const}(C, \alpha)$, such that if $\gamma$ satisfies $$\gamma^2 \geq \mathrm{const}(C, \alpha) (2 + \epsilon) p N K \ln N,$$ and $p \geq C \ln N / N$ for some $C > 0$, then,
$$M(\bm{\Theta}, \hat{\bm{\Theta}}) \leq \mathrm{const}(C, \alpha) \frac{(2 + \epsilon)}{\gamma^2} p N \ln N,$$
for every $\epsilon > 0$ with probability at least $1 - 2 N^{-\alpha}$ when a $(1 + \epsilon)$-approximate algorithm for $k$-means clustering is used in Step 5 of Algorithm \ref{alg:urepsc}.
\end{theorem}
\begin{theorem}[Error bound for \textsc{NRepSC}]
\label{theorem:consistency_result_normalized}
Let $\rank{\mathbf{R}} \leq N - K$ and assume that all clusters have equal sizes. Let $\mu_1 \leq \mu_2 \leq \dots \leq \mu_{N - r}$ denote the eigenvalues of $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1}$, where $\mathcal{Q} = \sqrt{\mathbf{Y}^\intercal \mathcal{D} \mathbf{Y}}$ and $\mathbf{Y}$ was defined in Section \ref{section:unnormalized_repsc}. Define $\gamma = \mu_{K + 1} - \mu_{K}$ and $\lambda_1 = qd + s(N - d) + (p - q) \frac{d}{K} + (r - s) \frac{N - d}{K}$. Under Assumption \ref{assumption:R_is_d_regular}, there are universal constants $\mathrm{const}_1(C, \alpha)$, $\mathrm{const}_4(C, \alpha)$, and $\mathrm{const}_5(C, \alpha)$ such that if:
\begin{enumerate}
\item $\left(\frac{\sqrt{p N \ln N}}{\lambda_1 - p}\right) \left(\frac{\sqrt{p N \ln N}}{\lambda_1 - p} + \frac{1}{6\sqrt{C}}\right) \leq \frac{1}{16(\alpha + 1)}$,
\item $\frac{\sqrt{p N \ln N}}{\lambda_1 - p} \leq \mathrm{const}_4(C, \alpha)$, and
\item $16(2 + \epsilon)\left[ \frac{8 \mathrm{const}_5(C, \alpha) \sqrt{K}}{\gamma} + \mathrm{const}_1(C, \alpha)\right]^2 \frac{p N^2 \ln N}{(\lambda_1 - p)^2} < \frac{N}{K}$,
\end{enumerate}
and $p \geq C \ln N / N$ for some $C > 0$, then,
$$M(\bm{\Theta}, \hat{\bm{\Theta}}) \leq 32(2 + \epsilon)\left[ \frac{8 \mathrm{const}_5(C, \alpha) \sqrt{K}}{\gamma} + \mathrm{const}_1(C, \alpha)\right]^2 \frac{p N \ln N}{(\lambda_1 - p)^2},$$
for every $\epsilon > 0$ with probability at least $1 - 2 N^{-\alpha}$ when a $(1 + \epsilon)$-approximate algorithm for $k$-means clustering is used in Step 6 of Algorithm \ref{alg:nrepsc}.
\end{theorem}
Next, we discuss our assumptions and use the error bounds above to establish the weak consistency of our algorithms.
\subsection{Discussion}
\label{section:discussion}
Note that $\mathbf{I} - \bm{1} \bm{1}^\intercal / N$ is a projection matrix and $\bm{1}$ is its eigenvector with eigenvalue $0$. Any vector orthogonal to $\bm{1}$ is also an eigenvector with eigenvalue $1$. Thus, $\rank{\mathbf{I} - \bm{1} \bm{1}^\intercal / N} = N - 1$. Because $\rank{\mathbf{R} (\mathbf{I} - \bm{1} \bm{1}^\intercal / N)} \leq \min(\rank{\mathbf{R}}, \rank{\mathbf{I} - \bm{1} \bm{1}^\intercal / N})$, requiring $\rank{\mathbf{R}} \leq N - K$ ensures that $\rank{\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal / N)} \leq N - K$, which is necessary for \eqref{eq:optimization_problem} and \eqref{eq:optimization_problem_normalized} to have a solution.
The assumption on the size of the clusters, together with the $d$-regularity assumption on $\mathcal{R}$, allows us to compute the smallest $K$ eigenvalues of the Laplacian matrix in the expected case. This is a crucial step in the proof of our main consistency results. The additional assumptions in Theorem \ref{theorem:consistency_result_normalized} are easy to satisfy as $\lambda_1$ scales linearly with $N$ for appropriate values of $p, q, r,$ and $s$. Similar assumptions were also used in \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}.
\begin{remark}
In practice, Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} only require the rank assumption on $\mathbf{R}$ to ensure the existence of solutions to the corresponding optimization problems. The assumptions on the size of clusters and $d$-regularity of $\mathcal{R}$ are only needed for our theoretical analysis.
\end{remark}
The next two corollaries are direct consequences of Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized}, respectively.
\begin{corollary}[Weak consistency of \textsc{URepSC}]
\label{corollary:weak_consistency_urepsc}
Under the same setup as Theorem \ref{theorem:consistency_result_unnormalized}, for \textsc{URepSC}, $M(\bm{\Theta}, \hat{\bm{\Theta}}) = o(1)$ with probability $1 - o(1)$ if $\gamma = \omega(\sqrt{pN K \ln N})$.
\end{corollary}
\begin{corollary}[Weak consistency of \textsc{NRepSC}]
\label{corollary:weak_consistency_nrepsc}
Under the same setup as Theorem \ref{theorem:consistency_result_normalized}, for \textsc{NRepSC}, $M(\bm{\Theta}, \hat{\bm{\Theta}}) = o(1)$ with probability $1 - o(1)$ if $\gamma = \omega(\sqrt{pN K \ln N} / (\lambda_1 - p))$.
\end{corollary}
Thus, under the assumptions in the corollaries above, Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} are weakly consistent \citep{Abbe:2018:CommunityDetectionAndStochasticBlockModels}. The conditions on $\gamma$ are satisfied in many interesting cases. For example, when there are $P$ protected groups, as is the case for statistical-level constraints, the equivalent representation graph has $P$ cliques that are not connected to each other (see Appendix \ref{appendix:constraint}). \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} show that $\gamma = \theta(N/K)$ in this case (for the unnormalized variant), which satisfies the criterion given above if $K$ is not too large.
Finally, Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized} require a $(1 + \epsilon)$-approximate solution to $k$-means clustering. Several efficient algorithms have been proposed in the literature for this task \citep{KumarEtAl:2004:ASimpleLinearTimeApproximateAlgorithmForKMeansClusteringInAnyDimension,ArthurVassilvitskii:2007:KMeansTheAdvantagesOfCarefulSeeding,AhmadianEtAl:2017:BetterGuaranteesForKMeansAndEuclideanKMedianByPrimalDualAlgorithms}. Such algorithms are also available in commonly used software packages like MATLAB and scikit-learn. The assumption that $p \geq C \ln N / N$ controls the sparsity of the graph, and is required in the the consistency proofs for standard spectral clustering as well \citep{LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM}.
\subsection{Proof of Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized}}
\label{section:proof_of_theorems}
The proof of Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized} follow the commonly used template for such results \citep{RoheEtAl:2011:SpectralClusteringAndTheHighDimensionalSBM, LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM}. In the context of \textsc{URepSC} (similar arguments work for \textsc{NRepSC} as well), we
\begin{enumerate}
\item Compute the expected Laplacian matrix $\mathcal{L}$ under $\mathcal{R}$-SBM and show that its top $K$ eigenvectors can be used to recover the ground-truth clusters (Lemmas \ref{lemma:introducing_uks}--\ref{lemma:orthonormal_eigenvectors_y2_yK}).
\item Show that these top $K$ eigenvectors lie in the null space of $\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal / N)$, and hence are also the top $K$ eigenvectors of $\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}$ (Lemma \ref{lemma:first_K_eigenvectors_of_L}). This implies that Algorithm \ref{alg:urepsc} returns the ground truth clusters in the expected case.
\item Use matrix perturbation arguments to establish a high probability mistake bound in the general case when the graph $\mathcal{G}$ is sampled from a $\mathcal{R}$-SBM (Lemmas \ref{lemma:bound_on_D-calD}--\ref{lemma:k_means_error}).
\end{enumerate}
We begin with a series of lemmas that highlight certain useful properties of eigenvalues and eigenvectors of the expected Laplacian $\mathcal{L}$. These lemmas will be used in Sections \ref{section:proof_consistency_unnormalized} and \ref{section:proof_consistency_normalized} to prove Theorem \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized}, respectively. See the supplementary material \citep{ThisPaperSupp} for the proofs of all technical lemmas. For the remainder of this section, we assume that all appropriate assumptions made in Theorems \ref{theorem:consistency_result_unnormalized} and \ref{theorem:consistency_result_normalized} are satisfied.
The first lemma shows that certain vectors that can be used to recover the ground-truth clusters indeed satisfy the representation constraint in \eqref{eq:matrix_fairness_criteria} and \eqref{eq:normalized_matrix_fairness_criteria}.
\begin{lemma}
\label{lemma:introducing_uks}
The $N$-dimensional vector of all ones, denoted by $\bm{1}$, is an eigenvector of $\mathbf{R}$ with eigenvalue $d$. Define $\mathbf{u}_k \in \mathbb{R}^N$ for $k \in [K - 1]$ as,
\begin{equation*}
u_{ki} = \begin{cases}
1 & \text{ if }v_i \in \mathcal{C}_k \\
-\frac{1}{K - 1} & \text{ otherwise,}
\end{cases}
\end{equation*}
where $u_{ki}$ is the $i^{th}$ element of $\mathbf{u}_k$. Then, $\bm{1}, \mathbf{u}_1, \dots, \mathbf{u}_{K - 1} \in \nullspace{\mathbf{R}(\mathbf{I} - \frac{1}{N}\bm{1}\bmone^\intercal)}$. Moreover, $\bm{1}, \mathbf{u}_1, \dots, \mathbf{u}_{K - 1}$ are linearly independent.
\end{lemma}
Recall that we use $\mathcal{A} \in \mathbb{R}^{N \times N}$ to denote the expected adjacency matrix of the similarity graph $\mathcal{G}$. Clearly, $\mathcal{A} = \tilde{\mathcal{A}} - p \mathbf{I}$, where $\tilde{\mathcal{A}}$ is such that $\tilde{\mathcal{A}}_{ij} = P(A_{ij} = 1)$ if $i \neq j$ (see \eqref{eq:sbm_specification}) and $\tilde{\mathcal{A}}_{ii} = p$ otherwise. Note that
\begin{equation}
\label{eq:eigenvector_A_tildeA}
\tilde{\mathcal{A}} \mathbf{x} = \lambda \mathbf{x} \,\,\,\, \Leftrightarrow \,\,\,\, \mathcal{A} \mathbf{x} = (\lambda - p) \mathbf{x}.
\end{equation}
Simple algebra shows that $\tilde{\mathcal{A}}$ can be written as
\begin{equation}
\label{eq:tilde_cal_A_def}
\tilde{\mathcal{A}} = q \mathbf{R} + s (\bm{1} \bm{1}^\intercal - \mathbf{R}) + (p - q)\sum_{k = 1}^K \mathbf{G}_k \mathbf{R} \mathbf{G}_k + (r - s) \sum_{k = 1}^K \mathbf{G}_k (\bm{1} \bm{1}^\intercal - \mathbf{R}) \mathbf{G}_k,
\end{equation}
where, for all $k \in [K]$, $\mathbf{G}_k \in \mathbb{R}^{N \times N}$ is a diagonal matrix such that $(\mathbf{G}_k)_{ii} = 1$ if $v_i \in \mathcal{C}_k$ and $0$ otherwise. The next lemma shows that $\bm{1}, \mathbf{u}_1, \dots, \mathbf{u}_{K - 1}$ defined in Lemma \ref{lemma:introducing_uks} are eigenvectors of $\tilde{\mathcal{A}}$.
\begin{lemma}
\label{lemma:uk_eigenvector_of_tildeA}
Let $\bm{1}, \mathbf{u}_1, \dots, \mathbf{u}_{K - 1}$ be as defined in Lemma \ref{lemma:introducing_uks}. Then,
\begin{eqnarray*}
\tilde{\mathcal{A}} \bm{1} &=& \lambda_1 \bm{1} \text{ where } \lambda_1 = qd + s(N - d) + (p - q) \frac{d}{K} + (r - s) \frac{N - d}{K}, \text{ and } \\
\tilde{\mathcal{A}} \mathbf{u}_k &=& \lambda_{1 + k} \mathbf{u}_k \text{ where } \lambda_{1 + k} = (p - q) \frac{d}{K} + (r - s) \frac{N - d}{K}.
\end{eqnarray*}
\end{lemma}
Let $\mathcal{L} = \mathcal{D} - \mathcal{A}$ be the expected Laplacian matrix, where $\mathcal{D}$ is a diagonal matrix with $\mathcal{D}_{ii} = \sum_{j=1}^N \mathcal{A}_{ij}$ for all $i \in [N]$. It is easy to see that $\mathcal{D}_{ii} = \lambda_1 - p$ for all $i \in [N]$ as $\mathcal{A} \bm{1} = (\lambda_1 - p) \bm{1}$ by \eqref{eq:eigenvector_A_tildeA} and Lemma \ref{lemma:uk_eigenvector_of_tildeA}. Thus, $\mathcal{D} = (\lambda_1 - p) \mathbf{I}$ and hence any eigenvector of $\tilde{\mathcal{A}}$ with eigenvalue $\lambda$ is also an eigenvector of $\mathcal{L}$ with eigenvalue $\lambda_1 - \lambda$. That is, if $\tilde{\mathcal{A}} \mathbf{x} = \lambda \mathbf{x}$,
\begin{equation}
\label{eq:eigenvectors_of_L}
\mathcal{L} \mathbf{x} = (\mathcal{D} - \mathcal{A})\mathbf{x} = ((\lambda_1 - p) \mathbf{I} - (\tilde{\mathcal{A}} - p \mathbf{I})) \mathbf{x} = (\lambda_1 - \lambda) \mathbf{x}.
\end{equation}
Hence, the eigenvectors of $\mathcal{L}$ corresponding to the $K$ smallest eigenvalues are the same as the eigenvectors of $\tilde{\mathcal{A}}$ corresponding to the $K$ largest eigenvalues.
Recall that the columns of the matrix $\mathbf{Y}$ used in Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} contain the orthonormal basis for the null space of $\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal/N)$. To solve \eqref{eq:optimization_problem} and \eqref{eq:optimization_problem_normalized}, we only need to optimize over vectors that belong to this null space. By Lemma \ref{lemma:introducing_uks}, $\bm{1}, \mathbf{u}_1, \dots, \mathbf{u}_{K - 1} \in \nullspace{\mathbf{R}(\mathbf{I} - \bm{1} \bm{1}^\intercal/N)}$ and these vectors are linearly independent. However, we need an orthonormal basis to compute $\mathbf{Y}$. Let $\mathbf{y}_1 = \bm{1} / \sqrt{N}$ and $\mathbf{y}_2, \dots, \mathbf{y}_K$ be orthonormal vectors that span the same space as $\mathbf{u}_1, \dots, \mathbf{u}_{K - 1}$. The next lemma computes such $\mathbf{y}_2, \dots, \mathbf{y}_K$. The matrix $\mathbf{Y} \in \mathbb{R}^{N \times N - r}$ contains these vectors $\mathbf{y}_1, \dots, \mathbf{y}_K$ as its first $K$ columns.
\begin{lemma}
\label{lemma:orthonormal_eigenvectors_y2_yK}
Define $\mathbf{y}_{1 + k} \in \mathbb{R}^N$ for $k \in [K - 1]$ as
\begin{equation*}
y_{1 + k, i} = \begin{cases}
0 & \text{ if } v_i \in \mathcal{C}_{k{'}} \text{ s.t. } k{'} < k \\
\frac{K - k}{\sqrt{\frac{N}{K}(K - k)(K - k + 1)}} & \text{ if } v_i \in \mathcal{C}_k \\
-\frac{1}{\sqrt{\frac{N}{K}(K - k)(K - k + 1)}} & \text{ otherwise.}
\end{cases}
\end{equation*}
Then, for all $k \in [K - 1]$, $\mathbf{y}_{1 + k}$ are orthonormal vectors that span the same space as $\mathbf{u}_1, \mathbf{u}_2, \dots, \mathbf{u}_{K - 1}$ and $\mathbf{y}_1^\intercal \mathbf{y}_{1 + k} = 0$. As before, $y_{1 + k, i}$ refers to the $i^{th}$ element of $\mathbf{y}_{1 + k}$.
\end{lemma}
Let $\mathbf{X} \in \mathbb{R}^{N \times K}$ be such that it has $\mathbf{y}_1, \dots, \mathbf{y}_{K}$ as its columns. If two nodes belong to the same cluster, the rows corresponding to these nodes in $\mathbf{X} \mathbf{U}$ will be identical for any $\mathbf{U} \in \mathbb{R}^{K \times K}$ such that $\mathbf{U}^\intercal \mathbf{U} = \mathbf{U} \mathbf{U}^\intercal = \mathbf{I}$. Thus, any $K$ orthonormal vectors belonging to the span of $\mathbf{y}_1, \dots, \mathbf{y}_K$ can be used to recover the ground truth clusters. With the general properties of the eigenvectors and eigenvalues established in the lemmas above, we next move on to the proof of Theorem \ref{theorem:consistency_result_unnormalized} in the next section and Theorem \ref{theorem:consistency_result_normalized} in Section \ref{section:proof_consistency_normalized}.
\subsubsection{Proof of Theorem \ref{theorem:consistency_result_unnormalized}}
\label{section:proof_consistency_unnormalized}
Let $\mathcal{Z} \in \mathbb{R}^{N - r \times K}$ be a solution to the optimization problem \eqref{eq:optimization_problem} in the expected case with $\mathcal{A}$ as input. The next lemma shows that columns of $\mathbf{Y} \mathcal{Z}$ indeed lie in the span of $\mathbf{y}_1, \dots, \mathbf{y}_K$. Thus, the $k$-means clustering step in Algorithm \ref{alg:urepsc} will return the correct ground truth clusters when $\mathcal{A}$ is passed as input.
\begin{lemma}
\label{lemma:first_K_eigenvectors_of_L}
Let $\mathbf{y}_1 = \bm{1} / \sqrt{N}$ and $\mathbf{y}_{1 + k}$ be as defined in Lemma \ref{lemma:orthonormal_eigenvectors_y2_yK} for all $k \in [K - 1]$. Further, let $\mathcal{Z}$ be the optimal solution of the optimization problem in \eqref{eq:optimization_problem} with $\mathbf{L}$ set to $\mathcal{L}$. Then, the columns of $\mathbf{Y} \mathcal{Z}$ lie in the span of $\mathbf{y}_1, \mathbf{y}_2, \dots, \mathbf{y}_K$.
\end{lemma}
Next, we use arguments from matrix perturbation theory to show a high-probability bound on the number of mistakes made by the algorithm. In particular, we need an upper bound on $\norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}}$, where $\mathbf{L}$ is the Laplacian matrix for a graph randomly sampled from $\mathcal{R}$-SBM and $\norm{\mathbf{P}} = \sqrt{\lambdamax{\mathbf{P}^\intercal \mathbf{P}}}$ for any matrix $\mathbf{P}$. Note that $\norm{\mathbf{Y}} = \norm{\mathbf{Y}^\intercal} = 1$ as $\mathbf{Y}^\intercal \mathbf{Y} = \mathbf{I}$. Thus,
\begin{equation}
\label{eq:reducing_YLY_to_L}
\norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}} \leq \norm{\mathbf{Y}^\intercal} \,\, \norm{\mathbf{L} - \mathcal{L}} \,\, \norm{\mathbf{Y}} = \norm{\mathbf{L} - \mathcal{L}}.
\end{equation}
Moreover, $$\norm{\mathbf{L} - \mathcal{L}} = \norm{\mathbf{D} - \mathbf{A} - (\mathcal{D} - \mathcal{A})} \leq \norm{\mathbf{D} - \mathcal{D}} + \norm{\mathbf{A} - \mathcal{A}}.$$ The next two lemmas bound the two terms on the right hand side of the inequality above, thus providing an upper bound on $\norm{\mathbf{L} - \mathcal{L}}$, and hence on $\norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}}$ by \eqref{eq:reducing_YLY_to_L}.
\begin{lemma}
\label{lemma:bound_on_D-calD}
Assume that $p \geq C \frac{\ln N}{N}$ for some constant $C > 0$. Then, for every $\alpha > 0$, there exists a constant $\mathrm{const}_1(C, \alpha)$ that only depends on $C$ and $\alpha$ such that $$\norm{\mathbf{D} - \mathcal{D}} \leq \mathrm{const}_1(C, \alpha) \sqrt{p N \ln N}$$ with probability at-least $1 - N^{-\alpha}$.
\end{lemma}
\begin{lemma}
\label{lemma:bound_on_A-calA}
Assume that $p \geq C \frac{\ln N}{N}$ for some constant $C > 0$. Then, for every $\alpha > 0$, there exists a constant $\mathrm{const}_2(C, \alpha)$ that only depends on $C$ and $\alpha$ such that $$\norm{\mathbf{A} - \mathcal{A}} \leq \mathrm{const}_2(C, \alpha) \sqrt{p N}$$ with probability at-least $1 - N^{-\alpha}$.
\end{lemma}
From Lemmas \ref{lemma:bound_on_D-calD} and \ref{lemma:bound_on_A-calA}, we conclude that there is always a constant $\mathrm{const}_3(C, \alpha) = \max\{\mathrm{const}_1(C, \alpha), \mathrm{const}_2(C, \alpha)\}$ such that, for any $\alpha > 0$, with probability at least $1 - 2N^{-\alpha}$,
\begin{equation}
\label{eq:L-calL_bound}
\norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}} \leq \norm{\mathbf{L} - \mathcal{L}} \leq \mathrm{const}_3(C, \alpha) \sqrt{p N \ln N}.
\end{equation}
Let $\mathcal{Z}$ and $\mathbf{Z}$ denote the optimal solution of \eqref{eq:optimization_problem} in the expected ($\mathbf{L}$ replaced with $\mathcal{L}$) and observed case.
We use \eqref{eq:L-calL_bound} to show a bound on $\norm{\mathbf{Y} \mathcal{Z} - \mathbf{Y} Z}[F]$ in Lemma \ref{lemma:bound_on_eigenvector_diff} and then use this bound to argue that Algorithm \ref{alg:urepsc} makes a small number of mistakes when the graph is sampled from $\mathcal{R}$-SBM.
\begin{lemma}
\label{lemma:bound_on_eigenvector_diff}
Let $\mu_1 \leq \mu_2 \leq \dots \leq \mu_{N - r}$ be eigenvalues of $\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}$. Further, let the columns of $\mathcal{Z} \in \mathbb{R}^{N - r \times K}$ and $\mathbf{Z} \in \mathbb{R}^{N - r \times K}$ correspond to the leading $K$ eigenvectors of $\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}$ and $\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}$, respectively. Define $\gamma = \mu_{K + 1} - \mu_{K}$. Then, with probability at least $1 - 2N^{-\alpha}$, $$\inf_{\mathbf{U} \in \mathbb{R}^{K \times K} : \mathbf{U}\bfU^\intercal = \mathbf{U}^\intercal \mathbf{U} = \mathbf{I}} \norm{\mathbf{Y} \mathcal{Z} - \mathbf{Y} \mathbf{Z} \mathbf{U}}[F] \leq \mathrm{const}_3(C, \alpha) \frac{4\sqrt{2K}}{\gamma} \sqrt{p N \ln N},$$ where $\mathrm{const}_3(C, \alpha)$ is from \eqref{eq:L-calL_bound}.
\end{lemma}
Recall that $\mathbf{X} \in \mathbb{R}^{N \times K}$ is a matrix that contains $\mathbf{y}_1, \dots, \mathbf{y}_K$ as its columns. Let $\mathbf{x}_i$ denote the $i^{th}$ row of $\mathbf{X}$. Simple calculation using Lemma \ref{lemma:orthonormal_eigenvectors_y2_yK} shows that,
\begin{equation*}
\norm{\mathbf{x}_i - \mathbf{x}_j}[2] = \begin{cases}
0 & \text{ if }v_i \text{ and }v_j \text{ belong to the same cluster} \\
\sqrt{\frac{2K}{N}} & \text{ otherwise.}
\end{cases}
\end{equation*}
By Lemma \ref{lemma:first_K_eigenvectors_of_L}, $\mathcal{Z}$ can be chosen such that $\mathbf{Y} \mathcal{Z} = \mathbf{X}$. Let $\mathbf{U}$ be the matrix that solves $\inf_{\mathbf{U} \in \mathbb{R}^{K \times K} : \mathbf{U}\bfU^\intercal = \mathbf{U}^\intercal \mathbf{U} = \mathbf{I}} \norm{\mathbf{Y} \mathcal{Z} - \mathbf{Y} \mathbf{Z} \mathbf{U}}[F]$. As $\mathbf{U}$ is orthogonal, $\norm{\mathbf{x}_i^\intercal \mathbf{U} - \mathbf{x}_j^\intercal \mathbf{U}}[2] = \norm{\mathbf{x}_i - \mathbf{x}_j}[2]$. The following lemma is a direct consequence of Lemma 5.3 in \citep{LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM}.
\begin{lemma}
\label{lemma:k_means_error}
Let $\mathbf{X}$ and $\mathbf{U}$ be as defined above. For any $\epsilon > 0$, let $\hat{\bm{\Theta}} \in \mathbb{R}^{N \times K}$ be the assignment matrix returned by a $(1 + \epsilon)$-approximate solution to the $k$-means clustering problem when rows of $\mathbf{Y} \mathbf{Z}$ are provided as input features. Further, let $\hat{\bm{\mu}}_1$, $\hat{\bm{\mu}}_2$, \dots, $\hat{\bm{\mu}}_K \in \mathbb{R}^{K}$ be the estimated cluster centroids. Define $\hat{\mathbf{X}} = \hat{\bm{\Theta}} \hat{\bm{\mu}}$ where $\hat{\bm{\mu}} \in \mathbb{R}^{K \times K}$ contains $\hat{\bm{\mu}}_1, \dots, \hat{\bm{\mu}}_K$ as its rows. Further, define $\delta = \sqrt{\frac{2K}{N}}$, and $S_k = \{v_i \in \mathcal{C}_k : \norm{\hat{\mathbf{x}}_i - \mathbf{x}_i} \geq \delta/2\}$. Then,
\begin{equation}
\label{eq:num_mistakes_bound}
\delta^2 \sum_{k = 1}^K \abs{S_k} \leq 8(2 + \epsilon) \norm{\mathbf{X} \mathbf{U}^\intercal - \mathbf{Y} \mathbf{Z}}[F][2].
\end{equation}
Moreover, if $\gamma$ from Lemma \ref{lemma:bound_on_eigenvector_diff} satisfies $\gamma^2 > \mathrm{const}(C, \alpha) (2 + \epsilon) p NK \ln N$ for a universal constant $\mathrm{const}(C, \alpha)$, there exists a permutation matrix $\mathbf{J} \in \mathbb{R}^{K \times K}$ such that
\begin{equation}
\label{eq:correct_solution_on_non-mistakes}
\hat{\bm{\theta}}_i^\intercal \mathbf{J} = \bm{\theta}_i^\intercal, \,\,\,\, \forall \,\, i \in [N] \backslash (\cup_{k=1}^K S_k).
\end{equation}
Here, $\hat{\bm{\theta}}_i \mathbf{J}$ and $\bm{\theta}_i$ represent the $i^{th}$ row of matrix $\hat{\bm{\Theta}}\mathbf{J}$ and $\bm{\Theta}$ respectively.
\end{lemma}
By the definition of $M(\bm{\Theta}, \hat{\bm{\Theta}})$, for the matrix $\mathbf{J}$ used in Lemma \ref{lemma:k_means_error}, $M(\bm{\Theta}, \hat{\bm{\Theta}}) \leq \frac{1}{N} \norm{\bm{\Theta} - \hat{\bm{\Theta}} \mathbf{J}}[0]$. But, according to Lemma \ref{lemma:k_means_error},
$\norm{\bm{\Theta} - \hat{\bm{\Theta}} \mathbf{J}}[0] \leq 2 \sum_{k = 1}^K \abs{S_k}$. Using Lemma \ref{lemma:bound_on_eigenvector_diff} and \ref{lemma:k_means_error}, we get:
\begin{eqnarray*}
M(\bm{\Theta}, \hat{\bm{\Theta}}) \leq \frac{1}{N} \norm{\bm{\Theta} - \hat{\bm{\Theta}} \mathbf{J}}[0] \leq \frac{2}{N} \sum_{k = 1}^K \abs{S_k}
&\leq& \frac{16(2 + \epsilon)}{N \delta^2} \norm{\mathbf{X} \mathbf{U}^\intercal - \mathbf{Y} \mathbf{Z}}[F][2] \\
&\leq& \mathrm{const}_3(C, \alpha)^2 \frac{512(2 + \epsilon)}{N \delta^2 \gamma^2} p N K \ln N.
\end{eqnarray*}
Noting that $\delta = \sqrt{\frac{2K}{N}}$ and setting $\mathrm{const}(C, \alpha) = 256 \times \mathrm{const}_3(C, \alpha)^2$ finishes the proof.
\subsubsection{Proof of Theorem \ref{theorem:consistency_result_normalized}}
\label{section:proof_consistency_normalized}
Recall that $\mathbf{Q} = \sqrt{\mathbf{Y}^\intercal \mathbf{D} \mathbf{Y}}$ and analogously define $\mathcal{Q} = \sqrt{\mathbf{Y}^\intercal \mathcal{D} \mathbf{Y}}$, where $\mathcal{D}$ is the expected degree matrix. It was shown after Lemma \ref{lemma:uk_eigenvector_of_tildeA} that $\mathcal{D} = (\lambda_1 - p) \mathbf{I}$. Thus, $\mathcal{Q} = \sqrt{\lambda_1 - p} \;\mathbf{I}$ as $\mathbf{Y}^\intercal \mathbf{Y} = \mathbf{I}$. Hence $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} = \frac{1}{\lambda_1 - p} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}$. Therefore, $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} \mathbf{x} = \frac{\lambda}{\lambda_1 - p} \mathbf{x} \; \Longleftrightarrow \; \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathbf{x} = \lambda \mathbf{x}$. Let $\mathcal{Z} \in \mathbb{R}^{N - r \times K}$ contain the leading $K$ eigenvectors of $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1}$ as its columns. Algorithm \ref{alg:nrepsc} will cluster the rows of $\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z}$ to recover the clusters in the expected case. As $\mathcal{Q}^{-1} =\frac{1}{\sqrt{\lambda_1 - p}} \mathbf{I}$, we have $\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z} = \frac{1}{\sqrt{\lambda_1 - p}} \mathbf{Y} \mathcal{Z}$. By Lemma \ref{lemma:first_K_eigenvectors_of_L}, $\mathcal{Z}$ can always be chosen such that $\mathbf{Y} \mathcal{Z} = \mathbf{X}$, where recall that $\mathbf{X} \in \mathbb{R}^{N \times K}$ has $\mathbf{y}_1, \dots, \mathbf{y}_K$ as its columns. Because the rows of $\mathbf{X}$ are identical for nodes that belong to the same cluster, Algorithm \ref{alg:nrepsc} returns the correct ground truth clusters in the expected case.
To bound the number of mistakes made by Algorithm \ref{alg:nrepsc}, we show that $\mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z}$ is close to $\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z}$. Here, $\mathbf{Z} \in \mathbb{R}^{N - r \times K}$ contains the top $K$ eigenvectors of $\mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}$. As in the proof of Lemma \ref{lemma:bound_on_eigenvector_diff}, we use Davis-Kahan theorem to bound this difference. This requires us to compute $\norm{\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} - \mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}}$. Note that:
\begin{align*}
\norm{\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} - \mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}} = &\norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}} \cdot \norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}} \cdot \norm{\mathcal{Q}^{-1}} + \\
&\norm{\mathbf{Q}^{-1}} \cdot \norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}} \cdot \norm{\mathcal{Q}^{-1}} + \\
&\norm{\mathbf{Q}^{-1}} \cdot \norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}} \cdot \norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}}.
\end{align*}
We already have a bound on $\norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}}$ in \eqref{eq:L-calL_bound}. Also, note that $\norm{\mathcal{Q}^{-1}} = \frac{1}{\sqrt{\lambda_1 - p}}$ as $\mathcal{Q}^{-1} = \frac{1}{\sqrt{\lambda_1 - p}} \mathbf{I}$. Similarly, as $\mathbf{Y}^\intercal \mathbf{Y} = \mathbf{I}$, $\norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}} \leq \norm{\mathcal{L}} = \lambda_1 - \bar{\lambda}$, where $\bar{\lambda} = \lambdamin{\tilde{\mathcal{A}}}$. Finally,
\begin{align*}
\norm{\mathbf{Q}^{-1}} &\leq \norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}} + \norm{\mathcal{Q}^{-1}} = \norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}} + \frac{1}{\sqrt{\lambda_1 - p}} \text{, and} \\
\norm{\mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}} &\leq \norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}} + \norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y}} = \norm{\mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} - \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y}} + \lambda_1 - \bar{\lambda}.
\end{align*}
Thus, to compute a bound on $\norm{\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} - \mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}}$, we only need a bound on $\norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}}$. The next lemma provides this bound.
\begin{lemma}
\label{lemma:calQ-Q_bound}
Let $\mathcal{Q} = \sqrt{\mathbf{Y}^\intercal \mathcal{D} \mathbf{Y}}$, $\mathbf{Q} = \sqrt{\mathbf{Y}^\intercal \mathbf{D} \mathbf{Y}}$, and assume that
$$\left(\frac{\sqrt{pN \ln N}}{\lambda_1 - p}\right)\left(\frac{\sqrt{pN \ln N}}{\lambda_1 - p} + \frac{1}{6\sqrt{C}} \right) \leq \frac{1}{16(\alpha + 1)},$$
where $C$ and $\alpha$ are used in $\mathrm{const}_1(C, \alpha)$ defined in Lemma \ref{lemma:bound_on_D-calD}. Then,
$$\norm{\mathcal{Q}^{-1} - \mathbf{Q}^{-1}} \leq \sqrt{\frac{2}{(\lambda_1 - p)^3}} \norm{\mathbf{D} - \mathcal{D}}.$$
\end{lemma}
\noindent
Using the lemma above with \eqref{eq:L-calL_bound}, we get
{
\small
\begin{align}
\label{eq:calQYcalLYcalQ-QYLYQ_bound}
\begin{aligned}
\norm{\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1} - \mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} &\mathbf{Y} \mathbf{Q}^{-1}} \leq \frac{2(\lambda_1 - \bar{\lambda})}{(\lambda_1 - p)^2} \left[ \sqrt{2} + \frac{\norm{\mathbf{D} - \mathcal{D}}}{\lambda_1 - p}\right] \norm{\mathbf{D} - \mathcal{D}} + \\
& \frac{\mathrm{const}_3(C, \alpha)}{\lambda_1 - p} \left[\frac{2\sqrt{2} \norm{\mathbf{D} - \mathcal{D}}}{\lambda_1 - p} + \frac{2 \norm{\mathbf{D} - \mathcal{D}}[][2]}{(\lambda_1 - p)^2} + 1 \right] \sqrt{p N \ln N}.
\end{aligned}
\end{align}
}
\noindent
The next lemma uses the bound above to show that $\mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z}$ is close to $\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z}$.
\begin{lemma}
\label{lemma:eigenvector_diff_bound_normalized}
Let $\mu_1 \leq \mu_2 \leq \dots \leq \mu_{N - r}$ be eigenvalues of $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1}$. Further, let the columns of $\mathcal{Z} \in \mathbb{R}^{N - r \times K}$ and $\mathbf{Z} \in \mathbb{R}^{N - r \times K}$ correspond to the leading $K$ eigenvectors of $\mathcal{Q}^{-1} \mathbf{Y}^\intercal \mathcal{L} \mathbf{Y} \mathcal{Q}^{-1}$ and $\mathbf{Q}^{-1} \mathbf{Y}^\intercal \mathbf{L} \mathbf{Y} \mathbf{Q}^{-1}$, respectively. Define $\gamma = \mu_{K + 1} - \mu_{K}$ and let there be a constant $\mathrm{const}_4(C, \alpha)$ such that
$\frac{\sqrt{p N \ln N}}{\lambda_1 - p} \leq \mathrm{const}_4(C, \alpha)$. Then, with probability at least $1 - 2N^{-\alpha}$, there exists a constant $\mathrm{const}_5(C, \alpha)$ such that
\begin{align*}
\inf_{\mathbf{U} : \mathbf{U}^\intercal \mathbf{U} = \mathbf{U}\bfU^\intercal = \mathbf{I}} \norm{\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z} - &\mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z} \mathbf{U}}[F] \leq \\
&\left[\frac{16 K \mathrm{const}_5(C, \alpha)}{\gamma (\lambda_1 - p)^{3/2}} + \frac{2 \mathrm{const}_1(C, \alpha) \sqrt{K}}{(\lambda_1 - p)^{3/2}} \right] \sqrt{p N \ln N},
\end{align*}
where $\mathrm{const}_1(C, \alpha)$ is defined in Lemma \ref{lemma:bound_on_D-calD}.
\end{lemma}
Recall that, by Lemma \ref{lemma:first_K_eigenvectors_of_L}, $\mathcal{Z}$ can always be chosen such that $\mathbf{Y} \mathcal{Z} = \mathbf{X}$, where $\mathbf{X}$ contains $\mathbf{y}_1, \dots, \mathbf{y}_K$ as its columns. As $\mathcal{Q}^{-1} = \frac{1}{\sqrt{\lambda_1 - p}} \mathbf{I}$, one can show that:
\begin{equation*}
\norm{(\mathcal{Q}^{-1} \mathbf{X})_i - (\mathcal{Q}^{-1} \mathbf{X})_j}[2] = \begin{cases}
0 & \text{ if } v_i \text{ and } v_j \text{ belong to the same cluster} \\
\sqrt{\frac{2K}{N(\lambda_1 - p)}} & \text{ otherwise.}
\end{cases}
\end{equation*}
Here, $(\mathcal{Q}^{-1} \mathbf{X})_i$ denotes the $i^{th}$ row of the matrix $\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z}$. Let $\mathbf{U}$ be the matrix that solves $\inf_{\mathbf{U} \in \mathbb{R}^{K \times K} : \mathbf{U}\bfU^\intercal = \mathbf{U}^\intercal \mathbf{U} = \mathbf{I}} \norm{\mathbf{Y} \mathcal{Q}^{-1} \mathcal{Z} - \mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z} \mathbf{U}}[F]$. As $\mathbf{U}$ is orthogonal, $\norm{(\mathcal{Q}^{-1} \mathbf{X})_i^\intercal \mathbf{U} - (\mathcal{Q}^{-1} \mathbf{X})_j^\intercal \mathbf{U}}[2] = \norm{(\mathcal{Q}^{-1} \mathbf{X})_i - (\mathcal{Q}^{-1} \mathbf{X})_j}[2]$. As in the previous case, the following lemma is a direct consequence of Lemma 5.3 in \citep{LeiEtAl:2015:ConsistencyOfSpectralClusteringInSBM}.
\begin{lemma}
\label{lemma:k_means_error_normalized}
Let $\mathbf{X}$ and $\mathbf{U}$ be as defined above. For any $\epsilon > 0$, let $\hat{\bm{\Theta}} \in \mathbb{R}^{N \times K}$ be the assignment matrix returned by a $(1 + \epsilon)$-approximate solution to the $k$-means clustering problem when rows of $\mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z}$ are provided as input features. Further, let $\hat{\bm{\mu}}_1$, $\hat{\bm{\mu}}_2$, \dots, $\hat{\bm{\mu}}_K \in \mathbb{R}^{K}$ be the estimated cluster centroids. Define $\hat{\mathbf{X}} = \hat{\bm{\Theta}} \hat{\bm{\mu}}$ where $\hat{\bm{\mu}} \in \mathbb{R}^{K \times K}$ contains $\hat{\bm{\mu}}_1, \dots, \hat{\bm{\mu}}_K$ as its rows. Further, define $\delta = \sqrt{\frac{2K}{N(\lambda_1 - p)}}$, and $S_k = \{v_i \in \mathcal{C}_k : \norm{\hat{\mathbf{x}}_i - \mathbf{x}_i} \geq \delta/2\}$. Then,
\begin{equation*}
\delta^2 \sum_{k = 1}^K \abs{S_k} \leq 8(2 + \epsilon) \norm{\mathbf{X} \mathbf{U}^\intercal - \mathbf{Y} \mathbf{Q}^{-1} \mathbf{Z}}[F][2].
\end{equation*}
Moreover, if $\gamma$ from Lemma \ref{lemma:eigenvector_diff_bound_normalized} satisfies $$16(2 + \epsilon)\left[ \frac{8 \mathrm{const}_5(C, \alpha) \sqrt{K}}{\gamma} + \mathrm{const}_1(C, \alpha)\right]^2 \frac{p N^2 \ln N}{(\lambda_1 - p)^2} < \frac{N}{K},$$ then, there exists a permutation matrix $\mathbf{J} \in \mathbb{R}^{K \times K}$ such that
\begin{equation*}
\hat{\bm{\theta}}_i^\intercal \mathbf{J} = \bm{\theta}_i^\intercal, \,\,\,\, \forall \,\, i \in [N] \backslash (\cup_{k=1}^K S_k).
\end{equation*}
Here, $\hat{\bm{\theta}}_i \mathbf{J}$ and $\bm{\theta}_i$ represent the $i^{th}$ row of matrix $\hat{\bm{\Theta}}\mathbf{J}$ and $\bm{\Theta}$ respectively.
\end{lemma}
The proof of Lemma \ref{lemma:k_means_error_normalized} is similar to that of Lemma \ref{lemma:k_means_error}, and has been omitted. The result follows by using a similar calculation as was done after Lemma \ref{lemma:k_means_error} in Section \ref{section:proof_consistency_unnormalized}.
\section{Numerical Results}
\label{section:numerical_results}
We perform three types of experiments. In the first two cases, we use synthetically generated data to validate our theoretical results using $d$-regular representation graphs (Section \ref{chapter:fairness:section:d_reg_experiments}) and non-$d$-regular representation graphs (Section \ref{chapter:fairness:section:sbm_experiments}). In the third case, we demonstrate the effectiveness of the proposed algorithms on
a real-world dataset (Section~\ref{chapter:fairness:section:trade_experiments}). Notably, our experiments in Sections \ref{chapter:fairness:section:sbm_experiments} and \ref{chapter:fairness:section:trade_experiments} demonstrate that the $d$-regularity assumption on $\mathcal{R}$ is not needed in practice. Before proceeding further, we mention two important details below: \textbf{(i)} How do we compare with the algorithms presented in \cite{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}? and \textbf{(ii)} What do we do when the rank assumption on $\mathcal{R}$ is not satisfied?
\begin{figure}[t]
\centering
\subfloat[][Accuracy vs no. of nodes]{\includegraphics[width=0.33\textwidth]{Images/U_d_reg_vs_N.pdf}\label{fig:d_reg_unnorm:vs_N}}%
\subfloat[][Accuracy vs no. of clusters]{\includegraphics[width=0.33\textwidth]{Images/U_d_reg_vs_K.pdf}\label{fig:d_reg_unnorm:vs_K}}%
\subfloat[][Accuracy vs degree of $\mathcal{R}$]{\includegraphics[width=0.33\textwidth]{Images/U_d_reg_vs_d.pdf}\label{fig:d_reg_unnorm:vs_d}}
\caption{Comparing \textsc{URepSC} with other ``unnormalized'' algorithms using synthetically generated $d$-regular representation graphs.}
\label{fig:d_reg_unnorm}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[][Accuracy vs no. of nodes]{\includegraphics[width=0.33\textwidth]{Images/N_d_reg_vs_N.pdf}\label{fig:d_reg_norm:vs_N}}%
\subfloat[][Accuracy vs no. of clusters]{\includegraphics[width=0.33\textwidth]{Images/N_d_reg_vs_K.pdf}\label{fig:d_reg_norm:vs_K}}%
\subfloat[][Accuracy vs degree of $\mathcal{R}$]{\includegraphics[width=0.33\textwidth]{Images/N_d_reg_vs_d.pdf}\label{fig:d_reg_norm:vs_d}}
\caption{Comparing \textsc{NRepSC} with other ``normalized'' algorithms using synthetically generated $d$-regular representation graphs.}
\label{fig:d_reg_norm}
\end{figure}
\paragraph*{Comparison with \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}} We refer to the algorithms proposed in \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints} as \textsc{UFairSC} and \textsc{NFairSC}, corresponding to the unnormalized and normalized fair spectral clustering, respectively. These algorithms assume that each node belongs to one of the $P$ protected groups $\mathcal{P}_1, \dots, \mathcal{P}_P \subseteq \mathcal{V}$ that are observed by the learner. Recall that these algorithms are special cases of our algorithms when $\mathcal{R}$ is block diagonal (Section \ref{section:constraint}). To demonstrate the generality of our algorithms, we only experiment with representation graphs that are not of the form specified above. Naturally, \textsc{UFairSC} and \textsc{NFairSC} are not directly applicable in this setting. Nonetheless, to compare with these algorithms, we approximate the protected groups by clustering the nodes in $\mathcal{R}$ using standard spectral clustering. Each discovered cluster is then treated as a protected group.
\paragraph*{Approximate \textsc{URepSC} and \textsc{NRepSC}} Recall that the rank assumption on $\mathcal{R}$ requires $\rank{\mathbf{R}} \leq N - K$. It is not possible to find $K$ orthonormal eigenvectors in Algorithms \ref{alg:urepsc} and \ref{alg:nrepsc} if $\mathbf{R}$ violates this assumption. Unlike other assumptions in our theoretical analysis, this assumption is necessary in practice. If a graph $\mathcal{R}$ violates the rank assumption, we instead use the best rank $R$ approximation of its adjacency matrix $\mathbf{R}$ ($R \leq N - K$). This approximation does not have binary elements, but it works well in practice. Whenever this approximation is used, we refer to \textsc{URepSC} and \textsc{NRepSC} as \textsc{URepSC (approx.)} and \textsc{NRepSC (approx.)}, respectively.
\begin{figure}[t]
\centering
\subfloat[][Unnormalized case]{\includegraphics[width=0.4\textwidth]{Images/U_d_reg_vs_rank_groups.pdf}\label{fig:d_reg_unnorm:rank_groups}}%
\hspace{1cm}\subfloat[][Normalized case]{\includegraphics[width=0.4\textwidth]{Images/N_d_reg_vs_rank_groups.pdf}\label{fig:d_reg_norm:rank_groups}}%
\caption{Accuracy vs the values of $P$ and $R$ used by \textsc{U/NFairSC} and \textsc{U/NRepSC}, respectively, for $d$-regular representation graphs.}
\end{figure}
\subsection{Experiments with $d$-regular representation graphs}
\label{chapter:fairness:section:d_reg_experiments}
For these experiments, we sampled $d$-regular representation graphs using $p=0.4$, $q=0.3$, $r=0.2$, and $s=0.1$, for various values of $d$, $N$, and $K$. We ensured that the sampled $\mathcal{R}$ satisfies Assumption \ref{assumption:R_is_d_regular} and $\rank{\mathbf{R}} \leq N - k$. Further, the ground-truth clusters have equal size and are representation-aware by construction as described in Section \ref{section:consistency_results}. Figure \ref{fig:d_reg_unnorm} compares the performance of \textsc{URepSC} with unnormalized spectral clustering (\textsc{USC}) (Algorithm \ref{alg:unnormalized_spectral_clustering}) and \textsc{UFairSC}. Figure \ref{fig:d_reg_unnorm:vs_N} shows the effect of varying $N$ for a fixed $d = 40$ and $K=5$. Figure \ref{fig:d_reg_unnorm:vs_K} varies $K$ and keeps $N = 1200$ and $d = 40$ fixed. Similarly, Figure \ref{fig:d_reg_unnorm:vs_d} keeps $N = 1200$ and $K = 5$ fixed and varies $d$. In all cases, we use $R = P = N/10$, where recall that $R$ is the rank used for approximation in \textsc{URepSC (approx.)} and $P$ is the number of protected groups discovered in $\mathcal{R}$ for running \textsc{UFairSC}. The figures plot the accuracy on $y$-axis and report the mean and standard deviation across $10$ independent executions of the algorithms in each case.
As the ground truth clusters satisfy Definition \ref{def:representation_constraint} by construction, a high accuracy of cluster recovery implies that the algorithm returns representation-aware clusters. Figure \ref{fig:d_reg_norm} shows the corresponding results for \textsc{NRepSC}, where we compare it with the normalized variants of other algorithms. In Figures \ref{fig:d_reg_unnorm:vs_N} and \ref{fig:d_reg_norm:vs_N}, it appears that even the standard spectral clustering algorithm will return representation-aware clusters for a large enough graph. However, Figures \ref{fig:d_reg_unnorm:vs_K} and \ref{fig:d_reg_norm:vs_K} show that this is not true if the number of clusters also increases with $N$, as is more common in practice.
It may also be tempting to think that \textsc{UFairSC} and \textsc{NFairSC} may perform well with a more carefully chosen value of $P$, the number of protected groups. However, Figures \ref{fig:d_reg_unnorm:rank_groups} and \ref{fig:d_reg_norm:rank_groups} show that this is not true. These figures plot the performance of \textsc{UFairSC} and \textsc{NFairSC} as a function of the number of protected groups $P$. Also shown in these plots is the performance of the approximate variants of our algorithms for various values of rank $R$. As expected, the accuracy increases with $R$ as the approximation of $\mathbf{R}$ becomes better.
\begin{figure}[t]
\centering
\subfloat[][$N = 1000$, $K = 4$]{\includegraphics[width=0.48\textwidth]{Images/U_AccVsGroupRankSBM_1000_4.pdf}}%
\hspace{0.5cm}\subfloat[][$N = 3000$, $K = 4$]{\includegraphics[width=0.48\textwidth]{Images/U_AccVsGroupRankSBM_3000_4}}
\subfloat[][$N = 1000$, $K = 8$]{\includegraphics[width=0.48\textwidth]{Images/U_AccVsGroupRankSBM_1000_8}}%
\hspace{0.5cm}\subfloat[][$N = 3000$, $K = 8$]{\includegraphics[width=0.48\textwidth]{Images/U_AccVsGroupRankSBM_3000_8}}
\caption{Comparing \textsc{URepSC (approx.)} with \textsc{UFairSC} using synthetically generated representation graphs sampled from an SBM.}
\label{fig:sbm_comparison_unnorm}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[][$N = 1000$, $K = 4$]{\includegraphics[width=0.48\textwidth]{Images/N_AccVsGroupRankSBM_1000_4.pdf}}%
\hspace{0.5cm}\subfloat[][$N = 3000$, $K = 4$]{\includegraphics[width=0.48\textwidth]{Images/N_AccVsGroupRankSBM_3000_4}}
\subfloat[][$N = 1000$, $K = 8$]{\includegraphics[width=0.48\textwidth]{Images/N_AccVsGroupRankSBM_1000_8}}%
\hspace{0.5cm}\subfloat[][$N = 3000$, $K = 8$]{\includegraphics[width=0.48\textwidth]{Images/N_AccVsGroupRankSBM_3000_8}}
\caption{Comparing \textsc{NRepSC (approx.)} with \textsc{NFairSC} using synthetically generated representation graphs sampled from an SBM.}
\label{fig:sbm_comparison_norm}
\end{figure}
\subsection{Experiments with representation graphs sampled from SBM}
\label{chapter:fairness:section:sbm_experiments}
In this case, we divide the nodes into $P = 5$ protected groups and sample a representation graph $\mathcal{R}$ using a Stochastic Block Model. Nodes in $\mathcal{R}$ are connected with probability $p_{\mathrm{in}} = 0.8$ (resp. $p_{\mathrm{out}} = 0.2$) if they belong to the same (resp. different) protected group(s). Conditioned on $\mathcal{R}$, we then sample an adjacency matrix from $\mathcal{R}$-SBM as before. As an $\mathcal{R}$ generated this way may violate the rank assumption, we only experiment with the approximate variants of \textsc{URepSC} and \textsc{NRepSC} in this case. Moreover, as such an $\mathcal{R}$ may not be $d$-regular, the notion of accuracy no longer conveys information about the representation awareness of an algorithm. Thus, we instead compute the individual balance $\rho_i$ with respect to each node, as defined in \eqref{eq:balance}.
Recall that $0 \leq \rho_i \leq 1$ and higher values indicate that the representatives of node $v_i$ are well spread out across clusters $\hat{\mathcal{C}}_1$, \dots, $\hat{\mathcal{C}}_K$. We use average balance $\bar{\rho} = \frac{1}{N} \sum_{i = 1}^N \rho_i$ to measure the representation-awareness of the clusters.
While average balance measures the representation awareness of the clusters, we also need to ensure that they have a high quality. Thus, we compute the ratio of the average balance to the ratio-cut objective. A high value indicates balanced clusters with a high quality (low ratio-cut score). Figure \ref{fig:sbm_comparison_unnorm} fixes the value of $P = 5$ and shows the variation of the metric described above on $y$-axis as a function of the number of protected groups used by \textsc{UFairSC} and rank $R$ used by \textsc{URepSC (approx.)}, for various values of $N$ and $K$. We used the same values of parameters $p$, $q$, $r$, and $s$, as in Section \ref{chapter:fairness:section:d_reg_experiments}. The plots in Figure \ref{fig:sbm_comparison_unnorm} show a trade-off between clustering accuracy and representation awareness. One can choose an appropriate value of $R$ and use \textsc{URepSC (approx.)} to get good quality clusters with a high balance. Figure \ref{fig:sbm_comparison_norm} presents analogous results for \textsc{NRepSC (approx.)}.
\begin{figure}[t]
\centering
\subfloat[][$K = 2$]{\includegraphics[width=0.48\textwidth]{Images/U_Trade_2}}%
\hspace{0.5cm}\subfloat[][$K = 4$]{\includegraphics[width=0.48\textwidth]{Images/U_Trade_4}}
\subfloat[][$K = 6$]{\includegraphics[width=0.48\textwidth]{Images/U_Trade_6}}%
\hspace{0.5cm}\subfloat[][$K = 8$]{\includegraphics[width=0.48\textwidth]{Images/U_Trade_8}}
\caption{Comparing \textsc{URepSC (approx.)} with \textsc{UFairSC} on FAO trade network.}
\label{fig:real_data_comparison_unnorm}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[][$K = 2$]{\includegraphics[width=0.48\textwidth]{Images/N_Trade_2}}%
\hspace{0.5cm}\subfloat[][$K = 4$]{\includegraphics[width=0.48\textwidth]{Images/N_Trade_4}}
\subfloat[][$K = 6$]{\includegraphics[width=0.48\textwidth]{Images/N_Trade_6}}%
\hspace{0.5cm}\subfloat[][$K = 8$]{\includegraphics[width=0.48\textwidth]{Images/N_Trade_8}}
\caption{Comparing \textsc{NRepSC (approx.)} with \textsc{NFairSC} on FAO trade network.}
\label{fig:real_data_comparison_norm}
\end{figure}
\subsection{Experiments with a real-world network}
\label{chapter:fairness:section:trade_experiments}
For the final set of experiments, we use the FAO trade network \citep{DomenicoEtAl:2015:StructuralReducibilityOfMultilayerNetworks}, which is a multiplex network based on the data made available by the Food and Agriculture Organization (FAO) of the United Nations. It has $214$ nodes representing countries and $364$ layers corresponding to commodities like coffee, banana, barley, etc. An edge between two countries in a layer indicates the volume of the corresponding commodity traded between these countries. We convert the weighted graph in each layer to an unweighted graph by connecting every node with its five nearest neighbors. We then make all the edges undirected and use the first $182$ layers to construct the representation graph $\mathcal{R}$. Nodes in $\mathcal{R}$ are connected if they are linked in either of these layers. Similarly, the next $182$ layers are used to construct the similarity graph $\mathcal{G}$. Note that $\mathcal{R}$ constructed this way is not $d$-regular. The goal is to find clusters in $\mathcal{G}$ that satisfy Definition \ref{def:representation_constraint} with respect to $\mathcal{R}$.
To motivate this further, note that clusters based only on $\mathcal{G}$ only consider the trade of commodities $183$--$364$. However, countries also have other trade relations in $\mathcal{R}$, leading to shared economic interests. Assume that the members of each cluster would jointly formulate the economic policies for that cluster. However, the policies made in one cluster affect everyone, even if they are not part of the cluster, as they all share a global market. This incentivizes the countries to influence the economic policies of all the clusters. Being representation aware with respect to $\mathcal{R}$ entails that each country has members in other clusters with shared interests. This enables a country to indirectly shape the policies of other clusters.
As before, we use the low-rank approximation for the representation graph in \textsc{URepSC (approx.)} and \textsc{NRepSC (approx.)}. Figure \ref{fig:real_data_comparison_unnorm} compares \textsc{URepSC (approx.)} with \textsc{UFairSC}, and has the same semantics as Figure \ref{fig:sbm_comparison_unnorm}. Different plots in Figure \ref{fig:real_data_comparison_unnorm} correspond to different choices of $K$. \textsc{URepSC (approx.)} achieves a higher ratio of average balance to ratio-cut. In practice, a user would choose $R$ by assessing the relative importance of a quality metric like ratio-cut and representation metric like average balance. Figure \ref{fig:real_data_comparison_norm} presents analogous results for \textsc{NRepSC (approx.)}.
\section{Conclusion}
\label{section:conclusion}
The primary focus of this work has been on studying the consistency of constrained spectral clustering under an individual-level representation constraint. The proposed representation constraint naturally generalizes similar population-level constraints \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets} by using auxiliary information encoded in a representation graph $\mathcal{R}$. We showed that the constraint can be expressed as a linear expression that when added to the optimization problem solved by spectral clustering results in the representation-aware variants of the algorithm. An interesting consequence of this problem is a variant of the stochastic block model that plants the properties of $\mathcal{R}$ in a similarity graph in addition to the given clusters to provide a hard problem instance to our algorithms. Under this model, we derive a high-probability upper bound on the number of mistakes made by the algorithms and establish conditions under which they are weakly consistent. To the best of our knowledge, these are the first consistency results for constrained spectral clustering under individual-level constraints. Next, we make a few additional remarks.
\paragraph*{The $d$-regularity assumption} The $d$-regularity assumption on $\mathcal{R}$ ensures the representation-awareness of the ground-truth clusters in our analysis. Note that the representation graph that recovers the statistical-level constraint is also $d$-regular (see Appendix \ref{appendix:constraint}), hence our analysis strictly generalizes the previously known results. It would be interesting to study the performance of our algorithms under weaker assumptions on $\mathcal{R}$. One can also use a similar strategy as ours to modify more expressive variants of the stochastic block model, like the degree-corrected SBM \citep{KarrerNewman:2011:StochasticBlockmodelsAndCommunityStructureInNetworks}, to establish the consistency of the algorithms on more realistic similarity graphs.
\paragraph*{Computational complexity} Our current approach involves finding the null space of a $N \times N$ matrix. This operation has $O(N^3)$ complexity. Existing methods for speeding up the standard spectral clustering algorithm focus on making the eigen-decomposition and/or the $k$-means step faster (see \citet{TremblayEtAl:2016:CompressiveSpectralClustering} and the references within). However, even with these modifications, the null space computation would still be the computationally dominant step. \citet{XuEtAl:2009:FastNormalizedCutWithLinearConstraints} proposed an efficient algorithm for solving the normalized cut problem under a linear constraint. However, their algorithm assumes that $K = 2$. Developing similar algorithms for general values of $K$ and exploring their theoretical guarantees will enable the application domains that involve very large graphs to utilise our ideas.
Other possible extensions of our work include similar algorithms for weighted similarity graphs, overlapping clusters, and other types of graphs such as hypergraphs. This paper provides the first step towards consistency analysis of spectral clustering under individual-level constraints.
\begin{appendix}
\section{Representation constraint: Additional details}
\label{appendix:constraint}
In this section, we make two additional remarks about the properties of the proposed constraint, both in the context of fairness.
\paragraph*{Statistical fairness as a special case} Recall that our constraint specifies an individual fairness notion. Contrast this with several existing approaches that assign each node to one of the $P$ \textit{protected groups} $\mathcal{P}_1, \dots, \mathcal{P}_P \subseteq \mathcal{V}$ \citep{ChierichettiEtAl:2017:FairClusteringThroughFairlets}, and require these protected groups to have a proportional representation in all clusters, i.e.,
\begin{equation*}
\frac{\abs{\mathcal{P}_i \cap \mathcal{C}_j}}{\abs{\mathcal{C}_j}} = \frac{\abs{\mathcal{P}_i}}{N}, \,\, \forall i \in [P],\,\, j \in [K].
\end{equation*}
This is an example of \textit{statistical fairness}. In Example \ref{example:statistical_vs_individual_fairness}, we argued that statistical fairness may not be enough in some cases. We now show that the constraint in Definition \ref{def:representation_constraint} is equivalent to a statistical fairness notion for an appropriately constructed representation graph $\mathcal{R}$ from the given protected groups $\mathcal{P}_1, \dots, \mathcal{P}_P$. Namely, let $\mathcal{R}$ be such that $R_{ij} = 1$ if and only if $v_i$ and $v_j$ belong to the same protected group. In this case, it is easy to verify that the constraint in Definition \ref{def:representation_constraint} reduces to the statistical fairness criterion given above. In general, for other configurations of the representation graph, we strictly generalize the statistical fairness notion. We also strictly generalize the approach presented in \citet{KleindessnerEtAl:2019:GuaranteesForSpectralClusteringWithFairnessConstraints}, where the authors use spectral clustering to produce statistically fair clusters. Also noteworthy is the assumption made by statistical fairness, namely that every pair of vertices in a protected group can represent each others' interests ($R_{ij} = 1 \Leftrightarrow v_i$ and $v_j$ are in the same protected group), or they are very similar with respect to some sensitive attributes. This assumption becomes unreasonable as protected groups grow in size.
\paragraph*{Sensitive attributes and protected groups} Viewed as a fairness notion, the proposed constraint only requires a representation graph $\mathcal{R}$. It has two advantages over existing fairness criteria: \textbf{(i)} it does not require observable sensitive attributes (such as age, gender, and sexual orientation), and \textbf{(ii)} even if sensitive attributes are provided, one need not specify the number of protected groups or explicitly compute them. This ensures data privacy and helps against individual profiling. Our constraint only requires access to the representation graph $\mathcal{R}$. This graph can either be directly elicited from the individuals or derived as a function of several sensitive attributes. In either case, once $\mathcal{R}$ is available, we no longer need to expose any sensitive attributes to the clustering algorithm. For example, individuals in $\mathcal{R}$ may be connected if their age difference is less than five years and if they went to the same school. Crucially, the sensitive attributes used to construct $\mathcal{R}$ may be numerical, binary, categorical, etc.
\end{appendix}
\bibliographystyle{plainnat}
| a5d0710afa5c0b16a9aefb184bb0ec06dbdad698 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this paper, we consider the problem of \emph{point counting} for
principally polarized (p.p.)~abelian varieties over finite fields:
given a p.p.~abelian variety~$A$ over~$\mathbb{F}_q$, we aim to compute the
characteristic polynomial of Frobenius~$\chi(A)\in \mathbb{Z}[X]$. If~$A$ is
an elliptic curve, this is equivalent to computing~$\# A(\mathbb{F}_q)$.
The motivation behind this challenge comes from different
directions. Counting points is a prerequisite for elliptic and
hyperelliptic-curve
cryptography~\cite{koblitz_EllipticCurveCryptosystems1987,
koblitz_HyperellipticCryptosystems1989}. More recently, the hardness
of the point counting problem itself was proposed as a source of
cryptographic
protocols~\cite{dobson_TrustlessUnknownorderGroups2021}. From a
more mathematical point of view, if~$A$ is defined over a number
field, then counting points on~$A$ modulo primes of good reduction
determines the Euler factors of the $L$-function attached to~$A$.
To this date, Schoof's polynomial-time
algorithm~\cite{schoof_EllipticCurvesFinite1985,pila_FrobeniusMapsAbelian1990}
remains the central approach to point counting for abelian varieties
of dimension~$2$ or more over finite fields of large characteristic,
and much work has been devoted to making this algorithm
practical~\cite{gaudry_CountingPointsHyperelliptic2000,
gaudry_CountingPointsGenus2011, gaudry_GenusPointCounting2012,
abelard_ImprovedComplexityBounds2019}. In the case of abelian
surfaces over~$\mathbb{F}_q$, the complexity of Schoof's method
is~$\Otilde(\log^8q)$ binary operations in general,
and~$\Otilde(\log^5q)$ binary operations if the abelian surfaces
have explicit real multiplication~(RM) by a fixed quadratic
field. Note however that Schoof's approach is in competition with
cohomological algorithms, surveyed
in~\cite{kedlaya_ComputingZetaFunctions2004}, when the base field has
small characteristic; in the context of computing $L$-functions, it is
in competition with average polynomial time algorithms based on
Hasse-Witt matrices~\cite{harvey_CountingPointsHyperelliptic2014,
harvey_ComputingHasseWittMatrices2016,
sutherland_CountingPointsSuperelliptic2020}.
Schoof's approach is multi-modular: for a series of small
primes~$\ell \neq p$, the reduction of~$\chi(A)$ mod~$\ell$ is
computed as the characteristic polynomial of Frobenius on
the~$\ell$-torsion subgroup~$A[\ell]$, the latter being defined by
explicit polynomial equations. The algorithm stops when sufficient
information is collected to reconstruct~$\chi(A)$ using the Weil
bounds and the Chinese remainder theorem.
In the case of elliptic curves,
Elkies~\cite{elkies_EllipticModularCurves1998} showed how to
accelerate Schoof's algorithm by replacing~$A[\ell]$, the kernel of
the \emph{endomorphism}~$[\ell]$, by the kernel of
an~\emph{isogeny}~$\phi\colon A\to A'$ of degree~$\ell$. Such an
isogeny will exist as soon as~$\chi(A)$ splits in linear factors
modulo~$\ell$; we say that~$\ell$ is~\emph{Elkies} in this case.
Heuristically, about half of the small primes are Elkies for a
given~$A$; this heuristic is true on average, either for all elliptic
curves over a given finite
field~\cite{shparlinski_DistributionAtkinElkies2014}, or, assuming
GRH, for all reductions of a given elliptic curve over a number field
modulo primes of good
reduction~\cite{shparlinski_DistributionAtkinElkies2015}. Then the
resulting point counting algorithm will run in~$\Otilde(\log^4q)$
binary operations, instead of~$\Otilde(\log^5q)$. Elkies's method is
an essential part of the SEA
algorithm~\cite{schoof_CountingPointsElliptic1995}, implemented in
both Pari/GP~\cite{theparigroup_PariGPVersion2019} and
Magma~\cite{bosma_MagmaAlgebraSystem1997}.
The central player in Elkies's method is the classical modular
polynomial~$\Phi_\ell$ of level~$\ell$, an explicit polynomial
equation cutting out the moduli space of pairs of~$\ell$-isogenous
elliptic curves. More precisely, Elkies's method relies on three main
ingredients: first, upper bounds on the degree and height
of~$\Phi_\ell$~\cite{cohen_CoefficientsTransformationPolynomials1984,
broker_ExplicitHeightBound2010}; second, an evaluation
algorithm, to compute~$\Phi_\ell(j, Y)$ as well as its derivative
$\partial_X \Phi_\ell(j, X)$ for a given value
of~$j\in\mathbb{F}_q$~\cite{enge_ComputingModularPolynomials2009,
broker_ModularPolynomialsIsogeny2012,
sutherland_EvaluationModularPolynomials2013}; and third, an isogeny
algorithm to recover~$\phi$ as an explicit rational map from this
data~\cite{elkies_EllipticModularCurves1998,
bostan_FastAlgorithmsComputing2008}.
The recent series of papers~\cite{kieffer_DegreeHeightEstimates2022,
kieffer_EvaluatingModularEquations2021,
kieffer_ComputingIsogeniesModular2019} extend all these three
ingredients to the context of p.p.~abelian surfaces. In this setting,
the classical modular polynomials are replaced by modular equations
for abelian surfaces, as described
in~\cite{broker_ModularPolynomialsGenus2009,
milio_QuasilinearTimeAlgorithm2015,
martindale_HilbertModularPolynomials2020,
milio_ModularPolynomialsHilbert2020}. Here we reap the benefits of
these works and describe their consequences on the point counting
problem under heuristics related to the distribution of Elkies
primes. We separate two cases depending on the moduli space of abelian
surfaces we wish to consider. In the \emph{Siegel case}, we assume
nothing a priori on our abelian surfaces; in the \emph{Hilbert case},
we fix a real quadratic field~$F$ and we only consider p.p.~abelian
surfaces with RM by its ring of integers~$\mathbb{Z}_F$ (but the action
of~$\mathbb{Z}_F$ is not assumed to be explicitly computable). In the Hilbert
case, we reach the same asymptotic complexity as the SEA algorithm up
to constant factors depending on~$F$.
\begin{thm}
\label{thm:main-hilbert}
Let~$F$ be a real quadratic field, and let~$\varepsilon>0$. Then there
exists an algorithm which, given a prime power~$q = p^r$
with~$r = o(\log p)$, and given the Igusa invariants of a
p.p.~abelian surface~$A$ over~$\mathbb{F}_q$ with real multiplication
by~$\mathbb{Z}_F$ for which a proportion~$\varepsilon$ of primes are Elkies
(see \cref{def:elkies-hilbert}), computes~$\chi(A)\in \mathbb{Z}[X]$ in
$\Otilde_{F,\varepsilon}(\log^4 q)$ binary operations.
\end{thm}
In the Siegel case, it turns out that Elkies's method brings no
complexity improvement (except perhaps for logarithmic factors) over
Schoof's method for a general abelian surface~$A$
over~$\mathbb{F}_q$. However, it does bring an improvement when the invariants
of~$A$ admit lifts in characteristic zero of small heights. The
exponent in the complexity estimate is further decreased if we wish to
count points modulo sufficiently many primes at once.
\begin{thm}
\label{thm:main-siegel}
Let~$K$ be a number field, and let~$\varepsilon>0$. Then:
\begin{enumerate}
\item There exists an algorithm which, given~$H\geq 0$, given a
p.p.~abelian surface~$A$ over~$K$ whose Igusa invariants are
well-defined and have height at most~$H$, and given a prime
ideal~$\mathfrak{p}$ of~$K$ of norm~$q$ such that~$A$ has good reduction
at~$\mathfrak{p}$ and a proportion~$\varepsilon$ of primes are Elkies for its
reduction~$A_\mathfrak{p}$ (see \cref{def:elkies-siegel}),
computes~$\chi({A_\mathfrak{p}})\in \mathbb{Z}[X]$ in $\Otilde_{K,\varepsilon}(H \log^7 q)$ binary
operations.
\item There exists an algorithm which, given~$H\geq 0$ and
$q\geq 1$, given a p.p.~abelian surface~$A$ over~$K$, and
given~$\Theta(H\log q)$ many distinct primes $\mathfrak{p}_1,\ldots,\mathfrak{p}_n$
of~$K$ such that~$\log N(\mathfrak{p}_i) = O(\log q)$, such that~$A$ has
good reduction at all primes~$\mathfrak{p}_i$, and such that a
proportion~$\varepsilon$ of primes are Elkies for each of its
reductions~$A_{\mathfrak{p}_i}$, computes all characteristic
polynomials~$\chi({A_{\mathfrak{p}_i}})\in \mathbb{Z}[X]$
using~$\Otilde_{K,\varepsilon}(\log^6 q)$ binary operations on average
for each~$i$.
\end{enumerate}
\end{thm}
We have released an implementation of the key step of the above
algorithms in terms of running time, namely the evaluation of modular
equations~\cite{kieffer_HDMELibraryEvaluation}, building on the C
libraries Flint~\cite{hart_FLINTFastLibrary} and
Arb~\cite{johansson_ArbEfficientArbitraryprecision2017}. This allows
us to roughly estimate the total cost of the above point-counting
algorithms in practice.
As remarked
in~\cite[§4.2.1]{dobson_TrustlessUnknownorderGroups2021},
dimension~$2$ is the largest dimension where Elkies's method can be
superior to Schoof's algorithm for generic abelian varieties, at least
in the asymptotic sense. However, Elkies's method still seems
promising in the context of counting points on p.p.~abelian varieties
with fixed RM in any dimension.
This paper is organized as follows. In \cref{sec:background}, we
quickly review previous results on Schoof's method for abelian
surfaces. In \cref{sec:siegel,sec:hilbert}, we describe Elkies's
method for p.p.~abelian surfaces in the Siegel and Hilbert case
respectively. Experimental results appear in \cref{sec:exp}. Finally,
\cref{sec:implem} presents possible directions to further reduce the
cost of point counting for abelian surfaces in practice.
\subsection*{Acknowledgements} I am deeply indebted to his former
advisors Damien Robert and Aurel Page for suggesting the thesis
project that led to this work. I also thank Noam Elkies, John Voight
and Andrew Sutherland for their insightful comments on this
work. Finally, I thank the LMFDB team for allowing me to access their
computational resources.
\section{Background on point-counting algorithms}
\label{sec:background}
\subsection{The characteristic polynomial of Frobenius}
Let~$A$ be a p.p.~abelian surface over~$\mathbb{F}_q$, and denote its
Frobenius endomorphism by~$\pi_A$. The
characteristic polynomial~$\chi(A)$ of~$\pi_A$ takes the form
\begin{equation}
\label{eq:charpoly-schoof}
\chi(A) = X^4 - s_1X^3 + (s_2+2q) X^2 - q s_1 X + q^2,
\end{equation}
where~$s_1,s_2$ are integers satisfying the following
inequalities~\cite{weil_CourbesAlgebriquesVarietes1948},
\cite[Lem.~3.1]{ruck_AbelianSurfacesJacobian1990}:
\begin{equation}
\label{eq:weil-rueck}
\abs{s_1}\leq 4\sqrt{q},\quad \abs{s_2}\leq 4q,
\quad s_1^2 - 4s_2\geq 0,\quad s_2 + 4q \geq 2\abs{s_1}.
\end{equation}
Denote the Rosati involution on~$\End(A)$ induced by the principal
polarization of~$A$ by~$\dagger$. Then the \emph{real
Frobenius}~$\psi_A = \pi_A+\ros{\pi_A}$ is an element of the
subgroup~$\Rend(A)$ of real endomorphisms of~$A$. Its characteristic
polynomial is
\begin{equation}
\label{eq:charpoly-rm}
\xi(A) = X^2 - s_1X + s_2.
\end{equation}
If~$\ell\neq q$ is a prime, then~$A[\ell] \subset A$ is a finite étale
group scheme isomorphic to~$(\mathbb{Z}/\ell\mathbb{Z})^4$, and we identify it with
its set of points over an algebraic closure of~$\mathbb{F}_q$. The reduction
of~$\chi(A)$ modulo~$\ell$ is the characteristic polynomial of~$\pi_A$
acting on~$A[\ell]$. Recall that~$A[\ell]$ is endowed with the Weil
pairing, an alternating and nondegenerate bilinear form induced by the
principal polarization of~$A$; we denote it by
$(x,y)\mapsto \pair{x,y} \in \mathbb{Z}/\ell\mathbb{Z}$. For all~$x,y\in A[\ell]$, we
have
\begin{equation}
\label{eq:frob-pairing}
\pair{\pi_A(x),\pi_A(y)} = q\pair{x,y}.
\end{equation}
The Rosati involution is equal to adjunction with respect to the Weil
pairing, so~\eqref{eq:frob-pairing} translates to the equality
$\pi_A\ros{\pi_A} = q$.
Assume now that~$A$ has RM by~$\mathbb{Z}_F$, where~$F$ is a real quadratic
field; this means that~$A$ is equipped with an
embedding~$\mathbb{Z}_F\hookrightarrow \Rend(A)$. Let~$\ell\in\mathbb{Z}$ be a prime which
splits in~$F$ in a product of two principal ideals, generated
by~$\beta,\conj{\beta}\in \mathbb{Z}_F$. We then have an orthogonal
decomposition
\begin{equation}
\label{eq:rm-splitting}
A[\ell] = A[\beta]\oplus A[\conj{\beta}],
\end{equation}
and both~$A[\beta]$ and~$A[\conj{\beta}]$ are stable
under~$\pi_A$,~$\ros{\pi_A}$ and~$\psi_A$. Since~$A[\beta]$
and~$A[\conj{\beta}]$ are not isotropic, the determinant of~$\pi_A$ on
both of these subspaces is~$q$.
\subsection{Schoof's method in dimension 2}
\label{subsec:schoof}
Recall that any p.p.~abelian surface~$A$ over~$k = \mathbb{F}_q$ is either a
product of two elliptic curves or the Jacobian of a hyperelliptic
genus~$2$ curve~$\mathcal{C}$ defined over~$k$. In the point-counting
context, we only have to consider this second case.
The fundamental building block of Schoof's
method~\cite{gaudry_CountingPointsHyperelliptic2000,
gaudry_GenusPointCounting2012} is to be able to work with the
torsion subgroups~$A[\ell]$ in a computationally efficient way. Since
Elkies's method ultimately involves computations with subgroups of
Jacobians as well, the computational techniques developed for Schoof's
method will still apply in our context.
The first step is to choose birational coordinates on~$A$. A popular
choice is to consider Mumford coordinates. Let~$y^2 = P(x)$ be an
equation of~$\mathcal{C}$, and let~$K_\mathcal{C}$ be the canonical divisor
of~$\mathcal{C}$. A generic point of~$A$ is linearly equivalent to~$D-K_\mathcal{C}$
for a unique degree-two divisor~$D$ on~$\mathcal{C}$; in turn, a generic
divisor of degree two can be written as the zero locus of polynomials
of the form~$x^2 + u_1 x + u_0$ and~$y - v_1x - v_0$ in a unique
way. This defines the Mumford coordinates~$(u_0,u_1,v_0,v_1)$ as a
rational map from~$A$ to the affine space~$\mathbb{A}_k^4$, and~$A$ is
birational to its image. Denote the coordinate ring of~$\mathbb{A}_k^4$
by~$k[U_0,U_1,V_0,V_1]$.
If~$S$ is any finite subgroup of~$A$, then~$S$ is stable
by~$D\mapsto i(D)$ where~$i(D)$ denotes the hyperelliptic involution,
and hence by change of sign of~$v$-coordinates. Assume that~$S$ is
\emph{generic} in the sense that Mumford coordinates are well-defined
at all points of~$S$ and that all pairs~$\set{D,i(D)}$ in~$S$ have
distinct~$u_1$-coordinates. Then the Gröbner basis cutting out~$S$ in
terms of Mumford coordinates, in the monomial
ordering~$U_1 < U_0 < V_1 < V_0$, will take the convenient form
\begin{equation}
\label{eq:groebner-basis}
\begin{cases}
V_0 - V_1 S_0(U_1) = 0,\\
V_1^2 - S_1(U_1) = 0, \\
U_0 - R_0(U_1) = 0, \\
R_1(U_1) = 0
\end{cases}
\end{equation}
for some univariate polynomials~$R_1,R_0,S_1,S_0\in k[U_1]$. The
methods of~\cite[§3]{gaudry_GenusPointCounting2012} describe how to
compute this Gröbner basis when $S = A[\ell]$, assuming that the
equation of~$\mathcal{C}$ is given by a polynomial~$P$ of degree five, in
other words that~$\mathcal{C}$ admits a rational Weierstrass
point~$\infty$. The input of their algorithm is given by Cantor's
division polynomials, which provide an explicit description of the
following composition as a rational map:
\begin{displaymath}
\begin{tikzcd}
\mathcal{C} \ar[rr, "{p \mapsto [p]-[\infty]}"] && A \ar[r, "{[\ell]}"] &
A \ar[rr, "{(u_0,u_1,v_0,v_1)}"] && \mathbb{A}^4.
\end{tikzcd}
\end{displaymath}
More generally, if~$A'$ is the Jacobian of another hyperelliptic
genus~$2$ curve~$\mathcal{C}'$ over~$k$, is~$f\colon A\to A'$ is any isogeny,
and if~$p_0\in \mathcal{C}(k)$ is any rational point, then the same methods
will compute a Gröbner basis describing~$\ker(f)\subset A$ given the
explicit expression of the composed map
\begin{equation}
\label{eq:rational-map}
\begin{tikzcd}
\mathcal{C} \ar[rr, "{p \mapsto [p]-[p_0]}"] && A \ar[r, "{f}"] &
A' \ar[rr, "{(u_0,u_1,v_0,v_1)}"] && \mathbb{A}^4.
\end{tikzcd}
\end{equation}
If the degrees of these rational fractions is bounded above by~$d$,
then the whole Gröbner basis computation takes~$\Otilde(d^3)$
operations in~$k$, hence~$\Otilde(d^3\log q)$ binary operations. Note
that a complexity~$\Otilde(d^{3-1/\omega} \log q)$, where~$\omega$
denotes the exponent of matrix multiplication, could probably be
achieved by computing bivariate resultants using an algorithm of
Villard~\cite{villard_ComputingResultantGeneric2018a} instead of the
more classical evaluation-interpolation method. The resulting
polynomials in~\eqref{eq:groebner-basis} have degree~$O(d^2)$.
Once the Gröbner basis~\eqref{eq:groebner-basis} is known, computing
the Frobenius endomorphism on~$S$ is simply a matter of
computing~$(U_0^q,U_1^q,V_0^q,V_1^q)$ using a square-and-multiply
algorithm, reducing the result modulo the defining ideal of~$S$ at
each step, for a total cost of~$\Otilde(d^2\log^2 q)$ binary
operations.
When running Schoof's method in the generic case, one takes $S = A[\ell]$ and
$d = O(\ell^2)$. It only remains to find the correct values of~$s_1$
and~$s_2$ in~$\mathbb{Z}/\ell\mathbb{Z}$ such that Frobenius characteristic
equation~\eqref{eq:charpoly-schoof} holds on~$A[\ell]$
\cite[Alg.~1]{gaudry_GenusPointCounting2012}. The dominant step in the
whole method is the Gröbner basis computation, which accounts for the
final complexity of~$\Otilde(\log^8q)$ binary operations. In the RM
case, one can take~$S = A[\beta]$ instead provided that~$\ell$ splits
correctly in~$\mathbb{Z}_F$~\cite{gaudry_CountingPointsGenus2011}. Then one
has~$d = O(\ell)$, giving a total point-counting complexity
of~$\Otilde(\log^5 q)$ binary operations; both the Gröbner step and
the Frobenius computation are asymptotically
dominant. In~\cite{gaudry_CountingPointsGenus2011}, the real
multiplication action of~$\mathbb{Z}_F$ is assumed to be explicitly
computable; in this case the Chinese remainder theorem can be used to
recover~$\psi_A\in \mathbb{Z}_F$ directly.
\subsection{Non-generic cases}
The above genericity assumption on~$S$ does not necessarily hold in
general. As detailed
in~\cite[§5]{abelard_ImprovedComplexityBounds2019}, it could fail in a
finite number of different ways: certain elements of~$S$ might be of
the form~$[p]-[\infty]$ or $2[p]-K$ for some~$p\in \mathcal{C}(k)$, so that
their Mumford coordinates are not defined; or they might take the
generic form~$[p_1] + [p_2] - K$, but the Mumford coordinates
of~$\ell([p_1]-[\infty])$ might not be defined. Each of these possible
degeneracy types can be managed by writing another polynomial system
with a smaller number of variables or lower degrees than the original
one; therefore, considering the generic case is sufficient from a
complexity-theoretic point of view. It is also sufficient from a
practical point of view: in large characteristics,~$S$ will be generic
with overwhelming probability, and any particular~$\ell$ causing
problems can simply be skipped.
\section{Elkies's method for abelian surfaces: the Siegel case}
\label{sec:siegel}
\subsection{Polarized isogenies between abelian surfaces}
Modular polynomials describing~$\ell$-isogenies between elliptic
curves play a central role in Elkies's method in
dimension~$1$. Similarly, explicit equations for moduli spaces of
suitably isogenous abelian surfaces will play a central role in
Elkies's method for p.p.~abelian surfaces. Such moduli spaces only
exist for certain isogeny types that are directly related to the
polarizations and endomorphism rings of the abelian varieties we
consider.
Recall that~$\NS(A)$ denotes the Néron--Severi group of~$A$,
consisting of line bundles on~$A$ defined over an algebraic closure of
the base field, up to algebraic equivalence.
\begin{thm}[{\cite[Prop.~17.2]{milne_AbelianVarieties1986}}]
\label{thm:End-NS}
Let~$A$ be a principally polarized abelian variety over a
field~$k$. Then there is a natural isomorphism of abelian groups
\begin{displaymath}
\L_A \colon (\Rend(A), +) \to (\NS(A), \otimes).
\end{displaymath}
The line bundle~$\L_A(1)$ is the unique line bundle (up to algebraic
equivalence) defining the principal polarization of~$A$; moreover,
for each~$\beta\in \Rend(A)$, the line bundle~$\L_A(\beta)$ is ample
if and only if~$\beta$ is totally positive.
\end{thm}
If~$A$ and~$A'$ are p.p.~abelian varieties of the same dimension
over~$k$, and if $\beta\in \Rend(A)$, we say that $f\colon A\to A'$ is
a~\emph{$\beta$-isogeny} if $f^*\L_{A'}(1) = \L_A(\beta)$. In
particular,~$\beta$ must be totally positive. If~$\beta$ is moreover
prime to~$\chr(k)$, an isogeny~$f$ is a~$\beta$-isogeny if and only if
the following conditions are
satisfied~\cite[Thm.~1.1]{dudeanu_CyclicIsogeniesAbelian2017}:
\begin{enumerate}
\item \label{it:isotropic} $\ker(f)\subset A[\beta]$ and~$\ker(f)$ is
maximal isotropic for the Weil pairing attached to this subgroup;
\item The image~$A'$ is endowed with the natural principal
polarization of~$A/\ker(f)$ coming from the
conditions~\eqref{it:isotropic}.
\end{enumerate}
Since~$\Rend(A)$ always contains a copy of~$\mathbb{Z}$, it makes sense to
talk about~$\ell$-isogenies for any~$\ell\in \mathbb{Z}_{\geq 1}$. In the case
of elliptic curves, this corresponds to the usual notion of cyclic
isogenies of degree~$\ell$; but in dimension~$g$ the degree of
an~$\ell$-isogeny is~$\ell^g$. In general, the degree of
a~$\beta$-isogeny is $(\deg\beta)^{1/2}$.
We say that two isogenies~$f_1\colon A\to A'$ and~$f_2\colon A\to A''$
are~\emph{equivalent} if there exists an isomorphism of p.p.~abelian
varieties~$\eta\colon A'\to A''$ such that~$f_2 = \eta\circ f_1$.
\subsection{Elkies primes}
The goal of Elkies's method in the Siegel case is, given a
p.p.~abelian surface~$A$ over~$\mathbb{F}_q$ and a prime~$\ell$, to obtain
information on~$\chi(A)\mod\ell$ using an~$\ell$-isogeny~$f$, with
domain~$A$, defined over~$\mathbb{F}_p$. We say that~$\ell$ is~\emph{Elkies}
for~$A$ if such an~$f$ exists. Before explaining how to obtain such
an~$f$, let us describe how~$\chi(A)$ can be computed from the action
of~$\pi_A$ on~$\ker(f)$; and conversely, how the splitting behavior
of~$\chi(A)$ modulo~$\ell$ can guarantee that~$\ell$ is Elkies.
If~$P$ is a monic polynomial of degree~$d$ whose constant
coefficient~$a_0$ is invertible, we denote by~$\Rec_q(P)$ the monic
polynomial~$a_0^{-1} X^d P(q/X)$.
\begin{prop}
\label{prop:siegel-elkies-charpoly}
Assume that~$\ell$ is Elkies for~$A$, and let~$f\colon A\to A'$ be
an~$\ell$-isogeny defined over~$\mathbb{F}_q$. Let~$P\in \mathbb{Z}/\ell\mathbb{Z}[X]$ be
the characteristic polynomial of~$\pi_A$ on~$\ker(f)$.
Then~$\chi(A) = P \Rec_q(P)$ modulo~$\ell$.
\end{prop}
\begin{proof}
Choose a symplectic basis of~$A[\ell]$ whose first two vectors
generate~$\ker(f)$. By~\eqref{eq:frob-pairing}, the matrix
of~$\pi_A$ in this basis is of the form
\begin{displaymath}
\mat{M}{*}{0}{qM^{-t}}
\end{displaymath}
for some~$M\in \GL_2(\mathbb{Z}/\ell\mathbb{Z})$; here~$M^{-t}$ denotes the inverse
transpose of~$M$. The characteristic polynomial of~$q M^{-t}$
is~$\Rec_q(P)$.
\end{proof}
\begin{prop}
\label{prop:siegel-elkies-sufficient}
Let~$\ell$ be a prime, and assume that one of the following holds:
\begin{enumerate}
\item \label{it:coprime-split} $\chi(A)$ splits modulo~$\ell$ as a
product of the form~$P\Rec_q(P)$ where the polynomials~$P$
and~$\Rec_q(P)$ are coprime;
\item \label{it:total-split} $\chi(A)$ is totally split modulo~$\ell$.
\end{enumerate}
Then~$\ell$ is Elkies for~$A$.
\end{prop}
Recall that the roots of~$\chi(A)$ over~$\mathbb{C}$ take the
form~$\lambda_1,q/\lambda_1,\lambda_2,q/\lambda_2$
where~$\lambda_1,\lambda_2$ are complex numbers of modulus~$\sqrt{q}$;
hence assumption~\eqref{it:coprime-split} means that~$\chi(A)$ splits
modulo~$\ell$ in two coprime degree~$2$ factors whose roots
are~$\set{\lambda_1,\lambda_2}$ and~$\set{q/\lambda_1,q/\lambda_2}$
respectively (up to a possible renaming
of~$q/\lambda_2\mapsto \lambda_2$). Merely assuming that~$\chi(A)$
splits in degree~$2$ factors is not sufficient to ensure that~$\ell$
is Elkies: for instance,~$A$ might be product of two elliptic curves
over~$\mathbb{F}_q$ for which~$\ell$ is an Atkin prime.
\begin{proof}
In case~\eqref{it:coprime-split}, define~$a = P(\pi_A)$
and~$b = \Rec_q(P)(\pi_A)$ as endomorphisms of~$A[\ell]$. We have a
decomposition of~$A[\ell]$ as~$\ker(a)\oplus \ker(b)$, and both
subspaces have dimension~2. Let us show that~$\ker(a)$ is
isotropic: by~\eqref{eq:frob-pairing},~$b$ is the adjoint of~$a$,
hence
\begin{displaymath}
\pair{\ker(a), \ker(a)} = \pair{ \im(b), \im(b)}
= \pair{A[\ell],\im (a b)} = 0.
\end{displaymath}
In case~\ref{it:total-split}, if~$v\in A[\ell]$ is an eigenvector
of~$\pi_A$, then~$v^\perp \subset A[\ell]$ is still~$\pi_A$-stable.
Therefore, there exists~$w\in v^\perp$ such
that~$\gen{v}\oplus \gen{w}\subset A[\ell]$ is a $\pi_A$-stable
subspace of dimension~$2$; it is isotropic by construction.
\end{proof}
Given \cref{prop:siegel-elkies-sufficient}, it is reasonable to expect
that a positive proportion of small primes~$\ell$ (in general, about
three in eight) will be Elkies for a given~$A$. This heuristic will
motivate the more precise \cref{def:elkies-siegel} later on.
\subsection{Modular equations of Siegel type}
\label{subsec:siegel-modeq}
Denote by~$\mathcal{A}_g$ the Siegel moduli space of p.p.~abelian varieties of
dimension~$g$, considered as an algebraic variety over~$\mathbb{Q}$. Then for
each prime~$\ell$, we have the following diagram of~$\mathbb{Q}$-varieties:
\begin{equation}
\label{diag:siegel}
\begin{tikzcd}
& \mathcal{A}_g^0(\ell) \ar[ld, "p_1", swap] \ar[rd, "p_2"] & \\
\mathcal{A}_g & & \mathcal{A}_g
\end{tikzcd}
\end{equation}
where~$\mathcal{A}_g^0(\ell)$ denotes the coarse moduli space of pairs~$(A,K)$
where~$A$ is a p.p.~abelian variety of dimension~$g$
and~$K\subset A[\ell]$ is the kernel of an~$\ell$-isogeny. The
morphisms~$p_1$ and~$p_2$ are~$(A,K)\mapsto A$ and~$(A,K)\mapsto A/K$
respectively. Both~$p_1$ and~$p_2$ and are finite coverings;
moreover~$(p_1,p_2)$ realizes a birational isomorphism
between~$\mathcal{A}_g^0(\ell)$ and its image in~$\mathcal{A}_g\times\mathcal{A}_g$.
If~$g=2$, the graded~$\mathbb{Q}$-algebra of Siegel modular forms is free over
four
generators~$I_4,I_6',I_{10},I_{12}$~\cite{igusa_ArithmeticVarietyModuli1960}.
Therefore~$\mathcal{A}_2$ is a rational variety. The zero locus of~$I_{10}$
exactly corresponds to the locus of products of elliptic curves;
moreover, it is computationally convenient to work with coordinates
on~$\mathcal{A}_2$ which share a common, small-degree denominator, so a common
choice of coordinates~$j = (j_1,j_2,j_3)$ on~$\mathcal{A}_g$ is given by the
\emph{Igusa invariants}~\cite{streng_ComputingIgusaClass2014}, which
are scalar multiples of
\begin{equation}
\label{eq:igusa}
\frac{I_4 I_6'}{I_{10}}, \quad \frac{I_4 I_{12}}{I_{10}^2}, \quad \text{and}\quad
\frac{I_4^5}{I_{10}^2}.
\end{equation}
The Igusa invariants define a local isomorphism from~$\mathcal{A}_2$
to~$\mathbb{A}^3$ at every point where both~$I_4$ and~$I_{10}$ are
nonzero. Hitting this singular locus may cause problems during the
point-counting algorithm. In this case, one can always consider
another set of coordinates on~$\mathcal{A}_2$; we postpone this
discussion to~§\ref{subsec:siegel-degenerate} below.
The \emph{Siegel modular equations} of level~$\ell$ are explicit
equations for the image of~$\mathcal{A}_2^0(\ell)$ in~$\mathcal{A}_2\times \mathcal{A}_2$. They
take the form of three multivariate rational
fractions~$\Psi_{\ell,k}\in \mathbb{Q}(J_1,J_2,J_3)[Y]$ for~$1\leq k\leq
3$. Writing~$j\circ p_1 = (j_1,j_2,j_3)$
and~$j\circ p_2 = (j_1',j_2',j_3')$, the equations of~$\mathcal{A}_2^0(\ell)$
are the following:
\begin{equation}
\label{eq:siegel-modeq}
\begin{cases}
\Psi_{\ell,1}(j_1,j_2,j_3,j_1') = 0,\\
\partial_X \Psi_{\ell,1}(j_1,j_2,j_3,j_1')\cdot j_k' -
\Psi_{\ell,k}(j_1,j_2,j_3,j_1') = 0
\quad \text{for $k = 2,3$}.
\end{cases}
\end{equation}
Let~$\dd(\ell) = \ell^3+\ell^2+\ell+1$ be the degree
of~$p_1\colon\mathcal{A}_2^0(\ell)\to\mathcal{A}_2$. Then, for any p.p.~abelian
surface~$A$ over~$\mathbb{C}$, there are exactly~$d(\ell)$
non-equivalent~$\ell$-isogenies with domain~$A$. Let~$A'_i$
for~$1\leq i\leq \dd(\ell)$ be their codomains. Assume that
neither~$A$ nor any of the isogenous surfaces~$A_i'$ lie in the
singular locus of~$j$ as defined above; assume moreover that all
coordinates~$j_1(A_i')$ are distinct. Then the denominator of Siegel
modular equations does not vanish at~$j(A)$; moreover the roots of
Siegel modular equations evaluated at~$A$, by which we mean all
tuples~$(j_1',j_2',j_3')\in \mathbb{C}^3$ such that the
equations~\eqref{eq:siegel-modeq} are satisfied with~$j_k = j_k(A)$
for~$1\leq k\leq 3$, are precisely the~$d(\ell)$ tuples of the
form~$j(A_i')$ for~$1\leq i\leq d(\ell)$. The same properties holds if
we replace~$\mathbb{C}$ by an algebraic closure of the finite field~$\mathbb{F}_p$ for
any prime~$p\neq \ell$ such that~$p>3$; this can be deduced either
from the classical lifting theorems, or from the fact that both~$\mathcal{A}_2$
and~$\mathcal{A}_2^0(\ell)$ are actually smooth stacks
over~$\mathbb{Z}[1/\ell]$~\cite[Chap.~I,~§4]{faltings_DegenerationAbelianVarieties1990}.
\subsection{Algorithms for Siegel modular equations}
We now present the prerequisites of Elkies's method in the Siegel
case, namely an upper bound on the size of modular equations of Siegel
type, as well as algorithms that allow us to evaluate modular
equations at a given point and compute the associated
$\ell$-isogenies.
In the following result, the \emph{total degree} of a
multivariate rational fraction~$F$ over~$\mathbb{Q}$ means the maximum between
the total degrees of its numerator and denomiator; moreover, the
\emph{height} of~$F$ is the maximum of~$\log\abs{c}$, where~$c\in \mathbb{Z}$
runs through the coefficients of~$F$ written in irreducible form. The
notion of height generalizes to arbitrary number fields.
\begin{thm}[{\cite[Thm.~1.1 and
Prop.~4.11]{kieffer_DegreeHeightEstimates2022}}]
\label{thm:height-siegel} The degree of~$\Psi_{\ell,k}$ in~$X$
equals~$\dd(\ell)$ for~$k=1$, and equals~$\dd(\ell)-1$
for~$k = 2,3$. The total degree of~$\Psi_{\ell,k}$ in~$J_1,J_2,J_3$
is bounded above by~$5\dd(\ell)/3$ for~$k=1$, and~$10\dd(\ell)/3$
if~$k=2,3$. The height of~$\Psi_{\ell,k}$ is~$O(\ell^3\log \ell)$
as~$\ell$ tends to infinity.
\end{thm}
\begin{thm}[{\cite{kieffer_EvaluatingModularEquations2021}}]
\label{thm:eval-siegel}
Let~$K$ be a fixed number field. There exists an algorithm which,
given a prime~$\ell$, given~$H\geq 0$, and
given~$(j_1,j_2,j_3)\in K^3$ where the denominator of Siegel modular
equations of level~$\ell$ does not vanish, computes the polynomials
\begin{displaymath}
\Psi_{\ell,k}(j_1,j_2,j_3,X) \quad\text{and}\quad
\partial_{J_i} \Psi_{\ell,k}(j_1,j_2,j_3,X)
\end{displaymath}
for all indices~$1\leq i,k\leq 3$ as elements of~$K[X]$ in quasi-linear
time, in other words in~$\Otilde_K(\ell^6 H)$ binary operations.
\end{thm}
\begin{thm}[{\cite[Thm.~1.1]{kieffer_ComputingIsogeniesModular2019}}]
\label{thm:isog-siegel}
Let~$k$ be a field. Then there exists an algorithm which, given:
\begin{itemize}
\item a prime~$\ell$ such that~$\chr(k) > 8\ell +7$ if it is finite;
\item the Igusa invariants of two~$\ell$-isogenous p.p.~abelian
varieties~$A$ and~$A'$ defined over~$k$, lying outside the
singular locus of~$j$, such that~$\Aut(A)$ and~$\Aut(A')$ are both
equal to~$\set{\pm 1}$, and such that the subvariety
of~$\mathbb{A}^3\times\mathbb{A}^3$ cut out by the
equations~\eqref{eq:siegel-modeq} is normal at~$(j(A),j(A'))$;
\item the nine values
\begin{displaymath}
\partial_{J_i}\Psi_{\ell,k}(j_1(A),j_2(A),j_3(A),j_1(A'))\in k
\end{displaymath}
for~$1\leq i, k\leq 3$;
\end{itemize}
computes the following data:
\begin{itemize}
\item a tower~$k'/k$ of at most three quadratic extensions;
\item equations for two genus~$2$ hyperelliptic curves~$\mathcal{C}$
and~$\mathcal{C}'$ over~$k'$ whose Jacobians are isomorphic to $A$
and~$A'$ respectively;
\item a point~$p_0\in \mathcal{C}(k')$; and
\item four rational fractions in~$k'(x,y)$ of total degree~$O(\ell)$
describing an~$\ell$-isogeny $f\colon A\to A'$ in the sense
of~\eqref{eq:rational-map} using~$p_0\in \mathcal{C}(k')$ as a base
point;
\end{itemize}
for the cost of~$\Otilde(\ell)$ elementary operations and~$O(1)$
square roots in~$k'$.
\end{thm}
\subsection{Complexity bounds for Elkies's method}
\label{subsec:siegel-complexity}
Let~$K$ be a fixed number field, and let~$A_0$ be a p.p.~abelian
surface over~$K$. Let~$\mathfrak{p}$ be a prime of~$K$, of residue field~$\mathbb{F}_q$,
where~$A_0$ has good reduction; we denote the reduced abelian variety
over~$\mathbb{F}_q$ by~$A$. In order to count points on~$A$, we apply
Elkies's method in the following way.
\begin{enumerate}
\item For a series of small primes~$\ell$, we apply
\cref{thm:eval-siegel} to evaluate the Siegel modular equations
at~$j(A_0)$, and reduce the result to~$\mathbb{F}_q$. This step
costs~$\Otilde_K(\ell^6H)$ binary operations, and will dominate the
rest of the algorithm; however, it needs to be done only once if we
with to count points on~$A$ modulo~$\mathfrak{p}_i$ for several primes~$\mathfrak{p}_i$.
\item We then attempt to find a root of the reduced Siegel modular
equations~\eqref{eq:siegel-modeq} over~$\mathbb{F}_q$; this
costs~$\Otilde(\ell^3\log^2 q)$ binary operations. If there are
none, we simply skip~$\ell$.
\item If we find one, then we hope that it corresponds to the Igusa
invariants of a p.p.~abelian surface~$A'$ over~$\mathbb{F}_q$ for which the
genericity conditions of \cref{thm:isog-siegel} hold. If they do,
then~$\ell$ is Elkies for~$A$, and we are able to compute an
explicit rational representation of such an~$f$
using~$\Otilde(\ell\log q)$ binary operations.
\item At this point, the methods described in~§\ref{subsec:schoof}
allow us to compute with formal points of~$\ker(f)$, and hence
recover the characteristic polynomial of Frobenius on this subgroup;
the Gröbner basis and Frobenius computations
cost~$\Otilde(\ell^3\log q)$ and~$\Otilde(\ell^2\log^2 q)$ binary
operations respectively. The polynomial~$\chi(A)\mod\ell$ itself is
finally computed using \Cref{prop:siegel-elkies-charpoly}.
\end{enumerate}
By the Hasse-Weil bounds, carrying out this algorithm successfully for
a series of primes~$\ell_i$ such that~$\prod\ell_i > 8q$ is sufficient
to recover~$\chi(A)\in \mathbb{Z}[X]$ using the Chinese remainder theorem.
\subsection{Degenerate cases}
\label{subsec:siegel-degenerate}
We now analyze the different failure cases of the algorithm sketched
above. The following issues may arise for any Elkies prime~$\ell$:
\begin{enumerate}
\item \label{pb:singular-locus} One or more of the p.p.~abelian
surfaces~$\ell$-isogenous to~$A$ over an algebraic closure of~$\mathbb{F}_q$
may lie on the singular locus of~$j$.
\item \label{pb:same-j1} Several of these abelian surfaces may have
the same~$j_1$-coordinate.
\item \label{pb:elliptic} Either~$A$ or~$A'$ may be the product of two
elliptic curves.
\item \label{pb:automorphisms} Either~$A$ or~$A'$ may have extra automorphisms.
\item \label{pb:not-normal} The subvariety of~$\mathbb{A}^3\times\mathbb{A}^3$ cut
out by the Siegel modular equations of level~$\ell$ may not be
normal at~$(j(A),j(A'))$.
\end{enumerate}
Both failure cases~\eqref{pb:singular-locus} and~\eqref{pb:same-j1}
can be detected during the execution of the algorithm of
\cref{thm:eval-siegel}; they are easily solved by taking different
birational coordinates on~$\mathcal{A}_2$. For instance, we can apply a
projective linear transformation~$m$ with integer coefficients on the
weight~$20$
coordinates~$(I_4^5 : I_4^2 I_6'^2 : I_4^2 I_{12} : I_4 I_6' I_{10} :
I_{10}^2)$ before taking the quotients~\eqref{eq:igusa}. When
choosing~$m$, we must make sure that~$O(\ell^6)$ non-equalities in the
algebraic closure of~$\mathbb{F}_q$ are satisfied. This can always be achieved
provided that we choose coefficients in~$m$ of height~$O(\log
\ell)$. The degree and height bounds of \cref{thm:height-siegel} still
hold for the modified modular equations using this new set of
invariants.
In case~\eqref{pb:elliptic}, we can apply the SEA algorithm on both
factors.
In case~\eqref{pb:automorphisms}, we obtain a lot of new information
about~$A$: either~$A$ is a twist of the Jacobian of the hyperelliptic
curve~$y^2 = x^5-1$ with complex multiplication by~$\mathbb{Q}(\zeta_5)$, so
that~$\chi(A)$ can be determined by the CM
method~\cite{weng_ConstructingHyperellipticCurves2003}; or we can find
an explicit isogeny from~$A$ to the product of two elliptic
curves~\cite[§8]{igusa_ArithmeticVarietyModuli1960}.
Finally, in case~\eqref{pb:not-normal}, a geometric argument shows
that~$(A,A')$ must be the reduction to~$\mathbb{F}_q$ of a singular point in
characteristic
zero~\cite[Rem.~4.12]{kieffer_ComputingIsogeniesModular2019}. Using
the complex-analytic uniformization of~$\mathcal{A}_2$ then shows that such
singular points either admit extraneous automorphisms, or have the
property that there exist two non-equivalent~$\ell$-isogenies from~$A$
to~$A'$. This implies that~$A$ possesses a non-integral endomorphism
of norm~$\ell^4$. Unlike the elliptic curve case, where we would
switch to the CM method straighaway, this new piece of information
seems insufficient to describe~$\End(A)$ in a precise way in higher
dimensions. Instead, we simply skip~$\ell$ and carry on with Elkies's
method for other primes; we make the heuristic assumption that
sufficiently many Elkies primes still exist.
\begin{defn}
\label{def:elkies-siegel}
Let~$\varepsilon>0$, and let~$A$ be a p.p.~abelian surface over~$\mathbb{F}_q$. We
say that~$A$ has \emph{a proportion~$\varepsilon$ of primes are Elkies
for~$A$} if the following holds: for
every~$X \geq \frac{1}{\varepsilon}\log q$, the proportion of
primes~$\ell\leq X$ such that~$\ell$ is Elkies for~$A$ and~$\End(A)$
admits no non-integral endomorphism of norm~$\ell^4$ is at
least~$\varepsilon$.
\end{defn}
If a proportion~$\varepsilon$ of primes are Elkies for~$A$, then we are able
to collect sufficiently many Elkies primes~$\ell_i$ such
that~$\ell_i\in O_\varepsilon(\log q)$. Thus, \Cref{thm:main-siegel}
directly follows from the complexity estimates given
in~§\ref{subsec:siegel-complexity}.
\begin{rem}
It is known that for any fixed~$\varepsilon>0$, there exists an elliptic
curve~$E$ over some finite field~$\mathbb{F}_q$ for which it does not hold
that a proportion~$\varepsilon$ of primes are Elkies
for~$E$~\cite{shparlinski_ProductSmallElkies2015}.
Following~\cite{sutherland_EvaluationModularPolynomials2013}, we
could relax \cref{def:elkies-siegel} by taking an upper bounds of
the form $X = \frac1\varepsilon\log q\log\log(q)^n$ for some
fixed~$n\geq 1$. Then we can hope that there exists a
positive~$\varepsilon>0$ such that all abelian surfaces over finite fields
have a proportion~$\varepsilon$ of Elkies primes. This more permissive
definition does not modify the complexity estimates of
\cref{thm:main-siegel}.
\end{rem}
\section{Elkies's method for abelian surfaces: the Hilbert case}
\label{sec:hilbert}
\subsection{Elkies primes}
In this section, we fix a real quadratic field~$F$ of
discriminant~$\Delta_F$, and we consider the point counting problem
for a p.p.~abelian surface~$A$ over~$\mathbb{F}_q$ with real multiplication
by~$\mathbb{Z}_F$. By \cref{thm:End-NS}, this real multiplication structure
guarantees the presence of supplementary isogenies compared to the
Siegel case: for each totally positive~$\beta\in \mathbb{Z}_F$, we can look
for~$\beta$-isogenies $f\colon A\to A'$ defined over~$\mathbb{F}_q$. Assume
further that~$\beta\in \mathbb{Z}_F$ is prime, prime to~$\Delta_F$, and
that~$\ell = N_{F/\mathbb{Q}}(\beta)\in \mathbb{Z}$ is also prime; in other
words~$\ell$ is a prime that splits in~$F$ in a product of two
ideals~$(\beta)\cdot(\conj{\beta})$ that are trivial in the narrow
class group of~$\mathbb{Z}_F$. By the \v{C}ebotarev density theorem, this kind of
splitting will occur for a positive proportion of
primes~$\ell\in\mathbb{Z}$. We say that~$\beta$ is \emph{Elkies} for~$A$ if
a~$\beta$-isogeny~$f$ with domain~$A$ exists
over~$\mathbb{F}_q$. Then~$\ker(f)\subset A[\ell]$ is a~$\pi_A$-stable
subgroup of order~$\ell$; therefore we can hope to obtain information
on~$\chi(A)$ mod~$\ell$ by manipulating polynomials of
degree~$O(\ell)$ only, as in Elkies's original method for elliptic curves.
In the Hilbert case, the Chinese remaindering step is formulated in
terms of the real Frobenius endomorphism~$\psi_A = \pi_A+\ros{\pi_A}$
as an element of~$\mathbb{Z}_F$. By the Weil bounds~\eqref{eq:weil-rueck}, we
have
\begin{equation}
\label{eq:weilbounds-rm}
|\Tr(\psi_A)|\leq 4\sqrt{q} \quad \text{and}\quad \Disc(\mathbb{Z}[\psi_A])\leq 4q.
\end{equation}
Assume that~$\beta$ is Elkies, and the
eigenvalue~$\lambda\in \mathbb{Z}/\ell\mathbb{Z}$ of~$\pi_A$ on~$\ker(f)$ has been
computed. Then we have
\begin{displaymath}
\psi_A = \lambda + q/\lambda \mod\beta
\end{displaymath}
under the canonical isomorphism~$\mathbb{Z}_F/\beta\mathbb{Z}_F\simeq \mathbb{Z}/\ell\mathbb{Z}$.
In the algorithm, we consider a series totally positive Elkies
primes~$\beta_i$ in $\mathbb{Z}_F$, with norms~$\ell_i\in \mathbb{Z}$. We collect the
values of $\psi_A$ modulo~$\beta_i$ as elements of $\mathbb{Z}/\ell_i\mathbb{Z}$ as
above. The Chinese remainder theorem in~$\mathbb{Z}_F$ allows us to
reconstruct the value of $\psi_A$ modulo the ideal
$\id{B} = \prod_i (\beta_i)$. The cost of this reconstruction is
negligible when compared to the rest of the algorithm.
\begin{prop}
\label{prop:crt-hilbert}
Assume that $N(\id{B})> 16q$. Then $\psi_A$ is uniquely determined by
equations~\eqref{eq:weilbounds-rm} and the data of $\psi_A\mod{\id{B}}$.
\end{prop}
\begin{proof}
Assume that~$\id{B}$ contains a nonzero~$\alpha\in\mathbb{Z}_F$ such
that $|\Tr_{F/\mathbb{Q}}(\alpha)|\leq 8\sqrt{q}$ and
$\Disc(\mathbb{Z}[\alpha])\leq 16q$. Then we have
\begin{displaymath}
N(\id{B})\leq |N_{F/\mathbb{Q}}(\alpha)| = \tfrac{1}{4}|\Tr(\alpha)^2 - \Disc(\mathbb{Z}[\alpha])| \leq 16q.
\qedhere
\end{displaymath}
\end{proof}
Once~$\psi_A\in \mathbb{Z}_F$ has been determined, its characteristic
polynomial completely describes~$\chi(A)$, as
equations~\eqref{eq:charpoly-schoof} and~\eqref{eq:charpoly-rm}
show. Heuristically, we expect that roughly half of the suitable
primes~$\beta$ will be Elkies: indeed~$\beta$ is Elkies if and only if
the characteristic polynomial of~$\pi_A$ on~$A[\beta]$, a polynomial
of degree~$2$, splits in~$\mathbb{Z}/\ell\mathbb{Z}$.
\subsection{Modular equations of Hilbert type} The key fact that
allows us to reach similar point-counting complexities in the RM case
and the case of elliptic curves is that the associated \emph{Hilbert
modular equations} have a reasonable size.
Denote by~$\mathcal{A}_{2,F}$ the Hilbert moduli space of p.p.~abelian surfaces
with real multiplication by~$\mathbb{Z}_F$, seen as an algebraic variety
over~$\mathbb{Q}$. For each~$\beta$ as above, we have a diagram
of~$\mathbb{Q}$-varieties
\begin{equation}
\label{diag:hilbert}
\begin{tikzcd}
& \mathcal{A}_{2,F}^0(\beta) \ar[ld, "p_1", swap] \ar[rd, "p_2"] & \\
\mathcal{A}_{2,F} & & \mathcal{A}_{2,F}
\end{tikzcd}
\end{equation}
where~$\mathcal{A}_{2,F}^0(\beta)$ denotes the coarse moduli space of
pairs~$(A,K)$ where~$A$ is a p.p.~abelian surface with real
multiplication by~$\mathbb{Z}_F$, and~$K\subset A[\beta]$ is the kernel of
a~$\beta$-isogeny. The Hilbert modular equations of level~$\beta$ are
explicit equations for the image of~$\mathcal{A}_{2,F}^0(\beta)$
in~$\mathcal{A}_{2,F}\times\mathcal{A}_{2,F}$. To define them, we make a choice of
coordinates~$j = (j_1,j_2,j_3)$ on~$\mathcal{A}_{2,F}$, related by an explicit
equation of the form
\begin{displaymath}
E(j_1,j_2,j_3) = 0.
\end{displaymath}
Assume further that~$j_1$ and~$j_2$ are algebraically independent, and
write~$e = \deg_{j_3}(E)$. The Hilbert modular equations are then the
data of the three multivariate rational
fractions~$\Psi_{\beta,k}\in \mathbb{Q}(J_1,J_2)[J_3,X]$ of degree at
most~$e-1$ in~$J_3$ such that the system of
equations~\eqref{eq:siegel-modeq} holds with~$\ell$ replaced
by~$\beta$.
In concrete cases, it is sometimes convenient to modify this
definition and consider \emph{symmetric} modular equations on Humbert,
rather than Hilbert, surfaces. Recall that there exists a forgetful
map~$\mathcal{A}_{2,F}\to \mathcal{A}_2$ which is generically~$2$-$1$. The image~$\H_F$
of~$\mathcal{A}_{2,F}$, called the Humbert surface attached to~$\mathbb{Z}_F$, is often
less geometrically complicated than~$\mathcal{A}_{2,F}$. Explicit coordinates
on~$\H_F$ are also easier to describe: for instance, the Igusa
invariants~\eqref{eq:igusa} are always a valid choice. If the
discriminant of~$F$ is less than~$100$, then~$\H_F$ is rational, and
explicit parametrizations appear
in~\cite{elkies_K3SurfacesEquations2014}. In the
case~$F = \mathbb{Q}(\sqrt{5})$, the Gundlach invariants denoted by~$g_1,g_2$
(see \cite[Satz~6]{gundlach_BestimmungFunktionenZur1963}, although
other normalizations are also used) are convenient coordinates
on~$\H_F$ derived from an explicit description of the associated
graded ring of symmetric Hilbert modular forms.
We will denote the (symmetric) Hilbert modular equations in Igusa
invariants by~$\Psi_{\beta,k}^J$ for~$1\leq k\leq 3$; they are equal
for the prime~$\beta$ and its real
conjugate~$\conj{\beta}$. Similarly, we denote the Hilbert modular
equations of level~$\beta$ in Gundlach invariants
for~$F = \mathbb{Q}(\sqrt{5})$ by~$\Psi_{\beta,k}^G$ for~$k = 1,2$; they are
multivariate rational fractions in~$\mathbb{Q}(G_1,G_2)[Y]$. Modular
equations on Hilbert surfaces describe~$\beta$-isogenies between
abelian surfaces with~RM by~$\mathbb{Z}_F$, in a similar way as
in~§\ref{subsec:siegel-modeq} for modular equations of Siegel type. In
the symmetric case, modular equations describe~$\beta$-
and~$\conj{\beta}$-isogenies simultaneously.
\subsection{Algorithms for Hilbert modular equations}
Let~$\dd(\beta) = \ell +1$ be the degree of~$p_1$ in
diagram~\eqref{diag:hilbert}. We fix a choice of
coordinates~$(j_1,j_2,j_3)$ on~$\mathcal{A}_{2,F}$.
\begin{thm}[{\cite[Thm.~1.1 and
Prop.~4.13]{kieffer_DegreeHeightEstimates2022}}]
\label{thm:height-hilbert}
The degree of~$\Psi_{\beta,k}$ in~$X$ is~$\dd(\beta)$ for~$k=1$,
and~$\dd(\beta-1)$ for~$k>1$. The total degrees of~$\Psi_{\beta,k}$
are~$O_F(\ell)$, and their heights are~$O_F(\ell \log \ell)$. In
the case of $F = \mathbb{Q}(\sqrt{5})$, the total degree of~$\Psi_{\beta,k}^G$
in~$G_1,G_2$ is at most~$10 \dd(\beta)/3$ for~$k = 1,2$.
\end{thm}
\begin{thm}[{\cite[Thm.~5.3]{kieffer_EvaluatingModularEquations2021}}]
\label{thm:eval-hilbert}
Let~$q = p^r$ be a prime power, and let~$F = \mathbb{Q}(\sqrt{5})$. There
exists an algorithm which, given~$(g_1,g_2)\in \mathbb{F}_q^2$ where the
denominator of~$\Psi_{\beta,k}^G$ for~$k = 1,2$ does not vanish,
computes the modular equations~$\smash{\Psi_{\beta,k}^G}(g_1,g_2,X)$
as well as their
derivatives~$\partial_{G_i} \Psi_{\beta,k}^G(g_1,g_2,X)$
for~$1\leq i,k\leq 2$ as elements of~$\mathbb{F}_q[X]$
in~$\Otilde(\ell^2 r^2\log p)$ binary operations.
\end{thm}
This result generalizes to any other real quadratic field~$F$ for
which explicit generators of the graded rings of Hilbert modular forms
over~$\mathbb{Z}$ are known. Otherwise, the evaluation algorithm can still be
run, but it involves a heuristic reconstruction of rational numbers
from their complex approximations.
\begin{thm}[{\cite[Thm.~6.3]{kieffer_ComputingIsogeniesModular2019}}]
\label{thm:isog-hilbert}
Let~$F$ be a fixed real quadratic field. Then there exists an open
subvariety~$U\subset \mathcal{A}_{2,F}^0(\beta)$ and an algorithm which, for any
field~$k$, given:
\begin{itemize}
\item a totally positive~$\beta\in \mathbb{Z}_F$ such
that~$\chr(k) > 4\Tr_{F/\mathbb{Q}}(\beta)+7$ if it is positive;
\item the Igusa invariants of two~$\beta$-isogenous p.p.~abelian
surfaces~$A$ and~$A'$ defined over~$k$ with real multiplication
by~$\mathbb{Z}_F$, such that this~$\beta$-isogeny comes from a point
of~$U$;
\item The nine values
\begin{displaymath}
\partial_{J_i}\Psi_{\beta,k}^J(j_1(A),j_2(A),j_3(A),j_1(A'))\in k
\end{displaymath}
for~$1\leq i,k\leq 3$;
\end{itemize}
computes the following data:
\begin{itemize}
\item a tower~$k'/k$ of at most three quadratic extensions;
\item equations for two genus~$2$ hyperelliptic curves~$\mathcal{C}$
and~$\mathcal{C}'$ whose Jacobians are isomorphic to~$A$ and~$A'$ over an
algebraic closure of~$k$;
\item a point~$p_0\in \mathcal{C}(k')$; and
\item at most four possible tuples of rational fractions
in~$k'(x,y)$ of total degree $O(\Tr_{F/\mathbb{Q}}(\beta))$, such that one
of these tuples describes a~$\beta$-isogeny~$f\colon A\to A'$ in the sense
of~\eqref{eq:rational-map} using~$p_0\in \mathcal{C}(k')$ as a base
point;
\end{itemize}
using~$\Otilde(\Tr_{F/\mathbb{Q}}(\beta)) + O_F(1)$ elementary operations
and~$O(1)$ square roots in~$k'$.
\end{thm}
As in \cref{thm:isog-siegel}, the open
subvariety~$U\subset \mathcal{A}_{2,F}^0(\beta)$ in \cref{thm:isog-hilbert} can
be described explicitly. It is sufficient to impose the following
conditions~\cite[§4.2.3]{kieffer_ComputingIsogeniesModular2019}:
\begin{itemize}
\item Both~$A$ and~$A'$ have no extraneous automorphisms as
p.p.~abelian surfaces;
\item There exists only one isogeny~$f\colon A\to A'$ over an algebraic
closure of~$k$ whose kernel is cyclic of degree~$\ell$, up to
equivalence;
\item There exists only one possible real multiplication embedding
of~$\mathbb{Z}_F$ inside both~$\End(A)$ up to real conjugation on~$\mathbb{Z}_F$,
and the same holds for~$A'$;
\item Both~$A$ and~$A'$ lie outside of the singular locus of Igusa
invariants.
\end{itemize}
\subsection{Complexity bounds} Let~$q = p^r$ be a prime power, and
let~$A$ be a p.p.~abelian surface over~$\mathbb{F}_q$ with RM by~$\mathbb{Z}_F$. We
apply Elkies's method as follows.
\begin{enumerate}
\item Let~$\ell\in \mathbb{Z}$ be a prime with the correct splitting behavior
in~$\mathbb{Z}_F$, and let $\beta\in \mathbb{Z}_F$ be a totally positive prime
above~$\ell$. By~\cite[Lem.~1]{gaudry_CountingPointsGenus2011}, it
is possible to choose~$\beta$ such
that~$\Tr_{F/\mathbb{Q}}(\beta)\in O_F(\sqrt{\ell})$. We evaluate the
corresponding modular equations using \cref{thm:eval-hilbert}, for
instance in Igusa invariants. Assuming that~$r = o(\log p)$, this
costs~$\Otilde_F(\ell^2 \log q)$ binary operations.
\item We then attempt to find a root of these modular equations
over~$\mathbb{F}_q$; this costs~$\Otilde(\ell\log^2 q)$ binary
operations. If there are none, we skip~$\ell$.
\item If we find one, we attempt to compute a cyclic
isogeny~$f\colon A\to A'$ of degree~$\ell$ using
\cref{thm:isog-hilbert}. If this computation is successful, we
obtain the rational representation of~$f$ as
in~\eqref{eq:rational-map} in terms of rational functions of
degree~$O_F(\sqrt{\ell})$. This costs~$\Otilde_F(\sqrt{\ell})$
binary operations.
\item After that, computing a Gröbner basis describing~$\ker(f)$
costs~$\Otilde(\ell^{3/2}\log q)$ binary operations; the resulting
polynomials have degree~$O(\ell)$. Computing the Frobenius
eigenvalue on~$\ker(f)$ costs~$\Otilde(\ell\log^2 q)$ binary
operations.
\end{enumerate}
By \cref{prop:crt-hilbert}, the total algorithm
costs~$\Otilde_F(\log^4 q)$ binary operations provided that
sufficiently many Elkies primes exist and the computations of
\cref{thm:isog-hilbert} succeed sufficiently often.
\subsection{Degenerate cases}
The treatment of degenerate cases which may occur the algorithm above
is similar to the Siegel case. The only new possible problems are the
following:
\begin{enumerate}
\item \label{pb:rm-cm} The algorithm may involve a p.p.~abelian
surface~$A$ that corresponds to a point where the map from~$\mathcal{A}_{2,F}$
to~$\mathcal{A}_2$ is not étale;
\item \label{pb:rm-end} $\End(A)$ may contain an element of
norm~$\ell^2$ outside~$\mathbb{Z}_F$;
\item \label{pb:parasites} The isogeny algorithm of
\cref{thm:isog-hilbert} may output several possibilities for the
rational representation of~$f$.
\end{enumerate}
In case~\eqref{pb:rm-cm}, we can always consider coordinates on the
Hilbert surface~$\mathcal{A}_{2,F}$ instead. As in the Siegel case, we
incorporate the assumtion that case~\eqref{pb:rm-end} does not happen
too often into \cref{def:elkies-hilbert} below. Finally, in
case~\eqref{pb:parasites}, we can check which of the candidate values
actually describes a group morphism between Jacobians; if more than
one candidate passes this test, we are led back to
case~\eqref{pb:rm-end}.
\begin{defn}
\label{def:elkies-hilbert}
Let~$\varepsilon>0$, and let~$A$ be a p.p.~abelian surface over~$\mathbb{F}_q$ with
real multiplication by~$\mathbb{Z}_F$. We say that \emph{a proportion~$\varepsilon$
of primes are Elkies for~$A$} if the following holds: for
every~$X\geq \frac{1}{\varepsilon}\log q$, the proportion of
primes~$\ell\leq X$ such that
\begin{itemize}
\item $\ell$ splits in the form~$(\beta)(\conj{\beta})$ for some
totally positive~$\beta\in \mathbb{Z}_F$,
\item one of~$\beta$ or~$\conj{\beta}$ is Elkies, and
\item $\End(A)$ admits no non-real endomorphism of norm~$\ell^2$,
\end{itemize}
is at least~$\varepsilon$.
\end{defn}
\begin{rem}
Let~$\ell$ be a prime that splits in~$\mathbb{Z}_F$, but whose prime
factors~$\mathfrak l\cdot \conj{\mathfrak l}$ are nontrivial in the
narrow class group of~$\mathbb{Z}_F$. Instead of skipping~$\ell$ altogether
in the point-counting algorithm, we can compute a totally positive
generator~$\beta$ of~$\mathfrak l \cdot \mathfrak c$
where~$\mathfrak c$ denotes a small representative of the relevant
class in the narrow class group of~$\mathbb{Z}_F$. Elkies's method will also
apply to the non-prime~$\beta$.
\end{rem}
\section{Experimental results}
\label{sec:exp}
We have implemented algorithms to evaluate modular equations for
p.p.~abelian surfaces over~$\mathbb{Q}$ and prime finite fields in
C~\cite{kieffer_HDMELibraryEvaluation}. The experiments presented here
can be reproduced by downloading the library and running \texttt{make
reproduce}. In practice, we expect that the evaluation of modular
equations will exceed the cost of other polynomial manipulations in
Elkies's method by a large margin.
In the Siegel case, we consider the ``random'' tuple of Igusa
invariants of small height given by~$(159/239,-19/28,-193/246)$. We
record the time to evaluate Siegel modular equations at this point for
prime levels~$\ell \leq 17$ on a single core (AMD EPYC 7713), and
compare it with the cost estimation
of~$0.002\, \ell^6 \log(\ell)^3\log\log(\ell)$ seconds.
{\small
\begin{center}
\begin{tabular}{c|ccccccc}
$\ell$ & 2 & 3 & 5 & 7 & 11 & 13 & 17 \\
Time (s) & $1.34$
& $5.12$
& $96.7$
& $1.23\cdot 10^3$
& $3.97\cdot 10^4$
& $1.57\cdot 10^5$
& $1.12\cdot 10^6$ \\
Estimation
& -
& -
& $62$
& $1.2\cdot 10^3$
& $4.3\cdot 10^4$
& $1.5\cdot 10^5$
& $1.1\cdot 10^6$
\end{tabular}
\end{center}
}
In light of these results, a point-counting approach based exclusively
on Elkies's method for general p.p.~abelian surfaces would be unlikely
to beat Schoof's method in practical instances. However, using modular
equations would still allow one to introduce several improvements
inspired from the SEA algorithm (see~§\ref{sec:implem}).
In the Hilbert case for~$F = \mathbb{Q}(\sqrt{5})$, we consider the pair of
``random'' Gundlach invariants of small height given by
$(-117/64, -199/172)$. We evaluate Hilbert modular equations of
level~$\beta$, where~$\beta\in \mathbb{Z}_F$ is totally positive of prime
norm~$\ell\leq 500$ at that point, and compare it with the estimation
of~$2\cdot 10^{-4}\, \ell^2 \log(\ell)^3 \log\log\ell $ seconds.
{
\small
\begin{center}
\begin{tabular}{c|cccccccc}
$\ell$ & 11 & 19 & 29 & 31 & 41 & 59 & 61 & 71 \\
Time (s) & $2.45$
& $4.14$
& $9.66$
& $11.1$
& $25.6$
& $59.6$
& $64.0$
& $107$
\\
Estimation
& -
& $2.0$
& $7.8$
& $9.6$
& $23$
& $66$
& $73$
& $113$
\\[1pt] \hline \\[-1em]
$\ell$ & 79 & 89 & 101 & 109 & 131 & 139 & 149 & 151 \\
Time (s)
& $141$
& $190$
& $256$
& $315$
& $562$
& $673$
& $794$
& $709$ \\
Estimation
& $153$
& $215$
& $307$
& $379$
& $630$
& $741$
& $896$
& $929$
\end{tabular}
\end{center}
}
\noindent
For larger primes, counting in core-hours is perhaps more readable.
{
\small
\begin{center}
\setlength{\tabcolsep}{4pt}
\begin{tabular}{c|cccccccccc}
$\ell$ & 179 & 181 & 191 & 199 & 211 & 229 & 239 & 241 & 251 & 269\\
Time (h) & $0.299$
& $0.307$
& $0.348$
& $0.388$
& $0.452$
& $0.556$
& $0.613$
& $0.623$
& $0.697$
& $0.948$
\\
Estimation
& $0.41$ & $0.42$ & $0.49$ & $0.54$ & $0.64$ & $0.79$ & $0.89$ & $0.91$ & $1.0$ & $1.2$
\\[1pt] \hline \\[-1em]
$\ell$ & 271 & 281 & 311 & 331 & 349 & 359 & 379 & 389 & 401 & 409 \\
Time (h) & $0.965$
& $1.06$
& $1.37$
& $1.56$
& $1.78$
& $1.93$
& $2.37$
& $2.80$
& $3.03$
& $3.26$
\\
Estimation
& $1.2$ & $1.4$ & $1.8$ & $2.1$ & $2.4$ & $2.6$ & $3.0$ & $3.2$ & $3.4$ & $3.6$
\\[1pt] \hline \\[-1em]
$\ell$ & 419 & 421 & 431 & 439 & 449 & 461 & 479 & 491 & 499\\
Time (h) & $3.47$
& $3.50$
& $3.71$
& $4.03$
& $4.04$
& $5.46$
& $4.67$
& $4.98$
& $6.14$
\\
Estimation
& $3.9$ & $3.9$ & $4.2$ & $4.4$ & $4.6$ & $4.9$ & $5.5$ & $5.8$ & $6.1$
\end{tabular}
\end{center}
}
Consider the problem of counting points on a p.p.~abelian surface~$A$
over a prime field~$\mathbb{F}_q$ with~$\log q\simeq 512$ given by these
Gundlach invariants. Assuming that half of the primes~$\beta$ are
Elkies, a strategy based purely on Elkies's method would involve all
primes~$\ell$ up to roughly~$800$ with the correct splitting behavior
in~$\mathbb{Z}_F$. We can roughly estimate a total cost of a few core-weeks
for this computation, thus indicating that Schoof's method can perhaps
be beaten in this context~\cite[§5.2]{gaudry_CountingPointsGenus2011}.
Finally, we report on the time (in seconds) to evaluate modular
equations for Hilbert type in Igusa invariants for~$\ell\leq 50$, at
the point given by the parameters~$(-239/152, 224/193)$ in the
different parametrizations of Humbert surfaces found
in~\cite{elkies_K3SurfacesEquations2014}, for all real quadratic
fields of discriminants up to~$100$.
{
\small
\begin{center}
\setlength{\tabcolsep}{2.5pt}
\begin{tabular}{c|ccccccccccccccc}
$\Delta_F,\ell$
& 2 & 3 & 5 & 7 & 11 & 13 & 17 & 19 & 23 & 29 & 31 & 37 & 41 & 43 & 47\\\hline
5 & - & - & - & - & $9.76$ & - & - & $30.2$ & - & $78.0$ & $91.1$ & - & $205$ & - & - \\
8 & - & - & - & $2.67$ & - & - & $26.9$ & - & $48.2$ & - & $95.1$ & - & $208$ & - & $290$ \\
12 & - & - & - & - & - & $40.9$ & - & - & - & - & - & $451$ & - & - & - \\
13 & - & $2.02$ & - & - & - & - & $27.5$ & - & $51.1$ & $83.9$ & - & - & - & $240$ & - \\
17 & $4.34$ & - & - & - & - & $57.2$ & - & $126$ & - & - & - & - & - & $735$ & $825$ \\
21 & - & - & - & - & - & - & - & - & - & - & - & $363$ & - & $505$ & - \\
24 & - & - & - & - & - & - & - & $37.5$ & - & - & - & - & - & $235$ & - \\
28 & - & - & - & - & - & - & - & - & - & $211$ & - & $413$ & - & - & -\\
29 & - & - & $4.75$ & $5.76$ & - & $15.0$ & - & - & $48.3$ & - & - & - & - & - & -\\
33 & - & - & - & - & - & - & - & - & - & - & $255$ & $430$ & - & - & -\\
37 & - & $3.04$ & - & $13.5$ & $29.5$ & - & - & - & - & - & - & - & $503$ & - & $680$\\
40 & - & - & - & - & - & - & - & - & - & - &$212$ & - & $462$ & - & -\\
41 & $5.27$ & - & $10.5$ & - & - & - & - & - & $326$ & - & $625$ & $1050$ & - & $1450$ & -\\
44 & - & - & $4.77$ & - & - & - & - & - & - & - & - & $403$ & - & - & -\\
53 & - & - & - & $24.7$ & $49.6$ & $59.7$ & $107$ & - & - & $281$ & - & $524$ & - & $700$ & $854$ \\
56 & - & - & - & - & $27$ & - & - & - & - & - & - & - & - & $1360$ & -\\
57 & - & - & - & $30.0$ & - & - & - & - & - & - & - & - & - & $1980$ & -\\
60 & - & - & - & - & - & - & - & - & - & - & - & - & - & - & -\\
61 & - & $8.34$ & $13.9$ & - & - & $49.1$ & - & $108$ & - & - & - & - & $580$ & - & $772$ \\
65 & - & - & - & - & - & - & - & - & - & $1500$ & - & - & - & - & -\\
69 & - & - & - & - & - & $41.8$ & - & - & - & - & $253$ & - & - & - & -\\
73 & $6.40$ & $14.9$ & - & - & - & - & - & $241$ & $342$ & - & - & $1090$ & $1370$ & - & - \\
76 & - & - & $6.9$ & - & - & - & $55.8$ & - & - & - & - & - & - & - & -\\
77 & - & - & - & - & - & - & - & - & $133$ & - & - & $429$ & - & - & -\\
85 & - & - & - & - & - & - & - & $87.1$ & - & - & - & - & - & - & -\\
88 & - & $18.1$ & - & - & - & - & - & - & - & - & - & - & - & - & -\\
89 & $3.55$ & - & $15.4$ & - & $57.3$ & - & $147$ & - & - & - & - & - & - & - & $1600$\\
92 & - & - & - & - & - & $104$ & - & - & - & $567$ & - & - & $1370$ & - & -\\
93 & - & - & - & $16.5$ & - & - & - & $95$ & - & - & - & - & - & - & -\\
97 & $7.01$ & $8.48$ & - & - & $58.1$ & - & - & - & - & - & $1370$ & - & - & $3210$ & $4020$\\
\end{tabular}
\end{center}
}
On the first line, we observe that Hilbert modular equations in Igusa
invariants for~$F = \mathbb{Q}(\sqrt{5})$ are indeed more expensive to
evaluate than their counterparts in Gundlach invariants.
\section{Perspectives}
\label{sec:implem}
In this final section, we sketch possible improvements to Elkies's
method for abelian surfaces as described above, following existing
works in the dimension~$1$ case. They would either reduce the constant
factors hidden in complexity estimates by large amounts, or introduce
exponential-time gains.
\subsection{Smaller modular equations}
The modular equations of Siegel and Hilbert type presented above are
higher-dimensional analogues of the classical modular
polynomials~$\Phi_\ell$ in dimension~$1$. It is well-known that other
kinds of modular polynomials provide explicit equations for
essentially the same modular curve which are much smaller, despite
sharing the same~$O(\ell^3\log\ell)$ size asymptotic: see for
instance~\cite[§3]{enge_ClassInvariantsCRT2010} and the data available
at~\cite{sutherland_DatabaseModularPolynomials}.
In the dimension~$2$ case, modular equations written in terms of theta
constants are considerably smaller than Siegel or Hilbert modular
equations as defined
above~\cite{milio_DatabaseModularPolynomials}. One can ask whether
this choice of coordinates is the optimal one.
More generally, it could well be that systems of equations of the
form~\eqref{eq:siegel-modeq} inherently force modular equations to
have large coefficients; other ways of describing the
diagram~\eqref{diag:siegel} might lead to smaller polynomials -- for
instance, a more intrinsic equation for~$\mathcal{A}_2^0(\ell)$ along with the
Atkin--Lehner involution exchanging~$p_1$ and~$p_2$. Such equations do
not even have to be defined by a formula valid for each~$\ell$; all we
need is an algorithm to compute such equations when~$\ell$ is given,
perhaps by computing a basis of Siegel modular forms of
level~$\Gamma^0(\ell)$, or Hilbert modular forms of
level~$\Gamma^0(\beta)$, on the fly.
\subsection{Other SEA strategies}
In the case of elliptic curves, there is more to the SEA algorithm
than applying Elkies's method to a series of distinct primes. We list
some of the possible improvements below. Due to the larger implied
constants in complexity estimates about modular equations, we expect
these improvements to have an even larger impact on practical running
times in higher dimensions.
\begin{enumerate}
\item \emph{Isogeny chains.} In favorable situations, modular
polynomials of level~$\ell$ can be used to compute not only
an~$\ell$-isogeny~$E\to E'$, but a chain
of~$\ell$-isogenies~$E \to E_1\to \cdots \to E_r$ whose composition
is an~$\ell^r$-isogeny, for
some~$r\geq 2$~\cite{fouquet_IsogenyVolcanoesSEA2002}. This yields
the value of~$\chi(E)$ modulo~$\ell^r$. In order to remain within
the same complexity bound, one should take~$r$ no greater
than~$\log\log(q)/\log(\ell) + O(1)$. The existence of an isogeny
chain of the desired length over~$\mathbb{F}_q$ depends on the shape of the
connected component of the~$\ell$-isogeny graph on which~$E$ lies.
Note that a chain of length~$r=2$ can be constructed by evaluating
modular polynomials only once, at~$E_1$. In dimension~$2$, this
strategy seems easier to apply in the Hilbert case, since isogeny
graphs are still volcanoes in this
case~\cite{ionica_IsogenyGraphsMaximal2020} and the composition of a
non-backtracking chain of~$\beta$-isogenies will always be
a~$\beta^r$-isogeny. This property does not hold
for~$\ell$-isogenies in the Siegel case, and the shape of the
associated isogeny graphs is also more
complicated~\cite{brooks_IsogenyGraphsOrdinary2017}.
\item \emph{Atkin's method.} It is known that studying the
factorization patterns of modular equation of level~$\ell$
over~$\mathbb{F}_q$, even in the absence of rational roots, restrict the
possible Frobenius eigenvalues
modulo~$\ell$~\cite[§6]{schoof_CountingPointsElliptic1995},
\cite{ballentine_IsogeniesPointCounting2016}. This allows one to
take advantage of Atkin (i.e.~non-Elkies) primes as well. This
information can be used at the end of the point-counting algorithm
in an exponential-time sieve, whose practical effect is to reduce
the number of necessary Elkies primes. In general, if we can
compute~$n\geq 2$ possible values of~$\chi(A)$ modulo~$\ell$, the
``value'' of~$\ell$ as an Atkin prime is~$\log(\ell)/\log(n)$, and
one should only keep the highest-valued primes for the final
sieve. Thus, once a few Atkin primes have been collected, it only
makes sense to look for low-degree factors of modular equations;
this is cheaper than computing the full factorization.
\item \emph{Schoof's method.} When~$\ell = O(\sqrt{\log q})$ is not
Elkies, it is usually more interesting to apply Schoof's original
method to compute~$\chi(A)$ mod~$\ell$ than attempting to
keep~$\ell$ as part of the Atkin data; this makes space for larger
primes in the final sieve. If~$\ell$ is very small (for
instance~$\ell=2$), then Schoof's method can also yield~$\chi(A)$
modulo a suitable power
of~$\ell$~\cite[§4]{gaudry_GenusPointCounting2012}.
\end{enumerate}
\bibliographystyle{abbrv}
| 710bebd9b458734799aaecf223875c20a916cc12 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Superconformal field theories (SCFTs) in five spacetime dimensions have received much deserved attention in recent years. Their mere existence forces us to expand our view of quantum field theories, as these fixed points cannot be reached by traditional means of perturbing around free field Lagrangians. Yet, by now, there is overwhelming evidence to support their existence, mostly due to stringy constructions starting with the seminal papers \cite{Seiberg:1996bd,Intriligator:1997pq,Ganor:1996pc,Brandhuber:1999np}. Broadly speaking, there are three independent, yet complementary points of view for studying these SCFTs, namely their embedding into type IIB brane webs \cite{Aharony:1997bh,Aharony:1997ju,Bergman:2013ala,Bergman:2013aca,Bergman:2014kza,Bergman:2015dpa,Kim:2015jba,Zafrir:2015ftn,Hayashi:2015vhy,Zafrir:2015rga,Hayashi:2015zka,Zafrir:2016jpu,Hayashi:2016abm,Hayashi:2016jak,Hayashi:2017jze,Hayashi:2017btw,Hayashi:2018bkd,Hayashi:2018lyv,Hayashi:2019yxj,Kim:2020hhh,Kim:2021fxx}, geometric engineering \cite{Douglas:1996xp,Morrison:1996xf,Intriligator:1997pq,Jefferson:2018irk,Closset:2018bjz,Bhardwaj:2019jtr,Apruzzi:2019kgb,Bhardwaj:2020gyu,Tian:2021cif,Acharya:2021jsp} and holography \cite{ Bergman:2018hin,Fluder:2018chf,Kaidi:2018zkx,Uhlemann:2019lge,Uhlemann:2019ypp,Legramandi:2021uds,Legramandi:2021aqv}.
Many 5d SCFTs, admit supersymmetry preserving mass deformations, which trigger an RG flow, whose low energy dynamics is effectively captured by an $\mathcal{N}=1$ gauge theory. Such deformations, preserve the SU(2)$_R$ symmetry, while breaking the flavour symmetry. An important dynamical question is therefore to determine the full global symmetry of the parent SCFT of a given 5d gauge theory. 5d SCFTs and gauge theories, can possess a Higgs branch of their moduli space of vacua, which in the gauge theory regime can be constructed as the hyperK\"ahler quotient \cite{Gray:2008yu,Hanany:2008kn}. In the SCFT limit, the hyperK\"ahler quotient, is no longer accessible, due to a lack of a Lagrangian description, making the study of the Higgs branch in this limit more challenging.
There are, by now, a plethora of techniques to determine the enhanced global symmetry of the SCFT parent of a given 5d gauge theory, such as 7-brane analysis \cite{Sen:1996vd,Gaberdiel:1997ud,Gaberdiel:1998mv,DeWolfe:1998eu,DeWolfe:1998pr,DeWolfe:1999hj}, superconformal indices \cite{Kim:2012gu,Bergman:2013koa, Hwang:2014uwa}, as well as geometric approaches \cite{Katz:1996xe,Witten:1996qb,Morrison:1996xf,Douglas:1996xp,Intriligator:1997pq, Jefferson:2018irk,Apruzzi:2019opn,Apruzzi:2019enx,Bhardwaj:2020ruf,Bhardwaj:2020avz}. One particularly elegant approach to determine the SCFT flavour symmetries as well as the Higgs branch of the SCFT, pioneered in \cite{Ferlito:2017xdq}, is to consider their magnetic quivers. The magnetic quiver (MQ), of a given 5d theory, is a 3d $\mathcal{N}=4$ quiver gauge theory, whose Coulomb branch is isomorphic to the Higgs branch of the 5d theory in question. In many cases, though not always, one can show that the magnetic quiver of a given 5d theory, is the 3d mirror of its torus compactification. This leads to an interesting interplay between 5d $\mathcal{N}=1$ theories and 3d $\mathcal{N}=4$ theories, and has prompted many recent studies \cite{Akhond:2020vhc,Akhond:2021knl,Akhond:2021ffo,vanBeest:2020civ,vanBeest:2020kou,vanBeest:2021xyt,Bourget:2019rtl,Bourget:2020asf,Bourget:2020gzi,Bourget:2020mez,Bourget:2020xdz,Bourget:2021xex,Closset:2020afy,Closset:2020scj,Cabrera:2019izd,Ferlito:2017xdq,Carta:2021whq,Carta:2021dyx}.
That 3d magnetic quivers are advantageous, is due to recent advances in extracting the algebraic geometry of 3d $\mathcal{N}=4$ Coulomb branches \cite{Cremonesi:2013lqa,Cremonesi:2014kwa,Cremonesi:2014vla,Cremonesi:2014xha,Hanany:2014dia}. Coulomb branches of 3d $\mathcal{N}=4$ theories, are parameterised by BPS monopole operators, whose R-charges can be determined from the R-charges of the fermionic content of the \cite{Borokhov:2002cg}. Following the terminology of Gaiotto and Witten \cite{Gaiotto:2008ak}, we refer to a 3d $\mathcal{N}=4$ theory as good, ugly or bad, depending on the conformal dimension of the monopole operators in that theory. A theory is said to be good, if the conformal dimension of all monopole operators is greater than $\frac{1}{2}$, which is the dimension of a scalar in (2+1) dimensions. A bad theory is one in which some monopole operator has a conformal dimension less than $\frac{1}{2}$. Finally, in an ugly theory, there are some monopole operators whose conformal dimension is exactly $\frac{1}{2}$, but none that are smaller. For good or ugly theories, the monopole formula \cite{Cremonesi:2013lqa}, can correctly produce the Hilbert series of the Coulomb branch, using the UV gauge theory data. However, for bad theories, the monopole formula fails, due to the fact that the UV R-symmetries of the theory are different from the superconformal R-symmetry, and hence the UV R-charge cannot be used to predict the conformal dimension of the SCFT operators. The full structure of the Coulomb branch of some bad theories was recently understood by Assel and Cremonesi in \cite{Assel:2017jgo,Assel:2018exy}, building on earlier results in \cite{Yaakov:2013fza,Ferlito:2016grh}. In particular, it was found that for USp$(2N)$ SQCD with $2N$ flavours in the fundamental representation, the Coulomb branch has two most singular points, where the theory flows to an interacting SCFT in the deep infra-red, with a local mirror Lagrangian in the vicinity of each singular point. This is unlike the situation for good theories, where the most singular point on the moduli space is unique, namely the origin, with an SCFT at the bottom of the RG flow in that vacuum. The Coulomb branch Hilbert series of the local geometry around one of the two most singular points of the USp($2N$) SQCD with $2N$ flavours was computed in \cite{Assel:2018exy}, and shown to agree with the Higgs branch Hilbert series of the local mirror.
The goal of this paper is to illuminate the Higgs branch of the SCFT parents of 5d gauge theories whose gauge group is either SO(6), or SO(8), and with matter in the vector, spinor, and conjugate spinor representations. Our motivation is partly due to the fact that, among classical simple gauge groups, the magnetic quivers for $A_N$ and $C_N$ cases are well studied \cite{Cabrera:2018jxt,Ferlito:2017xdq}, while those of $B_N$ and $C_N$ remain relatively unexplored. The magnetic quivers for SO$(N)$ gauge theories with matter in vector representation were recently constructed in \cite{Akhond:2021ffo}, using their embedding into brane webs with O7$^+$-planes. The current work complements this study by adding matter in the spinor and the conjugate spinor representations. The reason for restricting the rank of the gauge group is mostly for practical reasons. The construction of magnetic quivers from brane webs with O5-planes, is at present, still not completely systematic. It is, therefore, useful to limit the discussion to situations where a consistency check of our computations is available. For SO(6) gauge theories, this is achieved by comparing the orthosymplectic magnetic quivers obtained from brane webs with an O5-plane, with the unitary magnetic quivers obtained from the brane webs of SU(4) gauge theories. Similarly, for SO(8) gauge theories, one can set up a consistency check, by exploiting SO(8) triality. As we shall see, in the rest of this paper, such considerations also lead to some interesting results for 3d $\mathcal{N}=4$ theories. In particular, in Section \ref{SO6 theories}, we exploit the isomorphism between SO(6) and SU(4), to conjecture exact highest weight generating functions for orthosymplectic magnetic quivers, simply by carrying over the results from their unitary counterparts. Similarly, in Section \ref{SO8 theories}, we uncover several intriguing equalities of moduli spaces, of naively unrelated quivers.
Throughout the paper, we encounter magnetic quivers which contain bad USp$(2N)$ nodes, where the effective number of hypermultiplets is exactly $2N$. We devise a method to associate a Hilbert series to these quivers, by using the local mirror description around one of their two most singular loci. We will refer to this procedure of associating the Coulomb branch Hilbert series of a good theory, to the local geometry of the Coulomb branch near a singular locus of a bad theory as ``B2G". For the specific case of USp$(2)$ with 2 flavours, the prescription is first encountered in section \ref{sec:SO6+1S1C1V}. A similar prescription for USp($2N$) theories for $N>2$ is used throughout the paper, though the technical details will be published elsewhere \cite{HSbad}. The validity of our prescription is confirmed, by comparing the results with those computed using the Hall-Littlewood technique \cite{Cremonesi:2014kwa}. In addition, many of the magnetic quivers that involve bad nodes, can have a good ``dual" description, since we have several inequivalent brane constructions for each 5d theory we consider. The Hilbert series computed using our proposed prescription, is also consistent with the good ``dual" quivers.
The content of the remainder of this paper is organised as follows: Section \ref{SO6 theories} contains MQs for SO(6) gauge theories derived from brane webs with O5-planes, as well as MQs for SU(4) theories, derived from ordinary brane webs. Here we will encounter MQs with bad symplectic gauge nodes and devise various methods to extract the good interacting part. Section \ref{SO8 theories} is dedicated to studying the MQs for SO(8) theories, and a comparison is made between MQs of SO(8) theories related by triality, exchanging spinor, conjugate spinor, and vector representations of SO(8). We conclude with our wish list for future projects in \ref{conclusions}. Appendix \ref{appendixA} contains the unrefined Hilbert series results for the OSp quivers.
\section{SO(6) vs SU(4)}\label{SO6 theories}
In this section we consider magnetic quivers for 5d SCFTs that admit a mass deformation, such that their low-energy dynamics is captured by SO(6) gauge theory with $s$ hypermultiplets in the spinor representation \textbf{S}, $c$ hypermultiplets in the conjugate spinor representation \textbf{C} and $v$ hypermultiplets in the vector representation \textbf{V} of SO(6). Since the Lie algebra of SO(6), is isomorphic to the Lie algebra of SU(4), the low-energy theory may also be thought of as SU(4) gauge theory, with $(s+c)$ hypermultiplets in the fundamental representation \textbf{F}, and $v$ hypermultiplets in the 2nd rank antisymmetric representation \textbf{AS} of SU(4). The Chern-Simons level for the SU(4)$_\kappa$ gauge theory is given by $\kappa = \frac{s-c}{2}$:
\begin{equation}
\text{SO}(6)+s\textbf{S}+c\textbf{C}+v\textbf{V}\leftrightarrow \text{SU}(4)_{\frac{s-c}{2}} +(s+c)\textbf{F}+v\textbf{AS}\;.
\end{equation}
One can engineer SO(6)+$s$\textbf{S}+$c$\textbf{C}+$v$\textbf{V} using a five-brane web with an O5-plane.
In particular, (conjugate) spinors are introduced as a separate web on the left or right side of an SO(6) five-brane web as given in Figure \ref{fig:webwithspinors}. This web can be viewed as a quiver theory USp(0)--SO(6)--USp(0), where the USp(0) instanton, plays the role of the spinor matter \cite{Zafrir:2015ftn}. %
\begin{figure}[ht]
\centering
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(1,0.5);
\draw[thick](1,.5)--(1.5,1);
\draw[thick](1.5,1)--(1.5,1.5);
\draw[thick](1.5,1)--(2.5,1);
\draw[thick](2.5,1)--(2.5,1.5);
\draw[thick](1.5,1.5)--(2.5,1.5);
\draw[thick](1.5,1.5)--(1,2);
\draw[thick](2.5,1.5)--(3,2);
\draw[thick](.25,2)--(1,2);
\draw[thick](3.75,2)--(3,2);
\node[label=right:{1\textbf{V}}]at(3.75,2){};
\node[label=left:{1\textbf{V}}]at(0.25,2){};
\draw[thick](3,2.5)--(3,2);
\draw[thick](1,2.5)--(1,2);
\draw[thick](2.5,1)--(3,.5);
\draw[thick](1,.5)--(3,.5);
\draw[thick](3,.5)--(4,0);
\draw[thick](5,0)--(6,.5);
\draw[thick](6.5,1)--(6,.5);
\draw[thick](6.5,1)--(6.5,1.5);
\draw[thick](6,2)--(6.5,1.5);
\draw[thick](7,1.5)--(6.5,1.5);
\draw[thick](6.5,1)--(7,1);
\node[label=right:{1\textbf{S}+1\textbf{C}}]at (7,.5){};
\node[label=right:{2\textbf{S}+2\textbf{C}}]at (7,1){};
\node[label=right:{4\textbf{S}+4\textbf{C}}]at (7,1.5){};
\node at (5,.5){$\begin{matrix}1\textbf{S}\\(\text{or} 1\textbf{C})\end{matrix}$};
\draw[thick](7,.5)--(6,.5);
\draw[thick,dashed](-4,0)--(8,0);
\node[label=below:{O5$^-$}]at(2,0){};
\node[label=below:{O5$^+$}]at(4.5,0){};
\node[label=below:{O5$^-$}]at(6,0){};
\node[label=below:{O5$^+$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(-2,0){};
\draw[thick](-1,0)--(-2,.5);
\draw[thick](-3,0.5)--(-2,.5);
\draw[thick](-2.5,1)--(-2,.5);
\draw[thick](-2.5,1)--(-3,1);
\draw[thick](-2.5,1)--(-2.5,1.5);
\draw[thick](-3,1.5)--(-2.5,1.5);
\draw[thick](-2,2)--(-2.5,1.5);
\node[label=left:{1\textbf{S}+1\textbf{C}}]at(-3,.5){};
\node[label=left:{2\textbf{S}+2\textbf{C}}]at(-3,1){};
\node[label=left:{4\textbf{S}+4\textbf{C}}]at(-3,1.5){};
\node at (-1,.5){1\textbf{S} (or 1\textbf{C})};
\end{tikzpicture}
\end{scriptsize}
\caption{A web diagram for SO(6)+2\textbf{V} with various spinor matters. }
\label{fig:webwithspinors}
\end{figure}
Notice that in Figure \ref{fig:webwithspinors}, only restricted configurations of spinors and conjugate spinors are possible. For instance, 4\textbf{S}+3\textbf{C} cannot be depicted as a five-brane web. Furthermore, the distinction between spinors and conjugate spinors needs some caution. For instance, consider a five-web with two spinors, as shown in Figure \ref{fig:SorC}. This web can have three different interpretations, depending on how the configuration undergoes the ``generalized flop'' transition \cite{Hayashi:2017btw} when two spinor webs are brought towards the centre of the Coulomb branch. In particular, 1\textbf{S} is modified into the web for the $E_1$ theory, while 1\textbf{C} is modified into the web for the $\tilde{E}_1$ theory \cite{Hayashi:2017btw,Akhond:2020vhc}. In Figure \ref{fig:SorC}, we depict three possible configurations, 1\textbf{S}+1\textbf{C}, 2\textbf{S}, and 2\textbf{C}. We stress that the corresponding webs in the infinite coupling limit all look the same, and hence one needs to distinguish them by hand.
\begin{figure}[ht]
\centering
\begin{scriptsize}
\begin{tikzpicture}[scale=.75]
\draw[thick](0,0)--(1,0.5);
\draw[thick](1,.5)--(1.5,1);
\draw[thick](1.5,1)--(1.5,1.5);
\draw[thick](1.5,1)--(2.5,1);
\draw[thick](2.5,1)--(2.5,1.5);
\draw[thick](1.5,1.5)--(2.5,1.5);
\draw[thick](1.5,1.5)--(1,2);
\draw[thick](2.5,1.5)--(3,2);
\draw[thick](2.5,1)--(3,.5);
\draw[thick](1,.5)--(3,.5);
\draw[thick](3,.5)--(4,0);
\draw[thick](5,0)--(7,1);
\draw[thick,dashed](-4,0)--(8,0);
\draw[thick](-1,0)--(-3,1);
\draw[thick](0.5,-3.75)--(1,-3.5);
\draw[thick](0.5,-3.75)--(.25,-4);
\draw[thick](0,-3.75)--(.25,-4);
\draw[thick](0,-3.75)--(.5,-3.75);
\draw[thick](0,-3.75)--(-1,-3.25);
\draw[thick](1,-3.5)--(1.5,-3);
\draw[thick](1.5,-3)--(1.5,-2.5);
\draw[thick](1.5,-3)--(2.5,-3);
\draw[thick](2.5,-3)--(2.5,-2.5);
\draw[thick](1.5,-2.5)--(2.5,-2.5);
\draw[thick](1.5,-2.5)--(1,-2);
\draw[thick](2.5,-2.5)--(3,-2);
\draw[thick](2.5,-3)--(3,-3.5);
\draw[thick](1,-3.5)--(3,-3.5);
\draw[thick](3,-3.5)--(3.5,-3.75);
\draw[thick](3.75,-4)--(3.5,-3.75);
\draw[thick](3.75,-4)--(4,-3.75);
\draw[thick](3.5,-3.75)--(4,-3.75);
\draw[thick](5,-3.25)--(4,-3.75);
\node at (3.75,-3.5){1\textbf{S}};
\node at (0.25,-3.5){1\textbf{S}};
\draw[thick, dashed](-1,-4)--(5,-4);
\draw[thick](7.5,-3.75)--(8,-3.5);
\draw[thick](7.5,-3.75)--(7.5,-4);
\draw[thick](7.4,-3.75)--(7.4,-4);
\draw[thick](7.4,-3.75)--(7.5,-3.75);
\draw[thick](7.4,-3.75)--(6.4,-3.25);
\draw[thick](8,-3.5)--(8.5,-3);
\draw[thick](8.5,-3)--(8.5,-2.5);
\draw[thick](8.5,-3)--(9.5,-3);
\draw[thick](9.5,-3)--(9.5,-2.5);
\draw[thick](8.5,-2.5)--(9.5,-2.5);
\draw[thick](8.5,-2.5)--(8,-2);
\draw[thick](9.5,-2.5)--(10,-2);
\draw[thick](9.5,-3)--(10,-3.5);
\draw[thick](8,-3.5)--(10,-3.5);
\draw[thick](10,-3.5)--(10.5,-3.75);
\draw[thick](10.5,-4)--(10.5,-3.75);
\draw[thick](10.6,-4)--(10.6,-3.75);
\draw[thick](10.5,-3.75)--(10.6,-3.75);
\draw[thick](11.6,-3.25)--(10.6,-3.75);
\node at (10.5,-3.45){1\textbf{C}};
\node at (7.5,-3.45){1\textbf{C}};
\draw[thick, dashed](6,-4)--(12,-4);
\draw[thick](-6.5,-3.75)--(-6,-3.5);
\draw[thick](-6.5,-3.75)--(-6.5,-4);
\draw[thick](-6.6,-3.75)--(-6.6,-4);
\draw[thick](-6.5,-3.75)--(-6.6,-3.75);
\draw[thick](-7.6,-3.25)--(-6.6,-3.75);
\draw[thick](-6,-3.5)--(-5.5,-3);
\draw[thick](-5.5,-3)--(-5.5,-2.5);
\draw[thick](-5.5,-3)--(-4.5,-3);
\draw[thick](-4.5,-3)--(-4.5,-2.5);
\draw[thick](-5.5,-2.5)--(-4.5,-2.5);
\draw[thick](-5.5,-2.5)--(-6,-2);
\draw[thick](-4.5,-2.5)--(-4,-2);
\draw[thick](-4.5,-3)--(-4,-3.5);
\draw[thick](-6,-3.5)--(-4,-3.5);
\draw[thick](-4,-3.5)--(-3.5,-3.75);
\draw[thick](-3.25,-4)--(-3.5,-3.75);
\draw[thick](-3.25,-4)--(-3,-3.75);
\draw[thick](-3.5,-3.75)--(-3,-3.75);
\draw[thick](-2,-3.25)--(-3,-3.75);
\node at (-3.25,-3.5){1\textbf{S}};
\node at (-6.5,-3.45){1\textbf{C}};
\draw[thick, dashed](-2,-4)--(-8,-4);
\draw[thick](2,-8)--(4,-7);
\draw[thick](2,-8)--(3,-7);
\draw[thick](2,-8)--(1,-7);
\draw[thick](2,-8)--(0,-7);
\draw[thick,dashed](-1,-8)--(5,-8);
\draw[thick,->](2,-.5)--(2,-1.5);
\draw[thick,->](5,-.5)--(6,-1.5);
\draw[thick,->](-1.5,-.5)--(-2.5,-1.5);
\draw[thick,->](2,-4.75)--(2,-5.75);
\draw[thick,->](6,-4.75)--(5,-5.75);
\draw[thick,->](-2.5,-4.75)--(-1.5,-5.75);
\end{tikzpicture}
\end{scriptsize}
\caption{5-brane webs for the SO(6) with three possible spinor configurations. The web diagram on the top is generic 5-brane web depicted with an O5-plane (dotted line) where spinors are introduced on the left and right. The 5-brane webs in the middle are three possible brane configurations, 1\textbf{S}+1\textbf{C}, 2\textbf{S}, and 2\textbf{C}. The web on the bottom is the 5-brane configuration at infinite coupling.
}
\label{fig:SorC}
\end{figure}
One can engineer SU$(4)_\frac{s-c}{2}$+$(s+c)$\textbf{F}+$v$\textbf{AS}, using a five-brane web description without an O5-plane \cite{Bergman:2015dpa}. Taking the SCFT limit of the two brane webs, of the SO(6) and SU(4) theories, and reading off the corresponding magnetic quivers following the techniques developed in \cite{Akhond:2020vhc,Akhond:2021knl,Cabrera:2018jxt} yields two dual descriptions for the Higgs branch of the 5d SCFT in question. The magnetic quivers obtained from the brane web with an O5-plane will have qualitatively different features, compared to the magnetic quivers obtained from the brane web without an O5-plane. Typical features of magnetic quivers obtained from brane webs with an O5-plane are the presence of orthogonal, and symplectic gauge nodes, hypermultiplets in the fundamental-fundamental representation of two unitary nodes, and hypermultiplets in the second rank symmetric or antisymmetric representation of unitary nodes. In contrast, the magnetic quivers obtained from ordinary brane webs will only contain unitary nodes and bi-fundamental matter, with the possibility of multiple links (bi-fundamental hypermultiplets) connecting two gauge nodes.
In what follows, we will often encounter theories, whose Higgs branch is given as the union of several cones, where each cone is described by a distinct magnetic quiver. We will denote by $\MQ_i^{s,c,v}$, the magnetic quiver for the $i$-th cone of the SCFT parent of SO(6) +$s$\textbf{S}+$c$\textbf{C}+$v$\textbf{V}, i.e.,
\begin{equation}
\mathcal{H}^{5\text{d}}_\infty\left(\text{SO}(6)+s\textbf{S}+c\textbf{C}+v\textbf{V}\right)=\bigcup_{i}\mathcal{C}^{3\text{d}}\left(\MQ_i^{s,c,v}\right)\ .
\end{equation}
Our notation for the magnetic quivers follows those of \cite{Akhond:2020vhc, Akhond:2021ffo}, which we now briefly review for the reader's convenience. We use a circular white, red, or blue node to denote a unitary, orthogonal, or symplectic gauge algebra (and hence a 3d $\mathcal{N}=4$ vector multiplet) respectively. Square nodes with the same colour-coding are used to denote flavour symmetries. We label each node by the dimension of the fundamental representation of the corresponding algebra. A solid line connecting two nodes, represents a hypermultiplet transforming in the bi-fundamental representation of the nodes that it connects. An exception to the previous statement is the case in which a solid line connects a symplectic and orthogonal node, in that instance one has to impose a reality condition which leads to a half hypermultiplet in the fundamental-fundamental representation. We use a dashed line connecting two unitary gauge nodes to represent hypermultiplets transforming under the fundamental-fundamental representation. A link starting and ending on a unitary gauge node represents hypermultiplets in the second rank antisymmetric representation, the number of such hypermultiplets will be written explicitly in the diagrams. Hypermultiplets in the second rank symmetric representation of unitary gauge nodes are denoted by jagged lines.
The rest of this section is further subdivided to the cases with $v=0,1,2$.
\subsection{SO(6) theories without vector matter}
We will first consider SO(6) theories, with (conjugate) spinors and without any vector hypermultiplets. The unitary magnetic quivers correspond to SU$(4)_\kappa$ with fundamental hypermultiplets, and were given explicitly in \cite{Cabrera:2019izd}. All the magnetic quivers appearing in this section have a simple highest weight generating function (HWG) \cite{Hanany:2014dia} for their Coulomb branch Hilbert series. We verify that the same HWG can correctly reproduce the unrefined Hilbert series of the Coulomb branch of the OSp quivers.
\subsubsection{SO(6)+2\textbf{S}/(1\textbf{S}+1\textbf{C})\, \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_{1/0}$}{TEXT}+2\textbf{F}}
\label{subsubsec:SO62S}
The brane web for SO(6)+2\textbf{S} at infinite coupling is given by
\begin{equation}\label{SO(6)+2s orientifold web}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(4,0);
\draw[thick](2,0)--(4,1);
\draw[thick](2,0)--(0,1);
\draw[thick](2,0)--(3,1);
\draw[thick](2,0)--(1,1);
\node[label=above:{$(2,-1)$}][7brane]at(0,1){};
\node[label=above:{$(1,-1)$}][7brane]at(1,1){};
\node[label=above:{$(1,1)$}][7brane]at(3,1){};
\node[label=above:{$(2,1)$}][7brane]at(4,1){};
\node[label=below:{O5$^-$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(3.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
This diagram can actually represent two distinct theories, namely, SO(6)+2\textbf{S}, as well as SO(6)+1\textbf{S}+1\textbf{C}. The easiest way to see this, is to realise that \eqref{SO(6)+2s orientifold web} contains the following subweb
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-2,-1)--(2,1);
\draw[thick](-2,1)--(2,-1);
\draw[thick,dashed](-3,0)--(3,0);
\node[label=below:{O5$^-$}]at(-2.5,0){};
\node[label=below:{O5$^-$}]at(2.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
which can either be the strong coupling limit of USp$(2)_0$, or that of USp$(2)_\pi$, depending on how one performs a generalized flop in going to the gauge theory phase \cite{Hayashi:2017btw, Akhond:2020vhc}.
Therefore, this decomposition gives rise to two magnetic quivers, depending on whether one treats the subweb comprised of the $(2,1)$ and $(2,-1)$ 5-branes as the strongly coupled limit of $\text{USp}(2)_0$ or $\text{USp}(2)_\pi$ gauge theory. In the former case, the subweb in question is dynamical, and gives rise to a gauge node of rank one in the magnetic quiver, while in the latter we treat this subweb as frozen, and it only contributes as a flavour node in the magnetic quiver, whose rank is determined by taking its intersection number with the remaining subweb made up of the $(1,1)$ and $(1,-1)$ 5-branes. In addition, we can consider the subweb decomposition
\begin{equation}
\begin{array}{c}
\begin{tikzpicture}
\draw[thick] (-1,-1)--(0,0);
\draw[thick] (1,-1)--(0,0);
\draw[thick] (-2,1)--(0,0);
\draw[thick] (2,1)--(0,0);
\draw[thick] (1,1.3)--(0,0.3);
\draw[thick] (-1,1.3)--(0,0.3);
\draw[thick] (2,-.7)--(0,0.3);
\draw[thick] (-2,-.7)--(0,0.3);
\end{tikzpicture}
\end{array}\;,
\end{equation}
giving rise to a total of three magnetic quivers that one can extract from the diagram \eqref{SO(6)+2s orientifold web}. A comparison with the unitary magnetic quivers, obtained from the brane webs for SU$(4)_0$+2\textbf{F}, and SU$(4)_1$+2\textbf{F}, allows us to recognise the magnetic quiver for SO(6)+1\textbf{S}+1\textbf{C} to be
\begin{equation}
\MQ_1^{1,1,0}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](11)[right of=1]{};
\node[label=above:{1}][uf](f)[above of=11]{};
\draw[dashed](1) to[out=30, in=150]node[above]{3}(11);
\draw (1) to[out=-30, in=-150](11);
\path [draw,snake it](11)--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,\\
\label{SO6:MQ110}
\end{equation}
while the magnetic quivers for SO(6)+2\textbf{S} are given by
\begin{equation}
\MQ_1^{2,0,0}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\draw(1)--node[above]{4}++(1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,\qquad
\MQ_2^{2,0,0}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=above:{2}][uf](1f)[above of=1]{};
\path [draw,snake it](1)--(1f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ200}
\end{equation}
To confirm this, let us consider the brane web for SU$(4)_0$+2\textbf{F} at infinite coupling
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}[scale=.7]
\draw[thick](-2,2)--(2,-2);
\draw[thick](-2,-1)--(2,1);
\draw[thick](-2,0)--(2,0);
\node[label=above:{$(1,-1)$}][7brane]at(-2,2){};
\node[label=above:{$(2,1)$}][7brane]at(2,1){};
\node[label=below:{$(1,-1)$}][7brane]at(2,-2){};
\node[label=below:{$(2,1)$}][7brane]at(-2,-1){};
\node[label=right:{$(1,0)$}][7brane]at(2,0){};
\node[label=left:{$(1,0)$}][7brane]at(-2,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver one reads from this brane web is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=below:{1}][u](2)[below left of=1]{};
\node[label=below:{1}][u](3)[below right of=1]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--node[above]{3}++(1);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\label{SU4_0+2F}
\end{equation}
which we claim to be the same theory, in a sense to be clarified shortly, as MQ$_1^{1,1,0}$.
In addition, we can consider, yet another web diagram, for the infinite gauge coupling limit of SO(6)+1\textbf{S}+1\textbf{C}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](-1,0)--(0,0);
\draw[thick,dashed](2,0)--(3,0);
\draw[thick](0,0)--(2,0);
\draw[thick](-1,1)--(0,0);
\draw[thick](2,2)--(0,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,1){};
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\node[label=above:{$(2,1)$}][7brane]at(2,2){};
\node[label=below:{O5$^+$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(2.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(.5,0){};
\node[label=above:{$1$}]at(1.5,1.5){};
\node[label=above:{$1$}]at(-.5,.5){};
\node[label=above:{$2$}]at(.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver that one obtains from this brane system is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](2){};
\node[label=below:{1}][u](3)[right of=2]{};
\node[label=above:{1}][uf](2f)[above of =2]{};
\node[label=above:{3}][uf](3f)[above of =3]{};
\draw(2)--(3);
\draw(2)--(2f);
\draw(3)--(3f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
which is trivially the same as \eqref{SU4_0+2F}, upon removing the overall decoupled U(1).
Consider now, the brane web for the SCFT limit of SU$(4)_1$+2\textbf{F}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick] (-1,-1)--(0,0);
\draw[thick] (1,-1)--(0,0);
\draw[thick] (-2,1)--(0,0);
\draw[thick] (2,1)--(0,0);
\draw[thick](-2,0)--(2,0);
\draw[thick] (1,1)--(0,0);
\draw[thick] (-1,1)--(0,0);
\draw[thick] (2,-1)--(0,0);
\draw[thick] (-2,-1)--(0,0);
\node[7brane]at(-2,0){};
\node[7brane]at(2,0){};
\node[7brane]at(2,1){};
\node[label=left:{$(2,-1)$}][7brane]at(-2,1){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(-1,1){};
\node[7brane]at(2,-1){};
\node[label=left:{$(2,1)$}][7brane]at(-2,-1){};
\node[label=below:{$(1,-1)$}][7brane]at(1,-1){};
\node[7brane]at(-1,-1){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From here, one can read off the following pair of magnetic quivers
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\draw(1)--node[above]{4}++(1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\draw(1)--node[above]{2}++(1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\label{SU4_1+2F}
\end{equation}
which can easily be identified with $\MQ_1^{2,0,0}\cup\MQ_2^{2,0,0}$ \eqref{SO6:MQ200}, modulo field redefinitions. The Coulomb branch and Higgs branch Hilbert series of $\MQ_1^{1,1,0}$, $\MQ_1^{2,0,0}$ and $\MQ_2^{2,0,0}$ match with their unitary counterparts and the unrefined results are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG for these magnetic quivers takes a very simple form and is given by
\begin{align}
\HWG_\mathcal{C}\left(\MQ_1^{1,1,0}\right)&= \HWG_\mathcal{C}\left(\ref{SU4_0+2F}\right) =\PE\left[(1+\mu^2)t^2+(q+q^{-1})\mu t^4-\mu^2t^8\right]\ , \nonumber \\
\HWG_\mathcal{C}\left(\MQ_1^{2,0,0}\right)&=\HWG_\mathcal{C}^1\left(\ref{SU4_1+2F}\right)=\PE\left[t^2+(q+q^{-1})t^4-t^8\right]\ , \nonumber \\
\HWG_\mathcal{C}\left(\MQ_2^{2,0,0}\right)&=\HWG_\mathcal{C}^2\left(\ref{SU4_1+2F}\right)=\PE\left[\mu^2t^2\right] \ ,
\end{align}
where in the above, we use $\mu$ to denote the highest weight fugacity for SU$(2)$, and $q$ as a U(1) fugacity.
We will now present an intuitive argument, for the agreement of the Hilbert series of the two apparently different quivers in \eqref{SO6:MQ110} and \eqref{SU4_0+2F}. These quivers actually describe the same $3$d $\mathcal{N}=4$ theory. The only reason for which they look different, is that the two abelian quivers are written considering U$(1)$ charges of the hypermultiplets under two different sets of U$(1)$s. Said in other words, a simple change of basis in the Lie algebra of U$(1)^2$ can recast the table of charges of \eqref{SU4_0+2F} into that of \eqref{SO6:MQ110}. We collect the charges of the various hypermultiplets in tables \ref{tab:1} and \ref{tab:2}.
\begin{table}[h]
\centering
\begin{tabular}{c|c|c}
& U$(1)_1$ & U$(1)_2$ \\
\hline
$Q_1$ & $1$ & $-1$\\
\hline
$Q_2$ & $1$ & $1$\\
\hline
$Q_3$ & $2$ & $0$
\end{tabular}
\caption{Table of charges for the hypermultiplets in theory \eqref{SO6:MQ110}}
\label{tab:1}
\end{table}
\begin{table}[h]
\centering
\begin{tabular}{c|c|c|c|c}
& $\tilde{\text{U}}(1)_1$ & $\tilde{\text{U}}(1)_2$ & $a\tilde{\text{U}}(1)_1+b\tilde{\text{U}}(1)_2$ & $c\tilde{\text{U}}(1)_1+d\tilde{\text{U}}(1)_2$ \\
\hline
$\tilde{Q}_1$ & $1$ & $0$ & $a$ & $c$ \\
\hline
$\tilde{Q}_2$ & $1$ & $-1$ & $a-b$ & $c-d$\\
\hline
$\tilde{Q}_3$ & $0$ & $1$ & $b$ & $d$
\end{tabular}
\caption{Table of charges for the hypermultiplets in theory \eqref{SU4_0+2F}. In the last two columns we define two different U$(1)$s, whose Lie algebra generators are a linear combination of the generators of $\tilde{\text{U}}(1)_1$ and $\tilde{\text{U}}(1)_2$.}
\label{tab:2}
\end{table}
We notice that we can reproduce the U$(1)$ charges of table \ref{tab:1} by taking
$a=1, \ b=c=d=-1$ in table \ref{tab:2}. We will then identify $Q_1$ with $\tilde{Q}_1$, $Q_3$ with $\tilde{Q}_2$, and finally $Q_2$ with the charge conjugate of $\tilde{Q}_3$. In particular we remark that it is consistent to identify $Q_2$ and the charge conjugate of $\tilde{Q}_3$, as each hypermultiplet consists of two chiral multiplets in conjugated gauge representations.
\subsubsection{SO(6)+2\textbf{S}+1\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_\frac{1}{2}$}{TEXT}+3\textbf{F}}
The brane web for SO(6)+2\textbf{S}+1\textbf{C} at infinite gauge coupling limit is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](-1,0)--(2,0);
\draw[thick](2,0)--(.5,.75);
\node[label=left:{(2,-1)}][7brane]at (.5,.75){};
\draw[thick](2,0)--(4,2);
\draw[thick](2,0)--(1,1);
\node[label=above:{(1,-1))}][7brane]at(1,1){};
\node[label=above:{(1,1)}][7brane]at (4,2){};
\node[7brane]at (3,1){};
\draw[thick](2,0)--(4,0);
\node[7brane]at (4,0){};
\node[7brane]at (3,0){};
\draw[thick,dashed](5,0)--(4,0);
\node[label=below:{O5$^-$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\node[label=below:{1}]at(2.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=below:{1}]at(.75,.75){};
\node[label=above:{2}]at(2.5,.5){};
\node[label=above:{1}]at(3.5,1.5){};
\node[label=above:{1}]at(1.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
from which we read off the following magnetic quiver
\begin{equation}
\MQ_1^{2,1,0}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=left:{1}][u](1){};
\node[label=above:{1}][uf](uf)[above of=1]{};
\node[label=below:{1}][u](2)[below right of=1]{};
\node[label=below:{1}][u](3)[below left of=1]{};
\draw(2)--(3);
\draw[dashed](3)--(1);
\draw(1)--node[above] {3}++(2);
\path [draw,snake it](1)--(uf);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ210}
\end{equation}
Next, we consider the web diagram for SU$(4)_0$+3\textbf{F}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-2,0)--(2,0);
\draw[thick](-1,1)--(1,-1);
\draw[thick](-2,-1)--(0,0);
\draw[thick](1,1)--(0,0);
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\node[label=below:{$(1,-1)$}][7brane]at(1,-1){};
\node[label=above:{(1,1)}][7brane]at(1,1){};
\node[label=below:{(2,1)}][7brane]at(-2,-1){};
\node[7brane]at(-2,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[label=below:{1}]at(1.75,0.125){};
\node[label=below:{2}]at(.75,0.125){};
\node[label=below:{1}]at(-1.75,0.125){};
\node[label=below:{1}]at(-1.5,-.75){};
\node[label=below:{1}]at(.5,-.5){};
\node[label=above:{1}]at(-.5,.5){};
\node[label=above:{1}]at(.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
whose magnetic quiver is readily obtained to be
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](2)[right of=1]{};
\node[label=above:{1}][u](3)[above of=2]{};
\node[label=above:{1}][u](4)[left of=3]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(4)--node[left] {3}++(1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_0+3F}
\end{equation}
The unrefined computations of the Coulomb branch and Higgs branch Hilbert series for the two quivers match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The Coulomb branch and the Higgs branch of the above two quivers are identical with the following HWG:
\begin{equation}
\HWG_\mathcal{C}=\HWG_\mathcal{H}=\PE\left[(1+\mu_1\mu_2)t^2+\left(\mu_1q+\mu_2q^{-1}\right)t^4-\mu_1\mu_2t^8\right]~,
\end{equation}
where $\mu_i$ are the highest weight fugacities of SU(3) and $q$ is the U(1) charge. We remark that the duality between \eqref{SU4_0+3F} and \eqref{SO6:MQ210} generalises to the case in which the three bifundamental hypermultiplets among the U(1) nodes are replaced with an arbitrary number $n>3$ of them.
\subsubsection{SO(6)+2\textbf{S}+2\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_0$}{TEXT}+4\textbf{F}}
There are two inequivalent brane webs with O5-plane that realise SO(6)+2\textbf{S}+2\textbf{C}. The first web diagram, involves a configuration with the orientifold plane asymptotically an O5$^-$ on both ends, corresponding to the 5d electric quiver $[1\textbf{S}+1\textbf{C}]-\text{SO}(6)-[1\textbf{S}+1\textbf{C}]$
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(5,0);
\draw[thick,dashed](5,0)--(6,0);
\draw[thick](3,0)--(5,2);
\draw[thick](3,0)--(1,2);
\node[7brane] at (1,0){};
\node[7brane] at (2,0){};
\node[7brane] at (4,0){};
\node[7brane] at (5,0){};
\node[label=above:{$(1,-1)$}][7brane] at (1,2){};
\node[label=above:{$(1,1)$}][7brane] at (5,2){};
\node[7brane] at (2,1){};
\node[7brane] at (4,1){};
\node[label=below:{O5$^-$}]at(0.5,0){};
\node[label=below:{O5$^-$}]at(5.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$1$}]at(3.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(4.5,0){};
\node[label=above:{2}]at(2.5,.5){};
\node[label=above:{1}]at(1.5,1.5){};
\node[label=above:{2}]at(3.5,.5){};
\node[label=above:{1}]at(4.5,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The corresponding magnetic quiver is
\begin{equation}
\MQ_1^{2,2,0}\cup\MQ_2^{2,2,0}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so4){};
\node[label=below left:{2}][u](u2)[above of=so4]{};
\node[label=above:{1}][u](u1)[left of=u2]{};
\node[label=above:{1}][u](u1')[right of=u2]{};
\draw(so4)--(u2);
\draw(u2)--(u1);
\draw(u2)--(u1');
\draw (u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{$2\;\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](u1){};
\node[label=above:{4}][spf](u1')[above of=u1]{};
\draw(u1)--(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ220}
\end{equation}
Alternatively, one can consider a brane web with an O5-plane which is asymptotically an O5$^+$-plane on one end, and an O5$^-$-plane on the other end
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(7,0);
\draw[thick,dashed](8,0)--(7,0);
\draw[thick](1,0)--(0,1);
\draw[thick](1,0)--(1,1);
\node[7brane][label=above:{$(1,-1)$}]at(0,1){};
\node[7brane]at (1,1){};
\node[7brane]at (2,0){};
\node[7brane]at (3,0){};
\node[7brane]at (4,0){};
\node[7brane]at (5,0){};
\node[7brane]at (6,0){};
\node[7brane]at (7,0){};
\node[label=below:{O5$^+$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=below:{$3$}]at(1.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(2.5,0){};
\node[label=below:{$2$}]at(3.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(4.5,0){};
\node[label=below:{$1$}]at(5.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(6.5,0){};
\node[label=above:{$1$}]at(.5,0.5){};
\node[label=right:{$2$}]at(1,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
corresponding to the 5d IR quiver $\text{SO}(6)-[2\textbf{S}+2\textbf{C}]$. From this web diagram we find an alternative magnetic quiver
\begin{equation}
\widehat{\MQ}_1^{2,2,0}\cup\widehat{\MQ}_2^{2,2,0}=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=below:{2}][sp](sp2)[right of=u1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so3]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=above:{2}][uf](uf)[above of=u1]{};
\node[label=above:{1}][sof](sof)[above of=sp2]{};
\node[label=above:{1}][sof](sof')[above of=sp2']{};
\draw(u1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp2');
\draw(sp2')--(so2);
\draw(u1)--(uf);
\draw(sp2)--(sof);
\draw(sp2')--(sof');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=above:{6}][sof](sof)[above of=sp2]{};
\draw(sp2)--(sof);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQtilde220} ~.
\end{equation}
Let us now consider the fixed point limit of the web diagram for SU$(4)_0$+4\textbf{F}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](1,1)--(-1,-1);
\draw[thick](-1,1)--(1,-1);
\draw[thick](0,0)--(3,0);
\draw[thick](0,0)--(-3,0);
\node[7brane]at(1.5,0){};
\node[7brane]at(3,0){};
\node[7brane]at(-1.5,0){};
\node[7brane]at(-3,0){};
\node[label=above:{(1,1)}][7brane]at(1,1){};
\node[label=above:{(1,-1)}][7brane]at(-1,1){};
\node[label=below:{(1,-1)}][7brane]at(1,-1){};
\node[label=below:{(1,1)}][7brane]at(-1,-1){};
\node[label=below:{1}]at(-2,0){};
\node[label=below:{2}]at(-1,0){};
\node[label=below:{1}]at(2,0){};
\node[label=below:{2}]at(1,0){};
\node[label=above:{1}]at(-.5,.5){};
\node[label=above:{1}]at(.5,.5){};
\node[label=below:{1}]at(-.5,-.5){};
\node[label=below:{1}]at(.5,-.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
The magnetic quivers for this brane system are
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{1}][u](1')[right of=2]{};
\node[label=above:{1}][u](11)[above left of=2]{};
\node[label=above:{1}][u](11')[above right of=2]{};
\draw(1)--(2);
\draw(2)--(1');
\draw(2)--(11);
\draw(2)--(11');
\draw[double distance=2pt](11)--(11');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=above:{1}][u](u1')[above of=u1]{};
\draw(u1)--node[right] {4}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SU4_0+4F}
\end{equation}
The unrefined Coulomb branch and Higgs branch Hilbert series for the orthosymplectic quivers in \eqref{SO6:MQ220} and \eqref{SO6:MQtilde220} match with their corresponding unitary quivers in \eqref{SU4_0+4F} and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. Moreover, we know the HWG of the Coulomb branch of the two cones of \eqref{SU4_0+4F} as
\begin{align}
\text{HWG}_{\mathcal{C}}^1(\ref{SU4_0+4F}) &= \text{PE}[t^2(1+\mu_1 \mu_3)+t^4(\mu_2^2+q_1 \mu_2+q_1^{-1}\mu_2)-t^8 \mu_{2}^2] \nonumber \\
\text{HWG}_{\mathcal{C}}^2(\ref{SU4_0+4F}) &= \text{PE}[t^2+t^4(q_2+q_2^{-1})-t^8] ~,
\end{align}
where $\mu_i$ are the highest weight fugacities of SU(4) and $q_1$ and $q_2$ are U(1) charges. Thus, we can write the HWG for the orthosymplectic quivers as:
\begin{align}
\text{HWG}_{\mathcal{C}}(\MQ_1^{2,2,0}) = \text{HWG}_{\mathcal{C}}(\widehat{\MQ}_1^{2,2,0}) &= \text{HWG}_{\mathcal{C}}^1(\ref{SU4_0+4F}) \nonumber \\
\text{HWG}_{\mathcal{C}}(\MQ_2^{2,2,0}) = \text{HWG}_{\mathcal{C}}(\widehat{\MQ}_2^{2,2,0}) &= \text{HWG}_{\mathcal{C}}^2(\ref{SU4_0+4F}) ~.
\end{align}
\subsubsection{SO(6)+3\textbf{S}+2\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_\frac{1}{2}$}{TEXT}+5\textbf{F}}
The brane web for SO(6)+3\textbf{S}+2\textbf{C} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(-1,1);
\draw[thick](1,0)--(7,0);
\draw[thick,dashed](8,0)--(7,0);
\draw[thick](1,0)--(0,1);
\draw[thick](1,0)--(1,1);
\node[7brane][label=above:{$(1,-1)$}]at(0,1){};
\node[7brane][label=above:{$(2,-1)$}]at(-1,1){};
\node[7brane]at (1,1){};
\node[7brane]at (2,0){};
\node[7brane]at (3,0){};
\node[7brane]at (4,0){};
\node[7brane]at (5,0){};
\node[7brane]at (6,0){};
\node[7brane]at (7,0){};
\node[label=below:{O5$^-$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=below:{$3$}]at(1.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(2.5,0){};
\node[label=below:{$2$}]at(3.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(4.5,0){};
\node[label=below:{$1$}]at(5.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(6.5,0){};
\node[label=above:{$1$}]at(.5,0.5){};
\node[label=right:{$2$}]at(1,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From this brane web we obtain the following magnetic quivers
\begin{equation}
\MQ_1^{3,2,0}\cup\MQ_2^{3,2,0}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=right:{1}][u](u1)[above of=sp2']{};
\node[label=left:{1}][u](u1')[above of=so4]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(u1);
\draw(so4)--(u1'); \draw[dashed](u1)--node[above]{2}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][so](so2')[right of=sp2]{};
\node[label=left:{1}][u](u1)[above of=sp2]{};
\node[label=above:{2}][sp](sp2')[above of=u1]{};
\draw(so2)--(sp2);
\draw(sp2)--(so2');
\draw(sp2)--(u1);
\draw(u1)--node[left]{3}++(sp2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQ320}
\end{equation}
On the other hand, the brane web for SU$(4)_\frac{1}{2}$+5\textbf{F} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-3,0);
\draw[thick](0,0)--(2,0);
\draw[thick](0,0)--(1,1);
\draw[thick](1,-1)--(-1,1);
\draw[thick](0,0)--(0,-1);
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(0,-1){};
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[label=below:{$(1,-1)$}][7brane]at(1,-1){};
\node[label=below:{1}]at(-2.5,0){};
\node[label=below:{2}]at(-1.5,0){};
\node[label=below:{3}]at(-.5,0){};
\node[label=above:{2}]at(.75,0){};
\node[label=above:{1}]at(1.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver one reads off from this ordinary brane web is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{2}][u](2')[right of=2]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{1}][u](u1)[above of=2]{};
\node[label=above:{1}][u](u1')[above of=2']{};
\draw(1)--(2);
\draw(2)--(2');
\draw(2')--(1');
\draw(2)--(u1);
\draw(2')--(u1');
\draw(u1)--node[above]{2}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\node[label=left:{1}][u](2)[above left of=1]{};
\node[label=right:{1}][u](2')[above right of=1']{};
\node[label=above:{1}][u](3)[above right of=2]{};
\node[label=above:{1}][u](3')[above left of=2']{};
\draw(1)--(1');
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--node[above]{3}++(3');
\draw(3')--(2');
\draw(2')--(1');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_1/2+5F}
\end{equation}
The unrefined Coulomb branch and Higgs branch Hilbert series for the orthosymplectic quivers in \eqref{SO6:MQ320} match with their corresponding unitary quivers in \eqref{SU4_1/2+5F} and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG of the Coulomb branch of the two cones of \eqref{SO6:MQ320} can therefore be predicted to be the same as the HWG of the unitary quivers in \eqref{SU4_1/2+5F}
\begin{align}
\HWG_\mathcal{C}^1(\ref{SU4_1/2+5F}) &=\PE\left[t^2+(\mu_2q+\mu_3q^{-1})t^4+\sum_{k=1}^2\mu_k\mu_{5-k}t^{2k}-\mu_2\mu_3t^8\right]\ , \nonumber \\
\HWG_\mathcal{C}^2(\ref{SU4_1/2+5F}) &=\PE\left[(1+\mu_1\mu_4)t^2+(\mu_1q+\mu_4q^{-1})t^4-\mu_1\mu_4t^8\right] \ ,
\end{align}
where $\mu_i$ are highest weight fugacities of SU(5). Thus from the above discussion, we can conjecture that:
\begin{align}
\HWG_\mathcal{C}(\MQ_1^{3,2,0}) &= \HWG_\mathcal{C}^1(\ref{SU4_1/2+5F}) \ ,\nonumber \\
\HWG_\mathcal{C}(\MQ_2^{3,2,0}) &= \HWG_\mathcal{C}^2(\ref{SU4_1/2+5F}) \ .
\end{align}
\subsubsection{SO(6)+3\textbf{S}+3\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_0$}{TEXT}+6\textbf{F}}
The brane web for SO(6)+3\textbf{S}+3\textbf{C} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick, dashed](-2,0)--(-1,0);
\draw[thick, dashed](7,0)--(8,0);
\draw[thick](-1,0)--(7,0);
\node[7brane]at(-1,0){};
\node[7brane]at(0,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,1){};
\draw[thick](5,0)--(7,2);
\draw[thick](5,0)--(5,1);
\node[7brane]at(5,1){};
\node[label=above:{$(1,1)$}][7brane]at(7,2){};
\node[label=below:{O5$^-$}]at(-1.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(-.5,0){};
\node[label=below:{$1$}]at(.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(5.5,0){};
\node[label=below:{$3$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=above:{1}]at(6.5,1.5){};
\node[label=above:{2}]at(5.5,.5){};
\node[label=left:{2}]at(5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From here we read off the magnetic quiver to be
\begin{equation}
\MQ_1^{3,3,0}\cup\MQ_2^{3,3,0}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{2}][so](so2')[right of=sp4]{};
\node[label=left:{2}][u](u2)[above of=sp4]{};
\node[label=above:{1}][u](u1)[above of=u2]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so2');
\draw(sp4)--(u2);
\draw(u2)--(u1);
\draw(u2)to[out=45,in=-45,loop,looseness=10](u2);
\node[right of=u2]{1 $\wedge^2$};
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{2}][sp](sp22)[above of=so4]{};
\node[label=above:{2}][sp](sp22')[above of=so2']{};
\node[label=left:{1}][u](u1)[above of=sp22]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2');
\draw(so4)--(sp22);
\draw(so2')--(sp22');
\draw(sp22)--(sp22');
\draw(sp22)--(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQ330}
\end{equation}
On the other hand, the brane web for SU$(4)_0$+6\textbf{F} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(6,0);
\node[7brane]at(0,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\draw[thick](3,-1)--(3,1);
\draw[thick](2,1)--(4,-1);
\node[7brane]at(3,-1){};
\node[7brane]at(3,1){};
\node[label=above:{$(1,-1)$}][7brane]at(2,1){};
\node[label=below:{$(1,-1)$}][7brane]at(4,-1){};
\node[label=below:{1}]at(.5,0){};
\node[label=below:{2}]at(1.5,0){};
\node[label=below:{3}]at(2.5,0){};
\node[label=above:{3}]at(3.5,0){};
\node[label=above:{2}]at(4.5,0){};
\node[label=above:{1}]at(5.5,0){};
\node[label=right:{1}]at(3,-.75){};
\node[label=left:{1}]at(3,.75){};
\node[label=right:{1}]at(3.7,-.7){};
\node[label=left:{1}]at(2.3,.7){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
from which we obtain the magnetic quiver
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{3}][u](3)[right of=2]{};
\node[label=below:{2}][u](2')[right of=3]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{1}][u](u1)[above left of=3]{};
\node[label=above:{1}][u](u1')[above right of=3]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(2');
\draw(2')--(1');
\draw(3)--(u1);
\draw(3)--(u1');
\draw(u1)--(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{2}][u](3)[right of=2]{};
\node[label=below:{2}][u](2')[right of=3]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{1}][u](u1)[above of=2]{};
\node[label=above:{1}][u](u1')[above of=2']{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(2');
\draw(2')--(1');
\draw(2)--(u1);
\draw(2')--(u1');
\draw[double distance=2pt](u1)--(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SU4_0+6F}
\end{equation}
The unrefined Coulomb branch and Higgs branch Hilbert series for the orthosymplectic quivers in \eqref{SO6:MQ330} match with their corresponding unitary quivers in \eqref{SU4_0+6F} and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG of the Coulomb branch of the two cones of \eqref{SU4_0+6F} are given by\cite{Bourget:2019rtl}:
\begin{align}
\HWG_\mathcal{C}^1(\ref{SU4_0+6F}) &=\PE\left[t^2+(q+q^{-1})\mu_3t^4+\sum_{k=1}^3\mu_k\mu_{6-k}t^{2k}-\mu_3^2t^8\right] \nonumber \\
\HWG_\mathcal{C}^2(\ref{SU4_0+6F}) &=\PE\left[t^2+(\mu_4q+\mu_2q^{-1})t^4+\sum_{k=1}^2\mu_k\mu_{6-k}t^{2k}-\mu_2\mu_4t^8\right] ~,
\end{align}
where $\mu_i$ are the highest weight fugacities of SU(6). Thus, from the above discussion, we can conjecture that:
\begin{align}
\HWG_\mathcal{C}(\MQ_1^{3,3,0}) &= \HWG_\mathcal{C}^1(\ref{SU4_0+6F}) \nonumber \\
\HWG_\mathcal{C}(\MQ_2^{3,3,0}) &= \HWG_\mathcal{C}^2(\ref{SU4_0+6F}) ~.
\end{align}
\subsubsection{SO(6)+4\textbf{S}+4\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_0$}{TEXT}+8\textbf{F}}
The brane web for SO(6)+4\textbf{S}+4\textbf{C} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick,dashed](14,0)--(13,0);
\draw[thick](1,0)--(13,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(8,0){};
\node[7brane]at(9,0){};
\node[7brane]at(10,0){};
\node[7brane]at(11,0){};
\node[7brane]at(12,0){};
\node[7brane]at(13,0){};
\draw[thick](7,0)--(7,2);
\node[7brane]at(7,1){};
\node[7brane]at(7,2){};
\node[label=below:{O5$^-$}]at(.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(5.5,0){};
\node[label=below:{$3$}]at(6.5,0){};
\node[label=below:{$3$}]at(7.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(8.5,0){};
\node[label=below:{$2$}]at(9.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(10.5,0){};
\node[label=below:{$1$}]at(11.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(12.5,0){};
\node[label=below:{O5$^-$}]at(13.5,0){};
\node[label=right:{$4$}]at(7,.5){};
\node[label=right:{$2$}]at(7,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
whose corresponding magnetic quiver is
\begin{equation}
\MQ_1^{4,4,0}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{6}][so](so6)[right of=sp4]{};
\node[label=below:{4}][sp](sp4')[right of=so6]{};
\node[label=below:{4}][so](so4')[right of=sp4']{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{4}][sp](sp4'')[above of=so6]{};
\node[label=left:{2}][u](u2)[above of=sp4'']{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so6);
\draw(so2')--(sp2');
\draw(sp2')--(so4');
\draw(so4')--(sp4');
\draw(sp4')--(so6);
\draw(so6)--(sp4'');
\draw(sp4'')--(u2);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ440}
\end{equation}
The brane web for SU$(4)_0$+8\textbf{F} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(8,0);
\node[7brane]at(0,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(8,0){};
\draw[thick](4,2)--(4,-2);
\node[7brane]at(4,2){};
\node[7brane]at(4,1){};
\node[7brane]at(4,-2){};
\node[7brane]at(4,-1){};
\node[label=below:{1}]at(.5,0){};
\node[label=below:{2}]at(1.5,0){};
\node[label=below:{3}]at(2.5,0){};
\node[label=below:{4}]at(3.5,0){};
\node[label=above:{4}]at(4.5,0){};
\node[label=above:{3}]at(5.5,0){};
\node[label=above:{2}]at(6.5,0){};
\node[label=above:{1}]at(7.5,0){};
\node[label=left:{1}]at(4,1.5){};
\node[label=left:{2}]at(4,.5){};
\node[label=right:{1}]at(4,-1.5){};
\node[label=right:{2}]at(4,-.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
from here we obtain the following magnetic quiver
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{3}][u](3)[right of=2]{};
\node[label=below:{4}][u](4)[right of=3]{};
\node[label=below:{3}][u](3')[right of=4]{};
\node[label=below:{2}][u](2')[right of=3']{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{2}][u](22)[above of=4]{};
\node[label=above:{1}][u](11)[left of=22]{};
\node[label=above:{1}][u](11')[right of=22]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(4)--(3');
\draw(3')--(2');=
\draw(2')--(1');
\draw(4)--(22);
\draw(22)--(11);
\draw(11')--(22);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SU4_0+8F}
\end{equation}
The perturbative results of unrefined Coulomb branch and Higgs branch Hilbert series for the quiver \eqref{SO6:MQ440} match with that of \eqref{SU4_0+8F} and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG of the Coulomb branch of the unitary quiver was computed in \cite{Ferlito:2017xdq}. Thus, we can conjecture the following HWG for the orthosymplectic quiver:
\begin{equation}
\HWG_{\mathcal{C}}(\MQ_1^{4,4,0})=\PE\left[\sum_{i=1}^4\mu_i\mu_{8-i}t^{2i}+\left(\nu^2+\lambda^2\right)t^2+t^4+\mu_4\nu\lambda (t^4+t^6)-\nu^2\lambda^2\mu_4^2t^{12}\right]\;,
\end{equation}
where $\mu_i$ are SU$(8)$ highest weight fugacities, while $\nu$ and $\lambda$ are SU$(2)$ highest weight fugacities.
\subsection{SO(6) theories with a single vector hypermultiplet}
We will now begin to examine SO(6) gauge theories with a single hypermultiplet in the vector representation, as well as hypermultiplets in the (conjugate) spinor representations. Accordingly, we will be looking at the magnetic quivers for SU$(4)_\kappa$ gauge theory with a single second rank antisymmetric hypermultiplet, as well as fundamental hypermultiplets. As we shall see, the magnetic quivers do not always have a simple HWG, unlike the previously examined cases. Nevertheless, some magnetic quivers still have a HWG that can be written down in closed form.
\subsubsection{SO(6)+2\textbf{S}/(1\textbf{S}+1\textbf{C})+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_{1/0}$}{TEXT}+2\textbf{F}+1\textbf{AS}}\label{sec:SO6+1S1C1V}
Consider SO(6) gauge theory with two spinors or one spinor and a conjugate spinor, in addition to a vector.
Both of the theories can be realized by a brane web of the same shape (see the discussion around figure \ref{fig:SorC}) and the brane web
at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(-2,0);
\draw[thick](0,0)--(2,0);
\draw[thick,dashed](2,0)--(3,0);
\draw[thick](0,0)--(-2,1);
\draw[thick](0,0)--(-1,1);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(0,1);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[label=right:{$(1,1)$}][7brane]at(1,1){};
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\node[label=left:{$(2,-1)$}][7brane]at(-2,1){};
\node[7brane]at(0,1){};
\node[label=below:{O5$^-$}]at(-1,0){};
\node[label=below:{2}]at(.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{O5$^-$}]at(2.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From here
we find three subdivisions and the magnetic quivers obtained from the subdivisions are
\begin{equation}
\MQ_1^{*,*,1}\cup\MQ_2^{*,*,1} \cup \MQ_3^{*,*,1}=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=right:{2}][sp](sp2){};
\node[label=above:{1}][u](u1)[above left of=sp2]{};
\node[label=below:{1}][u](u1')[below left of=sp2]{};
\node[label=left:{1}][uf](f)[left of=u1']{};
\draw(sp2)--(u1);
\draw(sp2)--(u1');
\draw(u1)--(u1');
\draw[dashed](u1)to[out=-135, in=135]node[left]{2}(u1');
\path[draw, snake it](u1')--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u1){};
\node[label=above:{2}][sp](sp2)[right of=u1]{};
\node[label=left:{2}][uf](f)[below of=u1]{};
\draw(u1)--node[above]{2}(sp2);
\path[draw, snake it](u1)--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=below:{1}][u](2)[below of=1]{};
\draw[dashed](1)to[out=-135, in=135]node[left]{2}(2);
\draw(1)to[out=-45, in=45]node[right]{3}(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ201starred}
\end{equation}
Then a question is which of the magnetic quivers are for SO(6)+2\textbf{S}+1\textbf{V} or SO(6)+1\textbf{S}+1\textbf{C}+1\textbf{V}. In section \ref{subsubsec:SO62S}, we encountered a similar situation and at that time some subdivisions corresponded to the magnetic quivers of one theory and the others corresponded to those of the other theory. In this case it turns out that the situation is different due to the presence of bad gauge nodes in the magnetic quivers.
In order to resolve this issue, we consider the magnetic quivers from their unitary counterparts. SO(6)+1\textbf{S}+1\textbf{C}+1\textbf{V} corresponds to SU$(4)_0$+2\textbf{F}+1\textbf{AS} and its brane web at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-2,0);
\draw[thick](0,0)--(2,-2);
\draw[thick](0,0)--(-2,-1);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(0,1);
\node[label=below:{$(2,1)$}][7brane]at(-2,-1){}; \node[label=below:{$(1,-1)$}][7brane]at(2,-2){}; \node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(0,1){};
\node[7brane]at(1,-1){};
\node[label=above:{1}]at(-1.5,0){};
\node[label=above:{2}]at(-.5,0){};
\node[label=below:{1}]at(-1.5,-.75){};
\node[label=below:{2}]at(.5,-0.5){};
\node[label=below:{1}]at(1.5,-1.5){};
\node[label=below:{2}]at(.75,0.75){};
\node[label=right:{1}]at(0,0.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The diagram admits two subdivisions and the magnetic quivers that one reads off from this brane web are
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](2)[right of=1]{};
\node[label=above:{1}][u](3)[above of=2]{};
\node[label=above:{1}][u](4)[above of=1]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(4)--(1);
\draw(1)--node[above left]{2}++(3);
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad\cup\quad\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=left:{1}][u](1){};
\node[label=right:{1}][u](2)[below right of=1]{};
\node[label=right:{1}][u](3)[above right of=1]{};
\draw(1)--(2);
\draw(2)--node[right]{2}++(3);
\draw(3)--(1);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{SU4_0+2F+1AS}
\end{equation}
On the other hand SO(6)+2\textbf{S}+1\textbf{V} corresponds to SU$(4)_1$+2\textbf{F}+1\textbf{AS} and its brane web at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[7brane]at(-1,1){};
\node[7brane]at(-2,2){};
\node[7brane]at(-1,-1){};
\node[7brane]at(1,1){};
\node[7brane]at(1,-1){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\draw[thick](0,0)--(2,0);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(1,-1);
\draw[thick](0,0)--(-1,-1);
\draw[thick](0,0)--(-2,2);
\node[label=above:{2}]at(0.5,-0.1){};
\node[label=above:{1}]at(1.5,0){};
\node[label=above:{1}]at(0.5,0.5){};
\node[label=below:{1}]at(0.5,-0.5){};
\node[label=below:{2}]at(-0.5,-0.5){};
\node[label=above:{2}]at(-0.5,0.5){};
\node[label=above:{1}]at(-1.5,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
This brane web has three subdivisions and the corresponding magnetic quivers are
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](2)[right of=1]{};
\node[label=above:{1}][u](3)[above of=2]{};
\node[label=above:{1}][u](4)[above of=1]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--node[above]{2}++(4);
\draw(4)--(1);
\draw(1)--(3);
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad\cup\quad\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=above:{1}][u](2)[right of=1]{};
\node[label=above:{1}][u](3)[right of=2]{};
\draw(1)--node[above]{2}++(2);
\draw(2)--node[above]{2}++(3);
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad\cup\quad\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=above:{1}][u](2)[right of=1]{};
\node[label=above:{1}][u](3)[right of=2]{};
\draw(1)--node[above]{3}++(2);
\draw(2)--node[above]{2}++(3);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{SU4_1+2F+1AS}
\end{equation}
Hence we have five subdivisions for the two theories in total and the five subdivisions need to fit into the three subdivisions in \eqref{SO6:MQ201starred}.
Among the three magnetic quivers in \eqref{SO6:MQ201starred}, the last quiver $\MQ_3^{*,*,1} $ is a good theory and one can compute the Coulomb branch Hilbert series by using the monopole formula. Then it is possible to see that the Hilbert series agrees with that of the rightmost magnetic quiver in \eqref{SU4_1+2F+1AS}.
When the magnetic quivers contain bad gauge nodes the correspondence could be more involved. In bad theories, the Coulomb branch may not be simply a cone. For example the Coulomb branch of USp($2N$) SQCD with $2N$ flavours has two most singular points and the theory flows to an interacting SCFT at low energies at each point \cite{Assel:2018exy}. Both the first and the second magnetic quivers in \eqref{SO6:MQ201starred} contain a bad USp(2) gauge node with 2 effective flavours (which is the $N=1$ case of aforementioned example), it is natural to expect that the full Coulomb branch of the bad magnetic quivers is not a single cone. On the other hand, all the magnetic quivers in \eqref{SU4_1+2F+1AS} are good theories and so their Coulomb branch is expected to be a single cone. Hence, the Coulomb branch of the first and the second magnetic quivers in \eqref{SO6:MQ201starred} seem to be different from the magnetic quivers in \eqref{SU4_0+2F+1AS} and \eqref{SU4_1+2F+1AS}.
From the viewpoint of the Higgs branch of 5d theories, we expect that the Higgs branch of 5d SO(6)+2\textbf{S}+1\textbf{V}/SO(6)+1\textbf{S}+1\textbf{C}+1\textbf{V} or SU$(4)_1$+2\textbf{F}+1\textbf{AS}/SU$(4)_0$+2\textbf{F}+1\textbf{AS} at the infinite coupling is a union of cones and the origins of the cones are located at the same point where a 5d SCFT is realized. While this picture is consistent with the expectation for the Coulomb branch of the magnetic quivers in \eqref{SU4_0+2F+1AS} and \eqref{SU4_1+2F+1AS}, this does not agree with expectation for the Coulomb branch of the bad magnetic quivers in \eqref{SO6:MQ201starred}. Hence we argue that the Higgs branch of the 5d theories is given by the Coulomb branch of the good unitary quivers in \eqref{SU4_0+2F+1AS} and \eqref{SU4_1+2F+1AS}. As for the bad quivers in \eqref{SO6:MQ201starred}, which contain the bad USp(2) gauge node, we need to extract a part which corresponds to the Coulomb branch of the good unitary quivers from the Coulomb branch of the bad quivers in order to reproduce the Higgs branch of the 5d theories.
We hereby propose a prescription (which we denote as `B2G') that extracts the magnetic quivers whose Coulomb branch agrees with that of the unitary counterparts from magnetic quivers which contain the USp(2) gauge node with two effective flavours. A similar prescription can be given to extract the good magnetic quivers from bad magnetic quivers which contain a USp(2$N$) gauge node with $2N$ effective flavours. We will discuss more about the $N>1$ case in an upcoming work \cite{HSbad}. To motivate the prescription, note that the Coulomb branch of the USp(2) gauge theory with two flavours has two most singular points. An SCFT is realized at low energies at each point and it is the same SCFT which arises from the U(1) gauge theory with two flavours \cite{Seiberg:1996nz, Ferlito:2016grh, Assel:2018exy}. Hence the local geometry in the vicinity of one of the most singular points of the Coulomb branch of USp(2)$+2\textbf{F}$, which is given by $\mathbb{C}^2/\mathbb{Z}_2$, can be described by the Coulomb branch of U(1)$+2\textbf{F}$. Then a natural way to change quivers which contain the bad USp(2) gauge nodes into good theories, is to replace the USp(2)$+2\textbf{F}$ with U(1)$+2\textbf{F}$. Namely, we use the local geometry near one of the most singular points of the Coulomb branch of USp(2)$+2\textbf{F}$ to construct the whole Coulomb branch moduli space. In order to embed U(1)$+2\textbf{F}$ into the original quiver, we need to gauge the flavour symmetry. However we need to be careful of the gauging since the flavour symmetry of U(1)$+2\textbf{F}$ is different from that of USp(2)$+2\textbf{F}$. The flavour symmetry of the USp(2) gauge theory with two flavours is SO(4) and either all or a part of the flavour symmetry is gauged in a magnetic quiver. We propose matter contributions to the dimension of the monopole operators in various gaugings. It turns out that there are two ways of implementing the matter contributions which can give different Coulomb branch Hilbert series. We will see that these different Coulomb branch Hilbert series agree with the Coulomb branch Hilbert series of the unitary quivers in \eqref{SU4_0+2F+1AS} and \eqref{SU4_1+2F+1AS}.
When SO(2)$\times$SO(2) is gauged, the prescription of replacing the bad USp(2) gauge node is given by the following two ways,
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=left:$\cdots$](0){};
\node[label=above:{2}][so](1)[right of=0]{};
\node[label=above:{2}][sp](2)[right of=1]{};
\node[label=above:{2}][so](3)[right of =2]{};
\node[label=right:$\cdots$](4)[right of=3]{};
\node[label=right:\contour{black}{$\xrightarrow{\text{B2G}}$}](5)[right of =4]{};
\node[](6)[right of=5]{};
\node[label=left:$\cdots$](7)[right of=6]{};
\node[label=above:$2$][so](8)[right of =7]{};
\node[label=above:$1$][u](9)[right of =8]{};
\node[label=above:$2$][so](10)[right of =9]{};
\node[label=right:$\cdots$](11)[right of=10]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(1)--(0);
\draw(3)--(4);
\draw(7)--(8);
\draw(10)--(11);
\draw(8)--node[above]{$\frac{1}{2}$}++(9);
\draw(9)--node[above]{$\frac{1}{2}$}++(10);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,\label{prescription1-}
\end{equation}
or
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=left:$\cdots$](0){};
\node[label=above:{2}][so](1)[right of=0]{};
\node[label=above:{2}][sp](2)[right of=1]{};
\node[label=above:{2}][so](3)[right of =2]{};
\node[label=right:$\cdots$](4)[right of=3]{};
\node[label=right:\contour{black}{$\xrightarrow{\text{B2G}}$}](5)[right of =4]{};
\node[](6)[right of=5]{};
\node[label=left:$\cdots$](7)[right of=6]{};
\node[label=above:$2$][so](8)[right of =7]{};
\node[label=above:$1$][u](9)[right of =8]{};
\node[label=above:$2$][so](10)[right of =9]{};
\node[label=right:$\cdots$](11)[right of=10]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(1)--(0);
\draw(3)--(4);
\draw(7)--(8);
\draw(10)--(11);
\draw(8)--node[above]{$\frac{1}{2}$}++(9);
\draw[dashed](9)--node[above]{$\frac{1}{2}$}++(10);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{prescription1+}
\end{equation}
Here the solid / dashed line between SO(2) and U(1), with $\frac{1}{2}$ above it, implies the bifundamental / fundamental-fundamental matter. In the Coulomb branch Hilbert series computations, the contribution to the dimension of the monopole operators from the matter is given by\footnote{Appropriate modifications are required while using the Molien-Weyl integral to compute the Higgs branch Hilbert series. We do not give explicit expressions, but it should be implicit from the expression of the monopole formula. \label{footnote1}}
\begin{align}
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][so](1){};
\node[label=above:$1$][u](2)[right of =1]{};
\draw(1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{2}|m_1 - m_2|,\\
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][so](1){};
\node[label=above:$1$][u](2)[right of =1]{};
\draw[dashed](1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{2}|m_1 + m_2|,
\end{align}
where $m_1$ is the magnetic flux for SO(2) and $m_2$ is the magnetic flux for U(1). Note that this is different from the matter represented by a straight line without $\frac{1}{2}$ between SO(2) and U(1) since in that case U(1) gauge node arises by gauging U(1) part of USp(2). When we consider the local part then the two configurations can be equivalently written as
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](2)[right of=1]{};
\node[label=above:{2}][so](3)[right of =2]{};
\node[label=right:{=}](4)[right of =3]{};
\node[](5)[right of=4]{};
\node[label=above:$1$][u](7)[right of =5]{};
\node[label=above:$1$][u](8)[right of =7]{};
\draw(2)--node[above]{$\frac{1}{2}$}++(3);
\draw(7)--(8);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,\label{u1u1-}
\end{equation}
and
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](2)[right of=1]{};
\node[label=above:{2}][so](3)[right of =2]{};
\node[label=right:{=}](4)[right of =3]{};
\node[](5)[right of=4]{};
\node[label=above:$1$][u](7)[right of =5]{};
\node[label=above:$1$][u](8)[right of =7]{};
\draw[dashed](2)--node[above]{$\frac{1}{2}$}++(3);
\draw[dashed](7)--(8);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{u1u1+}
\end{equation}
When SO(3) inside SO(4) is gauged, the prescription is the given by the following replacement,
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][so](1)[right of=0]{};
\node[label=above:{2}][sp](2)[right of=1]{};
\node[label=above:{3}][so](3)[right of =2]{};
\node[label=right:$\cdots$](4)[right of=3]{};
\node[label=right:\contour{black}{$\xrightarrow{\text{B2G}}$}](5)[right of =4]{};
\node[](6)[right of=5]{};
\node[label=above:$1$][so](7)[right of =6]{};
\node[label=above:$1$][u](8)[right of =7]{};
\node[label=above:$3$][so](9)[right of =8]{};
\node[label=right:$\cdots$](10)[right of=9]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(7)--node[above]{$\frac{1}{2}$}++(8);
\draw(8)--node[above]{$\frac{1}{2}$}++(9);
\draw(9)--(10);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{prescription2}
\end{equation}
Here the contribution from the matter represented by the solid line with $\frac{1}{2}$ on it between U(1) and SO(3) to the dimension of the monopole operators in the Coulomb branch Hilbert series computations is given by:
\begin{align}
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=above:$3$][so](2)[right of =1]{};
\draw(1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{2}|m_1 - m_2| + \frac{1}{4}|m_1|,\\
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][so](1){};
\node[label=above:$1$][u](2)[right of =1]{};
\draw(1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{4}|m_1|,
\end{align}
where $m_1$ is the magnetic flux of U(1) and $m_2$ is the magnetic flux of SO(3).
Finally when the full SO(4) is gauged then the prescription is given by the following replacement,
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][sp](2)[right of=1]{};
\node[label=above:{4}][so](3)[right of =2]{};
\node[label=right:$\cdots$](4)[right of=3]{};
\node[label=right:\contour{black}{$\xrightarrow{\text{B2G}}$}](5)[right of =4]{};
\node[](6)[right of=5]{};
\node[label=above:$1$][u](7)[right of =6]{};
\node[label=above:$4$][so](8)[right of =7]{};
\node[label=right:$\cdots$](9)[right of=8]{};
\draw(2)--(3);
\draw(3)--(4);
\draw(7)--node[above]{$\frac{1}{2}$}++(8);
\draw(8)--(9);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,\label{prescription3-}
\end{equation}
or
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][sp](2)[right of=1]{};
\node[label=above:{4}][so](3)[right of =2]{};
\node[label=right:$\cdots$](4)[right of=3]{};
\node[label=right:\contour{black}{$\xrightarrow{\text{B2G}}$}](5)[right of =4]{};
\node[](6)[right of=5]{};
\node[label=above:$1$][u](7)[right of =6]{};
\node[label=above:$4$][so](8)[right of =7]{};
\node[label=right:$\cdots$](9)[right of=8]{};
\draw(2)--(3);
\draw(3)--(4);
\draw[dashed](7)--node[above]{$\frac{1}{2}$}++(8);
\draw(8)--(9);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.\label{prescription3+}
\end{equation}
The contribution from the matter represented by the solid/dashed line with $\frac{1}{2}$ on it between U(1) and SO(4) to the dimension of the monopole operators in the Coulomb branch Hilbert series computations is given by (see footnote \ref{footnote1}):
\begin{align}
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=above:$4$][so](2)[right of =1]{};
\draw(1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{2}\left(|m_1 - m_{2,1}| + |m_1 - m_{2,2}|\right),\\
\Delta_{\text{hyp}}\left(
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=above:$4$][so](2)[right of =1]{};
\draw[dashed](1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\right) &= \frac{1}{2}\left(|m_1 - m_{2,1}| + |m_1 + m_{2,2}|\right),
\end{align}
where $m_1$ is the magnetic flux of U(1) and $m_{2,1}, m_{2,2}$ is the magnetic flux of SO(4).
We can apply the above prescription to the first and second magnetic quivers of \eqref{SO6:MQ201starred}. Each quiver gives rise to two magnetic quivers whether we use the replacement \eqref{prescription1-} or \eqref{prescription1+}. From the explicit evaluation of the Hilbert series of the moduli spaces, we find that the replacement \eqref{prescription1-} corresponds to the magnetic quivers of \eqref{SU4_1+2F+1AS} and the replacement \eqref{prescription1+} corresponds to the magnetic quivers of \eqref{SU4_0+2F+1AS}. Namely the correspondence is given by
\begin{align}
\MQ_1^{2,0,1}\cup\MQ_2^{2,0,1} \cup \MQ_3^{2,0,1} & =
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=right:{1}][u](sp2){};
\node[label=above:{1}][u](u1)[above left of=sp2]{};
\node[label=below:{1}][u](u1')[below left of=sp2]{};
\node[label=left:{1}][uf](f)[left of=u1']{};
\draw(sp2)--(u1);
\draw(sp2)--(u1');
\draw(u1)--(u1');
\draw[dashed](u1)to[out=-135, in=135]node[left]{2}(u1');
\path[draw, snake it](u1')--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u1){};
\node[label=above:{1}][u](sp2)[right of=u1]{};
\node[label=left:{2}][uf](f)[below of=u1]{};
\draw(u1)--node[above]{2}(sp2);
\path[draw, snake it](u1)--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=below:{1}][u](2)[below of=1]{};
\draw[dashed](1)to[out=-135, in=135]node[left]{2}(2);
\draw(1)to[out=-45, in=45]node[right]{3}(2);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\label{SO6:MQ201} \\
\MQ_1^{1,1,1}\cup\MQ_2^{1,1,1} &=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=right:{1}][u](sp2){};
\node[label=above:{1}][u](u1)[above left of=sp2]{};
\node[label=below:{1}][u](u1')[below left of=sp2]{};
\node[label=left:{1}][uf](f)[left of=u1']{};
\draw(sp2)--(u1);
\draw[dashed](sp2)--(u1');
\draw(u1)--(u1');
\draw[dashed](u1)to[out=-135, in=135]node[left]{2}(u1');
\path[draw, snake it](u1')--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u1){};
\node[label=above:{1}][u](sp2)[right of=u1]{};
\node[label=left:{2}][uf](f)[below of=u1]{};
\draw(u1)--(sp2);
\draw[dashed](u1)to[out=45, in=135](sp2);
\path[draw, snake it](u1)--(f);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;. \label{SO6:MQ111}
\end{align}
We have explicitly computed the unrefined Coulomb branch Hilbert series of the orthosymplectic quivers and they agree with their unitary counterparts. The results are given in table \ref{CoulombHSOSp}. Moreover, since we know the HWG of the Coulomb branch of some of the unitary quivers \cite{Bourget:2019rtl,Cremonesi:2013lqa}, we can conjecture the same for the orthosymplectic quivers:
\begin{align}
&\HWG_{\mathcal{C}}(\MQ_2^{2,0,1})=\HWG_{\mathcal{C}}^2(\ref{SU4_1+2F+1AS})=\PE\left[\left(\mu^2+\nu^2\right)t^2\right] \nonumber\\
&\HWG_{\mathcal{C}}(\MQ_3^{2,0,1})=\HWG_{\mathcal{C}}^3(\ref{SU4_1+2F+1AS})=\PE\left[\left(\mu^2+1\right)t^2+\left(q+q^{-1}\right)t^3-t^6\right] \nonumber\\
&\HWG_{\mathcal{C}}(\MQ_1^{1,1,1})=\HWG_{\mathcal{C}}^1(\ref{SU4_0+2F+1AS})=\PE\left[(1+\mu^2+\nu^2)t^2+\mu\nu(q+q^{-1})t^4-\mu^2\nu^2t^8\right] \nonumber \\
&\HWG_{\mathcal{C}}(\MQ_2^{1,1,1})=\HWG_{\mathcal{C}}^2(\ref{SU4_0+2F+1AS})=\PE\left[(1+\mu^2)t^2+\mu(q+q^{-1})t^3-\mu^2t^6\right]~,
\end{align}
where $\mu$ and $\nu$ are highest weight fugacities of SU$(2)$, and $q$ is the U(1) charge.
In fact, we found that the same prescription also works for the Higgs branch. The Higgs branch Hilbert series of the magnetic quivers in \eqref{SO6:MQ201} and \eqref{SO6:MQ111} agree with the Higgs branch Hilbert series of the magnetic quivers in \eqref{SU4_1+2F+1AS} and \eqref{SU4_0+2F+1AS} respectively. The result is summarized in table \ref{HiggsHSOSp}.
\subsubsection{SO(6)+2\textbf{S}+1\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_\frac{1}{2}$}{TEXT}+3\textbf{F}+1\textbf{AS}}
The brane web for SO(6)+2\textbf{S}+1\textbf{C}+1\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(5,0);
\draw[thick,dashed](5,0)--(6,0);
\draw[thick](3,0)--(4,1);
\draw[thick](3,0)--(1,2);
\node[7brane] at (1,0){};
\node[7brane] at (2,0){};
\node[7brane] at (4,0){};
\node[7brane] at (5,0){};
\node[label=above:{$(1,-1)$}][7brane] at (1,2){};
\node[label=above:{$(1,1)$}][7brane] at (4,1){};
\node[7brane] at (2,1){};
\node[label=below:{O5$^-$}]at(0.5,0){};
\node[label=below:{O5$^-$}]at(5.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$2$}]at(3.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(4.5,0){};
\node[label=above:{2}]at(2.5,.5){};
\node[label=above:{1}]at(1.5,1.5){};
\node[label=above:{1}]at(3.5,.5){};
\draw[thick](3,0)--(3,2);
\node[7brane]at(3,2){};
\node[label=right:{1}]at(3,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver extracted from this brane web is
\begin{equation}
\MQ_1^{2,1,1}\cup\MQ_2^{2,1,1}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u11){};
\node[label=left:{1}][u](u1)[below left of=u11]{};
\node[label=right:{1}][u](u1')[below right of=u11]{};
\node[label=below:{2}][so](so2)[below of=u1]{};
\node[label=below:{2}][sp](sp2)[below of=u1']{};
\draw(u11)--(u1);
\draw(u11)--(u1');
\draw(u1)--(so2);
\draw(u1')--(sp2);
\draw(so2)--(sp2);
\draw[dashed](u1)--node[below]{2}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad\cup \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=below:{2}][sp](sp2)[right of=u1]{};
\node[label=below:{1}][u](u1')[right of=sp2]{};
\draw(u1)--(sp2);
\draw(sp2)--node[above]{3}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ211}
\end{equation}
The magnetic quiver $\MQ_1^{2,1,1}$ contains a bad USp(2) gauge node and
we
apply the prescription proposed in section \ref{sec:SO6+1S1C1V} to extract magnetic quivers which capture the moduli spaces of the unitary counterpart. The magnetic quiver after applying the prescription is given by
\begin{equation}
\MQ^{2,1,1}_1 \mbox{}\quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u11){};
\node[label=left:{1}][u](u1)[below left of=u11]{};
\node[label=right:{1}][u](u1')[below right of=u11]{};
\node[label=below:{2}][so](so2)[below of=u1]{};
\node[label=below:{1}][u](sp2)[below of=u1']{};
\draw(u11)--(u1);
\draw(u11)--(u1');
\draw(u1)--(so2);
\draw(u1')--(sp2);
\draw(so2)--node[above]{$\frac{1}{2}$}++(sp2);
\draw[dashed](u1)--node[below]{2}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u11){};
\node[label=left:{1}][u](u1)[below left of=u11]{};
\node[label=right:{1}][u](u1')[below right of=u11]{};
\node[label=below:{1}][u](so2)[below of=u1]{};
\node[label=below:{1}][u](sp2)[below of=u1']{};
\draw(u11)--(u1);
\draw(u11)--(u1');
\draw(u1)--(so2);
\draw[dashed](u1)to[out=-45,in=45](so2);
\draw(u1')--(sp2);
\draw(so2)--(sp2);
\draw[dashed](u1)--node[below]{2}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQ211.modified}
\end{equation}
Next, consider the brane web for SU$(4)_\frac{1}{2}$+3\textbf{F}+1\textbf{AS}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-3,0);
\draw[thick](0,0)--(2,-2);
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(-1,-1);
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(0,1){};
\node[label=below:{$(1,1)$}][7brane]at(-1,-1){};
\node[label=below:{$(1,-1)$}][7brane]at(2,-2){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(1,-1){};
\node[label=above:{1}]at(-2.5,0){};
\node[label=above:{2}]at(-1.5,0){};
\node[label=above:{3}]at(-.5,0){};
\node[label=above:{1}]at(1.75,-1.75){};
\node[label=above:{2}]at(.75,-.75){};
\node[label=below:{2}]at(.75,.75){};
\node[label=above:{1}]at(-.75,-.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The corresponding unitary magnetic quiver is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=below:{1}][u](22)[below of=1]{};
\node[label=left:{1}][u](2)[left of=22]{};
\node[label=right:{1}][u](2')[right of=22]{};
\node[label=below:{1}][u](3)[below of=2]{};
\node[label=below:{1}][u](3')[below of=2']{};
\draw(1)--node[above left]{2}++(2);
\draw(1)--(2');
\draw(2)--(22);
\draw(2')--(22);
\draw(2)--(3);
\draw(2')--(3');
\draw(3)--(3');
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad\cup\quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=below:{2}][u](sp2)[right of=u1]{};
\node[label=below:{1}][u](u1')[right of=sp2]{};
\draw(u1)--(sp2);
\draw(sp2)--node[above]{3}++(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_1/2+3F+1AS}
\end{equation}
The unrefined Coulomb branch and Higgs branch Hilbert series for the quivers in \eqref{SO6:MQ211.modified} match with the corresponding unitary quivers in \eqref{SU4_1/2+3F+1AS} and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG of the Coulomb branch for the second cone on unitary side is known \cite{Hanany:2011db} and so we can write the same for the OSp quiver:
\begin{equation}
\HWG_{\mathcal{C}}(\MQ^{2,1,1}_2) = \HWG_{\mathcal{C}}^2\eqref{SU4_1/2+3F+1AS} = \PE\left[\nu_1\nu_2(t^2+t^4)+(\nu_1^3+\nu_2^3)t^6-\nu_1^3\nu_2^3t^{12}\right] ~,
\end{equation}
where $\nu_1$ and $\nu_2$ are the highest weight fugacities of SU(3) group.
\subsubsection{SO(6)+2\textbf{S}+2\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_0$}{TEXT}+4\textbf{F}+1\textbf{AS}}
There are two inequivalent web diagrams for SO(6)+2\textbf{S}+2\textbf{C}+1\textbf{V}. The first web diagram, corresponding to the quiver $[1\textbf{V}]-\text{SO}(6)-[2\textbf{S}+2\textbf{C}]$, is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(9,0);
\draw[thick,dashed](9,0)--(10,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\draw[thick](3,0)--(3,2);
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(8,0){};
\node[7brane]at(9,0){};
\node[7brane]at(3,1){};
\node[7brane]at(3,2){};
\node[label=below:{O5$^+$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(9.5,0){};
\node[label=below:{$1$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$3$}]at(3.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(4.5,0){};
\node[label=below:{$2$}]at(5.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(6.5,0){};
\node[label=below:{$1$}]at(7.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(8.5,0){};
\node[label=right:{3}]at(3,.5){};
\node[label=right:{1}]at(3,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The corresponding magnetic quiver for this theory is
\begin{align}
\MQ_1^{2,2,1}&=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=left:{2}][sp](sp2')[above of=so3]{};
\node[label=left:{1}][u](u1)[above of=sp2']{};
\node[label=right:{1}][uf](uf)[right of=u1]{};
\node[label=right:{1}][sof](soof)[right of=sp2']{};
\node[label=below:{2}][sp](sp2'')[right of=so3]{};
\node[label=below:{3}][so](so3')[right of=sp2'']{};
\node[label=below:{2}][sp](sp22)[right of=so3']{};
\node[label=below:{2}][so](so2)[right of=sp22]{};
\node[label=above:{1}][sof](so1)[above of=sp22]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp2');
\draw(sp2')--(u1);
\draw(u1)--(uf);
\draw(so3)--(sp2'');
\draw(sp2'')--(so3');
\draw(so3')--(sp22);
\draw(sp22)--(so2);
\draw(sp22)--(so1);
\draw(soof)--(sp2');
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_2^{2,2,1}&=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][sof](o1){};
\node[label=below:{2}][sp](sp2)[below of=o1]{};
\node[label=below:{2}][so](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp4)[right of=so3]{};
\node[label=below:{4}][so](so4)[right of=sp4]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=above:{2}][spf](spf)[above of=so4]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp4);
\draw(sp4)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\draw(so4)--(spf);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQ221}
\end{align}
Note that the magnetic quivers $\MQ_1^{2,2,1}$ and $\MQ_2^{2,2,1}$ have a bad USp(2) gauge node.
Hence, we use the prescription proposed in section \ref{sec:SO6+1S1C1V} to extract magnetic quivers which capture the moduli spaces of the unitary counterparts:
\begin{align}
\MQ_1^{2,2,1} &\quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{1}][u](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=left:{2}][sp](sp2')[above of=so3]{};
\node[label=left:{1}][u](u1)[above of=sp2']{};
\node[label=right:{1}][uf](uf)[right of=u1]{};
\node[label=right:{1}][sof](soof)[right of=sp2']{};
\node[label=below:{2}][sp](sp2'')[right of=so3]{};
\node[label=below:{3}][so](so3')[right of=sp2'']{};
\node[label=below:{2}][sp](sp22)[right of=so3']{};
\node[label=below:{2}][so](so2)[right of=sp22]{};
\node[label=above:{1}][sof](so1)[above of=sp22]{};
\draw(o1)--node[above]{$\frac{1}{2}$}++(sp2);
\draw(sp2)--node[above]{$\frac{1}{2}$}++(so3);
\draw(so3)--(sp2');
\draw(sp2')--(u1);
\draw(u1)--(uf);
\draw(so3)--(sp2'');
\draw(sp2'')--(so3');
\draw(so3')--(sp22);
\draw(sp22)--(so2);
\draw(sp22)--(so1);
\draw(soof)--(sp2');
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_2^{2,2,1}&\quad\contour{black}{$\xrightarrow{\text{B2G}}$}\quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][uf](o1){};
\node[label=below:{1}][u](sp2)[below of=o1]{};
\node[label=below:{2}][so](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp4)[right of=so3]{};
\node[label=below:{4}][so](so4)[right of=sp4]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=above:{2}][spf](spf)[above of=so4]{};
\draw(o1)--(sp2);
\draw(sp2)--node[above]{$\frac{1}{2}$}++(so3);
\draw(so3)--(sp4);
\draw(sp4)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\draw(so4)--(spf);
\end{tikzpicture}
\end{scriptsize}
\end{array} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][uf](o1){};
\node[label=below:{1}][u](sp2)[below of=o1]{};
\node[label=below:{1}][u](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp4)[right of=so3]{};
\node[label=below:{4}][so](so4)[right of=sp4]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=above:{2}][spf](spf)[above of=so4]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp4);
\draw(sp4)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\draw(so4)--(spf);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQ221modified}
\end{align}
An alternative web diagram, for 5d SO(6)+2\textbf{S}+2\textbf{C}+1\textbf{V} can be obtained by considering ${[1\textbf{S}+1\textbf{C}]-\underset{\underset{\text{\large$\left[1\textbf{V}\right]$}}{\textstyle\vert}}{\text{SO}(6)}}-[1\textbf{S}+1\textbf{C}]$
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick, dashed](0,0)--(1,0);
\draw[thick, dashed](7,0)--(8,0);
\draw[thick](1,0)--(7,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,1){};
\draw[thick](5,0)--(7,2);
\draw[thick](5,0)--(5,2);
\node[7brane]at(5,1){};
\node[7brane]at(5,2){};
\node[label=above:{$(1,1)$}][7brane]at(7,2){};
\node[label=below:{O5$^-$}]at(.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(5.5,0){};
\node[label=below:{$3$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=above:{1}]at(6.5,1.5){};
\node[label=above:{2}]at(5.5,.5){};
\node[label=left:{2}]at(5,.5){};
\node[label=left:{1}]at(5,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
which gives the following magnetic quivers:
\begin{equation}
\widehat{\MQ}_1^{2,2,1}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{4}][sp](sp4)[right of=so2]{};
\node[label=below:{2}][so](so2')[right of=sp4]{};
\node[label=below left:{2}][u](u2)[above of=sp4]{};
\node[label=above:{1}][u](u1)[left of=u2]{};
\node[label=above:{1}][u](u1')[right of=u2]{};
\draw(so2)--(sp4);
\draw(so2')--(sp4);
\draw(sp4)--(u2);
\draw(u1)--(u2);
\draw(u1')--(u2);
\draw (u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{$1\;\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad;\quad \widehat{\MQ}_2^{2,2,1} =
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=above:{1}][u](u1)[above of=sp2]{};
\node[label=above:{1}][u](u1')[above of=sp2']{};
\node[label=below:{2}][sp](sp2'')[below of=so2]{};
\node[label=below:{2}][so](so2')[below of=sp2]{};
\draw(so2)--(sp2);
\draw(sp2)--(sp2');
\draw(u1)--(sp2);
\draw(u1')--(sp2');
\draw(so2')--(sp2');
\draw(so2)--(sp2'');
\draw(sp2'')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQhat221}
\end{equation}
The magnetic quiver $\widehat{\MQ}_2^{2,2,1}$ has a bad USp(2) gauge node and we
again replace it with a U(1) node by using the prescription in section \ref{sec:SO6+1S1C1V}. The magnetic quiver after applying the prescription is given by
\begin{equation}
\widehat{\MQ}^{2,2,1}_2 \quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=above:{1}][u](u1)[above of=sp2]{};
\node[label=above:{1}][u](u1')[above of=sp2']{};
\node[label=below:{1}][u](sp2'')[below of=so2]{};
\node[label=below:{2}][so](so2')[below of=sp2]{};
\draw(so2)--(sp2);
\draw(sp2)--(sp2');
\draw(u1)--(sp2);
\draw(u1')--(sp2');
\draw(so2')--(sp2');
\draw(so2)--node[left]{$\frac{1}{2}$}++(sp2'');
\draw(sp2'')--node[above]{$\frac{1}{2}$}++(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=above:{1}][u](u1)[above of=sp2]{};
\node[label=above:{1}][u](u1')[above of=sp2']{};
\node[label=below:{1}][u](sp2'')[below of=so2]{};
\node[label=below:{1}][u](so2')[below of=sp2]{};
\draw(so2)--(sp2);
\draw(sp2)--(sp2');
\draw(u1)--(sp2);
\draw(u1')--(sp2');
\draw(so2')--(sp2');
\draw(so2)--(sp2'');
\draw(sp2'')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQhat221.modified}
\end{equation}
On the other hand the brane web for SU$(4)_0$+4\textbf{F}+1\textbf{AS} is given as
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(4,0);
\draw[thick](4,0)--(6,-2);
\draw[thick](4,0)--(6,2);
\draw[thick](4,0)--(3,-1);
\node[7brane]at(0,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(3,-1){};
\node[7brane]at(5,-1){};
\node[7brane]at(6,-2){};
\node[7brane]at(5,1){};
\node[7brane]at(6,2){};
\node[label=above:{1}]at(0.5,0){};
\node[label=above:{2}]at(1.5,0){};
\node[label=above:{3}]at(2.5,0){};
\node[label=above:{4}]at(3.5,0){};
\node[label=above:{2}]at(4.5,-.5){};
\node[label=above:{1}]at(5.5,-1.5){};
\node[label=below:{1}]at(3.5,-.5){};
\node[label=above:{3}]at(4.5,.5){};
\node[label=above:{1}]at(5.5,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
which gives the following two unitary quivers:
\begin{equation}\label{eq:MQ_su4_0_1as_4f}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=above:{1}][u](11)[above of=2]{};
\node[label=below:{2}][u](22)[right of=2]{};
\node[label=below:{1}][u](111)[right of=22]{};
\node[label=above:{1}][u](1111)[above of=22]{};
\node[label=above:{1}][u](11111)[above of=111]{};
\draw(1)--(2);
\draw(2)--(11);
\draw(2)--(22);
\draw(22)--(111);
\draw(22)--(1111);
\draw(22)--(11111);
\draw(111)--(11111);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\quad\cup\quad
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{2}][u](22)[right of=2]{};
\node[label=below:{1}][u](11)[right of=22]{};
\node[label=above:{1}][u](111)[above of=2]{};
\node[label=above:{1}][u](1111)[above of=11]{};
\draw(1)--(2);
\draw(2)--(22);
\draw[double distance=2pt](22)--(11);
\draw(11)--(1111);
\draw(2)--(111);
\draw(111)--(1111);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\end{equation}
The unrefined Coulomb branch and Higgs branch Hilbert series of the orthosymplectic quivers match with their unitary counterparts and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively.
\subsubsection{SO(6)+3\textbf{S}+2\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_{\frac{1}{2}}$}{TEXT}+5\textbf{F}+1\textbf{AS}}
The brane web for the SCFT limit of SO$(6)$+3\textbf{S}+2\textbf{C}+1\textbf{V} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-6,0);
\draw[thick](0,0)--(2,0);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(0,2);
\draw[thick,dashed](-7,0)--(-6,0);
\draw[thick,dashed](3,0)--(2,0);
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-5,0){};
\node[7brane]at(-6,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(0,1){};
\node[7brane]at(0,2){};
\node[label=below:{$\frac{1}{2}$}]at(-5.5,0){};
\node[label=below:{$1$}]at(-4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-3.5,0){};
\node[label=below:{$2$}]at(-2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(-1.5,0){};
\node[label=below:{$3$}]at(-.5,0){};
\node[label=below:{$2$}]at(.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{O5$^-$}]at(2.5,0){};
\node[label=below:{O5$^-$}]at(-6.5,0){};
\node[label=left:{$1$}]at(0,1.5){};
\node[label=left:{$3$}]at(0,.5){};
\node[label=right:{$1$}]at(0.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\end{equation}
The magnetic quiver associated to this brane system is
\begin{equation}
\MQ_1^{3,2,1} =\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u1){};
\node[label=above:{1}][u](u11)[left of=u1]{};
\node[label=above:{2}][sp](sp22)[right of=u1]{};
\node[label=below:{4}][so](so4')[below of=sp22]{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{4}][sp](sp4)[below of=u11]{};
\node[label=below:{4}][so](so4)[left of=sp4]{};
\node[label=below:{2}][sp](sp2)[left of=so4]{};
\node[label=below:{2}][so](so2)[left of=sp2]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so4');
\draw(so4')--(sp2');
\draw(so4')--(sp22);
\draw(sp4)--(u11);
\draw(u1)--(u11);
\draw(u1)--(sp22);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQ321}
\end{equation}
On the other hand, the brane web for the SCFT parent of SU$(4)_\frac{1}{2}$+5\textbf{F}+1\textbf{AS} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-2,2);
\draw[thick](0,0)--(-2,-2);
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(5,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(-2,2){};
\node[7brane]at(-1,1){};
\node[7brane]at(-2,-2){};
\node[7brane]at(-1,-1){};
\node[7brane]at(0,1){};
\node[label=below:{1}]at(4.5,0){};
\node[label=below:{2}]at(3.5,0){};
\node[label=below:{3}]at(2.5,0){};
\node[label=below:{4}]at(1.5,0){};
\node[label=below:{5}]at(.5,0){};
\node[label=right:{1}]at(0,.5){};
\node[label=right:{3}]at(-.5,-.5){};
\node[label=right:{1}]at(-1.5,-1.5){};
\node[label=left:{2}]at(-.5,.5){};
\node[label=left:{1}]at(-1.5,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~,
\end{equation}
which gives the following unitay quiver:
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](u1){};
\node[label=above:{2}][u](u2)[right of=u1]{};
\node[label=above:{1}][u](u11)[left of=u1]{};
\node[label=below:{2}][u](2)[below of=u11]{};
\node[label=below:{1}][u](1)[left of=2]{};
\node[label=below:{3}][u](3)[below of=u2]{};
\node[label=below:{2}][u](2')[right of=3]{};
\node[label=below:{1}][u](1')[right of=2']{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(2');
\draw(2')--(1');
\draw(2)--(u11);
\draw(3)--(u2);
\draw(u11)--(u1);
\draw(u1)--(u2);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SU4_0+5F+1AS}
\end{equation}
Note that the quiver \eqref{SO6:MQ321} has a bad USp(2) gauge node and we can use our B2G prescription to extract the good quiver. The Coulomb and Higgs branch Hilbert series after doing the B2G replacement agrees with that of the unitary quiver \eqref{SU4_0+5F+1AS} and the results are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively.
\subsubsection{SO(6)+3\textbf{S}+3\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_0$}{TEXT}+6\textbf{F}+1\textbf{AS}}
The brane web for SO(6)+3\textbf{S}+3\textbf{C}+1\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-6,0);
\draw[thick](0,0)--(4,0);
\draw[thick,dashed](-7,0)--(-6,0);
\draw[thick,dashed](5,0)--(4,0);
\draw[thick](0,0)--(0,3);
\node[7brane]at(-6,0){};
\node[7brane]at(-5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(4,0){};
\node[7brane]at(3,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[7brane]at(0,3){};
\node[7brane]at(0,2){};
\node[7brane]at(0,1){};
\node[label=below:{O5$^-$}]at(-6.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(-5.5,0){};
\node[label=below:{1}]at(-4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-3.5,0){};
\node[label=below:{2}]at(-2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(-1.5,0){};
\node[label=below:{3}]at(-.5,0){};
\node[label=below:{3}]at(.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(1.5,0){};
\node[label=below:{1}]at(2.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=right:{4}]at(0,0.5){};
\node[label=right:{2}]at(0,1.5){};
\node[label=right:{1}]at(0,2.5){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From here we read off the following magnetic quiver
\begin{equation}
\MQ_1^{3,3,1} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{6}][so](so6)[right of=sp4]{};
\node[label=below:{4}][sp](sp4')[right of=so6]{};
\node[label=below:{2}][so](so2')[right of=sp4']{};
\node[label=right:{4}][sp](sp4'')[above of=so6]{};
\node[label=right:{2}][u](u2)[above of=sp4'']{};
\node[label=right:{1}][u](u1)[above of=u2]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so6);
\draw(so6)--(sp4');
\draw(sp4')--(so2');
\draw(so6)--(sp4'');
\draw(sp4'')--(u2);
\draw(u2)--(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ331}
\end{equation}
Note that this quiver has a bad USp(4) gauge node. We use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series which are presented in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. These Hilbert series are in agreement with the Hilbert series of the following unitary quiver:
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{3}][u](3)[right of=2]{};
\node[label=below:{4}][u](4)[right of=3]{};
\node[label=below:{2}][u](2')[right of=4]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{3}][u](33)[above of=4]{};
\node[label=above:{1}][u](11)[left of=33]{};
\node[label=above:{1}][u](11')[right of=33]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(4)--(2');
\draw(2')--(1');
\draw(4)--(33);
\draw(33)--(11);
\draw(33)--(11');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~,
\end{equation}
which is associated with the following brane web for SU$(4)_0$+6\textbf{F}+1\textbf{AS} at infinite coupling:
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-5,0);
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(0,-1);
\draw[thick](0,0)--(2,0);
\draw[thick](0,0)--(2,-2);
\node[7brane]at(-5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,-1){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,-1){};
\node[label=below:{$(1,-1)$}][7brane]at(2,-2){};
\node[label=above:{1}]at(-4.5,0){};
\node[label=above:{2}]at(-3.5,0){};
\node[label=above:{3}]at(-2.5,0){};
\node[label=above:{4}]at(-1.5,0){};
\node[label=above:{5}]at(-.5,0){};
\node[label=above:{3}]at(.5,0){};
\node[label=above:{1}]at(1.5,0){};
\node[label=left:{1}]at(0,-.5){};
\node[label=below:{1}]at(1.5,-1.5){};
\node[label=below:{2}]at(.5,-.5){};
\node[label=right:{3}]at(0,.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\end{equation}
\subsubsection{SO(6)+4\textbf{S}+4\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_0$}{TEXT}+8\textbf{F}+1\textbf{AS}}
The brane web for SO(6)+4\textbf{S}+4\textbf{C}+1\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}[scale=.75]
\draw[thick](-8,0)--(8,0);
\draw[thick,dashed](8,0)--(9,0);
\draw[thick,dashed](-8,0)--(-9,0);
\draw[thick](0,0)--(0,1.5);
\node[7brane]at(-8,0){};
\node[7brane]at(-7,0){};
\node[7brane]at(-6,0){};
\node[7brane]at(-5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(8,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,0){};
\node[7brane]at(5,0){};
\node[7brane]at(4,0){};
\node[7brane]at(3,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[label=below:{$\frac{1}{2}$}]at(-7.5,0){};
\node[label=below:{$1$}]at(-6.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-5.5,0){};
\node[label=below:{$2$}]at(-4.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(-3.5,0){};
\node[label=below:{$3$}]at(-2.5,0){};
\node[label=below:{$\frac{9}{2}$}]at(-1.5,0){};
\node[label=below:{$10$}]at(-.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(7.5,0){};
\node[label=below:{$1$}]at(6.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(5.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(3.5,0){};
\node[label=below:{$3$}]at(2.5,0){};
\node[label=below:{$\frac{9}{2}$}]at(1.5,0){};
\node[label=below:{$10$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(8.5,0){};
\node[label=below:{O5$^-$}]at(-8.5,0){};
\node[7brane]at(0,1.5){};
\node[label=right:{4}]at(0,1){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver associated with this brane web is
\begin{equation}
\MQ_1^{4,4,1} =\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{6}][so](so6)[right of=sp4]{};
\node[label=below:{8}][sp](sp8)[right of=so6]{};
\node[label=below:{10}][so](so10)[right of=sp8]{};
\node[label=below:{8}][sp](sp8')[right of=so10]{};
\node[label=below:{6}][so](so6')[right of=sp8']{};
\node[label=below:{4}][sp](sp4')[right of=so6']{};
\node[label=below:{4}][so](so4')[right of=sp4']{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=above:{4}][sp](sp44)[above of=so10]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so6);
\draw(so6)--(sp8);
\draw(sp8)--(so10);
\draw(so2')--(sp2');
\draw(sp2')--(so4');
\draw(so4')--(sp4');
\draw(sp4')--(so6');
\draw(so6')--(sp8');
\draw(sp8')--(so10);
\draw(so10)--(sp44);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ441}
\end{equation}
Now let us consider the brane web for SU$(4)_0$+8\textbf{F}+1\textbf{AS} at infinite coupling
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-6,0);
\draw[thick](0,0)--(5,0);
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(0,-1);
\node[7brane]at(-6,0){};
\node[7brane]at(-5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,-1){};
\node[label=below:{1}]at(-5.5,0){};
\node[label=below:{2}]at(-4.5,0){};
\node[label=below:{3}]at(-3.5,0){};
\node[label=below:{4}]at(-2.5,0){};
\node[label=below:{5}]at(-1.5,0){};
\node[label=below:{6}]at(-.5,0){};
\node[label=above:{6}]at(.5,0){};
\node[label=above:{4}]at(1.5,0){};
\node[label=above:{3}]at(2.5,0){};
\node[label=above:{2}]at(3.5,0){};
\node[label=above:{1}]at(4.5,0){};
\node[label=left:{3}]at(0,.75){};
\node[label=right:{3}]at(0,-.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver for this brane system is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{3}][u](3)[right of=2]{};
\node[label=below:{4}][u](4)[right of=3]{};
\node[label=below:{5}][u](5)[right of=4]{};
\node[label=below:{6}][u](6)[right of=5]{};
\node[label=below:{4}][u](4')[right of=6]{};
\node[label=below:{3}][u](3')[right of=4']{};
\node[label=below:{2}][u](2')[right of=3']{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{3}][u](33)[above of=6]{};
\draw(1)--(2);
\draw(2)--(3);
\draw(3)--(4);
\draw(4)--(5);
\draw(5)--(6);
\draw(6)--(4');
\draw(4')--(3');
\draw(3')--(2');
\draw(2')--(1');
\draw(6)--(33);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_0+1AS+8F} ~.
\end{equation}
The magnetic quiver in equation \eqref{SU4_0+1AS+8F} (resp. \eqref{SO6:MQ441}) can also be seen as the 3d mirror of a class-S theory of $A_4$ (resp. $D_5$) type on the sphere, with three regular untwisted punctures. The three (Nahm) punctures defining the $A$-type theory are $[1^6], \ [3,3],\ [2,1^4]$, while the three (Nahm) punctures defining the $D$-type theory are $[2^2,1^6], \ [2^2,1^6], \ [5,5]$. Following the techniques explained in \cite{Chacaltana:2010ks} and \cite{Chacaltana:2011ze}, we computed the Coulomb branch spectrum of the two four-dimensional theories, as well as their superconformal central charges. Let us denote by $d_i$ the number of Coulomb branch operators of conformal dimension $i$. We find for both theories an identical spectrum given by $\{d_4=1,\, d_5=1,\, d_6=1\}$, and identical central charges given by $a=65/8$ and $c=19/2$. We take this match as a further evidence for the agreement of the magnetic quivers in \eqref{SU4_0+1AS+8F} and \eqref{SO6:MQ441}.
Note that the quiver in \eqref{SO6:MQ441} has two bad USp(8) gauge nodes. We use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series. The perturbative results up to first few orders are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. These results agree with the Hilbert series of the unitary quiver \eqref{SU4_0+1AS+8F}.
\subsection{SO(6) theories with two vector hypermultiplets}
We now move on to study the Higgs branch of SO(6) theories with two hypermultiplets in the vector representation, as well as the hypermultiplets in (conjugate) spinor representations.
\subsubsection{SO(6)+2\textbf{S}/(1\textbf{S}+1\textbf{C})+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_{1/0}$}{TEXT}+2\textbf{F}+2\textbf{AS}}
The brane web for SO(6)+1\textbf{S}+1\textbf{C}+2\textbf{V} at infinite coupling is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(7,0);
\draw[thick,dashed](7,0)--(8,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\draw[thick](3,0)--(3,3);
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(3,1){};
\node[7brane]at(3,2){};
\node[7brane]at(3,3){};
\node[label=below:{O5$^+$}]at(.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=below:{$1$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$3$}]at(3.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(4.5,0){};
\node[label=below:{$1$}]at(5.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(6.5,0){};
\node[label=right:{3}]at(3,.5){};
\node[label=right:{2}]at(3,1.5){};
\node[label=right:{1}]at(3,2.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver for this brane system is given by the union of three cones
\begin{align}
\MQ_1^{1,1,2}&= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so3]{};
\node[label=left:{2}][sp](sp2'')[above of=so3]{};
\node[label=left:{2}][u](u2)[above of=sp2'']{};
\node[label=left:{1}][u](u1)[above of=u2]{};
\node[label=right:{1}][uf](uf)[right of=u2]{};
\node[label=below:{1}][so](sof)[right of=sp2']{};
\node[label=right:{1}][sof](sof')[right of=sp2'']{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp2');
\draw(so3)--(sp2'');
\draw(sp2'')--(u2);
\draw(u1)--(u2);
\draw(uf)--(u2);
\draw(sp2')--(sof);
\draw(sp2'')--(sof');
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_2^{1,1,2}&=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][so](so2')[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so2']{};
\node[label=above:{2}][spf](spf)[above of=so2]{};
\node[label=above:{2}][sof](sof)[above of=sp2']{};
\draw(so2)--(sp2);
\draw(sp2)--(so2');
\draw(sp2')--(so2');
\draw(so2)--(spf);
\draw(sp2')--(sof);
\end{tikzpicture}
\end{scriptsize}
\end{array} \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\node[label=above:{1}][uf](f)[above of=1]{};
\node[label=above:{1}][uf](f')[above of=1']{};
\draw(1)--(1');
\draw(1)--(f);
\draw(1')--(f');
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{1}][u](sp2)[right of=so2]{};
\node[label=below:{2}][so](so2')[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so2']{};
\node[label=above:{2}][spf](spf)[above of=so2]{};
\node[label=above:{2}][sof](sof)[above of=sp2']{};
\draw(so2)--node[above]{$\frac{1}{2}$}++(sp2);
\draw(sp2)--node[above]{$\frac{1}{2}$}++(so2');
\draw(sp2')--node[above]{$\frac{1}{2}$}++(so2');
\draw(so2)--(spf);
\draw(sp2')--node[right]{$\frac{1}{2}$}++(sof);
\end{tikzpicture}
\end{scriptsize}
\end{array} \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{1}][u](1')[right of=1]{};
\node[label=above:{1}][uf](f)[above of=1]{};
\node[label=above:{1}][uf](f')[above of=1']{};
\draw(1)--(1');
\draw(1)--(f);
\draw(1')--(f');
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\\MQ_3^{1,1,2}&=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so3]{};
\node[label=below:{2}][so](so2)[right of=sp4]{};
\node[label=above:{3}][sof](sof)[above of=sp4]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp4);
\draw(sp4)--(so2);
\draw(sp4)--(sof);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ112}
\end{align}
We can alternatively consider the brane web for SO(6)+2\textbf{S}/(1\textbf{S}+1\textbf{C})+2\textbf{V} at infinite coupling
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(5,0);
\draw[thick,dashed](5,0)--(6,0);
\draw[thick](3,0)--(4,1);
\draw[thick](3,0)--(3,2);
\node[7brane]at(3,1){};
\node[7brane]at(3,2){};
\draw[thick](3,0)--(2,1);
\node[7brane] at (1,0){};
\node[7brane] at (2,0){};
\node[7brane] at (4,0){};
\node[7brane] at (5,0){};
\node[label=above:{$(1,-1)$}][7brane] at (2,1){};
\node[label=above:{$(1,1)$}][7brane] at (4,1){};
\node[label=below:{O5$^-$}]at(0.5,0){};
\node[label=below:{O5$^-$}]at(5.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(2.5,0){};
\node[label=below:{$2$}]at(3.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(4.5,0){};
\node[label=right:{1}]at(2.85,1.75){};
\node[label=right:{2}]at(2.85,.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
which gives the following magnetic quivers
\begin{equation}
\MQ_1^{*,*,2}\cup\MQ_2^{*,*,2} \cup \MQ_3^{*,*,2}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=left:{1}][u](u1)[above left of=so4]{};
\node[label=right:{2}][sp](sp22)[above right of=so4]{};
\node[label=above:{1}][u](u11)[above of=sp22]{};
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(so4)--(sp22);
\draw(so4)--(u1);
\draw(u1)--(sp22);
\draw(sp22)--(u11);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=left:{1}][u](2)[below left of=1]{};
\node[label=right:{1}][u](2')[below right of=1]{};
\node[label=below:{2}][sp](sp2)[below left of=2]{};
\node[label=below:{2}][sp](sp2')[below right of=2']{};
\node[label=below:{2}][so](so2)[below left of=2']{};
\draw(1)--(2);
\draw(1)--(2');
\draw(2)--(sp2);
\draw(2')--(sp2');
\draw(sp2)--(so2);
\draw(sp2')--(so2);
\draw[dashed](2)--node[above]{2}++(2');
\draw(2)to[out=-45,in=-135](2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=above:{1}][u](u1)[above of=so4]{};
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(so4)--node[right]{2}++(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\;.
\label{SO6:MQ202starred}
\end{equation}
To distinguish, the magnetic quivers for SO(6)+1\textbf{S}+1\textbf{C}+2\textbf{V}, from those of SO(6)+2\textbf{S}+2\textbf{V}, we must apply the B2G procedure to the above magnetic quivers, and compare the Hilbert series results with the unitary magnetic quivers for SU$(4)_{0/1}$+2\textbf{F}+2\textbf{AS} (similar to what we did in section \ref{sec:SO6+1S1C1V}). The end result can be summarised as
\begin{equation}
\widehat{\MQ}_1^{1,1,2}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so4]{};
\node[label=left:{1}][u](u1)[above left of=so4]{};
\node[label=right:{2}][sp](sp22)[above right of=so4]{};
\node[label=above:{1}][u](u11)[above of=sp22]{};
\draw(sp2)--node[below]{$\frac{1}{2}$}++(so4);
\draw(so4)--node[below]{$\frac{1}{2}$}++(sp2');
\draw(so4)--(sp22);
\draw(so4)--(u1);
\draw(u1)--(sp22);
\draw(sp22)--(u11);
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad \widehat{\MQ}_2^{1,1,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=left:{1}][u](2)[below left of=1]{};
\node[label=right:{1}][u](2')[below right of=1]{};
\node[label=below:{1}][u](sp2)[below left of=2]{};
\node[label=below:{1}][u](sp2')[below right of=2']{};
\node[label=below:{1}][u](so2)[below left of=2']{};
\draw(1)--(2);
\draw(1)--(2');
\draw(2)--(sp2);
\draw(2')--(sp2');
\draw[dashed](sp2)--(so2);
\draw(sp2')--(so2);
\draw[dashed](2)--node[above]{2}++(2');
\draw(2)to[out=-45,in=-135](2');
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad \widehat{\MQ}_3^{1,1,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so4]{};
\node[label=above:{1}][u](u1)[above of=so4]{};
\draw[dashed](sp2)--node[above]{$\frac{1}{2}$}++(so4);
\draw(so4)--node[above]{$\frac{1}{2}$}++(sp2');
\draw(so4)--node[right]{2}++(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO6:MQtilde112}
\end{equation}
\begin{equation}
\MQ_1^{2,0,2}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so4]{};
\node[label=left:{1}][u](u1)[above left of=so4]{};
\node[label=right:{2}][sp](sp22)[above right of=so4]{};
\node[label=above:{1}][u](u11)[above of=sp22]{};
\draw(sp2)--node[below]{$\frac{1}{2}$}++(so4);
\draw[dashed](so4)--node[below]{$\frac{1}{2}$}++(sp2');
\draw(so4)--(sp22);
\draw(so4)--(u1);
\draw(u1)--(sp22);
\draw(sp22)--(u11);
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad \MQ_2^{2,0,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=left:{1}][u](2)[below left of=1]{};
\node[label=right:{1}][u](2')[below right of=1]{};
\node[label=below:{1}][u](sp2)[below left of=2]{};
\node[label=below:{1}][u](sp2')[below right of=2']{};
\node[label=below:{1}][u](so2)[below left of=2']{};
\draw(1)--(2);
\draw(1)--(2');
\draw(2)--(sp2);
\draw(2')--(sp2');
\draw(sp2)--(so2);
\draw(sp2')--(so2);
\draw[dashed](2)--node[above]{2}++(2');
\draw(2)to[out=-45,in=-135](2');
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad \MQ_3^{2,0,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so4]{};
\node[label=above:{1}][u](u1)[above of=so4]{};
\draw(sp2)--node[above]{$\frac{1}{2}$}++(so4);
\draw(so4)--node[above]{$\frac{1}{2}$}++(sp2');
\draw(so4)--node[right]{2}++(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array} \;.
\label{SO6:MQ202}
\end{equation}
To confirm these results, we can consider the brane web at infinite gauge coupling for SU$(4)_0$+2\textbf{F}+2\textbf{AS}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-.5,0)--(2.5,0);
\draw[thick](0,-1)--(2,1);
\draw[thick](1,-2)--(1,2);
\node[7brane]at(1,1){};
\node[7brane]at(1,2){};
\node[7brane]at(1,-1){};
\node[7brane]at(1,-2){};
\node[7brane]at(-.5,0){};
\node[7brane]at(2.5,0){};
\node[label=above:{(1,1)}][7brane]at(2,1){};
\node[label=below:{(1,1)}][7brane]at(0,-1){};
\node[label=below:{2}]at(2,0){};
\node[label=above:{2}]at(0,0){};
\node[label=left:{1}]at(1,1.5){};
\node[label=left:{2}]at(1,.75){};
\node[label=right:{1}]at(1,-1.5){};
\node[label=right:{2}]at(1,-.75){};
\node[label=above:{2}]at(0.1,-1){};
\node[label=below:{2}]at(1.9,1){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The corresponding magnetic quiver is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{1}][u](1')[right of=2]{};
\node[label=above:{2}][u](u1)[above left of=2]{};
\node[label=above:{2}][u](u1')[above right of=2]{};
\draw(1)--(2);
\draw(2)--(1');
\draw(2)--(u1);
\draw(u1')--(2);
\draw(u1)--(u1');
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad\cup\quad
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=right:{1}][u](1r)[right of=1]{};
\node[label=left:{1}][u](1l)[left of=1]{};
\node[label=below:{1}][u](1br)[below right of=1]{};
\node[label=below:{1}][u](1bl)[below left of=1]{};
\node[label=above:{1}][u](1ar)[above right of=1]{};
\node[label=above:{1}][u](1al)[above left of=1]{};
\draw(1)--(1l);
\draw(1)--(1r);
\draw(1)--(1al);
\draw(1)--(1ar);
\draw(1al)--(1ar);
\draw(1r)--(1br);
\draw(1l)--(1bl);
\draw[double distance=2pt](1bl)--(1br);
\end{tikzpicture}
\end{scriptsize}
\end{array}\quad\cup\quad
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{2}][u](2')[right of=2]{};
\node[label=below:{1}][u](1')[right of=2']{};
\draw(1)--(2);
\draw(1')--(2');
\draw[double distance=2pt](2)--(2');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SU4_0+2F+2AS}
\end{equation}
On the other hand the brane web for SU$(4)_1$+2\textbf{F}+2\textbf{AS} at infinite coupling takes the form
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(0,-3);
\draw[thick](-2,0)--(2,0);
\draw[thick](0,0)--(-2,2);
\draw[thick](0,0)--(1,1);
\node[7brane]at(0,-1){};
\node[7brane]at(0,-2){};
\node[7brane]at(0,-3){};
\node[7brane]at(-1,1){};
\node[label=left:{$(1,-1)$}][7brane]at(-2,2){};
\node[label=left:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(-2,0){};
\node[7brane]at(2,0){};
\node[label=right:{1}]at(0,-2.5){};
\node[label=right:{4}]at(0,-.5){};
\node[label=right:{2}]at(0,-1.5){};
\node[label=below:{2}]at(1.5,0){};
\node[label=below:{2}]at(-1.5,0){};
\node[label=above:{1}]at(-1.5,1.5){};
\node[label=above:{2}]at(-.5,.5){};
\node[label=above:{2}]at(.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
whose magnetic quiver is the union of the following three cones
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{2}][u](2')[right of=2]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{2}][u](22)[above right of=2]{};
\draw(1)--(2);
\draw(2)--(2');
\draw(2')--(1');
\draw(2)--(22);
\draw(22)--(2');
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=left:{1}][u](2)[below left of=1]{};
\node[label=right:{1}][u](2')[below right of=1]{};
\node[label=below:{1}][u](sp2)[below left of=2]{};
\node[label=below:{1}][u](sp2')[below right of=2']{};
\node[label=above:{1}][u](so2)[below left of=2']{};
\node[label=below:{1}][u](down)[below of=so2]{};
\draw(1)--(2);
\draw(1)--(2');
\draw(2)--(sp2);
\draw(2')--(sp2');
\draw(sp2)--(so2);
\draw(sp2')--(so2);
\draw(2)--(2');
\draw[double distance=2pt](so2)--(down);
\end{tikzpicture}
\end{scriptsize}
\end{array}\cup\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{1}][u](1')[right of=2]{};
\node[label=above:{2}][u](2')[above of=2]{};
\draw(1)--(2);
\draw(2)--(1');
\draw[double distance=2pt](2)--(2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_1+2F+2AS}
\end{equation}
The Coulomb branch and Higgs branch Hilbert series of the quivers in \eqref{SO6:MQ112} and \eqref{SO6:MQtilde112} match with the corresponding unitary quivers in \eqref{SU4_0+2F+2AS}, while those of \eqref{SO6:MQ202} match with that of \eqref{SU4_1+2F+2AS}. The results are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. Further, the HWG of the Coulomb branch of the second cone in \eqref{SU4_0+2F+2AS} is known \cite{Bourget:2019rtl}:
\begin{equation}
\HWG_{\mathcal{C}}^2\eqref{SU4_0+2F+2AS}=\PE\left[(1+\mu_1\mu_3)t^2+(\mu_1q+\mu_3q^{-1})t^3-\mu_1\mu_3t^6\right]\PE\left[\nu_1\nu_2t^2\right]\;,
\end{equation}
where $\mu_i$ and $\nu_i$ are the highest weight fugacities associated with the groups SU(4) and SU(3) respectively, and $q$ is the U(1) charge. Thus, the HWG of the Coulomb branch of the corresponding orthosymplectic quivers can be given as:
\begin{equation}
\HWG_{\mathcal{C}}(\MQ_2^{1,1,2}) = \HWG_{\mathcal{C}}(\widehat{\MQ}_2^{1,1,2}) = \HWG_{\mathcal{C}}^2\eqref{SU4_0+2F+2AS} ~.
\end{equation}
\subsubsection{SO(6)+2\textbf{S}+1\textbf{C}+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SU\texorpdfstring{$(4)_\frac{1}{2}$}{TEXT}+3\textbf{F}+2\textbf{AS}}
The brane web for SO(6)+2\textbf{S}+1\textbf{C}+2\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-2,0);
\draw[thick](0,0)--(4,0);
\draw[thick,dashed](-3,0)--(-2,0);
\draw[thick,dashed](4,0)--(5,0);
\draw[thick](0,0)--(0,3);
\draw[thick](0,0)--(-1,1);
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,2){};
\node[7brane]at(0,3){};
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\node[label=below:{O5$^-$}]at(-2.5,0){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{$3$}]at(.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=right:{$1$}]at(0,2.5){};
\node[label=right:{$2$}]at(0,1.5){};
\node[label=right:{$3$}]at(0,.5){};
\node[label=above:{$1$}]at(-.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver for this brane system is
\begin{equation}
\MQ_1^{2,1,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=above:{1}][u](1){};
\node[label=left:{2}][u](u2)[below of=1]{};
\node[label=left:{2}][sp](sp2)[below left of=u2]{};
\node[label=right:{1}][u](u1)[below right of=u2]{};
\node[label=below:{4}][so](so4)[below of=sp2]{};
\node[label=below:{4}][sp](sp4)[below of=u1]{};
\node[label=below:{2}][sp](sp2')[left of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp4]{};
\draw(1)--(u2);
\draw(u2)--(sp2);
\draw(u2)--(u1);
\draw(u1)--(sp4);
\draw(sp2)--(so4);
\draw(so4)--(sp4);
\draw(sp4)--(so2);
\draw(so4)--(sp2');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQ212}
\end{equation}
Now, let us consider the infinite coupling web diagram for SU$(4)_\frac{1}{2}$+3\textbf{F}+2\textbf{AS}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-1,0);
\draw[thick](0,0)--(2,0);
\draw[thick](0,0)--(-3,-3);
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(0,2);
\node[label=below:{$(1,1)$}][7brane]at(-3,-3){};
\node[7brane]at(-2,-2){};
\node[7brane]at(-1,-1){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[7brane]at(0,1){};
\node[7brane]at(0,2){};
\node[7brane]at(-1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[label=below:{1}]at(-2.5,-2.5){};
\node[label=below:{2}]at(-1.5,-1.5){};
\node[label=below:{4}]at(-.5,-.5){};
\node[label=below:{2}]at(1.5,0){};
\node[label=below:{4}]at(.5,0){};
\node[label=above:{2}]at(-.75,0){};
\node[label=left:{1}]at(0,1.5){};
\node[label=left:{2}]at(0,.75){};
\node[label=above:{1}]at(0.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
which gives the following magnetic quiver:
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=above:{2}][u](22)[above of=2]{};
\node[label=above:{2}][u](222)[right of=22]{};
\node[label=above:{2}][u](22')[right of=222]{};
\node[label=below:{2}][u](2')[below of=22']{};
\node[label=below:{1}][u](1')[right of=2']{};
\draw(1)--(2);
\draw(2)--(2');
\draw(2')--(1');
\draw(2)--(22);
\draw(22)--(222);
\draw(222)--(22');
\draw(2')--(22');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SU4_1/2+3F+2AS}
\end{equation}
We would like to point out that the quiver in \eqref{SO6:MQ212} has a bad USp(2) and a bad USp(4) node. For the bad USp(2) node, we can use our B2G prescription discussed before. For the bad USp(4) node, we use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series, which are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. These results are in agreement with that of the unitary quiver in \eqref{SU4_1/2+3F+2AS}.
\subsubsection{SO(6)+2\textbf{S}+2\textbf{C}+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} \texorpdfstring{SU$(4)_0$}{TEXT}+4\textbf{F}+2\textbf{AS}}
There are two inequivalent web diagrams that engineer the SO(6)+2\textbf{S}+2\textbf{C}+2\textbf{V} theory. The two brane webs differ in their asymptotic O5-plane charges. The first brane configuration, corresponding to the quiver ${[1\textbf{S}+1\textbf{C}]-\underset{\underset{\text{\large$\left[2\textbf{V}\right]$}}{\textstyle\vert}}{\text{SO}(6)}}-[1\textbf{S}+1\textbf{C}]$ is given at infinite coupling by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-4,0);
\draw[thick](0,0)--(4,0);
\draw[thick,dashed](-5,0)--(-4,0);
\draw[thick,dashed](5,0)--(4,0);
\draw[thick](0,0)--(0,4);
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(4,0){};
\node[7brane]at(3,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[7brane]at(0,4){};
\node[7brane]at(0,3){};
\node[7brane]at(0,2){};
\node[7brane]at(0,1){};
\node[label=below:{O5$^-$}]at(-4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(-3.5,0){};
\node[label=below:{$1$}]at(-2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(-1.5,0){};
\node[label=below:{$3$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(1.5,0){};
\node[label=below:{$3$}]at(.5,0){};
\node[label=right:{1}]at(0,3.5){};
\node[label=right:{2}]at(0,2.5){};
\node[label=right:{3}]at(0,1.5){};
\node[label=right:{4}]at(0,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver one reads from this brane web is
\begin{equation}
\MQ_1^{2,2,2}=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{4}][sp](sp4)[right of=so2]{};
\node[label=below:{6}][so](so6)[right of=sp4]{};
\node[label=below:{4}][sp](sp4')[right of=so6]{};
\node[label=below:{2}][so](so2')[right of=sp4']{};
\node[label=right:{4}][sp](sp44)[above of=so6]{};
\node[label=right:{3}][u](u3)[above of=sp44]{};
\node[label=right:{2}][u](u2)[above of=u3]{};
\node[label=right:{1}][u](u1)[above of=u2]{};
\draw(so2)--(sp4);
\draw(sp4)--(so6);
\draw(so6)--(sp4');
\draw(sp4')--(so2');
\draw(so6)--(sp44);
\draw(sp44)--(u3);
\draw(u3)--(u2);
\draw(u2)--(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQ222}
\end{equation}
The quiver in \eqref{SO6:MQ222} has two bad USp(4) gauge nodes. We use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series, which are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively.
The second web diagram, realising SO(6)+2\textbf{S}+2\textbf{C}+2\textbf{V}, corresponds to the infinite coupling limit of the gauge theory $[2\textbf{V}]-\text{SO}(6)-[2\textbf{S}+2\textbf{C}]$, and is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-4,0)--(8,0);
\draw[thick,dashed](-4,0)--(-5,0);
\draw[thick,dashed](8,0)--(9,0);
\draw[thick](0,0)--(0,1);
\node[7brane]at(8,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,0){};
\node[7brane]at(5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(0,1){};
\node[label=below:{O5$^+$}]at(-4.5,0){};
\node[label=below:{O5$^-$}]at(8.5,0){};
\node[label=below:{$1$}]at(-3.5,0){};
\node[label=below:{$1$}]at(-2.5,0){};
\node[label=below:{$2$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{$4$}]at(.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(1.5,0){};
\node[label=below:{$3$}]at(2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(3.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(5.5,0){};
\node[label=below:{$1$}]at(6.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(7.5,0){};
\node[label=right:{3}]at(0,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The corresponding magnetic quiver is
\begin{equation}
\widehat{\MQ}_1^{2,2,2}=\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{4}][sp](sp6)[right of=so3]{};
\node[label=below:{5}][so](so7)[right of=sp6]{};
\node[label=below:{4}][sp](sp6')[right of=so7]{};
\node[label=below:{5}][so](so5)[right of=sp6']{};
\node[label=below:{4}][sp](sp4)[right of=so5]{};
\node[label=below:{4}][so](so4)[right of=sp4]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=right:{2}][sp](sp22)[above of=so7]{};
\node[label=above:{1}][sof](soff)[above of=sp22]{};
\node[label=above:{1}][sof](sof)[above of=sp4]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp6);
\draw(sp6)--(so7);
\draw(so7)--(sp6');
\draw(sp6')--(so5);
\draw(so7)--(sp22);
\draw(so5)--(sp4);
\draw(sp4)--(sof);
\draw(sp4)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\draw(sp22)--(soff);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQtilde222}
\end{equation}
This quiver has a bad USp(4) gauge node and we use upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series as given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively.
Now let us consider the brane web for the infinite gauge coupling limit of SU$(4)_0$+4\textbf{F}+2\textbf{AS}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-3,0)--(3,0);
\draw[thick](0,-2)--(0,2);
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(3,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,2){};
\node[7brane]at(0,-1){};
\node[7brane]at(0,-2){};
\node[label=below:{1}]at(-2.5,0){};
\node[label=below:{2}]at(-1.5,0){};
\node[label=below:{4}]at(-.5,0){};
\node[label=above:{1}]at(2.5,0){};
\node[label=above:{2}]at(1.5,0){};
\node[label=above:{4}]at(.5,0){};
\node[label=left:{2}]at(0,1.5){};
\node[label=left:{4}]at(0,.5){};
\node[label=right:{2}]at(0,-1.5){};
\node[label=right:{4}]at(0,-.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver for this web diagram is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{4}][u](4)[right of=2]{};
\node[label=below:{2}][u](2')[right of=4]{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{4}][u](44)[above of=4]{};
\node[label=above:{2}][u](22)[left of=44]{};
\node[label=above:{2}][u](22')[right of=44]{};
\draw(1)--(2);
\draw(2)--(4);
\draw(4)--(2');
\draw(2')--(1');
\draw(4)--(44);
\draw(44)--(22);
\draw(44)--(22');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SU4_0+4F+2AS}
\end{equation}
The Hilbert series results of \eqref{SO6:MQ222} and \eqref{SO6:MQtilde222} mentioned in the appendix match with the Hilbert series results of the quiver \eqref{SU4_0+4F+2AS}.
\subsubsection{SO(6)+3\textbf{S}+2\textbf{C}+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT}
SU\texorpdfstring{$(4)_\frac{1}{2}$}{TEXT}+5\textbf{F}+2\textbf{AS}}
The brane web for the SCFT limit of SO$(6)$+3\textbf{S}+2\textbf{C}+2\textbf{V} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}[scale=.85]
\draw[thick](0,0)--(-4,0);
\draw[thick](0,0)--(8,0);
\draw[thick](0,0)--(0,2);
\draw[thick,dashed](-5,0)--(-4,0);
\draw[thick,dashed](8,0)--(9,0);
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(8,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,2){};
\node[label=below:{$\frac{3}{2}$}]at(-3.5,0){};
\node[label=below:{$2$}]at(-2.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(-1.5,0){};
\node[label=below:{$4$}]at(-.5,0){};
\node[label=below:{$4$}]at(.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(1.5,0){};
\node[label=below:{$3$}]at(2.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(3.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(5.5,0){};
\node[label=below:{$1$}]at(6.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(7.5,0){};
\node[label=left:{$1$}]at(0,1.5){};
\node[label=left:{$4$}]at(0,.5){};
\node[label=below:{O5$^-$}]at(8.5,0){};
\node[label=below:{O5$^-$}]at(-5.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
The magnetic quiver associated to this brane system is given by
\begin{equation}
\MQ_1^{3,2,2} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{6}][sp](sp6)[right of=so4]{};
\node[label=below:{8}][so](so8)[right of=sp6]{};
\node[label=below:{6}][sp](sp6')[right of=so8]{};
\node[label=below:{6}][so](so6')[right of=sp6']{};
\node[label=below:{4}][sp](sp4')[right of=so6']{};
\node[label=below:{4}][so](so4')[right of=sp4']{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{4}][sp](sp44)[above of=so8]{};
\node[label=left:{1}][u](u1)[above of=sp44]{};
\draw(sp2)--(so4);
\draw(so4)--(sp6);
\draw(sp6)--(so8);
\draw(so8)--(sp44);
\draw(sp44)--(u1);
\draw(so8)--(sp6');
\draw(sp6')--(so6');
\draw(so6')--(sp4');
\draw(sp4')--(so4');
\draw(so4')--(sp2');
\draw(sp2')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO6:MQ322}
\end{equation}
If, on the other hand, we consider the brane web for the SCFT limit of SU$(4)_\frac{1}{2}$+5\textbf{F}+2\textbf{AS}
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-3,0);
\draw[thick](0,0)--(4,0);
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(0,-3);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(0,1){};
\node[7brane]at(0,-1){};
\node[7brane]at(0,-2){};
\node[7brane]at(0,-3){};
\node[label=below:{1}]at(3.5,0){};
\node[label=below:{2}]at(2.5,0){};
\node[label=below:{4}]at(1.5,0){};
\node[label=below:{6}]at(.5,0){};
\node[label=above:{6}]at(-.5,0){};
\node[label=above:{4}]at(-1.5,0){};
\node[label=above:{2}]at(-2.5,0){};
\node[label=right:{4}]at(0,.5){};
\node[label=left:{4}]at(0,-.5){};
\node[label=left:{2}]at(0,-1.5){};
\node[label=left:{1}]at(0,-2.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~,
\end{equation}
we get the following magnetic quiver:
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{4}][u](4)[right of=2]{};
\node[label=below:{6}][u](6)[right of=4]{};
\node[label=below:{4}][u](4')[right of=6]{};
\node[label=below:{2}][u](2')[right of=4']{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=left:{4}][u](44)[above of=6]{};
\node[label=left:{2}][u](22)[above of=44]{};
\draw(1)--(2);
\draw(2)--(4);
\draw(4)--(6);
\draw(6)--(4');
\draw(4')--(2');
\draw(2')--(1');
\draw(6)--(44);
\draw(44)--(22);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{MQ232_u}
\end{equation}
The magnetic quiver in \eqref{MQ232_u} (resp. \eqref{SO6:MQ322}) can also be seen as the 3d mirror of a class-S theory of $A_5$ (resp. $D_4$) type on the sphere, with three regular untwisted punctures. The three (Nahm) punctures defining the $A$-type theory are $[2^2,1^2], \ [2^2,1^2],\ [2^3]$, while the three (Nahm) punctures defining the $D$-type theory are $[1^8], \ [2^4], \ [3^3,1^2]$. We find that both the four-dimensional theories have an identical spectrum given by $\{d_3=1,\, d_4=1,\, d_6=1\}$, and identical central charges given by $a=163/24$ and $c=47/6$. We take this match as a further evidence for the agreement of the magnetic quivers in \eqref{MQ232_u} and \eqref{SO6:MQ322}.
Note that the quiver in \eqref{SO6:MQ322} has a bad USp(6) gauge node. We use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series. The perturbative results up to first few orders are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. These results agree with the Hilbert series of the unitary quiver \eqref{MQ232_u}.
\subsubsection{SO(6)+3\textbf{S}+3\textbf{C}+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} \texorpdfstring{SU$(4)_0$}{TEXT}+6\textbf{F}+2\textbf{AS}}
The brane web for SO(6)+3\textbf{S}+3\textbf{C}+2\textbf{V} at infinite gauge coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}[scale=.75]
\draw[thick](10,0)--(-6,0);
\draw[thick](0,0)--(0,1.5);
\draw[thick,dashed](10,0)--(11,0);
\draw[thick,dashed](-6,0)--(-7,0);
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-5,0){};
\node[7brane]at(-6,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(8,0){};
\node[7brane]at(9,0){};
\node[7brane]at(10,0){};
\node[label=below:{$\frac{1}{2}$}]at(-5.5,0){};
\node[label=below:{$1$}]at(-4.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(-3.5,0){};
\node[label=below:{$3$}]at(-2.5,0){};
\node[label=below:{$\frac{9}{2}$}]at(-1.5,0){};
\node[label=below:{$5$}]at(-.5,0){};
\node[label=below:{$5$}]at(.5,0){};
\node[label=below:{$\frac{9}{2}$}]at(1.5,0){};
\node[label=below:{$4$}]at(2.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(3.5,0){};
\node[label=below:{$3$}]at(4.5,0){};
\node[label=below:{$\frac{5}{2}$}]at(5.5,0){};
\node[label=below:{$2$}]at(6.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(7.5,0){};
\node[label=below:{$1$}]at(8.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(9.5,0){};
\node[label=right:{$4$}]at(0,.75){};
\node[7brane]at(0,1.5){};
\node[label=below:{O5$^-$}]at(-6.5,0){};
\node[label=below:{O5$^-$}]at(10.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver one reads off from here is given by
\begin{equation}
\MQ_1^{3,3,2}= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{4}][sp](sp4)[right of=so2]{};
\node[label=below:{6}][so](so6)[right of=sp4]{};
\node[label=below:{8}][sp](sp8)[right of=so6]{};
\node[label=below:{10}][so](so10)[right of=sp8]{};
\node[label=below:{8}][sp](sp8')[right of=so10]{};
\node[label=below:{8}][so](so8)[right of=sp8']{};
\node (dots)[right of=so8]{$\cdots$};
\node[label=below:{2}][sp](sp2)[right of=dots]{};
\node[label=below:{2}][so](so2')[right of=sp2]{};
\node[label=above:{4}][sp](sp4')[above of=so10]{};
\draw(so2)--(sp4);
\draw(sp4)--(so6);
\draw(so6)--(sp8);
\draw(sp8)--(so10);
\draw(so10)--(sp4');
\draw(so10)--(sp8');
\draw(sp8')--(so8);
\draw(so8)--(dots);
\draw(dots)--(sp2);
\draw(sp2)--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO6:MQ332}
\end{equation}
On the other hand, the brane web for SU$(4)_0$+6\textbf{F}+2\textbf{AS} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-5,0)--(5,0);
\draw[thick](0,-1.5)--(0,1.5);
\node[7brane]at(-5,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(0,1.5){};
\node[7brane]at(0,-1.5){};
\node[label=below:{1}]at(-4.5,0){};
\node[label=below:{2}]at(-3.5,0){};
\node[label=below:{4}]at(-2.5,0){};
\node[label=below:{6}]at(-1.5,0){};
\node[label=below:{8}]at(-.5,0){};
\node[label=above:{1}]at(4.5,0){};
\node[label=above:{2}]at(3.5,0){};
\node[label=above:{4}]at(2.5,0){};
\node[label=above:{6}]at(1.5,0){};
\node[label=above:{8}]at(.5,0){};
\node[label=left:{4}]at(0,0.75){};
\node[label=right:{4}]at(0,-0.75){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
The magnetic quiver associated with this brane web is
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\node[label=below:{4}][u](4)[right of=2]{};
\node[label=below:{6}][u](6)[right of=4]{};
\node[label=below:{8}][u](8)[right of=6]{};
\node[label=below:{6}][u](6')[right of=8]{};
\node[label=below:{4}][u](4')[right of=6']{};
\node[label=below:{2}][u](2')[right of=4']{};
\node[label=below:{1}][u](1')[right of=2']{};
\node[label=above:{4}][u](44)[above of=8]{};
\draw(1)--(2);
\draw(2)--(4);
\draw(4)--(6);
\draw(6)--(8);
\draw(8)--(6');
\draw(6')--(4');
\draw(4')--(2');
\draw(2')--(1');
\draw(8)--(44);
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{MQ:323_u}
\end{equation}
The magnetic quiver in \eqref{MQ:323_u} (resp. \eqref{SO6:MQ332}) can also be seen as the 3d mirror of a class-S theory of $A_7$ (resp. $D_5$) type on the sphere, with three regular untwisted punctures. The three (Nahm) punctures defining the $A$-type theory are $[2^3,1^2], \ [2^3,1^2],\ [4^2]$, while the three (Nahm) punctures defining the $D$-type theory are $[1^{10}], \ [2^4,1^2], \ [5^2]$. We find for both theories an identical Coulomb branch spectrum given by $\{d_4=1,\, d_5=1,\, d_8=1\}$, and identical central charges given by $a=223/24$ and $c=65/6$. We take this match as a further evidence for the agreement of the magnetic quivers in \eqref{MQ:323_u} and \eqref{SO6:MQ332} .
Note that the quiver in \eqref{SO6:MQ332} has a bad USp(4) and a bad USp(8) gauge node. We use our upcoming prescription \cite{HSbad} to compute the Coulomb branch and Higgs branch Hilbert series. The perturbative results up to first few orders are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. These results agree with the Hilbert series of the unitary quiver \eqref{MQ:323_u}.
\section{SO(8) triality}\label{SO8 theories}
In this section we consider SO(8) gauge theories with matter in the vector \textbf{V}, spinor \textbf{S} and conjugate spinor \textbf{C} representations. We then use SO(8) triality to produce several equivalent magnetic quivers for a given theory. We will encounter theories whose Higgs branch is given as the union of several cones, where each cone is described by a distinct magnetic quiver. We will denote by $\MQ_{i;\, \text{SO}(8)}^{s,c,v}$, the magnetic quiver for the $i$-th cone of the SCFT parent of SO(8) +$s$\textbf{S}+$c$\textbf{C}+$v$\textbf{V}, i.e.
\begin{equation}
\mathcal{H}^{5\text{d}}_\infty\left(\text{SO}(8)+s\textbf{S}+c\textbf{C}+v\textbf{V}\right)=\bigcup_{i}\mathcal{C}^{3\text{d}}\left(\MQ_{i;\, \text{SO}(8)}^{s,c,v}\right) \ .
\end{equation}
\subsection{SO(8)+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+1\textbf{S}}
The web diagram for SO(8)+1\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(2,0);
\draw[thick](4,0)--(2,0);
\draw[thick,dashed](4,0)--(5,0);
\draw[thick](2,0)--(3,1);
\draw[thick](2,0)--(0,1);
\node[7brane]at(4,0){};
\node[7brane]at(3,0){};
\node[label=above:{$(2,-1)$}][7brane]at(0,1){};
\node[label=above:{$(1,1)$}][7brane]at(3,1){};
\node[label=below:{O5$^+$}]at(1,0){};
\node[label=below:{O5$^+$}]at(4.5,0){};
\node[label=below:{$1$}]at(3.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=above:{$1$}]at(1,.5){};
\node[label=above:{$1$}]at(2.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{equation}
The magnetic quiver that we obtain from here is given by
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{0,0,1} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[left of=o1]{};
\node[label=below:{1}][u](u1)[left of=sp2]{};
\node[label=below:{1}][uf](uf1)[left of=u1]{};
\node[label=above:{3}][uf](uf2)[above of=u1]{};
\node[label=above:{1}][sof](sof)[above of=sp2]{};
\draw(o1)--(sp2);
\draw(sp2)--(u1);
\draw(u1)--(uf2);
\draw(sp2)--(sof);
\path [draw,snake it](u1)--(uf1);
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u11){};
\node[label=below:{1}][u](u1)[left of=u11]{};
\node[label=below:{1}][uf](uf1)[left of=u1]{};
\node[label=above:{3}][uf](uf2)[above of=u1]{};
\node[label=above:{1}][uf](uf3)[above of=u11]{};
\draw(u11)--(u1);
\draw(u1)--(uf2);
\draw(u11)--(uf3);
\path [draw,snake it](u1)--(uf1);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO8:MQ001}
\end{equation}
On the other hand, the brane web for the infinite coupling limit of SO(8)+1\textbf{S} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](3,0)--(-2,0);
\draw[thick](0,0)--(-2,1);
\draw[thick](0,0)--(3,1.5);
\node[label=above:{$(2,-1)$}][7brane]at(-2,1){};
\node[label=above:{$(2,1)$}][7brane]at(3,1.5){};
\node[7brane]at(2,1){};
\node[label=below:{O5$^+$}]at(-1,0){};
\node[label=below:{O5$^-$}]at(1.5,0){};
\node[label=above:{1}]at(-1,.5){};
\node[label=above:{2}]at(1,.5){};
\node[label=above:{1}]at(2,1){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
whose magnetic quiver is given by
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{1,0,0} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u11){};
\node[label=below:{1}][u](u1)[left of=u11]{};
\node[label=below:{1}][uf](uf1)[left of=u1]{};
\node[label=above:{3}][uf](uf2)[above of=u1]{};
\node[label=above:{1}][uf](uf3)[above of=u11]{};
\draw(u11)--(u1);
\draw(u1)--(uf2);
\draw(u11)--(uf3);
\path [draw,snake it](u1)--(uf1);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO8:MQ100}
\end{equation}
This coincides with the magnetic quiver obtained in \cite{Akhond:2021knl} from the brane web with an O7$^+$-plane. The unrefined Coulomb and Higgs branch Hilbert series of the two quivers agree and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. We further propose the following highest weight generating function of the Coulomb branch:
\begin{equation}
\HWG_{\mathcal{C}}(\ref{SO8:MQ001})=\HWG_{\mathcal{C}}(\ref{SO8:MQ100})=\PE\left[\left(1+\mu^2\right)t^2+\left(q+q^{-1}\right)\mu t^6-\mu^2t^{12}\right] ~,
\end{equation}
where $\mu$ is the SU(2) fugacity and $q$ is the U(1) charge.
\subsection{SO(8)+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+2\textbf{S}}
The web diagram for SO(8)+2\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](0,0)--(1,0);
\draw[thick](1,0)--(5,0);
\draw[thick,dashed](5,0)--(6,0);
\draw[thick](3,0)--(4,1);
\draw[thick](3,0)--(2,1);
\node[7brane] at (1,0){};
\node[7brane] at (2,0){};
\node[7brane] at (4,0){};
\node[7brane] at (5,0){};
\node[label=above:{$(1,-1)$}][7brane] at (2,1){};
\node[label=above:{$(1,1)$}][7brane] at (4,1){};
\node[label=below:{O5$^+$}]at(0.5,0){};
\node[label=below:{O5$^+$}]at(5.5,0){};
\node[label=below:{$1$}]at(1.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$1$}]at(3.5,0){};
\node[label=below:{$1$}]at(4.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
and the corresponding magnetic quiver is given by
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{0,0,2} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp1)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp1]{};
\node[label=below:{2}][sp](sp1')[right of=so3]{};
\node[label=below:{1}][so](o1')[right of=sp1']{};
\node[label=left:{1}][u](u1)[above of=so3]{};
\node[label=above:{3}][uf](uf)[above of=u1]{};
\draw(o1)--(sp1);
\draw(sp1)--(so3);
\draw(so3)--(sp1');
\draw(sp1')--(o1');
\draw(so3)--(u1);
\draw(u1)--(uf);
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\label{SO8:MQ002}
\end{equation}
By SO(8) triality, the 5d theory SO(8)+2\textbf{V} should be identical to SO(8)+2\textbf{S}, the latter has a web at infinite coupling given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](-1,0)--(5,0);
\draw[thick](2,0)--(5,1.5);
\draw[thick](2,0)--(-1,1.5);
\node[label=above:{(2,1)}][7brane]at (5,1.5){};
\node[label=above:{(2,-1)}][7brane]at (-1,1.5){};
\node[7brane]at (4,1){};
\node[7brane]at (0,1){};
\node[label=below:{O5$^-$}]at (-.5,0){};
\node[label=below:{O5$^-$}]at (4.5,0){};
\node[label=above left:{2}]at(3,.5){};
\node[label=above right:{2}]at(1,.5){};
\node[label=above left:{1}]at(4.5,1.25){};
\node[label=above right:{1}]at(-.5,1.25){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
Here, we encounter a 5-brane web diagram that includes more than one coincident subwebs that intersect with O5-plane.
In this case, we need to consider the two types of contribution:
One is the contribution coming from one of the subweb and its mirror image, which corresponds to the factor $|2m_i|$ in the monopole formula. The other is the contribution coming from the mirror pair of different subwebs among coincident ones, which corresponds to the factor $|m_i + m_j|$ $(i < j)$ in the monopole formula. As discussed in \cite{Akhond:2020vhc}, the number of each contribution is given by
\begin{align}
\# \text{ of } |2m_i|&:~~ \frac{1}{2} (\text{Self SI}) - (\text{SI with O5} )\ ,
\cr
\# \text{ of } |m_i + m_j| &:~~ (\text{Self SI}) \qquad (i < j)\ .
\end{align}
Here, ``SI'' denotes the stable intersection number, which is used to compute the number of hypermultiplets, and ``Self SI'' means the SI with the mirror images of the considered subwebs.
Both types of contributions need to be part of the weight of representations of the unitary group. Such weights are included in two representations:
One is in rank 2 symmetric tensor representation, which we denote as ``Sym$^2$'', and the other is in rank 2 antisymmetric tensor, which we denote as ``$\wedge^2$''.
The weight $|2m_i|$ is included only in Sym$^2$ while the weight $ |m_i + m_j| $ is included in both of the representations.
In order to reproduce the contribution mentioned above, we claim that the number of these hypermultiplets are given by
\begin{align}
\# \text{ of Sym$^2$}:\quad \frac{1}{2} (\text{Self SI}) - (\text{SI with O5} )\ ,
\cr
\# \text{ of } \wedge^2:\quad \frac{1}{2} (\text{Self SI}) + (\text{SI with O5} )\ .
\end{align}
Note that ``Sym$^2$'' is interpreted as the charge 2 hypermultiplets when the gauge group is U(1). Indeed, the proposal for Sym$^2$ above is the generalization of the rule proposed in \cite{Akhond:2020vhc} regarding the number of the charge 2 hypermultiplets. Also, it was proposed in \cite{Akhond:2020vhc} that the number of $\wedge^2$ is simply given by Self SI, which now turned out to be valid only when the number of Sym$^2$ is 0. For generic case, we need to modify the rule as mentioned above.
If we apply this proposal to coincident $(p,q)$ 5-branes intersecting with O5$^-$-plane, we have
\begin{align}
\# \text{ of Sym$^2$}:\quad pq - q\ ,
\cr
\# \text{ of } \wedge^2:\quad pq + q\ .
\end{align}
By using this, we claim that the magnetic quiver for the SCFT limit of SO(8)+2\textbf{S} is given by
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{2,0,0} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=above left:{2}][u](2)[right of=1]{};
\node[label=below:{1}][u](1')[right of=2]{};
\node[label=above:{1}][uf](uf)[above of=2]{};
\draw(1)--(2);
\draw(2)--(1');
\path [draw,snake it](2)--node[right] {$\;\text{Sym}^2$}++(uf);
\draw (2)to[out=-45,in=225,loop,looseness=10](2);
\node[label=above:{$3\;\wedge^2$}][below of=2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO8:MQ200}
\end{equation}
The unrefined Coulomb and Higgs branch Hilbert series of \eqref{SO8:MQ002} and \eqref{SO8:MQ200} match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. We further propose the following highest weight generating function of the Coulomb branch:
\begin{equation}
\HWG_{\mathcal{C}}(\ref{SO8:MQ002})=\HWG_{\mathcal{C}}(\ref{SO8:MQ200})=\PE\left[\left(1+\mu_1^2\right)t^2+\mu_2^2t^4+\left(q+q^{-1}\right)\mu_2 t^6-\mu_2^2t^{12}\right] ~,
\end{equation}
where $\mu_i$ are highest weight fugacities for USp$(4)$, and $q$ is the fugacity for U(1).
\subsection{SO(8)+2\textbf{S}+1\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+2\textbf{S}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+1\textbf{S}+2\textbf{V}}
The brane web for SO(8)+2\textbf{S}+1\textbf{C} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](-1,0)--(2,0);
\draw[thick](2,0)--(-1,1.5);
\node[label=above:{(2,-1)}][7brane]at (-1,1.5){};
\node[7brane]at (.5,.75){};
\draw[thick](2,0)--(3,1);
\node[label=above:{(1,1)}][7brane]at (3,1){};
\draw[thick](2,0)--(6,0);
\node[7brane]at (4,0){};
\node[7brane]at (3,0){};
\node[7brane]at (5,0){};
\node[7brane]at (6,0){};
\draw[thick,dashed](7,0)--(6,0);
\node[label=below:{O5$^-$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(6.5,0){};
\node[label=below:{2}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(5.5,0){};
\node[label=above:{1}]at(0,1){};
\node[label=above:{2}]at(1,.5){};
\node[label=above:{2}]at(2.25,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
from which we read off the magnetic quiver
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{2,1,0} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=above left:{2}][u](u2)[right of=sp2]{};
\node[label=below:{1}][u](u1)[right of=u2]{};
\node[label=above:{1}][uf](uf)[above of =u2]{};
\draw(so2)--(sp2);
\draw(sp2)--(u2);
\draw[dashed](u1)--(u2);
\path [draw,snake it](u2)--node[right] {$\;\text{Sym}^2$}++(uf);
\draw (u2)to[out=-45,in=225,loop,looseness=10](u2);
\node[label=above:{$2\wedge^2$}][below of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO8:MQ210}
\end{equation}
The brane web for SO(8)+2\textbf{S}+1\textbf{V} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick,dashed](-1,0)--(2,0);
\draw[thick](2,0)--(-1,1.5);
\node[label=above:{(2,-1)}][7brane]at (-1,1.5){};
\node[7brane]at (.5,.75){};
\draw[thick](2,0)--(4,2);
\node[label=above:{(1,1)}][7brane]at (4,2){};
\draw[thick](2,0)--(4,0);
\node[7brane]at (4,0){};
\node[7brane]at (3,1){};
\node[7brane]at (3,0){};
\draw[thick,dashed](4,0)--(5,0);
\node[label=below:{O5$^-$}]at(-.5,0){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\node[label=below:{2}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=above:{1}]at(0,1){};
\node[label=above:{2}]at(1,.5){};
\node[label=above:{2}]at(2.25,.5){};
\node[label=above:{2}]at(3.25,1.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
From here we obtain the following magnetic quiver
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{2,0,1} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=left:{2}][u](u2)[above right of=u1]{};
\node[label=below:{1}][u](u1')[right of=u2]{};
\node[label=above:{2}][sp](sp2)[above left of=u2]{};
\node[label=above:{1}][uf](uf)[above of=u2]{};
\draw(u1)--(u2);
\draw(u2)--(u1');
\draw(u2)--(sp2);
\path [draw,snake it](u2)--node[right] {$\;\text{Sym}^2$}++(uf);
\draw (u2)to[out=-45,in=225,loop,looseness=10](u2);
\node[label=above:{$2\wedge^2$}][below of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad \contour{black}{$\xrightarrow{\text{B2G}}$} \quad
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](u1){};
\node[label=left:{2}][u](u2)[above right of=u1]{};
\node[label=below:{1}][u](u1')[right of=u2]{};
\node[label=above:{1}][u](sp2)[above left of=u2]{};
\node[label=above:{1}][uf](uf)[above of=u2]{};
\draw(u1)--(u2);
\draw(u2)--(u1');
\draw[dashed](u2)--node[left]{$\frac{1}{2}$}(sp2);
\path [draw,snake it](u2)--node[right] {$\;\text{Sym}^2$}++(uf);
\draw (u2)to[out=-45,in=225,loop,looseness=10](u2);
\node[label=above:{$2\wedge^2$}][below of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} ~,
\label{SO8:MQ201}
\end{equation}
where the contribution of the dashed line to the monopole formula is given by (see footnote \ref{footnote1})
\begin{equation}
\Delta_\text{hyp}\left(\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](1){};
\node[label=below:{2}][u](2)[right of=1]{};
\draw[dashed](1)--node[above]{$\frac{1}{2}$}++(2);
\end{tikzpicture}
\end{scriptsize}\end{array}\right)=\frac{1}{2}\left(|m_1-m_{2,1}|+|m_1+m_{2,2}|\right)\;,
\end{equation}
where $m_1$ is the magnetic flux of the U(1) node, while $m_{2,i}$ are magnetic fluxes for the U(2) gauge node.
The brane web for SO(8)+1\textbf{S}+2\textbf{V} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick, dashed](2,0)--(3,0);
\draw[thick, dashed](7,0)--(8,0);
\draw[thick](3,0)--(7,0);
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,1){};
\draw[thick](5,0)--(7,2);
\draw[thick](5,0)--(4,1);
\node[label=above:{$(1,1)$}][7brane]at(7,2){};
\node[label=above:{$(1,-1)$}][7brane]at(4,1){};
\node[label=below:{O5$^+$}]at(2.5,0){};
\node[label=below:{$1$}]at(3.5,0){};
\node[label=below:{$2$}]at(5.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=above:{1}]at(6.5,1.5){};
\node[label=above:{2}]at(5.5,.5){};
\node[label=above:{1}]at(4.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From here we read off the following magnetic quiver:
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{1,0,2} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=left:{2}][sp](sp2')[above of=so3]{};
\node[label=below:{1}][u](u1)[right of=so3]{};
\node[label=below:{1}][u](u1')[right of=u1]{};
\node[label=above:{2}][uf](uf)[above of=u1]{};
\node[label=above:{1}][uf](uf')[above of=u1']{};
\node[label=above:{1}][so](sof)[above of=sp2']{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp2');
\draw(sp2')--(sof);
\draw(so3)--(u1);
\draw(u1)--(u1');
\draw(u1)--(uf);
\draw(u1')--(uf');
\end{tikzpicture}
\end{scriptsize}
\end{array} ~.
\label{SO8:MQ102}
\end{equation}
The unrefined Coulomb and Higgs branch Hilbert series of \eqref{SO8:MQ210}, \eqref{SO8:MQ201} and \eqref{SO8:MQ102} match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. We further propose the following highest weight generating function of the Coulomb branch of these quivers:
\begin{equation}
\HWG_{\mathcal{C}}=\PE\left[\left(1+\mu_1^2+\nu^2\right)t^2+\mu_2^2t^4+\left(q+q^{-1}\right)\mu_2\nu t^6-\mu_2^2\nu^2t^{12}\right]\;,
\end{equation}
where $\mu_i$ are USp$(4)$ highest weight fugacities, $\nu$ is an SU$(2)$ highest weight fugacity, and $q$ is the U(1) charge.
\subsection{SO(8)+2\textbf{S}+2\textbf{C} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+2\textbf{S}+2\textbf{V}}
The brane web for SO(8)+2\textbf{S}+2\textbf{C} at infinite coupling is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick, dashed](0,0)--(1,0);
\draw[thick, dashed](9,0)--(10,0);
\draw[thick](1,0)--(9,0);
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(8,0){};
\node[7brane]at(9,0){};
\draw[thick](5,0)--(6,1);
\draw[thick](5,0)--(4,1);
\node[label=above:{$2(1,1)$}][7brane]at(6,1){};
\node[label=above:{$2(1,-1)$}][7brane]at(4,1){};
\node[label=below:{O5$^-$}]at(0.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(1.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$2$}]at(5.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(6.5,0){};
\node[label=below:{$1$}]at(7.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(8.5,0){};
\node[label=below:{O5$^-$}]at(9.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;.
\end{equation}
From this, we obtain the following magnetic quiver:
\begin{equation}
\MQ_{1;\, \text{SO}(8)}^{2,2,0} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{2}][u](u2)[above of=so4]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2');
\draw(so4)--(u2);
\draw (u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{$2\;\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad;\quad
\MQ_{2;\, \text{SO}(8)}^{2,2,0} = \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\draw(so2)--(sp2);
\draw[double distance=2pt](sp2)--(sp2');
\draw(sp2')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\;.
\label{SO8:MQ220}
\end{equation}
On the other hand, SO(8) triality implies the equivalence of SO(8)+2\textbf{S}+2\textbf{C} with SO(8)+2\textbf{S}+2\textbf{V}. The latter theory has a brane web at infinite coupling given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick, dashed](2,0)--(3,0);
\draw[thick, dashed](7,0)--(8,0);
\draw[thick](3,0)--(7,0);
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(6,0){};
\node[7brane]at(7,0){};
\node[7brane]at(6,1){};
\node[7brane]at(4,1){};
\draw[thick](5,0)--(7,2);
\draw[thick](5,0)--(3,2);
\node[label=above:{$(1,1)$}][7brane]at(7,2){};
\node[label=above:{$(1,-1)$}][7brane]at(3,2){};
\node[label=below:{O5$^-$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$2$}]at(5.5,0){};
\node[label=below:{$2$}]at(4.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(7.5,0){};
\node[label=above:{1}]at(6.5,1.5){};
\node[label=above:{2}]at(5.5,.5){};
\node[label=above:{1}]at(3.5,1.5){};
\node[label=above:{2}]at(4.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
From which we obtain the corresponding magnetic quiver to be
\begin{align}
\MQ_{1;\, \text{SO}(8)}^{2,0,2} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below left:{2}][u](u2)[above of=so4]{};
\node[label=above:{1}][u](u1)[left of=u2]{};
\node[label=above:{1}][u](u1')[right of=u2]{};
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(so4)--(u2);
\draw(u2)--(u1);
\draw(u2)--(u1');
\draw (u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{$2\;\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad\contour{black}{$\xrightarrow{\text{B2G}}$}\quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{1}][u](sp2')[right of=so4]{};
\node[label=below left:{2}][u](u2)[above of=so4]{};
\node[label=above:{1}][u](u1)[left of=u2]{};
\node[label=above:{1}][u](u1')[right of=u2]{};
\draw(sp2)--node[below]{$\frac{1}{2}$}++(so4);
\draw(so4)--node[below]{$\frac{1}{2}$}++(sp2');
\draw(so4)--(u2);
\draw(u2)--(u1);
\draw(u2)--(u1');
\draw (u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{$2\;\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_{2;\, \text{SO}(8)}^{2,0,2} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=below:{1}][u](u1)[right of=sp2']{};
\draw(so2)--(sp2);
\draw[double distance=2pt](sp2)--(sp2');
\draw(sp2')--(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO8:MQ202}
\end{align}
The unrefined Coulomb and Higgs branch Hilbert series of $\MQ_{i;\, \text{SO}(8)}^{2,2,0}$ and $\MQ_{i;\, \text{SO}(8)}^{2,0,2}$ match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG's of the Coulomb branches for the second cones take a very simple form and are given as
\begin{equation}
\HWG_{\mathcal{C}}(\MQ_{2;\, \text{SO}(8)}^{2,2,0}) = \HWG_{\mathcal{C}}(\MQ_{2;\, \text{SO}(8)}^{2,0,2}) = \PE\left[\mu_1^2t^2+\left(\mu_2+\mu_2^2\right)t^4 +\mu_1^2t^6+\mu_1^2\mu_2t^8-\mu_1^4\mu_2^2t^{16}\right]
\end{equation}
where $\mu_i$ are USp$(4)$ highest weight fugacities.
\subsection{SO(8)+1\textbf{S}+1\textbf{C}+2\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+2\textbf{S}+1\textbf{C}+1\textbf{V}}
The brane web for the SCFT limit of SO$(8)$+1\textbf{S}+1\textbf{C}+2\textbf{V} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-4,0);
\draw[thick](0,0)--(4,0);
\draw[thick,dashed](-5,0)--(-4,0);
\draw[thick,dashed](5,0)--(4,0);
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-4,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(0,1){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\draw[thick](0,0)--(1,1);
\draw[thick](0,0)--(0,1);
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(.5,0){};
\node[label=below:{$1$}]at(-3.5,0){};
\node[label=below:{$1$}]at(-2.5,0){};
\node[label=below:{$2$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{O5$^+$}]at(-4.5,0){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\node[label=left:{$1$}]at(0,.5){};
\node[label=right:{$2$}]at(0.5,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
The magnetic quivers associated to this brane system are
\begin{align}
\MQ_{1;\, \text{SO}(8)}^{1,1,2} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so3]{};
\node[label=below:{3}][so](so3')[right of=sp4]{};
\node[label=below:{2}][sp](sp2')[right of=so3']{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=left:{1}][u](u1)[above of=sp4]{};
\node[label=above:{2}][uf](uf)[above of=u1]{}; \node[label=above:{1}][sof](sof)[above of=sp2']{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp4);
\draw(sp4)--(u1);
\draw(u1)--(uf);
\draw(sp4)--(so3');
\draw(so3')--(sp2');
\draw(sp2')--(sof);
\draw(sp2')--(so2);
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_{2;\, \text{SO}(8)}^{1,1,2} &=
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][so](o1){};
\node[label=below:{2}][sp](sp2)[right of=o1]{};
\node[label=below:{3}][so](so3)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so3]{};
\node[label=left:{2}][sp](sp22)[above of=so3]{};
\node[label=below:{1}][so](sof)[right of=sp2']{};
\node[label=above:{5}][sof](soff)[above of=sp22]{};
\draw(o1)--(sp2);
\draw(sp2)--(so3);
\draw(so3)--(sp2');
\draw(sp2')--(sof);
\draw(so3)--(sp22);
\draw(sp22)--(soff);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO8:MQ112}
\end{align}
The brane web for the SCFT limit of SO$(8)$+2\textbf{S}+1\textbf{C}+1\textbf{V} is given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](0,0)--(-2,0);
\draw[thick](0,0)--(4,0);
\draw[thick](0,0)--(-2,2);
\draw[thick](0,0)--(1,1);
\draw[thick,dashed](5,0)--(4,0);
\draw[thick,dashed](-3,0)--(-2,0);
\node[7brane]at(-1,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(-1,1){};
\node[label=above:{$(1,-1)$}][7brane]at(-2,2){};
\node[label=above:{$(1,1)$}][7brane]at(1,1){};
\node[label=below:{O5$^-$}]at(-2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(.5,0){};
\node[label=above:{1}]at(-1.5,1.5){};
\node[label=above:{2}]at(-.5,.5){};
\node[label=above:{2}]at(.5,.5){};
\node[label=below:{O5$^-$}]at(4.5,0){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
The magnetic quiver that we read off here is
\begin{align}
\MQ_{1;\, \text{SO}(8)}^{2,1,1} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=right:{2}][u](u2)[above of=so4]{};
\node[label=right:{1}][u](u1)[above of=u2]{};
\draw(u2)to[out=135,in=-135,loop,looseness=10](u2);
\node[left of=u2]{2 $\wedge^2$};
\draw(u1)--(u2);
\draw(u2)--(so4);
\draw(sp2)--(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\end{tikzpicture}
\end{scriptsize}
\end{array} \quad\contour{black}{$\xrightarrow{\text{B2G}}$}\quad \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{1}][u](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{2}][sp](sp2')[right of=so4]{};
\node[label=below:{2}][so](so2)[right of=sp2']{};
\node[label=right:{2}][u](u2)[above of=so4]{};
\node[label=right:{1}][u](u1)[above of=u2]{};
\draw(u2)to[out=135,in=-135,loop,looseness=10](u2);
\node[left of=u2]{2 $\wedge^2$};
\draw(u1)--(u2);
\draw(u2)--(so4);
\draw(sp2)--node[below]{$\frac{1}{2}$}++(so4);
\draw(so4)--(sp2');
\draw(sp2')--(so2);
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_{2;\, \text{SO}(8)}^{2,1,1} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{2}][sp](sp2')[right of=sp2]{};
\node[label=above:{1}][u](u1)[above of=sp2]{};
\draw(so2)--(sp2);
\draw[double distance=2pt](sp2)--(sp2');
\draw(sp2)--(u1);
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO8:MQ211}
\end{align}
The unrefined Coulomb and Higgs branch Hilbert series of $\MQ_{i;\, \text{SO}(8)}^{1,1,2}$ and $\MQ_{i;\, \text{SO}(8)}^{2,1,1}$ match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively. The HWG's for the Coulomb branches for the second cones takes a very simple form
\begin{equation}
\HWG_{\mathcal{C}}(\MQ_{2;\, \text{SO}(8)}^{1,1,2}) = \HWG_{\mathcal{C}}(\MQ_{2;\, \text{SO}(8)}^{2,1,1}) = \PE\left[\mu_1^2t^2+\left(1+\mu_2+\mu_2^2\right)t^4 +\mu_2t^6-\mu_2^2t^{12}\right]\;,
\end{equation}
where the $\mu_i$ are USp(4) highest weight fugacities.
\subsection{SO(8)+2\textbf{S}+2\textbf{C}+1\textbf{V} \texorpdfstring{$\longleftrightarrow$}{TEXT} SO(8)+2\textbf{S}+1\textbf{C}+2\textbf{V}}
The brane web for the SCFT limit of SO$(8)$+2\textbf{S}+2\textbf{C}+1\textbf{V} takes the form
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](6,0)--(-4,0);
\draw[thick,dashed](-5,0)--(-4,0);
\draw[thick,dashed](6,0)--(7,0);
\node[7brane]at(-4,0){};
\node[7brane]at(-3,0){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(6,0){};
\node[7brane]at(5,0){};
\node[7brane]at(4,0){};
\node[7brane]at(3,0){};
\node[7brane]at(2,0){};
\node[7brane]at(1,0){};
\node[7brane]at(0,1){};
\node[label=above:{$(1,-1)$}][7brane]at(-1,1){};
\draw[thick](0,0)--(0,1);
\draw[thick](0,0)--(-1,1);
\node[label=below:{$\frac{1}{2}$}]at(-3.5,0){};
\node[label=below:{$1$}]at(-2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{$4$}]at(.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(5.5,0){};
\node[label=below:{O5$^-$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(-4.5,0){};
\node[label=right:{$2$}]at(0,0.5){};
\node[label=left:{$2$}]at(-.5,0.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}
\end{equation}
The magnetic quivers associated to this brane system are
\begin{align}
\MQ_{1;\, \text{SO}(8)}^{2,2,1} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{6}][sp](sp6)[right of=so4]{};
\node[label=below:{4}][so](so4')[right of=sp6]{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{2}][u](u2)[above of=sp6]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp6);
\draw(sp6)--(so4');
\draw(so4')--(sp2');
\draw(sp2')--(so2');
\draw(sp6)--(u2);
\draw(u2)to[out=45,in=135,loop,looseness=10](u2);
\node[label=below:{1 $\wedge^2$}][above of=u2]{};
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_{2;\, \text{SO}(8)}^{2,2,1} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][so](so2){};
\node[label=below:{2}][sp](sp2)[right of=so2]{};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{4}][so](so4')[right of=sp4]{};
\node[label=above:{2}][sp](sp22)[above of=so4]{};
\node[label=above:{2}][sp](sp22')[above of=so4']{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp22);
\draw(sp22)--(sp22');
\draw(sp22')--(so4');
\draw(so4)--(sp4);
\draw(so4')--(sp4);
\draw(so4')--(sp2');
\draw(sp2')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO8:MQ221}
\end{align}
If, instead, we take the SCFT limit of the brane web for SO$(8)$+2\textbf{S}+1\textbf{C}+2\textbf{V}, given by
\begin{equation}
\begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\draw[thick](-2,0)--(6,0);
\draw[thick,dashed](-2,0)--(-3,0);
\draw[thick,dashed](6,0)--(7,0);
\draw[thick](0,0)--(-2,2);
\draw[thick](0,0)--(0,1);
\node[label=above:{$(1,-1)$}][7brane]at(-2,2){};
\node[7brane]at(-1,1){};
\node[7brane]at(0,1){};
\node[7brane]at(-2,0){};
\node[7brane]at(-1,0){};
\node[7brane]at(1,0){};
\node[7brane]at(2,0){};
\node[7brane]at(3,0){};
\node[7brane]at(4,0){};
\node[7brane]at(5,0){};
\node[7brane]at(6,0){};
\node[label=below:{$\frac{3}{2}$}]at(-1.5,0){};
\node[label=below:{$2$}]at(-.5,0){};
\node[label=below:{$4$}]at(.5,0){};
\node[label=below:{$\frac{7}{2}$}]at(1.5,0){};
\node[label=below:{$2$}]at(2.5,0){};
\node[label=below:{$\frac{3}{2}$}]at(3.5,0){};
\node[label=below:{$1$}]at(4.5,0){};
\node[label=below:{$\frac{1}{2}$}]at(5.5,0){};
\node[label=below:{O5$^-$}]at(6.5,0){};
\node[label=below:{O5$^-$}]at(-2.5,0){};
\node[label=left:{$1$}]at(-1.5,1.5){};
\node[label=left:{$2$}]at(-.5,.5){};
\node[label=right:{$2$}]at(0,.5){};
\end{tikzpicture}
\end{scriptsize}
\end{array}\;,
\end{equation}
we obtain the following alternative magnetic quivers
\begin{align}
\MQ_{1;\, \text{SO}(8)}^{2,1,2} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{6}][sp](sp6)[right of=so4]{};
\node[label=below:{4}][so](so4')[right of=sp6]{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=left:{2}][u](u2)[above of=sp6]{};
\node[label=left:{1}][u](u1)[above of=u2]{};
\draw(so2)--(sp2);
\draw(sp2)--(so4);
\draw(so4)--(sp6);
\draw(sp6)--(so4');
\draw(so4')--(sp2');
\draw(sp2')--(so2');
\draw(sp6)--(u2);
\draw(u2)--(u1);
\draw(u2)to[out=45,in=-45,loop,looseness=10](u2);
\node[right of=u2]{1 $\wedge^2$};
\end{tikzpicture}
\end{scriptsize}
\end{array} \nonumber \\
\MQ_{2;\, \text{SO}(8)}^{2,1,2} &= \begin{array}{c}
\begin{scriptsize}
\begin{tikzpicture}
\node[label=below:{2}][sp](sp2){};
\node[label=below:{4}][so](so4)[right of=sp2]{};
\node[label=below:{4}][sp](sp4)[right of=so4]{};
\node[label=below:{4}][so](so4')[right of=sp4]{};
\node[label=above:{2}][sp](sp22)[above of=so4]{};
\node[label=above:{2}][sp](sp22')[above of=so4']{};
\node[label=below:{2}][sp](sp2')[right of=so4']{};
\node[label=below:{2}][so](so2')[right of=sp2']{};
\node[label=above:{1}][u](u1)[right of=sp22']{};
\draw(u1)--(sp22');
\draw(sp2)--(so4);
\draw(so4)--(sp22);
\draw(sp22)--(sp22');
\draw(sp22')--(so4');
\draw(so4)--(sp4);
\draw(so4')--(sp4);
\draw(so4')--(sp2');
\draw(sp2')--(so2');
\end{tikzpicture}
\end{scriptsize}
\end{array}
\label{SO8:MQ212}
\end{align}
The unrefined Coulomb and Higgs branch Hilbert series of $\MQ_{i;\, \text{SO}(8)}^{2,2,1}$ and $\MQ_{i;\, \text{SO}(8)}^{2,1,2}$ match and are given in table \ref{CoulombHSOSp} and table \ref{HiggsHSOSp} respectively.
\section{Unrefined Hilbert series results for OSp quivers}\label{appendixA}
\captionsetup{width=15cm}
\begin{longtable}{|c|C{6.11cm}|C{6.10cm}|}
\caption{Unrefined Coulomb branch Hilbert series for the orthosymplectic quivers.}
\label{CoulombHSOSp} \\ \hline
& \multicolumn{2}{c|}{$\text{HS}_{\mathcal{C}}(t)=\text{HS}_{\mathcal{C}}(t;\vec{m} \in \mathbb{Z})+\text{HS}_{\mathcal{C}}(t;\vec{m} \in \mathbb{Z}+\tfrac{1}{2})$} \\ \cline{2-3}
\multirow{-2}{*}{Quiver} & $\text{HS}_{\mathcal{C}}(t;\vec{m} \in \mathbb{Z})$ & $\text{HS}_{\mathcal{C}}(t;\vec{m} \in \mathbb{Z}+\tfrac{1}{2})$ \\[0.1cm] \hline
\hyperref[SO6:MQ110]{$\MQ_1^{1,1,0}$} & \CBHS{1 + 3 t^4 + 2 t^6 + 3 t^8 + t^{12}}{(1-t^2)^2\,(1-t^4) \,(1-t^8)}{1 + 2 t^2 + 7 t^4 + 14 t^6 + 29 t^8 + 46 t^{10}} & \CBHS{2 t^2 + 2 t^4 + 2 t^6 + 2 t^8 + 2 t^{10}}{(1-t^2)^2\,(1-t^4) \,(1-t^8)}{2 t^2 + 6 t^4 + 14 t^6 + 26 t^8 + 46 t^{10}} \\ \hline
\hyperref[SO6:MQ200]{$\MQ_1^{2,0,0}$} & \CBHS{1 + t^4}{(1 - t^2)\,(1 - t^4)}{1 + t^2 + 3 t^4 + 3 t^6 + 5 t^8 + 5 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQ200]{$\MQ_2^{2,0,0}$} & \CBHS{1 + t^4}{(1 - t^2)\,(1 - t^4)}{1 + t^2 + 3 t^4 + 3 t^6 + 5 t^8 + 5 t^{10}} & \CBHS{2t^2}{(1 - t^2)\,(1 - t^4)}{2 t^2 + 2 t^4 + 4 t^6 + 4 t^8 + 6 t^{10}} \\ \hline
\hyperref[SO6:MQ210]{$\MQ_1^{2,1,0}$} & \CBHS{P_1(t)}{(1-t^2)^3\,(1-t^4)^2 \,(1-t^8)}{1 + 5 t^2 + 24 t^4 + 72 t^6 + 189 t^8 + 413 t^{10}} & \CBHS{4 t^2 + 6 t^4 + 14 t^6 + 14 t^8 + 14 t^{10} + 6 t^{12} + 4 t^{14}}{(1-t^2)^3\,(1-t^4)^2 \,(1-t^8)}{4 t^2 + 18 t^4 + 64 t^6 + 168 t^8 + 388 t^{10}} \\ \hline
\hyperref[SO6:MQ220]{$\MQ_1^{2,2,0}$} & \CBHS{P_2(t)}{(1-t^2)^5\,(1-t^4)^4 \,(1-t^8)}{1 + 8 t^2 + 72 t^4 + 371 t^6 + 1598 t^8 + 5510 t^{10}} & \CBHS{P_3(t)}{(1-t^2)^5\,(1-t^4)^4 \,(1-t^8)}{8 t^2 + 60 t^4 + 364 t^6 + 1536 t^8 + 5464 t^{10}} \\ \hline
\hyperref[SO6:MQ220]{$\MQ_2^{2,2,0}$} & \CBHS{1+t^4}{(1-t^2)\,(1-t^4)}{1 + t^2 + 3 t^4 + 3 t^6 + 5 t^8 + 5 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQtilde220]{$\widehat{\MQ}_1^{2,2,0}$} & \CBHS{P_4(t)}{(1-t^2)^5\,(1-t^4)^5}{1 + 16 t^2 + 132 t^4 + 735 t^6 + 3134 t^8 \\+ 10974 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQtilde220]{$\widehat{\MQ}_2^{2,2,0}$} & \CBHS{1+t^4}{(1-t^2)\,(1-t^4)}{1 + t^2 + 3 t^4 + 3 t^6 + 5 t^8 + 5 t^{10}} & \footnotesize{not required} \\\hline
\hyperref[SO6:MQ320]{$\MQ_1^{3,2,0}$} & \footnotesize{$1 + 17 t^2 + 184 t^4 + 1446 t^6 + 8758 t^8 + 43000 t^{10} +
178362 t^{12} + 644654 t^{14} + 2079047 t^{16} + 6092795 t^{18} +
16458838 t^{20}+\ldots$} & \footnotesize{$8 t^2 + 136 t^4 + 1248 t^6 + 8072 t^8 + 40952 t^{10} + 172888 t^{12} +
631376 t^{14} + 2049176 t^{16} + 6030000 t^{18} + 16333832 t^{20}+\ldots$} \\\hline
\hyperref[SO6:MQ320]{$\MQ_2^{3,2,0}$} & \CBHS{P_5(t)}{(1-t^2)^5\,(1-t^4)\,(1-t^8)^4}{1 + 17 t^2 + 139 t^4 + 751 t^6 + 3148 t^8\\ + 10894 t^{10}} & \CBHS{P_6(t)}{(1-t^2)^5\,(1-t^4)\,(1-t^8)^4}{8 t^2 + 96 t^4 + 624 t^6 + 2832 t^8 + 10232 t^{10}} \\ \hline
\hyperref[SO6:MQ330]{$\MQ_1^{3,3,0}$} & \footnotesize{$1 + 20 t^2 + 350 t^4 + 4199 t^6 + 38358 t^8 + 278738 t^{10} +
1683601 t^{12} + 8713628 t^{14} + 39600362 t^{16} + 161030946 t^{18} +
594866176 t^{20}+\ldots$} & \footnotesize{$16 t^2 + 320 t^4 + 4096 t^6 + 37920 t^8 + 277504 t^{10} + 1679744 t^{12} + 8704320 t^{14} + 39576496 t^{16} + 160979760 t^{18} +
594751936 t^{20}+\ldots$} \\\hline
\hyperref[SO6:MQ330]{$\MQ_2^{3,3,0}$} & \footnotesize{$1 + 20 t^2 + 340 t^4 + 3926 t^6 + 33526 t^8 + 224534 t^{10} +
1240034 t^{12} + 5850463 t^{14} + 24223718 t^{16} + 89832720 t^{18} +
303191840 t^{20}+\ldots$} & \footnotesize{$16 t^2 + 320 t^4 + 3872 t^6 + 33344 t^8 + 224128 t^{10} +
1238960 t^{12} + 5848352 t^{14} + 24218944 t^{16} + 89824160 t^{18} +
303174592 t^{20}+\ldots$} \\ \hline
\hyperref[SO6:MQ440]{$\MQ_1^{4,4,0}$} & \footnotesize{$1 + 37 t^2 + 1350 t^4 + 34389 t^6 + 668310 t^8 + 10281564 t^{10} +
129992857 t^{12} + 1388266357 t^{14} + 12803207039 t^{16} +
103789879656 t^{18} + 750444248396 t^{20}+\ldots$} & \footnotesize{$32 t^2 + 1280 t^4 + 34080 t^6 + 666016 t^8 + 10272864 t^{10} +
129948064 t^{12} + 1388117408 t^{14} + 12802606720 t^{16} +
103788095648 t^{18} + 750438227104 t^{20}+\ldots$} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_1^{2,0,1}$} & \CBHS{P_7(t)}{(1-t^2)^2\,(1-t^4)\,(1-t^6)^3}{1 + 5 t^2 + 18 t^4 + 58 t^6 + 149 t^8 + 325 t^{10}} & \CBHS{2t^3 P_8(t)}{(1-t^2)\,(1-t^4)^2\,(1-t^6)^3}{6 t^3 + 26 t^5 + 78 t^7 + 198 t^9} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_2^{2,0,1}$} & \CBHS{1+t^4}{(1-t^2)^4}{1 + 4 t^2 + 11 t^4 + 24 t^6 + 45 t^8 + 76 t^{10}} & \CBHS{2t^2}{(1-t^2)^4}{2 t^2 + 8 t^4 + 20 t^6 + 40 t^8 + 70 t^{10}} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_3^{2,0,1}$} & \CBHS{1 - 2 t + 3 t^2 - 4 t^3 + 6 t^4 - 4 t^5 + 3 t^6 - 2 t^7 + t^8}{(1-t)^2\,(1-t^4)\,(1-t^6)}{1 + 2 t^2 + 5 t^4 + 4 t^5 + 10 t^6 + 8 t^7 + 17 t^8\\ + 16 t^9 + 28 t^{10}} & \CBHS{2 t^2 - 2 t^3 + 2 t^4 - 2 t^5 + 2 t^6}{(1-t)^2\,(1-t^4)\,(1-t^6)}{2 t^2 + 2 t^3 + 4 t^4 + 4 t^5 + 8 t^6 + 10 t^7 + 16 t^8\\ + 18 t^9 +
26 t^{10}} \\ \hline
\hyperref[SO6:MQ111]{$\MQ_1^{1,1,1}$} & \CBHS{P_9(t)}{(1-t^2)^3\,(1-t^4)\,(1-t^8)^2}{1 + 5 t^2 + 20 t^4 + 60 t^6 + 157 t^8 + 345 t^{10}} & \CBHS{(1-t^4)P_{10}(t)}{(1-t^2)^5\,(1-t^8)^2}{2 t^2 + 14 t^4 + 50 t^6 + 136 t^8 + 314 t^{10}} \\ \hline
\hyperref[SO6:MQ111]{$\MQ_2^{1,1,1}$} & \CBHS{1 + 2 t^2 + 2 t^4 + 6 t^6 + 2 t^8 + 2 t^{10} + t^{12}}{(1-t^2)^2\,(1-t^6)^2}{1 + 4 t^2 + 9 t^4 + 22 t^6 + 41 t^8 + 66 t^{10}} & \CBHS{4 t^3 + 4 t^5 + 4 t^7 + 4 t^9}{(1-t^2)^2\,(1-t^6)^2}{4 t^3 + 12 t^5 + 24 t^7 + 48 t^9} \\ \hline
\hyperref[SO6:MQ211]{$\MQ_1^{2,1,1}$} & \footnotesize{$1 + 9 t^2 + 4 t^3 + 57 t^4 + 52 t^5 + 291 t^6 + 312 t^7 + 1172 t^8 +
1360 t^9 + 3932 t^{10}+\ldots$} & \footnotesize{$4 t^2 + 10 t^3 + 36 t^4 + 82 t^5 + 208 t^6 + 426 t^7 + 920 t^8 +
1708 t^9 + 3276 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ211]{$\MQ_2^{2,1,1}$} & \CBHS{1 + t^2 + 7 t^4 + 6 t^6 + 7 t^8 + t^{10} + t^{12}}{(1-t^2)^3\,(1-t^4)^3}{1 + 4 t^2 + 19 t^4 + 55 t^6 + 146 t^8 + 317 t^{10}} & \CBHS{4 t^2 + 4 t^4 + 8 t^6 + 4 t^8 + 4 t^{10}}{(1-t^2)^3\,(1-t^4)^3}{4 t^2 + 16 t^4 + 56 t^6 + 140 t^8 + 320 t^{10}} \\ \hline
\hyperref[SO6:MQ221]{$\MQ_1^{2,2,1}$} & \footnotesize{$1 + 20 t^2 + 28 t^3 + 220 t^4 + 456 t^5 + 1905 t^6 + 4132 t^7 +
13026 t^8 + 27600 t^9 + 72438 t^{10}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQ221]{$\MQ_2^{2,2,1}$} & \footnotesize{$1 + 19 t^2 + 24 t^3 + 188 t^4 + 368 t^5 + 1396 t^6 + 2968 t^7 +
8302 t^8 + 17168 t^9 + 40474 t^{10} + 79648 t^{11} + 167230 t^{12} +
312656 t^{13} + 603338 t^{14} + 1074896 t^{15} + 1943919 t^{16} +
3316912 t^{17} + 5693149 t^{18} + 9351600 t^{19} + 15372660 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQhat221]{$\widehat{\MQ}_1^{2,2,1}$} & \footnotesize{$1 + 12 t^2 + 8 t^3 + 124 t^4 + 200 t^5 + 1033 t^6 + 1912 t^7 +
6834 t^8 + 13192 t^9 + 37406 t^{10}+\ldots$} & \footnotesize{$8 t^2 + 20 t^3 + 96 t^4 + 256 t^5 + 872 t^6 + 2220 t^7 + 6192 t^8 +
14408 t^9 + 35032 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQhat221]{$\widehat{\MQ}_2^{2,2,1}$} & \footnotesize{$1 + 11 t^2 + 8 t^3 + 100 t^4 + 176 t^5 + 724 t^6 + 1448 t^7 +
4206 t^8 + 8496 t^9 + 20394 t^{10}+\ldots$} & \footnotesize{$8 t^2 + 16 t^3 + 88 t^4 + 192 t^5 + 672 t^6 + 1520 t^7 + 4096 t^8 +
8672 t^9 + 20080 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ321]{$\MQ_1^{3,2,1}$} & \footnotesize{$1 + 21 t^2 + 28 t^3 + 292 t^4 + 710 t^5 + 3576 t^6 + 10234 t^7 +
37720 t^8 + 107242 t^9 + 331772 t^{10}+\ldots$} & \footnotesize{$8 t^2 + 24 t^3 + 192 t^4 + 688 t^5 + 2992 t^6 + 10144 t^7 +
34904 t^8 + 106960 t^9 + 320096 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ331]{$\MQ_1^{3,3,1}$} & \footnotesize{$1 + 24 t^2 + 36 t^3 + 497 t^4 + 1544 t^5 + 9096 t^6 + 32876 t^7 +
141780 t^8 + 501144 t^9 + 1831783 t^{10}+\ldots$} & \footnotesize{$16 t^2 + 48 t^3 + 432 t^4 + 1648 t^5 + 8656 t^6 + 33680 t^7 +
139248 t^8 + 505904 t^9 + 1818896 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ441]{$\MQ_1^{4,4,1}$} & \footnotesize{$1 + 46 t^2 + 96 t^3 + 1836 t^4 + 8256 t^5 + 66981 t^6+\ldots$} & \footnotesize{$32 t^2 + 128 t^3 + 1664 t^4 + 8704 t^5 + 64960 t^6+\ldots$} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_1^{1,1,2}$} & \footnotesize{$1 + 18 t^2 + 205 t^4 + 1591 t^6 + 9499 t^8 + 45959 t^{10} +
188535 t^{12} + 675381 t^{14} + 2163400 t^{16} + 6305809 t^{18} +
16961842 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_2^{1,1,2}$} & \footnotesize{$ 1 + 24 t^2 + 8 t^3 + 255 t^4 + 144 t^5 + 1716 t^6 + 1256 t^7 +
8594 t^8 + 7312 t^9 + 34872 t^{10} + 32640 t^{11} + 120628 t^{12} +
120336 t^{13} + 367968 t^{14} + 383440 t^{15} + 1013621 t^{16} +
1088720 t^{17} + 2565512 t^{18} + 2814744 t^{19} + 6045369 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_3^{1,1,2}$} & \footnotesize{$1 + 11 t^2 + 8 t^3 + 65 t^4 + 80 t^5 + 295 t^6 + 432 t^7 + 1122 t^8 +
1720 t^9 + 3666 t^{10} + 5640 t^{11} + 10564 t^{12} + 16024 t^{13} +
27460 t^{14} + 40752 t^{15} + 65445 t^{16} + 94888 t^{17} + 144935 t^{18} +
205440 t^{19} + 301571 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_1^{1,1,2}$} & \footnotesize{$1 + 14 t^2 + 129 t^4 + 895 t^6 + 5071 t^8 + 23887 t^{10} + 96579 t^{12} +
343049 t^{14} + 1093344 t^{16} + 3176685 t^{18} + 8527154 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 76 t^4 + 696 t^6 + 4428 t^8 + 22072 t^{10} + 91956 t^{12} +
332332 t^{14} + 1070056 t^{16} + 3129124 t^{18} + 8434688 t^{20}+\ldots$} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_2^{1,1,2}$} & \footnotesize{$1 + 20 t^2 + 179 t^4 + 32 t^5 + 1092 t^6 + 416 t^7 + 5142 t^8 +
2816 t^9 + 20024 t^{10} + 13600 t^{11} + 67328 t^{12} + 52544 t^{13} +
201208 t^{14} + 172576 t^{15} + 545849 t^{16} + 500352 t^{17} +
1365524 t^{18} + 1313248 t^{19} + 3188473 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 8 t^3 + 76 t^4 + 112 t^5 + 624 t^6 + 840 t^7 + 3452 t^8 +
4496 t^9 + 14848 t^{10} + 19040 t^{11} + 53300 t^{12} + 67792 t^{13} +
166760 t^{14} + 210864 t^{15} + 467772 t^{16} + 588368 t^{17} +
1199988 t^{18} + 1501496 t^{19} + 2856896 t^{20}+\ldots$} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_3^{1,1,2}$} & \footnotesize{$1 + 11 t^2 + 65 t^4 + 295 t^6 + 1122 t^8 + 3666 t^{10} + 10564 t^{12} +
27460 t^{14} + 65445 t^{16} + 144935 t^{18} + 301571 t^{20}+\ldots$} & \footnotesize{$8 t^3 + 80 t^5 + 432 t^7 + 1720 t^9 + 5640 t^{11} + 16024 t^{13} +
40752 t^{15} + 94888 t^{17} + 205440 t^{19}+\ldots$} \\ \hline
\hyperref[SO6:MQ202]{$\MQ_1^{2,0,2}$} & \footnotesize{$1 + 14 t^2 + 119 t^4 + 806 t^6 + 4480 t^8 + 20886 t^{10} + 83778 t^{12}+\ldots$} & \footnotesize{$16 t^3 + 208 t^5 + 1568 t^7 + 8736 t^9 +
39552 t^{11}+\ldots$} \\ \hline
\hyperref[SO6:MQ202]{$\MQ_2^{2,0,2}$} & \footnotesize{$1 + 20 t^2 + 175 t^4 + 32 t^5 + 1060 t^6 + 416 t^7 + 4994 t^8 +
2784 t^9 + 19432 t^{10} + 13376 t^{11} + 65340 t^{12} + 51584 t^{13} + 195360 t^{14} + 169120 t^{15} + 530141 t^{16} +
489760 t^{17} + 1326692 t^{18} + 1284544 t^{19} + 3098889 t^{20} + \ldots$} & \footnotesize{$2 t^2 + 16 t^3 + 40 t^4 + 208 t^5 + 348 t^6 + 1408 t^7 + 2080 t^8 + 6896 t^9 + 9640 t^{10} + 27184 t^{11} + 36784 t^{12} + 91248 t^{13} + 121040 t^{14} + 270672 t^{15} + 353936 t^{16} + 726832 t^{17} + 939242 t^{18} + 1797536 t^{19} + 2299448 t^{20} + \ldots$} \\ \hline
\hyperref[SO6:MQ202]{$\MQ_3^{2,0,2}$} &\footnotesize{$1 + 11 t^2 + 68 t^4 + 313 t^6 + 1202 t^8 + 3953 t^{10} + 11453 t^{12} + 29842 t^{14} + 71275 t^{16} + 158094 t^{18} + 329343 t^{20} + \ldots$} & \footnotesize{$2 t^2 + 32 t^4 + 214 t^6 + 972 t^8 + 3472 t^{10} + 10544 t^{12} + 28260 t^{14} + 68662 t^{16} + 153948 t^{18} + 323034 t^{20} + \ldots$} \\ \hline
\hyperref[SO6:MQ212]{$\MQ_1^{2,1,2}$} & \footnotesize{$1 + 20 t^2 + 8 t^3 + 275 t^4 + 296 t^5 + 3045 t^6 + 4800 t^7 +
27790 t^8 + 51128 t^9 + 211871 t^{10}+\ldots$} & \footnotesize{$6 t^2 + 24 t^3 + 150 t^4 + 504 t^5 + 2144 t^6 + 6528 t^7 +
22284 t^8 + 61992 t^9 + 182902 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ222]{$\MQ_1^{2,2,2}$} & \footnotesize{$1 + 27 t^2 + 16 t^3 + 537 t^4 + 992 t^5 + 9266 t^6 + 23904 t^7 +
133582 t^8 + 383712 t^9 + 1626798 t^{10}+\ldots$} & \footnotesize{$12 t^2 + 48 t^3 + 384 t^4 + 1440 t^5 + 7700 t^6 + 28192 t^7 +
120740 t^8 + 417440 t^9 + 1536816 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQtilde222]{$\widehat{\MQ}_1^{2,2,2}$} & \footnotesize{$1 + 39 t^2 + 64 t^3 + 921 t^4 + 2432 t^5 + 16966 t^6 + 52096 t^7 +
254322 t^8 + 801152 t^9 + 3163614 t^{10}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO6:MQ322]{$\MQ_1^{3,2,2}$} & \footnotesize{$1 + 40 t^2 + 64 t^3 + 1104 t^4 + 3520 t^5 + 26972 t^6 + 108032 t^7 +
587528 t^8 + 2428480 t^9 + 11018073 t^{10}+\ldots$} & \footnotesize{$16 t^2 + 64 t^3 + 736 t^4 + 3520 t^5 + 22848 t^6 + 108032 t^7 +
549792 t^8 + 2428480 t^9 + 10724912 t^{10}+\ldots$} \\ \hline
\hyperref[SO6:MQ332]{$\MQ_1^{3,3,2}$} & \footnotesize{$1 + 57 t^2 + 88 t^3 + 2432 t^4 + 9024 t^5 + 91715 t^6+\ldots$} & \footnotesize{$32 t^2 + 128 t^3 + 2016 t^4 + 9984 t^5 + 85856 t^6+\ldots$} \\ \hline
\hyperref[SO8:MQ001]{$\MQ_{1;\, \text{SO}(8)}^{0,0,1}$} & \CBHS{1 + 2 t^2 + 2 t^4 + 4 t^6 + 2 t^8 + 2 t^{10} + t^{12}}{(1 - t^2)^2\, (1 - t^6)^2}{1 + 4 t^2 + 9 t^4 + 20 t^6 + 37 t^8 + 60 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ100]{$\MQ_{1;\, \text{SO}(8)}^{1,0,0}$} & \CBHS{1 + 2 t^2 + 2 t^4 + 4 t^6 + 2 t^8 + 2 t^{10} + t^{12}}{(1 - t^2)^2\, (1 - t^6)^2}{1 + 4 t^2 + 9 t^4 + 20 t^6 + 37 t^8 + 60 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ002]{$\MQ_{1;\, \text{SO}(8)}^{0,0,2}$} & \CBHS{P_{11}(t)}{(1 - t^2)^4\, (1 - t^6)^4}{1 + 11 t^2 + 60 t^4 + 235 t^6 + 745 t^8 + 2016 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ200]{$\MQ_{1;\, \text{SO}(8)}^{2,0,0}$} & \CBHS{P_{12}(t)}{(1 - t^2)^2\, (1 - t^4)^2\, (1 - t^6)^4}{1 + 7 t^2 + 36 t^4 + 133 t^6 + 409 t^8 + 1082 t^{10}} & \CBHS{P_{13}(t)}{(1 - t^2)^2\, (1 - t^4)^2\, (1 - t^6)^4}{4 t^2 + 24 t^4 + 102 t^6 + 336 t^8 + 934 t^{10}} \\ \hline
\hyperref[SO8:MQ210]{$\MQ_{1;\, \text{SO}(8)}^{2,1,0}$} & \footnotesize{$1 + 10 t^2 + 62 t^4 + 291 t^6 + 1102 t^8 + 3556 t^{10} + 10104 t^{12} +
25904 t^{14} + 60965 t^{16} + 133590 t^{18} + 275450 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 36 t^4 + 196 t^6 + 824 t^8 + 2840 t^{10} + 8448 t^{12} + 22392 t^{14} + 54040 t^{16} + 120684 t^{18} + 252596 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ201]{$\MQ_{1;\, \text{SO}(8)}^{2,0,1}$} & \footnotesize{$1 + 10 t^2 + 62 t^4 + 291 t^6 + 1102 t^8 + 3556 t^{10} + 10104 t^{12} +
25904 t^{14} + 60965 t^{16} + 133590 t^{18} + 275450 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 36 t^4 + 196 t^6 + 824 t^8 + 2840 t^{10} + 8448 t^{12} + 22392 t^{14} + 54040 t^{16} + 120684 t^{18} + 252596 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ102]{$\MQ_{1;\, \text{SO}(8)}^{1,0,2}$} & \footnotesize{$1 + 14 t^2 + 98 t^4 + 487 t^6 + 1926 t^8 + 6396 t^{10} + 18552 t^{12} +
48296 t^{14} + 115005 t^{16} + 254274 t^{18} + 528046 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ220]{$\MQ_{1;\, \text{SO}(8)}^{2,2,0}$} & \footnotesize{$1 + 13 t^2 + 143 t^4 + 1106 t^6 + 6918 t^8 + 35792 t^{10} +159285 t^{12} + 623177 t^{14} + 2187539 t^{16} + 6992878 t^{18} +
20617582 t^{20}+\ldots$} & \footnotesize{$8 t^2 + 104 t^4 + 936 t^6 + 6200 t^8 + 33384 t^{10} + 151776 t^{12} +
602624 t^{14} + 2134800 t^{16} + 6868552 t^{18} + 20339496 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ220]{$\MQ_{2;\, \text{SO}(8)}^{2,2,0}$} & \footnotesize{$1 + 6 t^2 + 30 t^4 + 110 t^6 + 339 t^8 + 900 t^{10} + 2140 t^{12} +
4644 t^{14} + 9365 t^{16} + 17754 t^{18} + 31962 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 24 t^4 + 100 t^6 + 320 t^8 + 872 t^{10} + 2096 t^{12} + 4584 t^{14} + 9280 t^{16} + 17644 t^{18} + 31816 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ202]{$\MQ_{1;\, \text{SO}(8)}^{2,0,2}$} & \footnotesize{$1 + 17 t^2 + 161 t^4 + 1178 t^6 + 7146 t^8 + 36564 t^{10} + 161447 t^{12} + 628953 t^{14} + 2201473 t^{16} + 7025402 t^{18} +
20688074 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 86 t^4 + 864 t^6 + 5972 t^8 + 32612 t^{10} + 149614 t^{12} +
596848 t^{14} + 2120866 t^{16} + 6836028 t^{18} + 20269004 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ202]{$\MQ_{2;\, \text{SO}(8)}^{2,0,2}$} & \footnotesize{$1 + 6 t^2 + 30 t^4 + 110 t^6 + 339 t^8 + 900 t^{10} + 2140 t^{12} +
4644 t^{14} + 9365 t^{16} + 17754 t^{18} + 31962 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 24 t^4 + 100 t^6 + 320 t^8 + 872 t^{10} + 2096 t^{12} + 4584 t^{14} + 9280 t^{16} + 17644 t^{18} + 31816 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ112]{$\MQ_{1;\, \text{SO}(8)}^{1,1,2}$} & \footnotesize{$1 + 17 t^2 + 16 t^3 + 151 t^4 + 248 t^5 + 1065 t^6 + 2032 t^7 +
6375 t^8 + 12320 t^9 + 32229 t^{10} + 61368 t^{11} + 140928 t^{12} +
261264 t^{13} + 545817 t^{14} + 978968 t^{15} + 1902162 t^{16} +
3301248 t^{17} + 6048732 t^{18} + 10175424 t^{19} + 17765022 t^{20}+\ldots$} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ112]{$\MQ_{2;\, \text{SO}(8)}^{1,1,2}$} & \CBHS{P_{14}(t)}{(1 - t^2)^4\, (1 - t^4)^4}{1 + 10 t^2 + 55 t^4 + 215 t^6 + 679 t^8 + 1831 t^{10}} & \footnotesize{not required} \\ \hline
\hyperref[SO8:MQ211]{$\MQ_{1;\, \text{SO}(8)}^{2,1,1}$} & \footnotesize{$1 + 13 t^2 + 8 t^3 + 99 t^4 + 124 t^5 + 645 t^6 + 1016 t^7 +
3631 t^8 + 6160 t^9 + 17605 t^{10} + 30684 t^{11} + 74988 t^{12} +
130632 t^{13} + 285329 t^{14} + 489484 t^{15} + 982434 t^{16} +
1650624 t^{17} + 3098316 t^{18} + 5087712 t^{19} + 9046590 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 8 t^3 + 52 t^4 + 124 t^5 + 420 t^6 + 1016 t^7 + 2744 t^8 +
6160 t^9 + 14624 t^{10} + 30684 t^{11} + 65940 t^{12} + 130632 t^{13} +
260488 t^{14} + 489484 t^{15} + 919728 t^{16} + 1650624 t^{17} +
2950416 t^{18} + 5087712 t^{19} + 8718432 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ211]{$\MQ_{2;\, \text{SO}(8)}^{2,1,1}$} & \footnotesize{$1 + 6 t^2 + 31 t^4 + 111 t^6 + 351 t^8 + 927 t^{10} + 2222 t^{12} +
4811 t^{14} + 9745 t^{16} + 18463 t^{18} + 33309 t^{20}+\ldots$} & \footnotesize{$4 t^2 + 24 t^4 + 104 t^6 + 328 t^8 + 904 t^{10} + 2168 t^{12} +
4768 t^{14} + 9640 t^{16} + 18376 t^{18} + 33128 t^{20}+\ldots$} \\ \hline
\hyperref[SO8:MQ221]{$\MQ_{1;\, \text{SO}(8)}^{2,2,1}$} & \footnotesize{$1 + 16 t^2 + 16 t^3 + 192 t^4 + 384 t^5 + 2037 t^6 + 5072 t^7 +
19123 t^8 + 49824 t^9 + 156259 t^{10}+\ldots$} & \footnotesize{$8 t^2 + 16 t^3 + 136 t^4 + 384 t^5 + 1688 t^6 + 5072 t^7 +
17400 t^8 + 49824 t^9 + 148624 t^{10}+\ldots$} \\ \hline
\hyperref[SO8:MQ221]{$\MQ_{2;\, \text{SO}(8)}^{2,2,1}$} & \footnotesize{$1 + 15 t^2 + 16 t^3 + 175 t^4 + 368 t^5 + 1840 t^6 + 4768 t^7 +
16982 t^8 + 45408 t^9 + 135469 t^{10}+\ldots$} & \footnotesize{$8 t^2 + 16 t^3 + 136 t^4 + 368 t^5 + 1640 t^6 + 4768 t^7 +
16208 t^8 + 45408 t^9 + 132744 t^10+\ldots$} \\ \hline
\hyperref[SO8:MQ212]{$\MQ_{1;\, \text{SO}(8)}^{2,1,2}$} & \footnotesize{$1 + 20 t^2 + 16 t^3 + 224 t^4 + 384 t^5 + 2189 t^6 + 5072 t^7 +
19815 t^8 + 49824 t^9 + 159015 t^{10}+\ldots$} & \footnotesize{$4 t^2 + 16 t^3 + 104 t^4 + 384 t^5 + 1536 t^6 + 5072 t^7 +
16708 t^8 + 49824 t^9 + 145868 t^{10}+\ldots$} \\ \hline
\hyperref[SO8:MQ212]{$\MQ_{2;\, \text{SO}(8)}^{2,1,2}$} & \footnotesize{$1 + 19 t^2 + 16 t^3 + 211 t^4 + 368 t^5 + 2020 t^6 + 4768 t^7 +
17758 t^8 + 45408 t^9 + 138385 t^{10}+\ldots$} & \footnotesize{$4 t^2 + 16 t^3 + 100 t^4 + 368 t^5 + 1460 t^6 + 4768 t^7 +
15432 t^8 + 45408 t^9 + 129828 t^{10}+\ldots$} \\ \hline
\end{longtable}
\captionsetup{width=15cm}
\begin{longtable}{|c|C{12.25cm}|}
\caption{Unrefined Higgs branch Hilbert series for the orthosymplectic quivers.}
\label{HiggsHSOSp} \\ \hline
Quiver & Higgs branch HS: $\text{HS}_{\mathcal{H}}(t)$ \\ \hline
\hyperref[SO6:MQ110]{$\MQ_1^{1,1,0}$} & \HBHS{(1-t)\,Q_1(t)}{(1-t^2)^4\,(1-t^3)^3}{1 + 9 t^2 + 6 t^3 + 36 t^4 + 36 t^5 + 112 t^6 + 120 t^7 + 285 t^8 \\ +
320 t^9+ 621 t^{10}} \\ \hline
\hyperref[SO6:MQ200]{$\MQ_1^{2,0,0}$} & \HBHS{1 + 9 t^2 + 9 t^4 + t^6}{(1-t^2)^6}{1 + 15 t^2 + 84 t^4 + 300 t^6 + 825 t^8 + 1911 t^{10}} \\ \hline
\hyperref[SO6:MQ200]{$\MQ_2^{2,0,0}$} & \HBHS{1 + t^2}{(1-t^2)^2}{1 + 3 t^2 + 5 t^4 + 7 t^6 + 9 t^8 + 11 t^{10}} \\ \hline
\hyperref[SO6:MQ210]{$\MQ_1^{2,1,0}$} & \HBHS{Q_2(t)}{(1-t^2)^3\,(1-t^4)^3}{1 + 9 t^2 + 42 t^4 + 136 t^6 + 357 t^8 + 801 t^{10}} \\ \hline
\hyperref[SO6:MQ220]{$\MQ_1^{2,2,0}$} & \HBHS{Q_3(t)}{(1-t^2)\,(1-t^3)^2\,(1-t^4)\,(1-t^5)^2}{1 + 4 t^2 + 4 t^3 + 11 t^4 + 16 t^5 + 31 t^6 + 44 t^7\\ + 72 t^8 +
104 t^9 + 155 t^{10}} \\ \hline
\hyperref[SO6:MQ220]{$\MQ_2^{2,2,0}$} & \HBHS{1 + 9 t^2 + 9 t^4 + t^6}{(1-t^2)^6}{1 + 15 t^2 + 84 t^4 + 300 t^6 + 825 t^8 + 1911 t^{10}} \\ \hline
\hyperref[SO6:MQtilde220]{$\widehat{\MQ}_1^{2,2,0}$} & \HBHS{Q_3(t)}{(1-t^2)\,(1-t^3)^2\,(1-t^4)\,(1-t^5)^2}{1 + 4 t^2 + 4 t^3 + 11 t^4 + 16 t^5 + 31 t^6 + 44 t^7\\ + 72 t^8 +
104 t^9 + 155 t^{10}} \\ \hline
\hyperref[SO6:MQtilde220]{$\widehat{\MQ}_2^{2,2,0}$} & \HBHS{1 + 9 t^2 + 9 t^4 + t^6}{(1-t^2)^6}{1 + 15 t^2 + 84 t^4 + 300 t^6 + 825 t^8 + 1911 t^{10}} \\\hline
\hyperref[SO6:MQ320]{$\MQ_1^{3,2,0}$} & \HBHS{Q_4(t)}{(1-t^2)\,(1-t^4)^3\,(1-t^6)^2}{1 + 4 t^2 + 14 t^4 + 37 t^6 + 86 t^8 + 176 t^{10}} \\ \hline
\hyperref[SO6:MQ320]{$\MQ_2^{3,2,0}$} & \HBHS{Q_5(t)}{(1-t^2)^3\,(1-t^6)^3}{1 + 9 t^2 + 36 t^4 + 106 t^6 + 261 t^8 + 561 t^{10}} \\ \hline
\hyperref[SO6:MQ330]{$\MQ_1^{3,3,0}$} & \HBHS{(1-t)Q_6(t)}{(1-t^2)\,(1-t^3)\,(1-t^4)^2\,(1-t^5)\,(1-t^6)\,(1-t^7)}{1 + t^2 + 2 t^3 + 3 t^4 + 4 t^5 + 8 t^6\\ + 10 t^7 + 15 t^8 + 20 t^9 +
30 t^{10}} \\\hline
\hyperref[SO6:MQ330]{$\MQ_2^{3,3,0}$} & \HBHS{Q_7(t)}{(1-t^2)\,(1-t^5)^2\,(1-t^7)^2\,(1-t^8)}{1 + 4 t^2 + 10 t^4 + 4 t^5 + 20 t^6\\ + 16 t^7 + 36 t^8 + 40 t^9 +
67 t^{10}} \\ \hline
\hyperref[SO6:MQ440]{$\MQ_1^{4,4,0}$} & \footnotesize{$1 + t^4 + 2 t^6 + 5 t^8 + 5 t^{10} + 12 t^{12}+\ldots$} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_1^{2,0,1}$} & \HBHS{(1-t)Q_8(t)}{(1-t^2)^2\,(1-t^4)^2\,(1-t^3)^3}{1 + 5 t^2 + 6 t^3 + 18 t^4 + 26 t^5 + 58 t^6 \\+ 78 t^7 + 149 t^8 +
198 t^9 + 325 t^{10}} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_2^{2,0,1}$} & \HBHS{(1+t^2)^2}{(1-t^2)^4}{1 + 6 t^2 + 19 t^4 + 44 t^6 + 85 t^8 + 146 t^{10}} \\ \hline
\hyperref[SO6:MQ201]{$\MQ_3^{2,0,1}$} & \HBHS{1 + 5 t^2 + 5 t^4 + t^6}{(1-t^2)^6}{1 + 11 t^2 + 56 t^4 + 192 t^6 + 517 t^8 + 1183 t^{10}} \\ \hline
\hyperref[SO6:MQ111]{$\MQ_1^{1,1,1}$} & \HBHS{Q_9(t)}{(1-t)\,(1-t^2)\,(1-t^3)^3\,(1-t^4)}{1 + 5 t^2 + 8 t^3 + 16 t^4 + 32 t^5\\ + 58 t^6 + 88 t^7 + 151 t^8 +
224 t^9 + 329 t^{10}} \\ \hline
\hyperref[SO6:MQ111]{$\MQ_2^{1,1,1}$} & \HBHS{1 + 2 t^2 + 2 t^3 + 2 t^4 + t^6}{(1-t^2)^2\,(1-t^3)^2}{1 + 4 t^2 + 4 t^3 + 9 t^4 + 12 t^5\\ + 22 t^6 + 24 t^7 + 41 t^8 +
48 t^9 + 66 t^{10}} \\ \hline
\hyperref[SO6:MQ211]{$\MQ_1^{2,1,1}$} & \HBHS{(1-t)\,Q_{10}(t)}{(1-t^2)^2\,(1-t^4)^2\,(1-t^5)^3}{1 + 5 t^2 + 18 t^4 + 6 t^5 + 46 t^6\\ + 26 t^7 + 101 t^8 + 78 t^9 +
205 t^{10}} \\ \hline
\hyperref[SO6:MQ211]{$\MQ_2^{2,1,1}$} & \HBHS{1 + 2 t^2 + 2 t^4 + t^6}{(1-t^2)^6}{1 + 8 t^2 + 35 t^4 + 111 t^6 + 286 t^8 + 637 t^{10}} \\ \hline
\hyperref[SO6:MQ221]{$\MQ_1^{2,2,1}$} & \HBHS{Q_{11}(t)}{(1-t^4)^2\,(1-t^5)\,(1-t^6)^2\,(1-t^7)}{1 + t^2 + 2 t^3 + 4 t^4 + 4 t^5 + 9 t^6 + 12 t^7 \\+ 20 t^8 + 26 t^9 +
39 t^{10}} \\\hline
\hyperref[SO6:MQ221]{$\MQ_2^{2,2,1}$} & \HBHS{Q_{16}(t)}{(1 - t^2)^2\, (1 - t^5)^2\, (1 - t^6)^2}{1 + 4 t^2 + 10 t^4 + 4 t^5 + 23 t^6 + 16 t^7 + 46 t^8 + 40 t^9 +
88 t^{10}} \\\hline
\hyperref[SO6:MQhat221]{$\widehat{\MQ}_1^{2,2,1}$} & \footnotesize{$1 + t^2 + 4 t^5 + 4 t^5 + 9t^6 + 12t^7 + 20t^8 +\ldots$} \\\hline
\hyperref[SO6:MQhat221]{$\widehat{\MQ}_2^{2,2,1}$} & \HBHS{Q_{16}(t)}{(1 - t^2)^2\, (1 - t^5)^2\, (1 - t^6)^2}{1 + 4 t^2 + 10 t^4 + 4 t^5 + 23 t^6 + 16 t^7 + 46 t^8 + 40 t^9 +
88 t^{10}} \\\hline
\hyperref[SO6:MQ321]{$\MQ_1^{3,2,1}$} & \footnotesize{$1 + t^2 + 2 t^4 + 2 t^5 + 4 t^6 + 4 t^7 + 8 t^8 + 8 t^9 + 13 t^{10} + \ldots$} \\ \hline
\hyperref[SO6:MQ331]{$\MQ_1^{3,3,1}$} & \footnotesize{$1 + t^4 + 2 t^6 + 5 t^8+\ldots$} \\ \hline
\hyperref[SO6:MQ441]{$\MQ_1^{4,4,1}$} & \footnotesize{$1+\mathcal{O}(t^7)$} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_1^{1,1,2}$} & \HBHS{Q_{12}(t)}{(1 - t)\, (1 - t^3)\, (1 - t^4)\, (1 - t^5)^2\, (1 - t^6)}{1 + t^2 + 2 t^3 + 4 t^4 + 6 t^5\\ + 10 t^6 + 14 t^7 + 22 t^8 + 32 t^9 +
46 t^{10}} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_2^{1,1,2}$} & \HBHS{Q_{13}(t)}{(1 - t)\, (1 - t^2)^2\, (1 - t^3)\, (1 - t^5)^2}{1 + 5 t^2 + 2 t^3 + 14 t^4 + 14 t^5\\ + 32 t^6 + 44 t^7 + 73 t^8 +
102 t^9 + 157 t^{10}} \\ \hline
\hyperref[SO6:MQ112]{$\MQ_3^{1,1,2}$} & \footnotesize{$1+3t^2+14t^4+34t^6+89t^8+\cdots$} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_1^{1,1,2}$} & \HBHS{Q_{12}(t)}{(1 - t)\, (1 - t^3)\, (1 - t^4)\, (1 - t^5)^2\, (1 - t^6)}{1 + t^2 + 2 t^3 + 4 t^4 + 6 t^5\\ + 10 t^6 + 14 t^7 + 22 t^8 + 32 t^9 +
46 t^{10}} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_2^{1,1,2}$} & \HBHS{1 - t + 3 t^2 - 2 t^3 + 4 t^4 + 2 t^6 + 4 t^8 - 2 t^9 +
3 t^{10} - t^{11} + t^{12}}{(1 - t)\, (1 - t^2)^2\, (1 - t^3)\, (1 - t^5)^2}{1 + 5 t^2 + 2 t^3 + 14 t^4 + 14 t^5 + 32 t^6 + 44 t^7 + 73 t^8 +
102 t^9 + 157 t^{10}} \\ \hline
\hyperref[SO6:MQtilde112]{$\widehat{\MQ}_3^{1,1,2}$} & \HBHS{1 + 5 t^4 + 5 t^8 + t^{12}}{(1 - t^2)^3\, (1 - t^4)^3}{1 + 3 t^2 + 14 t^4 + 34 t^6 + 89 t^8 + 179 t^{10}} \\ \hline
\hyperref[SO6:MQ212]{$\MQ_1^{2,1,2}$} &\footnotesize{$1+t^2+3t^4+2t^5+4t^6+6t^7+8t^8+\ldots$} \\ \hline
\hyperref[SO6:MQ222]{$\MQ_1^{2,2,2}$} & \footnotesize{$1+t^4+2t^6+5t^8+\ldots$} \\ \hline
\hyperref[SO6:MQtilde222]{$\widehat{\MQ}_1^{2,2,2}$} & \footnotesize{$1+t^4+2t^6+5t^8+\ldots$} \\ \hline
\hyperref[SO6:MQ322]{$\MQ_1^{3,2,2}$} & \footnotesize{$1+t^6+\ldots$} \\ \hline
\hyperref[SO6:MQ332]{$\MQ_1^{3,3,2}$} & \footnotesize{$1 + 2 t^2 + 8 t^4+\ldots$} \\ \hline
\hyperref[SO8:MQ001]{$\MQ_{1;\, \text{SO}(8)}^{0,0,1}$} & \HBHS{(1-t)\,Q_{13}(t)}{(1 - t^2)^4\, (1 - t^3)^4\, (1 - t^5)}{1 + 10 t^2 + 18 t^3 + 52 t^4 + 116 t^5 + 250 t^6 + 454 t^7 +
889 t^8 + 1490 t^9 + 2538 t^{10}} \\ \hline
\hyperref[SO8:MQ100]{$\MQ_{1;\, \text{SO}(8)}^{1,0,0}$} & \HBHS{(1-t)\,Q_{13}(t)}{(1 - t^2)^4\, (1 - t^3)^4\, (1 - t^5)}{1 + 10 t^2 + 18 t^3 + 52 t^4 + 116 t^5 + 250 t^6 + 454 t^7 +
889 t^8 + 1490 t^9 + 2538 t^{10}} \\ \hline
\hyperref[SO8:MQ002]{$\MQ_{1;\, \text{SO}(8)}^{0,0,2}$} & \HBHS{1 + 6 t^2 + 27 t^4 + 48 t^6 + 84 t^8 + 86 t^{10} + 84 t^{12} + 48 t^{14} +
27 t^{16} + 6 t^{18} + t^{20}}{(1 - t^2)^3\, (1 - t^4)^4\, (1 - t^6)}{1 + 9 t^2 + 55 t^4 + 212 t^6 + 688 t^8 + 1852 t^{10}} \\ \hline
\hyperref[SO8:MQ200]{$\MQ_{1;\, \text{SO}(8)}^{2,0,0}$} & \HBHS{1 + 6 t^2 + 27 t^4 + 48 t^6 + 84 t^8 + 86 t^{10} + 84 t^{12} + 48 t^{14} +
27 t^{16} + 6 t^{18} + t^{20}}{(1 - t^2)^3\, (1 - t^4)^4\, (1 - t^6)}{1 + 9 t^2 + 55 t^4 + 212 t^6 + 688 t^8 + 1852 t^{10}} \\ \hline
\hyperref[SO8:MQ210]{$\MQ_{1;\, \text{SO}(8)}^{2,1,0}$} & \footnotesize{$1 + 5 t^2 + 4 t^3 + 25 t^4 + 22 t^5 + 86 t^6 + 86 t^7 + 254 t^8 +
270 t^9 + 648 t^{10} + 716 t^{11} + 1499 t^{12} + 1686 t^{13} + 3177 t^{14}+\ldots$} \\ \hline
\hyperref[SO8:MQ201]{$\MQ_{1;\, \text{SO}(8)}^{2,0,1}$} & \HBHS{(1-t)\,Q_{14}(t)}{(1 - t^2)^3\, (1 - t^3)^2\, (1 - t^4)^2\, (1 - t^5)\, (1 - t^6)}{1 + 5 t^2 + 4 t^3 + 25 t^4 + 22 t^5 + 86 t^6 + 86 t^7 + 254 t^8 +
270 t^9 + 648 t^{10}} \\ \hline
\hyperref[SO8:MQ102]{$\MQ_{1;\, \text{SO}(8)}^{1,0,2}$} & \HBHS{(1-t)\,Q_{14}(t)}{(1 - t^2)^3\, (1 - t^3)^2\, (1 - t^4)^2\, (1 - t^5)\, (1 - t^6)}{1 + 5 t^2 + 4 t^3 + 25 t^4 + 22 t^5 + 86 t^6 + 86 t^7 + 254 t^8 +
270 t^9 + 648 t^{10}} \\ \hline
\hyperref[SO8:MQ220]{$\MQ_{1;\, \text{SO}(8)}^{2,2,0}$} & \footnotesize{$1 + 4 t^2 + 11 t^4 + 8 t^5 + 31 t^6 + 32 t^7 + 70 t^8 + 88 t^9 +
170 t^{10}+\ldots$} \\ \hline
\hyperref[SO8:MQ220]{$\MQ_{2;\, \text{SO}(8)}^{2,2,0}$} & \HBHS{1 + 2 t^2 + 2 t^4 + 2 t^6 + t^8}{(1 - t^2)^8}{1 + 10 t^2 + 54 t^4 + 210 t^6 + 659 t^8 + 1772 t^{10}} \\ \hline
\hyperref[SO8:MQ202]{$\MQ_{1;\, \text{SO}(8)}^{2,0,2}$} & \footnotesize{$1 + 4 t^2 + 11 t^4 + 8 t^5 + 31 t^6 + 32 t^7 + 70 t^8 + 88 t^9 +
170 t^{10}+\ldots$} \\ \hline
\hyperref[SO8:MQ202]{$\MQ_{2;\, \text{SO}(8)}^{2,0,2}$} & \HBHS{1 + 2 t^2 + 2 t^4 + 2 t^6 + t^8}{(1 - t^2)^8}{1 + 10 t^2 + 54 t^4 + 210 t^6 + 659 t^8 + 1772 t^{10}} \\ \hline
\hyperref[SO8:MQ112]{$\MQ_{1;\, \text{SO}(8)}^{1,1,2}$} & \footnotesize{$1+4t^2+10t^4+8t^5+28t^6+28t^7+63t^8+\cdots$} \\ \hline
\hyperref[SO8:MQ112]{$\MQ_{2;\, \text{SO}(8)}^{1,1,2}$} & \HBHS{1 + 6 t^2 + 17 t^4 + 27 t^6 + 32 t^8 + 27 t^{10} + 17 t^{12} +
6 t^{14} + t^{16}}{(1 - t^2)^4\, (1 - t^4)^4}{1 + 10 t^2 + 55 t^4 + 215 t^6 + 679 t^8 + 1831 t^{10}} \\ \hline
\hyperref[SO8:MQ211]{$\MQ_{1;\, \text{SO}(8)}^{2,1,1}$} & \HBHS{(1 - t)\, Q_{15}(t)}{(1 - t^3)^2\, (1 - t^4)\, (1 - t^5)^3\, (1 - t^6)^2\, (1 - t^8)}{1 + 4 t^2 + 10 t^4 + 8 t^5 + 28 t^6 + 28 t^7 + 63 t^8 + 76 t^9 + 148 t^{10}} \\ \hline
\hyperref[SO8:MQ211]{$\MQ_{2;\, \text{SO}(8)}^{2,1,1}$} & \HBHS{1 + 6 t^2 + 17 t^4 + 27 t^6 + 32 t^8 + 27 t^{10} + 17 t^{12} +
6 t^{14} + t^{16}}{(1 - t^2)^4\, (1 - t^4)^4}{1 + 10 t^2 + 55 t^4 + 215 t^6 + 679 t^8 + 1831 t^{10}} \\ \hline
\hyperref[SO8:MQ221]{$\MQ_{1;\, \text{SO}(8)}^{2,2,1}$} & \footnotesize{$1+t^2+2t^3+2t^4+2t^5+6t^6+8t^7+13t^8+\ldots$} \\ \hline
\hyperref[SO8:MQ221]{$\MQ_{2;\, \text{SO}(8)}^{2,2,1}$} & \footnotesize{$1+3t^2+8t^4+2t^5+16t^6+10t^7+32t^8+\ldots$} \\ \hline
\hyperref[SO8:MQ212]{$\MQ_{1;\, \text{SO}(8)}^{2,1,2}$} & \footnotesize{$1+t^2+2t^3+2t^4+2t^5+6t^6+8t^7+13t^8+\ldots$} \\ \hline
\hyperref[SO8:MQ212]{$\MQ_{2;\, \text{SO}(8)}^{2,1,2}$} & \footnotesize{$1+3t^2+8t^4+2t^5+16t^6+10t^7+32t^8+\ldots$} \\ \hline
\end{longtable}
\captionsetup{width=15cm}
\begin{longtable}{|c|C{14cm}|}
\caption{Palindromic polynomials appearing in the main sections.}
\label{PalPol} \\ \hline
Label & Palindromic Polynomial \\ \hline
$P_1(t)$ & \footnotesize{$1 + 2 t^2 + 10 t^4 + 10 t^6 + 16 t^8 + 10 t^{10} + 10 t^{12} +
2 t^{14} + t^{16}$} \\ \hline
$P_2(t)$ & \footnotesize{$1 + 3 t^2 + 38 t^4 + 69 t^6 + 225 t^8 + 240 t^{10} + 372 t^{12} +
240 t^{14} + 225 t^{16} + 69 t^{18} + 38 t^{20} + 3 t^{22} + t^{24}$} \\ \hline
$P_3(t)$ & \footnotesize{$8 t^2 + 20 t^4 + 112 t^6 + 156 t^8 + 328 t^{10} + 276 t^{12} + 328 t^{14} + 156 t^{16} + 112 t^{18} + 20 t^{20} + 8 t^{22}$} \\ \hline
$P_4(t)$ & \footnotesize{$1 + 11 t^2 + 57 t^4 + 170 t^6 + 324 t^8 + 398 t^{10} + 324 t^{12} +
170 t^{14} + 57 t^{16} + 11 t^{18} + t^{20}$} \\ \hline
$P_5(t)$ & \footnotesize{$1 + 12 t^2 + 63 t^4 + 204 t^6 + 550 t^8 + 1094 t^{10} + 1906 t^{12} +
2708 t^{14} + 3432 t^{16} + 3596 t^{18} + 3432 t^{20} + 2708 t^{22} +
1906 t^{24} + 1094 t^{26} + 550 t^{28} + 204 t^{30} + 63 t^{32} +
12 t^{34} + t^{36}$} \\ \hline
$P_6(t)$ & \footnotesize{$8 t^2 + 56 t^4 + 216 t^6 + 536 t^8 + 1136 t^{10} + 1888 t^{12} +
2752 t^{14} + 3344 t^{16} + 3664 t^{18} + 3344 t^{20} + 2752 t^{22} +
1888 t^{24} + 1136 t^{26} + 536 t^{28} + 216 t^{30} + 56 t^{32} + 8 t^{34}$} \\ \hline
$P_7(t)$ & \footnotesize{$1 + 3 t^2 + 8 t^4 + 21 t^6 + 33 t^8 + 34 t^{10} + 33 t^{12} + 21 t^{14} +
8 t^{16} + 3 t^{18} + t^{20}$} \\ \hline
$P_8(t)$ & \footnotesize{$3 + 10 t^2 + 20 t^4 + 31 t^6 + 38 t^8 + 31 t^{10} + 20 t^{12} + 10 t^{14} +
3 t^{16}$} \\ \hline
$P_9(t)$ & \footnotesize{$1 + 2 t^2 + 7 t^4 + 12 t^6 + 22 t^8 + 16 t^{10} + 22 t^{12} + 12 t^{14} +
7 t^{16} + 2 t^{18} + t^{20}$} \\ \hline
$P_{10}(t)$ & \footnotesize{$2 t^2 + 4 t^4 + 2 t^6 + 10 t^8 + 2 t^{10} + 4 t^{12} + 2 t^{14}$} \\ \hline
$P_{11}(t)$ & \footnotesize{$1 + 7 t^2 + 22 t^4 + 53 t^6 + 94 t^8 + 129 t^{10} + 148 t^{12} + 129 t^{14} + 94 t^{16} + 53 t^{18} + 22 t^{20} + 7 t^{22} + t^{24}$} \\ \hline
$P_{12}(t)$ & \footnotesize{$1 + 5 t^2 + 21 t^4 + 54 t^6 + 114 t^8 + 182 t^{10} + 248 t^{12} + 270 t^{14} + 248 t^{16} + 182 t^{18} + 114 t^{20} + 54 t^{22} + 21 t^{24} + 5 t^{26} + t^{28}$} \\ \hline
$P_{13}(t)$ & \footnotesize{$4 t^2 + 16 t^4 + 50 t^6 + 108 t^8 + 188 t^{10} + 252 t^{12} + 284 t^{14} + 252 t^{16} + 188 t^{18} + 108 t^{20} + 50 t^{22} + 16 t^{24} + 4 t^{26}$} \\ \hline
$P_{14}(t)$ & \footnotesize{$1 + 6 t^2 + 17 t^4 + 27 t^6 + 32 t^8 + 27 t^{10} + 17 t^{12} +
6 t^{14} + t^{16}$} \\ \hline
$Q_1(t)$ & \footnotesize{$1 + t + 6 t^2 + 9 t^3 + 15 t^4 + 12 t^5 + 15 t^6 + 9 t^7 +
6 t^8 + t^9 + t^{10}$} \\ \hline
$Q_2(t)$ & \footnotesize{$1 + 6 t^2 + 15 t^4 + 18 t^6 + 15 t^8 + 6 t^{10} + t^{12}$} \\ \hline
$Q_3(t)$ & \footnotesize{$1 + 3 t^2 + 2 t^3 + 6 t^4 + 4 t^5 + 10 t^6 + 6 t^7 + 9 t^8 + 6 t^9 +
10 t^{10} + 4 t^{11} + 6 t^{12} + 2 t^{13} + 3 t^{14} + t^{16}$} \\ \hline
$Q_4(t)$ & \footnotesize{$1 + 3 t^2 + 7 t^4 + 12 t^6 + 16 t^8 + 16 t^{10} + 16 t^{12} + 12 t^{14} + 7 t^{16} + 3 t^{18} + t^{20}$} \\ \hline
$Q_5(t)$ & \footnotesize{$1 + 6 t^2 + 12 t^4 + 21 t^6 + 24 t^8 + 24 t^{10} + 21 t^{12} + 12 t^{14} +
6 t^{16} + t^{18}$} \\ \hline
$Q_6(t)$ & \footnotesize{$1 + t + t^2 + 2 t^3 + 2 t^4 + 3 t^5 + 5 t^6 + 6 t^7 + 7 t^8 + 7 t^9 + 7 t^{10} + 7 t^{11} + 8 t^{12} + 7 t^{13} + 7 t^{14} + 7 t^{15} + 7 t^{16} + 6 t^{17} + 5 t^{18} + 3 t^{19} + 2 t^{20} + 2 t^{21} + t^{22} + t^{23} + t^{24}$} \\ \hline
$Q_7(t)$ & \footnotesize{$1 + 3 t^2 + 6 t^4 + 2 t^5 + 10 t^6 + 4 t^7 + 15 t^8 + 6 t^9 +
21 t^{10} + 8 t^{11} + 22 t^{12} + 10 t^{13} + 25 t^{14} + 10 t^{15} + 22 t^{16} +
8 t^{17} + 21 t^{18} + 6 t^{19} + 15 t^{20} + 4 t^{21} + 10 t^{22} + 2 t^{23} +
6 t^{24} + 3 t^{26} + t^{28}$} \\ \hline
$Q_8(t)$ & \footnotesize{$1 + t + 4 t^2 + 7 t^3 + 14 t^4 + 19 t^5 + 25 t^6 + 24 t^7 + 25 t^8 +
19 t^9 + 14 t^{10} + 7 t^{11} + 4 t^{12} + t^{13} + t^{14}$} \\ \hline
$Q_9(t)$ & \footnotesize{$1 - t + 4 t^2 + t^3 + 5 t^4 + 2 t^5 + 5 t^6 + t^7 + 4 t^8 - t^9 + t^{10}$} \\ \hline
$Q_{10}(t)$ & \footnotesize{$1 + t + 4 t^2 + 4 t^3 + 11 t^4 + 14 t^5 + 23 t^6 + 28 t^7 + 38 t^8 +
37 t^9 + 44 t^{10} + 37 t^{11} + 38 t^{12} + 28 t^{13} + 23 t^{14} + 14 t^{15} +
11 t^{16} + 4 t^{17} + 4 t^{18} + t^{19} + t^{20}$} \\ \hline
$Q_{11}(t)$ & \footnotesize{$1 + t^2 + 2 t^3 + 2 t^4 + 3 t^5 + 5 t^6 + 6 t^7 + 9 t^8 + 11 t^9 +
12 t^{10} + 13 t^{11} + 16 t^{12} + 12 t^{13} + 16 t^{14} + 13 t^{15} +
12 t^{16} + 11 t^{17} + 9 t^{18} + 6 t^{19} + 5 t^{20} + 3 t^{21} + 2 t^{22} +
2 t^{23} + t^{24} + t^{26}$} \\ \hline
$Q_{12}(t)$ & \footnotesize{$1 - t + t^2 + 2 t^4 + 3 t^6 + t^7 + 2 t^8 + t^9 + 2 t^{10} + t^{11} +
3 t^{12} + 2 t^{14} + t^{16} - t^{17} + t^{18}$} \\ \hline
$Q_{13}(t)$ & \footnotesize{$1 + t + 7 t^2 + 21 t^3 + 39 t^4 + 58 t^5 + 90 t^6 + 110 t^7 +
118 t^8 + 110 t^9 + 90 t^{10} + 58 t^{11} + 39 t^{12} + 21 t^{13} +
7 t^{14} + t^{15} + t^{16}$} \\ \hline
$Q_{14}(t)$ & \footnotesize{$1 + t + 3 t^2 + 5 t^3 + 16 t^4 + 21 t^5 + 34 t^6 + 34 t^7 + 53 t^8 +
56 t^9 + 73 t^{10} + 64 t^{11} + 73 t^{12} + 56 t^{13} + 53 t^{14} + 34 t^{15} +
34 t^{16} + 21 t^{17} + 16 t^{18} + 5 t^{19} + 3 t^{20} + t^{21} + t^{22}$} \\ \hline
$Q_{15}(t)$ & \footnotesize{$1 + t + 5 t^2 + 3 t^3 + 12 t^4 + 9 t^5 + 32 t^6 + 30 t^7 + 68 t^8 +
65 t^9 + 119 t^{10} + 114 t^{11} + 187 t^{12} + 181 t^{13} + 268 t^{14} +
248 t^{15} + 330 t^{16} + 287 t^{17} + 354 t^{18} + 287 t^{19} + 330 t^{20} +
248 t^{21} + 268 t^{22} + 181 t^{23} + 187 t^{24} + 114 t^{25} + 119 t^{26} +
65 t^{27} + 68 t^{28} + 30 t^{29} + 32 t^{30} + 9 t^{31} + 12 t^{32} + 3 t^{33} +
5 t^{34} + t^{35} + t^{36}$} \\ \hline
$Q_{16}(t)$ & \footnotesize{$1 + 2 t^2 + 3 t^4 + 2 t^5 + 5 t^6 + 4 t^7 + 6 t^8 + 6 t^9 + 6 t^{10} +
6 t^{11} + 6 t^{12} + 4 t^{13} + 5 t^{14} + 2 t^{15} + 3 t^{16} + 2 t^{18} + t^{20}$} \\ \hline
\end{longtable}
\section{Conclusion}\label{conclusions}
In this paper, we studied the Higgs branches of 5d SCFTs, that admit deformations to SO(6) or SO(8) gauge theories with matter in either the vector, spinor, or conjugate spinor representation of the gauge group. We used the brane configurations engineering these theories, to find the corresponding magnetic quivers. The magnetic quivers for SO(6) theories were verified by comparison with those of SU(4) theories, while we used SO(8) triality, to provide consistency checks of the magnetic quivers proposed for the SO(8) theories. The agreement of the different magnetic quivers for a given SCFT, were shown to hold at the level of the unrefined Hilbert series, for all cases, on both the Coulomb and the Higgs branch of the moduli space. The unrefined Hilbert series for both the Coulomb and the Higgs branch Hilbert series of all magnetic quivers studied in this work are collected in appendix \ref{appendixA}. Some magnetic quivers encountered in this work, involved bad symplectic gauge nodes. We were able to assign a Hilbert series for both Coulomb and Higgs branches of these theories, developing a new technique which is referred to as B2G in the main text, which uses a local mirror description of the effective theory around the most singular locus of the moduli space. Details of these computational techniques will be the content of a future publication \cite{HSbad}. Since the Coulomb branch of a bad 3d $\mathcal{N}=4$ USp($2N$) theory is comprised of multiple singular loci, one may wonder whether a 5d Higgs branch whose magnetic quiver contains bad USp($2N$) nodes, has a similar structure on its Higgs branch. However the output of our study seems to suggest otherwise. In particular, we found that for the theories under our construction, the local geometry near one of the two singular points on the Coulomb branch of the magnetic quiver, can capture the full Higgs branch of the 5d theory. This conclusion follows because for every magnetic quiver with a bad node, we have a dual magnetic quiver without any bad nodes. It would be interesting to find further checks of this claim.
We encountered brane webs whose magnetic quivers involve second rank symmetric and second rank antisymmetric tensors. We provided a formula for computing the number of hypermultiplets in the (anti)symmetric representations, which generalises previous rules for extracting magnetic quivers from brane webs with O5-planes in \cite{Bourget:2020gzi,Akhond:2020vhc}. Despite these improvements, there is still no completely systematic algorithm to extract magnetic quivers for 5d SCFTs engineered using O5-planes. Perhaps the most urgent question to address, is the magnetic quivers for SO($2r+1$) theories, which are engineered using $\widetilde{\text{O5}}$-planes. One might be able to use the results of \cite{Akhond:2021ffo} as a guide for such a study. Another subtle issue, that will eventually need to be addressed, is developing a more systematic method for distinguishing different theories whose 5-brane web at the SCFT limit is identical.
| c8f3a106df9a03d7529b66169421ec161c38e5b1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Acknowledgments}
This material is based upon work supported by the U.S.~Department of Energy, Office of Science, Office of Nuclear Physics under contract / award numbers DE-AC02-05CH11231, DE-AC05-00OR22725, DE-AC05-76RL0130, DE-FG02-97ER41020, DE-FG02-97ER41033, DE-FG02-97ER41041, DE-SC0012612, DE-SC0014445, DE-SC0018060, and LANLEM77/LANLEM78. We acknowledge support from the Particle Astrophysics Program and Nuclear Physics Program of the National Science Foundation through grant numbers MRI-0923142, PHY-1003399, PHY-1102292, PHY-1206314, PHY-1614611, PHY-1812409, PHY-1812356, and PHY-2111140. We gratefully acknowledge the support of the Laboratory Directed Research \& Development (LDRD) program at Lawrence Berkeley National Laboratory for this work. We gratefully acknowledge the support of the U.S.~Department of Energy through the Los Alamos National Laboratory LDRD Program and through the Pacific Northwest National Laboratory LDRD Program for this work. We gratefully acknowledge the support of the South Dakota Board of Regents Competitive Research Grant. We acknowledge support from the Russian Foundation for Basic Research, grant No.~15-02-02919. We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada, funding reference number SAPIN-2017-00023, and from the Canada Foundation for Innovation John R.~Evans Leaders Fund. This research used resources provided by the Oak Ridge Leadership Computing Facility at Oak Ridge National Laboratory and by the National Energy Research Scientific Computing Center, a U.S.~Department of Energy Office of Science User Facility. We thank our hosts and colleagues at the Sanford Underground Research Facility for their support.
| 5cf7e5bda1b953d74f4da101b1607cb3cea9adb0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\noindent Semileptonic rare charm decays are strongly Glashow-Iliopoulos-Maiani (GIM) suppressed in the Standard Model (SM) due to light down-type quarks running in one-loop diagrams~\cite{Burdman:2001tf}. Therefore, Beyond Standard Model (BSM) effects can be probed in $c\to u\ell^+\ell^-$ modes. However, several impediments need to be overcome in BSM searches in rare charm decays. These include the lack of sufficient theoretical control on decay amplitudes due to strong pollution from intermediate QCD resonances and poor convergence of the heavy quark expansion at the charm mass scale~\cite{Feldmann:2017izn}. Null test observables, already part of the ongoing precision programs of $b\to s\ell^+\ell^-$ transitions, vanish in the SM and hence any signal indicates BSM physics, which makes null tests inevitable in New Physics (NP) searches in $\vert \Delta c \vert=\vert \Delta u \vert=1$ processes.
In particular, the spin structure in rare charm baryon decays offers a rich angular distribution, including several clean null tests~\cite{Golz:2021imq, Golz:2022alh}.
Recent experimental and theoretical progress of rare charm decays is compiled in Ref.~\cite{Gisbert:2020vjx}. Research interest in the field is rapidly increasing, see \textit{e.g.}~\cite{Burdman:2001tf, Gisbert:2020vjx, deBoer:2015boa, Fajfer:2015mia, Feldmann:2017izn, Bause:2019vpr, Bharucha:2020eup, Golz:2021imq, Golz:2022alh}.
The plan of this paper is as follows. The effective field theory framework and SM phenomenology in rare charm baryon decays is presented in Sec.~\ref{sec:eft_pheno}. Sec.~\ref{sec:angular} discusses the NP sensitivity from null tests in angular observables of $\Lambda_c\to p \mu^+\mu^-$. We discuss further opportunities in rare charm four-body decay modes based on recent results in Ref.~\cite{Golz:2022alh} and conclude in Sec.~\ref{sec:concl}.
\section{Standard Model phenomenology of rare charm baryon decays}\label{sec:eft_pheno}
\noindent The effective Hamiltonian at the charm mass scale $\mu_c$ inducing $c\to u\mu^+\mu^-$ processes reads
\begin{equation}
\mathcal{H}_{\rm eff} \supset -\frac{4G_F}{\sqrt2} \frac{\alpha_e}{4\pi} \sum_{k=7,9,10} \bigl( C_kO_k + C_k^\prime O_k^\prime \bigr)\,,
\label{eq:Heff}
\end{equation}
where the dimension 6 operators relevant for this work are defined as follows:
\begin{equation}
\begin{split}
O_7 &= \frac{m_c}{e} (\overline{u}_L \sigma_{\alpha\beta} c_R) F^{\alpha\beta} \,,\\
O_9 &= (\overline{u}_L \gamma_\mu c_L) (\overline{\mu} \gamma^\mu \mu) \,,\\
O_{10} &= (\overline{u}_L \gamma_\mu c_L) (\overline{\mu} \gamma^\mu \gamma_5 \mu) \,, \\
\end{split}
\quad\quad
\begin{split}
O^\prime_7 &= \frac{m_c}{e} (\overline{u}_R \sigma_{\alpha\beta} c_L) F^{\alpha\beta}\,, \\
O^\prime_9 &= (\overline{u}_R \gamma_\mu c_R) (\overline{\mu} \gamma^\mu \mu) \,, \\
O^\prime_{10} &= (\overline{u}_R \gamma_\mu c_R) (\overline{\mu} \gamma^\mu \gamma_5 \mu) \,, \\
\end{split}
\label{eq:operators}
\end{equation}
with $\sigma_{\alpha\beta}=\frac{\text{i}}{2}\left[\gamma_\alpha,\,\gamma_\beta\right]$ and the electromagnetic field strength tensor $F^{\alpha\beta}$.
Short distance contributions to operators in Eq.~\eqref{eq:operators} in the SM are calculated consistently up to (partly) NNLLO in RG improved perturbation theory in Ref.~\cite{deBoer:thesis}. All contributions originate solely from four-quark operators at the $W$ mass scale and are contained in effective $q^2$ dependent coefficients $C_7^{\text{eff}}(q^2),\,C_9^{\text{eff}}(q^2)$, where $q^2$ is the dimuon invariant mass squared and $C_7^{\text{eff}}\lesssim 10^{-3}$ and $C_9^{\text{eff}}(q^2>0.1\,\text{GeV}^2)\lesssim 10^{-2}$, negligible for most phenomenological applications, see results in Ref.~\cite{deBoer:thesis} and discussions in~\cite{Golz:2021imq}. SM contributions to all primed operators and $C_{10}$ vanish.
Further contributions to $\vert \Delta c\vert=\vert\Delta u \vert = 1 $ come from intermediate resonances, which we parametrize in terms of a sum of Breit-Wigner distributions fit to data. These resonance effects are then added as a contribution to $O_9$ in $C_9^R(q^2)$ and are the main source of uncertainty due to unknown strong phases entering the parametrization. Dominant vector resonances include $M=\rho(770)$, $\omega(782)$ and $\phi(1020)$ and parameters are fixed via $\mathcal{B}_{C_9^R}(\Lambda_{c} \to p\mu^{+}\mu^{-}) = \mathcal{B}(\Lambda_{c} \to pM)\mathcal{B}(M \to \mu^{+}\mu^{-})\,$. Contributions from pseudoscalar mesons $\eta,\,\eta^\prime$ are found to be negligible on the level of branching ratios in rare charm baryon decays~\cite{Golz:2021imq}.
We demonstrate the dominance of long-range resonances over non-resonant short-distance SM contributions in Fig.~\ref{fig:br}, where the $q^2$ differential branching ratio for $\Lambda_c \to p \mu^+\mu^-$ is shown in orange and blue, respectively.
The main sources of uncertainty are from form factors~\cite{Meinel:2017ggx} and due to varying strong phases independently from $-\pi$ to $\pi$ for resonant contributions and due to varying the charm mass scale $\mu_c$ for short-distance SM effects.
\begin{figure}[!t]\centering
\includegraphics[width=0.45\textwidth]{BR_SM_eta}
\caption{The resonant (orange) and non-resonant (blue) $q^2$-differential branching ratio in the SM for the rare charm baryon decay $\Lambda_c\to p \mu^+\mu^-$. The bandwidths indicate uncertainties from form factors and strong phase variation (charm mass scale $\mu_c$ variation) in the (non-)resonant case. Figure taken from Ref.~\cite{Golz:2021imq}.}
\label{fig:br}
\end{figure}
\section{Null test observables in angular distributions and beyond}\label{sec:angular}
\noindent The angular distribution of three-body rare charm baryon decays can be described by the dilepton invariant mass $q^2$ and a single angle $\theta_\ell$ associated to the angle between the $\ell^+$ momentum and the negative direction of flight of the $\Lambda_c$ in the dilepton rest frame. Then, contributions are split into angular coefficients as
\begin{equation}
\frac{\text{d}^2\Gamma}{\text{d}q^2\text{d}\cos\theta_\ell}=\frac{3}{2}\,(K_{1ss}\,\sin^2\theta_\ell\,+\,K_{1cc}\,\cos^2\theta_\ell\,+\,K_{1c}\,\cos\theta_\ell)\,
\label{eq:angl_distr}
\end{equation}
where the dependence on Wilson coefficients and form factors of $K_{1ss},\,K_{1cc}$ and $K_{1c}$ is given in Ref.~\cite{Golz:2021imq}. Next to the differential branching ratio, two angular observables can be defined, the fraction of longitudinally polarized dimuons $F_L$ and the forward-backward asymmetry in the leptons $A_{\text{FB}}$, defined as
\begin{align}\label{eq:angulars}
F_L=\frac{2\,K_{1ss}-K_{1cc}}{2\,K_{1ss}+K_{1cc}}\,, \quad A_{\text{FB}}=\frac{3}{2}\,\frac{K_{1c}}{2\,K_{1ss}+K_{1cc}}\,,
\end{align}
where $\frac{\text{d}\Gamma}{\text{d}q^2}=2\,K_{1ss}+K_{1cc}$, and one can also define $\tilde{A}_{\text{FB}}=\frac{3}{2}\frac{K_{1c}}{\Gamma}$ with $\Gamma=\int_{q_{\text{min}}^2}^{q_{\text{max}}^2}(2\,K_{1ss}+K_{1cc})\text{d}q^2$.
We show $F_L$ (top row), $\tilde{A}_{\text{FB}}$ (bottom left) and ${A}_{\text{FB}}$ (bottom right) in the SM and in several BSM benchmark scenarios in Fig.~\ref{fig:ang}.
\begin{figure}[!t]\centering
\includegraphics[width=0.42\textwidth]{FL_NP}
\includegraphics[width=0.42\textwidth]{F_L_sensitivityc7prime}
\includegraphics[width=0.42\textwidth]{AFB_TILDE}
\includegraphics[width=0.42\textwidth]{AFB}
\caption{Angular observables in $\Lambda_c\to p\mu^+\mu^-$. The upper row shows the fraction of longitudinally polarized dimuons $F_L$ in various NP scenarios (left) and the sensitivity for decreasing values of the dipole operator $C_7^\prime$. The lower row shows the forward-backward asymmetry $\tilde{A}_{\text{FB}}$ (left) and ${A}_{\text{FB}}$ (right) in NP scenarios with decreasing $C_{10}$ Wilson coefficient, see text for details. Upper left and lower row plots are taken from Ref.~\cite{Golz:2021imq}.}
\label{fig:ang}
\end{figure}
For $F_L$, we learn that uncertainties associated with resonances drop out in the SM and $F_L=\frac{1}{3}$ model-independently at both kinematic endpoints~\cite{Hiller:2021zth,Golz:2021imq}. The $q^2$-shape is significantly altered from the one in the SM in scenarios involving dipole couplings $C_7^{(\prime)}$. There is sensitivity to values as low as $\sim 0.02$, as illustrated in the upper right plot. $A_{\text{FB}}$ is a SM null test, which is why no orange SM curve is shown (it is exactly zero). Depending on the normalization, the axial vector coupling $C_{10}$ can also be probed down to $\sim 0.01$. In $\tilde{A}_{\text{FB}}$, the normalization is a constant in $q^2$, and the signal is resonance enhanced, due to large values of $C_9^R$ around the resonance masses interfering with the BSM coupling $C_{10}$. This resonance enhancement can not be seen in $A_{\text{FB}}$ with differential normalization as the peaking contributions in the differential branching ratio around the resonances suppress the signal in the numerator.
\section{Opportunities with self-analyzing four-body decays}\label{sec:fourbody}
\noindent Recently, further null test opportunities were presented for (quasi-) four-body decay chains of charmed baryons, where the final state baryon further decays weakly with sizable polarization parameter $\alpha$~\cite{Golz:2022alh}. These \textit{self analyzing} four-body final states offer a rich angular distribution with complementary sensitivities to NP couplings. We refer to Ref.~\cite{Golz:2022alh} for further details, but exemplary discuss for the decay chain $\Xi_c^+\to \Sigma^+(\to p\pi^0) \mu^+\mu^-$ the forward-backward asymmetry of the final state hadron system $A_{\text{FB}}^{\text{H}}$, which can be defined in close analogy to $A_{\text{FB}}$ in Eq.~\eqref{eq:angulars}. $A_{\text{FB}}^{\text{H}}$ is not a null test, but very sensitive to right-handed quark currents $C_{7,\,9,\,10}^\prime$. This can be seen in Fig.~\ref{fig:ang2}, where the upper plot shows the SM in orange and contributions for sizable right-handed currents in red and green, compared to small deviations induced by NP in $C_7$ in blue, whereas NP benchmarks with $C_9$ or $C_{10}$ are not shown, as they are SM--like. The lower plots again illustrate sensitivities, the plot to the left for $C_7^\prime$ and the plot to the right for $C_{10}^\prime$, but a plot for $C_9^\prime$ would be equal to the latter within uncertainties. As evident from the lower row plots, $C_{7}^\prime$ can be probed down to $\sim 0.01$, whereas the sensitivity in $C_{9,\,(10)}^\prime$ is $\mathcal{O}(0.1)$.
\begin{figure}[!t]\centering
\includegraphics[width=0.41\textwidth]{afbH}\\
\includegraphics[width=0.41\textwidth]{afbH_sensc7prime}
\includegraphics[width=0.41\textwidth]{afbH_sensc10prime}
\caption{$A_{\text{FB}}^{\text{H}}$ for $\Xi_c^+\to \Sigma^+(\to p\pi^0) \mu^+\mu^-$. The upper plot is taken from Ref.~\cite{Golz:2022alh} and shows several BSM scenarios (see legend) and the SM. The lower row plots further illustrate the sensitivity to $C_7^\prime$ and $C_{9,\,(10)}^\prime$ via decreasing benchmark values. The SM, shown in orange, includes form factor uncertainties, whereas uncertainties from resonance parameters cancel. In NP scenarios, the dominant uncertainties are due to the strong phases associated to $C_9^R$ and entering via interference with the BSM Wilson coefficients.}
\label{fig:ang2}
\end{figure}
Beyond $A_{\text{FB}}^{\text{H}}$, the angular distribution offers three additional null tests, one of which is the combined forward-backward asymmetry of the lepton and hadron system. As discussed in detail in Ref.~\cite{Golz:2022alh}, $F_L$ and the three forward-backward asymmetries are already sufficient to qualitatively disentangle NP effects from $C_7,\,C_7^\prime, C_9^\prime, C_{10}$ and $C_{10}^\prime$.
\section{Conclusions}\label{sec:concl}
\noindent We presented null tests in angular observables of three- and four-body decays of charmed baryons and investigated their NP sensitivity. Along with the fraction of longitudinally polarized dimuons $F_L$ and the hadron-side forward-backward asymmetry $A_{\text{FB}}^{\text{H}}$, dipole operators $O_{7}^{(\prime)}$ can be probed down to values as small as $\sim0.01$ and right-handed vector couplings $C_{9}^{\prime}$ down to $\mathcal{O}(0.1)$, whereas the null tests probe axial vector couplings $C_{10}^{(\prime)}$, which are zero in the standard model, with a significance of few$\,\times\,0.01$.
This null test strategy exploits synergies between three- and four-body modes and complements null test searches in rare charm meson decays. Contributions to different Wilson coefficients are disentangled and thus the issue of dominating resonance effects in rare charm decays is overcome. The presented strategy therefore constitutes an excellent road towards a future global fit, crucial for BSM tests with flavor changing neutral currents in the up-type flavor sector.
\acknowledgments
\noindent MG would like to thank the organizers for the wonderful web-format of the conference. This work is supported by the \textit{Studienstiftung des Deutschen Volkes} (MG) and in part by the \textit{Bundesministerium f\"ur Bildung und Forschung} (BMBF) under project number
05H21PECL2 (GH).
| 965940d30da3c143d89164700d0d7be93502202a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Photometric redshift estimates (photo-$z$s) represent a fundamental ingredient in studies of both galaxy and active galactic nuclei (AGN) evolution, and in an increasing number of cosmological experiments.
Broadly, methods of estimating photo-$z$s fall into two categories: template fitting approaches that fit model or empirical template libraries to the observed photometry \cite[most effective and popular for deep survey fields with large numbers of available filters;][]{Dahlen:2013eu}, and machine learning (ML) approaches that use classification or regression algorithms trained on reference samples to predict new redshifts for the science sample of interest \citep[e.g.][]{Collister:2004fx, CarrascoKind:2013kd, CarrascoKind:2014gb}.
Thanks to the ever increasing depth and fidelity of photometric surveys, the samples of optical sources for which these photo-$z$ estimates are required has grown significantly - requiring an increasing dependence on machine learning photo-$z$ approaches due to their potential efficiency and scalability \citep[see e.g.][]{2019NatAs...3..212S}.
In addition to the disadvantage in speed, reliable template-fitting photo-$z$s are more reliant on robust photometry and good photometric calibration \citep{Hildebrandt:2012du}, as well the use of appropriate template libraries that are able to encompass the range of colours present in the population of interest \citep[e.g.][]{2019MNRAS.489.3351B, 2017ApJ...850...66A} - something which becomes increasingly challenging as sample sizes grow.
Conversely, ML methods are less impacted by systematic errors within photometry that can limit template fitting as they do not rely on physically consistent models.
Rather, ML methods learn the correlations between the apparent colour or magnitude and redshift for the specific dataset in question.
Furthermore, they are also able to incorporate additional valuable information such as source size and morphology \citep{2018MNRAS.475..331G, 2020ApJ...888...83W} that can break redshift degeneracies.
However, given the requirement for training samples, machine learning photo-$z$ estimates perform best for populations where large spectroscopic samples are available for training and testing.
Many machine learning photo-$z$ estimates are therefore tailored towards the particular population of interest by design \citep{2007MNRAS.375...68C,2012ApJ...749...41B,2018MNRAS.478..592H}.
Even in the case where a broad range of populations are included within the training samples, the optimisation process for many of the regression or classification algorithms employed will be strongly weighted towards the most densely populated regions of the training space.
A natural consequence of this is that estimates for rare or extreme populations can be extremely biased or imprecise \citep[see e.g.][]{2019PASP..131j8004N,2020A&A...644A..31E}.
Alternatively, studies of samples which span a broad range of optical properties must rely on multiple different sets of photo-$z$ estimates for different sub-populations, each with their own systematic biases and limitations.
These potential limitations are particularly relevant for samples selected through other wavelength regimes, such as radio continuum and X-ray surveys, where the range of optical properties can be extremely broad.
For example, radio-selected samples from the new generation of wide area radio continuum surveys with MeerKAT \citep{Booth:2009wx}, the Australian SKA Pathfinder \citep[ASKAP;][]{2007PASA...24..174J} and the Low Frequency Array \citep[LOFAR;][]{vanHaarlem:2013gi} can range from massive elliptical galaxies at low redshift \citep{2019A&A...622A..17S} through to luminous quasars at $z > 5$ \citep{2021arXiv211006222G}.
\citet[][see also \citeauthor{Duncan:21}~\citeyear{Duncan:21}]{2019A&A...622A...3D} tried to solve this problem using a `hybrid' machine learning and template fitting approach over 400 deg$^{2}$ - with the template estimates contributing to good estimates for a wide range of source properties, while machine learning estimates provided better precision and reduced bias for the population where larger training samples were available \citep[see][for a similar approach to combining ML and template estimates]{2017MNRAS.466.2039C}.
While the hybrid approach provides photo-$z$ estimates with sufficient quality to enable a broad range of science \citep[e.g.][]{2019MNRAS.488.2701M,2019A&A...622A..12H, Smith:2020di}, the processing requirements of the template fitting estimates prohibit extension of this method to much larger datasets.
Extensions of data driven photo-$z$ methods to better cope with a diverse range of populations have been proposed.
\citet{2020MNRAS.498.5498H} present one solution, whereby unsupervised clustering of sources is used to divide the population into distinct regions of parameter space (specifically, colour-magnitude space), with machine learning photo-$z$ estimates then trained separately for each region.
By treating each region of parameter space independently, the resulting photo-$z$ estimates are significantly less biased and more robust than comparable estimates that include the full population within a single estimator.
However, while promising, the methods outlined in \citet{2020MNRAS.498.5498H} have so far only been applied to relatively small, but deep photometric datasets.
In this paper, we aim to build upon the methods of \citet{2019A&A...622A...3D}, \citet{Duncan:21} and \citet{2020MNRAS.498.5498H} to provide photo-$z$ estimates for the deepest photometric survey available over the full extragalactic sky, namely the Dark Energy Spectroscopic Instrument (DESI) Legacy Imaging Surveys \citep[][colloquially known as the `Legacy Surveys']{2019AJ....157..168D}.
While photo-$z$ estimates have been produced for the Legacy Surveys eighth data release (LS DR8 hereafter) and previous data releases, these estimates have crucially been limited to subsets of the optical population.
For example, \citet{2019ApJS..242....8Z} presents photo-$z$ and stellar mass estimates for Legacy Surveys DR6 and 7 (since updated to also include DR8).
By construction these estimates are limited to the redshift range $z < 1$, with morphological cuts imposed to exclude unresolved sources (i.e. those modelled by a point-spread function) from the analysis.
Similarly, \citet{2021MNRAS.501.3309Z} presents photo-$z$ estimates primarily targeted at the luminous red galaxy (LRG) population, with strict magnitude and colour cuts imposed to the limit the analysis to the target population.
In both cases, the selection criteria are well motivated by the data and respective scientific goals and the resulting photo-$z$ estimates are of a high quality for the chosen sub-populations.
However, given the explicit colour, morphology and redshift ranges for which the \citet{2019ApJS..242....8Z} and \citet{2021MNRAS.501.3309Z} photo-$z$s are designed, there remain key subsets of the optical population without valid photo-$z$ estimates, most notably unresolved optical sources at all redshifts (e.g. quasars) as well as resolved optical sources with significant AGN contributions and wider $z > 1$ populations.
The goal of this paper is therefore to provide a set of photo-$z$ estimates that encompass all populations for which robust photo-$z$s can be derived (given the observational limits of the photometry), regardless of redshift, colour or morphology.
Our photo-$z$ methodology will build upon the approach presented in \citet{2020MNRAS.498.5498H}, using the data itself to divide the empirical photo-$z$ training and prediction into different regions of optical parameter space with the goal of both improved results and computational performance.
The remaining sections of this paper are set out as follows.
In Section~\ref{sec:data}, we present details of the optical catalogues used in this analysis, the processing and homogenisation steps applied to them and the spectroscopic redshift samples used for photo-$z$ training and testing.
Section~\ref{sec:method} then outlines the photo-$z$ methodology employed, including testing and analysis of the variety of possible approaches explored and details of the final photo-$z$ training and estimation steps.
Next, Section~\ref{sec:results} presents a quantitative analysis of the resulting photo-$z$ estimates, including their performance as a function of optical properties.
We also analyse the photo-$z$ performance relative to existing estimates using the Legacy DR8 catalogues and for key subsets of the optical population (radio continuum and X-ray).
Finally, Section~\ref{sec:catalogues} provides details of the photo-$z$ catalogues presented in this analysis before Section~\ref{sec:summary} presents a summary of our work.
Throughout this paper, all magnitudes are quoted in the AB system \citep{1983ApJ...266..713O} unless otherwise stated. We also assume a $\Lambda$ Cold Dark Matter cosmology with $H_{0} = 70$ km\,s$^{-1}$\,Mpc$^{-1}$, $\Omega_{m}=0.3$ and $\Omega_{\Lambda}=0.7$.
\section{Observational Data}\label{sec:data}
\subsection{Photometric data}
\subsubsection{DESI Legacy Imaging Surveys}\label{sec:ls_data}
Initially motivated by target selection requirements for the forthcoming DESI spectroscopic survey, the Legacy Surveys comprise a set of optical photometric surveys providing imaging and source catalogues in the $g$, $r$ and $z$ bands.
Specifically, the Legacy Surveys consists of the Beijing Arizona Sky Survey \citep[BASS, $g/r$;][]{2017AJ....153..276Z} and the Mayall $z$-band Legacy Survey \citep[MzLS, $z$;][]{2016AAS...22831702S} providing imaging at declinations $\delta \gtrsim 32$ in the North Galactic cap (LS DR8 North).
Imaging over the remainder of the extragalactic sky (LS DR8 South) is provided by the Dark Energy Camera Legacy Survey (DeCaLS, PIs: D. Schlegel and A. Dey), supplemented by all publicly available imaging in the $g$, $r$ and $z$ bands \citep[predominantly the Dark Energy Survey,][]{2005astro.ph.10346T}.
Combined, the optical imaging in LS DR8 covers over 19\,000 deg$^{2}$ and is typically at least $\sim1$ mag deeper than the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS) 3$\pi$ Steradian Survey \citep{2016arXiv161205560C} in each band, reaching median $5\sigma$ depths $m_{g}\sim 24.3-24.9$, $m_{r}\sim 23.7-24.3$, and $m_{z}\sim 23.3-23.4$.
In addition to photometry from the dedicated optical imaging, the Legacy Surveys catalogues also incorporate mid-infrared (mid-IR) photometry from the Wide-field Infrared Survey Explorer mission \citep[WISE;][]{Wright:2010in} at 3.4, 4.6, 12 and 22\,$\mu$m (W1, W2, W3 and W4 respectively).
WISE photometry is provided for all optically detected sources based on the combined data from the cryogenic and post-cryogenic \citep[NEOWISE;][]{2011ApJ...731...53M} and NEOWISE-reactivation \citep[or NEOWISER;][]{2017AJ....153...38M, 2017AJ....154..161M} missions.
Full details of photometric calibration, source detection and photometry are provided in \citet[][and references therein]{2019AJ....157..168D}, but here we summarise the key details.
After self-consistent astrometric and photometric calibration, source detection is performed using the three individual bands and two combined $grz$ stacks (an un-weighted stack optimised for "flat" spectral energy distributions, and a weighted sum optimised for red sources).
In all five detection images, sources are selected using a simple thresholding algorithm, retaining unique detections above $6\sigma$ in any individual band.
Photometry is then performed using \textit{The Tractor} \citep{2016ascl.soft04008L}, which simultaneously models each source across all filters using a set of parametric light-profiles.
Specifically, \textit{The Tractor} models sources as either point-sources (i.e. with a delta function; $\texttt{type} =$ `PSF'), or as resolved sources with either a round exponential profile (`REX'), an exponential disk (`EXP'), a de Vaucouleurs $r^{-1/4}$ power-law (`DEV'), or a "composite" model with both an exponential and de Vaucouleurs component (`COMP').
During fitting of $grz$ photometry, the respective models are convolved with the corresponding point-spread function (PSF), with the PSF for individual exposures computed for each CCD separately using stars in the field \citep[see][for further details]{2019AJ....157..168D}.
\textit{The Tractor} assumes that the same model applies across all optical bands, i.e. if a source is spatially extended, the same light-profile is fitted to all images so that shape and size measurements are consistent.
Therefore, in addition to providing robust total flux measurements, morphological type, shape and size information is available for every source in the LS DR8 catalogue.
Mid-IR photometry is performed in a similar way, with \textit{The Tractor} forcing the location and shape of sources based on the optical information and convolving with the respective WISE PSF to model the flux of the sources in each WISE band.
Using the $grz$ information allows for deblending of confused WISE sources, reliably probing to significantly fainter mid-IR fluxes than possible with blind mid-IR selected photometric catalogues \citep{Wright:2010in}.
The additional mid-IR sensitivity provided by this deblending is particularly valuable for photo-$z$ estimation with LS DR8 due to the small number of optical filters available.
All of the Legacy surveys source detection and photometry is performed in small $0.25^{\circ} \times 0.25^{\circ}$ regions known as ``bricks".
For this work, we restrict our analysis to only those bricks with at least one exposure in all three optical bands.
When both downloading the LS DR8 catalogue data and using them for photo-$z$ prediction, we make the decision to group bricks into larger regions using Hierarchical Equal Area isoLatitude Pixelation \citep[\textsc{HEALPix};][]{2005ApJ...622..759G}\footnote{\href{http://healpix.sourceforge.net}{http://healpix.sourceforge.net}}.
Bricks are grouped based on the \textsc{HEALPix} order 3 pixel (with area $\approx 53.7$ deg) in which the right ascension and declination of the brick centre lies.
This results in 123 and 349 distinct \textsc{HEALPix} regions for LS DR8 North and South respectively, which together cover the total footprint of $\sim19\,400$ deg$^{2}$ (note that there is overlap between North and South and not all \textsc{HEALPix} regions are 100\% filled, particularly around the survey edges).
After restricting to primary sources for which photometric measurements are available, the final source catalogues for analysis comprise 323\,213\,867 sources in LS DR8 North and 1\,252\,523\,992 sources in the South.
\subsubsection{$\emph{asinh}$ magnitudes}
The LS DR8 photometric catalogues provide optical ($grz$) and WISE total flux density measurements (and associated uncertainties) for all sources.
Previous studies, however, have shown that photo-$z$ estimates derived using magnitudes yield improved results for methodologies such as those employed in this study \citep{2020MNRAS.498.5498H,Duncan:21}.
To enable training and estimation in log-flux space whilst also allowing for the possibility of non-detections and negative flux measurements, we follow the approach outlined in \citet{Duncan:21} and derive \emph{asinh} magnitudes \citep[][also colloquially known as `luptitudes']{1999AJ....118.1406L} for the purposes of photo-$z$ training and estimation.
For a given flux, $f$ (with a flux zeropoint $f_{0} = 3631~\rm{Jy}$), our \emph{asinh} magnitudes are defined as
\begin{equation}
m = \frac{-2.5}{\log(10)} \times \sinh^{-1}\left ( \frac{f/f_{0}}{2b} \right ) + \log(b)
\end{equation}
where the softening parameter, $b$, is unit-less (normalised by $1/f_{0}$). As defined in \citet{1999AJ....118.1406L}, the optimal choice for $b$ is approximately equal to the noise in the flux.
The corresponding magnitude uncertainties are then given by
\begin{equation}
\sigma_{m} = \frac{-2.5}{\log(10)} \times \frac{(\sigma_{f}/|f|)}{ \sqrt{\left(1 + (2b / (f/f_{0}))^{2}\right )}}
\end{equation}
where $\sigma_{f}$ is the 1-$\sigma$ flux uncertainty derived during model-fitting photometry (converted from inverse variance; \texttt{flux\_ivar\_[g/r/z]}).
When calculating \emph{asinh} magnitudes, we correct both the flux and flux uncertainties for Galactic extinction based on the transmission provided in the LS DR8 catalogues \citep[which combine the measured $E(B-V)$ of][ and the respective extinction coefficients, $A_{\lambda}/E(B-V)$, for each filter]{1998ApJ...500..525S}.
Previously \citet{2019MNRAS.489..820B} have shown that photo-$z$ estimates derived using \emph{asinh} magnitudes are not sensitive to the exact definition of softening parameter ($b$) assumed, provided the data are of similar depth across the field in question.
However, due to the variation in depth between LS DR8 North and South optical catalogues and the significant variation within each dataset, the assumption of homogeneity and a single value for $b$ for each filter is not justified.
We therefore make the decision to define the softening parameter $b$ based on the local 1$\sigma$ PSF depth for each source in the respective bands (e.g. $b_{g} = \sigma_{\text{PSF},g}/f_{0}$, where $\sigma_{\text{PSF},g}$ is \texttt{psfdepth\_g} from the LS DR8 catalogue converted to Jy).
Based on direct comparison of photo-$z$ estimates derived for sources in common between both LS DR8 North and South (which differ in optical depth, and hence softening parameter), we find the estimates are not sensitive to the exact choice of $b$ (Appendix~\ref{app:pz_offsets}), in line with previous studies.
\subsubsection{Data homogenisation}\label{sec:data_homogenisation}
Although both Legacy North and South imaging and catalogues have been processed and calibrated following the same procedures, the difference in effective filter response curves between the sets of $g/r/z$ observations means that the two datasets are not completely homogenous.
To avoid systematic biases or offsets between datasets (and their associated regions of the sky), photo-$z$ training and estimation must either be done separately for the two catalogues \citep[as in][]{2021MNRAS.501.3309Z}, or colour corrections applied to homogenise to a consistent photometric system.
While spectroscopic training samples sufficient for good photo-$z$ estimates are available within both datasets, Legacy South benefits from significantly improved training samples for complete magnitude selected samples (due to the presence of equatorial deep fields).
Therefore, in this work, we take the approach of homogenising the photometric datasets and using the full spectroscopic training sample to train photo-$z$ models that can be applied uniformly to both hemispheres.
Our colour corrections exploit the fact that there is a significant region of overlap between the two sets of imaging and catalogues.
After selecting a subset of \textit{The Tractor} bricks from the overlapping region, we positionally cross-matched the two sets of catalogues, keeping only those sources with matches within 0.35\arcsec (the pixel scale).
For a large set of bright, high S/N point-sources ($>10\sigma$ detections in each dataset), we then measured the \emph{asinh} magnitude offset between the measurements through each pair of filters as a function of observed $g - z$ colour in the Legacy North catalogue.
The resulting magnitude offsets for each of the three optical bands as a function of colour are illustrated in Fig.~\ref{fig:north-south}, both for individual sources (blue points) and the robust 3-$\sigma$ clipped mean and standard deviation within bins of $(g - z)_{\text{N}}$.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{figures/legacy_north_south_mag_offsets.pdf}
\caption{Observed magnitude offsets between Legacy North and South datasets as a function of $g - z$ colour for a sample of $\sim9\times 10^{5}$ objects where both datasets are available. Blue dots demonstrate individual measurements used in the analysis. Black circles show the 3$\sigma$-clipped mean offset in bins of $(g-z)_{\text{N}}$, with error bars showing the corresponding sigma-clipped standard deviation. For each filter ($g / r/ z$), the trend is fitted with a cubic polynomial (dashed line).}
\label{fig:north-south}
\end{figure*}
The largest magnitude offsets between North and South catalogues are observed in their respective $g$-band filters, with the systematic offset reaching $+0.1$ dex for the reddest sources.
Smaller offsets are observed in the $r$ and $z$ bands, reaching $\approx0.03$ and $\approx -0.02$ dex offsets respectively.
The more significant offsets observed for the $g$ filters are line with expectations based on the differences in the respective filter response curves \citep[see Fig.~3 of][]{2019AJ....157..168D}.
To homogenise the photometry before photo-$z$ training and estimation, we parametrise the trend as a function of $(g - z)_{\text{N}}$ using simple cubic polynomial fits (black dashed line in Fig.~\ref{fig:north-south}).
All sources in the Legacy North photometric catalogues are then corrected to match the Legacy South system based on their observed $(g - z)_{\text{N}}$ colour, with the observed colours clipped to the range $0.3 < (g - z)_{\text{N}} < 3.3$ to prevent overcorrection due to extreme (potentially unphysical) colours resulting from photometric scatter in lower S/N sources.
Given the average measured colour within the full catalogued population is $ (g - z)_{\text{N}} = 1.59$, the average magnitude offsets applied to the Legacy North photometry ar 0.056, 0.012 and -0.010 in the $g / r/ z$ bands respectively.
The impact (or lack thereof) of any residual colour systematics on the final resulting photo-$z$ estimates are explored in further detail in Appendix~\ref{app:pz_offsets}.
\subsection{Spectroscopic training data}\label{sec:data-specz}
Thanks to the large spatial coverage of the combined Legacy photometric datasets, extensive spectroscopic training samples are available over a very broad range of parameter space.
We compile our full training sample from two primary datasets.
Firstly, we make use of the combined samples presented in the 14th data release of the Sloan Digital Sky Survey \citep[SDSS DR14;][]{2018ApJS..235...42A}, which includes the SDSS main galaxy sample \citep{2002AJ....124.1810S} as well as the subsequent Baryon Oscillation Spectroscopic Survey \citep[BOSS][]{2013AJ....145...10D} and the Extended BOSS \citep{2016AJ....151...44D} samples.
For any SDSS spectroscopic source which was also in the separate DR14 quasar catalogue \citep[][DR14Q hereafter]{2018A&A...613A..51P}, we use the redshift measurement provided by DR14Q in place of the standard SDSS pipeline redshift.
For both samples, the SDSS catalogues are joined with the LS DR8 photometry using the row matched catalogues provided by Legacy Surveys, which provides the closest matching sources up to a maximum separation of 1.5\arcsec.
Since the combined SDSS spectroscopic sample extends to over 2 million individual sources, for our training and testing we make use of a random subset of 50\% of the SDSS sample, reserving the second half for potential future testing of the spatial variation in photo-$z$ quality or bias.
We next supplement the large samples provided by SDSS with an additional sample of spec-$z$s compiled over a large number of deeper survey fields for the Herschel Extragalactic Legacy Project \citep[HELP][]{2021arXiv210505659S}.
The full HELP spectroscopic redshift sample consists of publicly available spectroscopic redshifts from 101 individual spectroscopic surveys, with all input catalogues homogenised to consistent formats and coordinate systems.\footnote{The spec-$z$ compilation can be downloaded in its entirety or queried through standard Virtual Observatory tools at the \href{https://herschel-vos.phys.sussex.ac.uk/browse/specz/q}{HELP Virtual Observatory Server}.}
As part of the HELP compilation procedure, whenever multiple spec-$z$ measurements exist for the same individual source, the `best' available measurement has been retained based on visual inspection and duplicate entries have been removed \citep[see][for further details]{2021arXiv210505659S}.
The HELP spec-$z$ sample was cross-matched to LS DR8 using simple nearest neighbour match, employing a more conservative maximum separation than employed for the SDSS samples ($<1\arcsec$).
In total, 618\,606 sources in the HELP compilation are matched with the LS DR8 photometry catalogues.
However, this number includes a fraction of sources with unreliable spectroscopic redshifts, as well as some duplication of sources already contained in the SDSS sample.
For any source which is present in both the SDSS and HELP samples, we therefore remove the duplicate source to prevent over-weighting the source in training, retaining the spec-$z$ measurement provided by HELP.
Additionally, we include only sources deemed to have reliable redshift measurements, with HELP sources requiring a quality flag of 3 or higher, corresponding to ``good'', ``reliable'' or high confidence (>90\% confidence) redshifts depending on individual surveys definition \citep[see Section. 4.5 of][for additional details]{2021arXiv210505659S}.
For SDSS spectroscopic redshifts, we also require all training sources to have no redshift warning flags (\texttt{ZWARNING} = 0).
Finally, to ensure that our training sample is robust and as complete as possible for quasars out to the highest redshifts, we supplement the SDSS quasar samples with the compilation of $z > 5$ quasars presented by \cite{2020MNRAS.494..789R}.
The Very High-$z$ Quasar (VH$z$Q) catalogue\footnote{\url{www.github.com/d80b2t/VHzQ}} consists of 488 $z>5$ quasars with robust spectroscopic confirmations compiled from the literature, ensuring accurate positions with consistent astrometry.
Matching to the LS DR8 catalogues with the same 1\arcsec\, limit used for the HELP sample, we find matches for 375 sources extending out to redshifts as high as $z = 7.0842$.
To ensure a clean training dataset we apply additional photometric quality criteria based on the Legacy photometry catalogues.
We require $\texttt{fracflux\_[g/r/z]}$ (the profile-weighted fraction of the measured flux from other sources divided by the total flux) in each optical bands to be $ < 0.2$, ensuring that the source is not significantly impacted by blending with neighbouring sources.
Similarly, $\texttt{fracmasked\_[g/r/z]}$ (the profile-weighted fraction of pixels masked from all observations of this object) is required to be $ < 0.1$, such that the source is not affected by imaging artefacts.
Additionally, we require $\texttt{maskbits} = 0 $, such that the source does not overlap with regions included in the Legacy bitmasks.\footnote{See \href{https://www.legacysurvey.org/dr8/bitmasks/}{https://www.legacysurvey.org/dr8/bitmasks/} for additional information on LS DR8 bitmask definitions.}
Together, these criteria ensure that the training sample is free from significant source blending or contamination from bright nearby sources (e.g. artefacts around bright stars or extended galaxies) that would impact the observed colours, and hence bias the resulting photo-$z$ estimates.
Since automated source classification and redshifting pipelines employed for large surveys can never be 100\% reliable, small numbers of stars can contaminate even notionally `robust' spec-$z$s samples.
As a final cut, we exclude from the spectroscopic training sample potential stellar contaminants that exhibit either significant measured proper motions \citep[Gaia parallax S/N > 3;][]{2016A&A...595A...1G} or have colours consistent with stars ($P_{\text{star}} > 0.2$; see Section~\ref{sec:star-gal}).
In Fig.~\ref{fig:specz_training_dist} we present the redshift distribution for the final spectroscopic sample, split by best-fitting morphological type from the LS DR8 photometry.
The total sample statistics are also summarised in Table~\ref{tab:specz_summary}.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/specz_hist_zdist_all.pdf}
\caption{Redshift distribution of the final spectroscopic training sample, split by optical morphology classification within the LS DR8 catalogue (\texttt{type}). As would be expected given the optical photometry resolution and depths, resolved sources (EXP, DEV, REX, COMP) are largely confined to $z < 1$, while unresolved sources (PSF; predominantly quasars) extend to $z > 6$.}
\label{fig:specz_training_dist}
\end{figure}
\begin{table}
\centering
\caption{Total sample sizes for the spectroscopic training, validation and test subsets used in the photo-$z$ estimation. Training and validation samples are defined at the point of training (see Section~\ref{sec:method}), with a ratio of 3:1.}
\begin{tabular}{lrr}
\hline
Morphology & Training \& Validation & Test \\
\hline
PSF & 260\,930 & 65\,237 \\
EXP & 188\,792 & 47\,202 \\
DEV & 775\,309 & 193\,832 \\
REX & 162\,426 & 40\,612 \\
COMP & 35\,032 & 8\,762 \\
\hline
All & 1\,422\,489 & 355\,645 \\
\hline
\end{tabular}\label{tab:specz_summary}
\end{table}
\subsection{Star-QSO classification}\label{sec:star-gal}
Although robust classification of photometric point-sources in the LS DR8 catalogues is not a primary goal of this analysis, the inclusion of known stars within various steps in our methodology could negatively impact both the accuracy and precision of estimates for the extra-galactic point-source population.
We therefore first use the existing spectroscopic samples to train a star-galaxy classifier using the colour information available in the optical catalogues.
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{figures/star_prob_col_v2.pdf}
\caption{Median star probability ($P_{\text{star}}$, colour-scale) as a function of optical and mid-infrared colour in a sample of $\sim3.5\times10^{6}$ sources based on the GMM classification. A minimum of 5 sources per cell are required for plotting (with empty cells shown in white). To illustrate the colours of key samples of interest, also plotted are a random subset of 5000 known SDSS QSOs (blue points) and 5000 Legacy catalogue sources identified as stars independently from their colour based on high-significance parallax measurements from \textit{Gaia} (green points). The importance of WISE observations in cleanly separating the stellar population from QSOs can be seen in the middle and right panels.}
\label{fig:star_prob}
\end{figure*}
Motivated by previous successful approaches \citep{2011ApJ...729..141B,2012ApJ...749...41B}, we employ a Gaussian Mixture Model (GMM) to derive the densities of different source classes within the available photometric colour-space.
GMMs assume that a given dataset can be modelled by a mixture of a finite number of multi-dimensional Gaussian distributions and can be used both for the purposes of clustering to identify sub-populations within the sample, or as in this case to flexibly modelling the probability density of a distribution of data (i.e. the colours of stars).
To derive the star-QSO classification we construct GMMs for two different source classes.
Firstly, we use morphological point-sources (\texttt{type} = `PSF') from the spec-$z$ training sample to model the colour-space of the galaxy/QSO population.
Note that this sample includes all PSF sources, regardless of whether they were explicitly targeted as QSOs.
Next, we then use sources from the SDSS DR14 spectroscopic sample \citep{2018ApJS..235...42A} classified as `STAR' to model the stellar population.
We model each GMM using four colours derived from the LS DR8 catalogues; $g - r$, $r - z$, $z - W1$, $W1 - W2$, with 20 components (see Section~\ref{sec:method-divide} for further discussion of the choice of how many components are used to construct the GMM).
The resulting star probability is then defined as
\begin{equation}
P_{\text{star}}(\mathbf{\hat{m}}) \propto \frac{\mathcal{L}_{\text{star}}(\mathbf{\hat{m}})}{\mathcal{L}_{\text{star}}(\mathbf{\hat{m}}) + \mathcal{L}_{\text{QSO}}(\mathbf{\hat{m}})}
\end{equation}\label{eq:pstar}
\noindent where $\mathcal{L}(\mathbf{\hat{m}})$ is the probability density in a given GMM for a set of observed colours, $\mathbf{\hat{m}}$.
Note that strictly, to derive a true probability Eq.~\ref{eq:pstar} should be amended to incorporate prior probabilities on the relative source densities of the star and quasar populations \citep[also including the dependence on observed magnitude;][]{2011ApJ...729..141B}.
However, for the purposes of this work, it is sufficient to identify and remove the likely stellar population as the photo-$z$ estimates themselves can be used to further identify true QSO.
An additional caveat to our star-QSO separation procedure is that it is naturally biased towards stellar types present within the SDSS DR14 training sample and therefore may be incomplete for some key populations with colours similar to some QSO populations (e.g. cool dwarf stars and high redshift QSOs).
The full stellar training sample does however include stars initially targeted as quasar candidates based on their optical colours \citep{2015ApJS..219...12A}, providing valuable training data in the key regions of optical parameter space where the two populations are most similar (and where infrared colours can offer key information).
To illustrate the effectiveness of our star-QSO classifications for bright point sources, in Fig.~\ref{fig:star_prob} we show how the derived $P_{\text{star}}$ varies as a function of source colour.
We first calculate $P_{\text{star}}$ for a sample of $\sim3.5\times10^{6}$ PSF sources selected randomly from the LS DR8 catalogues.
For three projections of the 4-dimensional colour space used in the GMM, we then calculate the average $P_{\text{star}}$ as a function of observed colour.
For reference, we also illustrate the observed colours of a random subsample of spectroscopically confirmed QSOs from SDSS DR 14, as well as a set of confirmed stars that were identified independently of their observed photometric colours (or spectroscopic classification) - with robust ($>5\sigma$) measurements of parallax from \emph{Gaia} \citep[][see \citeauthor{2019AJ....157..168D}~\citeyear{2019AJ....157..168D} for details of cross matching to LS DR8 photometry]{2016A&A...595A...1G}.
Note that given the sensitivity limits of Gaia, parallax measurements are present only for bright stars within the LS DR8 catalogues ($m_{r}\lesssim 20$, corresponding to an average $13\%$ of PSF sources across the full catalogue).
Nevertheless, a stellar locus is clearly visible in each combination of observed colours, which is well traced by the region of high $P_{\text{star}}$.
The importance of WISE mid-IR information on robustly separating the star and QSO populations is also clearly demonstrated, with significant overlap between the two populations when only optical information is used ($g - r$ vs $r - z$), but a clear separation between the two loci when optical and mid-IR colours are combined.
\section{Photo-$z$ Methodology}\label{sec:method}
As with the previous studies upon which our methodology builds \citep{2019A&A...622A...3D,Duncan:21}, the primary algorithm we employ to derive photo-$z$ estimates is the sparse Gaussian process (GP) photometric redshift code, \textsc{GPz} \citep{2016MNRAS.455.2387A, 2016MNRAS.462..726A}, which extends the standard GP method to incorporate key features suited to photo-$z$ estimation.
Overall, \textsc{GPz} provides several benefits over standard GP implementations, including the use sparse of GPs that reduces computational requirements, with no loss in accuracy of the resulting models.
Additionally, \textsc{GPz} accounts for non-uniform and variable noise (heteroscedastic) within the input data - modelling both the intrinsic noise within the photometric data and model uncertainties due to limited training data.
Finally, \textsc{GPz} also allows for cost-sensitive learning (CSL), allowing different parts of parameter space to be weighted more or less importantly based on the specific scientific requirements.
In a key change from previous studies where we applied \textsc{GPz} to a range of datasets \citep{Duncan:2017ul,2019A&A...622A...3D,Duncan:21}, in this analysis we employ a new C++ implementation of \textsc{GPz}, \texttt{gpz++}\footnote{\url{https://github.com/cschreib/gpzpp}}, which provides substantial speed improvements and also implements support for missing data.
Furthermore, \citet{2020MNRAS.498.5498H} recently outlined a number of key augmentations to the photo-$z$ approach, which together lead to significant improvements in the resulting estimates compared to more straightforward applications of \textsc{GPz}.
In the following subsections, we present details of our implementation for a number of these photo-$z$ training augmentations and other additional steps in our training methodology, before summarising the final training and prediction pipeline in Section~\ref{sec:final-method}.
\subsection{Population division with GMM}\label{sec:method-divide}
Amongst the approaches suggested by \citet{2020MNRAS.498.5498H}, one of the most effective additions was found to be the use of GMM to divide the population into $N$ distinct regions of colour-magnitude space, with \textsc{GPz} trained separately on each mixture.
In addition to providing improved estimates overall, this `GMM-Divide' technique was also found to result in improved training and prediction speeds due to the order of complexity intrinsic to the sparse GP algorithm.
In Section~\ref{sec:star-gal}, we found that modelling the population in 4 colour dimensions provides sufficient information to separate key populations and model the complex distribution of the optical samples.
However, the optimal way to divide the galaxy population for the purposes of photo-$z$ estimation represents a different and non-trivial problem, with even the limited available photometric bands available in LS DR8 providing an extremely complex high-dimensional parameter space.
This available parameter space increases when morphological information and apparent magnitudes (in addition to colour) are included, with both potentially incorporating more redshift information than a given colour.
Similarly, in addition to selecting the parameter space used to divide the population, when generating a GMM one must also define the number of Gaussian model components.
Unlike in Section~\ref{sec:star-gal}, when dividing the population for photo-$z$ training purposes we must also balance the accuracy of modelling the populations within the chosen parameter space with minimising the potential for empty or under-sampled training sets for individual populations (i.e. no spec-$z$ training sources belonging to mixtures defined from the full photometric sample).
Although the optimal number of components required to model the desired parameter space can be derived iteratively, i.e. choosing the number of components which minimises the Bayesian Information Criterion (BIC), this must be balanced against the size of the resulting training sample for each GMM component in the subsequent \textsc{GPz} estimation and the corresponding photo-$z$ performance.
Optimally modelling the observed colour space with the GMM therefore does not necessarily result in optimal photo-$z$ estimates for the available data.
Based on initial tests, we find that including more than four dimensions within the GMM component results in unacceptable numbers of empty training sets for the current spectroscopic training samples.
For four GMM dimensions, the number of possible combinations for which data to use as input can not be practically be explored in full.
Our final approach to modelling the LS DR8 populations for the GMM-Divide methodology is therefore informed by prior knowledge on the most effective probes of redshift available within the optical catalogues, the distribution of expected signal to noise ratios for that observable and the results of manual experimentation.
First, the LS DR8 sample is separated into its individual morphological type (i.e. PSF, round exponential etc.), which implicitly separates the population into different parts of colour space.
For each morphological type, we then divide the population using GMMs constructed from the LS DR8 parameter space, with $g - r$ and $r - z$ colours and $m_z$ \emph{asinh} magnitudes used as the primary three dimensions for all sources.
For the resolved population (DEV, REX, EXP and COMP), the final GMM dimension was then selected to be the measured half-light radius, motivated by the positive impact of including size information in previous \textsc{GPz} photo-$z$ estimates \citep{2018MNRAS.475..331G}.\footnote{For the DEV and REX/EXP populations, we use the half-light radius of the deVaucouleurs (\texttt{SHAPEDEV\_R}) and exponential models respectively (\texttt{SHAPEEXP\_R}). For the composite population (COMP), which are modeled by the sum of deVaucouleurs and exponential models, we found that the half-light radii of the exponential components span a greater dynamic range and was therefore selected as the optimal size parameter for subsequent analysis.}
With size information unavailable for the unresolved (PSF) sources, $z-W1$ colour was chosen as the final GMM dimension due to its importance in separating stellar and extragalactic sources (Section~\ref{sec:star-gal}).
To ensure that the GMM is also representative of the \emph{extragalactic} PSF population for which we want to optimise photo-$z$ estimation, we only include PSF sources with $P_{\textup{star}} < 0.5$ and with parallax measurements $<3\sigma$ (if available).
\begin{figure}
\centering
\includegraphics[width=1.02\columnwidth]{figures/gmm_divide_col_mag_EXP.pdf}
\includegraphics[width=1.02\columnwidth]{figures/gmm_divide_col_mag_DEV.pdf}
\caption{Optical colour-colour (left) and colour-magnitude (right) distributions of EXP and DEV morphologies within the Legacy dataset. The colour-scale shows the density of sources in each cell (logarithmically scaled). The means and covariances of the 10 components in the respective GMMs are over-plotted as ellipses.}
\label{fig:gmm-divide}
\end{figure}
To illustrate how the observed parameter space can be modelled with the trained GMMs, in Fig.~\ref{fig:gmm-divide} we show two projections of full the colour-magnitude space for the EXP and DEV morphological populations.
Plotted over the density distributions for the full photometric sample are the means and covariances of the individual mixture components for the corresponding trained GMMs, where each GMM consists of 10 components.
In Section~\ref{sec:method-optimisation} we further explore how the number of GMM components impacts the overall quality of photo-$z$ predictions.
As outlined by \citet{2020MNRAS.498.5498H}, once the population has been modelled with a particular GMM, the subsequent population division can be incorporated into subsequent photo-$z$ analysis in a number of ways.
Firstly, sources can be assigned to their best-matching component based on which individual mixture model has the highest probability for the corresponding position in the 4-dimensional parameter space, with the photo-$z$ estimate taken from the \textsc{GPz} model trained on this component.
Alternatively, the photo-$z$ for a given source can be calculated as a weighted sum of the predicted photo-$z$s from the \textsc{GPz} models for all components, with the weighting derived from the relative probability of a source belonging to each component.
Unlike \citet{2020MNRAS.498.5498H}, who found that the majority of galaxies are assigned component probabilities near unity (and hence the weighted sum would produce identical results to the above approach), we find that a significant fraction of the catalogue sources have non-negligible probabilities of belonging to multiple GMM components.
The probabilistic approach could therefore potentially yield significant improvements in the overall photo-$z$ accuracy for the LS DR8 dataset.
However, due to the linearly increased computational requirements of predicting photo-$z$s for the full photometric sample ($k$ times longer prediction times for $k$ GMM components), such an approach was not deemed practical for the current analysis and the potential gains from using the linear superposition approach will be further explored in future studies.
Similar to the division of the full population, there is also flexibility in how the GMM is used to divide the spectroscopic training sample for training the respective instances of \textsc{GPz}.
As above, the simplest approach, and that adopted in this paper, is to assign each training source to its best-matching GMM component.
Alternatively, when training \textsc{GPz} for a given GMM component we can include all training sources with probability of belonging to the component greater than some non-negligible threshold, on the basis that including additional training sources on the periphery of the relevant parameter space should provide useful information.
\subsection{Weighted training samples for cost-sensitive learning}\label{sec:method-weight}
A second key approach suggested by \citet{2020MNRAS.498.5498H} employs the same GMM idea to calculate CSL weights for the training sample based on the relative probability density for a given set of features, $\mathbf{\hat{m}}$ (i.e. magnitudes and colours), in both the training sample and the full parent photometric sample.
Cost-sensitive learning allows for minimising potential biases in training from spectroscopic training samples that are not fully representative of the photometric sample.
For details of how CSL weights are incorporated into the objective function of \textsc{GPz} during training, we refer the reader to Section~4.2 of \citet{2016MNRAS.455.2387A}.
Similar to the nearest-neighbour based method employed by \citet{Lima:2008eu} and \cite{Duncan:2017ul}, the use of GMM to model the desired parameter space minimises the potentially negative impacts of small number statistics in some regions of parameter space.
Following \citet{2020MNRAS.498.5498H} we define the weighting for a given source as:
\begin{equation}
w_{i} = \frac{p_{\textup{full}}(\mathbf{\hat{m}}_{i}) + \epsilon}{p_{\textup{train}}(\mathbf{\hat{m}}_{i}) + \epsilon},
\end{equation}
where $p_{\textup{full}}(\mathbf{\hat{m}}_{i})$ and $p_{\textup{train}}(\mathbf{\hat{m}}_{i})$ correspond to the probability density for a source with photometric properties, $\mathbf{\hat{m}}_{i}$, in the GMMs for the full photometric population and spectroscopic training sets respectively.
The additional constant, $\epsilon$, is included to prevent sparsely populated regions of parameter space (typically in the smaller training sample GMM) from producing extreme weights that negatively impact the photo-$z$ training.
Due to the significantly increased training sample size in this work compared to that available in \citet{2020MNRAS.498.5498H}, we choose a smaller value of $\epsilon = 0.001$ (c.f. 0.01) which enables greater dynamic range in the resulting CSL weights.
Similarly, the maximum CSL weight allowed in our training was set to an increased value of 100 (c.f. 20).
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/gmm_weight_EXP.pdf}
\includegraphics[width=1.0\columnwidth]{figures/gmm_weight_DEV.pdf}
\caption{Illustration of the average CSL weight as a function of galaxy colour for two different morphologies in the photometric catalogues. The regions of dark colours (with medians weights $\lesssim10^{-1}$) highlight areas of parameter space with extensive training data such as the SDSS and BOSS/eBOSS ELG and LRG.}
\label{fig:gmm-weight}
\end{figure}
For self-consistency and computational efficiency, we employ the same GMM models of the colour-magnitude magnitude space as generated in Section~\ref{sec:method-divide}, combining the $g - r$ and $r - z$ colours and $m_z$ \emph{asinh} magnitudes for all sources, with the addition of either $z-W1$ colours for unresolved (PSF) sources or the measured half-light radius for resolved sources (DEV, REX, EXP and COMP).
Fig.~\ref{fig:gmm-weight} illustrates the resulting distribution of weights derived for the EXP and DEV populations using this method, showing the average CSL weights for a 2D ($g - r$ versus $r - z$) projection of the respective colour-magnitude(-size) space.
Across $g - r$/$r - z$ colour space, the median weights span over three orders of magnitude, demonstrating the significant variation in spectroscopic training samples available over the parameter space probed by LS DR8 photometry.
Regions of parameter space which are significantly over-represented in the spectroscopic samples can be also clearly seen for both morphological classes, resulting in very low CSL weights.
These over-dense regions are a result of large colour-selected samples from SDSS BOSS and eBOSS surveys, such as the emission line galaxy (ELG) and luminous red galaxy (LRG) samples for the EXP and DEV morphologies respectively.
\subsection{Optimisation of photo-$z$ method hyper-parameters}\label{sec:method-optimisation}
As outlined in Section~\ref{sec:method-divide}, the precise impacts of how we choose to model (and hence divide) the population with GMMs on the resulting bulk photo-$z$ estimates represents a complex choice.
Furthermore, how best to combine the population division and CSL weighting across the full range of parameter space also presents an open question.
To quantitatively test the impacts of different methodology choices on the resulting final photo-$z$ estimates and find the optimal approach for our dataset (and scientific objectives), we therefore perform a systematic test of a range of approaches.
Specifically, we investigate two primary choices: the number of mixtures used when fitting GMMs to observational parameter space, and the inclusion (or not) of colour/magnitude/size dependent CSL weights during photo-$z$ training.
Using the full spectroscopic training sample for each morphological type, we perform the `GMM-Divide' step outlined in Section~\ref{sec:method-divide} and train the associated \textsc{GPz} photo-$z$ models for GMMs with a number of components, $N$.
For all iterations, the total number of basis functions used for training is 500 (with 500/$N$ used to train the \textsc{GPz} model for each GMM component).
As input properties for training \textsc{GPz}, we use the seven \emph{asinh} magnitudes available for all sources ($g$, $r$, $z$ \& WISE 1-4), with half-light radius also included for resolved sources.
We evaluate the resulting photo-$z$ performance by calculating bulk quality statistics for the spectroscopic test samples for each group, using the robust scatter, $\sigma_{\textup{NMAD}}$, and an absolute outlier fraction, $\text{OLF}_{0.15}$.
Following common literature definitions \citep[e.g. ][]{Dahlen:2013eu}, we define
\begin{equation}
\sigma_{\textup{NMAD}} =1.48 \times \text{median} ( \left | \delta z \right | / (1+z_{\textup{spec}})),
\end{equation}
where $\delta z = z_{\textup{phot}} - z_{\textup{spec}}$.
Similarly, we define outliers as sources where
\begin{equation}
\left | \delta z \right | / (1+z_{\text{spec}}) > 0.15 .
\end{equation}
Since we are interested in only the relative performance of different approaches in this comparison, to minimise the impact of the least well constrained photo-$z$s on the overall statistics, for these comparison tests we exclude the 20\% of sources with the worst relative photo-$z$ uncertainty for each type ($\sigma_{z}/(1+z_{\text{phot}})$).
We then derive a balanced estimate of the photo-$z$ quality that is representative of the full population, calculating $\sigma_{\textup{NMAD}}$ and $\text{OLF}_{0.15}$ for 1000 sources that are drawn randomly (with replacement) from the respective test sample, with probabilities proportional to the CSL weight for each source.
We then repeat the random sampling 100 times and take the mean of the resulting distribution as the final estimate of $\sigma_{\textup{NMAD}}$/$\text{OLF}_{0.15}$ and the standard deviation of the distribution as an estimate of the corresponding uncertainty due to sample statistics.
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{figures/stats_vs_gmm_ncomp.pdf}
\caption{Representative photo-$z$ performance as a function of the number of components, $N$, used to divide the population with GMMs. For $N=1$, \textsc{GPz} is trained for the full population at once. Filled symbols (solid lines) show the results when CSL weights are included in the photo-$z$ training while open symbols (dotted lines) show the corresponding statistics for when \textsc{GPz} is trained with uniform weighting.}
\label{fig:gmm_ncomp_stats}
\end{figure}
Figure~\ref{fig:gmm_ncomp_stats} presents the resulting photo-$z$ performance as a function of $N$ for each optical morphology sample (filled symbols and solid lines).
We find that for all morphologies, the resulting scatter and outlier fractions are reduced when the population is split using the GMM, with the most significant improvement seen in the resulting outlier fractions.
However, we see that once the population has been sufficiently divided, there is no universal improvement to be gained from more detailed modelling of the paramater space (i.e. larger $N$).
Results for some morphologies are further improved with more GMM components (e.g. improved scatter and outlier fraction for REX sources, or outlier fractions for DEV sources), but others deteriorate, most notably PSF and EXP sources.
When examining the performance for rarer populations of interest, we find that the conclusions drawn from the full test population are still valid.
For example, calculating the scatter and outlier fraction for $z > 3$ QSOs in the test sample (with no weighting included in the statistics) we find that using 10 GMM components results in substantially improved outlier fractions compared to training all PSF sources simultaneously, with $\text{OLF}_{0.15}$ reducing from a very poor 0.71 to $\text{OLF}_{0.15} = 0.2$ (with $\sigma_{\textup{NMAD}}$ improving from 0.12 to 0.08).
For 15 or 20 GMM components the $\text{OLF}_{0.15}$ worsens to $\approx 0.32$, confirming that for the current available training samples 10 GMM components represents an optimal choice over a wide range of parameter space, although all options with larger $N$ would yield comparable results.
The second key methodology decision we explore is the inclusion of CSL weights derived from the GMM distributions (Section~\ref{sec:method-weight}).
When no population division is employed (i.e. $N = 1$), the incorporation of CSL weights during training leads to improvement in robust scatter and outlier fraction for all source types (open symbols and dotted lines in Fig.~\ref{fig:gmm_ncomp_stats}).
However, when combining population division and CSL weights the results are more varied, with some morphologies seeing better performance (both scatter and outlier fractions) with CSL weights included and others performing better with uniform training weights.
Motivated by the improved overall performance (or negligible loss in performance) for the resolved optical morphologies when CSL weights are included alongside sample division, we opt to include CSL weights as standard for our final methodology.
Finally, as discussed in Section~\ref{sec:method-divide}, an additional variation that can be explored is how spectroscopic training sources are assigned to GMM components derived in the GMM-Divide step.
However, when training \textsc{GPz} including all training sources with $>10\%$ probability of belonging to a given mixture and training including sources with $>50\%$ probability, we find no significant difference in the resulting photo-$z$ statistics.
For simplicity and speed, we therefore train \textsc{GPz} for GMM components using only the training samples which are best-matched to the given component.
In addition to the steps outlined in Sections~\ref{sec:method-divide} and \ref{sec:method-weight}, \citet{2020MNRAS.498.5498H} also found some improvement in accuracy of the photo-$z$ posterior predictions could be gained from training multiple iterations of \textsc{GPz} where the training data had been perturbed based on the photometric noise (resulting in non-Gaussian posteriors).
However, we determined that the small statistical improvement would not be of sufficient benefit to justify the linearly increased training and prediction times associated.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/pdfcalib_10_weight_05_v3.pdf}
\caption{Panels illustrate the cumulative distribution of threshold credible intervals, $c$, ($\hat{F}(c)$) for the spectroscopic test sample before (dotted lines) and after (solid lines) uncertainty calibration. The cumulative distribution is weighted using the CSL weights to provide a representative view of the overall uncertainties. Lines that rise above the 1:1 relation illustrate under-confidence in the photo-$z$ uncertainties (uncertainties overestimated) while lines that fall under illustrate over-confidence (uncertainties underestimated).}
\label{fig:error_calibration}
\end{figure*}
\subsection{Calibration of photo-$z$ uncertainties}\label{sec:method-errorcalib}
One of the key strengths of the \textsc{GPz} photo-$z$ algorithm is its ability to predict accurate uncertainties, with the accuracy of the uncertainty estimates fully incorporated into the objective function during model training.
However, due to the fact that the spectroscopic training sets cannot be fully representative (and CSL weights cannot perfectly account for these biases), input photometric uncertainties may be over or under-estimated, and its photo-$z$ posteriors are not multi-modal, previous studies have found that additional calibration of the uncertainties after training is still necessary \citep{Duncan:2017ul,2019A&A...622A...3D}.
We therefore include an additional post-training error calibration step in our method.
To quantify the over- or under-confidence of our photo-$z$ estimates, we follow the method outlined in Section~3.3.1 of \citet[][itself based on the proposal by \citeauthor{2016MNRAS.457.4005W} \citeyear{2016MNRAS.457.4005W}]{Duncan:2017ul}, whereby we calculate the distribution of threshold credible intervals, $c$ (where the spectroscopic redshift intersects the redshift posterior).
For perfectly accurate uncertainty estimates, the cumulative distribution of credible intervals, $\hat{F}(c)$, should follow a straight 1:1 relation, i.e. a quantile-quantile (or $Q-Q$) plot.
Curves that fall below this 1:1 relation indicate that the photo-$z$ errors are underestimated (i.e. the $P(z)$s are too sharply peaked), while curves that rise above indicate uncertainties are over-estimated and the predictions are under-confident.
As the uncertainties predicted by \textsc{GPz} are Gaussian, we calculate the threshold credible interval for a given source, $c_{i}$, following
\begin{equation}
c_{i} = \textup{erf} \left (\frac{| z_{i,\textup{spec}} - z_{i,\textup{phot}} |}{\sqrt{2} \sigma_{i}} \right),
\end{equation}
where $z_{i,\textup{spec}}$ and $z_{i,\textup{phot}}$ are the true (spectroscopic) and predicted redshift estimates and $ \sigma_{i}$ the corresponding predicted 1-$\sigma$ photo-$z$ uncertainty.
For the majority of source morphologies and apparent magnitudes in our trained \textsc{GPz} models (independent of the precise augmentation steps applied), we find that the photo-$z$ uncertainties are typically over-estimated.
When calibrating the uncertainties, we scale the original uncertainties as a function of magnitude, $m$, following
\begin{equation}\label{eq:gpz_err_scale}
\sigma_{\textup{new},i} = \sigma_{\textup{old},i} \times \alpha(m_{i}),
\end{equation}
\noindent where $\alpha(m)$ is defined as
\begin{equation}\label{eq:smoothing_2}
\alpha(m) = \begin{cases}
\alpha_{\eta} & m \leq m_{\eta}\\
\alpha_{\eta} + \kappa \times(m-m_{\eta}) & m > m_{\eta},
\end{cases}
\end{equation}
\noindent with $\alpha(m)$ constant ($\alpha_{\eta}$) below a chosen characteristic apparent magnitude, $m_{\eta}$, and scaling linearly with magnitude for sources fainter than $\alpha_{\eta}$ \citep{2009ApJ...690.1236I}.
We use the $r$-band optical \emph{asinh} magnitudes for calculating the magnitude dependence of the error scaling and assume a characteristic magnitude of $m_{\eta} = 16$.
The parameters $\alpha_{\eta}$ and $\kappa$ are then fitted for each morphology type individually (combining all respective GMM components) using the \textsc{emcee} Markov chain Monte Carlo fitting tool \citep[MCMC;][]{2013PASP..125..306F} to minimise the Euclidean distance between the measured and ideal $\hat{F}(c)$ distributions.
In this work, we include the additional step of weighting the measured $\hat{F}(c)$ distribution using the CSL weights derived in Section~\ref{sec:method-weight} to mitigate for biases in the spectroscopic training and validation samples used to calibrate photo-$z$ uncertainties.
Fig.~\ref{fig:error_calibration} illustrates the accuracy of the photo-$z$ uncertainty estimates of one set of photo-$z$ estimates (10 GMM components, CSL weights included; see below) before and after uncertainty calibration.
In Fig.~\ref{fig:error_calibration} we present an estimate of the overall photo-$z$ uncertainty based on the spectroscopic test sample weighted by the CSL weights derived in Section~\ref{sec:method-weight}.
A general trend visible in our results is that prior to the additional calibration, the photo-$z$ uncertainties predicted by \texttt{gpz++} are under-confident, overestimating the uncertainties across a wide range of magnitudes and across all morphologies.
Following the calibration step, the overall weighted uncertainty distributions are significantly improved, very closely following the optimal 1:1 relation.
However, for some morphologies (PSF, DEV, REX) the $\hat{F}(c)$ distribution falls slightly below the 1:1 relation at credible intervals of $0.8 < c < 1$, indicating that the tails of the photo-$z$ posteriors are underestimated.
This effect illustrates a limitation of the assumption of purely Gaussian posteriors in our analysis, which additional \textsc{GPz} augmentation steps could further mitigate in future.
Nevertheless, Fig.~\ref{fig:error_calibration} demonstrates that the overall accuracy of the corrected final photo-$z$ uncertainties are extremely accurate.
\subsection{Final photo-$z$ training and prediction pipeline}\label{sec:final-method}
Based upon the systematic tests outlined in Section~\ref{sec:method-optimisation}, the final photo-$z$ training methodology for our analysis is as follows:
\noindent the training sample and full population for each morphological type is modelled in four dimensions ($g - r$, $r - z$, $m_z$ and $z-W1$ or half-light radius) using 10 GMM components.
A total of 500 basis functions for each morphological type are then used to train \textsc{GPz} (50 per mixture model), with CSL weights included in the training for all components.
As above, we use the seven \emph{asinh} magnitudes available for all sources ($g$, $r$, $z$ \& WISE 1-4), plus the half-light radius for resolved sources, as inputs for training and prediction.
Sources within the training sample are used to train only their best-matching mixture model.
For all \textsc{GPz} training steps, the training sample is split such that 60\% is used for training and 20\% used for validation during training (to prevent over-fitting).
The remaining 20\% is retained for testing and quantitative analysis in Section~\ref{sec:results}.
For the final production run, we train \textsc{GPz} models for each GMM component in two stages, with the first pass trained using only the simpler variable diagonal covariances (GPVD) between parameters.
The trained GPVD model for each GMM component is then used as a starting point for a second training iteration using the full variable covariances (GPVC).
The two-stage training process allows us to make full use of the complex covariance modelling implemented by \textsc{GPz}, improving the overall accuracy of the photo-$z$ estimates and the estimated uncertainties (with smaller adjustments required in the uncertainty calibration stage), while also avoiding local minima during training and poor convergence during optimisation.
Due to the reduced complexity of the GP optimisation for the GPVD covariances, the increase in model training time is negligible.
Overall, this two-stage training process generally results in improved efficiency due to the prevention of training failures resulting from \textsc{GPz} models not reaching convergence.
The final trained GMM for each morphology and the full set of associated \texttt{gpz++} models (one for each GMM component) are stored for future predictions.
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{figures/ls_help_photoz_morph_v2.pdf}
\caption{Distribution of estimated photo-$z$ as a function of true spectroscopic redshift for each morphological type in the spectroscopic test sample (i.e. sources not included in any training or validation stage). Only sources with well-constrained estimates ($\sigma_{z} / (1+z_{\text{phot}}) < 0.2$) are included in the plot. Colour scales illustrate the total number of sources in each cell with a logarithmic colour scale and the blue dashed and dotted lines illustrate the 1:1 and $\pm 0.1\times(1+z_{\text{spec}})$ relations respectively.}
\label{fig:gpz_specz_photz}
\end{figure*}
For subsequent photo-$z$ prediction of new galaxy samples from LS DR8, we implement a pipeline which takes as input any catalogues in the format of the Legacy Surveys \texttt{tractor} catalogue.
The pipeline performs all the steps required to produce photo-$z$ predictions, including:
\begin{enumerate}
\item Deriving \emph{asinh} magnitudes and reformatting of available size information.
\item Homogenisation of photometry for LS DR 8 North sources (if required).
\item Calculation of star-QSO probability for PSF sources. Note that photo-$z$ predictions are derived for all PSF sources regardless of $P_{\text{star}}$, with the associated uncertainties and $P_{\text{star}}$ used to inform a final star-QSO classification alongside the photo-$z$ uncertainties.
\item Calculation of GMM component membership and CSL weights based on the source properties for the appropriate morphology (Sections~\ref{sec:method-divide} and ~\ref{sec:method-weight}).
\item Production of \texttt{gpz++} input catalogues and prediction with the associated trained \texttt{gpz++} model for the corresponding GMM component.
\item Adjustment of the output photo-$z$ uncertainties as a function of apparent magnitude and morphological type (Section~\ref{sec:method-errorcalib}).
\item Combination and formatting of all separate component and morphology catalogues into a single output catalogue for a given input catalogue.
\end{enumerate}
By design, the full pipeline is ambivalent to the specific choices made regarding the photo-$z$ training (i.e. the number of GMM components, inclusion of CSL weights, threshold for training sample model membership), accounting for all necessary book-keeping and merging.
When used to predict photo-$z$ estimates for the full LS DR8 catalogue in bulk, the processing is split into individual \textsc{HEALPix} order 3 regions.\footnote{We note that when using 10 cores (20 threads) on an Intel Xeon W-2175 14C 2.5GHz processor, the full prediction pipeline is able to process $\sim3\times 10^{6}$ sources per hour.}
In the following section we present a detailed quantitative analysis of the resulting photo-$z$ estimates for the spectroscopic test sample, including comparison to existing estimates in the literature and investigation of the photo-$z$ performance of individual sub-populations of interest.
\section{Photo-$z$ Quality}\label{sec:results}
The goal of this work is to produce photo-$z$ estimates that perform well over the full range of source types and redshifts present in the LS DR8 photometric catalogues.
Fig.~\ref{fig:gpz_specz_photz} presents a qualitative illustration that demonstrates how well the full \textsc{GPz} methodology is able to achieve this goal, showing the distribution of photo-$z$ estimates as a function of spec-$z$ for the spectroscopic test sample where the photo-$z$ is deemed to be relatively reliable, i.e. that the uncertainty on individual estimates are small.
Specifically, here and in the remainder of this work we define a permissive criteria that photo-$z$ estimates are \emph{well-constrained} if
\begin{equation}\label{eq:err_cut}
\sigma_{z} / (1+z_{\text{phot}}) < 0.2.
\end{equation}
For all resolved morphologies, the photo-$z$ predictions are extremely precise and reliable at $z < 1$, with no evidence of systematic biases and relatively few catastrophic outlier.
Although exhibiting visibly increased scatter, robust estimates for PSF sources (predominantly QSOs) span $0.5 < z < 3$ and beyond.
\subsection{Overall photo-$z$ quality statistics}\label{sec:stats_overall}
To provide a more quantitative analysis of the photo-$z$ performance and evaluate the range of parameter space for which they can be used for different scientific analyses, it is important to characterise the statistical properties of the photo-$z$ estimates.
In addition to calculating the $\sigma_{\text{NMAD}}$ and $\text{OLF}_{0.15}$ outlined above, in the following section we also examine the photo-$z$ performance including two additional statistics: the relative outlier fraction, $\text{OLF}_{3\sigma}$, defined as $\left | \delta z \right | / (1+z_{\text{spec}}) > 3\sigma_{\text{NMAD}}$, and the bias, $\Delta_{z} = \text{median} ( \delta z / (1+z_{\text{spec}}) )$.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{figures/gpz_mag_z_stats.pdf}
\caption{Robust scatter ($\sigma_{\textup{NMAD}}$; top), outlier fraction ($\textup{OLF}$; middle) and bias ($\Delta_{z}$; bottom) as a function of spectroscopic redshift and $z$-band magnitude for the full spectroscopic test sample. For a cell to be plotted we require a minimum of five galaxies.}
\label{fig:gpz_stats_z_mag}
\end{figure}
In Fig.~\ref{fig:gpz_stats_z_mag} we show how $\sigma_{\text{NMAD}}$, $\text{OLF}_{0.15}$ and $\Delta_{z}$ vary as a function of true redshift and apparent magnitude, $m_{z}$, for the full spectroscopic training sample (including all morphologies).
As in Fig.~\ref{fig:gpz_specz_photz}, analysis is restricted to photo-$z$ that are estimated to be well-constrained (satisfying Eq.~\ref{eq:err_cut}) and free from significant source blending or optical artefacts.
We find that both $\sigma_{\text{NMAD}}$ and $\text{OLF}_{0.15}$ follow a similar trend, with average photo-$z$ quality being a strong function of redshift and magnitude - typically with accuracy, precision and outlier fraction deteriorating for fainter magnitudes and the highest redshifts as expected.
The increase in scatter and outlier fraction towards fainter magnitudes is most significant at $z_{\text{spec}} < 1$, where the LS DR8 population is dominated by resolved morphologies.
At $z > 1$, where the catalogue becomes dominated by luminous QSOs, the overall photo-$z$ precision and reliability is reduced compared to low redshift but remains relatively consistent over a very broad range in redshift and magnitude.
When examining the bias, we find that the bias remains negligible across the majority of the redshift-magnitude space probed by LS DR8.
However, two distinct areas of parameter space exhibit significant bias: very faint sources at $z < 0.3$ which have photo-$z$ estimates that are significantly over-estimated, and QSOs at $2.5 \lesssim z \lesssim 4$ that are biased towards low redshifts.
In both cases, we postulate that the poorer performance in these regions of parameter space is primarily driven by the limitations in the available photometric information, rather than a failure of the photo-$z$ methodology itself or lack of training data.
In particular, for faint sources where no size or mid-IR information are available, the lack of $u$-band photometry in LS DR8 means that we are unable to probe key features such as the Balmer or Lyman break \citep[see e.g. Fig.~2 and 3 of][for a systematic test on the impact of $u$-band photometry on photo-$z$ estimates]{2008MNRAS.387..969A}.
Additionally, given the physical lower limit on redshifts within the training sample ($z=0$, excluding proper motion) and the overall redshift distribution of sources with $m_{z} > 20$, predictions for faint low-redshift sources that are poorly constrained will naturally be biased towards the higher redshifts at which the majority of the training sources with comparable magnitudes reside.
\subsection{Comparison with existing Legacy DR8 Photo-$z$s}\label{sec:zhou_comp}
As outlined in Section~\ref{sec:intro}, photo-$z$ estimates for the full LS DR8 catalogues do already exist within the literature.
The most recent and extensive of these estimates, presented by \citet{2021MNRAS.501.3309Z}, employs a random-forest machine learning approach applied to a combination of colour, magnitude and shape information based on the LS catalogues.
To enable a direct comparison with the estimates from \citet{2021MNRAS.501.3309Z}, we extract the photo-$z$ predictions from both datasets within two regions of the sky with dense spec-$z$ sampling coverage down to magnitude limits sufficient to provide a representative sample.
Within the LS DR8 North footprint, we use the region around the NOAO Deep Wide Field Survey \citep[NDWFS][]{{Jannuzi:1999wu}} in Bo\"{o}tes, which incorporates spectroscopic surveys such as the AGN and Galaxy Evolution Survey \citep[AGES;][]{Kochanek:jy}.
In LS DR8 South we use the region around the XMM-LSS deep field, which contains numerous redshift surveys across a broad range of parameter space.
After matching both sets of photo-$z$ estimates to the combined spectroscopic sample, we restrict the sample to only those sources which are free from source blending, are unaffected by masked regions in the LS DR8 imaging and have low probability of being a star ($P_{\text{star}} < 0.2$).
We then calculate the photo-$z$ quality statistics for both sets of estimates as a function of spec-$z$ using identical samples.
To estimate the potential variation in statistics given the smaller samples used in this comparison, we calculate the quality statistics for 500 bootstrap samples of a maximum of 100 sources in each bin, taking the median and 16/84th percentiles of the resulting distribution as our estimate and corresponding uncertainties.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{figures/gpz_zhou_z_stats.pdf}
\caption{Photo-$z$ performance metrics as a function for spec-$z$ for the estimates of \citet[][green circles]{2021MNRAS.501.3309Z} compared to those of this work (gold squares). Figure panels show $\sigma_{\text{NMAD}}$ (top), $\text{OLF}_{0.15}$/$\text{OLF}_{3\sigma}$ (middle) and $\Delta_{z}$ (bottom). Error bars encompass the 16 to 84th percentiles of 500 bootstrap samples, each containing a maximum of 100 sources. For the definition of $\Delta_{z}$ used in this work, positive values mean that photo-$z$s are biased to lower redshifts than the true redshift.}
\label{fig:zhou_stats_comparison}
\end{figure}
In Fig.\ref{fig:zhou_stats_comparison} we show the evolution of the key photo-$z$ statistics for both estimates as a function of spectroscopic redshift.
Overall, both sets of estimates provide excellent results at $z < 1$, with low scatter, outlier fraction and negligible bias.
We note that the estimates of \citet{2021MNRAS.501.3309Z} produce excellent results even beyond the parameter space for which they were designed.
However, at the highest redshifts where the LS DR8 photometric sample becomes dominated by QSOs, our \textsc{GPz} estimates perform significantly better, with substantially reduced bias (up to $>10\times$ lower) and absolute outlier fraction ($\text{OLF}_{0.15}$; $2-3\times$ lower).
At $1 < z < 2$, where LS DR8 contains a mix of luminous galaxies and QSOs, the photo-$z$ estimates in this work are significantly less biased and have substantially reduced fraction of catastrophic outliers ($\text{OLF}_{3\sigma}$) at the expense of marginally increased scatter.
These overall statistics demonstrate the difference in the respective intended use case and methodologies, with the photo-$z$ estimates produced here trading a small amount of performance for bright and well sampled populations (e.g. $z < 1$ and $m_{z} < 21$) for much more robust estimates of rarer or faint populations in LS DR8.
For studies at low redshift, where the lowest possible scatter and bias for a magnitude selected and resolved population is essential (e.g. clustering analysis of luminous red galaxies), the \citet{2021MNRAS.501.3309Z} therefore offer some increased performance compared to those in this work.
However, due to the significant bias observed for high-$z$ sources (incorrectly estimated to be $z < 1$), purely photo-$z$ selected samples could be contaminated by interlopers.
For studies that require reliable photo-$z$ estimates for the widest range of potential source types, the photo-$z$ estimates provided here should therefore provide substantially more reliable and less biased results - with negligible loss of precision for $z < 1$ population.
\subsection{Performance for rare or extreme populations}
In the following section, we explore the performance of our photo-$z$ estimates for a number of rare but scientifically import subsets of the LS DR8 photometric catalogues.
\subsubsection{High redshift quasars}\label{sec:results_highz}
As partially illustrated in Section~\ref{sec:zhou_comp}, a common problem for many empirical photo-$z$ estimates derived for samples that span a wide range of source properties and redshifts is that predictions for the highest redshift sources can become extremely biased towards low redshift.
These biases can be present even when the photometry involved is able to probe key redshift features such as the Lyman break and training samples extend beyond the relevant redshift range \citep[see e.g.][for examples across multiple datasets and machine learning algorithms]{2019PASP..131j8004N,2020A&A...644A..31E}.
To verify the performance of our photo-$z$ estimates out to the highest redshifts probed by the LS DR8 optical data, we select all sources from the spectroscopic redshift sample retained for testing that have both a predicted redshift of $z_{\text{phot}} > 4.5$ and have uncertainties that satisfy Eq.~\ref{eq:err_cut}.
The resulting $z_{\text{phot}}$ vs $z_{\text{spec}}$ distribution and associated sample statistics are presented in Fig.~\ref{fig:hzq}.
\begin{figure}
\centering
\includegraphics[width=0.92\columnwidth]{figures/gpz_gmm_highz_photz.pdf}
\caption{Comparison of predicted, $z_{\text{phot}}$, vs true redshift, $z_{\text{spec}}$, for the highest redshifts probed by the spectroscopic test sample. Sources selected to have $z_{\text{phot}} > 4.5$ with well-constrained estimates (Eq.~\ref{eq:err_cut}) are plotted as solid blue circles with associated 1$\sigma$ uncertainties. Additional spectroscopic test sample sources with $z_{\text{spec}} > 3.5$ are shown as orange pentagons (1$\sigma$ uncertainties are also shown for $z_{\text{spec}} > 4.5$ that are not selected through the photo-$z$ criteria).}
\label{fig:hzq}
\end{figure}
We find that our photo-$z$ estimates perform extremely well out to the highest redshifts (even extending to $z\sim7$), with the measured robust scatter $\sigma_{\text{NMAD}} = 0.05$ and outlier fraction of $\text{OLF}_{0.15} = 2.50\%$.
Although a small number of sources have $z_{\text{phot}}$ estimates significantly below the true redshift (see e.g. the $z_{\text{spec}}\sim6.5$ source with an estimate of $z_{\text{phot}}\sim 4$), the accurate photo-$z$ errors produced by our methodology means that such catastrophic failures can be identified and excluded from scientific samples.
Given that the majority of current spectroscopically confirmed QSOs at $z > 5$ used for both training and testing were selected using traditional Lyman-break colour selection techniques, we cannot reliably quantify how our photo-$z$ estimates will perform outside of this traditional parameter space.
The fact that the photo-$z$ remain reliable at $z\sim6$, where QSOs are typically selected based on $i-z$ colours not available in the LS DR8 catalogues, suggests that the inclusion of WISE magnitude and colour information may enable such extensions.
However, spectroscopic confirmation of sources selected in this way are required to verify the full potential of the full photo-$z$ sample.
Nevertheless, these results demonstrate that our photo-$z$ estimates can potentially be used to select robust samples of QSOs at $z > 4$ across the full $\sim19\,400$ deg$^{2}$ of LS DR8, or aid in the prioritisation or validation of candidates selected through other criteria.
\subsubsection{X-ray sources}
With the launch of the extended ROentgen Survey with an Imaging Telescope Array (eROSITA) on the Spectrum-Roentgen-Gamma telescope \citep{2021A&A...647A...1P}, we are now in a new era of sensitive all-sky X-rays.
In the coming years, spectroscopic surveys targeting X-ray point sources and X-ray selected cluster members \citep{2019Msngr.175...42M,2019Msngr.175...39F} will yield highly complete spec-$z$ samples discovered in the eROSITA All-Sky Survey (eRASS).
However, precise and unbiased photo-$z$s still remain a key tool for exploiting eRASS across a wide range of scientific objectives.
With LS imaging offering the deepest available all-sky (or full extragalactic sky) optical photometry available for eRASS source identification \citep{2021arXiv210614520S}, the photo-$z$s produced in this analysis offer a potentially valuable resource.
We therefore investigate how our photo-$z$ estimates perform for X-ray selected population present in the LS DR8 catalogues.
For bright X-ray sources over a large area, we consider sources from the Second Rosat all-sky survey \citep[2RXS;][]{2016A&A...588A.103B} and the XMM-Newton slew survey (XMMSL2)\footnote{\hyperlink{https://www.cosmos.esa.int/web/xmm-newton/xmmsl2-ug}{https://www.cosmos.esa.int/web/xmm-newton/xmmsl2-ug}}.
Robust optical host-identifications and spec-$z$s for the subset of these with optical spectroscopy are taken from the SPectroscopic IDentification of eROSITA Sources \citep[SPIDERS;][]{2017MNRAS.469.1065D} survey, completed as part of SDSS-IV.
To supplement the bright X-ray catalogues, we also include X-ray sources from the deep XMM-XXL North \citep{2016MNRAS.457..110M,2016MNRAS.459.1602L} and Chandra Deep Wide Field Survey \citep[CDWFS;][]{2020ApJS..251....2M} fields.
In total, the test sample of X-ray soruces consists of 16,004 sources spanning a broad range of X-ray flux densities, including significant statistics down to X-ray fluxes as faint as $F_{0.5-10~\text{keV}} \sim 3\times10^{-15}$ erg\,cm$^{-2}$\,s$^{-1}$, below the expected final eROSITA survey depths \citep[eRASS:8;][]{2021A&A...647A...1P}.
Of this full sample, 14\,207 have LS DR8 photometry free from artefacts or significant source blending.
Overall, we find that the fraction of sources for which we can derive well-constrained photo-$z$ estimates is significantly lower than for the general population, with only 55\% of sources with clean photometry having $\sigma_{\text{phot}}/(1+z_{\text{phot}}) < 0.2$.
However, when examining the demographics of the X-ray spectroscopic sample, we find that a significant fraction of the sample are PSF sources at $z_{\text{spec}} < 0.5$ where our photo-$z$ estimates are known to be less reliable (Section~\ref{sec:stats_overall}).
The increased uncertainty at $z < 0.5$ can be seen in the distribution of $z_{\text{phot}}$ estimates as a function of $z_{\text{spec}}$ presented in Fig.~\ref{fig:xr_spz_phz}, where all sources in the X-ray sample with clean photometry are shown in the background, while sources with well-constrained photo-$z$ estimates are over-plotted in solid colour.
Crucially, we find that by imposing cuts based solely on the predicted photo-$z$ uncertainties (Eq.~\ref{eq:err_cut}), we can reliably select samples with accurate and unbiased photo-$z$ estimates.
In this regard, the performance for luminous X-ray sources is comparable to that of X-ray faint sources with equivalent optical properties.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{figures/xray_specz_photz.pdf}
\caption{Distribution of predicted photo-$z$ ($z_{\text{phot}}$) vs spec-$z$ ($z_{\text{spec}}$) for a sample of X-ray detected sources from a compilation of deep X-ray surveys within the LS DR8 footprint. The colour scale illustrates the density of sources (with logarithmic scaling), with all sources with clean photometry shown as semi-transparent cells in the background and photo-$z$ estimates satisfying Eq.~\ref{eq:err_cut} over-plotted as solid colours.}
\label{fig:xr_spz_phz}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{figures/xray_stats_z_flux_sigma.pdf}
\includegraphics[width=0.95\columnwidth]{figures/xray_stats_z_flux_olf15.pdf}
\caption{Robust scatter, $\sigma_{\text{NMAD}}$ (top), and absolute outlier fraction, $\text{OLF}_{0.15}$ (bottom), as a function of X-ray flux, $F_{X}$ ($ = F_{0.5-10~\text{keV}}$ or $F_{0.5-10~\text{keV}}$), and true redshift for the compilation of X-ray detected sources. Statistics are calculated only for sources with $z_{\text{phot}}$ estimates deemed as well-constrained, with a minimum of ten sources required per bin.}
\label{fig:xr_stats_z_flux}
\end{figure}
To further investigate the dependence of our photo-$z$ quality on X-ray properties, we calculate the photo-$z$ quality as a function of observed X-ray flux, $F_{X}$ (specifically, $F_{0.5-10~\text{keV}}$ or $F_{0.5-7~\text{keV}}$ for CDWFS), and spectroscopic redshift.
The resulting statistics are presented in Fig.~\ref{fig:xr_stats_z_flux}, with the sources spanning 3 orders of magnitude in X-ray flux and extending to $z > 3$.
At $z < 0.5$, we find that the photo-$z$ predictions with low uncertainties exhibit low scatter ($\sigma_{\text{NMAD}} < 0.05$) and low outlier fraction ($<10\%$) across a very wide dynamic range in X-ray flux densities.
As redshift increases, we find that both scatter and outlier fraction increase in line with trends observed for the wider optically selected population.
Overall, our photo-$z$ estimates perform comparably to dedicated template based estimated derived using the same photometric data \citep{2021arXiv210614520S}, with the key additional benefit of also providing high quality estimates for the full optical population (Section~\ref{sec:stats_overall}).
Our photo-$z$ catalogues therefore offer the ability to e.g. robustly separate luminous background X-ray point-sources from foreground cluster members \citep{2021arXiv210614519K,2021arXiv211009544B} within one single homogeneous resource across the full sky.
\subsubsection{Radio continuum detected sources}
As with eROSITA and the X-ray sky, a new generation of wide area radio continuum surveys with facilities such as LOFAR and ASKAP are probing an order of magnitude fainter than previous all-sky surveys \citep[e.g.][]{Shimwell:2017ch}.
Together, the LOFAR Two-metre Sky Survey \citep[LoTSS;][]{Shimwell:2018to} in the northern hemisphere and the Evolutionary Map of the Universe \citep[EMU;][]{Norris:2011de} in the south will detect tens of millions of faint radio continuum sources.
The extreme diversity in optical properties for radio continuum selected sources means that these new samples will range from low-luminosity radio AGN and star-forming galaxies in the nearby Universe through to the most distant luminous quasars at high redshift \citep{2021arXiv211006222G}.
To measure the quality of the photo-$z$ estimates as a function of their radio properties we construct a sample of radio continuum selected sources with spectroscopic redshifts from the LOFAR Two-metre Sky Survey First Data Release \citep[LoTSS DR1][]{Shimwell:2018to,2019A&A...622A...2W} over $>400\,\text{deg}^{2}$.
The wide area sample from LoTSS DR1 is then supplemented with additional faint radio sources from the Bo\"{o}tes field of the LoTSS Deep Fields First Data Release \citep[][]{Tasse:2020tr,Sabater:2020ur,Kondapally2020,Duncan:21}.
The combined sample contains 32\,495 sources at $0.01 < z < 5$, of which 26\,957 have LS DR8 photometry free from artefacts and significant blending and the radio continuum flux densities range from $S_{\nu, 144\text{MHz}} <80\mu$Jy to $>10$\,Jy (>5$\sigma$ detections) - spanning the full dynamic range of the radio continuum population.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{figures/radio_specz_photz.pdf}
\caption{Distribution of predicted photo-$z$ ($z_{\text{phot}}$) vs spec-$z$ ($z_{\text{spec}}$) for a sample of low-frequency radio continuum detected sources within the LS DR8 North catalogues. The colour scale illustrates the density of sources (with logarithmic scaling), with all sources with clean photometry shown as semi-transparent cells in the background and well-constrained photo-$z$ estimates over-plotted as solid colours.}
\label{fig:rad_spz_phz}
\end{figure}
In Fig.~\ref{fig:rad_spz_phz} we show the distribution of photo-$z$ vs spec-$z$ distribution for the radio continuum sample.
As above, we show all radio sources sample with clean photometry shown in the semi-transparent background, while sources photo-$z$ uncertainties below our permissive cut of $\sigma_{\text{phot}}/(1+z_{\text{phot}}) < 0.2$ are plotted in solid colour.
Over the full spectroscopic sample, we find that 89\% of the radio spec-$z$ sample meet this additional uncertainty criteria, significantly more than for the X-ray selected population and indicative of the fact that the radio spec-$z$ population is dominated by resolved sources at $z < 1$ for which our photo-$z$ estimates are highly reliable.
Similar to the X-ray population, we find that the additional photo-$z$ precision criteria excludes a significant number of $z_{\text{spec}} < 0.5$ sources with photo-$z$ estimates $z_{\text{phot}} > 1$.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{figures/radio_stats_z_flux_sigma.pdf}
\includegraphics[width=0.95\columnwidth]{figures/radio_stats_z_flux_olf15.pdf}
\caption{Robust scatter, $\sigma_{\text{NMAD}}$ (top), and absolute outlier fraction, $\text{OLF}_{0.15}$ (bottom), as a function of radio continuum flux density and true redshift for the LoTSS 144 MHz radio continuum detected sources with available spec-$z$. Statistics are calculated only for sources with photo-$z$ estimates deemed as well-constrained, with a minimum of ten sources required per bin.}
\label{fig:rad_stats_z_flux}
\end{figure}
Fig~\ref{fig:rad_stats_z_flux} presents the $\sigma_{\textup{NMAD}}$ and $\text{OLF}_{0.15}$ as a function of both spec-$z$ and 144\,MHz flux density, $S_{\nu, 144\text{MHz}}$.
Similar to trends observed by \citet{2019A&A...622A...3D}, we find a clear evolution in the scatter and OLF of the radio source population with $z_{\textup{spec}}$, with photo-$z$ estimates of high redshift sources being significantly worse than for sources with similar radio flux density at low redshift.
However, for a given $z_{\textup{spec}}$ bin we find no evidence of strong variation in photo-$z$ precision or reliability as a function of radio continuum flux density (and hence luminosity).
Compared directly to the photo-$z$ estimates provided by \citet{2019A&A...622A...3D} for the same sample, we find that our photo-$z$s for the optically luminous QSO population perform marginally worse.
For sources in LoTSS DR1 estimates at $1 < z_{\text{spec}} < 3$ with reliable photo-$z$ in this work, the estimates from \citet{2019A&A...622A...3D} have $\sigma_{\textup{NMAD}} = 0.091$ and $\text{OLF}_{0.15} = 19.5\%$, while the photo-$z$s in this work have $\sigma_{\textup{NMAD}} = 0.103$ and $\text{OLF}_{0.15} = 22.8\%$.
We postulate that the reasons for \citet{2019A&A...622A...3D} performing better in this regime are two-fold: firstly, the inclusion of additional $i$ and $y$-band photometry in the \textsc{GPz} estimates, and secondly, due to training \textsc{GPz} estimates specifically for identified optical QSO population (cf. the more generalised approach in this work).
Outside of this specific parameter space, where the overall reduction in quality is only of order $\sim10\%$, the photo-$z$s produced in this work significantly out-perform those provided by \citet{2019A&A...622A...3D}.
The estimates in this work have $\sim35\%$ lower scatter and outlier fraction at $z_{\text{spec}} <1$ and $\sim3-4$ times lower scatter and bias at $z_{\text{spec}} > 3$.
Given the overall performance of our photo-$z$s for radio continuum selected sources, the catalogues produced in this work offer a valuable resource for exploiting the new generation of wide area radio continuum surveys.
\section{Photo-$z$ predictions for the full LS DR8}\label{sec:catalogues}
Following the validation of the final photo-$z$ performance and reliability, photo-$z$ predictions for the full LS DR8 sample were produced by processing each of the 123 and 349 distinct \textsc{HEALPix} regions for LS DR8 North and South respectively with the pipeline steps outlined in Section~\ref{sec:final-method}.
As outlined in Section~\ref{sec:ls_data}, catalogues were limited to LS DR8 bricks with at least one exposure in all three optical bands and luminous bright \textit{Gaia} sources (\texttt{type} = `DUP') were excluded.
In total, 323\,213\,867 and 1\,252\,523\,992 sources LS DR8 North and South respectively were processed through the pipeline.
However, these full optical photometry catalogues will contain many sources for which photo-$z$ estimates are not valid, or reliable.
This may be the case either because the source is a star, or because the photometric is adversely effected by imaging artefacts, neighbouring bright stars (e.g. through diffraction spikes) or due to very significant blending between close objects.
To allow for high quality scientific samples to be easily selected from the full photo-$z$ catalogues, we define a set of simple criteria to select sources for which photo-$z$ estimates are expected to be relatively unbiased and accurate enough for general astrophysical studies.
Photometry for a given object is deemed to be `clean' if the following criteria are satisfied:
\begin{equation}\label{eq:blend}
\texttt{fracflux\_}x < 0.33 \quad \text{or} \quad S/N(x) < 2
\end{equation}
and
\begin{equation}\label{eq:mask}
\texttt{maskbits} = 0
\end{equation}
where $x$ is each of the three optical bands, $g$, $r$ and $z$.
Allowing $\texttt{fracflux\_}x > 0.33$ for objects which have $S/N(x) < 2$ ensures that objects which are genuinely non-detections in a particular band (e.g. high-redshift sources) are not excluded due to low levels of contaminating flux from neighbouring objects.
In addition, to exclude likely stars we require
\begin{equation}\label{eq:pstar_cut}
P_{\text{star}} < 0.2,
\end{equation}
for unresolved optical sources.
Finally, estimates are deemed plausible if they satisfy the uncertainty criteria outlined in Eq.~\ref{eq:err_cut}.
One of the key advantages of the \textsc{GPz} algorithm is its ability to account for noise from the training data when deriving uncertainty estimates.
Sources with properties not well modelled by the spectroscopic training sample (i.e. stars) therefore have large predicted uncertainties, even if the noise on the input magnitudes are extremely small.
Examining the \textsc{GPz} predictions for spectroscopically confirmed stars in LS DR8 (from SDSS DR14, Section~\ref{sec:star-gal}), we find that using solely the photo-$z$ uncertainty cut is able to successfully remove 93.9\% of SDSS stars from the sample.
Requiring $P_{\text{star}} < 0.2$ alone successfully excludes 98.9\% of stars.
However, when combined the criteria in Eq.~\ref{eq:pstar_cut} and \ref{eq:err_cut} are able to exclude 99.9\% of stars in the SDSS spectroscopic sample.
Exploring the effectiveness of these cuts as a function of optical magnitude we find that the purity does decline for fainter sources, ranging from 100\% of SDSS stars excluded at bright magnitudes down to $97.0^{+1.3}_{-2.2}\,\%$ at $m_{z} = 21.5$ (where the uncertainty is derived solely from binomial uncertainties from sample statistics).
\begin{table}
\centering
\caption{Total sample sizes for the full optical catalogues processed in this analysis, the subset which pass the criteria for `clean' photometry (Eq.~\ref{eq:blend} and \ref{eq:mask}), and the number predicted to have good quality photo-$z$ estimates based on the combination of quality criteria outlined in the text.}
\begin{tabular}{lrrr}
\hline
Morphology & Total & Clean Photometry & `Good' Photo-$z$\\
\hline
& \multicolumn{3}{c}{North} \\
\hline
PSF & 135\,630\,162 & 109\,369\,165 & 41\,938\,327 \\
DEV & 15\,445\,582 & 13\,058\,724 &13\,050\,326 \\
EXP & 24\,354\,533 & 21\,777\,404 & 21\,388\,924 \\
REX & 147\,576\,416 & 132\,511\,841 & 123\,654\,049 \\
COMP & 207\,174 & 145\,098 & 127\,708 \\
\hline
All & 323\,213\,867 & 276\,862\,232 & 200\,159\,334 \\
\hline
& \multicolumn{3}{c}{South} \\
\hline
PSF & 571\,512\,918 & 489\,858\,010 & 205\,831\,876 \\
DEV & 42\,818\,253 & 37\,032\,451 & 37\,006\,782 \\
EXP & 114\,955\,799 & 105\,469\,819 & 100\,327\,246 \\
REX & 522\,373\,297 & 477\,763\,969 & 406\,110\,027 \\
COMP & 863\,725 & 660\,308 & 595\,708 \\
\hline
All & 1\,252\,523\,992 & 1\,110\,784\,557 & 749\,871\,639 \\
\hline
\end{tabular}\label{tab:photz_summary}
\end{table}
Table~\ref{tab:photz_summary} summarises the total photometric sample processed in this analysis for both LS DR8 North and South, as well as the number of sources that satisfy the quality criteria outlined above - deemed to be `good' photo-$z$ predictions.
Combining the two samples following the approach taken by the LS DR8 team (where LS DR8 North sources are only included at $\delta > 32.375\degr$ and in the North galactic plane), the total number of unique sources with `good' photo-$z$ estimates is 936\,131\,489.
Averaged over all source morphologies, the fraction of sources with good photo-$z$ estimates is similar between the two datasets (61.9\% and 59.9\% in the North and South respectively).
As expected, significantly fewer sources with PSF morphology have good photo-$z$ predictions, consistent with the majority of bright point-sources in the LS DR8 optical catalogues being stars.
The fraction of PSF sources with good photo-$z$s is also seen to vary strongly with galactic latitude, ranging from only $\sim15\%$ of PSF sources with good estimates at galactic latitudes of $| b | < 20\degr$, up to $\approx 45\%$ at $| b | > 80\degr$.
Similar trends with galactic latitude are seen for resolved morphologies, but with weaker evolution.
For example, the fraction of exponential morphologies with good photo-$z$ estimates ranges from $\approx 70\%$ to $\approx 90\%$ over the same range.
The decline in photo-$z$ reliability for resolved sources is driven primarily by the impacts of increasing source density (particularly of bright sources), with half of the increase in poor photo-$z$s attributed to the increased fraction of sources affected by masked pixels (e.g. due to bright stars, diffraction spikes etc.).
\begin{figure}
\centering
\includegraphics[width=0.99\columnwidth]{figures/nzdz_sqdeg_all_north_south.pdf}
\caption{Distribution of photo-$z$ point-estimates for sources in the LS DR8 catalogues with good quality photo-$z$ estimates. Solid and dashed lines represent the median source density per unit redshift across all \textsc{HEALPix} regions in LS DR8 North and South respectively, with shaded regions illustrate the corresponding inter-quartile ranges. Thick lines show the median source density per unit redshift for sources with $m_{z} < 21$, illustrating that the two datasets produce consistent predictions for the same magnitude limit.}
\label{fig:photoz_dist}
\end{figure}
When comparing statistics for the two datasets, the most significant difference occurs for round exponential morphologies (REX), which constitute the majority of resolved sources in both datasets.
Given the excellent consistency between North and South for both spec-$z$ test samples and with each other (see Appendix~\ref{app:pz_offsets}), we attribute the lower fraction of good estimates in LS DR8 South to the increased optical depth resulting in the detection greater numbers of higher redshift ($z > 1$) sources, for which the photo-$z$ estimates are less precise.
In Fig.~\ref{fig:photoz_dist} we show the distribution of $z_{\text{phot}}$ per unit area of sources with good photo-$z$ estimates in both datasets (not accounting for the full posterior predictions of individual sources).
At $z < 1$, where resolved morphologies constitute a large fraction of the extragalactic population, the photo-$z$ distributions (averaged over all \textsc{HEALPix} regions) for North and South are in extremely strong agreement.
Above $z > 1$, LS DR8 South provides additional well-constrained photo-$z$ estimates compared to North, but follows a similar decline in the source density of robust estimates as redshift increases -- in both datasets the good photo-$z$ estimates extend beyond $z > 6$ (see Section~\ref{sec:results_highz}).
The extremely tight agreement in the predicted redshift distribution of bright optical samples ($m_{z} < 21$; thick solid and dashed lines respectively) supports the conclusion that the primary cause of the difference in predictions for LS DR8 North and South is the depths of the optical photometry.
\subsection{Photo-$z$ catalogues}
Table~\ref{tab:cat_desc} presents information on the columns available in the photo-$z$ catalogues, while Table~\ref{tab:cat_example} provides an example of the catalogues for the first ten sources in LS DR8.
Key object identifier and positional information from the underlying LS DR8 photometric catalogues is included to enable unambiguous matching to other columns available in LS DR8 (including underlying flux information and additional morphological properties).
The catalogues provided in this work contain information for all optical sources processed through the \textsc{GPz} pipeline, regardless of the reliability of the resulting estimates.
Sources that appear in both LS DR8 North and South optical catalogues will have separate photo-$z$ predictions based on the two sets of photometry and we provide all available predictions.
For consistency with the convention used by the Legacy Survey collaboration we recommend users follow the approach used above, using LS DR8 North predictions for sources at $\delta > 32.375\degr$ in the North Galactic Cap ($ b > 0 \degr$) and LS DR8 South predictions otherwise.
In addition to the main photo-$z$ prediction and calibrated 1-$\sigma$ uncertainties, a separate flag is provided for all sources indicating whether they satisfy the criteria for `good' photo-$z$ prediction (Eq.10--13).
As outlined above, this photo-$z$ quality flag is able to reliably exclude most bright stars and extragalactic sources with unreliable estimates, allowing for robust samples over a wide range of parameter space.
However, we caution that with the full sample reaching over $\gtrsim10^{9}$ sources, no single assessment criteria can offer 100\% reliability for all potential scientific uses.
We therefore encourage users of the photo-$z$ catalogues to incorporate additional sample selection criteria appropriate for their specific target population and scientific goals.
\begin{table*}
\caption{Format of the LS DR8 North and South photo-$z$ catalogues produced in this work. Full catalogues can be accessed online through the \textit{Vizier} catalogue service as well as the NOIRLab \textit{Astro Data Lab}.}\label{tab:cat_desc}
\begin{tabular}{lll}
\hline
Column & Units & Description \\
\hline
\texttt{id} & & Unique object identifier combining \texttt{release}, \texttt{brickid} and \texttt{objid} \\
\texttt{release} & & LS DR8 release number - denoting the camera and filter set used \\
\texttt{brickid} & & LS DR8 brick number [1,662174] \\
\texttt{objid} & & LS DR 8 catalogue object number within this brick\\
\texttt{ra} & deg & Right ascension (J2000) \\
\texttt{dec} & deg & Declination (J2000) \\
\texttt{type} & & Optical morphological model: PSF, REX, DEV, EXP or COMP\\
\texttt{pstar} & & $P_{\text{star}}$ - star likelihood based on colours from GMM star-QSO classification (Valid only for PSF sources) \\
\texttt{gmmcomp} & & Best-matching GMM component used to derive photo-$z$ estimate\\
\texttt{zphot} & & Photo-$z$ estimate (mean of the normally distributed photo-$z$ posterior) \\
\texttt{zphot\_err} & & Uncertainty on photo-$z$ estimate (standard deviation of the normally distributed photo-$z$ posterior) \\
\texttt{flag\_clean} & & Photometry reliability flag, $=1$ for sources free of blending or imaging artefacts (Eq.~\ref{eq:blend} and \ref{eq:mask}).\\
\texttt{flag\_qual} & & Photo-$z$ reliability flag, 1 for sources expected to have well-constrained estimates (\texttt{flag\_clean} and satisfying Eq.~\ref{eq:err_cut} and \ref{eq:pstar_cut}).\\
\hline
\end{tabular}\\
\end{table*}
\setlength\tabcolsep{3pt}
\begin{table*}
\caption{Example of the LS DR8 photo-$z$ catalogues produced in this work, showing the first 10 entries of the North catalogue. Column descriptions are as outlined in Table~\ref{tab:cat_example}.}\label{tab:cat_example}
\footnotesize
\begin{tabular}{ccccccccccccc}
\hline
\texttt{id} & \texttt{release} & \texttt{brickid} & \texttt{objid} & \texttt{ra} & \texttt{dec} & \texttt{type} & \texttt{pstar} & \texttt{gmmcomp} & \texttt{zphot} & \texttt{zphot\_err} & \texttt{flag\_clean} & \texttt{flag\_qual} \\
\hline
8001655327000001 & 8001 & 655327 & 1 & 134.491665 & 78.375823 & PSF & 0.056 & P3 & 1.099 & 0.202 & 1 & 1 \\
8001655327000002 & 8001 & 655327 & 2 & 134.485347 & 78.376252 & REX & 0.005 & R4 & 0.940 & 0.129 & 1 & 1 \\
8001655327000003 & 8001 & 655327 & 3 & 134.494718 & 78.376295 & REX & 0.006 & R8 & 0.949 & 0.148 & 1 & 1 \\
8001655327000004 & 8001 & 655327 & 4 & 134.533553 & 78.375083 & PSF & 0.225 & P1 & 0.808 & 0.272 & 0 & 0 \\
8001655327000006 & 8001 & 655327 & 6 & 134.517881 & 78.377070 & PSF & 0.002 & P0 & 0.676 & 0.702 & 0 & 0 \\
8001655327000007 & 8001 & 655327 & 7 & 134.516587 & 78.376066 & PSF & 0.001 & P1 & 0.715 & 0.228 & 0 & 0 \\
8001655327000008 & 8001 & 655327 & 8 & 134.488446 & 78.377107 & REX & 0.001 & R5 & 0.909 & 0.242 & 1 & 1 \\
8001655327000013 & 8001 & 655327 & 13 & 135.463593 & 78.375506 & REX & 0.003 & R0 & 0.927 & 0.252 & 1 & 1 \\
8001655327000021 & 8001 & 655327 & 21 & 134.586454 & 78.376363 & REX & 0.000 & R4 & 0.515 & 0.079 & 1 & 1 \\
8001655327000023 & 8001 & 655327 & 23 & 134.575900 & 78.377192 & DEV & 0.003 & D3 & 0.414 & 0.098 & 1 & 1 \\
\hline
\end{tabular}
\end{table*}
\subsection{LS Data Release 9 and future prospects}
The photo-$z$ catalogues produced in this work are based on DR8 of the LS optical catalogues, the latest available at the point predictions for the full optical samples commenced (Section~\ref{sec:catalogues}).
Subsequent to this analysis, Legacy Surveys DR9\footnote{\href{https://www.legacysurvey.org/dr9/}{https://www.legacysurvey.org/dr9/}} was made public.
While LS DR9 does not include significant new optical imaging, it does include additional NEOWISER observations that increase sensitivity in WISE bands as well incorporating several improvements to the pipeline processing.
Future applications of the method to LS DR9 therefore offers the potential for additional small improvements in photo-$z$ quality and greater sample sizes.
Independent of any changes or expansion to the underlying photometric catalogues available, the commencement of a new generation of galaxy and cosmology spectroscopic surveys means that spectroscopic training samples available for photo-$z$ estimation will grow rapidly in the coming years.
Most relevant in this landscape is the DESI survey \citep{2016arXiv161100036D}, which will provide both magnitude selected samples at low redshift and colour selected samples of cosmological tracers out to higher redshift (including emission line galaxies at $z\sim1$ and QSOs at $z > 2$).
The WEAVE-LOFAR survey \citep{Smith:2016vw} will also soon provide spec-$z$ measurements for $>10^{5}$ radio continuum selected sources, providing training samples hugely complementary to those derived from optical selection.
Additionally, surveys in the southern hemisphere with the forthcoming the 4-metre Multi-Object Spectroscopic Telescope \citep[4MOST;][]{2019Msngr.175....3D} will provide large spectroscopic samples across a broad range of galaxy and AGN types, including dedicated follow-up of eROSITA selected X-ray sources will further expand the available training samples in the coming years \citep{2019Msngr.175...42M}.
There is therefore potential for significant improvements in the precision and reliability of photo-$z$ estimates for the LS datasets (and future large area optical surveys).
The methodology outlined in this work is well suited to benefit from these new training samples, with the GMM population division and CSL weighting steps able to maximise the impact of training samples in new or previously under-sampled regions of parameter space.
As training samples increase, the optical parameter space can be further divided (or better modelled) with GMMs, either with additional components or extension to higher dimensions, while still providing sufficient training and test samples for \textsc{GPz}.
Further optimisation of the photo-$z$ methodology itself is also still possible, for example through dynamic allocation of additional GP basis functions to GMM components with significantly larger training samples (reallocating from models with significant over-fitting found from the validation sample).
\section{Summary and conclusions}\label{sec:summary}
In this paper we present new photometric redshift (photo-$z$) estimates derived for the DESI Legacy Imaging Surveys Data Release 8 (LS DR8), covering almost the full extra-galactic low-extinction sky ($\sim 19\,400$\,deg$^{2}$).
By design, our photo-$z$ estimation methodology aims to produce robust photo-$z$ predictions for all populations present in the optical catalogues, ensuring that predictions for rare but scientifically valuable populations are given comparable weight to the more numerous populations that can typically dominate traditional empirical (or machine-learning) photo-$z$ training procedures.
Our method employs Gaussian mixture models (GMMs) derived from the colour, magnitude and size properties of both the training and full optical samples.
Using these purely data-driven models, we are able to divide the observed population into different regions of parameter space and to weight the training samples based on the relative density of the training and full populations.
The sparse Gaussian processes redshift code, \textsc{GPz}, is then used to derive photo-$z$ estimates for individual regions of observed parameter space, including cost-sensitive learning weights derived from the GMMs to mitigate against biases in the spectroscopic training sample.
For unresolved optical sources (PSF morphologies in the LS DR8 catalogue), separate GMMs are used to calculate tentative star-QSO classifications based on literature spectroscopic training samples - ensuring that the primary GMM based population division and weights are not biased by contamination from stars in the optical sample.
Subsequent to photo-$z$ prediction, we perform an additional calibration step on the predicted uncertainties, significantly improving the population averaged accuracy of the uncertainty estimates.
In comparison to other photo-$z$ predictions available in the literature for the same optical population, we find that our photo-$z$ estimates offer substantially improved reliability and precision at $z > 1$, with negligible loss in accuracy for brighter, resolved populations at $z < 1$.
Examining the photo-$z$ predictions for key sub-populations of the full optical sample, we find that our photo-$z$s are able to provide extremely high quality predictions for some of the rarest populations, including QSOs out to $z > 6$ and luminous star-forming galaxies and AGN selected via deep radio continuum or X-ray observations.
With appropriate quality cuts, our photo-$z$ predictions for X-ray and radio continuum selected populations can be used over a wide range in parameter space - with low robust scatter ($\sigma_{\text{NMAD}} < 0.02 - 0.10$) and outlier fraction ($\textup{OLF}_{0.15} <10\%$) at $z < 1$ across a broad range of X-ray or radio continuum flux (densities).
The photo-$z$ catalogues provided in this work therefore offer enormous potential value for new generations of all-sky X-ray (eROSITA) and radio continuum surveys (LOFAR, ASKAP), whilst still offering extremely high precision and reliability for broader optically selected samples.
Alongside the photo-$z$ prediction and associated uncertainty for all $1.6\times10^{9}$ optical sources processed in this analysis, we provide a simple photo-$z$ reliability flag that incorporates photometric quality flags, source contamination, star-QSO probability and photo-$z$ precision.
Combining both LS DR8 North and South datasets, the catalogues provided in this work offer high quality photo-$z$ predictions for $>9\times10^{8}$ galaxies across the bulk of the extragalactic sky, making it one of the most extensive samples of redshift estimates ever produced.
\section*{Acknowledgements}
The author thanks the referee for their feedback and contributions to improving the manuscript, Philip Best for valuable feedback on various stages of the work and Peter Hatfield and Matt Jarvis for helpful discussions early in the development of the methodology.
The author acknowledges funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement No. 892117 (HIZRAD).
The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID \#2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop. ID \#2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall z-band Legacy Survey (MzLS; Prop. ID \#2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF's NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. The Legacy Surveys project is honored to be permitted to conduct astronomical research on Iolkam Du'ag (Kitt Peak), a mountain with particular significance to the Tohono O'odham Nation.
NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A\&M University, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cientifico e Tecnologico and the Ministerio da Ciencia, Tecnologia e Inovacao, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energeticas, Medioambientales y Tecnologicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenossische Technische Hochschule (ETH) Zurich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ciencies de l'Espai (IEEC/CSIC), the Institut de Fisica d'Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig Maximilians Universitat Munchen and the associated Excellence Cluster Universe, the University of Michigan, NSF's NOIRLab, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A\&M University.
BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program "The Emergence of Cosmological Structures" Grant \# XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant\# 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant \# 11433005).
The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration.
The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH1123, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.
\section*{Data Availability}
The data underlying this article are available through the \href{https://www.legacysurvey.org/dr8/}{DESI Legacy Imaging Surveys consortium website}, as well as through the \href{https://datalab.noirlab.edu}{NOIRLab Astro Data Lab} service.
All catalogues produced in this work will be made available through \href{https://vizier.cfa.harvard.edu/viz-bin/VizieR}{the VizieR repository} and \href{https://datalab.noirlab.edu}{NOIRLab Astro Data Lab}.
\bibliographystyle{mnras}
| df95b47d6df458994d2cf2539edf6e3c7a9f9f3c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The Transiting Exoplanet Survey Satellite (TESS; \citealt{Ricker_TESS}) has vastly increased the number of planets available for detailed characterization; more than 5000 TESS Objects of Interest (TOI; \citealt{Guerrero2021}) have now been detected \footnote{exofop.ipac.caltech.edu/tess}. TOIs are candidate or confirmed transiting planets. With the launch of JWST and increasing numbers of space and ground based measurements of exoplanet atmospheres (e.g. \citealt{Benneke2019,DiamondLowe2020}), exoplanet science is moving beyond discovery into detailed characterization. Exoplanet characterization efforts require a careful treatment of starspots and flares to avoid false positives in atmospheric line or bio-signature detections \citep{Barclay2021, Luger2015} and to probe non-equilibrium atmospheric states resulting from high rates of flaring \citep{Chen2021}.
\par Stellar flares occur when magnetic reconnection in the stellar atmosphere accelerates charged particles into the photosphere, heating the plasma and releasing energy across the electromagnetic spectrum. Low mass stars can remain active for $\sim$10 Gyr \citep{France2020}, frequently emitting superflares 10-1000$\times$ larger than flares from our Sun \citep{Howard2018}. Superflares may impinge on the habitability of terrestrial planets orbiting in the liquid-water habitable zones (HZ) of these stars \citep{tarter2007}, although it remains uncertain if they are a net positive for surface life. A high flare rate may photo-dissociate atmospheric volatiles such as ozone, allowing lethal doses of UV surface radiation \citep{Tilley2019}. On the other hand, the UV energy needed for pre-biotic chemistry on planets orbiting low mass stars can only come from flares \citep{Ranjan2017}.
\par Although flares impact both exoplanet characterization and habitability, no comprehensive survey of the flare rates of the TOI catalog exists. The impact of flaring on each confirmed TOI has been assessed largely on a case-by-case basis. Usually, discovery papers acknowledge whether flares were observed by TESS or not but do not measure the flare frequency distribution or place upper limits on the flare rate. As a result, $\sim$15 TOI host stars have measured flare rates (e.g. \citealt{Gunter2020, Pope2021, Gilbert2021, Bogner2021, Colombo2022arXiv}). Most notably, \citet{Medina2020} performed a flare search on all single low mass stars in the southern sky within 15 pc. By employing injection and recovery testing, they measured flare rates or placed upper limits for 5 of the most promising TOIs for future characterization.
\par Most of the sky has now been observed continuously for at least 2 months during the TESS Prime and Extended Mission, sufficient to place strong upper limits on the flare rates of low mass TOIs. We carry out the first comprehensive search for flares for all 2250 non-retired TOIs with 2 min light curves (Fig. \ref{fig:flaring_cartograph}). Non-retired TOIs are those that have not been classified as false positives by the TESS mission. Following \citet{Medina2020}, we use injection testing to obtain upper limits on the flare rates of all non-flaring TOIs in order to determine ``worst case scenarios" for atmospheric characterization.
\begin{figure*}
\centering
{
\includegraphics[trim= 0 0 0 0, width=0.92\textwidth]{TOI_survey_overview.png}
}
\caption{Flare rates or upper limits of all 2 min cadence non-retired TOIs. Upper limits are obtained with flare injection and recovery testing. The colorbar gives the number of large flares yr$^{-1}$, defined here as the rate of flares with an ED$\geq$100 sec. We use ED instead of energy in the color scaling to visually account for the differing contrasts of the same size flare for low mass and high mass stars.}
\label{fig:flaring_cartograph}
\end{figure*}
\section{Building the TOI Flare Catalog}\label{TOIsample}
TESS searches the entire sky for both transiting exoplanets and astrophysical variability in a tiling-based approach. The sky is split into 24$\times$96 deg$^{2}$ sectors, which are each observed for 28 days at a time with four on-board 10.5 cm telescopes in a red (600-1000 nm) bandpass. A given sector is therefore revisited with a $\sim2$ year cadence, although overlaps between sectors can increase coverage. Targets in the Continuous Viewing Zone are observed with almost unbroken coverage for a year. Since 2018, TESS has identified more than 5000 exoplanet candidates, or TOIs. We select all TOIs with 2 min photometry for our flare analysis, except for those TOIs with a TESS disposition indicating known false positives (FP). TOIs from 10-30 min cadence light curves are excluded due to a need to resolve flares. In sum, we download Simple Aperture Photometry (SAP) light curves for 2250 TOI planet candidates and 2096 host stars meeting these criteria from the Mikulski Archive for Space Telescopes (MAST).
\par Long-term variability is carefully removed from each light curve before searching for flares. First, the period of variability is estimated from a Lomb-Scargle periodogram of the light curve. The highest peak with a signal/noise ratio (S/N) above 50 is selected as the best period; otherwise the period is set to $\sim$2 d to ensure any out-of-flare variability at longer timescales is removed. Not all long-term variability is strictly periodic. Next, we fit and remove the out-of-flare variability using a Savitzky–Golay (SG) filter with a window size determined by the estimated variability period as described in \citet{Howard_MacGregor2021}. An example of this fitting process is shown in Fig. \ref{fig:methods_overview_fig}. Light curves are then converted into fractional flux units as in \citet{hawley2014}.
\par Flares are identified by an automated algorithm described in \citet{Howard_MacGregor2021} as $\geq$4.5$\sigma$ excursions above the photometric scatter in the detrended light curve. Flare start and stop times are given as the first and last timestamps with fluxes more than 1$\sigma$ above the noise and are subsequently adjusted by eye. Candidates are inspected for FPs. The equivalent duration (ED) of each flare is obtained by integrating the fractional flux between the start and stop times in seconds. The quiescent luminosity of each star $Q_0$ is computed in $T$ in erg $s^{-1}$ using the $T$=0 to flux calibration \citep{Sullivan2015}, the stellar distance, and the $T$ mag using values primarily from the TESS Input Catalog \citep{Bailer_Jones2018, Stassun2019}. Bolometric flare energies are computed as $E_\mathrm{bol} = Q_0\times$ED$\times f_\mathrm{bol}^{-1}$ assuming a canonical $\sim$9000 K flare with a $T$-$E_\mathrm{bol}$ conversion factor of $f_\mathrm{bol}$=0.19 \citep{Howard_MacGregor2021}. In all, we observe 697 flares from 87 unique host stars to 93 TOIs.
\begin{figure*}
\centering
{
\includegraphics[trim= 0 0 0 0, width=0.99\textwidth]{brief_flare_methods.png}
}
\vspace{-0.3cm}
\caption{Top: Example TOI light curve illustrating the fitting of non-flare variability, flare detections, and the sometimes close occurrence in time of flares and transits. Bottom: Example FFDs of TOIs illustrating the cumulative flare rates and uncertainties of each individual energy y-value, the best fit in red, and fit errors in navy. The \citet{Tilley2019} ozone depletion and abiogenesis regions of \citet{Rimmer2018, Ducrot2020} are over-plotted.}
\label{fig:methods_overview_fig}
\end{figure*}
\subsection{Flare rates for each TOI}\label{tess_FFDs}
Flare emission from the sun and other stars is described by a power law in which higher energy flares are emitted less frequently than lower energy flares. Flare frequency distributions (FFDs) are computed for each TOI given the cumulative rate at which flares of energy $E$ or greater are observed per day and the total observing time. FFDs are fit in log-log space in the usual way:
\begin{equation}
\log{\nu}=(1-\alpha)\log{E} + \beta
\end{equation}
Here, $\nu$ is the number of flares with an energy greater than or equal to $E$ erg per day, 1-$\alpha$ describes the frequency at which flares of various energies occur, and $\beta$ determines the overall rate of flaring. We calculate the uncertainty in the cumulative occurrence rate for each energy value with a binomial 1$\sigma$ confidence interval (CI). Rates and uncertainties of the smallest flares are corrected/propagated when below the 100\% completeness fraction as computed below. Previous studies find $\alpha\approx 2$ (e.g. \citealt{Ilin2021a, Medina2020}). We use this information to constrain 1-$\alpha$ when fitting FFDs to stars with less than 5 flares. There is some debate in the literature whether $\alpha=2$ \citep{Medina2020} or depends on factors such as stellar rotation \citep{Seligman2022}. We therefore fit 200 trials for each FFD as shown in Fig. \ref{fig:methods_overview_fig}, varying the y-axis values within their uncertainties and allowing fits across the 1.3$<\alpha<$2.3 literature range (e.g. \citealt{Howard2019,Ilin2021a,Feinstein2022}).
Injection and recovery tests are performed for both flaring and non-flaring TOIs to determine the completeness fraction of flares recovered from each TOI as a function of flare energy following \citet{Glazier2020}. Batches of 100 flares of a given size are randomly injected into separate instances of each light curve and the number of flares recovered is recorded. The process is repeated for increasingly large sets of flares in order to determine when the recovery fraction converges to one. The 10\% completeness fractions (i.e. a proxy for the minimum fraction with sufficient signal to correct for incompleteness), and 100\% completeness levels are reported in Table \ref{table:all_toi_tab}, respectively. Injection testing of the non-flaring TOIs enables upper limits to be placed on their flare rates when combined with the total observation time. First, the upper 1$\sigma$ binomial uncertainty on the number of flares of energy $E_\mathrm{min}$ observed across the total observation time in days is measured. Then the rate of flares per day of $\geq E_\mathrm{min}$ is fit with 200 FFD trials across the full $\alpha$ range. When no flares are observed, we report the number of flares as zero and the FFD fit parameters for the upper limit in Table \ref{table:all_toi_tab}.
\begin{table*}
\renewcommand{\arraystretch}{1.6}
\caption{Catalog of all 2250 TOIs with 2 min cadence TESS data and a non-FP TESS disposition}
\begin{tabular}{p{0.7cm} p{0.5cm} p{0.5cm} p{0.5cm} p{0.6cm} p{1.3cm} p{0.7cm} p{0.7cm} p{0.6cm} p{0.9cm} p{1.1cm} p{0.7cm} p{0.7cm} p{1.3cm}}
\hline
TOI & R$_p$ & TSM & FRM & TFOP & TIC-ID & M$_*$ & $d$ & Bin. & N$_{fl}$ & log $E_\mathrm{100}$ & 1-$\alpha$ & $\beta$ & N$_{super}$ \\
& [$R_\oplus$] & & & & & [$M_\odot$] & [pc] & [bool] & & [erg] & & & [fl. yr$^{-1}$] \\
\hline
134.01 & 1.74 & 135 & 0.94 & Y & 234994474 & 0.59 & 25.1 & 0 & 0 & 33.44 & -0.80 & 25.2 & 3.0 \\
177.01 & 2.13 & 150.0 & 0.932 & Y & 262530407 & 0.52 & 22.5 & 0 & 2 & 33.32 & -0.80 & 25.18 & 4.2 \\
218.01 & 1.83 & 133.0 & 0.980 & Y & 32090583 & 0.26 & 52.3 & 0 & 152 & 33.27 & -1.67 & 54.84 & 4.4 \\
540.01 & 0.97 & 38.0 & 0.998 & Y & 200322593 & 0.17 & 14.0 & 0 & 14 & 32.45 & -1.66 & 52.9 & 0.07 \\
\hline
\end{tabular}
\label{table:all_toi_tab}
{\newline\newline \textbf{Notes.} A subset of the flaring parameters of 2250 non-retired TOIs observed at 2 min cadence. The full table is available in machine-readable form. Columns shown are TOI, planet radius, TSM, FRM, TFOP disposition, TIC ID, stellar mass, stellar distance, a flag whether binarity in Gaia DR2 \citep{Kervella2019} or source contamination is likely present, number of flares observed, smallest energy flare that would be recovered 100\% of the time, the FFD fit parameters 1-$\alpha$, $\beta$, and the rate of 10$^{34}$ erg superflares. Other columns in the machine readable table not shown here include orbital distance, orbital period, coordinates, Gaia ID, observation time, rate of ED=100 sec flares yr$^{-1}$, the energy and ED of 10\% completeness limits, and uncertainties where applicable.}
\end{table*}
\section{Suitability for atmospheric characterization and flare rates}\label{sorted_by_flaring}
TOIs are ranked by flaring on their suitability for follow-up, where the most promising targets for follow-up are low mass stars with low flare rates. Each TOI is assigned a ranking between zero and one, with one being highly-favorable for follow-up and zero very unfavorable. This TOI flare ranking metric (FRM) sorts TOIs by increasing flare rate within distinct groupings organized by stellar mass and distance, with nearby, low mass stars ranked higher than distant, high mass stars. The quiescent brightness of higher mass stars reduces the contrast of a given size flare and induces a flare detection bias toward higher energies from larger stars. Grouping the TOIs by stellar mass minimizes this bias and facilitates comparisons of the relative suitability for follow-up between stars of different masses. Grouping by distance accounts for the increased priority for follow-up of stars in the solar neighborhood. The first several groups are ordered as follows: 0.1-0.3$M_\odot$ late M-dwarfs within 15 pc, 0.3-0.6$M_\odot$ early M-dwarfs within 15 pc, 0.1-0.3$M_\odot$ late M-dwarfs beyond 15 pc, 0.3-0.6$M_\odot$ early M-dwarfs beyond 15 pc, 0.6-0.75$M_\odot$ late K-dwarfs within 36 pc, 0.75-0.9$M_\odot$ early K-dwarfs within 36 pc, 0.6-0.75$M_\odot$ late K-dwarfs beyond 36 pc, 0.75-0.9$M_\odot$ early K-dwarfs beyond 36 pc, and so on.
The grouping order continues to higher-mass stars until all TOIs are included. \citet{Winters2021} gives 15 pc as the demarcation distance between nearby and distant M-dwarfs. We compute an equivalent demarcation distance for the larger mass bins such that the apparent magnitude of stars in each bin is approximately comparable to M-dwarfs observed from 15 pc.
\par Sorting by flare rates/upper limits within each of these mass-distance groupings broadly agrees with our intuitions that a nearby rocky planet orbiting an M-dwarf flare star is less suitable for characterization than a a nearby rocky planet orbiting a non-flaring M-dwarf, while still having a higher priority for follow-up than a planet orbiting a distant G-dwarf from which no flares were observed. The catalog ranking for the top 10\% of targets is shown in Fig. \ref{fig:results_fig} to illustrate the effects of flare rate, stellar mass, and distance on the flare ranking. Specific use cases for our catalog may benefit from sorting the flare rates of the TOIs with different criteria or in a different order, which may be done using the columns provided in the machine-readable version of Table \ref{table:all_toi_tab}. The final flare ranking is generated by splitting the catalog into three categories based on TFOP working group disposition in ExoFOP-TESS. TFOP-confirmed planets and active candidates are placed first, then ambiguous candidates second, and candidates retired as FPs third. Within each TFOP category, the mass, distance, and flare rate ordering is preserved.
\par Once each TOI is assigned its final flare ranking, we cross-correlate the flare catalog with those targets for which a 10 hr JWST program could detect exoplanet atmospheres. Following \citet{Kempton2018}, we compute a transmission spectroscopy metric (TSM) for each TOI in our sample as described by Equation 1 of \citet{Kempton2018}, reproduced below:
\begin{equation}
\mathrm{TSM} = f_s \times \frac{{R_p}^3 T_\mathrm{eq}}{M_p {R_{*}}^2} \times 10^{-m_J/5}
\end{equation}
Here, $f_s$ is the scale factor as computed in Table 1 of \citet{Kempton2018}. $R_p$ is the ExoFOP planet radius, $T_\mathrm{eq}$ is the estimated planet equilibrium temperature, $M_p$ is the estimated planet mass, $R_*$ is the estimated stellar radius \citep{Torres2010, Mann2015}, and $m_J$ is the $J$ magnitude estimated from \citet{Kraus2007}. We refer the reader to \citet{Kempton2018} for details. The TSM scales with the signal-to-noise of exoplanet atmospheric features, assuming a cloud-free atmosphere and 10 hr of JWST observations. \citet{Kempton2018} define a detectable atmosphere to have TSM$>$10 for terrestrial planets and TSM$>$90 for larger planets. Since our inputs to the TSM equation are estimated from ExoFOP, we caution that terrestrial planets just above or below TSM=10 may be consistent with slightly lower or higher values across the cutoff. Highly-ranked TOIs in our flare catalog that also have high TSM values are particularly compelling for follow-up.
\section{Key results of the TOI flare catalog for atmospheric follow-up}\label{key_results_sect}
Across the sample, we find small ($<$1.5$R_\oplus$) TOIs are more likely to orbit flare stars than larger planets as shown in Fig. \ref{fig:results_fig}. This is a result of the high signal to noise of small planets orbiting low mass stars and the oftentimes high flare rates of these stars.
\subsection{Terrestrial planets potentially suitable for follow-up}\label{some_exciting_targets}
M-dwarf planets or likely candidates of $<$1.5$R_\oplus$ flagged as suitable for JWST transmission spectroscopy make up 35 out of 2250 TOIs in the catalog. Of these, nine or 25.7$\substack{+12 \\ -7}$\% orbit stars observed to flare. These TOIs are LHS 3844 b, the L 98-59 system, the LTT 1445 system, TOI 486.01, TOI 540 b, the GJ 1132 system, TOI 1450.01, and TOI 2267.01 as shown in Fig. \ref{fig:results_fig}. Three orbit flare stars with FFDs previously reported from \citet{Medina2020}. The flare rate of LTT 1445 was previously measured by \citet{Howard2019} but the flares likely come from the BC component \citep{Winters2019}. \citet{Bogner2021} fits dozens of FFDs which includes the host stars to LHS 1815 b, the L 98-59 system, TOI 486, and TOI 1450 but does not report FFD fit parameters for these stars.
\par We do not detect flares from the remaining 26 terrestrial TOIs amenable to characterization, but place upper limits on their flare rates in Table \ref{table:all_toi_tab}. More than 90\% of these did not have previous upper limits. Each of the 35 flagged TOIs is in the top 10\% of our flare suitability ranking, and even the most active do not reach the ozone depletion zone of \citet{Tilley2019}, although some disequilibrium chemistry is consistent with the rates/upper limits of several TOIs \citep{Howard2018}.
\subsection{Other flaring and non-flaring TOIs of note}\label{other_exciting_targets}
The flare rates of the remaining TOIs in our sample enable characterization of super-earths, mini-Neptunes, and several planets in the HZ. Terrestrial HZ planets in our sample include TOI 700 d, LHS 1140 b, TOI 715.01, and TOI 4312.01 (Fig. \ref{fig:results_fig}). The first two are among the most exciting planets for large follow-up programs \citep{Gilbert2020_TOI, Wunderlich2021}. We report upper limits on the rate of 10$^{34}$ erg superflare yr$^{-1}$ of 0.25 and 0.71 for these TOIs, respectively. TOI 715.01 is a 1.7$R_\oplus$ super-earth orbiting a likely M-dwarf. It emits no more than 0.15 10$^{34}$ erg superflare yr$^{-1}$. TOI 4312.01 is a 2.4$R_\oplus$ candidate with an upper limit of 4.7 10$^{34}$ erg superflare yr$^{-1}$.
\begin{figure*}
\centering
{
\includegraphics[trim= 0 0 0 0, width=0.99\textwidth]{TOI_flaring_key_results.png}
}
\caption{Top left: All non-retired TOIs orbiting flaring (red) and non-flaring (grey) host stars of $\leq$2M$_\odot$. TOI 1224, DS Tuc Ab \citep{Newton2019_DSTuc}, and TOIs in the HZ of \citep{Kopparapu2013} are noted. Top right: The flare ranking metric of \S \ref{sorted_by_flaring}. Bottom left: All non-retired TOIs around M-dwarfs amenable to transmission spectroscopy, with the same color scheme as above. Note the HZ planets from the top left panel fail the \citet{Kempton2018} TSM criteria for a 10 hr JWST program. Bottom right: smaller TOIs orbit flare stars more frequently than larger TOIs, with 1$\sigma$ binomial CIs.}
\label{fig:results_fig}
\end{figure*}
\section{Conclusions}\label{conclude}
We present the first comprehensive TOI flare survey. A new flare ranking metric is created to aid proposers in selecting planets for further characterization. The metric of every TOI is cross-matched against a S/N metric for a 10 hr JWST program. As a result, we discover that $\leq$1.5$R_\oplus$ TOIs orbit flare stars more frequently than do larger planets, and find 1/4 of all $<$1.5$R_\oplus$ TOIs around M-dwarfs accessible to transmission spectroscopy with JWST orbit flare stars. Even though none of the hosts of terrestrial planets suitable for characterization reach the flares rates needed for complete ozone loss, it has been shown that flare rates a factor of $\sim$10$\times$ lower may still induce significant dis-equilibrium states \citep{Howard2018}. These lower rates are comparable to the flare rates and upper limits of the TOIs presented here. Therefore, spectral modeling of dis-equilibrium chemistry and atmospheric survival will be needed on a case-by-case basis and population level to understand atmospheric signals. In addition to JWST, the upcoming 2029 launch of the Ariel mission to characterize the atmospheric properties of exoplanets \citep{Tinetti2013_Ariel} would benefit from our catalog. Finally, our sample of flaring TOIs provides an excellent dataset for optical counterparts to radio star-planet interactions \citep{Pope2021}.
\section*{Acknowledgements}\label{acknowledge}
We would like to thank the anonymous referee for improving the work. WH would like to thank Meredith MacGregor for helpful conversations on TOI flaring. WH acknowledges funding support by the TESS Guest Investigator Program GO 3174. This paper includes data collected by the TESS mission. Funding for the TESS mission is provided by the NASA Explorer Program. This research has made use of the Exoplanet Follow-up Observation Program website, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program. This work has made use of data from the European Space Agency (ESA) mission {\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia} Data Processing and Analysis Consortium (DPAC, \url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the {\it Gaia} Multilateral Agreement.
{\it Facilities: TESS}
\section*{Data availability}\label{availdata}
The data underlying this article are available in the MAST archive at Space Telescope Science Institute, at https://doi.org/10.17909/t9-nmc8-f686
\bibliographystyle{mnras}
| 368aa2f934e127eab05bcee916c85c8c28967609 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Motivation and significance}
Scanned probe techniques such as scanning tunneling microscopy (STM) and non-contact atomic force microscopy (nc-AFM) have now made direct investigation of on-surface reactions a routine experimental technique \cite{Barth2007, Jelinek2017}. As the diverse ecosystem of organic molecules accessible by these techniques continues to grow so does the need for more sophisticated tools to extract quantitative information from larger and more complicated datasets of molecule imaging.
Extracting statistics from STM images is often done by hand \cite{Capsoni2016, Prinz2015, Stetsovych2016, Goll2022}. This necessarily limits the size and complexity of statistical problems that can be tackled using these datasets. To address this limitation in extracting and categorizing molecules in STM images, we developed an automated scheme utilizing existing image processing libraries written using the Python programming language.
The existing widely used programs for the analysis of scanned probe data (WSxM \cite{Horcas2007} and Gwyddion \cite{Necas2012}) do not presently incorporate tools for automated image processing and feature extraction. ImageJ \cite{Rueden2017} is a powerful and widely used piece of software in the biology community, but lacks compatibility and relevance for the datasets this package is designed to address. Another hindrance is that these image processing tools for biology applications are often written using commercial, closed source MATLAB code. Digital Surf has feature categorization tools but is also commercial and closed source \cite{Cognard2020}.
There has been recent progress in developing machine-learning based tools for the automated assignment of atomic structure and defect information from scanned probe datasets, \cite{Scherbela2017, Ziatdinov2017, Li2021} as well as automated routines for improving the probe condition in STM \cite{Rashidi2018, Krull2020}. However these solutions address comparatively homogeneous datasets of more simple structures compared the present case, or require the computational resources and data (real or artificial) to train a convolutional neural network \cite{Krull2020, Li2021}. We wanted to avoid this approach and instead develop a lightweight tool to quickly identify statistical trends in a large number of STM images with a diverse manifold of initially unknown molecular species; the functions offered in this package were used to prepare Fig. 3 in Hellerstedt \emph{et. al} \cite{Hellerstedt2019}.
Our approach addresses some specific obstacles. There are practical limitations to the data resolution versus the number of species sampled. The sorting method needs account for different adsorption configurations of the same species (\emph{e.g.}, rotations and chirality). While explicit pairwise template matching has advantages in robustness, it suffers from the computational demands scaling in quadrature with the number of molecules to be sorted.
Our solution to these problems relies on the Zernike polynomial basis set to provide a ``fingerprint" of coefficients representing every molecule \cite{Khotanzad1990, Coelho2013}. These coefficients are robust in response to rotations and noise in the data. In conjunction with other physically motivated coefficients, they provide the input to clustering algorithms for sorting the molecules. There are many different algorithms that can be applied to this type of data for the purposes of grouping like kinds of molecules together. We discuss some of the approaches we applied and their relative efficacy.
\section{Software description}
This package consists of a set of functions written using the Python libraries Numpy \cite{Oliphant2007}, scikit-image \cite{VanderWalt2014}, scikit-learn \cite{Pedregosa2011}, Mahotas \cite{Coelho2013} and Matplotlib \cite{Hunter2007}. In practice, it is a series of functions called sequentially as shown in Fig. \ref{flowchart}. We developed the current implementation to be useful `out-of-the-box' with only minimum skill prerequisites required (\emph{e.g.}, installing a python distribution and necessary packages). The modular nature of the data processing flow allows for a great deal of flexibility to address different datasets. The open source, repository based distribution of this package allows for further customization, and the possibility for community based development and improvement on these core functionalities.
\subsection{Software Architecture}
We first read in the image data and apply filters to make the image suitable for molecule template extraction using an adaptive threshold. After identifying closed contours corresponding to the perimeters of each molecule, we use the interior data of each contour as a representative template image. We then extract numerical features from each molecule, consisting of the calculated Zernike moments \cite{Khotanzad1990, Coelho2013} of the template images as well as the maximum topographic heights and the contour perimeter lengths. Finally, we use clustering algorithms \cite{Pedregosa2011} to categorize these features. Further functionality is provided to visualize the sorted categories, and make manual corrections.
\begin{figure}
\centering
\includegraphics[keepaspectratio=true, width=.8\linewidth]{counting_molecules_flowchart}
\caption{Architecture and usage of main functions. Yellow boxes indicate optional user input parameters, blue boxes are visual outputs of the program.}
\label{flowchart}
\end{figure}
Figure \ref{flowchart} shows the typical use of the provided functions. We rely on the Nanonispy library to read in data from one of the prevalent formats (Nanonis SXM files). This generates an array of pixel data and the rescaling factors to convert to real-space distances.
\subsection{Software Functionalities}
The default filtering function provided performs a Gaussian filter and plane fit subtraction of the image data. We calculate the global Otsu threshold \cite{Otsu1979} and use that to scale the offset value for feature extraction using a local thresholding method. We include diagnostic functions that can plot the filtered image data with the extracted contours, labeled by number, as well as a grid view of all the extracted molecule templates.
Zernike moments are the coefficients for representing an image decomposed into the orthogonal Zernike polynomial basis set \cite{Khotanzad1990}. This method has attracted attention and effort in the computer vision community for being rotationally invariant. This property is particularly useful in the context of our application because we wish to match molecules regardless of how they are absorbed on the sample surface. We utilize the Mahotas library for calculating the Zernike moments for each molecule template \cite{Coelho2013}. We set the median template diagonal length as the default value of the Zernike radius input, an assumption that could fail depending on the homogeneity of the templates.
In addition to rotational invariance, these moments are insensitive to translation, mirroring, and rescaling. To account for differences in the real space footprint of each molecule, we additionally calculate the length, as well as the maximum pixel value within each contour.
With these characteristic moments and physical length scales, we perform a clustering analysis using the algorithms available in the Scikit-learn library \cite{Pedregosa2011}. For the datasets studied in this work, we found the Birch algorithm \cite{Zhang1996} with a threshold factor between 0.1 to 0.4 to be most effective at sorting images with no \emph{a priori} knowledge of the number of categories. When the number of sorting categories is known, hierarchical clustering was found to be more accurate. The most effective sorting was accomplished using affinity propagation, where the cluster center preferences were defined using a hand-selected set of exemplar molecules.
The invariance of the Zernike moments to mirror symmetry in particular, makes them insensitive to differentiating between chiral molecules. Absolute quantifications of chirality are surprisingly difficult to define \cite{Buda1992}. We developed a function to do a pairwise comparison of molecules within each sorted category. By comparing each molecule and its mirror image, they can be sorted into right- and left- handed categories.
All these image, contour, correlations, and categorization data can be exported/ saved for later use. Also included is a function using the interactive features of Matplotlib to manually categorize all the molecules.
\section{Illustrative Examples}
\begin{figure}
\centering
\includegraphics[keepaspectratio=true, width=1\linewidth]{APT_044}
\caption{9-azidophenanthrene molecules on Ag(111) sorted into 6 groups using hand-selected exemplars. 80\x80 nm, 1024\x1024 pixels.}
\label{APT-044-counted}
\end{figure}
\begin{figure}
\centering
\includegraphics[keepaspectratio=true, width=1\linewidth]{APT_111}
\caption{9-azidophenanthrene molecules on Ag(111) sorted into 9 groups using hand-selected exemplars. 50\x50 nm, 512\x512 pixels.}
\label{APT-111-counted}
\end{figure}
\begin{figure}
\centering
\includegraphics[keepaspectratio=true, width=1\linewidth]{Helicene_008}
\caption{Helicene data sorted using hand-selected exemplars. The dimers (categories 1 and 4) were subsequently sorted by chirality. 80\x80 nm, 1024\x1024 pixels. Compare to Fig. 4 of Stetsovych \emph{et. al}. \cite{Stetsovych2016}.}
\label{helicene-counted}
\end{figure}
Representative outputs for the three example datasets from hand-selected exemplar molecules are shown in Figs. \ref{APT-044-counted}, \ref{APT-111-counted} and \ref{helicene-counted}. The work on the 9-azidophenanthrene system which was the source of the data in Figs. \ref{APT-044-counted} and \ref{APT-111-counted} provided the original impetus to develop this package \cite{Hellerstedt2019}. The helicene data shown in Fig. \ref{helicene-counted} has been previously described in the literature \cite{Stetsovych2016}. These data were acquired via scanning tunneling microscopy at low temperature ($\sim$ 5~K), with a pixel density of 10 nm$^{-1}$ to 13 nm$^{-1}$.
\section{Expected Impact}
This is the first effort to our knowledge to create an open source scheme to automate the task of counting and sorting hundreds of molecules, specifically tailored for scanning probe microscopy datasets. We expect this package to be immediately useful for anyone tasked with extracting population statistics from images comparable to those included as examples. This tool has already proven useful by making it easy to quickly extract and visualize population statistics in image datasets. By scaling the statistics and quantitative information that can be extracted from these types of datasets, we anticipate that this will facilitate the ability to answer novel and more nuanced questions about how chemical processes unfold on a surface. We hope to reveal and address limitations by having this set of functions applied to the increasingly diverse ecosystem of on-surface chemistry being explored by the greater scanning probe community.
\section{Conclusions}
Here we have presented our efforts to automate the counting and sorting of molecules in topographic images acquired with a scanning tunneling microscope. The example datasets, images with several hundred molecules, can be sorted on a personal computer in seconds. The python based, open source, modular design of the components strives to be immediately accessible to a non-expert user, and allow for significant modification/ customization to suit individual needs. We hope that use and feedback from the wider community will allow this tool to continue to be developed and provide utility for advancing the understanding of on-surface chemistry experiments.
\section{Software/ Data availability}
\noindent This package is available on GitHub and archived at:\newline
\url{https://doi.org/10.5281/zenodo.6324850}\newline
The data in Figures \ref{APT-044-counted}, \ref{APT-111-counted}, \ref{helicene-counted} is available on figshare: \newline
\url{https://doi.org/10.6084/m9.figshare.19217556}
\section*{Acknowledgements}
S. Edalatmanesh and A. Caparr\'{o}s provided useful discussion in developing the solution to the chirality sorting problem. We thank Dr. Karth{\"a}user for critical reading of this manuscript. J.H. acknowledges the support of the CIES/ Czech Fulbright commission during the development of this package.
\bibliographystyle{elsarticle-num}
| 7faad185537b32a4988e2f57a574c3aaaa4975c4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
Active galactic nuclei (AGN) are amongst the most luminous objects
(L = 10$^{42}$ $-$ 10$^{46}$ erg s$^{-1}$; \citealt{1999PNAS...96.4749F})
in the Universe that emit radiation over a wide range of wavelengths. They are believed to be powered by
accretion of matter onto SMBH (10$^5$ $-$
10$^9$ $M_{\odot}$) situated at the center of galaxies \citep{1984ARA&A..22..471R,2008ApJ...689..762D}. The SMBH is generally supposed to be surrounded by an optically thick
and geometrically thin accretion disk \citep{1973A&A....24..337S}.
The observed X-ray emission from the nuclear region of the radio-quiet
category of AGN is believed to be produced by inverse Compton process, caused
by the interaction of the seed ultra-violet (UV) photons from the accretion disk with the
thermal electrons in a hot ($\sim$10$^{8-9}$ K) region called the corona that is situated close to the
accretion disk \citep{1991ApJ...380L..51H,1994ApJ...432L..95H}.
This X-ray continuum gets reprocessed in the accretion disk giving
rise to the reflection hump at around 15$-$30 keV and also the
broad FeK$\alpha$ line at 6.4 keV \citep{1991MNRAS.249..352G, 1993MNRAS.262..179M}. Soft excess between
0.1 $-$ 2 keV is ubiquitously observed in Type I AGN \citep{1998MNRAS.301..179M, 2002MNRAS.331L..35F, 2006MNRAS.365.1067C,2009A&A...495..421B,2020MNRAS.491..532G}, although the physical origin of this component remains highly debated \citep{2019ApJ...871...88G, 2021ApJ...913...13X}, different analyses showed that a two temperature Comptonization process agrees well with such a component either from an observational (e.g. \citealt{2012MNRAS.420.1825J, 2018A&A...609A..42P,2018A&A...611A..59P,2020A&A...640A..99M,2020MNRAS.497.2352M}) or theoretical (e.g. \citealt{2015A&A...580A..77R, 2020A&A...634A..85P, 2020MNRAS.491.3553B, 2020MNRAS.496.4255B}) point of view. Analysis of these spectral features (reflection, FeK$\alpha$ line, soft excess) will help in providing
strong constraints on the nature of the X-ray emitting region.
From X-ray reverberation studies \citep{2009Natur.459..540F,2012MNRAS.422..129Z}
AGN corona is believed to be a compact region situated above the
accretion disk typically within 3$-$10 $R_G$, where $R_G$ is the gravitational
radius and it is defined as, $R_G = GM_{BH}/c^2$, here M$_{BH}$ is the SMBH
mass and $G$ is the gravitational constant. However, there are strong debates concerning the geometry of the corona. The lamp post is one such possibility but other models also exist (e.g. \citealt{1994ApJ...432L..95H, 2012MNRAS.420.1848D, 2013A&A...549A..73P}).Also, rapid X-ray flux variability studies
\citep{2005MNRAS.359.1469M}, the observed small time scales of
X-ray eclipses \citep{2005ApJ...623L..93R,2011MNRAS.410.1027R}
and microlensing studies \citep{2009ApJ...693..174C} point to the small size
of the X-ray corona of 5$-$10 $R_G$.
The observed shape of the X-ray continuum
can be described by a power law with an exponential cut-off ($\rm{E_{cut}}$),
and the spectral shape depends on the optical depth ($\tau$),
temperature of the coronal plasma ($\rm{kT_{e}}$), seed
photon temperature and the viewing angle. From a study of the Seyfert galaxy NGC 5548, \cite{2000ApJ...540..131P} showed the existence of an approximate relation between $\rm{E_{cut}}$ and $\rm{kT_{e}}$ as $\rm{E_{cut}}$ = 2$-$3 $\rm{kT_{e}}$. On analysis of a sample of Seyfert galaxies, according to \cite{2001ApJ...556..716P}, for an optically thin corona with $\tau$ $<$ 1; $\rm{E_{cut}}$ $\approx$ 2 $\rm{kT_{e}}$, while for an optically thick corona with $\tau$ $>$ 1, $\rm{E_{cut}}$ $\approx$ 3 $\rm{kT_{e}}$. However, by fitting Comptonized spectra simulated using a range of $\tau$ and $\rm{kT_{e}}$ with a power law with an exponential cut-off model, \cite{2019A&A...630A.131M}, showed that the commonly adopted relation of $\rm{E_{cut}}$ = 2$-$3 $\rm{kT_{e}}$ is not valid for all values of $\tau$ and $\rm{kT_{e}}$, instead valid for only low values of $\tau$ and $\rm{kT_{e}}$.
\begin{table*}
\caption{Details of the sources analyzed in this work. The columns are (1) name of the source, (2) right
ascension (h:m:s), (3) declination (d:m:s), (4) redshift, (5) galactic hydrogen column density $\rm{N_H}$ in units of $10^{22}$ atoms $\rm{cm^{-2}}$ obtained from \cite{2013MNRAS.431..394W}, (6) type of the
source, (7) observation ID, (8) epoch, (9) date of observation, and (10) exposure
time in sec. The information such as the right ascension, declination and $z$ are from \cite{2010A&A...518A..10V}. The OBSIDs that are analysed for the first time are shown in bold.} \label{table-1}
\centering
\begin{tabular}{lllccclclc}
\hline
Name & $\alpha_{2000}$ & $\delta_{2000}$ & $z$ & $\rm{N_H}$ & Type & OBSID & Epoch & Date & Exposure \\
\hline
NGC 3227 & 10:23:30.60 & $+$19:51:56 & 0.004 & 0.021 & Sy1.5 & 60202002002 & A & 09-11-2016 & 49800 \\
& & & & & & 60202002004 & B & 25-11-2016 & 42457 \\
& & & & & & 60202002006 & C & 29-11-2016 & 39685 \\
& & & & & & 60202002008 & D & 01-12-2016 & 41812 \\
& & & & & & 60202002010 & E & 05-12-2016 & 40887 \\
& & & & & & 60202002012 & F & 09-12-2016 & 39277 \\
& & & & & & 60202002014 & G & 21-01-2017 & 47602 \\
& & & & & & 80502609002 & H & 15-11-2019 & 28782 \\
& & & & & & 80502609004 & I & 05-12-2019 & 27690 \\
NGC 5548 & 14:17:59.53 & $+$25:08:12 & 0.017 & 0.017 & Sy1.5 & 60002044002 & A & 11-07-2013 & 24096 \\
& & & & & & 60002044003 & B & 12-07-2013 & 27272 \\
& & & & & & 60002044005 & C & 23-07-2013 & 49521 \\
& & & & & & 60002044006 & D & 10-09-2013 & 51460 \\
& & & & & & 60002044008 & E & 20-12-2013 & 50102 \\
& & & & & & 90701601002 & F & 26-01-2021 & 38719 \\
MR 2251$-$178 & 22:54:05.90 & $-$17:34:55 & 0.064 & 0.027 & Sy1.5 & 60102025002 & A & 18-05-2015 & 23112 \\
& & & & & & 60102025004 & B & 17-06-2015 & 23185 \\
& & & & & & 60102025006 & C & 10-11-2015 & 20588 \\
& & & & & & 60102025008 & D & 11-12-2015 & 21707 \\
& & & & & & 90601637002 & E & 16-12-2020 & 23620 \\
\hline
\end{tabular}
\end{table*}
Observations from high energy X-ray missions such as {\it CGRO}
\citep{2000ApJ...542..703Z,1997ApJ...482..173J},
{\it BeppoSAX} \citep{2000ApJ...536..718N,2007A&A...461.1209D},
{\it INTEGRAL} \citep{2014ApJ...782L..25M,2010MNRAS.408.1851L,2016MNRAS.458.2454L,
2011A&A...532A.102R}, {\it Swift}-BAT \citep{2013ApJ...770L..37V,2017ApJS..233...17R};
and {\it Suzaku} \citep{2011ApJ...738...70T} showed that the corona in Seyfert galaxies has a wide range of temperature with $\rm{E_{cut}}$ ranging from 50 $-$
500 keV. However, observations from those missions are limited to bright and
nearby sources. Thus, it is very clear that many efforts were put to measure
$\rm{E_{cut}}$ in the X-ray spectra of several AGN. However, a major transformation
in the studies of the Comptonization spectrum of AGN to determine $\rm{E_{cut}}$
from an epoch of observation happened after the launch of
the {\it Nuclear Spectroscopic Telescope Array} ({\it NuSTAR}; \citealt{2013ApJ...770..103H}) in the year 2012, owing to its broad spectral coverage of 3$-$79 keV and its high sensitivity
beyond 10 keV. Since the launch of {\it NuSTAR}, $\rm{E_{cut}}$ values were
obtained for many AGN \citep{2017MNRAS.467.2566F,2018A&A...614A..37T,2018JApA...39...15R,2018ApJ...856..120R,2019MNRAS.484.5113R,2019ApJ...875L..20L,2020ApJ...905...41B, 2021MNRAS.500.1974R,2021MNRAS.502...80K}. Importantly, in addition to the determination of $\rm{E_{cut}}$ values
(and thereby constraining $\rm{kT_{e}}$), in less than half-a-dozen
sources there are also reports of variation in the $\rm{E_{cut}}$ values pointing
to variations in $\rm{kT_{e}}$.
For example in seven sources, namely MCG-5-23-16 \citep{2017ApJ...836....2Z}, 3C 382
\citep{2014ApJ...794...62B}, NGC 4593 \citep{2016MNRAS.463..382U},
NGC 5548 \citep{2015A&A...577A..38U}, Mrk 335 \citep{2016MNRAS.456.2722K}, NGC 3227 and SWIFT J2127.4+5654 \citep{2021MNRAS.502...80K} variations in the $\rm{E_{cut}}$ values are available in the literature. Recently,
from a reanalysis of the {\it NuSTAR} spectra of five sources using a model
independent approach, \cite{2018ApJ...863...71Z}
confirmed the $\rm{E_{cut}}$ variation in three of the five sources
namely 3C 382, NGC 5548 and Mrk 335. Most of these inferences were based on variations in $\rm{E_{cut}}$
obtained from phenomenological model fits to the data. However, to know changes in
$\rm{kT_{e}}$ it is imperative to fit physical models to the data, because, it is known
recently, that the relation $\rm{E_{cut}}$ = 2-3 $\rm{kT_{e}}$ does not always hold true \citep{2019A&A...630A.131M}.
Therefore it is imperative to fit physical model fits to the observed spectra to get $\rm{kT_{e}}$.
Though $\rm{E_{cut}}$ is known to vary, we do not yet know the causes for its variation. Despite that, it is important to increase the number of sources that show variation in the temperature of the corona. This is now achievable
owing to the multiple-epochs of observation available on a large number of AGN in
the {\it NuSTAR} archives\footnote{https://heasarc.gsfc.nasa.gov/cgi-bin/W3Browse/w3browse.pl}. The primary motivation here is, therefore to increase
the number of AGN that show variation in $\rm{kT_{e}}$. We are in the
process of a careful and systematic investigation of
$\rm{kT_{e}}$ variation in a large number of AGN.
Here, we present the results from the multiple-epoch spectral analysis of
three AGN, namely NGC 3227, NGC 5548 and MR 2251$-$178. This also includes few new observations not published so far.
NGC 3227,at a redshift of z = 0.004 and powered by a black hole of mass 4.79 $\times$ $10^{6}$ $M_{\odot}$ \citep{2015PASP..127...67B}, has been extensively studied in the X-ray band. Signatures of warm absorbers are evident in this source from observations with {\it ASCA}
\citep{1994ApJ...435..106N,1998ApJ...509..146G}, {\it ROSAT} \citep{1997A&A...327..483K}
and {\it XMM-Newton} \citep{2009ApJ...691..922M}. In the {\it XMM-Newton} observations, FeK$\alpha$ line
was evident \citep{2009ApJ...691..922M}. It has also been recently
studied by \cite{2020MNRAS.494.5056L} for flux variations
combining {\it XMM-Newton} and {\it NuSTAR} observations. It has complex
absorption features, which are also variable \citep{2018MNRAS.481.2470T}. Recently \cite{2021A&A...652A.150M} has reported the broad band spectral modeling of the source using multi-wavelength data from {\it XMM-Newton}, {\it NuSTAR} and {\it Hubble Space Telescope} {\it (HST)}.
NGC 5548 is a galaxy located
at $z$ = 0.017 with a black hole of mass
5.0 $\times $ 10$^7$ $M_{\odot}$ \citep{2015PASP..127...67B}. It has been
extensively studied in the X-ray band using data from various satellites
and has also been found to be strongly absorbed in
soft X-rays \citep{2016A&A...592A..27C, 2015A&A...577A..38U, 2015A&A...575A..22M,2014Sci...345...64K}.
MR 2251$-$178, with a black hole mass of 2.0 $\times $ 10$^8$ $M_{\odot}$ \citep{2009AJ....137.3388W}, first discovered by its strong X-ray emission
\citep{1978Natur.271...35R} was found to be a low redshift AGN at
$z$ = 0.06 \citep{1983MNRAS.202..125B}.
From VLA observations, \cite{1990ApJ...356..389M}
found the source to show weak radio emission
with elongated morphology resembling a FRI source.
Details of the observations and
reduction of the data are presented in Section 2, analysis is presented
in Section 3, results and discussion are presented in Section 4
followed by the summary in the final section.
\section{Observations and Data Reduction}
\begin{table*}
\caption{Results of the fit to the spectra of NGC 3227. The models are Model I: {\it const*TBabs*zTBabs(zpo+zgauss)}, Model II: {\it const*TBabs*zTBabs(pexrav+zgauss)}, Model III: {\it const*TBabs*zTBabs(xillver)} and Model IV: {\it const*TBabs*zTBabs(xillverCP)}. The {\it zTBabs} component is added with all the models to fit the epoch G, H and I spectra. The fluxes are in units of 10$^{-10}$ erg cm$^{-2}$ s$^{-1}$ in the 4$-$60 keV band except in epoch I where the flux was calculated in 4$-$50 keV band. Energy (E) of the FeK$\alpha$ line, equivalent width (EW) of the line, $\rm{E_{cut}}$ and $\rm{kT_{e}}$ are expressed in units of keV, model normalization is in units of 10$^{-4}$ photons keV$^{-1}$ cm$^{-2}$s$^{-1}$ and $\rm{N_H}$(zTBabs) is the host galaxy hydrogen column density in units of $10^{22}$ atoms $\rm{cm^{-2}}$. $\rm{C_{FPMA/FPMB}}$ is the cross-calibration constant. The width of the FeK$\alpha$ line was fixed to 0.1 keV during the fitting.}\label{table-2}
\centering
\begin{tabular}{p{0.09\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}p{0.06\linewidth}}
\hline
Parameter & epoch A & epoch B & epoch C & epoch D & epoch E & epoch F & epoch G & epoch H & epoch I \\
\hline\hline
\multicolumn{10}{c}{Model I : {\it const*TBabs*zTBabs(zpo+zgauss)}} \\
\hline
$\Gamma$ & 1.56$^{+0.01}_{-0.01}$ & 1.55$^{+0.01}_{-0.01}$ & 1.61$^{+0.01}_{-0.01}$ & 1.63$^{+0.01}_{-0.01}$ & 1.66$^{+0.01}_{-0.01}$ & 1.63$^{+0.01}_{-0.01}$ & 1.61$^{+0.02}_{-0.02}$ & 1.64$^{+0.04}_{-0.04}$ & 1.48$^{+0.07}_{-0.07}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & -& - & 2.91$^{+1.04}_{-1.03}$ & 4.66$^{+1.83}_{-1.80}$ & 4.57$^{+3.73}_{-3.62}$ \\
E & 6.35$^{+0.03}_{-0.03}$ & 6.38$^{+0.03}_{-0.04}$ & 6.33$^{+0.04}_{-0.04}$ & 6.32$^{+0.04}_{-0.03}$ & 6.30$^{+0.05}_{-0.05}$ & 6.34$^{+0.05}_{-0.05}$ & 6.23$^{+0.04}_{-0.04}$ & 6.28$^{+0.07}_{-0.07}$ & 6.38$^{+0.05}_{-0.05}$ \\
EW & 149$^{+23}_{-32}$ & 188$^{+32}_{-42}$ & 142$^{+30}_{-35}$ & 129$^{+22}_{-24}$ & 86$^{+26}_{-30}$ & 115$^{+26}_{-38}$ & 115$^{+23}_{-23}$ & 123$^{+44}_{-40}$ & 269$^{+66}_{-81}$ \\
norm & 71$^{+2}_{-2}$ & 58$^{+2}_{-2}$ & 73$^{+2}_{-2}$ & 90$^{+3}_{-3}$ & 101$^{+3}_{-3}$ & 92$^{+3}_{-3}$ & 122$^{+7}_{-6}$ & 76$^{+8}_{-7}$ & 17$^{+4}_{-3}$ \\
$\chi^2/dof$ & 930/826 & 752/706 & 834/704 & 909/777 & 815/790 & 684/717 & 1081/935 & 600/595 & 257/247 \\
$\rm{C_{FPMA/FPMB}}$ & 1.04$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.00$^{+0.01}_{-0.01}$ & 0.99$^{+0.01}_{-0.01}$ & 1.04$^{+0.01}_{-0.01}$ & 1.06$^{+0.03}_{-0.03}$ \\
\hline
\multicolumn{10}{c}{Model II : {\it const*TBabs*zTBabs(pexrav+zgauss)}} \\
\hline
$\Gamma$ & 1.68$^{+0.05}_{-0.05}$ & 1.64$^{+0.06}_{-0.05}$ & 1.77$^{+0.05}_{-0.05}$ & 1.83$^{+0.04}_{-0.04}$ & 1.88$^{+0.05}_{-0.05}$ & 1.82$^{+0.05}_{-0.05}$ & 1.92$^{+0.01}_{-0.07}$ & 1.85$^{+0.08}_{-0.08}$ & 1.70$^{+0.04}_{-0.11}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & -& - & 6.43$^{+0.92}_{-1.42}$ & 7.71$^{+1.93}_{-1.95}$ & 8.52$^{+4.30}_{-5.18}$ \\
$\rm{E_{cut}}$ & 212$^{+140}_{-63}$ & 163$^{+118}_{-50}$ & $>$254 & $>$1147 & $>$411 & $>$406 & $>$571 & $>$775 & $>$126 \\
R & 0.52$^{+0.15}_{-0.13}$ & 0.48$^{+0.18}_{-0.15}$ & 0.53$^{+0.18}_{-0.15}$ & 0.55$^{+0.16}_{-0.14}$ & 0.73$^{+0.19}_{-0.17}$ & 0.55$^{+0.18}_{-0.16}$ & 0.83$^{+0.16}_{-0.15}$ & 0.45$^{+0.20}_{-0.18}$ & 0.38$^{+0.49}_{-0.34}$ \\
E & 6.35$^{+0.03}_{-0.03}$ & 6.38$^{+0.04}_{-0.04}$ & 6.33$^{+0.04}_{-0.04}$ & 6.32$^{+0.04}_{-0.04}$ & 6.31$^{+0.06}_{-0.06}$ & 6.34$^{+0.05}_{-0.05}$ & 6.23$^{+0.04}_{-0.04}$ & 6.28$^{+0.07}_{-0.07}$ & 6.38$^{+0.06}_{-0.06}$ \\
EW & 137$^{+32}_{-20}$ & 177$^{+37}_{-42}$ & 129$^{+26}_{-33}$ & 116$^{+25}_{-23}$ & 74$^{+26}_{-21}$ & 102$^{+32}_{-22}$ & 85$^{+23}_{-19}$ & 102$^{+34}_{-31}$ & 231$^{+82}_{-74}$ \\
norm & 84$^{+6}_{-6}$ & 66$^{+6}_{-5}$ & 91$^{+8}_{-7}$ & 119$^{+8}_{-7}$ & 139$^{+10}_{-10}$ & 120$^{+9}_{-9}$ & 207$^{+19}_{-25}$ & 112$^{+18}_{-16}$ & 25$^{+9}_{-9}$ \\
$\chi^2/dof$ & 856/824 & 711/704 & 784/702 & 831/775 & 714/788 & 625/715 & 920/933 & 577/593 & 253/245 \\
$\rm{C_{FPMA/FPMB}}$ & 1.04$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.00$^{+0.01}_{-0.01}$ & 0.99$^{+0.01}_{-0.01}$ & 1.04$^{+0.01}_{-0.01}$ & 1.06$^{+0.03}_{-0.03}$ \\
\hline
\multicolumn{10}{c}{Model III : {\it const*TBabs*zTBabs(xillver)}}\\
\hline
$\Gamma$ & 1.69$^{+0.01}_{-0.01}$ & 1.69$^{+0.01}_{-0.01}$ & 1.77$^{+0.02}_{-0.01}$ & 1.83$^{+0.01}_{-0.01}$ & 1.80$^{+0.01}_{-0.01}$ & 1.79$^{+0.01}_{-0.01}$ & 1.77$^{+0.01}_{-0.01}$ & 1.87$^{+0.02}_{-0.02}$ & 1.91$^{+0.03}_{-0.03}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & -& - & 4.73$^{+0.57}_{-0.56}$ & 8.58$^{+0.98}_{-1.08}$ & 12.12$^{+1.86}_{-1.82}$ \\
$\rm{E_{cut}}$ & 127$^{+14}_{-12}$ & 92$^{+10}_{-8}$ & 204$^{+46}_{-34}$ & $>$439 & 378$^{+152}_{-91}$ & 326$^{+123}_{-78}$ & 248$^{+67}_{-34}$ & $>$510 & $>$198 \\
R & 0.86$^{+0.12}_{-0.12}$ & 1.09$^{+0.17}_{-0.16}$ & 0.85$^{+0.14}_{-0.13}$ & 0.76$^{+0.14}_{-0.12}$ & 0.61$^{+0.12}_{-0.09}$ & 0.71$^{+0.14}_{-0.11}$ & 0.68$^{+0.08}_{-0.08}$ & 0.64$^{+0.16}_{-0.14}$ & 1.36$^{+0.47}_{-0.40}$ \\
norm & 1.98$^{+0.02}_{-0.03}$ & 1.46$^{+0.02}_{-0.05}$ & 2.01$^{+0.03}_{-0.03}$ & 2.98$^{+0.04}_{-0.04}$ & 2.97$^{+0.03}_{-0.03}$ & 2.75$^{+0.04}_{-0.04}$ & 3.73$^{+0.04}_{-0.04}$ & 2.72$^{+0.03}_{-0.04}$ & 0.68$^{+0.03}_{-0.03}$ \\
$\chi^2/dof$ & 868/826 & 728/706 & 785/704 & 837/777 & 735/790 & 627/717 & 976/935 & 579/595 & 260/247 \\
$\rm{C_{FPMA/FPMB}}$ & 1.04$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.00$^{+0.01}_{-0.01}$ & 0.99$^{+0.01}_{-0.01}$ & 1.04$^{+0.01}_{-0.01}$ & 1.06$^{+0.03}_{-0.03}$ \\
\hline
\multicolumn{10}{c}{Model IV : {\it const*TBabs*zTBabs(xillverCP)}} \\
\hline
$\Gamma$ & 1.78$^{+0.01}_{-0.01}$ & 1.80$^{+0.01}_{-0.01}$ & 1.83$^{+0.01}_{-0.01}$ & 1.84$^{+0.01}_{-0.01}$ & 1.83$^{+0.01}_{-0.01}$ & 1.83$^{+0.01}_{-0.01}$ & 1.82$^{+0.01}_{-0.01}$ & 1.87$^{+0.02}_{-0.02}$ & 1.91$^{+0.04}_{-0.03}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & -& - & 4.97$^{+0.59}_{-0.55}$ & 8.17$^{+0.99}_{-0.82}$ & 12.21$^{+1.73}_{-1.65}$ \\
$\rm{kT_{e}}$ & 33$^{+9}_{-7}$ & 28$^{+9}_{-6}$ & 56$^{+131}_{-18}$ & $>$80 & $>$45 & $>$47 & 50$^{+39}_{-10}$ & $>$85 & $>$36 \\
R & 0.79$^{+0.11}_{-0.11}$ & 0.96$^{+0.18}_{-0.10}$ & 0.83$^{+0.15}_{-0.15}$ & 0.78$^{+0.13}_{-0.12}$ & 0.62$^{+0.11}_{-0.11}$ & 0.71$^{+0.14}_{-0.13}$ & 0.67$^{+0.09}_{-0.10}$ & 0.65$^{+0.16}_{-0.15}$ & 1.28$^{+0.60}_{-0.30}$ \\
norm & 1.94$^{+0.02}_{-0.02}$ & 1.52$^{+0.02}_{-0.02}$ & 1.89$^{+0.03}_{-0.03}$ & 2.56$^{+0.02}_{-0.03}$ & 2.55$^{+0.03}_{-0.05}$ & 2.50$^{+0.04}_{-0.04}$ & 3.32$^{+0.03}_{-0.07}$ & 2.24$^{+0.07}_{-0.05}$ & 0.66$^{+0.02}_{-0.02}$ \\
$\chi^2/dof$ & 881/826 & 746/706 & 790/704 & 838/777 & 733/790 & 629/717 & 970/935 & 581/595 & 261/247 \\
$\rm{C_{FPMA/FPMB}}$ & 1.04$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.01$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.00$^{+0.01}_{-0.01}$ & 0.99$^{+0.01}_{-0.01}$ & 1.04$^{+0.01}_{-0.01}$ & 1.06$^{+0.03}_{-0.03}$ \\
Flux & 1.04$^{+0.02}_{-0.01}$ & 0.86$^{+0.01}_{-0.01}$ & 0.94$^{+0.01}_{-0.02}$ & 1.11$^{+0.01}_{-0.01}$ & 1.15$^{+0.01}_{-0.02}$ & 1.13$^{+0.01}_{-0.01}$ & 1.55$^{+0.01}_{-0.01}$ & 0.88$^{+0.01}_{-0.01}$ & 0.27$^{+0.01}_{-0.01}$ \\
\hline\hline
\end{tabular}
\end{table*}
\subsection{Data reduction}
We reduced {\it NuSTAR} data in the 3$-$79 keV band using the standard {\it NuSTAR} data reduction software
NuSTARDAS\footnote{https://heasarc.gsfc.nasa.gov/docs/nustar/analysis/nustar swguide.pdf} distributed by HEASARC within HEASoft v6.29. Considering the passage of the satellite through the South Atlantic Anomaly (SAA) we selected, SAACALC \say{2}, SAAMODE \say{optimized} and also excluded the tentacle region. The calibrated, cleaned,
and screened event files were generated by running the {\tt nupipeline}
task using the CALDB release 20210701. To extract the source counts we chose a circular region of radius 60 arcsec
centered on the source. Similarly, to extract the background counts,
we selected a circular region of the same radius away
from the source on the same chip to avoid contamination from source photons. We then used the
{\tt nuproducts} task to generate energy spectra,
response matrix files (RMFs) and auxiliary response files (ARFs), for both
the hard X-ray detectors housed inside the corresponding focal plane modules FPMA and FPMB.
For spectral analysis, using XSPEC version 12.12.0 \citep{1996ASPC..101...17A},
we fitted the background subtracted spectra from
FPMA and FPMB simultaneously (without combining them) allowing the
cross normalization factor to vary freely during spectral fits. The spectra were binned to have a S/N ratio greater than 5 in each spectral channel using the {\it NuSTAR}-specific Python script {\it snrgrppha}\footnote{https://sites.astro.caltech.edu/$\sim$mislavb/}. To get an estimate of the model parameters that best describe the observed data, we
used the chi-square ($\chi^2$) statistics and for calculating the errors
in the model parameters we used the $\chi^2$ = 2.71 criterion i.e. 90 per cent confidence
range in XSPEC.
\begin{table*}
\caption{Results of the model fits to the spectra of MR 2251$-$178. The models are Model I: {\it const*TBabs*zTBabs(zpo)}, Model II: {\it const*TBabs*zTBabs(pexrav)}, Model III: {\it const*TBabs*zTBabs(xillver)} and Model IV: {\it const*TBabs*zTBabs(xillverCP)}. The {\it zTBabs} component is added with all the models to fit the epoch E spectra. In epoch D the width of the Fek$\alpha$ line was fixed to 0.1 keV. In the case where no line was detected, the upper limit on the EW was calculated by fixing the line energy to 6.4 keV.The fluxes are in units of 10$^{-10}$ erg cm$^{-2}$ s$^{-1}$ in the 4$-$60 keV band except in epoch E where the flux was calculated in 4$-$50 keV band. Columns and parameters have the same meaning as given in Table \ref{table-2}.}\label{table-3}
\centering
\begin{tabular}{p{0.12\linewidth}p{0.12\linewidth}p{0.12\linewidth}p{0.12\linewidth}p{0.12\linewidth}p{0.12\linewidth}}
\hline
Parameter & epoch A & epoch B & epoch C & epoch D & epoch E \\
\hline\hline
\multicolumn{6}{c}{Model I : {\it const*TBabs*zTBabs(zpo+zgauss)}} \\
\hline
$\Gamma$ & 1.75$^{+0.02}_{-0.02}$ & 1.79$^{+0.01}_{-0.01}$ & 1.79$^{+0.02}_{-0.02}$ & 1.79$^{+0.02}_{-0.02}$ & 1.83$^{+0.05}_{-0.05}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & 7.82$^{+2.99}_{-2.93}$ \\
E & - & - & - & 6.49$^{+0.37}_{-0.27}$ & - \\
EW & $<$35 & $<$49 & $<$34 & $<$70 & $<$46 \\
norm & 142$^{+5}_{-5}$ & 177$^{+6}_{-6}$ & 162$^{+6}_{-6}$ & 162$^{+6}_{-6}$ & 84$^{+13}_{-11}$ \\
$\chi^2/dof$ & 583/601 & 656/633 & 574/551 & 516/576 & 446/400 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.00$^{+0.01}_{-0.01}$ & 1.03$^{+0.02}_{-0.02}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.02}_{-0.02}$ \\
\hline
\multicolumn{6}{c}{Model II : {\it const*TBabs*zTBabs(pexrav+zgauss)}} \\
\hline
$\Gamma$ & 1.65$^{+0.05}_{-0.05}$ & 1.72$^{+0.05}_{-0.03}$ & 1.76$^{+0.07}_{-0.07}$ & 1.79$^{+0.06}_{-0.06}$ & 1.82$^{+0.11}_{-0.10}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & 7.66$^{+3.26}_{-3.02}$ \\
$\rm{E_{cut}}$ & 125$^{+96}_{-39}$ & 185$^{+200}_{-69}$ & 110$^{+70}_{-32}$ & 193$^{+417}_{-80}$ & $>$175 \\
R & $<$0.07 & $<$0.11 & 0.29$^{+0.22}_{-0.18}$ & 0.19$^{+0.20}_{-0.17}$ & $<$0.23 \\
E & - & - & - & 6.48$^{+0.67}_{-0.31}$ & - \\
EW & $<$49 & $<$43 & $<$33 & $<$63 & $<$50 \\
norm & 111$^{+8}_{-8}$ & 144$^{+11}_{-6}$ & 139$^{+15}_{-13}$ & 144$^{+15}_{-13}$ & 75$^{+19}_{-14}$\\
$\chi^2/dof$ & 569/599 & 647/631 & 552/549 & 509/574 & 446/398 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.00$^{+0.01}_{-0.01}$ & 1.03$^{+0.02}_{-0.02}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.02}_{-0.02}$ \\
\hline
\multicolumn{6}{c}{Model III : {\it const*TBabs*zTBabs(xillver)}} \\
\hline
$\Gamma$ & 1.65$^{+0.02}_{-0.02}$ & 1.72$^{+0.02}_{-0.02}$ & 1.70$^{+0.02}_{-0.02}$ & 1.77$^{+0.02}_{-0.02}$ & 1.82$^{+0.02}_{-0.03}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & 7.56$^{+1.69}_{-1.38}$ \\
$\rm{E_{cut}}$ & 124$^{+22}_{-18}$ & 169$^{+45}_{-30}$ & 103$^{+18}_{-14}$ & 163$^{+46}_{-30}$ & $>$366 \\
R & $<$0.10 & $<$0.16 & 0.17$^{+0.14}_{-0.13}$ & 0.22$^{+0.15}_{-0.13}$ & $<$0.11 \\
norm & 2.92$^{+0.02}_{-0.04}$ & 3.54$^{+0.03}_{-0.05}$ & 2.79$^{+0.05}_{-0.05}$ & 3.05$^{+0.02}_{-0.05}$ & 2.10$^{+0.03}_{-0.52}$ \\
$\chi^2/dof$ & 569/599 & 647/631 & 556/549 & 510/576 & 447/398 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.00$^{+0.01}_{-0.01}$ & 1.03$^{+0.02}_{-0.02}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.02}_{-0.02}$ \\
\hline
\multicolumn{6}{c}{Model IV : {\it const*TBabs*zTBabs(xillverCP)}} \\
\hline
$\Gamma$ & 1.76$^{+0.01}_{-0.01}$ & 1.79$^{+0.01}_{-0.01}$ & 1.80$^{+0.02}_{-0.01}$ & 1.83$^{+0.02}_{-0.01}$ & 1.83$^{+0.02}_{-0.02}$ \\
$\rm{N_H}$(zTBabs) & - & - & - & - & 7.54$^{+1.42}_{-1.45}$ \\
$\rm{kT_{e}}$ & 25$^{+26}_{-6}$ & 35$^{+149}_{-11}$ & 21$^{+8}_{-4}$ & 35$^{+67}_{-11}$ & $>$32 \\
R & $<$0.06 & $<$0.11 & $<$0.25 & 0.17$^{+0.14}_{-0.13}$ & $<$0.09 \\
norm & 2.68$^{+0.02}_{-0.03}$ & 3.26$^{+0.03}_{-0.05}$ & 2.66$^{+0.04}_{-0.04}$ & 2.85$^{+0.04}_{-0.04}$ & 1.71$^{+0.02}_{-0.04}$ \\
$\chi^2/dof$ & 576/599 & 650/631 & 555/549 & 511/576 & 447/398 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.00$^{+0.01}_{-0.01}$ & 1.03$^{+0.02}_{-0.02}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.02}_{-0.02}$ \\
Flux & 1.07$^{+0.01}_{-0.01}$ & 1.22$^{+0.01}_{-0.01}$ & 1.08$^{+0.01}_{-0.01}$ & 1.09$^{+0.01}_{-0.01}$ & 0.47$^{+0.02}_{-0.01}$ \\
\hline\hline
\end{tabular}
\end{table*}
\begin{figure*}
\hbox{
\includegraphics[scale=0.28]{spectra3227.eps}
\includegraphics[scale=0.28]{spectra5548.eps}
\includegraphics[scale=0.28]{spectramr.eps}
}
\caption{The unfolded spectra of the nine observations for
NGC 3227 (left), six observations for NGC 5548 (middle) and five observations
for MR 2251$-$178 (right) fitted with a simple power-law. The bottom panels show the
ratio of the observed spectra to the model. For clarity, we used only FPMA data. The spectra are rebinned for visualization purpose only. }\label{figure-1}
\end{figure*}
\begin{figure}
\vbox{
\includegraphics[scale=0.95]{a_ngc3227.eps}
\includegraphics[scale=0.95]{b_ngc3227.eps}
\includegraphics[scale=0.95]{c_ngc3227.eps}
}
\vspace{-1.0cm}
\caption{Ratio of data to the model for the model fits {\it const*TBabs(zpo+zgauss)},
{\it const*TBabs(pexrav+zgauss)} and {\it const*TBabs(xillverCP)} to the FPMA (blue triangle) and FPMB (yellow triangle) spectra of OBID 60202002002 of NGC 3227. The spectra are rebinned for visualization purpose.}
\label{figure-2}
\end{figure}
\begin{figure}
\vbox{
\includegraphics[scale=0.95]{a_ngc5548.eps}
\includegraphics[scale=0.95]{b_ngc5548.eps}
\includegraphics[scale=0.95]{c_ngc5548.eps}
}
\vspace{-1.0cm}
\caption{Ratio plots for the model fits {\it const*TBabs*zTBabs*(zpo+zgauss)}, {\it const*TBabs*zTBabs*(pexrav+zgauss)} and {\it const*TBabs*zTBabs(xillverCP)} to the FPMA (blue star) and FPMB (yellow star) spectra of OBID 60002044006 of NGC 5548. The spectra are rebinned for visualization purpose only.}
\label{figure-3}
\end{figure}
\begin{figure}
\vbox{
\includegraphics[scale=0.95]{a_mr.eps}
\includegraphics[scale=0.95]{b_mr.eps}
\includegraphics[scale=0.95]{c_mr.eps}
}
\vspace{-1.0cm}
\caption{Ratio of data to model for the model fits {\it const*TBabs*zTBabs*(zpo)}, {\it const*TBabs*zTBabs*(pexrav)} and {\it const*TBabs*zTBabs(xillverCP)} to the FPMA (blue dot) and FPMB (yellow dot) spectra of OBID 90601637002 of MR 2251$-$178. We re-binned the spectra for visualization purpose only.}
\label{figure-4}
\end{figure}
\section{Analysis of the data}
For few epochs of the sources studied in this work, we do have observations in the soft band from telescopes such as {\it XMM-Newton} for NGC 3227 and {\it XMM-Newton}, {\it Chandra} for NGC 5548 and {\it XMM-Newton} for MR 2251$-$178. However, for this work, we decided to use only {\it NuSTAR} data as (a) the good sensitivity of {\it NuSTAR} over the 3$-$79 keV energy band captures all the key reflection features of an AGN spectrum and as the main goal of this work is to model the high energy rollover of the Comptonized spectra we did not want the absorption in the soft band affecting our analysis in the determination of $\rm{kT_{e}}$ and (b) observations in the soft band are not available simultaneous to the {\it NuSTAR} observations for all the epochs and for all the sources. However we note here, the inclusion of soft X-ray data in the fitting might have an effect on the photon index ($\Gamma$) obtained from using the {\it NuSTAR} data alone. However, a simplest approach to constrain $\rm{kT_{e}}$ (which is the aim of this work) is the use of only {\it NuSTAR} data, but for better constraining the other physical characteristics of the sources broad band spectral analysis including data from UV to hard X-ray band is more appropriate. While analyzing only {\it NuSTAR} data, we ignored the 3$-$4 keV band to limit the effect of absorption if any and also did not consider data in the energy range beyond 60 keV due to lack of source photons. Thus, we carried out spectral fits to the {\it NuSTAR} data in the 4$-$60 keV energy band for all the OBSIDS except for epoch I of NGC 3227, epoch F of NGC 5548 and epoch E of MR 2251$-$178. Due to the unavailability of photons beyond 50 keV we restricted the spectral fit in the 4$-$50 keV energy band to the epoch I and epoch E spectra of NGC 3227 and MR 2251$-$178 respectively. Similarly, for the epoch F spectrum of NGC 5548, we used the FPMA/FPMB data in the 4$-$55 keV range. (see Fig \ref{figure-1}).
\subsection{Phenomenological spectral fits}
For our spectral fits, to model the line of sight galactic absorption, the value of the neutral hydrogen column density ($\rm{N_H}$) for all the sources were frozen to the values obtained from \cite{2013MNRAS.431..394W}. These $\rm{N_H}$ values are given in Table \ref{table-1}. Similarly, the redshifts of the sources were frozen to their corresponding values given in Table \ref{table-1}. Also, we used the solar abundances from \cite{2000ApJ...542..914W} and the photoelectric cross sections from \cite{1996ApJ...465..487V}. For models that require inclination angle ($i$) , we used $i$ = 50$^{\circ}$ for NGC 3227 (\citealt{1997ApJ...477..623S}, \citealt{2019A&A...628A..65A}, \citealt{2016MNRAS.457.1568M}), $i$ = 30$^{\circ}$ for NGC 5548 \citep{2015A&A...577A..38U} and $i$ = 60$^{\circ}$ (i.e., the default value) for MR 2251$-$178.
\subsubsection{Absorbed power law}
Firstly, to understand the continuum emission in our sample of sources we
fitted the observed X-ray spectra with the baseline phenomenological absorbed power law
model and having the following form in XSPEC
\begin{equation}
const*TBabs(zpo)
\end{equation}
The first component of this model is the constant used to calibrate the two focal plane modules of {\it NuSTAR}. The second component, {\it TBabs} \citep{2000ApJ...542..914W} was used to model the line
of sight galactic absorption. The parameters
that were kept free are $\Gamma$ and the normalization (i.e. photons keV$^{-1}$ cm$^{-2}$ s$^{-1}$). We found evidence of intrinsic absorption present in the lower energy end for all the epochs in NGC 5548, epoch G, H and I in NGC 3227 and epoch E in MR 2251$-$178. We therefore included a {\it zTBabs} component with the absorbed power law model to fit their spectra and the model looks like
\begin{equation}
const*TBabs*zTBabs(zpo)
\end{equation}
in XSPEC. For all epochs in NGC 5548, two epochs in NGC 3227 and one epoch of MR 2251$-$178 non-inclusion of {\it zTBabs} returned a poor fit with $\chi^2/dof$ larger than 1.2. Inclusion of {\it zTBabs} with $\rm{N_H}$({\it zTBabs}) kept free improved the fit with $\chi^2/dof$ close to unity.
\begin{figure}
\vbox{
\hspace{-1.0cm}\includegraphics[scale=0.40]{xrat14nh.eps}
}
\caption{The unfolded spectra with the ratio of data to model for the model fits {\it const*TBabs(xillverCP)} to the FPMA (blue) and FPMB (yellow) spectra of OBID 60202002014 (epoch G) of NGC 3227.}
\label{figure-5}
\end{figure}
From the residual spectra obtained from the simple absorbed power law model fit
to the observations, we noticed the presence of the fluorescent FeK$\alpha$ line
in NGC 3227, NGC 5548 and in epoch D spectra of MR 2251$-$178, but in the other observations of MR 2251$-$178, the residual spectra do not show conspicuous FeK$\alpha$ line. The spectral fits are shown in Fig. \ref{figure-1} for NGC 3227 and NGC 5548 and MR 2251$-$178.
Though this line is common in most of the
X ray spectra of AGN \citep{1993ARA&A..31..717M, 2007MNRAS.382..194N, 2010HEAD...11.0708D} there are exceptions \citep{2011MNRAS.416..629B}.
The apparent non-detection of FeK$\alpha$ line in the spectra of MR 2251$-$178
could be due to weaker reflection owing to larger viewing angle \citep{2011MNRAS.416..629B},
low signal-to-noise ratio spectra, very high ionized accretion disk
\citep{1993MNRAS.261...74R,1994MNRAS.266..653Z} or a combination of all.
However, the presence of a weak FeK$\alpha$ line in epoch D of MR 2251$-$178 could point to the physical characteristic of MR 2251$-$178 being different from the other two sources. A thorough analysis is needed to know the exact reasons for the absence/weakness of this line
but this issue is beyond the scope of this work as we are
in this work mainly interested in the changes in the temperature of the corona.
To model the FeK$\alpha$ line seen in the residual spectra in NGC 3227,
NGC 5548 and in one observation (epoch D) of MR 2251$-$178, we included a gaussian component and with this inclusion the model takes the form
{\it const*TBabs(zpo + zgauss)} and the quality of the fit improved.
After the inclusion of a gaussian component with the power law in the case of NGC 3227, in all nine epochs the $\chi^2$ value reduced in the range between 60 and 242 for a reduction of 2 dof. In epochs C, D, G and I, non-inclusion of the gaussian component, resulted in a reduced $\chi^2$ greater than 1.30. After the inclusion of the gaussian component the reduced $\chi^2$ ranged between 1.04$-$1.16 in these epochs. For the other epochs, the $\chi^2/dof$ was $>$ 1.1 before and it became $\sim$ 1.0 after the line inclusion. For MR2251$-$178, gaussian component was used to fit the FeK$\alpha$ line only in epoch D. Inclusion of the line component lead to a change in $\chi^2$ of 5 for a reduction of 2 dof. This negligible change in $\chi^2$ over 2 dof did not improve the fit quality significantly in this case. For NGC 5548, on inclusion of gaussian line component in all the epochs, the value of the $\chi^2$ reduced in the range 23 to 105 for a reduction of 2 dof. Just an absorbed power law fit to the data produced the reduced $\chi^2$ greater than 1.2 for the epochs C, D, E and F. Adding the Gaussian with the power law lead to a reduced $\chi^2$ of $\sim$ 1.1 for epoch C and D and $\sim$ 1.0 for epoch E and F. Similarly, for epochs A and B, the $\chi^2/dof$ changed from 1.10 to 1.06 and 1.02 to 0.97 respectively. The width of the FeK$\alpha$ line was frozen to the value of 0.1 keV during the fitting, letting it free to vary did not improve the fit significantly. The best fit
parameters for the sources are given in
Table \ref{table-2}, \ref{table-3} and \ref{table-4} for NGC 3227, MR 2251$-$178 and NGC 5548.
\subsubsection{Pexrav}
In the residuals of the simple power law fit to all the spectra of the sources (see Fig. \ref{figure-1}) we found
the signature of a high energy turn over and a reflection hump beyond 15 keV. To appropriately model both the high energy cut off and the reflection feature present in the spectra we replaced the {\it zpo} component in our earlier
model with {\it pexrav} and the new model has the form
\begin{equation}
const*TBabs(pexrav).
\end{equation}
While modelling the reflection component of NGC 5548, epoch G, H and I spectra of NGC 3227 and epoch E spectrum of MR 2251$-$178 an intrinsic absorption component, {\it zTBabs} was added with the above model. The intrinsic hydrogen column density, $\rm{N_H}$({\it zTBabs}) was kept free during the fit.
\begin{figure}
\vbox{
\includegraphics[scale=0.45]{cutchi_ngc3227.eps}
}
\caption{Variation of $\rm{E_{cut}}$ and $\rm{kT_{e}}$ with the observation epochs as obtained from {\it xillver} and {\it xillverCP} model fit respectively to the spectra NGC 3227. The plotted errors were calculated using the $\chi^2$ = 2.71 criterion i.e. 90\% confidence range. The black dashed lines in each panel are fits of constant (mean of $\rm{E_{cut}}$ and $\rm{kT_{e}}$) to the data points. For the fitting, epochs in which we were unable to constrain $\rm{E_{cut}}$ and $\rm{kT_{e}}$ were dropped (indicated with red point).} \label{figure-6}
\end{figure}
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x02_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x04_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x06_3227.eps}
}
\vspace{-0.5cm}
\end{figure*}
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x08_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x10_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x12_3227.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra along with {\it const*TBabs(xillverCP)} model fits and the data to model ratio for the epoch A (top left panel), epoch B (top middle panel), epoch C (top right panel), epoch D (bottom left panel), epoch E (bottom middle panel) and epoch F (bottom right panel) for the source NGC 3227. blue and yellow data points refer to FPMA and FPMB respectively.}
\label{figure-7}
\end{figure*}
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x14_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x82_3227.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x84_3227.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra and model fits along with the data to model ratio for {\it const*TBabs(xillverCP)} fits to the epoch G (left), epoch H (middle) and epoch I (right) of NGC 3227. For epoch G, epoch H and epoch I {\it zTBabs} was added to the model. blue and yellow data points refer to FPMA and FPMB respectively.}
\label{figure-8}
\end{figure*}
This model implements both photoelectric absorption and Compton scattering,
however, it does not include fluorescence. Therefore to model the FeK$\alpha$
line seen in the residual spectra in NGC 3227, NGC 5548 and epoch D in MR 2251$-$178, we included a gaussian component. So, the model takes the following form,
\begin{equation}
const*TBabs(pexrav+zgauss)
\end{equation}
in XSPEC. The model {\it pexrav} improved the fitting compared to {\it zpo} as evident in
the reduced $\chi^2$ in Tables \ref{table-2}, \ref{table-3} and
\ref{table-4}. This model includes primary emission
having a power law form with an exponential cut-off and a reflection component.
The reflector is considered to be an optically thick medium in an infinite plane
geometry and covering a larger fraction of the X-ray source. A parameter that comes as an output in the model fit is the reflection parameter $R$. This parameter gives a measure of the reflection component
present in the observed spectra of the sources. For an isotropic source, this
parameter is related to the solid angle ($\Omega$) as $R \sim \Omega/2\pi$, and it is
dependent on the angle of inclination ($i$) between the perpendicular
to the accretion disk and the line of sight to the observer. The width of the FeK$\alpha$ line was fixed at 0.1 keV, treating the parameter free did not improve the fit. For all the sources, the parameters that were left free in the model fits are $\rm{E_{cut}}$, $\Gamma$, R and normalization.
Using {\it (pexrav+zgauss)} in all the nine epochs of NGC 3227, the $\chi^2$ reduced in the range between 4 and 161 for a reduction of 2 dof compared to the {\it (zpo+zgauss)} fit. In MR 2251$-$178, the reduction in $\chi^2$ is in the range between 0 and 22 with a reduction of 2 in dof. For all the six epochs in NGC 5548, the {\it (pexrav+zgauss)} fit produced a reduction in the $\chi^2$ values between 9$-$60 with a reduction of 2 dof compared to (zpo+zgauss). The best fit parameters are given
in Tables \ref{table-2} , \ref{table-3} and \ref{table-4}.
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x02_5548.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x03_5548.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x05_5548.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra with the {\it const*TBabs*zTBabs(xillverCP)} model fits and the data to model ratio to the spectra of NGC 5548. Here epoch A is shown in the left panel, epoch B is shown in the middle panel and epoch C is shown in the right panel. Data from FPMA and FPMB are shown in blue and yellow respectively.}
\label{figure-9}
\end{figure*}
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x06_5548.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x08_5548.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x92_5548.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra with the {\it const*TBabs*zTBabs(xillverCP)} model fits and the data to model ratio to the spectra of NGC 5548. Here the left, middle and right panels show the results for epochs D, E and F respectively. Data from FPMA and FPMB are shown in blue and yellow respectively.}
\label{figure-10}
\end{figure*}
\subsection{Physical model fits}
\subsubsection{xillver and xillverCP}
Accretion disk reflection features in the form of narrow FeK$\alpha$ line was
conspicuously present in the residuals of NGC 3227, NGC 5548 and epoch D spectra of MR 2251$-$178 and the line was modelled using a gaussian component in {\it pexrav} model fits. For MR 2251$-$178 we have seen signatures of reflection (though weak) in few epochs. This is evident from the best fit values of the reflection fraction obtained using {\it pexrav} (see Table \ref{table-3}) and in the residuals of the {\it zpo} model fits (see Fig. \ref{figure-4}: top panel). We therefore modelled the spectra of the sources with the self consistent ionized reflection model {\it xillver}\citep{2010ApJ...718..695G,2013ApJ...768..146G}. The emergent reflected spectrum from the surface of the X-ray illuminated accretion disk is calculated in the model by solving the equations of radiative transfer, energy balance and ionization equilibrium in a Compton thick plane parallel medium \citep{2013ApJ...768..146G}. The model fit to the spectra has
the following form in XSPEC,
\begin{equation}
const*TBabs(xillver)
\end{equation}
Here also, to account for the effect of intrinsic absorption, {\it zTBabs} was used for all epochs of NGC 5548, three epochs (G, H and I) of NGC 3227 and epoch E of MR 2251$-$178. The intrinsic hydrogen column densities,$N_{\rm H}$(zTBabs) were kept as free parameters. In this model fit, the parameters describing the properties of the corona such as $\Gamma$, $\rm{E_{cut}}$ and R were made to vary, while the inclination angle was frozen to the value of 50$^{\circ}$ for
NGC 3227, 30$^{\circ}$ for NGC 5548 and 60$^{\circ}$ for MR 2251$-$178. The other
parameter that was kept free is the normalization. To account for a fair comparison between the $\rm{E_{cut}}$ values obtained from {\it pexrav} and {\it xillver} the reflector was assumed to be neutral i.e. log$\zeta$ = 0 erg cm s$^{-1}$. Treating the ionization as a free parameter too, returned results consistent with the ones obtained using log$\zeta$ = 0 erg cm s$^{-1}$. The value of Fe abundance was fixed to the solar value.
The results of the model fits are given in Tables \ref{table-2}, \ref{table-3} and \ref{table-4}.
Fitting the spectra using {\it xillver} yielded values of $\rm{E_{cut}}$ similar to that obtained using {\it pexrav}, although the errors in the $\rm{E_{cut}}$ measurements using {\it xillver} are relatively smaller.
As we are interested in the determination of $\rm{kT_{e}}$ and its variation we decided to use the thermal Comptonization model {\it xillverCP} \citep{2014ApJ...782...76G}. This model that takes into account the emission lines assuming it to be originating from disk reflection includes the Comptonization model {\it nthcomp} \citep{1996MNRAS.283..193Z,1999MNRAS.309..561Z} plus the ionized reflection model {\it xillver}\citep{2010ApJ...718..695G,2013ApJ...768..146G}. The model fit to the spectra has
the following form in XSPEC,
\begin{equation}
const*TBabs(xillverCP)
\end{equation}
We used {\it xillverCP} to model the primary continuum and the reflection spectrum simultaneously. Also, to take care of the intrinsic absoption present in all epochs of NGC 5548, epoch G, H and I of NGC 3227 and epoch E of MR 2251$-$178 spectra the {\it zTBabs} component was added with the described model. The results of the model fits are given in Tables \ref{table-2}, \ref{table-3} and \ref{table-4}.
\section{Results and Discussion}
Here, we aimed to find changes of $\rm{kT_{e}}$ in NGC 3227, NGC 5548 and MR 2251$-$178 . We
discuss below the results obtained on them.
\begin{table*}
\caption{Results of the model fits to the spectra of NGC 5548. The models are Model I: {\it const*TBabs*zTBabs(zpo+zgauss)}, Model II: {\it const*TBabs*zTBabs(pexrav+zgauss)}, Model III: {\it const*TBabs*zTBabs(xillver)} and Model IV: {\it const*TBabs*zTBabs(xillverCP)}. The width of the FeK$\alpha$ line was frozen to the value of 0.1 keV during the fitting. The fluxes are in units of 10$^{-10}$ erg cm$^{-2}$ s$^{-1}$ in the 4$-$60 keV band except in epoch F where the flux was derived in 4$-$55 keV band. Columns and parameters have the same meaning as given in Table \ref{table-2}}.\label{table-4}
\centering
\begin{tabular}{p{0.12\linewidth}p{0.12\linewidth}p{0.10\linewidth}p{0.10\linewidth}p{0.10\linewidth}p{0.10\linewidth}p{0.10\linewidth}}
\hline
Parameter & epoch A & epoch B & epoch C & epoch D & epoch E & epoch F\\
\hline\hline
\multicolumn{7}{c}{Model I : {\it const*TBabs*zTBabs(zpo+zgauss)}} \\
\hline
$\Gamma$ & 1.62$^{+0.03}_{-0.03}$ & 1.60$^{+0.03}_{-0.03}$ & 1.59$^{+0.02}_{-0.02}$ & 1.64$^{+0.02}_{-0.02}$ & 1.57$^{+0.03}_{-0.02}$ & 1.57$^{+0.03}_{-0.01}$ \\
$\rm{N_H}$(zTBabs) & 3.89$^{+1.76}_{-1.73}$ & $<3.21$ & 6.29$^{+1.40}_{-1.32}$ & 2.70$^{+1.28}_{-1.27}$ & 8.97$^{+1.65}_{-1.63}$ & $<$1.28 \\
E & 6.37$^{+0.09}_{-0.09}$ & 6.33$^{+0.08}_{-0.08}$ & 6.32$^{+0.04}_{-0.04}$ & 6.38$^{+0.04}_{-0.04}$ & 6.66$^{+0.04}_{-0.04}$ & 6.31$^{+0.04}_{-0.04}$ \\
EW & 70$^{+42}_{-23}$ & 89$^{+27}_{-31}$ & 117$^{+34}_{-34}$ & 107$^{+28}_{-29}$ & 114$^{+26}_{-44}$ & 137$^{+24}_{-45}$ \\
norm & 100$^{+10}_{-9}$ & 86$^{+11}_{-11}$ & 79$^{+8}_{-8}$ & 93$^{+7}_{-6}$ & 77$^{+6}_{-5}$ & 67$^{+5}_{-2}$\\
$\chi^2/dof$ & 684/643 & 612/634 & 880/811 & 890/813 & 840/810 & 692/692 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$\\
\hline
\multicolumn{7}{c}{Model II : {\it const*TBabs*zTBabs(pexrav+zgauss)}} \\
\hline
$\Gamma$ & 1.75$^{+0.02}_{-0.10}$ & 1.81$^{+0.02}_{-0.08}$ & 1.62$^{+0.10}_{-0.10}$ & 1.79$^{+0.09}_{-0.09}$ & 1.61$^{+0.10}_{-0.11}$ & 1.81$^{+0.08}_{-0.09}$\\
$\rm{N_H}$(zTBabs) & 6.06$^{+1.58}_{-2.46}$ & 4.77$^{+1.31}_{-2.05}$ & 5.61$^{+2.34}_{-2.36}$ & 4.46$^{+2.13}_{-2.16}$ & 6.91$^{+2.45}_{-2.47}$ & 3.66$^{+1.18}_{-2.10}$ \\
$\rm{E_{cut}}$ & $>$345 & $>$502 & 152$^{+158}_{-54}$ & $>$170 & 160$^{+178}_{-58}$ & $>$414 \\
R & 0.19$^{+0.14}_{-0.11}$ & 0.35$^{+0.16}_{-0.14}$ & 0.28$^{+0.11}_{-0.10}$ & 0.37$^{+0.12}_{-0.11}$ & 0.33$^{+0.12}_{-0.10}$ & 0.42$^{+0.14}_{-0.13}$\\
E & 6.37$^{+0.10}_{-0.10}$ & 6.32$^{+0.09}_{-0.10}$ & 6.32$^{+0.04}_{-0.04}$ & 6.38$^{+0.05}_{-0.05}$ & 6.35$^{+0.04}_{-0.04}$ & 6.31$^{+0.05}_{-0.05}$ \\
EW & 55$^{38}_{-27}$ & 64$^{+31}_{-46}$ & 114$^{+24}_{-38}$ & 89$^{+25}_{-20}$ & 113$^{+29}_{-34}$ & 104$^{+44}_{-27}$ \\
norm & 126$^{+7}_{-23}$ & 125$^{+15}_{-19}$ & 79$^{+18}_{-15}$ & 118$^{+23}_{-20}$ & 73$^{+17}_{-14}$ & 103$^{+3}_{-17}$ \\
$\chi^2/dof$ & 675/641 & 588/632 & 834/809 & 840/811 & 780/808 & 652/690 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ \\
\hline
\multicolumn{7}{c}{Model III : {\it const*TBabs*zTBabs(xillver)}} \\
\hline
$\Gamma$ & 1.75$^{+0.02}_{-0.02}$ & 1.78$^{+0.02}_{-0.04}$ & 1.68$^{+0.02}_{-0.01}$ & 1.77$^{+0.01}_{-0.01}$ & 1.66$^{+0.01}_{-0.02}$ & 1.81$^{+0.01}_{-0.01}$ \\
$\rm{N_H}$(zTBabs) & 6.47$^{+0.98}_{-0.91}$ & 4.83$^{+0.95}_{-0.91}$ & 7.13$^{+0.73}_{-0.73}$ & 4.52$^{+0.69}_{-0.67}$ & 8.16$^{+0.76}_{-0.75}$ & 4.25$^{+0.83}_{-0.81}$ \\
$\rm{E_{cut}}$ & $>$487 & $>$480 & 129$^{+15}_{-13}$ & 179$^{+40}_{-23}$ & 133$^{+15}_{-13}$ & $>$395 \\
R & 0.25$^{+0.08}_{-0.08}$ & 0.36$^{+0.10}_{-0.09}$ & 0.49$^{+0.09}_{-0.08}$ & 0.49$^{+0.08}_{-0.08}$ & 0.51$^{+0.09}_{-0.08}$ & 0.54$^{+0.11}_{-0.10}$ \\
norm & 4.35$^{+0.10}_{-0.06}$ & 3.61$^{+0.47}_{-0.14}$ & 2.20$^{+0.03}_{-0.03}$ & 2.47$^{+0.03}_{-0.03}$ & 2.07$^{+0.04}_{-0.04}$ & 2.70$^{+0.04}_{-0.04}$ \\
$\chi^2/dof$ & 677/643 & 595/634 & 851/811 & 849/813 & 792/810 & 666/692 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ \\
\hline
\multicolumn{7}{c}{Model IV : {\it const*TBabs*zTBabs(xillverCP)}} \\
\hline
$\Gamma$ & 1.77$^{+0.02}_{-0.02}$ & 1.80$^{+0.02}_{-0.03}$ & 1.79$^{+0.01}_{-0.01}$ & 1.85$^{+0.01}_{-0.01}$ & 1.77$^{+0.01}_{-0.01}$ & 1.84$^{+0.01}_{-0.02}$ \\
$\rm{N_H}$(zTBabs) & 6.47$^{+0.99}_{-0.91}$ & 4.63$^{+0.94}_{-0.92}$ & 8.72$^{+0.73}_{-0.74}$ & 5.66$^{+0.69}_{-0.68}$ & 9.79$^{+0.76}_{-0.75}$ & 4.30$^{+0.87}_{-0.92}$ \\
$\rm{kT_{e}}$ & $>$53 & $>$54 & 39$^{+14}_{-10}$ & 65$^{+147}_{-24}$ & 38$^{+12}_{-9}$ & $>$65 \\
R & 0.25$^{+0.08}_{-0.09}$ & 0.35$^{+0.11}_{-0.08}$ & 0.43$^{+0.08}_{-0.08}$ & 0.46$^{+0.09}_{-0.08}$ & 0.45$^{+0.08}_{-0.08}$ & 0.55$^{+0.11}_{-0.11}$ \\
norm & 3.46$^{+0.05}_{-0.05}$ & 2.95$^{+0.34}_{-0.04}$ & 2.23$^{+0.03}_{-0.03}$ & 2.46$^{+0.03}_{-0.04}$ & 2.07$^{+0.03}_{-0.03}$ & 2.34$^{+0.08}_{-0.03}$ \\
$\chi^2/dof$ & 677/643 & 595/634 & 855/811 & 852/813 & 795/810 & 667/692 \\
$\rm{C_{FPMA/FPMB}}$ & 1.02$^{+0.02}_{-0.02}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ \\
Flux & 1.23$^{+0.01}_{-0.01}$ & 1.11$^{+0.01}_{-0.01}$ & 1.02$^{+0.01}_{-0.01}$ & 1.05$^{+0.01}_{-0.01}$ & 0.96$^{+0.01}_{-0.01}$ & 0.92$^{+0.01}_{-0.01}$ \\
\hline\hline
\end{tabular}
\end{table*}
\begin{figure*}
\vbox{
\hspace{-0.05cm}\includegraphics[scale=0.40]{x02_mr.eps}
\hspace{-0.05cm}\includegraphics[scale=0.40]{x04_mr.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra along with the data to model ratio for {\it const*TBabs(xillverCp)} fit to epoch A (left panel) and epoch B (right panel) observations of MR 2251$-$178. Here blue points are for FPMA and yellow points are for FPMB.}
\label{figure-11}
\end{figure*}
\vspace{-0.5cm}
\begin{figure*}
\vbox{
\hspace{-0.2cm}\includegraphics[scale=0.30]{x06_mr.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x08_mr.eps}
\hspace{-0.2cm}\includegraphics[scale=0.30]{x92_mr.eps}
}
\vspace{-0.5cm}
\caption{The unfolded spectra along with the data to model ratio for {\it const*TBabs(xillverCp)} fit to epoch C (left panel), epoch D (middle panel) and epoch E (right panel) observations of MR 2251$-$178. Here blue points are for FPMA and yellow points are for FPMB. In epoch E the {\it zTBabs} component was added to the model.}
\label{figure-12}
\end{figure*}
\subsection{NGC 3227}
{\it NuSTAR} observed this source nine times between November 2016 and
December 2019. Of these, results on seven epochs were reported by \cite{2021MNRAS.502...80K}, wherein they were
able to constrain $\rm{E_{cut}}$ in three epochs using phenomenological fits and $\rm{kT_{e}}$
in two epochs using physical model fits. Here, we report results for two additional
epochs and for all the nine epochs, we used both phenomenological
and physical model fits to model the spectra.
Ratios of the model {\it const*TBabs(zpo + zgauss)}, {\it const*TBabs(pexrav + zgauss)} and {\it const*TBabs(xillverCP)} fits to the observed FPMA and FPMB spectra carried out on the OBSID 60202002002, the one with the longest exposure time are shown in Fig. \ref{figure-2}. Weak FeK$\alpha$ line is present in all the OBSIDs and therefore in the {\it const*TBabs(zpo)} and {\it const*TBabs(pexrav)} model fits we included a {\it zgauss} component to account for the line.
All the above models could fit the spectra of all epochs of NGC 3227
reasonably well except for epochs G, H and I, where the $\chi^2/dof$ was
greater than 1.2 and there is evidence for the presence of
significant absorption at the low energy end (see Fig. \ref{figure-5}). Addition of an extra absorption component
{\it zTBabs} that takes into account the effects of host
galaxy absorption to all the models improved the fit significantly
(see left, middle and right panel of Fig. \ref{figure-8}). The {\it const*TBabs(pexrav + zgauss)}
model fit to the FPMA/FPMB spectra showed evidence for variation in $\rm{E_{cut}}$. To analyze the variation of $\rm{E_{cut}}$ with time as well as to find the relation between $\rm{E_{cut}}$ and $\rm{kT_{e}}$ we used the $\rm{E_{cut}}$ values obtained from {\it xillver} model fits. The lowest value of $\rm{E_{cut}}$ of 92$^{+10}_{-8}$ keV was obtained in epoch B and the
highest value of $\rm{E_{cut}}$ was observed in
epoch E (378$^{+152}_{-91}$ keV) while intermediate values of
$\rm{E_{cut}}$ were obtained during other epochs. These observed variation of $\rm{E_{cut}}$ indicates that the coronal temperature of NGC 3227 must be changing with time. This
is also evident in Fig. \ref{figure-6} (upper panel) where
the values of $\rm{E_{cut}}$ are plotted against epochs. To quantify the significance of the variation
of $\rm{E_{cut}}$ we fitted a constant (mean of all the $\rm{E_{cut}}$ measurements)
to the variation of $\rm{E_{cut}}$ with epoch and calculated the $\chi^2$ and the
null hypothesis probability (p) that $\rm{E_{cut}}$ does not change with time (shown
as a dashed line in the top panel of Fig. \ref{figure-6}). We found $\chi^2$/dof $>>$ 10, and a
$p$ value of 0.0 that the $\rm{E_{cut}}$ does not change with time. The values of $\rm{E_{cut}}$
obtained during the first two epochs (A and B) are similar to the value of
$\rm{E_{cut}}$ = 90$\pm$20 keV reported by \cite{2009ApJ...691..922M}.
\cite{2013ApJ...763..111V} from an analysis of XMM and BAT spectra pointed
$\rm{E_{cut}}$ to lie outside the BAT range at $>$ 636 keV. Recently \cite{2021MNRAS.502...80K} too reported $\rm{E_{cut}}$ values for epoch A, B and G, and lower limits for epochs C,D,E and F from {\it pexrav} model fits. The values of $\rm{E_{cut}}$ obtained here using {\it xillver} for epochs A, B and G are in agreement with that of \cite{2021MNRAS.502...80K} from using {\it pexrav} model. Also, our model fits could constrain $\rm{E_{cut}}$ during epochs E and F using {\it xillver}.
Comptonization model fits using {\it xillverCP} provided values of $\rm{kT_{e}}$ which too was found to vary between epochs. We could constrain $\rm{kT_{e}}$ for epochs A, B, C and G, and obtain lower limits for epochs D, E, F, H and I. \cite{2021MNRAS.502...80K} using the same model used here could constrain $\rm{kT_{e}}$ for only epochs A and B. Our results for epochs A and B are in agreement with that of \cite{2021MNRAS.502...80K}. The variation of $\rm{kT_{e}}$ is shown in the lower panel of Fig. \ref{figure-6}. The trend seen in the variation of $\rm{kT_{e}}$ with epochs is
similar to the variation of $\rm{E_{cut}}$. From Comptonization model fits we found
the lowest value of $\rm{kT_{e}}$ =
28$^{+9}_{-6}$ keV during epoch B while the highest value of
$\rm{kT_{e}}$ = 56$^{+131}_{-18}$ keV was obtained for epoch E. We obtained lower limits of $\rm{kT_{e}}$ of 80, 45, 47, 85 and 36 keV during epochs D, E, F, H and I respectively. Spectral fits using {\it xillverCP} along with the data to model ratio for all the epochs are shown in Figures \ref{figure-7} and \ref{figure-8}. The results of the model fits are given in Table \ref{table-2}. To quantify the significance of the variation of $\rm{kT_{e}}$ we fitted a constant (mean of all the $\rm{kT_{e}}$ measurements) to the variation of $\rm{kT_{e}}$ with epoch and calculated the $\chi^2$ and the
null hypothesis probability (p) that $\rm{kT_{e}}$ does not change with time (shown as a dashed line in Fig. \ref{figure-6}). We found $\chi^2$/dof of 12.86, and a $p$ value of 0.005 that $\rm{kT_{e}}$ does not change with time. The contour plots between $\Gamma$ against $\rm{kT_{e}}$ and R against $\rm{kT_{e}}$ for NGC 3227 for all the epochs are shown in Fig. \ref{figure-14} and \ref{figure-15} (left panels).From the contour plots it is evident that the coronal temperature variation is prominent in NGC 3227. We therefore, conclude that we found variation in the temperature of the corona in NGC 3227. According to \cite{2001ApJ...556..716P} for an optically thick corona ($\tau$ $>$ 1) $\rm{E_{cut}}$ = 3$\rm{kT_{e}}$. However, the relation between $\rm{E_{cut}}$ and $\rm{kT_{e}}$ can be complicated in the case of a non-static corona such as the one with outflows \citep{2014ApJ...783..106L}. Also, according to \cite{2019A&A...630A.131M}, the relation of $\rm{E_{cut}}$ = 2$-$3 $\rm{kT_{e}}$ is valid only for low values of $\tau$ and $\rm{kT_{e}}$. For NGC 3227 using all the five epochs in which we could constrain both $\rm{E_{cut}}$ and $\rm{kT_{e}}$ we found $\rm{E_{cut}}$ = 3.94$\pm$0.62 $\rm{kT_{e}}$ which is similar to that of \cite{2001ApJ...556..716P} and \cite{2019A&A...630A.131M}
\begin{table*}
\caption{Results of the correlation analysis between different parameters for NGC 3227. Provided are the slope (m), intercept (c), Pearson's correlation coefficient (r) and the probability (p) for null hypothesis (no correlation) from OLS fit and the least squares fit from simulated points. See text in Section 4.4.}\label{table-5}
\centering
\begin{tabular}{lrrrcrrrc}
\hline
Parameter & \multicolumn{4}{c}{OLS} & \multicolumn{4}{c}{Simulated} \\ \\
\hline
& m & c & r & p & m & c & r & p \\
\hline
$\Gamma$/Flux & -0.07$\pm$0.03 & 1.90$\pm$0.03 & -0.63 & 0.07 & -0.07$\pm$0.03 & 1.91$\pm$0.04 & -0.64 & 0.06 \\
$\rm{kT_{e}}$/Flux & 18$\pm$28 & 22$\pm$31 & 0.42 & 0.58 & 18$\pm$150 & 13$\pm$169 & 0.08 & 0.52 \\
$\Gamma$/$\rm{kT_{e}}$ & 0.001$\pm$0.0006 & 1.75$\pm$0.03 & 0.86 & 0.14 & 0.0005$\pm$0.0002 & 1.78$\pm$0.02 & 0.80 & 0.20 \\
$R$/$\rm{kT_{e}}$ & -0.005$\pm$0.005 & 1.03$\pm$0.22 & -0.57 & 0.43 & -0.001$\pm$0.002 & 0.90$\pm$0.17 & -0.36 & 0.60 \\
$R$/Flux & -0.51$\pm$0.13 & 1.31$\pm$0.13 & -0.84 & 0.01 & -0.04$\pm$0.15 & 0.78$\pm$0.16 & -0.09 & 0.68 \\
$\tau$/Flux & -0.29$\pm$0.62 & 2.35$\pm$0.65 & -0.17 & 0.66 & -0.98$\pm$1.88 & 3.15$\pm$2.12 & -0.35 & 0.65 \\
$y$/Flux & -0.15$\pm$0.70 & 13.06$\pm$0.79 & -0.15 & 0.85 & -0.53$\pm$1.50 & 14.01$\pm$1.71 & -0.21 & 0.66 \\
$\tau$/$\rm{kT_{e}}$ & -0.043$\pm$0.002 & 4.19$\pm$0.09 & -0.99 & 0.00 & -0.02$\pm$0.00 & 3.44$\pm$0.28 & -0.97 & 0.03 \\
\hline\hline
\end{tabular}
\end{table*}
\subsection{MR 2251$-$178}
This source has five epochs of observations that are public and having exposure $>$ 20 ks. In this work we analyzed all of them. Simple power law fits to the FPMA spectra of all the epochs is shown in the right panel of Fig. \ref{figure-1}. From this figure, noticeable
change in the spectra could not be ascertained. The ratio of the model fits {\it const*TBabs(zpo)}, {\it const*TBabs(pexrav)} and {\it const*TBabs(xillverCP)} to the observed FPMA and FPMB spectra on OBSID 90601637002, the one with the longest exposure time are shown in Fig. \ref{figure-4}. From two sets of observations from {\it Einstein} separated by about a year \cite{1984ApJ...281...90H} found evidence of variable X-ray absorption
in MR 2251$-$178, with the column density changing from
$<$ 5 $\times$ 10$^{21}$ cm$^{-2}$ to 2 $\times$ 10$^{22}$ cm$^{-2}$,
suggested to be due to the presence of a warm absorber. {\it EXOSAT} and
{\it Ginga} observations revealed strong correspondence between the absorbing
column density and the flux of the source with the low energy absorption
decreasing with the increasing flux of the source. These observations were
explained by variable absorption in photo-ionized gas along the line of sight
\citep{1993MNRAS.262..817M,1990MNRAS.242..177P}. However,
\cite{1992A&A...266...57W} from an analysis of the {\it EXOSAT} data argued
that the variability seen in the source can be explained without invoking the
presence of a warm absorber. From the ratio of the observed data to the model fit we did not find any signature of absorption that could affect the source spectra in all observations but in epoch E and to take care of this we added the {\it zTBabs} component to all the four models in epoch E(see Fig \ref{figure-11} and \ref{figure-12}). Model fits to the four sets of observations that span about five years using {\it const*TBabs(zpo)} do not reveal the presence of FeK$\alpha$ line in the spectra, the reflection bump was also found to be either negligible or weak (see top panel of Fig. \ref{figure-4}), more likely due to poor S/N. However, the ratio plot for the model {\it const*TBabs(zpo)} to the epoch D spectra revealed the presence of the FeK$\alpha$ line at around 6.4 keV. We therefore added a {\it zgauss} component with the model and found the energy of the line at 6.49$^{+0.37}_{-0.27}$ keV with a fixed width of 0.1 keV, letting the parameter free did not significantly improve the fit. For this source previously the FeK$\alpha$ line was reported to be present in the {\it Ginga} observations with an equivalent width of 125$^{+100}_{-105}$ eV \citep{1993MNRAS.262..817M}. Relatively strong FeK$\alpha$ was also reported to be present in the {\it BeppoSaX} observation \citep{2001A&A...376..413O} and a narrow FeK$\alpha$ line was present in the {\it Chandra} observations \citep{2005ApJ...627...83G}. From {\it BeppoSAX} observations in the 0.1$-$200 keV band, \cite{2001A&A...376..413O} found a $\rm{E_{cut}}$ value of around 100 keV which is similar to that obtained here.
\begin{figure}
\vbox{
\includegraphics[scale=0.48]{NGC3227_flux.eps}
}
\vspace{-0.34cm}
\caption{Correlation of $\Gamma$ and $\rm{kT_{e}}$ with the X-ray brightness of
NGC 3227. The orange dashed line is the ordinary linear
least squares (OLS) fit to the data points. The black dashed line is the linear line drawn using the median values of the simulated parameters. The grey shaded region indicates the upper and lower errors in the fit parameters for OLS and the green region indicates the errors in the fit parameter obtained from the simulation. In the least square analysis, epochs in which we were unable to constrain $\rm{kT_{e}}$ were dropped (indicated with red point).} \label{figure-13}
\end{figure}
To find evidence for the change in $\rm{kT_{e}}$ if any, we carried out
fitting of the observations with the physical model, {\it const*TBabs(xillverCP)}.
Examination of the results of the fit (Table \ref{table-3}) shows that
the $\rm{kT_{e}}$ obtained during all the epochs agree
within errors. Though we could not find any signature of $\rm{kT_{e}}$ variation with epochs from the {\it xillverCP} fit, the $\rm{E_{cut}}$ values obtained from the {\it xillver} model fit during epochs A, C and D agree within errors except that of epoch B. This could be due to the
quality of the data in epoch B, as the values of $\rm{E_{cut}}$ and $\rm{kT_{e}}$ obtained during that epoch also have large error bars. To test for the $\rm{kT_{e}}$ variation if any, we plotted the 90 per cent contours between $\rm{kT_{e}}$ and $\Gamma$. The 90 percent contours of $\rm{kT_{e}}$ against $\Gamma$ overlap (see Fig. \ref{figure-14}) and we conclude that in MR 2251$-$178 we did not find any variation of $\rm{kT_{e}}$ with time. The 90 per cent contours of R against $\rm{kT_{e}}$ are also shown in the bottom panel of Fig. \ref{figure-15}. Considering all the four epochs where $\rm{E_{cut}}$ and $\rm{kT_{e}}$ could be constrained we found $\rm{E_{cut}}$ = 4.84$\pm$0.11 $\rm{kT_{e}}$. This is deviant from the generally adopted $\rm{E_{cut}}$ = 2$-$3 $\rm{kT_{e}}$ \citep{2001ApJ...556..716P}. It is likely the relation between $\rm{E_{cut}}$ and $\rm{kT_{e}}$ is complex and may depend on other physical properties of the sources. Homogeneous analysis of a large number of sources are needed to establish the relation between $\rm{E_{cut}}$ and $\rm{kT_{e}}$ as well as on its dependence on other physical properties. Spectral fits using {\it xillverCP} along with the residuals for all the epochs are shown in Fig. \ref{figure-11} and \ref{figure-12}.
\begin{figure*}
\vbox{
\hspace{-0.4cm}\includegraphics[scale=0.58]{xg_ngc3227_new.eps}
\hspace{-0.4cm}\includegraphics[scale=0.58]{xg_mr_new.eps}
\hspace{-0.4cm}\includegraphics[scale=0.58]{xg_ngc5548_new.eps}
}
\vspace{-0.5cm}
\caption{The 90 per cent confidence level contours between $\rm{kT_{e}}$ and
$\Gamma$ for the {\it xillverCP} model fit to the FPMA/FPMB spectra of
NGC 3227(left), MR 2251$-$178(middle) and NGC 5548(right). The colours indicate different OBSIDs. Here for MR 2251$-$178 the limits of $\rm{kT_{e}}$ do not have a correspondence with the values given in Table \ref{table-3}. This is because the contours were generated by freezing the $R$ parameter so as to put all the five OBSIDs together.}
\label{figure-14}
\end{figure*}
\begin{figure*}
\vbox{
\hspace{-0.4cm}\includegraphics[scale=0.58]{xr_ngc3227_new.eps}
\hspace{-0.4cm}\includegraphics[scale=0.58]{xr_mr_new.eps}
\hspace{-0.4cm}\includegraphics[scale=0.58]{xr_ngc5548_new.eps}
}
\vspace{-0.5cm}
\caption{The 90 per cent confidence level contours between R and $\rm{kT_{e}}$ for the {\it xillverCP} model fit to the FPMA/FPMB spectra of NGC 3227(left), MR 2251$-$178(middle) and NGC 5548(right).}
\label{figure-15}
\end{figure*}
\begin{figure}
\includegraphics[scale=0.45]{Gamma_R_kTe_3227.eps}
\caption{Correlation between $\Gamma$ and $R$ with $\rm{kT_{e}}$ of
NGC 3227. The dashed lines and shaded regions have the same meaning as in Fig. \ref{figure-13}. Epochs in which $\rm{kT_{e}}$ could not be constrained (shown as red points) were not included in the linear least square analysis.} \label{figure-16}
\end{figure}
\subsection{NGC 5548}
NGC 5548 was observed by {\it NuSTAR} six times between July 2013 to January 2021. Of these, \cite{2018ApJ...863...71Z} have reported results for five epochs. In this work we carried out both phenomenological and physical model fits for all the six epochs. We fit the spectra using {\it const*TBabs(zpo+zgauss)} and {\it const*TBabs(pexrav+zgauss)} to model the primary power law emission and
the reflection component along with the FeK$\alpha$ line with the width of the line being frozen to the value of 0.1 keV. From the ratio of the observed data to the model fit using
{\it TBabs} we found the presence of absorption component in the
spectra. So unlike the other sources, for NGC 5548, we added an extra
component {\it zTBabs} in each model to account for the absorption by the host
galaxy. Moreover, as {\it xillver} self consistently models the reflected spectrum and its associated FeK$\alpha$ line, we also fit the spectra using {\it xillver} and $\rm{E_{cut}}$ values obtained using {\it xillver} model were used to find variation in $\rm{E_{cut}}$. The $\rm{E_{cut}}$ obtained from {\it xillver} model fits are $>487$ keV, $>480$ keV, 129$^{+15}_{-13}$ keV, 179$^{+40}_{-23}$ keV, 133$^{+15}_{-13}$ and $>$394 keV for epochs A, B, C, D, E and F respectively. The residuals of the fit to the spectra for the epoch D that has the maximum exposure time, for various models are given in Fig. \ref{figure-3} and the results of the fit are given in Table \ref{table-4}. From {\it xillver} model fits, we could constrain $\rm{E_{cut}}$ only for epochs C, D and E, while \cite{2018ApJ...863...71Z} could constrain $\rm{E_{cut}}$ in the four epochs (A, C, D and E). For the epochs in common, though our values of $\rm{E_{cut}}$ are in agreement with \cite{2018ApJ...863...71Z} the results do not match for epoch E. Though \cite{2018ApJ...863...71Z} claims to have detected $\rm{E_{cut}}$ variation in NGC 5548, our analysis could not confirm changes in $\rm{E_{cut}}$. This could be due to differences in the choice of binning and the energy ranges used in both the work. \cite{2015A&A...577A..38U} via the joint fitting of {\it XMM-Newton}, {\it Chandra}, {\it NuSTAR}, and {\it INTEGRAL} data obtained lower limits for $\rm{E_{cut}}$ in all the epochs except for epoch D. Using the simultaneous {\it XMM-Newton} and {\it NuSTAR} data from 2013 campaign \cite{2016A&A...592A..27C} also fitted the 4$-$79 keV epoch A/B , epoch D and epoch E spectra using cut-off power-law and {\it pexmon}. Their values of $\rm{E_{cut}}$ are in agreement with the $\rm{E_{cut}}$ values obtained in this work.
Fitting the physical Comptonization model {\it xillverCP} to ascertain the change in coronal temperature, we got the highest $\rm{kT_{e}}$ value of 65$^{+147}_{-24}$ keV during epoch D, for epochs A, B and F we obtained lower limits of 53, 54 and 71 keV, while for the remaining two epochs C and E, we obtained similar values of 39$^{+14}_{-10}$ keV and 38$^{+12}_{-9}$ keV respectively. The values of $\rm{kT_{e}}$ between epochs agree to each other within error bars. The model fits to the data along with the data to model ratio for all the epochs of observations are shown in Fig. \ref{figure-9} and \ref{figure-10}. The $\rm{E_{cut}}$ values obtained here points to variability. However, the derived $\rm{kT_{e}}$ values are consistent with each other within error bars. The contour plots between $\rm{kT_{e}}$ against $\Gamma$ as well as R against $\rm{kT_{e}}$ are shown in the right most panels of Fig. \ref{figure-14} and \ref{figure-15}. From these plots, the change in $\rm{kT_{e}}$ is not evident. We thus conclude that we have not found evidence for variation in $\rm{kT_{e}}$ in NGC 5548. Considering all the epochs in which we could constrain both $\rm{E_{cut}}$ and $\rm{kT_{e}}$, we found $\rm{E_{cut}}$ = 3.19$\pm$0.32 $\rm{kT_{e}}$ which is in agreement with the empirical relation of \cite{2001ApJ...556..716P}.
\subsection{Correlation between different parameters}
We discuss below the correlation between various parameters. This is restricted to the source NGC 3227 as the temperature of its corona is found to vary in this work. Since the errors in the measured $\rm{E_{cut}}$ and $\rm{kT_{e}}$ values are not symmetric and there is no conventional way to take care of such errors during correlation study we adopted two procedures to find the correlation between various parameters.
In first case, we neglected the corresponding errors and considered only the best fit values of the parameters and performed the ordinary linear least square (OLS) fit between them. The Pearson's correlation coefficient (r) and the null hypothesis probability (p) for no correlation were also estimated using the best fit values.
In the second case, to take into account the non-symmetric errors we simulated $10^{5}$ points from each rectangular box around the best fit values (x,y) with x and y boundaries of ($\rm{x_{low}}$,$\rm{x_{high}}$) and ($\rm{y_{low}}$,$\rm{y_{high}}$) respectively. Here $\rm{x_{low}}$, $\rm{y_{low}}$ and $\rm{x_{high}}$, $\rm{y_{high}}$ refer to the lower and upper errors in x and y values. Linear least squares fit was done for $10^{5}$ times yielding distribution of the slope (m), the intercept (c), the Pearson's rank correlation coefficient and the probability of no correlation. The median values of the distributions were taken to represent the best fit values of the correlation. All the values and the errors for the unweighted as well as for the simulated correlation are given in Table \ref{table-5}.
\subsubsection{$\Gamma$ v/s Flux}
In Seyfert galaxies, the X-ray spectra are generally found to be softer with increasing X-ray flux \citep{2003ApJ...593...96M}. We show in Fig. \ref{figure-13}, the correlation between $\Gamma$ and the brightness of NGC 3227. For the source, each point in the Figure corresponds to $\Gamma$ and flux obtained by {\it xillverCP} model fits to each epoch of spectra. Ordinary linear least squares fit to the data are shown by orange dashed lines. The black dashed lines show the linear least square fit using the median values of the simulated points using the lower and upper errors in $\Gamma$ and flux values. The grey and green shaded regions are the area bounded by the errors in the least square fit parameters. For NGC 3227, we found anti correlation between $\Gamma$ and flux (see Fig. \ref{figure-13} and Table \ref{table-5}) between epochs of observations significant at the greater than 90 per cent level.
\subsubsection{$\rm{kT_{e}}$ v/s Flux}
Correlation between the changes in the temperature of the corona with other physical parameters of the sources such as its apparent brightness as well as its spectral shape can provide important
constraints in enhancing our understanding of AGN corona. The correlation between $\rm{kT_{e}}$ and flux is shown in Fig. \ref{figure-13}. Also, shown in the figure are the ordinary and simulated linear least squares fit to the data. No correlation is found between $\rm{kT_{e}}$ and flux in NGC 3227.
\subsubsection{$\Gamma$ v/s $\rm{kT_{e}}$}
In Fig. \ref{figure-16} the correlation between $\Gamma$ and $\rm{kT_{e}}$ is shown. The orange and black dashed lines represent the ordinary and simulated least square fit to the data. The correlation between $\Gamma$ and $\rm{kT_{e}}$ is found to be not significant.
\subsubsection{$R$ v/s $\rm{kT_{e}}$}
The correlation between the distant reflection fraction $R$ and $\rm{kT_{e}}$ is presented in the bottom panel of Fig. $\ref{figure-16}$. We did not find any correlation between these two parameters.
\subsection{Nature of corona in AGN}
The primary X-ray emission from thermal Comptonization
depends on the optical depth $\tau$ and $\rm{kT_{e}}$ as
\citep{1996MNRAS.283..193Z,1999MNRAS.309..561Z},
\begin{equation}
\tau = \sqrt{\frac{9}{4} + \frac{3}{\theta\Big[\Big(\Gamma + \frac{1}{2}\Big)^2 - \frac{9}{4}\Big]}} - \frac{3}{2} \\
\end{equation}
where, $\theta = {kT_e}/{m_{e}c^2}$. We show in Fig. \ref{figure-18} the variation
of the reflection fraction, the optical depth and the Compton \say{$y$} parameter
with flux. The Compton $y$ parameter is defined
as \citep{2001ApJ...556..716P},
\begin{equation}
y \simeq 4\Bigg(\frac{4kT_e}{mc^2}\Bigg)\Bigg[1+\Bigg(\frac{4kT_e}{mc^2}\Bigg)\Bigg] \tau(1+\tau)
\end{equation}
According to \cite{1995ApJ...449L..13S}, a Comptonized corona must have a constant
$y$ in equilibrium. We too found no correlation of $y$ with the flux of the source. The parameter $\tau$ is found not to show any statistically significant variation with flux. $R$ is found to be not correlated with flux(see Fig. \ref{figure-17}).
\begin{figure}
\vbox{
\includegraphics[scale=0.50]{change_NGC3227_lsf.eps}
}
\caption{Distant reflection fraction $R$, the optical
depth $\tau$ and Compton-y parameter as a function of flux for the sources NGC 3227. The dashed lines and the shaded regions have the same meaning as in Fig. \ref{figure-13}. Points shown as red in the Figure were not included in the linear least squares fit.} \label{figure-17}
\end{figure}
Multiple epochs of observations on few AGN available today point to increased
$\rm{E_{cut}}$ with the flux of the sources. For example,
from analysis of {\it NuSTAR} and XMM-Newton observations of NGC 5548,
\cite{2018ApJ...863...71Z} found $\rm{E_{cut}}$ to be positively correlated
with the flux of the source. Similarly in Mrk 335 too, \cite{2016MNRAS.456.2722K}
through joint fitting of {\it Suzaku} and {\it NuSTAR} found $\rm{E_{cut}}$ to be positively
correlated with flux. Recently from flux resolved spectroscopy of Ark 564,
\cite{2020MNRAS.492.3041B} found the temperature of the corona to decrease with
increasing flux. Also, for the source ESO 103$-$035 \cite{2021ApJ...921...46B} found a positive correlation between $\rm{kT_{e}}$ and flux. Scenarios that could cause the change in the temperature of the corona or $\rm{E_{cut}}$ in
AGN are (a) Compton cooling and (b) expanding corona. In the Compton
cooling scenario, as the source brightens, there will be increased
seed UV photons from the disk for Comptonization leading to a cooling
of the corona. This leads to a
\say{cooler when brighter behaviour}.
In the
expanding corona scenario, the increase in $\rm{kT_{e}}$ during high flux
states of the AGN could be due to changes in the geometry of the corona.
This has been invoked to explain the $\rm{E_{cut}}$ changes in
Mrk 335 \citep{2016MNRAS.456.2722K} and NGC 5548 \citep{2018ApJ...863...71Z} .
According to the expanding corona model, at low flux state,
the corona is warm, optically thick, compact and close to the black hole. This
causes more illumination of the disk leading to larger reflection
fraction. As the source increases in
brightness, the corona
expands, the optical depth drops and the temperature rises. A reduced
reflection fraction during this period argues that the corona expands
vertically from the disk. In NGC 3227 too, the corona temperature is
found to vary with time. We found a negative correlation between the optical depth $\tau$ and $\rm{kT_{e}}$, with $\tau$ decreasing towards higher temperatures (see Fig. \ref{figure-18}). We calculated $\tau$ using Equation 7 and according to \cite{1996MNRAS.283..193Z} $\tau$ is geometry dependent and equals the radial optical depth in an uniform sphere. The negative correlation between $\tau$ and $\rm{kT_{e}}$ argues for a change in
the geometry of the corona \citep{2014ApJ...794...62B}. \cite{2018A&A...614A..37T} too from the analysis of a sample of AGN found a negative correlation between $\tau$ and $\rm{kT_{e}}$. According to the authors this negative correlation could not be explained with a fixed disk-corona configuration in radiative balance. The possible explanation for this kind of behaviours could be (a) due to the change in the geometry and position of the corona and/or (b) variation in the fraction of the intrinsic disk emission to the total disk emission. We note that broad band spectral energy distribution fits to simultaneous UV to hard X-ray data alone will be able to provide strong constrain on $\tau$. This in turn can put constrains on the role of accretion disk emission to $\rm{kT_{e}}$ changes. In NGC 3227 we found no statistically significant correlation between $\rm{kT_{e}}$ and flux, $R$ and flux, $\Gamma$ and flux as well as $\Gamma$ and $\rm{kT_{e}}$.
\section{Summary}
In this work, we carried out spectral analysis of the {\it NuSTAR} data
for three Seyfert type AGN namely NGC 3227, NGC 5548 and MR 2251$-$178 including a few OBSIDs not analyzed yet in comparison to previous works. We summarize our results below
\begin{enumerate}
\item All the sources were found to show moderate variations in their average brightness during the epochs analyzed in this work.
\item In NGC 3227, we found unambiguous evidence for the change in the
temperature of the corona. This change in $\rm{kT_{e}}$ is also reflected in the variation in $\rm{E_{cut}}$.
For NGC 5548 and MR 2251$-$178 we found no evidence for the variation in the temperature of the corona.
\end{enumerate}
Our knowledge on the variation in the temperature of the corona is known only for less than half a dozen sources. Details on such coronal temperature variation in more AGN are needed to pin point the reasons for the temperature of the corona to vary and its effect on the other physical properties of the sources.
\section{Acknowledgements}
We thank the anonymous referee for her/his useful comments and suggestions which improved the quality and the clarity of the paper. We thank the {\it NuSTAR} Operations, Software and Calibration teams for support with the execution and analysis of these observations. This research has made use of the {\it NuSTAR} Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (USA). This research has made use of data and/or software provided by
the High Energy Astrophysics Science Archive Research Center (HEASARC), which is a service of the Astrophysics Science Division at NASA/GSFC.
\begin{figure}
\vbox{
\includegraphics[scale=0.54]{odkt_ngc3227.eps}
}
\caption{Correlation between $\tau$ and $\rm{kT_{e}}$ of
NGC 3227. The dashed lines and the shaded regions have the same meaning as in Fig. \ref{figure-13}. Points shown as red in the Figure were not considered for the linear least squares fit.} \label{figure-18}
\end{figure}
\bibliographystyle{aa}
| 9a611fc17f00f8d2625a97931c06829cf0b136b8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Todo for revision}
\section{Introduction}
Zero-shot cross-lingual transfer is the ability of a model to learn from labeled data in one language and transfer the learning to another language without any labeled data. Transformer ~\cite{NIPS2017_3f5ee243} based multilingual models pre-trained on unlabeled data from multiple languages are the state-of-the-art means for cross-lingual transfer \cite{ruder-etal-2019-transfer,Devlin2019}. While pre-training based cross-lingual transfer holds great promise for low web-resource languages (LRLs), such techniques are found to be more effective for transfer within high web-resource languages (HRLs) \cite{wu-dredze-2020-languages}.
Vocabulary generation is an important step in multilingual model training, where vocabulary size directly impacts model capacity. Usually, the vocabulary is generated from a union of HRL{} and LRL{} data. This often results in under-allocation of vocabulary bandwidth to LRL{}s, as LRL{} data is significantly smaller in size compared to HRL{}. This under-allocation of model capacity results in lower LRL{} performance \cite{wu-dredze-2020-languages}, as mentioned previously. In response, prior research has explored development of region-specific models \cite{antoun2020arabert,khanuja2021muril}, generating vocabulary specific to language clusters \cite{chung-etal-2020-improving}, and exploring relatedness among languages to build better LMs for LRL{}s \cite{khemchandani-etal-2021-exploiting}. However, none of these methods have utilized relatedness among languages for better vocabulary generation during multilingual pre-training.
\begin{table}[!t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|p{2.5cm}|p{6cm}|} \hline
Language and Token frequencies & English: {\color{red}Universit}y (10), versity (6); German: {\color{red}Universit}aten (2); Dutch: {\color{red}Universit}eit (1); Western Frisian: {\color{red}Universit}eiten (1) \\
\hline
Starting Vocab & Uni, versit, U,n,i,v,e,r,s,i,t,y,a \\
\hline
BPE Vocab & versity, Uni, versit, U,n,i,v,e,r,s,i,t,y,a \\ \hline
\textsc{OBPE}{} Vocab & {\color{red}Universit}, Uni, versit, U,n,i,v,e,r,s,i,t,y,a \\
\hline
\end{tabular}
}
\caption{\label{tab:overlap-example}First row shows lexically overlapping tokens in four different languages with their corpus frequencies (in brackets), with English (En) as the High Web-Resource Language (HRL{}). From a starting vocabulary shown in the second row, BPE merges tokens based on greater overall frequency, adding new vocabulary item \textit{versity} as it has the highest overall frequency (16). \textsc{OBPE}{} instead adds \textit{Universit} since it also rewards cross-lingual overlap, even though \textit{Universit} has lower overall frequency (15).
}
\end{table}
In this paper, we hypothesize that exploiting language relatedness can result in an overall more effective vocabulary, which is also better representative of LRLs.
Closely related languages (e.g., languages belonging to a single family) have common origins for words with similar meanings. We show some examples across three different families of related languages in Table~\ref{tab:overlap-exampleAll}.
Morphological inflections of the root word lead to lexically overlapping tokens across languages. Learning representations for such subwords in lexically overlapping words
shared across HRL and its related LRLs can enable better transfer of supervision from HRL to LRLs.
During Masked Language Modelling (MLM) pretraining \cite{Devlin2019}, the shared tokens can serve as anchors in learning contextual representations of neighboring tokens. However, choosing the correct granularity of sharing automatically is tricky. On one extreme, we can choose a vocabulary which favours longer units frequent in HRL without regard for sharing,
thereby leading to better semantic representation of the tokens but no cross-lingual transfer. On the other extreme, we can choose character-level vocabulary \cite{ma-etal-2020-charbert}, where every token is shared across languages but have no semantic significance.
%
Given
text from a mix of high and low Web-resource languages (HRL and LRL, respectively), Byte Pair Encoding (BPE)~\cite{sennrich-etal-2016-neural} and its variants like Wordpiece \cite{schuster2012japanese} and Sentencepiece \cite{kudo-richardson-2018-sentencepiece}
prefer frequent tokens, most of those
from the HRLs.
This would cause most long HRL tokens to get included, leaving only a limited budget of short tokens for the LRL. Any sub-token level overlap between HRL and LRL could get lost in this process. In a zero-shot setting, since available supervision is HRL based, this creates a bottleneck when transferring supervision from HRL to LRLs. Oversampling LRLs is a common strategy to offset this imbalance but that hurts HRL performance as shown in \cite{conneau-etal-2020-unsupervised}.
In this paper, we propose Overlap BPE (\textsc{OBPE}{}). \textsc{OBPE}{} chooses a vocabulary
%
by giving token overlap among HRL and LRLs a primary consideration. \textsc{OBPE}{} prefers vocabulary units which are shared across multiple languages, while also encoding the input corpora compactly. Thus, \textsc{OBPE}{} tries to balance the trade-off between cross-lingual subword sharing and the need for robust representation of individual languages in the vocabulary. This results in a more balanced vocabulary, resulting in improved performance for LRLs without hurting HRL accuracy. Table~\ref{tab:overlap-example} shows an example to highlight this difference between \textsc{OBPE}\ and BPE.
Recently ~\citet{K2020Cross-Lingual,conneau-etal-2020-emerging} concluded that token overlap is unimportant for cross-lingual transfer. However, they studied language pairs where either both languages had a large corpus, or where the languages were not sufficiently related. We focus on related languages within a family and observe drastic drop in zero-shot accuracy when we synthetically reduce the overlap to zero (58\% F1 drops to 17\% for NER, 71\% drops to 30\% for text classification).
This paper offers the following contributions
\begin{itemize}
\item
We present \textsc{OBPE}{}, a simple yet effective modification to the popular BPE algorithm to promote overlap between LRLs and a related HRL during vocabulary generation. \textsc{OBPE}{} uses a generalized mean based formulation to quantify token overlap among languages.
\item
We evaluate \textsc{OBPE}{} on twelve languages across three related families, and show consistent improvement in zero-shot transfer over state-of-the art
baselines on four NLP tasks.
We analyse the reasons behind the gains obtained by OBPE and show that OBPE increases the percentage of LRL tokens in the vocabulary without reducing HRL tokens. This is unlike over-sampling strategies where increasing one reduces the other.
\item
Through controlled experiments on the amount of token overlap on a related HRL-LRL pair, we show that token overlap is extremely important in the low-resource, related language setting. Recent literature which conclude that token overlap is unimportant may have overlooked this important setting.
\end{itemize}
The source code for our experiments is available
at \href{https://github.com/Vaidehi99/OBPE}{https://github.com/Vaidehi99/OBPE}.
\section{Related Work}
Transformer-based multilingual language models such as mBERT \cite{devlin-etal-2019-bert} and XLM-R \cite{conneau-etal-2020-unsupervised} are now established as the de-facto method for zero-shot cross-lingual transferability, and thus hold promise for low resource domains.
However, recent studies have indicated
that even the current state-of-the-art models such
as XLM-R (Large) do not yield reasonable
transfer performance across low resource target languages with limited data~\cite{wu-dredze-2020-languages}.
This has led to a surge of interest in enhancing cross-lingual transfer of multilingual models to the low-resource setting. We categorize existing work based on the stage of the pre-training pipeline where it is relevant:
\noindent{\bf Input Data}
In the data creation stage, \citet{conneau-etal-2020-unsupervised} propose over-sampling of LRL documents to improve LRL representation in the vocabulary and pre-training steps. \citet{khemchandani-etal-2021-exploiting} specifically target related languages and propose transliteration of LRL documents to the script of related HRL for greater lexical overlap. We deploy both these tricks in this paper.
\noindent{\bf Tokenization}
\citet{rust-etal-2021-good} study that even the tokenization step could have a crucial impact on performance accrued to each language in a multilingual models. They propose the use of dedicated tokenizer for each language instead of the automatically generated multilingual mBERT tokenizer. However, they continue to use the default mBERT vocabulary generator.
\noindent{\bf Vocabulary Generation}
\citet{sennrich-etal-2016-neural} highlighted the importance of subword tokens in the vocabulary and proposed use of the BPE algorithm~\cite{gage1994new} for efficiently growing such a vocabulary incrementally. Variants like Wordpiece \cite{schuster2012japanese} and Sentencepiece \cite{kudo-richardson-2018-sentencepiece} either build on top of BPE or follow a very similar process.
\citet{kudo-2018-subword} is a variant method that chooses tokens based on unigram LM score. We obtained better results with BPE and continued with that. All these BPE variants incrementally add subwords based on overall frequency in the combined corpus, and they all ignore language boundaries.
\citet{chung-etal-2020-improving} observed that such a combined approach could under-represent several languages, and proposed instead to separately create vocabularies for clusters of related languages and take a union of each cluster-specific vocabulary. However, within each cluster they continue to use the default vocabulary generator. Our approach can be used as a drop-in replacement to further enhance the quality of the cluster-specific vocabulary that they obtain.
\citet{wang2018multilingual,gao-etal-2020-improving} propose a soft-decoupled encoding approach for exploiting subword overlap between LRLs and HRLs. However, their focus is NMT models and does not easily integrate in existing multilingual models such as mBERT. \cite{maronikolakis-etal-2021-wine-v} targets tokenization compatibility based purely on vocabulary size and does not focus on choosing the tokens that go in the vocabulary.
\noindent{\bf Pre-Training and Adaptation}
Several previous works have proposed to include additional alignment loss between parallel~\cite{DBLP:conf/iclr/CaoKK20} or pseudo-parallel~\cite{khemchandani-etal-2021-exploiting} sentences to co-embed HRLs and LRLs. Another approach is to design language-specific Adapter layers~\cite{pfeiffer-etal-2020-adapterhub, pfeiffer-etal-2020-mad,artetxe-etal-2020-cross,ustun-etal-2020-udapter} that can be easily fine-tuned for each new language.
~\citet{pfeiffer-etal-2021-unks} leverages the pre-trained embeddings of lexically overlapping tokens between the vocabulary of pre-trained model and that of unseen target language to initialize the corresponding embeddings of target language. However, they did not attempt to increase the fraction of such tokens in the vocabulary.
We are not aware of any prior work that explicitly promotes overlapping tokens between LRLs and HRLs in the vocabulary of multilingual models.
\section{Overlap-based Vocabulary Generation}
We are given monolingual data ${D_{1},...,D_n}$ in a set of $n$ languages $\mathcal{L}=\{{L_{1},...,L_n}\}$ and a vocabulary budget $\text{V}$. Our goal is to generate a vocabulary $\mathcal{V}$ that when used to tokenize each $D_i$ in a multilingual model would provide cross-lingual transfer to LRL s from related HRL s. We use $\cL_\text{\lrl}$ to denote the subset of the $n$ languages that are low-resource, the remaining languages $\mathcal{L}-\cL_\text{\lrl}$ are denoted as the set $\cL_\text{\hrl}$ of high resource languages.
Existing methods of vocabulary creation start with a union $D$ of monolingual data ${D_{1},...,D_n}$, and choose a vocabulary $\mathcal{V}$ that most compactly represents $D$. We first present an overview of BPE, a popular algorithm for vocabulary generation
\subsection{Background: BPE}
Byte Pair Encoding (BPE) \cite{gage1994new} is a simple data compression technique that chooses a vocabulary $\mathcal{V}$ that minimizes total size of $D=\cup_i D_i$ when encoded using $\mathcal{V}$.
\begin{equation}
\label{eq:bpe:goal}
\mathcal{V} = \argmin\limits_{S: |S|=\text{V}}\sum_{i=1}^n \lvert \text{encode}(D_{i},S) \rvert
\end{equation}
The size of the encoding $\lvert \text{encode}(D_{i},S) \rvert$ can be alternately expressed as the sum of frequency of tokens in $S$ when $D_i$ is tokenized using $S$.
This motivates the following efficient greedy algorithm to implement the above optimization~\cite{sennrich-etal-2016-neural}. Let $f_{ki}$ denote the frequency of a candidate token $k$ in the corpus $D_i$ of language $L_i$.
The BPE algorithm grows $\mathcal{V}$ incrementally. Initially, $\mathcal{V}$ comprises of characters in $D$. Then, until $|\mathcal{V}| \le \text{V}$, it chooses the token $k$ obtained by merging two existing tokens in $\mathcal{V}$ for which the frequency in $D$ is maximum.
\vspace{-0.3cm}
\begin{equation}
\label{eq:bpe:step}
\mathcal{V} = \mathcal{V} \cup arg\,max_{k=[u,v]:u,v\in \mathcal{V}} \sum_i f_{ki}
\end{equation}
A limitation of BPE on multilingual data is that tokens that appear largely in low-resource $D_i$ may not get added to $\mathcal{V}$, leading to sentences in $L_i$ being over-tokenized.
For a low resource language, the available monolingual data $D_i$ is often orders of magnitude smaller than another high-resource language. Models like mBERT and XLM-R address this limitation by over-sampling documents of low-resource languages. However, over-sampling LRL s might compromise learned representation of HRL s where task-specific labeled data is available.
We propose an alternative strategy of vocabulary generation called \textsc{OBPE}{} that seeks to maximize transfer from HRL\ to LRL.
\begin{algorithm}[t]
\begin{small}
\caption{Overlap based BPE (\textsc{OBPE}{})}
\begin{algorithmic}
\For{$i \in \{1,2,...,n\}$}
\State Split words in $D_{i}$ into characters $C_{i}$ with a special marker after every word
\EndFor\\
$\mathcal{V}$ = $\cup_{i=1}^n C_{i}$
\While{$\lvert\mathcal{V}\rvert < \text{V}$}
\State Update token and pair frequency on $\{D_i\},\mathcal{V}$
\State Add to $\mathcal{V}$ token $k$ formed by merging pairs $u,v \in \mathcal{V}$ \hspace{1cm}with the largest value of
\vspace{-0.5cm}
\begin{equation*}
(1-\alpha)\sum_{j}f_{kj} + \alpha\sum\limits_{i\in \cL_\text{\lrl}{}}\max\limits_{h \in \cL_\text{\hrl}{}}\left(\frac{f^p_{ki} + f^p_{kh}}{2}\right)^{\frac{1}{p}}
\end{equation*}
\EndWhile
\end{algorithmic}
\end{small}
\end{algorithm}
\subsection{Our Proposal: OBPE}
The key idea in OBPE{} is to maximize the overlap between an LRL\ and a closely related HRL\ while simultaneously encoding the input corpora compactly as in BPE. When labeled data $D^T_h$ for a task $T$ is available in an HRL\ $L_h$, then a multilingual model fine-tuned with $D^T_h$ is likely to transfer better to a related LRL\ $L_i$ when $L_i$ and $L_h$ share several tokens in common. Thus, the objective that OBPE\ seeks to optimize when creating a vocabulary is:
\begin{equation}
\begin{aligned}
\label{eq:obpe:goal}
\mathcal{V} = & \argmin \limits_{S: |S|=\text{V}} \left[ (1-\alpha) \sum_{i=1}^n \lvert \text{encode}(D_{i},S) \rvert \right. \\ &-\left. \alpha\sum\limits_{i\in \cL_\text{\lrl}}\max\limits_{j\in \cL_\text{\hrl}}\text{overlap}(L_{i}, L_{j}, S) \right]
\end{aligned}
\end{equation}
where $0 \le \alpha \le 1$ determines importance of the two terms. The first term in the objective compactly represents the total corpus, as in BPE's (Eq~\eqref{eq:bpe:goal}). The second term additionally biases towards vocabulary with greater overlap of each LRL to one HRL\ where we expect task-specific labeled data to be present. There are several ways in which we can measure the overlap between two languages with respect to a current vocabulary. First, we encode each of $D_i$ and $D_j$ using the vocabulary $S$, which then yields a multiset of tokens in each corpus. Inspired by the literature on fair allocation~\cite{barman2021universal}, we explore a continuously parameterized function that expresses overlap between two languages' encoding as a generalized mean function as follows:
\begin{equation}
\text{overlap}(L_{i}, L_{h}, S) = \sum_{k \in S} \left(\frac{f_{ki}^{p} + f_{kh}^{p}}{2}\right)^\frac{1}{p},~~p \le 1
\label{Bilingual overlap defn}
\end{equation}
where $f_{ki}$ denotes the frequency of token $k$ when $D_i$ is encoded with $S$. For different values of $p$, we get different tradeoffs between fairness to each language and overall goodness. When $p=-\infty$, generalized mean reduces to the minimum function, and we get the most egalitarian allocation. However, this ignores the larger of the two frequencies. When $p=1$, we get a simple average which is what the first term in Equation~\eqref{eq:obpe:goal} already covers. For $p=0,-1$, we get the geometric and harmonic means respectively. Due to smaller size of LRL monolingual data, the frequency of a token which is shared across languages is likely to be much higher in HRL monolingual data as compared to that in LRL monolingual data, Hence, setting $p$ to large negative values will increase the weight given to LRLs and thus increase overlap. We will present an exploration of the effect of $p$ on zero-shot transfer in the experiment section.
\begin{table*}[h]
\begin{small}
\centering
\begin{tabular}{|l|l|l|r|r|} \hline
Family & HRL & LRLs & \multicolumn{2}{|c|}{Number of HRL Docs} \\
& & & {\sc balanced} & {\sc skewed} \\ \hline
West Germanic & English (en) & German (de), Dutch (nl), Western\ Frisian (fy) & 0.16M & 1.00M \\ \hline
Romance & French (fr) & Spanish (es), Portuguese (pt), Italian (it) & 0.16M & 0.50M \\ \hline
Indo-Aryan & Hindi (hi) & Marathi (mr), Punjabi (pa), Gujarati (gu) & 0.16M & 0.16M \\ \hline
\end{tabular}
\caption{Twelve Languages \emph{simulated} as HRLs and LRLs across with two different corpus distribution: {\sc balanced}\ and {\sc skewed}. Number of documents in languages simulated as LRLs is 20K.}
\label{tab:langs}
\end{small}
\end{table*}
The greedy version of the above objective that controls the
candidate vocabulary item to be
inducted in each iteration of OBPE is thus:
\begin{equation}
\label{eq:obpe:step}
\begin{split}
\mathcal{V} = \mathcal{V} \cup arg\,max_{k=[u,v]:u,v\in \mathcal{V}}
(1-\alpha)\sum_{j}f_{kj} \\ + \alpha\sum_{i\in \cL_\text{\lrl}}\max_{h \in \cL_\text{\hrl}}\left(\frac{f_{ki}^{p} + f_{kh}^{p}}{2}\right)^\frac{1}{p}
\end{split}
\end{equation}
The data structure maintained by BPE to efficiently conduct such merges can be applied with little changes to the OBPE algorithm. The only difference is that we need to separately maintain the frequency in each language in addition to overall frequency. Since the time and resources used to create the vocabulary is significantly smaller than the model pre-training time, this additional overhead to the pre-training step is negligible.
\section{Experiments}
\label{sec:expts}
We evaluate by measuring the efficacy of zero-shot transfer from the HRL\ on four different tasks: named entity recognition
(NER), part of speech tagging (POS), text classification(TC), and Cross-lingual Natural Language Inference (XNLI).
Through our experiments, we evaluate the following questions:
\begin{enumerate}
\item Is OBPE more effective than BPE for zero-shot transfer? (\refsec{sec:effective-obpe})
\item What is the effect of token overlap on overall accuracy? (\refsec{sec:analysis})
\item How does increased LRL representation in the vocabulary impact accuracy? (\refsec{sec:samp})
\end{enumerate}
We report additional ablation and analysis experiments in \refsec{sec:ablation}.
\noindent
\subsection{Setup}
{\bf Pre-training Data and Languages}
As our pre-training dataset $\{D_i\}$, we use
the Wikipedia dumps of all the languages as used in mBERT. We pre-train with 12 languages grouped into three families of four related languages as shown in Table~\ref{tab:langs}. In each family, we simulate as HRL the most populous language, and call the remaining as LRLs. The number of documents for languages simulated as LRLs is set to 20K. For the HRLs, we consider two corpus distributions:
\begin{itemize}
\item {\sc balanced}\ : all three HRLs get 160K documents each
\item {\sc skewed}\ : English gets one million, French half million, and Hindi 160K documents
\end{itemize}
We evaluate twelve-language models in each of these settings, and present results for separate four language models per family in Table \ref{tab:4_lang} in the Appendix. For the Indo-Aryan languages set, the monolingual data of Punjabi and Gujarati is transliterated to Devanagari, the script of Hindi and Marathi. We use libindic’s indictrans library ~\cite{Bhat:2014:ISS:2824864.2824872} for
transliteration. Languages in the other two sets do not require transliteration as they have a common script. Thus, all four languages in each set are in the same script so their lexical overlap can be leveraged.
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|l|l|r|r|r|r|}
\hline
\multirow{3}{*}{Dataset split} & Lang & \multicolumn{4}{|c|}{Number of sentences} \\ \hline
~ & ~ & NER & POS & TC & XNLI \\ \hline
\multirow{3}{*}{Train:HRL} & hi & 5.0 & 53.0 & 25.0 & ~ \\
& en & 10.5 & 18.0 & 10.0 & 393.0 \\
& fr & 7.5 & 16.5 & 10.0 & 393.0 \\ \hline
\multirow{3}{*}{Validation:HRL} & hi & 1.0 & 3.0 & 4.0 & \\
~ & en & 6.0 & 4.0 & 10.0 & 2.5 \\
~ & fr & 4.0 & 2.0 & 10.0 & 2.5 \\ \hline
\multirow{3}{*}{Test data} & hi & 0.2 & 12.0 & 7.0 & ~ \\
~ & en & 6.0 & 4.6 & 10.0 & 5.0 \\
~ & fr & 4.0 & 4.1 & 10 & 5.0 \\
~ & mr & 0.8 & 9.5 & 6.5 & - \\
~ & pa & 0.2 & 13.4 & 7.9 & - \\
~ & gu & 0.3 & 14.0 & 8.0 & - \\
~ & de & 12.0 & 19.3 & 10.0 & 5.0 \\
~ & nl & 8.0 & 1.0 & - & - \\
~ & fy & 0.8 & - & - & - \\
~ & es & 5.0 & 3.1 & 10.0 & 5.0 \\
~ & pt & 4.0 & 2.5 & - & - \\
~ & it & 5.0 & 3.4 & - & - \\ \hline
\end{tabular}
}
\caption{\label{tab:task:stats}Task-specific data sizes. Number of sentences in thousands.}
\end{table}
\noindent
{\bf Pre-Training Details}
To ensure that LRLs are not under-represented, we over-sample using exponentially smoothed weighting similar to multilingual BERT \cite{devlin-etal-2019-bert} with exponentiation factor
0.7. We perform MLM pretraining on a BERT base model with 110M parameters from scratch. We generate a vocabulary of size of 30k.
We chose batch size as 2048, learning rate as 3e-5 and maximum sequence length as 128. Pre-training of BERT
was done with duplication factor 5 for
for 64k iterations for HRLs.
For all LRLs, duplication factor was 20 and training was done for 24K iterations. MLM pre-training was done on Google
v3-8 Cloud TPUs where 10K iterations required 2.1 TPU hours.
\begin{table*}[!ht]
\centering
\begin{adjustbox}{max width=1.0\textwidth,center}
\begin{tabular}{|l|l| l|l| l|l |l|l |l|}
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c|}{LRL Performance ($\uparrow$)} &\multicolumn{4}{c|}{HRL Performance ($\uparrow$)} \\
& NER & TC & XNLI & POS & NER & TC & XNLI & POS \\
\hline
BPE \cite{sennrich-etal-2016-neural} & 64.48 & 65.52 & 52.07 & 84.64 & 83.26
& \textbf{82.07}
& 62.71
& \textbf{95.20}
\\
BPE-dp \cite{provilkov-etal-2020-bpe} & 63.92 & 64.15 & 52.66 & 84.75 & 81.73
& 81.07
& 63.74
& 94.61
\\
CV \cite{chung-etal-2020-improving} & 59.58 & 61.91 & 49.30 & 81.68 & 81.15 & 80.93& 64.51 & 94.47
\\
TokComp \cite{maronikolakis-etal-2021-wine-v} & 63.79 & 65.77 & 53.94 & \textbf{85.49} & 82.43 & 80.93& 66.10 & 94.86
\\
{\textsc{OBPE}{} (This paper)} & \textbf{65.72}& \textbf{68.02} & \textbf{54.03} & 85.26
& \textbf{83.98} & 81.91 & \textbf{66.27} & 95.09\\
\hline
\end{tabular}
\end{adjustbox}
\caption{Zero-shot performance of models in the Balanced-12 setting trained on 9 LRL and 3 HRL languages. Performance is measured on four tasks: NER (F1), Text Classification (Accuracy), POS (Accuracy), and XNLI (Accuracy). For all metrics, higher is better ($\uparrow$). Zero-shot transfer to LRL improves without hurting HRL\ accuracy. P-value of paired-t-test between BPE and OBPE LRL gains has values $0.01,0.04, 0.02,0.01$ for each of the 4 tasks establishing statistical significance. Detailed results for each language is pesented in Table~\ref{tab:varying_p}. \refsec{sec:effective-obpe} has further discussion.}
\label{tab:overall}
\end{table*}
\begin{table*}[htb]
\centering
\begin{adjustbox}{max width=1.1\textwidth,center}
\begin{tabular}{|l|l| l|l| l|l |l|l |l|}
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c|}{LRL Performance ($\uparrow$)} &\multicolumn{4}{c|}{HRL Performance ($\uparrow$)} \\
& NER & TC & XNLI & POS & NER & TC & XNLI & POS \\
\hline
BPE \cite{sennrich-etal-2016-neural} & 52.91 & 51.68 & 48.57 & 74.79 & 81.78
& 80.04
& 64.96
& 95.03
\\
CV \cite{chung-etal-2020-improving} & 52.73 & 54.40 & 44.28 & 76.70
& 79.84 & 77.74 & 57.18 & 94.60
\\
{\textsc{OBPE}{}\ (This paper)} & {\bf 55.09} & {\bf 55.37} & {\bf 50.01} & {\bf 75.05} & {\bf 82.94} & {\bf 80.31} & {\bf 65.57} & {\bf 95.09}\\
\hline
\end{tabular}
\end{adjustbox}
\caption{Zero-shot performance of models in the Skewed-12 setting of Table~\ref{tab:langs} on same four tasks as Table~\ref{tab:overall}. \textsc{OBPE}{} shows gains here too. Detailed numbers in Table~\ref{tab:12_lang:skew} of Supplementary. \refsec{sec:effective-obpe} has further discussion.}
\label{tab:overall:skew}
\end{table*}
\noindent
{\bf Task-specific Data} We evaluate on four down-stream tasks: (1) NER: data from WikiANN
\cite{pan-etal-2017-cross} and XTREME \cite{pmlr-v119-hu20b}, (2) XNLI: data from \cite{conneau-etal-2018-xnli}, (3) POS: data from XTREME \cite{pmlr-v119-hu20b} and TDIL\footnote{Technology Development for Indian Languages
(TDIL), https://www.tdil-dc.in}, and (4) Text Classification (TC): data from TDIL and XGLUE \cite{liang-etal-2020-xglue}. We downsampled the
TDIL data for each
language to make them class-balanced. The POS
tagset for Indo-Aryan languages used was the BIS Tagset \cite{sardesai-etal-2012-bis}. Table~\ref{tab:task:stats} presents a summary.
The test set to compute LRL perplexity was formed by sampling 10K sentences from Samanantar corpus\cite{ramesh2021samanantar} for Indic languages and from Tatoeba corpus\footnote{Tatoeba
, https://tatoeba.org} for other languages. The perplexity reported for a language is the average of sentence perplexity over all the sentences sampled from that language's corpus.\\
\noindent
{\bf Task-specific fine-tuning details}
We perform task-specific fine-tuning of pre-trained BERT on the task-specific training data of HRL\ and evaluate on all languages in the same family. Here we used
learning-rate 2e-5 and batch size 32, with training duration as 16 epochs for NER, 8 epochs for
POS and 3200 iterations for Text Classification and XNLI. The models were evaluated on a separate validation dataset of the HRL and the model with the minimum validation loss, maximum F1-score, accuracy and minimum validation loss was selected
for final evaluation for XNLI, NER, POS and Text Classification respectively. All fine-tuning experiments
were performed on Google Colaboratory.
The results
reported for all the experiments are an average of 3
independent runs.
\subsection{Effectiveness of \textsc{OBPE}{}}
\label{sec:effective-obpe}
We evaluate the impact of \textsc{OBPE}{} on improving zero-shot transfer from HRLs to LRLs within the same family across four different tasks. We compare with four
existing methods that represent different methods of vocabulary creation and allocation of budget across languages:
\noindent
{\bf Methods compared}
\begin{enumerate}
\item \textbf{BPE} \cite{sennrich-etal-2016-neural}, the existing default method of vocabulary generation.
\item \textbf{Clustered vocabulary (CV)} \cite{chung-etal-2020-improving} Since the paper uses a SentencePiece unigram for vocabulary, we followed the same approach for this comparison. We allocate each family equal number of vocabulary tokens which is \text{V}/3.
\item \textbf{BPE-dropout (BPE-dp)} \cite{provilkov-etal-2020-bpe} uses
the vocabulary generated by BPE but tokenizes the text using a dropout rate of 0.1. This allows the training of tokens that are subsumed by larger tokens in the vocabulary.
\item \textbf{Compatibility of Tokenizations (TokComp)} \cite{maronikolakis-etal-2021-wine-v} uses a method to select meaningful vocabulary sizes in an automated manner for all language using compression rates. Since their best performances are found, when the compression rates are similar, we choose a size for each language corresponding to compression rate of 0.5. The tokenizer used in this method is WordPiece.
.
\item \textbf{\textsc{OBPE}{}} (Ours) with default $\alpha=0.5, p = -\infty$. We also do ablation on these.
\end{enumerate}
In Table~\ref{tab:overall} we observe that across all four tasks, zero-shot LRL accuracy improves compared to BPE. For example, the average accuracy on XNLI for the LRL languages improves from 55.6 to 58.1 just by changing the set of tokens in the vocabulary. These gains are obtained without compromising HRL performance on the tasks. The Clustered Vocabulary (CV) approach is much worse than BPE. These experiments are on the Balanced-12 model. In the supplementary section, we report the results on the Skewed-12 (Table~\ref{tab:overall:skew}) and Balanced-4 models (Table~\ref{tab:4_lang}) and show similar gains even with these models.
In this table, we averaged the gains over nine LRLs, and in the Supplementary Table~\ref{tab:varying_p} we show consistent gains for individual languages.
In addition to improving zero-shot transfer from HRLs to LRLs on downstream tasks, OBPE\ also leads to better intrinsic representation of LRLs. We validate that by measuring the pseudo-perplexity~\cite{salazar-etal-2020-masked} of a test set of LRL sentences. We find that average perplexity of LRL sentences drops by 2.6\%
when we go from the BPE to OBPE vocabulary. More details on this experiment appear in Figure~\ref{fig:ppl}.
\begin{figure}[t]
\centering
\includegraphics[scale =0.32]{PPL.pdf}
\caption{Percentage reduction in Pseudo perplexity~\cite{salazar-etal-2020-masked} for different LRLs as we go from BPE to OBPE\ vocabulary. (Section \ref{sec:effective-obpe})}
\label{fig:ppl}
\end{figure}
\input{samplegraph}
In order to investigate the reasons behind the OBPE gains, we first inspected the percentage of tokens in the vocabulary that belong to LRLs, HRLs, and in their overlap. We find that with OBPE both LRL tokens and overlapping tokens increase. Either of these could have led to the observed gains. We analyze the effect of each of these factors in the following two sections.
\subsection{Effect of Token Overlap}
\label{sec:analysis}
\begin{table}[!ht]
\begin{small}
\centering
\begin{tabular}{|l|c|c|}
\hline
~ &\multicolumn{2}{c|}{en-es} \\ \hline
~ & \multicolumn{1}{c|}{High (es: 1 GB)} & \multicolumn{1}{c|}{Low: (es: 20K)} \\ \hline
NER & -1.4 & -11.7 \\ \hline
XNLI & 0.7 & -1.3 \\ \hline
~ &\multicolumn{2}{c|}{hi-mr} \\ \hline
~ & \multicolumn{1}{c|}{High (mr: 110K)} & \multicolumn{1}{c|}{Low (mr: 20K)} \\ \hline
NER & -12.2 & -41.6 \\ \hline
TC & -2.7 & -41.3 \\ \hline
POS & -6.6 & -7.8 \\ \hline
\end{tabular}
\caption{Drop in Accuracy of Zero-shot transfer when we synthetically reduce token overlap to zero. Transfer is from English (en) as HRL to Spanish (es) and from Hindi (hi) as HRL to Marathi (mr) in two settings: (1) High where es, mr have sizes comparable to the HRL and (2) Low where their sizes are only 20K. Token overlap is important in the low-resource and related language setting (Section~\ref{sec:analysis})}
\label{tab:overlapHL}
\end{small}
\end{table}
We present the impact of token overlap via two sets of experiments: first, a controlled setup where we synthetically vary the fraction of overlap and second where we measure correlation between overlap and gains of OBPE on the data as-is.
For the controlled setup we follow \cite{K2020Cross-Lingual} for synthetically controlling the amount of overlap between HRL and LRL.
We trained a bilingual model between Hindi (HRL 160K) and Marathi (LRL 20K) --- two closely related languages in the Indo-Aryan family.
To find the set of overlapping tokens between Hindi and Marathi, we first run \textsc{OBPE}{} on Hindi-Marathi language pair to generate a vocabulary and label all tokens present in both languages as
\emph{overlapping tokens}. We then incrementally sample 10\%, 40\%, 50\%, 90\% of the tokens from this set. We shift the Unicode of the entire Hindi monolingual data except the set of sampled tokens so that there are no overlapping tokens between Hindi (hi) and Marathi (mr) monolingual data other than the sampled tokens. Let us call this Hindi data \textbf{SynthHindi}. We then run \textsc{OBPE}{} on SynthHindi-Marathi language pair to generate a vocabulary
to pretrain the model. The task-specific Hindi data is also converted to SynthHindi during fine-tuning and testing of the model.
\reffig{fig:unicode} shows results with increasing overlap. We observe increasing gains in LRL accuracy as we go from no overlap to full overlap on all three tasks. NER accuracy increases from 17\% to 58\% for the LRL (mr) even while the HRL (hi) accuracy stays unchanged. For TC we observe similar gains. For POS, even without token overlap, we get good cross-lingual transfer because POS tags are more driven by structural similarity, and Hindi and Marathi follow similar structure.
\begin{table*}[h]
\centering
\parbox{0.55\textwidth}{
\begin{small}
\setlength\tabcolsep{3.0pt}
\begin{tabular}{|l|l| l|l| l|l |l|l |l|}
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c|}{LRL Performance ($\uparrow$)} &\multicolumn{4}{c|}{HRL Performance ($\uparrow$)} \\
& NER & TC & XNLI & POS & NER & TC & XNLI & POS \\
\hline
BPE & 64.5 & 65.5 & 52.1 & 84.6 & 83.3
& \textbf{82.1}
& 62.7
& \textbf{95.2}
\\
+overSample & 64.4 & 67.6 & 52.1 & 84.6 & 82.4 & 82.0 & 62.0 & 95.2
\\ \hline
{\textsc{OBPE}{} } & \textbf{65.7}& \textbf{68.0} & \textbf{54.0} & \textbf{85.3}
& \textbf{84.0} & 81.9 & \textbf{66.3} & 95.1\\
+overSample & 64.6 & 67.9 & 53.5 & 85.1 & 82.7 & 81.7 & 65.7 & 94.8
\\
\hline
\end{tabular}
\caption{\label{tab:sample}Zero-shot performance of models in the same setting as Table~\ref{tab:overall} but comparing default sampling with oversampling (exponentiation factor S=0.5). Note, even if BPE\_overSamp improves LRL somewhat, it causes HRL to drop. OBPE with default sampling is best for both LRLs and HRLs. Also OBPE\_overSampled is better than BPE\_overSampled (Section~\ref{sec:samp}). }
\end{small}
}
\vspace{-0.6 cm}
\qquad
\begin{minipage}[c]{0.38\textwidth}%
\centering
\includegraphics[scale =0.3]{chart.pdf}
\caption{Percentage rise over BPE in representation of LRL, HRL and Shared (percentage of tokens shared between HRL and LRL weighted by frequency) in vocabulary generated by OBPE and BPE\_overSample and OBPE\_overSample (Section~\ref{sec:samp}). }
\label{img:vocabstats}
\end{minipage}
\end{table*}
Our results contradict the conclusions of \cite{K2020Cross-Lingual} which claimed that token overlap is unimportant for cross-lingual transfer. However, there are two key differences with our setting: (1)
unlike \cite{K2020Cross-Lingual}, we explore low-resource settings, and (2) except for English-Spanish, the other language pairs they considered are not linguistically related. To explain the importance of both these factors, in \reftbl{tab:overlapHL} we present accuracy of English-Spanish in a simulated low-resource setting where we sample 20K Spanish documents and 160K English documents.
Also, we repeat our Hindi-Marathi experiments where Marathi is not low-resource. We observe that
(1) Spanish as LRL benefits significantly on overlap with English.
(2) Marathi gains from token overlap with Hindi even in the high resource setting.
Thus, we conclude that as long as languages are related, token overlap is important and the benefit from overlap is higher in the low resource setting.
\paragraph{Overlap Vs Gain: Real data setup}
\label{sec:real}
\begin{table}[t]
\begin{small}
\centering
\begin{tabular}{|l|l|c|}
\hline
Lang family & Task & Pearson Correlation\\ \hline
\multirow{2}{*}{Indo-Aryan} & NER & 0.835 \\
& POS & 0.690 \\ \hline
\multirow{2}{*}{West Germanic} & NER & 0.387 \\
& POS & 0.348 \\ \hline
\multirow{2}{*}{Romance} & NER & 0.946 \\
& POS & 0.595 \\ \hline
\end{tabular}
\caption{\label{tab:corr}Correlation coefficient between performance gain and overlap gain within languages in a family for various tasks.
(Section~\ref{sec:real}).}
\end{small}
\end{table}
\vspace{-0.2cm}
We further substantiate our hypothesis that the shared tokens across languages favoured by \textsc{OBPE}{} enable transfer of supervision from HRL to LRL via statistics on real-data. In Table~\ref{tab:corr} we show the Pearson product-moment correlation coefficient between overlap gain and performance gain within LRLs of the same family and task. We get a high positive correlation coefficient, with an average of 0.644.
\subsection{Effect of Increased LRL representation}
\label{sec:samp}
We next investigate the impact of increased representation of LRL tokens in the vocabulary. OBPE increases LRL representation by favoring overlapping tokens, but LRL tokens can also be increased by just over-sampling LRL documents.
We train another {\sc balanced} 12 model but with further over-sampling LRLs with exponentiation factor of 0.5 instead of 0.7. We observe in Figure~\ref{img:vocabstats} that this increases LRL fraction but reduces HRL tokens in the vocabulary. Table~\ref{tab:sample} also shows the comparison of zero-shot transfer accuracy with over-sampled BPE against over-sampled OBPE. We find that OBPE even with default exponentiation factor achieves highest LRL gains, whereas aggressively over-sampled BPE hurts HRL accuracy. Within the same sampling setting, OBPE is better than corresponding BPE.
\subsection{Ablation study}
\label{sec:ablation}
\vspace{-0.2cm}
We conducted experiments for different values of $p$ that controls the amount of overlap in the generalized mean function (\refeqn{eq:obpe:step}). Figure \ref{fig:p} and Table \ref{tab:varying_p} show the results for various $p$.
Setting $p=1$ gives the original BPE algorithm. Setting $p=0,-1$ gives geometric and harmonic mean respectively, setting $p=-\infty$ gives minimum. We compare the task-specific results for different values of $p$ as shown in Table~\ref{tab:varying_p} and find that the gains we obtain are highest in the $p=-\infty$ (minimum) setting (Figure~\ref{fig:p}).
\input{p_graph}
We also experiment with $\alpha = 0.7$, and find that for most languages the results were not better than our default $\alpha = 0.5$.
\section{Conclusion}
In this paper, we address the problem of cross-lingual transfer from HRL{}s to LRL{}s by exploiting relatedness among them. We focus on lexical overlap during the vocabulary generation stage of multilingual pre-training. We propose Overlap BPE (\textsc{OBPE}{}), a simple yet effective modification to the BPE algorithm, which chooses a vocabulary that maximizes overlap across languages. \textsc{OBPE}{} encodes input corpora compactly while also balancing the
trade-off between cross-lingual subword sharing and language-specific vocabularies. We focus on three sets of closely related languages from diverse language families. Our experiments provide evidence that \textsc{OBPE}{} is effective in leveraging overlap across related languages to improve LRL{} performance. In contrast to prior work, through controlled experiments on the amount of token overlap between two related HRL{}-LRL{} language pairs, we establish that token overlap is important when a LRL is paired with a related HRL.
\paragraph{Acknowledgements}
We thank Yash Khemchandani and Sarvesh Mehtani for participating in the early phases of this research. We thank Dan Garrette and Srini Narayanan for comments on the draft.
We thank Technology Development for Indian Languages (TDIL) Programme initiated by the Ministry of Electronics
Information Technology, Govt. of India for providing us datasets used in this study. The experiments
reported in the paper were made possible by a Tensor Flow Research Cloud (TFRC) TPU grant. The
IIT Bombay authors thank Google Research India
for supporting this research.
| 09f5cdaf1e66fc0ab908efb675affe2ae094d522 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Euclidean quantum gravity has proven quite powerful in understanding non-perturbative and thermodynamic aspects of gravitational systems (e.g.\ see \cite{Gibbons:1994cg}). For non-gravitational theories, Wick rotation transforms an oscillatory integral in Lorentzian signature into an exponentially suppressed integral in Euclidean signature. However, the same effect on convergence of path integrals is not guaranteed for gravitational theories as the Euclidean action for gravity is not positive definite. Despite various potential difficulties, including the conformal factor problem~\cite{Gibbons:1978ac}, it remains a quite useful tool in understanding black hole thermodynamics, quantum cosmology and AdS/CFT.
The question of which (if any) complex metrics should contribute to the gravitational path integral remains a question of great importance. The recent ``allowability'' criterion of~\cite{Kontsevich2021} (see also~\cite{Witten2021,Lehners:2021mah}) provides a simple diagnostic identifying to which complex metrics all possible $p$-form gauge theories may be consistently coupled. It remains to be seen if this is a necessary or sufficient criterion for consistently coupling any matter theory to gravity. There is also evidence that complex saddle points are \emph{required} to make sense of pure 3D gravity~\cite{Maloney:2007ud}. One approach is to avoid performing the Wick rotation to Euclidean signature at all and aim to understand the gravitational path integral from a Lorentzian perspective~\cite{Marolf:1996gb, Dasgupta:2001ue, Feldbrugge:2017fcc, DiazDorronsoro:2017hti,Brown:2017wpl, Marolf:2020rpm, Marolf:2021ghr}. We will adopt this approach as well and consider Lorentzian path integrals, using Picard-Lefschetz theory in the particularly visualizable case of a single complex variable to define oscillatory integrals. This provides a consistent framework in which Lorentzian, complex and Euclidean saddle points may be treated democratically, and has previously been fruitfully applied in the study of quantum cosmology and beyond~\cite{Cherman:2014sba,Tanizaki:2014xba, Behtash:2015loa,Feldbrugge:2017kzv,Feldbrugge:2017mbc,Feldbrugge:2017fcc,DiazDorronsoro:2017hti,Brown:2017wpl,Feldbrugge:2019fjs,Rajeev:2021yyl}.
In axion gravity the Giddings-Strominger (GS) wormhole~\cite{Giddings:1989bq} appears as a highly symmetric saddle point. Determining whether the GS wormhole represents a true minimum of the Euclidean action is critical in understanding its role and interpretation in the gravitational path integral. The nature of the GS critical point can be analyzed by considering fluctuations around the background geometry and looking for tachyonic directions in which the Euclidean action is decreased. There has been previous work in analyzing this saddle point in the axion picture~\cite{Hertog:2018kbz} (see also~\cite{Rubakov:1996cn, Kim:2003js, Alonso:2017avz, Hertog:22toappear}); here we approach the question of perturbative stability in the 3-form picture, which we argue provides a more transparent application of the correct boundary conditions. We find that no negative modes are present and the GS wormhole represents a true minimum of the Euclidean path integral.
Our findings have some bearing on the existence of Euclidean wormholes in string theory. Previous analysis of perturbative stability \cite{Hertog:2018kbz,Rubakov:1996cn,Kim:2003js} cast doubts on them being genuine saddle points of the path integral. While such analyses have so far only been carried out for the GS wormhole, it was argued in \cite{Hertog:2018kbz} that any perturbative instabilty may explain away the puzzle with embedding Euclidean wormholes into AdS/CFT \cite{Maldacena:2004rf, Arkani-Hamed:2007cpn}. The correlation functions across the two boundary CFTs should factorize while the Euclidean wormhole saddle contributions (if they exist) would seem to indicate that they do not. Extending our work to include one or more dilaton would be necessary to establish the existence of Euclidean wormhole saddles in string theory (where the axions are accompanied by their dilaton partners). Our results, if they can be successfully extrapolated, would seem to suggest that the difficulty may not be in finding Euclidean wormhole saddles but in embedding them into AdS/CFT~\cite{Arkani-Hamed:2007cpn, Hertog:2017owm, Astesiano:2022qba, Marolf:2021kjc}. We note also in passing that establishing the existence of axionic Euclidean wormhole saddles is of importance to the Weak Gravity conjecture for axions~\cite{Arkani-Hamed:2006emk, Rudelius:2015xta, Brown:2015iha, Brown:2015lia, Montero:2015ofa, Heidenreich:2015nta, Hebecker:2016dsw, Hebecker:2018ofv, Andriolo:2020lul}.
The remainder of this paper is organized as follows. In Sec.~\ref{sec:duality} we review the duality between a 2-form gauge field/3-form flux and axion in 4D, emphasizing how the duality prescribes corresponding boundary conditions for the fields. In Sec.~\ref{sec:pathintegrals} we compute Lorentzian transition amplitudes on either side of the duality in saddle-point approximation and demonstrate how the analysis can be qualitatively different in the presence of a bilocal operator. Picard-Lefschetz theory is used to make sense of any oscillatory integrals encountered. In Sec.~\ref{sec:stability} we address the perturbative stability of the Giddings-Strominger wormhole in the 3-form picture. Finally, we conclude in Sec.~\ref{sec:disc}.
\section{Review of the duality}
\label{sec:duality}
The two systems we consider are axion coupled to Einstein gravity and a 2-form gauge field coupled to Einstein gravity, which are dual to one another in 4D. In this section we review this duality; the familiar reader can jump straight to Sec.~\ref{sec:pathintegrals}. See~\cite{Collinucci:2005opa} for a nice discussion of these ideas in Euclidean signature.
The duality is most clearly seen by going to a first-order formalism where the 3-form flux is the dynamical field. Write
\begin{equation}\label{eq:gHth_action}
\begin{aligned}
S[g,H,\theta] &= S_\text{grav}[g] + S_\text{m}[g,H,\theta] \,,\\
S_\text{grav}[g] &= \frac{1}{2}\int_\mathcal{M}\d[4]{x}\,\sqrt{-g}\,R + \epsilon\int_{\partial\mathcal{M}}\d[3]{x}\,\sqrt{|h|}\,K \,,\\
S_\text{m}[g,H,\theta] &= \int_\mathcal{M}\left(-\frac{1}{2}H\wedge{\star H} + \theta\,\d{H}\right) \,,
\end{aligned}
\end{equation}
where $\epsilon=\pm1$ for time-like/space-like boundaries. The variational problem for the above action is well-posed if one chooses Dirichlet boundary conditions for both the metric (due to the Gibbons-Hawking-York term) and 3-form (restricting to a particular choice of flux). There are no boundary conditions on the Lagrange multiplier $\theta$ which has been introduced to impose $\d{H}=0$ as a constraint. Path integrals on a fixed manifold $\mathcal{M}$ take the form
\begin{equation}
Z_\mathcal{M}[h,J] = \int_{g|_\partial=h,\;H|_\partial = J}\hspace{-50pt}\mathcal{D}g\mathcal{D}H\mathcal{D}\theta\;e^{iS[g,H,\theta]} \,,
\end{equation}
where $h$, $J$ indicate the chosen boundary conditions for $g$, $H$ respectively. Of course, with dynamical gravity one can imagine specifying only the boundary manifold and performing a sum over bulk manifolds with the specified boundary:
\begin{equation}
Z_\mathcal{B}[h,J] = \sum_{\mathcal{M}\;:\;\partial\mathcal{M}=\mathcal{B}}Z_\mathcal{M}[h,J] \,.
\end{equation}
Such a topological expansion has been well-explored in the context of 2D dilaton gravity, e.g.\ in Jackiw-Teitelboim gravity~\cite{Saad:2019lba}.
On one hand, the Lagrange multiplier may be integrated out exactly, resulting in the constraint that $H$ is closed (i.e.\ $H=\d{B}$ locally) and the action for a 2-form gauge field. On the other, the 3-form flux may be integrated out by completing the square in $H$:
\begin{equation}
S_\text{m}[g,H,\theta] = \int_\mathcal{M}\left(-\frac{1}{2}(H-{\star\d{\theta}})\wedge{\star(H-{\star\d{\theta}})} - \frac{1}{2}\d{\theta}\wedge{\star\d{\theta}}\right) + \int_{\partial\mathcal{M}}\theta H \,.
\end{equation}
We have used that $\star^2=(-1)^{p+1}$ for 4D Lorentzian manifolds when acting on $p$-forms. The integral over the bulk is Gaussian and $H|_\partial = J$ is fixed as a boundary condition. This allows us to write
\begin{equation}
\begin{aligned}
Z_\mathcal{M}[h,J] &= \int_{g|_\partial = h}\hspace{-20pt}\mathcal{D}g\mathcal{D}\theta\;e^{iS[g,\theta]} \,,\\
S[g,\theta] &= S_\text{grav}[g] + \int_\mathcal{M}\left(-\frac{1}{2}\d{\theta}\wedge{\star\d{\theta}}\right) + \int_{\partial\mathcal{M}}\theta J \,.
\end{aligned}
\end{equation}
If we split the axion into bulk and boundary degrees of freedom then the above has the structure of a Fourier transform:
\begin{equation}
\begin{aligned}
Z_\mathcal{M}[h,J] &= \int\mathcal{D}\theta_\text{bdy}\;e^{i\int_{\partial\mathcal{M}}\theta_\text{bdy} J}\int_{g|_\partial = h,\;\theta|_\partial = \theta_\text{bdy}}\hspace{-50pt}\mathcal{D}g\mathcal{D}\theta \;e^{iS_\text{grav}[g] + i\int_\mathcal{M}\big(-\frac{1}{2}\d{\theta}\wedge{\star\d{\theta}}\big)} \,,\\
&\equiv \int\mathcal{D}\theta_\text{bdy}\;e^{i\int_{\partial\mathcal{M}}\theta_\text{bdy} J} Z_\mathcal{M}[h,\theta_\text{bdy}] \,.
\end{aligned}
\end{equation}
Alternatively, one can view this as imposing Neumann boundary conditions on the axion.
To summarize, the path integral $Z_\mathcal{M}[h,J]$ may be computed in one of two ways. In the 3-form picture one simply computes
\begin{equation}
Z_\mathcal{M}[h,J] = \int_{g|_\partial=h,\;H|_\partial = J}\hspace{-50pt}\mathcal{D}g\mathcal{D}H\;\delta[\d{H}]\;e^{iS_\text{grav}[g] + i\int_\mathcal{M}\big(-\frac{1}{2}H\wedge{\star H}\big)} \,,
\end{equation}
while in the axion picture one can structure the computation as
\begin{equation}
\begin{aligned}
Z_\mathcal{M}[h,J] &= \int\mathcal{D}\theta_\text{bdy}\;e^{i\int_{\partial\mathcal{M}}\theta_\text{bdy} J} Z_\mathcal{M}[h,\theta_\text{bdy}] \,,\\
Z_\mathcal{M}[h,\theta_\text{bdy}] &= \int_{g|_\partial = h,\;\theta|_\partial = \theta_\text{bdy}}\hspace{-50pt}\mathcal{D}g\mathcal{D}\theta \;e^{iS_\text{grav}[g] + i\int_\mathcal{M}\big(-\frac{1}{2}\d{\theta}\wedge{\star\d{\theta}}\big)} \,.
\end{aligned}
\end{equation}
In the next section we will compute one such $Z_\mathcal{M}[h,J]$ in some detail on either side of this duality.
\section{Lorentzian path integrals}
\label{sec:pathintegrals}
\subsection{Transition amplitudes}
\label{sec:amplitudes}
In this section we will examine transition amplitudes between two boundaries of $S^3$ topology, restricting attention to $\mathcal{M}=\mathbb{R}\times S^3$ as the leading contribution in a topological expansion. We can take the coordinate time to be $t\in[0,1]$ so that the boundaries lie at $t=0$ and $t=1$. For boundary conditions we take
\begin{equation}
\begin{aligned}
\d{s^2}\big|_{t=0} &= q_0\,\d{\Omega_3^2} \,, & \qquad J\big|_{t=0} &= \frac{n_0\,\mathrm{vol}_3}{2\pi^2} \,,\\
\d{s^2}\big|_{t=1} &= q_1\,\d{\Omega_3^2} \,, & J\big|_{t=1} &= \frac{n_1\,\mathrm{vol}_3}{2\pi^2} \,,
\end{aligned}
\end{equation}
where $\d{\Omega_3^2}$ is the round metric on $S^3$ and $\mathrm{vol}_3$ is the corresponding volume form normalized to $\int_{S^3}\mathrm{vol}_3 = 2\pi^2$. Let's begin with the 3-form computation; without loss of generality, we will consider only $q_1\geq q_0\geq 0$ and work with the following homogeneous, isotropic ansatz
\begin{equation}\label{eq:metricHansatz}
\begin{aligned}
\d{s^2} &= -\frac{N^2}{q(t)}\,\d{t^2} + q(t)\,\d{\Omega_3^2} \,,\\
H &= \mathfrak{h}(t)\,\mathrm{vol}_3 \,,
\end{aligned}
\end{equation}
with $N$ constant. The transition amplitude of interest is then
\begin{equation}
\begin{aligned}
Z[q_0,q_1;n_0,n_1] &\simeq \int_0^\infty\d{N}\int_{q(0)=q_0}^{q(1)=q_1}\mathcal{D}q\int_{\mathfrak{h}(0)=n_0}^{\mathfrak{h}(1)=n_1}\mathcal{D}\mathfrak{h}\;\delta[\dot{\mathfrak{h}}(t)]\;e^{iS[N,q,\mathfrak{h}]} \,,\\
S[N,q,\mathfrak{h}] &= 2\pi^2\int_0^1\d{t}\,\left( 3N - \frac{3\dot{q}^2}{4N} - \frac{\mathfrak{h}^2N}{2q^2} \right) \,,
\end{aligned}
\end{equation}
where $\dot{A}\equiv\de{A}{t}$. The path integral over $\mathfrak{h}$ may immediately be done since the constraint $\d{H}=0$ fixes $\dot{\mathfrak{h}}=0$. The transition amplitude is nonzero only if $n_0=n_1\equiv n$, in which case one simply has $\mathfrak{h}(t)=\frac{n}{2\pi^2}$ with $n$ integer-quantized:
\begin{equation}
n = \int_{S^3}H \in\mathbb{Z} \,.
\end{equation}
To reduce notational clutter it is convenient to introduce the rescaled flux $\tilde{n}=\frac{n}{2\pi^2\sqrt{6}}$ and restrict attention to $\tilde{n}>0$. Stripping off the flux-conserving $\delta$-function, we have
\begin{equation}\label{eq:Hpath}
Z[q_0,q_1;n] \simeq \int_0^\infty\d{N}\int_{q(0)=q_0}^{q(1)=q_1}\mathcal{D}q\;e^{iS[N,q,n]} \,.
\end{equation}
Our approach will be to evaluate the $q$ path integral for any boundary conditions $q_0,q_1$ and then use Picard-Lefschetz theory to evaluate the remaining one-dimensional integral over $N$. The equation of motion for $q$ is
\begin{equation}
\ddot{q} + \frac{4\tilde{n}^2N^2}{q^3} = 0 \,,
\end{equation}
for which there are two solutions satisfying the boundary conditions,
\begin{equation}
q_\pm(t) = \sqrt{q_0^2(1-t)^2 + q_1^2t^2 + 2q_0q_1\Gamma_\pm\,t(1-t)} \,, \qquad \Gamma_\pm = \pm\sqrt{1 + \frac{4\tilde{n}^2N^2}{q_0^2q_1^2}} \,,
\end{equation}
where we take the principle branch of the square-root so that $q_\pm(0)=\sqrt{q_0^2}=q_0$ and $q_\pm(1)=\sqrt{q_1^2}=q_1$. It is straightforward to show that $q_+^2$ is nowhere zero on $t\in(0,1)$ whereas $q_-^2$ always passes through zero. Consequently, we take only $q(t)=q_+(t)$ as a viable classical saddle point. One finds
\begin{equation}
S[N,q_+,n] = 2\pi^2\left[3N - \frac{3(q_0^2+q_1^2)}{4N}+\frac{3q_0q_1}{2N}\,f\Big(\frac{2\tilde{n}N}{q_0q_1}\Big)\right] \,,
\end{equation}
where
\begin{equation}
f(z) = \sqrt{z-i}\sqrt{z+i} - z\log{\big(z+\sqrt{z-i}\sqrt{z+i}\big)} \,.
\end{equation}
There are branch points at $z=\pm i$ and we have chosen the branch cuts to extend parallel to the negative real axis for sake of presentation (see Fig.~\ref{fig:PL_H}). We arrive at
\begin{equation}
Z[q_0,q_1;n] \simeq \int_0^\infty\d{N}\,e^{2\pi^2i\left[3N - \frac{3(q_0^2+q_1^2)}{4N}+\frac{3q_0q_1}{2N}f\big(\frac{2\tilde{n}N}{q_0q_1}\big) \right]} \,,
\end{equation}
and we evaluate this oscillatory function using Picard-Lefschetz theory. There are four saddle points of $S[N,q_+,n]$, which occur where
\begin{equation}
16N^4 - 8N^2(2\tilde{n}^2-q_0^2-q_1^2) + (q_0^2-q_1^2)^2 = 0 \,,
\end{equation}
namely at
\begin{equation}\label{eq:saddles}
N_{\pm\pm} = \frac{1}{2}\left(\pm\sqrt{\tilde{n}^2 - q_0^2} \pm \sqrt{\tilde{n}^2 - q_1^2}\right)
\end{equation}
with the two signs being chosen independently. The initial contour $N\in(0,\infty)$ is deformed to intersect one or more of these saddle points in such a way that $\Im(iS)$ is (piecewise) constant and $\Re(iS)$ decreases as one moves on the contour away from a saddle -- this ensures that the resulting integrals are absolutely convergent. The required contour deformation depends on the relative values of $q_0,q_1,\tilde{n}$ and falls into the following three cases:
\begin{figure}[t]
\centering
\vspace{35pt}
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case1.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.6,0.9,1)$}
\label{fig:PL_H_1}
\end{subfigure}\hfill
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case2.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.6, 1.1, 1)$}
\label{fig:PL_H_2}
\end{subfigure}\\
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case3.png}
\subcaption{$(q_0,q_1,\tilde{n})=(1.03, 1.1, 1)$ with $\hbar=1-0.3i$}
\label{fig:PL_H_3}
\end{subfigure}\hfill
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case4.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.4, 0.8, 0.25)$ with $\hbar=1-0.5i$}
\label{fig:PL_H_4}
\end{subfigure}
\caption{Structure of $iS[N,q_+,n]$ in the complex-$N$ plane for (a) $q_0,q_1<\tilde{n}$, (b) $q_0<\tilde{n}<q_1$, (c) $q_0,q_1>\tilde{n}$ with $\frac{1}{q_0}+\frac{1}{q_1}>\frac{1}{\tilde{n}}$ and (d) $q_0,q_1>\tilde{n}$ with $\frac{1}{q_0}+\frac{1}{q_1}<\frac{1}{\tilde{n}}$. In (c) and (d) we have set $\Im(\hbar)<0$ to resolve ambiguities in the Picard-Lefschetz prescription as described in the main text. Regions of red have $\Re(iS)<0$ and regions of blue have $\Re(iS)>0$: several level-sets of $\Re(iS)$ are shown with dotted lines. Steepest ascent/descent contours (level-sets of $\Im(iS)$ emanating from the saddle points) are shown with solid black lines. The initial and deformed $N$-contours are shown with dashed and solid green lines, respectively.\vspace{35pt}}
\label{fig:PL_H}
\end{figure}
\begin{enumerate}[label=\roman*)]
\item $q_0< q_1<\tilde{n}$: all saddles are on the real axis and the deformed contour intersects the two positive saddle points with angles $\pm45^\circ$ (see Fig.~\ref{fig:PL_H_1}).
\item $q_0<\tilde{n}<q_1$: all saddles are complex and the deformed contour intersects the one saddle point in the positive quadrant (see Fig.~\ref{fig:PL_H_2}).
\item $\tilde{n}<q_0<q_1$: all saddles are on the imaginary axis and the steepest ascent/descent contours intersect the branch points of $S[N,q_+,n]$. To resolve this ambiguity we shift $S\to \frac{S}{\hbar}=\frac{S}{1-i\epsilon}$ and then take $\epsilon\to0^+$. The deformed contour intersects one saddle point. For $\frac{1}{q_0}+\frac{1}{q_1}= \frac{1}{\tilde{n}}$ two of the four saddles collide with the branch points and for $\frac{1}{q_0}+\frac{1}{q_1} < \frac{1}{\tilde{n}}$ they move to another sheet; however, these saddle points are never included under the $\epsilon$-deformation and so this curiosity will not concern us (see Fig.~\ref{fig:PL_H_3} and Fig.~\ref{fig:PL_H_4}).
\end{enumerate}
Having identified which saddles contribute to the transition amplitude, we may then evaluate the transition amplitude in each of the three cases above. These are compactly written in terms of
\begin{equation}
\begin{aligned}
g_1(z) &= \sqrt{1-z^2} - \operatorname{arcsech}{z} & z\leq1\\
g_2(z) &= ig_1(z+i\epsilon) = \sqrt{z^2-1} - \arcsec{z} & \quad z\geq1
\end{aligned}
\end{equation}
as
\begin{equation}\label{eq:Z3form}
Z[q_0,q_1;n] \simeq \begin{cases}
e^{-i\pi/4}e^{6\pi^2i\tilde{n}[g_1(\frac{q_0}{\tilde{n}}) - g_1(\frac{q_1}{\tilde{n}})]} + e^{i\pi/4}e^{6\pi^2i\tilde{n}[g_1(\frac{q_0}{\tilde{n}}) + g_1(\frac{q_1}{\tilde{n}})]} & q_0 < q_1<\tilde{n}\\
e^{i\alpha}e^{6\pi^2i\tilde{n}g_1(\frac{q_0}{\tilde{n}})}e^{-6\pi^2\tilde{n}g_2(\frac{q_1}{\tilde{n}})} & q_0<\tilde{n}<q_1\\
e^{i\pi/2}e^{-6\pi^2\tilde{n}[g_2(\frac{q_1}{\tilde{n}})-g_2(\frac{q_0}{\tilde{n}})]} & \tilde{n}<q_0<q_1
\end{cases}
\end{equation}
where $\alpha$ gives the angle of the steepest-descent contours through the sole contributing saddle point -- an unilluminating function of $\frac{q_0}{\tilde{n}}$ and $\frac{q_1}{\tilde{n}}$. For $q_0,q_1<\tilde{n}$ there are two Lorentzian saddles which contribute as phases and lead to interference effects. For $q_0<\tilde{n}<q_1$ we can interpret the $q_0$-dependent phase as describing Lorentzian evolution from $q=q_0$ to $q=\tilde{n}$ and then the $q_1$-dependent exponential suppression as due to the tunnelling into the classically forbidden region $q>\tilde{n}$. For $\tilde{n}<q_0<q_1$ the saddle point is Euclidean and the bulk geometry is a segment of the full GS wormhole: the exponential suppression can be identified with its action.
\bigskip
\begin{figure}[t]
\centering
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.95\linewidth]{PL_axion_case1.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.6,0.95,1)$}
\label{fig:PL_axion_1}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_axion_case2.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.7,1.5,1)$}
\label{fig:PL_axion_2}
\end{subfigure}\hfill
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_axion_case3.png}
\subcaption{$(q_0,q_1,\tilde{n})=(1.1,2.5,1)$ with $\hbar=1-0.1i$}
\label{fig:PL_axion_3}
\end{subfigure}
\caption{Structure of $\Phi$ of Eqn.~\eqref{eq:FT} in the complex-$\Delta\vartheta$ plane for (a) $q_0,q_1<\tilde{n}$, (b) $q_0<\tilde{n}<q_1$ and (c) $\tilde{n}<q_0<q_1$. In (c) we have set $\Im(\hbar)<0$ to resolve ambiguities in the Picard-Lefschetz prescription. Regions of red have $\Re(i\Phi)<0$ and regions of blue have $\Re(i\Phi)>0$: several level-sets of $\Re(i\Phi)$ are shown with dotted lines. Steepest ascent/descent contours are shown with solid black lines. The initial and deformed $\Delta\vartheta$-contours are shown with dashed and solid green lines, respectively.}
\label{fig:PL_axion}
\end{figure}
Turning now to the axion side of the duality, we see that the restriction to $n_0=n_1$ arises due to the shift symmetry of the theory:
\begin{equation}
\begin{aligned}
Z[q_0,q_1;n_0,n_1] &\simeq \int\d{\theta_0}\d{\theta_1} e^{i(\theta_1n_1-\theta_0n_0)}Z[q_0,q_1;\theta_0,\theta_1]\\
&= \int\d{\overline{\theta}}\d{\Delta\theta}\,e^{i[\overline{\theta}(n_1-n_0) + \Delta\theta(\frac{n_0+n_1}{2})]}Z[q_0,q_1;0,\Delta\theta]\\
&= 2\pi\,\delta(n_1-n_0)\int\d{\Delta\theta}\,e^{i\Delta\theta(n_0+n_1)/2}Z[q_0,q_1;0,\Delta\theta] \,.
\end{aligned}
\end{equation}
Calculating $Z[q_0,q_1;0,\Delta\theta]$ is relatively easy. We have
\begin{equation}
\begin{aligned}
Z[q_0,q_1;0,\Delta\theta] &\simeq \int_0^\infty\d{N}\int_{q(0)=q_0}^{q(1)=q_1}\mathcal{D}q\int_{\theta(0)=0}^{\theta(1)=\Delta\theta}\mathcal{D}\theta\,e^{iS[N,q,\theta]} \,,\\
S[N,q,\theta] &= 2\pi^2\int_0^1\d{t}\,\left(3N - \frac{3\dot{q}^2}{4N} + \frac{q^2\dot{\theta}^2}{2N}\right) \,,
\end{aligned}
\end{equation}
and the resulting equations of motion for $q$ and $\theta$ are (introducing $\vartheta=\sqrt{\frac{2}{3}}\theta$),
\begin{equation}
\ddot{q} + q\dot{\vartheta}^2 = 0 \,, \qquad \de{}{t}\big(q^2\dot{\vartheta}\big) = 0 \,,
\end{equation}
which have solution
\begin{equation}
\begin{aligned}
q(t) &= \sqrt{q_0^2(1-t)^2 + q_1^2t^2 + 2q_0q_1\cosh{(\Delta\vartheta)}\,t(1-t)} \,,\\
\vartheta(t) &= \frac{1}{2}\log{\left( \frac{q_0(1-t)+q_1e^{\Delta\vartheta}t}{q_0(1-t)+q_1e^{-\Delta\vartheta}t} \right)} \,.
\end{aligned}
\end{equation}
Much like for the 3-form, there is another candidate solution satisfying the boundary conditions which is discarded because it results in $q^2<0$ in the interval $t\in(0,1)$. Because the above solution is independent of $N$, the remaining one-dimensional integral is particularly simple:
\begin{equation}
\begin{aligned}
Z[q_0,q_1;0,\Delta\theta] &\simeq \int_0^\infty\d{N}\,e^{6\pi^2i\left(N-\frac{q_0q_1}{2N}\big[\cosh{\big(\log{\frac{q_1}{q_0}}\big)}-\cosh{(\Delta\vartheta)}\big]\right)}\\
&= i\sqrt{b-i\epsilon}\,K_1\big(6\pi^2\sqrt{b-i\epsilon}\big) \,,
\end{aligned}
\end{equation}
where $b = 2q_0q_1\big[\cosh{\big(\log{\tfrac{q_1}{q_0}}\big)} - \cosh{(\Delta\vartheta)}\big]$ and $K_1(z)$ is a modified Bessel function of the second kind. For our purposes it will suffice to use the approximation
\begin{equation}
K_1(z) \sim \sqrt{\frac{\pi}{2}}\,\frac{e^{-z}}{\sqrt{z}}
\end{equation}
to arrive at
\begin{equation}
Z[q_0,q_1;0,\Delta\theta] \simeq \frac{i}{\sqrt{12\pi}}(b-i\epsilon)^{1/4}e^{-6\pi^2\sqrt{b-i\epsilon}} \simeq e^{-6\pi^2\sqrt{b-i\epsilon}} \,,
\end{equation}
where again we drop all prefactors. Finally, the remaining Fourier transform takes the form
\begin{equation}\label{eq:FT}
\begin{aligned}
Z[q_0,q_1;n] &\simeq \int\d{\Delta\theta}\,e^{in\Delta\theta}e^{-6\pi^2\sqrt{b-i\epsilon}} \simeq \int\d{\Delta\vartheta}\,e^{i\Phi} \,,\\
\Phi &= 6\pi^2\Big(\tilde{n}\Delta\vartheta + i\sqrt{2q_0q_1}\sqrt{\cosh{(\log{\tfrac{q_1}{q_0}})} - \cosh{(\Delta\vartheta)} - i\epsilon}\Big) \,.
\end{aligned}
\end{equation}
This oscillatory integral may be performed in saddle-point approximation using Picard-Lefschetz theory. One finds that the locations of the saddle-points in $\Delta\vartheta$ and thus the deformation of the initial $\Delta\vartheta$-contour falls into three cases exactly matching those found in the 3-form analysis: see Fig.~\ref{fig:PL_axion}. Evaluating Eqn.~\eqref{eq:FT} on the selected saddle(s) leads to perfect agreement with Eqn.~\eqref{eq:Z3form}, as expected.
\subsection{Higher topologies}
\label{sec:top}
\begin{figure}[h]
\centering
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case5.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.6,1.1,1)$ and $c=0.1$}
\end{subfigure}\hfill
\begin{subfigure}{0.49\textwidth}
\centering
\includegraphics[width=\linewidth]{PL_H_case6.png}
\subcaption{$(q_0,q_1,\tilde{n})=(0.6,1.1,1)$ and $c=-0.1$}
\end{subfigure}
\caption{Structure of $iS[N,q_+,n;c]$ in the complex-$N$ plane for (a) $c>0$ and (b) $c<0$. The new saddles of Eqn.~\eqref{eq:newSaddles} are far from the other saddles near the origin and can change the structure of the deformed contour for large $|N|$ (cf.\ Fig.~\ref{fig:PL_H_2}).}
\label{fig:topPL}
\end{figure}
In the dilute gas approximation wormholes can be argued to contribute non-local terms in the action with coefficients exponentially suppressed by the wormhole action~\cite{Coleman:1988tj, Preskill:1988na, Hebecker:2018ofv}. One can choose to subsequently bring the action to a local form at the cost of introducing $\alpha$-parameters. Determining the details of this non-local action is a monumental task; indeed recently it has been proposed that such contributions should vanish~\cite{McNamara:2020uza, VanRiet:2020pcn, Schlenker:2022dyo}. In this section we show that \emph{if} such terms are present, \emph{then} they can lead to a qualitative change in the Picard-Lefschetz analysis of the transition amplitudes found above. We demonstrate this by working in the 3-form picture where the difference is most clearly seen.
Schematically, bilocal terms induced by a dilute gas of wormholes will contribute terms to the action of the form
\begin{equation}
S \supset \frac{1}{2}\int\d[4]{x}\,\sqrt{-g}\int\d[4]{y}\,\sqrt{-g}\;C_{ij}\mathcal{O}_i(x)\mathcal{O}_j(y) \,.
\end{equation}
Restricting attention to operators of the lowest dimension, such terms will renormalize the couplings of Eqn.~\eqref{eq:gHth_action} but the largest qualitative change comes with the introduction of an effective cosmological constant (for $\mathcal{O}_i=\mathcal{O}_j=1$),
\begin{equation}
S \supset \int\d[4]{x}\,\sqrt{-g}\,\left(-\Lambda_\text{eff}\right) \,, \qquad \Lambda_\text{eff} = \frac{c}{2\pi^2}\int\d[4]{y}\,\sqrt{-g} \,,
\end{equation}
where $|c|\sim e^{-S_\text{wh}}\ll 1$ (the factor of $2\pi^2$ will be convenient). Taking the same ansatz as in Eqn.~\eqref{eq:metricHansatz}, the transition amplitude of Eqn.~\eqref{eq:Hpath} now reads
\begin{equation}
\begin{aligned}
Z[q_0,q_1;n] &\simeq \int_0^\infty\d{N}\int_{q(0)=q_0}^{q(1)=q_1}\mathcal{D}q\, e^{iS[N,q,n;c]} \,,\\
S[N,q,n;c] &= 2\pi^2\int_0^1\d{t}\,\left(3N-\frac{3\dot{q}^2}{4N} - \frac{3\tilde{n}^2N}{q^2} - cN^2q\ev{q}\right) \,,
\end{aligned}
\end{equation}
where
\begin{equation}
\ev{q} = \int_0^1\d{t'}\,q(t') \,.
\end{equation}
The equation of motion for $q$ is an integro-differential equation,
\begin{equation}
\ddot{q} + \frac{4\tilde{n}^2N^2}{q^3} + \frac{4}{3}cN^3\ev{q} = 0 \,,
\end{equation}
which can be solved perturbatively as $q(t) = q_+(t) + c\,\delta q(t) + \mathcal{O}(c^2)$. The correction $\delta q(t)$ can be found exactly in terms of hypergeometric functions, but its detailed form will not be important for our purposes. Indeed, using the zeroth-order equations of motion the action only depends on $q_+$ at leading-order in $c$:
\begin{equation}
S[N,q_+,n;c] = 2\pi^2\left[3N - \frac{3(q_0^2+q_1^2)}{4N}+\frac{3q_0q_1}{2N}\,f\Big(\frac{2\tilde{n}N}{q_0q_1}\Big) - cN^3\ev{q_+}^2 + \mathcal{O}(c^2)\right] \,.
\end{equation}
We have seen previously that there are several saddle points around $N\sim q_0,q_1,\tilde{n}$ (see Eqn.~\eqref{eq:saddles}). For these saddles the $c$-term is subdominant and shifts the locations of the saddle points. A qualitative change to the Picard-Lefschetz analysis comes from the appearance of new saddles points which occur for $N\gg q_0,q_1,\tilde{n}$. In this limit one has $\ev{q_+}^2\approx \frac{\pi^2\tilde{n}N}{16}$ and
\begin{equation}\label{eq:newSaddles}
S \approx 2\pi^2\left(3N - \frac{\pi^2c\tilde{n}}{16}N^4\right) \qquad\implies\qquad N_\ast^3 \approx \frac{12}{\pi^2c\tilde{n}} \,.
\end{equation}
One can check that even with this large value of $N$ the $c$-expansion of $q(t)$ is under control:
\begin{equation*}
\left|\frac{c\,\delta q(t)}{q_+(t)}\right| \xrightarrow{N\to\infty} \left|\frac{\pi c}{64}\, \frac{\frac{3\pi}{8}\big(1+(1-2t)^2\big) - {}_2F_1\big({-\frac{3}{2}},-\frac{1}{2},\frac{1}{2};(1-2t)^2\big)}{t(1-t)} \right| \leq \frac{\pi(3\pi-8)|c|}{128} \ll 1 \,.
\end{equation*}
If $c>0$ then $N_\ast^3>0$ and a new saddle appears on the positive real axis and always contributes as a Lorentzian saddle under the contour deformation for any $q_0,q_1$. If $c<0$ then $N_\ast^3<0$ and these new saddles \emph{never} contribute under the contour deformation. See Fig.~\ref{fig:topPL} for two representative cases.
\section{Boundary conditions \& stability}
\label{sec:stability}
Gravitational path integrals famously suffer from issues of convergence. Candidate saddle points of the Euclidean path integral should be minima so that the action at the critical point truly represents the dominant contribution from configurations near this point in field space. Saddle points (with their unstable directions) can be interpreted as mediating decay. Of course, statements of stability should only refer to gauge-invariant degrees of freedom.
In the previous sections we have restricted attention to spatially-uniform fields which obscures whether the contributing saddle points are truly stable in the appropriate sense. In order to address the question of stability we will analyze scalar perturbations around the GS wormhole in the 3-form picture, the spectrum of which depends intimately on the chosen boundary conditions; it is natural to choose Dirichlet boundary conditions for the 3-form because of flux quantization. As we saw in some detail in Sec.~\ref{sec:duality}, the duality which relates the 3-form and axion includes a correspondence between boundary conditions in the two frames: Dirichlet boundary conditions for the 3-form correspond to Neumann boundary conditions for the axion (equivalently, the Fourier transform of Dirichlet boundary conditions for the axion, in the sense discussed in Sec.~\ref{sec:duality}). Normalizable perturbations of the 3-form, namely those with finite energy for which
\begin{equation}
\int\delta H\wedge{\star\delta H} < \infty \,,
\end{equation}
correspond, via $H\leftrightarrow{\star\d{\theta}}$, to perturbations of the axion which approach constant values at the boundaries and which have finite energy
\begin{equation}
\int\d{\delta\theta}\wedge{\star\d{\delta\theta}} < \infty \,,
\end{equation}
even if they are not normalizable in the sense that
\begin{equation}
\int\star(\delta\theta^2) \to \infty \,.
\end{equation}
This is natural in view of the axion's shift symmetry; a constant shift to the background field profile can be implemented with a constant perturbation which has zero energy but divergent ``norm''. Consequently the discussion of fluctuations around the GS wormhole is most transparent in the 3-form picture where Dirichlet boundary conditions are required by flux quantization and, perhaps more importantly, the criteria of normalizability and finite energy coincide.
In similar spirit to the previous sections we will remain in Lorentzian signature until absolutely necessary. We parametrize the fields as
\begin{equation}
\begin{aligned}
\d{s^2} &= a(\eta)^2\Big\{ {-(1+2\phi)}\,\d{\eta^2} + 2\partial_iB\,\d{x^i}\d{\eta} + \big[(1-2\psi)\gamma_{ij} + 2\nabla_i\partial_jE\big]\,\d{x^i}\d{x^j} \Big\} \,,\\
H &= \sqrt{6}\,\tilde{n}\left[(1+s)\,\mathrm{vol}_3 + \d{\eta}\wedge\left(\frac{1}{2}\sqrt{\gamma}\gamma^{ij}\epsilon_{jkl}\partial_iw\,\d{x^k}\wedge\d{x^l}\right)\right] \,,
\end{aligned}
\end{equation}
where $\gamma_{ij}$ is the (fixed) round metric on $S^3$ and $\nabla$ the corresponding covariant derivative. Note that we are using conformal time $\eta$, so now $\dot{A}=\de{A}{\eta}$. It will be useful to introduce $\mathcal{H}=\frac{\dot{a}}{a}$, in terms of which the zeroth-order Einstein equations amount to
\begin{equation}\label{eq:hubbleEOM}
1+\mathcal{H}^2 = \frac{\tilde{n}^2}{a^4} > 0 \,.
\end{equation}
Returning to the action of Eqn.~\eqref{eq:gHth_action} and using the above parametrization results in the following quadratic action for the perturbations,
\begin{equation}
\begin{aligned}
S_2 &= \int\d{\eta}\d[3]{x}\,\sqrt{\gamma}\,a^2\Big\{ {-3}\big(\dot{\psi} + \mathcal{H}\phi\big)^2 + \big(B-\dot{E}\big)\Delta\big(B-\dot{E}\big) - 2\big(\dot{\psi}+\mathcal{H}\phi\big)\Delta\big(B-\dot{E}\big)\\
&\qquad\qquad - 3\big(1+\mathcal{H}^2\big)\big[(\phi+3\psi-\Delta E+s)^2 -\phi^2+ (B-w)\Delta(B-w)\big]\\
&\qquad\qquad + (2\phi-\psi)(\Delta+3)\psi\Big\} + \sqrt{6}\tilde{n}\int\d{\eta}\d[3]{x}\,\sqrt{\gamma} \,(\dot{s}-\Delta w)\theta \,,
\end{aligned}
\end{equation}
where we have used integration by parts on $S^3$ liberally. The individual perturbations are not gauge-invariant; under a diffeomorphism $\xi=\zeta^0\partial_\eta + \gamma^{ij}(\partial_i\zeta)\partial_j$ parametrized by the two scalar functions $\zeta^0,\zeta$ the perturbations transform according to $\mathcal{L}_\xi g$ and $\mathcal{L}_\xi H$:
\begin{equation}
\begin{aligned}
\delta_\xi\phi &= \dot{\zeta}^0+\mathcal{H}\zeta^0 \,, & \qquad \delta_\xi B &= -\zeta^0+\dot{\zeta} \,, & \qquad \delta_\xi s &= \Delta\zeta \,,\\
\delta_\xi\psi &= -\mathcal{H}\zeta^0 \,, & \delta_\xi E &= \zeta \,, & \delta_\xi w &= \dot{\zeta} \,.
\end{aligned}
\end{equation}
Physically meaningful statements can only be made about linear combinations of perturbations which are gauge-invariant.
To proceed it is useful to reduce to 1D by writing all fields in terms of hyperspherical harmonics, e.g.
\begin{equation}
\phi(\eta,x^i) = \sum_{j\geq0} \phi_j(\eta)Y_j(x^i)
\end{equation}
where
\begin{equation}
\Delta Y_j = -\lambda_jY_j \,, \qquad \int_{S^3}\d[3]{x}\,\sqrt{\gamma}\,Y_jY_{j'} = \delta_{jj'} \,.
\end{equation}
The degeneracy of the eigenvalue $\lambda_j=j(j+2)\in\{0,3,8,15,\ldots\}$ is $(j+1)^2$; we suppress labels which distinguish degenerate states for simplicity. The action decomposes into sectors labeled by the integer $j$:
\begin{equation}
\begin{aligned}
S_2 &= \int\d{\eta}\,\sum_{j\geq0}\mathcal{L}_j \,,\\
\mathcal{L}_j &= a^2\Big\{ {-3}\big(\dot{\psi}_j + \mathcal{H}\phi_j\big)^2 - \lambda_j\big(B_j-\dot{E}_j\big)^2 + 2\lambda_j\big(\dot{\psi}_j+\mathcal{H}\phi_j\big)\big(B_j-\dot{E}_j\big)\\
&\qquad\qquad - 3\big(1+\mathcal{H}^2\big)\big[(\phi_j+3\psi_j + \lambda_jE_j+s_j)^2 -\phi_j^2 - \lambda_j(B_j-w_j)^2\big]\\
&\qquad\qquad - (\lambda_j-3)(2\phi_j-\psi_j)\psi_j \Big\} + \sqrt{6}\tilde{n}(\dot{s}_j+\lambda_jw_j)\theta_j \,.
\end{aligned}
\end{equation}
For now let us focus on the sectors with $j\geq2$ ($\lambda_j\geq8$) which can all be treated simultaneously. Performing the path integral over the Lagrange multiplier $\theta_j$ results in a $\delta$-function imposing the (gauge-invariant) condition $\dot{s}_j+\lambda_jw_j=0$. Using this to eliminate $w_j$ leads to
\begin{equation}
\begin{aligned}
\mathcal{L}_j &= a^2\Big\{ {-3}\big(\dot{\psi}_j + \mathcal{H}\phi_j\big)^2 - \lambda_j\big(B_j-\dot{E}_j\big)^2 + 2\lambda_j\big(\dot{\psi}_j+\mathcal{H}\phi_j\big)\big(B_j-\dot{E}_j\big)\\
&\qquad\qquad - 3\big(1+\mathcal{H}^2\big)\big[(\phi_j+3\psi_j + \lambda_jE_j+s_j)^2 -\phi_j^2 - \lambda_j^{-1}(\dot{s}_j+\lambda_jB_j)^2\big]\\
&\qquad\qquad - (\lambda_j-3)(2\phi_j-\psi_j)\psi_j \Big\} \,.
\end{aligned}
\end{equation}
All of $\psi_j,E_j,s_j$ are dynamical and have conjugate momenta given by
\begin{equation}
\begin{aligned}
\Pi_j^\psi &= 2a^2\Big[ {-3}\big(\dot{\psi}_j+\mathcal{H}\phi_j\big) + \lambda_j\big(B_j-\dot{E}_j\big) \Big] \,,\\
\Pi_j^E &= 2a^2\Big[ {-\lambda_j}\big(\dot{\psi}_j+\mathcal{H}\phi_j\big) + \lambda_j\big(B_j-\dot{E}_j\big) \Big] \,,\\
\Pi_j^s &= 6\lambda_j^{-1}a^2\big(1+\mathcal{H}^2\big)(\dot{s}_j+\lambda_jB_j) \,.
\end{aligned}
\end{equation}
In terms of these we may write the Lagrangian in first-order form:
\begin{equation}
\begin{aligned}
\mathcal{L}_j &= \Pi_j^\psi\dot{\psi}_j + \Pi_j^E\dot{E}_j + \Pi_j^s\dot{s}_j - (\Pi_j^E-\lambda_j\Pi_j^s)B_j\\
&\qquad + \big[\mathcal{H}\Pi_j^\psi - 6a^2\big(1+\mathcal{H}^2\big)(3\psi_j+\lambda_jE_j+s_j) - 2a^2(\lambda_j-3)\psi_j\big]\phi_j\\
&\qquad + a^{-2}\left[ -\frac{(\Pi_j^\psi)^2}{4(\lambda_j-3)} + \frac{\Pi_j^\psi\Pi_j^E}{2(\lambda_j-3)} - \frac{3(\Pi_j^E)^2}{4\lambda_j(\lambda_j-3)} - \frac{\lambda_j(\Pi_j^s)^2}{12(1+\mathcal{H}^2)} \right]\\
&\qquad - a^2\Big[ 3\big(1+\mathcal{H}^2\big) (3\psi_j+\lambda_jE_j+s_j)^2 -(\lambda_j-3)\psi_j^2 \Big] \,.
\end{aligned}
\end{equation}
This is linear in both of the non-dynamical fields, $\phi_j$ and $B_j$, and performing the path integral over them produces two more (gauge-invariant) $\delta$-function constraints. Using these to integrate out $\Pi_j^\psi$ and $\Pi_j^E$ gives
\begin{equation}
\begin{aligned}
\mathcal{L}_j &= \Pi_j^S\dot{S}_j - \frac{\lambda_j}{12a^2}\left(\frac{9}{\lambda_j-3} + \frac{1}{1+\mathcal{H}^2} \right)(\Pi_j^S)^2\\
&\qquad + \frac{3\lambda_j(1+\mathcal{H}^2)}{(\lambda_j-3)\mathcal{H}}\Pi_j^SS_j - \frac{3a^2(1+\mathcal{H}^2)^2}{\mathcal{H}^2}\left(\frac{\lambda_j}{\lambda_j-3} - \frac{1}{1+\mathcal{H}^2}\right)S^2
\end{aligned}
\end{equation}
(plus a total derivative involving $\psi_j$), where we have introduced the gauge-invariant field $S_j$ and its gauge-invariant conjugate momentum $\Pi_j^S$ defined by
\begin{equation}
\begin{aligned}
S_j &= s_j + \lambda_jE_j \,,\\
\Pi_j^S &= \Pi_j^s - 6a^2(1+\mathcal{H}^2)\mathcal{H}^{-1}\psi \,.
\end{aligned}
\end{equation}
The fields $\psi_j,E_j$ no longer appear; we can interpret the path integral over these as giving the gauge-orbit volume. Finally, integrating out $\Pi_j^S$ returns the Lagrangian to second-order form, now for the sole physical, gauge-invariant scalar perturbation:
\begin{equation}
\mathcal{L}_j = \frac{3a^2}{\lambda_j\big(\frac{9}{\lambda_j-3}+\frac{1}{1+\mathcal{H}^2}\big)}\left[ \dot{S}_j^2 + \frac{6\lambda_j\big(1+\mathcal{H}^2\big)}{(\lambda_j-3)\mathcal{H}}S_j\dot{S}_j - \lambda_j\left(\frac{\lambda_j-9}{\lambda_j-3}\big(1+\mathcal{H}^2\big) - 1\right)S_j^2 \right] \,.
\end{equation}
A similar analysis for $j=0$ and $j=1$ reveals that these two sectors are pure-gauge and as such do not correspond to any physical perturbations.\footnote{For $j=0$ the Lagrange multiplier imposes $\dot{s}_0=0$, so $s_0=0$ for the given boundary conditions. Additionally, all terms involving $B_0,E_0,w_0$ vanish and of the two remaining fields, $\phi_0$ and $\psi_0$, only $\psi_0$ is dynamical. Transforming to gauge-invariant variables and integrating out the remaining non-dynamical field shows that $\mathcal{L}_0$ is a total derivative. Similarly, for $j=1$ one can replace $w_1\to-\dot{s}_1$ using the constraint to find that there are two dynamical fields, $\psi_1+E_1$ and $s_1$, and two non-dynamical fields, $\phi_1$ and $B_1$ (the linear combination $\psi_1-E_1$ does not appear). Integrating out the non-dynamical fields imposes two gauge-invariant constraints which result in $\mathcal{L}_1$ being a total derivative.} In particular, there is no conformal factor problem in the homogeneous ($j=0$) sector (the same observation was made in~\cite{Hertog:2018kbz}). This can be understood as a consequence of the zeroth-order Einstein equations of Eqn.~\eqref{eq:hubbleEOM} being first-order. In terms of the canonically normalized field
\begin{equation}
Q_j = \sqrt{\frac{6a^2}{\lambda_j\big(\frac{9}{\lambda_j-3}+\frac{1}{1+\mathcal{H}^2}\big)}}\;S_j
\end{equation}
(note that the ratio in the square-root is positive since $1+\mathcal{H}^2>0$ by Eqn.~\eqref{eq:hubbleEOM} and $\lambda_j>3$) the Lagrangian is
\begin{align}
\mathcal{L}_j &= \frac{1}{2}\dot{Q}_j^2 - \frac{1}{2}\big[U_j + (\lambda_j+1)\big] Q_j^2 + \de{}{\eta}\left[ \left(\frac{\mathcal{H}}{2}\frac{9(1+\mathcal{H}^2)-(\lambda_j-3)}{9(1+\mathcal{H}^2)+(\lambda_j-3)}+\frac{3\lambda_j(1+\mathcal{H}^2)}{2\mathcal{H}(\lambda_j-3)}\right)Q_j^2 \right] \,, \notag\\
U_j &= -\big(1+\mathcal{H}^2\big)\left(4(\lambda_j+6)\,\frac{9(1+\mathcal{H}^2)+4(\lambda_j-3)}{[9(1+\mathcal{H}^2)+(\lambda_j-3)]^2} - 1 \right) \,.
\end{align}
At this point there is no obstacle to performing the standard Wick rotation to Euclidean signature, $\eta\to-ir$. Under this rotation it is useful to introduce $A'\equiv\de{A}{r}$ and $\mathcal{H}_\text{E}(r) = -i\mathcal{H}(ir) = \frac{a'(r)}{a(r)}$ for clarity; the replacement $\mathcal{H}^2\to-\mathcal{H}_\text{E}^2$ should be made in both $Q_j$ and $U_j$. The equation of motion Eqn.~\eqref{eq:hubbleEOM} now reads
\begin{equation}
1 \geq 1 - \mathcal{H}_\text{E}^2 = \frac{\tilde{n}^2}{a^4} > 0
\end{equation}
and has solution
\begin{equation}
a(r) = \sqrt{\tilde{n}\cosh{(2r)}} \,, \qquad \mathcal{H}_\text{E}(r) = \tanh{(2r)} \,.
\end{equation}
This is the GS wormhole; the wormhole throat is at $r=0$ and the two asymptotically flat regions are $r\to\pm\infty$. The Euclidean action becomes
\begin{align}
S_\text{E} &= \int\d{r}\sum_{j\geq2}\left\{\frac{1}{2}Q_j\Big[\hat{\mathcal{O}}_j + (\lambda_j+1)\Big]Q_j + G_j\right\} \,, \notag\\
\hat{\mathcal{O}}_j &= -\de[2]{}{r} + U_j \,, \\
G_j &= \frac{1}{2}(Q_jQ_j')' + \mathcal{H}_\text{E}\left( \frac{9(1-\mathcal{H}_\text{E}^2)-(\lambda_j-3)}{9(1-\mathcal{H}_\text{E}^2)+(\lambda_j-3)}-\frac{3\lambda_j(1-\mathcal{H}_\text{E}^2)}{(\lambda_j-3)\mathcal{H}_\text{E}^2} \right)Q_jQ_j' \notag\\
&\quad + (1-\mathcal{H}_\text{E}^2)\left( \frac{[9(1-\mathcal{H}_\text{E}^2)+2(\lambda_j-3)]^2-(\lambda_j-3)(5\lambda_j+21)}{[9(1-\mathcal{H}_\text{E}^2)+(\lambda_j-3)]^2} + \frac{3\lambda_j(1+\mathcal{H}_\text{E}^2)}{(\lambda_j-3)\mathcal{H}_\text{E}^2} \right)Q_j^2 \notag\\
&= \frac{1}{2}\left[Q_jQ_j' + \mathcal{H}_\text{E}\left(\frac{9(1-\mathcal{H}_\text{E}^2)-(\lambda_j-3)}{9(1-\mathcal{H}_\text{E}^2)+(\lambda_j-3)}-\frac{3\lambda_j(1-\mathcal{H}_\text{E}^2)}{(\lambda_j-3)\mathcal{H}_\text{E}^2}\right)Q_j^2\right]' \,, \notag
\end{align}
where the constant $\lambda_j+1$ has been stripped off in defining $U_j$ so that $U_j\to0$ in the two asymptotic regions (see Fig.~\ref{fig:potentials}); note, however, that $U_j+(\lambda_j+1)>0$ everywhere. The term $G_j$ is a total derivative and has no effect on the equations of motion for $Q_j$. Nevertheless, once solutions have been found one has to check that $S_\text{E}$ is positive definite when including the contributions from integrating over $G_j$.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{potentials.png}
\caption{The potentials $U_j$ for the Schr\"odinger problem of Eqn.~\eqref{eq:Schrodinger}.}
\label{fig:potentials}
\end{figure}
Finding the spectrum of scalar fluctuations amounts to finding eigenvalues of the fluctuation operators $\hat{\mathcal{O}}_j$ given appropriate boundary conditions. For each $j$ this is a standard Schr\"odinger-type problem with potential $U_j$:
\begin{equation}\label{eq:Schrodinger}
\hat{\mathcal{O}}_jQ_j^{(k)} = \left(-\de[2]{}{r} + U_j\right)Q_j^{(k)} = \omega_j^{(k)}Q_j^{(k)} \,.
\end{equation}
What boundary conditions should be required for the canonically normalized fields? Although physical perturbations should vanish at the boundaries, the normalization of $Q_j$,
\begin{equation}
Q_j = \sqrt{\frac{6\tilde{n}\cosh{(2r)}}{\lambda_j\big[\frac{9}{\lambda_j-3}+\cosh^2{(2r)}\big]}}\;S_j \quad\xrightarrow{|r|\to\infty}\quad \sqrt{\frac{12\tilde{n}}{\lambda_j}}\;e^{-|r|}S_j \,,
\end{equation}
shows that we should insist that $Q_j\to0$ \emph{faster} than $e^{-|r|}$ for $|r|\to\infty$ in order that the corresponding perturbation $S_j$ goes to zero in both asymptotic regions. Any bound states of energy \emph{strictly less than} ${-1}$ will have exponential tails in the classically forbidden region which die off quickly enough. Note also that the normalizing factor in $Q_j$ is well-behaved near $r=0$, so no additional conditions need be imposed there.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{eigenfunctions.png}
\caption{(Left) Even and odd bound states for $j=2$ and $j\to\infty$. (Right) The functions $G_j(r)$ for the even and odd $j=2$ bound states. For the even bound state the integral over $G_j$ diverges while for the odd bound state it vanishes.}
\label{fig:eigen}
\end{figure}
For $j\gg1$ the functions $U_j$ approach a universal, P\"oschl--Teller form,
\begin{equation}
U_\infty = -\frac{15}{\cosh^2{(2r)}} \,,
\end{equation}
for which there are only two bound states, one even in $r$ and the other odd:
\begin{equation}
\begin{aligned}
Q_\infty^{\text{(e)}} &= \frac{2}{\sqrt{\pi}}\frac{1}{[\cosh{(2r)}]^{3/2}} \,, & \qquad \omega_\infty^{\text{(e)}} &= -9 \,,\\
Q_\infty^{\text{(o)}} &= \frac{2}{\sqrt{\pi}}\frac{\sinh{(2r)}}{[\cosh{(2r)}]^{3/2}} \,, & \omega_\infty^{\text{(o)}} &= -1 \,.
\end{aligned}
\end{equation}
One can check numerically that exactly two bound states, one even and one odd, exist for all $j\geq2$. These bound states, $Q_j^\text{(e)}$ and $Q_j^\text{(o)}$, are perturbed versions of their $j\to\infty$ counterparts: see Fig.~\ref{fig:eigen}. Although $Q_\infty^{\text{(o)}}$ marginally violates the required fall-off conditions, one can use standard time-independent perturbation theory to show that
\begin{equation}\label{eq:lambdaapprox}
\begin{aligned}
\omega_j^\text{(e)} &= -9 + \frac{99}{2\lambda_j} + \mathcal{O}(\lambda_j^{-2}) \,,\\
\omega_j^\text{(o)} &= -1 - \frac{9}{2\lambda_j} + \mathcal{O}(\lambda_j^{-2}) \,,
\end{aligned}
\end{equation}
so that the required fall-off conditions at $|r|\to\infty$ are in fact satisfied by the odd bound states for finite $j$. In evaluating the action for these states one finds that $G_j$ diverges near $r=0$ as $G_j\sim \frac{1}{r^2}>0$ for the even bound states $Q_j^{\text{(e)}}$; the Euclidean action diverges $S_\text{E}\to+\infty$ and these states are discarded. In contrast, the odd states survive as finite-action perturbations since $G_j$ is everywhere finite and integrates to zero: see Fig.~\ref{fig:eigen}. All told there is (up to the $(j+1)^2$ degeneracy) a unique admissible scalar perturbation for each $j\geq2$ with
\begin{equation}
S_\text{E} = \frac{1}{2}\int\d{r}\,Q_j\Big[\hat{\mathcal{O}}_j+(\lambda_j+1)\Big]Q_j = \frac{1}{2}\big(\omega_j^\text{(o)} + \lambda_j+1\big) > 0 \,.
\end{equation}
The low-lying spectrum is found using finite-difference techniques and presented in Tab.~\ref{tab:spectrum}. We conclude that the GS wormhole is perturbatively stable.
\begin{table}[t]
\centering
\begin{tabular}{c|ccc||c|ccc}
$j$ & $-1-\frac{9}{2\lambda_j}$ & $\omega_j^\text{(o)}$ & $\omega_j^\text{(o)}+\lambda_j+1$ & $j$ & $-1-\frac{9}{2\lambda_j}$ & $\omega_j^\text{(o)}$ & $\omega_j^\text{(o)}+\lambda_j+1$\\ \hline
2 & $-1.5625$ & $-1.5335$ & $\phantom{0}7.4665$ & 6 & $-1.0938$ & $-1.0921$ & $47.9079$\\
3 & $-1.3000$ & $-1.2873$ & $14.7127$ & 7 & $-1.0714$ & $-1.0705$ & $62.9295$\\
4 & $-1.1875$ & $-1.1817$ & $23.8183$ & 8 & $-1.0563$ & $-1.0556$ & $79.9444$\\
5 & $-1.1286$ & $-1.1256$ & $34.8744$ & 9 & $-1.0455$ & $-1.0450$ & $98.9550$
\end{tabular}
\caption{Low-lying spectrum of finite-action scalar perturbations, along with the approximation of Eqn.~\eqref{eq:lambdaapprox}.}
\label{tab:spectrum}
\end{table}
\section{Discussion}
\label{sec:disc}
In this paper, we have explored several aspects of the dual axion-gravity and gravity$+$3-form theories directly with the Lorentzian path integral. Using Picard-Lefschetz theory we are able to identify which subset of (complex) saddle points contribute to the Lorentzian path integrals computing simple transition amplitudes. As the boundary values are adjusted the interpretation in terms of Lorentzian-time evolution vs.\ Euclidean wormhole changes, but nevertheless they may all be treated democratically in the complex-$N$ plane. All four saddle points are compatible with the K-S criterion~\cite{Kontsevich2021} but only ever one or two are selected by the contour deformation. It would be interesting to understand if another principle could be used to omit the saddles that do not contribute according to Picard-Lefschetz theory.
We considered scalar perturbations to the GS wormhole in the 3-form picture, where it was found that (i) there is no conformal factor problem in the homogeneous sector and (ii) there are no negative modes amongst the physical perturbations of nonzero angular momentum. The perturbative stability of the GS wormhole presents several puzzles. For one, our results disagree with the conclusion of~\cite{Hertog:2018kbz} which analyzed stability in the axion picture and found that there are infinitely many negative modes. Here the application of boundary conditions is more straightforward since both metric and 3-form scalar perturbations are both subject to Dirichlet boundary conditions, whereas the axion is subject to Neumann boundary conditions and the gauge-invariant field chosen in~\cite{Hertog:2018kbz} mixes metric and axion perturbations. We have also seen the importance of keeping all boundary terms for the gauge-invariant modes $Q_j$, namely in ruling out the even-parity perturbations for which the total derivative terms $G_j$ are not integrable.
Our findings contribute to the long-standing debate on whether Euclidean wormholes can be embedded into string theory. Previous studies found negative modes \cite{Hertog:2018kbz,Rubakov:1996cn,Kim:2003js} suggesting that axionic Euclidean wormholes are not saddle points of the gravitational path integral. Here we showed that with appropriate boundary conditions, the spectrum of quadratic fluctuations contains only modes with positive eigenvalues. This suggests the interpretation that rather than mediating decay these wormholes compute the non-perturbative energy splitting for the degenerate vacua with flux $\pm n$ (note that for $n=0$ this 2-fold degeneracy disappears, as does the GS wormhole).
While we have analyzed the stability of axionic wormholes in asymptotically-flat space, the conclusion of stability may carry over to AdS since the perturbations we find are localized to the wormhole's throat whose size is much less than the AdS curvature.
These Euclidean wormholes, if embeddable into AdS compactifications, pose a puzzle for AdS/CFT as they seem to jeopardize factorization of the two boundary CFTs.
Given this potential tension with AdS/CFT the embeddability of these solutions into string theory is called into question. There are a few logically possible resolutions:
\begin{enumerate}[label=\roman*)]
\item The stability of axionic Euclidean wormholes changes qualitatively with the inclusion of a dilaton (which always comes along for the ride with an axion) or nonzero cosmological constant. It would be interesting to extend our current work to analyze the spectrum of gauge invariant perturbations of such axio-dilatonic Euclidean wormholes. It may also be that these wormholes are non-perturbatively unstable to brane nucleation, such as was found in~\cite{Marolf:2021kjc}.
\item The existence of a regular wormhole solution requires finding a long-enough timelike geodesic in the scalar moduli space \cite{Arkani-Hamed:2007cpn}. This in turn implies conditions on the dilaton couplings \cite{Arkani-Hamed:2007cpn, Astesiano:2022qba}. It may happen that such couplings are not realizable in controlled limits of string theory. These embeddability criteria are reminiscent of swampland criteria such as the distance conjecture \cite{Ooguri:2006in} which limits the field range for a single effective field theory description to be valid (albeit for positive definite field space metric). Our Lorentzian, 3-form approach may allow us to formulate these embeddability criteria with a more familiar positive-definite moduli space metric.
\item The axio-dilatonic Euclidean wormholes, if shown to be stable and embeddable into AdS compactifications, suggest a more general interpretation of the AdS/CFT duality in $D\geq3$, for example involving an ensemble-average. This would be at odds with~\cite{McNamara:2020uza} and~\cite{Schlenker:2022dyo} (see also~\cite{Heckman:2021vzx}).
\end{enumerate}
Further studies along these lines may teach us useful lessons about quantum gravity. We plan to investigate these possibilities in the future.
\acknowledgments
We thank Thomas Van Riet for useful discussions. The work of GL and GS is supported in part by the DOE grant DE-SC0017647.
\newpage
\bibliographystyle{utphys}
| 980a10145864f41e6655cbe2c1de967c9035eebe | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Robots supporting people in their daily activities at home or at the workplace need to accurately and robustly perceive objects, such as containers, and their physical properties, for example when they are manipulated by a person prior to a human-to-robot handover~\cite{Sanchez-Matilla2020,Medina2016,Rosenberger2021RAL,Ortenzi2021TRO,Yang2021ICRA}. Audio-visual perception should adapt -- on-the-fly and with limited or no prior knowledge -- to changing conditions in order to guarantee the correct execution of the task and the safety of the person. For assistive scenarios at home, audio-visual perception should accurately and robustly estimate the physical properties (e.g., weight and shape) of household containers, such as cups, drinking glasses, mugs, bottles, and food boxes~\cite{Sanchez-Matilla2020,Ortenzi2021TRO,Liang2020MultimodalPouring,Modas2021ArXiv,Xompero2021_ArXiv}. However, the material, texture, transparency and shape can vary considerably across containers and also change with their content, which may not be visible due to the opaqueness of the container or occlusions, and hence should be inferred through the behaviour of the human~\cite{Sanchez-Matilla2020,Modas2021ArXiv,Xompero2021_ArXiv,Mottaghi2017ICCV,Duarte2020ICDL_EpiRob}.
In this paper, we present the tasks and the results of the CORSMAL challenge at IEEE ICASSP 2022, supporting the design and evaluation of audio-visual solutions for the estimation of the physical properties of a range of containers manipulated by a person prior to a handover (see Fig.~\ref{fig:avsamples}).
The specific containers and fillings are not known in advance, and the only priors are the sets of object categories ({drinking glasses}, {cups}, {food boxes}) and filling types ({water}, {pasta}, {rice}). The estimation of the mass and dimensions of the containers are novel tasks of this challenge, and complement the tasks of its previous version~\cite{Xompero2021_ArXiv}, such as the estimation of the container capacity and the type, mass and amount of the content. We carefully defined a set of performance scores to directly evaluate and systematically compare the algorithms on each task. Moreover, to assess the accuracy of the estimations and visualise the safeness of human-to-robot handovers, we implemented a real-to-simulation framework~\cite{Pang2021ROMAN} that provides indirect high-level evaluations on the impact of these tasks (see Fig.~\ref{fig:challengetasksdiagram}). The source code of the entries to the challenge and the up-to-date leaderboards are available at
\mbox{\url{http://corsmal.eecs.qmul.ac.uk/challenge.html}}.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{challenge_image.png}
\caption{Sample video frames and audio spectrograms of people manipulating objects prior to handing them over to a robot.}
\label{fig:avsamples}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{diagram_tasks.eps}
\caption{The challenge tasks feeding into the CORSMAL simulator~\cite{Pang2021ROMAN} to evaluate the impact of estimation errors. Given video frames and audio signals from the CORSMAL Containers Manipulation (CCM) dataset~\cite{Xompero2021_ArXiv,Xompero_CCM}), the results of T1 (filling level), T2 (filling type), and T3 (container capacity) are used to compute the filling mass, which is added to T4 (container mass) for estimating the mass of the object (container + filling). The estimated dimensions (T5) are used to visualise the container. The simulator also uses object annotations, such as 6D poses over time, the true weight (container + filling), a 3D mesh model reconstructed offline with a vision baseline~\cite{Pang2021ROMAN}, and the frame where the object is ready to be grasped by the simulated robot arm, for performing and visualising the handover.
}
\label{fig:challengetasksdiagram}
\end{figure*}
\section{The tasks}
\label{sec:tasks}
In the scope of the challenge and based on the reference dataset~\cite{Xompero2021_ArXiv,Xompero_CCM}, containers vary in shape and size, and may be empty or filled with an unknown content at 50\% or 90\% of its capacity.
We define a configuration as the manipulation of a container with a filling type and amount under a specific setting (i.e., background, illumination, scenario).
The challenge features five tasks (Ts), each associated with a physical property to estimate for each configuration $j$.
\begin{description}
\item[Filling level classification (T1).] The goal is to classify the filling level ($\tilde{\lambda}^j$) as empty, 50\%, or 90\%.
\item[Filling type classification (T2).] The goal is to classify the type of filling ($\tilde{\tau}^j$), if any, as one of these classes: 0 (no content), 1 (pasta), 2 (rice), 3 (water).
\item[Container capacity estimation (T3).] The goal is to estimate the capacity of the container ($\tilde{\gamma}^j$, in mL).
\item[Container mass estimation (T4).] The goal is to estimate the mass of the (empty) container ($\tilde{m}_{c}^j$, in g).
\item[Container dimensions estimation (T5).] The goal is to estimate the width at the top ($\tilde{w}_t^j$, in mm) and at the bottom ($\tilde{w}_b^j$, in mm), and height ($\tilde{h}^j$, in mm) of the container.
\end{description}
Algorithms designed for the challenge are expected to estimate these physical properties to compute the mass of the filling as
\begin{equation}
\tilde{m}_f^j = \tilde{\lambda}^j \tilde{\gamma}^j D(\tilde{\tau}^j),
\label{eq:fillingmass}
\end{equation}
where $D(\cdot)$ selects a pre-computed density based on the classified filling type. The mass of the object $\tilde{m}$ is calculated as the sum of the mass of the empty container and the mass of the content, if any.
\section{The evaluation}
\label{sec:evaluation}
\subsection{Data}
CORSMAL Containers Manipulation (CCM)~\cite{Xompero2021_ArXiv,Xompero_CCM} is the reference dataset for the challenge and consists of 1,140 visual-audio-inertial recordings of people interacting with 15 container types: 5 drinking cups, 5 drinking glasses, and 5 food boxes.
These containers are made of different materials, such as plastic, glass, and cardboard. Each container can be empty or filled with water, rice or pasta at two different levels of fullness: 50\% or 90\% with respect to the capacity of the container. In total, 12 subjects of different gender and ethnicity\footnote{An individual who performs the manipulation is referred to as \textit{subject}. Ethical approval (QMREC2344a) was obtained at Queen Mary University of London, and consent from each person was collected prior to data collection.} were invited to execute a set of 95 configurations as a result of the combination of containers and fillings, and for one of three manipulation scenarios. The scenarios are designed with an increasing level of difficulty caused by occlusions or subject motions, and recorded with two different backgrounds and two different lighting conditions to increase the visual challenges for the algorithms. The annotation of the data includes the capacity, mass, maximum width and height (and depth for boxes) of each container, and the type, level, and mass of the filling. The density of pasta and rice is computed from the annotation of the filling mass, capacity of the container, and filling level for each container. Density of water is 1 g/mL.
For validation, CCM is split into a training set (recordings of 9 containers), a public test set (recordings of 3 containers), and a private test set (recordings of 3 containers). The containers for each set are evenly distributed among the three categories.
The annotations are provided publicly only for the training set.
\subsection{Real-to-sim visualisation}
The challenge adopts a real-to-simulation framework~\cite{Pang2021ROMAN} that complements the CCM dataset with a human-to-robot handover in the PyBullet simulation environment~\cite{coumans2019pybullet}. The framework uses the physical properties of a manipulated container estimated by a perception algorithm.
The handover setup recreated in simulation consists of a 6 DoF robotic arm (UR5) equipped with a 2-finger parallel gripper (Robotiq 2F-85), and two tables.
The simulator renders a 3D object model reconstructed offline by a vision baseline in manually selected frames with no occlusions~\cite{Pang2021ROMAN}. The weight of the object used by the simulator is the true, annotated value. We manually annotated the poses of the containers for each configuration of CCM every 10 frames and interpolated the intermediate frames. We also annotated the frame where the person started delivering the object to the robot arm. We use the annotated and interpolated poses to render the motion of the object in simulation and control the robot arm to approach the object at the annotated frame for the handover. If the robot is not able to reach the container before the recording ends, the last location of the container is kept for 2~s.
When reaching the container, the simulated robot arm closes the gripper to 2~cm less than the object width to ensure good contact with the object, and applies an amount of force determined by the estimated weight of the object to grasp the container. Note that in the scope of the challenge, we avoid simulating the human hands so that the object is fully visible and can be grasped by the robot arm. The simulator visualises whether the estimations enable the robot to successfully grasp the container without dropping it or squeezing it. After grasping the container, the robot delivers it to a target area on a table via a predefined trajectory.
\subsection{Scores}
To provide sufficient granularity into the behaviour of the various components of the audio-visual algorithms and pipelines, we compute {13 performance scores} individually for the public test set (no annotations available to the participants), the private test set (neither data nor annotations are available to the participants), and their combination. All scores are in the range $[0,1]$. With reference to Table~\ref{tab:scores}, the first 7 scores quantify the accuracy of the estimations for the 5 main tasks and include filling level, filling type, container capacity, container width at the top, width at the bottom, and height, and container mass.
Other 3 scores evaluate groups of tasks and assess filling mass, joint filling type and level classification, joint container capacity and dimensions estimation. The last 2 scores are an indirect evaluation of the impact of the estimations (i.e., the object mass) on the quality of human-to-robot handover and delivery of the container by the robot in simulation.
\textbf{T1 and T2.} For filling level and type classification, we compute precision, recall, and F1-score for
each class $k$ across all the configurations of that class, $J_k$. \textit{Precision} is the number of true positives divided by the total number of true positives and false positives for each class $k$ ($P_k$). \textit{Recall} is the number of true positives divided by the total number of true positives and false negatives for each class $k$ ($R_k$). \textit{F1-score} is the harmonic mean of precision and recall, defined as
\begin{equation}
F_k = 2\frac{P_k R_k}{P_k + R_k}.
\end{equation}
We compute the weighted average F1-score across $K$ classes as,
\begin{equation}
\bar{F}_1 = \sum_{k=1}^K \frac{J_k F_k}{J},
\label{eq:wafs}
\end{equation}
where $J$ is the total number of configurations (for either the public test set, the private test set, or their combination). Note that $K=3$ for the task of filling level classification and $K=4$ for the task of filling type classification.
\textbf{T3, T4 and T5.} For container capacity and mass estimation, we compute the relative absolute error between the estimated measure, $a \in \{\tilde{\gamma}^j, \tilde{m}_c^j \}$, and the true measure, $b \in \{\gamma^j, m_c^j \}$:
\begin{equation}
\varepsilon(a, b) = \frac{|a - b |}{b}.
\label{eq:ware}
\end{equation}
For container dimensions estimation, where $a \in \left\{\tilde{w}_t^j,\tilde{w}_b^j,\tilde{h}^j\right\}$ and $b$ is the corresponding annotation, we use the normalisation function $\sigma_1(\cdot,\cdot)$~\cite{Sanchez-Matilla2020}:
\begin{equation}
\sigma_1(a,b)=
\begin{cases}
1 - \frac{|a - b |}{b} & \text{if} \quad | a - b | < b, \\
0 & \text{otherwise}.
\end{cases}
\end{equation}
For filling mass estimation\footnote{Note that an algorithm with lower scores for T1, T2 and T3, may obtain a higher filling mass score than other algorithms due to the multiplicative formula to compute the filling mass for each configuration.}, we compute the relative absolute error between the estimated, $\tilde{m}_{f}^j$, and the true filling mass, $m_{f}^j$, unless the annotated mass is zero (empty filling level),
\begin{equation}
\epsilon(\tilde{m}_f^j, m_f^j) =
\begin{cases}
0, & \text{if } m_f^j = 0 \land \tilde{m}_f^j=0, \\
\tilde{m}_f^j & \text{if } m_f^j = 0 \land \tilde{m}_f^j \neq 0, \\
\frac{|\tilde{m}_f^j - m_f^j |}{m_f^j} & \text{otherwise}.
\end{cases}
\label{eq:ware2}
\end{equation}
With reference to Table~\ref{tab:scores}, we compute the score, $s_i$, with \mbox{$i=\left\{3,\dots,8\right\}$}, across all the configurations $J$ for each measure as:
\begin{equation}
\noindent
s_i =
\begin{cases}
\frac{1}{J}\sum_{j=1}^{J}{ \mathds{1}_j e^{-\varepsilon(a, b)}} & \text{if} \, a \in \left\{\tilde{\gamma}^j, \tilde{m}_c^j\right\},\\
\frac{1}{J}\sum_{j=1}^{J}{\mathds{1}_j \sigma_1(a,b)} & \text{if} \, a \in \left\{\tilde{w}^j,\tilde{w}_b^j,\tilde{h}^j\right\},\\
\frac{1}{J}\sum_{j=1}^{J}{ \mathds{1}_j e^{-\epsilon(a, b)}} & \text{if} \, a=\tilde{m}_f^j.\\
\end{cases}
\end{equation}
The value of the indicator function, \mbox{$\mathds{1}_j \in \{0,1\}$}, is 0 only when \mbox{$a \in \left\{ \tilde{\gamma}^j, \tilde{m}_c^j, \tilde{w}_t^j,\tilde{w}_b^j,\tilde{h}^j, \tilde{m}_f^j \right\}$} is not estimated in configuration $j$. Note that estimated and annotated measures are strictly positive, $a>0$ and $b>0$, except for filling mass in the empty case (i.e., $\tilde{\lambda}^j = 0$ or $\tilde{\tau}^j = 0$).
\begin{table*}[t!]
\centering
\scriptsize
\renewcommand{\arraystretch}{1.2}
\setlength\tabcolsep{1.3pt}
\caption{Results of the CORSMAL challenge entries on the combination of the public and private CCM test sets~\cite{Xompero2021_ArXiv,Xompero_CCM}. For a measure $a$, its corresponding ground-truth value is $\hat{a}$. All scores are normalised and presented in percentages. $\bar{F}_1(\cdot)$ is the weighted average F1-score. Filling amount and type are sets of classes (no unit).
}
\begin{tabular}{ccccclllllccrrrrrrrrr}
\specialrule{1.2pt}{3pt}{0.6pt}
T1 & T2 & T3 & T4 & T5 & Description & Unit & Measure & Score & Weight & Type & R2S & RAN & AVG & \cite{Donaher2021EUSIPCO_ACC} & \cite{Liu2020ICPR} & \cite{Ishikawa2020ICPR} & \cite{Iashin2020ICPR} & \cite{Apicella_GC_ICASSP22} & \cite{Matsubara_GC_ICASSP22} & \cite{Wang_GC_ICASSP22} \\
\specialrule{1.2pt}{3pt}{1pt}
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & Filling level & & $\lambda^j$ & $s_1 = \bar{F}_1(\lambda^1, \ldots, \lambda^J, \hat{\lambda}^1, \ldots, \hat{\lambda}^J)$ & $\pi_1 = 1/8$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 37.62 & 33.15 & \textbf{80.84} & 43.53 & 78.56 & 79.65 & \multicolumn{1}{c}{--} & 65.73 & 77.40 \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & Filling type & & $\tau^j$ & $s_2 = \bar{F}_1(\tau^1, \ldots, \tau^J, \hat{\tau}^1, \ldots, \hat{\tau}^J)$ & $\pi_2 = 1/8$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 24.38 & 23.01 & 94.50 & 41.83 & 96.95 & 94.26 & \multicolumn{1}{c}{--} & 80.72 & \textbf{99.13} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & Capacity & mL & $\gamma^j$ & $s_3 = \frac{1}{J} \sum_{j=1}^{J} \mathds{1}_j e^{-\varepsilon^j(\gamma^j, \hat{\gamma}^j)}$ & $\pi_3 = 1/8$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 24.58 & 40.73 & \multicolumn{1}{c}{--} & 62.57 & 54.79 & 60.57 & \multicolumn{1}{c}{--} & \textbf{72.26} & 59.51 \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & Container mass & g & $m_c^j$ & $s_4 = \frac{1}{J} \sum_{j=1}^{J} \mathds{1}_j e^{-\varepsilon^j(m_c^j, \hat{m}_c^j)}$ & $\pi_4 = 1/8$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 29.42 & 22.06 & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & 49.64 & 40.19 & \textbf{58.78} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Width at top & mm & $w_t^j$ & $s_5 = \frac{1}{J}\sum_{j=1}^{J}{\mathds{1}_j \sigma_1(w_t^j, \hat{w_t}^j)}$ & $\pi_5 = 1/24$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 32.33 & 76.89 & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & 69.09 & \textbf{80.01} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Width at bottom & mm & $w_b^j$ & $s_6 = \frac{1}{J}\sum_{j=1}^{J}{\mathds{1}_j \sigma_1(w_b^j, \hat{w_b}^j)}$ & $\pi_6 = 1/24$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 25.36 & 58.19 & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & 59.74 & \textbf{76.09} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Height & mm & $h^j$ & $s_7 = \frac{1}{J}\sum_{j=1}^{J}{ \mathds{1}_j \sigma_1(h^j, \hat{h}^j)}$ & $\pi_7 = 1/24$ & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 42.48 & 64.32 & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & \multicolumn{1}{c}{--} & 70.07 & \textbf{74.33} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & Filling mass & g & $m_f^j$ & $s_8 = \frac{1}{J} \sum_{j=1}^{J} \mathds{1}_j e^{-\epsilon^j(m_f^j, \hat{m}_f^j)}$ & $\pi_8 = 1/8$* & I & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 35.06 & 42.31 & 25.07 & 53.47 & 62.16 & 65.06 & \multicolumn{1}{c}{--} & \textbf{70.50} & 65.25 \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Object mass & g & $m^j$ & $s_9 = \frac{1}{J}\sum_{j=1}^{J}{\mathds{1}_j \psi^j(m^j, \hat{F}^j)}$ & $\pi_9 = 1/8$* & I & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & 56.31 & 58.30 & 55.22 & 64.13 & 66.84 & 65.04 & 53.54 & 60.41 & \textbf{71.19} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Pose at delivery & (mm, $^\circ$) & ($\alpha^j$,$\beta^j$) & $s_{10} = \frac{1}{J}\sum_{j=1}^{J}{\Delta_j(\alpha^j,\beta^j,\eta,\phi)}$ & $\pi_{10} = 1/8$* & I & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & 72.11 & 70.01 & 73.94 & 78.76 & 72.91 & \textbf{80.40} & 60.54 & 73.17 & 79.32 \\
\midrule
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \multicolumn{3}{l}{Joint filling type and level} & $s_{11} = \bar{F}_1(\lambda^1, \tau^1, \ldots, \hat{\lambda}^1, \hat{\tau}^1, \ldots)$ & \multicolumn{1}{c}{--} & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 10.49 & 8.88 & 77.15 & 24.32 & 77.81 & 76.45 & \multicolumn{1}{c}{--} & 59.32 & \textbf{78.16} \\
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \multicolumn{3}{l}{Container capacity and dimensions} & $s_{12} = {s_3}/{2} + (s4 + s5 + s6)/{6}$ & \multicolumn{1}{c}{--} & D & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=white] (1,1) circle (0.5ex);} & 28.99 & 53.60 & \multicolumn{1}{c}{--} & 31.28 & 27.39 & 30.28 & \multicolumn{1}{c}{--} & \textbf{69.28} & 68.16 \\
\specialrule{1.2pt}{3pt}{1pt}
\protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & \protect\raisebox{1pt}{\protect\tikz \protect\draw[black,fill=black] (1,1) circle (0.5ex);} & Overall score & & & $S = \sum_{l=1}^{10} \pi_l s_l$ & \multicolumn{1}{c}{--} & I & \multicolumn{1}{c}{--} & 39.11 & 44.51 & 31.52 & 35.89 & 47.04 & 48.35 & 9.05 & 66.16 & \textbf{73.43} \\
\specialrule{1.2pt}{3pt}{1pt}
\multicolumn{21}{l}{\scriptsize{Best performing results for each row highlighted in bold. Results of tasks not addressed shown with a hyphen (--).}}\\
\multicolumn{21}{l}{\scriptsize{For $s_9$ and $s_{10}$, configurations with failures in grasping and/or delivering the containers in simulation using true physical properties as input are annotated and discarded.}}\\
\multicolumn{21}{l}{\scriptsize{For fairness, the residual between 100 and the scores obtained with true measures of the physical properties are added to $s_9$ and $s_{10}$ to remove the impact of the simulator.}}\\
\multicolumn{21}{l}{\scriptsize{KEY -- T:~task, D:~direct score, I:~indirect score, R2S:~measured in the real-to-simulation framework, RAN:~random estimation, AVG:~average from the training set.}}\\
\multicolumn{21}{l}{\scriptsize{* weighted by the number of performed tasks.}}\\
\end{tabular}
\label{tab:scores}
\vspace{-10pt}
\end{table*}
\textbf{Object safety and accuracy of delivery.}
Object safety is the probability that the force applied by the robot, $\tilde{F}$, enables a gripper to hold the container without dropping it or breaking it~\cite{Pang2021ROMAN}. We approximate the force required to hold the container as
\begin{equation}
\hat{F} \approx \frac{\hat{m} (g + a_{max})}{\mu},
\label{equ:grasp_force_theoretical}
\end{equation}
where $\hat{m}$ is the annotated object mass; $g=9.81$~m/s$^2$ is the gravitational earth acceleration; $a_{max}$ is the maximum acceleration of the robot arm when carrying the object; and $\mu$ is the coefficient of friction between the container and the gripper ($\mu=1.0$~\cite{Pang2021ROMAN}).
The value of the force applied by the robot to grasp the object is calculated with Eq.~\ref{equ:grasp_force_theoretical} using the predicted object mass $\tilde{m}$. We compute object safety as an exponential function that accounts for the difference between the applied normal force $\tilde{F}^j$ (measured in simulation) and the required normal force, $\hat{F}^j$:
\begin{equation}
\psi^j = e^{\frac{| \tilde{F}^j - \hat{F}^j |}{\hat{F}^j} \ln{(1-c)}} = {\ln{(1-c)}}^{\frac{| \tilde{F}^j - \hat{F}^j |}{\hat{F}^j}},
\label{eq:forcesafety}
\end{equation}
where the normal force is the component of the contact force perpendicular to the contact surface and $c$ controls the sensitivity of $\psi^j$~\cite{Pang2021ROMAN}. A negative difference represents a higher probability of dropping the container, and a positive difference represents a higher probability of breaking the container.
We quantify the accuracy of delivering the container upright and within the target area as
\begin{equation}
\Delta_j =
\begin{cases}
1 - \frac{\alpha}{\eta} & \text{if } (\alpha < \eta) \text{ and } (\beta < \phi), \\
0 & \text{otherwise}, \\
\end{cases}
\end{equation}
where $\alpha$ is the distance from the centre of the base of the container to the target location $\boldsymbol{d}$; $\eta$ is the maximum distance allowed from the delivery target location; $\beta$ is the angle between the vertical axis of the container and the vertical axis of the world coordinate system; and $\phi$ is the value of $\beta$ at which the container would tip over.
We compute the score for object safety, $s_9$, as
\begin{equation}
s_{9} = \frac{1}{J}\sum_{j=1}^{J}{ \mathds{1}_j \psi^j(m^j, \hat{F}^j)},
\label{eq:totobjsafetyscore}
\end{equation}
where the value of the indicator function, $\mathds{1}_j$, is 0 only when either the filling mass or the containers mass is not estimated for each configuration $j$;
and the score for the delivery accuracy, $s_{10}$, as
\begin{equation}
s_{10} = \frac{1}{J}\sum_{j=1}^{J}{ \Delta_j(\alpha^j,\beta^j,\eta,\phi)}.
\label{eq:deliveryscore}
\end{equation}
The scores $s_9$ and $s_{10}$ are partially influenced by the simulator conditions (e.g, friction, contact, robot control), but we aimed at making the simulated handover reproducible across different algorithms through the annotated object trajectory, starting handover frame, and reconstructed 3D model.
{\bf Group tasks and overall score.} For joint filling type and level classification ($s_{11}$), estimations and annotations of both filling type and level are combined in $K=7$ feasible classes, and $\bar{F}_1$ is recomputed based on these classes. For joint container capacity and dimensions estimation, we compute the following weighted average:
\begin{equation}
s_{12} = \frac{s_3}{2} + \frac{s4 + s5 + s6}{6}.
\end{equation}
Finally, the overall score is computed as the weighted average of the scores from $s_1$ to $s_{10}$. Note that $s_8$, $s_9$, and $s_{10}$ may use random estimations for either of the tasks not addressed by an algorithm.
\subsection{IEEE ICASSP 2022 Challenge entries}
Nine teams registered for the IEEE ICASSP 2022 challenge; three algorithms were submitted for container mass estimation (T4), two algorithms were submitted for classifying the filling level (T1) and type (T2), and two other algorithms were submitted for estimating the container properties (T3, T4, T5) by three teams. We refer to the submissions of the three teams as A1~\cite{Apicella_GC_ICASSP22}, A2~\cite{Matsubara_GC_ICASSP22}, and A3~\cite{Wang_GC_ICASSP22}.
A1 solved only the task of container mass estimation (T4) using RGB-D data from the fixed frontal view and by regressing the mass with a shallow Convolutional Neural Network (CNN)~\cite{Christmann2020NTNU}. To increase the accuracy, A1 extracted a set of patches of the detected container from automatically selected frames in a video, and averaged their predicted masses.
To classify the filling level (T1) and type (T2), A2 used Vision Transformers~\cite{Dosovitskiy2021ICLR}, whereas A3 used pre-trained CNNs (e.g., Mobilenets~\cite{Howard2017Arxiv}) combined with Long Short-Term Memory units or majority voting~\cite{Hochreiter1997LSTM}. Only audio or audio with visual information (RGB) from the fixed, frontal view is preferred as input. To estimate the container properties (T3, T4, T5), A2 used RGB data from the three fixed views, and A3 used RGB-D data from the fixed frontal view.
A2 used a modified multi-view geometric approach that iteratively fits a hypothetical 3D model~\cite{Xompero2020ICASSP_LoDE}.
A3 fine-tunes multiple Mobilenets via transfer learning from the task of dimensions estimation (T5) to the tasks of container capacity (T3) and mass (T4) estimation~\cite{Wang_GC_ICASSP22}.
These Mobilenets regress the properties using patches extracted from automatically selected frames where the container is mostly visible~\cite{Wang_GC_ICASSP22}. To overcome over-fitting of the limited training data and improve generalisation on novel containers, these Mobilenets are fine-tuned with geometric-based augmentations and variance evaluation~\cite{Wang_GC_ICASSP22}. Overall, A3 is designed to process a continuous stream (online), thus being more suitable for human-to-robot handovers.
Table~\ref{tab:scores} shows the scores of the submissions on the combined CCM test sets. As reference, we provide the results for the properties estimated by a pseudo-random generator (RAN), by using the average (AVG) of the training set for container capacity, mass, and dimensions; or by the algorithms of four earlier entries to the challenge~\cite{Donaher2021EUSIPCO_ACC,Liu2020ICPR,Ishikawa2020ICPR,Iashin2020ICPR}. A3 achieves the highest $\bar{F}_1$ for filling type classification ($s_2 = 99.13$), and joint filling type and level classification ($s_{11} = 78.16$). A3 is also the most accurate in estimating the container mass ($s_4 = 58.78$), followed by A1 ($s_4=49.64$), and the container dimensions.
A2 is the most accurate in estimating the capacity ($s_3 = 72.26$). A2 is also the most accurate for filling mass ($s_8 = 70.50$). A3 has a high accuracy for filling level and type classification, but is affected by its lower accuracy
for capacity estimation.
Among the entries of the challenge at IEEE ICASSP 2022, A3 achieves the best score for object safety ($s_9 = 71.19$) and delivery accuracy ($s_{10} = 79.32$). In conclusion, A3 reaches the highest overall score ($S = 73.43$), followed by A2 ($S = 66.16$).
\section{Conclusion}
\label{sec:conclusion}
Recent, fast advances in machine learning and artificial intelligence have created an expectation on the ability of robots to seamlessly operate in the real world by accurately and robustly perceiving and understanding dynamic environments, including the actions and intentions of humans.
However, several challenges in audio-visual perception and modelling humans with their hard-to-predict behaviours hamper the deployment of robots in real-world scenarios.
We presented the tasks, the real-to-simulation framework, the scores and the entries to the CORSMAL challenge at IEEE ICASSP 2022 .
These new entries complement the algorithms previously submitted to the challenge~\cite{Donaher2021EUSIPCO_ACC,Liu2020ICPR,Ishikawa2020ICPR,Iashin2020ICPR}.
\bibliographystyle{IEEEbib}
| 28aefb059d9f6613e5ff09536977455486e365d4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The problem of developing control policies that guarantee stability and safety of nonlinear control systems has received significant attention in recent years. In particular, the unification of Control Lyapunov Functions (CLFs) \cite{SontagSCL89,AmesTAC14} and Control Barrier Functions (CBFs) \cite{AmesTAC17,AmesECC19} has provided a pathway towards safe and stable control of complex nonlinear systems such as autonomous vehicles \cite{AmesCDC14,WeiAutomatica21}, multi-agent systems \cite{AmesTRO17}, and bipedal robots \cite{AmesCDC16}. Although powerful, the guarantees afforded by these approaches are model-based, hence the success in transferring such guarantees to real-world systems is inherently tied to the fidelity of the underlying system model. Inevitably, such models are only an approximation of the true system due to parametric uncertainties and unmodeled dynamics, thus there is strong motivation to study the synthesis of CLF and CBF-based controllers in the presence of model uncertainty. Although robust approaches \cite{JankovicAutomatica18,AmesIEEEA20} have demonstrated success in this regard, in general, such techniques can be highly conservative. On the other hand, data-driven approaches have demonstrated the ability to reduce uncertainty and yield high-performance controllers in terms of both safety and stability. Popular data-driven approaches {\color{black} for} reducing uncertainty include work based on episodic learning \cite{TaylorL4DC20,SreenathADHS21} or by modeling the uncertainty using Gaussian processes (GPs) \cite{SreenathCDC21,DhimanTAC21}. However, providing strong guarantees in an episodic learning setting is challenging and, although {\color{black} GP-based approaches account for very general classes of uncertainties, GPs can be computationally intensive and the generality offered by GPs generally results in probabilistic, rather than deterministic, guarantees on stability and safety.}
As adaptive control \cite{Krstic} has a long history of success in controlling nonlinear systems with parametric uncertainty, there is also a rich line of work that unites CLFs and CBFs with techniques from adaptive control. The authors of \cite{TaylorACC20} extend the adaptive CLF (aCLF) paradigm \cite[Ch. 4.1]{Krstic} to CBFs, yielding the first instance of an \emph{adaptive} CBF\footnote{The term aCBF was also used in \cite{WeiTAC21-adaptive-cbf} to refer to a class of CBFs that account for time-varying control bounds.} (aCBF) that allows for the safe control of uncertain nonlinear systems with parametric uncertainty. The authors of \cite{LopezLCSS21,DixonACC21,SanfeliceACC21} extend the aCBF techniques from \cite{TaylorACC20} using set-membership identification, concurrent learning (CL) \cite{Chowdhary,DixonIJACSP19}, and hybrid techniques, respectively, which were shown to reduce the conservatism of original aCBF formulation. Nevertheless, all of the aforementioned aCBF techniques are limited to CBFs with relative degree one. In practice, however, many safety-critical constraints have relative degrees larger than one (e.g., constraints on the configuration of a mechanical system generally have at least relative degree two). The unification of CLFs/CBFs with techniques from CL adaptive control was also presented in \cite{AzimiCDC18,Azimi}; however, the resulting CLF controllers either only guarantee uniformly ultimately bounded stability or are limited to single-input feedback linearizable systems. Importantly, the CL-based aCBF controllers from \cite{AzimiCDC18,Azimi} do not provide strong safety guarantees since the CBF-based control inputs are generated using the estimated dynamics without accounting for estimation errors, leading to potential safety violations that can be understood through the notion of input-to-state-safety \cite{AmesLCSS19}.
To address high relative degree safety constraints for systems with \emph{known} dynamics, the authors of \cite{SreenathACC16,WeiCDC19,WeiTAC21-hocbf,DimosTAC21-hocbf} introduce exponential and high order CBFs (HOCBFs), which provide a systematic framework to construct CBFs that account for high relative degree constraints. Importantly, as noted in \cite{DimosTAC21-hocbf}, HOCBFs can also be used to simplify the search for valid CBFs since the dependence on higher order dynamics need not be directly encoded through the definition of the CBF itself (as in the relative degree one case). Rather, {\color{black} the} dependence {\color{black} on} higher order dynamics is \emph{implicitly} encoded through conditions on higher order derivatives of the CBF candidate. Despite the advancements of both aCBFs and HOCBFs, to our knowledge, the intersection of these two techniques has yet to be explored in the literature.
In this paper we unite aCBFs and HOCBFs to develop control policies satisfying high relative degree safety constraints for nonlinear systems with parametric uncertainty. Similar to \cite{DixonACC21}, our approach leverages the concurrent learning technique presented in \cite{DixonIJACSP19}, which identifies uncertain parameters of the nonlinear system online by exploiting sufficiently rich data collected along the system trajectory. A key insight enabling our high relative degree approach is that if the relative degrees of the CBF with respect to the control and uncertain parameters are the same (in a sense to be clarified later in this paper), then the sufficient conditions for safety can be encoded through affine constraints on the control input, allowing control synthesis to be performed in a computationally efficient quadratic programming framework. Furthermore, unlike existing aCBF formulations for relative degree one constraints \cite{TaylorACC20,LopezLCSS21,DixonACC21,SanfeliceACC21}, we show that our High Order Robust Adaptive Control Barrier Functions (HO-RaCBFs) inherit the robustness properties of zeroing CBFs \cite{AmesADHS15} in the sense that our developed aCBFs not only render the safe set forward invariant, but also \emph{asymptotically stable} when solutions begin outside the safe set. We then introduce a novel class of aCLF, termed exponentially stabilizing aCLFs (ES-aCLFs), that extends the CL paradigm from \cite{DixonIJACSP19} to a CLF setting by exploiting the same history stack used to reduce conservatism of the safety controller to endow a nominal CLF-based control policy with \emph{exponential} stability guarantees. The efficacy of the combined HO-RaCBF/ES-aCLF controller is demonstrated through simulations of a robotic navigation task and an inverted pendulum, both of which involve high relative degree safety constraints that cannot be addressed by existing aCBF approaches.
\section{Mathematical Preliminaries}
Consider a nonlinear control affine system of the form
\begin{equation}\label{eq:dyn}
\dot{x}=f(x) + g(x)u,
\end{equation}
with state $x\in\mathbb{R}^n$ and control $u\in\mathcal{U}\subseteq\mathbb{R}^m$, where $f\,:\,\mathbb{R}^n\rightarrow\mathbb{R}^n$ and $g\,:\,\mathbb{R}^n\rightarrow\mathbb{R}^{n\times m}$ are locally Lipschitz vector fields modeling the drift and control directions, respectively. Given a feedback law $u=k(x,t)$, locally Lipschitz in $x$ and piecewise continuous in $t$, the closed-loop vector field $f_{\text{cl}}(x,t)\coloneqq f(x)+g(x)k(x,t)$ is also locally Lipschitz in $x$ and piecewise continuous in $t$, implying that \eqref{eq:dyn} admits a unique solution $x\,:\,\mathcal{I}\rightarrow\mathbb{R}^n$ starting from $x(0)\in\mathbb{R}^n$ on some maximal interval of existence $\mathcal{I}\subset\mathbb{R}_{\geq0}$. A closed set $\mathcal{C}\subset\mathbb{R}^n$ is said to be \emph{forward invariant} for the closed-loop system $\dot{x}=f_{\text{cl}}(x,t)$ if $x(0)\in\mathcal{C}\implies x(t)\in\mathcal{C}$ for all $t\in\mathcal{I}$. In this paper (and in the related literature) forward invariance is used to formalize the abstract notion of safety. Hence, if a given ``safe" set $\mathcal{C}$ is forward invariant for $\dot{x}=f_{\text{cl}}(x,t)$, then we say the closed-loop system is \emph{safe} with respect to $\mathcal{C}$. It will be assumed throughout this paper that any safe set $\mathcal{C}$ can be expressed as the zero-superlevel set of a continuously differentiable function $h\,:\,\mathbb{R}^n\rightarrow\mathbb{R}$ as
\begin{equation}\label{eq:C}
\mathcal{C}=\{x\in\mathbb{R}^n\,|\,h(x)\geq0\}.
\end{equation}
A popular tool for developing controllers that render \eqref{eq:C} forward invariant for \eqref{eq:dyn} is the concept of a CBF \cite{AmesTAC17,AmesECC19}, which places Lyapunov-like conditions on the derivative of $h$ to guarantee safety. A limitation of traditional CBFs from \cite{AmesTAC17}, however, is that their effectiveness is conditioned upon the assumption that the function $h$ has relative degree one. Yet, many relevant systems and safe sets fail to satisfy such a condition, which has motivated the introduction of exponential and higher order CBFs \cite{SreenathACC16,WeiCDC19,WeiTAC21-hocbf,DimosTAC21-hocbf} to account for safety constraints with \emph{high relative degree}. Before proceeding, we recall that the \emph{Lie derivative} of a differentiable function $h\,:\,\mathbb{R}^n\rightarrow\mathbb{R}$ along a vector field $f\,:\,\mathbb{R}^n\rightarrow\mathbb{R}^n$ is defined as $L_fh(x)\coloneqq \tfrac{\partial h}{\partial x}f(x)$. This notation allows us to denote higher order Lie derivatives along an additional vector field $g$ as $L_gL_f^{i-1}h(x)=\tfrac{\partial(L_f^{i-1}h)}{\partial x}g(x)$ (see e.g. \cite[Ch. 13.2]{Khalil}).
\begin{definition}[\cite{Khalil}]
A sufficiently smooth function $h\,:\,\mathbb{R}^n\rightarrow\mathbb{R}$ is said to have \emph{relative degree} $r\in\mathbb{N}$ with respect to \eqref{eq:dyn} on a set $\mathcal{R}\subset\mathbb{R}^n$ if 1) for all $1\leq i\leq r-1$, $L_gL_f^{i-1}h(x)\equiv0$; 2) $L_gL_f^{r-1}h(x)\neq0$ for all $x\in\mathcal{R}$.
\end{definition}
To account for high relative degree safety constraints, the authors of \cite{WeiCDC19,WeiTAC21-hocbf,DimosTAC21-hocbf} introduce the notion of a HOCBF. Before stating the definition, we recall from \cite{AmesTAC17} that a continuous function $\alpha\,:\,(-b,a)\rightarrow(-\infty,\infty)$, for some $a,b\in\mathbb{R}_{>0}$, is said to be an \emph{extended class} $\mathcal{K}$ \emph{function} if it is strictly increasing and $\alpha(0)=0$.
\begin{definition}[\cite{WeiCDC19,WeiTAC21-hocbf,DimosTAC21-hocbf}]\label{def:HOCBF}
Consider system \eqref{eq:dyn} and a set $\mathcal{C}\subset\mathbb{R}^n$ as in \eqref{eq:C}. Let $\{\mathcal{C}_i\}_{i=1}^r$ be a collection of sets of the form $\mathcal{C}_i\coloneqq\{x\in\mathbb{R}^n\,|\,\psi_{i-1}(x)\geq0\}$, where $\psi_{0}(x)\coloneqq h(x)$ and
\begin{equation}\label{eq:psi}
\begin{aligned}
\psi_{i}(x)\coloneqq & \dot{\psi}_{i-1}(x) + \alpha_{i}(\psi_{i-1}(x)),\; i\in\{1,\dots,r-1\},\\
\psi_{r}(x,u)\coloneqq & \dot{\psi}_{r-1}(x,u) + \alpha_r(\psi_{r-1}(x)),
\end{aligned}
\end{equation}
where $\{\alpha_{i}\}_{i=1}^r$ is a collection of differentiable extended class $\mathcal{K}$ functions. Then, the function $h$ is said to be a HOCBF of order $r$ for \eqref{eq:dyn} on an open set $\mathcal{D}\supset\cap_{i=1}^r\mathcal{C}_i$ if $h$ has relative degree $r$ on some nonempty $\mathcal{R}\subseteq\mathcal{D}$ and there exists a suitable choice of $\{\alpha_{i}\}_{i=1}^r$ such that for all $x\in\mathcal{D}$
\[
\sup_{u\in\mathcal{U}}\{\underbrace{L_{f}\psi_{r-1}(x) + L_g\psi_{r-1}(x)u + \alpha_r(\psi_{r-1}(x))}_{\psi_r(x,u)}\} \geq0.
\]
\end{definition}
If $\mathcal{U}=\mathbb{R}^m$, the above states that $h$ is a HOCBF if $\|L_g\psi_{r-1}h(x)\|=0$ $\implies$ $L_f\psi_{r-1}h(x) \geq - \alpha_r(\psi_{r-1}(x))$, implying that $h$ need not have \emph{uniform} relative degree on $\mathcal{D}$ as illustrated in \cite{DimosTAC21-hocbf}, provided the unforced dynamics satisfy the above condition at points where $\|L_g\psi_{r-1}h(x)\|=0$. The following result provides higher order conditions for safety.
\begin{theorem}[\cite{DimosTAC21-hocbf}]\label{theorem:HOCBF}
Let $h$ be a HOCBF for \eqref{eq:dyn} on $\mathcal{D}\subset\mathbb{R}^n$ as in Def. \ref{def:HOCBF}. Then, any locally Lipschitz controller $u=k(x)\in K_{\text{cbf}}(x)$, where $K_{\text{cbf}}(x)\coloneqq \{u\in\mathcal{U}\,|\,\psi_r(x,u)\geq0\}$, renders $\cap_{i=1}^r\mathcal{C}_i^r$ forward invariant for the closed-loop system.
\end{theorem}
\section{High Order Robust Adaptive Control Barrier Functions}
This section introduces the concept of a High Order Robust Adaptive Control Barrier Function (HO-RaCBF), which provides a tool to synthesize controllers that guarantee the satisfaction of high relative degree safety constraints for nonlinear systems with parametric uncertainty.
To this end, we now turn our attention to systems of the form
\begin{equation}\label{eq:dyn2}
\dot{x}=f(x) + Y(x)\theta + g(x)u,
\end{equation}
where $f$ and $g$ are known locally Lipschitz vector fields as in \eqref{eq:dyn}, $Y\,:\,\mathbb{R}^n\rightarrow\mathbb{R}^{n\times p}$ is a known locally Lipschitz regression matrix, and $\theta\in\mathbb{R}^p$ is a constant vector of uncertain parameters. We assume that $f(0)=0$ and $Y(0)=0$ so that 0 is an equilibrium point of the unforced system. Our main objective is to synthesize controllers for \eqref{eq:dyn2} that guarantee safety under the presumption that the function $h$ defining the safe set $\mathcal{C}$ as in \eqref{eq:C} has a high relative degree with respect to \eqref{eq:dyn2}. To facilitate our approach, we make the following assumption on the structure of the uncertainty in \eqref{eq:dyn2}.
\begin{assumption}\label{assumption:rd}
Consider a set $\mathcal{C}$ as in \eqref{eq:C} and an open set $\mathcal{D}$ as in Def. \ref{def:HOCBF}. If $h$ has relative degree $r$ on $\mathcal{R}\subseteq\mathcal{D}$ with respect to \eqref{eq:dyn2} (i.e., if there exist some nonempty $\mathcal{R}\subseteq\mathcal{D}$ such that $L_gL_f^{i-1}h(x)\equiv 0$ for all $1\leq i\leq r-1$ and $L_gL_f^{r-1}h(x)\neq0$ for all $x\in\mathcal{R}$), then there exists $\mathcal{R}'\subseteq\mathcal{D}$ such that $L_YL_f^{i-1}h(x)\equiv 0$ for all $1\leq i\leq r-1$ and $L_YL_f^{r-1}h(x)\neq0$ for all $x\in\mathcal{R}'$.
\end{assumption}
\begin{remark}
The above assumption requires that the uncertainty in \eqref{eq:dyn2} does not appear before the control when taking higher order derivatives of $h$. Although this may seem restrictive, a variety of physical systems satisfy Assumption \ref{assumption:rd}. Examples include Lagrangian mechanical systems, where $h$ is a function of only the system's configuration. If the above assumption is not made, then the uncertain parameters $\theta$ will appear alongside the control input $u$ in higher order terms, complicating the formulation of the affine constraints on $u$ developed in this paper. {\color{black} From an adaptive control perspective, Assumption \ref{assumption:rd} is similar to the assumption that the uncertain parameters satisfy the matching condition.}
\end{remark}
According to Theorem \ref{theorem:HOCBF}, the forward invariance of $\mathcal{C}$ can be enforced by ensuring that the control input is selected such that the HOCBF condition from Def. \ref{def:HOCBF} is satisfied for all $x\in\cap_{i=1}^{r}\mathcal{C}_i$; however, the presence of model uncertainty in \eqref{eq:dyn2} makes it impossible to directly enforce such a condition. To address this challenge, we aim to take a data-driven approach and update the estimates of the uncertain parameters online using techniques from adaptive control \cite{Krstic,Chowdhary,DixonIJACSP19,DixonACC21} while guaranteeing safety at all times. Following the approach from \cite{DixonIJACSP19}, observe that integrating \eqref{eq:dyn2} over a time interval $[t-\Delta T, t]\subset\mathbb{R}$ using the Fundamental Theorem of Calculus allows \eqref{eq:dyn2} to be equivalently represented as
\begin{equation}\label{eq:intdyn}
\Delta x(t)=x(t)-x(t-\Delta T)=\mathcal{F}(t) + \mathcal{Y}(t)\theta + \mathcal{G}(t),
\end{equation}
where $\mathcal{F}(t)\coloneqq \int_{t-\Delta T}^{t}f(x(\tau))d\tau$, $\mathcal{Y}(t)\coloneqq \int_{t-\Delta T}^{t}Y(x(\tau))d\tau$, $\mathcal{G}(t)\coloneqq \int_{t-\Delta T}^{t}g(x(\tau))u(\tau)d\tau$. Now let $\mathcal{H}\coloneqq\{(t_j,x_j,x_{j}^{-},u_j) \}_{j=1}^M$ be a history stack of $M\in\mathbb{N}$ instances of input-output data, where $t_j\in[\Delta T, t]$ denotes a sampling time, $x_j\coloneqq x(t_j)$, $x_j^{-}\coloneqq x(t_j-\Delta t)$, $u_j\coloneqq u(t_j)$, and define\footnote{The function $\Lambda$ is implicitly a function of time as data is added/removed from the history stack $\mathcal{H}$ along the system trajectory.}
\begin{equation}\label{eq:Lambda}
\Lambda(t)\coloneqq\sum_{j=1}^{M}\mathcal{Y}_j^\top\mathcal{Y}_j,\quad \lambda(t)\coloneqq \lambda_{\min}(\Lambda(t)),
\end{equation}
where $\lambda_{\min}(\Lambda)$ denotes the minimum eigenvalue of $\Lambda$. As noted in \cite{DixonIJACSP19,DixonACC21}, the function $\lambda$ is piecewise constant between sampling times and nonnegative since $\Lambda(t)$ is at least positive semidefinite at all times. The following assumption will be used to ensure safety for all possible realizations of the uncertain parameters.
{\color{black}
\begin{assumption}\label{assumption:estimation_error}
The uncertain parameters $\theta$ belong to a known convex polytope $\Theta\subset\mathbb{R}^p$.
\end{assumption}
The above assumption implies that for any given parameter estimate $\hat{\theta}\in\Theta$ there exists some maximum possible estimation error $\tilde{\vartheta}\in\mathbb{R}^p$ in the sense that $\|\theta-\hat{\theta}\|\leq\|\tilde{\vartheta}\|$ for all $\hat{\theta}\in\Theta$. Given that $\Theta$ is a convex polytope, each component of $\tilde{\vartheta}$ can be computed as
\begin{equation*}
\tilde{\vartheta}_i=\max\big\{\big\vert\min_{\theta,\hat{\theta}\in\Theta}\theta_i - \hat{\theta}_i\big\vert,\, \big\vert\max_{\theta,\hat{\theta}\in\Theta}\theta_i - \hat{\theta}_i\big\vert\big\},
\end{equation*}
where $|\cdot|$ denotes aboslute value and $\theta_i$ denotes the $i$th component of $\theta$, which requires solving a pair of linear programs for each parameter.} The following lemma, adapted from \cite{DixonACC21}, provides a verifiable bound on the parameter estimation error.
\begin{lemma}[\cite{DixonACC21}]\label{lemma:theta_bound}
Consider system \eqref{eq:dyn2} and suppose the estimated parameters are updated according to
\begin{equation}\label{eq:theta_dot_cbf}
\dot{\hat{\theta}}=\gamma\sum_{j= 1}^M\mathcal{Y}_j^\top (\Delta x_j - \mathcal{F}_j - \mathcal{Y}_j\hat{\theta} - \mathcal{G}_j),
\end{equation}
where $\Delta x_j\coloneqq x_j-x_j^-$, $\mathcal{Y}_j\coloneqq \mathcal{Y}(t_j)$, $\mathcal{F}_j\coloneqq \mathcal{F}(t_j)$, $\mathcal{G}_j\coloneqq \mathcal{G}(t_j)$, and $\gamma\in\mathbb{R}_{>0}$ is an adaptation gain. Provided Assumption \ref{assumption:estimation_error} holds and $\hat{\theta}(0)\in\Theta$, then the parameter estimation error $\tilde{\theta}=\theta-\hat{\theta}$ is bounded for all $t\in\mathcal{I}$ as
\begin{equation}\label{eq:theta_bound}
\|\tilde{\theta}(t)\|\leq\nu(t)\coloneqq \|\tilde{\vartheta}\|e^{-\gamma\int_{0}^t\lambda(\tau)d\tau}.
\end{equation}
\end{lemma}
The above lemma implies that, under the update law in \eqref{eq:theta_dot_cbf}, the parameter estimation error is always bounded by a known value provided the initial parameter estimates are selected such that $\hat{\theta}(0)\in\Theta$. Moreover, if there exists some time $T$ such that $\lambda(t)>0$ for all $t>T$, then the bound in \eqref{eq:theta_bound} implies all estimated parameters exponentially converge to their true values\footnote{See \cite[Ch. 3]{Chowdhary} for a discussion on the relation between traditional persistence of excitation conditions and the milder \emph{finite} excitation conditions leveraged in concurrent learning adaptive control required to achieve $\lambda>0$. We also refer the reader to \cite{ChowdharyACC11} for various algorithms that record data points in $\mathcal{H}$ so as to ensure $\lambda(t)$ is always nondecreasing.\label{footnote:data}}. We now have the {\color{black}necessary} tools in place to introduce a new class of aCBF that allows for the consideration of high relative degree safety constraints.
\begin{definition}\label{def:HO-RaCBF}
Consider system \eqref{eq:dyn2} and a safe set $\mathcal{C}\subset\mathbb{R}^n$ as in \eqref{eq:C}. Consider a collection of sets $\{\mathcal{C}_i\}_{i=1}^r$ of the form $\mathcal{C}_i\coloneqq\{x\in\mathbb{R}^n\,|\,\psi_{i-1}(x)\geq0\}$, where $\psi_{0}(x)\coloneqq h(x)$ and $\{\psi_{i}\}_{i=1}^r$ are defined as in \eqref{eq:psi}. The sufficiently smooth function $h$ is said to be a \emph{high order robust adaptive control barrier function} (HO-RaCBF) of order $r$ for \eqref{eq:dyn2} on an open set $\mathcal{D}\supset\cap_{i=1}^r\mathcal{C}_i$ if $h$ has relative degree $r$ on some nonempty $\mathcal{R}\subseteq\mathcal{D}$ and there exists a suitable choice of $\{\alpha_{i}\}_{i=1}^r$ as in \eqref{eq:psi} such that for all $x\in\mathcal{D}$, $\theta\in\Theta$, and $t\in\mathcal{I}$
\begin{equation}\label{eq:HO-RaCBF}
\begin{aligned}
\sup_{u\in\mathcal{U}}\left\{L_{f}\psi_{r-1}(x) + L_Y\psi_{r-1}(x)\theta + L_g\psi_{r-1}(x)u \right\} \\
\geq -\alpha_r(\psi_{r-1}(x)) + \|L_Y\psi_{r-1}(x)\|\nu(t),
\end{aligned}
\end{equation}
where $\nu$ is defined as in \eqref{eq:theta_bound}.
\end{definition}
Intuitively, the above condition adds a buffer to the original HOCBF condition from Def. \ref{def:HOCBF} to account for all possible realizations of the uncertain parameters given the set $\Theta$. This buffer may shrink over time as the uncertain parameters are identified and exponentially converges to zero in the limit as $t\rightarrow\infty$ provided there exists a time $T$ for which $\lambda(t)>0$ for all $t\geq T$. Furthermore, Def. \ref{def:HO-RaCBF} allows us to consider the set of all control values satisfying \eqref{eq:HO-RaCBF} as
\begin{equation}\label{eq:Kcbf}
\begin{aligned}
\hat{K}_{\text{cbf}}(x,\theta,t)\coloneqq & \{u\in\mathcal{U}\,|\, L_{f}\psi_{r-1}(x) + L_Y\psi_{r-1}(x)\theta \\ & + L_g\psi_{r-1}(x)u +\alpha_r(\psi_{r-1}(x)) \\ &- \|L_Y\psi_{r-1}(x)\|\nu(t) \geq0\}.
\end{aligned}
\end{equation}
The following theorem shows that any well-posed control law $u=k(x,\hat{\theta},t)$ satisfying $k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ renders $\cap_{i=1}^r\mathcal{C}_i$ forward invariant for \eqref{eq:dyn2}.
\begin{theorem}\label{theorem:HO-RaCBF}
Consider system \eqref{eq:dyn2}, a set $\mathcal{C}$ defined by a sufficiently smooth function $h$ as in \eqref{eq:C}, and let $h$ be a HO-RaCBF on $\mathcal{D}$. Provided Assumptions \ref{assumption:rd}-\ref{assumption:estimation_error} hold and the estimated parameters are updated according to \eqref{eq:theta_dot_cbf}, then any controller $u=k(x,\hat{\theta},t)$ locally Lipschitz in $(x,\hat{\theta})$ and piecewise continuous in $t$ satisfying $k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ renders $\cap_{i=1}^r\mathcal{C}_i$ forward invariant for \eqref{eq:dyn2}.
\end{theorem}
\begin{proof}
According to Theorem \ref{theorem:HOCBF}, to guarantee forward invariance of $\cap_{i=1}^r\mathcal{C}_i$, it is sufficient to show that for each $x\in\cap_{i=1}^r\mathcal{C}_i$ the input is selected such as $\psi_r(x,u)\geq0$. For the dynamics in \eqref{eq:dyn2}, if Assumption \ref{assumption:rd} holds, then $\psi_r(x,u)=L_{f}\psi_{r-1}(x) + L_{Y}\psi_{r-1}(x)\theta + L_g\psi_{r-1}(x)u + \alpha_r(\psi_{r-1}(x))$. Hence, it is our aim to show that any $u=k(x,\hat{\theta},t)$ satisfying $k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ satisfies $\psi_r(x,k(x,\hat{\theta},t))\geq0$ for all $x\in\cap_{i=1}^r\mathcal{C}_i$, all $\hat{\theta}\in\Theta$, and all $t\in\mathcal{I}$. To this end, observe that under control $u=k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ (omitting functional arguments for ease of presentation)
\[
\begin{aligned}
\psi_r
=& L_{f}\psi_{r-1} + L_{Y}\psi_{r-1}\hat{\theta} + L_{Y}\psi_{r-1}\tilde{\theta} \\ & + L_g\psi_{r-1}k + \alpha_r(\psi_{r-1}) \\
\geq & L_{f}\psi_{r-1} + L_{Y}\psi_{r-1}\hat{\theta} + L_g\psi_{r-1}k \\ &+ \alpha_r(\psi_{r-1}) - \|L_Y\psi_{r-1}\|\|\tilde{\theta}\| \\
\geq & L_{f}\psi_{r-1} + L_{Y}\psi_{r-1}\hat{\theta} + L_g\psi_{r-1}k \\ &+ \alpha_r(\psi_{r-1}) - \|L_Y\psi_{r-1}\|\nu \\
\geq & 0,
\end{aligned}
\]
for all $x\in\cap_{i=1}^r\mathcal{C}_i$, all $\hat{\theta}\in\Theta$, and all $t\in\mathcal{I}$.
In the above, the first inequality follows from the fact that $L_{Y}\psi_{r-1}(x)\tilde{\theta}\geq - \|L_{Y}\psi_{r-1}(x)\|\|\tilde{\theta}\|$, the second from the bound in \eqref{eq:theta_bound}, and the third from \eqref{eq:Kcbf}. Since $\psi_r(x,k(x,\hat{\theta},t))\geq0$ holds for all $x\in\cap_{i=1}^r\mathcal{C}_i$, all $\hat{\theta}\in\Theta$, and all $t\in\mathcal{I}$, it follows from Theorem \ref{theorem:HOCBF} that $\cap_{i=1}^r\mathcal{C}_i$ is forward invariant for the closed-loop system, as desired.
\end{proof}
Definition \ref{def:HO-RaCBF} and Theorem \ref{theorem:HO-RaCBF} generalize the ideas introduced in \cite{DixonACC21} to constraints with high relative degree, thereby facilitating the application of such ideas to more complex systems and safe sets. Although the results of \cite{DixonACC21} apply to safe sets defined by multiple barrier functions, whereas ours apply only to those defined by a single barrier function, there exist various approaches in the CBF literature \cite{LarsLCSS19,EgerstedtLCSS17} to formally combine multiple barrier functions\footnote{In practice, it is common to simply include multiple CBF-based constraints in a quadratic program such as the one proposed in \eqref{eq:HO-RaCBF-QP}, which, as we demonstrate empirically in Sec. \ref{sec:sim}, allows one to consider safe sets defined by multiple barrier functions.} using smooth approximations of min/max operators \cite{LarsLCSS19} or nonsmooth analysis \cite{EgerstedtLCSS17}. We now show that if the conditions of Theorem \ref{theorem:HO-RaCBF} hold on $\mathcal{D}$ and $x(0)\in\mathcal{D}\backslash\cap_{i=1}^r\mathcal{C}_i$, then any controller satisfying $k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ also guarantees asymptotic stability of $\mathcal{C}$, which ensures the developed controller is robust to perturbations \cite{AmesADHS15}.
\begin{corollary}\label{corollary:HO-RaCBF}
Let the conditions of Theorem \ref{theorem:HO-RaCBF} hold and suppose that $x(0)\in\mathcal{D}\backslash\cap_{i=1}^r\mathcal{C}_i$. Provided $u=k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ renders the closed-loop dynamics \eqref{eq:dyn2} forward complete, then the set $\cap_{i=1}^r\mathcal{C}_i$ is asymptotically stable for the closed-loop system.
\end{corollary}
\begin{proof}
It was shown in Theorem \ref{theorem:HO-RaCBF} that the proposed controller $k(x,\hat{\theta},t)\in \hat{K}_{\text{cbf}}(x,\hat{\theta},t)$ satisfies $\psi_r(x,k(x,\hat{\theta},t))\geq0$ for all $x\in\mathcal{D}$, $\hat{\theta}\in\Theta$ and $t\in\mathcal{I}$. Hence, by definition, $k(x,\hat{\theta},t)\in K_{\text{cbf}}(x)$ for all $x\in\mathcal{D}$, $\hat{\theta}\in\Theta$, and $t\in\mathcal{I}$. It then follows from \cite[Prop. 3 and Rem. 2]{DimosTAC21-hocbf} that $\cap_{i=1}^r\mathcal{C}_i$ is asymptotically stable for the closed-loop system.
\end{proof}
A controller satisfying the conditions of Theorem \ref{theorem:HO-RaCBF} can be computed through the use of quadratic programming (QP). Specifically, given an estimate of the uncertain parameters $\hat{\theta}$ and a nominal feedback control policy $k_d(x,\hat{\theta},t)$, a minimally invasive safe controller can be computed using the following HO-RaCBF-QP
\begin{equation}\label{eq:HO-RaCBF-QP}
\begin{aligned}
\min_{u\in\mathcal{U}} &\quad \tfrac{1}{2}\|u-k_d(x,\hat{\theta},t)\|^2 \\
\text{s.t.} & \quad L_{f}\psi_{r-1}(x) + L_Y\psi_{r-1}(x)\hat{\theta} + L_{g}\psi_{r-1}(x)u \\ & \geq -\alpha_{r}(\psi_{r-1}(x)) + \|L_Y\psi_{r-1}(x)\|\nu(t),
\end{aligned}
\end{equation}
which enforces the conditions of Theorem \ref{theorem:HO-RaCBF} provided the resulting controller is Lipschitz continuous and $h$ is a valid HO-RaCBF with $u\in\mathcal{U}$. That is, the above QP \eqref{eq:HO-RaCBF-QP} allows the nominal policy $k_d$ to be executed on \eqref{eq:dyn2} if $k_d$ can be formally verified as safe and intervenes in a minimally invasive fashion to guarantee safety only if $k_d$ cannot be certified as safe. We illustrate in the following section how one can leverage the same history stack used to reduce uncertainty in the parameter estimates to synthesize a desired policy $k_d$ with exponential stability guarantees.
\section{Exponentially Stabilizing Adaptive Control Lyapunov Functions}
In this section we introduce the concept of an exponentially stabilizing adaptive control Lyapunov function (ES-aCLF) as a tool to exponentially stabilize uncertain nonlinear systems in the presence of parametric uncertainty.
\begin{definition}\label{def:ES-aCLF}
A continuously differentiable positive definite function $V\,:\,\mathbb{R}^n\rightarrow\mathbb{R}_{\geq0}$ is said to be an \emph{exponentially stabilizing adaptive control Lyapunov function} (ES-aCLF) for \eqref{eq:dyn2} if there exist positive constants $c_1,c_2,c_3\in\mathbb{R}_{>0}$ such that for all $x\in\mathbb{R}^n$ and $\theta\in\mathbb{R}^p$
\begin{subequations}
\begin{equation}
c_1\|x\|^2\leq V(x)\leq c_2\|x\|^2,
\end{equation}
\begin{equation}\label{eq:aclf}
\inf_{u\in\mathcal{U}}\{L_{f}V(x) + L_YV(x)\theta + L_gV(x)u \}\leq -c_3V(x).
\end{equation}
\end{subequations}
\end{definition}
Given the above definition, let
\begin{equation}\label{eq:Kclf}
\begin{aligned}
K_{\text{clf}}(x,\theta)\coloneqq & \{u\in\mathcal{U}\,|\ L_{f}V(x) \\ &+ L_YV(x)\theta + L_gV(x)u\leq -c_3V(x)\},
\end{aligned}
\end{equation}
denote the point-wise set of all control values satisfying \eqref{eq:aclf}. The following lemma provides a parameter update law that can be combined with any locally Lipschitz control policy satisfying $k(x,\hat{\theta})\in K_{\text{clf}}(x,\hat{\theta})$ to guarantee stability.
\begin{lemma}\label{lemma:aclf}
Consider system \eqref{eq:dyn2}. Let $V$ be an ES-aCLF as in Def. \ref{def:ES-aCLF} and define $z\coloneqq\begin{bmatrix} x^\top & \tilde{\theta}^\top \end{bmatrix}^\top$. Provided the estimates of the unknown parameters are updated according to
\begin{equation}\label{eq:theta_dot}
\dot{\hat{\theta}}=\Gamma L_YV(x)^\top + \gamma\Gamma\sum_{j=1}^M\mathcal{Y}_j^\top(\Delta x_j - \mathcal{F}_j - \mathcal{Y}_j\hat{\theta} - \mathcal{G}_j),
\end{equation}
where $\Gamma\in\mathbb{R}^{p\times p}$ is a positive definite gain matrix and $\gamma\in\mathbb{R}_{>0}$ is a user-defined adaptation gain, then any locally Lipschitz controller $u=k(x,\hat{\theta})$ satisfying $k(x,\hat{\theta})\in K_{clf}(x,\hat{\theta})$
ensures that the composite system trajectory $t\mapsto z(t)$ remains bounded in the sense that for all $t\in[0,\infty)$
\begin{equation}\label{eq:z_bound}
\|z(t)\|\leq\sqrt{\tfrac{\eta_2}{\eta_1}}\|z(0)\|,
\end{equation}
where $\eta_1\coloneqq\min\{c_1,\tfrac{1}{2}\lambda_{\min}(\Gamma^{-1})\}$ and $\,\eta_2\coloneqq\max\{c_2,\tfrac{1}{2}\lambda_{\max}(\Gamma^{-1})\}$ are positive constants. Moreover
\[
\lim_{t\rightarrow\infty}x(t)=0.
\]
\end{lemma}
\begin{proof}
Consider the Lyapunov function candidate $V_a(z)\coloneqq V(x) + \frac{1}{2}\tilde{\theta}^\top\Gamma^{-1}\tilde{\theta}$, which can be bounded for all $z\in\mathbb{R}^{n+p}$ as $\eta_1\|z\|^2\leq V_a(z)\leq\eta_2\|z\|^2$. Taking the derivative of $V_a$ along the composite system trajectory yields
\begin{equation}\label{eq:Vadot}
\begin{aligned}
\dot{V}_a= & L_fV(x) + L_YV(x)\theta + L_gV(x)u - \tilde{\theta}^\top L_YV(x)^\top \\
& - \gamma\tilde{\theta}^\top\sum_{j=1}^M\mathcal{Y}_j^\top(\Delta x_j - \mathcal{F}_j - \mathcal{Y}_j\hat{\theta} - \mathcal{G}_j)\\
= & L_fV(x) + L_YV(x)\hat{\theta} + L_gV(x)u -\gamma\tilde{\theta}^\top\Lambda(t)\tilde{\theta},
\end{aligned}
\end{equation}
where $\Lambda$ is from \eqref{eq:Lambda}. Using the fact that $\Lambda(t)$ is at least positive semi-definite for all time implies $\dot{V}_a$ can be bounded as $\dot{V}_a\leq L_fV(x) + L_YV(x)\hat{\theta} + L_gV(x)u$. Choosing $u=k(x,\hat{\theta})\in K_{\text{clf}}(x,\hat{\theta})$ and the hypothesis that $V$ is a valid ES-aCLF allows $\dot{V}_a$ to be further bounded as $\dot{V}_a\leq-c_3V(x)\leq0$, revealing that $\dot{V}_a$ is negative semi-definite. Hence, $V_a$ is nonincreasing and $V_a(z(t))\leq V_a(z(0))$ for all $t\in[0,\infty)$, which can be combined with the bounds on $V_a$ to yield \eqref{eq:z_bound}. Since $V$ is continuous and $\dot{V}_a\leq-c_3V(x)\leq0$, it follows from the LaSalle-Yoshizawa theorem \cite[Thm. A.8]{Krstic} that $\lim_{t\rightarrow\infty}c_3V(x)=0$, implying $\lim_{t\rightarrow\infty}x(t)=0$.
\end{proof}
The following theorem shows that if sufficiently rich data is collected along the system trajectory, then $x(t)$ and $\tilde{\theta}(t)$ both exponentially converge to the origin.
\begin{theorem}\label{theorem:ES-aCLF}
Under the assumption that the conditions of Lemma \ref{lemma:aclf} hold, suppose that there exists a time $T\in\mathbb{R}_{\geq0}$ and a positive constant $\underline{\lambda}\in\mathbb{R}_{>0}$ such that $\lambda(t)\geq\underline{\lambda}$ for all $t\in[T,\infty)$. Then, for all $t\in[0,T)$, $z(t)$ is bounded in the sense that \eqref{eq:z_bound} holds. Furthermore, for all $t\in[T,\infty)$, $z(t)$ exponentially converges to the origin at a rate proportional to $\eta_3\coloneqq \min\{\gamma\underline{\lambda},c_1c_3\}$ in the sense that for all $t\in[T,\infty)$
\begin{equation}\label{eq:z_exp_bound}
\|z(t)\|\leq\sqrt{\tfrac{\eta_2}{\eta_1}}\|z(T)\|e^{-\frac{\eta_3}{2\eta_2}(t-T)}.
\end{equation}
\end{theorem}
\begin{proof}
Since $\lambda(t)\geq0$ for all $t\in[0,T)$ the conclusions of Lemma \ref{lemma:aclf} hold, implying $z(t)$ is bounded as in \eqref{eq:z_bound} for all $t\in[0,T)$. Provided $\lambda(t)\geq\underline{\lambda}$ for all $t\in[T,\infty)$ and $u=k(x,\hat{\theta})\in K_{clf}(x,\hat{\theta})$ for all $(x,\hat{\theta})\in\mathbb{R}^{n+p}$, then \eqref{eq:Vadot} can be bounded as
\[
\dot{V}_a\leq -c_3V(x) - \gamma\underline{\lambda}\|\tilde{\theta}\|^2\leq -\eta_3\|z\|^2\leq -\frac{\eta_3}{\eta_2}V_a.
\]
Invoking the comparison lemma \cite[Lem. 3.4]{Khalil} implies $t\mapsto V_a(z(t))$ is bounded for all $t\in[T,\infty)$ as
\[
V_a(z(t))\leq V_a(z(T))e^{-\frac{\eta_3}{\eta_2}(t-T)},
\]
which can be combined with the bounds on $V_a$ to yield \eqref{eq:z_exp_bound}.
\end{proof}
Similar to the previous section, given an estimate of the uncertain parameters $\hat{\theta}$, control inputs satisfying \eqref{eq:aclf} can be computed using the following ES-aCLF-QP
\begin{equation}\label{eq:ES-aCLF-QP}
\begin{aligned}
\min_{u\in\mathcal{U}}\quad & \tfrac{1}{2}u^\top u \\
\text{s.t.}\quad & L_fV(x) + L_YV(x)\hat{\theta} + L_gV(x)u\leq -c_3V(x).
\end{aligned}
\end{equation}
A controller guaranteeing stability and safety\footnote{Similar to works such as \cite{TaylorACC20,LopezLCSS21}, the parameter update laws for the proposed adaptive CBF and CLF are different | if one wishes to combine the two in a single QP-based controller, separate estimates of the uncertain parameters must be maintained. Despite this, note that the data from a single history stack can be used in both update laws.} can then be synthesized by either taking the solution to \eqref{eq:ES-aCLF-QP} as $k_d$ in \eqref{eq:HO-RaCBF-QP} or by forming a single QP with the HO-RaCBF constraint from \eqref{eq:HO-RaCBF-QP} and a relaxed version of the ES-aCLF constraint from \eqref{eq:ES-aCLF-QP} to guarantee feasibility {\color{black} provided the resulting controller is Lipschitz continuous}.
\begin{remark}
Note that asymptotic stability of the origin for the closed-loop system \eqref{eq:dyn2} with $u=k(x,\hat{\theta})\in K_{\text{clf}}(x,\hat{\theta})$ is guaranteed by Lemma \ref{lemma:aclf} regardless of whether or not the richness of data condition $\lambda(t)\geq\underline{\lambda}>0$ is satisfied. In this situation, the stability guarantees induced by the ES-aCLF reduce to those of the classically defined adaptive CLF from \cite[Ch. 4.1]{Krstic} (see also \cite{TaylorACC20}). In this regard, neither safety nor stability is predicated upon collecting sufficiently rich data | this data is exploited only to reduce conservatism of the HO-RaCBF controller and to endow the ES-aCLF controller with exponential convergence guarantees.
\end{remark}
\begin{remark}
The concept of an ES-aCLF generalizes the adaptive control designs from \cite{DixonIJACSP19} in that the Lyapunov functions used to verify stability of the controllers proposed in \cite{DixonIJACSP19} meet the criteria of an ES-aCLF posed in Def. \ref{def:ES-aCLF}. We refer the reader to works such as \cite{AmesTAC14} for a discussion on the potential advantages of using an optimization-based CLF-based control law as in \eqref{eq:ES-aCLF-QP} over a traditional closed-form feedback control law such as those posed in \cite{DixonIJACSP19}.
\end{remark}
\section{Case Studies}\label{sec:sim}
\subsection{Robotic Navigation}
Our first example involves a robotic navigation task for a system modeled as a double integrator with uncertain friction effects. The system is in the form of \eqref{eq:dyn2}:
\[
\begin{aligned}
\underbrace{
\begin{bmatrix}
\dot{x}_1 \\ \dot{x}_2 \\ \dot{x}_3 \\ \dot{x}_4
\end{bmatrix}
}_{\dot{x}}
=
\underbrace{
\begin{bmatrix}
x_3 \\ x_4\\ 0\\ 0
\end{bmatrix}
}_{f(x)}
+
\underbrace{
\begin{bmatrix}
0 & 0 \\ 0 & 0 \\ -\tfrac{x_3}{\mathrm{m}} & 0 \\ 0 & -\tfrac{x_4}{\mathrm{m}}
\end{bmatrix}
}_{Y(x)}
\underbrace{
\begin{bmatrix}
\mu_1 \\ \mu_2
\end{bmatrix}
}_{\theta}
+
\underbrace{
\begin{bmatrix}
0 & 0\\ 0 & 0 \\ \tfrac{1}{\mathrm{m}} & 0 \\ 0 & \tfrac{1}{\mathrm{m}}
\end{bmatrix}
}_{g(x)}
\underbrace{
\begin{bmatrix}
u_1 \\ u_2
\end{bmatrix}}_{u},
\end{aligned}
\]
where $\mathrm{m}\in\mathbb{R}_{>0}$ is a known mass and $\mu_1,\mu_2\in\mathbb{R}_{>0}$ are uncertain viscous friction coefficients. For simplicity, we set $\mathrm{m}=\mu_1=\mu_2=1$. The objective is to stabilize the system to the origin while avoiding a set of static obstacles in the state space. To achieve the stabilization task, we construct an ES-aCLF as $V(x)\coloneqq x^\top Px$ with $P=[2,\, 0,\, 1,\, 0; 0,\, 2,\, 0,\, 1; 1,\, 0,\, 1,\, 0; 0,\, 1,\, 0,\, 1]$ and $c_3=1$. The safety objective is achieved by considering a collection of safe sets of the form \eqref{eq:C} with $h^{(i)}(x)\coloneqq (x_1 - p_{1,i})^2 + (x_2 - p_{2,i})^2 - R^2$, where $(p_{1,i},p_{2,i})\in\mathbb{R}^2$ denotes the center of a circular disk and $R_i\in\mathbb{R}_{>0}$ its radius. With straightforward calculations one can verify that this CBF candidate has relative degree $r=2$ with respect to the dynamics and that Assumption \ref{assumption:rd} is satisfied. Although the dynamics and safe set are relatively simple, the aCBF techniques from \cite{TaylorACC20,LopezLCSS21,DixonACC21,SanfeliceACC21} cannot be applied to solve the problem as currently constructed since the relative degree of $h$ is larger than one\footnote{Note that it may be possible to construct a relative degree one CBF for such a problem by including dependence on velocity into the definition of $h$. As noted earlier, an advantage of HOCBFs is that the design of $\mathcal{C}$ is greatly simplified since dependence on velocity is naturally encoded by higher order terms rather than through the explicit definition of $h$.}. To illustrate the efficacy of the developed approach, we simulate the system under the HO-RaCBF/ES-aCLF control architecture presented herein and compare the results to controllers {\color{black} take a purely robust approach}. The HO-RaCBF/ES-aCLF controller is computed by taking the solution to \eqref{eq:ES-aCLF-QP} as $k_d$ in \eqref{eq:HO-RaCBF-QP}. For each simulation we construct two HOCBFs as outlined previously with $(p_{1,1},p_{2,1})=(-1.75,2)$, $(p_{1,2},p_{2,2})=(-1,0.5)$, and $R_1=R_2=0.5$. For each simulation, the system is initialized as $x(0)=(-2.5,2.5,0,0)$ and all parameter estimates are initialized at zero. All extended class $\mathcal{K}$ functions used are defined as $\alpha(r)=r$. For the learning-based approach, we maintain a history stack with $M=20$ entries, the integration window is chosen as $\Delta T=0.5$, and the learning rates are chosen as $\Gamma=I_2$ and $\gamma=10$.
To demonstrate the relationship between adaptation and safety, we run set of simulations comparing system performance using the HO-RaCBF/ES-aCLF controller to that under a purely robust approach (i.e., accounting for the maximum possible parameter estimation error without adaptation), where the set of possible parameters $\Theta$ is varied, the results of which are presented in Fig. \ref{fig:double_int_traj}-\ref{fig:theta}. The controller used in the robust approach is the same as that of the adaptive approach but with $\gamma=0$. As shown in Fig. \ref{fig:double_int_traj} and Fig. \ref{fig:cbf_traj}, each trajetory is safe; however, for larger levels of uncertainty the purely robust controller is overly conservative, causing the system trajectory to diverge from the origin. In constrast, the adaptive controller reduces the uncertainty online and achieves dual objectives of stability and safety. In fact, the convergence of the trajectory to the origin under the adaptive controller is minimally affected by the initial level of uncertainty, whereas the trajectory under the robust controller fails to converge to the origin in the presence of large uncertainty. The ability of the learning scheme to identify the uncertain parameters is showcased in Fig. \ref{fig:theta}, which shows the trajectory of the estimated parameters used in the ES-aCLF QP and HO-RaCBF QP, both of which converge to their true values in just under 15 seconds.
\begin{figure}
\centering
\includegraphics{dbl_int_traj.pdf}
\caption{System trajectory under each controller across four different uncertainty sets with solid lines denoting trajectories under the adaptive controller and dashed lines denoting trajectories under the purely robust controller. The gray disks represent the obstacles.}
\label{fig:double_int_traj}
\end{figure}
\begin{figure}
\centering
\includegraphics{cbf_traj.pdf}
\caption{Minimum value among the two HOCBFs point-wise in time along each system trajectory. The solid and dashes curves have the same interpretation as those in Fig. \ref{fig:double_int_traj} and the dashed black line denotes $h(x)=0$.}
\label{fig:cbf_traj}
\end{figure}
\begin{figure}
\centering
\includegraphics{th_cbf.pdf}
\caption{Estimates of the uncertain parameters used in the ES-aCLF QP and HO-RaCBF QP for the simulation corresponding to the uncertainty set $\Theta=[0,3]^2$}
\label{fig:theta}
\end{figure}
\subsection{Inverted Pendulum}
To demonstrate the applicability of the developed results to an unstable nonlinear system with uncertain parameters, we now consider an inverted pendulum of the form \eqref{eq:dyn2} as
\[
\begin{aligned}
\underbrace{
\begin{bmatrix}
\dot{x}_1 \\ \dot{x}_2
\end{bmatrix}
}_{\dot{x}}
=
\underbrace{
\begin{bmatrix}
x_2 \\ 0
\end{bmatrix}
}_{f(x)}
+
\underbrace{
\begin{bmatrix}
0 & 0 \\ \tfrac{1}{\ell}\sin(x_1) & -\tfrac{1}{\mathrm{m}}x_2
\end{bmatrix}
}_{Y(x)}
\underbrace{
\begin{bmatrix}
\mathrm{g} \\ \mathrm{c}
\end{bmatrix}
}_{\theta}
+
\underbrace{
\begin{bmatrix}
0 \\ \tfrac{1}{\mathrm{m}\ell^2}
\end{bmatrix}
}_{g(x)}
u
\end{aligned},
\]
with length $\ell=0.7$, mass $\mathrm{m}=0.7$, gravitational acceleration $\mathrm{g}=9.8$, and damping coefficient $\mathrm{c}=0.2$. The objective is to regulate $x$ to the origin while satisfying $-\tfrac{\pi}{4}\leq x_1(t)\leq \tfrac{\pi}{4}$ for all time. To achieve the stabilization objective we select the aCLF $V(x)=x^\top Px$ with $P=[1,\,0.5;\,0.5,\,0.5]$ and $c_3=2.5\tfrac{\lambda_{\min}(Q)}{\lambda_{\max}(P)}$, where $Q=[2,\,1;1,\,1]$. The safety objective is achieved by constructing two safe sets defined by $h_1(x)=x_1 + \tfrac{\pi}{4}$ and $h_2(x)=\tfrac{\pi}{4} - x_1$, both of which have relative degree 2 with respect to the dynamics. Again, the results of \cite{TaylorACC20,LopezLCSS21,DixonACC21,SanfeliceACC21} are not applicable to this example since $h$ has relative degree larger than one. The uncertain parameters $\theta=(\theta_1,\theta_2)=(\mathrm{g},\mathrm{c})$ are assumed to take values from $\Theta=[7,13]\times[0,3]\subset\mathbb{R}^2$ and higher order terms from \eqref{eq:psi} are defined using $\alpha_1(r)=\alpha_2(r)=5r$.
To compare the performance between various control architectures, the system is simulated under the influence of 1) a HO-RaCBF/ES-aCLF controller as in \eqref{eq:HO-RaCBF-QP} with $k_d$ obtained from solving the ES-aCLF QP \eqref{eq:ES-aCLF-QP}; 2) an ES-aCLF controller as in \eqref{eq:ES-aCLF-QP}; 3) an ``open-loop" HO-RaCBF controller as in \eqref{eq:HO-RaCBF-QP} with $k_d\equiv 0$ and an initial condition outside the safe set. For each simulation all parameters associated with the learning scheme remain the same as in the previous example. The results the simulations are presented in Fig. \ref{fig:inv_pen_traj} and Fig. \ref{fig:inv_pen_theta}. Although the ES-aCLF controller achieves the stabilization objective (orange curves in Fig. \ref{fig:inv_pen_traj}), it does so at the cost of violating the safety constraints. In essence, for the ES-aCLF controller to achieve the stabilization objective, it must first allow the pendulum to tip over to collect sufficiently rich data for identifying the dynamics, which, once collected, allows for rapid stabilization to the origin. In contrast, by augmenting the ES-aCLF controller with a HO-RaCBF (blue curves in Fig. \ref{fig:inv_pen_traj}), the stabilization objective is achieved \emph{safely} in an exponential fashion. As predicted by Corollary \ref{corollary:HO-RaCBF}, the HO-RaCBF controller is capable of stabilizing the system to the safe set even when starting from an unsafe initial condition (purple curve in Fig. \ref{fig:inv_pen_traj}). Furthermore, all conservatism associated with the initial parameter uncertainty has essentially been eliminated | once the system is stabilized to the safe set, the HO-RaCBF controller allows the pendulum to lie on the boundary of the safe set without crossing it. Similar to the previous example, the uncertainty in the parameter estimates is exponentially driven to zero (see Fig. \ref{fig:inv_pen_theta}), implying that after only a few seconds of learning, the HO-RaCBF and ES-aCLF conditions from Def. \ref{def:HO-RaCBF} and Def. \ref{def:ES-aCLF}, respectively, closely approximate the original HOCBF condition from Def. \ref{def:HOCBF} and ES-CLF condition from e.g. \cite{AmesTAC14,AmesTAC17,AmesECC19}.
\begin{figure}
\centering
\includegraphics{inv_pen_traj.pdf}
\caption{Evolution of the pendulum's orientation under each control architecture. The dashed black line denote the boundary of the safe set.}
\label{fig:inv_pen_traj}
\end{figure}
\begin{figure}
\centering
\includegraphics{inv_pen_th.pdf}
\caption{Norm of the parameter estimation error under each control architecture. The color of each curve shares the same interpretation of those in Fig. \ref{fig:inv_pen_traj}.}
\label{fig:inv_pen_theta}
\end{figure}
\section{Conclusions}
In this paper we introduced HO-RaCBFs and ES-aCLFs as a means to synthesize safe and stable control policies for uncertain nonlinear systems with high relative degree safety constraints. The novel class of HO-RaCBF is, to the best of our knowledge, the first to extend the aCBF paradigm from \cite{TaylorACC20} to CBFs with arbitrary relative degree under mild assumptions regarding the structure of the uncertainty and, unlike existing formulations, the proposed HO-RaCBF inherits desirable robustness properties of zeroing CBFs \cite{AmesADHS15}. The class of ES-aCLFs introduced herein builds upon the classical aCLF formulation \cite[Ch. 4.1]{Krstic} by leveraging data-driven techniques from CL adaptive control to guarantee exponential stability. The advantages of the proposed HO-RaCBFs/ES-aCLFs were illustrated through two numerical examples that cannot be addressed by existing approaches. Directions for future research include relaxing Assumption \ref{assumption:rd} and extending the approach to systems with nonparametric and actuation uncertainty.
\section*{Acknowledgements}
We thank the anonymous reviewers for their helpful comments and suggestions, which have improved the presentation of the material in paper.
\bibliographystyle{ieeetr}
| b00e62009d9a95aaee42615daf619128477ce7b4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Diffuse, non-thermal emission has been observed in more than 100 clusters of galaxies, revealing the existence of magnetic fields and relativistic particles on scales as large as a few megaparsecs. In the last decade, the advent of low-frequency, sensitive radio observations has brought about a major advance in the discovery of these objects and the characterization of their properties.\\
Synchrotron emission from the intra-cluster medium (ICM) has been observed in the form of giant radio halos, mini halos, and radio relics, depending on their location and morphology. Giant radio halos are found at the centres of merging galaxy clusters, co-spatial with the X-ray emitting gas, and with sizes of 1-2\,Mpc. Mini halos only have sizes of few hundreds of kpc and are found mostly in cool-core clusters. Finally, radio relics are arc-like sources located in cluster outskirts where they trace merger shock waves. We refer to \cite{vanWeeren19} for a recent review.\\
Since their discovery, it has been proposed that radio halos and relics are generated by shocks and turbulence driven in the ICM by cluster mergers (see e.g. \citealt{BJ14} for a review, and ref. therein). Although the details of the proposed mechanisms are not understood yet, radio halos have preferentially been found in merging clusters, supporting a connection between mergers and radio emission \citep{Cuciti15}, and radio relics are often coincident with gas discontinuities detected in the X-ray and Sunyaev-Zeldovich' (SZ) images \citep[e.g.][]{PlanckComa,OgreanToothbrish}. The origin of mini halos is still debated. They could either originate from
turbulent motions that develop in the cluster core \citep[e.g.][]{ZuHone13} or from hadronic collisions between cosmic ray protons and thermal protons \citep[e.g][]{Pfrommer04}.\\
In the last years, the picture has become more complicated, and it has been found that some clusters with mini halos also host a larger-scale radio component, resembling a dimmer version of the giant radio halos but in non-merging objects \citep[e.g.,][]{Savini_2019,Raja20}. Moreover, a giant halo has also been found in a strong cool-core cluster CL1821+643 \citep{Bonafede14b}. In addition, relics and halos are sometimes connected through low-brightness radio bridges \citep{vanweeren12,Bonafede21}, that could be powered by mechanisms similar to those that are currently used to explain radio halos.
On even larger scales, giant bridges of radio emission have been discovered, connecting massive clusters in a pre-merging state \citep[e.g.,][]{Govoni19,Botteon20bridge}.\\
Despite the differences between these types of sources, they all have a low surface brightness ($\sim 1 \mu \rm{Jy/arcsec^2}$) at GHz frequencies. Also they have steep radio spectra \footnote{Throughout this paper, we define the spectral index $\alpha$ as $S(\nu) \propto \nu^{\alpha}$, where $S$ is the flux density at the frequency $\nu$}, with a spectral index $\alpha < -1$ that make them brighter at low radio frequencies.
Hence, the advent of deep, low-frequency radio surveys, such as the LOFAR Two-metre Sky Survey \citep[LoTSS;][]{Shimwell17,Shimwell19} has both increased the number of new detections (see e.g. \citealt{Biava21b,Riseley21}; Botteon et al., accepted; Hoang et al, submitted) and allowed the study of known objects with unprecedented sensitivity and detail.\\
The Coma cluster hosts the most famous and best studied radio halo, as well as a radio relic and a radio bridge connecting the two (see Fig. \ref{fig:ComaRGB}). The emission from the Coma field has been the subject of many studies since its discovery \citep[e.g.][]{Large59, Ballarati81, Giovannini91, Venturi90, Kronberg07,BrownRudnick11}.
In paper I \citep{Bonafede21}, we have analysed the properties of the radio bridge and
we have shown under which conditions it can be powered by turbulent acceleration.
In this paper, we focus on the radio halo and relic of the Coma cluster. New LOFAR data give us information in regions that have been so far inaccessible, providing important clues on the origin of the radio emission.\\
\cite{Giovannini93} first found that the spectrum of the halo between 326 MHz and 1.28 GHz is characterized by two different regions: a central one with $\alpha \sim -0.8$ and a peripheral one with $\alpha \sim -1.2$. The integrated spectrum, computed between 30 MHz and 4.8 GHz shows a high-frequency steepening consistent with homogeneous in-situ re-acceleration models.
They also reported a smooth distribution of the radio surface brightness, with no evidence for substructures at the resolution of $\sim$50$^{\prime \prime} \,$.
Their results have been confirmed later on by several authors \citep[e.g][]{Thierbach03} who complemented the analysis with observations up to 4.8 GHz, where the halo is barely detected.
The steepening of the halo at high frequencies has been subject to debate, as the decrement due to the SZ effect was initially not accounted for. After the observation of the SZ effect by the Planck satellite \citep{Planck11,Planck14} it has been possible to confirm the steepening of the radio emission \citep{Brunetti13}.
However, we note that data below 300 MHz, being taken in the 80s, lack the sensitivity, resolution, calibration and imaging accuracy that are allowed by present instruments and techniques. In particular, the observation at 151 MHz \citep{Cordey85}, i.e. the closest in frequency to LOFAR HBA, has a resolution of $\sim$70$^{\prime \prime} \,$, and a sensitivity of a few 10 mJy/beam, that allowed the detection of the central part of the halo only.
Similarly, observations at 43 and 73 MHz by \cite{HanischErickson} did not allow to properly subtract the emission from radiogalaxies present in the cluster. \\
In addition to integrated spectral studies, better resolved spatial analysis of radio halos spectral properties yields important information about the distribution of the non-thermal component in the ICM \citep[e.g.][]{Kamlesh20,Rajpurohit21a,Rajpurohit21}.
The halo in the Coma cluster offers a unique chance to perform spatially resolved studies of the halo brightness and of its connection between thermal and non-thermal plasma \citep[e.g][]{Govoni01,BrownRudnick11}.
So far, these studies have been inhibited by the lack of resolution, as
it is often necessary to convolve the radio images with large Gaussian beams to recover the full halo emission.\\
LOFAR \citep{vanHaarlem}, thanks to its sensitivity and resolution, provides a step forward for precise measurements of the halo size and flux density, and for resolved studies. In this paper, we study the radio emission from the Coma cluster at 144\,MHz, using data from LoTSS (\citealt{Shimwell19}, Shimwell et al. in press) after ad-hoc reprocessing.
We also use published data, and reprocessed archival X-ray and radio observations, to perform a multi-wavelength study of the emission.\\
This paper is organized as follows: In Sec.~\ref{sec:data} we describe the observations and data reduction procedures, in Sec.~\ref{sec:diffuse} we present the main sources of diffuse emission, both known and newly discovered. The radio halo is analysed in Sec.~\ref{sec:halo_fit}, its spectrum in Sec. ~\ref{sec:haloSpectrum} and its correlation with the thermal gas are shown in Sec.~\ref{sec:radioXcorrelation} and \ref{sec:radioSZ}. In particular, in Sec. 5.3 and 6.4, the halo properties as discussed in the framework of turbulent re-acceleration models, and constraints to the model parameters are derived. In Sec.~\ref{sec:simulations}, we use cosmological MHD simulations to reproduce the thermal/non-thermal properties of the Coma cluster. In Sec. \ref{sec:haloFront}, the halo front is analysed and its origin is discussed in connection with the shock wave found by X-ray and SZ studies. We also analyse the emission detected between the relic and the radiogalaxy NGC4789 in Sec.\ref{sec:relicNAT}. Finally, results are discussed in
sec.~\ref{sec:discussion}.
Throughout this paper, we use a $\Lambda$CDM cosmological model, with $\Omega_{\Lambda}=0.714$, $H_0=$69.6 km/s/Mpc. At the Coma redshift ($z=0.0231$) the angular to linear scale is 0.469 kpc/arcsec and the Luminosity distance $D_L=101.3$ Mpc.
\begin{figure*}
\includegraphics[width=2\columnwidth]{result_rgb_add.png}
\caption{Composite IR-radio image of the Coma cluster field. In white the IR image in band 1, 2, and 3 of WISE (\emph{Wide-field Infrared Survey Explorer}) are shown. The red-orange color scale shows the composite radio image of the diffuse emission at 1$^{\prime}$ for the diffuse emission and of 20$^{\prime \prime} \,$ for the sources in the field.}
\label{fig:ComaRGB}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{Coma_1arcmin_allfield.pdf}
\caption{LOFAR emission of the Coma field at the resolution of 1 arcmin. The candidate accretion relic, halo front, bridge, relic, and NAT-relic connection are labelled.}
\label{fig:Coma_fov}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=2.1\columnwidth]{Coma_zoom_filaments.pdf}
\caption{Left panel: zoom into the radio halo from the 20$^{\prime \prime} \,$ image, which is made imposing an inner UV cut (image LOFAR as LoTSS 20 in Tab. \ref{tab:images}). The halo core and the radio galaxies, both, of Coma and of the field are visible. The black bowl around the core indicates that large-scale diffuse emission is filtered out.
White arrows mark the filaments (3 arrows on the left) and the radio-loop (arrow on top-right). Right panel: same as left panel but from the 35$^{\prime \prime} \,$ image, where all the baselines have been included in the image and the discrete sources have been subtracted (see text for details)
The outer halo and the halo front are well visible. Contours start at 3$\sigma_{\rm rms}$ and are spaced by a factor of 2.}
\label{fig:zoomHalo}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=2.1\columnwidth]{Coma_zoom_Relic.pdf}
\caption{Zoom into the radio relic region. Left panel:
6$^{\prime \prime} \,$ image showing the complex emission that links the source NGC 4789 to the relic emission. Right panel: 35$^{\prime \prime} \,$ image. The source NGC 4789 is labelled. Arrows mark the position of the stripes that depart from the relic towards NGC4839. White dashed circles indicate residuals from the subtraction of the lobes of NGC4827, a Coma radiogalaxy, and Coma A. Contours start at 3 $\sigma_{\rm rms}$ and are spaced by a factor of 2.}
\label{fig:zoomRelic}
\end{figure*}
\section{Observations used in this work}
\label{sec:data}
\subsection{LOFAR data and data reduction}
The data used in this work are part of the LoTSS (\citealt{Shimwell19}, Shimwell et al. 2022) and consist of 2 pointings of 8~hr each taken with LOFAR (the LOw Frequency ARray, \citealt{vanHaarlem}) High Band Array antennas, in the DUAL\_INNER mode configuration. Each pointing is 8~hr long, book-ended by 10~min observations of a calibrator (3C196), used to correct for the ionospheric Faraday rotation, clock offsets, instrumental XX and YY phase offsets, and time-independent amplitude solutions. The pointings are specified in Tab. \ref{tab:obs} together with the distance from the central source NGC~4874, at the cluster centre.
Observations are centered at 144\,MHz and have a 48\,MHz total bandwidth. After pre-processing and direction-independent calibration, data are averaged into 24 visibility files, each having a bandwidth of 1.953\,MHz with a frequency resolution of 97.6\,kHz and a time resolution of 8s. For details about pre-processing and direction-independent calibration steps, we refer to \cite{Shimwell19}, where LoTSS data acquisition and processing are both explained in detail.
The direction-dependent calibration has been made using the LoTSS DR2 pipeline, but with a slightly different procedure to account for the large-scale emission present in the Coma field. Specifically, we have included all baselines in imaging and calibration, while the LoTSS DR2 pipeline applies an inner \emph{uv}-cut of the visibilities below 100~m to eliminate radio frequency interferences on the shortest baselines and filter out large-scale Galactic emission that would make the process of calibration and imaging more difficult. The two pointings have been calibrated separately and unrelated sources have been subtracted from the visibilities through a multi-step procedure, for which we refer the reader to \cite{Bonafede21}. Briefly, we first subtracted all the sources outside a radius of 1.5$^{\circ}$ centered on the cluster, using the model components obtained from an image at 20$^{\prime \prime} \,$ resolution with no inner \emph{uv}-cut, and using a threshold of 5$\sigma$ (0.75 mJy/beam). The subtracted data have been reimaged using an \emph{uv}-cut of 300~m to filter out the diffuse emission from the ICM, and the model components have been subtracted using a threshold of 0.6 mJy/beam. Data have been reimaged again to check for the presence of residual emission from discrete sources using an \emph{uv}-cut of 100 m. Residual emission associated with Active Galacic Nuclei (AGN) and sources unrelated to the halo and relic emission has been identified and subtracted by the data. The sources NGC4789 and NGC4839 have not been subtracted in order to study how their emission is connected to the relic and bridge, respectively.
Two diffuse patches of emission to the West of Coma are still present and have been deconvolved in the imaging runs. They are discussed in the following sections.\\
The images used in this work have been done in a different way than in \cite{Bonafede21}, taking advantage of the new implementations that have been added to the imaging software in the last months. Imaging has been done with DDFacet \citep{Tasse18}, using the recently added features that allow to deconvolve large-scale emission in joint-deconvolution mode.
We used the Sub Space Deconvolution (SSD) algorithm (\citealt{Tasse18} and ref. therein) to better model the clean components. Typically, four major cycles were needed to achieve a noise-like residual map.
The beam correction has been applied in the image plane, interpolating the beams of the two different pointings in the direction of Coma.
Different images have been produced, as listed in Tab. \ref{tab:images}, at various resolution to highlight the emission from discrete sources and from the diffuse emission.
To align the flux scale to LoTSS, we have extracted the fluxes from the sources in the Coma field, and followed the same bootstrap procedure described in \citet{Hardcastle16}, which is based on the NVSS (NRAO VLA Sky Survey, \citealt{NVSS}).
\subsection{XMM-Newton and ROSAT data}
We used the \emph{XMM-Newton} Science Analysis System (SAS) v18.0.0 for data reduction. The ObsIDs we used are listed in Appendix \ref{appendix:obsid}. Event files from the MOS and pn detectors were generated from the observation data files with the tasks \texttt{emproc} and \texttt{epproc}. The out-of-time (OoT) events of pn were corrected following the user guide\footnote{\url{http://xmm-tools.cosmos.esa.int/external/xmm_user_support/documentation/sas_usg/USG/removingOoTimg.html}}.
We used stacked Filter Wheel Closed (FWC) event files to generate non X-ray background (NXB) maps. For each ObsID, the FWC event files were reprojected using task \texttt{evproject} to match the observation. NXB maps were scaled to match the NXB level of each count image. For MOS, the scale factor was calculated using the ratio of out-of-field-of-view (OoFoV) count rates. For pn, even the OoFoV area can be contaminated by soft protons, and cannot be used for an accurate rescaling of the NXB \citep{Gastaldello17,Zhang20,Marelli21}. On the other hand, the pn instrumental background shows a similar long term variability as that of \emph{Chandra}'s ACIS-S3 chip; therefore we used this information for rescaling the pn NXB level in the observations (details are provided in Zhang et al. in prep).\\
The corresponding exposure maps were generated using the task \texttt{eexpmap} with parameter \texttt{withvignetting=yes}. Point sources were detected and filled by the tasks \texttt{wavdetect} and \texttt{dmfilth} in the $Chandra$ Interactive Analysis of Observations (CIAO) v4.13 package. We stacked individual count maps, exposure maps, and NXB maps, respectively, using the 0.5--2.0 keV energy band. We divided the NXB subtracted count image by the exposure map to generate the flux map. After removing the instrumental background and correcting for telescope vignetting, a constant sky background was further subtracted from the images. The level of this background was estimated from the median flux in an annulus spanning radii of 60-70 arcmin. Note that \citet{Mirakhor20} do detect a signal from the Coma ICM even at these large radii, albeit in a narrower energy interval of 0.7-1.2 keV where the signal to background is optimized. The sky background used here may therefore be slightly overestimated, but this does not have an impact on the radii of interest considered in our \emph{XMM-Newton} analysis (limited in this case mostly within the cluster's R$_{500}$). \\
We furthermore used the four archival observations of the Coma cluster performed in June 1991 by the ROSAT Position Sensitive Proportional Counter (PSPC). These pointings
extend out to radii of 60-70 arcmin, totaling a clean exposure time of 78 ks. The data reduction was performed exactly as already described in \citet{Simionescu13,Bonafede21}. The low instrumental background of ROSAT makes these data an important complement to \emph{XMM-Newton} in the faint cluster outskirts.
\subsection{Planck data}
We used six Planck frequency maps acquired by the High Frequency Instrument (HFI) and the corresponding energy responses released in 2018 by the Planck collaboration \citep{Planck20}. The average frequencies of the HFI maps are 100, 143, 217, 353, 545 and 857 GHz. The nominal angular resolution of the maps are 9.69, 7.30, 5.02, 4.94, 4.83, and 4.64 arcminutes, respectively. Following prescriptions of Galactic thermal dust studies performed by the Planck collaboration (Planck collaboration, 14), we corrected HFI maps for individual offsets that maximise their spatial correlation with neutral hydrogen density column measurements performed in regions of the sky characterised by their low dust emissivity and by the exclusion of prominent Sunyaev-Zel'dovich (SZ) sources. \\
This multi-frequency data set allowed us to map the thermal SZ Compton parameter in a four square degree sky area centred on Coma, following the spectral-imaging algorithm described in \citet{Baldi19}. Briefly, the Compton parameter, $y$, is jointly mapped in the wavelet space with the Cosmic Microwave Background anisotropies and with the Galactic thermal dust emissivity. This is achieved via a spatially-weighted likelihood approach that includes the smoothing effect of individual HFI beams in the reconstruction of B3-spline wavelet coefficients. In order to best restore anisotropic details, a curvelet transform is eventually computed from these wavelet coefficients, and denoised via a soft-thresholding that we parametrise as a function of local values of the noise standard deviation. The resulting $y$-map has a resolution (FWHM of the PSF) of 5 arcmin, i.e. a factor 2 higher than the one published in \cite{PlanckComa}. \\
\subsection{WSRT data}
We use data from the Westerbork Synthesis Radio Telescope (WSRT) at $\sim$ 325 MHz. Part of these data come from the observations published in \cite{Venturi90,Giovannini91}, that we have reimaged as explained in \cite{Bonafede21}.
In this work, we also use more recent WSRT observations published by \cite{BrownRudnick11} that recover a larger fraction of the radio halo, but are more affected by imaging and calibration artefacts due to the presence of the source Coma A, North of the relic.
\cite{BrownRudnick11} observations are used here to study the spectral properties of the halo, while \cite{Venturi90,Giovannini91} observations are used here to study the spectral index in the relic region. In Tab. \ref{tab:images}, these images and their main properties are listed. We refer to the image published by \citet{BrownRudnick11} as WSRT H, as it is used to study the halo emission, and to the image published by \cite{Venturi90,Giovannini91} as WSRT R, as it is used to study the relic region.
\begin{table}
\centering
\caption{ LOFAR observations details}
\begin{tabular}{l c c c c } \hline
LoTSS & RA & DEC & time & dist of NGCC4874 \\
pointing ° & deg & h & deg \\
P192$+$27 & 192.945 & 27.2272 & 8 & 1.88 \\
P195$+$27 & 195.856 & 27.2426 & 8 & 1.11 \\
\hline
\end{tabular}
\label{tab:obs}
\end{table}
\begin{figure*}
\includegraphics[width=2.1\columnwidth]{Coma_zoom_AccretionRelic.pdf}
\caption{Zoom into the region of the accretion relic. Left panel: 20$^{\prime \prime} \,$ image from LoTSS showing the diffuse emission of the accretion relic (top left), the candidate remnant source
(bottom left) and all the radio sources in the field. Right panel: same as left panel but from the 2$^{\prime}$ image. Contours are plotted at (3,4,8,16,32,64)$\sigma_{\rm{rms}}$. The white circle is centered on the Coma cluster centre and has a radius $r=R_{100}$. }
\label{fig:zoomAR}
\end{figure*}
\begin{figure*}
\includegraphics[width=2.1\columnwidth]{Coma_zoom_NGC4849.pdf}
\caption{Zoom into the region of NGC~4849, to the South of Coma. Left panel: 6$^{\prime \prime} \,$ image from LoTSS showing the core of NGC~4849. Right panel: 1$^{'}$ image showing the full extent of the tail. The white box marks the region shown in the left panel. Contours start at 3 $\sigma$ and are spaced by a factor 2.
}
\label{fig:NGC4849}
\end{figure*}
\section{Components of the diffuse emission}
\label{sec:diffuse}
The radio emission from the Coma field consists of several different components, either associated with the ICM or originating from the interaction of the radio galaxies with the environment.
In Fig.~\ref{fig:Coma_fov}, the diffuse emission from the Coma cluster field is shown, after the subtraction of radio galaxies and point sources.
The most relevant features are labelled. As already known, the Coma cluster hosts a radio halo, a radio relic, and a bridge of low surface brightness emission connecting the two. In addition to these components, several new features are detected in our LOFAR observations. In this Section, we present the emission from the diffuse sources as imaged by LOFAR at 144\,MHz. An analysis of their properties that includes radio images at other frequencies and the comparison with the X-ray emission from the ICM is presented in the following sections for each source separately.
\subsection{The radio halo}
At 144\,MHz, the radio halo appears larger than at higher frequencies, with a LAS of 1.2 deg, measured East West, corresponding to $\sim$ 2\,Mpc (see Tab. \ref{tab:sources}). The halo appears to be composed of a central, bright core and a larger, weaker component that is asymmetrical and more pronounced towards the West (see Fig. \ref{fig:zoomHalo}).
The inner portion of the halo, is what we define the ``halo core"\footnote{The halo core and outer halo are labelled in Fig. \ref{fig:corr_core_outer}}.
This part of the halo is the one visible in the LoTSS images which impose an inner \emph{uv}-cut of $80 \lambda$, corresponding to $\sim 43'$. The 20$^{\prime \prime} \,$ resolution of LoTSS, and here reproduced using the same \emph{uv}-range restriction, provides a detailed image of the inner portion of the halo. Clearly, the emission from the halo core is brighter than the rest, and its surface brightness is characterized by bright filaments of radio emission, marked with arrows in Fig.~\ref{fig:zoomHalo}.
Despite being the best studied radio halo, this is the first time that features such as these filaments are detected in its diffuse emission.\\
Outside the halo core, low surface brightness emission is detected, which is brighter towards the West. We call this emission the ``outer halo'' and discuss in Secs.~\ref{sec:halo_fit} and \ref{sec:radioXcorrelation} whether the halo core and the outer halo show different properties, as recently proposed for some cool-core clusters (\citealt{Savini_2018,Savini_2019,Biava21b,Riseley21}).
The emission from the outer halo is only visible when baselines shorter than 80$\lambda$ are included in the image, and hence it is filtered out in the LoTSS images.\\
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{plot_profile_anuli_front.png}
\caption{Radio profile of the Coma halo form the LOFAR image at 2$^{\prime}$ resolution across the halo front in comparison with the rest of the halo. Errorbars represent statistical errors only. }
\label{fig:profile_front}
\end{figure}
In the right panel of Fig.~\ref{fig:zoomHalo}, we show the halo in the LOFAR 35$^{\prime \prime} \,$ image, after the subtraction of the unrelated radio sources.
To the West, the halo has a sharp edge coincident with the halo front (see Fig.~\ref{fig:Coma_fov} for labelling) already found by \citet{BrownRudnick11} and coincident with the shock front detected in the X-rays \citep{Simionescu13} and SZ \citep{PlanckComa}.
In Fig. \ref{fig:profile_front} we show the radio profile computed in annuli across the halo front in comparison with the rest of the cluster. The two sectors have azimuth angles from 47$^{\circ}$ to 287$^{\circ}$ (Coma) and from 323$^{\circ}$ to 48$^{\circ}$ (Front, centered in RA=12:59:07, DEC=+28:01:31). The Southwest part of the halo, towards the radio bridge, has been excluded. While the average cluster profile shows a smooth radial decline, a sharp edge is visible around $\sim$1900\arcsec from the cluster centre, coincident with the radio front found by \cite{BrownRudnick11}.\\
Towards the southwest, the spherical front continues and its brightness becomes weaker and merges with the emission from the radio bridge. The total flux density of the radio halo, measured from the images above 2$\sigma_{\rm rms}$ is 10$\pm$2 Jy at 144\,MHz. The main properties of the radio halo are listed in Tab.~\ref{tab:sources}.
\subsection{The radio relic and the NAT-relic connection}
To the South-West of the Coma cluster, a radio relic has been discovered by \citet{Ballarati81} and studied by several authors afterwards \citep{Venturi90,Giovannini93,BrownRudnick11, OgreanComa, Bonafede13}. However, the presence of the bright source ComaA, at the north-west of the relic, has always made its study difficult. Residual calibration errors remain in the LOFAR image (see Fig.~\ref{fig:zoomRelic}) and make it difficult to determine whether the relic extension to the NW is real. We note that the same extension has also been detected by single dish observations at 1.4 GHz, where it also appears polarized \citep{BrownRudnick11}. Despite this, because of its questionable nature, we adopt a conservative approach and do not include it in our discussion of the relic.\\
The radio relic has a LAS of 38$'$, corresponding to $\sim$ 1.1\,Mpc at the Coma redshift.
Its emission is connected on both sides to two head-tail radio galaxies, namely NGC~4839 to the North-East, and NGC~4789 to the South-West of the relic. From NGC4839, we detect diffuse emission that blends into the bridge and relic emission \citep{Bonafede21}. The surface brightness of the relic is not uniform, but composed of patchy and filamentary sub-structures (see Fig.~\ref{fig:zoomRelic}). Stripes of radio emission depart from the relic and are directed towards the bridge and NGC 4839.
A similar stripe is detected also near the tail of NGC 4839, directed towards the relic. These features could be regions where the magnetic field has been amplified and/or where the plasma has been stripped from the tail of NGC~4789.
The surface brightness of these stripes is a factor 2-3 higher than the nearby emission, and their size is $\sim$ 11.7$'$ that corresponds to 300 kpc at the Coma's redshift.\\
Beyond the relic, towards the SW, the narrow angle tail galaxy (NAT) NGC 4789 appears connected to the relic, as already found by e.g. \citealt{Giovannini91}. The connection between the relic and NGC 4789 has been interpreted as the tail of NGC 4789 feeding the relic with radio plasma \citep[e.g.][]{EnsslinGK01}. In Sec.~\ref{sec:relicNAT}, we will investigate the possible origin of this emission. \\
\subsection{Diffuse emission from NGC4849}
To the South of Coma, we detect diffuse emission, elongated in the North-South direction, with a size of $\sim$27$^{\prime}$. This emission seems associated with the radiogalaxy NGC~4839, located at RA$\rm{=12h58m12.679s}$, DEC$\rm{=+26d23m48.77s}$, at redshift $z=0.01966$. Its angular size translates into a linear size of $\sim$650 kpc (see Fig. \ref{fig:NGC4849}). From the 6$^{\prime \prime} \,$ resolution image, the core of NGC~4849 is visible, and a hint of jet emission in the North-South direction is also present. A second bright component is located at the South of the core, which could either be an unrelated radio source seen in projection, or the lobes of NGC~4849. The maximum extension of this tail in the WSRT image is $\sim$7$^{\prime}$. Using the WSRT H and ``LOFAR as WSRT H" images (see Sec. \ref{sec:data} and Tab. \ref{tab:images}) and taking into account the higher noise of the WSRT image in that region, we derive a 2-$\sigma$ limit for the spectral index of the tail $\alpha < -2$. Analysing the emission of this source is not the aim of this paper, we note that similar steep spectrum tails have been found in the outskirts of other clusters (e.g. Abell~1132 \citealt{Wilber17}, Abell~2255 \citealt{Botteon20}) likely tracing the motion of the galaxy in the ICM. The long tail of aged cosmic ray electrons (CRe) left behind by these sources during their motion provides seed electrons that could be re-accelerated by turbulence and shocks.
\subsection{Accretion relic}
After the subtraction of unrelated sources, two extended patches of diffuse emission are visible at the periphery of Coma, to its North-East and East (see Fig.~\ref{fig:zoomAR}). The diffuse component at the Eastern side of Coma is not associated to any Coma cluster galaxy \citep{PizzoTesi}. \citet{PizzoTesi}
also noticed that this source lies at the crossroad of two filaments of galaxies, pointing towards the clusters A~2197 and A~2199, respectively, and suggest that this emission could be due to the accretion of matter towards the Coma cluster. Our images have a higher resolution and allow us to recognize a double-lobe structure that resembles a radiogalaxy. However, no core is obviously associated with it in the 6$^{\prime \prime} \,$ and 20$^{\prime \prime} \,$ images (see Fig.~\ref{fig:zoomAR}). We tentatively label this source as a ``candidate remnant", but we note that its classification remains uncertain. Indeed, given its large angular size ($\sim$ 27$^{'}$) and the lack of an optical identification at the Coma cluster redshift, it would have a very large linear size if it was in the background of Coma. \\
The other diffuse patch in the NE part of Coma is entirely new and has an arc-like morphology, and a weak uniform brightness. This emission does not appear connected to any discrete source (see Fig.~\ref{fig:zoomAR}).
When convolving the image to a resolution of 2 arcmin, it has a largest angular size of $\sim$ 55', that would correspond to 1.5\,Mpc at the cluster redshift, and it is located at $\sim \, 2.1^{\circ}$ from the cluster centre.
At the Coma redshift, this distance corresponds to $\sim$3.6 Mpc, while the cluster virial radius\footnote{$R_{100}$ is considered here to be a good approximation of the virial radius, as the cluster redshift is 0.023} is $\sim 2.9$ Mpc.
Since recent studies \citep[e.g.,][]{Malavasi20} have found an intergalactic filament of galaxies in the NW direction from the Coma cluster, this source could be connected to the cluster and its large-scale environment. Hence, we consider it plausible that this is an ``accretion relic", though more data supporting our hypothesis are needed.
Considering that accretion shocks on these scales should be characterized by a strong Mach number ($\mathcal{M} \gg 5$, e.g. \citealt{Hong14}), a strong prediction in this case would be that the radio spectrum of this emission should be $\alpha \propto \nu^{-1}$, and also characterized by a large degree of polarisation.
The properties of this accretion relic are listed in Table \ref{tab:sources}.
\begin{table*}
\centering
\caption{Images used in this work}
\begin{tabular}{l c c c c}\hline
Image name & Freq & Resolution & $\sigma_{\rm rms}$ & Fig. \\
& MHz & & mJy/beam & \\
LOFAR as LoTSS 6 & 144 & 6 $^{\prime \prime} \,$ $\times$ 6 $^{\prime \prime} \,$ & 0.1 & Fig. \ref{fig:zoomRelic} \\
LOFAR as LoTSS 20 & 144 & 20 $^{\prime \prime} \,$ $\times$ 20 $^{\prime \prime} \,$ & 0.15 & Fig. \ref{fig:zoomHalo} \\
LOFAR 35$^{\prime \prime} \,$ & 144 & 35 $^{\prime \prime} \,$ $\times$ 35 $^{\prime \prime} \,$ & 0.2 & Fig. \ref{fig:zoomHalo} (top panel), Fig. \ref{fig:zoomRelic}. \\
LOFAR 1$'$ & 144 & 60 $^{\prime \prime} \,$ $\times$ 60 $^{\prime \prime} \,$ & 0.4 & Fig. \ref{fig:Coma_fov}\\
&&&& \\
WSRT H & 342 & 134$^{\prime \prime} \,$ $\times$ 68 $^{\prime \prime} \,$ & 0.4 & \citealt{BrownRudnick11} \\
LOFAR as WSRT H & 144 & 134$^{\prime \prime} \,$ $\times$ 68 $^{\prime \prime} \,$ & 1.5 & Fig. \ref{fig:spix_front} \\
WSRT R & 326 & 150$^{\prime \prime} \,$ $\times$ 100 $^{\prime \prime} \,$ & 1.2 & \citealt{Giovannini91,Bonafede21}\\
LOFAR as WSRT R & 144 & 150$^{\prime \prime} \,$ $\times$ 100 $^{\prime \prime} \,$ & 1 & \\
\hline
\end{tabular}
\label{tab:images}
\end{table*}
\begin{table*}
\centering
\caption{Source properties at 144 MHz}
\begin{tabular}{l c c c c}
\hline
Source name & dist & LAS & $S_{\rm{144 \, MHz}}$ & $P_{\rm{144 \, MHz}}$ \\
& & & Jy & W/Hz \\
Halo & - & 71$'$ - 2.00 Mpc & 12$\pm$2 & 1.5 $\pm 0.2 \times 10^{25}$ \\
Accretion relic & 2.1$^{\circ}$ - 3.55 Mpc & 55$'$ - 1.5 Mpc & 0.47 $\pm$ 0.07 & 5.7$\pm 0.9 \times 10^{23}$\\
Relic & 73$'$ - 2.0 Mpc & 38$'$ - 1.1 Mpc & 2.4 $\pm$ 0.4 & 3.0$\pm 0.4 \times 10^{24}$ \\
Relic-NAT connection & 79$'$ - 2.3 Mpc& 10$'$ - 280 kpc & 0.7$\pm$0.1 & 9$\pm 1 \times 10^{23}$ \\
\hline
\multicolumn{5}{l}{Col. 1: Name of the diffuse source; Col 2: Distance from the cluster centre; Col 3: Largest angular size measured }\\
\multicolumn{5}{l}{above the 2$\sigma_{mrs}$ contour Col. 4: FLux density measured above the 2$\sigma_{\rm rms}$ contour; Col 5: Radio power at 144 MHz}\\
\multicolumn{5}{l}{The K-correction is applied assuming $\alpha=-1, -1, -1.2, -1.4$}\\
\multicolumn{5}{l}{for the halo, accretion relic, relic, relic-NAT connection, respectively.}\\
\end{tabular}
\label{tab:sources}
\end{table*}
\section{The radio halo profile}
\label{sec:halo_fit}
To characterize the halo properties, we have fitted its surface brightness profile adopting the approach firstly proposed by \citet{Murgia09} and more recently generalized by \citet{Boxelaar_2021} to account for asymmetric halo shapes.
The main novelty of this procedure is that the profiles are fitted to a two-dimensional image directly, using MCMC to explore the parameter space, rather than to a radially averaged profile. In addition, the fitting procedure by \citet{Boxelaar_2021} allows one to fit also elliptical and skewed (asymmetric) models.
We refer to \citet{Boxelaar_2021} for a detailed explanation and summarize here the relevant parameters.
The surface brightness model is given by:
\begin{equation}
I(r) = I_0 \exp^{-G(r)} ,
\label{eq:exp}
\end{equation}
where $I_0$ is the central surface brightness and $G(r)$ a radial function. $G(r)= \left( \frac{|r^2|}{r_e^2}\right)^{0.5}$ for the circular model, while $G(r)= \left( \frac{x^2}{r_1^2} + \frac{y^2}{r_2^2}\right)^{0.5}$ for the elliptical models, where $r_e$ is the characteristic e-folding radius, and $r^2=x^2 + y^2$. The skewed model allows for an off-center maximum of the brightness distribution, and is characterized by four different scale radii ($r_1$, $r_2$, $r_3$, $r_4$) and by an angle to describe the asymmetric brightness distribution.\\
We have investigated these three profiles (circular, elliptical, and skewed), and results of the fit are given in Table \ref{tab:fit_halo}.
In this table, we only list the statistical error, while the systematic one is 15\% of the listed flux density, and due to the uncertainty on the absolute flux scale.
All models give a consistent total radio power ($P_{144 \rm{MHz}} \sim 1.47 \times 10^{25} \, \rm{W \, Hz^{-1}}$ computed within 3 times the e-folding radius) and central radio brightness ($I_0 \sim 5 \, \mu \rm{Jy/arcsec^{2}}$). We note that the flux density and consequently the power measured by the fit are in perfect agreement with the estimate derived from the images above 2$\sigma_{\rm rms}$ (see Sec. \ref{sec:diffuse}).
As the reduced-$\chi^2$ values are similar in the skewed and elliptical model, we consider the latter in the following analysis.\\
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{plot_radio_6re_noSectors_chWidth.png}
\includegraphics[width=\columnwidth]{plot_profile_exp_anuli_panel2.png}
\caption{Left panel: bullets are the radio average brightness profile of the radio halo, computed within elliptical annuli having major and minor axis sub-multiples and multiples of $r_1$ and $r_2$, i.e. the minor and major axis of the elliptical exponential fit (see Tab. \ref{tab:fit_halo}). The SW sector of the halo has been blanked for ellipses with major and minor axis larger than 3$r_1$ and 3$r_2$, respectively, to exclude the bridge region.
Errorbars represent the error on the mean, upper limits at 3$\sigma $ are plotted as arrows. The width of the annuli goes from 1$^{\prime}$ in the center to 4$^{\prime}$ in the outer annuli, to improve the sensitivity to low surface brightness emission in the outer parts of the halo.
The blue line refers to the best fit elliptical model, see Tab. \ref{tab:fit_halo}.
Right panel: radio image of the Coma cluster at 1$^{\prime}$ resolution. The inner cyan ellipse has major and minor axis equal to 3$r_{1}$ and 3$r_{2}$ respectively, and elliptical annuli are spaced by 1$^{\prime}$. The outer elliptical bins have a width of 4$^{\prime}$ and and trace ellipses out to 6$r_{1}$ and 6$r_{2}$, showing the SW region that has been excluded from the analysis because of the bridge. }
\label{fig:halo_profile}
\end{figure*}
In Fig. \ref{fig:halo_profile}, we show the radial profile of the halo brightness at 144 MHz. We have computed the mean of the radio brightness and its error within elliptical annuli having a width than changes progressively from 1$^{'}$ in the centre to 4$^{'}$ in the outer regions, to maximise the resolution at the centre and the sensitivity to low surface-brightness emission in the halo peripheral regions. The bridge region s excluded starting from elliptical annuli with major and minor axis larger than 3$r_1$ and 3$r_2$, respectively.
We considered upper limits the values where we have a mean smaller than 3 times the rms noise. From this plot, we derive that we have detected the halo emission up to $r\sim 1.3$ Mpc, that corresponds to $\sim$ 4.5$r_1$ and $r_2$ (see Tab. \ref{tab:fit_halo}).\\
In the following, we refer to the halo core as the emission contained within an ellipse having as major and minor axis $r_1$ and $r_2$, respectively, and to the
outer halo as the emission contained within 4$r_1$ and 4$r_2$, excluding the halo core. We note that the halo core and outer halo are well represented by a single exponential model and do not need to be considered as two separate components, if we consider the average surface brightness of the halo emission. However, as we show in the following analysis, they are characterized by different properties.\\
We applied the same fitting procedure to the WSRT image (WSRT H in Tab. \ref{tab:images}), after blanking the sources that were contaminating the emission. We only attempted an elliptical and circular model fit.
Results are listed in the bottom part of Table \ref{tab:fit_halo}. Both fits give the same total flux density and radio power, and have the same $\chi^2_r$ value, smaller than 1, possibly indicating that the errors are being overestimated.
We note that the $r_1$ and $r_2$ values are slightly smaller than those found in the LOFAR image, indicating a more peaked profile of the radio emission, hence of the emitting CRe at higher frequencies.
\begin{table*}
\caption{Radio halo 2D fit}
\begin{tabular}{lccccccccc}
\hline
\hline
\multicolumn{10}{l}{LOFAR - 144 MHz}\\
\hline
Halo model & $\chi_r^2$ & $I_0$ & $r_1$ & $r_2$& $r_3$& $r_4$ & angle & S$_ {144, \rm{MHz}}$ & $P_{144, \rm{MHz}}$\\
& &$\mu \rm{Jy}/\rm{arcsec}^2$ & kpc&kpc&kpc&kpc& deg & Jy & $\rm{ 10^{25} W \, Hz^{-1}}$ \\
Circular & 1.7 & 5.42 $\pm$ 0.04 & 310 $\pm$ 1 &&&&& 12.20 $\pm$ 0.04 & 1.470 $\pm$ 0.002 \\
Elliptical & 1.6 & 5.50$\pm$ 0.02 & 355 $\pm$ 1 & 268 $\pm$ 1 & & & & 12.21 $\pm$ 0.04 &1.470 $\pm$ 0.005 \\
Skewed & 1.6 & 5.46 $\pm$ 0.01 & 337$\pm$2 & 368$\pm$2 & 207$\pm$2 & 342$\pm$ 2 & 2.99$\pm$ 0.001 & 12.35 $\pm$ 0.05 & 1.490$\pm$ 0.05 \\
\hline\hline
\multicolumn{10}{l}{WSRT - 342 MHz}\\
\hline
Halo model & $\chi_r^2$ & $I_0$ & $r_1$ & $r_2$& $r_3$& $r_4$ & angle & S$_ {342, \rm{MHz}}$ & $P_{342, \rm{MHz}}$\\
& &$\mu \rm{Jy}/\rm{arcsec}^2$ & kpc&kpc&kpc&kpc& deg & Jy & $\rm{ 10^{25} W \, Hz^{-1}}$ \\
Circular &0.7 & 3.52$\pm$0.08 & 255$\pm$4 &&&&& 5.3 $\pm$ 0.1 & 0.64 $\pm$ 0.01 \\
Elliptical & 0.7 & 3.54$\pm$0.08 & 268$\pm$5 & 240$\pm$6 & & & & 5.3$\pm$ 0.1 & 0.64 $\pm$ 0.01 \\
\hline\hline
\multicolumn{10}{l}{Col 1: Model used; Col 2: Reduced $\chi^2$ value; Col 3: Central brightness of the fit; Col 4 -7: e-folding radii; }\\
\multicolumn{10}{l}{Col 8: Angle for the skewed halo model fit; Col 9: Total halo power at 144 MHz computed within 3$r_e$.}\\
\multicolumn{10}{l}{Only statistical fit errors are shown in the table}
\end{tabular}
\label{tab:fit_halo}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{Spectrum_halo.png}
\includegraphics[width=1.01\columnwidth]{spix_halo_sectors.png}
\caption{Left: Integrated spectrum of the radio halo from literature data (green points) from \cite{Thierbach03} and ref. therein, as corrected and re-scaled by \cite{Brunetti13}. The new LOFAR measurement (red point) and the WSRT measurement (orange box) are shown, measured within 3 times the effective radii as derived by the halo fit (see text for details). The continuous line shows expectations from a model where secondary particles are reaccelerated by compressive turbulence \citep{BL11}. The dashed line shows the spectrum computed from secondary emission, after 200 Myrs since turbulent re-acceleration has been switched off \citep{BL11}. Right: spectral index radial profile computed in the different sectors, as listed in the legend. Arrows are 3$\sigma$ upper limits, only statistical errors are shown, the values of the spectral index are also affected by the flux calibration uncertainties of WSRT and LOFAR (10\% and 15\%, respectively) that would contribute with an additional error of 0.2.}
\label{fig:haloSpectrum}
\end{figure*}
\section{Spectral properties of the radio halo}
\label{sec:haloSpectrum}
\subsection{The integrated spectrum}
Using literature results, and the halo flux density at 144 and 342\,MHz, we can constrain the low-frequency part of the radio halo spectrum. The spectrum of radio halos provides important information about the underlying re-acceleration mechanism. Indeed, it allowed in the past to conclude that a very steep spectrum ($\alpha < -1.5$, also called ultra-steep spectrum halos) cannot be produced by hadronic models \citep[e.g.,][]{Brunetti08}.\\
Using the flux densities that result from the fits, we can measure the spectral index between 144\,MHz and 342\,MHz yielding $\alpha = - 1.0 \pm 0.2$. Though residuals from sources could still be present, the fitting procedure we have used should minimise that contribution \citep{Boxelaar_2021}, hence, a considerable impact to the whole halo emission is unlikely. We find that the spectrum is slightly flatter than previously reported in the literature, although still consistent within the errors.
The spectrum of the halo in the Coma cluster has been extensively studied in the literature \citep[e.g.][]{Giovannini93, Thierbach03}, using both interferometric images and single-dish data. At low frequencies, the contribution of radio galaxies is difficult to account for, for mainly two reasons: (i) the low resolution of observations published so far, (ii) the larger extent of tailed radio galaxies whose emission blends with the halo emission. Our sensitive and high-resolution LOFAR images allow us to alleviate both of these problems. In addition, we now have a more accurate method to estimate the halo flux density.
We have collected the data available in the literature, i.e., those presented in \citet{Thierbach03}, and re-scaled to the same absolute flux scale as in \citet{Brunetti13},
and added our measurement at 144\,MHz. In Fig.~\ref{fig:haloSpectrum} we show the spectrum of the halo.
Green bullets refer to values taken from \citet{Brunetti13}, and do not refer to the same aperture radius.
The red and orange point refer to the LOFAR and WSRT map computed within the same aperture (corresponding to 3 e-folding radii of the LOFAR image, see Tab. \ref{tab:fit_halo}).
In Fig. \ref{fig:haloSpectrum}, left panel, we show one possible spectrum that would be produced by re-acceleration models.
However, for a proper derivation of the halo spectral properties, one should compare the flux densities from images done with the same \emph{uv}-range, the same procedure for compact source subtraction, and using the same aperture radius. We also note that using a fitting algorithm, as the one proposed by \cite{Boxelaar_2021} and used here, would also minimise the effect of different noise obtained at different frequencies.\\
Recently, \cite{Rajpurohit2022} have shown that the hint for a spectral break claimed in the radio relic of Abell 2256 below 1.4\,GHz \citep{Trasatti15} is not confirmed once the analysis is performed with matching \emph{uv}-coverage and unrelated sources are subtracted properly. In the case of the Coma halo, we note that the high-frequency measurements are almost a factor 10 below the value extrapolated from a power-law at low frequencies. However, the exact shape of the spectrum could be affected by the effects mentioned above. In addition, it is important to get a precise flux density at frequencies below 100\,MHz to characterize the integrated radio spectrum. Here, data from the LOFAR Low-Band Antennas (LBA) will provide powerful constraints.
\subsection{Spectral index profile}
Using WSRT and LOFAR HBA data, we can obtain a radial profile of the halo spectral index. A steepening of the radial profile has been found by \citet{Giovannini93} using data between 326 MHz and 1.38 GHz. They found that the radio halo has a smooth spectrum in the central regions (inner 480 $^{\prime \prime} \,$ radius) with $\alpha \sim -0.8$, and a steeper value ($\alpha \sim -1.2$ down to $-1.8$ in the outer regions).\\
Using WSRT and LOFAR HBA images, we can now compute the spectral index profile out to larger distances.
We have re-imaged the LOFAR data using only baselines larger than 40m, i.e. the shortest WSRT baseline, and convolved the LOFAR image to the same resolution as the WSRT image. The sources embedded in the diffuse emission are subtracted from the WSRT image, however, as some residuals were still present, we applied the multi-filtering technique described in \citet{Rudnick2002} and blanked the WSRT image wherever the filtered image was above 3 mJy/beam. We blanked the LOFAR image accordingly.
We have divided the radio halo into 4 sectors (NW, NE, SE, and SW) and computed the mean spectral index $\alpha$ in elliptical annuli having the major and minor axis proportional to $r_1$ and $r_2$, up to the maximum distance where the halo is detected (see Sec. \ref{sec:halo_fit} and Fig. \ref{fig:halo_profile}). \\
The 4 outermost annuli of the SW sector have been removed to exclude the bridge region. We have computed the spectral index in each annulus, and considered upper limits the annuli that have a total flux density smaller than 3$\sigma_{\rm rms} \times \sqrt(N_{\rm{Beams}}) $, with $N_{\rm{beams}}$ being the number of independent beams sampled in each annulus. As the radio halo is more extended in the LOFAR image than in the WSRT image, we mainly derive upper limits at distances larger than $\sim$ 2000$^{\prime \prime} \,$ (950 kpc).
The radial trend of the spectral index is shown in Fig. \ref{fig:haloSpectrum}, right panel. We note that values in some annuli are surprisingly flat, possibly because of residual contamination from unrelated sources. The increasing values of the upper limit in the outer annulli are due to the larger area sampled by the annuli.
All the sectors show a spectral index that becomes steeper at distances of $\sim$1500$^{\prime \prime} \,$ from the cluster centre.
In the SW sector, we also clearly detect a steepening towards the cluster centre. A similar, though less pronounced, trend is also observed in the other sectors, though we note that in the NE sector the radial profile of $\alpha$ is more complex, and no clear trend at distances smaller than $r \sim 2000$$^{\prime \prime} \,$ can be established. The SW sector is affected by the passage of the NGC~4839 group, hence is it possible that different physical conditions are present there. Overall, we can cocnlude that the spectral trend is characterized by clear steepening at the cluster outskirts, and a mild steepening towards the cluster centre. \\
In the next Sec., we will discuss the physical implications of these results in the framework of turbulent re-acceleration models.
\subsection{Radial variations of the spectral index and re-acceleration models}
In the presence of a break in the integrated spectrum of radio halos, homogeneous re-acceleration models predict that increasingly steeper spectra will be seen at increasing distance from the cluster center \citep{Brunetti01}.
Assuming homogeneous conditions, the frequency at which steepening occurs, $\nu_s$, is proportional to:
\begin{equation}
\nu_s \propto \tau_{acc}^{-2} \frac{B}{ (B^2+ B_{IC}^2)^2}
\end{equation}
$\tau_{acc}$ is the re-acceleration time (see \citealt{CassanoBrunetti05}, \citealt{BrunettiLazarian07}) that depends on the assumed turbulent properties and re-acceleration mechanism. $B_{IC}$ is the inverse Compton equivalent magnetic field.
In the case of a constant $\tau_{acc}$, the steepening frequency depends only on the magnetic field strength.
To better follow the discussion below, let us define a critical magnetic field value
$$
B_{cr}=\frac{B_{IC}}{ \sqrt{3}} \sim 2 \mu G.
$$
As long as $B<B_{cr}$, one expects to see a radial steepening of the spectral index at distances larger and larger from the cluster centre as we move towards lower observing frequencies. This is what we observe beyond a radius of $\sim 30 ^{\prime}$. \citet{Giovannini93} also detected a steepening between the higher frequency pair 325~MHz and 1.38~GHz, beyond a radius of $\sim 8^{\prime}$ from the cluster centre. While we do expect more dramatic steepening at the higher frequencies, whether these two sets of measurements are consistent with a single physical model requires further investigation.\\
Moreover, we detect for the first time a steepening of the spectral index towards the cluster centre. This can also be explained in the framework of homogeneous re-acceleration models, if in the cluster centre we have $B>B_{cr}$. Assuming a magnetic field profile as derived from RM studies ( $B\propto B_0 n_e^{0.5)}$ , \citealt{Bonafede10}), we have indeed a central magnetic field of $B_0 \sim 5 \mu$G, hence $B>B_{cr}$.
We note that a similar steepening towards the cluster centre should be visible also in the higher frequency spectral index map by \citet{Giovannini93}. However, in that work the authors only report a spectral index trend through a line passing from the cluster from SE to NW. It is possible that a radial analysis similar to the one we present here would show the same trend. Alternatively, one should think of ad-hoc re-acceleration conditions that make this steepening visible only at low frequency. Future observations at higher ferquencies could shed light on this point.\\
We can conclude that the data presented in this work, together with literature data by \citet{Giovannini93} and \citet{Bonafede10} provide a coherent picture with the expectations from homogeneous turbulent re-acceleration models \citep{Brunetti01}, though the spectral index trend in the cluster centre leaves some open questions that could be addressed by future observations.\\
\subsubsection{Towards a constrain of the re-acceleration model parameters}
\label{sec:haloSpectrum_theo}
Constraining the model parameters, such as $\tau_{acc}$, would require a detailed 3D modeling and a precise constrain on the radial position of $\nu_s$. However, we can try to make first order calculations to see whether the qualitative coherent picture outlined above is quantitatively supported.
\citet{Giovannini93} have detected a spectral steepening at $\sim$8$^{\prime}$ between 326 MHz and 1.38 GHz.
Assuming a central magnetic field $B_0=4.7 \mu$G, one would expect to detect the steepening between 144 and 342 MHz where the magnetic field is $B\sim 0.7 \mu$G, i.e. at $r \sim 2$ Mpc ($\sim 71^{\prime}$) from the cluster centre. Instead, the steepening is detected at $r\sim$ 30$^{\prime}$ from the cluster centre, i.e. a factor 2 closer to the expected location
This may suggest that $\tau_{acc}$ is not constant and increasing with the distance from the cluster centre.
Though this result is not surprising from a theoretical point of view, it would be the first time that data support this claim.
Instead of a constant $\tau_{acc}$, we can make a step further and assume a constant turbulent Mach number $\mathcal{M}_t$. In this case, the steepening frequency is proportional to:
\begin{equation}
\nu_s \propto T^{2a} \frac{B}{ (B^2+ B_{ICM}^2)^2}
\end{equation}
where T is the cluster temperature and $a$ is a constant that is $a=1$ for
re-acceleration via Transit-Time-Damping mechanism with compressive turbulence, and $a=1.5$ for non-resonant second-order acceleration with solenoidal turbulence\footnote{in this case we consider also a constant Alfv\'en velocity, based on the scaling $B^2 \propto n$.} \citep{BrunettiLazarian16}.
In order to explain the spectral index steepening at $r\sim 30^{\prime}$ from the cluster centre, the temperature should be $\sim$35-45\% lower at a distance of 30$^{\prime'}$ than at a distance of 8$^{\prime}$, where the steepening is detected at higher frequency.
We note that this temperature drop is consistent with the temperature profile found by \citet{Simionescu13}\footnote{we take as a reference the profile extracted along the E sector of their analysis, which is the only one not contaminated by the W shock and by the merger with the NGC4839 group.}.
We stress again that the calculations we have just performed do not allow us to make any claim, as long as $\nu_s$ is not precisely determined from data and projection effects are not taken into account. However, they show that the qualitative picture outlined above is not in at odds with first order quantitative estimates.\\
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{corr5arcmin.pdf}
\caption{$I_R - I_X$ correlation computed from the 6$'$ image using cells equally spaced by 6$'$. Arrows mark the 2$\sigma_{\rm rms}$ upper limits, the black continuous line shows the best-fit line, dotted lines show the 10\% and 90\% slopes for the posterior distribution of $\beta$. The magenta dotted horizontal line marks the 1$\sigma_{\rm rms}$. All data are computed within 2400$^{\prime \prime} \,$ from the cluster centre. Here and in the following correlation plots, errorbars are plotted every second point for clarity reason.}
\label{fig:corr_trend}
\end{figure}
\begin{table}
\caption{Radio - X-ray correlation for different Gaussian smoothing lengths.}
\begin{tabular}{c c c c}
\hline\hline
Gaussian Beam & $\beta$ & 10\% -- 90\% & $\rho_{P}$ \\
FWHM & && \\
1$^{\prime}$ &0.64 & 0.65 -- 0.63 & 0.86 \\
2 $^{\prime}$ & 0.65 & 0.66 -- 0.63 & 0.88 \\
3$^{\prime}$ & 0.68 & 0.70 -- 0.66 & 0.89 \\
4 $^{\prime}$ & 0.70 & 0.73 -- 0.68 & 0.88 \\
5 $^{\prime}$ & 0.74 & 0.78 -- 0.70 & 0.89 \\
6 $^{\prime}$ & 0.76 & 0.81 -- 0.72 & 0.88 \\
\hline\hline
\multicolumn{4}{l}{Col 1: FWHM of the smoothing Gaussian or restoring beam; }\\
\multicolumn{4}{l}{Col 2: Best-fit slope; Col 3: 10th and 90th percentile of the }\\
\multicolumn{4}{l}{posterior distribution for $\beta$. Col 4: Pearson correlation}\\
\multicolumn{4}{l}{coefficient}\\
\end{tabular}
\label{tab:radioX}
\end{table}
\section{Thermal and non-thermal correlations in the radio halo }
\label{sec:radioXcorrelation}
\subsection{Point-to-point analysis}
Investigating the point-to-point correlation between the radio and the X-ray surface brightness can give important information about the relation between the thermal and non-thermal components of the ICM. Also, it has the potential of constraining the mechanism responsible for the radio emission. \citet{Govoni01} have first investigated this correlation for a small sample of radio halos, finding a sub-linear scaling of the radio brightness with respect to the X-ray brightness. More recently, \citet{Botteon20, Rajpurohit21, Rajpurohit21a, Ignesti20} have investigated the same correlation for radio halos and mini halos, respectively, finding that halos tend to have a sub-linear or linear scaling, and mini halos show linear or super-linear behaviours.\\
Since the Coma radio halo is the one for which most spatially resolved and multi-wavelength data are available, it is important to establish the statistical relation between its thermal and non-thermal components.\\
We have investigated the thermal to non-thermal correlation for the Coma cluster, fitting the radio ($I_R$) and X-ray ($I_X$) surface brightness in log-log space, according to:
\begin{equation}
\log I_R= \beta \log I_X + \gamma
\end{equation}
where $\beta$ is the correlation slope.
We have used a hierarchical Bayesian model \citep{Kelly07}, that allows us to perform linear regression of $I_R$ on $I_X$ accounting for
intrinsic scatter in the regression relationship, possibly correlated measurement errors, and selection effects (e.g., Malmquist bias). Using this method, we have derived a likelihood function for the data. We consider the mean of the posterior distribution as the best-fit slope. Following \cite{Botteon20,Bonafede21}, we considered as upper limits the radio values that are below 2$\sigma_{\rm rms}$.\\
Despite the fact that we detect the radio halo up to a distance of $\sim$1.3 Mpc, the analysis of the radio-X ray correlation is limited by the extent of the \emph{XMM-Newton} mosaic. In particular, because of soft-proton contamination, the analysis is restricted to a distance of 2400$^{\prime \prime} \,$ ($\sim$ 1.1 Mpc) from the cluster centre.
To gain sensitivity towards the low surface brightness emission of the halo outskirts, we have convolved the radio image with Gaussian beams having FWHM of 1,2,3,4,5, and 6 arcminutes. We have computed the mean of the radio and X-ray brightness in square boxes having an area equal to a Gaussian beam of the radio image, and computed the fit using X-ray images smoothed at the same resolution. We have considered upper limits the values that are below 2$\sigma_{\rm rms}$, i.e. twice the noise of the radio image.
The results of the fits are listed in Tab. \ref{tab:radioX}.
The image at 6$'$ resolution allows to recover the outermost regions of the halo keeping the highest possible resolution.
$I_X$ and $I_R$ are positively correlated with a slope $\beta=0.76^{+0.05}_{-0.04}$ and a Pearson correlation coefficient $\rho_p=0.89$.
We note that the finest grid (boxes spaced by 1$^{\prime}$) recovers a slope similar to the one initially found by \citet{Govoni01}. The slope increases as we gain sensitivity to the low surface brightness emission that characterizes the outermost regions of the radio halo. This is suggesting that the slope of the correlation is not constant throughout the radio halo, with the slope $\beta$ increasing when low surface brightness emission is added. \\
\subsection{Correlations in the halo core and outer halo}
Our analysis suggests that the point-to-point $I_R-I_X$ correlation may be different in the halo central regions and in its outskirts.\\
Recent low-frequency observations have found radio emission in galaxy clusters that can be interpreted as the coexistence of a mini halo in the cluster core and a giant halo on larger scales (\citealt{Savini_2019}, \citealt{Biava21b}).
A different $I_R - I_X$ trend is found in the core and in the outer part of the halo of these clusters, with a super-linear scaling in the mini halo region and a sub-linear scaling in the outer part (\citealt{Biava21b}, Lusetti et al., in prep).\\
Although those clusters have a cool-core and a radio halo with a steep spectrum, we note that a change in the $I_R$ - $I_X$ correlation slope for radio halos has never been explicitly investigated in the literature.
In the cluster Abell 2142, a radio halo with two components has been found \citep{Venturi17}, yet no analysis of the radio and X-ray correlation has been performed so far. Also, in the cluster Abell 2744, a multi-component halo has been discovered with a different radio-X ray correlation slope for the northern and southern components \citep{Kamlesh2744}.\\
Since a point-to-point analysis of radio and X-ray surface brightness has been so far presented only for a few clusters, we investigate whether or not the different trends observed in the core and in the outer part of the Coma halo can be a common property of radio halos.
We have repeated the analysis described in Sec.~\ref{sec:radioXcorrelation} considering the ``halo core" and the ``outer halo" separately (see Fig.~\ref{fig:corr_core_outer}).
For the halo core, we have used the finest grid of 1$^{\prime}$ cell size, while for the outer halo we have used the 6$^{\prime}$ grid, to recover the faintest emission.\\
We find that the slope of the ``halo core'' is $\beta=0.41^{+0.04}$, while for the ``outer halo"
we find $\beta=0.76 \pm {+0.05}$.\\
Hence, we can conclude that in the ``halo core" we find a flatter slope for the $I_R - I_X$ correlation, with respect to the outer part, which is the opposite trend found in cool-core clusters that host a mini halo and a halo-type component. (e.g. RXC~J1720.1$+$2638, \citealt{Biava21b}; Abell 1413, Lusetti et al, in prep). We note that this is the first time that a change in the $I_R-I_X$ correlation has been investigated, hence it could be a common property of radio halos.
This result indicates that the inner part of the Coma radio halo has different properties than mini halos observed in more relaxed clusters, likely indicating different local plasma conditions. In particular, the detection of a sub-linear trend in the brightest part of the halo hints at a negligible contribution to the halo central emission from the hadronic mechanism.
\begin{figure*}
\centering
\includegraphics[width=1.5\columnwidth]{Coma_halo_coreOuter_grid.pdf}
\caption{Left panel: Colors are the X-ray emission from the Coma field from XMM-Newton observations. Contours show the radio emission at 1$^{'}$ resolution, starting at 3$\sigma_{\rm rms}$ and increasing by a factor 2. The major axes of the ellipses are $r_1$ and $r_2$ (inner ellipse) and $3 r_1$ and 3 $r_2$ (outer ellipse). $r_1$ and $r_2$ are listed in Tab. \ref{tab:fit_halo}. The inner ellipse is the region considered for the ``halo core". The cyan dotted circle mark the region that is not contaminated by soft X-ray protons.}
\label{fig:corr_core_outer}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{plot_radioX_radial_sectors_chWidth.png}
\caption{Radial profile of the X-ray brightness and radio brightness
in the NW, NE, SE sectors (top to bottom). The profiles are computed in elliptical annuli with different width from the centre to the outermost regions as specified in Fig. \ref{fig:halo_profile}, but excluding the SW sector and dividing the remaining area in 3 sectors. The outermost annulus in the NW sector shows contamination from soft X-ray protons and it has been excluded from the following analysis. Statistical errorbars do not appear as they are smaller than the points. }
\label{fig:plot_profile_radioX}
\end{figure}
\begin{figure}
\vspace{-1pt}
\includegraphics[width=0.9\columnwidth]{plot_RadioOverX_sectors_chWidth.png}
\includegraphics[width=0.9\columnwidth]{plot_RadioOverXtoBeta_sectors_chWidth.png}
\includegraphics[width=0.9\columnwidth]{plot_RadioOverXtoBeta_sectors_chWidth_core.png}
\caption{Top: Ratio of the radio to X-ray brightness, computed in elliptical annuli as in Fig. \ref{fig:plot_profile_radioX}. Middle: Ratio of $I_R/I_X^{\beta}$, where $\beta =0.76$ is the best fit correlation coefficient found for the whole halo. Bottom: Ratio of $I_R/I_X^{\beta_{core}}$, where $\beta_{core}=0.39$ is the best fit correlation coefficient found for the halo core. In the three panels, Bullets, stars and squares refer to the NW, NE, and SE sector, respectively. Vertical dashed lines mark the position of $r_1$ and $r_2$, used to separate the halo core from the outer halo. Statistical errorbars do not appear as they are smaller than the points.}
\label{fig:RadioVersusX}
\end{figure}
\subsection{Radial analysis}
The point-to-point analysis above is important to understand the local connection between thermal and non-thermal plasma, and allows one to understand if regions with higher non-thermal energy are traced by high X-ray brightness. The point-to-point analysis has also been used to probe the radial scaling of the radio and X-ray brightness \citep[e.g.][]{Govoni01,Botteon20}.
To better analyse the radial trend of $I_X$ and $I_R$ and to investigate a possible change with radius, we have performed an additional analysis by comparing the X-ray and radio brightness in elliptical annuli and dividing the radio halo in sectors.
Specifically, we have excluded from the analysis the SW sector, where the bridge is, and we have considered separately the NW, NE, and SE sectors.
The profiles are computed in elliptical annuli with progressively increasing width to gain sensitivity towards low surface brightness emission in the halo outermost regions. X-ray profiles have been computed using the same regions. These plots, shown in Fig. \ref{fig:plot_profile_radioX}, show a different radio profile in the three sectors we have considered, and show clearly a shallower decline of the radio brightness with respect to the X-ray brightness.
In the NW sector, we detect a sharp decline of the radio brightness at $\sim$1900$^{\prime \prime} \,$ i.e. the location of the radio front (see sec. \ref{sec:haloFront}). In the NE sector a sharp decline of the radio emission is detected at $\sim$ 1400$^{\prime \prime} \,$ (670 kpc). This is due to the presence of the filaments of the radio halo (Fig. \ref{fig:zoomHalo}) and to the asymmetry of the halo emission, that is more pronounced towards the W. The radio profile in the SE sector shows, instead, a smooth decline.\\
To inspect the possible change in the relative radial trend of $I_R$ and $I_X$, we show in Fig. \ref{fig:RadioVersusX}, top panel, the ratio $I_R/I_X$ in Logarithmic base 10 scale, for the different sectors (NW, NE, and SW). The trend is similar in the 3 sectors, with a ratio that is smaller in the central part of the halo and that progressively increases. In the 3 outermost annuli (approximatively at 1600$^{\prime \prime} \,$, i.e. $\sim$ 750 kpc ) the ratio decreases in the NE and SE sectors. This plot shows that the ratio of the two quantities is not constant throughout the halo, consistent with the results obtained in the point-to-point analysis.
In the middle panel of Fig. \ref{fig:RadioVersusX}, we show the ratio of $I_R$ to $I_X^{\beta}$, using the value of $\beta$ obtained from the point-to-point analysis ($\beta=0.76$). If the same value of $\beta$ were representative of the whole halo, we would expect to see a horizontal line. Instead, the plot shows that a single value of $\beta$ does not represent the whole halo emission.
The change in the slope of $\beta$ is shown more clearly in the bottom panel of Fig. \ref{fig:RadioVersusX}, where we plot $I_R/I_X^{\beta_{core}}$, where $\beta_{core}=0.41$ is the best fit value obtained from the point-to-point analysis restricted to the halo core. \\
Hence, we conclude that the $I_R-I_X$ correlation has a slope that changes with the radial distance from the cluster centre, being flatter in the halo core (where we find $\beta\sim0.41$) and steeper in the outer halo (where we find $\beta \sim 0.76$. While the slope in the core seems to remain constant, the ratio $I_R-I_X^{\beta}$, shown in the middle panel of Fig. \ref{fig:RadioVersusX} indicates a progressive steepening of the correlation with increasing distance from the cluster centre.\\
\begin{figure}
\includegraphics[width=0.9\columnwidth]{Beta_versus_B.png}
\caption{Trend of the radio-X ray correlation slope $\beta$ ($I_R=I_X^{\beta}$) with radial distance from the cluster centre. $\beta$ is computed
as the ratio between the logarithm of the theoretical projected radio emissivity to the logarithm of the X-ray projected emissivity. The radio emissivity is computed for different values of the central magnetic field $B_0$, from 3 (top curve) to 10 $\mu$G (bottom curve), as specified in the legend. Shaded area refer to the best-fit $\beta$ obtained from data.}
\label{fig:beta_th_core_outer}
\end{figure}
\subsection{Modelling the $I_R-I_X$ correlation and its radial trend}
A steepening of the $I_R-I_X$ correlation outside the core is expected as a consequence of the different relative weight of IC and synchrotron losses in a magnetic field declining with radius. Hence, under some assumptions on the magnetic field profile, we can investigate whether the expected radial drop of magnetic field can be entirely responsible for the steepening of the correlation, or whether additional effects are required.
We assume a magnetic field profile scaling with the thermal gas as $B(r)\propto B_0 \cdot n_e(r)^{0.5}$, consistent with \cite{Bonafede10}. Note that
in this case the Alfv\'en velocity in the ICM is constant.
The radio emissivity in turbulent re-acceleration models can be expressed as:
\begin{equation}
\label{eq:epsilonr}
\epsilon_R \propto F \eta_e {{B^2}\over{B^2 +
B_{IC}^2}}
\end{equation}
where $\eta_e$ is the acceleration efficiency, $B_{IC}$ is the CMB equivalent magnetic field, and $F$ is the turbulent energy flux:
\begin{equation}
\label{eq:fluxTurb}
F \sim \frac{1}{2} \rho \frac{\sigma_v^3}{ L }
\end{equation}
Here, $\sigma_v$ is the velocity dispersion on scale L, and $\rho$ is the gas density.
Assuming an isotropic distribution of electrons in the momentum space, $f(p)$, the acceleration efficiency is \citep[e.g.][]{BrunettiLazarian07}:
\begin{equation}
\eta_e \sim F^{-1} \int d^3p {{E}\over{p^2}}
{{\partial }\over{\partial p}} \big(
p^2 D_{pp} {{\partial f}\over{\partial p}}
\big) \approx {{U_{CRe}}\over{F}} ( D_{pp}/p^2 )
\end{equation}
where $U_{CRe}$ is the energy density of reaccelerated electrons,
and $D_{pp}/p^2$ has different expressions in different re-acceleration models. Specifically,
\begin{equation}
\frac{D_{pp}}{p^2} \propto \frac{c_s^2 \mathcal{M}_t^4}{L}
\end{equation}
in the case of Transit-Time-Damping acceleration with compressive modes \citep{BrunettiLazarian07, Miniati15}, and:
\begin{equation}
\frac{D_{pp}}{p^2} \propto \frac{ c_s^3 \mathcal{M}_t^3 }{L v_A}
\end{equation}
in case of non-resonant second-order Fermi acceleration with solenoidal modes \citep{BrunettiLazarian16}.
If we also assume a constant temperature and a scenario based on a constant turbulent Mach number, the synchrotron emissivity is :
\begin{equation}
\epsilon_R(r) =\epsilon_R(0)
{{X(r)}\over{X(0)}}
\left( {{\epsilon_X(r)}\over{\epsilon_X(0)}} \right)^{1 \over 2}
{{
1+ \left({{B_{IC}}\over{B_0}}\right)^2 }\over{1+\left({{B_{IC}}\over{B(r)}}\right) ^2} \left({{\epsilon_X(0)}\over{\epsilon_X(r)}}\right)^{1 \over 2} }
\end{equation}
where $\epsilon_X \propto n^2$ is the X-ray (bremsstrahlung) emissivity and
$X= \frac{U_{CRe}}{U_{th}}$ the ratio of the energy density of CRe to thermal gas.
From the radio and X-ray emissivity, we have computed the projected radio brightness assuming constant $X$, and different values of the central magnetic field, $B_0$, ranging from 3 $\mu G$ to 10 $\mu G$, and the X-ray brightness using the parameters obtained by \citet{Briel01}. Then, we have computed the slope of the correlation $I_X-I_R$ as a function of the radial distance from the cluster centre. The trend of $\beta$ versus the radial distance is shown in Fig. \ref{fig:beta_th_core_outer} for different values of $B_0$. We have overplotted as shaded areas the values obtained in the halo core and in the outer halo. \\
Although all the models predict a steepening of the radio-X-ray scaling with radius, none of them is able to reproduce all observed values. In particular, except for the profile with $B_0=5\mu$G, the magnetic field profiles that would be compatible with the value of $\beta$ in the outer halo are higher than those derived from RM studies \citep{Bonafede10}.
A possible scenario to explain the observed trend, is to assume that the ratio of the CRe to thermal gas energy densities ($X$) increases with radial distance from the cluster centre. This is indeed expected, as the lifetime of CRe in the ICM increases with the distance from the cluster center due to the lower Coulomb and synchrotron losses. However, it is the first time that a radially decreasing $X$ seems to be suggested by observational data. \\
As for the calculations done in Sec. \ref{sec:haloSpectrum_theo}, we stress that the calculations outlined here are subject to several assumptions and the decrease of $X$ with the distance from the cluster centre is only one of the possible causes of the observed $I_R-I_X$ slope.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{LOFAR_SZ_corr.pdf}
\caption{Radio-SZ correlation for the radio halo. The radio image has been convolved to a resolution of 5$'$ and the source NGC4839 has been masked. The errorbars refer to the statistical errors of the two quantities. Arrows mark the $2\sigma_{\rm rms}$ upper limits. The vertical and horizontal dotted lines marks $1\sigma_{\rm rms}$ for the $y$-parameter and radio image, respectively. Errorbars are plotted every second value.}
\label{fig:corr_radioSZ}
\end{figure}
\section{Radio - thermal pressure correlation in the radio halo}
\label{sec:radioSZ}
Resolved SZ maps of the Coma cluster \citep{PlanckComa} can be used to understand the connection between the thermal gas and the radio emission.
Since the comptonization parameter $y$ is proportional to the gas pressure integrated along the line of sight ($y \propto n_e T$), it is less contaminated by cold gas clumps with respect to X-ray emission.
We have computed the $I_{SZ} - I_R$ point to point correlation, following the same approach as described above for the $I_R - I_X$ correlation. We have used the radio image convolved to 5 arcmin resolution, to match the $y$-map resolution.\\
Initially, we have fit $I_{SZ}$ versus $I_R$ in the same region as the one considered in the $I_R-I_X$ correlation, finding a super-linear slope
$\beta_{SZ}= 1.76\pm0.08$.
Assuming an isothermal model, one would expect that the value found from the $I_{SZ} - I_R$ correlation would be twice the scaling of the $I_R-I_X$. Hence, given the values found from the $I_R-I_X$ correlation, one would expect $\beta_{SZ}= 1.4 - 1.6$. There is a small tension between SZ and X-ray, that could be due e.g. to dense and cold X-ray clumps in regions of low surface brightness emission. Understanding the details of this small tension is not the scope of this work, what is relevant for our analysis is that regions of high non-thermal energy are regions of high thermal energy, as probed by the $I_{SZ} - I_R$ correlation. \\
The $y$-map is more extended than the \emph{XMM-Newton} mosaic, and covers regions
at the cluster outskirts where we have a radio halo detection. Hence, we can use
the $y$ parameter map to investigate the correlation between thermal and non-thermal regions up to larger radii from the cluster centre, though we miss the resolution given by X-ray data.
We have fitted $I_{SZ}$ versus $I_R$ out to a distance of $\sim$1.3 Mpc ($\sim$2800$^{\prime \prime} \,$ radius) from the cluster centre, that is the maximum distance where we have detected the radio halo. As shown in sec. \ref{sec:radioXcorrelation}, we have indication from the X-ray analysis that the slope would further increase when outer regions of the halo are considered. We find $\beta=1.78 \pm 0.08$, which is slightly steeper than the slope found in the inner 2400$^{\prime \prime} \,$ radius, though consistent within the errors. The $I_{SZ} - I_R$ correlation is shown in Fig. \ref{fig:corr_radioSZ}. Both X-ray and SZ analysis show that when outer regions of the halo are considered, the correlation slope increases, i.e. the outermost regions of the radio halo have a different ratio of thermal/non-thermal energy than the inner ones.\\
A correlation between the radio emission and the $y$-signal has been firstly obtained by \cite{PlanckComa}, who fitted $I_{SZ}$ versus $I_R$, finding a quasi-linear relation: $I_{SZ} \propto I_{R}^{0.92 \pm 0.04}$, that would correspond to our $\beta \sim 1.1$. They used the WSRT 325\,MHz map and $y$ images at 10 arcmin resolution, and extracted the radio and $y$-signal from $r<50$ arcmin region.
As we find a steeper slope, using a lower frequency radio image we investigate in the following the possible causes of the different trend.
First of all, we are looking at the $I_R - I_{SZ}$ correlation using a radio image at a different frequency than used in \citet{PlanckComa}.
In addition, we are able to perform a more accurate subtraction of the contaminating sources, as the highest LOFAR resolution is 6$^{\prime \prime} \,$, and we are less affected from calibration artefacts than the WSRT image. Residuals from contaminating sources would increase the values of the radio brightness in each box, and this effect could be more prominent for boxes at the halo periphery. In this case, the effect would be a flattening of the correlation.
The data we have at the moment of writing do not allow to exclude that this is the cause of the different slope found with LOFAR and WSRT. However, we note that a higher value of the correlation slope for the $I_R - I_X $ at lower frequencies has recently been found by Rajpurohit et al. (submitted) in the halo of the cluster Abell 2256, and by \cite{Hoang21} in the halo of CIG 0217+70.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{image_FturbSIm.pdf}
\caption{Projected $F_{comp}$ (left) and $F_{sol}$ (right) from MHD cosmological simulations of a Coma-like cluster, from \citet{va18mhd}. Each map has a size of 2560 kpc.}
\label{fig:sim_images}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{corrV_comp.im.pdf}
\includegraphics[width=\columnwidth]{corrV_sol.im.pdf}
\caption{Left panel: $F_{Comp} - I_X$ correlation from simulated data. Right panel: $F_{Sol}- I_X$correlation from simulated data.}
\label{fig:corr_sim}
\end{figure*}
\section{Comparison with numerical simulations}
\label{sec:simulations}
Having the radio halo resolved in great detail, we can try to understand its origin with the help of numerical simulations. Specifically, we try in this Section to understand whether the observed scaling $I_R - I_X$ can provide useful information to better constrain the particle acceleration mechanism.
\subsubsection{Simulated thermal to non-thermal correlations}
We use a Coma-like galaxy cluster, simulated at high resolution and with ideal Magneto-Hydrodynamics using the cosmological Eulerian code ENZO (enzo-project.org) by \citet{Vazza18}. This system has a $z \approx 0.02$ total mass comparable with the real Coma cluster, and it shows a radial decline of the magnetic field compatible with the estimates from Rotation Measures \citep{Bonafede10,va18mhd}. The simulation includes eight levels of Adaptive Mesh Refinement (AMR) to increase the spatial and force resolution in most of the innermost cluster volume, down to $\Delta x=3.95 \rm ~ kpc/cell$. While this simulation assumes
an initial volume-filling background of weak magnetic field, $B_0=10^{-4} \rm \mu G$ (comoving) at $z=40$, the low redshift properties of the magnetic field are fairly independent of the exact origin scenario, due to the effect of the efficient small-scale dynamo amplification \citep[e.g.][]{2009MNRAS.392.1008D,va21b}.
Using the filtering technique described in \citet{va17turb}, we have computed the turbulent energy flux, $F_{Comp,Sol}$, associated to the compressive and solenoidal velocity components:
\begin{equation}
F_{Comp,Sol}= \rho \frac{\sigma_{v_{Comp,Sol}}^3}{L}\times \frac{B^2}{B^2 + B_{IC}^2},
\end{equation}
where, similarly as in Eq. \ref{eq:fluxTurb}, $\sigma_{v_{Comp,Sol}}$ is the dispersion of the compressive, solenoidal velocity field on scale $L$, different for solenoidal and compressive modes.
From $F$, one can compute the simulated synchrotron luminosity (see Eq. \ref{eq:epsilonr}) as:
\begin{equation}
L_{Comp,Sol}= \eta_e F_{Comp,Sol}
\end{equation}
The constant $\eta_e \leq 1$ gives the dissipation efficiencies for solenoidal and compressive modes into cosmic ray acceleration, which depend on the complex physics of cosmic ray acceleration via Fermi II process \citep[see e.g.][]{Miniati15,BrunettiLazarian16,BV20}. However, since in this application to our new LOFAR observations of Coma we are only concerned in the relative distribution of the two energy fluxes, we fix $\eta$ for simplicity, acknowledging that both fluxes will represent an overestimate (likely by a factor $\sim \geq 10^2$) of the effective dissipation onto cosmic ray acceleration. Hence, we have used $F_{Comp,Sol}$ as a proxy for the synchrotron luminosity instead of $L_{Comp,Sol}$.
The maps of projected $F_{Comp}$ and $F_{Sol}$ are shown in Fig. \ref{fig:sim_images}.
We have computed the correlation between $F_{Comp,Sol}$ and the simulated X-ray brightness $I_X$ as we have done for the radio and X-ray emission.
We note that some assumptions must be done to compare the simulated $F_{Comp,Sol}$ to the observed radio emission. Specifically, we have assumed that the halo emission is generated by turbulent re-acceleration, and that $\eta_e$ is constant throughout the cluster volume. Furthermore, as we are not modeling the CRe component, we have averaged $F_{Comp,Sol}$ on scales larger than the electron diffusion length at 144 MHz.
The diffusion length $l_e$ of electrons emitting at 144 MHz can be estimated at first order as
\begin{equation}
l_e \sim 2 \sqrt{D \tau_{rad}} \sim 100 \rm kpc.
\end{equation}
Here, $\tau_{rad}$ is the electrons' radiative age at 144 MHz ($\tau_{rad} \approx 200$ Myr) and
$D$ is the spatial diffusion coefficient. A simple estimate for $D$ (i.e. ignoring pitch angle scattering along magnetic field lines) can be obtained from the typical scale for MHD turbulence in the ICM ($l_A \sim 0.1-0.5$ kpc) as $D \approx \frac{1}{3} l_A c$.\\
We have computed the mean of $F_{Comp,Sol}$ and $I_X$ in boxes of 160 kpc side, that is larger than $l_e$ and comparable to the size of the boxes used to compute the $I_R-I_X$ correlation from data (see Sec. \ref{sec:radioXcorrelation}). \\
In Fig. \ref{fig:corr_sim}, the two correlations obtained with simulated data are shown. Both $F_{Comp}$ and $F_{Sol}$ are positively correlated with the simulated $I_X$. The correlation slope is similar, though a bit steeper in the case of $F_{Sol}$-$I_X$ (see Tab. \ref{tab:corr_sim}. Both slopes are consistent with the observed $I_R - I_X$ correlation, supporting the connection between turbulence and radio diffuse emission. We note that the power in the solenoidal energy flux is $\sim$ 10\% higher than in the compressive energy flux. Once the CRe emission is properly modelled, this could be used to disentangle the role played by the two modes for CRe re-acceleration.
\subsection{Correlations in the centre and peripheral regions}
As shown in the previous section, the global $I_R-I_X$ in the Coma cluster can be recovered using both compressive and solenoidal energy fluxes as a tracer for the radio emission.
MHD simulations show that solenoidal modes are dominant in the cluster central regions, and compressive modes dominate the cluster outskirts \citep{Miniati15,Vazza17}. We have seen in Sec. \ref{sec:radioXcorrelation}, that the different $\beta$ obtained in the halo core and in the outer halo can not be explained by the decline of the magnetic field only. Here, we investigate whether different turbulent modes could play a role. We have fitted the simulated $F_{comp,sol}$ versus the simulated X-ray brightness $I_X$ in the cluster core and in the cluster external regions, separately. We used boxes of 32 kpc size to have a better sampling of the halo core.
The results are listed in Table \ref{tab:corr_sim}. The best-fit slope in the halo core is flatter than in the outer halo for both solenoidal and turbulent energy flux. The Pearson correlation coefficient, though, indicates that the correlation between $I_X$ and $F_{comp,sol}$ is weak in the halo core. \\
Hence, we conclude that globally the trends are in agreement with the scenario where turbulent re-acceleration produces the radio halo. In order to understand the process in more detail, the CRe distribution needs to be modelled.
\begin{table}[]
\caption{Correlation between thermal and non-thermal simulated quantities}
\centering
\begin{tabular}{c c c c}
\hline
\hline
& $\beta$ & 10\% -- 90\% & $\rho_{P}$\\
$F_{Sol} -I_X$ & 0.88 & 0.94 -- 0.81 & 0.8\\
$F_{Comp} - I_X $ & 0.85 & 0.92 -- 0.77 & 0.8\\
\hline
\multicolumn{4}{c}{Halo core} \\
$F_{sol} - I_X $ & 0.5 &0.6 - 0.4 & 0.2\\
$F_{Comp} - I_X $ & 0.5 &0.6 - 0.3 & 0.3\\
\hline
\multicolumn{4}{c}{Outer halo} \\
$F_{sol} - I_X $ & 0.83 &0.91 - 0.75 & 0.8\\
$F_{Comp} - I_X $ & 0.78 &0.87 - 0.69 & 0.8\\
\hline \hline
\end{tabular}
\label{tab:corr_sim}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{front_spix_anuli.pdf}
\includegraphics[width=\columnwidth]{plot_blankSources_11anuli_front_preshock.txt_spix.png}
\caption{Left panel: annuli used to compute the spectral index trend in the halo front region (white - dashed) and in the outer halo (red). Right: Spectral index trend computed in the annuli shown in the left panel. Arrows indicate 3$\sigma$ upper limits. The vertical dashed line indicates the position of the halo front.
Errorbars only show statistical errors, the values of the spectral index are also affected by the flux calibration uncertainties of WSRT and LOFAR (10\% and 15\%, respectively) that would contribute with an additional error of 0.2.}
\label{fig:spix_front}
\end{figure*}
\section{The halo front}
\label{sec:haloFront}
Previous studies pointed out the presence of discontinuities in the thermal gas cluster properties located at the western edge of the radio halo \citep{Markevitch10,PlanckComa,Simionescu13}. This discontinuity, detected in both the Comptonisation parameter $y$ and in the temperature and deprojected density profiles, is consistent with an outwardly-propagating shock front, located at $\sim$ 33 arcmin (910 kpc) from the cluster centre. More recently, \citet{Churazov21} used X-ray data from eROSITA and confirmed the presence of a shock wave in the West, with a Mach number $M \sim 1.5$. This shock is interpreted as a secondary shock, or ``mini accretion shock" driven by the first passage of the NGC 4839 group through the cluster before reaching the apocenter and inverting its orbital motion. According to this scenario, both the relic and the W shock would be caused by the merger of NGC 4839 with the Coma cluster. During its first passage NGC4839 would have driven a first shock that should now be located at the position marked by the radio relic. The gas displaced by the merger would settle back into hydrostatic equilibruim, forming a ``mini accretion" shock.
We refer the reader to \cite{Burns94}, \cite{Lyskova19}, \cite{Zhang21}, \cite{Churazov21} for a more detailed explanation of the proposed merging scenario.\\
Our LOFAR image confirms the presence of an edge of the radio halo towards the West (see Fig. \ref{fig:profile_front}), that could be in line with the scenario explained above.
Shocks in the ICM are often associated with radio relics, where they leave a clear imprint on the spectral index distribution.
The spectral index profile is flatter where particles are freshly re-accelerated and steeper moving towards the downstream region where particles radiate their energy via synchrotron and Inverse Compton losses. We have investigated whether a similar trend is found in the ``halo front", computing the spectral index profile in annuli that follow the front (Fig. \ref{fig:spix_front}). To derive the spectral index values, we have used the same images presented in Sec. \ref{sec:haloSpectrum}\footnote{we recompute the spectral index radial trend in this section to better highlight the differences between the front region and the rest of the halo, while in Fig. 9 we have divided the halo in 4 identical sectors. For the same reason, the spectral index is computed out to a smaller distance from the cluster centre than in Fig. 9.}.
In Fig. \ref{fig:spix_front}, we show the spectral index profile obtained in annuli that follow the front, in comparison with the global spectral index trend obtained in annuli centred on the cluster and excluding the halo front region (Fig. \ref{fig:spix_front}, left panel).
We note that in the direction of the front, the spectral index is flatter, but no steepening in the putative downstream region is present. However, the outermost annuli seem to follow a different trend than the global halo profile: the spectral index remains almost constant, while a radial steepening is detected when the whole halo is considered. It is possible that we are limited by the resolution, as the width of the annuli is $\sim$2$'$, corresponding to $\sim$60 kpc, that is larger than the electron cooling length in the post shock region \citep[e.g.][]{KangRyu12}.
Hence, it is possible that we do not have the resolution to separate the shock front from the post-shock region, where particles have already experienced strong radiative losses.
An alternative scenario is that the radio front could be radio plasma re-accelerated by the same mechanism responsible for the halo emission, and then dragged by the shock passage and compressed by it.
Possibly, future high-resolution observations at higher or lower frequency will allow us to understand whether the halo front shows the typical shocked
spectral index profile that we detect in radio relics or not.\\
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{spix_profile_relic-WAT.pdf}
\caption
Spectral index profile between 326 MHz and 144 MHz of the emission called relic-NAT connection. Images at the WSRT resolution (100$^{\prime \prime} \,$ $ \times $ 150$^{\prime \prime} \,$) have been used to compute the spectral index in each box.}
\label{fig:Relic_NAT_profiles}
\end{figure}
\section{The relic-NAT connection}
\label{sec:relicNAT}
The diffuse emission connecting the relic to the NAT radio galaxy NGC4789 is imaged here with unprecedented resolution, that allows us to detect substructures in its surface brightness distribution.
In particular, the bent jets of NGC4789 do not blend smoothly into the diffuse emission, as observed in other cases of radio galaxies nearby relics \citep{Bonafede14b, vanWeeren17_Nature, Stuardi19}, and two discontinuities between the endpoint of the jets and the diffuse emissions are detected.
At the centre of the diffuse source region, there is a bright transverse bar.
Such a bar has been detected in other bent tails (e.g. in Abell 2443, \citealt{Clarke13}, and recently in the Shapley supercluster \citealt{Venturi22}) and are predicted by simulations of interacting AGN tails and shocks \citep{Nolting19}.
The length of this source, measured from the relic's edge down to the endpoints of the jets, is 10 arcmin, corresponding to 280 kpc at the Coma redshift.\\
To investigate the possible connection of the radio plasma with the thermal gas, we have used the ROSAT image and the radio image at 35 $^{\prime \prime} \,$ resolution, and investigated the existence of a correlation between the thermal and non-thermal plasma, as done already for the radio halo (see Sec. \ref{sec:radioXcorrelation}), and for the Coma bridge \citep{Bonafede21}.
The average brightness profile of the relic-NAT connection is largely constant, increasing close to the relic edge where the radio bar is located. In addition, the low counts in the X-ray image do not allow us to make a proper analysis. Hence, though no correlation or anti-correlation between the two quantities seems to be present, no firm conclusions can be derived, and we can not rule out that the radio emisison originates from phenomena similar to those responsible for the bridge.
However, given its morphology, we will investigate below an alternative
scenario.\\
The spectral index trend along the Relic-NAT connection main axis is shown in Fig. \ref{fig:Relic_NAT_profiles}. We have used the WSRT R image (see Tab. \ref{tab:images}, \citealt{Giovannini91}) that has been used already in \citet{Bonafede21} to analyse the bridge. Although the higher resolution of the image by \cite{BrownRudnick11} would provide a better description of the spectral index trend, calibration errors from ComaA are strongly affecting that region.
The spectrum of the radio emission computed between 144 MHz and 326 MHz shows a gradual steepening from regions close to the AGN core towards the relic's outer edge, reaching values of $\alpha = -1.6 \pm 0.2$.
At the relic's outer edge, the spectral index flattens to $\alpha=-1.2 \pm 0.2$, and it steepens towards the NE, i.e. the putative post-shock region, reaching again $\alpha= -1.6 \pm 0.2$. We note that the spectrum starts to flatten already in front
of the relic edge (at $\sim$77$^{\prime}$ from the cluster centre), though within the error that value is consistent with the steepest point. This apparent flattening could be due to projection effects, as if the relic has a velocity component along the line of sight, some relic emission could appear in front of the relic edge in projection.
\\
Overall, the spectral trend detected along the tail of NCG4789 and in the relic-NAT connection is consistent with AGN particle ageing. NGC4789 would be moving towards the SW of the cluster injecting particles into the ICM. Thus more recently injected particles are closer to the AGN core than the older ones that have been left behind. The shape of the source and the connection with the radio relic make this source peculiar and suggest a link between the relic-NAT connection and the shock wave that would power the radio relic.\\
However, according to the merging scenario outlined in several papers \citep{Venturi90,Giovannini91,Churazov21}, NGC4789 would be in the pre-shock region. Hence, it is puzzling to understand how features like the bar could have been formed, because that emission has not yet interacted with the shock.
\cite{EnsslinGK01} have proposed that the plasma from NGC4789 is dragged by the infalling matter (falling into Coma's cluster) to the location of the relic, where it is re-energised adiabatically by the shock wave.\\
Three possible scenarios could in principle explain the emission of the NAT and its connection with the relic, and we briefly outline them here:
\begin{itemize}
\item (i) NGC4789 is in the pre-shock region, and the shock wave responsible for the relic is moving towards the southwest and approaching the tail. In this scenario, the emission from the tail
in the Relic-NAT connection region would be unaffected by the shock passage. However, the plasma injected in the ICM by the tail would play an important role to explain the relic emission, as it would furnish energetic electrons that the low Mach number shock wave would reaccelerate.
\item (ii) The relic is powered by a quasi-stationary accretion shock, and that NGC4789 is moving supersonically and crossing the shock region from NE to SW. Hence, NGC4789 would be in the pre-shock region, but its tail would have crossed the shock wave responsible for the relic emission.
\item (iii) The relic is powered by a shock wave that is moving towards the cluster centre. NGC4789 has been crossed by the shock and it is now in the post-shock region.
\end{itemize}
Scenario (iii) has been recently studied by \cite{Nolting19} using
numerical simulations. They have analysed the case of a shock wave that crosses an AGN when the shock normal is perpendicular to the jets. They find that, when a shock passes through the active jets and lobes of an AGN, jets are distorted by the shock passage, creating a NAT morphology, and a structure similar to the radio bar, as observed here. The spectral index trend resulting from the simulations by \citet{Nolting19} is in agreement with the one observed here along the tail and relic-NAT connection, though the values are slightly flatter after 200 Myr from the shock passage, when the simulation stops. This scenario would require a shock wave moving towards the cluster centre, that seems disfavoured by some authors \citep{Feretti05, Akamatsu13,Churazov21} and also by the spectral index trend across the relic that we detect, which
is steepening NE, in agreement with expectation from an outwards moving shock wave.
In scenario (ii), the situation would have some similarities with the simulation by \cite{Nolting19}, as also in this scenario the AGN would be crossed by a shock. However, in this case, the AGN would be in the pre-shock region, and its tail would be interacting with a less dense environment than the one simulated by \cite{Nolting19}.
Scenario (ii) seems also supported by the analysis of \citet{Adami05}, where they detect a relative velocity of $\sim$1000 km/s between NGC~4789 and NGC~4839, at the cluster centre.
Our analysis does not allow us to discard scenario (i), where the shock is overtaking the pre-existing tail from the back. We can see that the tail and shock must be interacting, because the tail abruptly stops at the position of the relic. This situation has not been simulated yet, but it is likely that it would require some fine-tuning of the parameters.
We indicate scenario (ii) as the favoured one, given the data we have now, though it remains to be seen whether the interaction of the AGN tail with a pre-shocked environment at the cluster outskirts would create the substructures that we observe.
In all scenarios, the NAT NGC~4789 would be providing seed electrons that are (re)accelerated by the shock and originate the radio relic. In the literature, there are a few other cases where a link between an AGN tail and a relic has been established (see \citealt{Bonafede14b,vanWeeren17_Nature,Stuardi19}.
\section{Discussion and conclusions}
\label{sec:discussion}
In this work, we have used new data at 144\, MHz from LOFAR to analyse the emission from the Coma cluster.
We summarise our findings in the following and discuss how these observations allow us to advance our understanding of the non-thermal emission in clusters of galaxies.\\
We have focused our analysis on the properties of the radio halo which - detected at 144\,MHz with the resolution and sensitivity allowed by LOFAR - presents new interesting features and allows us to perform detailed resolved studies of the radio emission. We find that:
\begin{itemize}
\item the radio halo at 144 MHz appears larger than previously reported in the literature, with a largest angular scale of $\sim$71$'$, corresponding to $\sim$ 2 Mpc. The halo is connected to the relic through a low surface brightness radio bridge, and the relic is connected to the AGN NGC\,4879 to the southwest. In total, the radio emission from the halo to the head of NGC4879 spans $\sim$2$^{\circ}$, corresponding to $\sim$3.4 Mpc.
\item The halo brightness profile is well fitted by an exponential elliptical profile. At 144\,MHz, it is characterized by e-folding radii $r_1$= 355 kpc and $r_2$= 268 kpc. At 325 MHz, the profile is more peaked, with $r_1$= 268 kpc and $r_2$= 240 kpc. This is consistent with a spectral steepening of the radio emission towards the halo outskirts, in agreement with the results by \citet{Giovannini93}.
It would be useful to perform these fits also at other frequencies in order to study how the halo size changes with frequency.
\item The spectrum of the radio halo between 144\,MHz and 342\,MHz
is flatter than previously reported, though consistent within the errors. We find $\alpha=-1.0 \pm 0.2$, while previous studies indicate $\alpha \sim - 1.2$. We have computed the flux density of the halo at both frequencies from the best-fit exponential model, that is less affected by the different sensitivities of the two images and the possible presence of residuals from unrelated sources.
\item We have computed the radial trend of the spectral index $\alpha$ dividing the halo in four symmetric sectors. Our estimates could still be affected by residual contamination from unrelated source and calibration errors, but these should not have a major impact on the global results.
We detect for the first time a moderate steepening towards the cluster center of the spectral index, and confirm the steepening towards the cluster outskirt found at higher frequencies by \citet{Giovannini93}. This trend is in agreement with the expectations of homogeneous turbulent re-acceleration models. Though a detailed modeling is needed to understand the effect of projection affects and the exact location of the steepening frequency, we argue that the steepening detected at the cluster outskirts could indicate a non-constant acceleration time, and hint to a constant turbulent Mach number.
The spectral index steepening towards the cluster centre is more or less pronounced in the different sectors, being prominent in the SW sector and not clearly detected in the NE sector. It is possible that the SW sector has been perturbed by the passage of the NGC4839 group, and shows now different properties.
\item The point-to-point analysis between radio and X-ray surface brightness indicates a sub-linear slope of non-thermal plasma with respect to the thermal plasma. We obtain $I_R\propto I_X^{0.64}$ when the correlation is computed on images at 1$^{\prime}$ resolution. We detect a steeper, yet sublinear, correlation, when the radio image is convolved with Gaussian kernels of 2, 3, 4, 5 and 6 arcmin, and it converges to $I_R \propto I_X^{0.76}$. Indeed, images at lower resolution are more sensitive to the weak emission in the halo outermost regions. We note that the total halo flux density does not change when computed from the 35$^{\prime \prime} \,$ or 6$^{\prime}$ image, because the outermost regions of the halo yield a very minor contribution to the total halo flux density. However, these regions affect the radio- X-ray correlation making it steeper.
\item We have investigated whether the radio-X-ray correlation has a different slope in the halo core than in the outer halo, finding that the correlation is flatter in the core ($I_R\propto I_X^{0.41}$) than in the outer halo ($I_R\propto I_X^{0.76}$). By investigating the radial trend of
the quantity $I_R-I_X$, we have confirmed that this trend can be interpreted as a radial trend of $I_R$ versus $I_X$ being flatter in the halo core.
An opposite slope-change has been recently found in some cool-core clusters, where the mini halo emission is surrounded by a weaker and more extended component \cite{Biava21b}. A flatter slope in the halo core is inconsistent with a major contribution of secondary electrons produced through hadronic interaction between thermal protons and cosmic ray protons in the ICM.
\item In the framework of homogeneous re-acceleration models, the change of the slope of the $I_R-I_X$ correlation can be only partially accounted for by a declining magnetic field profile. We have investigated the role of $X$, i.e. the ratio of CRe to thermal energy density, and find that a radially increasing value of $X$ would provide a better match with data. Although more detailed modeling should be done to derive firm conclusions, we note that an increasing value of $X$ with the distance from the cluster centre is also expected from a theoretical point of view.
\item With the help of MHD cosmological simulations, we have computed the turbulent energy flux associated to the compressive and solenoidal velocity components in a Coma-like cluster, and we find that both quantities show a sub-linear scaling with the simulated X-ray emission, that is in agreement with the observed scaling of $I_R$ versus $I_X$.
Assuming the same efficiency for both modes, the flux associated to the solenoidal velocity component is a factor 10 higher than the flux associated to the compressive component. Hence, once the CRe distribution throughout the volume is assumed, it would be possible to constrain the relative importance of the two modes in the process of particle acceleration.
\end{itemize}
From the analysis of the Coma field, we also conclude that:
\begin{itemize}
\item To the northeast of Coma, at a projected distance of $\sim$ 1.2 $R_{vir}$, an arc-like diffuse patch of emission is discovered. As a large-scale filament of galaxies is detected in that direction, we tentatively propose that this emission is due to particles re-accelerated by an accretion shock, and name this emission ``accretion relic". If confirmed, this would be the first detection of particle acceleration from an accretion shock.
\item The halo front, already reported by \citet{BrownRudnick11} is here confirmed coincident with the position of a shock front detected in both X-ray and SZ studies \citep{PlanckComa,Churazov21}. The radio spectral index does not seem to follow the typical trend found in radio relics. However, the large errors due to the small frequency range used to compute the spectral index do not allow us to exclude such a trend. It is possible that the halo front is caused by the radio halo plasma dragged along by the shock wave amd compressed by it. In any case, we can conclude that the properties of the halo in the W region are affected by the shock passage.
\item The radio relic in the Coma cluster is here imaged with unprecedented detail, thanks to both the sensitivity of the LOFAR observations and to the calibration techniques that we have used, that allow us to minimise the artefacts from Coma A.
The relic emission is connected to the tail of NGC4879, that is likely moving towards southwest. The connection between the relic and NGC4879 is what we name relic-NAT connection. The 20$^{\prime \prime} \,$ resolution image shows that the endpoints of the NGC4879 jets are well distinct from the weak diffuse emission of the relic-NAT connection. A bright bar of radio emission is detected, similar to what has been found in other cluster tails \cite[e.g][]{Clarke13,Wilber17}. We have discussed three possible scenarios to explain the presence of the relic-NAT connection, and propose that NGC4879 is moving supersonically towards south-west. During its motion it has crossed the shock at the position of the relic. The shock has re-energised the particles injected by the tail in the ICM in the past and left behind during the galaxy's motion through the ICM.
\end{itemize}
Using literature information about the merging scenario of Coma and its large-scale structure environment, we can outline a global picture to explain the observed radio emission. The Coma cluster is currently accreting matter through filaments of galaxies that connect it to Abell 1367 \citep{Malavasi20}. A recent merger has happened between Coma and the galaxy group NGC~4839, that has passed the cluster from North-East to South-West, injecting a first shock wave in the ICM that is now powering the radio relic emission \citep{Lyskova19}. The cluster core has been perturbed by this merger, and possibly from previous less massive mergers, that have released thermal energy in the ICM. A small fraction of this energy has been dissipated in turbulent motions, that have re-accelerated a mildly relativistic population of CRe already present in the ICM originating the radio halo.\\
From the global spectral index of the halo, its radial trend, and the analysis of the radio-X ray correlation, we are able to derive a coherent picture where particles are re-accelerated by homogeneous turbulence in the ICM.
In this picture, we have made several working assumptions that indicate possible regimes for re-acceleration to operate (i.e. constant turbulent Mach number and a radial increase of the CRe energy density with respect to the thermal energy density).
We have attempted to distinguish between re-acceleration by Transit-Time-Damping with compressive modes and non-resonant second order Fermi acceleration with solenoidal modes, and although present data are not accurate enough, we have shown that observations are entering a regime where the details of the model can be in principle tested. It is possible that future observations, either with LOFAR 2.0 and/or with MeerKAT will be able to make an additional step forward and unveil the details of turbulent re-acceleration.\\
The cluster core has been perturbed by the passage of NGC4839, and its motion around its equilibrium position has caused a second shock wave \citep{Lyskova19,Churazov21} whose front is now coincident with the halo front. We find that the spectrum of the halo front has been affected by the shock passage, but we are not able to distinguish between shock re-acceleration or compression by the shock wave.\\ NGC4839 is now at its second passage towards the cluster centre \citep{Lyskova19,Churazov21}. Its motion might have injected additional turbulence in the ICM and a considerable amount of seed electrons, that originate to the radio bridge \citep{Bonafede21}. The NAT NGC~4789 that is now located to the South-West of the relic is moving away from the cluster center at a supersonic velocity after crossing the shock wave at the location of the relic. From this interaction, the radio plasma injected in the ICM by NGC~4789 has been re-energised, leading to the emission that we detect in the relic-NAT connection.\\
Finally, a filament of matter is detected to the Northeast of the Coma cluster. The matter accreting towards the cluster from that direction could result in the tentative ``accretion relic" that we have discovered. \\
The scenario that we outline here is not the only possible one, and it leaves open questions. However, our analysis shows that we are entering a new era for the physics of non-thermal ICM emission, where we can constrain the model parameters.
\section*{Acknowledgements}
AB, EB, NB, CJR acknowledge support from the ERC Starting Grant `DRANOEL', number 714245.
AB and CS acknowledge support from the MIUR grant FARE `SMS'
FV, KR, and MBrienza acknowledge support from the ERC Starting Grant `MAGCOW', number 714196.
HB and PM acknowledge financial contribution from the contracts ASI-INAF Athena 2019- 27-HH.0, ``Attivit\'a di Studio per la comunit\`a scientifica di Astrofisica delle Alte Energie e Fisica Astroparticellare" (Accordo Attuativo ASI-INAF n. 2017-14-H.0), from the European Union’s Horizon 2020 Programme under the AHEAD2020 project (grant agreement n. 871158) and support from INFN through the InDark initiative.
MBr\"uggen acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC 2121 ‘Quantum Universe’ – 390833306.
ABotteon acknowledges support from the VIDI research programme with project number 639.042.729, which is financed by the Netherlands Organisation for Scientific Research (NWO).
RJvW acknowledges support from the ERC Starting Grant ClusterWeb 804208.
XZ acknowledges support from Chinese Scholarship Council (CSC).
AS is supported by the Women In Science Excel (WISE) programme of the NWO.
MR and FG acknowledge support from INAF mainstream project "Galaxy Clusters science with LOFAR".
LOFAR, the Low Frequency Array designed and constructed by ASTRON, has facilities owned by various parties (each with their own funding sources), and that are collectively operated by the International LOFAR Telescope (ILT) foundation under a joint scientific policy. The LOFAR software and dedicated reduction packages on https://github.com/apmechev/GRID\_LRT were deployed on the e-infrastructure by the LOFAR e-infragroup, consisting of J.B.R.O. (ASTRON \& Leiden Observatory), A.P.M. (Leiden Observatory) and T.S. (Leiden Observatory) with support from N. Danezi (SURFsara) and C. Schrijvers (SURFsara). This research had made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This research made use of APLpy, an open-source plotting package for Python (Robitaille and Bressert, 2012).
This research made use of Astropy,\footnote{http://www.astropy.org} a community-developed core Python package for Astronomy \citep{astropy:2013, astropy:2018}
\bibliographystyle{aasjournal}
| 34771d3e8b1d09996b62bf2760884ab140ac4f1f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{intro.sec}
FBOTs have peak brightnesses comparable to those of (typical or superluminous) supernovae (SNe). As implied by their name, however, they exhibit very blue colors and evolve with much shorter timescales \citep{Ho2021}. Their rapid evolution is hard to explain with radioactive decay that powers most SNe, suggesting that they are a distinct new type of transients (e.g. \citealt{X21}, hereafter \citetalias{X21}). At a distance of only 63~Mpc, AT2018cow is the closest FBOT ever discovered. It has a peak luminosity of $\sim$10$^{44}$~erg~s$^{-1}$, rises to peak in just a few days and then declines dramatically (\citealt{P18}; \citealt{P19}, hereafter \citetalias{P19}; \citealt{X21}). The photospheric temperature is $\sim$30,000~K near peak and still as high as $\sim$10,000~K at $\sim$50~d after explosion (\citealt{Kuin2019}; \citetalias{P19, X21}).
It is still unclear what progenitor is responsible for AT2018cow and what process(es) powers its rapid evolution. Current models include TDE (\citealt{Liu2018, Kuin2019}; \citetalias{P19}), CSM interaction (\citealt{Fox2019}; \citetalias{X21}), magnetar (\citealt{P18}; \citealt{Fang2019}), pulsational pair-instability SN \citep[PPISN;][]{Leung2020}, jet-envelope interaction in a core-collapse SN \citep{jet.ref}, or a jet-driven SN impostor from a common-envelope binary system \citep{Soker2022}. It is worth mentioning that AT2018cow is located in an environment typical for core-collapse SNe \citep{L20}; more recently, \citet{Pasham2021} found evidence for a rapidly spinning compact object, which could be a neutron star or a black hole of mass $<$850~$M_\odot$.
For a number of SNe, late-time observations sometimes reveal significant brightness unexpected from their early-time evolution. The late-time sources contain important information for these SNe as their possible light echos \citep[e.g. SN~2011dh;][]{Maund2019}, binary companions \citep[e.g. SN~1993J and SN~2006jc;][]{Maund2004, Maund2016, Sun2020a}, host star clusters \citep[e.g. SN~2014C;][]{Sun2020b}, late CSM interaction \citep[e.g. SN~1993J;][]{Zhang2004} or something else. In comparison, no late-time sources have ever been reported for FBOTs, most of which are located in distant galaxies and observations become very difficult beyond several months after their explosions.
Thanks to its relative proximity, AT2018cow may be the only possible FBOT to study its brightness at significantly late times. In this paper, we report the discovery of a bright source at the position of AT2018cow at 2--3~years after its explosion. This is somewhat surprising for a fast evolving transient with a steeply declining light curve. We describe the observed features and then discuss its possible origin scenarios.
Throughout this paper, we use a redshift of 0.01406 and a distance of 63~Mpc for the host galaxy \citepalias[CGCG~137-068;][]{X21} and a Galactic reddening of $E(B-V)$~= 0.078~mag for AT2018cow \citep{galebv.ref}. Its internal reddening within the host galaxy is negligible \citepalias{P19, X21}. All epochs are relative to an estimated explosion date of MJD~= 58,284.79 \citepalias{X21}.
\section{Data}
\label{data.sec}
\begin{table}
\caption{HST/WFC3/UVIS observations of AT2018cow.}
\begin{tabular}{ccccc}
\hline
\hline
Program & Epoch$^{\rm d}$ & Filter & Exposure & Magnitude$^{\rm e}$ \\
ID & (day) & & Time (s) & (mag) \\
\hline
15600$^{\rm a}$
& 52 & F218W & 880 & 19.12 (0.02) \\
& 52 & F225W & 770 & 18.99 (0.01) \\
& 52 & F275W & 280 & 18.88 (0.02) \\
& 52 & F336W & 150 & 19.17 (0.02) \\
\hline
15600
& 57 & F218W & 880 & 19.58 (0.01) \\
& 57 & F225W & 770 & 19.48 (0.01) \\
& 57 & F275W & 280 & 19.34 (0.01) \\
& 57 & F336W & 150 & 19.58 (0.01) \\
\hline
15600
& 62 & F218W & 880 & 19.75 (0.03) \\
& 62 & F225W & 770 & 19.56 (0.01) \\
& 62 & F275W & 280 & 19.56 (0.02) \\
& 62 & F336W & 150 & 19.82 (0.02) \\
\hline
15974$^{\rm b}$
& 714 & F657N & 1119 & $>$24.65 (5$\sigma$) \\
& 714 & F665N & 1119 & 24.47 (0.24) \\
& 714 & F225W & 1116 & 22.55 (0.06) \\
& 714 & F336W & 1116 & 23.32 (0.05) \\
& 714 & F555W & 1044 & 25.64 (0.07) \\
& 714 & F814W & 1044 & 25.82 (0.19) \\
\hline
16179$^{\rm c}$
& 1136 & F555W & 710 & 25.63 (0.08) \\
& 1136 & F814W & 780 & 25.96 (0.24) \\
\hline
\end{tabular} \\
PIs: (a) Foley R.; (b) Levan A.; (c) Filippenko A. \\
(d) Epoch is relative to an explosion date of MJD~= 58,284.79 \citepalias{X21}. \\
(e) All magnitudes are in the Vega system.
\label{obs.tab}
\end{table}
AT2018cow was observed by three HST programs (Table~\ref{obs.tab}), all conducted with the Ultraviolet-Visible (UVIS) channel of the Wide Field Camera 3 (WFC3). The first program (ID: GO-15600) was performed at $t$~= 52, 57 and 62~d while the other two programs (IDs: 15974 and 16179) were carried out at significantly later times of $t$~= 714 and 1136~d, respectively. We retrieved the images from the Mikulski Archive for Space Telescopes (\url{https://archive.stsci.edu/index.html}) and re-drizzled them with \texttt{driz\_cr\_grow = 3} for better cosmic ray removal (all other parameters were left unchanged as in the standard calibration pipeline).
In this work, we also use AT2018cow's early-time light curves of \citetalias{P19} out to $t$~$\gtrsim$ 60~days. At optical wavelengths, the reported magnitudes in the Swift/UVOT and SDSS-like filters were converted into the Johnson-Cousins $UBVRI$ system, using the relative offsets shown in Fig.~2 of \citetalias{P19}. Magnitudes in the $V$ and $I$ bands are very similar to those in the F555W and F814W bands, respectively, with differences no larger than the photometric uncertainties \citep{Harris2018}. Light curves in the Swift/UVOT UVM2 and UVW1 filters and in the $U$ band were transformed to the HST F218W, F225W, F275W and F336W bands using relations derived with synthetic magnitudes of blackbody spectra of 7000--50,000~K (calculated with the \textsc{pysynphot} package). \citetalias{P19} used the AB magnitude system and we converted their magnitudes into the Vega magnitude system with $m_{\rm AB} - m_{\rm Vega}$ = 1.68, 1.66, 1.50, 1.19, $-$0.02 and 0.43~mag for the F218W, F225W, F275W, F336W, F555W and F814W bands, respectively. These offsets were also calculated with the \textsc{pysynphot} package and are valid for hot sources like AT2018cow.
\section{Detection of a late-time source}
\label{source.sec}
\begin{figure*}
\centering
\includegraphics[width=1\linewidth, angle=0]{c_stamp.stamp.pdf}
\caption{Example images of AT2018cow at early and late times of $t$~= 62~d (a--d) and 714~d (e--j), respectively; each panel has a dimension of 1"~$\times$ 1" and is centered on AT2018cow (shown by the cross hair). No source is detected at the position of AT2018cow in the F657N image (e). In the F814W image (j), the late-time source is not very obvious by eye due to a bright neighboring star to its northeast. (k) Three-color composite of the F336W, F555W and F814W images at $t$~= 714~d; the square corresponds to the extent of the other panels (a--j) and the circle has a radius of 1". All panels are aligned with North up and East to the left. An angular size of 1" corresponds to a linear size of 305~pc at the distance of AT2018cow.}
\label{image.fig}
\end{figure*}
We performed point-spread-function (PSF) photometry on the HST images with the \textsc{dolphot} package \citep{dolphot.ref}. At the position of AT2018cow, a point source is significantly detected in all the broad-band images, only marginally detected in the F665N image with a 4.6$\sigma$ significance, and not detected in the F657N image (Fig.~\ref{image.fig}). The magnitudes or magnitude limits (obtained with artificial star tests) for this source are listed in the last column of Table~\ref{obs.tab}. Note that the late-time source is not spatially resolved; therefore, its size should not significantly exceed that of the PSF, which corresponds to $\sim$20~pc at the distance of AT2018cow.
\paragraph*{Light curves}
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth, angle=0]{c_curve.curve.pdf}
\caption{HST photometry of AT2018cow (filled circles) with error bars smaller than the symbol size. The solid lines are AT~2018cow's early-time light curves reported by \citetalias{P19} (with line thickness showing the photometric uncertainties), which have been converted into the HST filters and the Vega magnitude system. The dashed lines correspond to linear extrapolations of the light curve tails or the brightness of the late-time source.}
\label{curve.fig}
\end{figure}
At the first three epochs of the HST observations ($t$~= 52, 57 and 62~d), the derived magnitudes are consistent with the light curves of \citetalias{P19} within the uncertainties (Fig.~\ref{curve.fig}). AT2018cow's brightness evolves very rapidly; if we simply extrapolate the light curve tails, the brightness should decline to a very low level within a few months. Somewhat surprisingly, there is still a bright source at the position of AT2018cow at $t$~= 714 and 1136~d, or $\sim$2--3 years after its explosion. It is also worth noting that the magnitudes of the late-time source are strikingly stable, in both F555W and F814W bands, over a time span of $\sim$1.2~yr between the last two epochs. The magnitude differences are only 0.01~mag and 0.14~mag, respectively, much smaller than the photometric uncertainties.
\paragraph*{SED}
\begin{figure}
\centering
\includegraphics[width=1\linewidth, angle=0]{c_srcsed.sed.pdf}
\caption{SED of AT2018cow's late-time source (black data points); the horizontal error bars correspond to the root-mean-square widths of the HST filters and the vertical error bars reflect photometric uncertainties (if not smaller than the symbol size). The F657N detection limit does not provide additional constraint and is omitted in the plot. For comparison we show model spectra (colored lines) for (a) blackbodies, (b) single stars \citep{ck04.ref}, and (c) star clusters \citep[from \textsc{bpass} v2.1 binary population synthesis;][]{bpass.ref}, all reddened with AT2018cow's Galactic reddening and normalized to the F555W band. In (a) we also show the early-time SED of AT2018cow at $t$~= 4 d (open triangles) and 50 d (open squares) \textbf{normalized to the F555W band}; their error bars are no larger than the symbol size.}
\label{sed.fig}
\end{figure}
The late-time source has an SED even bluer than AT2018cow at early times (\citealt{P18}; \citealt{Kuin2019}; \citetalias{P19, X21}). In Fig.~\ref{sed.fig} the observed SED is compared with model spectra for blackbodies, single normal stars \citep{ck04.ref} and star clusters with a \citet{imf.ref} initial mass function (IMF) from \textsc{bpass} v2.1 binary population synthesis models \citep{bpass.ref}. The late-time source has significant UV excess even compared with the hottest stars of 50,000~K\footnote{If a star has strong wind, its continuum would appear redder and contamination due to emission lines is more significant for the optical filters than for the UV filters \citep{Gotberg2017}. Therefore, this effect is not able to explain the UV excess. We note, however, that the helium-burning Wolf-Rayet stars may reach significantly higher temperatures, up to 10$^5$~K \citep{Crowther2007}, than the hottest stars included in the \citet{ck04.ref} models.} or the youngest star clusters of 1~Myr\footnote{We also checked the effect of stochastic sampling on star clusters' SEDs with the \textsc{slug} package \citep{slug1.ref, slug2.ref}. This effect may make the SEDs even redder, due to some massive stars having evolved into red supergiants, and thus more inconsistent with observations.}. We find a power-law solution of $f_\lambda$~$\propto$~$\lambda^{-4.1 \pm 0.1}$ by fitting to the data, consistent with the Rayleigh-Jeans tail of very hot blackbodies.
\paragraph*{Position on the HR diagram}
\begin{figure}
\centering
\includegraphics[width=1\linewidth, angle=0]{c_srchrd.hrd.pdf}
\caption{In this HR diagram, we use a color-shaded region, for each HST band, to show the possible positions of blackbodies whose synthetic magnitudes match the observed late-time brightness of AT2018cow within 5$\sigma$ uncertainties. The four regions should converge together if the observed SED is indeed a blackbody spectrum across all four bands. This is met at log($T_{\rm eff}$/K)~$>$ 4.7 and log($L$/$L_\odot$)~$>$ 7.0, i.e. the light blue-shaded area in the upper-left corner. The grey dashed lines correspond to constant blackbody radii and the labelled values are in units of solar radius ($R_\odot$). The thin black solid lines are \textsc{parsec} v1.2S single stellar isochrones \citep{parsec.ref} and the thick black dashed line is the \citet{hdlimit.ref} limit for stellar luminosity. The ``$+$" symbols show some of the most luminous stars in the Milky Way (NGC~3603-A1a, A1b, B and C; in blue) and the Large Magellanic Cloud (R136a1--3 and b; in black; \citealt{Crowther2010}).}
\label{hrd.fig}
\end{figure}
In Fig.~\ref{hrd.fig} we show the possible positions of blackbodies on the HR diagram, whose synthetic magnitudes can match the observed ones for the late-time source. The result suggests that, if the late-time source is indeed a blackbody, it should have a very high temperature of log($T_{\rm eff}$/K)~$>$ 4.7 and luminosity of log($L$/$L_\odot$)~$>$ 7.0, corresponding to a blackbody radius of several tens of solar radii. The derived luminosity is significantly higher than any typical single stars, which is obvious by comparing it with the \citet{hdlimit.ref} limit\footnote{Recently, \citet{Davies2018} found evidence for an even lower luminosity limit for cool supergiants in the Magellanic Clouds.}, the \textsc{parsec} single-stellar isochrones \citep{parsec.ref}, and some of the most luminous very massive stars of $>$100~$M_\odot$ in the Milky Way or in the Large Magellanic Cloud (\citealt{Crowther2010}; see also \citealt{Bestenlehner2020}). Note, however, that the bolometric luminosity remains very uncertain since it is not clear whether the late-time source deviates from a blackbody SED outside the observed wavelength range.
\paragraph*{H$\alpha$ emisssion}
The F665N detection indicates prominent H$\alpha$ emission from the late-time source (Fig.~\ref{sed.fig}). Note that the wavelength of H$\alpha$ is redshifted away from F657N into the F665N band due to the recession of the host galaxy. By subtracting a continuum interpolated between the F555W and F814W bands, we estimate a wavelength-integrated line flux of $\sim$ 10$^{-17}$~erg~s$^{-1}$~cm$^{-2}$ or an H$\alpha$ luminosity of $\sim$4~$\times$ 10$^{36}$~erg~s$^{-1}$ (corrected for extinction). It is not clear whether the H$\alpha$ emission arise from stellar sources (e.g. WNh stars; \citealt{Schaerer1998}) and/or from a modest and compact H~\textsc{ii} region at the position of AT2018cow (for comparison, H~\textsc{ii} regions have H$\alpha$ luminosities ranging from 10$^{37}$~erg~s$^{-1}$ for the Orion Nebula to $>$10$^{40}$~erg~s$^{-1}$ for 30~Doradus; \citealt{Crowther2013}). With integral-field spectroscopy, \citet{L20} found AT2018cow to be associated with a giant H~\textsc{ii} region. The F665N image, which has a much higher spatial resolution, shows that the giant H~\textsc{ii} region is actually to the northwest of AT2018cow with an offset of $\sim$100~pc. Therefore, the detected F665N source is not due to the giant H~\textsc{ii} region in the environment of AT2018cow.
\section{Possible origins}
\label{origin.sec}
In this section we discuss what objects or physical mechanism(s) may give rise to the late-time brightness of AT2018cow.
\paragraph*{Chance alignment}
The detected late-time source is less likely due to an unrelated object in chance alignment. We performed relative astrometry with 11 reference stars between the late-time images and the F218W image at $t$~= 62~d. The (transformed) positions have an offset of only 0.3~pixel between AT2018cow and the late-time source, smaller than the astrometric uncertainty of 0.5~pixel. Additionally, there are only 22 sources significantly detected within 1" from AT2018cow (Fig.~\ref{image.fig}k); the probability is $<$1\% for a randomly positioned source to coincide with AT2018cow within the 1$\sigma$ error radius. Moreover, very few sources in the surrounding region can match AT2018cow's late-time brightness and its very blue SED. Therefore, we suggest the late-time source is not in chance alignment but physically associated with AT2018cow.
\paragraph*{Light echo}
It is very difficult to explain AT2018cow's late-time brightness with a light echo. The H$\alpha$ emission is inconsistent with AT2018cow's featureless spectrum near peak \citepalias{P19, X21}, and it may require a very specific dust configuration to reproduce its very blue SED. For circumstellar dust from a steady wind, the brightness of its light echo declines continuously \citep{Chevalier1986}, inconsistent with the observed stable F555W/F814W brightness from $t$~= 714~d to 1136~d. A foreground dust sheet may produce a stable light echo brightness if it is distant enough from the transient \citep{Cappellaro2001}. However, AT2018cow has a negligible amount of dust in its foreground since observations find very little extinction from within the host galaxy (\citetalias{P19, X21}). For dust in the background, the scattering is much less efficient due to the large scattering angles \citep{Draine2003}. Therefore, it seems unlikely that the late-time source is a light echo.
\paragraph*{Stellar source(s)}
A late-time source with stable brightness is naturally expected if the transient's progenitor is located inside a binary/multiple system \citep{Maund2004, Maund2016, Sun2020a, Sun2022} or star cluster \citep{Sun2020b}. For AT2018cow, the late-time source is inconsistent with a single companion star; the very high luminosity requires it to be at least 2--3 very massive ($>$100~$M_\odot$) stars inside a multiple system (Fig.~\ref{hrd.fig}). Star clusters with standard IMFs are difficult to match the observed SED (Fig.~\ref{sed.fig}); if the late-time source were indeed a star cluster, it may have a very top-heavy IMF \citep[e.g.][]{30dor.ref} so that the brightness is dominated by the very massive stars.
It is unclear if FBOTs mark the death of massive stars or arise from their non-terminal explosions. In a PPISN, for example, material ejected at a later epoch may collide with earlier ejecta, producing a luminous transient, but the progenitor may not have completely destroyed itself and can still be observed at late times \citep{Woosley2007}. It is, however, still difficult to explain the high luminosity of AT2018cow's late-time source unless the progenitor is overluminous due to some special mechanisms. In a PPISN model of AT2018cow, \citet{Leung2020} found a progenitor mass of $M_{\rm ini}$~$\sim$ 80~$M_\odot$ in order to fit the fast evolving light curves.
In these scenarios, the H$\alpha$ emission may arise from stars with strong winds or from the local ionized gas, and AT2018cow is (probably) related to a very young and massive progenitor. Caution, however, that some SNe spatially aligned with young star clusters are found to have much older progenitors due to sequentially triggered star formation \citep[e.g. SN~2012P;][]{Sun2021}.
\paragraph*{Late emission from AT2018cow}
AT2018cow has been proposed to be a core-collapse SN powered by CSM interaction (\citealt{Fox2019}, \citetalias{X21}). In the model of \citetalias{X21}, AT2018cow was surrounded by CSM with a flat density profile from 3 to 1200~$R_\odot$. If the late-time source were also due to CSM interaction, then the CSM should extend out to a much larger distance of $\gtrsim$~5~$\times$ 10$^{16}$~cm (for an ejecta velocity of 5000~km/s); the CSM density may also be different at larger distances in order to explain the significant light curve flattening from early to late times. In this scenario, the late-time emission may be contributed by emission lines. However, it may be challenging to reproduce the very stable brightness from $t$~= 714 to 1136~d; most interacting SNe have time-varying brightness and their light curve plateaus, if any, last for much shorter periods (\citealt{Mauerhan2013a, Mauerhan2013b}; but see \citealt{Smith2009} for the exceptions of extremely enduring SN~1988Z and SN~2005ip).
In a magnetar-powered model of AT2018cow, \citet{P18} derived a spin period of 11~ms and a magnetic field strength of 2.0~$\times$ 10$^{15}$~G, corresponding to a magnetic dipole spin-down timescale of 0.3~d \citep{Fang2019}. In this case, the energy injection rate is expected to decline by 60\% from $t$~= 714~d to 1136~d, inconsistent with the observed very stable late-time brightness.
AT2018cow has also been suggested to arise from a low-mass star disrupted by an intermediate-mass black hole (\citealt{Liu2018, Kuin2019}; \citetalias{P19}). In such a TDE, the stellar debris is usually assumed to be swallowed within a few orbits, giving rise to a $t^{-5/3}$ declining light curve after peak \citep{Rees1988}. We might expect some late-time brightness if there is significant time delay for some of the debris to be accreted. In this case, a constant accretion rate is needed to explain the stable F555W/F814W brightness at 714--1136~d; the H$\alpha$ detection also requires the disrupted star to be hydrogen-rich.
\section{Summary}
In this paper we report the discovery of a late-time source of the FBOT AT2018cow at 714~d and 1136~d after its explosion. It has a very stable brightness between the two epochs and a very blue SED consistent with the Rayleigh-Jeans tail of a blackbody spectrum with temperature log($T$/K)~$>$ 4.7 and luminosity log($L$/$L_\odot$)~$>$ 7.0. Significant H$\alpha$ emission is also detected. This late-time source is unlikely an unrelated object in chance alignment, or due to a light echo of AT2018cow. We discussed other possible scenarios as stellar source(s) (companion stars, star cluster, or the survived progenitor star) or due to AT2018cow's late emission (CSM interaction, magnetar, or delayed accretion in a TDE). All these scenarios have some difficulties in explaining this late-time source and require some fine tuning or special mechanisms to match the observations. It cannot be ruled out, however, that the late-time brightness is contributed by multiple components and/or processes. We believe this late-time source may contain important information for AT2018cow, but long-term monitoring and multi-wavelength observations are required to resolve its origin.
\section*{Acknowledgements}
We thank the anonymous referee for the very helpful comments on our paper. Research of N-CS and JRM is funded by the Science and Technology Facilities Council through grant ST/V000853/1. L-DL is supported by the
National Key R\&D Program of China (2021YFA0718500). This paper is based on observations made with the NASA/ESA Hubble Space Telescope and has used the light curves published by \citetalias{P19}.
\section*{Data availability}
Data used in this work are all publicly available from the Mikulski Archive for Space Telescope (\url{https://archive.stsci.edu}) or from the paper by \citetalias{P19}.
| 478b805188bc81f1650a7fc25e05838b412472b1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Alternative LYSO Calorimeter Considerations}
A naturally segmented array of tapered LYSO crystals provides an attractive alternative to our proposed LXe-based calorimeter. We are exploring a geometry that matches that of the PEN pure CsI detector. The PEN crystals were limited to 12$X_0$ depth, which is insufficient for an ideal $\pi \rightarrow e \nu$ measurement. However, the segmentation and fast response allows for various trigger patterns and it is well-designed for the pion beta decay phase in the PiBeta configuration. This detector has an inner radius of 26\,cm, and an outer radius of 48\,cm. CsI has a radiation length of 1.86\,cm. With the compact ATAR geometry we are proposing, a sufficient volume exists to insert an inner array of crystals between the ATAR and the existing PEN CsI array, see Fig.~\ref{fig:detector_PEN} and Fig.~\ref{fig:LYSO-figs}a. On paper, LYSO crystals appear to be the ideal choice for such an array. LYSO is a Cerium doped Lutetium based scintillator whose light output is comparable to doped NaI(Tl). It has high density ($X_0$= 1.14\,cm, $R_{M}$=2.07 cm) and a light yield comparable to the highly luminous NaI(Tl), but with much faster light signals. Its 420\,nm typical scintillation light has a 40\,ns single exponential decay time and the spectrum is well matched to conventional SiPM photosensors. LYSO is both radiation hard and non-hygroscopic. To date, the main application for LYSO crystals has been in PET medical scanners, where small crystals are needed with high light output to resolve the 511\,keV gammas from positron annihilation. The attractive properties of these crystals suggest that larger arrays could be made for use in particle physics applications and, indeed, several groups have explored this possibility with small arrays and somewhat limited success.
Experts in the field, such as Ren-Yuan Zhu from Caltech, have advised our collaboration on the pros and cons and experiences of other groups, see for example \cite{Zhu:2021sij}. The growth of relatively long LYSO crystals is a fairly new and expensive R\&D effort, not easily justified without a demanding end-use application. The Mu2e Collaboration built a $5 \times 5$ array of $3 \times 3 \times 20\,$cm$^3$ crystals and subjected the array to test beams at Frascati and Mainz~\cite{Atanov:2016blu}. At 100\,MeV, the resolution was in the range of 3--4\%, dominated by a 2.5\% constant term indicative of non-uniformity in light transmission and possibly crystal to crystal calibrations and surface preparation. That performance met the specifications for Mu2e, but they did not push it further, choosing eventually to use undoped CsI based on an overall cost analysis.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{Figures/LYSO-figs.pdf}
\caption{Possible use of an inner array of tapered LYSO crystals within the open volume of the existing PEN CsI calorimeter. a) Opened view showing in blue the array of LYSO crystals that matches one-to-one to the existing geometry of the PEN crystals shaded in gray. b) An example array ideal for testing the concept. c) An individual pentagonal crystal, 16$X_0$ in depth. Each such crystal would be read out by a thin array of SiPMs.
}
\label{fig:LYSO-figs}
\end{figure}
Dr. Zhu has investigated crystals from various companies and published properties that will guide us going forward. An example of a study of a 25$X_0$ crystal's longitudinal transmission and how to improve it is found in \cite{Mao:2012dr}. It is imperative to improve on the uniformity of light production and transmission along the length of the crystal. Experience has shown -- as it has for CsI -- that custom surface preparation on tapered crystals is also important and can be realized with careful lab bench work. We aim to investigate the possibilities of using LYSO for PIONEER, but would not now claim this to be defensible until proper bench tests and manufacturer quotes are in hand. Figures\,\ref{fig:LYSO-figs}b and c indicated prototype geometries that are designed to fit inside the PEN calorimeter. We will issue requests to a variety of companies for quotes to see if these crystal shapes can be made to meet our specifications.
If tests are successful --- principally achieving the needed energy resolution below 3\% -- then a careful side-by-side comparison of costs and performance against our leading calorimeter candidate based on LXe can be made.
\section{In-target Decay-in-flight $\pi \rightarrow \mu \nu$ Suppression with BDT}
\label{sec:DIFsuppression_BDT_appendix}
\begin{itemize}
\item plots of input features (pienu vs DIF)
\item BDT score distribution
\item ROC curve
\item suppression factor with the consideration of energy smearing
\end{itemize}
\section{PiBeta}
{\bf Pion beta decay measurements}
The branching ratio for pion beta decay was most accurately measured by the PiBeta experiment\footnote{The PiBeta and PEN experiments shared much of the same apparatus.} at PSI \cite{Pocanic1,Frlez:2003vg,Pocanic:2003pf,Frlez:2003pe,Bychkov:2008ws} to be
$\frac{\Gamma(\pi^+ \to \pi^0 e^+ \nu)}{\Gamma\textrm{(Total)}}= [1.036 \pm 0.004 \textrm{(stat)} \pm 0.004\textrm{(syst)} \pm 0.003(\pi\to e\nu)] \times 10^{-8}$, where the first uncertainty is statistical, the second systematic, and the third is the $\pi\to e\nu$ branching ratio uncertainty.
Pion beta decay potentially provides the theoretically cleanest determination of the magnitude of the CKM matrix element \ensuremath{\left|V_{ud}\right|}\xspace. With current input one obtains $\ensuremath{\left|V_{ud}\right|}\xspace = 0.9739(28)_{\textrm{exp}}(1)_{\textrm{th}}$, where the experimental uncertainty comes almost entirely from the $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ branching ratio (BRPB).
The theory uncertainty has been reduced from $({\delta}V_{ud})_{\textrm{th}} = 0.0005$ \cite{Sirlin:1977sv, Cirigliano:2002ng, Passera:2011ae} to $({\delta}V_{ud})_{\textrm{th}} = 0.0001$ via a lattice QCD calculation of the radiative corrections \cite{Feng:2020zdc}. The current precision of \unit[0.3]{\%} on \ensuremath{\left|V_{ud}\right|}\xspace makes $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ not presently relevant for the CKM unitarity tests because super-allowed nuclear beta decays provide a nominal precision of
\unit[0.03]{\%}.
In order to make $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ important for CKM unitarity tests, two precision experimental stages can be identified:
(1) As advocated in Ref.~\cite{Czarnecki:2019mwq}, a three-fold improvement in BRPB precision compared to Ref.~\cite{Pocanic:2003pf} would allow for a 0.2\% determination of $\left|V_{us}/V_{ud}\right|$ improving on measurement of the following ratio being currently
$ R_V = \frac{\Gamma\left(\textrm{K} \rightarrow \pi l \nu (\gamma) \right)}{\Gamma\left(\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)\right)}=1.3367(25)$ ,
independent of the Fermi constant, short-distance, and structure-dependent radiative corrections. This
would match the precision of the current extraction of $\left|V_{us} / V_{ud}\right|$ from the axial channels~\cite{Marciano:2004uf}
$R_A = \frac{\Gamma\left(\textrm{K} \rightarrow \mu \nu (\gamma) \right)}{\Gamma\left(\pi \rightarrow \mu \nu (\gamma)\right)}=1.9884(115)(42)$,
providing a new competitive constraint on the \ensuremath{\left|V_{us}\right|}\xspace--\ensuremath{\left|V_{ud}\right|}\xspace\ plane and probing new physics that might affect vector and axial-vector channels in different ways.
The theoretical case for this approach was recently strengthened by improved analysis of radiative corrections in $K \to \pi e \nu $ decays \cite{Seng:2021nar}.
(2) In the second phase, an order of magnitude improvement in the
BRPB precision will be sought. This would provide the theoretically cleanest extraction of \ensuremath{\left|V_{ud}\right|}\xspace at the \unit[0.02]{\%} level.
\section{ PIENU and PEN}
{\bf Measurements of $R_{e/\mu}$ and associated exotic searches.}
\label{Appendix}
The PIENU experiment has provided the most precise measurement of the branching ratio $R_{e/\mu}=(1.2344\pm0.0023_{stat}\pm0.0019_{sys})\times10^{-4}$\cite{PiENu:2015seu}; a further factor two improvement in precision is anticipated. The $\pi\to e\nu$ branching ratio provides the best test of electron–muon universality in charged current weak interactions resulting in the ratio of weak interaction strengths $\frac{g_\mu}{g_e}=1.0010\pm0.0009$\cite{Bryman:2021teu}. The PEN experiment at PSI is aiming at comparable precision to PIENU.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.485]{Figures/Analysis.png}
\caption{The upper panel shows the positron energy spectrum with the red line indicating $E_{cut}$. The lower panels show the time distributions for events below and above $E_{cut}$. The black histograms are data, the red curve is the $\pi^+ \to \mu^+ \to e^+$ signal, and the blue line is the $\pi^+ \to e^+ \nu$ signal. The other histograms in various colors are the background terms related to pile-up, muon DIF, and other effects discussed in Ref.~\cite{PiENu:2015seu}.}
\label{fig:analysis}
\end{figure}
PIENU obtained the branching ratio by first separating events into high- and low-energy regions at an energy cut value ($E_{cut}$) as illustrated in Fig. \ref{fig:analysis}. The time spectra were fit in each region with the $\pi^+ \rightarrow e^+ \nu$ and $\pi^+ \rightarrow \mu^+ \rightarrow e^+$ shapes, plus backgrounds originating from different sources including pion decays in flight, contamination from old muon decays etc. The raw branching ratio $R^{raw}_{e/\mu}$ was the ratio of the $\pi^+ \rightarrow e^+ \nu$ amplitude to the $\pi^+ \rightarrow \mu^+ \rightarrow e^+$ amplitude. Corrections such as the tail correction for low energy events below the Michel spectrum, were subsequently applied to obtain the final value.
High precision pion decay experiments also provide a plethora of constraints on exotic phenomena including heavy neutrinos and dark sector processes.
Extensions of the Standard Model postulate the
existence of additional (sterile) neutrinos \cite{Boyarsky:2009ix, Bryman:2019bjg}. These additional states may
contribute to the solution of outstanding puzzles like
the nature of dark matter, early cosmological processes like small scale structure formation \cite{Bertoni:2014mva}, and Mesogenesis \cite{Elahi:2021jia}.
Massive neutrino states $\nu_H$ are sought in the two-body pion decays
$\pi^+\rightarrow e^+\nu_H$ \cite{PIENU:2017wbj} and $\pi^+\rightarrow \mu^+\nu_H$ \cite{PIENU:2019usb}. Exploiting
large data sets of pion decays and the resulting decay muons, exotic two-body muon decays like $\mu^+\rightarrow e^+X$ can
be sought \cite{PIENU:2020loi}, where X is a massive neutral boson (e.g. an axion or a Majoron).
Similarly, exotic particles have been searched for in three body decays like $\pi^+\rightarrow l^+ \nu X$ ($l=e^+,\mu^+$) \cite{PIENU:2021clt}.
The PIENU experiment also obtained upper limits for the rare decays $\pi^+\rightarrow e^+\nu_e\nu\bar{\nu}$ and
$\pi^+\rightarrow \mu^+\nu_{\mu}\nu\bar{\nu}$ at the $10^{-7}-10^{-6}$ level \cite{PIENU:2020las}.
PIONEER~ with two orders of magnitude more statistics has the potential to improve the existing limits by at least an order of magnitude.
Since the searches are based on fits to the energy spectra of the visible final state particles, an improved experiment
can bring significant additional advantages in lowering the limits and in reducing the systematic errors. For example,
the $\pi^+\rightarrow e^+\nu$ low energy tail represents the main background for the
$\pi\rightarrow e^+\nu_H$, $\pi^+\rightarrow e^+\nu X$, and $\pi^+\rightarrow e^+\nu_e\nu\bar{\nu}$ searches:
more precise knowledge of the tail and its further reduction will significantly improve the upper limits beyond the statistics.
The search for rare and exotic decays involving muons, like $\pi^+\rightarrow \mu^+\nu_H$, $\pi^+\rightarrow \mu^+\nu X$, and $\pi^+\rightarrow \mu^+\nu_{\mu}\nu\bar{\nu}$ will
benefit from an improved stopping target and faster electronics, which will allow better separation of muons from pions and thus further improve the sensitivity.
The PEN/PiBeta and PIENU experiments relied on inorganic scintillator calorimetry.
PIENU used a high-resolution ($\sigma=1\%$) crystal calorimeter consisting of a single crystal NaI(Tl) detector surrounded by an array of 97 pure CsI crystals for shower leakage containment. The PIENU detector is shown in Fig.~\ref{fig:detector_PIENU} and described in \cite{PiENu:2015pkq}. The large NaI(Tl) crystal was 19$X_0$ lengths thick and 19$X_0$ in diameter. The high energy resolution and long radiation-lengths of the Na(Tl) crystal were essential for reducing the low energy tail. However, the slow decay constant of NaI(Tl) limited pile-up detection and rejection.
The acceptance of the PIENU detector was relatively small $<\unit[20]{\%}$ which resulted in an important source of systematic uncertainty. The PEN experiment, on the contrary, adopted a high solid angle geometry (see Fig.~\ref{fig:detector_PEN}). Its key components were a highly segmented (240 elements) spherical pure CsI crystal calorimeter covering $\sim \unit[3\pi$]{sr} of solid angle around the pion stopping target. The key limitations were related to the imperfect separation of
$\pi\rightarrow \mu \rightarrow e$ and $\pi\rightarrow e\nu$ decays. The primary culprit was the 12$X_0$ thickness of the CsI calorimeter, which produced a substantial low energy tail for \unit[70]{MeV} positrons (and photons)
extending well under the $\pi\rightarrow \mu \rightarrow e$ spectrum.
The PIONEER approach using a high resolution, uniform response LXe calorimeter with fast timing and high solid angle combines the assets of both experiments.
\begin{figure}[!tbp]
\centering
\subfloat[]{\includegraphics[width=0.5\textwidth]{Figures/PIENUDetector.png}\label{fig:detector_PIENU}}
\hfill
\subfloat[]{\includegraphics[width=0.5\textwidth]{Figures/PEN_apparatus_gen.pdf}\label{fig:detector_PEN}}
\caption{(a) Schematic view of the PIENU detector. Plastic scintillators are shown in dark blue, wire chambers in green, silicon strip trackers in orange and the calorimeter in light blue and red. (b) Schematic cross section of the PEN detector, with a view of the CsI crystal calorimeter.}
\end{figure}
\subsubsection{Data acquisition system}
The PIONEER data acquisition system must handle the read out, event assembly and data storage for the active target, positron tracker, electromagnetic calorimeter and other detector sub-systems of the experimental setup. It must provide a deadtime-free, distortion-free record of the datasets identified by the various physics and calibration triggers. It must facilitate the monitoring needed to guarantee the overall integrity of data taking and provide the metadata needed to document the experimental configuration during data taking. Finally, it must enable the configuration of the readout electronics and the associated trigger, clock and control system.
The acquisition will be implemented as a modular, distributed system on a parallel, layered processor array consisting of networked, multi-core, commodity PC's running a scientific Linux operating system. The overall layout is depicted schematically in Fig.\ \ref{f:daqlayout}. It will comprise a frontend processor layer responsible for readout and processing of event fragments from the FPGA-based fast electronics instrumenting the various detector sub-systems, a backend layer responsible for event building and data storage, a slow control layer responsible for configuration and readback, and a data quality layer responsible for monitoring of data integrity.
To maximize the bandwidth the frontends will utilize memory-mapped, Gen3 PCI Express communication with the Apollo readout boards and the downstream DAQ layers will utilize memory-mapped, Gen3 PCI Express host / target adaptors for data transfer and sharing system resources (disks, GPUs, {\it etc}). PCI-Express 3.0 delivers an effective data transfer rate of nearly 1 GB/s per serial lane. Commercial host-to-target optical interfaces are now available for the Apollo FPGA to frontend computer communication ({\it e.g.}, www.dolphinics.com/products/PXH842.html). Commercial peer-to-peer communication is also now available for direct PCI-Express data transfer between devices ({\it e.g.}, www.dolphinics.com/products/PXH810.html). The setup will offer the capability of direct transfer of data to / from FPGA and CPU / GPU memory for rapid realtime processing.
\begin{figure}[t!]
\includegraphics[width=0.99\textwidth]{Figures/DAQschematic.png}
\caption{Schematic of the data acquisition system showing the frontend layer for data readout and experiment configuration, the backend layer for event assembly and data storage, and the analysis layer for data quality monitoring. The number of frontends and the topology of the FPGA-to-frontend and frontend-to-backend networks will be based on the calorimeter, ATAR and FPGA technology choices.}
\label{f:daqlayout}
\end{figure}
The DAQ software will be based on the MIDAS data acquisition package \cite{midas}, CUDA GPU toolkit \cite{cuda}, ROOT data analysis package \cite{root}, and Linux PCI Express system and utility libraries. The MIDAS software consists of library functions for data flow between different processes on local / remote devices as well as infrastructure for data logging, experimental configuration and local /remote run control. It also incorporates an integrated alarm system and slow control system. The devices drivers for the configuration and the readout of the Apollo board FPGAs will be based on the PCI Express communication protocols / libraries.
As discussed in Sec.~\ref{Trigger} and Table~\ref{tab:triggers}, we anticipate the various physics triggers to include a prescaled beam trigger \mbox{\fontfamily{lmss}\selectfont PI}, a beam particle - high energy \mbox{CALO}\ trigger \mbox{\fontfamily{lmss}\selectfont CaloH}, a prescaled beam particle - low threshold \mbox{CALO}\ trigger \mbox{\fontfamily{lmss}\selectfont PROMPT}, and a beam particle - outgoing positron trigger \mbox{\fontfamily{lmss}\selectfont TRACK} ~(as identified by either the ATAR or the tracker). The cumulative trigger rate in Table~\ref{tab:triggers} is estimated at approximately 8.8\,kHz.
The ATAR data will comprise above-threshold islands of contiguously digitized ADC samples from the fired detector strips of the active stopping target. Based on a 1 GSPS sampling rate, 4 Bytes per sample, and 66 above-threshold strips per trigger, we anticipate an \mbox{ATAR}\ data rate of about 70~MB/s for the aforementioned trigger types.
The CALO data will comprise islands of contiguously digitized ADC samples from the SiPM / PMT readout of the calorimeter. Based on a 1 GSPS sampling rate, 2 Bytes per sample, 200 samples per island and 1000 islands / channels per trigger, we anticipate a \mbox{CALO}\ data rate of about 3.5 GB/s for the aforementioned trigger types.
Other detector systems will include the various beam detectors upstream of the ATAR and the positron tracking detector surrounding the ATAR. The data rates from beam detectors and tracking detector are expected to be a small fraction of the ATAR / CALO data rate.
Separate frontend processors running MIDAS frontends will readout the ATAR, calorimeter, tracker and beam sub-systems via the PCI-express bus using commericial transparent host / target PCI Express adapters to the Virtex FPGAs on the Apollo readout boards. The memory-mapped readout from the Virtex FPGAs chip will permit both high data transfer rates and direct data transition to system CPU and GPU memories. We plan to utilize IPBus over PCI Express for the configuration of the readout electronics and the trigger, clock and control system.
The data acquisition system will process in real-time the data from a roughly 3.5 GB/sec raw data rate to a roughly $\sim$300 MB/sec processed data rate for data storage on PSI's Petabyte archive. One option for the data processing is the lossless compression of the slow decay-time calorimeter signals via a combination of delta encoding and Golomb coding. The delta encoding will exploit the correlations between sequential ADC samples to shrink the word usage distribution while Golomb coding is suited to situations in which the occurrences of small values are significantly more likely than the occurrences of large values. Other possibilities are zero suppression of \mbox{CALO}\ islands and realtime fitting of \mbox{CALO}\ pulses. These algorithms are well suited to parallel processing using GPUs.
We are initiating an R\&D effort to demonstrate both the technology for the FPGA-to-CPU / GPU communication via optical PCI-express links and the performance of the data compression schemes. The R\&D setup will allow for code development, testing and debugging as well as the evaulation of the rate capabilities and the compression capabilities of the system. We plan to use commercial PCI-Express FPGA development boards as data simulators for the detector sub-systems in order to prototype and benchmark the DAQ.
The PIONEER DAQ group has developed and operated similar architectures of distributed data acquisition systems for the MuLan, MuCap and MuSun experiments at the Paul Scherrer Institute and the g-2 experiment at Fermi National Accelerator Lab.
\subsubsection{Data Policy}
The PIONEER collaboration will comply with the \emph{data policy for PSI research data} \cite{PSIdataPolicy} and will publish the experiment's data under PSI's custody after a suitable embargo period. We are also committed to publish our software under an open source license at a suitable time.
\section{ATAR technical details}
\label{sec:ATAR_appendix}
As introduced in Sec.~\ref{sec:ATAR}, the highly segmented active target (ATAR) is a key new feature of the proposed PIONEER experiment which will define the fiducial pion stop region, provide high resolution timing information, and furnish selective event triggers. Advanced technical details and studies will be covered by this Appendix.
A brief summary of the baseline ATAR design follows.
The ATAR tentative design dimensions are 2 $\times$ 2\,cm$^2$ transverse to the beam, in the beam direction individual silicon sensors are tightly stacked with a total thickness of roughly 6\,mm.
Planes of the ATAR are composed of LGAD sensors, for which a thickness of 120\,$\mu$m and a strip geometry with a pitch of 200\,$\mu$m is foreseen.
The detectors are paired with the high-voltage facing each other to avoid ground and high voltage in proximity. These strip sensor pairs are oriented at 90$\degree$ to each other in subsequent staggered planes to provide measurement of both coordinates of interest and allow space for the readout and wirebonds.
In the preliminary design, the strips are wire bonded to a flex, alternating the connection on the four sides of the ATAR, that brings the signal to a readout chip.
This positions the chip a few cm away from the active volume and outside of the path of the exiting positrons, reducing the degradation of their energy resolution.
The readout ASIC sits on the first flex that tapers out to accommodate the additional traces. Then this flex is connected, via connector, to a PCB connected to a second flex that brings the amplified signal to the digitizers in the back end.
A schematic drawing of the ATAR is shown in Sec.~\ref{sec:ATAR}, Fig.~\ref{fig:ATAR_scheme}.
\subsection{LGAD technology}
\label{sec:LGAD_app}
The chosen technology for the ATAR is based on Low Gain Avalanche Detectors (LGAD) \cite{bib:LGAD}, thin silicon detectors with moderate internal gain. LGADs are composed by a low doped region referred to as `bulk', typically 50\,$\mu$m thick, and a highly doped thin region, a few $\mu$m from the electrodes, called `gain layer'. The electric field in the gain layer is high enough to generate charge multiplication from electrons but not from holes: this mechanism allows to have moderate charge multiplication (up to 50) without generating an avalanche. A schematic cross-section of an LGAD can be found in Fig.~\ref{fig:Pulse} (Left). Due to the internal gain and thin bulk, LGADs have fast rise time and short full charge collection time.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/LGAD_concept.png}
\includegraphics[width=0.54\textwidth]{Figures/120um_pulse.png}
\caption{Left: LGAD schematic cross-section with electric field on the left. Right: Simulated pulse shape for a MIP in a 120~$\mu$m thick LGAD (Weightfield2 \cite{WF2}). The black line is the response of a 2\,GHz bandwidth electronic readout.}
\label{fig:Pulse}
\end{figure}
The best estimate at present for the sensor thickness is around 120\,$\mu$m to avoid support structures for the sensor, which would introduce dead areas and inactive material within the beam. Normally LGADs have a thickness of 50\,$\mu$m or less, so the exact sensor performance of a non-conventional 120\,$\mu$m-thick LGAD will need to be established by tests on more mature prototypes.
Using fast electronics is expected to result in a pulse rise time of about 1\,ns for an LGAD of this thickness, as shown in Fig.~\ref{fig:Pulse} (Right). The time resolution on the rising edge should be less than 100\,ps for a minimum ionizing signal, down to much better time resolution for large $\pi/\mu$ signals.
Such a sensor would be able to separate two closely overlapping hits if they arrive more than 1.5\,ns apart.
The LGAD technology was chosen over standard Silicon technology because of the intrinsic gain and thin bulk.
Normal silicon detectors without gain need 300--500\,$\mu$m of active silicon to have enough collected charge for high MIP detection efficiency (unless for the case of very small pixels, where a lower capacitance and noise allows 150\,$\mu$m of thickness).
This would reduce the achievable granularity in the direction of the beam and increase significantly the pulse width and therefore reduce the temporal pulse separation discrimination.
The increased S/N of LGADs is also helpful for signal transmission through the first flex to the ASIC.
HV-CMOS was also considered as a possibility, however monolithic pixels usually have non-depleted regions between pixels, and furthermore to achieve the same level of time resolution, the power would need to be increased to a level that is not sustainable for the ATAR.
If the amplifier chip is removed from the compact region of the ATAR as in the case of the current design, heat dissipation from the readout electronics is less problematic.
3D sensors have a similar issue of non-depleted regions, furthermore the temporal pulse separation discrimination would be depending on the strip pitch.
On top of this, LGADs can be produced by several small and research foundries in addition to commercial vendors, making the prototyping phase easier and reducing the risk of production delays in case of large-scale availability bottlenecks at big companies.
Current standard LGADs are limited in terms of granularity and active area due to a protection structure (junction termination extension) at the edge of the high field area of the gain layer, necessary to avoid breakdown of the sensor.
A typical width for such an inter-pad gap in an LGAD is around 50--100\,$\mu$m, and thus limits the granularity to the mm scale for regular LGADs.
To achieve a ~100\% active area, several technologies still at prototype level are being evaluated for PIONEER, such as AC-LGADs~\cite{Apresyan:2020ipp} (studies shown in Appendix~\ref{sec:ACLGAD_app}), TI-LGADs~\cite{9081916} (studies shown in Appendix~\ref{sec:TILGAD_app}) and DJ-LGAD~\cite{Ayyoub:2021dgk}.
The DJ-LGAD is the newest high-granularity design and no prototype exists at the moment, however a prototype run is expected to be finished and tested by Q1 2022.
Radiation damage is not expected to be an issue for the initial phases of PIONEER. For the $\pi\to e \nu$~ phase a total fluence of $10^{13}$ is expected, this is at an order of magnitude away from where radiation damage would significantly degrade the performance of LGADs~\cite{CERN-LHCC-2020-007}.
However $10^7$ pions per second are expected in the pion beta decay phase, this adds up to $10^{14}$ for a full year of data taking.
The integrated radiation damage of several $10^{14}$ for the full run will start to affect LGADs performance~\cite{CERN-LHCC-2020-007}, although the fluence is still moderate in comparison to other LGAD applications.
A simple solution to reduce the radiation damage is the substitution of the ATAR at the mid-life of the experiment.
\subsection{AC-LGAD preliminary studies}
\label{sec:ACLGAD_app}
AC-LGADs overcome the granularity limitation of traditional LGADs and have been shown to provide spatial resolution of the order of tens of $\mu$m~\cite{Tornago:2020otn}. This remarkable feature is achieved with an unsegmented (p-type) gain layer and a resistive (n-type) N-layer. An insulating dielectric layer separates the metal readout pads from the N+ resistive layer. This design also allows to have a completely active sensor with no dead regions.
AC-LGADs have intrinsic charge sharing between AC metal pads, so the signal can be picked up by multiple channels at the same time.
In a low hit-density environment, such as PIONEER, this allows to have a sparse electrode distribution but with elevated position resolution together with a 100\% active area.
Charge depositions can be then reconstructed by calculating the fraction of signal present in all nearby electrodes. With a strip geometry, the hit reconstruction can be made using two neighboring strips.
The following studies were conducted on strip AC-LGAD prototypes from Brookhaven National Lab (BNL) (Fig.~\ref{fig:FNAL_data}, Left).
The sensors have been tested with a laboratory IR laser TCT station~\cite{Particulars} and at a Fermilab (FNAL) test beam~\cite{ACLGADpico}.
In both setups the sensors are mounted on fast analog amplifier boards (16 channels) with 1\,GHz of bandwidth (designed at FNAL), the board is read out by a fast oscilloscope (2\,GHz, 20Gs).
The response of two strips of a 200\,$\mu$m pitch BNL AC-LGAD as a function of position can be seen in Fig.~\ref{fig:FNAL_data} (Right).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{Figures/sensor.png}
\includegraphics[width=0.5\textwidth]{Figures/FNAL_TB.png}
\caption{Left: prototype BNL AC-LGAD strip sensor with 80\,$\mu$m wide strips and pitch of (left to right) 100, 150, 200\,$\mu$m. Right: sensor response ($P_{max}$) as a function of position (perpendicular to the strip) of two strips with 200~$\mu m$ of pitch~\cite{ACLGADpico}. The dashed lines highlight the position of the two strips in the plot. Data taken at the FNAL 120\,GeV proton test beam facility.}
\label{fig:FNAL_data}
\end{figure}
The position resolution is calculated from the signal fraction as function of position between adjacent strips:
\begin{equation}
\label{Pos_resolution}
\sigma(pos) = \sqrt{2} \frac{d(\textrm{Position})}{d(\textrm{Fraction})}\left(\frac{S}{N}\right)^{-1}
\end{equation}
The position resolution perpendicular to the strips for the studied 200\,$\mu$m pitch is 5--15\,$\mu$m across the sensor, a few \% of the pitch as seen in Fig.~\ref{fig:BNL_resolution}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{Figures/BNL_strips_combined.png}
\caption{Signal charge sharing and position resolution of AC-LGAD strip sensors fabricated by BNL, assuming a constant S/N of 60. Left: fraction of the maximum signal amplitude ($P_{max}$) measured in one channel with respect to the summed signal of two neighboring strips. Data taken at the FNAL 120\,GeV proton test beam facility.
Right: position resolution over the respective strip pitch, determined with Eq.~\ref{Pos_resolution}. The zero point is in the middle of the gap between two strips.
}
\label{fig:BNL_resolution}
\end{figure}
Strips were tested with readout connected at both ends, for this particular configuration the signal is additionally split between the two ends of the strip.
By applying the same fractional method of charge sharing, it is possible to reconstruct the hit position also in the direction parallel to the strip, as seen in Fig.~\ref{fig:double_side} (Left).
A precision of a few hundred $\mu$m was found with a 2\,mm strips, giving a position resolution of 10\% of the strip length.
Having a double-connection readout setup would be more challenging in terms of mechanics, but would allow X-Y charge deposition reconstruction.
The envisioned metal size for the 200\,$\mu$m pitch strips, as well as other parameters of interest for the sensor such as the doping profile, needs to be confirmed after a testing R\&D campaign and TCAD simulations as shown in Appendix~\ref{sec:TCAD_app}.
In addition, some non-typical detector setup configurations, such as zig-zag strips, can be explored to identify muons traveling parallel to a strip, schematic shown in Fig.~\ref{fig:double_side} (Right).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{Figures/double_side.png}
\includegraphics[width=0.3\textwidth]{Figures/zigzag.png}
\caption{Left: response of an AC-LGAD strip read out from both sides. Right: muon escaping parallet to the strip direction detected with zig zag strip geometry.}
\label{fig:double_side}
\end{figure}
As already stated, a dynamic range from MIP (positron) to several MeV (pion/muon) of deposited charge is expected in the ATAR.
High ionizing events might affect the charge sharing mechanism, inducing charge also in electrodes far away from the charge deposition.
Since the event reconstruction relies on temporal pulse separation, the response to successive MiP and high charge deposition have to be studied \cite{Pulserep}.
Furthermore the effect of gain suppression for large charge deposition in LGADs has to be taken into account \cite{gainsuppr}.
To study these effects a test beam will be organized at the ion beam line of the University of Washington (CENPA) in 2022 to study the response of the aforementioned sensors to high ionizing events.
Furthermore, laboratory tests will be conducted with an alpha source.
\subsection{TI-LGAD preliminary studies}
\label{sec:TILGAD_app}
Trench Isolated (TI) LGADs are a novel silicon sensor technology that utilizes a deep narrow trench to electrically isolate neighboring pixels to prevent breakdown, as opposed to standard LGADs which use a junction termination extension to prevent breakdown at the pixel edges \cite{9081916}. By utilizing the deep trench isolation technology, the no-gain region is reduced to a few micrometers, thus achieving a higher fill factor than regular LGADs.
The studies shown here were done on strip TI-LGADs sensors from Fondazione Bruno Kessler (FBK) \cite{9081916}. These strip sensors had a pitch of 100\,$\mu$m and had varied characteristics such as contact type, pixel border version, and one or two trenches.
The sensors have been tested using a laser TCT station, the setup is as described in Appendix~\ref{sec:ACLGAD_app}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\textwidth]{Figures/TI_LGAD_1D_profiles.png}
\caption{Left: $P_{max}$ vs position across the strips is plotted. Ch0, Ch1, Ch2 correspond to the different strips being read out. There is minimal amount of “cross-talk” between neighboring strips. Right: $P_{max}$ vs position along the strip is plotted. The response across the different strips remains constant along the strip.}
\label{fig:TILGAD_data1}
\end{figure}
The response of the sensor as a function of position can be observed in Fig.~\ref{fig:TILGAD_data1}. On Fig.~\ref{fig:TILGAD_data1} Left, $P_{max}$, a measure of charge collection, is plotted on the y-axis versus the position going across different strips being read out on the x-axis. It can be seen that the no-gain region between two adjacent strips (red and black in the plot) is less than 10\,$\mu$m.
The sensor has the standard response of a conventional LGAD and exhibits a small amount of “cross-talk”. Figure~\ref{fig:TILGAD_data1} Right, shows the response of the sensor is constant along the strip. The sensor is mounted at an angle with respect to the IR laser so the fall-off of the response is expected to slightly decrease as the laser scans for larger distances along the strip.
The measured signal across a TI-LGAD strip sensor is shown in Fig.~\ref{fig:TILGAD_data2}. The red pulse corresponds to a region of strong signal being for the strip being read out in channel 1 on the oscilloscope. The remaining signals (black and green) correspond to the position of the nearest strip being read out. The maximum values for the nearest strips correspond to ~3$\%$ of the maximum value of the red pulse showing there is good isolation between strips.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\textwidth]{Figures/TI_LGAD_pulses.png}
\caption{Signals across a TI-LGAD strip sensor. The red pulse corresponds to a region of strong signal for the strip being read out. The black and red signals correspond to the position of the nearest strips. It can be observed there is “little” pickup from neighboring strips.}
\label{fig:TILGAD_data2}
\end{figure}
\subsection{TCAD simulation}
\label{sec:TCAD_app}
AC-LGADs have several parameters that can be tuned to optimize the sensor response to the specific application.
The geometry of the electrodes in terms of pitch and pad dimension is the most important one, however also the N+ sheet resistivity and the dielectric thickness between N+ and electrodes influence the charge sharing mechanism.
These parameters have been studied with the TCAD Silvaco \cite{Silvaco} to have a good representation of the observed sensor performance. Simulations with TCAD software are important to compare with existing prototype data and to help in optimizing the design.
Figure~\ref{fig:simulation} (Left) shows the simulated waveforms of four metal pads using TCAD Silvaco The sensor is a 200\,$\mu$m pitch strip sensors with N+ sheet resistivity of 500\,$\Omega$.
Figure~\ref{fig:simulation} (Right) is the Pmax profile of 200\,$\mu$m pitch strip sensors with different sheet resistivity and the same geometry of the BNL strip prototype already introduced.
500\,$\Omega$ shows a reasonably good match with FNAL TB data from Fig.~\ref{fig:FNAL_data}, Right.
These simulations are for 50\,$\mu$m thick sensors, to simulate the final PIONEER prototype they need to be redone for 120\,$\mu$m of thickness.
The calibrated simulation will be used as input to prototype productions to optimize the sensor design for the PIONEER specific application.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{Figures/Simulation.png}
\includegraphics[width=0.55\textwidth]{Figures/PmaxSim.png}
\caption{Left: AC-LGAD strip simulated waveform with TCAD Silvaco, charge deposition is the strip with the red waveform, green is first neighbor, blue is second neighbor. Right: Pulse maximum (normalized) as a function of position for a strip with 80\,$\mu$m width and 200\,$\mu$m pitch. Left edge of the plot is the center of the readout strip. Curves are for data (FNAL testbeam) and TCAD Silvaco simulation with several N+ sheet resistivity. }
\label{fig:simulation}
\end{figure}
\subsection{Electronics and readout chain}
\label{sec:ATAR_electronics_app}
To read out the ATAR sensors, two crucial electronic components need to be identified: an amplifier chip and a digitizer board.
The ASIC needs to be fast enough for the sensor in use; for the signal rise time in the 120\,$\mu$m-thick prototype sensors, a bandwidth of 1\,GHz should be sufficient.
However, the high dynamic range (2000) requirement for the ATAR brings major complications to the readout.
Current fast readout chips usually have a dynamic range of $<$~1000, since they are targeted at MIPs-only detection in tracker sub-systems.
One possibility is to develop an amplifier chip with logarithmic response as well as a high enough bandwidth, currently no such chip exists with the necessary characteristics.
Another alternative is to stream to the digitizer both the amplified (for MIP) and non-amplified (non-MIP) signals, or use two different amplifiers with different gains.
Yet another option is to adopt a chip capable of dynamic gain switching.
Nevertheless already available integrated chips, such as FAST~\cite{OLAVE2021164615} and FAST2, will be evaluated (Fig.~\ref{fig:FAST2}).
Some new ASIC technologies that are being developed at UCSC in collaboration with external companies can run with 2.5V maximum signal, this allows for an increased dynamic range.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/FAST2_cali.png}
\includegraphics[width=0.45\textwidth]{Figures/FAST2_cali_saturated.png}
\caption{Left: FAST2 analog calibration input response. Right: FAST2 analog saturated response, the pulse becomes wider and the peak flattens. In both cases the return to baseline is with a 5~ns time range.}
\label{fig:FAST2}
\end{figure}
To successfully reconstruct the decay chains, the ATAR is expected to be fully digitized at each event.
To achieve this goal, a high bandwidth digitizer with sufficient bandwidth and sampling rate have to be identified.
The same issue afflicting the amplifier, the high dynamic range, is also problematic for the digitization stage.
A digitizer that would suit PIONEER's requirements needs to be identified, a ready commercial solution would be the best option but the cost per channel might be prohibitive.
For this reason the collaboration is exploring the possibility to develop a new kind of digitizer specific to this application.
Since the amplification chip has to be positioned away from the active region, the effect of placing a short (5\,cm) flex cable between the sensor and the amplification stage has to be studied.
The high S/N provided by LGADs would allow the signal to travel without compromising the transmitted information.
A prototype flex was produced (Fig.~\ref{fig:flex}) and the effect on LGAD signals will be studied.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/IMG_20211216_094551.jpg}
\caption{Picture of prototype flex cables with difference trace separations.}
\label{fig:flex}
\end{figure}
\subsection{Mechanics}
\label{sec:ATAR_mech_app}
The ATAR mechanical support needs to introduce as little dead material as possible to avoid degradation of the positron energy.
The wirebonds between the sensor and the flex are theoretically strong enough to hold the sensor, however additional support is necessary.
Figure~\ref{fig:mechanics} (Left) show the tentative U-shaped sensor support with 4 connection at the corners; the open side of the support is for flex connection.
Four staves at the corners would align the planes and hold them together.
Further, the entire ATAR structure is supported by a stave coming from the back plane.
On the HV side the chip are directly connected, on the strip side an interposer layer (Kapton) between chips avoids shorting between strips.
The materials for the support can be carbon fiber or Pyrolytic graphite, providing low density and heat conductivity.
The flex on the chip side can be supported in several ways:
\begin{itemize}
\item Extruding support from the U-shaped sensor support
\item Flex can be bent and glued on the side of the sensor
\item Thin kapton layer extruding from the flex glued on the back of the sensor
\end{itemize}
On the other side the flex is connected to the back support and connected to the interposer board between two flexes. Figure~\ref{fig:mechanics} (Right) shows the mechanical support.
The second flex runs to the back crate with the digitizers.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/ATAR_mech_2.png}
\includegraphics[width=0.45\textwidth]{Figures/ATAR_mech_1.png}
\caption{Left: U-shaped support and staves. Right: final mechanic support assembled.}
\label{fig:mechanics}
\end{figure}
A tentative mounting procedure for the ATAR is as follows:
\begin{itemize}
\item Glue ASIC to flex, wire bond the connections
\item Connect flex to sensor via wire bonds
\item Glue sensor to support
\item Slide support in staves while using a temporary flex support
\item Slide all sensor modules in the staves
\item Remove temporary support, bend wire bonds and flex, then connect to back wheel
\end{itemize}
Another issue to take into account is the thermal load of the sensors.
Before irradiation sensors will likely run at 300~V -- with ~nA current through each of 50 sensor layers, the total power dissipation would be a few tens of $\mu$W for the entire ATAR. This load can likely be extracted with the U-shape supports and staves. If needed, air flow can be enhanced to increase the heat extraction.
After irradiation the current and the voltage will increase due to radiation damage, but this is not a concern for the PieNU phase.
For the entire PieNU+PiBeta the sensors will likely reach 600V, with $\mu$A current multiplied by 50 and thus a total power dissipation of tens of mW (based on \cite{CERN-LHCC-2020-007}), i.e. a factor of 1000 higher than during initial operation. This might be problematic, but a simple solution would be the substitution of the ATAR at the mid-life of the experiment.
\subsection{Alternative designs}
\label{sec:ATAR_designs}
A few alternative designs for the ATAR can be considered:
\begin{itemize}
\item Use of thinner sensors (50\,$\mu$m) with pairs of sensors with same direction of strips, so that they would be read out by a single flex per pair. This would double the granularity in Z (although the X-Y information will not be provided at alternating planes) without doubling the number of flexes.
\item In the central part of the ATAR read out, double-sided strips to have X-Y tracking.
\item Using 8 instead of 4 angles for the sensors, strips are rotated 45 degrees each plane, giving more space to exiting flexes.
\item Instead of ATAR with positron tracker and degrader, construct an improved ATAR that does full tracking. To achieve this, different sensor thickness would be used depending on the position.
\end{itemize}
\subsection{LGAD energy resolution}
\label{sec:LGAD_energy_app}
Preliminary data taken at the Stanford Light source (SSRL) \cite{GALLOWAY20195} show that LGADs can detect low energy X-rays with a reasonable energy resolution (8\% to 15\%) thanks to the internal gain.
The beamline at SSRL had a 2\,ns repetition rate and single pulses were completely separated with 50\,$\mu$m thick LGADs.
Figure~\ref{fig:energy} shows some results of detection of X-rays with LGADs.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/LGAD_energy_distribution.png}
\includegraphics[width=0.45\textwidth]{Figures/LGAD_energy_linearity.png}\\
\includegraphics[width=0.45\textwidth]{Figures/LGAD_energy_resolution.png}
\caption{Top-Left: Event distribution of pulse maximum for different X-ray energy, data is for a HPK 50~$\mu$m LGAD at gain 10. Top-Right: Peak of the pulse maximum distribution vs X-ray energy, data is for a HPK 50~$\mu$m LGAD at different gain levels. Bottom: Fractional energy resolution as a function of X-ray energy, data is for three HPK LGADs of 35~$\mu$m, 50~$\mu$m and 80~$\mu$m of thickness. }
\label{fig:energy}
\end{figure}
\subsection{ATAR timeline}
\label{sec:ATAR_timeline_app}
PIONEER is set to start taking data on a timescale of 7 years at PSI (see Sec.~\ref{sec:timeline_estimate}), the ATAR is a small project in terms of production (roughly 0.01~m$^2$ of sensor area) but still requires significant R\&D for sensors, electronics, digitization and mechanics. An envisioned path forward for this project is the following:
\begin{itemize}
\item Sensor characterization and design optimization is the top priority, so the first interest is to continue the collaboration with Brookhaven National Laboratory in the development of AC-LGAD and DJ-LGAD prototypes. This will culminate in a PIONEER-specific prototype production that should happen within 2 years from now. To steer the design effort, simulations are crucial. At UCSC, TCAD (Silvaco and Sentaurus) simulation experts are setting up a reliable simulation of the available AC-LGAD prototypes. At the same time TI-LGADs produced by FBK will be fully evaluated as an alternative.
\item Having the amplifier chip several cm away from the LGAD sensor is rather unconventional and the effect on the response needs to be understood; a first connection flex prototype was produced and will be thoroughly tested within 2023. Likely a second flex production with the lesson learned will happen in the same year.
\item A crucial aspect to understand is the energy resolution of LGAD devices (which was never studied thoroughly by the community since it is not of interest for common HEP applications) as well as the gain suppression mechanism~\cite{gainsuppr}. UCSC is currently collaborating with the University of Washington to organize a test beam at the the ion beam line of CENPA to study the response of LGADs to high ionizing events. This will tentatively happen in Q2/Q3 2022 with already available sensor prototypes and analog amplifier boards.
\item Building of a first ATAR demonstrator (ATAR0) with available sensor prototypes: the current BNL AC-LGAD production has 2.5~cm long strips with a pitch of 500~$\mu$m, which is close to the final 200~$\mu$m 2x2~cm$^2$ ATAR design (however the sensor thickness is 50~$\mu$m instead of the final design 120~$\mu$m). Since a 50~$\mu$m thick sensor is fabricated on a support wafer of a few 100s $\mu$m, the devices would need to undergo an etching procedure to have full active volume, the thinning procedure can be executed at BNL. The BNL sensor group has experience with wafer lapping and chemical mechanical planarization (CMP). The sensor would be initially lapped to remove most of the thickness of the handling thick substrate, followed by CMP for final polish. The etching rate would be determined on dummy wafers and the process would be optimized to reach the required final thickness.
The prototype would have a few layers (5-10) with a reduced number of channels to detect the temporal development of a muon or pion decay. As the layers would need to be very close to each other, a suitable readout board needs to be developed. The board might be built with discrete components or using a chip such as FAST2.
The prototype would be then tested in a pion/muon beamline either at TRIUMF or a PSI. Hopefully such a prototype can be produced by the end of 2023.
\item Identification of a suitable chip for the analog amplification; the ideal path would be to find an already existing chip (FAST2) and characterize it with LGAD sensors similar to what would eventually be used in the ATAR. In parallel an effort to produce a new chip can be pursued through external companies (SBIR-like funding). A prototype readout chip needs to be ready by 2024.
\item Identification of a digitizer chip: currently available digitizers such as DRS4 are too expensive for the number of channels in the ATAR. Furthermore a complicated triggering system might be needed for AC-LGAD readout. Small companies might be available to develop a new chip with the needed performance by modifying existing designs. A suitable digitizer chip needs to be identified/developed by 2024.
\item The support mechanics and thermal transport calculation needs to be studied well for the success of the ATAR. A discussion is ongoing between technicians from UCSC and UW for the design of the mechanic support and mounting procedure. Thermal load tests can be conducted with an ATAR mock-up made with Silicon heaters, thermal calculations are also foreseen.
\end{itemize}
\clearpage
\section{ATAR technical details}
\label{sec:ATAR_appendix}
As introduced in Sec.~\ref{sec:ATAR}, the highly segmented active target (ATAR) is a key new feature of the proposed PIONEER experiment which will define the fiducial pion stop region, provide high resolution timing information, and furnish selective event triggers. Further technical details will be covered by this Appendix.
\subsection{ATAR R\&D plan}
\label{sec:ATAR_timeline_app}
The ATAR is a small project in terms of production (roughly 0.01~m$^2$ of sensor area) but still requires significant R\&D for sensors, electronics, digitization and mechanics. An envisioned path forward for this project is the following:
\begin{itemize}
\item Sensor characterization and design optimization is the top priority, LGAD prototypes from BNL and FBK are being studied. A PIONEER-specific prototype production is expected to happen within 2 years from now at BNL. Introduction to the LGAD technology is shown in Appendix~\ref{sec:LGAD_app} and studies on available high granularity LGAD prototypes are shown in Appendix~\ref{sec:GLGAD_app}. To steer the design effort, TCAD (Silvaco and Sentaurus) simulations are crucial (see Appendix~\ref{sec:ACLGAD_app}).
An important aspect to understand is the energy resolution of LGAD devices as well as the gain suppression mechanism~\cite{gainsuppr}. A test beam at the the ion beam line at the University of Washington (CENPA) will be used to study the response of LGADs to high ionizing particles. This will tentatively happen in 2022 with available sensor prototypes and analog amplifier boards.
\item Having the amplifier chip several cm away from the LGAD sensor is rather unconventional and the effect on the response needs to be understood; a first connection flex prototype was produced and will be thoroughly tested within 2023. Likely a second flex production with the lesson learned will happen in the same year.
\item The building of a first ATAR demonstrator (ATAR0) is foreseen within the end of 2023 with available sensor prototypes. The project is described in Appendix~\ref{sec:ATAR0}.
\item Identification of a suitable chip for the analog amplification; the ideal path would be to find an already existing chip (e.g. FAST2) and characterize it with LGAD prototypes. In parallel an effort to produce a new chip can be pursued through external companies (SBIR-like funding). A prototype readout chip needs to be ready by 2024.
\item Identification of a digitizer chip: currently available digitizers are too expensive for the number of channels in the ATAR. Small companies might be available to develop a new chip with the needed performance by modifying existing designs. A suitable digitizer chip needs to be identified/developed by 2024.
\item The support mechanics and thermal transport calculation needs to be studied well for the success of the ATAR. The support needs to introduce as little dead material as possible to avoid degradation of the positron energy. A discussion is ongoing between technicians from UCSC and UW for the design of the mechanic support and mounting procedure. Thermal load tests can be conducted with an ATAR mock-up made with Silicon heaters, thermal calculations are also foreseen. The radiation damage during the $\pi\to e \nu$~phase of the experiments is low enough that the heat dissipation of the sensors is not a concern.
\end{itemize}
\subsection{Sensor technology}
\label{sec:LGAD_app}
The chosen technology for the ATAR is based on Low Gain Avalanche Detectors (LGAD) \cite{bib:LGAD}, thin silicon detectors with moderate internal gain. Due to the internal gain and thin bulk, LGADs have fast rise time and short full charge collection time.
The best estimate at present for the sensor thickness is around 120\,$\mu$m to avoid support structures for the sensor, which would introduce dead areas and inactive material within the beam.
Using fast electronics is expected to result in a pulse rise time of about 1\,ns for an LGAD of this thickness. Such a sensor would be able to separate two closely overlapping hits if they arrive more than 1.5\,ns apart. The time resolution on the rising edge should be less than 100\,ps for a minimum ionizing signal, down to much better time resolution for large $\pi/\mu$ signals.
\iffalse
The LGAD technology was chosen over standard Silicon technology because of the intrinsic gain and thin bulk.
HV-CMOS was also considered as a possibility, however monolithic pixels usually have non-depleted regions between pixels, and furthermore to achieve the same level of time resolution, the power would need to be increased to a level that is not sustainable for the ATAR.
3D sensors have a similar issue of non-depleted regions, furthermore the temporal pulse separation discrimination would be depending on the strip pitch.
On top of this, LGADs can be produced by several small and research foundries in addition to commercial vendors, making the prototyping phase easier and reducing the risk of production delays in case of large-scale availability bottlenecks at big companies.
\fi
Current standard LGADs are limited in terms of granularity to the mm scale.
To achieve a ~100\% active area, several technologies still at prototype level are being evaluated for PIONEER, such as AC-LGADs~\cite{Apresyan:2020ipp} (studies shown in Appendix~\ref{sec:ACLGAD_app}), TI-LGADs~\cite{9081916} (studies shown in Appendix~\ref{sec:TILGAD_app}) and DJ-LGAD~\cite{Ayyoub:2021dgk} (a prototype run is expected to be finished and tested by Q1 2022).
Preliminary data taken at the Stanford Light source (SSRL) \cite{GALLOWAY20195} show that LGADs can detect low energy X-rays with a reasonable energy resolution (8\% to 15\%) thanks to the internal gain.
The beamline at SSRL had a 2\,ns repetition rate and single pulses were completely separated with 50\,$\mu$m thick LGADs.
\iffalse
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Figures/LGAD_energy_linearity.png}
\includegraphics[width=0.45\textwidth]{Figures/LGAD_energy_resolution.png}
\caption{Left: Peak of the pulse maximum distribution vs X-ray energy, data is for a HPK 50~$\mu$m LGAD at different gain levels.
Right: Fractional energy resolution as a function of X-ray energy, data is for three HPK LGADs of 35~$\mu$m, 50~$\mu$m and 80~$\mu$m of thickness. }
\label{fig:energy}
\end{figure}
\fi
A dynamic range from MIP (positron) to several MeV (pion/muon) of deposited charge is expected in the ATAR.
Since the event reconstruction relies on temporal pulse separation, the response to successive MiP and high charge deposition have to be studied \cite{Pulserep}.
Furthermore the effect of gain suppression for large charge deposition in LGADs has to be taken into account \cite{gainsuppr}.
To study these effects a test beam will be organized at the ion beam line of the University of Washington (CENPA) in 2022 to study the response of the aforementioned sensors to high ionizing events.
Furthermore, laboratory tests will be conducted with an alpha source.
\subsection{High granularity LGAD ongoing R\&D}
\label{sec:GLGAD_app}
\subsubsection{AC-LGADs}
\label{sec:ACLGAD_app}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{Figures/sensor.png}
\includegraphics[width=0.5\textwidth]{Figures/FNAL_TB.png}
\caption{Left: prototype BNL AC-LGAD strip sensor with 80\,$\mu$m wide strips and pitch of (left to right) 100, 150, 200\,$\mu$m. Right: sensor response ($P_{max}$) as a function of position (perpendicular to the strip) of two strips with 200~$\mu m$ of pitch~\cite{ACLGADpico}. The dashed lines highlight the position of the two strips in the plot. Data taken at the FNAL 120\,GeV proton test beam facility.}
\label{fig:FNAL_data}
\end{figure}
AC-LGADs overcome the granularity limitation of traditional LGADs and have been shown to provide spatial resolution of the order of tens of $\mu$m~\cite{Tornago:2020otn}. AC-LGAD design also allows to have a completely active sensor with no dead regions.
Studies were conducted on strip AC-LGAD prototypes from Brookhaven National Lab (BNL) (Fig.~\ref{fig:FNAL_data}, Left).
The sensors have been tested with a laboratory IR laser TCT station~\cite{Particulars} and at a Fermilab (FNAL) test beam~\cite{ACLGADpico}.
The response of two strips of a 200\,$\mu$m pitch BNL AC-LGAD as a function of position can be seen in Fig.~\ref{fig:FNAL_data} (Right). The position resolution of this prototype sensor varies between 5-15~$\mu m$ in the direction perpendicular to the strip.
In the next few years several prototypes will be tested in laboratories and test beams to identify the right parameter configuration, then a PIONEER-specific production will be made at Brookhaven.
The envisioned metal size for the 200\,$\mu$m pitch strips, as well as other parameters of interest for the sensor such as the doping profile, needs to be confirmed after a testing R\&D campaign and TCAD simulations.
These parameters have been studied with the TCAD Silvaco \cite{Silvaco} to have a good representation of the observed sensor performance. Simulations with TCAD software are important to compare with existing prototype data and to help in optimizing the design, currently the simulated sensor response show a reasonably good match with FNAL TB data.
Studies made with simulations tools will provide crucial input for the future PIONEER sensor production.
\iffalse
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{Figures/Simulation.png}
\includegraphics[width=0.55\textwidth]{Figures/PmaxSim.png}
\caption{Left: AC-LGAD strip simulated waveform with TCAD Silvaco. The sensor is a 200\,$\mu$m pitch strip sensors with N+ sheet resistivity of 500\,$\Omega$. Right: Pulse maximum (normalized) as a function of position for a strip with 80\,$\mu$m width and 200\,$\mu$m pitch. Left edge of the plot is the center of the readout strip. Curves are for data (FNAL testbeam) and TCAD Silvaco simulation with several N+ sheet resistivity. }
\label{fig:simulation}
\end{figure}
\fi
\subsubsection{TI-LGADs}
\label{sec:TILGAD_app}
Trench Isolated (TI) LGADs are a novel silicon sensor technology that utilizes a deep narrow trench to electrically isolate neighboring pixels to prevent breakdown, as opposed to standard LGADs which use a junction termination extension to prevent breakdown at the pixel edges \cite{9081916}. By utilizing the deep trench isolation technology, the no-gain region is reduced to a few micrometers, thus achieving a higher fill factor than regular LGADs.
Prototypes TI-LGADs sensors from Fondazione Bruno Kessler (FBK) \cite{9081916} were studied at UCSC.
TI-LGADs show the standard response of a conventional LGAD and exhibits a small amount of “cross-talk”. Furthermore the response of the sensor is constant along the strip.
The maximum values of the pulse shape for the neighbor strips correspond to ~3$\%$ of the maximum value of the red pulse showing there is good isolation between strips.
New TI-LGADs productions are envisioned at FBK and other vendors, the prototypes will be tested in laboratories and test beams to find a suitable alternative sensor technology for PIONEER.
\subsection{Electronics and readout chain}
\label{sec:ATAR_electronics_app}
To read out the ATAR sensors, two crucial electronic components need to be identified: an amplifier chip and a digitizer board.
The ASIC needs to be fast enough for the sensor in use; for the signal rise time in the 120\,$\mu$m-thick prototype sensors, a bandwidth of 1\,GHz should be sufficient.
However, the high dynamic range (2000) requirement for the ATAR brings major complications to the readout.
Current fast readout chips usually have a dynamic range of $<$~1000, since they are targeted at MIPs-only detection in tracker sub-systems.
One possibility is to develop an amplifier chip with logarithmic response or dynamic gain switching as well as a high enough bandwidth, currently no such chip exists with the necessary characteristics.
Already available integrated chips, such as FAST~\cite{OLAVE2021164615} and FAST2, are being evaluated.
Some new ASIC technologies that are being developed at UCSC in collaboration with external companies can run with 2.5V maximum signal, this allows for an increased dynamic range.
Since the amplification chip has to be positioned away from the active region, the effect of placing a short (5\,cm) flex cable between the sensor and the amplification stage has to be studied.
To study this a prototype flex was produced and the effect on LGAD signals will be studied.
To successfully reconstruct the decay chains, the ATAR is expected to be fully digitized at each event.
To achieve this goal, a high bandwidth digitizer with sufficient bandwidth and sampling rate have to be identified.
The same issue afflicting the amplifier, the high dynamic range, is also problematic for the digitization stage.
A digitizer that would suit PIONEER's requirements needs to be identified, a ready commercial solution would be the best option but the cost per channel might be prohibitive.
For this reason the collaboration is exploring the possibility to develop a new kind of digitizer specific to this application.
\subsection{ATAR0 prototype}
\label{sec:ATAR0}
A first ATAR demonstrator (ATAR0) is foreseen with available sensor prototypes: the current BNL AC-LGAD production has 2.5~cm long strips with a pitch of 500~$\mu$m, which is close to the final 200~$\mu$m 2x2~cm$^2$ ATAR design (however the sensor thickness is 50~$\mu$m instead of the final design 120~$\mu$m). Since a 50~$\mu$m thick sensor is fabricated on a support wafer of a few 100s $\mu$m, the devices would need to undergo an etching procedure to have full active volume, the thinning procedure can be executed at BNL. The BNL sensor group has experience with wafer lapping and chemical mechanical planarization (CMP). The sensor would be initially lapped to remove most of the thickness of the handling thick substrate, followed by CMP for final polish. The etching rate would be determined on dummy wafers and the process would be optimized to reach the required final thickness.
The prototype would have a few layers (5-10) with a reduced number of channels to detect the temporal development of a muon or pion decay. As the layers would need to be very close to each other, a suitable readout board needs to be developed. The board might be built with discrete components or using a chip such as FAST2.
The prototype would be then tested in a pion/muon beamline either at TRIUMF or a PSI. Hopefully such a prototype can be produced by the end of 2023.
\subsection{Beamline Simulations}
\subsubsection{Beamline Simulations and Optimization}
The PSI High Intensity Proton Accelerator (HIPA) chain consists of a Cockcroft-Walton pre-accelerator, delivering 870 keV protons to a set of two isochronous cyclotrons which further accelerate to the maximum 590 MeV.
In the Ring cyclotron, the protons are accelerated by four copper resonator cavities operated in continuous wave (CW) mode at a frequency of 50.6 MHz, giving the periodic 19.75~ns beam microstructure and their final energy of 590 MeV.
The $\pi$E5 channel is a high-intensity low-energy muon and pion beamline, with a maximum momentum of 120~MeV/c, that views the second graphite production target, Target~E, at 165$^{\circ}$ with respect to the proton beam axis.
The $\pi$E5 channel is used exclusively for particle physics experiments, and is currently home to the MEG~II and Mu3e experiments.
The channel length from Target~E to end of the QSB43 quadrupole magnet and the entrance to the experimental area is 13~m, giving approximately 7~m of free space between the concrete wall and current MEG~II infrastructure.
In this area the SEP41 E$\times$B Wien Filter for particle separation, QSK quadrupole triplet for focusing the beam on the ATAR, and the PIONEER installation could be placed.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{Figures/G4BLSIM_PiE5_Labeled.png}
\caption{The G4beamline model of the $\pi$E5 beamline including a stand-in calorimeter. Important elements along the beamline where rates are compared are labeled.}
\label{fig:g4blsim_pie5_labeled}
\end{figure}
Simulations for both pion production at Target~E and transport of particles down the $\pi$E5 channel are performed using the G4Beamline toolkit~\cite{Roberts:2008zzc}.
A figure of the $\pi$E5 beamline from the G4Beamline simulation, beginning at Target E and ending at the PIONEER calorimeter, is shown in \cref{fig:g4blsim_pie5_labeled}.
The pion production simulation utilizes the parameterized cross section developed for the High Intensity Muon Beamline (HiMB) project.
The absolute rates determined from the simulation for particles near the production target are shown in \cref{fig:productionrates}.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{Figures/TgEProdSim_HiMBHYBRID_10B-PoT_MomSpecFromTarget_SF100_positive.pdf}
\caption{The rates of positive charged particles produced at Target E with longitudinal momentum in the direction of the $\pi$E5 channel. }
\label{fig:productionrates}
\end{figure}
The G4Beamline simulation was originally tuned for 28~MeV/c muons and has been rescaled for the respective pion momenta.
Some tuning was performed for the last two dipoles and quadrupole triplets to optimize transmission through the SEP41 and on to the ATAR.
Rates for pions and muons reaching the entrance to experimental area for beamline central momenta ranging from 55~MeV/c to 75~MeV/c are shown in \cref{fig:piE5rates}.
The rates for those particles reaching the entrance to the experimental area (QSB43) and 4.8~m downstream at the entrance to the calorimeter volume (CALO Entrance) at the lower and upper end of the possible momentum ranges are listed in \cref{tbl:rates_55MeVc} and \cref{tbl:rates_75MeVc}.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{Figures/AtQSB43DSPiE5_Rates_PionsMuons.png}
\caption{The rates for pions and muons at the entrance to the experimental area (QSB43).}
\label{fig:piE5rates}
\end{figure}
\begin{center}
\begin{tabular}{p{4cm}||p{3cm}|p{3cm}|p{3cm}}
\multicolumn{4}{c}{Particle Rates ($\times 10^{8}$) in $\pi$E5 (p$_{\pi^{+}}$=55 MeV/c) at 2.2~mA I$_{p}$} \\
\hline
Beamline Position& $\pi^{+}$ & $\mu^{+}$ & $e^{+}$\\
\hline
QSB43 & 0.50 & 1.22 & x\\
CALO Entrance & 0.09 & 0.59 & x\\
\end{tabular}
\label{tbl:rates_55MeVc}
\end{center}
\begin{center}
\begin{tabular}{p{4cm}||p{3cm}|p{3cm}|p{3cm}}
\multicolumn{4}{c}{Particle Rates ($\times 10^{8}$) in $\pi$E5 (p$_{\pi^{+}}$=75 MeV/c) at 2.2~mA I$_{p}$} \\
\hline
Beamline Position& $\pi^{+}$ & $\mu^{+}$ & $e^{+}$\\
\hline
QSB43 & 4.66 & 2.68 & x\\
CALO Entrance & 1.10 & 1.34 & x\\
\end{tabular}
\label{tbl:rates_75MeVc}
\end{center}
The pion rates calculated in simulation at the entrance to the calorimeter are approximately a factor two lower than the latest measurements at $\pi$E5, nevertheless they exceed the requirements of the experiment by two orders of magnitude.
This high rate would allow for losses in the SEP41 or upstream collimation, necessary for reducing background events in the ATAR and calorimeter.
Verification of these rates from measurements and comparison to simulation will be critical in determining achievable pion rates at the ATAR.
\subsection{Beamline Studies}
{\it AS: editing offline}
\subsubsection*{Requirements}
Mechanical engineering considerations for the LXe calorimeter are discussed below.
\vspace{1mm}\begin{easylist} [itemize]
\ListProperties(Start1=1)
& The \mbox{CALO}\ mechanical design has to support a load of about 9 tons of LXe as well as
the required cryogenic envelope, consisting of a cold vessel and an insulation vacuum. Ideally, the heat load
of the new system should be compatible with existing MEG cooling and purification infrastructure.
& The beam region consist of an entrance cone to focus a beam with 10\degree\ divergence onto a beam spot of
about 1 cm$^2$. The entrance side will support the \mbox{BEAM}\ detectors. The exit side should be in air, allow the
insertion of \mbox{ATAR}\ and \mbox{TRACKER}\ and their associated electronics, and provide sufficient heat removal.
& The \mbox{ATAR}\ target region should be surrounded by thin beam pipes. The goal is to
minimize material in the \mbox{CALO}\ acceptance of close to degree=140\degree\ in polar angle, while covering
the full azimuth.
& The \mbox{ATAR}\ and \mbox{TRACKER}\ region is very space limited, while increasing the inner pipe radii is problematic, both in terms
of solid angle and window strengths. A compromise between these conflicting requirements needs to be optimized.
& The PIONEER detector is designed for a measurement program extending over a decade. Both routine services as well as major
repairs and upgrades must be supported by the design.
\end{easylist}\vspace{1mm}
\subsubsection*{Conceptual design}
The calorimeter is built of interactive functional blocks incorporating a top load concept with the intent to be modular and to allow for ease of maintenance. The functional blocks are described below and a drawing is shown in Figure~\ref{fig:LXe_Calo_XSec}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{Figures/LXECal.png}
\caption{Concept design of the liquid xenon calorimeter. For scale, the lid is 3.05 m diameter. The yellow circles are merely representative of the photosensors; they are not placed accurately.}
\label{fig:LXe_Calo_XSec}
\end{figure}
\ListProperties(Start1=1)
\textbf{1. Liquid xenon vessel}:
This vessel has a spherical bottom to follow the contour of the detector mounting surface. This both reduces bending stress in the walls and reduces the volume of xenon required to cover the detectors. Running through this vessel horizontally is a cone-tube pathway. The tube portion, which surrounds the target, is envisioned as beryllium. The cone may be of a variety of metals. The connection between the two components is designed to be a braze or weld.
\textbf{2. Detector mounting surface}:
Beyond the basic requirement of being spherical and holding detectors, the inner radius is set at 0.9\,m to achieve the required number of radiation lengths of xenon.
\textbf{3. Thermal insulation}:
To achieve the best insulation, the calorimeter has been designed to be a vacuum dewar. The support of the liquid xenon vessel is achieved by hanging it on turnbuckles made of material which does not conduct heat well. Stainless steel is seen as the most likely option. If additional reduction in heat load is needed the link could be changed to titanium, allowing for a reduced cross section. The more complicated portion of the thermal insulation is maintaining vacuum around the incoming beam and the target space. This is made easier by having a break in this tube. The beam tube loads from one side, the target tube from the other. Both terminate in thin vacuum barrier windows which allow beam passage. To accommodate both temperature induced dimensional changes and deflection from weight loading, there are two bellows present. The tube running through the xenon vessel has bellows on the target loading side. There is a large bellows on the top of the xenon vessel to allow cabling and piping to pass through the vacuum space.
\textbf{4. Target air space}:
The tube the target resides in is envisioned as beryllium. This tube is mounted on a flange on the vacuum vessel and protrudes past the centerline by approximately 175\,mm. The inside end of the tube is capped with a window that allows passage of the beam and also acts as a vacuum barrier. This tube also contains the tracking detector and the necessary cabling.
\textbf{5. Beam path to the target}:
The beam is shown focusing with a center line to the beam edge angle of 10 degrees.
The inner cone is between the beam (at atmospheric pressure) and the vacuum (insulation). The outer cone is the barrier between the vacuum and the xenon. Further studies might lead to a modification of the downstream exit pipe to mirror the entrance side, and the optimization of the ATAR holder,
to allow pass through of \mbox{$\pi$DIF}\ muons.
\subsubsection{ Toward a realistic design for PIONEER}
In subsequent sections we describe the beam,
target, positron detectors, calorimeter, DAQ, and electronics aspects of the experiment. The Appendices include additional technical information and priority R\&D plans.
\subsubsection{Requirements for measuring $R_{e/\mu}$ }
At rest, the pion lifetime is 26\,ns and the muon lifetime is 2197\,ns. The monoenergetic positron from $\pi \rightarrow e\nu$ has an energy of 69.3\,MeV. Positrons from ordinary muon decay form the Michel spectrum from 0 to an endpoint of 52.3\,MeV. In principle, the monoenergetic $e^+$ from $\pi\to e \nu$~ is well isolated above the Michel endpoint and can be easily identified using a high-resolution, hermetic calorimeter. To determine $R_{e/\mu}$ we measure the ratio of positrons emitted from $\pi\to e \nu$~ and $\pi\to\mu\to e$~ decays for which many systematic effects such as solid angle acceptance cancel to first order. However, counting all $\pi \rightarrow e$ events with a precision of one part in $10^4$ requires determining the low-energy tail of the electromagnetic shower and radiative decays that hide under the Michel spectrum from the $\pi \rightarrow \mu \rightarrow e$ chain, which has four orders of magnitude higher rate.
Figure~\ref{fig:Signal-Tail} illustrates the relationship between the two channels and their respective positron energy spectra. Here, we have modeled the spectrum from both channels assuming a high resolution, $25\,X_0$ calorimeter. There remains an unavoidable tail fraction below 53\,MeV that must be determined accurately in order to obtain the branching ratio.
That challenge was critical to previous generations of experiments and was responsible for the leading systematic uncertainty in the
PIENU experiment at TRIUMF.
PIONEER~ will minimize the intrinsic tail fraction
through the use of a $25\,X_0$ LXe calorimeter, the design of which is based on the considerable experience of the MEG Collaboration.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{Figures/Simulation/high_stats_tail_example_final.pdf}
\caption{The positron energy spectra from muon decays (blue) and from $\pi\to e \nu$~ decays (orange) for a calorimeter
resolution of 1.5\% and a
depth of $25\,X_0$.
The simulation includes energy losses owing to photonuclear interactions.
}
\label{fig:Signal-Tail}
\end{figure}
The $\pi\to e \nu$~ branching ratio $R_{e/\mu}$ will be obtained by first separating events into high- and low-energy regions at an energy cut value ($E_{cut}$) as discussed in the Appendices for the PIENU experiment. The time spectra will be fit in each region with the $\pi^+ \rightarrow e^+ \nu$ and $\pi^+ \rightarrow \mu^+ \rightarrow e^+$ timing distribution shapes, along with backgrounds originating from different sources including event pile-up effects, pion decays in flight, and effects from old muon decays.
PIONEER~ will incorporate improvements to the previous techniques including a deeper calorimeter.
However, it is also important to be able to create triggers that can isolate $\pi \rightarrow e$ from $\pi \rightarrow \mu \rightarrow e$ chains within the stopping target, identify pion and muon decays in flight, as well as identify pileup from long-lived muons remaining in the target from earlier pion stops.
Figure~\ref{fig:EventTypes} illustrates several key processes that will occur at different rates in the stopping target. The signal event (1) has a pion decay at rest; its Bragg peak energy deposition as well as its depth within the target identifies it as a pion stop. The much higher background (and normalization) channel (2) is illustrated by that same pion stop followed by a decay to a 4.1\,MeV muon, which travels through a number of planes and then stops, leaving behind an image of its own short trajectory and Bragg peak. If the muon decays in the prompt window being observed for the pion decays (typically 3 -- 50\,ns after the pion stop), its Michel positron will be recorded in the CALO. More rare but subtle processes
may also occur.
A pion can also decay in flight within the target to a muon before it stops (3).
In very rare cases, the emitted 4.1 MeV muon can decay in flight before stopping (4) providing a Lorentz boost to the emitted Michel positron, which contaminates the signal spectrum above the ordinary stopped muon endpoint.
To distinguish event types, we will use an active target that can provide 4D tracking (at the level of 150\,$\mu$m in space and $<$1\,ns in time) and energy measurements from the O(30)~keV signals for positrons to the 4000\,keV Bragg peaks of stopping pions and muons. As discussed in Sec.~\ref{sec:ATAR}, our collaboration is focusing on the new low gain avalanche detector (LGAD) sensors as a centerpiece of the experiment. Simulations using optimized LGAD parameters provide confidence that triggers can be constructed to isolate and measure all event types.
The calorimeter tail fraction for $\pi\to e \nu$~ events will be measured {\it in situ} by suppressing the $\pi\to\mu\to e$~ decays using information provided by the active target. $\pi\to\mu\to e$~ events can be identified and suppressed by the presence of the 4.1\,MeV pulse from $\pi \to \mu \nu$ decay, use of a narrow time window, $\pi-\mu$ particle identification, and tracking information to identify pion decay-in-flight (\mbox{$\pi$DIF}), and muon decay-in-flight (\mbox{$\mu$DIF}) following pion decay-at-rest (\mbox{$\pi$DAR}). As discussed in Sec.~\Ref{Sims}, we anticipate that $\mu-e$ backgrounds in the tail region can be suppressed to a level that will allow the uncertainty in the tail fraction to contribute $<0.01\%$ to the error in $R_{e/\mu}$.
The experiment will require a continuous wave low-momentum pion beam that can be focused to a small spot size and stop within the ATAR dimensions. Ideal characteristics include a relatively low momentum of $55\,$\,MeV/$c$ ($\pm2\%$) and a flux of 300\,kHz. At this low momentum, a separator is very effective to reduce background from beamline muons and positrons. The $\pi$E5 beam can provide the needed flux. We are also investigating use of the $\pi$E1 line.
Because of the high data rate, state-of-the-art triggering, fast digitizing electronics, and high bandwidth data acquisition systems are required.
\begin{figure}[h!]
\centering
\includegraphics[width = \textwidth]{Figures/EventTypes_vector.pdf}
\caption{Illustration of event types in the $\pi \rightarrow e$ and $\pi \rightarrow \mu \rightarrow e$ chains. The stars are indicative of the energy deposited at the Bragg peak for pions (red) or muons (purple) that stop in the segmented ATAR. The main channels of interest include 1) the $\pi \rightarrow e$ ``signal channel'' decay that emits a monoenergetic 69.3\,MeV positron; 2) the dominant $\pi \rightarrow \mu$ decay, where the 4.1\,MeV muon travels up to 0.8\,mm and also stops in the ATAR before emitting a positron. Events 3) and 4) represent situations that can confuse the classification of events into categories 1) or 2). In 3) the pion decays within the ATAR prior to stopping; the muon stop can then appear as a pion stop. In 4), the pion stops, and the decay muon (very rarely, but importantly) decays in the short time prior to stopping. Because it is a decay in flight, the Lorentz boost can push the positron energy beyond the 52.3\,MeV endpoint. The ATAR is being designed to distinguish these event patterns.
}
\label{fig:EventTypes}
\end{figure}
\subsubsection{Requirements for measuring pion beta decay }
In Phase II (III), pion beta decay $\pi^+\to\pi^0 e^+ \nu$ will be measured by observing the characteristic (nearly) back-to-back gammas from $\pi^0$ decay normalized to $\pi\to e \nu$~ decay as in \cite{Pocanic5,Frlez:2003vg,Pocanic:2003pf,Frlez:2003pe,Bychkov:2008ws}. In PIONEER\ we also expect to observe the low-energy positron absorbed in the ATAR in coincidence with the gammas in the calorimeter.
The Phase II (III) pion beta decay experiment will require $7\times 10^5$ ($7\times 10^6$) events at an intrinsic branching ratio of $ 10^{-8}$. This will require running at a significantly higher pion flux of $\geq 10$\,MHz. The beam momentum and emittance may be higher than for the $\pi\to e \nu$~ measurement to achieve the higher flux.
The higher rate can be handled because of the gamma ray coincidence identification and nearly fixed energy sum. The $\pi$E5 beamline appears to have the necessary properties for this measurement.
\subsubsection{Simulations Guiding the Design}
The discussion of the beamline and detector that follows is based on simulations reported in Section~\ref{Sims}. Briefly, there are three distinct efforts: beamline and upstream detectors, simulation of the ATAR and event topologies, and simulation of the calorimeter response.
A G4beamline~\cite{Roberts:2008zzc} model of $\pi$E5 is being used to extrapolate from the existing surface muon tunes used in the MEG program to the $55-70$\,MeV/$c$ range of interest for positive pions required for PIONEER.
Descriptions of the ATAR and CALO are included in the GEANT-4~\cite{GEANT4:2002zbu} models.
The models include inactive materials such as walls, cables, and windows and a
Tracker detector that surrounds the ATAR.
\subsubsection{Trigger design}
\label{Trigger}
Figure~\ref{fig:readout} shows the planned topology for readout and triggering of PIONEER. The design takes advantage of
the APOLLO~\cite{ApolloTWEPP21} platform designed for the trigger track finder and pixel readout in the CMS experiment at
the LHC. The Command Module (CM) component of the APOLLO board provides a highly configurable environment supporting a dense
high speed IO environment via
25 Gbps Samtec Firefly~\cite{firefly} optical inputs. The platform already
supports the flexible high speed synchronous command and triggering TCDS2 (Trigger and Timing Control and Distribution
System)~\cite{CMSP2TDR} for the CMS phase II upgrade, as well as options for direct IO for receiving timing signals or
generating trigger signals for devices expecting analog triggers. We anticipate using a single FPGA version of the board (utilizing the Xilinx VU13P) with an interposer in the second FPGA slot to bring all the serial links to the VU13P.
While by some measures this board likely provides more power than needed for PIONEER, it has already undergone significant prototyping with full production anticipated in 2024, and can satisfy several of the needs of the system with a single platform. The engineering group that designed the CM will also contribute to calorimetry readout, providing additional coherence.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.9\textwidth]{Figures/TriggerBackend.pdf}
\caption{Proposed topology of the trigger and subdetector readout systems.}
\label{fig:readout}
\end{center}
\end{figure}
PIONEER anticipates use of the 16.1 Gbps Firefly implementation. This protocol serializes 64 bit words with a 64b/66b encoding, resulting in an effective transmission of 64 bit words at rates up to 244\,MHz. Each channel supports both transmit and receive functions, which will allow for significant flexibility in the command and control system.
The Samtec Firefly system already supports PCI Express (PCIe) Gen 3 and Gen 4 integration, with commercial PCIe over firefly adapter cards already available for Gen 3. As discussed further in Sec.~\ref{subsec:daq}, PIONEER proposes use of PCIe as the protocol underlying the DAQ system.
The main trigger APOLLO processor will receive trigger inputs, either via inputs connected to the FPGA standard IO pins, or from the firefly inputs and apply the appropriate logic algorithms to generate triggers and guide sparsification. The processors encode the resulting trigger and configuration information into the TCDS (or equivalent) datastream, which gets transmitted to all subdetectors. The processor will also receive status and health information from the subdectector, communicating any alarm conditions to the DAQ system for appropriate action.
The trigger processor will receive trigger inputs from the BEAM counters and from the tracker to form the PI, Track and PROMPT triggers, with configurable prescale factors applied as noted earlier. The processor will also receive energy sums from subsections of the calorimeter (see Sec.~\ref{subsec:caloBE}) and complete the total energy sum used to form the CaloH trigger in coincidence with the (unprescaled) PI trigger. The processor can flexibly accommodate the needs of the beta decay (Phase II) of the experiment, completing, for example, opening angle or mass estimates based on energy-weighted position sums from the calorimeter electronics.
The processor will also interact with the ATAR, level 3 trigger system and DAQ system to coordinate additional trigger information from the patterns in the ATAR as described above.
\subsubsection{Calorimeter backend electronics}
\label{subsec:caloBE}
Figure~\ref{fig:calodaq} illustrates the digitization and readout electronics proposed for the calorimeter. The 12 channel digitization boards will utilize the Analog Devices AD 9234 dual channel, 12 bit 1 GSPS ADC chip, chosen for its low latency of 59\,ns from presentation of the signal at the front end to output of the digitized signal. These chips communicate the digitized output via the JESD204 protocol, which will present the data for each ADC channel as 64 bit words containing four samples at 250\,MHz. At these rates, the calorimeter information can be summed with a pipelined adder and potentially corrected with a running pedestal measurement as the first stage in a total energy measurement. By clocking the ADC at the slightly lower rate of 976\,MHz, the ADC information can be synchronized to the firefly data transfer rate of 244\,MHz, simplifying synchronization of the system.
The ADCs will sample continuously with samples stored initially in a ring buffer on the FPGA. Upon receipt of a trigger, a configurable time window will be stored in DIMM memory. By deploying a DIMM with a 128 bit data path, simultaneous reading and writing can be accomplished via two 64 bit pathways.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.9\textwidth]{Figures/CaloDAQ.pdf}
\caption{Proposed calorimeter digitization and readout. The 12 channel digitizer boards utlize the dual channel AD9234 12 bit, 1 GSPS ADCs, Xilinx Ultrascale FPGAs for contro, and Samtec Firefly\textsuperscript{TM} high speed communications to a CMS APOLLO board. The APOLLO board can receive up to 22 boards in this configuration -- instrumenting quadrant of an order 1000 channel calorimeter.}
\label{fig:calodaq}
\end{center}
\end{figure}
A single FPGA will control one pair of ADCs (four calorimeter channels). It can compare the energy sum from each channel against a channel activity threshold (or thresholds), as well as combine the four running energy sums as the first stage of a total energy sum for the high energy trigger. That FPGA will also drive a single firefly channel. The 64 bit word for the firefly communication provides much flexibility. After reserving 48 bits for readout of the four channels upon receipt of a trigger, 16 bits can provide triggering information at the full 244\,MHz rate. Possible configurations include a 12 bit sum of the four calorimeter channels along with a bit for each channel flagging a sum over threshold. With a coarse 8 bit sum, 2 bits could be devoted to each individual channel. For the beta decay phase, the trigger could pass energy-weighted position information for a mass calculation. Xilinx kintex ultrascale FPGAs provide sufficient numbers of high speed serial lines at a reasonable cost.
The digitizer boards will communicate with intermediate Apollo boards via the 16.1\,GHz firefly links, which come packaged with a minimum of four individual links. Three of these links will provide the TCDS (or equivalent) clock and control signals and send the trigger and channel readout information to each of three sets of 4 ADC channels. The fourth firefly line will provide PCIe or ethernet communication to allow board configuration and other slow control functions.
After taking into account communication links to the trigger processor and between the intermediate Apollo boards, each intermediate Apollo can accommodate 22 12-channel digitizer boards for a total of 264 channels. Four such boards would instrument a calorimeter with up to 1056 channels. With a more complex, run-state dependent scheme for communications, 16 channels per board could be accommodated.
The intermediate APOLLO boards will collate the waveforms read from all channels for each trigger, and send that information directly to the DAQ system. Each APOLLO can sustain a full readout rate of about 18.5 kHz for the 264 channels, allowing for significant headroom given the anticipated trigger rates (see Tab.~\ref{tab:triggers}) and use of sparsification for an order 1000 channel calorimeter.
The total calorimeter energy sum proceeds in several stages. Each quad-channel FPGA adds those four signals as noted above. Each intermediate Apollo will combine the those sums from the $22\times 3=66$ quad channels. The trigger process will provide the final combination of the four intermediate sums needed for the complete calorimeter sum.
The latency in the energy sum has three main contributions: the ADC latency, the FPGA summation time, and the serialization/deserialization overhead for the high speed links. The Xilinx GTY transceivers have a high degree of configurability, resulting in a total serialize/deserialize latency for the readout chain above that ranges from 65\,ns to 175\,ns at the 16.1\,GHz serial rate. Recent studies~\cite{OliveiraThesis} show that even the simpler, lower latency settings, can achieve reliable links. With 100~ns as a reasonable estimate, 50\,ns as a limit for the energy summation at 244\,MHz, and the 50\,ns ADC latency, we estimate a total latency of 200\,ns.
\subsection{ATAR R\&D}
A brief summary of ATAR R\&D follows. A more detailed plan is presented in Appendix~\ref{sec:ATAR_timeline_app}.
\begin{itemize}
\item After initial sensor characterization and design optimization a PIONEER specific prototype production should happen by the end of 2023. The characterization includes studies on LGAD energy resolution and gain suppression mechanism.
\item Building of a first ATAR demonstrator (ATAR0) with a few planes of available sensor prototypes (BNL strip LGADs, 2.5\,cm with 500\,$\mu$m pitch). An electronics board with suitable characteristics needs to be designed and produced. The prototype would be then tested in a pion/muon beamline either at TRIUMF or PSI. This prototype may be produced by the end of 2023.
\item Identification of a suitable chips for the analog amplification and digitization by 2024. The effect of a short flex between sensor and chip will be studied within 2022.
\item The support mechanics and thermal load needs to be studied well with mock-up prototypes and silicon heaters. These details needs to be fully understood by 2025.
\item Full production of sensors and readout ASIC, once identified, should take less than a year given the modest area of the ATAR. Therefore final production and subsequent assembly can start in 2025.
\end{itemize}
\subsubsection{Massive neutrino searches $\pi^+ \to \ell ^+ \nu _H$}
Searches for peaks in the positron energy spectum due to $\pi^+ \to \ell ^+ \nu _H$ decays were performed in the PIENU experiment \cite{PIENU:2017wbj,PIENU:2019usb} sensitive to masses $65<m_H<135$ MeV but no significant signal above statistical uncertainty was found. The measurement of $R_{e/\mu}$\cite{PiENu:2015seu} provides limits for $m_H<65$ MeV.
To estimate the expected sensitivities for PIONEER with $100\times$ the statistics (Phase I), reduced backgrounds, and improved detectors, toy MC simulations were performed.
For $\pi^+ \to e ^+ \nu _H$ decays, the peak search sensitivity was limited by residual $\pi^+ \to \mu^+ \to e^+$ background from pion and muon decay-in-flight ($\pi$DIF and $\mu$DIF).
The low energy calorimeter response tail and statistics of the $\pi^+ \to e^+ \nu_e$ decay also limits the sensitivity.
Using an active target and a larger electromagnetic calorimeter, the background $\pi^+ \to \mu^+ \to e^+$ will be significantly suppressed compared to PIENU, and a significantly smaller low energy $\pi^+ \to e^+ \nu_e$ tail is anticipated.
Figure \ref{pienuH} shows the result of a toy MC study for the expected sensitivity (90\% C.L. upper limits) in PIONEER, assuming $1 \times 10^8$ $\pi^+ \to e^+ \nu_e$ events, 1\% tail fraction below 52\,MeV, no $\pi^+ \to \mu^+ \to e^+$ events, and negligible acceptance corrections due to the larger detector acceptance.
Compared with PIENU (red in Fig.~\ref{pienuH}), the expected sensitivity in PIONEER (black) would be improved by one order of magnitude.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.33]{Figures/Beam/Exotics/pienuH.png}
\caption{Left: simulated $\pi^+ \to e^+ \nu_e$ energy spectrum. Right:Limits on $|U_{e i}|^2$ (90\% C.L.) from PIENU (red)\cite{PiENu:2015seu,PIENU:2017wbj}) and expected from PIONEER (black).
The lower region limits ($m_H$<65 MeV) come from the branching ratio measurement and the higher ($m_H$>65 MeV) region from the peak search.}
\label{pienuH}
\end{figure}
The sensitivity for $\pi^+ \to \mu^+ \nu_H$ decay will also be improved by the larger PIONEER statistics .
The dominant background is mainly due to the radiative pion decay $\pi^+ \to \mu^+ \nu_{\mu} \gamma$ with branching fraction $2\times10^{-4}$ \cite{PIMUNUG} (for $E_\gamma< 1$\,MeV).
A toy MC simulation was performed with $1 \times 10^9~{\pi}^+\to\mu^+\nu_{\mu}$ decays ($\times 100$ larger statistics than PIENU) including $\pi^+ \to \mu^+ \nu_{\mu} \gamma$ background considering the proper branching fraction, and assuming the same detector resolution as in PIENU.
Figure \ref{Result} shows the results of the simulation and the PIENU experiment \cite{PIENU:2019usb}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.34]{Figures/Beam/Exotics/pimuH.png}
\caption{Left: simulated $\pi^+ \to \mu^+ \nu_\mu$ kinetic energy spectrum. Right: The PIENU result (red triangles \cite{PIENU:2019usb}) and expected sensitivity with PIONEER (black circles) for the mixing matrix element $|U_{\mu i}|^2$ (90\% C.L. limit).}
\label{Result}
\end{figure}
\subsubsection{Two body muon decay $\mu^+ \to e^+ X_H$}
Massive or massless weakly interacting neutral bosons $X$ such as axions \cite{Axion1, Axion2, Axion3, Axion4} and Majorons \cite{Majoron, Majoron2, Majoron3} have been suggested to extend the SM including models with dark matter candidates, baryogenesis, and solutions to the strong $CP$ problem.
Wilczek suggested a model \cite{Wilczek} which may lead to charged lepton flavor violation (CLFV) where the boson X can be emitted in flavor changing interactions.
When decay products from a massive boson $X_H$ are not detected due to a long lifetime, flavor violating two-body muon decays involving a massive boson $\mu^+ \to e^+ X_H$ can be sought by searching for extra peaks in the Michel spectrum.
This search was performed with PIENU, resulting in the limit to the branching ratio $\Gamma (\mu^+ \to e^+ X_H)/\Gamma( \mu ^+ \to e^+ \nu \bar{\nu})$ at the level of $10^{-5}$ in mass range $47.8-95.1$ MeV/$c^2$ \cite{PIENU:2020loi}.
The statistics will improve by two orders of magnitude with respect to PIENU.
To estimate the expected sensitivity, $2 \times 10^{10}$ muon decays were simulated.
Figure \ref{FinalResult} shows the 90\% C.L. upper limits of the branching ratio from different experiments and the expected sensitivity for PIONEER.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.5]{Figures/Beam/Exotics/FinalResult.pdf}
\caption{90\% C.L. limit on the $\mu^+ \to e^+ X_H$ branching ratio anticipated for PIONEER (black circles) and other past experiments (see Refs. \cite{Exp1, Exp2, Exp3, Exp4} for more details).}
\label{FinalResult}
\end{figure}
\subsubsection{Other decays}
Three and four body pion decay modes can be analyzed with the same method as for the $\pi^+ \to \ell ^+ \nu_H$ searches. However, the signal shapes are in these cases are represented by continuous lepton energy spectra.
The expected sensitivities will also be improved by one order of magnitude. The current limits set by the PIENU experiment are described below.
Three body pion decays $\pi^+ \to \ell ^+ \nu X$, where $X$ is a massive or massless weakly interacting particle, were searched for in the PIENU experiment \cite{PIENU:2021clt}.
The decay $\pi^+ \to e^+ \nu X$ was sought; no signal beyond the statistical uncertainty was observed, and 90\% C.L. upper limits were set on the branching ratio $\Gamma (\pi^+ \to e^+ \nu X)/\Gamma (\pi^+ \to \mu^+ \nu)$ with $10^{-7}-10^{-8}$ level in the mass range of $0<m_X<120$ MeV/$c^2$.
The $\pi^+ \to \mu^+ \nu X$ decay was also searched for.
A 90\% C.L. upper limit was derived on the branching ratio $\Gamma (\pi^+ \to \mu^+ \nu X)/\Gamma (\pi^+ \to \mu^+ \nu)$ at the $10^{-5}-10^{-6}$ level in the mass region from 0 to 33.9 MeV/$c^2$.
The rare pion decays $\pi ^+ \to \ell^+ \nu_{\ell} \nu \bar{\nu}$ are highly suppressed. Thus, the experimental search for these processes could reveal small non-SM effects such as neutrino-neutrino interactions \cite{nunu} and six-fermion interactions \cite{6f,6f2}, which might compete with the SM processes at first order.
The rare pion decays, considering three models (SM, neutrino-neutrino interaction, and six-fermion) were also searched for in PIENU \cite{PIENU:2020las}, and a first result for $\Gamma(\pi^+ \to \mu^+ \nu_{\mu} \nu \bar{\nu})/\Gamma(\pi^+ \to \mu^+ \nu_{\mu})<8.6 \times 10^{-6}$ and an improved measurement $\Gamma(\pi^+ \to e^+ \nu_{e} \nu \bar{\nu})/\Gamma(\pi^+ \to \mu^+ \nu_{\mu})<1.6 \times 10^{-7}$ were obtained.
\section{Simulations} \label{Sims}
Each of the design elements discussed above is being actively studied
using GEANT4-based \cite{GEANT4:2002zbu} simulations.
The simulation efforts include
beamline and upstream detector simulations, simulation of the active target, and simulation of the calorimeter.
The $\pi$E5 beamline at PSI is simulated using G4Beamline \cite{Roberts:2008zzc} discussed in Section \ref{sec:beam_studies}. The remainder of the simulation is done primarily using GEARS,
an extension of GEANT4, which streamlines readout for rapidly iterating systems. The geometries for each of the experimental components are generated using a stand-alone Python script and geometry library, which takes as its input a \texttt{json} file of various parameters (e.g. diameters, number of elements, which detectors to implement, etc.) and exports a \texttt{GDML} file which is then read in by the Geant4 simulation. This file contains a full description of the physical geometry of the detectors, as well as their material properties (density, reflectivity, scintillation yield, etc.).
Because of this workflow, it becomes trivial to implement scans over various parameters to perform systematic studies.
In the following sections, we discuss the initial ATAR and calorimeter simulations.
\subsection{ATAR Simulations}
Simulations of the active target are done with both GEANT4, which allows us to model pions and their decay products through the full chain of detectors, and
TRIM\cite{ziegler1985stopping},
which allows us to track particle energy deposition precisely in a simulated detector readout.
Figure \ref{fig:atar_sim} (left) shows the ATAR concept and (right) a simulation of energy deposit by exiting positrons as a function of polar angle of emission.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{Figures/Simulation/atar_image.png}
\includegraphics[width=0.48\textwidth]{Figures/Simulation/atar_deposited_energy.pdf}
\caption{Left: An image of the simulated ATAR, showing the active planes (white) and readout strips (orange). Right: Total energy deposited in the ATAR by the decay positron as a function of the angle $\theta$ at which it enters the calorimeter. }
\label{fig:atar_sim}
\end{figure}
The active target is implemented in the simulation as alternating layers of cross-hatched silicon strips. The number of strips per layer, their thicknesses, and the number of layers is fully configurable. The nominal configuration consists of 48 \unit[120]{$\mu$m} thick planes of 100 strips each for a total size of $\unit[20 \times 20 \times 5.76]{mm^3}$. Optional `dead' zones can be implemented in each pixel, to simulate the effect of trenches between pixels within a plane or a dead backing layer. Each plane of the ATAR has a fully simulated readout cable, consisting of Kapton and Silicon, simulating the ultimate flex cable readout. These readouts introduce an asymmetry in the energy reconstructed by the calorimeter, and as such have the potential to introduce a systematic bias to our analysis. The design of the readouts will be informed by this simulation in order to minimize such an effect.
\subsubsection{ Background Suppression for the Measurement of the $\pi\to e \nu$\ tail fraction}
\input{Simulations and Analysis/SuppressedBackground}
There is an ongoing effort to apply machine learning tools to boost the sensitivity to $\pi^{+}\rightarrow e^{+}\nu$ events in the tail region and suppress the backgrounds. It has already been demonstrated with the in-ATAR $\pi\,\mathrm{DIF}$ events that gradient boosting decision trees (BDT) are able to outperform the manual cut-based methods.
Although the BDT model shows excellent classification performance, only manually-constructed physics-inspired high-level information has been used as input features for the model training. As deep neural networks, especially Convolutional Neural Networks (CNN), have shown extraordinary performance in various image processing and classification applications, it could be advantageous to make use of deep neural networks’ ability to automatically learn high-level features directly from the complete pixel-level energy deposit information.
\subsection{Calorimeter Simulations}
\label{sec:calo_simu}
\subsubsection{Pileup}
Simulation studies of the LXe calorimeter response have been performed using the GEANT4 package \cite{GEANT4:2002zbu} with optical photon tracking. A simple geometry (see Fig.~\ref{fig:geometry_MC}) that is representative of the current design was implemented. $\pi\rightarrow \mu\rightarrow e$ events were generated in the target at the center of the LXe sphere and optical photons originating from LXe scintillation induced by the showers generated by the positrons were tracked until the outside sensitive surface of the LXe sphere. The simulated
pulses obtained are similar to the ones reported by MEG.
A pulse fitting algorithm was employed to evaluate the possibility of separating events that overlap in time i.e.~pulse pileup. Pulse separation down to \unit[5]{ns} was achieved (see illustration in Fig.~\ref{fig:PSF}) across a wide range of amplitudes. This gives a first indication (without digitization) of the performance of the detector and subsequent data analysis with respect to dealing with pileup.
Further studies are anticipated to introduce and optimize optical qualities of the surfaces, optimize photo-sensing detectors' coverage,
and improve the simulation of the originating scintillation photons. Much of the input for these simulations will be provided by test measurements.
\begin{figure}[!tbp]
\centering
\subfloat[]{\includegraphics[width=0.22\textwidth]{
Figures/geometry_MC.pdf}\label{fig:geometry_MC}}
\hfill
\subfloat[]{\includegraphics[width=0.7\textwidth]{Figures/Simulation/lxe_pulse_fit_updated.pdf}\label{fig:PSF}}
\caption{(a) Rendering of the simplified geometry used in the first Monte-Carlo simulation of the PIONEER LXe calorimeter (b) Example of simulated pulse shapes of a single $\pi\rightarrow \mu \rightarrow e$ event (top) and 3 events happening closely in time (bottom) and recorded in the LXe calorimeter. Pulse shape fitting based on a template allows accurate identification and energy reconstruction of multiple pulses down to $\pm$\unit[5]{ns} separation.}
\end{figure}
Position resolution will also augment pile-up handling capabilities. The position resolution capabilities of the detector and its importance for achieving the targeted rate will be modeled and studies are envisaged in the apparatus which will also be used to test the photo-sensors in LXe.
Depending on the outcome of the pileup studies, segmentation of the LXe volume may be considered.
\subsubsection{Additional Calorimeter Studies}
At the current design stage and given the targeted level of precision, further simulations assessing the performance of the calorimeter will be carried out in conjunction with development of design parameters for the beam, ATAR, and tracking detectors.
One important aspect of the calorimeter design that is being informed by the simulations is understanding how energy from a decay positron is `lost' before it reaches the calorimeter. While energy lost in the ATAR is measured, losses occurring in inactive material e.g. the ATAR cabling, and calorimeter entrance windows, depend on the angle of emission. Simulations are being used to study these effects on the positron energy resolution.
Another important study involves modeling of photonuclear interactions. As decay positrons interact with atomic nuclei in the simulation, they will occasionally cause a nucleus to enter into an excited state with single or multiple neutron emission. When these neutrons escape the calorimeter without depositing their energy the observed energy is shifted down by multiples of the neutron binding energy (see Fig.~\ref{fig:Signal-Tail}). Photonuclear processes were studied in the PIENU experiment to determine the impact on the NaI(Tl) calorimeter lineshape. The results were used as input to the calorimeter simulations. Modeling and prototype studies will be pursued for evaluating these effects in LXe.
In summary, the robustness of the calorimeter performance for the determination of the low energy tail, the resolution, and the pileup rejection will be evaluated by studying the following aspects:
\begin{itemize}
\item impact on resolution and reconstruction of energy loss in inactive materials,
\item optical properties of the inner surfaces of the LXe detector (reflectivity, light absorption, etc),
\item granularity, layout (inner and outer surfaces) and specification of photo-sensors (dark current, PDE, etc.), and
\item characterizations of scintillation light in LXe; e.g., variations in the absorption length.
\end{itemize}
Understanding how these conditions affect the performance of the calorimeter will determine the critical parameters required for the next phase of the experimental design.
\subsection{Heavy Neutrinos and other Dark Sector Physics}
\label{sec:BSM}
The PIONEER experiment will also achieve improved sensitivity in probing for effects of heavy neutrinos. If a heavy neutrino has a sufficiently low
mass that it can be emitted, the signature will be a monochromatic peak
in the energy of the outgoing charged lepton at an anomalously low value.
Even if the heavy neutrino has a mass greater than the kinematic limit
for emission, there will, in general, still be an apparent violation of
$e$--$\mu$ universality, i.e., a deviation in the measured ratio $R_{e/\mu} \equiv BR(\pi^+ \to e^+ \nu_e)/BR(\pi^+ \to \mu^+ \nu_\mu)$ from its SM value
\cite{Shrock:1980vy,Shrock:1980ct,Shrock:1981wq}. Constraints from early experiments on $\pi^+_{\ell 2}$ decays were reported in
Refs.~\cite{Abela:1981nf,Minehart:1981fv,Bryman:1983cja,Azuelos:1986eg,Britton:1992pg}.
Recent results on $\pi^+_{\ell 2}$ decays include
\cite{PiENu:2015seu,PIENU:2017wbj,PIENU:2019usb}.
Constraints from the non-observation of neutrinoless double beta
decays imply that the heavy neutrinos of relevance here are Dirac fermions. This
can be arranged in various NP scenarios. There are also constraints
from primordial nucleosynthesis. Heavy neutrinos with masses of
relevance here are consistent with these primordial nucleosynthesis
constraints in several NP theories \cite{Arguelles:2021dqn}.
Recent limits include \cite{Bryman:2019ssi,Bryman:2019bjg}.
Looking beyond sterile neutrinos, dark sectors that consist of particles that interact very feebly with the SM are highly motivated extensions of the SM. There is an ever-growing interest in exploring the parameter space of such scenarios. The PIONEER experiment has unique capabilities to search for pion decays to various light NP states.
One example is the three-body decay $\pi^+ \to \ell^+ \nu_\ell X$, where $X$ could, for example, be an axion-like-particle (ALP) that mixes with the neutral pion~\cite{Altmannshofer:2019yji}, or a light gauge boson coupling to differences of lepton numbers~\cite{Dror:2020fbh}. Another interesting process is the three-body decay $\pi^+ \to \ell^+ X Y$, where both $X$ and $Y$ are light dark sector particles~\cite{Batell:2017cmf}. Existing limits from PIENU~\cite{PIENU:2021clt} already probe some of these models. Moreover, PIONEER can also look for lepton-flavor-violating decays of the muon into light NP particles $\mu \to e X$.
\subsection{Theory Overview}
While no particles or interactions beyond those of the SM have been observed so far, intriguing hints for LFUV have been accumulated in recent years~\cite{Crivellin:2021sff,Fischer:2021sqw,Bryman:2021teu}. In particular, the measurements of the ratios of branching ratios (Br) $R(D^{(*)})=Br[ B\to D^{(*)}\tau\nu_\tau$]/Br[$B\to D^{(*)}\ell\nu_\ell$]~\cite{Lees:2012xj,Aaij:2017deq,Abdesselam:2019dgh} , where $\ell=\mu, e$, and $R(K^{(*)})=Br[B\to K^{(*)} \mu^+ \mu^-$]$/$Br[$B\to K^{(*)} e^+ e^-$]~\cite{Aaij:2017vbb,LHCb:2019hip,LHCb:2021trn} deviate from the SM expectation by more than $3\sigma$~\cite{Amhis:2019ckw,Murgui:2019czp,Shi:2019gxi,Blanke:2019qrx,Kumbhakar:2019avh} and $4\sigma$~\cite{Alguero:2019ptt,Aebischer:2019mlg,Ciuchini:2019usw,Arbey:2019duh}, respectively. In addition, anomalous magnetic moments $(g-2)_\ell$ ($\ell=e,\mu,\tau$) of charged leptons are intrinsically related to LFUV, as they are chirality flipping quantities. Here, the longstanding discrepancy in $(g-2)_\mu$, just reaffirmed at the level of $4.2\sigma$~\cite{Bennett:2006fi,Muong-2:2021ojo,Aoyama:2020ynm}, can be considered as another hint of LFUV, since, if compared to $(g-2)_e$, the NP contribution scales with a power of the lepton mass~\cite{Davoudiasl:2018fbb,Crivellin:2018qmi}. In addition, there is a hint for LFUV in the difference of the forward-backward asymmetries ($\Delta A_{\rm FB}$) in $B\to D^*\mu\nu$ vs $B\to D^*e\nu$~\cite{Bobeth:2021lya,Carvunis:2021dss}. As another possible indication of LFUV, CMS observed an excess in non-resonant di-electron pairs with respect to di-muons~\cite{Sirunyan:2021khd}. Furthermore, the possible deficit in first-row unitarity of the Cabibbo-Kobayashi-Maskawa (CKM) matrix, known as the Cabibbo angle anomaly (CAA) (see Fig.~\ref{fig:CKM} (left)), can also be viewed as a sign of LFUV~\cite{Coutinho:2019aiy,Crivellin:2020lzu}. For these reasons, there is very strong motivation for an upgraded $R_{e/\mu}$ experiment whose precision matches that of the SM prediction (see Sec.~\ref{sec:LFUV}). Moreover, the significance of the CAA depends crucially on experimental input quantities used for the extraction of CKM matrix elements as well as a number of theory corrections. Here, an improved measurement of pion beta decay would allow one to extract \ensuremath{\left|V_{ud}\right|}\xspace in a theoretically pristine manner, requiring a gain of a factor $3$ in experimental precision when combined with $K_{\ell 3}$ decays~\cite{Czarnecki:2019iwz} and an order of magnitude for a stand-alone extraction (see Sec.~\ref{sec:CKM}). Finally, an experiment capable of addressing these physics goals would at the same time be able to improve sensitivity to a host of exotic decays, as described in Sec.~\ref{sec:BSM}.
\subsection{Lepton flavor universality tests and $R_{e/\mu}$}
\label{sec:LFUV}
The branching ratio
$R_{e/\mu} = \frac{\Gamma\left(\pi^+ \rightarrow e^+ \nu (\gamma) \right)}{\Gamma\left(\pi^+ \rightarrow \mu^+ \nu (\gamma)\right)}$
for pion decays to electrons over muons provides the best test of electron--muon universality in charged-current weak interactions. In the SM, $R_{e/\mu}$ has been calculated with extraordinary precision at the $10^{-4}$ level as \cite{Cirigliano:2007xi,Cirigliano:2007ga,Marciano:1993sh}
\begin{equation}
\label{Remu_SM}
R_{e/\mu} \hspace{0.1cm}\text{(SM)} = 1.23524(15)\times10^{-4},
\end{equation}
perhaps the most precisely calculated weak interaction observable involving quarks.\footnote{Reference~\cite{Bryman:2011zz} estimates the uncertainty due to the unknown non-leading-logarithmic contributions of $O(\alpha^2 \log (m_\mu/m_e)$ in a different way compared to Ref.~\cite{Cirigliano:2007xi}. This leads to a larger total uncertainty, i.e., $R_{e/\mu} \hspace{0.1cm}\text{(SM)} = 1.23524(19)\times10^{-4}$.} Because the
uncertainty of the SM calculation for $R_{e/\mu}$ is very small and the decay $\pi^+ \rightarrow e^+ \nu$ is helicity-suppressed by the $V-A$ structure of charged currents, a measurement of $R_{e/\mu}$ is extremely sensitive to the presence of pseudoscalar (and scalar) couplings absent from the SM; a disagreement with the theoretical expectation would unambiguously imply the existence of NP. With measurements of 0.01\% experimental precision, NP at the PeV scale can be probed~\cite{Bryman:2011zz}, even up to several PeV in specific models such as leptoquarks.
The uncertainty of the SM prediction~\eqref{Remu_SM} for $R_{e/\mu}$ arises from low-energy constants in chiral perturbation theory, which absorb the divergences in the two-loop calculation of Refs.~\cite{Cirigliano:2007xi,Cirigliano:2007ga}, but whose finite parts need to be determined by other means. Fortunately, in the case of $R_{e/\mu}$ these non-perturbative uncertainties only affect the SM prediction at the relative precision of $10^{-4}$, more than an order of magnitude beyond the current experimental precision~\cite{ParticleDataGroup:2020ssz,PiENu:2015seu,Bryman:1985bv,Britton:1993cj,Czapek:1993kc}
\begin{equation}
R_{e/\mu} \hspace{0.1cm}\text{(exp)} =1.2327(23)\times10^{-4}.
\end{equation}
$R_{e/\mu}$ thus provides a unique opportunity for a pristine test of LFU in the quark sector.
The comparison between theory and experiment provides a stringent test of the $e$--$\mu$ universality of the weak interaction. We express the results in terms of the effective couplings $A_\ell$ multiplying the low-energy charged current contact interaction
\begin{equation}
{\cal L}_{CC} = A_\ell [\bar u \gamma^\mu P_L d] [ \bar \nu_\ell \gamma_\mu P_L \ell ],
\end{equation}
where $P_L \equiv (1-\gamma_5)/2$. In the SM at tree level the couplings are given by $A_\ell = - 2 \sqrt{2} G_F V_{ud}$ and thus satisfy LFU, i.e., $A_\ell / A_{\ell^\prime} = 1$.
The measurement of $R_{e/\mu}$ results in
\begin{equation}
\left( \frac{A_\mu}{A_e} \right)_{ R_{e/ \mu} }= 1.0010(9)\,,\label{gpi}
\end{equation}
which is in agreement with the SM expectation and provides the best available test of LFU.
A deviation from $A_\ell / A_{\ell^\prime} = 1$ can originate from various mechanisms. In the literature it is common to interpret deviations from $A_\ell/A_{\ell^\prime} = 1$ in terms of flavor-dependent couplings $g_\ell$ of the $W$-boson to
the leptonic current, in which case $A_\ell \propto g_\ell$. We note that in the context of modified $W$-boson couplings LFU tested with $R_{e/\mu} $ probes the couplings of a longitudinally polarized $W$-boson, whereas tests using purely leptonic reactions such as $\tau\to \ell \nu_\tau \nu_{\ell}$ ($\ell=e,\mu$) test the couplings of transversely polarized $W$-boson and are thus complementary.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\textwidth]{Figures/Fig1-compact.pdf}
\caption{Left: Tensions in the first-row CKM unitarity test (see text)~\cite{Bryman:2021teu}. Right: Constraints (1$\,\sigma$) on modified $W\ell\nu$ couplings from CKM unitarity (green) and LFUV (red) (adapted from Ref.~\cite{Crivellin:2020lzu}). The light bands show the current status and the dark bands include the expected PIONEER sensitivity. The SM values of $A_\ell$ are assumed to be modified by $1+\varepsilon_{\ell\ell}$ where $\ell=e,\mu$.}
\label{fig:CKM}
\end{figure}
Assuming that LFUV originates from modified $W\ell\nu$ couplings, the determination of CKM elements will also be affected. Importantly, beta decays have an enhanced sensitivity to a modified $W\mu\nu$ coupling, due to a CKM enhancement by $\left|V_{ud} / V_{us}\right|^2 \sim 20$. Such a modification of the $W\ell\nu$ couplings would also affect $R_{e/\mu}$, albeit for a different flavor combination (see Fig.~\ref{fig:CKM} (right)). This connection provides further motivation for an improved $R_{e/\mu}$ measurement, especially because the sensitivity to LFUV would be comparable to future improved constraints from beta decays. Moreover, recent global fits to electroweak observables and tests of LFU show a preference for $R_{e/\mu}$ larger than its SM expectation \cite{Coutinho:2019aiy,Crivellin:2020ebi}. Furthermore, $R(K^{*})$ can be correlated to $R_{e/\mu}$~\cite{Capdevila:2020rrl} and a combined explanation of the deficit in the first-row CKM unitarity and the CMS excess in di-electrons even predicts that $R_{e/\mu}$ should be larger than its SM value~\cite{Crivellin:2021rbf}.
\subsection{CKM unitarity and pion beta decay}
\label{sec:CKM}
The detector optimized for a next-generation $R_{e/\mu}$ experiment will also be ideally suited for a high-precision measurement of pion beta decay. Precision measurements of beta decays of neutrons,
nuclei, and mesons provide very accurate determinations of the elements \ensuremath{\left|V_{ud}\right|}\xspace and \ensuremath{\left|V_{us}\right|}\xspace of the
CKM quark-mixing matrix \cite{Cabibbo:1963yz,Kobayashi:1973fv}. Recent theoretical
developments on radiative corrections and form factors have led to a $ 3\sigma$ tension with CKM
unitarity illustrated in
Fig.~\ref{fig:CKM} (left)~\cite{Bryman:2021teu,Cirigliano:2021yto}, and with specific assumptions on theory corrections and processes considered, higher significances have been obtained~\cite{Seng:2021nar,Aoki:2021kgd}. However, in these determinations hadronic~\cite{Marciano:2005ec,Seng:2018yzq,Seng:2018qru,Czarnecki:2019mwq,Seng:2020wjq,Hayen:2020cxh,Shiells:2020fqp} and nuclear~\cite{Miller:2008my,Miller:2009cg,Gorchtein:2018fxl,Hardy:2020qwl} corrections play an important role, to the extent that they dominate the systematic uncertainty of the \ensuremath{\left|V_{ud}\right|}\xspace extraction from superallowed beta decays~\cite{Hardy:2020qwl}. A determination from pion beta decay would be much cleaner, provided it could be measured to sufficient precision.
The branching ratio for pion beta decay was most accurately measured by the PiBeta experiment at PSI
\cite{Pocanic5,Frlez:2003vg,Pocanic:2003pf,Frlez:2003pe,Bychkov:2008ws} to be
\begin{equation}\frac{\Gamma(\pi^+ \to \pi^0 e^+ \nu)}{\Gamma(\text{Total})}= 1.036 \pm 0.004 (\text{stat}) \pm 0.004(\text{syst}) \pm 0.003(\pi\to e\nu) \times 10^{-8},
\end{equation}
where the first uncertainty is statistical, the second systematic, and the third is the $\pi\to e\nu$ branching ratio uncertainty.
Pion beta decay, $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$, provides the theoretically cleanest determination of the magnitude of the CKM matrix element \ensuremath{\left|V_{ud}\right|}\xspace. With current input one obtains $\ensuremath{\left|V_{ud}\right|}\xspace = 0.9739(28)_{\textrm{exp}}(1)_{\textrm{th}}$, where the experimental uncertainty comes almost entirely from the $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ branching ratio (BRPB) \cite{Pocanic:2003pf} (the pion lifetime contributes ${\delta}V_{ud} = 0.0001$), and the theory uncertainty has been reduced from $({\delta}V_{ud})_{\textrm{th}} = 0.0005$ \cite{Sirlin:1977sv,Cirigliano:2002ng,Passera:2011ae} to $({\delta}V_{ud})_{\textrm{th}} = 0.0001$ via a lattice QCD calculation of the radiative corrections \cite{Feng:2020zdc}. The current precision of 0.3\% on \ensuremath{\left|V_{ud}\right|}\xspace makes $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ not presently relevant for the CKM unitarity tests because superallowed nuclear beta decays provide a nominal precision of
0.03\%.
In order to make $\pi^+ \rightarrow \pi^0 e^+ \nu (\gamma)$ important for CKM unitarity tests, two precision experimental stages can be identified:
(1) As advocated in Ref.~\cite{Czarnecki:2019iwz}, a three-fold improvement in BRPB precision compared to Ref.~\cite{Pocanic:2014jka} would allow for a 0.2\% determination of $\left|V_{us}/V_{ud}\right|$, via improving the measurement of the ratio
\begin{equation}
\label{rv}
R_V = \frac{\Gamma\left(K \rightarrow \pi l \nu (\gamma) \right)}{\Gamma\left(\pi^+ \rightarrow \pi^0 e^+
\nu (\gamma)\right)}=1.9884(115)_{\pi}(93)_K\times 10^7, \end{equation}
where the uncertainties are due to the pion partial width and the $K_L$ lifetime and branching ratio. Equation~\eqref{rv} is
independent of the Fermi constant and SM short-distance corrections, but subject to structure-dependent long-distance radiative corrections. This
would match the precision of the current extraction of $\left|V_{us} / V_{ud}\right|$ from the axial channels~\cite{Marciano:2004uf}, which proceeds via
\begin{equation}
R_A = \frac{\Gamma\left(K \rightarrow \mu \nu (\gamma) \right)}{\Gamma\left(\pi \rightarrow \mu \nu (\gamma)\right)}=1.3367(25),
\end{equation}
(see Fig. \ref{fig:CKM}), thus providing a new competitive constraint on the \ensuremath{\left|V_{us}\right|}\xspace--\ensuremath{\left|V_{ud}\right|}\xspace plane and probing NP that might affect vector and axial-vector channels in different ways.
The theoretical case for this approach was recently strengthened by improved analysis of radiative corrections in $K \to \pi e \nu $ decays \cite{Seng:2021wcf}.
(2) In the second phase, an order of magnitude improvement in the
BRPB precision will be sought. This would provide the theoretically cleanest extraction of \ensuremath{\left|V_{ud}\right|}\xspace at the 0.02\% level, comparable to the current value from superallowed beta decays~\cite{Hardy:2020qwl}.
\subsection{Constraints on New Physics}
There are several ways that (heavy) NP can affect $R_{e/\mu}$ and the extraction of \ensuremath{\left|V_{ud}\right|}\xspace from beta decays, relevant for this proposal. Concerning the former, there are several possibilities: a modified $W\ell\nu$ coupling and a contribution to an $\ell\nu u d$ operator ($\ell=\mu,e$), while only modified $W\mu\nu$ couplings and an $e\nu u d$ operator affect the latter. However, \ensuremath{\left|V_{ud}\right|}\xspace is also sensitive to a direct NP contribution to muon decays via an $e\mu\nu \nu$ operator entering through the Fermi constant, necessary to extract \ensuremath{\left|V_{ud}\right|}\xspace from beta decays~\cite{Crivellin:2021njn}.
The following NP models (see Ref.~\cite{deBlas:2017xtg} for a complete categorization) give potential observable tree-level effects in $R_{e/\mu}$ and/or in the determination of \ensuremath{\left|V_{ud}\right|}\xspace from beta decays (direct, or indirect via muon decay):
\begin{itemize}
\item A $W^\prime$ bosons can generate tree-level effects in beta decays and in $R_{e/\mu}$ as well as modified $W\ell\nu$ couplings via mixing with the SM $W$.
\item Vector-like leptons (VLL) affect $W\ell\nu$ couplings via their mixing with SM leptons after EW symmetry breaking (see Ref.~\cite{Crivellin:2020ebi} for a recent analysis).
\item A singly charged $SU(2)_L$ singlet scalar can give a necessarily constructive tree-level effect in muon decay such that the CAA can be solved~\cite{Crivellin:2020klg,Marzocca:2021azj}.
\item An $SU(2)_L$ triplet scalar can give a necessarily destructive effect in muon decay.
\item A neutral vector boson ($Z^\prime$) can give a constructive effect in muon decay if it has flavor-violating couplings.
\item A leptoquark can generate a left-handed vector current and/or a scalar current affecting beta decays and $R_{e/\mu}$.
\item A charged Higgs gives rise to pseudoscalar operators such that chirally enhanced effects in $R_{e/\mu}$ can be generated.
\end{itemize}
Therefore, the PIONEER results will be important for a very wide range of SM extensions containing one or more of these particles.
\section{Introduction}
\input{Introduction}
\section {Theory}
\input{Theory/Theory}
\input{Theory/Exotics}
\section{PIONEER Experiment}
\subsection{Experiment Overview and Strategy}
\input{PioneerExperiment/ExptStrategy-dh}
\subsection{Beam}
\subsubsection{Beam requirements and target setup}
\input{PioneerExperiment/Beam/beam1}
\subsubsection{Beam line at $\pi$E5}
\input{PioneerExperiment/Beam/beam2}
\subsubsection{Beam optimization and studies} \label{sec:beam_studies}
\input{PioneerExperiment/Beam/beam3}
\subsection{Active target (ATAR)}
\input{PioneerExperiment/ATAR/ATAR}
\subsection{Cylindrical Tracker}
\input{PioneerExperiment/ATAR/positron_detectors}
\subsection{LXe Calorimeter}
\subsubsection {Overview}
\input{PioneerExperiment/Calorimeter/caloverview}
\subsubsection {Photosensors}\label{sec: LXe_photosensors}
\input{PioneerExperiment/Calorimeter/Photosensors}
\subsubsection{Cryogenics and Purification}\label{sec:LXe_Cryogenics-purification}
\input{PioneerExperiment/Calorimeter/cryogenics-purification}
\subsubsection{Calorimeter Mechanical Engineering}\label{sec:Calorimeter_Mech_Eng}
\input{PioneerExperiment/Calorimeter/Calorimeter_Mech_Eng}
\subsubsection{Other Calorimeter Options}
\input{PioneerExperiment/Calorimeter/Othercal}
\subsection{Trigger and Backend electronics}
\input{PioneerExperiment/Trigger/trigger}
\input{PioneerExperiment/Trigger/electronics}
\subsection{DAQ}
\label{subsec:daq}
\input{DAQ/daq.tex}
\input{Simulations and Analysis/Geant4simulations}
\section{Sensitivity}
\input{Sensitivity/Sensitivities}
\section{ Planning for Realization of PIONEER}
\input{Planning/Planning}
\clearpage
| 592c58a9ea3e91d2c9197e863f647ef6cab23611 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{Experimental Evaluations of Channel Pruning and Weight Allocation Techniques}\label{pruning sec}
\noindent\textbf{Empirical benefits of our proposed FLOP-aware pruning.} }%\textcolor{blue}{Both SplitCIFAR100 (Table~\ref{CIFARtable}) and MNIST (Figure~\ref{fig:mnist}) experiments show that our pruning algorithm performs well (by comparison between Individual-0.2/-1)}\ylm{Table~\ref{CIFARtable} shows that our algorithm performs well in our CL manner}, but missing additional experimental comparisons to alternative channel-pruning techniques. In this section, we evaluate our FLOP-aware pruning algorithm separately without continual learning and compare our results to $\ell_1$-based~\cite{liu2017learning} and Polarization-based~\cite{zhuang2020neuron} channel pruning methods. For all the experiments, we use the same ResNet18 architecture and hyperparameters as Table~\ref{CIFARtable} and train over CIFAR10 dataset. Results are presented in Figure~\ref{fig:CIFAR10_acc} \& \ref{fig:CIFAR10_ratio} and they are all averaged from 5 trials. We implement these three methods over different FLOPs constraints and also track the fraction of nonzero channels after pruning. Figure~\ref{fig:CIFAR10_acc} shows test accuracy after pruning. We observe our FLOP-aware pruning method (shown in the Blue curve) does well in maintaining high accuracy compared to the other methods. The benefit of our approach is most visible when the FLOPs constraint is more aggressive (e.g., around 1\% FLOPs). The green curve shows the Polarization-based method and demonstrate that it does not execute the pruning task successfully with less than 50\% FLOPs constraint. We found that this is because below a certain threshold, Polarization tends to prune a whole layer resulting in disconnectivity within the network. Figure~\ref{fig:CIFAR10_ratio} shows the fraction of nonzero channels that remained after pruning, and our method succeeds in keeping more channels while pruning the same number of FLOPs. Here, we emphasize keeping more channels is a measure of connectivity within the network (as we wish to avoid very sparse layers). We also observe that $\ell_1$ works well even for small FLOPs; however, it uses the same $\lambda_l$ penalization for all layers thus it is not FLOP-aware and performs uniformly worse than our FLOP-aware algorithm.
\begin{figure*}[t]
\vspace{-7pt}
\centering
\begin{subfigure}[t]{.33\textwidth}
\centering
\begin{tikzpicture}\hspace{-0pt}
\node at (0,0) [scale=1.] {\includegraphics[width=\linewidth]{figs/CIFAR10.pdf}};
\node at (0,-1.9) [scale=0.8] {{FLOPs constraint $\gamma$}};
\node at (-2.5,0) [scale=0.8,rotate=90] {Accuracy};
\end{tikzpicture}
\centering
\caption{Test accuracy}
\hspace{-0pt}\label{fig:CIFAR10_acc}
\end{subfigure}\hspace{0pt}\begin{subfigure}[t]{.33\textwidth}
\centering
\begin{tikzpicture}\hspace{-0pt}
\node at (-0,0) [scale=1.] {\includegraphics[width=\linewidth]{figs/CIFAR10_channel_num.pdf}};
\node at (0,-1.9) [scale=0.8] {FLOPs constraint $\gamma$};
\node at (-2.75,0) [scale=0.9,rotate=90] {Fraction of Nonzero Channels};
\end{tikzpicture}
\caption{{Fraction of nonzero channels
}}\label{fig:CIFAR10_ratio}
\end{subfigure}\hspace{0pt}\begin{subfigure}[t]{.33\textwidth}
\centering
\begin{tikzpicture}\hspace{-0pt}
\node at (0,0) [scale=1.] {\includegraphics[width=\linewidth]{figs/FLOP.pdf}};
\node at (0,-1.9) [scale=0.8] {FLOPs constraint $\gamma$};
\node at (-2.5,0) [scale=0.9,rotate=90] {Accuracy};
\end{tikzpicture}
\centering
\caption{{Impact of weight allocation $\alpha$
}}\label{fig:FLOP}
\end{subfigure}\vspace{-10pt}
\caption{{
\red{Experimental evaluations of our FLOP-aware channel pruning and weight allocation techniques.} Fig.~\ref{fig:CIFAR10_acc} \& \ref{fig:CIFAR10_ratio} show channel pruning results for different methods. We compare our FLOP-aware pruning algorithm to the $\ell_1$-based~\cite{liu2017learning} and Polarization-based~\cite{zhuang2020neuron} methods over ResNet18 model used in our CL experiments. These experiments are conducted for standard classification tasks on CIFAR10 dataset and only focus on pruning performance (rather than CL setting). Our FLOP-aware pruning is displayed as the Blue curve, and it broadly outperforms the alternative approaches ($\ell_1$-based and Polarization-based). The performance gap is most notable in the very-few FLOPs regime (which is of more interest for inference-time efficiency ).
In Fig.~\ref{fig:FLOP}, we display the impact of weight allocation $\alpha$ (Eq.~\eqref{wa-eq}) and FLOPs constraint $\gamma$ on \emph{task-averaged ESPN accuracy} for SplitCIFAR100. Setting $\alpha$ too small (Blue curve) degrades performance as tasks do not get enough nonzeros to train on. Setting both $\alpha$ and $\gamma$ to be large also degrades performance (Red \& Green curves), because the first few tasks get to occupy the whole supernetwork at the expense of future tasks. Finally $\alpha=0.1$ (Orange) achieves competitive performance for all $\gamma$ choices.
}}\vspace{-7pt}
\end{figure*}
\section{Application to Shallow Networks}\label{app:application}
As a concrete instantiation of Theorem \ref{cl thm} let us consider a realizable regression setting with a shallow network. More sophisticated examples are deferred to future work. Fix positive integers $d$ and $r_\text{frz}\leq r$. Here, $d$ is the raw feature dimension and $r$ is the representation dimension which is often much smaller than $d$. The ingredients of our neural net example are as follows.
\begin{myenumerate}
\item Let $\psi:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz activation function with $\psi(0)=0$ such as Identity or (parametric) ReLU.
}%\textcolor{blue}{\item Let $\sigma:\mathbb{R}\rightarrow[-1,1]$ be a Lipschitz link function such as logistic function $1/(1+e^{-x})$.}
\item Let $Z$ be a zero-mean noise variable taking values on $[-1,1]$.
\item Fix vectors $(\vct{v}^\star_t)_{t=1}^T\in\mathbb{R}^r$ with $\ell_2$ norms bounded by some $B>0$.
\item Fix matrix $\mtx{W}^\star\in\mathbb{R}^{r\times d}$ with spectral norm (maximum singular value) upper bounded by some $\bar{B}>0$.
\item Given input $\vct{x}\in\mathbb{R}^d$, task $t$ samples an independent $Z$ and assigns the label
\begin{align}
y=\sigma({\vct{v}^\star_t}^\top \psi(\mtx{W}^\star\vct{x}))+Z.\label{planted}
\end{align}
\item Fix $r_\text{frz}\leq r$. Let $\mtx{W}_\text{frz}\in\mathbb{R}^{r\times d}$ be the matrix where the first $r_\text{frz}$ rows }%\textcolor{blue}{are same as $\mtx{W}^\star$}\ylm{This assumes no mismatch} whereas the last $r_\text{new}:=r-r_\text{frz}$ rows are equal to zero.
\end{myenumerate}
This setting assumes that first $r_\text{frz}$ features are generated by $\mtx{W}_\text{frz}$ and \eqref{crl} should learn remaining features $\mtx{W}^\star_\text{new}:=\mtx{W}^\star-\mtx{W}_\text{frz}$ and the classifier heads $(\vct{v}^\star_t)_{t=1}^T$. We remark that above one can use arbitrary $[-C,C]$ limits rather than $[-1,1]$ or one can replace $[-1,1]$ limits with subgaussian tail conditions.
}%\textcolor{blue}{For some $\bar{B}\geq \|\mtx{W}^\star\|_F$}, we choose search space ${\cal{W}}$ to be the set of all matrices $\mtx{W}_\text{new}$ such that spectral norm obeys $\|\mtx{W}_\text{new}\|\leq \bar{B}$ and the first $r_\text{frz}$ rows of $\mtx{W}_\text{new}$ are zero. This way, we focus on learning the missing part of the representation $\mtx{W}^\star$. Let us fix the loss function $\ell$ to be quadratic and denote ${\mtx{V}}=(\vct{v}_t)_{t=1}^T$. Then, \eqref{crl} takes the following parametric form
\begin{align}
&\hat{{\mtx{V}}},\hat{\mtx{W}}_\text{new}=\underset{\mtx{W}=\mtx{W}_\text{frz}+\mtx{W}_\text{new}}{\underset{\tn{\vct{v}_t}\leq B,\mtx{W}_\text{new}\in {\cal{W}}}{\arg\min}}{\widehat{\cal{L}}}({\mtx{V}},\mtx{W}):=\frac 1 T\sum_{t=1}^T}%_{\text{new}} {\widehat{\cal{L}}}_{\mathcal{S}_t}(\vct{v}_t,\mtx{W}) \nonumber\\
&\text{WHERE}\quad {\widehat{\cal{L}}}_{\mathcal{S}_t}(\vct{v}_t,\mtx{W})=\frac{1}{N}\sum_{i=1}^N (y_{ti}-\sigma(\vct{v}_t^\top \psi(\mtx{W}\vct{x}_{ti})))^2. \n
\end{align}
We have the following result regarding this optimization. It is essentially a corollary of Theorem \ref{cl thm}. The proof is deferred to the Appendix \ref{app B}.
\begin{theorem}\label{cl thm3} Consider the problem above with $T$ tasks containing $N$ samples each with datasets $(\mathcal{S}_t)_{t=1}^T$ generated according to \eqref{planted}. Suppose input domain $\mathcal{X}\in\mathbb{R}^d$ has bounded $\ell_2$ norm. With probability at least $1-2e^{-\tau}$, the task-averaged population risk of the solution $(\hat{{\mtx{V}}},\hat{\mtx{W}}_\text{new})$ obeys
\begin{align}
{\cal{L}}(\hat{{\mtx{V}}},\hat{\mtx{W}}_\text{new})\leq \operatorname{\mathbb{E}}[Z^2]+ \sqrt{\frac{\ordet{T}%_{\text{new}} r+r_\text{new} d+\tau}}{T}%_{\text{new}} N}}.\nonumber
\end{align}
\end{theorem}
\textbf{Interpretation:} Observe that the minimal risk is ${\cal{L}}^\star=\operatorname{\mathbb{E}}[Z^2]$ which is the noise independent of features. The additional components are the excess risk due to finite samples. In light of Theorem \ref{cl thm}, we simply plug in $\cc{{\mtx{H}}}=r$, $\cc{{\boldsymbol{\Phi}}_{\text{new}}}=r_\text{new} d$ and ${\cal{L}}^\star_{\phi_{\text{frz}}}={\cal{L}}^\star$. The first two arise from counting number of trainable parameters: each classifier has $r$ parameters and representation ${\boldsymbol{\Phi}}_{\text{new}}$ has $r_\text{new} d$ parameters. ${\cal{L}}^\star_{\phi_{\text{frz}}}={\cal{L}}^\star$ arises from the fact that we chose $\mtx{W}_\text{frz}$ to be subset of $\mtx{W}^\star$ thus there is no mismatch. }%\textcolor{blue}{When $\mtx{W}_\text{frz}=0$} (i.e.~learning representation from scratch), this bound is comparable to prior works \cite{tripuraneni2020provable,du2020few}, and in fact, it leads to (slightly) improved sample-complexity bounds. }%\textcolor{blue}{For instance, when $\psi$ is identity activation (i.e.~linear setting) \cite{tripuraneni2020provable} requires $TN\gtrsim r^2d$ samples to learn the whereas our sample size grows only linear in $r$ and requires $TN\gtrsim rd$. Additionally, \cite{du2020few} requires per-task sample sizes to obey $N\gtrsim d$ samples whereas we only require $N\gtrsim r$.}
\section{Appendix B: Proof of Theorem \ref{cl thm}}
In this section, we prove our main theoretical result Theorem~\ref{cl thm}.
\begin{proof} To be start, let ${\mtx{H}}_\varepsilon$ and ${\boldsymbol{\Phi}}_{\text{NEW},\varepsilon}$ be $\varepsilon$-covers of ${\mtx{H}}$ and ${\boldsymbol{\Phi}}_{\text{new}}$ and define $\vct\mathcal{F}={\mtx{H}}_\varepsilon^T\times{\boldsymbol{\Phi}}_{\text{NEW},\varepsilon}$. Following the Def.~\ref{def:cov} we have
\begin{align}
\log|\vct\mathcal{F}|\leq D ~~~\text{where}~~~\red{D:=(T\mathcal{C}({\mtx{H}})+\mathcal{C}({\boldsymbol{\Phi}}_{\text{new}}))\log((\bar{C}/\varepsilon))}. \nonumber
\end{align}
Set $t=\sqrt{\frac{D+\tau}{cNT}}$ where $c>$ is an absolute constant. Since loss function $\ell(\cdot)$ is bounded in $[0,1]$, then following Hoeffding inequality we have
\begin{align}
\P(|{\cal{L}}({\vct{\hat{h}}},\hat{\phi})- {\cal{L}}^{\st}_{{\phi}_{\text{frz}}}|\geq t)&\leq2|\vct\mathcal{F}|e^{-cNTt^2}, \nonumber\\
&\leq2|\vct\mathcal{F}|e^{-D-\tau}, \nonumber\\
&\leq2e^{-\tau}. \label{prob}
\end{align}
Now consider about the task-averaged risk perturbation introduce by covered set. Let $\phi_{\text{new}},\phi_{\text{new}}'\in{\boldsymbol{\Phi}}_{\text{new}},{\boldsymbol{\Phi}}_{\text{NEW},\varepsilon}$ and $h,h'\in{\mtx{H}},{\mtx{H}}_{\varepsilon}$ be $L$-Lipschitz functions (in Euclidean distance). Then $\phi:=\phi_{\text{new}}+\phi_{\text{frz}}$ and $\phi':=\phi_{\text{new}}'+\phi_{\text{frz}}$ also satisfy $L$-Lipschitz constraint.
\begin{align}
{\cal{L}}({\vct{h}},\phi)- {\cal{L}}({\vct{h}}',\phi')&\leq|{\cal{L}}({\vct{h}},\phi)- {\cal{L}}({\vct{h}}',\phi')|, \nonumber\\
&\leq \sup_{t\in[T]}|{\cal{L}}_{\mathcal{S}_t}(h_t,\phi)-{\cal{L}}(h_t',\phi')|, \nonumber\\
&\leq \sup_{t\in[T], i\in[N]}|\ell(y_{ti},h_t\circ\phi(\vct{x}_{ti}))-\ell(y_{ti},h_t'\circ\phi'(\vct{x}_{ti}))|, \nonumber \\
&\leq \Gamma\sup_{t\in[T], i\in[N]}|h_t\circ\phi(\vct{x}_{ti})-h_t'\circ\phi'(\vct{x}_{ti})|, \nonumber\\
&\leq\Gamma\sup_{t\in[T], i\in[N]}|h_t\circ\phi(\vct{x}_{ti})-h_t\circ\phi'(\vct{x}_{ti})|+|h_t\circ\phi'(\vct{x}_{ti})-h_t'\circ\phi'(\vct{x}_{ti})|, \nonumber\\
&\leq \Gamma (L+1)\varepsilon. \label{pert}
\end{align}
Combine results from \ref{prob} and \ref{pert} by setting $\varepsilon=$, then we find that with probability at least $1-2e^{-\tau}$
\begin{align}
{\cal{L}}({\vct{\hat{h}}},\hat{\phi})- {\cal{L}}^{\st}_{{\phi}_{\text{frz}}}\leq\Gamma(L+1)\varepsilon+\sqrt{\frac{D+\tau}{cNT}}\nonumber.
\end{align}
\end{proof}
\section{Experimental Setting of Figure~\ref{fig:diversity}}\label{app:imagenet}
Following \cite{mallya2018packnet,hung2019compacting}, we conduct experiments in Sec.~\ref{sec:crl_exp} to study how task order and diversity benefits CL, we use $6$ image classification tasks, where ImageNet-1k~\cite{krizhevsky2012imagenet} is the first task, followed by CUBS~\cite{wah2011caltech}, Stanford Cars~\cite{krause20133d}, Flowers~\cite{nilsback2008automated}, WikiArt~\cite{saleh2015large} and Sketch~\cite{eitz2012humans}, Table~\ref{table:crl_dataset} provides the detail information of all datasets. In the experiment, we train a standard ResNet50 model on the last 5 tasks (CUBS to Sketch) to explore how the representation learned from ImageNet can benefit CRL. We follow the same learning hyperparameter setting as Table~\ref{CIFARtable}, where we use batch size of 128 and Adam optimizer with $(\beta_1,\beta_2)=(0,0.999)$. Also we train $60$ and $90$ epochs for pre-training and pruning with learning rate $0.01$, and then we fine-tune the free weights for $100$ epochs using cosine decay over learning rate starting from $0.01$.
Specifically, we employ weight allocation $\alpha=0.1$ (Sec.~\ref{appsec: espnalgo}) in both individual and ESPN to enable continue representation learning. Compared with Individual/ESPN without pretrain, the ESPN with ImageNet pretrain employs a sparse pretrained ImageNet model from \cite{Wortsman2019DiscoveringNW} with 20\% non-zero weights. Table~\ref{table:crl_exp} shows the test accuracy when learning the
last 5 tasks and performances get improved in all 5 tasks.
}
\section{Proof of Theorems \ref{cl thm} and \ref{cl thm3}}\label{app B}
In this section and in Appendix \ref{app C}, we provide two main theorems: Theorem \ref{cl thm} and Theorem \ref{seq thm}. We will first prove Theorem \ref{cl thm} which adds $T$ new tasks by leveraging a frozen feature extractor $\phi_{\text{frz}}$. Secondly, observe that Theorem \ref{cl thm} adds the $T$ new tasks in a multitask fashion which models the setting where new tasks arrive in batches of size $T$. This still captures continual learning due to the use of the frozen feature-extractor that corresponds to the features built by earlier tasks. Also, setting $T=1$ in Theorem \ref{cl thm} corresponds to adding a single new task and updating the representation. Using this observation, we provide an additional result Theorem \ref{seq thm} where the tasks are added to the super-network sequentially. Theorem \ref{seq thm}, provided in Appendix \ref{app C}, arguably better captures the CL setting. In essence, it follows as an iterative applications of Theorem \ref{cl thm} after introducing proper definitions that capture the impact of imperfections due to finite sample learning (see Definition \ref{def pop seq} and Assumption \ref{fin comp}).
\subsection{Proof of Theorem \ref{cl thm}}
The original statement of Theorem \ref{cl thm} does not introduce certain terms formally due to space limitations. Here, we first add some clarifying remarks on this. Let $\mathcal{B}(S)$ return the smallest Euclidean ball containing the set $S$ (that lies on an Euclidean space). For a function $f:\mathcal{Z}\rightarrow\mathcal{Z}'$, define the $\lin{\cdot}$ norm to be the Lipschitz constant $\lin{f}:=\lin{f}^{\mathcal{Z}}=\sup_{\vct{x},\vct{y}\in \mathcal{B}(\mathcal{Z})}\frac{\tn{f(\vct{x})-f(\vct{y})}}{\tn{\vct{x}-\vct{y}}}$. Below, we assume that $\phi_{\text{frz}}$ and all $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}$, $h\in{\mtx{H}}$ are $L$-Lipschitz (on their sets of input features). Suppose $\mathcal{X}$ is the set of feasible input features and $\mathcal{X}$ has bounded Euclidian radius $R=\sup_{\vct{x}\in\mathcal{X}}\tn{\vct{x}}$. This means that input features of the classifier $h$ lie on the set $\mathcal{X}'=\phi_{\text{frz}}\circ{\mtx{X}}+{\boldsymbol{\Phi}}_{\text{new}}\circ{\mtx{X}}$ with Euclidian radius bounded by $2LR$. Thus both raw features and intermediate features \ylm{(output of $\phi\in{\boldsymbol{\Phi}}$)} are bounded {and we use these sets in Def.~\ref{def:cov}.} {Thus, we set $\bar{C}:=\max(\bar{C}_{\mathcal{X}},\bar{C}_{\mathcal{X}'})$ below.}
\begin{theorem}[Theorem \ref{cl thm} restated] \label{cl thm2} Recall that $({\vct{\hat{h}}},\hat{\phi}=\hat{\phi}_{\text{new}}+{\phi}_{\text{frz}})$ is the solution of \eqref{crl}. Suppose that the loss function $\ell(y,\hat{y})$ takes values on $[0,B]$ and is $\Gamma$-Lipschitz w.r.t.~$\hat{y}$. Draw $T$ independent datasets $\{(\vct{x}_{ti},y_{ti})\}_{i=1}^N\subset \mathcal{X}\times \mathcal{Y}$ for $t\in [T]$ where each dataset is distributed i.i.d.~according to ${\cal{D}}_t$. Suppose the input set $\mathcal{X}$ has bounded Euclidean radius. Suppose $\lin{\phi_{\text{frz}}},\lin{\phi_{\text{new}}},\lin{h}\leq L$ for all $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}},h\in{\mtx{H}}$. With probability at least $1-2e^{-\tau}$, for some absolute constant $C>0$, the task-averaged population risk of the solution $({\vct{\hat{h}}},\hat{\phi})$ obeys
\begin{align}
{\cal{L}}({\vct{\hat{h}}},\hat{\phi})&\leq {\cal{L}}^{\st}_{{\phi}_{\text{frz}}}+ CB\sqrt{\frac{(T\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}})\log(\bar{C}\Gamma (L+1)NT)+\tau}{TN}},\nonumber\\
&\leq {\cal{L}}^{\st}+\text{MM}_{\text{frz}}}%^{\text{new}+CB\sqrt{\frac{(T\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}})\log(\bar{C}\Gamma (L+1)NT)+\tau}{TN}}\nonumber.
\end{align}
\end{theorem}
\begin{proof} Below $c,C>0$ denote absolute constants. For a scalar $a$, define $a_+=a+1$. The proof uses a covering argument following the definition of the covering dimension. Fix $1>\varepsilon>0$. To start with, let ${\boldsymbol{\Phi}}_{\text{new},\eps}$ and ${\mtx{H}}_\varepsilon$ be $\varepsilon$-covers (per Definition \ref{def:cov}) of the sets ${\boldsymbol{\Phi}}_{\text{new}}$ and ${\mtx{H}}$ respectively. Let ${\mtx{H}}_\varepsilon^T$ be $T$-times Cartesian product of ${\mtx{H}}_\varepsilon$. Our goal is bounding the supremum of the gap between the empirical and population risks to conclude with the result.
\noindent\textbf{$\bullet$ Step 1: Union bound over the cover.} Following Definition \ref{def:cov}, we have that $\log|{\mtx{H}}^T_\varepsilon|\leq T\cc{{\mtx{H}}}\log(\bar{C}/\varepsilon)$, $\log|{\boldsymbol{\Phi}}_{\text{new},\eps}| \leq \cc{{\boldsymbol{\Phi}}_{\text{new}}}\log(\bar{C}/\varepsilon)$. Define $\mathcal{S}={\boldsymbol{\Phi}}_{\text{new},\eps}\times {\mtx{H}}^T_{\varepsilon}$. These imply that
\[
\log|\mathcal{S}|\leq D\quad\text{where}\quad D:=(\cc{{\boldsymbol{\Phi}}_{\text{new}}}+T\cc{{\mtx{H}}})\log\left(\frac{\bar{C}}{\varepsilon}\right)
\]
Set $t=B\sqrt{\frac{D+\tau}{cNT}}$ for an absolute constant $c>0$ which corresponds to the concentration rate of the Hoeffding inequality that follows next. Using the fact that the loss function is bounded in $[0,B]$, applying a Hoeffding bound over all elements of the cover, we find that for all }%\textcolor{blue}{$(\vct{{\vct{h}}},\phi)\in \mathcal{S}$}
\begin{align}
\mathbb{P}(|{\cal{L}}(\vct{{\vct{h}}},\phi)-{\widehat{\cal{L}}}(\vct{{\vct{h}}},\phi)|\geq t)&\leq 2|\mathcal{S}|e^{-\frac{cNTt^2}{B^2}}\nonumber\\
&\leq 2|\mathcal{S}|e^{-D-\tau}=2e^{\log|\mathcal{S}|-D-\tau}\nonumber\\
&\leq 2e^{-\tau}.\label{prob bound}
\end{align}
\noindent\textbf{$\bullet$ Step 2: Perturbation analysis.} We showed the concentration on the cover. Now we need to relate the cover to the continuous set. Given any candidate $\phi=\phi_{\text{new}}+\phi_{\text{frz}}$ with $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}},\vct{{\vct{h}}}:=(h_t)_{t=1}^T\in{\mtx{H}}^T$, we draw a neighbor $\phi'=\phi_{\text{new}}'+\phi_{\text{frz}}$ with $\phi_{\text{new}}'\in{\boldsymbol{\Phi}}_{\text{new},\eps},\vct{{\vct{h}}}':=(h'_t)_{t=1}^T\in{\mtx{H}}_\varepsilon^T$.
Recall that $\lin{\phi_{\text{new}}},\lin{h}\leq L$ for all $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}},h\in{\mtx{H}}$. The task-averaged risk perturbation relates to the individual risks which in turn relates to individual examples as follows. Let $f_t=h_t\circ \phi$ for all $t\in[T]$.
\begin{align}
|{\cal{L}}(\vct{{\vct{h}}},\phi)-{\cal{L}}(\vct{{\vct{h}}}',\phi')|&\leq \sup_{1\leq t\leq T} |{\cal{L}}_t(h_t,\phi)-{\cal{L}}_t(h_t',\phi')|\\
&\leq\sup_{i\in[n],t\in[T]}|\ell(y_{ti},f_t(\vct{x}_{ti}))-\ell(y_{ti},f'_t(\vct{x}_{ti}))|\\
&\leq\Gamma\sup_{i\in[n],t\in[T]}|f_t(\vct{x}_{ti})-f'_t(\vct{x}_{ti})|.\label{pert1 bound}
\end{align}
The perturbations for the individual examples are bounded via triangle inequalities as follows.
\begin{align}
|f_t(\vct{x}_{ti})-f'_t(\vct{x}_{ti})|\leq &|h_t\circ\phi(\vct{x}_{ti})-h'_t\circ\phi'(\vct{x}_{ti})|\nonumber\\
\leq & (|h_t\circ\phi(\vct{x}_{ti})-h_t\circ\phi'(\vct{x}_{ti})|+|h_t\circ\phi'(\vct{x}_{ti})-h'_t\circ\phi'(\vct{x}_{ti})|)\nonumber\\
\leq& (L+1)\varepsilon:=L_+\varepsilon.\label{last line}
\end{align}
The last line follows from the triangle inequality via $\varepsilon$-covering and $L$-Lipschitzness of $h$ and ${\phi}$ as follows
\begin{itemize}
\item Since $\tn{\phi(\vct{x}_{ti})-\phi'(\vct{x}_{ti})}=\tn{\phi_{\text{new}}(\vct{x}_{ti})-\phi_{\text{new}}'(\vct{x}_{ti})}\leq \varepsilon\implies |h_t\circ\phi(\vct{x}_{ti})-h_t\circ\phi'(\vct{x}_{ti})|\leq L\varepsilon$.
\item Set $\vct{v}=\phi'(\vct{x}_{ti})$. Since $\tn{h(\vct{v})-h'(\vct{v})}\leq \varepsilon\implies |h'_t\circ\phi'(\vct{x}_{ti})-h_t\circ\phi'(\vct{x}_{ti})|\leq \varepsilon$.
\end{itemize}
Combining these, following \eqref{pert1 bound}, and repeating the identical perturbation argument \eqref{last line} for the empirical risk ${\widehat{\cal{L}}}$, we obtain
\begin{align}
\max(|{\cal{L}}(\vct{{\vct{h}}},\phi)-{\cal{L}}(\vct{{\vct{h}}}',\phi')|, |{\widehat{\cal{L}}}(\vct{{\vct{h}}},\phi)-{\widehat{\cal{L}}}(\vct{{\vct{h}}}',\phi')|)\leq \Gamma L_+\varepsilon.\label{pert bound}
\end{align}
\noindent\textbf{$\bullet$ Step 3: Putting things together.} Combining \eqref{prob bound} and \eqref{pert bound}, we found that, with probability at least $1-2e^{-\tau}$, for all ${\phi}=\phi_{\text{new}}+\phi_{\text{frz}}$ with $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}$ and all $\vct{{\vct{h}}}\in{\mtx{H}}^T$, we have that
\begin{align}
|{\cal{L}}(\vct{{\vct{h}}},\phi)-{\widehat{\cal{L}}}(\vct{{\vct{h}}},\phi)|\leq 2\Gamma L_+\varepsilon+B\sqrt{\frac{D+\tau}{cNT}}.
\end{align}
Setting }%\textcolor{blue}{$\varepsilon=\frac{1}{\Gamma L_+NT}$}, for an updated constant $C>0$, we find
\begin{align}
|{\cal{L}}(\vct{{\vct{h}}},\phi)-{\widehat{\cal{L}}}(\vct{{\vct{h}}},\phi)|\leq CB\sqrt{\frac{D+\tau}{NT}}.\label{unif conv}
\end{align}
where }%\textcolor{blue}{$D=(\cc{{\boldsymbol{\Phi}}_{\text{new}}}+T\cc{{\mtx{H}}})\log(\bar{C}\Gamma L_+NT)$} following the above definition of $D$.
Note that, the uniform concentration above also implies the identical bound for the minimizer of the empirical risk. Let $(\hat\vct{{\vct{h}}},\hat{\phi})$ be the minimizer of the empirical risk. Specifically, let $(\vct{{\vct{h}}}^{\star,\phi_{\text{frz}}},\phi_{\text{new}}^{\star,\phi_{\text{frz}}})$ be the optimal hypothesis in $({\mtx{H}},{\boldsymbol{\Phi}}_{\text{new}})$ minimizing the population risk subject to using frozen feature extractor $\phi_{\text{frz}}$, that is, ${\cal{L}}(\vct{{\vct{h}}}^{\star,\phi_{\text{frz}}},\phi_{\text{new}}^{\star,\phi_{\text{frz}}}+\phi_{\text{frz}})={\cal{L}}^\star_{\phi_{\text{frz}}}$. Then, we note that
\[
{\cal{L}}(\hat\vct{{\vct{h}}},\hat{\phi})\leq {\widehat{\cal{L}}}(\hat\vct{{\vct{h}}},\hat{\phi})+CB\sqrt{\frac{D+\tau}{NT}}\leq {\widehat{\cal{L}}}(\vct{{\vct{h}}}^{\star,\phi_{\text{frz}}},\phi_{\text{new}}^{\star,\phi_{\text{frz}}}+\phi_{\text{frz}})+CB\sqrt{\frac{D+\tau}{NT}}\leq {\cal{L}}^\star_{\phi_{\text{frz}}}+2CB\sqrt{\frac{D+\tau}{NT}}.
\]
This concludes the proof of the main statement (first line). The second statement follows directly from the definition of mismatch, that is, using the fact that $\text{MM}_{\text{frz}}}%^{\text{new}={\cal{L}}^\star_{\phi_{\text{frz}}}-{\cal{L}}^\star$.
\end{proof}
\subsection{Proof of Theorem \ref{cl thm3}}
\begin{proof}
Within this setting, classifier heads correspond to $h(\vct{a}):=h_{\vct{v}}(\vct{a})=\sigma(\vct{v}^\top\psi(\vct{a}))$ and ${\mtx{H}}=\{h_{\vct{v}}{~\big |~} \tn{\vct{v}}\leq B\}$. Similarly, feature representations correspond to $\phi(\vct{x})=\mtx{W}\vct{x}$, $\phi_{\text{frz}}(\vct{x})=\mtx{W}_\text{frz}\vct{x}$, $\phi_{\text{new}}(\vct{x}):=\phi_{\text{new}}^{\mtx{W}_\text{new}}(\vct{x})=\mtx{W}_\text{new}\vct{x}$ where $\mtx{W}=\mtx{W}_\text{frz}+\mtx{W}_\text{new}\in\mathbb{R}^{r\times d}$. The hypothesis set becomes ${\boldsymbol{\Phi}}_{\text{new}}=\{\phi_{\text{new}}^{\mtx{W}_\text{new}}{~\big |~} \mtx{W}_\text{new}\in{\cal{W}}\}$.
Here $\mtx{W}_\text{new}\in{\cal{W}}$ is the weights of the new feature representation to learn on top of $\mtx{W}_\text{frz}$. Importantly, $\mtx{W}_\text{new}$ only learns the last $r_\text{new}$ rows since first $r_\text{frz}$ rows are fixed by frozen feature extractor $\mtx{W}_\text{frz}$. For the proof, we simply need to plug in the proper quantities within Theorem \ref{cl thm}. First, observe that ${\cal{L}}^\star=\operatorname{\mathbb{E}}[Z]^2$ since $Z$ is independent zero-mean noise thus for any predictor using $\hat{y}:=\hat{y}(\vct{x})$ input features we have
\[
\operatorname{\mathbb{E}}[\ell(y,\hat{y})]=\operatorname{\mathbb{E}}[(y(\vct{x})+Z-\hat{y}(\vct{x}))^2]\geq \operatorname{\mathbb{E}}[Z^2],
\]
where $y(\vct{x})=\sigma({\vct{v}^\star_t}^\top \psi(\mtx{W}^\star\vct{x}))$ is the noiseless label.
We next prove that ${\cal{L}}^\star_\text{frz}={\cal{L}}^\star=\operatorname{\mathbb{E}}[Z^2]$. This simply follows from the fact that frozen representation $\mtx{W}_\text{frz}$ is perfectly compatible with ground-truth representation $\mtx{W}^\star$. Specifically, observe that $\mtx{W}^\star_\text{new}:=\mtx{W}^\star-\mtx{W}_\text{frz}$ lies within the hypothesis set ${\cal{W}}$ since by construction $\|\mtx{W}^\star_\text{new}\|\leq \|\mtx{W}^\star\|\leq \bar{B}$ and $\mtx{W}^\star_\text{new}$ is zero in the first $r_\text{frz}$ rows. Similarly $(\vct{v}^\star_t)_{t=1}^T$ obey the $\ell_2$ norm constraint $\tn{\vct{v}^\star_t}\leq B$. Thus, $\mtx{W}^\star_\text{new},{\mtx{V}}^\star=(\vct{v}^\star_t)_{t=1}^T$ are feasible solutions of the hypothesis space and since $\mtx{W}^\star_\text{new}+\mtx{W}_\text{frz}=\mtx{W}^\star$, for this choice we have that $\hat{y}(\vct{x})=y(\vct{x})$ thus task-specific risks induced by $(\vct{v}^\star_t,\mtx{W}^\star)$ obey ${\cal{L}}_t(\vct{v}^\star_t,\mtx{W}^\star)=\operatorname{\mathbb{E}}[Z^2]$ for all $t\in[T]$. Consequently, the task-averaged risk obeys ${\cal{L}}({\mtx{V}}^\star,\mtx{W}^\star)=\operatorname{\mathbb{E}}[Z^2]$ proving aforementioned claim.
The remaining task is bounding the covering dimensions of the hypothesis sets ${\mtx{H}}$ and ${\boldsymbol{\Phi}}_{\text{new}}$ and verifying Lipschitzness. The Lipschitzness of $h(\vct{a})=\sigma(\vct{v}^\top\psi(\vct{a}))\in{\mtx{H}}$, $\phi_{\text{frz}}(\vct{x})=\mtx{W}_\text{frz}\vct{x}$, $\phi_{\text{new}}(\vct{x})=\mtx{W}_\text{new}\vct{x}\in{\boldsymbol{\Phi}}_{\text{new}}$ follows from the fact that all $\mtx{W}_\text{new}\in{\cal{W}},\mtx{W}^\star,\mtx{W}_\text{frz}$ have spectral norms bounded by $\bar{B}$, and the fact that, the Lipschitz constant of $h$ (denoted by $\frac{5B^2\tn{\vct{y}}}{\laz^2}{\cdot}$) can be bounded as $\frac{5B^2\tn{\vct{y}}}{\laz^2}{h}\leq \frac{5B^2\tn{\vct{y}}}{\laz^2}{\sigma} \frac{5B^2\tn{\vct{y}}}{\laz^2}{\psi}\tn{\vct{v}}\leq \bar{L}^2 B$ where $\bar{L}=\max(\frac{5B^2\tn{\vct{y}}}{\laz^2}{\psi},\frac{5B^2\tn{\vct{y}}}{\laz^2}{\sigma})$. Recall that dependence on the Lipschitz constant is logarithmic.
{What remains is determining the covering dimensions of ${\mtx{H}},{\boldsymbol{\Phi}}_{\text{new}}$ which simply follows from covering the parameter spaces of $\vct{v}\in\mathcal{B}^d(B),\mtx{W}_\text{new}\in{\cal{W}}$. Here $\mathcal{B}^d(B)$ is defined to be the Euclidean ball of radius $B$. Suppose $\mathcal{X}$ lies on an Euclidean ball of radius $R$ and let $\mathcal{F}=(\{\phi_{\text{frz}}\}+{\boldsymbol{\Phi}}_{\text{new}})\circ \mathcal{X}$ be the feature representations. Since $\mtx{W}_\text{frz}$ and $\mtx{W}_\text{new}$ have spectral norm at most $\bar{B}$, $\mathcal{F}$ is subset of Euclidean ball of radius $2\bar{B} R$.}
{Fix an $\varepsilon_0=\frac{\varepsilon}{2\bar{B} R\bar{L}^2}$ $\ell_2$-cover of $\mathcal{B}^d(B)$. This cover has cardinality at most $(\frac{6B\bar{B} R\bar{L}^2}{\varepsilon})^d$ and induces an $\varepsilon$ cover of ${\mtx{H}}$. To see this, given any $\vct{f}\in\mathcal{F}$ and $\vct{v}\in \mathcal{B}^d(B)$, there exists a cover element $\vct{v}'$ with $\tn{\vct{v}'-\vct{v}}\leq \varepsilon_0$, as a result $|h_{\vct{v}'}(\vct{f})-h_{\vct{v}}(\vct{f})|\leq \bar{L}^2\tn{\vct{f}}\tn{\vct{v}'-\vct{v}}\leq \varepsilon$. Consequently $\cc{{\mtx{H}}}=d$. Similarly, since elements of ${\cal{W}}$ has $r_\text{new} d$ nonzero parameters (and recalling that ${\cal{W}}$ is also subset of spectral norm ball of radius $\bar{B}$) ${\cal{W}}$ admits a $\frac{\varepsilon}{R}$ Frobenius cover of cardinality $(\frac{3R\bar{B}\sqrt{r_\text{new}}}{\varepsilon})^{r_\text{new} d}$. Consequently, for any $\phi_{\mtx{W}}$ with $\mtx{W}=\mtx{W}_\text{new}+\mtx{W}_\text{frz}$, there exists a cover element $\phi_{\mtx{W}'}$ with $\|\mtx{W}'_{\text{new}}-\mtx{W}_\text{new}\|\leq \tf{\mtx{W}'_\text{new}-\mtx{W}_\text{new}}\leq \frac{\varepsilon}{R}$, such that for all $\vct{x}\in\mathcal{X}$, we have that $|\phi_{\mtx{W}}-\phi_{\mtx{W}'}|\leq \|\mtx{W}'_{\text{new}}-\mtx{W}_\text{new}\|\tn{\vct{x}}\leq \varepsilon$. Consequently $\cc{{\boldsymbol{\Phi}}_{\text{new}}}\leq r_\text{new} d$. These bounds on covering dimensions conclude the proof after applying Theorem \ref{cl thm}.}
\end{proof}
\section{Theoretical Analysis of Adding $T$ New Tasks Sequentially (Theorem \ref{seq thm})}\label{app C}
Theorem \ref{cl thm} adds $T$ tasks simultaneously on frozen feature extractor $\phi_{\text{frz}}$. Below, we consider the setting where we add these tasks sequentially via repeated applications of Theorem \ref{cl thm} and a new task $t$ builds upon the cumulative representation learned from tasks $1$ to $t-1$. This setting better reflects what actually happens in continual learning and in our experiments but is more involved because representation quality of an earlier task will impact the accuracy of the future tasks. To this end, we first describe the learning setting and assumptions on the representation mismatch.
\noindent\textbf{Sequential learning setting:} We will learn a new task with index $t$ over the hypothesis set ${\boldsymbol{\Phi}}_{\text{new}}^t$ for $t\in[T]$. Suppose we are at task $t$. That is, we assume that we have already built incremental (continual) feature-extractors $\phi_{\text{new}}^1,\dots,\phi_{\text{new}}^{t-1}$ for tasks $1$ through $t-1$ where each one obeys $\phi_{\text{new}}^\tau\in{\boldsymbol{\Phi}}_{\text{new}}^\tau$. Thus, the (cumulative) frozen representation at time $t$ is given by
\[
\phi_{\text{frz}}^t=\phi_{\text{frz}}+\sum_{i=1}^{t-1} \phi_{\text{new}}^i.
\]
Here $\phi_{\text{frz}}:=\phi_{\text{frz}}^1$ is the representation built before any new task arrived. Using $\phi_{\text{frz}}^t$, we solve the following (essentially identical) variation of \eqref{crl} where \textbf{we focus on task $t$ given the outcome of the continual learning procedure until task $t-1$.}
\begin{align}
h^t,&\phi_{\text{new}}^t=\underset{\phi=\phi_{\text{new}}+{\phi}_{\text{frz}}^t}{\underset{h\in{\mtx{H}},\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}^t}{\arg\min}} {\widehat{\cal{L}}}_{\mathcal{S}_t}(f)=\frac{1}{N}\sum_{i=1}^N \ell(y_{ti},f(\vct{x}_{ti}))\quad\text{where}\quad f=h\circ\phi. \tag{CRL-SEQ}\label{crlseq}
\end{align}
After obtaining $\phi_{\text{new}}^t$, the feature-extractor of task $t$ is given by $\phi^t=\phi_{\text{frz}}^t+\phi_{\text{new}}^t$ and prediction function becomes $f^t=h^t\circ \phi^t$. Finally, $\phi^t$ of task $t$ becomes the next frozen feature-extractor i.e.~$\phi_{\text{frz}}^{t+1}=\phi^t$.
In this sequential setting, intuitively $({\boldsymbol{\Phi}}_{\text{new}}^\tau)_{\tau=1}^T$ are less complex hypothesis spaces compared to ${\boldsymbol{\Phi}}_{\text{new}}$ of Theorem \ref{cl thm}. This is because we learn ${\boldsymbol{\Phi}}_{\text{new}}^\tau$ using a single task. In that sense, the proper scaling of hypothesis set complexity is $\cc{{\boldsymbol{\Phi}}_{\text{new}}^\tau}\propto \cc{{\boldsymbol{\Phi}}_{\text{new}}}/T$ for $\tau\in[T]$. Specifically, we assume that for some global value $\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}>0$
\begin{align}
\cc{{\boldsymbol{\Phi}}_{\text{new}}^\tau}\leq \mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}\quad \text{for all}\quad 1\leq\tau\leq T.\label{ccn decay}
\end{align}
Secondly, compared to Theorem \ref{cl thm}, we need to introduce a more intricate compatibility condition to assess the benefit of the representations learned from finite data $\phi_{\text{new}}^1,\dots,\phi_{\text{new}}^{t-1}$ for the new task $t$. This will be accomplished by first introducing population level compatibility and then introducing an assumption that controls the impact of finite sample learning on the new task. These definition and assumption arise naturally to control the learnability of a new task given features of earlier tasks. Related assumptions (e.g.~\emph{task diversity} condition) have been used by other works for transfer/meta learning purposes \cite{tripuraneni2020theory,oymak2021generalization,du2020few,xu2021representation}.
Let $(h^{\star,1},\phi_{\text{new}}^{\star,1}),\dots,(h^{\star,t},\phi_{\text{new}}^{\star,t}),\dots$ be the (classifier, representation) sequence obtained by solving \eqref{crlseq} using infinite samples $N=\infty$ (that is, solving the population-level optimization rather than finite-sample ERM). The following definition introduces the representation mismatch at task $t$ to capture the suitability of the population-level representations $(\phi_{\text{new}}^{\star,\tau})_{\tau=1}^{t-1}$ for a new task $t$. Set $\phi_{\text{frz}}^{\star,t}=\phi_{\text{frz}}+\sum_{\tau=1}^{t-1}\phi_{\text{new}}^{\star,\tau}$ and define the set of all feasible representations for task $t$ as }%\textcolor{blue}{${\boldsymbol{\Phi}}^t=\sum_{\tau=1}^t{\boldsymbol{\Phi}}_{\text{new}}^\tau+{\boldsymbol{\Phi}}_{\text{frz}}$\red{$\subseteq{\boldsymbol{\Phi}}$}.}
\begin{definition}[Population quantities and representation mismatch]\label{def pop seq}For task $t$, define the population (infinite-sample) risk as ${\cal{L}}_t(h,\phi)=\operatorname{\mathbb{E}}[{\widehat{\cal{L}}}_{\mathcal{S}_t}(h,\phi)]$. Define the optimal risk of task $t$ over all feasible representations in \red{${\boldsymbol{\Phi}}$} as $\Lci{t}=$ $\min_{h\in{\mtx{H}},\phi\in\red{{\boldsymbol{\Phi}}}}{\cal{L}}_t(h,\phi)$. Note that the optimal risk gets to choose the best representations within $({\boldsymbol{\Phi}}_{\text{new}}^\tau)_{\tau=1}^t$ and ${\boldsymbol{\Phi}}_{\text{frz}}$ \red{since $\sum_{\tau=1}^t{\boldsymbol{\Phi}}_{\text{new}}^\tau+{\boldsymbol{\Phi}}_{\text{frz}}\subseteq {\boldsymbol{\Phi}}$.} Finally, define the optimal risk with fixed frozen model $\phi_{\text{frz}}^{\star,t}=\phi_{\text{frz}}+\sum_{\tau=1}^{t-1}\phi_{\text{new}}^{\star,\tau}$ to be $\Lci{t}_\text{seq}=\min_{h\in{\mtx{H}},\phi_{\text{new}}^t\in{\boldsymbol{\Phi}}_{\text{new}}^t}{\cal{L}}_t(h,\phi)$ s.t.~$\phi=\phi_{\text{new}}^t+\phi_{\text{frz}}^{\star,t}$. The sequential representation mismatch at task $t$ is defined as
\begin{align}
\MS{t}=\Lci{t}_\text{seq}-\Lci{t}.\label{mst}
\end{align}
\end{definition}
This definition is the sequential counterpart of Definitions \ref{def pop} and \ref{def MM}. It quantifies the cost of continual learning with respect to choosing the best (oracle) representation. It also aims to capture the properties of the task distributions thus it uses infinite samples for tasks $1\leq \tau\leq t$. In practice, a new task $t$ is learned on top of finite sample tasks. We need to make a plausible assumption to formalize
\begin{quote}
\emph{With enough samples, representations learned from finite sample tasks are almost as useful as representations learned from infinite sample tasks.}
\end{quote}
We accomplish this by introducing empirical/population compatibility below. The basic idea is that, quality of the representation should decay gracefully as we move from infinite to finite samples.
\begin{assumption} [Empirical/population compatibility] \label{fin comp} For task $t$, define the population risk ${\cal{L}}_t(h,\phi)=\operatorname{\mathbb{E}}[{\widehat{\cal{L}}}_{\mathcal{S}_t}(h,\phi)]$. Recall the definitions of $\phi_{\text{frz}},(h^{\star,t},\phi_{\text{new}}^{\star,t})_{t\geq 1}$ from Def.~\ref{def pop seq}. Given a sequence of incremental feature-extractors ${\phi}=(\phi_{\text{new}}^\tau)_{t\geq 1}$, recall from \eqref{crlseq} that task $\tau-1$ uses the extractor $\phi_{\text{frz}}^{\tau}=\phi_{\text{frz}}+\sum_{i=1}^{\tau-1}\phi_{\text{new}}^{i}$. \emph{To quantify representation quality}, we introduce the risks ${\bar{\cal{L}}}^t_\text{seq}:={\bar{\cal{L}}}^{t,{\phi}}_\text{seq},{\cal{L}}^t_\text{seq}:={\cal{L}}^{t,{\phi}}_\text{seq}$ induced by ${\phi}$ (similar to Def.~\ref{def pop seq})
\begin{align}
&{\bar{\cal{L}}}^t_\text{seq}=\min_{h\in{\mtx{H}}}{\cal{L}}_t(h,\phi_{\text{frz}}^{t+1})\nonumber\\
&{\cal{L}}^t_\text{seq}=\min_{h\in{\mtx{H}},\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}^t}{\cal{L}}_t(h,\phi)~\text{s.t.}~\phi=\phi_{\text{new}}+\phi_{\text{frz}}^t.\label{ltseq}
\end{align}
Here ${\bar{\cal{L}}}^t_\text{seq}$ uses the given $\phi_{\text{new}}^t$ (within ${\phi}$) whereas ${\cal{L}}^t_\text{seq}$ chooses the optimal $\phi_{\text{new}}^t\in {\boldsymbol{\Phi}}_{\text{new}}^t$\footnote{${\cal{L}}^t_\text{seq}$ definition is needed to quantify the representation quality of a new task where incremental update $\phi_{\text{new}}^t$ has not been built yet. In contrast, ${\bar{\cal{L}}}^t_\text{seq}$ quantifies the representation quality of previous tasks for which (full) feature extractors are known.}. Thus ${\bar{\cal{L}}}^t_\text{seq}\geq {\cal{L}}^t_\text{seq}$. Based on these, define the mismatch between empirical and population-level optimizations
\begin{align}\nonumber
\ME{t}={\cal{L}}^t_\text{seq}-\Lci{t}_\text{seq}\quad\text{and}\quad\MA{t}={\bar{\cal{L}}}^t_\text{seq}-\Lci{t}_\text{seq}.
\end{align}
Again by construction {$\MA{t}\geq \ME{t}$}. We say \textbf{empirical and population representations} are compatible if there exists a constant $\eps_0>0$ such that, for all choices of $(\phi_{\text{new}}^\tau)_{\tau=1}^t\in {\boldsymbol{\Phi}}_{\text{new}}^1\times \dots{\boldsymbol{\Phi}}_{\text{new}}^t$, we have that
\
}%\textcolor{blue}{\underbrace{\ME{t}}_{\text{subopt on new task}}\leq\underbrace{\eps_0}_{\text{additive mismatch}}+\underbrace{\frac{1}{t-1} \sum_{\tau=1}^{t-1} \MA{\tau}}_{\text{avg subopt on first $t-1$ tasks}}}
\]
\end{assumption}
\textbf{Interpretation:} Here, }%\textcolor{blue}{$\frac{1}{t-1} \sum_{\tau=1}^{t-1} \MA{\tau}$} quantifies the suboptimality of the representations used by first $t-1$ tasks. Recall that task $\tau$ uses representation $\phi_{\text{frz}}^\tau$ for $\tau\leq t-1$. Verbally, this assumption guarantees that, task $t$ can find an (incremental) representation $\phi_{\text{new}}^t\in{\boldsymbol{\Phi}}_{\text{new}}^t$ and classifier $h\in{\mtx{H}}$ such that its suboptimality to population-optimal risk $\Lci{t}_\text{seq}$ is upper bounded in terms of the average of the suboptimalities over the first $t-1$ tasks. Note that, this can also be viewed as a \textbf{sequential task diversity} condition because we are assuming that good quality representations on the first $t-1$ tasks (w.r.t.~population minima) ensure a small excess risk (w.r.t.~population minima) on the new task.
Following this assumption, theorem below is our main guarantee on sequential CRL. It is obtained by stitching $T$ applications of Lemma \ref{lem seq} which adds a single new task. The statement of Lemma \ref{lem seq} is provided within the proof of Theorem \ref{seq thm} below.
\begin{theorem}\label{seq thm} Suppose we solve the sequential continual learning problem \eqref{crlseq} for each task $1\leq t\leq T$ to obtain hypothesis $(h^t,\phi_{\text{new}}^t)_{t=1}^T$. The $t$'th model uses the prediction $h^t\circ \phi_{\text{frz}}^{t+1}$ where $\phi_{\text{frz}}^t=\phi_{\text{frz}}+\sum_{\tau=1}^{t-1}\phi_{\text{new}}^t$. Consider the same core setting as in Theorem \ref{cl thm}: Namely, we assume Lipschitz hypothesis sets ${\boldsymbol{\Phi}}_{\text{new}}^t,{\mtx{H}}$, Lipschitz loss function $\ell:\mathbb{R}\times \mathbb{R}\rightarrow[0,1]$ and bounded input feature set $\mathcal{X}$ all with respect to Euclidean distance. Recall that $\Lci{t}$ is the optimal risk for task $t$. Suppose the complexity of each ${\boldsymbol{\Phi}}_{\text{new}}^t$ is upper bounded by $\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}$ as in \eqref{ccn decay} for $t\in [T]$. For some absolute constant $C>0$, with probability $1-2Te^{-\tau}$, the solutions $(h^t,\phi_{\text{new}}^t)_{t=1}^T$ of \eqref{crlseq} satisfies the following cumulative generalization bound (summed over all $T$ tasks)
\begin{align
\underbrace{\sum_{t=1}^T\left({\cal{L}}_t(h^t,\phi^t)-\Lci{t}\right)}_{\text{excess test risk w.r.t.~oracle}}&\leq \underbrace{\sum_{t=1}^T\MS{t}}_{\text{sequential representation mismatch}}+\underbrace{T^2(\eps_0+C\sqrt{\frac{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}{N}})}_{\text{cost of finite sample learning}}.\label{gen bound 3}
\end{align}
In light of Def.~\ref{def pop seq}, we can write the suboptimality with respect to solving sequential problems with $N=\infty$ as
\begin{align
\underbrace{\sum_{t=1}^T\left({\cal{L}}_t(h^t,\phi^t)-\Lci{t}_\text{seq}\right)}_{\text{excess test risk w.r.t.~sequential learning}}&\leq \underbrace{T^2(\eps_0+C\sqrt{\frac{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}{N}})}_{\text{cost of finite sample learning}}.\label{gen bound 4}
\end{align}
\end{theorem}
\noindent \textbf{Discussion.} Before providing the proof in the next section, a few remarks are in place. First, we state the sum of test errors rather than the average. Secondly, observe that \eqref{gen bound 3} compares the test errors to the optimal possible errors $\Lci{t}$. On the right hand side there are two terms: ``representation mismatch'' and ``cost of finite sample learning''.
``Representation mismatch'' quantifies the population-level error that arises even if each task had access to infinite samples. This is because, even if each task solved \eqref{crlseq} perfectly, the resulting sequence of representations does not have to be optimal for the next task and $\MS{t}$ precisely captures this suboptimality. Recall that this population-level gap arises from Definition \ref{def pop seq}. \red{This also emphasizes that we should train diverse tasks firstly, since diverse features learned from previous tasks help reduce $\MS{t}$ due to its highly relevant representation. }
The ``cost of finite sample learning'' term originally captures the finite sample effects, and it is proportional to the statistical error rate of solving \eqref{crlseq} for the first task-only i.e.~$\sqrt{\frac{\mathcal{C}({\mtx{H}})+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}}{N}}$. Here, recall from \eqref{ccn decay} that $\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}$ is an upper bound to the complexities of ${\boldsymbol{\Phi}}_{\text{new}}^1,\dots,{\boldsymbol{\Phi}}_{\text{new}}^T$. Perhaps unexpected dependence is the quadratic growth in $T$. This is in contrast to linear growth one would get from adding tasks simultaneously as in Theorem \ref{cl thm}. This quadratic growth arises from the accumulation of the finite-sample representation suboptimalities as we add more tasks. Specifically, Assumption \ref{fin comp} helps guarantee that feature-extractors of tasks $1,\dots,t-1$ are useful for task $t$; however, as more tasks are added they incur more divergence from $(\phi^{\star,t})_{t=1}^T$. Each new task has a finite sample and contributes to this divergence and our analysis leads to $\order{T^2}$ upper bound on the error. $\eps_0$ is an additional mismatch term that makes Assumption \ref{fin comp} significantly more flexible (albeit ideally, it is close to zero). Finally, it would be interesting to explore the tightness of these bounds for concrete analytical settings (e.g.~\ref{crlseq} with linear models or neural nets), running more experiments and further studying the role of finite sample effects \& representation divergences.
\subsection{Proof of Theorem \ref{seq thm}}
\begin{proof} Applying Lemma \ref{lem seq} for each task $1\leq t\leq T$ and union bounding, }%\textcolor{blue}{with probability $1-2Te^{-\tau}$}, for all $T$ applications of \eqref{crlseq}, we have that
\begin{align}
&{\cal{L}}(h^t,\phi_{\text{new}}^t+\phi_{\text{frz}}^t)-\Lci{t}\leq \MS{t}+\underbrace{\ME{t}+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}}{N}}}_{\text{excess empirical error}}\label{gen bound}\\
&\MA{t}\leq \sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}}{N}}+\ME{t}.\nonumber
\end{align}
}%\textcolor{blue}{We simply need to control $\ME{t}$ at time $t$. Following Assumption \ref{fin comp}, observe that at $t=1$ we simply have $\ME{1}=0$. For $\ME{t}$ we have that
\begin{align}
&\ME{t}\leq \eps_0+\frac{1}{t-1} \sum_{\tau=1}^{t-1} \MA{\tau}\nonumber\\
&\implies \MA{t}\leq \left[\eps_0+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}}{N}}\right]+\frac{\sum_{\tau=1}^{t-1}\MA{\tau}}{t-1}.\label{required eq}
\end{align}
Declare $B:=\eps_0+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}}{N}}$. We will inductively prove that for all $t$
\begin{align}
\MA{t}\leq (t-1) B.\label{induct}
\end{align}
For $t=1$, obviously it works. Suppose claim holds until time $t-1$. Using \eqref{required eq} this implies that
\[
\MA{t}\leq B+\frac{\sum_{\tau=1}^{t-1}\MA{\tau}}{t-1}\leq B+ \frac{\sum_{\tau=1}^{t-1}(\tau-1)}{t-1}B=B+\frac{t-2}{2}B=\frac{Bt}{2}\leq B(t-1).
\]
The last inequality holds since $t\geq 2$. Thus the claim holds for $t$ as well.
}Q
To proceed, using the fact that $\ME{t}\leq \MA{t}$ and using the upper bound \eqref{induct}, the excess error in \eqref{gen bound} is upper bounded by $tB=(t-1)B+B$ to obtain
\begin{align}
{\cal{L}}(h^t,\phi_{\text{new}}^t+\phi_{\text{frz}}^t)-\Lci{t}\leq \MS{t}+t (\eps_0+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\mathcal{C}^{{\boldsymbol{\Phi}}}_{\text{new}}+\tau}}{N}}).\label{bound fin gen}
\end{align}
To conclude, we simply sum up \eqref{bound fin gen} for $1\leq t\leq T$ to obtain the advertised bound where the total excess finite-sample error grows as $\frac{T(T+1)}{2}B\leq T^2B$. This yields \eqref{gen bound 3}. \eqref{gen bound 4} is identical to \eqref{gen bound 3} via \eqref{mst}.
\end{proof}
Following Def.~\ref{def pop seq} and Assumption \ref{fin comp}, the lemma below probabilistically quantifies the generalization risk when we add one task. Using this lemma, we will state our main result which quantifies the generalization risk when adding $T$ tasks.
\begin{lemma}\label{lem seq} Suppose we are given the output pairs $(h^\tau,\phi_{\text{new}}^\tau)_{\tau=1}^{t-1}$ of the first $t-1$ applications of sequential CRL problem \eqref{crlseq}. Now, we solve for the $t$'th solution denoted by the pair $h^t,\phi_{\text{new}}^t$. Under same conditions as in Theorem \ref{cl thm} (Lipschitz hypothesis ${\boldsymbol{\Phi}}_{\text{new}}^t,{\mtx{H}}$, Lipschitz loss $\ell:\mathbb{R}\times \mathbb{R}\rightarrow [0,1]$ and bounded input features $\mathcal{X}$), for some absolute constant $C>0$, with probability $1-2e^{-\tau}$, the solution $(h^t,\phi_{\text{new}}^t)$ of \eqref{crlseq} satisfies the following two properties
\begin{align}\label{eq mm}
&{\cal{L}}(h^t,\phi_{\text{new}}^t+\phi_{\text{frz}}^t)-\Lci{t}\leq \MS{t}+\ME{t}+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}}\\
&\MA{t}\leq \sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}}+\ME{t}.\nonumber
\end{align}
Here, $\MS{t},\ME{t},\MA{t}$ are mismatch definitions introduced in Def.~\ref{def pop seq} and Assumption \ref{fin comp}.
\end{lemma}
\begin{proof}
As introduced in \eqref{ltseq}, the representation quality of the new task will be captured by
\[
{\cal{L}}^t_\text{seq}=\min_{h\in{\mtx{H}},\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}^t}{\cal{L}}_t(h,\phi)~\text{s.t.}~\phi=\phi_{\text{new}}+\phi_{\text{frz}}^t.
\]
Similarly, recall the empirical mismatch $\ME{t}={\cal{L}}^t_\text{seq}-\Lci{t}_\text{seq}$.
Based on this and recalling $\MS{t}$ definition \eqref{mst}, applying Theorem \ref{cl thm} with $T=1$, we obtain that, with probability $1-2^{-\tau}$ we have the following bounds (on the same event)
\begin{align}
{\cal{L}}(h^t,\phi_{\text{new}}^t+\phi_{\text{frz}}^t)&\leq{\cal{L}}^t_\text{seq}+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}}\nonumber\\
{\cal{L}}(h^t,\phi_{\text{new}}^t+\phi_{\text{frz}}^t)&\leq \Lci{t}+\MS{t}+\ME{t}+\sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}}.\nonumber
\end{align}
The second equation establishes \eqref{eq mm}.
The remaining challenge is simply relating the population and empirical mismatches i.e.~controlling $\ME{t}$. Recall $h^t$ is the ERM solution of \eqref{crlseq} associated with $\phi_{\text{new}}^t$. Using the uniform concentration event (implied within the application of Theorem \ref{cl thm} via \eqref{unif conv}), we have
\[
|{\widehat{\cal{L}}}_{\mathcal{S}_t}(h,\phi_{\text{new}}+\phi_{\text{frz}}^t)-{\cal{L}}_t(h,\phi_{\text{new}}+\phi_{\text{frz}}^t)|\leq \sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}},
\]
Now, using the optimality of $h^t$ for $\phi_{\text{new}}^t$ and the fact that $(h^t,\phi_{\text{new}}^t)$ minimizes the empirical risk, observe that
\begin{align*}
\MA{t}&=\min_{h\in{\mtx{H}}}{\cal{L}}(h,\phi_{\text{frz}}^{t+1})-\Lci{t}_\text{seq}\\
&\leq [\min_{h\in{\mtx{H}}}{\cal{L}}(h,\phi_{\text{frz}}^{t+1})-{\cal{L}}^t_\text{seq}]+\ME{t}\\
&\leq [{\cal{L}}(h^t,\phi_{\text{frz}}^{t+1})-{\cal{L}}^t_\text{seq}]+\ME{t}\\
&\leq \sqrt{\frac{\ordet{\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}^t}+\tau}}{N}}+\ME{t},
\end{align*}
to conclude with the second line of \eqref{eq mm}.
\end{proof}
\section{Introduction}
\emph{Continual learning} (CL) or lifelong learning aims to build a model for a non-stationary and never-ending sequence of tasks, without access to previous or future data \cite{thrun1998lifelong,chen2018lifelong,parisi2019continual}. The main challenge in CL is that a standard learning procedure usually results in performance reduction on previously trained tasks if their data are not available. The phenomenon is termed as \emph{catastrophic forgetting} \cite{mccloskey1989catastrophic,kirkpatrick2017overcoming,pfulb2019comprehensive}.
Importance of continual learning in real-life inference and decision making scenarios led to a rich set of CL techniques~\cite{delange2021continual,van2019three}\ylm{methods including replay-based, regularization-based, and architecture-based strategies}. However, the theoretical principles of continual learning is relatively less understood and the progress is under-whelming compared to the algorithmic advances despite recent progress {(see \cite{bennani2020generalisation,doan2021theoretical,lee2021continual})}. In this work, for the first time, we investigate the problem of \emph{Continual Representation Learning (CRL)} to answer
\begin{center}
\fbox{\centering\begin{minipage}{0.84\textwidth}
\centering\textit{
What are the statistical benefits of previous feature representations for learning a new task? Can we build an insightful theory explaining empirical performance?}
\end{minipage}}
\end{center}
Our key contribution is addressing both questions affirmatively. We develop our algorithms and theory for architecture-based zero-forgetting CL which includes PackNet~\cite{mallya2018packnet}, CPG~\cite{hung2019compacting}, and RMN~\cite{wu2020understanding}. These methods eliminate forgetting by training a sub-network for each task and freezing the trained parameters. At a high level, all of these methods inherently have the potential of continual representation learning by allowing new tasks to reuse the frozen feature representations built for earlier tasks. However, quantifying the empirical benefits/performance of these representations and building the associated theory have been elusive. We overcome this via innovations in experiment design, theory, and algorithms:
\noindent$\bullet$ \textbf{Theoretical and empirical benefits of continual representations.} We establish theoretical results and sample complexity bounds for CRL by using tools from empirical process theory. Within our model, a new task uses frozen feature map $\phi_{\text{frz}}$ of previous tasks and learns an additional task-specific representation $\phi_{\text{new}}$. For PackNet \cite{mallya2018packnet}, $\phi_{\text{frz}}$ corresponds to all the nonzero weights so far and $\phi_{\text{new}}$ is the nonzeros allocated to the new task, thus $\phi_{\text{frz}}$ requires a lot more data to learn. Our theory (see Section~\ref{sec:crl_theory}) explains (1) how the new task can reuse the frozen feature map to greatly reduce the sample size and (2) how to quantify the representational compatibility between old and new tasks. Specifically, we fully avoid the statistical cost of learning $\phi_{\text{frz}}$ and replace it with a \emph{``representational mismatch''} term between the new task and frozen features. }%\textcolor{blue}{We also extend our results from a single task to learning a sequence of tasks by quantifying the aggregate impact of finite data on representation $\phi_{\text{frz}}$ which evolves as new tasks arrive (deferred to appendix)}.
An important conclusion is that ideally frozen representation $\phi_{\text{frz}}$ should contain \textbf{diverse and high-quality features} so that it has small \emph{mismatch} with new tasks. This is consistent with the results on transfer learning \cite{tripuraneni2020theory} and motivates the following CL principle supported by our experiments:
\begin{center}
\hspace{-10pt}\fbox{\begin{minipage}{0.84\linewidth}
\centering\textit{
First learn diverse and large-data tasks so that their representations help upcoming tasks.}
\end{minipage}}
\end{center}
Indeed, we show in Fig.~\ref{fig:LS} that a new task with small data achieves significantly higher accuracy if it is added later in the task sequence (so that $\phi_{\text{frz}}$ becomes more diverse). Then, we show in Fig.~\ref{fig:diversity} that choosing the first task to be diverse (such as ImageNet) helps all the downstream tasks. Finally, we show in Fig.~\ref{fig:Sample} that it is better to first learn tasks with large sample sizes to ensure high-quality features.
\red{Our results on the importance of task order also relate to curriculum learning \cite{bengio2009curriculum} where the agent gets to choose the order tasks are learned. However, instead of curriculum learning, which aims to learn one task from easy to hard, our work is based on a more general continual learning setup. Our conclusion on task diversity provides a new perspective on the learning order of curriculum learning that we should first learn more representative tasks instead of an easier task. }
\noindent$\bullet$ \textbf{Algorithms for inference-efficient CRL.} Zero-forgetting CL methods often need a large neural network (dubbed as supernetwork) to load numerous tasks into subnetworks, or dynamically expand the model to avoid forgetting \cite{delange2021continual}. Thus, CL/CRL may incur a large computational cost during inference. This leads us to ask whether one can retain the accuracy benefits of CRL while ensuring that each new task utilizes an \textbf{inference-efficient} sub-network, thus achieving the best of both worlds. We quantify inference-efficiency via floating point operations (FLOPs) required to compute the task subnetwork. To this end, we propose Efficient Sparse PackNet (ESPN) algorithm (Fig.~\ref{overview fig}). In a nutshell, ESPN guarantees inference-efficiency by incorporating a channel-pruning stage within PackNet-style approaches. Via extensive evaluations, we find that, ESPN incurs minimal loss of accuracy while greatly reducing FLOPs (as much as 80\% in our SplitCIFAR-100 experiments, Table \ref{CIFARtable}).
{In summary, this work makes key contributions to continual representation learning from empirical, theoretical, and algorithmic perspectives. In the remainder of the paper, we discuss related work, detail our empirical and theoretical findings on CRL, and present the ESPN algorithm and its evaluations.}
\section{Efficient Sparse PackNet (ESPN) Algorithm}\label{sec: espnalgo}
We will use PackNet and ESPN throughout the paper to study continual representation learning. Thus, we first introduce the high-level idea of our ESPN algorithm which augments PackNet. Suppose we have a single model referred to as supernetwork and a sequence of tasks. Our goal is to train and find optimal sparse sub-networks within the supernet that satisfy both FLOPs and sparsity restrictions without any performance reduction or forgetting of earlier tasks. Figure~\ref{overview fig} illustrates our proposed algorithm. We propose a joint channel and weight pruning strategy in which FLOPs constraint (in channel pruning) is important for efficient inference and sparsity constraint (in weight pruning) is important for \emph{packing} all tasks into the network even for a large number of tasks $T$. In essence, ESPN equips PackNet-type methods with inference-efficiency using an innovative FLOP-aware channel pruning strategy. }%\textcolor{blue}{Section \ref{sec: espn_detail} and appendix provide implementation details and inference-time evaluations on ESPN.
\noindent{\textbf{Notation.}} ESPN admits FLOPs constraint as an input parameter, which is a critical aspect of the experimental evaluations. Let MAX\_FLOPs be the FLOPs required for one forward propagation through the dense supernetwork. In our evaluations, for $\gamma\in (0,1]$, we use \textbf{ESPN-$\gamma$} to denote our CL Algorithm \ref{algo 1} in which each task obeys the FLOPs constraint $\gamma\times \text{MAX\_FLOPs}$. Similarly, \textbf{Individial-$\gamma$} will be the baseline that each task is trained individually (from scratch) on the full supernetwork while using at most $\gamma\times \text{MAX\_FLOPs}$. \textbf{Individual} is same as \textbf{Individual}-1 where we train the whole network without pruning.
\section{Conclusion and Discussion}
To summarize, our work elucidates the benefit of continual representation learning theoretically and empirically and sheds light on the role of task ordering, diversity and sample size. We also propose a new CL algorithm to learn good representations subject to inference considerations. Extensive experimental evaluations on the proposed Efficient Sparse PackNet demonstrate its ability to achieve good accuracy as well as fast inference.
\noindent \textbf{Limitations and future directions.}
Although we highlight the importance of task order in CRL, the task sequence is not always under our control. It would be desirable to develop adaptive learning schemes that can better identify an exploit diverse tasks and discover semantic connections across task pairs even for a predetermined task sequence.
Another potential direction is to develop similar inference-efficient continual learning schemes for other architectures by appropriately adapting our joint weight and channel pruning strategy. An example is transformer-based models where computation and memory efficiency is particularly critical.
\section*{Acknowledgements}\vspace{-10pt}
This work is supported in part by the National Science Foundation under grants CCF-2046816, CCF-2046293, CNS-1932254, and by the Army Research Office under grant W911NF-21-1-0312.
\section{Inference-efficient continual representations via Efficient Sparse PackNet (ESPN)} \label{sec: espn_detail}
In this section, we provide an in depth discussion of our ESPN algorithm and evaluate its benefit in inference efficiency.
ESPN (Algorithm~\ref{algo 1}) learns the new sub-network in three phases: pre-training (over all free weights), gradual pruning (to prune weights and channels), and fine-tuning (to refine the new allocated weights). While not shown in Algorithm~\ref{algo 1}, we also introduce the following innovations.
\input{sec/algo_oc}
\noindent$\bullet$ \textbf{FLOP-aware pruning.}
We propose a FLOP-aware pruning strategy that provides up to 80\% FLOPs reduction in our experiments. In practice, Algorithm~\ref{algo 1} minimizes a regularized objective ${\cal{L}}_{\mathcal{S}_t}({\boldsymbol{\theta}}\odot\vct{m})+\mathcal{R}({\boldsymbol{\Gamma}},\vct{m})$, with
\begin{align}
\label{eq:reg}\mathcal{R}({\boldsymbol{\Gamma}},\vct{m})&=\sum_{l\in [L]}\lambda_l\|{\boldsymbol{\Gamma}}_l\|_1,~~\lambda_l=g(\text{FLOP}_l(\vct{m})).
\end{align}
${\boldsymbol{\Gamma}}_l$ denotes the channel scores of $l^{\text{th}}$ layer, $L$ is the number of layers, and $g(\cdot)$ is a monotonically increasing function. In this manner, channels costing more FLOPs are assigned larger $\lambda_l$, pushed towards zero, and subsequently pruned. Additionally, $\text{FLOP}_l(\cdot)$ computes FLOPs of $l^{\text{th}}$ layer. Since we use gradual pruning (Line~\ref{algo:pruning} in Algorithm~\ref{algo 1}), FLOPs are changed over time and $\lambda_l$ is automatically tuned. In our experiments, we use $g(x)=C\sqrt{x}$ for a proper scaling choice $C>0$. }%\textcolor{blue}{The evaluation of our FLOP-aware channel pruning algorithm is presented in \ylm{Section~\ref{pruning sec} of} appendix.}
\som{discuss $\alpha$ better}
\noindent$\bullet$ \textbf{Weight allocation.}
{Inspired by our theoretical finding that CL improves data efficiency,} we introduce a simple weight allocation scheme to assign free weights for new task. Let $p$ be the total number of weights and $p_{t-1}$ be the number of used weights after training tasks $1$ to $t-1$. Then we assign $\lceil(p-p_{t-1})\cdot\alpha\rceil$ free weights to task $t$, where $0<\alpha<1$ is {the \emph{weight-allocation} level}. We emphasize that weight allocation controls the number of new nonzeros allocated to a task. A new task is allowed to use all of the (frozen) nonzeros that are allocated to the previous tasks (as long as FLOPs constraint is not violated). }%\textcolor{blue}{We evaluate the empirical performance of our weight allocation technique \ylm{in Section~\ref{sec:wa}} in appendix.}
\subsection{Experimental Setup}\label{sec:setting}
We evaluate the performance of our proposed ESPN in terms of accuracy and efficiency metrics on three datasets: SplitCIFAR100, RotatedMNIST, and PermutedMNIST. We compare ESPN to numerous baselines, which include training each task individually, multitask learning (MTL), PackNet~\cite{mallya2018packnet}, CPG~\cite{hung2019compacting}, RMN~\cite{kaushik2021understanding}, and SupSup~\cite{wortsman2020supermasks}. The last four methods are zero-forgetting CL methods. \som{Emphasize SUPSUP is learning full-dimensional parameter mask}
\som{Emphasize no one is efficient}
\som{Empirically verifying our FLOP constraint vs theirs}
\noindent{\textbf{Datasets. }} SplitCIFAR100, RotatedMNIST and PermutedMNIST are popular datasets for continual learning that we also use in our experiments. We follow the same setting as in \cite{wortsman2020supermasks}. For SplitCIFAR100 dataset, we randomly split CIFAR100 \cite{krizhevsky2009learning} into 20 tasks where each task contains $5$ classes, $2500$ training samples, and $500$ test samples. RotatedMNIST is generated by rotating all images in MNIST by the same degree. In our experiments, we generate $36$ tasks with $10,20,\ldots, 360$ degree rotations { and train in a random order}. PermutedMNIST dataset is created by applying a fixed pixel permutation to all images, and we created $36$ tasks with independent random permutations.
\input{sec/table_oc}
\noindent{\textbf{Models and implementation. }} For SplitCIFAR100, {following \cite{lopez2017gradient} we use a variation of ResNet18 model with fewer channels.} For each task, we use a batch size of 128 and Adam optimizer \cite{kingma2014adam} with hyperparameters $(\beta_1,\beta_2)=(0,0.999)$. As shown in Algorithm~\ref{algo 1}, for each task we apply pretraining, pruning, and finetuning strategies. First, we pretrain the model for $60$ epochs with learning rate $0.01$. Then, we gradually prune the channels and weights within $90$ epochs using the same learning rate. For the finetuning stage, we apply cosine decay \cite{loshchilov2016sgdr} over learning rate, starting from $0.01$, and train for $100$ epochs. Therefore, each task trains for $250$ epochs in total.
For RotatedMNIST and PermutedMNIST experiments, we use the same setting }%\textcolor{blue}{of FC1024 model} in \cite{wortsman2020supermasks}. This is a fully connected network with two hidden layers of size $1024$. We train for $10$ epochs with RMSprop optimizer \cite{graves2013generating}, batch size of $256$, and learning rate $0.001$.
The number of pretraining, pruning, and finetuning epochs are $3$, $4$, and $3$, respectively.
As for comparison baselines, MTL and individual training baselines follow the same configurations (e.g.~architecture, hyperparameters) as ESPN. In SplitCIFAR100 experiments, labels are not shared among different tasks. Thus, each task is assigned a separate classifier head while sharing the same backend supernet as a feature extractor. {Unlike SplitCIFAR100, all the tasks in RotatedMNIST and PermutedMNIST experiments share the same head via weight pruning.} MTL simultaneously trains all 20 tasks while sharing the backend supernet. In individual training, each task trains their own backend. Finally, CPG, RMN, and SupSup are all trained with their own publicly available codes but over the same ResNet18 model. To ensure fair comparison, we do not allow dynamic model expansion/enlargement in CPG. Similarly, for PackNet, instead of using its original setting with fixed pruning ratio for each task, we run it with our \emph{weight allocation} strategy to make it easier to compare. These baselines are all evaluated in Table~\ref{CIFARtable} which is discussed below.
\subsection{Investigation of Inference Efficiency}
\noindent{\textbf{SplitCIFAR100.}} Table~\ref{CIFARtable} presents results for our experiments on SplitCIFAR100. We use our FLOP-aware pruning technique to prune the channels and apply weight allocation with parameter $\alpha=0.1$.
}%\textcolor{blue}{Since the impact of classifier head is rather negligible as it contains $<0.1\%$ weights and $<0.01\%$ FLOPs of the backend, we evaluate the FLOPs for each task inside backend.} To compare the performance of different methods, we use the same random seed to generate $20$ tasks so that task sequence is exactly the same over all experiments. We conduct ESPN experiments with $20\%$, $50\%$}%\textcolor{blue}{, and $100\%$} of FLOPs constraints corresponding to ESPN-0.2, ESPN-0.5}%\textcolor{blue}{, and ESPN-1} in Table~\ref{CIFARtable}.
The results of Individual-0.2/-1 show that our FLOP-aware pruning technique can effectively reduce the computation requirements while maintaining the same level of model accuracy. Moreover, PackNet performing better than both CPG and RMN shows the benefit of our weight allocation technique. }%\textcolor{blue}{We remark that CPG method performs pruning and fine-tuning multiple times until it finds the optimal pruning ratio, which costs significantly more time in training compared to PackNet{/ESPN}.} ESPN-0.2/-0.5 results show that our continual learning algorithm outperforms both baselines in accuracy despite up to 80\% FLOP reduction. \ylm{Our accuracy improvement over PackNet arises from the trainable task-specific BatchNorm weights described under Algorithm \ref{algo 1}.}
\ylm{In practice, for all CPG, RMN, PackNet, and SupSup methods, task-specific running mean and running variance inside BatchNorm layers are essential to reconstruct the same performance. Therefore, despite our method applies additional BatchNorm weights for each task, since we prune BatchNorm layers and less replay memory is needed overall compared to the other methods (except ESPN-1).}
\noindent\textbf{RotatedMNIST and PermutedMNIST.} Figure~\ref{fig:mnist} presents the RotatedMNIST and PermutedMNIST results. We run experiments on ESPN, PackNet, and individual training and report the average accuracy over $5$ trials. Note that, for fully-connected layers, we prune neurons to reduce FLOPs (rather than channels of CNN). In our experiments, we assigned neurons with a pruning parameter ${\boldsymbol{\Gamma}}$. We use our FLOP-aware pruning technique to prune neurons based on ${\boldsymbol{\Gamma}}$ and use weight allocation with parameter $\alpha=0.05$. \ylm{${\boldsymbol{\Gamma}}$ is released after pruning.} Similar to channel-sharing in Figure \ref{overview fig}, different tasks are allowed to share the same neuron. Unlike SplitCIFAR100 experiments, all the tasks in both MNIST experiments share the same classifier head because they use the same 10 classes }%\textcolor{blue}{and FLOPs evaluation includes this head}. }%\textcolor{blue}{Therefore, \ylm{same as other zero-forgetting methods,} only binary mask is needed to restore performance.}
\input{sec/fig/mnist_oc}
Results of RotatedMNIST and PermutedMNIST experiments are shown in Figure~\ref{fig:mnist}. Blue curves show the results of ESPN-0.2. For fair comparison, PackNet (Orange curves) use the same weight allocation parameter $\alpha=0.05$. The Green and Red dashed lines show the task-averaged accuracy of Individual-0.2/-1 baselines that train each task with separate models for $20\%/100\%$ of FLOPs constraints. }%\textcolor{blue}{In both experiments, the gap between Individual-0.2 (Green) and Individual-1 (Red) curves is negligible and it again shows the benefit of our FLOP-aware pruning technique. While PackNet performs well for the first few tasks, it degrades gradually as more tasks arrive. Thus, when there are more than a few tasks, we see that ESPN algorithm works better than PackNet despite enforcing inference-efficiency and despite using the same weight allocation method. Figure \ref{fig:rotated} shows that ESPN-0.2 on RotatedMNIST exhibits mild accuracy degradation over $36$ tasks.} While in Figure \ref{fig:permuted}, test accuracy decreases more noticeably as the tasks are added. A plausible explanation is that, because each PermutedMNIST task corresponds to a random permutation, the tasks are totally unrelated. Thus, knowledge gained from earlier tasks are not useful for training new tasks and CRL does not really help in this case. In contrast, since tasks in RotatedMNIST are semantically relevant, ESPN and PackNet both achieve higher accuracy thanks to representation reuse across tasks. Specifically, the significant accuracy gap between the Blue curves in Figures \ref{fig:rotated} and \ref{fig:permuted} (especially for larger Task IDs) demonstrate the clear benefit of CRL.
\section{Efficient Sparse PackNet (ESPN) Algorithm}\label{appsec: espnalgo}
In this section, we introduce more implementation and evaluation details of our ESPN algorithm. Denote $[p]=\{1,2,\dots,p\}$. We use the phrases \emph{mask} and \emph{sub-network} interchangeably because we obtain the task sub-network by masking weights of the supernet. This sub-network is the nonzero support of the task, that is, the task-specific model is obtained by setting other weights to zero. We assume a sequence of tasks $\{\mathcal{T}_t$, $1\leq t\leq T\}$ are received during training time, where $t$ is task identifier and $T$ is the number of tasks. Let $\mathcal{S}_t=\{(\vct{x}_{ti},y_{ti})\}_{i=1}^{N_t}$ be labeled training pairs of $\mathcal{T}_t$, which consists of $N_t$ samples.
Given task sequence and a single model, our goal is to find optimal sparse sub-networks that satisfy both FLOPs and sparsity restrictions without performance reduction and knowledge forgetting. FLOPs constraint is important for efficient inference whereas sparsity constraint is important for adding all tasks into the network even for large number of tasks $T$. {To this end, we use joint channel and weight pruning strategy.} Let $\ell$ be a loss function, $f$ be a hypothesis (prediction function) and ${\boldsymbol{\theta}}\in\mathbb{R}^p$ denote the weights of $f$. We focus on task $\mathcal{T}_t$, assuming $\mathcal{T}_1,...,\mathcal{T}_{t-1}$ are already trained. Let $\vct{m}_t\subset[p]$ be the nonzero support of task $t$ and $\vct{m}^{\text{all}}_t=\cup_{\tau=1}^t \vct{m}_\tau$ be the combined support until task $t$. Initially $\vct{m}^{\text{all}}_0=\emptyset$. Let ${\boldsymbol{\theta}}_t\in\mathbb{R}^p$ be the model weights at time $t$. Note that all the trained weights of the tasks lie on the sub-network $\vct{m}^{\text{all}}_t$. We use the notation ${\boldsymbol{\theta}}\odot\vct{m}$ to set the weights of ${\boldsymbol{\theta}}$ outside of the mask $\vct{m}$ to zero.
Define the loss ${\cal{L}}_{\mathcal{S}_t}({\boldsymbol{\theta}})=\frac{1}{N_t}\sum_{i=1}^{N_t}\ell(f(\vct{x}_{ti};{\boldsymbol{\theta}}),y_{ti})$. The procedure for learning task $\mathcal{T}_t$ is formulated as the following optimization task that updates {supernet} weight/mask pair and {returns $({\boldsymbol{\theta}}_t,\vct{m}^{\text{all}}_t:=\vct{m}_t\cup \vct{m}^{\text{all}}_{t-1})$} given $({\boldsymbol{\theta}}_{t-1},\vct{m}^{\text{all}}_{t-1})$:
\begin{align}
{\boldsymbol{\theta}}_t,\vct{m}_t=\arg&\min_{{\boldsymbol{\theta}},\vct{m}}~~~{\cal{L}}_{\mathcal{S}_t}({\boldsymbol{\theta}}\odot\vct{m})\tag{ESPN-OPT}\label{cl_prob}\\
\text{s.t.}
&~~~\text{FLOP}(\vct{m})\leq \overline{\text{FLOP}}_t,\nonumber\\
&~~~\vct{m}_{\text{new}}=\vct{m}\setminus\vct{m}^{\text{all}}_{t-1},\nonumber\\
&~~~\text{NNZ}(\vct{m}_{\text{new}})\leq \overline{\text{NNZ}}_t,\nonumber\\
&~~~{{\boldsymbol{\theta}}\odot\vct{m}^{\text{all}}_{t-1}}={\boldsymbol{\theta}}_{t-1}\odot\vct{m}^{\text{all}}_{t-1}.\nonumber
\end{align}
Here $\vct{m}_t$ is the channel-constrained mask that corresponds to the sub-network of $\mathcal{T}_t$ and ${\boldsymbol{\theta}}_t\odot\vct{m}_t$ are the weights we use for task $\mathcal{T}_t$ and the prediction function is $f(\cdot;{\boldsymbol{\theta}}_t\odot\vct{m}_t)$. {The updated mask until task $t$ is obtained by $\vct{m}^{\text{all}}_t=\vct{m}_t\cup \vct{m}^{\text{all}}_{t-1}$.} $\text{FLOP}(\cdot)$ and $\text{NNZ}(\cdot)$ returns the FLOPs and nonzeros of a given mask $\vct{m}$. $\overline{\text{FLOP}}_t$ and $\overline{\text{NNZ}}_t$ are the FLOPs and nonzero constraints of task $t$. Observe that we only enforce NNZ constraint on the new weights $\vct{m}_{\text{new}}$ whereas FLOPs constraint applies on the whole sub-network. {The last equation in \eqref{cl_prob} highlights that the weights of earlier tasks on $\vct{m}^{\text{all}}_{t-1}$ are kept frozen}.
\som{Discuss algo in more detail} In practice, FLOPs \& NNZ constraints in \eqref{cl_prob} lead to a combinatorial problem. We propose Algorithm \ref{algo 1} to (approximately) solve this problem efficiently which learns the new sub-network in three phases: pre-training (over all free weights), gradual pruning (to satisfy constraints and obtain $\vct{m}_{\text{new},t}$), and fine-tuning (to refine the weights on $\vct{m}_{\text{new},t}$). While not shown in Algorithm \ref{algo 1}, we also introduce the following innovations.
\noindent$\bullet$ \textbf{Trainable task-specific BatchNorm.}
We train separate BatchNorm layers for each task. This has multiple synergistic benefits. First, our algorithm trains faster and generalizes better than PackNet which does not train BatchNorm weights (see Table \ref{CIFARtable}). Specifically, training BatchNorm weights allows ESPN to re-purpose the (frozen) weights of the earlier tasks with negligible memory cost. BatchNorm weights also guide our channel pruning scheme described next.
\noindent$\bullet$ \textbf{FLOP-aware pruning.} Many of the prior works on channel pruning \cite{liu2017learning,zhuang2020neuron} focus on reducing the number of channels rather than the computation/FLOPs cost of the channels which varies across layers. As \red{shown in Fig.~\ref{fig:CIFAR10_ratio} that to satisfy FLOPs constraint $\gamma$, numerous channels are pruned}. This results in unsatisfactory performance especially under aggressive FLOPs constraint \red{(shown in Fig.~\ref{fig:CIFAR10_acc})}. However in CL setup, in order to train a single model with many tasks, a significantly larger supernet with high capacity is needed compared to a single task requirement, and we aim to find a sub-network with very few FLOPs without compromising performance. To fit our specific needs for channel pruning, in this paper we present an innovative channel pruning algorithm called {FLOP-aware channel pruning} that preserves the performance up to 80\% FLOPs reduction in our SplitCIFAR100 experiments (Table \ref{CIFARtable}). Following \cite{liu2017learning}, we consider BatchNorm weights as a trainable saliency score for convolutional channels and prune all channels with scores lower than a certain threshold by setting them to zero. Let ${\boldsymbol{\Gamma}}$ be the BatchNorm vectors. Given dataset $\{(\vct{x}_i,y_i)\}_{i=1}^N$, in practice, rather than solving the constrained problem \eqref{cl_prob}, Algorithm~\ref{algo 1} minimizes a regularized objective ${\cal{L}}_{\mathcal{S}_t}({\boldsymbol{\theta}}\odot\vct{m})+\mathcal{R}({\boldsymbol{\Gamma}},\vct{m})$. Here $\mathcal{R}$ is a regularization term to promote channel sparsity in \eqref{app eq:reg}.
For an $L$ layer network, use ${\boldsymbol{\Gamma}}_l$ to denote the $l^{\text{th}}$ {BatchNorm weights} for $l\in[L]$. Intuitively, we wish to use $\ell_1$-regularization on ${\boldsymbol{\Gamma}}$. However, since layers of the network show variation, layerwise regularization parameters $(\lambda_l)_{l=1}^L$ are needed where $L$ is number of layers. Instead of designing $\lambda_l$ by trial-and-error -- which is time-consuming and expert-dependent -- we introduce a method that chooses $\lambda_l$ automatically, fine-tunes $\lambda_l$ during gradual pruning and adapts to the global FLOPs constraint.
Specifically, $\lambda_l$ is chosen based on the {FLOPs load of a channel} determined by its input feature dimensions and operations. Here an implicit goal is pruning as few channels as possible while achieving maximum FLOPs reduction. To achieve these goals, we use the following FLOP-weighted sparsity regularization (as also stated in the main body)
\begin{align}
\label{app eq:reg}\mathcal{R}({\boldsymbol{\Gamma}},\vct{m})&=\sum_{l\in [L]}\lambda_l\|{\boldsymbol{\Gamma}}_l\|_1,~~\lambda_l=g(\text{FLOP}_l(\vct{m}^{(l)})).
\end{align}
Here $\vct{m}^{(l)}$ denotes the restriction of the sub-network $\vct{m}$ to $l^{\text{th}}$ layer, $\text{FLOP}_l(\cdot)$ is the FLOPs load for a channel in the $l^{\text{th}}$ layer of the subnet, and $g(\cdot)$ is a monotonically increasing function. We use $\ell_1$-penalty to enforce unimportant elements to zeros and prune the channels with smallest weights over all layers. Since $g(\cdot)$ is increasing, channels costing more FLOPs are assigned with larger $\lambda_l$ and are pushed towards zero, thus they are easier to prune. Additionally, since $\text{FLOP}_l(\cdot)$ is based on subnet $\vct{m}^{(l)}$, $\lambda_l$ is automatically tuned while we use gradual pruning (Line~\ref{algo:pruning} in Algorithm \ref{algo 1}). In our experiments, we use $g(x)=C\sqrt{x}$ for a proper scaling choice $C>0$. \red{Here $C$ can be seen as a normalized term and in detail we have
\begin{align*}
\lambda_l=\frac{\sqrt{\text{FLOP}_l(\vct{m}^{(l)})}}{\sum_{i=1}^L\sqrt{\text{FLOP}_i(\vct{m}^{(i)})}}.
\end{align*}}
\noindent$\bullet$ \textbf{Weight allocation.} Since we do not modify the supernet architecture, without care, supernet might run out of free weights if there is a huge number of tasks. While original PackNet paper \cite{mallya2018packnet} also uses weight pruning, since they consider relatively fewer tasks, they don't develop an algorithmic strategy for allocating the free weights to new tasks. In our experiments, we introduce a simple weight allocation scheme to assign $\overline{\text{NNZ}}_t$ {depending on the number of remaining free weights}. Let $p$ be the total number of weights, $p_t$ be the total number of weights used by tasks $1$ to $t$ and $p_0=0$. We set
\begin{align}\label{wa-eq}
\overline{\text{NNZ}}_t=\lceil (p-p_{t-1})\cdot\alpha\rceil\quad\text{for some}\quad 0<\alpha<1.
\end{align} Here $\alpha$ is {the \emph{weight-allocation} level} and a new task gets to use $\alpha$ fraction of all unused weights in the supernet. We emphasize that weight allocation controls the number of new nonzeros allocated to a task. A new task is allowed to use all of the (frozen) nonzeros that are allocated to the previous tasks (as long as FLOPs constraint is not violated).
\section{Appendix C: Expanded Related Work}
\section{Expanded Related Work}\label{app:related}
Our contributions are most closely related to the continual learning literature. Our theory and algorithms also connect to representation learning and neural network pruning.
\noindent\textbf{Continual learning.}
A number of methods for continual and lifelong learning have been proposed to tackle the problem of catastrophic forgetting and existing approaches can be broadly categorized into three groups~\cite{delange2021continual}: replay-based~\cite{lopez2017gradient,rebuffi2017icarl,rolnick2018experience,buzzega2021rethinking,aljundi2019gradient}, regularization-based~\cite{kirkpatrick2017overcoming,zenke2017continual,li2017learning}, and parameter isolation methods~\cite{fernando2017pathnet,yoon2017lifelong,mallya2018piggyback,rusu2016progressive}. In our work, we focus on a branch of parameter isolation methods called zero-forgetting CL such as PackNet~\cite{mallya2018packnet}, CPG~\cite{hung2019compacting}, RMN~\cite{kaushik2021understanding}, and SupSup~\cite{wortsman2020supermasks} that completely eliminates forgetting by training a sub-network for each task and freezing the trained parameters (SupSup excepted). However, finding of \cite{frankle2018lottery} shows that a network can reduce by over 90\% of parameters without performance reduction. This inspires our weight-allocation strategy to adapt PackNet to more sparse sub-networks. Unlike PackNet which prunes the network by keeping largest absolute weights in each layer and reuses all the frozen weights, CPG and RMN apply real-valued mask to each fixed entry and prune by keeping the largest values of the mask. SupSup is motivated by \cite{ramanujan2020s,zhou2019deconstructing,malach2020proving} which show that a sufficiently over-parameterized random network contains a sub-network with roughly the same accuracy as the target network without training. In essence, it aims to find masks only over random network and it is adaptable to infinite tasks. However it leads to inefficient inference (due to using the full network) and potentially large memory requirements (as one has to store a mask as large as supernet rather than a subnet).
In CL, in order to load a network with a large number of tasks, often, a large model is needed, which naturally leads to inefficiency during inference-time without proper safeguards. Addressing this challenge appears to be an unexplored avenue as far as we are aware. While in this work, we present Efficient Sparse PackNet (ESPN) that implements zero-forgetting CL and achieves state-of-the-art accuracy with less computational demand.
We also emphasize that there are several interesting works on the theory of continual learning such as \cite{lee2021continual, doan2021theoretical,bennani2020generalisation,mirzadeh2020understanding,yin2020optimization}. These works focus on NTK-based analysis for deep nets, theoretical investigation of orthogonal gradient descent \cite{farajtabar2020orthogonal}, and task-similarity. However, to the best of our knowledge, ours is the first work on the representation learning ability and the associated data-efficiency.
\noindent\textbf{Representation learning theory.} The rise of deep learning motivated a growing interest in theoretical principles behind representation learning. Similar in spirit to this project, \cite{maurer2016benefit} provides generalization bounds for representation-based transfer learning in terms of the Rademacher complexities associated with the source and target tasks. Some of the earliest works towards this goal include \cite{baxter2000model}~and linear settings of \cite{lounici2011oracle,pontil2013excess,wang2016distributed,cavallanti2010linear}. More recent works \cite{hanneke2020no,lu2021power,kong2020meta,wu2020understanding,garg2020functional,gulluk2021sample,du2020few,tripuraneni2020theory,qin2022non,tripuraneni2020provable,sun2021towards,maurer2016benefit,arora2019theoretical} consider variations beyond supervised learning, concrete settings or established more refined upper/lower performance bounds. There is also a long line of works related to model agnostic meta-learning \cite{finn2017model,denevi2019online,balcan2019provable,khodak2019adaptive}. Unlike these works, we consider the CL setting and show how the representation learned by earlier tasks provably helps learning the new tasks with fewer samples and better accuracy.
\noindent\textbf{Neural network pruning.} Our work is naturally related to neural network pruning methods and compression techniques as we embed tasks into sub-networks. Large model sizes in deep learning have led to a substantial interest in model pruning/quantization \cite{han2015deep,hassibi1993second,lecun1990optimal}. DNN pruning has a diverse literature with various architectural, algorithmic, and hardware considerations \cite{sze2017efficient,han2015learning}. Here, we mention the ones related our work. \cite{frankle2018lottery} empirically shows that a large DNN contains a small subset of favorable weights (for pruning), which can achieve similar performance to the original network when trained with the same initialization. \cite{zhou2019deconstructing,malach2020proving,pensia2020optimal} demonstrate that there are subsets with good test performance even without any training and provide theoretical guarantees. In relation \cite{chang2021provable} establishes the theoretical benefits of training large over-parameterized networks to improve downstream pruning performance.
Although weight pruning is proven to be a good way to reduce model parameters and maintain performance, practically, it does not lead to compute efficiency except for some dedicated hardwares~\cite{han2016eie}. Unlike weight pruning, structured/channel pruning prunes the model at the channel level which results in a slim sub-network carrying much less FLOPs than the original dense model~\cite{liu2017learning, zhuang2020neuron, wen2016learning, ye2018rethinking}. For example~\cite{wen2016learning,zhou2016less,alvarez2016learning,lebedev2016fast,he2017channel} prune models by adding a sparse regularization over model weights whereas \cite{liu2017learning, zhuang2020neuron} only add regularization over channel factors, and prune channels with lower scaling factors. However, these prior works don't focus on the scenario where almost all of FLOPs pruned, for example with only 1\% of original FLOPs remained. To achieve this goal, we present an innovative channel pruning method based on FLOP-aware penalization. Our technique is inspired from \cite{liu2017learning, zhuang2020neuron} (it uses sparsity regularization over BatchNorm weights only) however it outperforms both methods as demonstrated in Appendix~\ref{pruning sec}.
\section{Theoretical Analysis of Data Efficiency and Continual Representation Learning}
\section{Empirical and Theoretical Insights for Continual Representation Learning}\label{sec:crl}
In this section, we discuss continual learning from the representation learning perspective. {We first present our experimental insights which show that (1) features learned from previous tasks help reduce the sample complexity of new tasks and (2) the order of task sequence (in terms of diversity and sample size) is critical for the success of CRL. In Section~\ref{sec:crl_theory}, we present our theoretical framework and a rigorous analysis in support of our experimental findings.}
\subsection{Empirical Investigation of CRL}\label{sec:crl_exp}
We further elucidate upon Figure \ref{figure 2 label} and discuss the role of sample size (for both past and new tasks) and task diversity.
\noindent\textbf{Investigating data efficiency (Fig \ref{fig:LS})} A good test to assess benefit of CRL is by constructing settings where new tasks have fewer samples. Consider SplitCIFAR100 for which $100$ classes are randomly partitioned into $20$ tasks. We partition the tasks into two sets: a continual learning set ${\cal{D}}_{cl}=\{\mathcal{T}_1,\dots,\mathcal{T}_{15}\}$ and a test set ${\cal{D}}_t=\{\mathcal{T}_{16},\dots,\mathcal{T}_{20}\}$. Test set is used to assess data efficiency and therefore, (intentionally) contains only 10\% of the original sample size (250 samples per task instead of 2500). We first train the network sequentially using ${\cal{D}}_{cl}$ via ESPN/PackNet and create checkpoints of the supernetwork at different task IDs: At time $t$, we get a supernet \ylm{$s_t$ that is} trained with $\mathcal{T}_1,\dots,\mathcal{T}_t$, for $t\leq15$. {$t=0$} stands for the initial supernet without any training. Then we assess the representation quality of different supernet by individually training tasks in ${\cal{D}}_t$ on it. Fig.~\ref{fig:LS} displays the test accuracy of tasks in ${\cal{D}}_t$ where we used SplitCIFAR100 setting detailed in Sec.~\ref{sec:setting}. Figure~\ref{fig:LS} shows that ESPN-0.2, ESPN-1, and PackNet methods all benefit from features trained by earlier tasks since the accuracy is above $75\%$ when we use supernets trained sequentially with multiple tasks. In contrast, individual learning trains separate models for each task where no knowledge is transferred; the accuracy is close to $68\%$. Note that the performance of ESPN gradually increases with the growing number of continual tasks. This reveals its ability to successfully transfer knowledge from previous learned tasks and reduce sample complexity.
\input{sec/fig/diversity}
\noindent\textbf{Importance of task order and diversity.} To study how task order and diversity benefits CL, similar to \cite{mallya2018packnet,hung2019compacting}, we use $6$ image classification tasks, where ImageNet-1k~\cite{krizhevsky2012imagenet} is the first task, followed by CUBS~\cite{wah2011caltech}, Stanford Cars~\cite{krause20133d}, Flowers~\cite{nilsback2008automated}, WikiArt~\cite{saleh2015large} and Sketch~\cite{eitz2012humans}. Intuitively, ImageNet should be trained first because of its higher diversity.
Figure~\ref{fig:diversity} shows the accuracy improvement on the $5$ tasks that follow ImageNet pretraining compared to individual training.
The results are displayed in Fig.~\ref{fig:diversity} where Green bars are CL with ImageNet as the first task, Orange is CL without ImageNet, and Blue is Individual training. In essence, this shows the importance of initial representation diversity in CL since results with the ImageNet pretraining (Green) are consistently and strictly better than no pretraining (Orange) and Individual (Blue).
We note that related findings for zero-forgetting CL have been reported in \cite{hung2019compacting,mallya2018piggyback,mallya2018packnet,mallya2018piggyback,tu2020extending} which further motivates our theory in Sec \ref{sec:crl_theory}. Unlike these works, our experiment aims to isolate the CRL benefit of ImageNet by training other 5 tasks continually without ImageNet. }%\textcolor{blue}{We defer experimental details to the supplementary material.}
\noindent\textbf{Importance of sample size.}
Finally, we show that the sample size is also critical for CRL because one can build higher-quality (less noisy) features with more data. To this end, we devise another experiment based on SplitCIFAR100 dataset. Instead of using original tasks each with 5 classes and 2,500 samples, we train the first task (task ID 1 in Fig.~\ref{fig:Sample}) with 2,500 samples, then decrease the sample size for all the following tasks using the rule $2,500\times (1/20)^{\text{(ID-1)}/19}$ until the last task (task ID 20 in Fig.~\ref{fig:Sample}) has only 125 training samples. Figure~\ref{fig:Sample} presents the results, where solid curves are obtained by training Task ID 1 to 20 with decreasing sample size, dashed curves are for training Task ID 20 to 1 with increasing sample size, and dotted curves are for individual training where task order does not matter. The accuracy curves are smoothed with a moving average and displayed in the decreasing order from Task ID 1 to 20. The results support our intuition that training large sample tasks first (decreasing order) performs better, as larger tasks build high quality representations that benefit generalization for future small tasks with less data. More strikingly, the dotted Individual line falls strictly between solid and dashed curves. This means that increasing order actually \emph{hurts accuracy} whereas decreasing order \emph{helps accuracy} compared to training from scratch (i.e.~no representation). Specifically, decreasing helps (solid$>$dotted) on the right side of the figure where tasks are small (thanks to good initial representations) whereas increasing hurts on the left side where tasks are large. The latter is likely due to the fact that, adding a large task requires a larger/better subnetwork to achieve high accuracy, however, since we train small tasks first, supernetwork runs out of sufficient free weights for a large subnetwork.
\subsection{Theoretical Analysis and Performance Bounds for Continual Representation Learning}\label{sec:crl_theory}
In this section, we provide theoretical analysis to explain how CRL provably promotes sample efficiency and benefits from initial tasks with large diversity and sample size.
We use $\ordet{\cdot}$ to denote equality up to a factor involving at most logarithmic terms. Following our experiments as well as \cite{maurer2016benefit}, a realistic model for deep representation learning is the compositional hypothesis $f=h\circ \phi$ where $h\in{\mtx{H}}$ is the classifier head and $\phi\in{\boldsymbol{\Phi}}$ is the shared backend feature extractor. In practice, $\phi$ has many more parameters than $h$. To model continual learning, let us assume that we already trained a frozen feature extractor $\phi_{\text{frz}}\in{\boldsymbol{\Phi}}_{\text{frz}}$\ylm{ on earlier tasks {from which $\phi$ can be learned faster}}. {Here ${\mtx{H}},{\boldsymbol{\Phi}},{\boldsymbol{\Phi}}_{\text{frz}}$ are the hypothesis sets to learn from.} Suppose we are now given a set of $T}%_{\text{new}}$ new tasks represented by independent datasets $\mathcal{S}_t=\{(\vct{x}_{ti},y_{ti})\}_{i=1}^N{\subset \mathcal{X}\times \mathcal{Y}}$, each drawn i.i.d.~from different distributions ${\cal{D}}_t$ for $1\leq t\leq T$. \ylm{$\mathcal{X},~\mathcal{Y}$ are the sets of feasible input features and labels respectively.} Our goal is to build the hypotheses $(f_t)_{t=1}^{T}%_{\text{new}}}:\mathcal{X}\rightarrow\mathbb{R}$ for these new tasks with small sample size $N$ while leveraging $\phi_{\text{frz}}\in{\boldsymbol{\Phi}}_{\text{frz}}$.
\textbf{CRL setting.} In a realistic CL setting the new tasks are allowed to learn new features. We will capture this with an \emph{incremental} feature extractor {$\phi_{\text{new}}\in {\boldsymbol{\Phi}}_{\text{new}}$, and represent the hypothesis of each task via the composition $f_t=h_t\circ \phi$ where $\phi=\phi_{\text{new}}+\phi_{\text{frz}}$. For PackNet/ESPN, ${\boldsymbol{\Phi}}_{\text{new}}$ corresponds to the free/trainable weights allocated to the new task, $\phi_{\text{frz}}$ corresponds to the trained weights of the earlier tasks and $\phi$ corresponds to the eventual task subnetwork and its weights. We will evaluate the quality of $\phi$ (which lies in the Minkowski sum ${\boldsymbol{\Phi}}_{\text{new}}+{\boldsymbol{\Phi}}_{\text{frz}}$) with respect to a global representation space ${\boldsymbol{\Phi}}$ which is chosen to be a superset: ${\boldsymbol{\Phi}}_{\text{new}}+{\boldsymbol{\Phi}}_{\text{frz}}\subseteq {\boldsymbol{\Phi}}$. For instance, in PackNet, ${\boldsymbol{\Phi}}_{\text{new}}+{\boldsymbol{\Phi}}_{\text{frz}}$ denotes the sparse sub-networks allocated to the \ylm{new and previous} tasks whereas ${\boldsymbol{\Phi}}$ corresponds to the full supernet.}
This motivates us to pose a CRL problem that builds a continual representation by searching for $\phi_{\text{new}}$ and combining with $\phi_{\text{frz}}$. Let ${\vct{h}}=(h_t)_{t=1}^T}%_{\text{new}}\in {\mtx{H}}^T}%_{\text{new}}$ denote all $T}%_{\text{new}}$ task-specific classifier heads, we solv
\begin{align}
\underset{\phi=\phi_{\text{new}}+\phi_{\text{frz}}}{\underset{{\vct{h}}\in{\mtx{H}}^T}%_{\text{new}},\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}}{\arg\min}}&{\widehat{\cal{L}}}({\vct{h}},\phi):=\frac{1}{T}%_{\text{new}}}\sum_{t=1}^T}%_{\text{new}} {\widehat{\cal{L}}}_{\mathcal{S}_t}(h_t\circ\phi)\nonumber\\
\text{WHERE}\quad &{\widehat{\cal{L}}}_{\mathcal{S}_t}(f):=\frac{1}{N}\sum_{i=1}^N \ell(y_{ti},f(\vct{x}_{ti})) \tag{CRL}\label{crl}.
\end{align}
\noindent \textbf{Intuition:} \eqref{crl} aims to learn the task-specific headers ${\vct{h}}$ and the shared incremental representation $\phi_{\text{new}}$. Let $\cc{\cdot}$ be a complexity measure for a function class (e.g.~VC-dimension). Intuitions from the MTL literature would advocate that when the total sample size obeys $N\times T}%_{\text{new}}\gtrsim T}%_{\text{new}}\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}}$, then \eqref{crl} would return generalizable solutions ${\vct{\hat{h}}},\hat{\phi}_{\text{new}}$. This is desirable as in practice $\phi_{\text{frz}}$ is a much more complex hypothesis obtained by training on many earlier tasks. Thus, from continual learning perspective, theoretical goals are:
\begin{enumerate}
\item The sample size should only depend on the complexity $\cc{{\boldsymbol{\Phi}}_{\text{new}}}$ of the incremental representation rather than the combined complexity that can potentially be much larger ($\cc{{\boldsymbol{\Phi}}_{\text{new}}}+\cc{{\boldsymbol{\Phi}}_{\text{frz}}}\gg \cc{{\boldsymbol{\Phi}}_{\text{new}}}$).
\item To explain Figure \ref{fig:LS}, we would like to quantify how frozen representation $\phi_{\text{frz}}$ can provably help accuracy. Ideally, thanks to $\phi_{\text{frz}}$, we can discover a near-optimal $\phi$ from the larger hypothesis set ${\boldsymbol{\Phi}}$.
\item Finally, we emphasize that, we add the $T$ new tasks to the network in one round for the sake of cleaner exposition. }%\textcolor{blue}{In appendix, we provide synergistic theory and detailed investigation of the scenario where tasks are learned sequentially in a continual fashion and frozen features $\phi_{\text{frz}}$ evolve as we add more tasks.} In a nutshell, this theory explains Figure \ref{fig:Sample} by quantifying the role of sample size in the quality of continual representations.
\end{enumerate}
Before stating our technical results, we need to introduce a few definitions. To quantify the complexities of the search spaces ${\boldsymbol{\Phi}}_{\text{new}},{\mtx{H}}$, we introduce \emph{metric dimension} \cite{mendelson2003few}, which is a generalization of the VC-dimension \cite{vapnik2015uniform}.
\begin{definition}[Metric dimension] \label{def:cov} Let ${\mtx{G}}:\mathcal{Z}\rightarrow\mathcal{Z}'$ be a set of functions. {Let $\bar{C}_{\mathcal{Z}}>0$ be a scalar that is allowed to depend on $\mathcal{Z}$.} Let ${\mtx{G}}_{\varepsilon}$ be a minimal-size $\varepsilon$-cover of ${\mtx{G}}$ such that for any $g\in{\mtx{G}}$ there exists $g'\in {\mtx{G}}_{\varepsilon}$ that ensures $\sup_{\vct{x}\in\mathcal{Z}}\tn{g(\vct{x})-g'(\vct{x})}\leq \varepsilon$. The metric dimension $\cc{{\mtx{G}}}$ is the smallest number that satisfies $\log|{\mtx{G}}_{\varepsilon}|\leq \cc{{\mtx{G}}}\log(\bar{C}_{\mathcal{Z}}/\varepsilon)$ for all $\varepsilon>0$.
\end{definition}
{$\bar{C}_{\mathcal{Z}}$ typically depends only logarithmically on the Euclidean radius of the feature space under mild Lipschizness conditions, thus, $\bar{C}_{\mathcal{Z}}$ dependence will be dropped for cleaner exposition.} In practice, for neural networks or other parametric hypothesis, metric dimension is bounded by the number of trainable weights up to logarithmic factors \cite{barron2018approximation}.
Metric dimension will help us characterize the sample complexity. However, we also would like to understand when ${\boldsymbol{\Phi}}_{\text{frz}}$ can help. To this end, we introduce definitions that capture the population loss (infinite data limit) of new tasks and the \emph{feature compatibility} between the new tasks and $\phi_{\text{frz}}$ of old tasks. These definitions help decouple the finite sample size $N$ and the distribution of the new tasks.
\begin{definition}[Distributional quantities]\label{def pop}Define the population (infinite-sample) risk as ${\cal{L}}({\vct{h}},\phi)=\operatorname{\mathbb{E}}[{\widehat{\cal{L}}}({\vct{h}},\phi)]$. Define the optimal risk over representation ${\boldsymbol{\Phi}}$ as ${\cal{L}}^{\st}=\min_{{\vct{h}}\in{\mtx{H}}^T}%_{\text{new}},\phi\in{\boldsymbol{\Phi}}}{\cal{L}}({\vct{h}},\phi)$. Note that the optimal risk can always choose the best representations within ${\boldsymbol{\Phi}}_{\text{new}}$ and ${\boldsymbol{\Phi}}_{\text{frz}}$ since ${\boldsymbol{\Phi}}_{\text{new}}+{\boldsymbol{\Phi}}_{\text{frz}}\subseteq {\boldsymbol{\Phi}}$. Finally, define the optimal population risk using frozen $\phi_{\text{frz}}$ to be ${\cal{L}}^{\st}_{{\phi}_{\text{frz}}}=\min_{{\vct{h}}\in{\mtx{H}}^T}%_{\text{new}},\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}}}{\cal{L}}({\vct{h}},\phi)$ s.t.~$\phi=\phi_{\text{new}}+\phi_{\text{frz}}$.
\end{definition}
Following this, the \emph{representation mismatch} introduced below assesses the suboptimality of $\phi_{\text{frz}}$ for the new task distributions compared to the optimal hypothesis within ${\boldsymbol{\Phi}}$.
\begin{definition}[New \& old tasks mismatch]\label{def MM}
{The representation mismatch between the frozen features $\phi_{\text{frz}}$ and the new tasks} is defined as \vspace{-8pt}
\begin{align*}
\text{MM}_{\text{frz}}}%^{\text{new}={\cal{L}}^{\st}_{\phi_{\text{frz}}}-{\cal{L}}^{\st}.
\end{align*}
\end{definition}
By construction, $\text{MM}_{\text{frz}}}%^{\text{new}$ is guaranteed to be non-negative. Additionally, $\text{MM}_{\text{frz}}}%^{\text{new}=0$ if we choose global space to be ${\boldsymbol{\Phi}}={\boldsymbol{\Phi}}_{\text{new}}+{\boldsymbol{\Phi}}_{\text{frz}}$ and $\phi_{\text{frz}}$ to be the optimal hypothesis wihin ${\boldsymbol{\Phi}}_{\text{frz}}$. With these definitions, we have the following generalization bound regarding \eqref{crl} problem. }%\textcolor{blue}{The proof is deferred to appendix\ylm{the Appendix \ref{app B}}.}
\begin{theorem}\label{cl thm} Let ${\vct{h}},{\vct{\hat{h}}}$ denote the set of classifiers $(h_t)_{t=1}^T}%_{\text{new}},({\hat{h}}_t)_{t=1}^T}%_{\text{new}}$ respectively and $({\vct{\hat{h}}},\hat{\phi}=\hat{\phi}_{\text{new}}+\phi_{\text{frz}})$ be the solution of \eqref{crl}. Suppose that the loss function $\ell(y,\hat{y})$ takes values on $[0,1]$ and is $\Gamma$-Lipschitz w.r.t.~$\hat{y}$. {Suppose that input set $\mathcal{X}$ is bounded and all $\phi_{\text{new}}\in{\boldsymbol{\Phi}}_{\text{new}},~h\in{\mtx{H}}$, \ylm{${\boldsymbol{\Phi}}_{\text{new}}\in\{{\boldsymbol{\Phi}}_{\text{new}}^i,1\leq i\leq k\}$,} and $\phi_{\text{frz}}$ have Lipschitz constants upper bounded with respect to Euclidean distance}. With probability at least $1-2e^{-\tau}$, the task-averaged population risk of the solution $({\vct{\hat{h}}},\hat{\phi})$ obeys
{
\begin{align}
{\cal{L}}({\vct{\hat{h}}},\hat{\phi})&\leq {\cal{L}}^{\st}_{\phi_{\text{frz}}}+ \sqrt{\frac{\ordet{T}%_{\text{new}}\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}}+\tau}}{T}%_{\text{new}} N}},\nonumber\\
&\leq {\cal{L}}^{\st}+{\text{MM}_{\text{frz}}}%^{\text{new}}+{\sqrt{\frac{\ordet{T}%_{\text{new}}\cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}}+\tau}}{T}%_{\text{new}} N}}}.\nonumber
\end{align}}
\end{theorem}
In words, this theorem shows that as soon as the total sample complexity obeys $T}%_{\text{new}} N\gtrsim T}%_{\text{new}} \cc{{\mtx{H}}}+\cc{{\boldsymbol{\Phi}}_{\text{new}}}$, we achieve small excess statistical risk and avoid the sample cost of learning $\phi_{\text{frz}}$ from scratch. {Importantly, the sample cost $\cc{{\boldsymbol{\Phi}}_{\text{new}}}$ of learning the incremental representation is shared between the tasks since per-task sample size $N$ only needs to grow with $\cc{{\boldsymbol{\Phi}}_{\text{new}}}/T$.}
Reusing ${\boldsymbol{\Phi}}_{\text{frz}}$ comes at the cost of a prediction bias $\text{MM}_{\text{frz}}}%^{\text{new}$ arising from the feature mismatch. Also, with access to a larger sample size (e.g.~$T}%_{\text{new}} N\gtrsim T}%_{\text{new}}\cc{{\mtx{H}}}+ \cc{{\boldsymbol{\Phi}}}$), new tasks can learn a near-optimal $\phi^\star\in{\boldsymbol{\Phi}}$ from scratch\footnote{In this statement, we ignore the continual nature of the problem and allow $\phi_{\text{frz}}$ to be overridden for the new tasks if necessary.}. Thus, the benefit of \ref{crl} on data-efficiency is most visible when the new tasks have few samples, which is exactly the setting in Figure \ref{fig:LS}.
\noindent$\bullet$ \textbf{${\boldsymbol{\Phi}}_{\text{new}}$ \& representation diversity.} Imagine the scenario where $\phi_{\text{frz}}$ is already very rich and approximately coincides with the optimal hypothesis within the global space ${\boldsymbol{\Phi}}$. This is precisely the ImageNet setting of Figure \ref{fig:diversity} where even fine-tuning $\phi_{\text{frz}}$ will achieve respectable results. Mathematically, this corresponds to the scenario where ${\boldsymbol{\Phi}}_{\text{new}}$ is empty set but the mismatch is $\text{MM}_{\text{frz}}}%^{\text{new}\approx 0$. In this case, our theorem reduces to the standard few-shot learning risk where the only cost is learning ${\mtx{H}}$ i.e.~${\cal{L}}({\vct{\hat{h}}},\hat{\phi})\leq {\cal{L}}^{\st}+ \sqrt{{\ordet{\cc{{\mtx{H}}}}}/{N}}$.
\noindent$\bullet$ \textbf{$\text{MM}_{\text{frz}}}%^{\text{new}$ \& initial sample size.} Note that $\phi_{\text{frz}}$ is built using previous tasks which has finite samples. }%\textcolor{blue}{The sequential CL analysis we develop in appendix decomposes mismatch as $\text{MM}_{\text{frz}}}%^{\text{new}\lesssim\text{MM}_{\text{frz}}}%^{\text{new}^\star+\ordet{\cc{{\boldsymbol{\Phi}}_{\text{frz}}}/N_{\text{prev}}}$ where $\text{MM}_{\text{frz}}}%^{\text{new}^\star$ is the mismatch if previous tasks had $N_{\text{prev}}=\infty$ samples and $\ordet{\cc{{\boldsymbol{\Phi}}_{\text{frz}}}/N_{\text{prev}}}$ is the excess mismatch due to finite samples shedding light on Figure \ref{fig:Sample}. }
Our analysis is related to the literature on representation learning theory \cite{maurer2016benefit,kong2020meta,wu2020understanding,du2020few,gulluk2021sample,tripuraneni2020theory,arora2019theoretical}. Unlike these works, we consider the CL setting and show how the representation learned by earlier tasks provably helps learning the new tasks with fewer samples and how initial representation diversity and sample size benefit CRL. }%\textcolor{blue}{Importantly, we also specialize our general results to neural networks (see Theorem \ref{cl thm3} and Appendix \ref{app:application}) to obtain tight sample complexity bounds (in the degrees of freedom).}
}%[1]{\textcolor{red}{#1}}% \textcolor{red}{
\noindent$\bullet$ \textbf{Adding tasks sequentially.} Theorem \ref{seq thm} in the appendix also provides guarantees for the practical setting where $T$ tasks are added sequentially to the network. Informally, this theorem provides the following cumulative generalization bound on $T$ tasks (see Line \eqref{gen bound 3})
\[
\sum_{i=1}^T \text{``excess risk of task $t$''} \leq \sum_{i=1}^T \text{``mismatch of task $t$''}+T^2 \times \text{``statistical error per task''}.
\]
Here, ``mismatch of task $t$'' is evaluated with respect to earlier $t-1$ tasks assuming those $t-1$ tasks have infinite samples. ``statistical error per task'' is the generalization risk arising from each task having finite samples. In our bound, the growth of statistical error is quadratic in $T$ because when the new task $t$ uses imperfect representations learned by the earlier $t-1$ tasks (which have finite samples), the statistical error may aggregate. The overall proof idea is decoupling the population risk from empirical risk by defining the `mismatch of task $t$'' in terms of the landscape of the population loss (see Definition \ref{def pop seq}).}
\section{Preliminaries}
\textbf{Notation.} Let $\{\mathcal{T}_\tau, \tau\in \mathbb N^+\}$ be task identifier and ${\cal{D}}_\tau$ be its sample distribution. $\{(\vct{x}_{\tau,i},y_{\tau,i})\}_{i=1}^{n_\tau}$ are input-output pairs sampled from ${\cal{D}}_\tau$ to generate $\mathcal{T}_\tau$ where $n_\tau$ is the sample size.
In the following discussion, we consider $\tau\leq T$ where $T\in\mathbb N^+$.
\subsection{}
\begin{align*}
{\boldsymbol{\theta}}^M&:=\arg\min_{{\boldsymbol{\theta}}}\sum_{\tau=1}^T\sum_{i=1}^{n_\tau}\ell(f(\vct{x}_{\tau,i};{\boldsymbol{\theta}}),y_{\tau,i})\\
{\boldsymbol{\theta}}^I_\tau&:=\arg\min_{{\boldsymbol{\theta}}}\sum_{i=1}^{n_\tau}\ell(f(\vct{x}_{\tau,i};{\boldsymbol{\theta}}),y_{\tau,i})\\
{\boldsymbol{\theta}}_\tau^{CL}&:=\arg\min_{{\boldsymbol{\theta}},f_\tau}\sum_{i=1}^{n_\tau}\ell(f_\tau(\vct{x}_{\tau,i};{\boldsymbol{\theta}}_{\tau-1}^{CL},{\boldsymbol{\theta}}),y_{\tau,i})
\end{align*}
\section{Related Work
Our contributions are closely related to the representation learning theory as well as continual learning methods.
\noindent\textbf{Representation learning theory.} The rise of deep learning motivated a growing interest in theoretical principles behind representation learning. Similar in spirit to this project, \cite{maurer2016benefit} provides generalization bounds for representation-based transfer learning in terms of the Rademacher complexities associated with the source and target tasks. Some of the earliest works towards this goal include \cite{baxter2000model}~and linear settings of \cite{lounici2011oracle,pontil2013excess,wang2016distributed,cavallanti2010linear}. More recent works \cite{hanneke2020no,lu2021power,kong2020meta,qin2022non,wu2020understanding,garg2020functional,gulluk2021sample,xu2021statistical,du2020few,tripuraneni2020theory,tripuraneni2020provable,maurer2016benefit,arora2019theoretical,chen2021weighted,sun2021towards} consider variations beyond supervised learning, concrete settings or established more refined upper/lower performance bounds. There is also a long line of works related to model agnostic meta-learning \cite{finn2017model,denevi2019online,balcan2019provable,khodak2019adaptive}. {Unlike these works, we consider the CL setting where tasks arrive sequentially and establish how the representations learned by earlier tasks help learning the new tasks with fewer samples and better accuracy.}
\noindent \textbf{Continual learning.}
A number of methods for continual and lifelong learning have been proposed to tackle the problem of catastrophic forgetting. Existing approaches can be broadly categorized into three groups: replay-based~\cite{lopez2017gradient,borsos2020coresets}, regularization-based~\cite{kirkpatrick2017overcoming, jung2020continual}, and architecture-based methods~\cite{yoon2017lifelong}. }%[1]{\textcolor{red}{#1}}% \textcolor{red}{Recent work \cite{ramesh2021model} explores statistical challenges associated with continual learning in terms of the relatedness across tasks through replay-based strategies. In comparison, we focus on the benefits of representation learning and establish how representation built for previous tasks can drastically reduce the sample complexity on new tasks in terms of representation mismatch without access to past data. Another key difference is that our analysis allows for learning multiple sequential tasks rather than a single task.} In our work, consistent with our theory, we focus on zero-forgetting CL~\cite{mallya2018packnet,hung2019compacting,mallya2018piggyback,wortsman2020supermasks,kaushik2021understanding,tu2020extending}, which is a sub-branch of architecture-based methods and completely eliminates forgetting. \cite{hung2019compacting, kaushik2021understanding, mallya2018packnet} train a sub-network for each task and implement zero-forgetting by freezing the trained parameters, while \cite{mallya2018piggyback} trains masks only over pretrained model. Inspired by \cite{ramanujan2020s}, SupSup \cite{wortsman2020supermasks} trains a binary mask for each task while keeping the underlying model fixed at initialization.
However in order to embed the super-network with many tasks or to achieve acceptable performance over a masked random network, sufficiently large networks are needed.
Additionally, without accounting for the network structure, the inference-time compute cost of these networks is high even for simple tasks. Our proposed method is designed to overcome these challenges. In addition to achieving zero-forgetting, our algorithm also provides efficient inference in terms of FLOPs.
\subsection{Linear regression}
Consider a linear feature matrix $\vct\Phi=[\vct \phi_1, \vct \phi_2,...,\vct \phi_r]\in\mathbb{R}^{d\times r}$ where $d\gg r$ and $\{\vct\phi_i\}_{i=1}^r$ are orthogonal. Given $\{(\vct{x}_{\tau,i},y_{\tau,i})\}_{i=1}^{n_\tau}\sim{\cal{D}}_\tau$, describe a linear regression problem as
\begin{align*}
y=\vct{x}^\top\vct\beta+\epsilon~~\text{where}~~\vct\beta=\vct\Phi \vct{w}=\sum_{i=1}^r\vct{w}_i\vct\phi_i
\end{align*}
where $\vct{w}\in\mathbb{R}^r$ and $\epsilon$ is random noise. Therefore we have $\vct\beta\in\text{Span}\{\vct\phi_1,...,\vct\phi_r\}$. Define linear continual regression problem as
\begin{align*}
\tau\leq r:&\min_{\vct\phi}\|{\mtx{X}}_\tau\vct\phi-Y_\tau\|_2\\
&\Longrightarrow\hat{\vct\phi}_\tau=({\mtx{X}}_\tau^\top {\mtx{X}}_\tau)^{-1}{\mtx{X}}_\tau^\top Y_\tau\\
\tau > r:&\min_{\vct{w}}\|{\mtx{X}}_\tau\hat\Phi\cdot\vct{w}-Y_\tau\|_2~~\text{where}~~\hat{\vct\Phi}=[\hat{\vct\phi}_1,...,\hat{\vct\phi}_r]\\
&\Longrightarrow\hat{\vct{w}}_\tau =(({\mtx{X}}_\tau\hat{\vct\Phi})^\top{\mtx{X}}_\tau\hat{\vct\Phi})^{-1}{\mtx{X}}_\tau\hat{\vct\Phi} Y_\tau
\end{align*}
\subsection{General case}
Consider feature extractor $\phi\in\Phi$ and task-specific function $h\in\mathcal{H}$. Denote task predictor $f:=h\circ\phi\in\mathcal{H}\times \Phi$, where $\mathcal{H}$ and $\Phi$ are continuous search spaces. Let $\phi^M$ be optimal representation that trained from multitasks learning.
\begin{align*}
\hat\phi^M&=\arg\min_{\phi\in\Phi}\frac{1}{T}\sum_{\tau=1}^T{\cal{L}}(h_\tau\circ\phi(\vct{x}_\tau),y_\tau)\\
&\text{s.t.}~~h_\tau=\arg\min_{\vct{h}_\tau\in\mathcal{H}}\frac{1}{n_\tau}\sum_{i=1}^{n_\tau}{\cal{L}}(h_\tau\circ\phi(\vct{x}_{\tau,i}),y_{\tau,i})
\end{align*}
In practical problem, $\Phi$ is not uniformly distributed. Here let $\Phi'=(\Phi_1\times\Phi_2\times...\times\Phi_T)$ and assume that for any $\phi\in\Phi$, there exists $\phi'\in\Phi'$ that ${\cal{L}}(\phi')-{\cal{L}}(\phi)<\varepsilon$.
Define our continual learning problem as
\begin{align*}
\hat\phi_1&=\arg\min_{\phi\in\Phi_1, h_1\in\mathcal{H}}{\cal{L}}_{{\cal{D}}_1}(h_1\circ\phi)\\
\hat\phi_\tau&=\arg\min_{\phi\in(\hat\phi_1\times...\times\hat\phi_{\tau-1})\times\Phi_\tau, h_\tau\in\mathcal{H}}{\cal{L}}_{{\cal{D}}_\tau}(h_\tau\circ\phi)
\end{align*}
We will show that for all $\phi\in(\hat\phi_1\times...\times\hat\phi_{\tau-1})\times\Phi_\tau$
\begin{align*}
Dist(\hat\phi^M,\phi) =\mathcal{O}()
\end{align*}
Consider about individual learning that
\begin{align*}
\hat\phi^I_\tau=\arg\min_{\phi\in\Phi,h_\tau\in\mathcal{H}}{\cal{L}}(h_\tau\circ\phi)
\end{align*}
Then we will have
\begin{align*}
Dist(\hat\phi^M,\hat\phi_\tau^I)=\mathcal{O}()
\end{align*}
\section{Experimental Evidences in CL}\label{sec:exp}
\subsection{Investigation of Data Efficiency
To investigate how features learned from earlier tasks help in training new tasks in continual representation learning, we introduce a new experimental setting based on SplitCIFAR100 described in Sec~\ref{sec:setting} where 100 classes are randomly partitioned into 20 5-class classification tasks. Let us denote the original $20$ tasks as $\mathcal{T}_1,...,\mathcal{T}_{20}$. First, we split the tasks into a continual learning task set ${\cal{D}}_{cl}=\{\mathcal{T}_1,..,\mathcal{T}_{15}\}$ and a test task set ${\cal{D}}_t=\{\mathcal{T}_{16},...,\mathcal{T}_{20}\}$. Tasks in ${\cal{D}}_{cl}$ {retain} all $2500$ training samples per task. To test the data efficiency of CL, ${\cal{D}}_t$ tasks only have 10\% of their original training samples (i.e., $250$ for each task). We use ResNet18 described in Sec~\ref{sec:setting}. To quantify the benefits of learned feature representations in the continual learning process, we train the network sequentially using tasks in ${\cal{D}}_{cl}$ and test knowledge transferability separately over ${\cal{D}}_t$ after each task in ${\cal{D}}_{cl}$ added to the network. \red{Our goal is to quantify the benefits of learned feature representations in the continual learning process. To this aim, we train the network using samples from test tasks in ${\cal{D}}_t$ after every task in ${\cal{D}}_{cl}$ and use the test accuracy of ${\cal{D}}_t$ as a proxy for quantifying the quality of representations learned in previous tasks}.
\input{sec/fig/cifar_oc}
First, we use the CL method to sequentially learn tasks ${\cal{D}}_{cl}$ and save the supernet after every task. In this manner, we get 16 different supernets $\vct{s}_0,\vct{s}_1,...,\vct{s}_{15}$, where $\vct{s}_0$ stands for the initial supernet with no task, and $\vct{s}_i$ stands for the supernet after training $\mathcal{T}_1,\dots,\mathcal{T}_i$ using CL algorithm for $i\leq 15$. The source of the knowledge in $\vct{s}_i$ is the first $i$ tasks of the ${\cal{D}}_{cl}$ dataset. The test dataset ${\cal{D}}_t$ is used to test the transferability of the knowledge contained in supernet $\vct{s}_i$. {We separately train each of the 5 tasks of ${\cal{D}}_t$ by adding them into $\vct{s}_i$ in a continual learning manner (i.e.~by preserving the weights trained for ${\cal{D}}_{cl}$ }%\textcolor{blue}{and following the same channel pruning and weight allocation rules}).} This is done for all $16$ supernets $\vct{s}_i$. Training at supernet $\vct{s}_0$ is equivalent to training the network from scratch. Training at supernet $\vct{s}_i$ shows how features learned from $\mathcal{T}_1,\dots,\mathcal{T}_i$ help the test tasks in terms of data efficiency. We plot the average test accuracy on }%\textcolor{blue}{$5$ trails in Fig.~\ref{fig:LS} for each $\vct{s}_i$ and test accuracy in each trail is averaged over $5$ test tasks.}
Figure~\ref{fig:LS} summarizes the results of our experiment. Orange and Blue curves display ESPN-1 and ESPN-0.2 results. We also run PackNet, SupSup, and Individual-0.2/-1 baselines using the same setting as in Table~\ref{CIFARtable}\som{ with only difference being weight allocation $\alpha=0.05$ rather than $\alpha=0.1$}. Since SupSup uses a fixed model with random initialization and learns a mask for each task separately, no knowledge is transferred. This is why dashed lines are strictly horizontal. Figure~\ref{fig:LS} shows that ESPN-0.2, ESPN-1, and PackNet methods benefit from features trained by earlier tasks since the test accuracy significantly improves when we train the new task on top of a network containing more continual tasks. Observe that, the performance of ESPN gradually increases with the growing number of continual tasks. This reveals its ability to successfully transfer knowledge from previous learned tasks. \ylm{We believe ESPN's advantage over PackNet arises from the Task-specific BatchNorm weights which allow ESPN to better re-purpose the features learned by earlier tasks. The small gap between ESPN-0.2 and ESPN-1 also shows that our ESPN algorithm performs well in finding the most relevant features despite channel sparsity restrictions.} Complementing these experiments, the next section provides theoretical insights into the representation learning ability of PackNet/ESPN.
\subsection{Importance of task order: diversity and sample size}
\input{sec/fig/sample_size}
\section{Appendix}
\subsection{Channel Pruning}
To prune channel effectively and iteratively, we introduce channel mask in continuous value ($\mathcal{M}_l\in\mathbb{R}^{n_l}, l\in [L]$) to induce sparsity where $L$ is layer count each with $n_l$ channels/neurons. In our experiments, inspired by \cite{liu2017learning} we adopt BatchNorm layers and extract their weights as masks. Then we have
\begin{align*}
\vct{x}_l=(\mathcal{M}_l\cdot \text{norm}(\text{conv}(\vct{x}_{l-1})))_+,~~\text{norm}({\vct{z}})=\frac{{\vct{z}}-\mu_{\vct{z}}}{\sqrt{\sigma_{\vct{z}}^2+\epsilon}}
\end{align*}
where $\vct{x}_l$ is the output feature of $l^{th}$ layer and $\mathcal{M}_l$ is trainable masked weights of BN layer.
Given a single task, structure pruning is trying to find and remove insignificant channels/neurons without hurting performance. To achieve this goal, we propose a FLOP-base sparsity regularization
\begin{align*}
\mathcal{R}(\mathcal{M})&=\sum_{l\in [L]}\lambda_l\|\mathcal{M}_l\|_1\\
&=\sum_{l\in[L]}\frac{g(\text{FLOP}_l(\mathcal{M}))}{\sum_{j\in[L]}g(\text{FLOP}_j(\mathcal{M}))}\|\mathcal{M}_l\|_1
\end{align*}
where $\lambda_l$, $\mathcal{M}_l$ and $\text{FLOP}_l(\cdot)$ denote regularization parameter, channel mask and FLOP calculation function of $l^{th}$ layer. We apply $\ell_1$-norm to achieve sparsity, since it enforces unimportant elements to zeros. $g(\cdot)$ is monotonically increasing function used to calculate $\lambda_l$ automatically.
In deep neural network, different layer has different FLOPs load corresponding to its input/output feature size and operations. Therefore, it is unfair to use global regularization parameters. We adopt layerwise regularization parameter $\lambda_l$ and relate it to FLOPs load. Our goal is to push layers with more FLOPs sparser. Results show that this approach achieves good results especially when the network is pruned to very sparse. In our experiments, we use $g(x)=\sqrt{x}$.
\begin{table}[t]
\small
\caption{\small{Continual learning results}}
\centering
\begin{tabular}{lcc}
\midrule
& RotatedMNIST & PermutedMNIST\\
\midrule
ESPN-0.2 &97.21$\pm$0.007 95.78$\pm$0.015\\
PackNet &95.66$\pm$0.014 & 92.88$\pm$0.024\\
Individual-0.2&95.48$\pm$0.370 & 95.36$\pm$0.303\\
Individual-1 &97.36$\pm$0.149 & 97.31$\pm$0.107\\
\midrule
\end{tabular}\label{MNISTtable}
\end{table}
\section*{Organization of the Appendix}
Supplementary material is organized as follows.
\begin{enumerate}
\item Appendix~\ref{appsec: espnalgo} discusses additional details of our ESPN algorithm. In Appendix~\ref{pruning sec} we evaluate the benefits of our FLOP-aware pruning and weight allocation strategies.
\item Appendix~\ref{app:application} discusses an application of CRL, that is we introduce a shallow network by setting feature extractor $\phi$ and classifier $h$ to be specific functions.
\item Appendix~\ref{app B} provides the proof of our Theorem \ref{cl thm} and also proves Theorem~\ref{cl thm3} proposed in Appendix~\ref{app:application}.
\item Appendix~\ref{app C} proposes Theorem \ref{seq thm} which provides guarantee for continual representation learning when the $T$ tasks are added into the super-network in a sequential fashion. This theorem adds further insights into the bias-variance tradeoffs surrounding sequential learning (e.g.~how representation mismatch might aggregate as we add more tasks).
\item Appendix~\ref{app:related} provides an expanded discussion of related work on continual learning, representation learning and neural network pruning.
\item Appendix~\ref{app:imagenet} discusses our implementation setting of experiment in Figure~\ref{fig:diversity}.
\end{enumerate}
} | 4332e396bf318cef0f6a4365860dcede3393a2a5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Faint blue white dwarfs in the GCNS color-magnitude diagram}
\textit{Gaia} color-magnitude diagrams (CMDs) reveal
small numbers of faint blue white dwarfs (FBWDs) with
absolute \textit{Gaia} magnitudes of $Gabs \gtrapprox 15$\,mag
and a wide range of blue \textit{Gaia} colors.
These FBWDs may constitute
of ''ultracool'' \citep{2020MNRAS.499.1890M}
or ''infrared-faint'' \citep{2020ApJ...898...84K},
but possibly also ''ultramassive" \citep{2021MNRAS.503.5397K} WDs.
The \textit{Gaia} catalogue of nearby stars
\citep[GCNS;][]{2021A&A...649A...6G} is based on
the \textit{Gaia} Early Data Release 3 \citep[EDR3;][]{2021A&A...649A...1G}.
Based on many \textit{Gaia} astrometric parameters and quality flags,
about 500000 EDR3 objects with measured magnitudes $Gmag$ and
colors $G-RP$ and parallaxes $Plx>10$\,mas were divided into a
GCNS-selected 100\,pc sample of $\approx$296000 stars (59\%) and a
GCNS-rejected sample of $\approx$204000 (41\%) objects.
Except for a few objects with extremely red $G-RP$ colors, all GCNS-selected
stars are shown in Fig.~\ref{fig:abcd}(a). In this CMD
the main sequence is well separated from the white dwarf (WD) sequence,
which is mainly located within the marked color-magnitude box defined
by \citet{2018MNRAS.480.3942H} for all nearby WDs within 20\,pc measured in
the previous \textit{Gaia} DR2. However, 60 FBWDs appear
below this box, with absolute magnitudes
\begin{equation}
Gabs>14.7+4.7(G-RP).
\label{Eq1}
\end{equation}
The GCNS-rejected sample is
dominated by very faint objects towards the Galactic centre (GC)
and the Magellanic
clouds (MCs), where the \textit{Gaia} photometry, in particular in the $RP$
and $BP$ bands, and astrometry is strongly affected by image crowding.
Therefore, the tails of their $G-RP$ color distribution extend even outside
the blue and red limits shown in Fig.~\ref{fig:abcd}(b). Interestingly, there
is an overlap of the blue tail of GCNS-rejected objects (411 objects
following Eq.~\ref{Eq1}) with the above mentioned CMD region of FBWDs
that may throw discredit on the 60 GCNS-selected stars. On the other
hand, only 55 WDs in the EDR3-based catalog of \citet{2021MNRAS.508.3877G}
are found with Eq.~\ref{Eq1}, and 8 of them are members of the GCNS-rejected
sample (Fig.~\ref{fig:abcd}(c)).
\section{Astrometric verification of FBWD candidates by proper motion check}
To investigate how much the apparent branch of FBWDs may be
caused by \textit{Gaia} measuring errors, I inspected their
optical finder charts (including
own measurements in corresponding FITS images) and catalogs from
different sources. The idea was to not only estimate the influence of crowding
or close companions but confirm EDR3 proper motions above a chosen limit
of 15\,mas/yr by combining the
available \textit{Gaia} DR1-DR3 (epochs 2015-2016) positions with additional
positional data. The presence of a confirmed proper motion was considered as
supporting the EDR3 parallax, and consequently the absolute magnitude
of a FBWD candidate. However, I took
also into account the occurrence of other EDR3 sources at small
($\lessapprox$2\,arcsec) separations and EDR3 common proper motion (CPM)
objects within 3\,arcmin and their available parallaxes. External data were
taken from the APM \citep{1994IEEES...2...14I} and SuperCOSMOS
\citep[SSS;][]{2001MNRAS.326.1279H} measurements of photographic Schmidt
plates, the Sloan Digital Sky Survey \citep[SDSS;][]{2009ApJS..182..543A} and
Pan-STARRS release 1 \citep[PS1;][]{2017yCat.2349....0C}. A proper motion
comparison was also made with the extended \textit{Gaia}-PS1-SDSS catalog
\citep[GPS1+;][]{2020ApJS..248...28T}. In addition, I used the DESI Legacy
Imaging Surveys \citep[][hereafter the Legacy Surveys]{2019AJ....157..168D},
but only in exceptional cases near-infrared data from the VISTA Hemisphere
Survey \citep[VHS;][]{2013Msngr.154...35M}.
The EDR3 total proper motions $PM$ and numbers of visibility periods $Nper$
(an important astrometric quality parameters, in particular in
crowded regions)
of 411 GCNS-rejected objects are both systematically smaller than those
of 60 GCNS-selected stars (Fig.~\ref{fig:abcd}(d)). Compared to the full
sample of $\approx$302000 GCNS-selected stars with $Plx>10$\,mas,
where 27\% have $PM$ around
the peak of the distribution at 50$\pm$20\,mas/yr, only 15\% of the 60
stars fall in this interval.
Consequently, their tangential velocities are relatively
high (median $\approx$50\,km/s, maximum $\approx$160\,km/s).
I confirmed
the relatively high proper motions of all 60 GCNS-selected stars,
including the brightest (B = \object{Gaia EDR3 4049566372499936640})
and faintest (F = \object{Gaia EDR3 1674805012263764352}) of them
(Figs.~\ref{fig:abcd}(c,d)). With $Gabs \approx 17.3$\,mag, the latter
is nearly as faint as, but much bluer than, the unusual cool WD candidate
\object{Gaia EDR3 6584418167391671808} \citep{2021RNAAS...5..229A}.
Only one of the 60 blue stars (H = \object{Gaia EDR3 1643819327188791040})
was previously classified as
high-mass WD by \citet{2019ApJ...886..100C}. It has another WD
(\object{Gaia EDR3 1643819331484091136}) as CPM companion
(separation 8.9\,arcsec) with similar parallax.
The 60 objects are uniformly distributed over the
sky, but one (W = \object{Gaia EDR3 4110333184616593280})
appears in PS1 and Legacy Surveys as a blue foreground object in the GC.
It has a redder
CPM companion separated by 4\,arcsec with a slightly smaller but more
precise parallax. Using the latter, the absolute magnitude
changes as indicated by an arrow in Fig.~\ref{fig:abcd}(c). Of 13
GCNS-selected stars missing in \citet{2021MNRAS.508.3877G} I confirmed
12 as FBWDs. As seen in Fig.~\ref{fig:abcd}(d), half of them
have $Nper < 12$ (one overlaps here with the unconfirmed object R, see below).
All eight WDs of \citet{2021MNRAS.508.3877G} in the GCNS-rejected sample
have very small $PM$ and $Nper$ (Fig.~\ref{fig:abcd}(d)) including five
objects in the GC and MCs
regions and a quasar candidate (Q = \object{Gaia EDR3 6696431605961218176})
from \citet{2019MNRAS.490.5615B}. Only one
(C = \object{Gaia EDR3 5155424067737827968}) has a relatively high
$PM \approx 25$\,mas/yr that I confirmed using SDSS, PS1, and Legacy
Surveys data. However, it has a large parallax error of about 3\,mas
and a CPM companion (separation 46\,arcsec) with a much smaller parallax
and ten times smaller error indicating a distance of 700\,pc. When
checking the proper motions of about 70 more GCNS-rejected objects with
$PM > 15$\,mas/yr, 85\% of which were in the crowded GC and MCs regions,
I did not find further FBWD candidates.
\section{Crowding and close companions affecting \textit{Gaia} color
measurements}
The error bars shown in Fig.~\ref{fig:abcd}(c) are dominated by
$RP$ magnitude errors and relative parallax uncertainties. For
GCNS-rejected objects (green) they are larger than for GCNS-selected
stars (black). Most (53 of 60) GCNS-selected stars were also measured
in PS1 or Legacy Surveys (22 in both). All but one of these appear blue
on the corresponding finder charts. Surprisingly, the exception is
the object at the blue end of the shown CMD
(R = \object{Gaia EDR3 6309477283343732096}), which in fact appears red
in PS1 and VHS and is probably not a WD. It has a close (separation 3\,arcsec)
brighter CPM companion
($\Delta G \approx 5.8$\,mag but $\Delta J \approx 3.0$\,mag !)
with a similar parallax that obviously affected the $RP$ photometry.
\begin{figure*}[ht!]
\includegraphics[scale=0.75,angle=0]{RN_FBWDs_fig_rev.pdf}
\caption{Color-magnitude diagram of full GCNS-selected \textbf{(a)}
and GCNS-rejected \textbf{(b)} 100\,pc samples.
The zoom in \textbf{(c)}
shows only the region of FBWDs (Eq.~\ref{Eq1}). The green and black
error bars represent mean uncertainties of GCNS-rejected (green pluses)
and GCNS-selected (black crosses) subsamples, respectively.
Their total proper motions ($PM$) and numbers of visibility periods
($Nper$) are displayed in the inserted panel \textbf{(d)}. Objects
labelled in \textbf{(c)} and \textbf{(d)} are discussed in the text.
\label{fig:abcd}}
\end{figure*}
| e3d343cc9b78d0966bb557937cf1c95ad20c6725 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Rotational bands are ubiquitous in the spectra of medium-mass and heavy nuclei. As has been known for seventy years~\cite{Bohr:1951zz}, they emerge in a description of the nucleus as a nearly rigid axially-symmetric rotor~\cite{Rowe}. For even-even nuclei the simplest rotational bands consist of $0^+$, $2^+$, \ldots states and their energies are described by an expansion in powers of $I(I+1)$~\cite{Papenbrock:2010,CoelloPerez:2015}. This behavior has recently been obtained in {\it ab initio} calculations of the Be isotope chain~\cite{Caprio:2013,Maris:2014,Jansen:2015,Caprio:2019yxh,McCoy:2020xhp} and $^{34}$Mg~\cite{Hagen:2022tqp}.
Odd-mass neighbors of a rotor nucleus can then be understood as a fermion coupled to the rotor. The fermion dynamics is simpler in the intrinsic frame in which the nucleus is not rotating, but this frame is non-inertial, so solving the problem there induces a Coriolis force proportional to $\vec{j} \cdot \vec{I}$, the dot product of the single-fermion angular momentum and the total angular momentum of the fermion-rotor system. When combined with other mechanisms, such as excitation of the fermion to higher-single particle states and the fermion disturbing the rotor, this induces a string of terms in the energy-level formula~\cite{BohrMottelson}. Odd powers of $I$ appear, and produce staggering between adjacent levels. Which powers of $I$ are present depends on the value of the quantum number, $K$, the projection of the fermion angular momentum on the rotor axis. For $K=1/2$ bands the energy-level formula is:
\begin{eqnarray}
&& E(I)=A_K I(I+1) + E_K
+ A_{1} (-1)^{I+1/2} (I+1/2) \nonumber\\
&& \quad + B_{1} I(I+1) (-1)^{I+1/2} (I+\tfrac{1}{2})
+ B_K [I(I+1)]^2
\label{eq:energies}
\end{eqnarray}
where $A_K$, $E_K$, $A_1$, $B_1$, and $B_K$ are parameters, related to rotor properties and single-particle matrix elements, that need to be either derived from a microscopic model or estimated from data.
Over the years a number of models have had success describing this pattern from underlying density functional theory~\cite{Dudek:1981,Cwiok:1980,Afanasjev:2013,Zhang:2020} or shell-model~\cite{Inglis:1954,Velazquez:1999,Liu:2004,Zhang:2020} dynamics. The models also predict specific values for the coefficients that appear in Eq.~(\ref{eq:energies}).
In Ref.~\cite{Alnamlah:2020} we took a different approach, organizing the formula (\ref{eq:energies}) as an effective field theory (EFT) expansion in powers of the small parameter $Q \equiv 1/I_{br}$, with $I_{br}$ the spin of the nuclear state at which dynamical effects associated with single-particle and/or vibrational degrees of freedom cause the polynomial expansion in powers of $I$ to break down. This description of rotational bands in odd-mass nuclei builds on the successful EFT developed for even-even nuclei in Refs.~\cite{Papenbrock:2010,CoelloPerez:2015}. Other efforts to develop an EFT for these rotational bands can be found in Refs.~\cite{Papenbrock:2020zhh,Chen:2020qbf}.
In the odd-mass rotor EFT, Eq.~(\ref{eq:energies}) is the next-to-next-to-next-to-next-to leading order (\NkLO{4}) result for the energies, and the first corrections to it are $\mathcal{O}(E Q^4)$. The EFT analysis of Eq.~(\ref{eq:energies}) organizes it in terms of increasingly accurate predictions: the N$^k$LO energy-level formula has accuracy $\mathcal{O}(E Q^{k})$. All short-distance/high-energy physical mechanisms that affect the energies up to that accuracy are subsumed into the parameters or low-energy constants (LECs) that multiply the $I$-dependent terms in Eq.~(\ref{eq:energies}). In Ref.~\cite{Alnamlah:2020} we determined these LECs by fitting the lowest levels in the different rotational bands we analyzed. However, this runs the risk of fine-tuning the values of the LECs to those levels, and it does not provide uncertainty estimates for them. Better parameter estimation would use all the data available on a particular band, and account for the $\mathcal{O}(E Q^{k})$ truncation uncertainty present at order N$^k$LO~\cite{Schindler:2009,Furnstahl:2014}.
Bayesian methods for EFT parameter estimation do just that~\cite{Schindler:2009,Wesolowski:2016,Wesolowski:2019,Wesolowski:2021}. Reference~\cite{Wesolowski:2019} showed that the effect of neglected terms in the EFT expansion could be included in the error model by modifying the likelihood so that the covariance matrix that appears there includes both experimental uncertainties and EFT truncation errors. More recently, Ref.~\cite{Wesolowski:2021} showed that MCMC sampling of that likelihood enabled the simultaneous determination of the LECs and the parameters of the error model, i.e., the value of $Q$ and the typical size of the ``order 1'' dimensionless coefficients that appear in the EFT expansion.
In this work we apply the EFT parameter estimation technology developed in Refs.~\cite{Schindler:2009,Wesolowski:2016,Wesolowski:2019,Wesolowski:2021} to the problem of rotational bands in odd-mass nuclei. We consider $K=1/2$ bands in $^{99}$Tc, ${}^{167, 169}$Er, ${}^{167, 169}$Tm, ${}^{183}$W, ${}^{235}$U and ${}^{239}$Pu as well as $K=3/2$ bands in ${}^{155,157}$Gd and ${}^{159}$Dy.
Section~\ref{sec:rotorfermionEFT} summarizes the elements of the EFT that are relevant for this paper. Section~\ref{sec:Bayesianmodel} then develops the Bayesian statistical model we use to analyze data on rotational bands. We first write down the likelihood that includes both experimental and theory uncertainties, and then explain how we use known information on the expected size of the LECs and the expansion parameter to set priors. A novel feature of this work, compared to earlier Bayesian EFT parameter-estimation studies, is that our statistical model incorporates the possibility that the LECs at even and odd orders have different typical sizes. This reflects the physics of odd-order LECs that are associated with matrix elements of the fermion spin, while even-order LECs contain a combination of effects from the rotor and the fermion. Section~\ref{sec:sampler} contains details of our Markov Chain Monte Carlo sampler, and then Sec.~\ref{sec:results} presents the results for LECs and the expansion parameter, $Q$, that we obtain from sampling the Bayesian posterior. We conclude in Sec.~\ref{sec:conclusion}. All the results and figures generated from this work can be reproduced using publicly available Jupyter notebooks~\cite{notebooks}.
\section{Rotational EFT Background}
\label{sec:rotorfermionEFT}
Here we summarize the results of the EFT for rotational bands in odd-mass nuclei that was developed up to fourth order in the angular velocity of the system in Ref.~\cite{Alnamlah:2020}. This theory constructs the Lagrangian of the particle-rotor system using its angular velocity and the angular momentum of the unpaired fermion, $\vec{j}$, as building blocks. The resulting Lagrangian corrects that of a rigid rotor with contributions arranged as a series in powers of a small expansion parameter, $Q$, according to a power-counting scheme that counts powers of the system's angular velocity. Naively, we expect $Q$ to be of order $E_{\rm rot}/ E_{\rm high}$, where $E_{\rm rot}$ is the energy scale at which rotational excitation take place and $E_{\rm high}$ is the scale of high-energy physics not explicitly taken into account by the EFT.
At leading order (LO), the energy of a rotational band on top of a bandhead with spin $K$ is
\begin{equation}
\label{Energy_LO}
E_{\rm LO}(I,K) =A_{\rm rot} I(I+1) + E_K,
\end{equation}
where $I$ is the spin of the rotational state (or, equivalently, the total angular momentum of the fermion-rotor system), and $A_{\rm rot}$ and $E_K$ are LECs that must be fitted to experimental data. $A_{\rm rot}$ is determined by the moment of inertia of the even-even nucleus (the rotor) to which the unpaired fermion is coupled.
At next-to-leading order (NLO) rotational bands with $K=1/2$ are affected by a term that takes the same $\vec{j} \cdot \vec{I}$ form as the Coriolis force. This produces:
\begin{equation}
\label{Energy_NLO}
\begin{aligned}
E_{\rm NLO}(I,K) =& A_{\rm rot} I(I+1) + E_K\\
& + A_{1} (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right) \delta^K_{1/2}.
\end{aligned}
\end{equation}
The LEC $A_1$ is expected to be of order $A_{\rm rot}$ times a sum of matrix elements involving the fermion's total angular momentum operator (for details see Ref.~\cite{Alnamlah:2020}). From previous studies we see that $A_1/A_{\rm rot}<1$. This correction, sometimes called the signature term, causes staggering between adjacent states in $K=1/2$ bands.
The energy of a rotational band at next-to-next-to-leading order (\ensuremath{{\rm N}{}^2{\rm LO}}\xspace) is
\begin{equation}
\begin{aligned}
E_{\ensuremath{{\rm N}{}^2{\rm LO}}\xspace}(I,K) =& A_K I(I+1) + E_K\\
&+ A_{1} (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right) \delta^K_{1/2}.
\end{aligned}
\end{equation}
The term proportional to $A_K$ combines the LO term proportional to $A_{\rm rot}$ and corrections entering at this order with the same spin dependence. From our power counting we expect the shift $\Delta A=A_{\rm rot}-A_K$ to be of order $A_{\rm rot} Q$. In contrast to $A_{\rm rot}$, $A_K$ is band dependent and so should be fitted to data on the rotational band of interest.
The \ensuremath{{\rm N}{}^3{\rm LO}}\xspace corrections to the energy of a rotational band are both $\sim I^3$ for $I \gg 1$, but take a different form in the $K=1/2$ and $K=3/2$ bands:
\begin{equation}
\begin{aligned}
\Delta E_{\ensuremath{{\rm N}{}^3{\rm LO}}\xspace}(&I,K)\\
= B_{1}& (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right)I(I+1) \delta^K_{1/2} \\
+& A_{3} (-1)^{I+3/2} \left(I+\tfrac{1}{2}\right)\left(I-\tfrac{1}{2}\right)\left(I+\tfrac{3}{2}\right) \delta^K_{3/2}.
\end{aligned}
\end{equation}
with $B_1$ and $ A_3$ expected to be of order $A_1 Q^2$.
Last, at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace we have the additional term:
\begin{equation}
\Delta E_{\ensuremath{{\rm N}{}^4{\rm LO}}\xspace}(I,K) = B_K [I(I+1)]^2.
\end{equation}
with $B_K$ expected to be of order $A_{\rm rot} Q^3$.
This pattern continues: at odd orders we add terms that correct the staggering term and have LECs of order $A_1Q^{n-1}$, while the even-order terms provide the overall trend with $I$ and have LECs of order $A_{\rm rot}Q^{n-1}$. (In both cases $n$ is the order of our expansion.) This difference in the expected sizes of odd and even LECs comes from the physics. Odd-order LECs are associated with operators in the effective Lagrangian that couple rotor and fermionic degrees of freedom, while even-order LECs encode both rotor-fermion interactions and effects coming from the non-rigidity of the rotor itself.
In what follows we denote the LECs $A_1$, $\Delta A$, $B_1$, and $B_K$ generically as $\{a_n:n=1,\ldots,k\} \equiv {\bf a}_k$, where $k$ is the order of the EFT calculation. (In the case of $K=3/2$ bands the set is $\Delta A$, $A_3$, and $B_K$, and $a_1=0$.)
We then divide the $n$th-order LEC, $a_n$, by the the reference scale and the power of the expansion parameter assigned to it by the EFT power counting, i.e., construct:
\begin{equation}
c_n=\frac{a_n}{A_{\rm rot} Q^{n-1}}.
\label{eq:cs}
\end{equation}
We expect these coefficients $c_n$ to be of order one, i.e., they should be natural coefficients. However, because sets of odd and even natural coefficients seem to have different sizes we will assume the even and odd $c_n$'s are drawn from two different distributions with different characteristic sizes that we denote by $\bar{c}_{even}$ and $\bar{c}_{odd}$.
\section{Building the Bayesian Model}
\label{sec:Bayesianmodel}
\subsection{Building the Posterior}
Our goal in this analysis is to use the information on the expected size of LECs to stablize the extraction of their values as we add more levels to the analysis, or as we use energy-level formulae computed at different EFT orders. At the same time, we want to estimate the expansion parameter, $Q$, of the theory, as well as the characteristic sizes for even and odd coefficients, $\bar{c}_{even}$ and $\bar{c}_{odd}$.
We want to obtain the posterior distribution for all the LECs that appear at order $k$, a set we collectively denote by ${\bf a}_k$. Here we will obtain the joint posterior pdf of ${\bf a}_k$, the expansion parameter $Q$, and the characteristic sizes. To do this we follow the successful endeavor by the BUQEYE collaboration in Refs.~\cite{Wesolowski:2016, Wesolowski:2019, Wesolowski:2021}, and write the posterior, given experimental data, $\vec{y}_{exp}$, and prior information on the model, $P*$, as
\begin{equation}
\begin{aligned}
\pr({\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd}|&\vec{y}_{exp},P*)\\
=&\pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(Q|\bar{c}_{even},\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(\bar{c}_{even}|\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(\bar{c}_{odd}|\vec{y}_{exp},P*).
\label{eq:master}
\end{aligned}
\end{equation}
Marginalization of this posterior distribution over $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$ yields the posterior distribution for ${\bf a}_k$. Other marginalizations can be carried out to obtain posteriors for $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$.
This joint posterior distribution tells us the probability of the LECs and the error model parameters given experimental data. We could use this posterior distribution to get other quantities or observables, such as the energy of a particular rotational level, which depend on the LECs or the error model parameters. These are now represented by distributions and not single numbers. Their distributions are called posterior predictive distributions (PPD). We write the PPD of an observable $\mathcal{O}$ as
\begin{equation}
\label{PPD}
\pr(\mathcal{O}|\vec{y}_{exp},P*)=\int d\vec{\theta}\delta(\mathcal{O}-\mathcal{O}(\vec{\theta})) \pr(\vec{\theta}|\vec{y}_{exp},P*)
\end{equation}
where $\vec{\theta}$ represents the LECs and the error model parameters. Calculating the observable at each point in the parameter space $\vec{\theta}$ and then integrating over the parameters $\vec{\theta}$ allows one to carefully account for correlations between the parameters.
Using Bayes' theorem, we can express the posterior (\ref{eq:master}) as
\begin{equation}
\label{posterior}
\begin{aligned}
\pr({\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd}|&\vec{y}_{exp},P*)\\
=&\pr(\vec{y}_{exp}|{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr(\bar{c}_{even}|P*)\pr(\bar{c}_{odd}|P*)\\
\times& \frac{1}{{\pr(\vec{y}_{exp}|P*)}}.
\end{aligned}
\end{equation}
The terms on the right-hand side of Eq.~(\ref{posterior}) have the following interpretations:
\begin{enumerate}
\item $\pr(\vec{y}_{exp}|{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)$ is the likelihood of the experimental data given specific values of both the LECs that appear in the energy formula at order $k$ and the parameters in our error model.
\item $\pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},P*)$ is the prior distribution of the LECs given the parameters encoding the systematic expansion of the EFT.
\item $\pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)$ is the prior distribution of the expansion parameter given the characteristic sizes of even and odd natural coefficients.
\item $\pr(\bar{c}_{even}|P*)$ and $\pr(\bar{c}_{odd}|P*)$ are the prior distributions of the even and odd characteristic sizes. (In Eq.~\eqref{posterior} we assume an uncorrelated prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$.)
\item $\pr(\vec{y}_{exp}|P*)$ is the evidence, which we drop in what follows as it does not depend on the parameters we are interested in extracting and functions only as a normalization constant.
\end{enumerate}
\subsection{Building the Likelihood}
We now build the likelihood function accounting for the expected error between the experimental and theoretical values, for data on $K=1/2$ rotational bands. The corresponding likelihood for $K=3/2$ bands is built analogously. Following~\cite{Wesolowski:2019} we start by writing our observable (the energy of a particular rotation level) at order $k$ as
\begin{equation}
\label{eq:energy}
\begin{aligned}
E(I) &= A_{\rm rot}I(I+1)\\
\Bigg\{1 &+ \sum_{n=odd}^{k} c_n Q^{n-1}
(-1)^{I+1/2}\left(I+\tfrac{1}{2}\right)[I(I+1)]^{(n-3)/2}\\
&+ \sum_{n=even}^{k} c_n Q^{n-1} [I(I+1)]^{(n-2)/2}\Bigg\}.
\end{aligned}
\end{equation}
We choose the leading-order energy for each level, $A_{\rm rot}I(I+1)$, to be the reference scale $E_{\rm ref}$ for the observable. The dimensionless coefficients $c_n$ (see Eq.~(\ref{eq:cs})) are assumed to be $\mathcal{O}(1)$. The theory error $\vec{\sigma}_{th}$ at any order is due to terms omitted from the summations in Eq.~\eqref{eq:energy}. Its most significant contribution comes from the first omitted term in the EFT expansion. Accounting only for this term yields an estimate for the theory error that is fully correlated across levels if $k+1$ is even, and anticorrelated for adjacent levels if $k+1$ is odd. To account for this correlation or anticorrelation we write the theory covariance matrix as the outer product of a vector representing the theory error, $\Sigma_{th} \equiv \vec{\sigma}_{th} \otimes \vec{\sigma}_{th}$.
The vector $\vec{\sigma}_{th}$ contains the value of the first omitted term for each of the $m$ energy levels that enter the likelihood.
We also account for experimental errors by writing the covariance matrix as
\begin{equation}
\Sigma=\Sigma_{th}+\Sigma_{exp}
\end{equation}
where we take $(\Sigma_{exp})_{ij} \equiv (\vec{\sigma}_{exp})_i^2\delta_{ij}$.
The likelihood function is then
\begin{equation}
\begin{aligned}
\pr(\vec{y}_{exp}|&{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
=& \sqrt{\frac{1}{(2\pi)^m |\Sigma|}} \exp\left( -\frac{1}{2}\vec{r}^T \Sigma^{-1} \vec{r} \right),
\end{aligned}
\end{equation}
where $\vec{r}\equiv\vec{y}_{exp}-\vec{y}_{th}$ is the residual between the central experimental energy for a level and the theory result (\ref{eq:energy}) and
$m$ is the number of levels included in the likelihood estimation.
We note that since the theory error is the outer product of the theory error with itself, the theory covariance $\Sigma_{th}$ is singular. Including the experimental error solves this singularity problem for the covariance $\Sigma$. However, $\Sigma$ can still become ill-conditioned for higher values of $Q$ if the experimental errors are too small; numerical issues then arise when we try to invert the covariance matrix.
Including more terms in the estimate for the theoretical error produces a steeper peak in the likelihood function, see Fig.~\ref{fig:likelihood}, which, in turn, restricts the values sampled for $Q$ to a narrower region. Because it precludes the sampler exploring large values of $Q$, this inclusion of more omitted terms in the model of the theoretical error solves the numerical problem of ill-conditioned matrices and gives a more accurate extraction of the LECs and the error-model parameters.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{likelihood.pdf}
\caption{Comparing the Log of the likelihood when accounting for different number of omitted terms, $p$, in the theory error. Apart from $Q$, the parameters that enter the likelihood were chosen to be the median parameters after we have had sampled the posterior distribution for $^{169}$Er.}
\label{fig:likelihood}
\end{figure}
In what follows we estimate the theory error including omitted terms up to a certain cutoff order $k_{max}$. Our theory error estimate for the level with spin $I$ is then
\begin{equation}
\sigma_{th}(I)=A_{\rm rot}\sum_{l=k+1}^{k_{max}}\bar{c}_{even,odd} Q^{l-1} P_l(I),
\label{eq:theoryerror}
\end{equation}
where the $\bar{c}$ that is used here is $\bar{c}_{even}$ for even values of $l$ and $\bar{c}_{odd}$ otherwise. The $I$-dependence of the $l$th term is chosen to match that in Eq.~\eqref{eq:energy}, and is denoted here by $P_l(I)$, a polynomial of power $l$.
We arrange the contributions to the theory error, (\ref{eq:theoryerror}) as the $p$ columns of a $m\times p$ matrix $\sigma_{th}$, where $p=k_{max}-k$ is the number of omitted terms.
Each column in this matrix then corresponds to the theory-error structure, while each row corresponds to a different energy level. To obtain $\Sigma_{th}$ we then again take the outer product of $\sigma_{th}$ with itself, i.e., we construct an outer product in our $m$-dimensional data space, while also taking an inner product in order space.
This results in the theory error associated with different orders being added in quadrature, while maintaining the correlation structure of the theory error across the data space.
\subsection{Building the Priors}
\label{priors}
The prior distributions for an order-$n$ LEC is taken to be a Gaussian with mean zero and standard deviation
\begin{equation}
\sigma_n= \left\{ \begin{array}{lc}
A_{\rm rot} \bar{c}_{even} Q^{n-1} & \mbox{if $n$ is even;}\\
A_{\rm rot} \bar{c}_{odd} Q^{n-1}
& \mbox {if $n$ is odd.} \end{array} \right.
\label{eq:sdeviation}
\end{equation}
encoding the EFT expectations for the sizes of the LECs arising from the power counting described in Sec.~\ref{sec:rotorfermionEFT}. The standard deviation in Eq.~\eqref{eq:sdeviation} allows the possibility for even and odd LECs to have different typical sizes. Combining the Gaussian priors for the LECs yields
\begin{equation}
\begin{aligned}
\pr({\bf a}_k|&Q,\bar{c}_{even},\bar{c}_{odd},P*)\\=&\frac{1}{\bar{E} \sqrt{2\pi}} \exp\left(-\frac{E_k^2}{2 \bar{E}^2}\right)
\prod_{n=1}^k \frac{1}{\sigma_n \sqrt{2\pi}}{\rm exp}\left(-\frac{a_n^2}{2\sigma_n^2}\right).
\end{aligned}
\end{equation}
The LEC $E_K$ is just an energy shift and its size is not determined by the EFT power counting. We set the prior on it to be Gaussian with mean zero and a standard deviation, $\bar{E}$, that is wide enough to capture its value. The value for $\bar{E}$ is determined from the energy of the bandhead and $A_{\rm rot}$ by means of Eq.~(\ref{Energy_NLO}).
We choose not to impose any expectations regarding the size of expansion parameter in the prior for $Q$ and so take it to be flat between two limits:
\begin{equation}
\pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)\propto
\begin{cases}
1 & Q \in (0,Q_{cut}) \\
0 & {\rm otherwise}.
\end{cases}
\end{equation}
Limiting $Q$ from above restricts the sampler from going to high values of $Q$, as they make the covariance matrix ill-conditioned and harder to invert. For all cases we check that the posterior for $Q$ is confined to values well below $Q_{cut}$.
The priors on the characteristic sizes $\bar{c}_{even}$ and $\bar{c}_{odd}$, are taken to be identical scaled-inverse-$\chi^2$ distributions
\begin{equation}
\pr(\bar{c}_l^2|P*) \propto
\begin{cases}
\chi^{-2}(\nu=1,\tau^2=1) & \bar{c}_l^2 \in (0, \bar{c}_{cut}^2) \\
0 & {\rm otherwise},
\end{cases}
\end{equation}
where the cutoff $\bar{c}_{cut}$ prevents numerical issues inverting the covariance matrix. The scaled-inverse-$\chi^2$ distribution, given by
\begin{equation}
\chi^{-2}(x; \nu, \tau^2)=
\frac{(\tau^2\nu/2)^{\nu/2}}{\Gamma(\nu/2)}~
\frac{\exp\left[-\frac{\nu \tau^2}{2 x}\right]}{x^{1+\nu/2}},
\end{equation}
is shown for different values of $\nu$ and $\tau$ in Fig.~\ref{fig:prior_c_bars}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{prior_c_bars.pdf}
\caption{Prior distribution of the size of the dimensionless natural coefficients, $\bar{c}$.}
\label{fig:prior_c_bars}
\end{figure}
We stress that we chose identical priors for $\bar{c}_{even}$ and $\bar{c}_{odd}$ even though we expect the former to be larger than the latter based on previous analyses of data on rotational bands~\cite{Alnamlah:2020}. We did not want to bias our analysis by imposing this hierarchy on the prior, instead anticipating that it will emerge naturally in the posteriors for those parameters.
The scaled-inverse-$\chi^2$ favors small values of $\bar{c}^2$ and has long tails. This allows the sampler to explore higher values of $\bar{c}^2$. The sharp decrease in this distribution for very small values of $\bar{c}^2$ could be a problem for cases where $\bar{c}_{odd}$ is much smaller than one. This is a concern in some $K=3/2$ bands where we expect smaller odd-order corrections to the leading-order energy than in $K=1/2$ bands.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{corner.pdf}
\caption{Corner plot for the marginalized distributions of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for $^{167}$Er including all adopted rotational levels ($I_{max}=16.5$) and accounting for 6 omitted terms in the theory error. The insert in the top right corner show the correlations between posterior parameters. The order of the parameters on the corner plot is the same on the correlations plot. (Here $E_K$ and all the EFT LECs are expressed in keV. The error-model parameters are dimensionless.)}
\label{fig:corner}
\end{figure*}
\section{Running the Sampler}
\label{sec:sampler}
To sample the posterior distribution in Eq.~\eqref{posterior} we use the Python ensemble sampling toolkit for affine-invariant MCMC (emcee) \cite{emcee}. We run the sampler for each nucleus at a certain EFT order using the $m$ rotational levels from the bandhead up to some $I_{max}$ and accounting for $p$ omitted terms in the theory error. We use 64 walkers to sample the posterior distribution for an initial 10000 steps. We then continue running the sampler with 3000 step increments. After every 3000 steps we calculate the autocorrelation time, $\tau_\alpha$, where $\alpha$ indexes an LEC or an error-model parameter. We declare the sampler to be converged if the sampler meets two criteria. First, the number of steps has to be more than 50 times the highest $\tau_\alpha$.
Second, the change in any of the $\tau_\alpha$'s has to be less than 2\% from its value after the last 3000 step increment.
To get the posterior distributions we discard $2\times {\rm max}(\tau_\alpha)$ steps from the beginning of the chain ({\it burn-in}) and $0.5\times {\rm min}(\tau_\alpha)$ steps in between steps we accept ({\it thinning}).
A sample corner plot of the marginalized distributions of the LECs and the error-model parameters $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$, for the case of ${}^{167}$Er is shown in Fig.~\ref{fig:corner}. This figure clearly shows that the posterior distributions for all parameters are fully converged. For this particular case we set $Q_{cut}=0.16$ and $\bar{c}_{cut}=22$ for both $\bar{c}_{even}$ and $\bar{c}_{odd}$. As explained in Sec.~\ref{priors}, the cutoffs on $Q$ and the characteristic sizes prevent the covariance matrix from being ill-conditioned. We also ran the sampler for $^{167}$Er at different values of $Q_{cut}$ and $\bar{c}_{cut}$ and found that different choices of these hyperparameters do not result in a significant change in the posterior distributions.
For some cases, namely $^{99}$Tc and $^{183}$W, the posterior distribution of $Q$ was initially at the upper limit of the prior. We then ran into numerical problems when increasing $Q_{cut}$ trying to encompass the entire posterior. This problem was solved by decreasing the number of levels included in the analysis, i.e., decreasing $I_{max}$. It was then possible to increase $Q_{cut}$ without encountering problems with degenerate matrices. This means that for $^{99}$Tc we were only able to extract the LECs and $Q$ at $I_{max}=11.5$. For $^{183}$W we needed to remove two levels from the upper end of the data set for the sampler to be numerically stable.
In Fig.~\ref{fig:corner} we see clear correlations between $E_K$, $A$, and $B$ and also between $A_1$ and $B_1$. (Here we have dropped the subscript $K$ on $A$ and $B$; it is to be understood that all LECs are band dependent.) The correlation coefficients given in the inset in the top-right corner of the figure make the block-diagonal structure of the covariance matrix clear. To a good approximation the correlation matrix can be decomposed into a correlation matrix for even-order LECs, one for odd-order LECs, and one for the error-model parameters.
We note that, as expected, $\bar{c}_{odd}$ is smaller than $\bar{c}_{even}$. Corrections to the energy levels carrying odd powers of $I$ are smaller than those carrying even powers of $I$. This size difference is connected to different physics correcting the effective Lagrangian at even and odd orders.
To see which of the parameters has the narrowest distribution and therefore places the strongest constraint on the posterior distribution, we did a Singular Value Decomposition (SVD) of the Hessian matrix. We found that the eigenvector with the highest eigenvalue, i.e., the parameter combination with smallest absolute error, is made up mostly of the highest-order LEC. This is unsurprising, since that LEC, $B$, is markedly smaller than the others (we note that its relative error is actually larger than that on, e.g., $A_1$).
We initially found a peculiar correlation between LECs in some cases where the rotational band was built on the ground state of the nucleus we were looking at. There we found the eigenvector with the highest eigenvalue was a very particular linear combination that involved all the LECs. We ultimately traced this correlation to the fact that the ground state experimental error had been set to zero, and so the combination of LECs that entered the formula for the ground-state energy was very well constrained (theory error is also very small there). This problem was solved by adding a small experimental error to the ground state. We chose it to be equal to the error that the NNDC quotes on the energy of the first excited state.
\section{Results}
\label{sec:results}
\begin{figure}[b!]
\centering
\includegraphics[width=\columnwidth]{wes_A_1169Er.pdf}
\caption{Posteriors for $A_1$ describing $^{169}$Er as a function of $I_{max}$ at different EFT orders. The solid line connects the median values and the error bands encompass the 16th and 84th percentiles of the marginalized distribution.}
\label{fig:wes_A1_above_orders}
\end{figure}
In this section we show results for our Bayesian analysis of the rotational energy levels in $^{99}$Tc, ${}^{155,157}$Gd, ${}^{159}$Dy, ${}^{167, 169}$Er, ${}^{167, 169}$Tm, ${}^{183}$W, ${}^{235}$U and ${}^{239}$Pu. The experimental data are taken from the National Nuclear Data Center (NNDC)~\cite{Baglin:2016vll,Browne:2017uto,Nica:2016elr,Nica:2019ykt,Reich:2012ouk,Browne:2014ukl,Browne:2014gwf,Baglin:2000ong,Baglin:2008hsa}. Except for the cases of ${}^{99}$Tc and ${}^{183}$W noted above, we included all levels in a certain rotational band according to the adopted level determination in the NNDC.
\subsection{Stable LEC Extraction Across EFT Orders and Additional Data}
In this subsection we show that lower-order LECs extracted for the selected rotational bands are stable across EFT orders and with the addition of high-energy data, provided that we account for enough omitted terms when treating the theory error. For most nuclei including omitted terms up to $k_{max}=10$, i.e., accounting for six omitted terms at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace, was enough to stabilize the extraction of the LECs.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_B_1239Pu.pdf}
\includegraphics[width=\columnwidth]{wes_B239Pu.pdf}
\caption{Posteriors for $B_1$ and $B$ describing $^{239}$Pu a function of $I_{max}$ for different values of $k_{max}$. The solid line connects the median values and the error bands encompass the 16th and 84th percentiles in the marginalized distribution.}
\label{fig:wes_B_above_errors}
\end{figure}
As an example, we show the stability of the extracted LEC, $A_1$, across number of levels included at different EFT orders in Fig.~\ref{fig:wes_A1_above_orders}. In this figure, $I_{max}$ is the spin of the highest-energy level included in a particular analysis. The central values of the resulting posteriors are consistent with each other within 68\% credible intervals, shown as error bars in the figure. Adding more levels to the analysis narrows the posteriors for the LECs up to a certain $I_{max}$, after which the widths of these distributions saturate.
Fig.~\ref{fig:wes_A1_above_orders} also demonstrates striking agreement between the distributions obtained at low and high EFT orders: they are almost identical as long as omitted terms up to the same $k_{max}$ are accounted for in both analyses.
The importance of including more than one omitted term in the theory error estimate is evident in Fig.~\ref{fig:wes_B_above_errors}.
The top and bottom panels of the figure show the way that posteriors for $B_1$ and $B$ evolve as $I_{max}$ increases. This is done using three error models that include different numbers of omitted terms. These results show that including more omitted terms in the model of the theory error removes the drifting and staggering of the central values.
For both cases the distributions at $k_{max}=10$ agree within errors as we go higher in $I_{max}$. The narrowing of the distribution as we go higher in $I_{max}$ is clearly seen in those two figures. In addition to having less data, the broadening of the error bands at low $I_{max}$ comes from the fact that including less levels in the analysis leads to highly correlated LECs. This allows the numerically larger errors on the lower-order LECs to contribute to the errors on the higher-order LECs, thereby enhancing them.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_corr.pdf}
\caption{Correlations between LECs and error-model parameters as a function of $I_{max}$, resulting from the analysis on the lowest $K=1/2$ rotational band in $^{239}$Pu at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace with $k_{max}=10$.}
\label{fig:wes_corr}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_EK_below.pdf}
\caption{The distribution of $E_K$ for $^{169}$Er at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ as we successively remove the lowest energy levels from the data set $D$. The solid blue line connects the median values and the error bands encompass uncertainties between on the 16th and 84th percentiles of the samples in the marginalized distribution. The solid black lines show to size of the standard deviation set with the Gaussian prior on $E_K$.}
\label{fig:wes_EK_below}
\end{figure}
In Fig.~\ref{fig:wes_corr} we show the decrease in the correlations between the LECs as $I_{max}$ increases. The high correlation between the LECs at low $I_{max}$ occurs because these analyses do not include enough data to constrain all LECs independently. Furthermore, the high correlation between the LECs at low $I_{max}$ also results in an unreliable extraction of the expansion parameter $Q$. This comes from the fact that at low energies the theory truncation error is very small compared to the experimental error. Indeed, adding more terms to our EFT error model (i. e., increasing $k_{max}$) leads to higher correlation between the LECs at low $I_{max}$. Thus, the number of levels required to reliably extract $Q$ increases with increasing $k_{max}$.
Starting instead at the low-$I$ end of the data: when we progressively remove the lowest-energy levels from the data set $D$ used to construct the likelihood we rapidly lose the ability to reliably extract the LECs. Figure \ref{fig:wes_EK_below} shows that the distribution for $E_K$ starts narrow and broadens as we remove levels from below. When we remove the six lowest energy levels the distribution of $E_K$ is exactly the same as the prior distribution: the likelihood is making no contribution to the $E_K$ posterior.
The previous results were nearly the same for all cases considered in this work. However, even for $k_{max}=10$, staggering and shifting of the LECs remains sizable for the $K=1/2$ bands in $^{183}$W, $^{167}$Tm and $^{235}$U. In $^{183}$W and $^{167}$Tm, these effects could be attributed to large expansion parameters, as they translate to large omitted contributions to the energies of the rotational levels. In $^{235}$U, the fermionic matrix elements could be larger than naively expected, causing the systematic expansion of the EFT to be questionable as discussed in Ref.~\cite{Alnamlah:2020}.
For $K=3/2$ bands, we were able to extract stable LECs from the $^{159}$Dy analysis by setting $I_{max}=15.5$. This extraction required us to consider omitted terms up to $k_{max}=12$. This is because the spin at which the EFT breaks in this nucleus is $I_{br}\approx 15.5$. (This, then, is the third case in which we do not use all the NNDC energy-level data available on a particular band.) $^{157}$Gd is stable across orders and $I_{max}$, while $^{155}$Gd exhibits shifting and staggering due to a larger expansion parameter, $Q \approx 0.07$.
\begin{table*}[t!]
\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
Nucleus&$E_K$ [keV]&$A_1$ [keV]&$ A$ [keV]&$(B_1,A_3)$ [keV]&$B$ [keV]&$Q$&$\bar{c}_{\rm even}$&$\bar{c}_{\rm odd}$\\ \hline
$^{99}$Tc&$147.3111_{-0.0142}^{0.0143}$&$70.1912_{-0.0103}^{0.0103}$&$82.2577_{-0.0086}^{0.0086}$&$-7.5851_{-0.0022}^{0.0022}$&$-2.99165_{-0.00079}^{0.00079}$&$0.204_{-0.016}^{0.021}$&$1.87_{-0.53}^{0.43}$&$1.35_{-0.39}^{0.54}$\\ \hline
$^{155}$Gd&$-45.2063_{-0.0025}^{0.0025}$&-&$12.0184_{-0.0009}^{0.0009}$&$-0.0081_{-0.0001}^{0.0001}$&$0.00571_{-0.00008}^{0.00008}$&$0.067_{-0.006}^{0.007}$&$3.54_{-1.11}^{2.00}$&$2.01_{-0.94}^{1.99}$\\ \hline
$^{157}$Gd&$-41.3241_{-0.0097}^{0.0093}$&-&$11.0240_{-0.0032}^{0.0034}$&$-0.0094_{-0.0002}^{0.0002}$&$-0.00500_{-0.00025}^{0.00024}$&$0.057_{-0.006}^{0.006}$&$5.23_{-2.33}^{5.18}$&$0.55_{-0.15}^{0.28}$\\ \hline
$^{159}$Dy&$-42.7714_{-0.0093}^{0.0092}$&-&$11.4076_{-0.0030}^{0.0030}$&$-0.0060_{-0.0003}^{0.0003}$&$-0.00307_{-0.00020}^{0.00020}$&$0.063_{-0.004}^{0.004}$&$3.52_{-1.16}^{2.22}$&$0.95_{-0.33}^{0.64}$\\ \hline
$^{167}$Er&$207.2088_{-0.0053}^{0.0053}$&$7.8383_{-0.0028}^{0.0027}$&$11.2398_{-0.0016}^{0.0016}$&$-0.0063_{-0.0003}^{0.0003}$&$-0.00820_{-0.00011}^{0.00010}$&$0.047_{-0.004}^{0.005}$&$4.04_{-1.44}^{2.63}$&$0.83_{-0.26}^{0.49}$\\ \hline
$^{169}$Er&$0.9673_{-0.0197}^{0.0195}$&$9.7774_{-0.0100}^{0.0102}$&$11.7625_{-0.0053}^{0.0058}$&$-0.0064_{-0.0009}^{0.0008}$&$-0.00309_{-0.00023}^{0.00016}$&$0.031_{-0.005}^{0.006}$&$5.57_{-2.51}^{5.31}$&$0.77_{-0.23}^{0.42}$\\ \hline
$^{167}$Tm&$-18.4635_{-0.0155}^{0.0156}$&$-9.1267_{-0.0057}^{0.0056}$&$12.5095_{-0.0022}^{0.0022}$&$0.0431_{-0.0003}^{0.0003}$&$-0.00906_{-0.00008}^{0.00008}$&$0.056_{-0.005}^{0.005}$&$2.16_{-0.76}^{1.37}$&$1.06_{-0.34}^{0.62}$\\ \hline
$^{169}$Tm&$-19.0532_{-0.0011}^{0.0011}$&$-9.7204_{-0.0008}^{0.0008}$&$12.4738_{-0.0007}^{0.0006}$&$0.0264_{-0.0002}^{0.0002}$&$-0.00497_{-0.00006}^{0.00007}$&$0.045_{-0.007}^{0.008}$&$2.59_{-1.12}^{2.30}$&$0.94_{-0.30}^{0.60}$\\ \hline
$^{183}$W&$-6.8750_{-0.0008}^{0.0008}$&$2.7512_{-0.0005}^{0.0005}$&$12.7754_{-0.0004}^{0.0004}$&$-0.0436_{-0.0001}^{0.0001}$&$0.01977_{-0.00003}^{0.00003}$&$0.076_{-0.006}^{0.007}$&$3.80_{-1.49}^{2.86}$&$0.67_{-0.20}^{0.35}$\\ \hline
$^{235}$U&$-6.1882_{-0.0008}^{0.0008}$&$-1.7298_{-0.0007}^{0.0007}$&$6.0508_{-0.0002}^{0.0002}$&$0.0025_{-0.0001}^{0.0001}$&$-0.00249_{-0.00001}^{0.00001}$&$0.040_{-0.003}^{0.004}$&$3.85_{-1.28}^{2.36}$&$0.82_{-0.27}^{0.53}$\\ \hline
$^{239}$Pu&$-8.3578_{-0.0019}^{0.0019}$&$-3.6559_{-0.0011}^{0.0011}$&$6.2726_{-0.0004}^{0.0004}$&$0.0041_{-0.0001}^{0.0001}$&$-0.00149_{-0.00003}^{0.00003}$&$0.029_{-0.003}^{0.004}$&$5.72_{-2.22}^{4.18}$&$0.93_{-0.31}^{0.61}$\\ \hline
\end{tabular}
\renewcommand{\arraystretch}{1}
\caption{The median value of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for the nuclei considered in this work. The uncertainties encompass the 16th and 84th percentiles of the samples in the marginalized distributions. $K=3/2$ rotational bands do not have a parameter $A_1$ and the parameters ($B_1,A_3$) refer to $K=1/2$ and $K=3/2$ bands respectively.}
\label{tab:LECs}
\end{table*}
The values of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for the nuclei considered in this work are given in Tabele \ref{tab:LECs}.
\subsection{Prior Sensitivity}
In addition to using the scaled-inverse-$\chi^2$ distribution as a prior for $\bar{c}_{even}$ and $\bar{c}_{odd}$ we tried truncated Gaussians with mean zero and standard deviations $\sigma=7$ and $\sigma=3$ respectively for all cases. These truncated Gaussian priors allow for smaller values of the characteristic sizes.
But the standard deviations were chosen to still allow values for $\bar{c}_{odd}$ and $\bar{c}_{even}$ larger than those resulting from scaled-inverse-$\chi^2$ priors with $\nu=1$ and $\tau^2=1$.
The change in prior for $\bar{c}_{odd}$ and $\bar{c}_{even}$ does not significantly change the posteriors for the LECs: the corresponding central values differ by less than 1\%, and are consistent with each other within the 68\% credible intervals.
Central values of the posteriors for $Q$ differ by less than 15\%, and were similarly consistent.
The strongest dependence on the prior is that exhibited by the posteriors for $\bar{c}_{odd}$ and $\bar{c}_{even}$: the central values differ in some cases by more than 50\%. However, even these values are consistent with each other within 68\% credible intervals, since the posteriors for the characteristic sizes are broad.
The changes in the posteriors of $Q$ on one hand and $\bar{c}_{odd}$ \& $\bar{c}_{even}$ on the other are anticorrelated. We only care about combinations of them to set the size of the theory error and the expected size of the LECs. Thus, the dependence of the theory error and the expected size of the LECs on the prior for the characteristic sizes is less profound. The difference in sizes of the theory error resulting from the chosen priors is less than 20\% for all cases except $^{157}$Gd, where the difference is about 40\%.
For all results that follow we used the scaled-inverse-$\chi^2$ distribution with $\nu=1$ and $\tau^2=1$ as the prior for both $\bar{c}_{odd}$ and $\bar{c}_{even}$, in keeping with the naturalness assumption.
\subsection{Posterior Predictive Distributions}
Figure \ref{fig:Er_ppd,U_ppd} shows the PPDs of the energy residuals as a function of the spin $I$ for two cases considered in this work. These distributions are calculated using Eq.~(\ref{PPD}).
In each figure, translucent blue lines connect energy residuals resulting from different LECs sets sampled from the posterior distribution in Eq.~\eqref{eq:master}. The solid black line represents the median of the PPD, and the dashed lines encompass the region between the 16th and 84th percentiles.
The dark and light red bands show the truncation error and the experimental error added in quadrature at 68\% and 95\% credible levels respectively. To calculate the truncation error, we consider a theory error that accounts for six omitted terms. They are added in quadrature, just as they are in the likelihood defined in Sec.~\ref{sec:Bayesianmodel}. This calculation was done using the median values of $Q, \bar{c}_{even}$, and $\bar{c}_{odd}$. The dependence of the size of the theory error on the prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$ is small in these cases: the theory error changes by about 10\% when the prior is changed.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{169Er_ppd.pdf}
\includegraphics[width=\columnwidth]{235U_ppd.pdf}
\caption{The posterior predictive distribution for energy-level residuals at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ in $^{169}$Er and $I_{max}=17.5$ (top panel) and in $^{235}$U and $I_{max}=23.5$ (bottom panel). The dark and light red bands show the truncation error plus the experimental error at 68\% and 95\% credible levels respectively. The lighter blue lines connect the energy residuals calculated from the distribution of the LECs. The solid black line represents the median of the distribution and the dashed lines indicate the 16th and 84th percentiles. The correlation shown on the plot is the highest correlation between any LEC and any error-model parameter. $I_{br}$ was determined from the distribution of $Q$. The dashed purple line shows the lower limit of $I_{br}$. The insert on the plot shows the residuals on the first 5 levels with an altered y-axis scale.}
\label{fig:Er_ppd,U_ppd}
\end{figure}
The correlation coefficient written in the legend in Fig. \ref{fig:Er_ppd,U_ppd} is the largest between any LEC and any error-model parameter for the shown analysis. When this value is small, the truncation error and the propagated LEC error could in principle be added together in quadrature.
In viewing Fig.~\ref{fig:Er_ppd,U_ppd} it is important to remember that
the truncation error on the energy residuals is highly correlated across levels. This comes from the high correlation between levels when building the correlation matrix that goes into the likelihood. This correlation also flows into a correlation between levels in the PPD of the energies. A correlation plot between two energy levels, like the ones in Fig. \ref{fig:Er_2d_ppd,U_2d_ppd}, gives a 2D cut of this multi-dimensional correlation.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{2d_169Er.pdf}
\includegraphics[width=\columnwidth]{2d_235U.pdf}
\caption{A 2D cut of the posterior predictive distribution at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ for $^{169}$Er and $I_{max}=17.5$ (top panel) and in $^{235}$U and $I_{max}=23.5$ (bottom panel). The blue dots show the energies calculated from the distribution of the LECs. The black cross shows the experimental value and the black lines and black ellipse shows the corresponding experimental uncertainty. The remaining ellipses and lines show the truncation error and the experimental error added in quadrature. (All the ellipses are centered at the experimental value.) The orange, red and green account for 1, 2 and 6 omitted terms in the theory error respectively. (In the top panel the red ellipse is completely covered by the green ellipse.)}
\label{fig:Er_2d_ppd,U_2d_ppd}
\end{figure}
In both panels we see the importance of accounting for more than one omitted term in the theory error. This is clearly shown in the reverse in the direction of the correlation from a negative to a positive correlation when going from the orange ellipse to the red ellipse. The orange ellipse is obtained when we account for only one omitted term, while the red ellipse includes the effect of two omitted terms. After accounting for six omitted terms the green ellipse is obtained and the 68\% ellipse in principle expands. This is more clearly seen when we go to high-energy levels plotted in the lower panel in Fig. \ref{fig:Er_2d_ppd,U_2d_ppd}. Note also that for lower-energy levels the correlation is smaller since the experimental error dominates over the truncation error, and we assumed that the experimental errors are not correlated across energy levels.
\subsection{Model Checking}
In Figs. \ref{fig:A_vs_Q}, \ref{fig:B1_vs_Q} and \ref{fig:B_vs_Q} we compare the marginalized posterior distributions of the LECs, on the y-axis, with their expected sizes from the EFT power counting, on the x-axis. Since we also extract the theory error parameters from the sampler and they are highly correlated among themselves, we calculate the expected size from the distributions of the error model parameters using Eq. (\ref{PPD}). We notice that the error on the distribution of the LECs is very small compared to the error on the expected sizes that comes from the distribution of the theory error parameters.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{A_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^2{\rm LO}}\xspace LEC, $\Delta A$, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The blue colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands. }
\label{fig:A_vs_Q}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{B1_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^3{\rm LO}}\xspace LEC, $B_1$ for $K=1/2$ bands and $A_3$ for $K=3/2$ bands, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass uncertainties between on the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The yellow colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:B1_vs_Q}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{B_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^4{\rm LO}}\xspace LEC, $B$, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass uncertainties between on the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The yellow colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:B_vs_Q}
\end{figure}
As these graphs are model-checking graphs, and since the estimates of LEC sizes plotted on the x-axis are meant as order-of-magnitude estimates, we do not expect perfect linear correlations.
Nevertheless, Fig.~\ref{fig:A_vs_Q} shows that, for all $K=1/2$ bands considered, the LEC $\Delta A$ agrees with its expected size within error bands. This result is surprisingly better than expected. In contrast, the size of $\Delta A$ for $K=3/2$ bands is larger than expected, especially for $^{155}$Gd (see blue points in Fig.~\ref{fig:A_vs_Q}). There are two factors that could contribute to this. First, the $K=3/2$ bands have larger fermionic matrix elements. This could hinder the systematic expansion of the EFT. Second, the $K=3/2$ bands have relatively larger expansion parameters, see Fig. \ref{fig:Q_vs_Q}.
The same discussion applies to the results in Figs. \ref{fig:B1_vs_Q} and \ref{fig:B_vs_Q}, where we see good agreement between the LECs and their expected sizes for $K=1/2$ bands. The disagreement with power-counting estimates for $K=3/2$ bands at \NkLO{3,4} is less of a concern than the one at \NkLO{2} seen in Fig.~\ref{fig:A_vs_Q}, since these higher-order LECs are smaller than their expected sizes. This doesn't undermine the convergence of the EFT expansion.
We also note here that the scale of the x-axis is prior dependent and could change by more than 50\% in some nuclei, depending on the choice of prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$.
For $^{157}$Gd changing the prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$ to a truncated normal allowed for smaller values of $\bar{c}_{odd}$ and $A_3$ was then equal to the expected size (i. e.,the point for $^{157}$Gd then falls exactly on the line in figure \ref{fig:B1_vs_Q}). This did not happen when the truncated normal is chosen as a prior for the analysis in $^{155}$Gd and $^{159}$Dy; this may occur because there is strong \NkLO{5} energy-level staggering present in the data for these nuclei.
The size of $\bar{c}_{even}$ and $\bar{c}_{odd}$ is constrained by both the sizes of the LECs and the size of the theory error. In a good systematic expansion the tension between those factors on setting the size of the $\bar{c}$'s would be small and one number apiece would suffice to represent the even and odd order corrections. However when the systematic expansion is hindered, as in the case for $K=3/2$ bands due to large fermionic matrix elements, this tension becomes clear. One example of this is seen in Figs. \ref{fig:A_vs_Q} and \ref{fig:B_vs_Q} for $K=3/2$ bands. There $\Delta A$ is large and favors large values of $\bar{c}_{even}$, however, $B$ is small and favors smaller values of $\bar{c}_{even}$. The eventual result is a compromise. This tension may be exacerbated by the truncation error also providing information on the size of the $\bar{c}$'s.
\subsection{Higher than Expected Break-down Scale}
In Fig. \ref{fig:Q_vs_Q} we see a clear correlation between the extracted values of $Q$ and those that are expected based on each nucleus' single-particle and vibrational energy scales, $E_{\rm sp}$ and $E_{\rm vib}$. The expected $Q$ is the larger of $E_{\rm rot}/E_{\rm sp}$ and $E_{\rm rot}/E_{\rm vib}$, while the extracted $Q$ comes from sampling the posterior in Eq. \ref{posterior}. This extracted $Q$ is what actually determines the convergence of the EFT expansion. It is markedly smaller than would be naively expected. The break-down scale of the theory is thus higher than naively expected: our rotational EFT works to much higher $I$ than energy-scale arguments would suggest. This could occur because coupling between the higher rotational states explicitly included in the EFT and the high-energy states that are not explicitly included in our EFT is hindered by the large difference in angular momentum between them.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{Q_vs_Q.pdf}
\caption{The extracted expansion parameter $Q$ from marginalized posterior distribution of our sampling compared to its naively expected size. The expected size is taken to be the maximum of $E_{\rm rot}/E_{\rm sp}$ and $E_{\rm rot}/E_{\rm vib}$. The dashed black line shows the best linear fit and its parameters are printed on the plot. The blue colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:Q_vs_Q}
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
We performed a Bayesian analysis to extract the LECs and expansion parameters $Q$ describing the rotational energy levels of diverse odd-mass nuclei within a recently developed EFT.
This analysis corroborates the EFT organization for energy-level formulae which results from the assumed power-counting scheme: the extracted LECs of order $k$ scale as $Q^{k-1}$, i.e., according to EFT expectations.
While our analysis reached this conclusion for both $K=1/2$ and $K=3/2$ rotational bands, the sizes of the LECs describing the latter exhibit larger deviations from their expected values than those describing the former. We attribute this behavior to the size of fermionic matrix elements, assumed to be of order one while organizing energy-level formulae. Since these matrix elements involve the angular momentum of the fermion, $\vec{j}$, we cannot exclude the possibility that the systematic behavior of the EFT is hindered in bands build on top of single-particle orbitals with larger values of $K$. For the $K=3/2$ bands studied in this work, however, this discrepancy does not destroy the systematic improvement of calculated energies up to \ensuremath{{\rm N}{}^4{\rm LO}}\xspace, as the sizes of extracted LECs are smaller that expected.
In order to ensure that the extracted values are independent of the EFT order and number of energy levels entering the analysis, we employed a theory error beyond the first-omitted-term approximation, considering omitted terms in the expansion for the energy of rotational levels up to order $k_{max}$. As we increased the number of omitted terms considered in the theory error, the corresponding log likelihood exhibited steeper and steeper peaks. Therefore, the `widths' of the sampled posteriors decrease as $k_{max}$ increases. Considering six omitted terms at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace enabled a stable extraction of the LECs and expansion parameters describing the levels of interest. The shapes of posteriors for low-order LECs extracted at this order and those extracted using lower-order energy formulae are, for all practical purposes, identical. On the other hand, the shapes of the posteriors depend strongly on the number of levels informing the model, narrowing as more levels are included. Nevertheless, the 68\% credible intervals of these posteriors possess significant overlap, facilitating reliable LEC extraction.
In addition to the posteriors for the LECs and expansion parameters, our analysis yielded distributions for the characteristic sizes of even and odd $c_n$'s, $\bar{c}_{even}$ and $\bar{c}_{odd}$. The values of $\bar{c}_{odd}$ are typically smaller than those for $\bar{c}_{even}$, in agreement with results from previous studies where the LECs were fitted to the smallest possible data sets. The difference of the characteristic sizes of even and odd LECs has its origin in the physics behind the corresponding contributions to the effective Lagrangian: while odd-order contributions correct the particle-rotor interaction, even-order contributions include terms that depend exclusively on the rotor degrees of freedom, thus correcting the physics of the core to which the particle is coupled. Here this conclusion was reached solely on the basis of experimental data; we assumed equal priors for both characteristic sizes.
Although the distributions of $\bar{c}_{odd}$ and $\bar{c}_{even}$ change depending on the choice of the their priors, that does not significantly change the distributions of the LECs. Altering the priors also does not have a large effect on the size of the theory error, which changes by less than 20\% for nearly all cases.
These considerations mean that our extractions of the LECs and the theory error parameters in the EFT of rotational bands in odd-mass nuclei are robust under the choice of prior. The formalism presented here also gives robust results for LECs across orders and as more data is added to the analysis. We conclude that a Bayesian framework that incorporates theory errors in the likelihood offers significant advantages for LEC extraction in EFTs. This methodology has already been used for the extraction of LECs in the NN potential from phase shifts~\cite{Wesolowski:2019} and to constrain parameters of the three-nucleon force~\cite{Wesolowski:2021}. But it is a very general approach which should improve the parameter estimation for LECs in any EFT.
\begin{acknowledgments}
We thank Dick Furnstahl and Jordan Melendez for useful discussions. We are also grateful for Daniel Odell's significant conceptual and computational assistance.
This work was supported by the US Department of Energy, contract DE-FG02-93ER40756 (IKA, DRP) by the National Science Foundation CSSI program Award OAC-2004601 (DRP), and under the auspices of the US Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 (EACP). IKA acknowledges the support of King Saud University and the Ministry of Education in Saudi Arabia.
\end{acknowledgments}
\section{Introduction}
Rotational bands are ubiquitous in the spectra of medium-mass and heavy nuclei. As has been known for seventy years~\cite{Bohr:1951zz}, they emerge in a description of the nucleus as a nearly rigid axially-symmetric rotor~\cite{Rowe}. For even-even nuclei the simplest rotational bands consist of $0^+$, $2^+$, \ldots states and their energies are described by an expansion in powers of $I(I+1)$~\cite{Papenbrock:2010,CoelloPerez:2015}. This behavior has recently been obtained in {\it ab initio} calculations of the Be isotope chain~\cite{Caprio:2013,Maris:2014,Jansen:2015,Caprio:2019yxh,McCoy:2020xhp} and $^{34}$Mg~\cite{Hagen:2022tqp}.
Odd-mass neighbors of a rotor nucleus can then be understood as a fermion coupled to the rotor. The fermion dynamics is simpler in the intrinsic frame in which the nucleus is not rotating, but this frame is non-inertial, so solving the problem there induces a Coriolis force proportional to $\vec{j} \cdot \vec{I}$, the dot product of the single-fermion angular momentum and the total angular momentum of the fermion-rotor system. When combined with other mechanisms, such as excitation of the fermion to higher-single particle states and the fermion disturbing the rotor, this induces a string of terms in the energy-level formula~\cite{BohrMottelson}. Odd powers of $I$ appear, and produce staggering between adjacent levels. Which powers of $I$ are present depends on the value of the quantum number, $K$, the projection of the fermion angular momentum on the rotor axis. For $K=1/2$ bands the energy-level formula is:
\begin{eqnarray}
&& E(I)=A_K I(I+1) + E_K
+ A_{1} (-1)^{I+1/2} (I+1/2) \nonumber\\
&& \quad + B_{1} I(I+1) (-1)^{I+1/2} (I+\tfrac{1}{2})
+ B_K [I(I+1)]^2
\label{eq:energies}
\end{eqnarray}
where $A_K$, $E_K$, $A_1$, $B_1$, and $B_K$ are parameters, related to rotor properties and single-particle matrix elements, that need to be either derived from a microscopic model or estimated from data.
Over the years a number of models have had success describing this pattern from underlying density functional theory~\cite{Dudek:1981,Cwiok:1980,Afanasjev:2013,Zhang:2020} or shell-model~\cite{Inglis:1954,Velazquez:1999,Liu:2004,Zhang:2020} dynamics. The models also predict specific values for the coefficients that appear in Eq.~(\ref{eq:energies}).
In Ref.~\cite{Alnamlah:2020} we took a different approach, organizing the formula (\ref{eq:energies}) as an effective field theory (EFT) expansion in powers of the small parameter $Q \equiv 1/I_{br}$, with $I_{br}$ the spin of the nuclear state at which dynamical effects associated with single-particle and/or vibrational degrees of freedom cause the polynomial expansion in powers of $I$ to break down. This description of rotational bands in odd-mass nuclei builds on the successful EFT developed for even-even nuclei in Refs.~\cite{Papenbrock:2010,CoelloPerez:2015}. Other efforts to develop an EFT for these rotational bands can be found in Refs.~\cite{Papenbrock:2020zhh,Chen:2020qbf}.
In the odd-mass rotor EFT, Eq.~(\ref{eq:energies}) is the next-to-next-to-next-to-next-to leading order (\NkLO{4}) result for the energies, and the first corrections to it are $\mathcal{O}(E Q^4)$. The EFT analysis of Eq.~(\ref{eq:energies}) organizes it in terms of increasingly accurate predictions: the N$^k$LO energy-level formula has accuracy $\mathcal{O}(E Q^{k})$. All short-distance/high-energy physical mechanisms that affect the energies up to that accuracy are subsumed into the parameters or low-energy constants (LECs) that multiply the $I$-dependent terms in Eq.~(\ref{eq:energies}). In Ref.~\cite{Alnamlah:2020} we determined these LECs by fitting the lowest levels in the different rotational bands we analyzed. However, this runs the risk of fine-tuning the values of the LECs to those levels, and it does not provide uncertainty estimates for them. Better parameter estimation would use all the data available on a particular band, and account for the $\mathcal{O}(E Q^{k})$ truncation uncertainty present at order N$^k$LO~\cite{Schindler:2009,Furnstahl:2014}.
Bayesian methods for EFT parameter estimation do just that~\cite{Schindler:2009,Wesolowski:2016,Wesolowski:2019,Wesolowski:2021}. Reference~\cite{Wesolowski:2019} showed that the effect of neglected terms in the EFT expansion could be included in the error model by modifying the likelihood so that the covariance matrix that appears there includes both experimental uncertainties and EFT truncation errors. More recently, Ref.~\cite{Wesolowski:2021} showed that MCMC sampling of that likelihood enabled the simultaneous determination of the LECs and the parameters of the error model, i.e., the value of $Q$ and the typical size of the ``order 1'' dimensionless coefficients that appear in the EFT expansion.
In this work we apply the EFT parameter estimation technology developed in Refs.~\cite{Schindler:2009,Wesolowski:2016,Wesolowski:2019,Wesolowski:2021} to the problem of rotational bands in odd-mass nuclei. We consider $K=1/2$ bands in $^{99}$Tc, ${}^{167, 169}$Er, ${}^{167, 169}$Tm, ${}^{183}$W, ${}^{235}$U and ${}^{239}$Pu as well as $K=3/2$ bands in ${}^{155,157}$Gd and ${}^{159}$Dy.
Section~\ref{sec:rotorfermionEFT} summarizes the elements of the EFT that are relevant for this paper. Section~\ref{sec:Bayesianmodel} then develops the Bayesian statistical model we use to analyze data on rotational bands. We first write down the likelihood that includes both experimental and theory uncertainties, and then explain how we use known information on the expected size of the LECs and the expansion parameter to set priors. A novel feature of this work, compared to earlier Bayesian EFT parameter-estimation studies, is that our statistical model incorporates the possibility that the LECs at even and odd orders have different typical sizes. This reflects the physics of odd-order LECs that are associated with matrix elements of the fermion spin, while even-order LECs contain a combination of effects from the rotor and the fermion. Section~\ref{sec:sampler} contains details of our Markov Chain Monte Carlo sampler, and then Sec.~\ref{sec:results} presents the results for LECs and the expansion parameter, $Q$, that we obtain from sampling the Bayesian posterior. We conclude in Sec.~\ref{sec:conclusion}. All the results and figures generated from this work can be reproduced using publicly available Jupyter notebooks~\cite{notebooks}.
\section{Rotational EFT Background}
\label{sec:rotorfermionEFT}
Here we summarize the results of the EFT for rotational bands in odd-mass nuclei that was developed up to fourth order in the angular velocity of the system in Ref.~\cite{Alnamlah:2020}. This theory constructs the Lagrangian of the particle-rotor system using its angular velocity and the angular momentum of the unpaired fermion, $\vec{j}$, as building blocks. The resulting Lagrangian corrects that of a rigid rotor with contributions arranged as a series in powers of a small expansion parameter, $Q$, according to a power-counting scheme that counts powers of the system's angular velocity. Naively, we expect $Q$ to be of order $E_{\rm rot}/ E_{\rm high}$, where $E_{\rm rot}$ is the energy scale at which rotational excitation take place and $E_{\rm high}$ is the scale of high-energy physics not explicitly taken into account by the EFT.
At leading order (LO), the energy of a rotational band on top of a bandhead with spin $K$ is
\begin{equation}
\label{Energy_LO}
E_{\rm LO}(I,K) =A_{\rm rot} I(I+1) + E_K,
\end{equation}
where $I$ is the spin of the rotational state (or, equivalently, the total angular momentum of the fermion-rotor system), and $A_{\rm rot}$ and $E_K$ are LECs that must be fitted to experimental data. $A_{\rm rot}$ is determined by the moment of inertia of the even-even nucleus (the rotor) to which the unpaired fermion is coupled.
At next-to-leading order (NLO) rotational bands with $K=1/2$ are affected by a term that takes the same $\vec{j} \cdot \vec{I}$ form as the Coriolis force. This produces:
\begin{equation}
\label{Energy_NLO}
\begin{aligned}
E_{\rm NLO}(I,K) =& A_{\rm rot} I(I+1) + E_K\\
& + A_{1} (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right) \delta^K_{1/2}.
\end{aligned}
\end{equation}
The LEC $A_1$ is expected to be of order $A_{\rm rot}$ times a sum of matrix elements involving the fermion's total angular momentum operator (for details see Ref.~\cite{Alnamlah:2020}). From previous studies we see that $A_1/A_{\rm rot}<1$. This correction, sometimes called the signature term, causes staggering between adjacent states in $K=1/2$ bands.
The energy of a rotational band at next-to-next-to-leading order (\ensuremath{{\rm N}{}^2{\rm LO}}\xspace) is
\begin{equation}
\begin{aligned}
E_{\ensuremath{{\rm N}{}^2{\rm LO}}\xspace}(I,K) =& A_K I(I+1) + E_K\\
&+ A_{1} (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right) \delta^K_{1/2}.
\end{aligned}
\end{equation}
The term proportional to $A_K$ combines the LO term proportional to $A_{\rm rot}$ and corrections entering at this order with the same spin dependence. From our power counting we expect the shift $\Delta A=A_{\rm rot}-A_K$ to be of order $A_{\rm rot} Q$. In contrast to $A_{\rm rot}$, $A_K$ is band dependent and so should be fitted to data on the rotational band of interest.
The \ensuremath{{\rm N}{}^3{\rm LO}}\xspace corrections to the energy of a rotational band are both $\sim I^3$ for $I \gg 1$, but take a different form in the $K=1/2$ and $K=3/2$ bands:
\begin{equation}
\begin{aligned}
\Delta E_{\ensuremath{{\rm N}{}^3{\rm LO}}\xspace}(&I,K)\\
= B_{1}& (-1)^{I+1/2} \left(I+\tfrac{1}{2}\right)I(I+1) \delta^K_{1/2} \\
+& A_{3} (-1)^{I+3/2} \left(I+\tfrac{1}{2}\right)\left(I-\tfrac{1}{2}\right)\left(I+\tfrac{3}{2}\right) \delta^K_{3/2}.
\end{aligned}
\end{equation}
with $B_1$ and $ A_3$ expected to be of order $A_1 Q^2$.
Last, at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace we have the additional term:
\begin{equation}
\Delta E_{\ensuremath{{\rm N}{}^4{\rm LO}}\xspace}(I,K) = B_K [I(I+1)]^2.
\end{equation}
with $B_K$ expected to be of order $A_{\rm rot} Q^3$.
This pattern continues: at odd orders we add terms that correct the staggering term and have LECs of order $A_1Q^{n-1}$, while the even-order terms provide the overall trend with $I$ and have LECs of order $A_{\rm rot}Q^{n-1}$. (In both cases $n$ is the order of our expansion.) This difference in the expected sizes of odd and even LECs comes from the physics. Odd-order LECs are associated with operators in the effective Lagrangian that couple rotor and fermionic degrees of freedom, while even-order LECs encode both rotor-fermion interactions and effects coming from the non-rigidity of the rotor itself.
In what follows we denote the LECs $A_1$, $\Delta A$, $B_1$, and $B_K$ generically as $\{a_n:n=1,\ldots,k\} \equiv {\bf a}_k$, where $k$ is the order of the EFT calculation. (In the case of $K=3/2$ bands the set is $\Delta A$, $A_3$, and $B_K$, and $a_1=0$.)
We then divide the $n$th-order LEC, $a_n$, by the the reference scale and the power of the expansion parameter assigned to it by the EFT power counting, i.e., construct:
\begin{equation}
c_n=\frac{a_n}{A_{\rm rot} Q^{n-1}}.
\label{eq:cs}
\end{equation}
We expect these coefficients $c_n$ to be of order one, i.e., they should be natural coefficients. However, because sets of odd and even natural coefficients seem to have different sizes we will assume the even and odd $c_n$'s are drawn from two different distributions with different characteristic sizes that we denote by $\bar{c}_{even}$ and $\bar{c}_{odd}$.
\section{Building the Bayesian Model}
\label{sec:Bayesianmodel}
\subsection{Building the Posterior}
Our goal in this analysis is to use the information on the expected size of LECs to stablize the extraction of their values as we add more levels to the analysis, or as we use energy-level formulae computed at different EFT orders. At the same time, we want to estimate the expansion parameter, $Q$, of the theory, as well as the characteristic sizes for even and odd coefficients, $\bar{c}_{even}$ and $\bar{c}_{odd}$.
We want to obtain the posterior distribution for all the LECs that appear at order $k$, a set we collectively denote by ${\bf a}_k$. Here we will obtain the joint posterior pdf of ${\bf a}_k$, the expansion parameter $Q$, and the characteristic sizes. To do this we follow the successful endeavor by the BUQEYE collaboration in Refs.~\cite{Wesolowski:2016, Wesolowski:2019, Wesolowski:2021}, and write the posterior, given experimental data, $\vec{y}_{exp}$, and prior information on the model, $P*$, as
\begin{equation}
\begin{aligned}
\pr({\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd}|&\vec{y}_{exp},P*)\\
=&\pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(Q|\bar{c}_{even},\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(\bar{c}_{even}|\bar{c}_{odd},\vec{y}_{exp},P*)\\
\times&\pr(\bar{c}_{odd}|\vec{y}_{exp},P*).
\label{eq:master}
\end{aligned}
\end{equation}
Marginalization of this posterior distribution over $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$ yields the posterior distribution for ${\bf a}_k$. Other marginalizations can be carried out to obtain posteriors for $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$.
This joint posterior distribution tells us the probability of the LECs and the error model parameters given experimental data. We could use this posterior distribution to get other quantities or observables, such as the energy of a particular rotational level, which depend on the LECs or the error model parameters. These are now represented by distributions and not single numbers. Their distributions are called posterior predictive distributions (PPD). We write the PPD of an observable $\mathcal{O}$ as
\begin{equation}
\label{PPD}
\pr(\mathcal{O}|\vec{y}_{exp},P*)=\int d\vec{\theta}\delta(\mathcal{O}-\mathcal{O}(\vec{\theta})) \pr(\vec{\theta}|\vec{y}_{exp},P*)
\end{equation}
where $\vec{\theta}$ represents the LECs and the error model parameters. Calculating the observable at each point in the parameter space $\vec{\theta}$ and then integrating over the parameters $\vec{\theta}$ allows one to carefully account for correlations between the parameters.
Using Bayes' theorem, we can express the posterior (\ref{eq:master}) as
\begin{equation}
\label{posterior}
\begin{aligned}
\pr({\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd}|&\vec{y}_{exp},P*)\\
=&\pr(\vec{y}_{exp}|{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)\\
\times& \pr(\bar{c}_{even}|P*)\pr(\bar{c}_{odd}|P*)\\
\times& \frac{1}{{\pr(\vec{y}_{exp}|P*)}}.
\end{aligned}
\end{equation}
The terms on the right-hand side of Eq.~(\ref{posterior}) have the following interpretations:
\begin{enumerate}
\item $\pr(\vec{y}_{exp}|{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)$ is the likelihood of the experimental data given specific values of both the LECs that appear in the energy formula at order $k$ and the parameters in our error model.
\item $\pr({\bf a}_k|Q,\bar{c}_{even},\bar{c}_{odd},P*)$ is the prior distribution of the LECs given the parameters encoding the systematic expansion of the EFT.
\item $\pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)$ is the prior distribution of the expansion parameter given the characteristic sizes of even and odd natural coefficients.
\item $\pr(\bar{c}_{even}|P*)$ and $\pr(\bar{c}_{odd}|P*)$ are the prior distributions of the even and odd characteristic sizes. (In Eq.~\eqref{posterior} we assume an uncorrelated prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$.)
\item $\pr(\vec{y}_{exp}|P*)$ is the evidence, which we drop in what follows as it does not depend on the parameters we are interested in extracting and functions only as a normalization constant.
\end{enumerate}
\subsection{Building the Likelihood}
We now build the likelihood function accounting for the expected error between the experimental and theoretical values, for data on $K=1/2$ rotational bands. The corresponding likelihood for $K=3/2$ bands is built analogously. Following~\cite{Wesolowski:2019} we start by writing our observable (the energy of a particular rotation level) at order $k$ as
\begin{equation}
\label{eq:energy}
\begin{aligned}
E(I) &= A_{\rm rot}I(I+1)\\
\Bigg\{1 &+ \sum_{n=odd}^{k} c_n Q^{n-1}
(-1)^{I+1/2}\left(I+\tfrac{1}{2}\right)[I(I+1)]^{(n-3)/2}\\
&+ \sum_{n=even}^{k} c_n Q^{n-1} [I(I+1)]^{(n-2)/2}\Bigg\}.
\end{aligned}
\end{equation}
We choose the leading-order energy for each level, $A_{\rm rot}I(I+1)$, to be the reference scale $E_{\rm ref}$ for the observable. The dimensionless coefficients $c_n$ (see Eq.~(\ref{eq:cs})) are assumed to be $\mathcal{O}(1)$. The theory error $\vec{\sigma}_{th}$ at any order is due to terms omitted from the summations in Eq.~\eqref{eq:energy}. Its most significant contribution comes from the first omitted term in the EFT expansion. Accounting only for this term yields an estimate for the theory error that is fully correlated across levels if $k+1$ is even, and anticorrelated for adjacent levels if $k+1$ is odd. To account for this correlation or anticorrelation we write the theory covariance matrix as the outer product of a vector representing the theory error, $\Sigma_{th} \equiv \vec{\sigma}_{th} \otimes \vec{\sigma}_{th}$.
The vector $\vec{\sigma}_{th}$ contains the value of the first omitted term for each of the $m$ energy levels that enter the likelihood.
We also account for experimental errors by writing the covariance matrix as
\begin{equation}
\Sigma=\Sigma_{th}+\Sigma_{exp}
\end{equation}
where we take $(\Sigma_{exp})_{ij} \equiv (\vec{\sigma}_{exp})_i^2\delta_{ij}$.
The likelihood function is then
\begin{equation}
\begin{aligned}
\pr(\vec{y}_{exp}|&{\bf a}_k,Q,\bar{c}_{even},\bar{c}_{odd},P*)\\
=& \sqrt{\frac{1}{(2\pi)^m |\Sigma|}} \exp\left( -\frac{1}{2}\vec{r}^T \Sigma^{-1} \vec{r} \right),
\end{aligned}
\end{equation}
where $\vec{r}\equiv\vec{y}_{exp}-\vec{y}_{th}$ is the residual between the central experimental energy for a level and the theory result (\ref{eq:energy}) and
$m$ is the number of levels included in the likelihood estimation.
We note that since the theory error is the outer product of the theory error with itself, the theory covariance $\Sigma_{th}$ is singular. Including the experimental error solves this singularity problem for the covariance $\Sigma$. However, $\Sigma$ can still become ill-conditioned for higher values of $Q$ if the experimental errors are too small; numerical issues then arise when we try to invert the covariance matrix.
Including more terms in the estimate for the theoretical error produces a steeper peak in the likelihood function, see Fig.~\ref{fig:likelihood}, which, in turn, restricts the values sampled for $Q$ to a narrower region. Because it precludes the sampler exploring large values of $Q$, this inclusion of more omitted terms in the model of the theoretical error solves the numerical problem of ill-conditioned matrices and gives a more accurate extraction of the LECs and the error-model parameters.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{likelihood.pdf}
\caption{Comparing the Log of the likelihood when accounting for different number of omitted terms, $p$, in the theory error. Apart from $Q$, the parameters that enter the likelihood were chosen to be the median parameters after we have had sampled the posterior distribution for $^{169}$Er.}
\label{fig:likelihood}
\end{figure}
In what follows we estimate the theory error including omitted terms up to a certain cutoff order $k_{max}$. Our theory error estimate for the level with spin $I$ is then
\begin{equation}
\sigma_{th}(I)=A_{\rm rot}\sum_{l=k+1}^{k_{max}}\bar{c}_{even,odd} Q^{l-1} P_l(I),
\label{eq:theoryerror}
\end{equation}
where the $\bar{c}$ that is used here is $\bar{c}_{even}$ for even values of $l$ and $\bar{c}_{odd}$ otherwise. The $I$-dependence of the $l$th term is chosen to match that in Eq.~\eqref{eq:energy}, and is denoted here by $P_l(I)$, a polynomial of power $l$.
We arrange the contributions to the theory error, (\ref{eq:theoryerror}) as the $p$ columns of a $m\times p$ matrix $\sigma_{th}$, where $p=k_{max}-k$ is the number of omitted terms.
Each column in this matrix then corresponds to the theory-error structure, while each row corresponds to a different energy level. To obtain $\Sigma_{th}$ we then again take the outer product of $\sigma_{th}$ with itself, i.e., we construct an outer product in our $m$-dimensional data space, while also taking an inner product in order space.
This results in the theory error associated with different orders being added in quadrature, while maintaining the correlation structure of the theory error across the data space.
\subsection{Building the Priors}
\label{priors}
The prior distributions for an order-$n$ LEC is taken to be a Gaussian with mean zero and standard deviation
\begin{equation}
\sigma_n= \left\{ \begin{array}{lc}
A_{\rm rot} \bar{c}_{even} Q^{n-1} & \mbox{if $n$ is even;}\\
A_{\rm rot} \bar{c}_{odd} Q^{n-1}
& \mbox {if $n$ is odd.} \end{array} \right.
\label{eq:sdeviation}
\end{equation}
encoding the EFT expectations for the sizes of the LECs arising from the power counting described in Sec.~\ref{sec:rotorfermionEFT}. The standard deviation in Eq.~\eqref{eq:sdeviation} allows the possibility for even and odd LECs to have different typical sizes. Combining the Gaussian priors for the LECs yields
\begin{equation}
\begin{aligned}
\pr({\bf a}_k|&Q,\bar{c}_{even},\bar{c}_{odd},P*)\\=&\frac{1}{\bar{E} \sqrt{2\pi}} \exp\left(-\frac{E_k^2}{2 \bar{E}^2}\right)
\prod_{n=1}^k \frac{1}{\sigma_n \sqrt{2\pi}}{\rm exp}\left(-\frac{a_n^2}{2\sigma_n^2}\right).
\end{aligned}
\end{equation}
The LEC $E_K$ is just an energy shift and its size is not determined by the EFT power counting. We set the prior on it to be Gaussian with mean zero and a standard deviation, $\bar{E}$, that is wide enough to capture its value. The value for $\bar{E}$ is determined from the energy of the bandhead and $A_{\rm rot}$ by means of Eq.~(\ref{Energy_NLO}).
We choose not to impose any expectations regarding the size of expansion parameter in the prior for $Q$ and so take it to be flat between two limits:
\begin{equation}
\pr(Q|\bar{c}_{even},\bar{c}_{odd},P*)\propto
\begin{cases}
1 & Q \in (0,Q_{cut}) \\
0 & {\rm otherwise}.
\end{cases}
\end{equation}
Limiting $Q$ from above restricts the sampler from going to high values of $Q$, as they make the covariance matrix ill-conditioned and harder to invert. For all cases we check that the posterior for $Q$ is confined to values well below $Q_{cut}$.
The priors on the characteristic sizes $\bar{c}_{even}$ and $\bar{c}_{odd}$, are taken to be identical scaled-inverse-$\chi^2$ distributions
\begin{equation}
\pr(\bar{c}_l^2|P*) \propto
\begin{cases}
\chi^{-2}(\nu=1,\tau^2=1) & \bar{c}_l^2 \in (0, \bar{c}_{cut}^2) \\
0 & {\rm otherwise},
\end{cases}
\end{equation}
where the cutoff $\bar{c}_{cut}$ prevents numerical issues inverting the covariance matrix. The scaled-inverse-$\chi^2$ distribution, given by
\begin{equation}
\chi^{-2}(x; \nu, \tau^2)=
\frac{(\tau^2\nu/2)^{\nu/2}}{\Gamma(\nu/2)}~
\frac{\exp\left[-\frac{\nu \tau^2}{2 x}\right]}{x^{1+\nu/2}},
\end{equation}
is shown for different values of $\nu$ and $\tau$ in Fig.~\ref{fig:prior_c_bars}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{prior_c_bars.pdf}
\caption{Prior distribution of the size of the dimensionless natural coefficients, $\bar{c}$.}
\label{fig:prior_c_bars}
\end{figure}
We stress that we chose identical priors for $\bar{c}_{even}$ and $\bar{c}_{odd}$ even though we expect the former to be larger than the latter based on previous analyses of data on rotational bands~\cite{Alnamlah:2020}. We did not want to bias our analysis by imposing this hierarchy on the prior, instead anticipating that it will emerge naturally in the posteriors for those parameters.
The scaled-inverse-$\chi^2$ favors small values of $\bar{c}^2$ and has long tails. This allows the sampler to explore higher values of $\bar{c}^2$. The sharp decrease in this distribution for very small values of $\bar{c}^2$ could be a problem for cases where $\bar{c}_{odd}$ is much smaller than one. This is a concern in some $K=3/2$ bands where we expect smaller odd-order corrections to the leading-order energy than in $K=1/2$ bands.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{corner.pdf}
\caption{Corner plot for the marginalized distributions of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for $^{167}$Er including all adopted rotational levels ($I_{max}=16.5$) and accounting for 6 omitted terms in the theory error. The insert in the top right corner show the correlations between posterior parameters. The order of the parameters on the corner plot is the same on the correlations plot. (Here $E_K$ and all the EFT LECs are expressed in keV. The error-model parameters are dimensionless.)}
\label{fig:corner}
\end{figure*}
\section{Running the Sampler}
\label{sec:sampler}
To sample the posterior distribution in Eq.~\eqref{posterior} we use the Python ensemble sampling toolkit for affine-invariant MCMC (emcee) \cite{emcee}. We run the sampler for each nucleus at a certain EFT order using the $m$ rotational levels from the bandhead up to some $I_{max}$ and accounting for $p$ omitted terms in the theory error. We use 64 walkers to sample the posterior distribution for an initial 10000 steps. We then continue running the sampler with 3000 step increments. After every 3000 steps we calculate the autocorrelation time, $\tau_\alpha$, where $\alpha$ indexes an LEC or an error-model parameter. We declare the sampler to be converged if the sampler meets two criteria. First, the number of steps has to be more than 50 times the highest $\tau_\alpha$.
Second, the change in any of the $\tau_\alpha$'s has to be less than 2\% from its value after the last 3000 step increment.
To get the posterior distributions we discard $2\times {\rm max}(\tau_\alpha)$ steps from the beginning of the chain ({\it burn-in}) and $0.5\times {\rm min}(\tau_\alpha)$ steps in between steps we accept ({\it thinning}).
A sample corner plot of the marginalized distributions of the LECs and the error-model parameters $Q$, $\bar{c}_{even}$ and $\bar{c}_{odd}$, for the case of ${}^{167}$Er is shown in Fig.~\ref{fig:corner}. This figure clearly shows that the posterior distributions for all parameters are fully converged. For this particular case we set $Q_{cut}=0.16$ and $\bar{c}_{cut}=22$ for both $\bar{c}_{even}$ and $\bar{c}_{odd}$. As explained in Sec.~\ref{priors}, the cutoffs on $Q$ and the characteristic sizes prevent the covariance matrix from being ill-conditioned. We also ran the sampler for $^{167}$Er at different values of $Q_{cut}$ and $\bar{c}_{cut}$ and found that different choices of these hyperparameters do not result in a significant change in the posterior distributions.
For some cases, namely $^{99}$Tc and $^{183}$W, the posterior distribution of $Q$ was initially at the upper limit of the prior. We then ran into numerical problems when increasing $Q_{cut}$ trying to encompass the entire posterior. This problem was solved by decreasing the number of levels included in the analysis, i.e., decreasing $I_{max}$. It was then possible to increase $Q_{cut}$ without encountering problems with degenerate matrices. This means that for $^{99}$Tc we were only able to extract the LECs and $Q$ at $I_{max}=11.5$. For $^{183}$W we needed to remove two levels from the upper end of the data set for the sampler to be numerically stable.
In Fig.~\ref{fig:corner} we see clear correlations between $E_K$, $A$, and $B$ and also between $A_1$ and $B_1$. (Here we have dropped the subscript $K$ on $A$ and $B$; it is to be understood that all LECs are band dependent.) The correlation coefficients given in the inset in the top-right corner of the figure make the block-diagonal structure of the covariance matrix clear. To a good approximation the correlation matrix can be decomposed into a correlation matrix for even-order LECs, one for odd-order LECs, and one for the error-model parameters.
We note that, as expected, $\bar{c}_{odd}$ is smaller than $\bar{c}_{even}$. Corrections to the energy levels carrying odd powers of $I$ are smaller than those carrying even powers of $I$. This size difference is connected to different physics correcting the effective Lagrangian at even and odd orders.
To see which of the parameters has the narrowest distribution and therefore places the strongest constraint on the posterior distribution, we did a Singular Value Decomposition (SVD) of the Hessian matrix. We found that the eigenvector with the highest eigenvalue, i.e., the parameter combination with smallest absolute error, is made up mostly of the highest-order LEC. This is unsurprising, since that LEC, $B$, is markedly smaller than the others (we note that its relative error is actually larger than that on, e.g., $A_1$).
We initially found a peculiar correlation between LECs in some cases where the rotational band was built on the ground state of the nucleus we were looking at. There we found the eigenvector with the highest eigenvalue was a very particular linear combination that involved all the LECs. We ultimately traced this correlation to the fact that the ground state experimental error had been set to zero, and so the combination of LECs that entered the formula for the ground-state energy was very well constrained (theory error is also very small there). This problem was solved by adding a small experimental error to the ground state. We chose it to be equal to the error that the NNDC quotes on the energy of the first excited state.
\section{Results}
\label{sec:results}
\begin{figure}[b!]
\centering
\includegraphics[width=\columnwidth]{wes_A_1169Er.pdf}
\caption{Posteriors for $A_1$ describing $^{169}$Er as a function of $I_{max}$ at different EFT orders. The solid line connects the median values and the error bands encompass the 16th and 84th percentiles of the marginalized distribution.}
\label{fig:wes_A1_above_orders}
\end{figure}
In this section we show results for our Bayesian analysis of the rotational energy levels in $^{99}$Tc, ${}^{155,157}$Gd, ${}^{159}$Dy, ${}^{167, 169}$Er, ${}^{167, 169}$Tm, ${}^{183}$W, ${}^{235}$U and ${}^{239}$Pu. The experimental data are taken from the National Nuclear Data Center (NNDC)~\cite{Baglin:2016vll,Browne:2017uto,Nica:2016elr,Nica:2019ykt,Reich:2012ouk,Browne:2014ukl,Browne:2014gwf,Baglin:2000ong,Baglin:2008hsa}. Except for the cases of ${}^{99}$Tc and ${}^{183}$W noted above, we included all levels in a certain rotational band according to the adopted level determination in the NNDC.
\subsection{Stable LEC Extraction Across EFT Orders and Additional Data}
In this subsection we show that lower-order LECs extracted for the selected rotational bands are stable across EFT orders and with the addition of high-energy data, provided that we account for enough omitted terms when treating the theory error. For most nuclei including omitted terms up to $k_{max}=10$, i.e., accounting for six omitted terms at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace, was enough to stabilize the extraction of the LECs.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_B_1239Pu.pdf}
\includegraphics[width=\columnwidth]{wes_B239Pu.pdf}
\caption{Posteriors for $B_1$ and $B$ describing $^{239}$Pu a function of $I_{max}$ for different values of $k_{max}$. The solid line connects the median values and the error bands encompass the 16th and 84th percentiles in the marginalized distribution.}
\label{fig:wes_B_above_errors}
\end{figure}
As an example, we show the stability of the extracted LEC, $A_1$, across number of levels included at different EFT orders in Fig.~\ref{fig:wes_A1_above_orders}. In this figure, $I_{max}$ is the spin of the highest-energy level included in a particular analysis. The central values of the resulting posteriors are consistent with each other within 68\% credible intervals, shown as error bars in the figure. Adding more levels to the analysis narrows the posteriors for the LECs up to a certain $I_{max}$, after which the widths of these distributions saturate.
Fig.~\ref{fig:wes_A1_above_orders} also demonstrates striking agreement between the distributions obtained at low and high EFT orders: they are almost identical as long as omitted terms up to the same $k_{max}$ are accounted for in both analyses.
The importance of including more than one omitted term in the theory error estimate is evident in Fig.~\ref{fig:wes_B_above_errors}.
The top and bottom panels of the figure show the way that posteriors for $B_1$ and $B$ evolve as $I_{max}$ increases. This is done using three error models that include different numbers of omitted terms. These results show that including more omitted terms in the model of the theory error removes the drifting and staggering of the central values.
For both cases the distributions at $k_{max}=10$ agree within errors as we go higher in $I_{max}$. The narrowing of the distribution as we go higher in $I_{max}$ is clearly seen in those two figures. In addition to having less data, the broadening of the error bands at low $I_{max}$ comes from the fact that including less levels in the analysis leads to highly correlated LECs. This allows the numerically larger errors on the lower-order LECs to contribute to the errors on the higher-order LECs, thereby enhancing them.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_corr.pdf}
\caption{Correlations between LECs and error-model parameters as a function of $I_{max}$, resulting from the analysis on the lowest $K=1/2$ rotational band in $^{239}$Pu at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace with $k_{max}=10$.}
\label{fig:wes_corr}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{wes_EK_below.pdf}
\caption{The distribution of $E_K$ for $^{169}$Er at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ as we successively remove the lowest energy levels from the data set $D$. The solid blue line connects the median values and the error bands encompass uncertainties between on the 16th and 84th percentiles of the samples in the marginalized distribution. The solid black lines show to size of the standard deviation set with the Gaussian prior on $E_K$.}
\label{fig:wes_EK_below}
\end{figure}
In Fig.~\ref{fig:wes_corr} we show the decrease in the correlations between the LECs as $I_{max}$ increases. The high correlation between the LECs at low $I_{max}$ occurs because these analyses do not include enough data to constrain all LECs independently. Furthermore, the high correlation between the LECs at low $I_{max}$ also results in an unreliable extraction of the expansion parameter $Q$. This comes from the fact that at low energies the theory truncation error is very small compared to the experimental error. Indeed, adding more terms to our EFT error model (i. e., increasing $k_{max}$) leads to higher correlation between the LECs at low $I_{max}$. Thus, the number of levels required to reliably extract $Q$ increases with increasing $k_{max}$.
Starting instead at the low-$I$ end of the data: when we progressively remove the lowest-energy levels from the data set $D$ used to construct the likelihood we rapidly lose the ability to reliably extract the LECs. Figure \ref{fig:wes_EK_below} shows that the distribution for $E_K$ starts narrow and broadens as we remove levels from below. When we remove the six lowest energy levels the distribution of $E_K$ is exactly the same as the prior distribution: the likelihood is making no contribution to the $E_K$ posterior.
The previous results were nearly the same for all cases considered in this work. However, even for $k_{max}=10$, staggering and shifting of the LECs remains sizable for the $K=1/2$ bands in $^{183}$W, $^{167}$Tm and $^{235}$U. In $^{183}$W and $^{167}$Tm, these effects could be attributed to large expansion parameters, as they translate to large omitted contributions to the energies of the rotational levels. In $^{235}$U, the fermionic matrix elements could be larger than naively expected, causing the systematic expansion of the EFT to be questionable as discussed in Ref.~\cite{Alnamlah:2020}.
For $K=3/2$ bands, we were able to extract stable LECs from the $^{159}$Dy analysis by setting $I_{max}=15.5$. This extraction required us to consider omitted terms up to $k_{max}=12$. This is because the spin at which the EFT breaks in this nucleus is $I_{br}\approx 15.5$. (This, then, is the third case in which we do not use all the NNDC energy-level data available on a particular band.) $^{157}$Gd is stable across orders and $I_{max}$, while $^{155}$Gd exhibits shifting and staggering due to a larger expansion parameter, $Q \approx 0.07$.
\begin{table*}[t!]
\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
Nucleus&$E_K$ [keV]&$A_1$ [keV]&$ A$ [keV]&$(B_1,A_3)$ [keV]&$B$ [keV]&$Q$&$\bar{c}_{\rm even}$&$\bar{c}_{\rm odd}$\\ \hline
$^{99}$Tc&$147.3111_{-0.0142}^{0.0143}$&$70.1912_{-0.0103}^{0.0103}$&$82.2577_{-0.0086}^{0.0086}$&$-7.5851_{-0.0022}^{0.0022}$&$-2.99165_{-0.00079}^{0.00079}$&$0.204_{-0.016}^{0.021}$&$1.87_{-0.53}^{0.43}$&$1.35_{-0.39}^{0.54}$\\ \hline
$^{155}$Gd&$-45.2063_{-0.0025}^{0.0025}$&-&$12.0184_{-0.0009}^{0.0009}$&$-0.0081_{-0.0001}^{0.0001}$&$0.00571_{-0.00008}^{0.00008}$&$0.067_{-0.006}^{0.007}$&$3.54_{-1.11}^{2.00}$&$2.01_{-0.94}^{1.99}$\\ \hline
$^{157}$Gd&$-41.3241_{-0.0097}^{0.0093}$&-&$11.0240_{-0.0032}^{0.0034}$&$-0.0094_{-0.0002}^{0.0002}$&$-0.00500_{-0.00025}^{0.00024}$&$0.057_{-0.006}^{0.006}$&$5.23_{-2.33}^{5.18}$&$0.55_{-0.15}^{0.28}$\\ \hline
$^{159}$Dy&$-42.7714_{-0.0093}^{0.0092}$&-&$11.4076_{-0.0030}^{0.0030}$&$-0.0060_{-0.0003}^{0.0003}$&$-0.00307_{-0.00020}^{0.00020}$&$0.063_{-0.004}^{0.004}$&$3.52_{-1.16}^{2.22}$&$0.95_{-0.33}^{0.64}$\\ \hline
$^{167}$Er&$207.2088_{-0.0053}^{0.0053}$&$7.8383_{-0.0028}^{0.0027}$&$11.2398_{-0.0016}^{0.0016}$&$-0.0063_{-0.0003}^{0.0003}$&$-0.00820_{-0.00011}^{0.00010}$&$0.047_{-0.004}^{0.005}$&$4.04_{-1.44}^{2.63}$&$0.83_{-0.26}^{0.49}$\\ \hline
$^{169}$Er&$0.9673_{-0.0197}^{0.0195}$&$9.7774_{-0.0100}^{0.0102}$&$11.7625_{-0.0053}^{0.0058}$&$-0.0064_{-0.0009}^{0.0008}$&$-0.00309_{-0.00023}^{0.00016}$&$0.031_{-0.005}^{0.006}$&$5.57_{-2.51}^{5.31}$&$0.77_{-0.23}^{0.42}$\\ \hline
$^{167}$Tm&$-18.4635_{-0.0155}^{0.0156}$&$-9.1267_{-0.0057}^{0.0056}$&$12.5095_{-0.0022}^{0.0022}$&$0.0431_{-0.0003}^{0.0003}$&$-0.00906_{-0.00008}^{0.00008}$&$0.056_{-0.005}^{0.005}$&$2.16_{-0.76}^{1.37}$&$1.06_{-0.34}^{0.62}$\\ \hline
$^{169}$Tm&$-19.0532_{-0.0011}^{0.0011}$&$-9.7204_{-0.0008}^{0.0008}$&$12.4738_{-0.0007}^{0.0006}$&$0.0264_{-0.0002}^{0.0002}$&$-0.00497_{-0.00006}^{0.00007}$&$0.045_{-0.007}^{0.008}$&$2.59_{-1.12}^{2.30}$&$0.94_{-0.30}^{0.60}$\\ \hline
$^{183}$W&$-6.8750_{-0.0008}^{0.0008}$&$2.7512_{-0.0005}^{0.0005}$&$12.7754_{-0.0004}^{0.0004}$&$-0.0436_{-0.0001}^{0.0001}$&$0.01977_{-0.00003}^{0.00003}$&$0.076_{-0.006}^{0.007}$&$3.80_{-1.49}^{2.86}$&$0.67_{-0.20}^{0.35}$\\ \hline
$^{235}$U&$-6.1882_{-0.0008}^{0.0008}$&$-1.7298_{-0.0007}^{0.0007}$&$6.0508_{-0.0002}^{0.0002}$&$0.0025_{-0.0001}^{0.0001}$&$-0.00249_{-0.00001}^{0.00001}$&$0.040_{-0.003}^{0.004}$&$3.85_{-1.28}^{2.36}$&$0.82_{-0.27}^{0.53}$\\ \hline
$^{239}$Pu&$-8.3578_{-0.0019}^{0.0019}$&$-3.6559_{-0.0011}^{0.0011}$&$6.2726_{-0.0004}^{0.0004}$&$0.0041_{-0.0001}^{0.0001}$&$-0.00149_{-0.00003}^{0.00003}$&$0.029_{-0.003}^{0.004}$&$5.72_{-2.22}^{4.18}$&$0.93_{-0.31}^{0.61}$\\ \hline
\end{tabular}
\renewcommand{\arraystretch}{1}
\caption{The median value of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for the nuclei considered in this work. The uncertainties encompass the 16th and 84th percentiles of the samples in the marginalized distributions. $K=3/2$ rotational bands do not have a parameter $A_1$ and the parameters ($B_1,A_3$) refer to $K=1/2$ and $K=3/2$ bands respectively.}
\label{tab:LECs}
\end{table*}
The values of the LECs and the error-model parameters at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace for the nuclei considered in this work are given in Tabele \ref{tab:LECs}.
\subsection{Prior Sensitivity}
In addition to using the scaled-inverse-$\chi^2$ distribution as a prior for $\bar{c}_{even}$ and $\bar{c}_{odd}$ we tried truncated Gaussians with mean zero and standard deviations $\sigma=7$ and $\sigma=3$ respectively for all cases. These truncated Gaussian priors allow for smaller values of the characteristic sizes.
But the standard deviations were chosen to still allow values for $\bar{c}_{odd}$ and $\bar{c}_{even}$ larger than those resulting from scaled-inverse-$\chi^2$ priors with $\nu=1$ and $\tau^2=1$.
The change in prior for $\bar{c}_{odd}$ and $\bar{c}_{even}$ does not significantly change the posteriors for the LECs: the corresponding central values differ by less than 1\%, and are consistent with each other within the 68\% credible intervals.
Central values of the posteriors for $Q$ differ by less than 15\%, and were similarly consistent.
The strongest dependence on the prior is that exhibited by the posteriors for $\bar{c}_{odd}$ and $\bar{c}_{even}$: the central values differ in some cases by more than 50\%. However, even these values are consistent with each other within 68\% credible intervals, since the posteriors for the characteristic sizes are broad.
The changes in the posteriors of $Q$ on one hand and $\bar{c}_{odd}$ \& $\bar{c}_{even}$ on the other are anticorrelated. We only care about combinations of them to set the size of the theory error and the expected size of the LECs. Thus, the dependence of the theory error and the expected size of the LECs on the prior for the characteristic sizes is less profound. The difference in sizes of the theory error resulting from the chosen priors is less than 20\% for all cases except $^{157}$Gd, where the difference is about 40\%.
For all results that follow we used the scaled-inverse-$\chi^2$ distribution with $\nu=1$ and $\tau^2=1$ as the prior for both $\bar{c}_{odd}$ and $\bar{c}_{even}$, in keeping with the naturalness assumption.
\subsection{Posterior Predictive Distributions}
Figure \ref{fig:Er_ppd,U_ppd} shows the PPDs of the energy residuals as a function of the spin $I$ for two cases considered in this work. These distributions are calculated using Eq.~(\ref{PPD}).
In each figure, translucent blue lines connect energy residuals resulting from different LECs sets sampled from the posterior distribution in Eq.~\eqref{eq:master}. The solid black line represents the median of the PPD, and the dashed lines encompass the region between the 16th and 84th percentiles.
The dark and light red bands show the truncation error and the experimental error added in quadrature at 68\% and 95\% credible levels respectively. To calculate the truncation error, we consider a theory error that accounts for six omitted terms. They are added in quadrature, just as they are in the likelihood defined in Sec.~\ref{sec:Bayesianmodel}. This calculation was done using the median values of $Q, \bar{c}_{even}$, and $\bar{c}_{odd}$. The dependence of the size of the theory error on the prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$ is small in these cases: the theory error changes by about 10\% when the prior is changed.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{169Er_ppd.pdf}
\includegraphics[width=\columnwidth]{235U_ppd.pdf}
\caption{The posterior predictive distribution for energy-level residuals at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ in $^{169}$Er and $I_{max}=17.5$ (top panel) and in $^{235}$U and $I_{max}=23.5$ (bottom panel). The dark and light red bands show the truncation error plus the experimental error at 68\% and 95\% credible levels respectively. The lighter blue lines connect the energy residuals calculated from the distribution of the LECs. The solid black line represents the median of the distribution and the dashed lines indicate the 16th and 84th percentiles. The correlation shown on the plot is the highest correlation between any LEC and any error-model parameter. $I_{br}$ was determined from the distribution of $Q$. The dashed purple line shows the lower limit of $I_{br}$. The insert on the plot shows the residuals on the first 5 levels with an altered y-axis scale.}
\label{fig:Er_ppd,U_ppd}
\end{figure}
The correlation coefficient written in the legend in Fig. \ref{fig:Er_ppd,U_ppd} is the largest between any LEC and any error-model parameter for the shown analysis. When this value is small, the truncation error and the propagated LEC error could in principle be added together in quadrature.
In viewing Fig.~\ref{fig:Er_ppd,U_ppd} it is important to remember that
the truncation error on the energy residuals is highly correlated across levels. This comes from the high correlation between levels when building the correlation matrix that goes into the likelihood. This correlation also flows into a correlation between levels in the PPD of the energies. A correlation plot between two energy levels, like the ones in Fig. \ref{fig:Er_2d_ppd,U_2d_ppd}, gives a 2D cut of this multi-dimensional correlation.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{2d_169Er.pdf}
\includegraphics[width=\columnwidth]{2d_235U.pdf}
\caption{A 2D cut of the posterior predictive distribution at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace and $k_{max}=10$ for $^{169}$Er and $I_{max}=17.5$ (top panel) and in $^{235}$U and $I_{max}=23.5$ (bottom panel). The blue dots show the energies calculated from the distribution of the LECs. The black cross shows the experimental value and the black lines and black ellipse shows the corresponding experimental uncertainty. The remaining ellipses and lines show the truncation error and the experimental error added in quadrature. (All the ellipses are centered at the experimental value.) The orange, red and green account for 1, 2 and 6 omitted terms in the theory error respectively. (In the top panel the red ellipse is completely covered by the green ellipse.)}
\label{fig:Er_2d_ppd,U_2d_ppd}
\end{figure}
In both panels we see the importance of accounting for more than one omitted term in the theory error. This is clearly shown in the reverse in the direction of the correlation from a negative to a positive correlation when going from the orange ellipse to the red ellipse. The orange ellipse is obtained when we account for only one omitted term, while the red ellipse includes the effect of two omitted terms. After accounting for six omitted terms the green ellipse is obtained and the 68\% ellipse in principle expands. This is more clearly seen when we go to high-energy levels plotted in the lower panel in Fig. \ref{fig:Er_2d_ppd,U_2d_ppd}. Note also that for lower-energy levels the correlation is smaller since the experimental error dominates over the truncation error, and we assumed that the experimental errors are not correlated across energy levels.
\subsection{Model Checking}
In Figs. \ref{fig:A_vs_Q}, \ref{fig:B1_vs_Q} and \ref{fig:B_vs_Q} we compare the marginalized posterior distributions of the LECs, on the y-axis, with their expected sizes from the EFT power counting, on the x-axis. Since we also extract the theory error parameters from the sampler and they are highly correlated among themselves, we calculate the expected size from the distributions of the error model parameters using Eq. (\ref{PPD}). We notice that the error on the distribution of the LECs is very small compared to the error on the expected sizes that comes from the distribution of the theory error parameters.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{A_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^2{\rm LO}}\xspace LEC, $\Delta A$, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The blue colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands. }
\label{fig:A_vs_Q}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{B1_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^3{\rm LO}}\xspace LEC, $B_1$ for $K=1/2$ bands and $A_3$ for $K=3/2$ bands, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass uncertainties between on the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The yellow colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:B1_vs_Q}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{B_vs_Q.pdf}
\caption{The size of the \ensuremath{{\rm N}{}^4{\rm LO}}\xspace LEC, $B$, (on the y-axis) compared to its expected size from the EFT power counting (on the x-axis). Error bands on the LEC distribution are small and can not been seen on the plot. The error bands on the x-axis encompass uncertainties between on the 16th and 84th percentiles. Different nuclei are labeled in the legend of the plot. The yellow colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:B_vs_Q}
\end{figure}
As these graphs are model-checking graphs, and since the estimates of LEC sizes plotted on the x-axis are meant as order-of-magnitude estimates, we do not expect perfect linear correlations.
Nevertheless, Fig.~\ref{fig:A_vs_Q} shows that, for all $K=1/2$ bands considered, the LEC $\Delta A$ agrees with its expected size within error bands. This result is surprisingly better than expected. In contrast, the size of $\Delta A$ for $K=3/2$ bands is larger than expected, especially for $^{155}$Gd (see blue points in Fig.~\ref{fig:A_vs_Q}). There are two factors that could contribute to this. First, the $K=3/2$ bands have larger fermionic matrix elements. This could hinder the systematic expansion of the EFT. Second, the $K=3/2$ bands have relatively larger expansion parameters, see Fig. \ref{fig:Q_vs_Q}.
The same discussion applies to the results in Figs. \ref{fig:B1_vs_Q} and \ref{fig:B_vs_Q}, where we see good agreement between the LECs and their expected sizes for $K=1/2$ bands. The disagreement with power-counting estimates for $K=3/2$ bands at \NkLO{3,4} is less of a concern than the one at \NkLO{2} seen in Fig.~\ref{fig:A_vs_Q}, since these higher-order LECs are smaller than their expected sizes. This doesn't undermine the convergence of the EFT expansion.
We also note here that the scale of the x-axis is prior dependent and could change by more than 50\% in some nuclei, depending on the choice of prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$.
For $^{157}$Gd changing the prior on $\bar{c}_{even}$ and $\bar{c}_{odd}$ to a truncated normal allowed for smaller values of $\bar{c}_{odd}$ and $A_3$ was then equal to the expected size (i. e.,the point for $^{157}$Gd then falls exactly on the line in figure \ref{fig:B1_vs_Q}). This did not happen when the truncated normal is chosen as a prior for the analysis in $^{155}$Gd and $^{159}$Dy; this may occur because there is strong \NkLO{5} energy-level staggering present in the data for these nuclei.
The size of $\bar{c}_{even}$ and $\bar{c}_{odd}$ is constrained by both the sizes of the LECs and the size of the theory error. In a good systematic expansion the tension between those factors on setting the size of the $\bar{c}$'s would be small and one number apiece would suffice to represent the even and odd order corrections. However when the systematic expansion is hindered, as in the case for $K=3/2$ bands due to large fermionic matrix elements, this tension becomes clear. One example of this is seen in Figs. \ref{fig:A_vs_Q} and \ref{fig:B_vs_Q} for $K=3/2$ bands. There $\Delta A$ is large and favors large values of $\bar{c}_{even}$, however, $B$ is small and favors smaller values of $\bar{c}_{even}$. The eventual result is a compromise. This tension may be exacerbated by the truncation error also providing information on the size of the $\bar{c}$'s.
\subsection{Higher than Expected Break-down Scale}
In Fig. \ref{fig:Q_vs_Q} we see a clear correlation between the extracted values of $Q$ and those that are expected based on each nucleus' single-particle and vibrational energy scales, $E_{\rm sp}$ and $E_{\rm vib}$. The expected $Q$ is the larger of $E_{\rm rot}/E_{\rm sp}$ and $E_{\rm rot}/E_{\rm vib}$, while the extracted $Q$ comes from sampling the posterior in Eq. \ref{posterior}. This extracted $Q$ is what actually determines the convergence of the EFT expansion. It is markedly smaller than would be naively expected. The break-down scale of the theory is thus higher than naively expected: our rotational EFT works to much higher $I$ than energy-scale arguments would suggest. This could occur because coupling between the higher rotational states explicitly included in the EFT and the high-energy states that are not explicitly included in our EFT is hindered by the large difference in angular momentum between them.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{Q_vs_Q.pdf}
\caption{The extracted expansion parameter $Q$ from marginalized posterior distribution of our sampling compared to its naively expected size. The expected size is taken to be the maximum of $E_{\rm rot}/E_{\rm sp}$ and $E_{\rm rot}/E_{\rm vib}$. The dashed black line shows the best linear fit and its parameters are printed on the plot. The blue colored points are results for rotational bands with bandheads $K=3/2$, all the others are $K=1/2$ bands.}
\label{fig:Q_vs_Q}
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
We performed a Bayesian analysis to extract the LECs and expansion parameters $Q$ describing the rotational energy levels of diverse odd-mass nuclei within a recently developed EFT.
This analysis corroborates the EFT organization for energy-level formulae which results from the assumed power-counting scheme: the extracted LECs of order $k$ scale as $Q^{k-1}$, i.e., according to EFT expectations.
While our analysis reached this conclusion for both $K=1/2$ and $K=3/2$ rotational bands, the sizes of the LECs describing the latter exhibit larger deviations from their expected values than those describing the former. We attribute this behavior to the size of fermionic matrix elements, assumed to be of order one while organizing energy-level formulae. Since these matrix elements involve the angular momentum of the fermion, $\vec{j}$, we cannot exclude the possibility that the systematic behavior of the EFT is hindered in bands build on top of single-particle orbitals with larger values of $K$. For the $K=3/2$ bands studied in this work, however, this discrepancy does not destroy the systematic improvement of calculated energies up to \ensuremath{{\rm N}{}^4{\rm LO}}\xspace, as the sizes of extracted LECs are smaller that expected.
In order to ensure that the extracted values are independent of the EFT order and number of energy levels entering the analysis, we employed a theory error beyond the first-omitted-term approximation, considering omitted terms in the expansion for the energy of rotational levels up to order $k_{max}$. As we increased the number of omitted terms considered in the theory error, the corresponding log likelihood exhibited steeper and steeper peaks. Therefore, the `widths' of the sampled posteriors decrease as $k_{max}$ increases. Considering six omitted terms at \ensuremath{{\rm N}{}^4{\rm LO}}\xspace enabled a stable extraction of the LECs and expansion parameters describing the levels of interest. The shapes of posteriors for low-order LECs extracted at this order and those extracted using lower-order energy formulae are, for all practical purposes, identical. On the other hand, the shapes of the posteriors depend strongly on the number of levels informing the model, narrowing as more levels are included. Nevertheless, the 68\% credible intervals of these posteriors possess significant overlap, facilitating reliable LEC extraction.
In addition to the posteriors for the LECs and expansion parameters, our analysis yielded distributions for the characteristic sizes of even and odd $c_n$'s, $\bar{c}_{even}$ and $\bar{c}_{odd}$. The values of $\bar{c}_{odd}$ are typically smaller than those for $\bar{c}_{even}$, in agreement with results from previous studies where the LECs were fitted to the smallest possible data sets. The difference of the characteristic sizes of even and odd LECs has its origin in the physics behind the corresponding contributions to the effective Lagrangian: while odd-order contributions correct the particle-rotor interaction, even-order contributions include terms that depend exclusively on the rotor degrees of freedom, thus correcting the physics of the core to which the particle is coupled. Here this conclusion was reached solely on the basis of experimental data; we assumed equal priors for both characteristic sizes.
Although the distributions of $\bar{c}_{odd}$ and $\bar{c}_{even}$ change depending on the choice of the their priors, that does not significantly change the distributions of the LECs. Altering the priors also does not have a large effect on the size of the theory error, which changes by less than 20\% for nearly all cases.
These considerations mean that our extractions of the LECs and the theory error parameters in the EFT of rotational bands in odd-mass nuclei are robust under the choice of prior. The formalism presented here also gives robust results for LECs across orders and as more data is added to the analysis. We conclude that a Bayesian framework that incorporates theory errors in the likelihood offers significant advantages for LEC extraction in EFTs. This methodology has already been used for the extraction of LECs in the NN potential from phase shifts~\cite{Wesolowski:2019} and to constrain parameters of the three-nucleon force~\cite{Wesolowski:2021}. But it is a very general approach which should improve the parameter estimation for LECs in any EFT.
\begin{acknowledgments}
We thank Dick Furnstahl and Jordan Melendez for useful discussions. We are also grateful for Daniel Odell's significant conceptual and computational assistance.
This work was supported by the US Department of Energy, contract DE-FG02-93ER40756 (IKA, DRP) by the National Science Foundation CSSI program Award OAC-2004601 (DRP), and under the auspices of the US Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 (EACP). IKA acknowledges the support of King Saud University and the Ministry of Education in Saudi Arabia.
\end{acknowledgments}
| da5ebc6492b2b29c00981a74401744570821112a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
The deviations from lepton flavor universality observed in neutral-current~\cite{LHCb:2014vgu,LHCb:2017avl,LHCb:2019hip,LHCb:2021trn} and charged-current~\cite{BaBar:2012obs,BaBar:2013mob,Belle:2015qfa,LHCb:2015gmp,LHCb:2017smo,LHCb:2017rln} semileptonic $B$~decays have stimulated intense model-building activity, triggering new ideas about the ultraviolet (UV) completion of the Standard Model (SM). Two key aspects have emerged quite clearly from the early attempts to provide a combined explanation of the two sets of anomalies: i)~a possible common origin of flavor anomalies and Yukawa hierarchies~\cite{Bordone:2017bld}, as hinted by the approximate $U(2)^n$ flavor structure of new physics ~\cite{Greljo:2015mma,Barbieri:2015yvd}, ii)~the necessity of new degrees of freedom at the TeV scale coupled mainly to the SM third generation, hinting at a possible link with the electroweak (EW) hierarchy problem~\cite{Barbieri:2017tuq,Fuentes-Martin:2020bnh}.
In this letter, we show how these two aspects can be consistently combined within a five-dimensional (5D) model. The three main assumptions of our construction, and their motivations, can be listed as follows:
\smallskip
{\bf I.} {\em 4321 gauge symmetry above the TeV scale}. The most effective mediator to address both sets of anomalies is a TeV-scale $U_1$ leptoquark~\cite{Alonso:2015sja,Calibbi:2015kma,Barbieri:2015yvd,Bhattacharya:2016mcc,Buttazzo:2017ixm}. This field can be identified with one of the broken generators of a fundamental $4321$ gauge group, where color is the diagonal subgroup of $SU(4)_h \times SU(3)_l$~\cite{DiLuzio:2017vat}. The labels $h$ and $l$ indicate the flavor non-universal assignment of the SM fermions under this part of the gauge group, resulting in a $U_1$ coupled mainly to third-generation fermions~\cite{Bordone:2017bld,Greljo:2018tuh}. Apart from the $U_1$, two EW-neutral gauge bosons acquire mass from the 4321 breaking: a color-octet, $G^\prime$, and a singlet, $Z^\prime$. The presence of these two mediators do not alter the $U_1$ solution of the anomalies~\cite{DiLuzio:2018zxy,Cornella:2019hct,Cornella:2021sby}.
\smallskip
{\bf II.} {\em Flavor hierarchies from a 3-brane structure in 5D}. The hierarchies in both the Yukawa and $U_1$ couplings, i.e.~the breaking of the approximate $U(2)^n$ flavor symmetry acting on the light families at the TeV scale, emerge from a multi-scale construction~\cite{Panico:2016ull,Bordone:2017bld,Allwicher:2020esa,Barbieri:2021wrc}
that, in turn, can be viewed as the effect of a 3-brane structure in 5D. The strong constraints on flavor-violating terms involving the light families naturally point toward a warped geometry~\cite{Fuentes-Martin:2020pww}.
The size of the Yukawa couplings implies $kL \approx 10$~\cite{Fuentes-Martin:2020pww}, where $L$ is the distance between the infrared (IR) (3$^{\rm rd}$ gen.) brane and the most UV (1$^{\rm st}$ gen.) brane, and $k$ is the 5D curvature constant.
\smallskip
{\bf III.} {\em Holographic Higgs}.
The SM Higgs can be realized as a pseudo-Nambu-Goldstone boson (pNGB) emerging from the same dynamics responsible for the breaking of $SU(4)_h \times SU(3)_l$~\cite{Fuentes-Martin:2020bnh}. In the warped 5D description, this can be achieved via gauge-Higgs unification~\cite{Contino:2003ve}, realized by extending the EW part of the bulk gauge symmetry. For the sake of simplicity and minimality, we assume
\begin{align}\label{eq:GbulkIR}
\begin{aligned}
\mathcal{G}^{\rm 23}_{{\rm bulk}} &\equiv SU(4)_h\times SU(3)_l\times U(1)_l \times SO(5)\,, \\
\mathcal{G}_{\rm IR} &\equiv SU(3)_c\times U(1)_{B-L}\times SO(4)\,,
\end{aligned}
\end{align}
where $SU(3)_c$ and $U(1)_{B-L}$ are flavor-universal
subgroups of $SU(4)_h\times SU(3)_l \times U(1)_l$, and the 23 bulk
is the most IR side of the bulk (see Fig.~\ref{fig:branes}).
The fifth component of the gauge fields associated to the $SO(5)/SO(4)$ coset contains four pNGB zero modes transforming as a ${\bf 4}$ of $SO(4)$ that we identify with the SM Higgs field, thus realizing the minimal composite Higgs scenario~\cite{Agashe:2004rs}.
\section{The 5D model}\label{sec:model}
We consider a 5D model with a warped compact extra dimension containing three branes (similar to the one explored in~\cite{Kogan:2000xc}) with two positive and one negative tension branes (++$-$). The metric is
\begin{align}
ds^{2} = e^{-2\sigma(y)}\, \eta_{\mu\nu}\, dx^{\mu} dx^{\nu} - dy^{2} \,,
\end{align}
where, in the absence of backreaction from scalar fields, the warp factor $\sigma(y)$ is
\begin{align}\label{eq:warp}
\sigma(y) = \begin{cases}
\sigma_1(y)=k_{1} y & 0\leq y \leq \ell \\
\sigma_2(y)=k_{1} \ell + k_{2}(y-\ell) & \ell \leq y \leq L
\end{cases} \,.
\end{align}
Here $L$ denotes the total length of the compact extra dimension and $\ell$ the location of the intermediate brane. Due to the tension of the branes, $k_1\leq k_2$.
For simplicity, in what follows we assume $k_{1} = k_{2} \equiv k$, corresponding to a zero middle brane tension. This multi-brane setup can be stabilized via a straight-forward extension of the Goldberger-Wise (GW) mechanism~\cite{Goldberger:1999uk,Lee:2021wau}, with suitable brane-localized potentials for the GW scalar~\bibnote{In the limit of $k_{1}\ell, k_{2}(L-\ell) > 1$ and small bulk mass for the GW scalar, the stabilization is achieved via two factorized solutions of the type found in~\cite{Goldberger:1999uk}, one for each distance}.
Beside the IR bulk and brane symmetries specified in \eqref{eq:GbulkIR}, we assume
\begin{align}
\mathcal{G}^{\rm 12}_{{\rm bulk}} &\equiv SU(4)_h\times SU(4)_l\times SO(5)\,, \\
\mathcal{G}_{\rm UV} & \equiv SU(4)_h\times SU(3)_l\times U(1)_l\times SU(2)_L\times U(1)_R\,, \nonumber
\end{align}
where $\mathcal{G}_{\rm UV}$ and $\mathcal{G}^{\rm 12}_{{\rm bulk}}$ are the gauge symmetries in the most UV brane and the UV side of the bulk, respectively. The middle brane is therefore a discontinuity corresponding to the symmetry-breaking pattern $SU(4)_l \rightarrow SU(3)_l \times U(1)_l$. Alternatively, we could have chosen $\mathcal{G}^{\rm 12}_{{\rm bulk}}=\mathcal{G}^{\rm 23}_{{\rm bulk}}$, with no difference in the low-energy phenomenology. In this case, light-family quark-lepton unification could take place in a bulk between a deeper UV brane (e.g. the Planck brane) and the first-family brane.
The chosen gauge symmetries yield $15+4$ pNGBs, 15 of which become the longitudinal components of the 4321 gauge bosons, $U_1$, $G'$ and $Z'$, which acquire a degenerate mass of $M_{15} \approx \Lambda_{\rm IR} \sqrt{2/(kL)}$~\bibnote{Assuming no boundary kinetic terms for $SU(4)_h\times SU(3)_\ell \times U(1)_{\ell+R}$ part of the gauge symmetry}, with $\Lambda_{\rm IR}\approx k \, e^{-kL}$. This mass generation mechanism (similar to the one in~\cite{Fuentes-Martin:2020bnh}), yields a mass gap between the 4321 gauge bosons and the lightest vector resonances of the Kaluza-Klein (KK) tower, namely $M_{\rm KK}/M_{15} \approx \sqrt{2kL}$. The remaining pNGBs correspond to the SM Higgs field: $H\sim(\mathbf{1},\mathbf{2})_{1/2}$.
\begin{table}[t]
\renewcommand{\arraystretch}{1.2}
\centering
\begin{tabular}{|c|ccc||cc|}
\hline
Field & $SU(4)_h$ & $SU(4)_l$ & $SO(5)$ & $U(1)_\Psi$ & $U(1)_\mathcal{S}$ \\
\hline
\hline
$\Psi^3,\Psi_d^3,\mathcal{X}^{({\prime})}$ & $\mathbf{4}$ & $\mathbf{1}$ & $\mathbf{4}$ & $1$ & $0$\\
$\Psi^j,\Psi_{u,d}^j$ & $\mathbf{1}$ & $\mathbf{4}$ & $\mathbf{4}$ & $1$ & $0$\\
$\mathcal{S}^i$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $0$ & $1$\\
\hline
\hline
$\Sigma$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{5}$ & $0$ & $0$\\
$\Omega$ & $\mathbf{1}$ & $\mathbf{4}$ & $\mathbf{4}$ & $1$ & $-1$\\
$\Phi$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $0$ & $2$\\
\hline
\end{tabular}
\caption{Matter content. Here, $i=1,2,3$ and $j=1,2$. The upper block refers to fermion fields, the lower block to scalars.}
\label{tab:content}
\end{table}
The matter fields and their corresponding transformation properties under the 12-bulk gauge symmetry are listed in Table~\ref{tab:content}. We embed all fermion fields (except $\mathcal{S}^i$) into the spinorial $\mathbf{4}$ representation of $SO(5)$, which contains two (complex) doublets, one transforming in the fundamental of $SU(2)_L$ and the other in the fundamental of $SU(2)_R$. These fermions are also embedded into fundamental representations of $SU(4)_{h,l}$ forming quark-lepton multiplets \`a la Pati-Salam. The boundary conditions (BCs) for the fermions are chosen as follows
\begin{align}\label{eq:FermionBCs}
\Psi^3&=
\begin{bmatrix}
\psi^3\,(+,+)\\[2pt]
\psi_u^3\,(-,-)\\[2pt]
\tilde\psi_d^3\,(+,-)\\
\end{bmatrix}
\,,&
\Psi^3_d&=
\begin{bmatrix}
\tilde\psi^3\,(+,-)\\[2pt]
\tilde\psi_u^3\,(+,-)\\[2pt]
\psi_d^3\,(-,-)\\
\end{bmatrix}
\,,\nonumber\\
\mathcal{X}^{(\prime)}&=
\begin{bmatrix}
\chi^{(\prime)} (\pm,\pm)\\[2pt]
\chi^{(\prime)}_u\,(\mp,\pm)\\[2pt]
\chi^{(\prime)}_d\,(\mp,\pm)\\
\end{bmatrix}
\,,&
\Psi^j&=
\begin{bmatrix}
\psi^j\,(+,+)\\[2pt]
\tilde\psi_u^j\,(-,+)\\[2pt]
\tilde\psi_d^j\,(-,+)\\
\end{bmatrix}
\,,\nonumber\\
\Psi_u^j&=
\begin{bmatrix}
\tilde\psi^j\,(+,-)\\[2pt]
\psi_u^j\,(-,-)\\[2pt]
\hat\psi_d^j\,(+,-)\\
\end{bmatrix}
\,,&
\Psi_d^j&=
\begin{bmatrix}
\hat\psi^j\,(+,-)\\[2pt]
\hat\psi_u^j\,(+,-)\\[2pt]
\psi_d^j\,(-,-) \\
\end{bmatrix}\,,
\end{align}
and $\mathcal{S}^i=S^i(+,+)$,
where we decomposed the spinorial $SO(5)$ multiplets into $SU(2)_L$, up-type $SU(2)_R$ and down-type $SU(2)_R$ components, as required by the $SU(2)_L\times U(1)_R$ symmetry in the UV. The resulting zero modes correspond to the SM field content (including three right-handed neutrinos), $\psi_L^i$ and $\psi_{uR,dR}^i$ with quarks and leptons unified in $SU(4)$ representations, one vector-like representation $\chi_L$ and $\chi_R^\prime$, and three chiral SM singlets $S_L^i$. The SM-singlet fermions, the scalars $\Omega$ and $\Phi$, and the (global) $U(1)_\mathcal{S}$ symmetry are needed to give neutrinos masses via an inverse see-saw mechanism~\cite{Fuentes-Martin:2020pww}. The $\Omega$ field is also responsible for the spontaneous breaking of the UV gauge symmetry down to the 4321 symmetry, where the $U(1)$ factor in 4321 is the diagonal subgroup of $U(1)_l\times U(1)_R$. Furthermore, the $\Omega$ field can play the role of the GW scalar by appropriately choosing its brane-localized potentials and bulk masses. The scalar $\Sigma$ plays a key role in generating the light Yukawa couplings, as discussed below. Finally, the (global) $U(1)_\Psi$ symmetry~\bibnote{These global $U(1)$ symmetries could instead be gauge bulk symmetries broken at a deeper UV brane (like the Planck brane). In this case, anomaly cancellation requires UV-brane fermions with masses at the symmetry breaking scale.} is introduced to forbid baryon and lepton number violating higher-dimensional operators, already present in the Randall-Sundrum (RS) model~\cite{Randall:1999ee} (see e.g.~\cite{Huber:2000ie}).
\subsection{Flavor hierarchies}
\begin{figure}[t]
\includegraphics[width=\columnwidth]{brane_fig_non_universal.pdf}
\caption{Schematic structure of the extra dimension including fermion zero-mode and scalar VEV profiles.}
\label{fig:branes}
\end{figure}
As we discuss in what follows, {\it all} flavor hierarchies in the Yukawas and vector-like masses are explained with $\mathcal{O}(1)$ parameters by assuming the fermion localizations illustrated in Fig.~\ref{fig:branes}, with $kL\approx 10$ and $k\ell \approx 4$. While there is some freedom in the choice of fermion bulk masses that do not affect the solution to the flavor hierarchies, we take the following benchmark for concreteness (with $c_i\equiv M_i/k$ and $M_i$ the corresponding fermion mass)
\begin{align}\label{eq:cCoeff}
c_{\Psi^3}^{(1,2)}&=c_\mathcal{X}^{(1,2)}=c_{\Psi_d^3}^{(1)}=c_{\Psi^2}^{(1)}=c_{\Psi_u^2}^{(1)}=0\,,\;
c_{\mathcal{X}^\prime}^{(1,2)}=\eminus1/2\,,\nonumber\\
c_{\Psi^1}^{(1,2)}&=c_{\Psi^2}^{(2)}=\vcenter{\hbox{\scalebox{0.75}[1]{$ - $}}} c_{\Psi_d^3}^{(2)}=\vcenter{\hbox{\scalebox{0.75}[1]{$ - $}}} c_{\Psi_d^2}^{(1)}=1,\;
c_{\Psi_{u,d}^1}^{(1,2)},c_{\Psi_{u,d}^2}^{(2)}\leq\eminus2\,.
\end{align}
Even though the gauge symmetry in the 23-bulk is $SU(4)_h \times SU(3)_l$, for simplicity we choose $SU(4)$ symmetric bulk masses for this benchmark. Furthermore, we fix $\Lambda_{\rm IR}=8$~TeV such that $M_{15}=3.6$~TeV, which provides a good benchmark for the explanation of the $B$ anomalies~\cite{DiLuzio:2018zxy,Cornella:2019hct,Cornella:2021sby}.
\paragraph{Vector-like masses and fermion mixing.}
The following mass terms are added on the IR brane
\begin{align}
\mathcal{L}_{\rm IR}& \supset \big(\bar{\mathcal{X}}_L \tilde M_\chi + \bar\Psi_L^3 \tilde M_{\Psi} + \bar \Psi_L^j \tilde m_\psi^j\big) \,\mathcal{P}_L \mathcal{X}^\prime_R\,,
\end{align}
in order to generate a vector-like mass among the zero modes. Here $\mathcal{P}_{L,R}$ is a projector into the $SU(2)_{L,R}$ components of the $SO(5)$ multiplets, and the IR masses decompose as $\tilde M_i=\mathrm{diag}(\tilde M_i^q,\tilde M_i^q,\tilde M_i^q,\tilde M_i^\ell)$ in $SU(4)$ space. These mass terms mix all the left-handed components of the zero modes, leaving the SM fermion content as massless chiral fields (plus the SM singlets needed for the inverse see-saw). The vector-like masses thus induced read
\begin{align}
\mathcal{L}&\supset-M_f\,\bar f_L\chi_R^\prime\,,\;\;
M_f=\Lambda_{\rm IR}\,\tilde M_f\, P\big(\{c_f^{(n)}\},\{c_{\mathcal{X}^\prime}^{(n)}\}\big)\,,
\label{VLmass}
\end{align}
where $P(\{c_1\},\{c_2\})$ is a fermion profile function (see~\ref{eq:ProfileFunction}).
The behavior of the profile function is such that, for the fermion profiles in~\eqref{eq:cCoeff}, we have
\begin{align}
\Lambda_{\rm IR}\,P\big(\{c_{\Psi^j}^{(n)}\},\{c_{\mathcal{X}^\prime}^{(n)}\}\big)&\sim \frac{\Lambda_{\rm IR}}{\sqrt{kL}}\,e^{k(\ell_j-L)/2}\approx 2~\mathrm{TeV}\times V_{j3}\,,\nonumber\\
\Lambda_{\rm IR}\,P\big(\{c_{\Psi^3,\mathcal{X}}^{(n)}\},\{c_{\mathcal{X}^\prime}^{(n)}\}\big)&\sim \frac{\Lambda_{\rm IR}}{\sqrt{kL}}\approx 2~\mathrm{TeV}\,,
\end{align}
where $\ell_j$ is the position of the $j$-th brane, and $V_{ij}$ are CKM matrix elements. We thus obtain TeV-scale vector-like masses with a $U(2)$-like mixing structure for the assumed benchmark, hence reproducing the required conditions for a successful explanation of the $B$-meson anomalies~\cite{DiLuzio:2018zxy,Cornella:2019hct,Cornella:2021sby}.
\paragraph{Yukawa couplings.}
Yukawa couplings with the Higgs are generated via three distinct mechanisms, depending on the fermions, giving rise to decreasing effective interactions:
\medskip
{\bf I.} {\it Top Yukawa.} The fermion BCs in~\eqref{eq:FermionBCs} have been chosen such that, in the absence of IR masses, only the third-generation up-type Yukawa is generated:
\begin{align}
\mathcal{L}&\supset - y_u^{33}\,\bar\psi_L^3 H \psi_{uR}^3\,,\,
y_u^{33}=\frac{g_*}{2\sqrt{2}}\, P\big(\{c_{\Psi^3}^{(n)}\},\{c_{\Psi^3}^{(n)}\}\big)\,,
\end{align}
where $g_*$ is the $SO(5)$ KK coupling. In the absence of fermion mixing effects, $y_u^{33}$ becomes the top Yukawa, $y_t$. Since $P(\{c_{\Psi^3}^{(n)}\},\{c_{\Psi^3}^{(n)}\}\big)\lesssim1$, we infer the lower bound $g_{*}\geq 2\sqrt{2}\,y_t(\Lambda_{\rm IR})\approx2.2$. The relation between the EW gauge couplings and the $SO(5)$ KK coupling is ($i=L,R$)
\begin{align}
g_i(\Lambda_{\rm IR})&=g_* /\sqrt{kL(1+r_{{\rm UV},i}^2+r_{{\rm IR},i}^2)}\,,
\end{align}
where $r_{{\rm UV (IR)},i}$ is the contribution from boundary kinetic terms for the corresponding gauge bosons at the UV (IR) brane evaluated at the IR scale~\cite{Csaki:2008zd}. Taking for simplicity $r_{{\rm IR},i}=r_{{\rm UV},i}\equiv r_i$, $kL=10$ and $g_*=2.5$, we find that the EW gauge couplings are reproduced for $r_L\approx0.5$ and $r_R\approx1.4$.
\medskip
{\bf II.} {\it Other third-family Yukawas.} $b$ and $\tau$ Yukawas, as well as the leading mixing among the light families and the third generation, are generated only after introducing IR-brane mass-mixing terms. The relevant IR masses are
\begin{align}
\mathcal{L}_{\rm IR}&\supset \bar\Psi_L^3 \tilde M_{\Psi d }^L \mathcal{P}_L \Psi_{dR}^3+\bar{\mathcal{X}_L} (\tilde M_{\chi d }^L \mathcal{P}_L + \tilde M_{\chi d }^R \mathcal{P}_R)\Psi_{dR}^3\nonumber\\
&+\bar\Psi^j_L \tilde m_{\Psi j}^R \mathcal{P}_R \Psi_R^3+\bar\Psi^j_L (\tilde m_{dj}^L \mathcal{P}_L+\tilde m_{dj}^R \mathcal{P}_R) \Psi_{dR}^3\,,
\end{align}
where we ignored the mass term between the $SU(2)_R$ components of $\Psi^3$ and $\chi$, as well as those with $\Psi_{u,d}^{1,2}$, which have a minor phenomenological impact. These generate Yukawas between zero modes of the form
\begin{align}
y_{f1 f2}=\frac{g_*}{2\sqrt{2}}\, (\tilde M_{12}^L-\tilde M_{12}^R)\, P\big(\{c_{f_1}^{(n)}\},\{c_{f_2}^{(n)}\}\big)\,,
\end{align}
$f_1$ and $f_2$ denoting two generic 5D fermions, and $\tilde M_{12}^{L(R)}$ a generic IR mass between their $SU(2)_{L(R)}$ components.
As anticipated, the hierarchies between these and the top Yukawa are fully explained by appropriate fermion localizations. For the benchmark in~\eqref{eq:cCoeff}, we find
\begin{align}
\begin{aligned}
P\big(\{c_{\Psi^j}^{(n)}\},\{c_{\Psi^3}^{(n)}\}\big)&\sim e^{k(\ell_j-L)/2} \approx V_{j3}\,,\\
P\big(\{c_{\Psi^3,\mathcal{X}}^{(n)}\},\{c_{\Psi_d^3}^{(n)}\}\big)&\sim e^{k(\ell-L)/2}\approx y_b\,,\\
P\big(\{c_{\Psi^j}^{(n)}\},\{c_{\Psi_d^3}^{(n)}\}\big)&\sim e^{k(\ell+
\ell_j-2L)/2}\approx V_{j3}\,y_b \,,
\end{aligned}
\end{align}
where we used that $y_b\approx V_{23}$. Interestingly, we obtain a down-aligned limit (i.e.~vanishing light-heavy entries in the down sector) in the $SO(5)$ symmetric limit, where $M_i^L=M_i^R$. Due to the chosen fermion BCs, an analogous limit in the up sector is not possible.
\medskip
{\bf III.} {\it Light-family Yukawas.} Due to the assumed strong UV localization of $\Psi_{u,d}^{1,2}$, light-family Yukawas are dominantly generated via their coupling with $\Sigma$. The relevant 5D proto-Yukawas are
\begin{align}
\mathcal{L}&\supset - Y_{u,d}^{ij}\,\bar\Psi^i\,\Sigma^a\,\Gamma^a\, P_R\Psi_{u,d}^j\,,
\end{align}
with $\Gamma^a$ ($a=1,\dots,5$) being the $SO(5)$ gamma matrices. These proto-Yukawas can be in the bulk as well as localized in the branes. The $\Sigma$ field decomposes under the EW symmetry as a Higgs $H^\prime$ and a singlet $S$, and acquires a vacuum expectation value (VEV) along the singlet direction with an IR-localized profile. This breaking of $SO(5)$ generates light-family Yukawas suppressed by the $\Sigma$ profile. Taking the $\Sigma$ bulk mass close to the Breitenlohner-Freedman stability
bound~\cite{Breitenlohner:1982jf,Breitenlohner:1982bm}, we find
\begin{align}\label{eq:LightYukawas}
y_{u,d}^{ij} &\approx \frac{g_*}{2\sqrt{2}}\, \tilde Y_{u,d}^{ij}\,
\frac{\langle \Sigma_{\rm IR}\rangle}{\Lambda_{\rm IR}}\,e^{-k (L-\ell_j)}\nonumber \\
&\quad\times e^{-k(c_i^{(1)}-\frac{1}{2})|y_i-\ell_j|}\,e^{k(c_j^{(1)}+\frac{1}{2})|y_j-\ell_j|}\,,
\end{align}
where $y_{i(j)}$ denotes the position of the brane where the left-handed (right-handed) field is dominantly localized, $c_{i(j)}^{(1)}$ is the left-handed (right-handed) 12-bulk mass in units of $k$, and $\tilde Y^{ij}_{u,d}$ are $\mathcal{O}(1)$ linear combinations of the proto-Yukawa couplings with coefficients depending on $k$, $\ell_i$, the fermion bulk masses, and the $\Sigma$ boundary masses (see~\ref{eq:tildeY}).
\subsection{Higgs potential and EW precision data}\label{subsec:HiggsPot}
The Higgs potential receives two types of contributions: i) a tree-level one resulting from the spontaneous breaking of the bulk gauge symmetry via $\Sigma$ and $\Omega$ VEVs, and ii) a loop-level one from a finite volume effect due to non-local operators generated by 5D loops stretching from one boundary to the other~\cite{HOSOTANI1983193}. In our model, the loop contribution comes dominantly from $\Psi^3$ and the EW gauge bosons. For small $h/f$, the Higgs potential is well approximated as
\begin{equation}
V(h) \approx \alpha(h) \cos\left(\frac{h}{f}\right) - \beta(h) \sin^2 \left(\frac{h}{f}\right) \,,
\end{equation}
where $f\approx 2\Lambda_{\rm IR}/g_{*}$ is the Higgs decay constant, $\alpha(h) \approx \alpha_\Omega + \alpha_{\Psi^3}(h)$ and $\beta(h) \approx \beta_\Sigma+\beta_{\rm EW} +\beta_{\Psi^3}(h)$. Using a holographic approach cross-checked by the spectral function method, we find the following expressions for the coefficients $\alpha_i$ and $\beta_i$~\bibnote{In the tree-level computation, we assume large $kL$ and neglect higher order terms in the boundary masses. For the loop via the spectral function, we include only $\Psi^3$ and EW gauge bosons, and we use approximate form factors as discussed in~\cite{Falkowski:2006vi}, which allow for an analytic computation.}
\begin{align}
\begin{aligned}
\alpha_{\Psi^3}(h) &\approx \frac{ 3N_c f^4 }{32\pi^2} \zeta(3)\, y_{t}^2 g_{*}^2 - 2\beta_{\Psi^3}(h)\,,
\\
\alpha_{\Omega}&\approx (\tilde M^R_{\Omega}-\tilde M^L_{\Omega})\,\Lambda_{\rm IR}^2 \langle \Omega_{\rm IR} \rangle^2\,, \\[5pt]
\beta_{\Psi^3}(h) &\approx \frac{ N_c f^4 }{16\pi^2} \, y_{t}^4 \left[\gamma + \log\frac{\Lambda_{\rm IR}^2}{m_{t}^2(h)} \right] \,, \\
\beta_{\rm EW} &\approx - \frac{9f^4}{512\pi^2}\, g_{*}^2\, \zeta(3) \left(3 g_{L}^2+g_{Y}^2\right) \,, \\
\beta_{\Sigma}&\approx \frac{1}{2}(\tilde M_{H^{\prime}}-\tilde M_{S}) \frac{\Lambda_{\rm IR}^2 }{(kL)^2} \langle \Sigma_{\rm IR} \rangle^2\,,
\end{aligned}
\end{align}
where $\gamma \approx 0.38$, $\zeta(3)\approx1.20$, $g_{L(Y)}$ is the $SU(2)_L$ ($U(1)_Y$) gauge coupling, $\langle \Omega_{\rm IR} \rangle$ and $\langle \Sigma_{\rm IR} \rangle$ are IR VEVs, $\tilde M^{L(R)}_{\Omega}$ are IR masses for the $SU(2)_{L(R)}$ components of $\Omega$, and $\tilde M_{H^{\prime},S}$ are UV masses for $\Sigma$ (all masses in units of $k$). Approximating $\alpha$ and $\beta$ as constants, the minimum of the potential and the Higgs mass are given by
\begin{align}
\cos(\langle h\rangle /f) &= -\frac{\alpha}{2\beta}\,\,,&
m_h^2\equiv 2\lambda \langle h\rangle^2 \approx \frac{2\beta\langle h\rangle^2}{f^4}\,.
\end{align}
As we can see, the loop contributions $\alpha_{\Psi^3},\beta_{\Psi^3},\beta_{\rm EW}$ are completely fixed once a value for $g_*$ (which also enters in the top Yukawa) is specified.
Interestingly, we find that the coefficients $\alpha_i$ and $\beta_i$ are of the right size such that the Higgs quartic comes out at the observed value for a natural choice of the undetermined tree-level parameters~\bibnote{We expect $\langle \Sigma_{\rm IR} \rangle \approx \mathcal{O}(\Lambda_{\rm IR})$. However, due to the localization of $\Omega$ in the UV, we can naturally have $\langle \Omega_{\rm IR} \rangle \lesssim \Lambda_{\rm IR }$.}. While the $\alpha_i$ and $\beta_i$ are all of the same order, obtaining the required hierarchy between $\langle h\rangle$ and $f$ implies a tuning of the parameters (the so-called little hierarchy problem), which is at the per mille level for $\Lambda_{\rm IR}=8$~TeV and $g_*=2.5$. We have verified that the results of this simplified computation hold up to small corrections when treating the full loop potential numerically including all fields dominantly localized in the IR, as well as the relevant IR boundary masses.
Additionally, our theory predicts a tower of KK vector resonances which couple dominantly to the IR localized fields. The most dangerous of these states are those that mix with the SM EW gauge bosons, as they induce $g_*/g_L$ enhanced modifications to their couplings with third generation SM fermions. Resumming the KK tower, we find corrections of the form
\begin{equation}
\frac{\delta g_{Z \Psi^3 \Psi^3}}{g_{Z \Psi^3 \Psi^3}} \approx -0.3 \frac{m_Z^2}{M_{\rm KK}^2} \frac{g_*^2}{g_L^2} \approx - \frac{0.3}{4c_W^2} \frac{\langle h\rangle^2}{f^2} \,,
\label{eq:deltagOng}
\end{equation}
where $c_W$ is the cosine of the Weinberg angle and the pre-factor of $0.3$ comes dominantly from the Higgs and $\Psi^3$ profiles. The strongest bound comes from $Z\rightarrow \tau_L \tau_L$, leading to a constraint on~\eqref{eq:deltagOng} at the per-mille level~\cite{Efrati:2015eaa}, which is well satisfied for our benchmark point.
\section{Conclusions}\label{sec:conclusions}
We have presented a UV extension of the SM that addresses two of its long-standing open issues: the origin of flavor hierarchies and the stabilization of the Higgs sector, while, at the same time, explaining the observed anomalies in $B$ decays. A coherent solution to these three problems is obtained by embedding the SM into a warped 5D construction with three (flat) four-dimensional branes, where each SM family is quasi-localized. The 3-brane structure, which lets us associate flavor indices to well-defined positions in the extra dimension, is a crucial distinction from previous explanations of the flavor hierarchies in the context of warped extra dimensions~\cite{Grossman:1999ra,Nelson:2000sn,Agashe:2004cp,Csaki:2008zd}. This structure results in an approximate $U(2)^n$ flavor symmetry with leading breaking in the left-handed sector, which is necessary in order to evade the tight bounds on new physics from flavor-changing processes while simultaneously addressing the $B$
anomalies~\cite{Buttazzo:2017ixm,Barbieri:2015yvd}.
We emphasize that the explicit model analyzed here is part of a larger class of theories, based on the three fundamental points presented in the Introduction. A few building blocks, such as the choice of the IR-bulk and UV-brane symmetries, are motivated by observations. Other aspects (especially those related to UV dynamics) are less constrained and could be modified. The geometry itself is a minimal choice, and the validity of the construction could be extended up to the Planck scale by adding an additional UV brane, solving the large EW hierarchy problem as in the original RS model~\cite{Randall:1999ee}.
By construction, the TeV-scale phenomenology of this model is equivalent to that of 4321 models discussed in the recent literature~\cite{DiLuzio:2018zxy,Cornella:2019hct,Cornella:2021sby}. However, deviations are expected around $M_{\rm KK} \sim 10$~TeV due to the tower of KK states. A further striking signature, specific of the multi-scale (multi-brane) structure of the theory, is a multi-peaked stochastic gravitational wave signal potentially within reach of future experiments, originating from a series of phase transitions in the early universe~\cite{Greljo:2019xan}.
\vspace{1cm}
\subsection*{Acknowledgments}
This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme under grant agreement 833280 (FLAY), and by the Swiss National Science Foundation (SNF) under contract 200020\_204428. The work of JF has been supported by the Spanish Ministry of Science and Innovation (MCIN) and the European Union NextGenerationEU/PRTR under grant IJC2020-043549-I, by the MCIN grant PID2019-106087GB-C22, and by the Junta de Andaluc\'ia grants P18-FR-4314 (FEDER) and FQM101.
\begin{widetext}
| c3f3e1ab9a2449e12457b4f2c845fe0eea9e5870 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:1}
Understanding how proto-planetary discs evolve and eventually dissipate is fundamental to building a comprehensive picture of the planet-formation process \citep[e.g.,][]{Manara2019_pps}. It is commonly assumed that proto-planetary discs evolve under the effect of viscosity, which, allowing for angular momentum re-distribution within the disc, drives accretion onto the forming star \citep{Shakura&Sunyaev1973,Lynden-Bell&Pringle1974}. In this framework, the mass accretion rate on the central star, $\dot{M}_\mathrm{acc}$, and the disc mass, $M_\mathrm{disc}$, are expected to be correlated at any given age \citep[e.g.,][]{Hartmann1998,Jones2012}:
\begin{equation}\label{eq:1}
M_\mathrm{disc}=t_\mathrm{acc}\dot{M}_\mathrm{acc}\propto(t+t_\nu)\dot{M}_\mathrm{acc},
\end{equation}
where $t_\nu$ is the initial viscous time scale and we call $t_\mathrm{acc}$ the accretion time scale\footnote{This is also known as the disc lifetime \citep[e.g.,][]{Lodato2017}.}.
The advent of new facilities, such as the Atacama Large Millimeter/submillimeter Array (ALMA) and the X-shooter spectrograph at the Very Large Telescope (VLT), has made it possible to test disc evolutionary models observationally. In the last years, disc masses, estimated from (sub-)millimetre dust fluxes under the assumption of optically thin emission and a constant gas-to-dust ratio \citep[e.g.,][]{Ansdell2016,Ansdell2018,Pascucci2016,Barenfeld2016}, as well as mass accretion rates \citep[e.g.,][]{Alcala2014,Alcala2017,Manara2016_cham,Manara2017,Manara2020}, have been inferred for a large number of young stellar objects.
Combining these data-sets, \citet{Manara2016} and \citet{Mulders2017} detected a slightly sub-linear correlation between mass accretion rates and proto-planetary disc masses in the young Lupus and Chamaeleon~I star-forming regions (aged $\approx1\text{ to }3\,\mathrm{Myr}$), that was interpreted as being compatible with viscous-diffusion models \citep{Rosotti2017,Lodato2017,Mulders2017}. A correlation was also identified by \citet{Manara2020} in the older Upper Scorpius OB association (aged $\approx5\text{ to }10\,\mathrm{Myr}$). However, at any given disc mass, the median accretion rate in Upper Sco is remarkably similar to that in young regions; moreover its scatter has not decreased with time. This is in contrast with the viscous evolution scenario, which predicts both lower $\dot{M}_\mathrm{acc}$ and a tighter relation at later times (\citealt{Rosotti2017,Lodato2017} and Eq.~\ref{eq:1}).
Nevertheless, all the previous analyses of the correlation only focused on \textit{gas} evolution. Instead, the disc mass estimates of \citet{Manara2016,Manara2020} and \citet{Mulders2017} are based on (sub-)millimetre \textit{dust} emission and are subject to systematic uncertainties (e.g., on dust opacity and the gas-to-dust ratio).
To try to explain the observations, both gas and dust ought to be considered when modelling disc evolution. Moreover, theoretical predictions and observational trends must be compared in the \textit{data space}, i.e. post-processing the model outputs to determine their (sub-)millimetre dust fluxes and masses as for real data. Such an exercise then removes the use of the restrictive assumptions about optical depth, opacities and the gas-to-dust ratio that are generally used to interpret (sub-)millimetre observations. \citet{Sellek2020} adopted such a forward modelling approach and demonstrated that viscous simulations provide a remarkable agreement with both Lupus and Upper Sco data (in particular when internal photo-evaporation is included to account for late-time disc dispersal). However, even though the models of \citet{Sellek2020} accounted for the underestimation of disc masses in current surveys, none of them were able to reproduce discs with accretion time scale shorter than $t_\mathrm{acc}\lesssim0.1\,\mathrm{Myr}$, which contribute to most of the large scatter/high accretion rates in Upper Sco.
As mass accretion rates are computed at a fixed snapshot in time, it might be thought that accretion variability is responsible for the outlying sources. Unfortunately, there are no studies addressing this issue at the age of Upper Sco \citep{Manara2020}, and very little is known about accretion variability on time scales longer than few days. In younger ($t\lesssim3\,{\rm to}\,4\,{\rm Myr}$) discs, $\dot{M}_{\rm acc}$ variations are $\lesssim0.4\,{\rm dex}$ on a time scale of days \citep{Costigan2014,Venuti2014,Manara2021}, not enough to account for the highest accretors in Upper Sco, with only a small fraction of the targets having larger variability \citep{Audard2014}.
All the previous considerations are based on the assumption that planet-forming discs evolve in isolation. This picture is clearly idealised: the majority of stars are born in clusters, where nearby sources can influence proto-planetary disc evolution history and affect their planet-formation potential \citep[e.g.,][]{Winter2020}. This commonly occurs by either external photo-evaporation, when the UV radiation of a massive nearby star dissipates the less bound gas in the outer disc regions \citep[e.g.,][]{Adams2004,Facchini2016}, or by tidal interactions, when a stellar companion in a gravitationally bound pair, or a flyby in a dense environment, truncates the disc \citep[e.g.,][]{Winter2018,Cuello2019,Cuello2020}. These processes have qualitatively similar effects, both reducing disc masses and sizes (e.g., \citealt{Harris2012,Akeson&Jensen2014,Cox2017,Akeson2019,Manara2019} in binaries and \citealt{Ansdell2017,Otter2021} for the effects of photo-evaporation), eventually hastening disc dispersal.
How external photo-evaporation and tidal truncation impact the correlation between mass accretion rates and disc masses was studied by \citet{Rosotti2017} and \citet{Rosotti&Clarke2018}, respectively. These studies showed that both discs exposed to strong UV fields and discs in multiple stellar systems are expected to have shorter accretion time scales than those (viscously) evolving in isolation. This means that the influence of the environment could be a potential explanation for those highly-accreting old discs with low masses that the models of \citet{Sellek2020} were not able to explain. However, these studies only focused on gas evolution. Since dust was not considered as a separate component, it was simply assumed that observationally inferred disc dust masses could be converted into total disc masses using a gas-to-dust ratio of 100. This motivates our present study based on modelling both gas and dust evolution and then comparing simulation-based synthetic observations with real data.
The required properties of star-forming environments for which external photo-evaporation and tidal encounters are effective were studied by \citet{Winter2018_trunc}. They showed that stellar densities $\gtrsim10^4\,{\rm pc}^{-3}$ are necessary for disc truncation by tidal encounters. This threshold is much higher than in the nearby star-forming regions, making fly-bys unlikely. As for external photo-evaporation, \citet{Winter2018_trunc} concluded that average UV fields $\gtrsim3\times10^3\,G_0$ are required to disperse a disc. (Here $G_0$ corresponds to an energy flux of $1.6\times10^{-3}\,{\rm erg}\,{\rm cm}^{-2}$ between $6\,{\rm and}\,13.6\,{\rm eV}$, \citealt{Habing1968}.) In Lupus the average radiation field is expected to be remarkably low (e.g., in IM~Lup it is $\lesssim4\,G_0$, \citealt{Cleeves2016}) while in Upper Sco it is somewhat higher (on average $42.9\,G_0$, \citealt{Trapman2020}). Since the median UV field in both regions is below the critical value for evaporation, we decided not to consider external photo-evaporation in this paper\footnote{In Upper Sco the radiation levels can grow up to $7\times10^3\,G_0$ \citep{Trapman2020}, suggesting that external photo-evaporation can influence the accretion time scales of some sources. We plan to study this effect in a future paper.}.
Instead, in this work we focus on the role played by stellar multiplicity in disc evolution, confronting our theoretical expectations with the mass accretion rates and sub-millimetre fluxes from nearby young star-forming regions. First of all, from the currently published catalogues, we collect a state-of-the-art sample of homogeneously determined disc masses and accretion rates in Lupus, Chamaeleon~I and Upper Sco. This sample is further updated making use of Gaia EDR3 distances \citep{GAIAeDR3} as in the PPVII chapter of Manara et al. (subm.). Furthermore, searching the literature allows us to identify several pairs, i.e. gravitationally bound stars with projected separation, $a_\mathrm{p}$, less than $300\,\mathrm{au}$, in binaries or higher order multiple stellar systems among the previously selected sources. Subsequently, we run models of tidally truncated circumstellar binary discs for a large number of initial disc parameters following the approach of \citet{Zagaria2021_theo}. To test our hypothesis that the accretion time scale is influenced by a companion, these are then compared with observations through modelling the predicted dust emission.
This paper is organised as follows. In Section~\ref{sec:2} we introduce the Lupus, Chamaeleon~I and Upper Sco samples. Discs in multiple stellar systems are identified (see Appendix~\ref{app:1}) and their properties discussed in comparison to those in single-star systems. In Section~\ref{sec:3} dust and gas modelling is introduced, while Section~\ref{sec:4} describes our results, first of all discussing their dependence on parameter choices and then comparing models and observations. In Section~\ref{sec:5} we consider the model limitations and possible improvements to the agreement with observational data. Finally, in Section~\ref{sec:6} we draw our conclusions.
\section{Sample selection and data analysis}\label{sec:2}
\subsection{Sample selection}
We focus on Lupus, Chamaeleon~I and Upper Sco because they are among the best studied nearby star-forming regions: (sub-)millimetre dust fluxes and accretion rates were measured for a large fraction of their young stellar objects (YSOs), and their multiplicity fraction is often well known.
The distances to individual targets are taken from inverting the parallaxes available in the Gaia EDR3 catalog \citep{GAIAeDR3}. The only exception are the cases where the Gaia parallaxes are unreliable, i.e. ${\rm RUWE}>1.8$ and/or distance differing more than $60\,{\rm pc}$ from the median distance to the region, or unavailable. In these cases, we assumed the median distance to the members of the region: $d=158\,{\rm pc}$ for Lupus, $d=190\,{\rm pc}$ for Chamaeleon~I, and $d=145\,{\rm pc}$ for Upper Sco.
In Lupus, we make use of the (sub-)millimetre dust emission measured by \citet{Ansdell2016}, with the addition of the Lupus completion survey sources of \citet{Sanchis2020}, Sz~82/IM~Lup \citep{Cleeves2016} and Sz~91 \citep{Tsukagoshi2019}.
We convert the ALMA (sub-)millimetre dust fluxes into dust masses assuming optically thin emission \citep{Hildebrand1983}:
\begin{equation}\label{eq:2}
M_\mathrm{dust}=\dfrac{d^2F_\nu}{B_\nu(T_\mathrm{dust})\kappa_\nu^\mathrm{abs}},\text{ where }\kappa_\nu^\mathrm{abs}=2.3\left(\dfrac{\nu}{230~\mathrm{GHz}}\right)\text{cm}^2\text{g}^{-1}.
\end{equation}
Here $F_\nu$ is the dust flux, $d$ the distance of the source, $B_\nu$ the black body emission at temperature $T_\mathrm{dust}=20\,\mathrm{K}$, and $\kappa_\nu^\mathrm{abs}$ the absorption opacity \citep{Beckwith1990}. Disc masses are then computed from dust masses assuming a standard ISM gas-to-dust ratio of 100 \citep{Bohlin1978}. A target is a non-detection if its dust flux does not exceed three times the continuum rms noise.
Mass accretion rates are inferred from the VLT/X-shooter spectroscopic measurements of \citet{Alcala2014,Alcala2017}. Briefly, stellar properties and the accretion parameters are derived self-consistently fitting the YSO spectra with the sum of the photospheric template of a non-accreting star and a slab model to reproduce the accretion luminosity excess in the Balmer continuum and also in the optical part of the spectrum, as described in \citet{Manara2013}. Following Manara et al. (subm.), the effective temperature, $T_{\rm eff}$, of the stars were re-calculated using the relation between spectral type and $T_{\rm eff}$ by \citet{Herczeg&Hillenbrand2014}, and the stellar luminosity and accretion luminosity were re-scaled to the distances from Gaia EDR3. To derive the stellar masses, needed to calculate the mass accretion rates, the non-magnetic evolutionary tracks by \citet{Feiden2016} and \citet{Baraffe2015} where used for targets hotter and colder than 3900 K, respectively, while the models by \citet{Siess2000} were used for sources appearing too young/old for the former tracks. Under-luminous objects (see \citealt{Alcala2014}), i.e., falling well below the main locus of pre-main sequence targets in the HR diagram, are excluded. If the excess continuum UV luminosity of a YSO with respect to the best-fit photospheric template is close to its chromospheric noise level, it is termed a non-accretor and its mass accretion rate is considered an upper-limit \citep[see][]{Alcala2017,Manara2017}. The accretion rate of MY~Lup is as computed from the HST observation results of \citet{Alcala2019}. After updating stellar distances, an average uncertainty of $0.35\,{\rm dex}$ is estimated.
In Chamaeleon~I and Upper Sco we follow a similar procedure. In the former region, we use the (sub-)millimetre dust emission of \citet{Pascucci2016} and \citet{Long2018_cham}, and the accretion luminosities of \citet{Manara2016_cham,Manara2017}. In the latter region, we rely on \citet{Manara2020} sample for mass accretion rates, while disc masses are computed using the (sub-)millimetre dust emission of \citet{Barenfeld2016}.
Transition discs (TDs) are identified as both discs with a ``resolved'' (imaged) cavity and ``unresolved'' discs, whose classification is based on their SED shape as in \citet{Manara2016,Manara2020} and \citet{Mulders2017}. Finally, we remark that Lupus and Chamaeleon~I samples are more than 90 per cent complete\footnote{In this paper, by saying that a given star-forming region is $x$ per cent complete, we mean that disc accretion rates were measured for a fraction $x/100$ of the YSOs with discs in that region.} \citep{Alcala2017,Manara2017}, while Upper Sco has high levels of completeness only for $M_\mathrm{disc}\gtrsim5\times10^{-5}\,M_\mathrm{Sun}$ \citep{Manara2020}\footnote{More specifically, \citet{Manara2020} claimed completeness to be 80 per cent in the mass range $4.8\times10^{-5}\,M_\mathrm{Sun}\leq M_\mathrm{disc}\leq6.47\times10^{-4}\,M_\mathrm{Sun}$ and 60 per cent in the mass range $6.47\times10^{-4}\,M_\mathrm{Sun}\leq M_\mathrm{disc}\leq1.55\times10^{-2}\,M_\mathrm{Sun}$ \citep{Manara2020}. Nevertheless, these are likely overestimates if the refined membership census of YSOs with discs in Upper Sco \citep{Luhman&Esplin2020} is considered.}.
We searched the literature for information on stellar multiplicity for the previously selected YSOs. Our main references are \citet{Zurlo2021} in Lupus, \citet{Lafreniere2008} in Chamaeleon~I and \citet{Barenfeld2019} in Upper Sco. A total of 9 binaries or higher-order multiple stellar systems with projected separation less than $300\,\mathrm{au}$ are found is Lupus, 16 in Chamaeleon and 9 in Upper Sco. 11 of these are in pairs closer than $40\,\mathrm{au}$ \citep{Kraus2012}. We refer the reader to Appendix~\ref{app:1} for a detailed discussion of the multiplicity detection methods and comparison among regions.
\subsection{Data analysis}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_2.1_alpha.pdf}
\caption{\textbf{Left panel:} Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the disc mass, $M_\mathrm{disc}=M_\mathrm{dust}\times100$, for Lupus, Chamaeleon~I and Upper Sco Class~II sources. Grey dots are used for isolated stars, orange and blue for binary ($a_\mathrm{p}<300\,\mathrm{au}$) and close binary ($a_\mathrm{p}<40\,\mathrm{au}$) discs, and green for circumbinary discs. Transition discs are identified by a wider circle, non-accretors by a downward-pointing triangle and non-detections by a left-pointing arrow. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1$ and $10\,\mathrm{Myr}$ are shown as dot-dashed lines. Letters A, B and C are used to label different binary disc populations (see main text). \textbf{Right panel:} Accretion time scale, $t_\mathrm{acc}$, cumulative distribution for single-star and binary discs in grey and orange, respectively. Solid lines are used for the full sample, dashed lines when non-accretors are excluded. The caption lists $p$-values for the null hypothesis of binaries and singles being drawn from the same $t_{\rm acc}$ distribution.}
\label{fig:2.1}
\end{figure*}
\paragraph*{General population trends} In the left-hand panel of Fig.~\ref{fig:2.1} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the disc mass inferred from (sub-)millimetre dust emission, $M_\mathrm{disc}=M_\mathrm{dust}\times100$, for all Lupus, Chamaeleon and Upper Sco discs in our sample. Grey dots are used for isolated stars, orange and blue for binary ($a_\mathrm{p}<300\,\mathrm{au}$) and close binary ($a_\mathrm{p}<40\,\mathrm{au}$) discs, while green for circumbinary discs. Non-accretors are displayed as downward-pointing triangles, while disc mass non-detections are identified by left-pointing arrows. Adopting the same convention of \citet{Manara2016}, transition discs can be recognised by a larger external circle of the same colour. In the background, lines of constant accretion time scale for $t_\mathrm{acc}=0.1,\,1\,\mathrm{and}\,10\,\mathrm{Myr}$ are displayed as grey dashed-dotted lines.
Single-star discs and discs in binary systems populate different regions of the data-space. The former span all the range of observed disc masses and accretion rates, while the latter are characterised by a systematically shorter accretion time scale and can be divided into three sub-regions. The bulk of discs in binaries (A) have $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$. The remainder have longer accretion time scales, with $t_\mathrm{acc}\gtrsim1\,\mathrm{Myr}$. Some of these are very faint (B), often undetected in the (sub-)millimetre, with $M_\mathrm{disc}\lesssim2\times10^{-4}\,M_\mathrm{Sun}$, while a small number are massive (C), with $M_\mathrm{disc}\gtrsim2\times10^{-3}\,M_\mathrm{Sun}$. Remarkably, very close binaries ($a_\mathrm{p}<40\,\mathrm{au}$) mainly live in region A, and $5/11\approx45$ per cent have accretion time scales much shorter than $0.1\,\mathrm{Myr}$, because of high accretion rates and (generally) low
disc masses.
To prove our inference of a shorter accretion time scale in discs in multiple systems we perform a Kolmogorov–Smirnov (K-S) test on the full sample made up of all data from the three star-forming regions together, excluding the three circumbinary discs in the sample. The null-hypothesis of single-star and binary discs being drawn from the same $t_\mathrm{acc}$ distribution is rejected with $p-$values of $7.00\times10^{-7}$ and $3.54\times10^{-6}$, when non-accretors are considered or excluded, respectively. This is visually displayed in the right-hand panel of Fig.~\ref{fig:2.1}, where the accretion time scale cumulative distributions are plotted for single-star and binary discs in grey and orange, respectively. Solid lines are used for the full sample, while dashed ones when non-accretors are not considered. Clearly, discs in multiple systems have systematically shorter $t_\mathrm{acc}$ than those evolving in isolation.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_2.2_alpha.pdf}
\caption{Same as in the left-hand panel of Fig.~\ref{fig:2.1}, with each star-forming region in a different sub-figure. \textbf{From left to right:} Lupus, Chamaeleon~I and Upper Sco. The 80 per cent completeness mass range \citep{Manara2020} is highlighted in grey in the latter region. Captions as in the right-panel of Fig.~\ref{fig:2.1}.}
\label{fig:2.2}
\end{figure*}
Combining the targets from different regions allows us to collect a large enough sample to make statistically significant comparisons between single-star and binary discs. This is a fair procedure as long as it does not introduce any multiplicity biases in the target selection. Although it can be argued that only Upper Sco discs with $5\times10^{-5}\lesssim M_\mathrm{disc}/M_\mathrm{Sun}\lesssim7\times10^{-4}$ have been surveyed for accretion rates with high completeness, such incompleteness does not discriminate between singles and binaries. Indeed, \citet{Barenfeld2019} showed that, in this region, discs in binaries and around singles span a similar (sub-)millimetre brightness range
Finally, we briefly comment on the treatment of (sub-)millimetre emission upper limits in our sample, highlighting that the significant difference between the accretion time scale in binary and single-star discs is not the result of uncertainties associated with how dust mass upper limits are assigned to non-detections. We test this claim in the worst case scenario, i.e. where we assign to non-detections values which make the data-sets for singles and binaries as similar as possible. Performing a K-S test on the new cumulative curve gives a $p-$values of $2.22\times10^{-4}$
\paragraph*{Analysis by star-forming region} While the combined data from Lupus, Chamaeleon~I and Upper Sco clearly demonstrates the difference between singles and binaries, it washes out possible evolutionary effects resulting from the different ages of the regions. For this reason, in Fig.~\ref{fig:2.2} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the disc mass, $M_\mathrm{disc}=M_\mathrm{dust}\times100$ for Lupus, Chamaeleon~I and Upper Sco individually.
Fig.~\ref{fig:2.2} shows that, while binary discs in the younger regions belong to all the previously identified families (A to C), only family A is represented in Upper Sco. To understand whether this is a disc evolution or sample-selection effect, further data are needed. We perform a K-S test on Lupus, Chamaeleon~I and Upper Sco separately, finding that the null-hypothesis that single-star and binary discs are drawn from the same $t_\mathrm{acc}$ distribution is rejected only in the latter (the K-S test $p-$values are reported in the bottom right of each panel in Fig.~\ref{fig:2.2}). Again, assessing whether this is an evolutionary or sample-selection effect needs a higher level of completeness in Upper Sco and a homogeneous analysis of the stellar multiplicity among the three star-forming regions (see Appendix~\ref{app:1}). \\
\noindent To sum up, our data show that binary discs have statistically shorter accretion time scales than those around single stars. Further data and information on disc multiplicity are needed to compare different star-forming regions from the evolutionary point of view.
\section{Numerical methods}\label{sec:3}
In the remainder of this paper our focus is on exploring if the observed trend of a shorter accretion time scale in binaries is compatible with the expectations of models for gas and dust evolution in viscously evolving discs subject to tidal truncation by a stellar companion. To do so, we run a number of binary disc simulations that are post-processed to compute their (sub-)millimetre dust emission profiles. Throughout our analysis we use these synthetic observations to evaluate model disc properties (e.g., masses and accretion time scales) employing the same standard assumptions used in the literature (cf \citealt{Sellek2020} in the case of single-star discs). Those properties are then compared with real data.
We model gas and dust evolution in circumstellar binary discs, neglecting re-supply of material from beyond the binary orbit and assuming that the disc cannot spread outside the tidal truncation limit imposed by the companion. For the gas we consider viscous evolution and omit the effects of photo-evaporation. For the dust the two-population model of \citet{Birnstiel2012} is employed. As a post-processing step, dust emission and accretion rates from our models are used to generate synthetic observations to be compared with real data.
The one-dimensional finite-differences code developed by \citet{Booth2017} is used, modified to take into account tidal truncation as \citet{Zagaria2021_theo} did. Briefly, the viscous-diffusion equation \citep{Lynden-Bell&Pringle1974} is solved on a grid made up of 250 cells equally spaced in $R^{1/2}$, assuming zero-flux \citep{Bath&Pringle1981,Rosotti&Clarke2018} at the outer boundary, $R=R_\mathrm{trunc}$. We choose an exponentially-tapered power-law initial condition:
\begin{equation}\label{eq:3}
\Sigma(R,t=0)=\dfrac{M_0}{2\pi R_0R}\exp\left(-\dfrac{R}{R_0}\right),
\end{equation}
where $R_0$ is a characteristic scale radius, and $M_0$ is the initial disc mass within $R_\mathrm{trunc}$. We explore the region of the parameter space corresponding to $R_0=10$ and $100\,\mathrm{au}$, and $M_0=1$, 3, 10, 30 and $100\,M_\mathrm{Jup}$ \citep{Sellek2020}.
A passively irradiated time-independent disc temperature profile \citep{Chiang&Goldreich1997} is considered, radially decaying as $R^{-1/2}$ and with reference temperature $T_0=88.23$~K at 10~au, calibrated on a Solar-mass star. The \citet{Shakura&Sunyaev1973} prescription is used for viscosity: we set $10^{-4}\leq\alpha\leq10^{-3}$, both for consistency with previous works \citep{Sellek2020} and because at these viscosities similar models reproduce the flux-radius correlation most closely \citep{Rosotti2019_fr,Zagaria2021_obs,Zormpas2022}.
For the dust, a uniform initial dust fraction of 0.01 is employed. The routine for grain growth implements the simplified treatment of \citet{Birnstiel2012}. In short, at each disc radius two dust populations are evolved in time: a population of small grains, whose size is the monomer grain size, $a_\mathrm{min}=0.1\,\mu\mathrm{m}$, and a population of large grains, dominating the mass, whose size, $a_\text{max}$, is determined by the combined effect of grain growth (with a grain-growth efficiency $f_\mathrm{grow}=1$, \citealt{Booth&Owen2020,Sellek2020}), fragmentation (with fragmentation velocity $u_\mathrm{f}=10\,\mathrm{m}\,\mathrm{s}^{-1}$, \citealt{Gundlach&Blum2015}) and radial drift. We refer to \citet{Birnstiel2012} for more information and to \citet{Booth2017} for the specific implementation. Finally, the dust fraction is advected along the gas flow, following the prescription of \citet{Laibe&Price2014} and taking into account the dust back-reaction on gas \citep[e.g.,][]{Dipierro2018,Garate2020}.
We compare models and observations in the data space. This means that our simulation results are post-processed to determine how they would look if observed with ALMA at the same wavelength as in \citet{Ansdell2016} and \citet{Barenfeld2016}. Their mass is then computed as in these surveys and compared with real data. We call this quantity the \textit{observer's equivalent} disc mass \citep{Sellek2020}. Doing so, we can be agnostic on the assumptions made to compute disc masses in the observations and our results would not be influenced by the latter. Our procedure is as follows. For each of our models, we first determine a synthetic surface brightness, $S_\mathrm{b}$, at frequency $\nu\approx338.75\,{\rm GHz}$ (ALMA band 7):
\begin{equation}\label{eq:4}
S_\text{b}(R)=B_\nu(T)\bigl\{1-\exp{(-\kappa_\nu\Sigma_\text{d})}\bigr\},
\end{equation}
where $\Sigma_\text{d}$ is the dust surface density, $B_\nu$ is the black body radiation spectrum at temperature $T$ and $\kappa_\nu$ the dust (absorption) opacity, computed as in \citet{Tazzari2016} and \citet{Rosotti2019_radii}. We assumed face-on discs. If one considers the distribution of disc inclinations on the sky, it can be shown that this approximation is correct within a factor of $\langle\cos i\rangle=\pi/4\approx0.8$. Moreover, \citet{Tazzari2017} showed that correcting for the disc inclination improves the disc masses in \citet{Ansdell2016} by less than a factor of 2. Then, the \textit{observer's equivalent} dust mass, $M_\mathrm{dust}$, is computed as in Eq.~\ref{eq:2}, where the disc luminosity, $L_\nu=d^2F_\nu$, is given by the surface integral of the synthetic surface brightness.
\section{Results}\label{sec:4}
In this Section we outline our results, discussing how the models are affected by the initial parameters. Subsequently, models and observations are confronted, in order to assess their compatibility.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_4.1_alpha.pdf}
\caption{Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the observer's equivalent disc mass (see Eq.~\ref{eq:2}), $M_\mathrm{dust}\times100$. The solid tracks identify the $1\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval. The dependence on the truncation radius, $R_\mathrm{trunc}$ (top left), the scale radius, $R_0$ (top right), and viscosity, $\alpha$ (bottom left), are explored. In the top-right panel, solid lines refer to singles, dashed to binaries. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10.0\,\mathrm{Myr}$. Lupus and Upper Sco data are shown in grey for singles and black for binaries. Larger circles identify transition discs and downward-pointing triangles are used for non-accretors.}
\label{fig:4.1}
\end{figure*}
\subsection{Model dependence on the initial disc parameters}
As a starting point, we illustrate how our results depend on the initial disc parameters. In Fig.~\ref{fig:4.1} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is shown as a function of the observer's equivalent disc mass, $M_\mathrm{dust}\times 100$. The colour dotted lines identify model evolutionary tracks and their solid sections highlight the $1\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ age interval. In the background, lines of constant accretion time scale for $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10\,\mathrm{Myr}$ are displayed as grey dashed-dotted lines. Lupus and Upper Sco discs are also plotted in grey for singles and black for binaries, respectively.
Let us begin with the upper-left panel of Fig.~\ref{fig:4.1}, where models with different values of the tidal truncation radius, $R_\mathrm{trunc}$, are plotted for $M_0=30\,M_\mathrm{Jup}$\footnote{We adjust the surface density normalisation so as to achieve a constant total disc mass, $M_0$, within $R_\mathrm{trunc}$. Our models are gravitationally stable, with \citet{Toomre1964} factor $Q\gtrsim1$, for $R_\mathrm{trunc}\gtrsim10\,\mathrm{au}$.}, $R_0=30\,\mathrm{au}$ and $\alpha=10^{-3}$. It can be seen from the plot that the models populate different regions of the plane. In particular, as the truncation radius decreases, the accretion time scale, $t_\mathrm{acc}$, also decreases. This reduction in $t_\mathrm{acc}$ with $R_\mathrm{trunc}$ is apparent even in the initial conditions and results from the fact that the initial accretion rate is higher for a compact, high density disc and the millimetre flux is somewhat smaller due to optical depth effects. For $t\gg0$, this trend is determined by binary discs both being fainter and accreting more. The former effect is due to the faster drift of grains for smaller values of $R_\mathrm{trunc}$ \citep{Zagaria2021_theo}, which precipitates mass depletion \citep{Zagaria2021_obs}; the latter is a consequence of the smaller viscous time scale of truncated discs \citep{Rosotti&Clarke2018}. Remarkably, while in the case of single-star models, evolutionary tracks are limited by $t_\mathrm{acc}=0.1\,\mathrm{Myr}$ for $R_0=10\,\mathrm{au}$ \citep{Sellek2020}, when $R_\mathrm{trunc}\leq100\,\mathrm{au}$, the binary models are able to populate the region of the data-plane with shorter accretion time scales, where many of the observed sources lie.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_4.2_alpha.pdf}
\caption{Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the observer's equivalent disc mass (see Eq.~\ref{eq:2}), $M_\mathrm{dust}\times 100$, for models with $\alpha=3\times10^{-4}$, $R_\mathrm{trunc}=51.48\,\mathrm{au}$ and $R_\mathrm{trunc}=18.13\,\mathrm{au}$ on the left and right panel, respectively. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10.0\,\mathrm{Myr}$. \textbf{Left:} The solid tracks identify the $1\,\mathrm{Myr}\leq t\leq3\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval; Lupus discs are shown in grey for singles and black for binaries. \textbf{Right:} The solid tracks identify the $5\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval; Upper Sco discs are shown in grey for singles and black for binaries. Larger circles identify transition discs and downward-pointing triangles are used for non-accretors. The colour bars refer to both panels.}
\label{fig:4.2}
\end{figure*}
We remark that the short $t_\mathrm{acc}$ in the models do {\it not} imply that the gas reservoir of the disc is severely depleted but instead that the discs have evolved to low (sub-)millimetre fluxes as a result of dust radial drift. The actual gas masses in our models are up to a factor of $1000$ larger than what would be obtained by assuming a standard dust-to-gas ratio. This effect is discussed extensively in \citet{Sellek2020} and we note that it is more marked in binary discs due to the way that tidal truncation accelerates radial drift.
In the upper-right panel of Fig.~\ref{fig:4.1}, models with different values of the initial disc scale radius, $R_0$, are plotted for $M_0=30\,M_\mathrm{Jup}$ and $\alpha=10^{-3}$, in the case of a single-star disc (solid lines) and a binary disc with $R_\mathrm{trunc}=50\,\mathrm{au}$ (dashed lines). The plot shows that the tidally truncated models are only slightly influenced by the scale radius. In fact, while in single-star models $R_0$ sets the initial viscous time scale, determining the scaling of $\dot{M}_\mathrm{acc}$ \citep[e.g.,][]{Lodato2017}, and the efficiency of radial drift \citep{Birnstiel2012}, such a time scale is only determined by $R_\mathrm{trunc}$ in close binary discs \citep{Rosotti&Clarke2018,Zagaria2021_theo}.
Finally, let us move on to the bottom, where the model dependence on viscosity, $\alpha$, is explored. The observer's equivalent dust mass is hardly affected by the viscous parameter when discs are young. However, by the age of Upper Sco, (sub-)millimetre fluxes are reduced more for higher values of $\alpha$, because the faster evolution of the gas raises the Stokes number and accelerates the radial drift of grains. This trend is observed also in single-star discs \citep{Sellek2020}, although to a lesser extent. Instead, the accretion rates depend more on viscosity and increase with $\alpha$, because it sets the velocity for gas transport thought the disc.
\subsection{Confronting models and data}
We compare binary disc models and observations in Lupus and Upper Sco. Although the Chamaeleon~I disc population and stellar multiplicity are well studied, the similar-age Lupus region is preferred because of its lower scatter in accretion time scale, better constrained (sub-)millimetre emission (85.7 \textit{vs} 79.8 per cent detections), and evenly censused stellar multiplicity (see details in Appendix~\ref{app:1}).
A systematic comparison between models and observations requires the disc truncation radius to be fixed to some value inferred from the data. As dealing with each source individually would be too complex (because of the difficulty of estimating individual truncation radii given the unknown binary eccentricity and projection effects: see Section 5.3), our models enforce a single value of the tidal truncation radius in each region. This is determined as $R_\mathrm{trunc}=0.33\times\langle a_\mathrm{p}\rangle$, where $\langle a_\mathrm{p}\rangle$ is the average binary separation in the data, and the previous relation applies to equal mass pairs in co-planar circular orbit \citep[e.g.,][]{Paczynski1977}. As a result of a systematically closer binary population in Upper Sco, the values of $R_\mathrm{trunc}$ are different in the two regions. For this reason, we do not discuss any evolutionary dependence between Lupus and Upper Sco discs in multiple stellar systems, which would be possible only if the binary separation distributions in the two regions were very similar.
Our results are displayed in Fig.~\ref{fig:4.2} where the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the observer's equivalent disc mass, $M_\mathrm{dust}\times 100$, for models with $\alpha=3\times10^{-4}$, and different values of the initial disc mass, $M_0$, and scale radius, $R_0$. Lines of constant $t_\mathrm{acc}$ are displayed as dashed-dotted lines. In the left-hand panel their solid sections refer to the $1\,\mathrm{Myr}\leq t\leq3\,\mathrm{Myr}$ age range, to be compared with Lupus data, while in the right-hand panel to the $5\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ age range to be compared with Upper Sco data. The dotted lines describe the evolution outside these time intervals.
As expected, tidal truncation causes models to pass through the region of short accretion time scale occupied by the bulk of the binary data (region A in Fig.~\ref{fig:2.1}). This is due to dust being depleted more rapidly, because truncation prevents the retention of solids at large radii, which can otherwise resupply the disc at late times (see \citealt{Sellek2020_photo} for a discussion of how external photo-evaporation similarly leads to a more rapid depletion of solids by radial drift). In fact, the highest accretors in Upper Sco can be explained using viscous models with radial dust drift as long as multiplicity effects are considered.
Nevertheless, both in Lupus and Upper Sco, model evolutionary tracks under-predict the observed disc masses. While in Lupus this issue concerns only few sources, particularly with long accretion time scale, the bulk of the Upper Sco population is roughly ten times brighter than our models. Searching the literature provides possible explanations for Lupus outliers: some of them are sub-structured (e.g., GQ~Lup and HT~Lup), while others could be affected by thermal winds (e.g., Lup~818s), as is discussed in detail in Appendix~\ref{app:2}.
In Upper Sco the under-prediction of the (sub-)millimetre fluxes by 5 to $10\,\mathrm{Myr}$ models is rather marked and applies to most of the sources in the region. Clearly, this also depends on uncertainties in opacity, although it should be noted that the assumptions we made on the grain composition correspond to an opacity at the upper end of the possible range of variation (cf. e.g., \citealt{Birnstiel2018}), and therefore with any other assumption the theoretical tracks would shift to even lower masses, exacerbating the problem. A further reason that the observed fluxes could be higher than in the models could be that (sub-)millimetre observations of closer binaries could contain emission from two discs, since more than a half of binary pairs are not resolved at the angular resolution ($\approx0.45\,\mathrm{arcsec}$) of \citet{Barenfeld2016}. This would however boost the emission by at most a factor of 2, which is not enough to account for the differences with our models. Alternatively, an evolutionary argument would suggest that only the most massive binary discs survive until the age of Upper Sco. However, models with $R_\mathrm{trunc}$ values that are too large for the observed binary separations would be required to match the data, as shown in the upper-left panel in Fig.~\ref{fig:4.1}.
To sum up, tidally-truncated models of viscously evolving discs can reproduce the short accretion time scales in the bulk of the observed binary population. Halting the fast drift of solids is required to account for the (sub-)millimetre brightness of Upper Sco and some sources in Lupus.
\section{Discussion}\label{sec:5}
\subsection{Sub-structures in binary discs}
The first detection of gaps and rings in HL~Tau \citep{ALMA_HLTau} provided striking evidence of sub-structures in planet-forming discs (see \citealt{Andrews2020} for a review). Despite such structures often being assumed to be ubiquitous \citep{Andrews2018_DSHARP,Long2018}, their presence/absence in compact or low mass discs is yet to be established
(e.g., \citealt{Long2019}, Jennings et al. subm.).
Sub-structures have been detected in binary discs as well: e.g., spirals in HT~Lup~A and AS~205~N \citep{Kurtovic2018}, inner cavities in XZ~Tau~B \citep{Osorio2016}, UZ~Tau~E and CIDA~9A \citep{Long2019,Manara2019}, gaps and rings in AS~205~S \citep{Kurtovic2018}, GQ~Lup~A \citep{Long2020} and T~Tau~N \citep{Yamaguchi2021}. However, it is not clear if they are as common as in isolated discs. In fact, since binary discs are systematically smaller than those in single stars \citep{Manara2019}, identifying sub-structures in the former is more challenging than in the latter.
Here we focus on axisymmetric annular features and discuss their influence on the correlation between mass accretion rates and disc masses. In the case of single-star discs, \citet{Sellek2020} suggested that bright rings, consistent with trapping large grains in local pressure maxima \citep{Pinilla2012,Dullemond2018}, could be an explanation for discs with long accretion time scales too massive for their smooth models (region C in Fig.~\ref{fig:2.1}). It is then tempting to hypothesise that binary discs in region C are likewise sub-structured and their longer accretion time scale compared with those in region A could be similarly explained. However, while for single-star discs this inference is supported by the presence of imaged gaps and rings, none of our binary sources in region C has clearly detected gaps or cavities, making such systems preferential targets for a first higher-resolution analysis of sub-structures in binary discs.
Noticeably, whereas discs around single stars fill the region between population B and C continuously, disc in binaries appear to be well separated: for accretion rates $\lesssim10^{-10} M_\mathrm{Sun}\,\mathrm{yr}^{-1}$ there is a lack of such discs over about an order of magnitude in disc mass (see Fig.~\ref{fig:2.1}). While it cannot be ruled out that this effect is driven by the smaller sample size of discs in binaries, it may hint at a dichotomy of evolutionary paths. Binary discs containing sub-structures may remain over several Myr in region C, comparable to the situation in structured single-star discs. Instead, in the absence of sub-structures, our models predict a more rapid decay of disc luminosity in binaries than in singles (see top-left panel of Fig.~\ref{fig:4.1}), particularly for discs with low initial disc masses \citep{Birnstiel2012}. The presence of dust traps could thus explain the apparent lack of binary discs with low accretion rates and intermediate disc masses.
Such a picture, both for binaries and singles, implies that traps maintain discs in region C over time scales of a few Myr since sources are found in region C in both Lupus and Chamaeleon I which are approximately this age. Notably, this region is almost devoid of sources in Upper Sco. This would suggest that sources do not evolve \textit{vertically} downwards from region C, since otherwise Upper Sco would contain several systems with high (sub-)millimetre fluxes but low/undetected accretion rates. Alternatively dust traps may be disrupted on a time scale of 5 to $10\,\mathrm{Myr}$ \citep{Rosotti2013,Sellek2020}, so that systems evolve out of region C \textit{horizontally} to the left. Evidence of remnant sub-structures in Upper Sco discs would support such an interpretation. Furthermore, our understanding of the late time trajectory of discs in the plane of accretion rates vs (sub-)millimetre fluxes would be improved by increasing the completeness of accretion rate determinations in Upper Sco for the discs with the highest (sub-)millimetre fluxes.
The most intriguing interpretation of dust gaps is that of being carved by planets in the act of formation \citep[e.g.,][]{Zhang2018}. If we retain the hypothesis that region C sources (whether binary or single) are sub-structured, then it is tempting to extend to binaries the argument of \citet{Manara2019_pps} that discs in region C are forming giant planets, whose deep gaps not only halt radial drift but also reduce accretion. The existence of some binaries in region C would thus support the idea that giant planet formation may be under way in at least some of these discs.
We also consider whether sub-structures can explain the fact that the observed binaries in Upper Sco are too bright compared with model predictions (see right-hand panel of Fig.~\ref{fig:4.2}, contrasting the data, shown as black dots, with the models, displayed as solid curves). If this were the case then we would need dust traps to be almost ubiquitous. However, apart from GQ~Lup~A \citep{Long2020}, a source $\approx10$ times more massive than those in Upper Sco, we have no evidence of imaged gaps and rings in any of our binary discs. This can be easily explained by the limited angular resolution of the relevant surveys\footnote{$0.25\,\mathrm{arcsec}$ \citep{Ansdell2018}, $0.60\,\mathrm{arcsec}$ \citep{Pascucci2016} and $0.45\,\mathrm{arcsec}$ \citep{Barenfeld2016}; to be compared with, e.g., the $0.12\,\mathrm{arcsec}$ resolution of \citet{Long2019}.}.
A more compelling problem is that in this picture sub-structured discs would live in region A, very far from region C where they would be expected according to our previous argument. However, looking at single-star sources in region C highlights that all transition discs with imaged cavities are clustered here, suggesting that these sub-structures play a role in maintaining conditions not only of high disc masses but also relatively low accretion rates. We can then hypothesise that, in the binary as in the single-star case, while discs in region C are characterised by prominent, large scale sub-structures affecting both gas and dust evolution, shallower, small scale dust traps would be required in the case of Upper Sco, halting dust drift but not impacting substantially the gas.
Testing this hypothesis quantitatively requires detailed modelling of secular binary evolution in the presence of sub-structures, as well as higher-resolution observations in Upper Sco.
\subsection{Coagulation efficiency in binary discs}
Shallow dust traps at small radial locations are not the only solution to the mass budget problem in Upper Sco. This can be alternatively explained in terms of a reduced grain growth efficiency in close binary discs.
Tidal interactions between gravitationally bound stars and their discs can be at the origin of asymmetries, spiral arms, eccentric modes, vertical hydraulic jumps, mass transfer between discs and relative shocks... (e.g., \citealt{Nelson2000,Muller&Kley2012,Picogna&Marzari2013}). It is commonly hypothesised that these higher-dimensional effects could halt grain growth in circumstellar discs orbiting close binaries ($a\leq30\text{ to }50\,\mathrm{au}$, \citealt{Nelson2000,Picogna&Marzari2013})\footnote{Given the average separation of Lupus and Upper Sco binaries, this argument applies preferentially to discs in the latter region.}. When spirals are triggered by tidal interactions, the disc temperature rises enough for water and other volatile species to be vaporised. As ices make up to 60 per cent of our grain composition \citep{Tazzari2016,Rosotti2019_radii}, vaporisation would lead to solid dis-aggregation and thus promote fragmentation\footnote{Adopting $u_\mathrm{f}=10\,\mathrm{m}\,\mathrm{s}^{-1}$ as in Section~\ref{sec:3} is only appropriate for ice-coated grains \citep{Gundlach&Blum2015}, while $u_\mathrm{f}=1\,\mathrm{m}\,\mathrm{s}^{-1}$ is more suited for silicate-rich grains \citep{Blum&Wurm2008}.}. Even though grain growth could take place between spirals, it would then need to begin from material in gaseous phase and would be disrupted by the frequent interactions with the spirals themselves (expected to occur on a time scale of less than $100\,\mathrm{yr}$, \citealt{Nelson2000}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_5.1_alpha.pdf}
\caption{Same as in the right panel of Fig.~\ref{fig:4.2}, assuming $f_\mathrm{grow}=10$. The green line identifies an additional model
($R_0=30\,\mathrm{au}$ and $M_0=100\,M_\mathrm{Jup}$) with a smaller truncation radius corresponding to the closest binary in Upper Sco: $R_\mathrm{trunc}=6.7\,\mathrm{au}$.}
\label{fig:5.1}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_5.2_alpha.pdf}
\caption{Accretion time scale, $t_\mathrm{acc}$, as a function of the pair projected separation, $a_\mathrm{p}$, in Lupus, Chamaeleon~I and Upper Sco. Orange dots are used for discs in binary systems and grey dots for discs evolving in isolation. The blue tracks highlight our models expectations for $M_0=30\,M_\mathrm{Jup},\,R_0=30\,\mathrm{au},\,\alpha=3\times10^{-4}$.}
\label{fig:5.2}
\end{figure*}
These effects can be taken into account in our one-dimensional models by reducing the grain growth efficiency. To do so, we introduce a coagulation efficiency factor, $f_\mathrm{grow}$ \citep{Booth&Owen2020,Sellek2020}. So far we assumed $f_\mathrm{grow}=1$, meaning that all grain collisions lead to coagulation. Instead, if $f_\mathrm{grow}>1$, only a fraction $f_\mathrm{grow}^{-1}$ of the collisions result in growth, and sticking is thus less efficient, as expected in binary discs. A lower coagulation efficiency would also increase the drift time scale, given the balance between dust growth and drift achieved in the drift-dominated regime \citep{Birnstiel2012}. Recently, \citet{Booth&Owen2020} suggested that assuming $f_\mathrm{grow}=10$ could be beneficial for explaining the depletion of refractory elements in the Sun relative to nearby Solar twins. \citet{Sellek2020} ran single-star disc models with $f_\mathrm{grow}=10$, showing that they were able to reproduce massive Lupus discs that they could not explain assuming more efficient growth. However $f_\mathrm{grow}=10$ single star models could not match the data in Upper Sco, being too massive at any given accretion rate.
We retain $f_\mathrm{grow}=10$ as in these works because it provides a balance between fast growth, where solids drift fast, and slow growth, where grains are fragmentation-dominated throughout and the dust follows the gas \citep{Booth&Owen2020}. The results of our binary disc models with $f_\mathrm{grow}=10$, in the case of $R_\mathrm{trunc}=18.13\,\mathrm{au}$ and $\alpha=3\times10^{-3}$, are shown in Fig.~\ref{fig:5.1}. Regardless of the initial disc parameters, our models evolve towards disc masses that are ten times higher than for $f_\mathrm{grow}=1$ (as seen by \citealt{Sellek2020}). This difference reconciles binary models and observations in Upper Sco, with the caveat that the highest accretors are still not reproduced. This is easily explained by our choice of an average binary separation. Enforcing $R_\mathrm{trunc}=6.7\,\mathrm{au}$, corresponding to the minimum binary separation in Upper Sco, allows models to reproduce high accretors as well. This is shown in the representative case of $R_0=30\,\mathrm{au}$ and $M_0=100\,M_\mathrm{Jup}$, depicted with a solid green line in Fig.~\ref{fig:5.1}.
\subsection{Accretion time scale \textit{vs} binary projected separation}
Using one-dimensional viscous models, we have shown that the fact that proto-planetary discs in multiple systems occupy region A in Fig.~\ref{fig:2.1} (where $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$) can be reproduced by the secular effects of outer disc truncation on dust and gas evolution. Such models suggest that a shorter accretion time scale in discs in closer binaries should be expected, owing to the larger accretion rates and lower dust masses predicted in those systems (see e.g., top-left panel of Fig.~\ref{fig:4.1}).
Let us therefore examine how the observationally-inferred accretion time scale, $t_\mathrm{acc}$, depends on the measured projected binary separation, $a_\mathrm{p}$. This is displayed in Fig.~\ref{fig:5.2} for Lupus, Chamaeleon~I and Upper Sco binary discs, using orange dots. Single-star discs are also shown for comparison, using grey dots. No particular trend of $t_\mathrm{acc}$ with $a_\mathrm{p}$ can be recognised in the data, at odds with our expectations. These are highlighted by the blue tracks and shaded area for the representative case with $M_0=30\,M_\mathrm{Jup},\,R_0=30\,\mathrm{au},\,\alpha=3\times10^{-4}$ and its $1\sigma$ spread.
However, what primarily sets the accretion time scale in our models is the location where truncation occurs. In fact, this is not determined by the binary projected separation only, but is expected to be affected by other effects, such as the relative inclination of the interacting pair with respect to the plane of the sky \citep{Manara2019}, the binary mass ratio and eccentricity \citep[e.g.,][]{Paczynski1977,Artymowicz&Lubow1994,Pichardo2005}, and the relative misalignment between the disc and the binary plane \citep{Lubow2015}.
In particular, comparing the binary separation distribution in our data with the observed relation between main sequence binaries eccentricity and orbital period \citep{Raghavan2010}, would imply that some discs in our sample could have a significant eccentricity, thus washing out the expected dependence of $t_\mathrm{acc}$ on $a_\mathrm{p}$. Recent evidence that eccentricity in wider binaries follows a thermal rather than uniform distribution (see \citealt{Hwang2021}, Fig.~6) would suggest that the most eccentric pairs would be those with larger $a_\mathrm{p}$. As long as the bulk of the population has relatively smaller eccentricities ($e\approx0.3$ as expected in the field, e.g., \citealt{Duchene&Kraus2013}), this is not in tension with the fact that the the dust disc sizes in Taurus and $\rho$~Ophiuchus can be well explained assuming low eccentric orbits (\citealt{Zagaria2021_obs} and \citealt{Rota2022}).
Even though the comparison between models and specific sources can be challenged by uncertainties on binary eccentricities, we do not expect them to substantially modify our results in Section~\ref{sec:4}, because they would reduce the average truncation radius only by a factor of a few and this is potentially compensated by the underestimation of the binary separation by its projection on the plane of the sky. Nevertheless, recovering the theoretically-expected trend of $t_\mathrm{acc}$ with $a_\mathrm{p}$ is only part of the problem, given the large scatter in Fig.~\ref{fig:5.2}. In fact, on top of the previously mentioned mechanisms, other effects related to grain coagulation efficiency, disc evolution and dispersal can influence the accretion time scale. We do not explore further these possibilities because, given our uncertainty on the determination of $R_\mathrm{trunc}$, modelling would be under constrained by the data. Finally, some of the sources in Fig.~\ref{fig:5.2} can also be components of undetected higher-order multiples.
\section{Summary and conclusions}\label{sec:6}
In this paper we made use of gas and dust one-dimensional simulations to address how stellar multiplicity can influence the correlation between mass accretion rates onto forming stars and the masses of their proto-planetary discs.
Hereafter our main conclusions are summarised:
\begin{itemize}
\item We searched the literature for information on the stellar multiplicity of Lupus, Chamaeleon~I and Upper Sco discs with both (sub-)millimetre emission and accretion rates in \citet{Manara2016,Manara2020} and \citet{Mulders2017}. We found that roughly 20 per cent of the targeted discs are in binary or higher-order multiple stellar systems;
\item We introduced the accretion time scale, $t_{\rm acc}$, as the ratio of observationally estimated disc mass to accretion rate (see Eq.~\ref{eq:1}) and found that discs in gravitationally bound pairs have a systematically shorter accretion time scale than discs evolving in isolation, clustering around $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$. The short accretion time scales is a consequence both of the well known fact that dust masses are lower in binaries compared with single stars \citep[e.g.,][]{Harris2012} {\it and} of the fact that accretion rates are somewhat higher than in single stars (see Fig.~\ref{fig:2.1});
\item To understand the trend in the data, we ran models of tidally truncated proto-planetary discs subject to viscous evolution, grain growth and radial drift. We found that our models show shorter accretion time scales for lower values of $R_\mathrm{trunc}$, as a consequence of the faster radial drift and reduced disc lifetime determined by tidal truncation (see upper-left panel of Fig.~\ref{fig:4.1});
\item Models and data agree reasonably well in Lupus (see left-hand panel of Fig.~\ref{fig:4.2}); we suggest that a small number of anomalously (sub-)millimetre bright discs may be explained by the existence of sub-structures, while a disc with anomalously long $t_\mathrm{acc}$, may potentially undergo internal photo-evaporation;
\item In Upper Sco our models systematically under-predict disc masses by up to a factor of ten (see right-hand panel of Fig.~\ref{fig:4.2}). It is unclear if this could be the effect of undetected traps halting the drift of solids. We also find that the agreement is improved if the grain growth efficiency is reduced (see Fig.~\ref{fig:5.1}) as has been hypothesised in the dynamically active environments of discs in binaries:
\item Finally, we find no clear evidence that the observationally-inferred accretion time scale increases with the binary projected separation (see Fig.~\ref{fig:5.2}), at odds with our theoretical expectation that the accretion time scale should increase with the tidal truncation radius of the disc. However, the relationship between projected separation and tidal truncation radius is complicated by projection effects and the orbital eccentricity: given the relatively limited dynamic range of projected separations in the sample, this lack of correlation is therefore not necessarily surprising.
\end{itemize}
\section*{Acknowledgements}
We thank the anonymous reviewer for their comments that helped to improve the manuscript. FZ acknowledges support from STFC and Cambridge Trust for a Ph.D. studentship. GR acknowledges support from the Netherlands Organisation for Scientific Research (NWO, program number
016.Veni.192.233) and from an STFC Ernest Rutherford Fellowship
(grant number ST/T003855/1). This work was partly supported by the Deutsche Forschungs-Gemeinschaft (DFG, German Research Foundation) - Ref no. FOR 2634/1 TE 1024/1-1. This work has also been supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska Curie grant agreement number 823823 (DUSTBUSTERS).
Software: \texttt{numpy} \citep{numpy20_2020Natur.585..357H}, \texttt{matplotlib} \citep{matplotlib_Hunter:2007}, \texttt{scipy} \citep{scipy_2020SciPy-NMeth}, \texttt{JupyterNotebook} \citep{Jupyter_nootbok}. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France \citep{Wenger2000}, and the Binary Star Database by Kendra Kellogg, Lowell Observatory. Colours in Fig.s~\ref{fig:4.2} and \ref{fig:5.1} are from \texttt{ColorBrewer.org} by Cynthia A. Brewer, Geography, Pennsylvania State University.
\section*{Data Availability}
The code used in this paper is publicly available on GitHub at
\texttt{github.com/rbooth200/DiscEvolution}. The data underlying
this paper are available in the ALMA archive as explained in the papers quoted in Section~\ref{sec:2} and Appendix~\ref{app:1}.
\bibliographystyle{mnras}
\section{Introduction}\label{sec:1}
Understanding how proto-planetary discs evolve and eventually dissipate is fundamental to building a comprehensive picture of the planet-formation process \citep[e.g.,][]{Manara2019_pps}. It is commonly assumed that proto-planetary discs evolve under the effect of viscosity, which, allowing for angular momentum re-distribution within the disc, drives accretion onto the forming star \citep{Shakura&Sunyaev1973,Lynden-Bell&Pringle1974}. In this framework, the mass accretion rate on the central star, $\dot{M}_\mathrm{acc}$, and the disc mass, $M_\mathrm{disc}$, are expected to be correlated at any given age \citep[e.g.,][]{Hartmann1998,Jones2012}:
\begin{equation}\label{eq:1}
M_\mathrm{disc}=t_\mathrm{acc}\dot{M}_\mathrm{acc}\propto(t+t_\nu)\dot{M}_\mathrm{acc},
\end{equation}
where $t_\nu$ is the initial viscous time scale and we call $t_\mathrm{acc}$ the accretion time scale\footnote{This is also known as the disc lifetime \citep[e.g.,][]{Lodato2017}.}.
The advent of new facilities, such as the Atacama Large Millimeter/submillimeter Array (ALMA) and the X-shooter spectrograph at the Very Large Telescope (VLT), has made it possible to test disc evolutionary models observationally. In the last years, disc masses, estimated from (sub-)millimetre dust fluxes under the assumption of optically thin emission and a constant gas-to-dust ratio \citep[e.g.,][]{Ansdell2016,Ansdell2018,Pascucci2016,Barenfeld2016}, as well as mass accretion rates \citep[e.g.,][]{Alcala2014,Alcala2017,Manara2016_cham,Manara2017,Manara2020}, have been inferred for a large number of young stellar objects.
Combining these data-sets, \citet{Manara2016} and \citet{Mulders2017} detected a slightly sub-linear correlation between mass accretion rates and proto-planetary disc masses in the young Lupus and Chamaeleon~I star-forming regions (aged $\approx1\text{ to }3\,\mathrm{Myr}$), that was interpreted as being compatible with viscous-diffusion models \citep{Rosotti2017,Lodato2017,Mulders2017}. A correlation was also identified by \citet{Manara2020} in the older Upper Scorpius OB association (aged $\approx5\text{ to }10\,\mathrm{Myr}$). However, at any given disc mass, the median accretion rate in Upper Sco is remarkably similar to that in young regions; moreover its scatter has not decreased with time. This is in contrast with the viscous evolution scenario, which predicts both lower $\dot{M}_\mathrm{acc}$ and a tighter relation at later times (\citealt{Rosotti2017,Lodato2017} and Eq.~\ref{eq:1}).
Nevertheless, all the previous analyses of the correlation only focused on \textit{gas} evolution. Instead, the disc mass estimates of \citet{Manara2016,Manara2020} and \citet{Mulders2017} are based on (sub-)millimetre \textit{dust} emission and are subject to systematic uncertainties (e.g., on dust opacity and the gas-to-dust ratio).
To try to explain the observations, both gas and dust ought to be considered when modelling disc evolution. Moreover, theoretical predictions and observational trends must be compared in the \textit{data space}, i.e. post-processing the model outputs to determine their (sub-)millimetre dust fluxes and masses as for real data. Such an exercise then removes the use of the restrictive assumptions about optical depth, opacities and the gas-to-dust ratio that are generally used to interpret (sub-)millimetre observations. \citet{Sellek2020} adopted such a forward modelling approach and demonstrated that viscous simulations provide a remarkable agreement with both Lupus and Upper Sco data (in particular when internal photo-evaporation is included to account for late-time disc dispersal). However, even though the models of \citet{Sellek2020} accounted for the underestimation of disc masses in current surveys, none of them were able to reproduce discs with accretion time scale shorter than $t_\mathrm{acc}\lesssim0.1\,\mathrm{Myr}$, which contribute to most of the large scatter/high accretion rates in Upper Sco.
As mass accretion rates are computed at a fixed snapshot in time, it might be thought that accretion variability is responsible for the outlying sources. Unfortunately, there are no studies addressing this issue at the age of Upper Sco \citep{Manara2020}, and very little is known about accretion variability on time scales longer than few days. In younger ($t\lesssim3\,{\rm to}\,4\,{\rm Myr}$) discs, $\dot{M}_{\rm acc}$ variations are $\lesssim0.4\,{\rm dex}$ on a time scale of days \citep{Costigan2014,Venuti2014,Manara2021}, not enough to account for the highest accretors in Upper Sco, with only a small fraction of the targets having larger variability \citep{Audard2014}.
All the previous considerations are based on the assumption that planet-forming discs evolve in isolation. This picture is clearly idealised: the majority of stars are born in clusters, where nearby sources can influence proto-planetary disc evolution history and affect their planet-formation potential \citep[e.g.,][]{Winter2020}. This commonly occurs by either external photo-evaporation, when the UV radiation of a massive nearby star dissipates the less bound gas in the outer disc regions \citep[e.g.,][]{Adams2004,Facchini2016}, or by tidal interactions, when a stellar companion in a gravitationally bound pair, or a flyby in a dense environment, truncates the disc \citep[e.g.,][]{Winter2018,Cuello2019,Cuello2020}. These processes have qualitatively similar effects, both reducing disc masses and sizes (e.g., \citealt{Harris2012,Akeson&Jensen2014,Cox2017,Akeson2019,Manara2019} in binaries and \citealt{Ansdell2017,Otter2021} for the effects of photo-evaporation), eventually hastening disc dispersal.
How external photo-evaporation and tidal truncation impact the correlation between mass accretion rates and disc masses was studied by \citet{Rosotti2017} and \citet{Rosotti&Clarke2018}, respectively. These studies showed that both discs exposed to strong UV fields and discs in multiple stellar systems are expected to have shorter accretion time scales than those (viscously) evolving in isolation. This means that the influence of the environment could be a potential explanation for those highly-accreting old discs with low masses that the models of \citet{Sellek2020} were not able to explain. However, these studies only focused on gas evolution. Since dust was not considered as a separate component, it was simply assumed that observationally inferred disc dust masses could be converted into total disc masses using a gas-to-dust ratio of 100. This motivates our present study based on modelling both gas and dust evolution and then comparing simulation-based synthetic observations with real data.
The required properties of star-forming environments for which external photo-evaporation and tidal encounters are effective were studied by \citet{Winter2018_trunc}. They showed that stellar densities $\gtrsim10^4\,{\rm pc}^{-3}$ are necessary for disc truncation by tidal encounters. This threshold is much higher than in the nearby star-forming regions, making fly-bys unlikely. As for external photo-evaporation, \citet{Winter2018_trunc} concluded that average UV fields $\gtrsim3\times10^3\,G_0$ are required to disperse a disc. (Here $G_0$ corresponds to an energy flux of $1.6\times10^{-3}\,{\rm erg}\,{\rm cm}^{-2}$ between $6\,{\rm and}\,13.6\,{\rm eV}$, \citealt{Habing1968}.) In Lupus the average radiation field is expected to be remarkably low (e.g., in IM~Lup it is $\lesssim4\,G_0$, \citealt{Cleeves2016}) while in Upper Sco it is somewhat higher (on average $42.9\,G_0$, \citealt{Trapman2020}). Since the median UV field in both regions is below the critical value for evaporation, we decided not to consider external photo-evaporation in this paper\footnote{In Upper Sco the radiation levels can grow up to $7\times10^3\,G_0$ \citep{Trapman2020}, suggesting that external photo-evaporation can influence the accretion time scales of some sources. We plan to study this effect in a future paper.}.
Instead, in this work we focus on the role played by stellar multiplicity in disc evolution, confronting our theoretical expectations with the mass accretion rates and sub-millimetre fluxes from nearby young star-forming regions. First of all, from the currently published catalogues, we collect a state-of-the-art sample of homogeneously determined disc masses and accretion rates in Lupus, Chamaeleon~I and Upper Sco. This sample is further updated making use of Gaia EDR3 distances \citep{GAIAeDR3} as in the PPVII chapter of Manara et al. (subm.). Furthermore, searching the literature allows us to identify several pairs, i.e. gravitationally bound stars with projected separation, $a_\mathrm{p}$, less than $300\,\mathrm{au}$, in binaries or higher order multiple stellar systems among the previously selected sources. Subsequently, we run models of tidally truncated circumstellar binary discs for a large number of initial disc parameters following the approach of \citet{Zagaria2021_theo}. To test our hypothesis that the accretion time scale is influenced by a companion, these are then compared with observations through modelling the predicted dust emission.
This paper is organised as follows. In Section~\ref{sec:2} we introduce the Lupus, Chamaeleon~I and Upper Sco samples. Discs in multiple stellar systems are identified (see Appendix~\ref{app:1}) and their properties discussed in comparison to those in single-star systems. In Section~\ref{sec:3} dust and gas modelling is introduced, while Section~\ref{sec:4} describes our results, first of all discussing their dependence on parameter choices and then comparing models and observations. In Section~\ref{sec:5} we consider the model limitations and possible improvements to the agreement with observational data. Finally, in Section~\ref{sec:6} we draw our conclusions.
\section{Sample selection and data analysis}\label{sec:2}
\subsection{Sample selection}
We focus on Lupus, Chamaeleon~I and Upper Sco because they are among the best studied nearby star-forming regions: (sub-)millimetre dust fluxes and accretion rates were measured for a large fraction of their young stellar objects (YSOs), and their multiplicity fraction is often well known.
The distances to individual targets are taken from inverting the parallaxes available in the Gaia EDR3 catalog \citep{GAIAeDR3}. The only exception are the cases where the Gaia parallaxes are unreliable, i.e. ${\rm RUWE}>1.8$ and/or distance differing more than $60\,{\rm pc}$ from the median distance to the region, or unavailable. In these cases, we assumed the median distance to the members of the region: $d=158\,{\rm pc}$ for Lupus, $d=190\,{\rm pc}$ for Chamaeleon~I, and $d=145\,{\rm pc}$ for Upper Sco.
In Lupus, we make use of the (sub-)millimetre dust emission measured by \citet{Ansdell2016}, with the addition of the Lupus completion survey sources of \citet{Sanchis2020}, Sz~82/IM~Lup \citep{Cleeves2016} and Sz~91 \citep{Tsukagoshi2019}.
We convert the ALMA (sub-)millimetre dust fluxes into dust masses assuming optically thin emission \citep{Hildebrand1983}:
\begin{equation}\label{eq:2}
M_\mathrm{dust}=\dfrac{d^2F_\nu}{B_\nu(T_\mathrm{dust})\kappa_\nu^\mathrm{abs}},\text{ where }\kappa_\nu^\mathrm{abs}=2.3\left(\dfrac{\nu}{230~\mathrm{GHz}}\right)\text{cm}^2\text{g}^{-1}.
\end{equation}
Here $F_\nu$ is the dust flux, $d$ the distance of the source, $B_\nu$ the black body emission at temperature $T_\mathrm{dust}=20\,\mathrm{K}$, and $\kappa_\nu^\mathrm{abs}$ the absorption opacity \citep{Beckwith1990}. Disc masses are then computed from dust masses assuming a standard ISM gas-to-dust ratio of 100 \citep{Bohlin1978}. A target is a non-detection if its dust flux does not exceed three times the continuum rms noise.
Mass accretion rates are inferred from the VLT/X-shooter spectroscopic measurements of \citet{Alcala2014,Alcala2017}. Briefly, stellar properties and the accretion parameters are derived self-consistently fitting the YSO spectra with the sum of the photospheric template of a non-accreting star and a slab model to reproduce the accretion luminosity excess in the Balmer continuum and also in the optical part of the spectrum, as described in \citet{Manara2013}. Following Manara et al. (subm.), the effective temperature, $T_{\rm eff}$, of the stars were re-calculated using the relation between spectral type and $T_{\rm eff}$ by \citet{Herczeg&Hillenbrand2014}, and the stellar luminosity and accretion luminosity were re-scaled to the distances from Gaia EDR3. To derive the stellar masses, needed to calculate the mass accretion rates, the non-magnetic evolutionary tracks by \citet{Feiden2016} and \citet{Baraffe2015} where used for targets hotter and colder than 3900 K, respectively, while the models by \citet{Siess2000} were used for sources appearing too young/old for the former tracks. Under-luminous objects (see \citealt{Alcala2014}), i.e., falling well below the main locus of pre-main sequence targets in the HR diagram, are excluded. If the excess continuum UV luminosity of a YSO with respect to the best-fit photospheric template is close to its chromospheric noise level, it is termed a non-accretor and its mass accretion rate is considered an upper-limit \citep[see][]{Alcala2017,Manara2017}. The accretion rate of MY~Lup is as computed from the HST observation results of \citet{Alcala2019}. After updating stellar distances, an average uncertainty of $0.35\,{\rm dex}$ is estimated.
In Chamaeleon~I and Upper Sco we follow a similar procedure. In the former region, we use the (sub-)millimetre dust emission of \citet{Pascucci2016} and \citet{Long2018_cham}, and the accretion luminosities of \citet{Manara2016_cham,Manara2017}. In the latter region, we rely on \citet{Manara2020} sample for mass accretion rates, while disc masses are computed using the (sub-)millimetre dust emission of \citet{Barenfeld2016}.
Transition discs (TDs) are identified as both discs with a ``resolved'' (imaged) cavity and ``unresolved'' discs, whose classification is based on their SED shape as in \citet{Manara2016,Manara2020} and \citet{Mulders2017}. Finally, we remark that Lupus and Chamaeleon~I samples are more than 90 per cent complete\footnote{In this paper, by saying that a given star-forming region is $x$ per cent complete, we mean that disc accretion rates were measured for a fraction $x/100$ of the YSOs with discs in that region.} \citep{Alcala2017,Manara2017}, while Upper Sco has high levels of completeness only for $M_\mathrm{disc}\gtrsim5\times10^{-5}\,M_\mathrm{Sun}$ \citep{Manara2020}\footnote{More specifically, \citet{Manara2020} claimed completeness to be 80 per cent in the mass range $4.8\times10^{-5}\,M_\mathrm{Sun}\leq M_\mathrm{disc}\leq6.47\times10^{-4}\,M_\mathrm{Sun}$ and 60 per cent in the mass range $6.47\times10^{-4}\,M_\mathrm{Sun}\leq M_\mathrm{disc}\leq1.55\times10^{-2}\,M_\mathrm{Sun}$ \citep{Manara2020}. Nevertheless, these are likely overestimates if the refined membership census of YSOs with discs in Upper Sco \citep{Luhman&Esplin2020} is considered.}.
We searched the literature for information on stellar multiplicity for the previously selected YSOs. Our main references are \citet{Zurlo2021} in Lupus, \citet{Lafreniere2008} in Chamaeleon~I and \citet{Barenfeld2019} in Upper Sco. A total of 9 binaries or higher-order multiple stellar systems with projected separation less than $300\,\mathrm{au}$ are found is Lupus, 16 in Chamaeleon and 9 in Upper Sco. 11 of these are in pairs closer than $40\,\mathrm{au}$ \citep{Kraus2012}. We refer the reader to Appendix~\ref{app:1} for a detailed discussion of the multiplicity detection methods and comparison among regions.
\subsection{Data analysis}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_2.1_alpha.pdf}
\caption{\textbf{Left panel:} Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the disc mass, $M_\mathrm{disc}=M_\mathrm{dust}\times100$, for Lupus, Chamaeleon~I and Upper Sco Class~II sources. Grey dots are used for isolated stars, orange and blue for binary ($a_\mathrm{p}<300\,\mathrm{au}$) and close binary ($a_\mathrm{p}<40\,\mathrm{au}$) discs, and green for circumbinary discs. Transition discs are identified by a wider circle, non-accretors by a downward-pointing triangle and non-detections by a left-pointing arrow. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1$ and $10\,\mathrm{Myr}$ are shown as dot-dashed lines. Letters A, B and C are used to label different binary disc populations (see main text). \textbf{Right panel:} Accretion time scale, $t_\mathrm{acc}$, cumulative distribution for single-star and binary discs in grey and orange, respectively. Solid lines are used for the full sample, dashed lines when non-accretors are excluded. The caption lists $p$-values for the null hypothesis of binaries and singles being drawn from the same $t_{\rm acc}$ distribution.}
\label{fig:2.1}
\end{figure*}
\paragraph*{General population trends} In the left-hand panel of Fig.~\ref{fig:2.1} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the disc mass inferred from (sub-)millimetre dust emission, $M_\mathrm{disc}=M_\mathrm{dust}\times100$, for all Lupus, Chamaeleon and Upper Sco discs in our sample. Grey dots are used for isolated stars, orange and blue for binary ($a_\mathrm{p}<300\,\mathrm{au}$) and close binary ($a_\mathrm{p}<40\,\mathrm{au}$) discs, while green for circumbinary discs. Non-accretors are displayed as downward-pointing triangles, while disc mass non-detections are identified by left-pointing arrows. Adopting the same convention of \citet{Manara2016}, transition discs can be recognised by a larger external circle of the same colour. In the background, lines of constant accretion time scale for $t_\mathrm{acc}=0.1,\,1\,\mathrm{and}\,10\,\mathrm{Myr}$ are displayed as grey dashed-dotted lines.
Single-star discs and discs in binary systems populate different regions of the data-space. The former span all the range of observed disc masses and accretion rates, while the latter are characterised by a systematically shorter accretion time scale and can be divided into three sub-regions. The bulk of discs in binaries (A) have $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$. The remainder have longer accretion time scales, with $t_\mathrm{acc}\gtrsim1\,\mathrm{Myr}$. Some of these are very faint (B), often undetected in the (sub-)millimetre, with $M_\mathrm{disc}\lesssim2\times10^{-4}\,M_\mathrm{Sun}$, while a small number are massive (C), with $M_\mathrm{disc}\gtrsim2\times10^{-3}\,M_\mathrm{Sun}$. Remarkably, very close binaries ($a_\mathrm{p}<40\,\mathrm{au}$) mainly live in region A, and $5/11\approx45$ per cent have accretion time scales much shorter than $0.1\,\mathrm{Myr}$, because of high accretion rates and (generally) low
disc masses.
To prove our inference of a shorter accretion time scale in discs in multiple systems we perform a Kolmogorov–Smirnov (K-S) test on the full sample made up of all data from the three star-forming regions together, excluding the three circumbinary discs in the sample. The null-hypothesis of single-star and binary discs being drawn from the same $t_\mathrm{acc}$ distribution is rejected with $p-$values of $7.00\times10^{-7}$ and $3.54\times10^{-6}$, when non-accretors are considered or excluded, respectively. This is visually displayed in the right-hand panel of Fig.~\ref{fig:2.1}, where the accretion time scale cumulative distributions are plotted for single-star and binary discs in grey and orange, respectively. Solid lines are used for the full sample, while dashed ones when non-accretors are not considered. Clearly, discs in multiple systems have systematically shorter $t_\mathrm{acc}$ than those evolving in isolation.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_2.2_alpha.pdf}
\caption{Same as in the left-hand panel of Fig.~\ref{fig:2.1}, with each star-forming region in a different sub-figure. \textbf{From left to right:} Lupus, Chamaeleon~I and Upper Sco. The 80 per cent completeness mass range \citep{Manara2020} is highlighted in grey in the latter region. Captions as in the right-panel of Fig.~\ref{fig:2.1}.}
\label{fig:2.2}
\end{figure*}
Combining the targets from different regions allows us to collect a large enough sample to make statistically significant comparisons between single-star and binary discs. This is a fair procedure as long as it does not introduce any multiplicity biases in the target selection. Although it can be argued that only Upper Sco discs with $5\times10^{-5}\lesssim M_\mathrm{disc}/M_\mathrm{Sun}\lesssim7\times10^{-4}$ have been surveyed for accretion rates with high completeness, such incompleteness does not discriminate between singles and binaries. Indeed, \citet{Barenfeld2019} showed that, in this region, discs in binaries and around singles span a similar (sub-)millimetre brightness range
Finally, we briefly comment on the treatment of (sub-)millimetre emission upper limits in our sample, highlighting that the significant difference between the accretion time scale in binary and single-star discs is not the result of uncertainties associated with how dust mass upper limits are assigned to non-detections. We test this claim in the worst case scenario, i.e. where we assign to non-detections values which make the data-sets for singles and binaries as similar as possible. Performing a K-S test on the new cumulative curve gives a $p-$values of $2.22\times10^{-4}$
\paragraph*{Analysis by star-forming region} While the combined data from Lupus, Chamaeleon~I and Upper Sco clearly demonstrates the difference between singles and binaries, it washes out possible evolutionary effects resulting from the different ages of the regions. For this reason, in Fig.~\ref{fig:2.2} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the disc mass, $M_\mathrm{disc}=M_\mathrm{dust}\times100$ for Lupus, Chamaeleon~I and Upper Sco individually.
Fig.~\ref{fig:2.2} shows that, while binary discs in the younger regions belong to all the previously identified families (A to C), only family A is represented in Upper Sco. To understand whether this is a disc evolution or sample-selection effect, further data are needed. We perform a K-S test on Lupus, Chamaeleon~I and Upper Sco separately, finding that the null-hypothesis that single-star and binary discs are drawn from the same $t_\mathrm{acc}$ distribution is rejected only in the latter (the K-S test $p-$values are reported in the bottom right of each panel in Fig.~\ref{fig:2.2}). Again, assessing whether this is an evolutionary or sample-selection effect needs a higher level of completeness in Upper Sco and a homogeneous analysis of the stellar multiplicity among the three star-forming regions (see Appendix~\ref{app:1}). \\
\noindent To sum up, our data show that binary discs have statistically shorter accretion time scales than those around single stars. Further data and information on disc multiplicity are needed to compare different star-forming regions from the evolutionary point of view.
\section{Numerical methods}\label{sec:3}
In the remainder of this paper our focus is on exploring if the observed trend of a shorter accretion time scale in binaries is compatible with the expectations of models for gas and dust evolution in viscously evolving discs subject to tidal truncation by a stellar companion. To do so, we run a number of binary disc simulations that are post-processed to compute their (sub-)millimetre dust emission profiles. Throughout our analysis we use these synthetic observations to evaluate model disc properties (e.g., masses and accretion time scales) employing the same standard assumptions used in the literature (cf \citealt{Sellek2020} in the case of single-star discs). Those properties are then compared with real data.
We model gas and dust evolution in circumstellar binary discs, neglecting re-supply of material from beyond the binary orbit and assuming that the disc cannot spread outside the tidal truncation limit imposed by the companion. For the gas we consider viscous evolution and omit the effects of photo-evaporation. For the dust the two-population model of \citet{Birnstiel2012} is employed. As a post-processing step, dust emission and accretion rates from our models are used to generate synthetic observations to be compared with real data.
The one-dimensional finite-differences code developed by \citet{Booth2017} is used, modified to take into account tidal truncation as \citet{Zagaria2021_theo} did. Briefly, the viscous-diffusion equation \citep{Lynden-Bell&Pringle1974} is solved on a grid made up of 250 cells equally spaced in $R^{1/2}$, assuming zero-flux \citep{Bath&Pringle1981,Rosotti&Clarke2018} at the outer boundary, $R=R_\mathrm{trunc}$. We choose an exponentially-tapered power-law initial condition:
\begin{equation}\label{eq:3}
\Sigma(R,t=0)=\dfrac{M_0}{2\pi R_0R}\exp\left(-\dfrac{R}{R_0}\right),
\end{equation}
where $R_0$ is a characteristic scale radius, and $M_0$ is the initial disc mass within $R_\mathrm{trunc}$. We explore the region of the parameter space corresponding to $R_0=10$ and $100\,\mathrm{au}$, and $M_0=1$, 3, 10, 30 and $100\,M_\mathrm{Jup}$ \citep{Sellek2020}.
A passively irradiated time-independent disc temperature profile \citep{Chiang&Goldreich1997} is considered, radially decaying as $R^{-1/2}$ and with reference temperature $T_0=88.23$~K at 10~au, calibrated on a Solar-mass star. The \citet{Shakura&Sunyaev1973} prescription is used for viscosity: we set $10^{-4}\leq\alpha\leq10^{-3}$, both for consistency with previous works \citep{Sellek2020} and because at these viscosities similar models reproduce the flux-radius correlation most closely \citep{Rosotti2019_fr,Zagaria2021_obs,Zormpas2022}.
For the dust, a uniform initial dust fraction of 0.01 is employed. The routine for grain growth implements the simplified treatment of \citet{Birnstiel2012}. In short, at each disc radius two dust populations are evolved in time: a population of small grains, whose size is the monomer grain size, $a_\mathrm{min}=0.1\,\mu\mathrm{m}$, and a population of large grains, dominating the mass, whose size, $a_\text{max}$, is determined by the combined effect of grain growth (with a grain-growth efficiency $f_\mathrm{grow}=1$, \citealt{Booth&Owen2020,Sellek2020}), fragmentation (with fragmentation velocity $u_\mathrm{f}=10\,\mathrm{m}\,\mathrm{s}^{-1}$, \citealt{Gundlach&Blum2015}) and radial drift. We refer to \citet{Birnstiel2012} for more information and to \citet{Booth2017} for the specific implementation. Finally, the dust fraction is advected along the gas flow, following the prescription of \citet{Laibe&Price2014} and taking into account the dust back-reaction on gas \citep[e.g.,][]{Dipierro2018,Garate2020}.
We compare models and observations in the data space. This means that our simulation results are post-processed to determine how they would look if observed with ALMA at the same wavelength as in \citet{Ansdell2016} and \citet{Barenfeld2016}. Their mass is then computed as in these surveys and compared with real data. We call this quantity the \textit{observer's equivalent} disc mass \citep{Sellek2020}. Doing so, we can be agnostic on the assumptions made to compute disc masses in the observations and our results would not be influenced by the latter. Our procedure is as follows. For each of our models, we first determine a synthetic surface brightness, $S_\mathrm{b}$, at frequency $\nu\approx338.75\,{\rm GHz}$ (ALMA band 7):
\begin{equation}\label{eq:4}
S_\text{b}(R)=B_\nu(T)\bigl\{1-\exp{(-\kappa_\nu\Sigma_\text{d})}\bigr\},
\end{equation}
where $\Sigma_\text{d}$ is the dust surface density, $B_\nu$ is the black body radiation spectrum at temperature $T$ and $\kappa_\nu$ the dust (absorption) opacity, computed as in \citet{Tazzari2016} and \citet{Rosotti2019_radii}. We assumed face-on discs. If one considers the distribution of disc inclinations on the sky, it can be shown that this approximation is correct within a factor of $\langle\cos i\rangle=\pi/4\approx0.8$. Moreover, \citet{Tazzari2017} showed that correcting for the disc inclination improves the disc masses in \citet{Ansdell2016} by less than a factor of 2. Then, the \textit{observer's equivalent} dust mass, $M_\mathrm{dust}$, is computed as in Eq.~\ref{eq:2}, where the disc luminosity, $L_\nu=d^2F_\nu$, is given by the surface integral of the synthetic surface brightness.
\section{Results}\label{sec:4}
In this Section we outline our results, discussing how the models are affected by the initial parameters. Subsequently, models and observations are confronted, in order to assess their compatibility.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_4.1_alpha.pdf}
\caption{Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the observer's equivalent disc mass (see Eq.~\ref{eq:2}), $M_\mathrm{dust}\times100$. The solid tracks identify the $1\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval. The dependence on the truncation radius, $R_\mathrm{trunc}$ (top left), the scale radius, $R_0$ (top right), and viscosity, $\alpha$ (bottom left), are explored. In the top-right panel, solid lines refer to singles, dashed to binaries. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10.0\,\mathrm{Myr}$. Lupus and Upper Sco data are shown in grey for singles and black for binaries. Larger circles identify transition discs and downward-pointing triangles are used for non-accretors.}
\label{fig:4.1}
\end{figure*}
\subsection{Model dependence on the initial disc parameters}
As a starting point, we illustrate how our results depend on the initial disc parameters. In Fig.~\ref{fig:4.1} the mass accretion rate, $\dot{M}_\mathrm{acc}$, is shown as a function of the observer's equivalent disc mass, $M_\mathrm{dust}\times 100$. The colour dotted lines identify model evolutionary tracks and their solid sections highlight the $1\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ age interval. In the background, lines of constant accretion time scale for $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10\,\mathrm{Myr}$ are displayed as grey dashed-dotted lines. Lupus and Upper Sco discs are also plotted in grey for singles and black for binaries, respectively.
Let us begin with the upper-left panel of Fig.~\ref{fig:4.1}, where models with different values of the tidal truncation radius, $R_\mathrm{trunc}$, are plotted for $M_0=30\,M_\mathrm{Jup}$\footnote{We adjust the surface density normalisation so as to achieve a constant total disc mass, $M_0$, within $R_\mathrm{trunc}$. Our models are gravitationally stable, with \citet{Toomre1964} factor $Q\gtrsim1$, for $R_\mathrm{trunc}\gtrsim10\,\mathrm{au}$.}, $R_0=30\,\mathrm{au}$ and $\alpha=10^{-3}$. It can be seen from the plot that the models populate different regions of the plane. In particular, as the truncation radius decreases, the accretion time scale, $t_\mathrm{acc}$, also decreases. This reduction in $t_\mathrm{acc}$ with $R_\mathrm{trunc}$ is apparent even in the initial conditions and results from the fact that the initial accretion rate is higher for a compact, high density disc and the millimetre flux is somewhat smaller due to optical depth effects. For $t\gg0$, this trend is determined by binary discs both being fainter and accreting more. The former effect is due to the faster drift of grains for smaller values of $R_\mathrm{trunc}$ \citep{Zagaria2021_theo}, which precipitates mass depletion \citep{Zagaria2021_obs}; the latter is a consequence of the smaller viscous time scale of truncated discs \citep{Rosotti&Clarke2018}. Remarkably, while in the case of single-star models, evolutionary tracks are limited by $t_\mathrm{acc}=0.1\,\mathrm{Myr}$ for $R_0=10\,\mathrm{au}$ \citep{Sellek2020}, when $R_\mathrm{trunc}\leq100\,\mathrm{au}$, the binary models are able to populate the region of the data-plane with shorter accretion time scales, where many of the observed sources lie.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_4.2_alpha.pdf}
\caption{Mass accretion rate, $\dot{M}_\mathrm{acc}$, as a function of the observer's equivalent disc mass (see Eq.~\ref{eq:2}), $M_\mathrm{dust}\times 100$, for models with $\alpha=3\times10^{-4}$, $R_\mathrm{trunc}=51.48\,\mathrm{au}$ and $R_\mathrm{trunc}=18.13\,\mathrm{au}$ on the left and right panel, respectively. In the background, lines of constant $t_\mathrm{acc}=0.1,\,1.0\,\mathrm{and}\,10.0\,\mathrm{Myr}$. \textbf{Left:} The solid tracks identify the $1\,\mathrm{Myr}\leq t\leq3\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval; Lupus discs are shown in grey for singles and black for binaries. \textbf{Right:} The solid tracks identify the $5\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ evolutionary interval, while the dotted lines describe the evolution outside this time interval; Upper Sco discs are shown in grey for singles and black for binaries. Larger circles identify transition discs and downward-pointing triangles are used for non-accretors. The colour bars refer to both panels.}
\label{fig:4.2}
\end{figure*}
We remark that the short $t_\mathrm{acc}$ in the models do {\it not} imply that the gas reservoir of the disc is severely depleted but instead that the discs have evolved to low (sub-)millimetre fluxes as a result of dust radial drift. The actual gas masses in our models are up to a factor of $1000$ larger than what would be obtained by assuming a standard dust-to-gas ratio. This effect is discussed extensively in \citet{Sellek2020} and we note that it is more marked in binary discs due to the way that tidal truncation accelerates radial drift.
In the upper-right panel of Fig.~\ref{fig:4.1}, models with different values of the initial disc scale radius, $R_0$, are plotted for $M_0=30\,M_\mathrm{Jup}$ and $\alpha=10^{-3}$, in the case of a single-star disc (solid lines) and a binary disc with $R_\mathrm{trunc}=50\,\mathrm{au}$ (dashed lines). The plot shows that the tidally truncated models are only slightly influenced by the scale radius. In fact, while in single-star models $R_0$ sets the initial viscous time scale, determining the scaling of $\dot{M}_\mathrm{acc}$ \citep[e.g.,][]{Lodato2017}, and the efficiency of radial drift \citep{Birnstiel2012}, such a time scale is only determined by $R_\mathrm{trunc}$ in close binary discs \citep{Rosotti&Clarke2018,Zagaria2021_theo}.
Finally, let us move on to the bottom, where the model dependence on viscosity, $\alpha$, is explored. The observer's equivalent dust mass is hardly affected by the viscous parameter when discs are young. However, by the age of Upper Sco, (sub-)millimetre fluxes are reduced more for higher values of $\alpha$, because the faster evolution of the gas raises the Stokes number and accelerates the radial drift of grains. This trend is observed also in single-star discs \citep{Sellek2020}, although to a lesser extent. Instead, the accretion rates depend more on viscosity and increase with $\alpha$, because it sets the velocity for gas transport thought the disc.
\subsection{Confronting models and data}
We compare binary disc models and observations in Lupus and Upper Sco. Although the Chamaeleon~I disc population and stellar multiplicity are well studied, the similar-age Lupus region is preferred because of its lower scatter in accretion time scale, better constrained (sub-)millimetre emission (85.7 \textit{vs} 79.8 per cent detections), and evenly censused stellar multiplicity (see details in Appendix~\ref{app:1}).
A systematic comparison between models and observations requires the disc truncation radius to be fixed to some value inferred from the data. As dealing with each source individually would be too complex (because of the difficulty of estimating individual truncation radii given the unknown binary eccentricity and projection effects: see Section 5.3), our models enforce a single value of the tidal truncation radius in each region. This is determined as $R_\mathrm{trunc}=0.33\times\langle a_\mathrm{p}\rangle$, where $\langle a_\mathrm{p}\rangle$ is the average binary separation in the data, and the previous relation applies to equal mass pairs in co-planar circular orbit \citep[e.g.,][]{Paczynski1977}. As a result of a systematically closer binary population in Upper Sco, the values of $R_\mathrm{trunc}$ are different in the two regions. For this reason, we do not discuss any evolutionary dependence between Lupus and Upper Sco discs in multiple stellar systems, which would be possible only if the binary separation distributions in the two regions were very similar.
Our results are displayed in Fig.~\ref{fig:4.2} where the mass accretion rate, $\dot{M}_\mathrm{acc}$, is plotted as a function of the observer's equivalent disc mass, $M_\mathrm{dust}\times 100$, for models with $\alpha=3\times10^{-4}$, and different values of the initial disc mass, $M_0$, and scale radius, $R_0$. Lines of constant $t_\mathrm{acc}$ are displayed as dashed-dotted lines. In the left-hand panel their solid sections refer to the $1\,\mathrm{Myr}\leq t\leq3\,\mathrm{Myr}$ age range, to be compared with Lupus data, while in the right-hand panel to the $5\,\mathrm{Myr}\leq t\leq10\,\mathrm{Myr}$ age range to be compared with Upper Sco data. The dotted lines describe the evolution outside these time intervals.
As expected, tidal truncation causes models to pass through the region of short accretion time scale occupied by the bulk of the binary data (region A in Fig.~\ref{fig:2.1}). This is due to dust being depleted more rapidly, because truncation prevents the retention of solids at large radii, which can otherwise resupply the disc at late times (see \citealt{Sellek2020_photo} for a discussion of how external photo-evaporation similarly leads to a more rapid depletion of solids by radial drift). In fact, the highest accretors in Upper Sco can be explained using viscous models with radial dust drift as long as multiplicity effects are considered.
Nevertheless, both in Lupus and Upper Sco, model evolutionary tracks under-predict the observed disc masses. While in Lupus this issue concerns only few sources, particularly with long accretion time scale, the bulk of the Upper Sco population is roughly ten times brighter than our models. Searching the literature provides possible explanations for Lupus outliers: some of them are sub-structured (e.g., GQ~Lup and HT~Lup), while others could be affected by thermal winds (e.g., Lup~818s), as is discussed in detail in Appendix~\ref{app:2}.
In Upper Sco the under-prediction of the (sub-)millimetre fluxes by 5 to $10\,\mathrm{Myr}$ models is rather marked and applies to most of the sources in the region. Clearly, this also depends on uncertainties in opacity, although it should be noted that the assumptions we made on the grain composition correspond to an opacity at the upper end of the possible range of variation (cf. e.g., \citealt{Birnstiel2018}), and therefore with any other assumption the theoretical tracks would shift to even lower masses, exacerbating the problem. A further reason that the observed fluxes could be higher than in the models could be that (sub-)millimetre observations of closer binaries could contain emission from two discs, since more than a half of binary pairs are not resolved at the angular resolution ($\approx0.45\,\mathrm{arcsec}$) of \citet{Barenfeld2016}. This would however boost the emission by at most a factor of 2, which is not enough to account for the differences with our models. Alternatively, an evolutionary argument would suggest that only the most massive binary discs survive until the age of Upper Sco. However, models with $R_\mathrm{trunc}$ values that are too large for the observed binary separations would be required to match the data, as shown in the upper-left panel in Fig.~\ref{fig:4.1}.
To sum up, tidally-truncated models of viscously evolving discs can reproduce the short accretion time scales in the bulk of the observed binary population. Halting the fast drift of solids is required to account for the (sub-)millimetre brightness of Upper Sco and some sources in Lupus.
\section{Discussion}\label{sec:5}
\subsection{Sub-structures in binary discs}
The first detection of gaps and rings in HL~Tau \citep{ALMA_HLTau} provided striking evidence of sub-structures in planet-forming discs (see \citealt{Andrews2020} for a review). Despite such structures often being assumed to be ubiquitous \citep{Andrews2018_DSHARP,Long2018}, their presence/absence in compact or low mass discs is yet to be established
(e.g., \citealt{Long2019}, Jennings et al. subm.).
Sub-structures have been detected in binary discs as well: e.g., spirals in HT~Lup~A and AS~205~N \citep{Kurtovic2018}, inner cavities in XZ~Tau~B \citep{Osorio2016}, UZ~Tau~E and CIDA~9A \citep{Long2019,Manara2019}, gaps and rings in AS~205~S \citep{Kurtovic2018}, GQ~Lup~A \citep{Long2020} and T~Tau~N \citep{Yamaguchi2021}. However, it is not clear if they are as common as in isolated discs. In fact, since binary discs are systematically smaller than those in single stars \citep{Manara2019}, identifying sub-structures in the former is more challenging than in the latter.
Here we focus on axisymmetric annular features and discuss their influence on the correlation between mass accretion rates and disc masses. In the case of single-star discs, \citet{Sellek2020} suggested that bright rings, consistent with trapping large grains in local pressure maxima \citep{Pinilla2012,Dullemond2018}, could be an explanation for discs with long accretion time scales too massive for their smooth models (region C in Fig.~\ref{fig:2.1}). It is then tempting to hypothesise that binary discs in region C are likewise sub-structured and their longer accretion time scale compared with those in region A could be similarly explained. However, while for single-star discs this inference is supported by the presence of imaged gaps and rings, none of our binary sources in region C has clearly detected gaps or cavities, making such systems preferential targets for a first higher-resolution analysis of sub-structures in binary discs.
Noticeably, whereas discs around single stars fill the region between population B and C continuously, disc in binaries appear to be well separated: for accretion rates $\lesssim10^{-10} M_\mathrm{Sun}\,\mathrm{yr}^{-1}$ there is a lack of such discs over about an order of magnitude in disc mass (see Fig.~\ref{fig:2.1}). While it cannot be ruled out that this effect is driven by the smaller sample size of discs in binaries, it may hint at a dichotomy of evolutionary paths. Binary discs containing sub-structures may remain over several Myr in region C, comparable to the situation in structured single-star discs. Instead, in the absence of sub-structures, our models predict a more rapid decay of disc luminosity in binaries than in singles (see top-left panel of Fig.~\ref{fig:4.1}), particularly for discs with low initial disc masses \citep{Birnstiel2012}. The presence of dust traps could thus explain the apparent lack of binary discs with low accretion rates and intermediate disc masses.
Such a picture, both for binaries and singles, implies that traps maintain discs in region C over time scales of a few Myr since sources are found in region C in both Lupus and Chamaeleon I which are approximately this age. Notably, this region is almost devoid of sources in Upper Sco. This would suggest that sources do not evolve \textit{vertically} downwards from region C, since otherwise Upper Sco would contain several systems with high (sub-)millimetre fluxes but low/undetected accretion rates. Alternatively dust traps may be disrupted on a time scale of 5 to $10\,\mathrm{Myr}$ \citep{Rosotti2013,Sellek2020}, so that systems evolve out of region C \textit{horizontally} to the left. Evidence of remnant sub-structures in Upper Sco discs would support such an interpretation. Furthermore, our understanding of the late time trajectory of discs in the plane of accretion rates vs (sub-)millimetre fluxes would be improved by increasing the completeness of accretion rate determinations in Upper Sco for the discs with the highest (sub-)millimetre fluxes.
The most intriguing interpretation of dust gaps is that of being carved by planets in the act of formation \citep[e.g.,][]{Zhang2018}. If we retain the hypothesis that region C sources (whether binary or single) are sub-structured, then it is tempting to extend to binaries the argument of \citet{Manara2019_pps} that discs in region C are forming giant planets, whose deep gaps not only halt radial drift but also reduce accretion. The existence of some binaries in region C would thus support the idea that giant planet formation may be under way in at least some of these discs.
We also consider whether sub-structures can explain the fact that the observed binaries in Upper Sco are too bright compared with model predictions (see right-hand panel of Fig.~\ref{fig:4.2}, contrasting the data, shown as black dots, with the models, displayed as solid curves). If this were the case then we would need dust traps to be almost ubiquitous. However, apart from GQ~Lup~A \citep{Long2020}, a source $\approx10$ times more massive than those in Upper Sco, we have no evidence of imaged gaps and rings in any of our binary discs. This can be easily explained by the limited angular resolution of the relevant surveys\footnote{$0.25\,\mathrm{arcsec}$ \citep{Ansdell2018}, $0.60\,\mathrm{arcsec}$ \citep{Pascucci2016} and $0.45\,\mathrm{arcsec}$ \citep{Barenfeld2016}; to be compared with, e.g., the $0.12\,\mathrm{arcsec}$ resolution of \citet{Long2019}.}.
A more compelling problem is that in this picture sub-structured discs would live in region A, very far from region C where they would be expected according to our previous argument. However, looking at single-star sources in region C highlights that all transition discs with imaged cavities are clustered here, suggesting that these sub-structures play a role in maintaining conditions not only of high disc masses but also relatively low accretion rates. We can then hypothesise that, in the binary as in the single-star case, while discs in region C are characterised by prominent, large scale sub-structures affecting both gas and dust evolution, shallower, small scale dust traps would be required in the case of Upper Sco, halting dust drift but not impacting substantially the gas.
Testing this hypothesis quantitatively requires detailed modelling of secular binary evolution in the presence of sub-structures, as well as higher-resolution observations in Upper Sco.
\subsection{Coagulation efficiency in binary discs}
Shallow dust traps at small radial locations are not the only solution to the mass budget problem in Upper Sco. This can be alternatively explained in terms of a reduced grain growth efficiency in close binary discs.
Tidal interactions between gravitationally bound stars and their discs can be at the origin of asymmetries, spiral arms, eccentric modes, vertical hydraulic jumps, mass transfer between discs and relative shocks... (e.g., \citealt{Nelson2000,Muller&Kley2012,Picogna&Marzari2013}). It is commonly hypothesised that these higher-dimensional effects could halt grain growth in circumstellar discs orbiting close binaries ($a\leq30\text{ to }50\,\mathrm{au}$, \citealt{Nelson2000,Picogna&Marzari2013})\footnote{Given the average separation of Lupus and Upper Sco binaries, this argument applies preferentially to discs in the latter region.}. When spirals are triggered by tidal interactions, the disc temperature rises enough for water and other volatile species to be vaporised. As ices make up to 60 per cent of our grain composition \citep{Tazzari2016,Rosotti2019_radii}, vaporisation would lead to solid dis-aggregation and thus promote fragmentation\footnote{Adopting $u_\mathrm{f}=10\,\mathrm{m}\,\mathrm{s}^{-1}$ as in Section~\ref{sec:3} is only appropriate for ice-coated grains \citep{Gundlach&Blum2015}, while $u_\mathrm{f}=1\,\mathrm{m}\,\mathrm{s}^{-1}$ is more suited for silicate-rich grains \citep{Blum&Wurm2008}.}. Even though grain growth could take place between spirals, it would then need to begin from material in gaseous phase and would be disrupted by the frequent interactions with the spirals themselves (expected to occur on a time scale of less than $100\,\mathrm{yr}$, \citealt{Nelson2000}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_5.1_alpha.pdf}
\caption{Same as in the right panel of Fig.~\ref{fig:4.2}, assuming $f_\mathrm{grow}=10$. The green line identifies an additional model
($R_0=30\,\mathrm{au}$ and $M_0=100\,M_\mathrm{Jup}$) with a smaller truncation radius corresponding to the closest binary in Upper Sco: $R_\mathrm{trunc}=6.7\,\mathrm{au}$.}
\label{fig:5.1}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_5.2_alpha.pdf}
\caption{Accretion time scale, $t_\mathrm{acc}$, as a function of the pair projected separation, $a_\mathrm{p}$, in Lupus, Chamaeleon~I and Upper Sco. Orange dots are used for discs in binary systems and grey dots for discs evolving in isolation. The blue tracks highlight our models expectations for $M_0=30\,M_\mathrm{Jup},\,R_0=30\,\mathrm{au},\,\alpha=3\times10^{-4}$.}
\label{fig:5.2}
\end{figure*}
These effects can be taken into account in our one-dimensional models by reducing the grain growth efficiency. To do so, we introduce a coagulation efficiency factor, $f_\mathrm{grow}$ \citep{Booth&Owen2020,Sellek2020}. So far we assumed $f_\mathrm{grow}=1$, meaning that all grain collisions lead to coagulation. Instead, if $f_\mathrm{grow}>1$, only a fraction $f_\mathrm{grow}^{-1}$ of the collisions result in growth, and sticking is thus less efficient, as expected in binary discs. A lower coagulation efficiency would also increase the drift time scale, given the balance between dust growth and drift achieved in the drift-dominated regime \citep{Birnstiel2012}. Recently, \citet{Booth&Owen2020} suggested that assuming $f_\mathrm{grow}=10$ could be beneficial for explaining the depletion of refractory elements in the Sun relative to nearby Solar twins. \citet{Sellek2020} ran single-star disc models with $f_\mathrm{grow}=10$, showing that they were able to reproduce massive Lupus discs that they could not explain assuming more efficient growth. However $f_\mathrm{grow}=10$ single star models could not match the data in Upper Sco, being too massive at any given accretion rate.
We retain $f_\mathrm{grow}=10$ as in these works because it provides a balance between fast growth, where solids drift fast, and slow growth, where grains are fragmentation-dominated throughout and the dust follows the gas \citep{Booth&Owen2020}. The results of our binary disc models with $f_\mathrm{grow}=10$, in the case of $R_\mathrm{trunc}=18.13\,\mathrm{au}$ and $\alpha=3\times10^{-3}$, are shown in Fig.~\ref{fig:5.1}. Regardless of the initial disc parameters, our models evolve towards disc masses that are ten times higher than for $f_\mathrm{grow}=1$ (as seen by \citealt{Sellek2020}). This difference reconciles binary models and observations in Upper Sco, with the caveat that the highest accretors are still not reproduced. This is easily explained by our choice of an average binary separation. Enforcing $R_\mathrm{trunc}=6.7\,\mathrm{au}$, corresponding to the minimum binary separation in Upper Sco, allows models to reproduce high accretors as well. This is shown in the representative case of $R_0=30\,\mathrm{au}$ and $M_0=100\,M_\mathrm{Jup}$, depicted with a solid green line in Fig.~\ref{fig:5.1}.
\subsection{Accretion time scale \textit{vs} binary projected separation}
Using one-dimensional viscous models, we have shown that the fact that proto-planetary discs in multiple systems occupy region A in Fig.~\ref{fig:2.1} (where $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$) can be reproduced by the secular effects of outer disc truncation on dust and gas evolution. Such models suggest that a shorter accretion time scale in discs in closer binaries should be expected, owing to the larger accretion rates and lower dust masses predicted in those systems (see e.g., top-left panel of Fig.~\ref{fig:4.1}).
Let us therefore examine how the observationally-inferred accretion time scale, $t_\mathrm{acc}$, depends on the measured projected binary separation, $a_\mathrm{p}$. This is displayed in Fig.~\ref{fig:5.2} for Lupus, Chamaeleon~I and Upper Sco binary discs, using orange dots. Single-star discs are also shown for comparison, using grey dots. No particular trend of $t_\mathrm{acc}$ with $a_\mathrm{p}$ can be recognised in the data, at odds with our expectations. These are highlighted by the blue tracks and shaded area for the representative case with $M_0=30\,M_\mathrm{Jup},\,R_0=30\,\mathrm{au},\,\alpha=3\times10^{-4}$ and its $1\sigma$ spread.
However, what primarily sets the accretion time scale in our models is the location where truncation occurs. In fact, this is not determined by the binary projected separation only, but is expected to be affected by other effects, such as the relative inclination of the interacting pair with respect to the plane of the sky \citep{Manara2019}, the binary mass ratio and eccentricity \citep[e.g.,][]{Paczynski1977,Artymowicz&Lubow1994,Pichardo2005}, and the relative misalignment between the disc and the binary plane \citep{Lubow2015}.
In particular, comparing the binary separation distribution in our data with the observed relation between main sequence binaries eccentricity and orbital period \citep{Raghavan2010}, would imply that some discs in our sample could have a significant eccentricity, thus washing out the expected dependence of $t_\mathrm{acc}$ on $a_\mathrm{p}$. Recent evidence that eccentricity in wider binaries follows a thermal rather than uniform distribution (see \citealt{Hwang2021}, Fig.~6) would suggest that the most eccentric pairs would be those with larger $a_\mathrm{p}$. As long as the bulk of the population has relatively smaller eccentricities ($e\approx0.3$ as expected in the field, e.g., \citealt{Duchene&Kraus2013}), this is not in tension with the fact that the the dust disc sizes in Taurus and $\rho$~Ophiuchus can be well explained assuming low eccentric orbits (\citealt{Zagaria2021_obs} and \citealt{Rota2022}).
Even though the comparison between models and specific sources can be challenged by uncertainties on binary eccentricities, we do not expect them to substantially modify our results in Section~\ref{sec:4}, because they would reduce the average truncation radius only by a factor of a few and this is potentially compensated by the underestimation of the binary separation by its projection on the plane of the sky. Nevertheless, recovering the theoretically-expected trend of $t_\mathrm{acc}$ with $a_\mathrm{p}$ is only part of the problem, given the large scatter in Fig.~\ref{fig:5.2}. In fact, on top of the previously mentioned mechanisms, other effects related to grain coagulation efficiency, disc evolution and dispersal can influence the accretion time scale. We do not explore further these possibilities because, given our uncertainty on the determination of $R_\mathrm{trunc}$, modelling would be under constrained by the data. Finally, some of the sources in Fig.~\ref{fig:5.2} can also be components of undetected higher-order multiples.
\section{Summary and conclusions}\label{sec:6}
In this paper we made use of gas and dust one-dimensional simulations to address how stellar multiplicity can influence the correlation between mass accretion rates onto forming stars and the masses of their proto-planetary discs.
Hereafter our main conclusions are summarised:
\begin{itemize}
\item We searched the literature for information on the stellar multiplicity of Lupus, Chamaeleon~I and Upper Sco discs with both (sub-)millimetre emission and accretion rates in \citet{Manara2016,Manara2020} and \citet{Mulders2017}. We found that roughly 20 per cent of the targeted discs are in binary or higher-order multiple stellar systems;
\item We introduced the accretion time scale, $t_{\rm acc}$, as the ratio of observationally estimated disc mass to accretion rate (see Eq.~\ref{eq:1}) and found that discs in gravitationally bound pairs have a systematically shorter accretion time scale than discs evolving in isolation, clustering around $t_\mathrm{acc}\approx0.1\,\mathrm{Myr}$. The short accretion time scales is a consequence both of the well known fact that dust masses are lower in binaries compared with single stars \citep[e.g.,][]{Harris2012} {\it and} of the fact that accretion rates are somewhat higher than in single stars (see Fig.~\ref{fig:2.1});
\item To understand the trend in the data, we ran models of tidally truncated proto-planetary discs subject to viscous evolution, grain growth and radial drift. We found that our models show shorter accretion time scales for lower values of $R_\mathrm{trunc}$, as a consequence of the faster radial drift and reduced disc lifetime determined by tidal truncation (see upper-left panel of Fig.~\ref{fig:4.1});
\item Models and data agree reasonably well in Lupus (see left-hand panel of Fig.~\ref{fig:4.2}); we suggest that a small number of anomalously (sub-)millimetre bright discs may be explained by the existence of sub-structures, while a disc with anomalously long $t_\mathrm{acc}$, may potentially undergo internal photo-evaporation;
\item In Upper Sco our models systematically under-predict disc masses by up to a factor of ten (see right-hand panel of Fig.~\ref{fig:4.2}). It is unclear if this could be the effect of undetected traps halting the drift of solids. We also find that the agreement is improved if the grain growth efficiency is reduced (see Fig.~\ref{fig:5.1}) as has been hypothesised in the dynamically active environments of discs in binaries:
\item Finally, we find no clear evidence that the observationally-inferred accretion time scale increases with the binary projected separation (see Fig.~\ref{fig:5.2}), at odds with our theoretical expectation that the accretion time scale should increase with the tidal truncation radius of the disc. However, the relationship between projected separation and tidal truncation radius is complicated by projection effects and the orbital eccentricity: given the relatively limited dynamic range of projected separations in the sample, this lack of correlation is therefore not necessarily surprising.
\end{itemize}
\section*{Acknowledgements}
We thank the anonymous reviewer for their comments that helped to improve the manuscript. FZ acknowledges support from STFC and Cambridge Trust for a Ph.D. studentship. GR acknowledges support from the Netherlands Organisation for Scientific Research (NWO, program number
016.Veni.192.233) and from an STFC Ernest Rutherford Fellowship
(grant number ST/T003855/1). This work was partly supported by the Deutsche Forschungs-Gemeinschaft (DFG, German Research Foundation) - Ref no. FOR 2634/1 TE 1024/1-1. This work has also been supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska Curie grant agreement number 823823 (DUSTBUSTERS).
Software: \texttt{numpy} \citep{numpy20_2020Natur.585..357H}, \texttt{matplotlib} \citep{matplotlib_Hunter:2007}, \texttt{scipy} \citep{scipy_2020SciPy-NMeth}, \texttt{JupyterNotebook} \citep{Jupyter_nootbok}. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France \citep{Wenger2000}, and the Binary Star Database by Kendra Kellogg, Lowell Observatory. Colours in Fig.s~\ref{fig:4.2} and \ref{fig:5.1} are from \texttt{ColorBrewer.org} by Cynthia A. Brewer, Geography, Pennsylvania State University.
\section*{Data Availability}
The code used in this paper is publicly available on GitHub at
\texttt{github.com/rbooth200/DiscEvolution}. The data underlying
this paper are available in the ALMA archive as explained in the papers quoted in Section~\ref{sec:2} and Appendix~\ref{app:1}.
\bibliographystyle{mnras}
| 830acdb79de2301bfaffd2b40519b91a3f45198a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Natural hazards associated with geophysical mass movements consist of a mixture of granular materials of different sizes of particles and the fluid with their respective physical properties. There have been rapid advancements in modeling shallow granular material (Savage and Hutter, 1989; Denlinger and Iverson, 2001; Pouliquen and Forterre, 2002, McDougall and Hungr, 2005; Pudasaini and Hutter, 2007; Luca et al., 2009; Kuo et al., 2011) and particle fluid mixture (Iverson, 1997; Pitman and Le, 2005; Pudasaini, 2012; Pudasaini and Mergili, 2019) mass flow modelling. These models are primarily based on the hydrostatic pressure assumptions. However, due to the centrifugal acceleration, the mass flows in curved channels also include some non-hydrostatic contributions to hydraulic pressure gradients and the Coulomb friction forces because of the enhanced normal load (Savage and Hutter, 1991; Gray et al., 1999; Pudasaini et al., 2005). Furthermore, Pailha and Pouliquen (2009), and Pudasaini (2012) showed that the pressure in mixture mass flows can be non-hydrostatic due to the Newtonian and non-Newtonian viscous contributions, the particle concentration distributions, and the relative velocity between particle and fluid.
\\[3mm]
Classically, modeling geophysical flows is usually based on the hydrostatic, depth-averaged mass and momentum balance equations. Hydrostatic flow models are based on the assumption that the slope parallel length scale is much larger than the length scale in the flow depth direction. However, if the similar length scalings are required in the slope parallel and the flow depth directions, then the gravity and the vertical acceleration can have the same order of magnitude effects (Denlinger and Iverson, 2004; Pudasaini and Hutter, 2007, Castro-Orgaz et al., 2015). This may call for the use of the full (without reducing to the hydrostatic condition) momentum equation also in the slope normal direction as in the slope parallel directions (Domnik et al., 2013). Denlinger and Iverson (2004) first mentioned that the vertical accelerations in granular mass flows can be of the same order of magnitude as the gravity acceleration. In this situation, the vertical acceleration can be as significant as the acceleration in the slope parallel direction. This is particularly so for steep, irregular and curved slopes where there is a substantial acceleration of the material in the flow depth direction. So, they suggested for the non-hydrostatic modeling of granular flows. This requires enhancements of the normal stress (in the slope normal or vertical direction) that results in the enhancements of the acceleration, friction and fluxes in the momentum balance equations (Castro-Orgaz et al., 2015; Yuan et al., 2018). Since the Coriolis acceleration is usually neglected in geophysical mass flows (such as landslide and avalanches), two-types of accelerations can be induced affecting the normal stress distributions of the free surface flows (Castro-Orgaz et al., 2015): First, the acceleration due to the real forces acting at the bed-normal direction. Second, the centripetal acceleration that arises due to the curved flow path (Savage and Hutter, 1991; Pudasaini and Hutter, 2007). The first is the main contributor of the Boussinesq-type models, while both combined result in the more comprehensive non-hydrostatic flows.
\\[3mm]
Following the work of Boussinesq (1872, 1877), the free surface water flow simulations are generally based on non-hydrostatic depth-averaged models. Fundamental further contributions in including Boussinesq-type non-hydrostatic and dispersive effects in water waves are also due to Serre (1953), Peregrine (1967) and Green et al. (1976). The recent advancements, analyses and applications of the dispersive wave characteristics of the Boussinesq system with advanced numerical schemes for real flow simulations include the works by
Nwogu (1993), Wei and Kirby (1995), Madsen and Sch\"affer (1998), Kennedy et al. (2000),
Stansby (2003), Chen et al. (2003), Erduran et al. (2005), Chen (2006), and Kim and Lynett (2011). For detailed review on it, we refer to Castro-Orgaz et al. (2015). However, for shallow granular flows the effect of nonzero vertical acceleration on depth-averaged momentum fluxes and stress states were first included by Denlinger and Iverson (2004) while modeling granular flows across irregular terrains. This was later extended by Castro-Orgaz et al. (2015) resulting in the novel Boussinesq-type theory for granular flows. Utilizing the non-hydrostatic Boussinesq-type water wave theory, Castro-Orgaz et al. (2015) rigorously developed a non-hydrostatic depth-averaged granular flow model. Considering the vertical motion of particles, they explicitly determined the vertical velocity, vertical acceleration, and vertical normal stresses from the mass and momentum conservation equations. They have shown that granular mass flow can be described by fully non-linear, Boussinesq-type gravity waves, generalizing the basic Boussinesq-type water wave theory used in civil and coastal engineering to granular mass flows. Later, Yuan et al. (2018) advanced further by presenting a refined and more complete non-hydrostatic shallow granular flow model. They also cast their model in to a usual Boussinesq-type water wave equations.
\\[3mm]
In developing the non-hydrostatic Boussinesq-type gravity wave models for granular flows, both Castro-Orgaz et al. (2015) and Yuan et al. (2018) considered the vertical momentum equation, assuming the shallowness of the flow depth and the constant velocity profiles of the horizontal velocity components. Along with these assumptions, there are three key aspects in their model development: Obtaining the vertical normal stress component from the vertical momentum equation, an expression for the vertical velocity component in terms of the horizontal mass flux (divergence), and the definition of the depth integration of the vertical velocity component from a generic elevation to the free surface. Finally, the depth averaged mass and momentum equations, together with these three considerations lead to a non-hydrostatic Boussinesq-type gravity wave models for granular flows. However, all these formulations are primarily based on the global horizontal-vertical Cartesian coordinate for a single-phase granular flows.
\\[3mm]
One- and two-phase models cannot appropriately represent many important aspects of very complex mass flows in terms of material composition and interactions among the involved phases. The rheological properties and flow dynamics are governed by coarse and fine solids, and viscous fluid, i.e., typically three phases (Iverson, 1997; Pierson, 1970, 2005; de Haas et al., 2015, 2016; Pudasaini and Fischer, 2020b). Consequently, the most complex model family for geophysical mass flows should aim at describing the flow as (typically) a three-phase mixture, as often observed in the field and experiments (Major and Iverson, 1999; McArdell et al., 2007; Conway et al., 2010; Schneider et al., 2011; Johnson et al., 2012; de Haas et al., 2015, 2016; Steinkogler et al., 2015). {In general terms, the mechanical components} in the mixture mass flow can be divided into three constituents: The fluid phase is a mixture of water and very fine particles (clay, silt, colloids), the fine-solid phase consists of sand and particles larger than clay and silt, and the solid phase represents the coarse material. These materials can be described as viscoplastic, Coulomb-viscoplastic, and Mohr-Coulomb continuum.
With this, Pudasaini and Mergili (2019) proposed a novel multi-phase, multi-mechanical mass flow model, by extending the two-phase viscous fluid and Coulomb solid model (Pudasaini, 2012) to additionally combine it with the fine-solid material. The Pudasaini and Mergili (2019) model can accurately simulate complex cascading multi-phase natural events (Mergili and Pudasaini, 2021; Mergili et al., 2020a, 2020b; Shugar et al., 2021).
\\[3mm]
Here, we extend and utilize the above mentioned ideas to the multi-phase mass flow model (Pudasaini and Mergili, 2019) to generate a new non-hydrostatic Boussinesq-type gravity wave model for multi-phase mass flows in a locally inclined Cartesian coordinate system (Pitman and Le, 2005; Pudasaini, 2012). The new non-hydrostatic multi-phase mass flow model includes enhanced gravity and dispersion effects as in the single-phase models by Denlinger and Iverson (2004), Castro-Orgaz et al. (2015) and Yuan et al. (2018). But, our new model further includes interfacial momentum transfers in the non-hydrostatic Boussinesq-type model formulation representing the complex multi-phase nature of the mass flow.
We delineate the fundamentally new contributions in the Boussinesq-type gravity waves in mass flows emerging from the phase-interactions. This includes buoyancy, drag, virtual mass and Newtonian plus non-Newtonian viscous effects.
We outline the first-ever application potential of the dispersive multi-phase mass flows. As in the effective gravity, the dispersive terms are strongly coupled, e.g., due to the interfacial drag and virtual mass contributions. There are direct and strong couplings between the solid, fine-solid and the fluid components among these dispersion relations. Interfacial drags bring completely new mechanisms in the non-hydrostatic, dispersion relations.
We discuss some particular situations where the non-hydrostatic dispersive effects are more pronounced in multi-phase particle-fluid mixture mass flows than in single-phase flows.
So, this contribution sets a foundation for a more comprehensive and general frame for the simulation of dispersive, multi-phase mass flows.
We also present simplified models that might be helpful in solving the equations with reduced complexity.
The reduced models already appeared to be the important generalizations and extensions of some mass flow models available in the literature.
We formally postulate a new, spatially varying dissipative force, called the prime-force, which can physically precisely control the mass flow dynamics, run-out and the deposition. We present a simple dispersion model and its solution. Dispersion produces a wavy velocity field about the reference state without dispersion. The dispersion increases greatly as the solid volume fraction or the basal friction decreases. These are new understandings for the motion of a dispersive landslide.
\section{Construction of the Model}
\subsection{Non-Hydrostatic Contributions}
Following Pudasaini and Mergili (2019) and Pudasaini and Fischer (2020a), first, we define the variables and parameters.
Let the solid, fine-solid and fluid phases be denoted by the suffices $s$, $fs$, $f$, respectively. The fluid phase is {governed} by its true density $\rho_f$, viscosity $\eta_f$, and isotropic stress distribution; the fine-solid and solid phases are characterized by their true densities $\rho_{fs}$, $\rho_s$; internal friction angles $\phi_{fs}$, $\phi_{s}$; basal friction angles $\delta_{fs}$, $\delta_{s}$; and anisotropic stress distribution, $K_{s}$ (lateral earth pressure coefficient); and the viscosity of the fine-solid $\eta_{fs}$. Furthermore,
$\gamma_s^f = \rho_f/\rho_s$,
$\gamma_s^{fs} = \rho_{fs}/\rho_s$,
$\gamma_{fs}^f = \rho_{f}/\rho_{fs}$
are the fluid to solid, fine-solid to solid and fluid to fine-solid density ratios, $\nu_f^e$ and $\nu_{fs}^e$ are the effective kinematic viscosities for the fluid and fine-solid, $\mu_s = \tan\delta_s$ and $\mu_{fs} = \tan\delta_{fs}$ are the friction coefficients for the solid and fine solid.
Let ${\bf u}_s = \lb u_s, v_s, w_s\rb$, ${\bf u}_{fs} = \lb u_{fs} , v_{fs} , w_{fs} \rb$, ${\bf u}_f = \lb u_f, v_f, w_f \rb$, and $\alpha_s$, $\alpha_{fs}$, $\alpha_f$ denote the velocities with their components along the flow directions ($x, y, z$), and the volume fractions for the solid, fine-solid, and fluid constituents. {Similarly,}
$p_{fs}$ and $p_{f}$ are the pressures, ${C}_{DG}$ and ${C}_{vm}$ constitute the interfacial force densities, namely, the drags and the virtual mass forces, and $C_{DV}$ are the viscous drag coefficients. The superscript-pair represents the considered phases, e.g., ${C}_{DG}^{s,f}$ means the drag force {exerted by fluid on solid}, ${\mathcal C}$ are the virtual mass coefficients, $T_{**}$ are the components of the Cauchy stress tensor, $\jmath = 1 \, \mbox{or}\, 2$ correspond to linear or quadratic drag coefficients, $g^x, g^y, g^z$ are the components of gravitational acceleration, basal- and the free-surface of the flow are denoted by $b = b(t,x,y)$ and $s = s(t,x,y)$, and $h = s - b$ is the flow depth.
\subsubsection{Derivation of Normal Stress Components}
The non-hydrostatic modelling framework includes two important and essential components: ($i$) enhanced gravity, and ($ii$) dispersive contributions (see, e.g., Castro-Orgaz et al., 2015; Yuan et al., 2018). Both emerge from the consideration of the momentum equation in the flow depth direction such that the normal component of the velocity is retained, that was neglected in simple hydrostatic model developments as discussed at Section 1. These contributions, however, are modelled in terms of the slope parallel velocity gradients or fluxes. For this, following Pudasaini (2012), and Pudasaini and Mergili (2019), first we consider the solid momentum balance in the flow depth direction:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial}{\partial t} \lb w_s
- \gamma^f_s \mathcal C^{s,f} \lb w_f - w_s \rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb w_{fs} - w_s \rb
\rb}\\[3mm]
\displaystyle{+\frac{\partial}{\partial x} \lb u_sw_s
- \gamma^f_s \mathcal C^{s,f} \lb u_fw_f - u_sw_s \rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb u_{fs}w_{fs} - u_{s}w_s \rb
\rb
}
\\[3mm]
\displaystyle{+\frac{\partial}{\partial y} \lb v_sw_s
- \gamma^f_s \mathcal C^{s,f} \lb v_fw_f - v_sw_s \rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb v_{fs}w_{fs} - v_sw_s \rb
\rb
}
\\[3mm]
\displaystyle{+\frac{\partial}{\partial z} \lb w_s^2
- \gamma^f_s \mathcal C^{s,f} \lb w_f^2 - w_s^2 \rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb w_{fs}^2 - w_s^2 \rb
\rb
}\\[3mm]
= -\lb 1 - \gamma^f_s\rb g^z-\lb
\mu_s\frac{\partial T_{zx_s}}{\partial x}
+\mu_s\frac{\partial T_{zy_s}}{\partial y}
+\frac{\partial T_{zz_s}}{\partial z}
\rb\\[3mm]
+\displaystyle{
\frac{1}{\alpha_s}\left [
C_{DG}^{s,f} \lb {w}_f -{w}_s \rb|{\bf u}_f -{\bf u}_s|^{j-1}
+C_{DG}^{s,fs} \lb {w}_{fs} -{w}_s \rb|{\bf u}_{fs} -{\bf u}_s|^{j-1}
- C_{DV}^s w_s|{\bf u}_s|\alpha_s\right ]
},
\label{w_solid_momentum}
\end{array}
\end{eqnarray}
\end{linenomath*}
where, for simplicity, $\alpha_s$ has been taken out. Note that since both $C_{DG}^{s,f}$ and $C_{DG}^{s,fs}$ contain $\alpha_s$ in their numerators (see, Appendix), appearance of $1/\alpha_s$ in (\ref{w_solid_momentum}) makes no problem. It is important to note that (\ref{w_solid_momentum}) contains the normal stress $T_{zz_s}$ from which we can construct the full description of the normal stress in the flow depth direction that includes all the essential components emerging from the flow dynamics and interfacial momentum transfers in excess to the usual hydrostatic normal load that is simply associated with the gravity load in the flow depth direction.
\\[3mm]
We define a new variable $\eta = z -b$, the relative flow depth. Then, following the procedure in Castro-Orgaz et al. (2015) and Yuan et al. (2018), integrating (\ref{w_solid_momentum}) from the generic elevation $z$ to the free surface $s$, neglecting the shear stresses, and using the tractionless condition at the free surface (Pudasaini, 2012; Pudasaini and Mergili, 2019), we obtain an expression for the normal stress in terms of $\eta$:
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\tau_{zz_s}(\eta) = \lb 1 - \gamma^f_s\rb g^z(h-\eta)
+ \frac{\partial}{\partial t}\left [ I_s
- \gamma^f_s \mathcal C^{s,f} \lb I_f - I_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb I_{fs} - I_s\rb
\right ]} \\[3mm]
\hspace{1.2cm}
\displaystyle{
+ \nabla\cdot\left [ I_s{\bf u}_s
- \gamma^f_s \mathcal C^{s,f} \lb I_f {\bf u}_f - I_s {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb I_{fs} {\bf u}_{fs} - I_s {\bf u}_s\rb
\right ]}\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \left [ w_s^2
- \gamma^f_s \mathcal C^{s,f} \lb w_f^2 - w_s^2\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb w_{fs}^2 - w_s^2\rb
\right ]}\\[3mm]
\hspace{1.2cm}
\displaystyle{
-\frac{1}{\alpha_s}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}\left [ I_f - I_s\right ]
+ C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}\left [ I_{fs} - I_s\right ]
- C_{DV}^s w_s|{\bf u}_s|\alpha_s I_s
\right ],}
\label{tau_zz_a}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-2.5mm}
where
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
I_s = \int_z^s w_s dz',\,\,\,\,\,
I_{fs} = \int_z^s w_{fs} dz',\,\,\,\,\,
I_f = \int_z^s w_f dz';\\[3mm]
w_s = w_{b_s} - \lb \nabla\cdot {\bf u}_s\rb\eta,\,\,\,\,\,
w_{fs} = w_{b_{fs}} - \lb \nabla\cdot {\bf u}_{fs}\rb\eta,\,\,\,\,\,
w_f = w_{b_f} - \lb \nabla\cdot {\bf u}_f\rb\eta;\\[3mm]
\nabla\cdot {\bf u}_{s} = \partial u_s/\partial x + \partial v_s/\partial y, \,\,\,\,\,
\nabla\cdot {\bf u}_{fs} = \partial u_{fs}/\partial x + \partial v_{fs}/\partial y,\,\,\,\,\,
\nabla\cdot {\bf u}_{f} = \partial u_{f}/\partial x + \partial v_{f}/\partial y.
\label{tau_zz_aa}
\end{array}
\end{eqnarray}
\end{linenomath*}
We depth-integrate $w_s$, and define $\hat I_s$ (similar structures hold for fine-solid and fluid):
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\bar w_s: = \frac{1}{h}\int_b^s w_s dz' = w_{b_s} -\lb \nabla\cdot {\bf u}_s\rb\frac{h}{2},\,\,\,\,\,
w_{b_s} = u_s\frac{\partial b}{\partial x} + v_s\frac{\partial b}{\partial y}};
\\[5mm]
\hat I_s: = \int_b^z w_s dz'= \int_b^s w_s dz' - \int_z^s w_s dz' = h \bar w_s -I_s,
\label{tau_zz_aaa}
\end{array}
\end{eqnarray}
\end{linenomath*}
where $b$ is the basal topography.
Equations (\ref{tau_zz_a})-(\ref{tau_zz_aaa}) constitute the fundamental basis for the non-hydrostatic dispersive model development.
With (\ref{tau_zz_aaa}), (\ref{tau_zz_a}) takes the form:
{\small
\\[-5mm]
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\tau_{zz_s}(\eta) = \lb 1 - \gamma^f_s\rb g^z(h-\eta)+ h\frac{\partial}{\partial t}\left [ \bar w_s
- \gamma^f_s \mathcal C^{s,f} \lb \bar w_f - \bar w_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb \bar w_{fs} - \bar w_s\rb
\right ]}
\\[3mm]
\hspace{1.2cm}
\displaystyle{
+h \left [
{\bf u}_s \cdot \nabla \bar w_s
- \gamma^f_s \mathcal C^{s,f}\lb {\bf u}_f \cdot \nabla \bar w_f - {\bf u}_s \cdot \nabla \bar w_s \rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb {\bf u}_{fs} \cdot \nabla \bar w_{fs} - {\bf u}_s \cdot \nabla \bar w_s \rb
\right ] }
\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \frac{\partial}{\partial t}\left [ \hat I_s
- \gamma^f_s \mathcal C^{s,f} \lb \hat I_f - \hat I_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb \hat I_{fs} - \hat I_s\rb
\right ] }\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \nabla\cdot\left [ \hat I_s{\bf u}_s
- \gamma^f_s \mathcal C^{s,f} \lb \hat I_f {\bf u}_f - \hat I_s {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb \hat I_{fs} {\bf u}_{fs} - \hat I_s {\bf u}_s\rb
\right ]}\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \left [ w_s^2
- \gamma^f_s \mathcal C^{s,f} \lb w_f^2 - w_s^2\rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb w_{fs}^2 - w_s^2\rb
\right ]}\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \frac{1}{\alpha_s}C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}\left [ \lb h\bar w_f- \hat I_f\rb - \lb h\bar w_s -\hat I_s\rb\right ]}\\[3mm]
\hspace{1.2cm}
\displaystyle{
- \frac{1}{\alpha_s} C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}\left [ \lb h\bar w_{fs} - \hat I_{fs}\rb - \lb h \bar w_{s} -\hat I_s\rb\right ]
+C_{DV}^s\left[\left( h\bar w_s-\hat I_s\right)\right]|{\bf u}_s|.}
\label{tau_zz_aaaa}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
In this representation, the first term on the right hand side contains the complementary relative flow depth, $(h-\eta)$, and indicates that at the bottom $(\eta = 0)$ it is $\lb 1 - \gamma^f_s\rb g^z h$, and at the free surface $(\eta = h)$ it is zero. So, that term is the usual hydrostatic normal load often used in shallow flow models together with the buoyancy effect $\lb 1 - \gamma^f_s\rb$. Thus, the appearance of $(h-\eta)$ in $\lb 1 - \gamma^f_s\rb g^z(h-\eta)$ implies its linear distribution from the bottom to the free surface, it is advantageous. Therefore, we should also try to transfer the other terms in (\ref{tau_zz_aaaa}) such that they contain some functions of $(h-\eta)$ and/or $\eta$. This will be achieved next.
\\[3mm]
With its definition in (\ref{tau_zz_aaa}), $\hat I_{s}$ (similar for fine-solid and fluid) can be obtained from (\ref{tau_zz_aa}) as:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\hat I_{s} = w_{b_s} \eta - \lb \nabla\cdot {\bf u}_s\rb\frac{\eta^2}{2}}.
\label{tau_zz_aaaaa}
\end{array}
\end{eqnarray}
\end{linenomath*}
This helps in producing desired terms with factors $h-\eta$ and/or $\eta$; see below.
\subsubsection{Effective Normal Loads}
{\bf A. The Solid Normal Load:} Now, define $D/Dt = \partial/\partial t + {\bf u}_s\cdot \nabla$ (similar for fine-solid and fluid). Then, with (\ref{tau_zz_aaaaa}), following the procedures in Yuan et al. (2018), after a lengthy calculations, (\ref{tau_zz_aaaa}) takes the form:
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
{\displaystyle
\tau_{zz_s} = \lb 1 - \gamma^f_s\rb g^z(h-\eta)
+\frac{D}{D t} \left[ \bar w_s
-\gamma^f_s \mathcal C^{s,f}\lb \bar w_f - \bar w_s\rb
-\gamma^{fs}_s \mathcal C^{s,fs}\lb \bar w_{fs} - \bar w_s\rb
\right](h-\eta)}\\[3mm]
\hspace{.7cm}
{\displaystyle-\frac{1}{2}\Bigg \{\frac{D}{D t} \left[
h \nabla\cdot {\bf u}_s
- \gamma^f_s \mathcal C^{s,f}\lb h \nabla\cdot {\bf u}_f - h \nabla\cdot {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb h \nabla\cdot {\bf u}_{fs} - h \nabla\cdot {\bf u}_s\rb
\right ]\eta}\\[5mm]
\hspace{.7cm}
{\displaystyle
-\frac{D}{D t} \left[
\nabla\cdot {\bf u}_s
- \gamma^f_s \mathcal C^{s,f}\lb \nabla\cdot {\bf u}_f - \nabla\cdot {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb \nabla\cdot {\bf u}_{fs} - \nabla\cdot {\bf u}_s\rb
\right ]\eta^2}\\[3mm]
\hspace{.7cm}
{\displaystyle
+ \left[
\lb\nabla\cdot {\bf u}_s\rb^2
- \gamma^f_s \mathcal C^{s,f}\lb \lb\nabla\cdot {\bf u}_f\rb^2 - \lb\nabla\cdot {\bf u}_s\rb^2\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb \lb \nabla\cdot {\bf u}_{fs}\rb^2 - \lb\nabla\cdot {\bf u}_s\rb^2\rb
\right ]\eta^2
\Bigg\}}
\\[3mm]
\hspace{.7cm}
{\displaystyle
- \frac{1}{\alpha_s}C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\left [
\lb \bar w_f - \bar w_s\rb \lb h - \eta\rb
- \lb \nabla\cdot\lb {\bf u}_f -{\bf u}_s\rb\rb\frac{1}{2}\eta \lb h -\eta\rb
\right ]}\\[3mm]
\hspace{.7cm}
{\displaystyle
- \frac{1}{\alpha_s}C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}\left [
\lb \bar w_{fs} - \bar w_s\rb \lb h - \eta\rb
- \lb \nabla\cdot\lb {\bf u}_{fs} -{\bf u}_s\rb\rb\frac{1}{2}\eta \lb h -\eta\rb
\right ]}\\[3mm]
\hspace{.7cm}
\displaystyle{
+\,C_{DV}^s|{\bf u}_s|\left[ \bar w_s\lb h-\eta\rb - \lb\nabla\cdot {\bf u}_s\rb\frac{1}{2}\eta\lb h-\eta\rb\right]
,}
\label{tau_zz_aaaaaa}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
which is the effective normal load for the solid component.
Note that (\ref{tau_zz_aaa}) and (\ref{tau_zz_aaaaa}) are utilized to obtain the structures associated with the drags.
$\tau_{zz_s}$ in (\ref{tau_zz_aaaaaa}) is written entirely in terms of the flow variables, flow dynamics and the phase-interaction terms. There are two types of terms in (\ref{tau_zz_aaaaaa}). First, the slope normal acceleration terms associated with $(h-\eta)$, which are linear in $\eta$. Second, the slope parallel (divergence, or flux) terms that are either linear or quadratic in $\eta$.
However, it is interesting to note that the interfacial drag contributions have two types of terms. First, in $C_{DG}^{s,f}$, the associated term $\lb \bar w_f - \bar w_s\rb \lb h - \eta\rb$ has a factor $\lb h - \eta\rb$ as in the usual gravity and the acceleration terms ($g^z$ and $D/Dt$). This term vanishes at the free surface. Second, $\lb \nabla\cdot\lb {\bf u}_{f} -{\bf u}_s\rb\rb\frac{1}{2}\eta \lb h -\eta\rb$ is quadratic in $\eta$, but has a special form. Such term with factor $\eta \lb h -\eta\rb$ does not appear in other contributions in $\tau_{zz_s}$. This vanishes both at the bottom and at the free surface of the flow and thus has maximum in between the flow depth. Similar analysis holds for the terms associated with $C_{DG}^{s,fs}$. So, the interfacial drags bring completely new mechanisms in the non-hydrostatic (dispersion) relations. The important point now is that, due to their structures, the first terms in the drag contributions must be (or better to) put together with the gravity and the acceleration terms, $g^z$ and $D/Dt$ (associated with $\bar w$). We consider these terms together in obtaining the enhanced gravity. Furthermore, the $D/Dt$ are due to the normal acceleration of the solid particles, and the relative acceleration of the solid particles with respect to the fine-solid and fluid. So, all $g^z, D/Dt$ and $C_{DG}$ terms (associated with $(h-\eta)$) basically represent the normal acceleration, or force. All the other remaining terms in (\ref{tau_zz_aaaaaa}) represent the dynamics and forcings in the slope parallel direction. For this reason, we re-write (\ref{tau_zz_aaaaaa}) as the first group of terms with the factor $(h-\eta)$, containing the usual gravity (including buoyancy, $\lb 1 - \gamma^f_s\rb g^z$), and the normal acceleration ($D/Dt$ terms including virtual mass) and drag terms ($C_{DG}$), and the second group of terms with $\eta$ and $\eta^2$ representing the slope parallel motion as:
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\tau_{zz_s} = \lb 1 - \gamma^f_s\rb g^z(h-\eta)
+\frac{D}{D t} \left[ \bar w_s
-\gamma^f_s \mathcal C^{s,f}\lb \bar w_f - \bar w_s\rb
-\gamma^{fs}_s \mathcal C^{s,fs}\lb \bar w_{fs} - \bar w_s\rb
\right](h-\eta)}\\[3mm]
\hspace{.7cm}
\displaystyle{
- \frac{1}{\alpha_s}\left [C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+ C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
-C_{DV}^s\bar w_s |{\bf u}_s|\alpha_s
\right] \lb h - \eta\rb,}
\\[3mm]
\hspace{.7cm}
\displaystyle{
-\frac{1}{2}\Bigg \{\frac{D}{D t} \left[
h \nabla\cdot {\bf u}_s
- \gamma^f_s \mathcal C^{s,f}\lb h \nabla\cdot {\bf u}_f - h \nabla\cdot {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb h \nabla\cdot {\bf u}_{fs} - h \nabla\cdot {\bf u}_s\rb
\right ]\eta}\\[3mm]
\hspace{.7cm}
\displaystyle{
-\frac{D}{D t} \left[
\nabla\cdot {\bf u}_s
- \gamma^f_s \mathcal C^{s,f}\lb \nabla\cdot {\bf u}_f - \nabla\cdot {\bf u}_s\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb \nabla\cdot {\bf u}_{fs} - \nabla\cdot {\bf u}_s\rb
\right ]\eta^2}\\[3mm]
\hspace{.7cm}
+ \left[
\lb\nabla\cdot {\bf u}_s\rb^2
- \gamma^f_s \mathcal C^{s,f}\lb \lb\nabla\cdot {\bf u}_f\rb^2 - \lb\nabla\cdot {\bf u}_s\rb^2\rb
- \gamma^{fs}_s \mathcal C^{s,fs}\lb \lb \nabla\cdot {\bf u}_{fs}\rb^2 - \lb\nabla\cdot {\bf u}_s\rb^2\rb
\right ]\eta^2
\Bigg\}
\\[3mm]
\hspace{.7cm}
\displaystyle{
+\!\frac{1}{\alpha_s}\left[ C_{DG}^{s,f}|{\bf u}_f\! -\!{\bf u}_s|^{j-1}
\lb \nabla\cdot\lb {\bf u}_f \!-\!{\bf u}_s\rb\rb
\!+\! C_{DG}^{s,fs}|{\bf u}_{fs} \!-\!{\bf u}_s|^{j-1} \lb \nabla\cdot\lb {\bf u}_{fs} \!-\!{\bf u}_s\rb\rb
\!-\!C_{DV}^s|{\bf u}_s|\lb\nabla\cdot {\bf u}_s\rb\alpha_s
\right ]\frac{1}{2}\eta \lb h \!-\!\eta\rb.}
\label{tau_zz_aaaaaa_group}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
So, it is legitimate to call the first group of terms (with factor $h-\eta$) the enhanced gravity, and the second group of terms (with factors $\eta, \eta^2$ and $\eta (h-\eta)$) the dispersion. Together, they constitute the (effective) non-hydrostatic normal load. This has been discussed in more detail later in Section 2.1.3 and Section 2.1.4. In (\ref{tau_zz_aaaaaa_group}), the components in the drag terms have been splitted in to normal and slope parallel-type components contributing to the enhanced gravity and dispersion relations.
\\[3mm]
To apply the normal loads in a depth-averaged formulation, we need to depth-average $\tau_{zz_s}$ in (\ref{tau_zz_aaaaaa_group}). For this, first we define the phase-divergence in slope parallel directions as:
$\displaystyle{U_s = \nabla \cdot {\bf u}_s,\,
U_{fs} = \nabla \cdot {\bf u}_{fs},\,
U_f = \nabla \cdot {\bf u}_f},$ then following Yuan et al. (2018), we
integrate (\ref{tau_zz_aaaaaa_group}) through the flow depth to obtain its mean:
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\bar\tau_{zz_s} = \lb 1 - \gamma^f_s\rb g^z\frac{1}{2}h^2
+\frac{D}{D t} \left[ \bar w_s
-\gamma^f_s \mathcal C^{s,f}\lb \bar w_f - \bar w_s\rb
-\gamma^{fs}_s \mathcal C^{s,fs}\lb \bar w_{fs} - \bar w_s\rb
\right]\frac{1}{2}h^2
}
\\[3mm]
\hspace{.75cm}
\displaystyle{
- \frac{1}{\alpha_s}\frac{h^2}{2}\left [C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
-C_{DV}^s\bar w_s|{\bf u}_s|\alpha_s
\right ]}
\\[3mm]
\hspace{.75cm}
\displaystyle{+\frac{h^3}{12}\left[ \lb U_s^2
-\gamma^f_s \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
-\gamma^{fs}_s \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
- \frac{D}{Dt} \lb U_s
- \gamma^f_s\mathcal C^{s,f}\lb U_f - U_s \rb
- \gamma^{fs}_s\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]} \\[3mm]
\hspace{.75cm}
\displaystyle{
+\frac{1}{\alpha_s}\frac{h^3}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s| U_s\alpha_s
\right ],}
\label{tau_zz_aaaaaab}
\end{array}
\end{eqnarray}
\end{linenomath*}
\small}
\hspace{-2mm}
which is the depth averaged effective solid normal load.
\\[3mm]
{\bf B. The Fine-solid and Fluid Normal Loads:} As in (\ref{w_solid_momentum}), we consider the normal components of the fine-solid and fluid momentum equations (Pudasaini, 2012; Pudasaini and Mergili, 2019). Then, following the procedure from (\ref{tau_zz_a}) to (\ref{tau_zz_aaaaaab}), we obtain the depth-averaged normal stresses for fine-solid and fluid, respectively:
\clearpage
{\footnotesize
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\bar\tau_{zz_{fs}} = \gamma^f_{fs} g^z\frac{1}{2}h^2
+\frac{D}{D t} \left[ \bar w_{fs}
-\gamma^f_{fs} \mathcal C^{fs,f}\lb \bar w_f - \bar w_{fs}\rb
+\alpha^{s}_{fs} \mathcal C^{s,fs}\lb \bar w_{fs} - \bar w_s\rb
\right]\frac{1}{2}h^2}\\[3mm]
\hspace{.75cm}
\displaystyle{
- \frac{1}{\alpha_{fs}}\frac{h^2}{2}\left [-\frac{1}{\gamma^{fs}_s}C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
+C_{DG}^{fs,f}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1}
\lb \bar w_{f} - \bar w_{fs}\rb
-C_{DV}^{fs}\bar w_{fs}|{\bf u}_{fs}|\alpha_{fs}
\right ]}
\\[3mm]
\hspace{.75cm}
\displaystyle{+\frac{h^3}{12}\left[ \lb U_{fs}^2
-\gamma^f_{fs} \mathcal C^{fs,f} \lb U_f^2-U_{fs}^2\rb
+\alpha^{s}_{fs} \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
- \frac{D}{Dt} \lb U_{fs}
- \gamma^f_{fs}\mathcal C^{fs,f}\lb U_f - U_{fs} \rb
+ \alpha^{s}_{fs}\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]} \\[3mm]
\hspace{.75cm}
\displaystyle{
+\frac{1}{\alpha_{fs}}\frac{h^3}{6}\left [-\frac{1}{\gamma^{fs}_s}
C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
+C_{DG}^{fs,{f}}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1} \lb U_{f} - U_{fs}\rb
-C_{DV}^{fs}|{\bf u}_{fs}| U_{fs}\alpha_{fs}
\right ],}
\label{tau_zz_aaaaaab_fs}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
{\footnotesize
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\bar\tau_{zz_f} = g^z\frac{1}{2}h^2
+\frac{D}{D t} \left[ \bar w_f
+\alpha^s_f \mathcal C^{s,f}\lb \bar w_f - \bar w_s\rb
+\alpha^{fs}_f \mathcal C^{fs,f}\lb \bar w_{f} - \bar w_{fs}\rb
\right]\frac{1}{2}h^2}\\[3mm]
\hspace{.75cm}
\displaystyle{
+ \frac{1}{\alpha_{f}}\frac{h^2}{2}\left [\frac{1}{\gamma^f_s}C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+\frac{1}{\gamma^f_{fs}}C_{DG}^{fs,f}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1}
\lb \bar w_{f} - \bar w_{fs}\rb
+C_{DV}^{f}\bar w_{f}|{\bf u}_{f}|\alpha_{f}
\right ]}
\\[3mm]
\hspace{.75cm}
\displaystyle{+\frac{h^3}{12}\left[ \lb U_f^2
+\alpha^s_f \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
+\alpha^{fs}_f \mathcal C^{fs,f} \lb U_{f}^2-U_{fs}^2\rb
\rb
- \frac{D}{Dt} \lb U_f
+ \alpha^s_f\mathcal C^{s,f}\lb U_f - U_s \rb
+ \alpha^{fs}_f\mathcal C^{fs,f}\lb U_{f} - U_{fs} \rb
\rb \right]} \\[3mm]
\hspace{.75cm}
\displaystyle{
-\frac{1}{\alpha_{f}}\frac{h^3}{6}\left [\frac{1}{\gamma^f_s}
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+\frac{1}{\gamma^f_{fs}}C_{DG}^{fs,{f}}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1} \lb U_{f} - U_{fs}\rb
+C_{DV}^{f}|{\bf u}_{f}| U_{f}\alpha_{f}
\right ].}
\label{tau_zz_aaaaaab_f}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
The first terms on the right hand sides in (\ref{tau_zz_aaaaaab})-(\ref{tau_zz_aaaaaab_f}) show the distinct scalings for the solid, fine-solid, and fluid-phases in the three-phase mixture flow. The solid and fine-solid pressures are reduced due to respective buoyancies by the factors $\lb 1- \gamma_{s}^{f}\rb$ and $\gamma_{fs}^f$. The buoyancy reduced normal load of the solid particles, $\lb 1- \gamma_{s}^{f}\rb$, is due to the {fluid} composed of water and very fine particles and the fine-solids, and thus $\gamma_{s}^{f}$ is the corresponding mixture fluid density normalized by the solid density. Similar statement holds for fine-solid. For more detail on this, see Pudasaini and Mergili (2019).
\\[3mm]
The mean values of the normal components of stresses are required to obtain the lateral (slope parallel) stress components, which for solid, fine-solid and fluid phases are given by:
$\alpha_s \bar \tau_{{xx}_s} = \alpha_s K_s^x\bar \tau_{{zz}_s},
\alpha_{fs} \bar \tau_{{xx}_{fs}} = \alpha_{fs}\bar \tau_{{zz}_{fs}},
\alpha_{f} \bar \tau_{{xx}_{f}} = \alpha_{f}\bar \tau_{{zz}_{f}},$
where only the solid-phase contains the earth pressure coefficient $K_s^x$ due to its Coulomb frictional behavior (Pudasaini, 2012; Pudasaini and Mergili, 2019). These lateral stresses enter the momentum balance equations as the sum of the enhanced hydraulic pressure gradients and dispersion relations. This is discussed later.
\subsubsection{Enhanced (Effective) Gravities}
From (\ref{tau_zz_aaaaaa}), or (\ref{tau_zz_aaaaaab}), (and similarly from (\ref{tau_zz_aaaaaab_fs}) and (\ref{tau_zz_aaaaaab_f})), we extract the enhanced (effective) gravity for solid, fine-solid and fluid components, respectively
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
{\acute g}^z_s = \lb 1- \gamma^f_s\rb{g}^z + \frac{D}{Dt} \left[ \bar w_s
- \gamma^f_s \mathcal C^{s,f} \lb \bar w_f - \bar w_s \rb
- \gamma^{fs}_s \mathcal C^{s,fs} \lb \bar w_{fs} - \bar w_s \rb
\right]}\\[3mm]
\hspace{.55cm}
\displaystyle{
- \frac{1}{\alpha_{s}}\left [C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
-C_{DV}^s\bar w_s|{\bf u}_s|\alpha_s
\right ],}
\\[7mm]
\displaystyle{
{\acute g}^z_{fs} = \gamma^f_{fs}{g}^z + \frac{D}{Dt} \left[ \bar w_{fs}
- \gamma^f_{fs} \mathcal C^{fs, f} \lb \bar w_{f} - \bar w_{fs} \rb
+ \alpha^s_{fs} \mathcal C^{s, fs} \lb \bar w_{fs} - \bar w_{s} \rb
\right]}\\[3mm]
\hspace{.6cm}
\displaystyle{
- \frac{1}{\alpha_{fs}}\left [-\frac{1}{\gamma^{fs}_s}C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
+C_{DG}^{fs,f}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1}
\lb \bar w_{f} - \bar w_{fs}\rb
-C_{DV}^{fs}\bar w_{fs}|{\bf u}_{fs}|\alpha_{fs}
\right ]
,}\\[7mm]
\displaystyle{{\acute g}^z_f = {g}^z + \frac{D}{Dt} \left[ \bar w_f
+ \alpha^s_f \mathcal C^{s,f} \lb \bar w_f - \bar w_s \rb
+ \alpha^{fs}_f \mathcal C^{fs,f} \lb \bar w_f - \bar w_{fs} \rb
\right]}\\[3mm]
\hspace{.55cm}
\displaystyle{
+ \frac{1}{\alpha_{f}}\left [\frac{1}{\gamma^f_s}C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+\frac{1}{\gamma^f_{fs}}C_{DG}^{fs,f}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1}
\lb \bar w_{f} - \bar w_{fs}\rb
+C_{DV}^f\bar w_f|{\bf u}_f|\alpha_f
\right ],}
\label{effective_gravity}
\end{array}
\end{eqnarray}
\end{linenomath*}
where the factors $h^2/2$ do not appear due to the definition of acceleration.
These expressions can be obtained directly from (\ref{tau_zz_aaaaaa_group}) by setting $\eta \to 0$, i.e., the normal loads at the bed. This clearly indicates which terms in (\ref{tau_zz_aaaaaa_group}) contribute to the enhanced gravity or the effective normal load at the bed, and which other terms contribute to dispersive effects. For vanishing fine-solid and fluid components, these reduce to the simple enhanced gravity in Denlinger and Iverson (2004), Castro-Orgaz et al. (2015) and Yuan et al. (2018) for single-phase granular flow equations. Our new multi-phase formulations include buoyancy reduced solid and fine-solid normal loads as indicated by the factors $\lb 1- \gamma^f_s\rb$ and $\gamma^f_{fs}$, and the virtual mass forces as indicated by $\mathcal C$. The virtual mass forces alter the solid, fine-solid and fluid accelerations in the flow normal direction (in $D/Dt$) that ultimately enhance the effective gravity of the solid, fine-solid, and fluid phases. Furthermore, the drags between the phases $\lb C_{DG} \rb$ and the viscous drags $\lb C_{DV} \rb$ appear only in our enhanced gravity. Depending on the values of $\gamma, \mathcal C, C_{DG}, C_{DV}$ and the relative phase-velocities in the flow depth direction, enhancements or reductions of the usual gravity loads can be substantial to dominant as compared to the usual gravity loads, $g^z$.
\\[3mm]
These enhanced gravity terms include the accelerations of the solid, fine-solid and fluid components in the slope normal direction indicated by $D/Dt$. Furthermore, (\ref{effective_gravity}) also includes the drag contributions in the slope normal direction.
The only common quantity in (\ref{effective_gravity}), is the usual gravity load, $g^z$. However, the enhanced gravities differ with the surface normal accelerations of the solid, fine-solid and fluid phases. Depending on the flow dynamics, interfacial momentum exchanges, viscous drags, and the boundary conditions, one or two of them could be substantially larger than the others. One prominent example is a landslide impacting a reservoir or a water body (Pudasaini, 2014; Kafle et al., 2019; Mergili et al., 2018, 2020b; Pudasaini and Mergili, 2019). In this situation, both the enhanced gravity and the dispersion (see below) of the water wave would be fundamentally different (can also be large) as compared to the enhanced gravity and the dispersion of the submarine landslide. As we will see later (\ref{effective_gravity}) are components of the full non-hydrostatic model formulation.
\subsubsection{Dispersive Contributions}
The main dispersive contributions for the solid, fine-solid and fluid are denoted by $\mathcal D_s^x, \mathcal D_{fs}^x, \mathcal D_f^x$ which are extracted from (\ref{tau_zz_aaaaaab})-(\ref{tau_zz_aaaaaab_f}). We call them dispersive (for simplicity of terminology, also, see Castro-Orgaz et al., 2015) and take the form:
{\footnotesize
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\mathcal D_s^x = K_s^x\frac{h^2}{12}\left[ \lb U_s^2
-\gamma^f_s \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
-\gamma^{fs}_s \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
- \frac{D}{Dt} \lb U_s
- \gamma^f_s\mathcal C^{s,f}\lb U_f - U_s \rb
- \gamma^{fs}_s\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]} \\[3mm]
\hspace{.55cm}
\displaystyle{
+\, K_s^x\frac{1}{\alpha_{s}}\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s|U_s\,\alpha_s
\right ],}
\\[7mm]
{\mathcal D_{fs}^x = \frac{h^2}{12}\left[ \lb U_{fs}^2
-\gamma^f_{fs} \mathcal C^{fs,f} \lb U_f^2-U_{fs}^2\rb
+\alpha^{s}_{fs} \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
- \frac{D}{Dt} \lb U_{fs}
- \gamma^f_{fs}\mathcal C^{fs,f}\lb U_f - U_{fs} \rb
+ \alpha^{s}_{fs}\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]}, \\[3mm]
\hspace{.65cm}
+\frac{1}{\alpha_{fs}}\frac{h^2}{6}\left [-\frac{1}{\gamma^{fs}_s}
C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
+C_{DG}^{fs,{f}}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1} \lb U_{f} - U_{fs}\rb
-C_{DV}^{fs}|{\bf u}_{fs}|U_{fs}\,\alpha_{fs}
\right ],
\\[7mm]
\displaystyle{\mathcal D_f^x = \frac{h^2}{12}\left[ \lb U_f^2
+\alpha^s_f \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
+\alpha^{fs}_f \mathcal C^{fs,f} \lb U_f^2-U_{fs}^2\rb
\rb
- \frac{D}{Dt} \lb U_f + \alpha^s_f\mathcal C^{s,f}\lb U_f - U_s \rb
+ \alpha^{fs}_f\mathcal C^{fs,f}\lb U_f - U_{fs} \rb
\rb \right]}\\[3mm]
\hspace{.6cm}
-\frac{1}{\alpha_{f}}\frac{h^2}{6}\left [\frac{1}{\gamma^f_s}
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+\frac{1}{\gamma^f_{fs}}C_{DG}^{fs,{f}}|{\bf u}_{f} -{\bf u}_{fs}|^{j-1} \lb U_{f} - U_{fs}\rb
+C_{DV}^f|{\bf u}_f|U_f\,\alpha_f
\right ].
\label{dispersion}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
So, (\ref{effective_gravity}) and (\ref{dispersion}) imply that $\bar \tau_{zz_s} = \acute g^z_s h^2/2 + \mathcal D^x_s h/K_s^x$,
$\bar \tau_{zz_{fs}} = \acute g^z_{fs} h^2/2 + \mathcal D^x_{fs} h$,
$\bar \tau_{zz_{f}} = \acute g^z_{f} h^2/2 + \mathcal D^x_{f} h$. Hence, the effective basal normal load is the sum of the effective gravity and (effective) dispersion. Note that the factor $h$ is taken out from the dispersion expressions to properly adjust the fluxes, because
$\alpha_s \bar\tau_{xx_s} = \alpha_s K_s^x\acute g^z_s h^2/2 + \alpha_s \mathcal D_s^x h
= \alpha_s h \left [ \acute \beta^x_s h/2 + \mathcal D_s^x \right ]$, etc., where $\acute \beta^x_s = K_s^x\acute g^z_s$. Later, such structures will appear in the lateral fluxes in the momentum balance equations, where $\acute \beta^x_s h/2$ and $ D_s^x$ correspond, respectively, to the enhanced hydraulic pressure gradient and dispersion.
\\[3mm]
In what follows, all the terms with ${\acute *}$ are the enhanced terms, while these and all the $\mathcal D$ terms are entirely new contributions to the Pudasaini and Mergili (2019) model. These reduce to the non-hydrostatic relations for single-phase granular flow in Denlinger and Iverson (2004), Castro-Orgaz (2015), and Yuan et al. (2018). It is important to note that the enhanced gravities (\ref{effective_gravity}) and the dispersion relations (\ref{dispersion}) are derived from the $w$ components of the momentum balances from the multi-phase phase mass flow model (Pudasaini and Mergili, 2019). So, there are direct and strong couplings between the solid, fine-solid and the fluid components among these dispersion relations. As in the effective gravity, the dispersive terms are strongly coupled, e.g., due to the interfacial drag and virtual mass contributions.
\subsection{The Non-Hydrostatic Multi-Phase Mass-Flow Model}
In what follows, we further develop the three-phase mass flow model (Pudasaini and Mergili, 2019) by incorporating the enhanced gravities (\ref{effective_gravity}) and the dispersion relations (\ref{dispersion}). The depth-averaged mass balance equations for the solid, fine-solid and fluid phases are:
\begin{linenomath*}
\begin{subequations}\label{Model_Final_Mass}
\begin{align}
&\displaystyle{\frac{\partial }{\partial t}{\lb \alpha_s h\rb} + \frac{\partial}{\partial x}{\lb \alpha_s h u_s\rb}
+ \frac{\partial}{\partial y}{\lb \alpha_s h v_s\rb}=0},\label{Model_Final_Mass_s}\\[1mm]
&\displaystyle{\frac{\partial }{\partial t}{\lb \alpha_{fs} h\rb} + \frac{\partial}{\partial x}{\lb \alpha_{fs} h u_{fs}\rb}
+ \frac{\partial}{\partial y}{\lb \alpha_{fs} h v_{fs}\rb}=0},\label{Model_Final_Mass_fs}\\[1mm]
&\displaystyle{\frac{\partial }{\partial t}{\lb \alpha_f h\rb} + \frac{\partial}{\partial x}{\lb \alpha_f h u_f\rb}
+ \frac{\partial}{\partial y}{\lb \alpha_f h v_f\rb}=0.\label{Model_Final_Mass_f}}
\end{align}
\end{subequations}
\end{linenomath*}
The $x$-directional depth-averaged momentum conservation equations for the solid, fine-solid and fluid phases are,
{\small
\begin{linenomath*}
\begin{subequations}\label{Model_Final_Momentum}
\begin{align}
&\hspace{-1.5cm}\displaystyle{\frac{\partial }{\partial t}\biggl [ \alpha_s h \lb u_s - u_{s}^{vm} \rb \biggr ]
+\frac{\partial }{\partial x}\biggl [ \alpha_s h \lb u_s^2 - uu_{s}^{vm}+ {\acute \beta}_{s}^x \frac{h}{2} + \mathcal D_s^x\rb \biggr ]
+\frac{\partial }{\partial y}\biggl[ \alpha_s h \lb u_sv_s - uv_{s}^{vm}\rb \biggr ]}
\displaystyle{= h\mathcal S_{s}^x}, \label{Model_Final_Momentum_s}\\[5mm]
&\hspace{-1.5cm}\displaystyle{\frac{\partial }{\partial t}\biggl [ \alpha_{fs} h \lb u_{fs} - u_{fs}^{vm} \rb \biggr ]
+\frac{\partial }{\partial x}\biggl [ \alpha_{fs} h \lb u_{fs}^2 - uu_{fs}^{vm}+ {\acute\beta}_{fs}^x \frac{h}{2} + \mathcal D_{fs}^x \rb \biggr ]
+\frac{\partial }{\partial y}\biggl[ \alpha_{fs} h \lb u_{fs}v_{fs} - uv_{fs}^{vm}\rb \biggr ]}
\displaystyle{= h\mathcal S_{fs}^x},\label{Model_Final_Momentum_fs}\\[5mm]
&\hspace{-1.5cm}\displaystyle{\frac{\partial }{\partial t}\biggl [ \alpha_f h \lb u_f + u_{f}^{vm} \rb \biggr ]
+\frac{\partial }{\partial x}\biggl [ \alpha_f h \lb u_f^2 + uu_{f}^{vm}+ {\acute\beta}_{f}^x \frac{h}{2} + \mathcal D_f^x\rb \biggr ]
+\frac{\partial }{\partial y}\biggl[ \alpha_f h \lb u_fv_f + uv_{f}^{vm}\rb \biggr ]}
\displaystyle{= h\mathcal S_{f}^x}.\label{Model_Final_Momentum_f
\end{align}
\end{subequations}
\end{linenomath*}
}
\hspace{-3mm}
It is evident that the enhancements of the momentum fluxes depend on ${\acute\beta}$ and $\mathcal D$. Since the flow depth $h$ is a common factor in the momentum fluxes, the terms associated with $\mathcal D$ are proportional to $h^3$, and the term associated with ${\acute\beta}$ are proportional to $h^2$. This, together with the structure of $\mathcal D$ and ${\acute\beta}$ in the fluxes in (\ref{Model_Final_Momentum}), signify the highly non-linear, non-hydrostatic (dispersion) contributions.
\\[3mm]
Due to symmetry, the $y$-directional momentum equations for the solid, fine-solid and fluid phases can be written similarly {here and in all the following considerations}. {This is achieved by formally utilizing the replacements: $x \longleftrightarrow y$ and $u \longleftrightarrow v$, whenever necessary, both for variables and associated parameters.}
Below, we present models for all the fluxes, and source terms {and forces in the momentum equations for multi-phase mass flows where we follow the structures in Pudasaini and Mergili (2019).} First, we write those terms that include the non-hydrostatic terms (enhanced gravity and dispersion). The other terms are as in Pudasaini and Mergili (2019) and are put in an Appendix for completeness. It is important to note that in structure (\ref{Model_Final_Mass})-(\ref{Model_Final_Momentum}) are the same as in Pudasaini and Mergili (2019). It is advantageous, because the similar analysis and numerical methods and tools as in Pudasaini and Mergili (2019) might be applied to solve the new system of non-hydrostatic multi-phase mass flow model. However, complexity arises due to the new non-hydrostatic terms, particularly associated with the higher order time and spatial derivatives.
\\[1mm]
{\bf The $x$-directional source terms} in (\ref{Model_Final_Momentum}) are
\begin{linenomath*}
\begin{subequations}\label{Source}
\begin{align}
\mathcal S_{s}^x &= \alpha_s\left [g^x - \frac{u_s}{|{\bf u}_s|}\tan\delta_s {\acute g}^z_s -{\acute g}^z_s\frac{\partial b}{\partial x}\right ] - \alpha_s {{g''}}^z_s \left [ \frac{\partial h}{\partial x} + \frac{\partial b}{\partial x}\right ]\nonumber\\[2mm]
&+ C_{DG}^{s,f} \lb u_f - u_s \rb{ |{\bf u}_f - {\bf u}_s|}^{\jmath-1}
+ C_{DG}^{s,fs} \lb u_{fs} - u_s \rb{ |{\bf u}_{fs} - {\bf u}_s|}^{\jmath-1}
-C_{DV}^s u_s |{\bf u}_s|\alpha_s,\label{Source_x_s}\\[2mm]
\mathcal S_{fs}^x &= \alpha_{fs}\biggl [g^x - \biggl [ - \frac{1}{2} {\acute g}^z_{fs} \frac{h}{\alpha_{fs}}\frac{\partial \alpha_{fs}}{\partial x} + {\acute g}^z_{fs} \frac{\partial b}{\partial x}\nonumber \\[2mm]
& -\left \{ 2\frac{\partial}{\partial x}\lb \nu_{fs}^e\frac{\partial u_{fs}}{\partial x}\rb
+ \frac{\partial}{\partial y}\lb \nu_{fs}^e\frac{\partial v_{fs}}{\partial x}\rb
+ \frac{\partial}{\partial y}\lb \nu_{fs}^e\frac{\partial u_{fs}}{\partial y}\rb
- \nu_{fs}^e\left [ \frac{\partial u_{fs}}{\partial z}\right ]_b\frac{1}{h} \right \} + \tau_{nN}^{fs^x} \biggr]\biggl ]\nonumber \\[2mm]
&-\displaystyle\frac{1}{\gamma_s^{fs}}{C_{DG}^{s,fs}\lb u_{fs} - u_s \rb{ |{\bf u}_{fs} - {\bf u}_s|}^{\jmath-1}}
+\displaystyle{C_{DG}^{fs,f}\lb u_f - u_{fs} \rb{ |{\bf u}_f - {\bf u}_{fs}|}^{\jmath-1}}
-C_{DV}^{fs} u_{fs} |{\bf u}_{fs}|\alpha_{fs},\label{Source_x_fs}
\\[2mm]
\mathcal S_{f}^x &= \alpha_f\biggl [g^x - \biggl [ - \frac{1}{2} {\acute g}^z_{f}\frac{h}{\alpha_f}\frac{\partial \alpha_f}{\partial x} + {\acute g}^z_{f}\frac{\partial b}{\partial x}\nonumber \\[2mm]
& -\left \{ 2\frac{\partial}{\partial x}\lb \nu_f^e\frac{\partial u_f}{\partial x}\rb
+ \frac{\partial}{\partial y}\lb \nu_f^e\frac{\partial v_f}{\partial x}\rb
+ \frac{\partial}{\partial y}\lb \nu_f^e\frac{\partial u_f}{\partial y}\rb
- \nu_{f}^e\left [ \frac{\partial u_f}{\partial z}\right ]_b\frac{1}{h} \right \} + \tau_{nN}^{f^x} \biggr]\biggl ]\nonumber \\[2mm]
&-\displaystyle{\frac{1}{\gamma_s^f}C_{DG}^{s,f}\lb u_f - u_s \rb{ |{\bf u}_f - {\bf u}_s|}^{\jmath-1}}
-\displaystyle{\frac{1}{\gamma_{fs}^f}C_{DG}^{fs,f}\lb u_f - u_{fs} \rb{ |{\bf u}_f - {\bf u}_{fs}|}^{\jmath-1}}
-C_{DV}^f u_f |{\bf u}_f|\alpha_f,\label{Source_x_f}
\end{align}
\end{subequations}
\end{linenomath*}
where ${{g''}}^z_s$ is obtained from ${\acute g}^z_s$ by replacing $\lb 1- \gamma^f_s\rb g^z$ by $\gamma^f_s g^z$ while the other terms remain unchanged. The expressions in (\ref{Source}) are more general than those in Pudasaini and Mergili (2019) as they include the non-hydrostatic effects together with the interfacial momentum transfers. The structure of ${\acute g}$ indicates that the enhancements of the forces associated with ${\acute g}$, including friction, buoyancy and basal and topographic pressure gradients, depend on the sign and magnitude of ${\acute g}$.
\\[3mm]
Due to the acceleration in the slope normal direction, in (\ref{Source}), the solid velocity is given by ${\bf u}_s = \lb u_s, v_s, w_s\rb$, where $w_s = u_s\partial b/\partial x + v_s\partial b/\partial y$ (Yuan et al., 2018). This indicates that for locally changing basal topography, the surface normal component of velocity is important. Similar expressions hold for the fine-solid and fluid components.
\\[3mm]
In (\ref{Model_Final_Momentum}) and (\ref{Source}),
$u^{vm}, uu^{vm}, uv^{vm}$ are the virtual mass induced mass and momentum enhancements,
${\acute \beta}$ are the hydraulic pressure coefficients,
$\nu^e$ are the effective kinematic viscosities, $\partial u/\partial z|_b$ are the $xz$- basal shear stresses, $\tau_{nN}$ are the enhanced non-Newtonian viscous stresses, and
$C_{DG}$ are the drag coefficients.
The momentum balances (\ref{Model_Final_Momentum}) and the sources (\ref{Source}) indicate that the effective gravity enhances the ``hydraulic pressure gradients'' (via the terms associated with ${\acute \beta}$) in the momentum flux, and the enhanced material loads at the base as indicated by the terms associated with ${\acute g}$ and $g''$, indicating their extensive effects in the source terms. In total, the lateral flux for solid is enhanced by
$\alpha_s \left [ {\acute \beta}_s^x - \beta_s^x\right ]h^2/2 +\alpha_s h {\mathcal D}_s^x$, where, $\beta_s^x = K_s^x\lb 1- \gamma_s^f \rb g^z$.
Similar flux enhancements emerge for the fine-solid and fluid phases.
\\[3mm]
{\bf The $x$-directional hydraulic pressure coefficients} for solid, fine-solid and fluid in (\ref{Model_Final_Momentum}) are:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}\label{hydraulic_pressure_x}
{\acute\beta}_{s}^x = K_{s}^x {\acute g}^z_s,\,\,\, {\acute\beta}_{fs}^x = {\acute g}^z_{fs},\,\,\,{\acute\beta}_{f}^x = {\acute g}^z_f,
\label{pressure_parameter_scaling}
\end{array}
\end{eqnarray}
\end{linenomath*}
where {$K_s^x$ is the earth pressure coefficient (Pudasaini and Hutter, 2007)} and ${\acute g}$ are given by (\ref{effective_gravity}). Above, we only wrote those terms that are new in the non-hydrostatic formulations, that are ${\acute \beta}, {\acute g}, g''$ and $\mathcal D$. Based on Pudasaini and Mergili (2019), all other terms appearing in the above model equations are explained in the Appendix.
\\[3mm]
{\bf A Closed System of Equations:} The model (\ref{Model_Final_Mass})-(\ref{Model_Final_Momentum}) constitutes a set of nine equations for mass and momentum balances (including the $y$-components) for three-phase mixture mass flows in nine unknowns, namely, the solid, fine-solid and fluid phase velocities in the down-slope $\lb u_s, u_{fs}, u_f\rb$, and cross slope $\lb v_s, v_{fs}, v_f\rb$ directions, and the respective phase depths $\lb h_s=\alpha_sh, h_{fs}=\alpha_{fs}h, h_f=\alpha_fh \rb$.
{Note that $h_s + h_{fs} + h_f = h$, the total material depth.}
The model is written in a well structured form of partial differential equations and may be solved numerically once appropriate initial and boundary conditions are prescribed (Pudasaini and Mergili, 2019).
\\[3mm]
{\bf Reduction to Existing Models:}
By setting the fine-solid and fluid fractions to zero ($\alpha_{fs} \to 0, \alpha_{f} \to 0$), the new non-hydrostatic multi-phase mass flow model reduces to the single-phase non-hydrostatic granular flow models by Castro-Orgaz et al. (2015) and Yuan et al. (2018). The major parts of $\acute g, g'', \acute \beta$ terms, and entirely the $\mathcal D$ terms in (\ref{Model_Final_Momentum})-(\ref{Source}) are new to Pudasaini and Mergili (2019) which are due to non-hydrostatic contributions. Furthermore, the Pudasaini and Mergili (2019) multi-phase mass flow model is obtained by neglecting all the non-hydrostatic contributions, i.e., by only considering
$\acute g_s^z: = \lb 1 - \gamma^f_s\rb g^z$, $\acute g_{fs}^z: = \gamma^f_{fs}g^z$, $\acute g_{f}^z: = g^z$; $\mathcal D_s^x = 0, \mathcal D_{fs}^x = 0, \mathcal D_{f}^x = 0$.
\section{Possible Simplifications}
As we saw in Section 2.2, relations (\ref{effective_gravity}) and (\ref{dispersion}) introduce higher order spatial and time derivatives in the momentum fluxes. The new enhanced gravity and dispersion may lead to a complexity in numerical integration of the model equations, and thus may require a fundamentally new and complicated numerical method to properly solve the model equations.
That was the case even for the simple single-phase granular flow models (Castro-Orgaz et al., 2015; Yuan et al., 2018).
\subsection{Reduced Normal Load - Ignoring the Time Derivatives in Dispersion}
One way to avoid computational difficulties, but still include the new effects, is to assume a negligible local time derivatives ($\partial/\partial t$) in (\ref{tau_zz_aaaaaab}). This can be a reasonable assumption, e.g., after the initial impact of the landslide at the water body and during continues impact. Another possibility is to ignore all the $D/Dt$ terms in (\ref{tau_zz_aaaaaab}). Yet, the reduced solid normal stress includes non-hydrostatic effects due to buoyancy, virtual mass, drags and slope parallel divergence and relative divergence,
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\bar\tau_{zz_{sR}} = \lb 1 - \gamma^f_s\rb g^z\frac{1}{2}h^2
- \frac{1}{2}\frac{1}{\alpha_{s}}h^2\left [C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
-C_{DV}^s\bar w_s |{\bf u}_s| \alpha_s
\right ]
\\[3mm]
\hspace{.9cm}
\displaystyle{+
\frac{h^3}{12}\left[ \lb U_s^2
-\gamma^f_s \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
-\gamma^{fs}_s \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
\right]} \\[3mm]
\hspace{.9cm}
\displaystyle{
+\frac{1}{\alpha_{s}}\frac{h^3}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s |{\bf u}_s| U_s\,\alpha_s
\right ],}
\label{tau_zz_aaaaaab_reduced}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
where $R$ in $\bar\tau_{zz_{sR}}$ stands for the reduced normal stress. And thus, the corresponding reduced enhanced gravity and reduced dispersion expressions are given, respectively, by
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
{\acute g}^z_{sR} = \lb 1 - \gamma^f_s\rb g^z
- \frac{1}{\alpha_{s}}\left [C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1}
\lb \bar w_f - \bar w_s\rb
+C_{DG}^{s,fs}|{\bf u}_{fs} -{\bf u}_s|^{j-1}
\lb \bar w_{fs} - \bar w_s\rb
-C_{DV}^s\bar w_s |{\bf u}_s| \alpha_s
\right ],
\\[5mm]
\displaystyle{
\mathcal D_{sR}^s = K_s^x\frac{h^2}{12}\left[ \lb U_s^2
-\gamma^f_s \mathcal C^{s,f} \lb U_f^2-U_s^2\rb
-\gamma^{fs}_s \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
\right]} \\[3mm]
\hspace{.75cm}
{\displaystyle
+\frac{1}{\alpha_{s}}K_s^x\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s |{\bf u}_s| U_s\,\alpha_s
\right ].}
\label{tau_zz_aaaaaab_reduce_g_d}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
From (\ref{effective_gravity}) and (\ref{dispersion}), similar reduced expressions can be obtained for the fine-solid and fluid components. For single-phase granular flow without the fine-solid and fluid components, (\ref{tau_zz_aaaaaab_reduce_g_d}) would further drastically reduce to
${\acute g}^z_{sR} = g^z + C_{DV}^s\bar w_s |{\bf u}_s|$ and
$\mathcal D_{sR}^x = K_s^xh^2 U_s^2/12 - \frac{1}{6}K_s^xh^2 C_{DV}^s |{\bf u}_s| U_s$. However, in general, as in (\ref{effective_gravity}) and (\ref{dispersion}), the full descriptions of ${\acute g}^z_{s}$ and $\mathcal D_{s}^x$ (similar for fine-solid and fluid components) should be considered in simulating non-hydrostatic mixture flows.
\subsection{Approximations to Time Derivatives in Dispersion and Enhanced Gravity}
One of the major difficulties associated with the non-hydrostatic model presented above is the presence of the time derivatives in enhanced gravity and dispersion. In the simple situation without interfacial drag and virtual mass, the dispersion in (\ref{Model_Final_Momentum_s}) is given by
{\footnotesize
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial}{\partial x}\left [\frac{K_s^x}{12}\alpha_sh^3
\left\{ U_s^2 - \frac{D U_s}{Dt}
-2 C_{DV}^s|{\bf u}_s| U_s
\right\}
\right] =
\frac{\partial}{\partial x}\left [\frac{K_s^x}{12}\alpha_sh^3
\left\{ U_s^2 - \lb \frac{\partial}{\partial t} + u_s\frac{\partial}{\partial x}
+ v_s\frac{\partial}{\partial y}
\rb U_s
-2 C_{DV}^s|{\bf u}_s| U_s
\right\}
\right]}\\[5mm]
= \displaystyle{
\frac{\partial}{\partial x}\left [\frac{K_s^x}{12}\alpha_sh^3
\left\{ U_s^2 -
\lb \frac{\partial}{\partial x}\frac{\partial u_s}{\partial t}
+ \frac{\partial}{\partial y}\frac{\partial v_s}{\partial t}
\rb
-\lb u_s\frac{\partial}{\partial x} + v_s\frac{\partial}{\partial y}\rb U_s
-2 C_{DV}^s|{\bf u}_s| U_s
\right\}
\right]}.
\label{Dispersion_T}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
From a computational point of view $\partial u_s/\partial t$ and $\partial v_s/\partial t$ in (\ref{Dispersion_T}) may pose great difficulties. So, it is desirable to find expressions for $\partial u_s/\partial t$ and $\partial v_s/\partial t$ in terms of spatial derivatives, flow variables, and parameters, but no direct involvement of (the time and) time derivatives. This is a challenging task. However, we can develop simplified expressions for these for non-inertial flows. This can be achieved, e.g., by combining the simple mass and momentum balance equation for solid from (\ref{Model_Final_Mass_s}) and (\ref{Model_Final_Momentum_s}), by ignoring all extra forces (which, however, could be considered to include more complex situations). Which is equivalent to assume that all the applied forces balance each other. This results in a simple expression as:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial u_s}{\partial t} = - u_s\frac{\partial u_s}{\partial x} - v_s\frac{\partial u_s}{\partial y}.
}
\label{Dispersion_TT}
\end{array}
\end{eqnarray}
\end{linenomath*}
Inserting (\ref{Dispersion_TT}) in to (\ref{Dispersion_T}), we technically remove $\partial u_s/\partial t$, which, however, is highly non-linear and very complex as it involves the fifth order terms (combining flow depth and velocities) and third order derivatives. Simplified expressions for the fine-solid and fluid components can be developed, and respectively take the form:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial u_{fs}}{\partial t} =
- u_{fs}\frac{\partial u_{fs}}{\partial x}
- v_{fs}\frac{\partial u_{fs}}{\partial y},
}
\label{Dispersion_TTT_fs}
\end{array}
\end{eqnarray}
\end{linenomath*}
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial u_{f}}{\partial t} =
- u_{f}\frac{\partial u_{f}}{\partial x}
- v_{f}\frac{\partial u_{f}}{\partial y}.
}
\label{Dispersion_TTT_f}
\end{array}
\end{eqnarray}
\end{linenomath*}
Similar expressions hold for $\partial v_s/\partial t, \partial v_{fs}/\partial t$ and $\partial v_f/\partial t$.
Then, the dispersion term containing the time derivatives, together with $U^2$ and the viscous drag in (\ref{Dispersion_T}), reduces, for solid-phase, to:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
U^2_s - \frac{DU_s}{Dt}
-2 C_{DV}^s|{\bf u}_s| U_s
= 2U_s^2
-2\frac{\partial u_s}{\partial x}\frac{\partial v_s}{\partial y}
+2\frac{\partial v_s}{\partial x}\frac{\partial u_s}{\partial y}
-2 C_{DV}^s|{\bf u}_s| U_s }.
\label{Dispersion_TTT_ss}
\end{array}
\end{eqnarray}
\end{linenomath*}
Expressions for $U_{fs}^2-{DU_{fs}}/{Dt}$ and $U_{f}^2-{DU_{f}}/{Dt}$ take analogous forms.
\\[3mm]
Similarly, with somewhat lengthy calculations, we can write the time derivative term, $D {\bar w}_s/Dt$, in the enhanced gravity (see, Section 2.1.3) as
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{D {\bar w}_s}{Dt}=}
\displaystyle{ -\frac{1}{2}\left [
-h\left\{\lb \frac{\partial u_s}{\partial x}\rb^2 + 2\frac{\partial v_s}{\partial x} \frac{\partial u_s}{\partial y} + \lb \frac{\partial v_s}{\partial y}\rb^2 \right\}
+\frac{\partial h}{\partial t}\lb \frac{\partial u_s}{\partial x}+\frac{\partial v_s}{\partial y}\rb
+
\lb
u_s\frac{\partial h}{\partial x}+v_s\frac{\partial h}{\partial y}\rb\lb\frac{\partial u_s}{\partial x}+\frac{\partial v_s}{\partial y}
\rb
\right ],
}
\label{Dispersion_TTT_ssn}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
where the topographic slope changes $(\partial b/\partial x, \partial b/\partial y)$ has been ignored, which could easily be included. Similar expressions as (\ref{Dispersion_TTT_ssn}) hold for fine-solid and fluid components, ${D {\bar w}_{fs}}/{Dt}$, ${D {\bar w}_{f}}/{Dt}$.
\\[3mm]
Due to the definition of ${\bar w}_s$, the time derivative of the flow depth, $\partial h/\partial t$, still remains in (\ref{Dispersion_TTT_ssn}). However, this can be obtained by summing-up the mass balance equations (\ref{Model_Final_Mass}) for the solid, fine-solid and fluid phases:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
\frac{\partial h}{\partial t} =
- \frac{\partial}{\partial x} \left [ h\lb \alpha_su_s + \alpha_{fs}u_{fs} + \alpha_{f}u_{f} \rb\right ]
- \frac{\partial}{\partial y} \left [ h\lb \alpha_sv_s + \alpha_{fs}v_{fs} + \alpha_{f}v_{f} \rb\right ] },
\label{EnhancedGravity_h}
\end{array}
\end{eqnarray}
\end{linenomath*}
where the hold up identity $\alpha_s + \alpha_{fs}+ \alpha_{f} = 1$ has been employed.
This way we can avoid the time derivatives in the terms associated with dispersion and enhanced gravity.
\section{Analysis of the Simplified Dispersion Relation}
Consider the dispersion for solid from (\ref{dispersion}):
{\footnotesize
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\mathcal D_s^x = K_s^x\frac{h^2}{12}\left[ \lb U_s^2
-\gamma^f_s \mathcal C^{s,f} \left( U_f^2-U_s^2\right)
-\gamma^{fs}_s \mathcal C^{s,fs} \lb U_{fs}^2-U_s^2\rb
\rb
- \frac{D}{Dt} \lb U_s
- \gamma^f_s\mathcal C^{s,f}\lb U_f - U_s \rb
- \gamma^{fs}_s\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]} \\[3mm]
\displaystyle{
\hspace{.65cm}+K_s^x\frac{1}{\alpha_{s}}\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s| U_s\,\alpha_s
\right ].}
\label{Dispersion_A}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
\hspace{-3mm}
The flux in the momentum balance shows that in total the dispersion relation contains third order terms in flow depth, and third order derivatives of the flow velocities. These are the highest order terms therein. So, it is important to analyze the terms appearing in the dispersion relation, and additionally seek its simplifications and consequences.
\subsection{The Role of Drag}
For the slowly varying slope parallel divergence, $U_s^2, U_{fs}^2, U_f^2$ can be neglected as compared to the other terms. Then, (\ref{Dispersion_A}) reduces to
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\mathcal D_s^x
= -K_s^x\frac{h^2}{12}\left[
\frac{D}{Dt} \lb U_s
- \gamma^f_s\mathcal C^{s,f}\lb U_f - U_s \rb
- \gamma^{fs}_s\mathcal C^{s,fs}\lb U_{fs} - U_s \rb
\rb \right]} \\[3mm]
\displaystyle{
\hspace{.9cm}+K_s^x\frac{1}{\alpha_{s}}\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s| U_s\,\alpha_s
\right ].}
\label{Dispersion_B}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
For negligible virtual mass force, (\ref{Dispersion_B}) simplifies to
{\small
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\mathcal D_s^x = -K_s^x\frac{h^2}{12}
\frac{DU_s}{Dt}}
\displaystyle{+K_s^x\frac{1}{\alpha_{s}}\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s| U_s\,\alpha_s
\right ].}
\label{Dispersion_BC}
\end{array}
\end{eqnarray}
\end{linenomath*}
}
Furthermore, for non-accelerating flows, the terms with $D/Dt$ vanish, and (\ref{Dispersion_BC}) further reduces to
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\mathcal D_s^x =
K_s^x\frac{1}{\alpha_{s}}\frac{h^2}{6}\left [
C_{DG}^{s,f}|{\bf u}_f -{\bf u}_s|^{j-1} \lb U_f - U_s\rb
+C_{DG}^{s,{fs}}|{\bf u}_{fs} -{\bf u}_s|^{j-1} \lb U_{fs} - U_s\rb
-C_{DV}^s|{\bf u}_s| U_s\,\alpha_s
\right ].}
\label{Dispersion_C}
\end{array}
\end{eqnarray}
\end{linenomath*}
So, the interfacial and viscous drag may play an important role in generating dispersion relation in mixture mass flows which was not the case in the single-phase mass flows (Castro-Orgaz et al., 2015; Yuan et al., 2018).
\subsection{Negligible Dispersion}
In the most simple case the interfacial drags and the virtual mass may be neglected. A situation can arise such that the dispersion effect could be ignored. Then, from (\ref{Model_Final_Momentum_s}) and (\ref{Dispersion_A}), by integrating
$\partial \left[\alpha_s h{\mathcal D_s^x}\right]/\partial x = 0$
with respect to $x$, we obtain:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{ \frac{K_s^x}{12}\alpha_sh^3\lb U_s^2 - \frac{D U_s}{Dt}
-2 C_{DV}^s|{\bf u}_s| U_s
\rb} = {\mathcal P_{f_0}},
\label{Dispersion_D}
\end{array}
\end{eqnarray}
\end{linenomath*}
where ${\mathcal P_{f_0}}$ is a constant of integration. However, determination of ${\mathcal P_{f_0}}$ may involve complex physical processes (explained in Section 4.3 - Section 4.5).
For simplicity, we assume a channelized flow, so the variation of the flow dynamic quantities with $y$ is negligible. For notational convenience we write $u = u_s$ and $\beta = C_{DV}^s$. Then, for $u_s > 0$, (\ref{Dispersion_D}) reduces to
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\frac{\partial^2u}{\partial x\partial t}
+\frac{\partial^2 \lb u^2/2\rb}{\partial x^2}
- 2\lb \frac{\partial u}{\partial x}\rb^2
+\beta\frac{\partial}{\partial x}\lb u^2\rb
= -{\mathcal P_{f}},}
\label{Dispersion_E}
\end{array}
\end{eqnarray}
\end{linenomath*}
where
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{{\mathcal P_f} = \frac{12{\mathcal P_{f_0}}}{K_s^x\alpha_sh^3}}.
\label{Lambda_1}
\end{array}
\end{eqnarray}
\end{linenomath*}
We call ${\mathcal P_f}$ the (dissipative) prime-force coefficient (or, simply the ${\mathcal P}$-force coefficient).
Equation (\ref{Dispersion_E}) can be solved analytically only with some further assumptions. And, the solutions are presented in Section 4.3. If the solid particle distribution is uniform and the flow height can be approximated (by a constant), e.g., for a smooth flow, then, ${\mathcal P_{f}}$ is a constant.
Equation (\ref{Dispersion_E}) can further be simplified as follows.
\\[3mm]
{\bf I. Negligible $\lb \partial u/\partial x\rb^2$:} First, assume that $\partial u/\partial x$ is small and thus $\lb \partial u/\partial x\rb^2$ can be neglected. Then, integrating (\ref{Dispersion_E}) with respect to $x$, we obtain:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\frac{\partial u}{\partial t}
+\frac{\partial \lb u^2/2\rb}{\partial x}
= -\beta u^2-{\mathcal P_{f}} x + \alpha,}
\label{Dispersion_F}
\end{array}
\end{eqnarray}
\end{linenomath*}
where $\alpha$ is a constant of integration, and we call $-{\mathcal P_{f}} x$ the prime-force (or, simply the ${\mathcal P}$-force), per unit mass. With this, we draw an important conclusion, that for spatially slowly varying velocity field, non-dispersive flows degenerate into an advective-dissipative system with a complex source term. Here, dissipation refers to the viscous dissipation due to the drag contribution $-\beta u^2$, and also $-{\mathcal P_{f}} x$, that we will elaborate later.
When ${\mathcal P_{f_0}} \to 0$, or $h$ is large (enough) then ${\mathcal P_{f}} \to 0$. Alternatively, consider sufficiently small $x$. In both situations, ${\mathcal P_{f}} x$ is negligible, and (\ref{Dispersion_F}) becomes an inviscid, dissipative Burgers' equation developed by Pudasaini and Krautblatter (2022):
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\frac{\partial u}{\partial t}
+\frac{\partial \lb u^2/2\rb}{\partial x}
= \alpha-\beta u^2}.
\label{Dispersion_F_1}
\end{array}
\end{eqnarray}
\end{linenomath*}
From a simple physical consideration, following Pudasaini and Krautblatter (2022), $\alpha$ can represent the net driving force for the landslide motion, defined later at Section 5.1. So, (\ref{Dispersion_F}) can be viewed as the formal extension of the Pudasaini and Krautblatter (2022) landslide velocity equation, who also
constructed numerous exact analytical solutions for (\ref{Dispersion_F_1}), including simple to very sophisticated ones.
\\[3mm]
{\bf The Super Inviscid Dissipative Burgers' Equation:}
There are two fascinating aspects of (\ref{Dispersion_F_1}). First, by setting the dispersion structure (which is internal to the new model developed here) to zero, we obtained the reduced equation of landslide motion without dispersion in Pudasaini and Krautblatter (2022). Second, the emergence of (\ref{Dispersion_F_1}) explicitly proves the consistency of our new model with dispersion.
However, when ${\mathcal P_{f}} x \neq 0$, (\ref{Dispersion_F}) is the extension of the inviscid, dissipative Burgers' equation in Pudasaini and Krautblatter (2022), for which, no exact analytical solutions have so far been developed. Yet, the model (\ref{Dispersion_E}) is more complex and general than (\ref{Dispersion_F}). For this reason,
we call (\ref{Dispersion_F}) the extension, and (\ref{Dispersion_E}) the super generalization of the inviscid, dissipative Burgers' equation.
\\[3mm]
{\bf II. Time Independent Flows:} Second, assume a time-independent (steady state) flow. Then, from (\ref{Dispersion_E}) we have
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\frac{\partial^2 }{\partial x^2}\lb u^2\rb
- 4\lb \frac{\partial u}{\partial x}\rb^2
+2\beta\frac{\partial}{\partial x}\lb u^2\rb
= -2{\mathcal P_{f}}.}
\label{Dispersion_F_2}
\end{array}
\end{eqnarray}
\end{linenomath*}
Since $\alpha_s, K_s^x$ and $h$ are positive, the nature of solution depends on the sign of ${\mathcal P_{f_0}}$ and its magnitude in ${\mathcal P_{f}}$ as given in (\ref{Lambda_1}).
\subsection{Analytical Solutions}
Physically meaningful exact solutions explain the true and entire nature of the problem associated with the model equation (Pudasaini, 2011; Faug, 2015). The exact analytical solutions to simplified cases of non-linear debris avalanche model equations provide important insights into the full flow behavior of the complex system (Pudasaini and Krautblatter, 2022), and are often needed to calibrate and validate the numerical solutions (Pudasaini, 2016) as a prerequisite before running numerical simulations based on complex numerical schemes. So, such solutions should be developed, analyzed and properly understood prior to numerical simulations. This is very useful to interpret complicated simulations and/or avoid mistakes associated
with numerical simulations.
Here, we construct some exact analytical solutions to (\ref{Dispersion_F_2}) for yet different simplified cases.
\\[3mm]
{\bf I. ${\mathcal P_{f}} = 0$, Vanishing Prime-force:} With this, the exact solution for (\ref{Dispersion_F_2}) takes the form:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
u(x) = C_2 \exp\left [{\frac{C_1 }{\beta}\exp(2 \beta x)}\right].
}
\label{Exact_1}
\end{array}
\end{eqnarray}
\end{linenomath*}
There are two integration parameters $C_1, C_2$ to be determined, e.g., with the value and the slope of $u$ at a given point.
\\[3mm]
{\bf II. $\beta = 0$, Vanishing Drag:} For this, the exact solution for (\ref{Dispersion_F_2}) becomes more complex:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
u(x) = \frac{\sqrt{-{\mathcal P_{f}}}\exp\lb -C_1\rb\tanh\left[ \exp\lb C_1\rb\lb C_2 +x\rb\right]}
{\sqrt{\tanh^2\left [ \exp\lb C_1\rb\lb C_2 +x\rb \right] -1}},
}
\label{Exact_2}
\end{array}
\end{eqnarray}
\end{linenomath*}
where the two integration parameters $C_1, C_2$ are to be determined. The solutions (\ref{Exact_1}) and (\ref{Exact_2}) with some parameter values are presented in Fig. \ref{Fig_1} showing the exponential increase in the velocity field as a function of the travel distance. Where, for comparison the solution (\ref{Exact_1}) has been shifted down by about 2.
However, more realistic solution is presented below when both ${\mathcal P_{f}}$ and $\beta$ cannot be ignored.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=13cm]{Analytical_Soln_NoDispersion_1Lambda0-C_DV0_N.eps}
\end{center}
\caption[]{Velocity fields represented by the solutions (\ref{Exact_1}) with parameters $C_1 = 0.05, C_2 = 1.0, \beta = 0.075$; and (\ref{Exact_2}) with parameters $C_1 = 0.01, C_2 = 0.005, {\mathcal P_{f}} = 2.5\times 10^{-6}$.}
\label{Fig_1}
\end{figure}
\\[3mm]
{\bf III. Small $\partial u/\partial x$:} Then, $\lb\partial u/\partial x\rb^2$ can be neglected in (\ref{Dispersion_F_2}) which, after integration, reduces to
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
u\frac{\partial u}{\partial x} = \alpha - \beta u^2 -{\mathcal P_{f}} x,}
\label{Dispersion_F_3}
\end{array}
\end{eqnarray}
\end{linenomath*}
where $\alpha$ is a constant (the net driving force, see, Section 5.1), and
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
T^s_f = \alpha - \beta u^2 -{\mathcal P_{f}} x,
\label{Dispersion_F_3t}
\end{array}
\end{eqnarray}
\end{linenomath*}
constitutes the total system force. The model (\ref{Dispersion_F_3}) includes both the parameters $\beta$ and ${\mathcal P_{f}}$ and extends the Pudasaini and Krautblatter (2022) landslide velocity equation for the time-independent motion for which their model corresponds to ${\mathcal P_{f}} = 0$.
With the initial condition $u(0) = 0$, the exact analytical solution for (\ref{Dispersion_F_3}) yields:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{
u(x) = \sqrt{\frac{\alpha}{\beta}}\sqrt{\left[ 1 - \exp\lb -2 \beta x\rb\right] - {\mathcal P_{u}} \left [ \lb 2 \beta x-1\rb + \exp\lb -2 \beta x\rb\right ]}
},
\label{Exact_3}
\end{array}
\end{eqnarray}
\end{linenomath*}
where ${\displaystyle{{\mathcal P_{u}} = \frac{1}{2}\frac{1}{\alpha\beta}{\mathcal P_{f}}}}$. We call ${\mathcal P_{u}}$ the unified prime-force coefficient, which is a dimensionless number (quantity). It is induced by the prime-force coefficient ${\mathcal P_{f}}$, and also includes other force components, the net driving force $\alpha$, and the viscous resistance, represented by $\beta$.
\subsection{Postulation of the Prime-force: $-{\mathcal P_{f}} x$}
The prime-force coefficient $-{\mathcal P_{f}}$ in (\ref{Dispersion_E}), and $x$ in the prime-force ${-\mathcal P_{f}} x$ in (\ref{Dispersion_F}) appear systematically. It emerged from our new modelling approach, with physical-mathematical foundation, from integrating the rate of acceleration, and the acceleration itself. This is exactly the reason why ${-\mathcal P_{f}} x$ is a dissipative (or anti-dissipative) force, and ${-\mathcal P_{f}}$ is the spatial rate of the prime-force along the slope. So, the new prime-force is physically meaningful. The values of ${\mathcal P_{f}}$ should be estimated with the dissipative processes taking place along the channel. It requires some extra and proper understanding of the flow dynamics to exactly determine ${\mathcal P_{f}}$ in (\ref{Dispersion_E}) and, thus, the force $-{\mathcal P_{f}} x$ itself. However, we have formally postulated (or invented) a new force mechanism, the prime-force $-{\mathcal P_{f}} x$, and have shown the physical ground for its existence.
Due to the presence of the term $-{\mathcal P_{f}} x$, the landslide velocity model (\ref{Dispersion_F_3}), and its solution (\ref{Exact_3}) are novel.
The term $-{\mathcal P_{f}} x$ in (\ref{Dispersion_F_3}) adds some dissipative force that results in the deviation of the solution from the reference solution, ${\mathcal P_{f}} = 0$, produced by the driving force $\alpha$ and the viscous resistance associated with $\beta$. We can perceive $-{\mathcal P_{f}} x$ in different ways. It can be seen as the congregate of space dependent dissipative forces. Yet, $-{\mathcal P_{f}} x$ can be realized as any additional force other than the driving force $\alpha$ and the viscous resistance $-\beta u^2$ in their classical forms, which, unlike $-{\mathcal P_{f}} x$, do not contain any spatially varying dissipative contributions. As it is a completely new term and conception, its physical meaning and significance is worth exclusive elaboration in (\ref{Dispersion_E}), (\ref{Dispersion_F}), (\ref{Dispersion_F_2}), (\ref{Dispersion_F_3}), and (\ref{Exact_3}).
As demonstrated below in Fig. \ref{Fig_2} and Fig. \ref{Fig_3}, the prime-force turned-out to be very useful in controlling the mass flow dynamics, or any other dynamical system, that can be described by the structure of the model equations presented here.
\subsubsection{Constraining ${\mathcal P_{f}}$}
We need to physically constrain ${\mathcal P_{f}}$ in (\ref{Dispersion_E}).
Here, we present two possible scenarios. Without loss of generality, we impose physically legitimate and mathematically consistent conditions on the velocity and its derivatives at some position $x_0$ somewhere along the channel, or at appropriately chosen near source location.
\\[3mm]
{\bf Scenario A:} First, consider plausible, but typical velocity and their gradients with magnitudes: $u\lb x_0\rb = 35, \lb\partial u/\partial x\rb \lb x_0\rb = 0.01, \lb\partial^2 u/\partial x^2\rb \lb x_0\rb = 0.00021$, and $\beta = 0.0019$. Then, from (\ref{Dispersion_E}), by neglecting the time variation of $\partial u/\partial x$, ${\mathcal P_{f}}$ assumes the value on the order of $-0.0085$ and ${\mathcal P_{u}} = -0.3$. However, similar values of ${\mathcal P_{f}}$ and ${\mathcal P_{u}}$ can be obtained with other physically admissible choices of $u\lb x_0\rb, \lb\partial u/\partial x\rb \lb x_0\rb, \lb\partial^2 u/\partial x^2\rb \lb x_0\rb$, and $\beta$.
\\[3mm]
{\bf Scenario B:}
Second, consider another plausible, but fundamentally different scenario, such that the velocity attains its local maximum somewhere at $x_0$ in the channel (e.g., a contracting flow). This is mathematically equivalent to $\lb\partial u/\partial x\rb \lb x_0\rb = 0$ and $\lb\partial^2 u/\partial x^2\rb \lb x_0\rb$ is negative, say $-0.00032$. With this, for the typical velocity of $u\lb x_0\rb = 35$, the estimated value of ${\mathcal P_{f}}$ is on the order of $0.0112$, and ${\mathcal P_{u}} = 0.4$. Again, similar values of ${\mathcal P_{f}}$ and ${\mathcal P_{u}}$ can be obtained with other physically admissible choices of $u\lb x_0\rb, \lb\partial u/\partial x\rb \lb x_0\rb$ and $\lb\partial^2 u/\partial x^2\rb \lb x_0\rb$.
\subsubsection{Dynamics of the Prime-force: $-{\mathcal P_{f}} x$}
Solutions presented in Fig. \ref{Fig_2} for {\bf Scenario A}, with parameters $\alpha = 7.0$ and $\beta = 0.0019$ (as in Pudasaini and Krautblatter, 2022),
show how the negative values of ${\mathcal P_{f}}$ (thus, the positive additional prime-force $-{\mathcal P_{f}} x$) enhances the motion from that discarding the effect of ${\mathcal P_{f}}$, i.e., ${\mathcal P_{f}} = 0$. As the value of ${\mathcal P_{u}}$ (or ${\mathcal P_{f}} = 2\alpha\beta\,{\mathcal P_{u}}$) decreases, the ${\mathcal P}$-force increases, and the velocity continuously deviates away from the reference (${\mathcal P_{f}} = 0$) state (solution).
Even a very small value of ${\mathcal P_{f}}$ pushes the system away from the reference state, and it continues to do so as ${\mathcal P_{f}}$ decreases. Thus, the term $-{\mathcal P_{f}} x$ with ${\mathcal P_{f}} < 0$ strongly weakens the drag force, adds to the pre-existing driving force, and thus the reference-state is never reached. It can be a possible scenario as the mass travels further downstream such that the drag force is always weaker than the net driving force and the additional force generated by the new term, $-{\mathcal P_{f}} x$, along the slope. This means that, as long as the condition $ \lb \alpha -{\mathcal P_{f}} x\rb > \beta u^2$ is satisfied, the system accelerates, always.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=13cm]{Analytical_Soln_NoDispersion_1GenaralLambda05_NN.eps}
\end{center}
\caption[]{The landslide motion enhanced by the prime-force $-{\mathcal P_{f}} x$, for ${\mathcal P_{f}} < 0$ given by the solution (\ref{Exact_3}), where ${\mathcal P_u} = {\mathcal P_{f}}/\lb 2 \alpha\beta\rb $.
For any value of ${\mathcal P_{f}} < 0$, no matter how close it is to $0$, the system continuously deviates away from the reference state ${\mathcal P_{f}} = 0$, as long as ${\mathcal P_{f}} < 0$.}
\label{Fig_2}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=13cm]{Analytical_Soln_NoDispersion_1GenaralLambda0-1_NN.eps}
\end{center}
\caption[]{The landslide motion controlled by the prime-force $ -{\mathcal P_{f}} x$, for ${\mathcal P_{f}} > 0$ given by the solution (\ref{Exact_3}), where
${\mathcal P_u} = {\mathcal P_{f}}/\lb 2 \alpha\beta\rb $.
The constrained velocity dome-curves and the reduced travel distances are shown.
For any value of ${\mathcal P_{f}} > 0$, no matter how close it is to $0$, the system continuously bends below the reference state ${\mathcal P_{f}} = 0$, as long as ${\mathcal P_{f}} > 0$.}
\label{Fig_3}
\end{figure}
\\[3mm]
Even more interesting, and perhaps physically more important, is the situation when ${\mathcal P_{f}} > 0$.
This induces a spatially varying additional dissipative force resulting in the reduction of the total system force $T^s_f$ in (\ref{Dispersion_F_3t}) than before with the reference state, the solution with ${\mathcal P_{f}} = 0$, which effectively means that the mass decelerates as it slides downstream. This results in the reduced motion of the landslide. Then, depending on the magnitude of ${\mathcal P_{u}}$ (or ${\mathcal P_{f}} = 2\alpha\beta\,{\mathcal P_{u}}$), both the velocity and the travel distance will be reduced significantly to dramatically. The solutions are presented in Fig. \ref{Fig_3} for {\bf Scenario B}, with parameters $\alpha = 7.0$ and $\beta = 0.0019$, showing differently architectured beautiful dome-like constrained velocity fields and the firmly reduced mobility with increasing values of ${\mathcal P_{f}} > 0$. Interestingly, no matter how small, the novel observation is that, any positive value of ${\mathcal P_{f}}$ results in the significantly reduced mobility (velocity) and the run-out. This can happen, if there emerges any (other) energy dissipation mechanism along the slope. This effectively means that the total system force $T^s_f$ is continuously reduced as the mass slides downslope. So, after a certain position, the situation may prevail such that
$ \beta u^2 > \lb \alpha -{\mathcal P_{f}} x\rb $, and the system decelerates along the slope, always, as long as ${\mathcal P_{f}} > 0$. This results in the reduced motion and the travel distance.
\\[3mm]
Both Fig. \ref{Fig_2} and Fig. \ref{Fig_3} demonstrate that the term $-{\mathcal P_{f}} x$ in (\ref{Dispersion_F_3}) can quickly and strongly compel the system away from its reference state (${\mathcal P_{f}} = 0$).
From the physical point of view, the ${\mathcal P}$-force $\lb-{\mathcal P_{f}} x\rb$ is associated with any possible spatially varying dissipative (or anti-dissipative) force. This may include any elements of forces that are not contained in $\alpha$ and $\beta$. The Coulomb-type force in $\alpha$ and the drag force associated with $\beta$ are almost exclusively used in mass flow simulations. However, the spatially dependent ${\mathcal P}$-force, postulated here, is entirely new, that was made possible with our modelling technique. Yet, as revealed by Fig. \ref{Fig_2} and Fig. \ref{Fig_3}, it helps to fundamentally and precisely control the dynamics, deposition and run-out of the landslide. We formally summarize these results in a Theorem.
\\[3mm]
{\bf The ${\mathcal P}$-force Theorem 4.1:} {\it There exists a unique number ${\mathcal P_{f}} > 0$ such that the landslide run-out (motion/dynamics) described by the dynamical equation}
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{\frac{\partial u}{\partial t}
+\frac{\partial \lb u^2/2\rb}{\partial x}
= \alpha -\beta u^2-{\mathcal P_{f}} x,}
\label{Dispersion_F_Th}
\end{array}
\end{eqnarray}
\end{linenomath*}
{\it can be precisely controlled as expected. Here, $t$ is time, $x$ is the position along the slope, $u$ is the landslide velocity, $\alpha$ is the net driving force, $\beta$ is the viscous drag coefficient, and ${\mathcal P_{f}}$ is the prime-force coefficient.}
\subsection{The Prime-force: Essence, Implication and use in Simulation}
Here, we further explain the essence and application potential of the new prime-force. Practitioners and applied researchers are frequently in trouble in controlling the motion and run-out of mass flows. One of the biggest problems in dealing with the natural mass flow events is the proper simulation of their flow velocities and the run-out distances. This also applies to industrial mass transports. We know that, more or less, until now, in real event simulations, the forces are used in a way that fits best to the data, sometimes very low (almost none) and sometimes substantially (much) higher than reality (Christen et al., 2010; Frank et al., 2015; Dietrich and Krautblatter, 2019; Mergili et al., 2020a, 2020b; Frimberger et al., 2021; Shugar et al., 2021). This clearly indicates that there are some physical processes operating in nature we were not aware of before. Now, we have formally proven that, in principle, such process exists, which can be quantified. The prime-force does exactly this by controlling the motion in a precise way. Our simple model, and particularly the emergence of the new prime-force, $-{\mathcal P_{f}} x$, can tremendously help to address this long standing problem. In this respect, the model (\ref{Dispersion_F_3}), and its exact analytical solution (\ref{Exact_3}), can be very useful for practitioners and engineers in efficiently and quickly simulating the motion of the landslide down the entire slope, accelerating and decelerating motions, and deposition as it comes to an standstill in a fully controlled manner.
\\[3mm]
There are two important aspects. ($i$) We have physically and mathematically proven that a new force structure, the prime-force, exists, which is extra to the known frictional or viscous forces. ($ii$) There are challenges related to the correct reproduction of field observations through simulations. Often, we have difficulties in adequately back-calculating the observed mass flow events.
The prime-force is induced by the rate of spatially varying dissipative forces, but not merely the spatially varying friction and viscosity parameters.
So, the prime-force ($i$) will help to overcome the challenges in ($ii$) and accordingly support the practitioners.
However, if it is only about the spatial distribution and evolution of friction and viscosity parameters, which we still do not at all understand, and also various numerical issues (e.g., cell size, topography and flow boundary), both do not involve the spatial rate of dissipative forces, the challenges in ($ii$) could still be addressed without the prime-force.
\\[3mm]
The Coulomb force cannot contain all the friction effects. The same applies to the viscous drag. As simulations often contrast the observations significantly, and none of the forces we know can reproduce the observation, there must be something extra to the Coulomb and viscous drag forces in the form we already know. The prime-force does exactly this. The prime force congregates all forces with spatially varying rate of dissipations that are not in $\alpha$ and $\beta$, and complement to what we know. The prime-force may even combine the Coulomb and viscous forces and generate a spatially varying rate of dissipation.
One may yet think of producing similar results, as done above by the prime-force, by means of other forces which we know already. However, we can not achieve this by changing basal friction and/or the viscous drag. First, it is not possible in a classical way with Coulomb friction.
The exact solution (\ref{Exact_3}) is constructed by assuming that $\alpha$ does not vary along the slope, while the ${\mathcal P}$-force, $-{\mathcal P_{f}} x$, by nature, does. The same is true for the drag force.
Second, even by spatially varying the Coulomb friction (i.e., $\delta$) and/or the viscous drag ($\beta$), the motion, as controlled by the prime-force in (\ref{Dispersion_F_3}), cannot be achieved to precisely reproduce the observed run-out distance. Physically, $\delta$ is bounded from above, so often it is not able to control the motion in an appreciable way. Moreover, by definition, the viscous drag cannot bring the motion to a halt. But now, we can formally accommodate any additional energy dissipation mechanism in the ${\mathcal P}$-force accomplishing the observed effect rather than changing the Coulomb friction, whose value (as mentioned above) is often used arbitrarily in simulation to fit the data, or it does not exhibit any admirable effect.
\\[3mm]
For granular, debris and particle-laden flows, several situations may arise where the dissipative (or anti-dissipative) force can increase (or decrease) as the mass moves downslope. There can be several factors aiding to the prime force. We mention some possible scenarios that may contribute to the spatial rate of the prime-force, i.e., ${\mathcal P_f}$. ($i$) Often the debris flow heads and lateral flanks become more and more granular dominated, or frictionally stronger due to phase-separation and/or particle sorting. These are observed phenomena (Johnson et al., 2012; de Haas et al., 2015, 2016; de Haas and van Woerkom, 2016; Pudasaini and Fischer, 2020b; Baselt et al., 2021, 2022).
($ii$) The collisional and viscous dissipations can increase as flow moves on, e.g., by added particles and fines (the situation prevails due to basal erosion and entrainment), and increased agitations (de Haas et al., 2015, 2016; Pudasaini and Mergili, 2019; Pudasaini and Krautblatter, 2021). The viscous resistance can also increase due to added fragmented fine particles, e.g., in rock-ice avalanche motion (Pudasaini and Krautblatter, 2014).
($iii$) The energy dissipation may increase in the downstream as the flow transits, e.g., from the glacial surface to the gravel-rich, or the rough moraine surface.
($iv$) Detailed topographic effects (Fischer et al., 2012), that could not be resolved otherwise, may also be included as an energy dissipation mechanism.
\\[3mm]
In reality, the prime-force coefficient, ${\mathcal P_{f}}$, can be a complex function of some or all of those physical phenomena described above, and any other permissible circumstances associated with the dissipative mass flows with the rate of dissipative forces along the slope.
Its admissible forms are yet to be determined. Still, ${\mathcal P_{f}}$ could also be constrained from laboratory experiments or from the field data with respect to the observed dynamics and the run-out. Alternatively, the practitioners may ascertain ${\mathcal P_{f}}$ in empirically adequate ways, if they prefer to do so.
This adds an additional uncertain parameter to the simulations, besides the existing ones. This may make parameter calibration and predictive simulations even more difficult, but helping to control the landslide as
observed.
However, we mention that, as the prime-force is a new concept, further intensive research would help to boost its clarity and expedite its practical applications.
\\[3mm]
Analytical solution presented in (\ref{Exact_3}) formally proves that the new dissipative force appreciably controls the motion and runout. Depending on its sign, it can enhance or control the motion, equivalently, stretch (Fig. \ref{Fig_2}) or reduce (Fig. \ref{Fig_3}) the travel distance (or coverage area).
With this, we can now formally include the new dissipative force $-{\mathcal P_{f}} x$ (similarly in other directions) in the list of forces in the momentum balance equations (\ref{Model_Final_Momentum}), and implement the prime-force in any simulation of mass flow. There are some technical aspects to consider while implementing the new force in computing.
$(i)$ Note that, ${\mathcal P_{f}}$ are relatively small numbers.
$(ii)$ In general, we can have different ${\mathcal P_{f}}$ for different phases.
$(iii)$ Because of the possible directional inhomogeneity, ${\mathcal P_{f}}$ can be different in $x$ and $y$ directions, say ${\mathcal P_{fx}}$ and ${\mathcal P_{fy}}$.
$(iv)$ We can formally include ${-\alpha_s\mathcal P_{fx}} x$ in the list of forces in (\ref{Source_x_s}), say at the end of it, similar for (\ref{Source_x_fs}) and (\ref{Source_x_f}) with ${\alpha_{fs}}$ and ${\alpha_{f}}$.
$(v)$ For the $y$-direction for solid, we should use ${-\alpha_s\mathcal P_{fy}} y$, but we should remember that the outward directions are the increasing directions. Similar for other phases in $y$-direction.
So, in principle, the prime-force can be relatively easily included in any computational softwares, such as the r.avaflow (Mergili and Pudasaini, 2021a,b) in a straightforward way.
\section{A Simple Dispersion Equation}
Reducing the sophistication, we consider a geometrically two-dimensional motion down a slope. We further assume that the relative velocity between coarse and fine solid
particles $(u_s, u_{fs})$ and the fluid phase $(u_f)$ in the landslide (debris) material is negligible, that
is, $u_s \approx u_{fs} \approx u_f =: u$, and so is the viscous deformation of the fluid.
This means, for simplicity, we are considering an effectively single-phase mixture (consisting of solid particles composed of coarse solid and fine solid, and viscous fluid) flow (Pudasaini and Krautblatter, 2021, 2022).
Then, by summing up the mass and momentum balance equations in Section 2.2, we obtain a single mass and
momentum balance equation describing the motion of a landslide (or a mass flow) with the non-hydrostatic contribution as:
\begin{equation}
\frac{\partial h}{\partial t} + \frac{\partial }{\partial x}\left ( hu\right ) = 0,
\label{Eqn_1}
\end{equation}
\begin{equation}
\frac{\partial}{\partial t}(hu) + \frac{\partial }{\partial x} \left [ h \left \{ u^2 + \left (\alpha_s\beta_s +\alpha_s\beta_f \right)\frac{h}{2} +\left (\alpha_s\mathcal D_s + \alpha_f\mathcal D_f\right)\right\}\right ]= h S,
\label{Eqn_2}
\end{equation}
where,\\ \indent
$\alpha_f = \left (1-\alpha_s \right)$,
$\displaystyle{
\alpha_s\beta_s +\alpha_f\beta_f = \left [ \left (1-\gamma_s^f \right)K_s\alpha_s + \left ( 1-\alpha_s\right)\right ]g^z
+ \left [ \alpha_s\left (K_s -1 \right)+1\right]
\lb \frac{D\bar w}{Dt} + C_{_{DV}}^s\bar w u \rb
}$,\\ \indent
$\displaystyle{\alpha_s\mathcal D_s+\alpha_f\mathcal D_f = \frac{h^2}{12}\left [\alpha_s\left(K_s-1 \right) +1\right] \left\{ \left ( \frac{\partial u}{\partial x}\right )^2
-\frac{D}{Dt}\lb \frac{\partial u}{\partial x}\rb - 2C_{_{DV}}^s u\frac{\partial u}{\partial x}
\right\}} $,\\ \indent
$\displaystyle{S = g^x
-\mu_s\alpha_s\left\{ \lb 1- \gamma_s^f\rb g^z + \frac{D\bar w}{Dt} + C_{_{DV}}^s\bar w u \right\}
-\alpha_s \left\{\gamma_s^f g^z + \frac{D\bar w}{Dt} + C_{_{DV}}^s\bar w u\right\}\frac{\partial h}{\partial h}
- C_{_{DV}}^su^2}$,\\[3mm]
are the fluid fraction in the mixture, the coefficient emerging from the hydraulic pressure gradients for the solid and fluid including the enhanced effects due to non-hydrostatic contributions, the dispersion contributions emerging from the non-hydrostatic consideration, and the source containing the forces.
Together with the mass balance (\ref{Eqn_1}), the momentum balance (\ref{Eqn_2}) can be written as:
{\small
\begin{eqnarray}
&&\frac{\partial u}{\partial t} + u\frac{\partial u}{\partial x}
+ \left [ \left \{\left( \left( 1-\gamma_s^f\right)K_s + \gamma_s^f\right)\alpha_s+\left ( 1-\alpha_s \right )\right\}g^z
+\alpha_s \left\{ \lb \frac{\partial}{\partial t}+u \frac{\partial}{\partial x}\rb {\bar w} + C_{_{DV}}^s\bar w u \right\}
\right ] \frac{\partial h}{\partial x}
\nonumber\\
&&+ \frac{1}{h}\frac{\partial }{\partial x}\left[\left \{ \alpha_s\left (K_s -1 \right)+1\right\}\left [\frac{h^3}{12} \left\{ \left ( \frac{\partial u}{\partial x}\right )^2 -\frac{\partial}{\partial x}\frac{\partial u}{\partial t} - u \frac{\partial^2u}{\partial x^2}
-2C_{_{DV}}^s u \frac{\partial u}{\partial x}
\right\}
+\frac{h^2}{2} \left\{ \lb \frac{\partial}{\partial t}+u \frac{\partial}{\partial x}\rb {\bar w} + C_{_{DV}}^s\bar w u \right\} \right]\right]\nonumber\\
&&= g^x
-\mu_s\alpha_s\left[ \lb 1- \gamma_s^f\rb g^z + \left\{ \lb \frac{\partial}{\partial t}+u \frac{\partial}{\partial x}\rb {\bar w} + C_{_{DV}}^s\bar w u \right\} \right]
- C_{_{DV}}^su^2.
\label{Eqn_3}
\end{eqnarray}
}
\hspace{-3mm}
The second term on the left hand side of (\ref{Eqn_3}) describes the advection, while the third term (in the square bracket) describes the extent of the local deformation that stems from the hydraulic pressure gradient of the free-surface of the landslide in which $\left ( 1-\alpha_s\right)g^z\partial h/\partial x$ emerges from the hydraulic pressure gradient associated with possible interstitial fluids in the landslide, and the terms associated with ${\bar w}$ are the components from enhanced gravity. The fourth term on the left hand side are extra contributions in the flux due to the non-hydrostatic contributions. Moreover, the third term on the left hand side and the other terms on the right hand side in the momentum equation (\ref{Eqn_3}) represent all the involved forces.
The first and second
terms on the right hand side of (\ref{Eqn_3}) are
the gravity
acceleration, effective Coulomb friction that includes
lubrication $\left ( 1- \gamma_s^f\right )$, liquefaction $\left (
\alpha_s\right )$ (because, if there is no or substantially low amount of solid, the mass is fully
liquefied, e.g., lahar flows), the third term with $\bar w$ emerges from enhanced gravity, and the fourth term is the viscous drag,
respectively. Note that the term with $1-\gamma_s^f$ or $\gamma_s^f$ originates from the buoyancy effect. By setting $\gamma_s^f = 0$ and $\alpha_s = 1$, we obtain a dry landslide, grain flow, or an avalanche motion. However, we keep $\gamma_s^f$ and $\alpha_s$ also to include possible fluid effects in the landslide (mixture).
\\[3mm]
Note that for $K_s = 1$ (which may prevail for extensional flows, Pudasaini and Hutter, 2007), the third term on the left hand side associated with $\partial h/\partial x$ simplifies drastically, because $\left \{ \left( \left( 1-\gamma_s^f\right)K_s + \gamma_s^f\right)\alpha_s+\left ( 1-\alpha_s \right )\right \}$ becomes unity.
So, the isotropic assumption (i.e., $K_s = 1$) loses some important information about the solid content and the buoyancy effect in the mixture.
\subsection{A Landslide Dispersion Equation}
For simplicity, we introduce the notations as: $b = \left \{ \alpha_s\left (K_s -1 \right)+1\right\}$, $\alpha = \left [\,g^x-(1-\gamma_s^f)\alpha_s\mu_s g^z\,\right]$, and $\beta = C_{_{DV}}^s$. Here, $b, \alpha$ and $\beta$ are the pressure parameter, net driving force and the viscous drag coefficient, respectively. Assume that the time-dependent terms in (\ref{Eqn_3}) can be ignored in relation to other terms. Moreover, let $hu = {\mathcal F}$ be a typical flux, and $\partial u/\partial x$ is a small quantity such that $\lb \partial u/\partial x\rb^2$ is negligible.
Consider the definition of $\bar w $ from (\ref{tau_zz_aaa}).
Then, with a long wave approximation (we suppose that $h$ can be approximated by a constant, or simply parameterize it, $h = h_0$), the momentum balance (\ref{Eqn_3}) can be reduced to yield a third-order inhomogeneous non-linear ordinary differential equation in $u$ with parameters ${\mathcal D_D}, {\mathcal D_{S1}}, {\mathcal D_{S2}}, \alpha, \beta$:
\begin{eqnarray}
{\mathcal D_P}\frac{\partial^3 u}{\partial x^3} + {\mathcal D_{S1}}\frac{\partial^2 u}{\partial x^2} + \lb u + {\mathcal D_{S2}}\rb \frac{\partial u}{\partial x}= \alpha - \beta u^2,
\label{Eqn_5}
\end{eqnarray}
where, $\displaystyle{{\mathcal D_P} = \pm\frac{1}{3} b h_0 {\mathcal F}}$,
$\displaystyle{{\mathcal D_{S1}} = \pm\frac{1}{2}\left [ \frac{5}{6}\beta b h_0 + \mu_s\alpha_s\right]{\mathcal F}}$,
${\displaystyle{{\mathcal D_{S2}} = \pm\frac{1}{2}\mu_s\beta\alpha_s{\mathcal F}}} $ are associated with dispersion. Here, the $\pm$ sign correspond to the primarily expanding or contracting flows, which can be obtained by separately analyzing the dispersive contributions in (\ref{Eqn_3}).
We call (\ref{Eqn_5}) the landslide dispersion equation in which ${\mathcal D_P}$ plays the primary role as it is associated with the highest order term therein, while ${\mathcal D_{S1}}$ and ${\mathcal D_{S2}}$ play the secondary role. So, ${\mathcal D_P}$ is termed as the prime dispersion parameter. This is a simple, yet very interesting, dispersion equation that characterizes the dispersion effect in the mass flow.
\subsection{Solution to the Dispersion Equation (\ref{Eqn_5})}
We analyze in detail the effect of dispersion in (\ref{Eqn_5}). Without the dispersive terms, (\ref{Eqn_5}) is the simple steady-state landslide velocity model developed in Pudasaini and Krautblatter (2022). We numerically solved (\ref{Eqn_5}) with the boundary conditions $\displaystyle{u(0) = 0.0, \frac{\partial u}{\partial x}(0) = 0.5, \frac{\partial^2 u}{\partial x^2}(0) = 0.0}$. The last two conditions are additionally required due to dispersion related dynamics. All conditions can be fixed based on the physics of the underlying problem. The results are shown in Fig. \ref{Fig_4} both with dispersion, ${\mathcal D_P = 327}, {\mathcal D_{S1} = 17}, {\mathcal D_{S2} = 0.03}$ (representing a realistic situation with $b = 1.0, h_0= 7.0, {\mathcal F} = 140, \alpha_s = 0.65, \mu_s = 0.36 \lb \delta_s = 20^\circ\rb, \alpha = 7.0, \beta = 0.0019$), and without dispersion ${\mathcal D_P = 0.0}, {\mathcal D_{S1} = 0.0}, {\mathcal D_{S2} = 0.0}$ effects. To demonstrate the influence of dispersion parameters ${\mathcal D}$ on the dynamics, we have amplified, downplayed, or ignored their values with different scales as $2.0\times{\mathcal D}, 1.0\times{\mathcal D}, 0.1\times{\mathcal D}, 0.0\times{\mathcal D}$, where the last value corresponds to the neglection of all dispersion effects. Figure \ref{Fig_4} clearly reveals fundamental effects of dispersion on the landslide dynamics. Moreover, the velocity distribution with dispersion is more complex due to its association with the higher-order derivative terms in (\ref{Eqn_5}). Dispersion produces a wavy velocity field of changing intensity about the simple reference state without dispersion.
Local surge developments and attenuations as well as enhanced or hindered motions are often observed dynamical spectacles in landslides and debris avalanches.
Such explicit description of the dispersive wave is the first of this kind for the avalanching debris mass. Once the landslide is triggered, the dispersive solution deviates significantly away from the non-dispersive one. However, after a sufficiently long distance, the dispersive solution tends to approach the non-dispersive state given by (\ref{Exact_3}) with ${\mathcal P_u = 0}$. Yet, significantly different scenarios can be generated with other sets of dispersion parameters. Alternatively, as ${\mathcal D_P \to 0.0}, {\mathcal D_{S1} \to 0.0}, {\mathcal D_{S2} \to 0.0}$, the dispersive wave coincides with the non-dispersive elementary solution. This proves the consistency of our model and also highlights the essence of dispersion in mass transport.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=13cm]{Dispersion_Depth_Paper_D_No_D_Full_Paper_NN_AllScale.eps}
\end{center}
\caption[]{The landslide velocity distribution with dispersion $\lb{\mathcal D_P = 327}, {\mathcal D_{S1} = 17}, {\mathcal D_{S2} = 0.03}\rb$ and without dispersion $\lb{\mathcal D_P} =0.0, {\mathcal D_{S1}} = 0.0, {\mathcal D_{S2}} = 0.0\rb$ described by (\ref{Eqn_5}). With dispersion ${\mathcal D}$, depending on its magnitude, the landslide behaves fundamentally differently by producing meanders of variable intensities around the reference state without dispersion.}
\label{Fig_4}
\end{figure}
\subsection{Influence of the Solid Volume Fraction in Dispersion}
The solid volume fraction $\alpha_s$ is the key (physical) parameter in the mixture that governs the landslide motion and deformation. The strength of the landslide material is directly related to $\alpha_s$. The solid volume fraction influences all the parameters ${\mathcal D_P}, {\mathcal D_{S1}}, {\mathcal D_{S2}}$ and $\alpha$ in the dispersion equation (\ref{Eqn_5}), except $\beta$. So, here we analyze how the solid volume fraction regulates the landslide dispersion.
Landslide velocity distributions with dispersion for different solid volume fractions in the mixture are presented in Fig. \ref{Fig_5}.
Dispersion is minimum for the fully dry material, and maximum for the vanishing solid fraction, akin to the fluid flow. The dispersion intensity increases energetically as the solid volume fraction decreases. This reveals that dispersion is related to the fluidness of the material. However, for higher values of $\alpha_s$ dispersion becomes weaker and weaker far downstream as compared to that near the source region.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=13cm]{Dispersion_Depth_Paper_D_No_D_Full_Paper_NN_a_s.eps}
\end{center}
\caption[]{Landslide velocity distributions with dispersion described by (\ref{Eqn_5}) for different solid volume fractions $\alpha_s$ in the landslide mixture. Dispersion increases firmly with decreasing solid volume fraction.}
\label{Fig_5}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=13cm]{Dispersion_Depth_Paper_D_No_D_Full_Paper_NNNN_delta.eps}
\end{center}
\caption[]{Landslide velocity distributions with dispersion described by (\ref{Eqn_5}) for different basal friction angles $\delta$. Dispersion increases strongly with decreasing basal friction angle.}
\label{Fig_6}
\end{figure}
\subsection{Influence of the Basal Friction in Dispersion}
The basal friction angle $\delta$ is a dominant physical parameter controlling the landslide dynamics. As for the solid volume fraction, the weaker material is associated with the lower friction angle.
However, unlike the solid volume fraction, basal friction influences only ${\mathcal D_{S1}}, {\mathcal D_{S2}}$ and $\alpha$ in the dispersion equation (\ref{Eqn_5}), but not ${\mathcal D_P}$ and $\beta$. Landslide velocity distributions with dispersion for different frictions in the mixture are presented in Fig. \ref{Fig_6}. Dispersion increases strongly with decreasing values of $\delta$, with highest dispersion taking place for the motion of a frictionless material $\lb \delta = 0^\circ\rb$, akin to a fluid flow. However, for higher values of $\delta$, dispersion becomes relatively weaker as the landslide continues to propagate downstream.
\\[3mm]
Both the solid volume fraction and the friction angle define the mechanical response of the landslide material against the applied forces, and govern the landslide motion and deformation. However, they regulate the landslide dynamics fundamentally differently, so are the dispersions with changing solid fractions and the basal frictions. These facts are demonstrated in Fig. \ref{Fig_5} and Fig. \ref{Fig_6}. Although at the first glance, they look similar, the dispersion intensity is higher with the change of the basal friction as compared to that with the solid volume fraction. This can be explained, because basal fiction is the main physical parameter determining the landslide dynamics. These results are in line with our intuition and experience, and indicate the consistency of our model. This also sheds light on the physical significance of the simple dispersion model derived here.
\section{Summary}
We considered the multi-phase mass flow model by Pudasaini and Mergili (2019) and extended it by including the non-hydrostatic contributions. This produces a novel non-hydrostatic multi-phase mass flow model. Effective normal stresses are constructed for all the solid, fine-solid and fluid phases in the mixture from the normal stress components, which include the interfacial momentum transfers such as the buoyancy, drag and virtual mass forces. Depending on the nature of the components in the effective normal stresses, the normal loads are separated into the enhanced gravity and the dispersion, which, respectively, correspond to the acceleration in the flow depth direction and the mass fluxes associated with the slope parallel directions. While drag and virtual mass forces appear in both, buoyancy is present only in the enhanced gravity for solid and fine-solid because it is associated with the reduced normal load of the solid particles in the mixture. As enhanced gravity and dispersion both emerge from the effective normal load, these enter into the lateral momentum fluxes via the hydraulic pressure gradients and additionally introducing the dispersion effects. This resulted in a complex and highly non-linear new contributions in the momentum fluxes. This may pose a great challenge in solving the model equations. This is mainly due to the involvements of the time derivatives in the fluxes that appear in the dispersion, and also in the enhanced gravity.
To reduce the complexity, we have also presented some simplifications and approximations for the time derivatives appearing in the enhanced non-hydrostatic contributions. Similarly, we have presented analysis of the dispersion relations showing the role of the drag force. We discussed some special situations where the non-hydrostatic dispersive effects are more pronounced in multi-phase particle-fluid mixture mass flow than in single-phase flows. We proved that the negligible dispersion leads to the generalization of the existing inviscid, dissipative Burgers’ equation with source term.
We also presented simplified models that can help in solving the equations with reduced complexity. The reduced models already appeared to be the important generalizations and extensions of several mass flow models available in the literature. We formally postulated a novel, spatially varying dissipative (or anti-dissipative) force, called the prime-force. The practitioners and engineers may find the prime-force very useful in solving technical problems as it precisely controls the dynamics, run-out and the deposition of mass flows. We elucidated the need of formally including this new, physically-founded force in momentum balance equations.
We constructed a simple dispersion model and its solution that highlighted the essence of dispersion on the flow dynamics. We consistently demonstrated that dispersion produces a wavy velocity field around the reference state without dispersion. Our results show that dispersion increases strongly as the solid volume fraction and the basal friction decreases. The explicit description of dispersive waves and their control by the solid volume fraction and the basal friction are seminal understanding in mass flows. So, this contribution sets a foundation for a more complete and general simulation of non-hydrostatic dispersive, multi-phase mass flows.
\\[3mm]
{\bf Acknowledgments:} Shiva P. Pudasaini acknowledges the financial support provided by the Technical University of Munich with the Visiting Professorship Program, and the international research project: AlpSenseRely $-$ Alpine remote sensing of climate‐induced natural hazards - from the Bayerisches Staatsministerium f\"ur Umwelt und Verbraucherschutz, Munich, Bayern.
\section*{Appendix}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
The expressions and discussions below are mainly based on Pudasaini and Mergili (2019).
\\[3mm]
{\bf A. The drag coefficients} are given by (Pudasaini, 2020):
\begin{linenomath*}
\begin{subequations}\label{Drags}
\begin{align}
& \displaystyle{C_{DG}^{s,f} = \frac{\alpha_s \alpha_f\lb 1-\gamma_s^f\rb g}{\left [\mathcal U_T^{s,f}\left\{{\cal P}^{s,f}\mathcal F^{s,f}\lb Re_p^{s,f}\rb + \lb 1-{\cal P}^{s,f}\rb\mathcal G^{s,f}\lb Re_p^{s,f}\rb\right\}
+ {\mathcal S}_{\mathcal P}^{s,f}
\right ]^{\jmath}}},\label{Drags_s,f}\\[3mm]
& \displaystyle{C_{DG}^{s,fs} = \frac{\alpha_s \alpha_{fs}\lb 1-\gamma_{s}^{fs}\rb g}{\left [\mathcal U_T^{s,fs}\left\{{\cal P}^{s,fs}\mathcal F^{s,fs}\lb Re_p^{s,fs}\rb + \lb 1-{\cal P}^{s,fs}\rb\mathcal G^{s,fs}\lb Re_p^{s,fs}\rb\right\}
+ {\mathcal S}_{\mathcal P}^{s,fs}
\right ]^{\jmath}}},\label{Drags_s,fs}\\[3mm]
& \displaystyle{C_{DG}^{fs,f} = \frac{\alpha_{fs} \alpha_{f}\lb 1-\gamma_{fs}^f\rb g}{\left [\mathcal U_T^{fs,f}\left\{{\cal P}^{fs,f}\mathcal F^{fs,f}\lb Re_p^{fs,f}\rb + \lb 1-{\cal P}^{fs,f}\rb\mathcal G^{fs,f}\lb Re_p^{fs,f}\rb\right\}
+ {\mathcal S}_{\mathcal P}^{fs,f}
\right ]^{\jmath}}},\label{Drags_fs,f}
\end{align}
\end{subequations}
\end{linenomath*}
where, ${\cal P}^{s,f} \in (0, 1)$ is a function of the solid volume fraction ${\cal P}^{s,f} = \alpha_s^m$, {where $m$ is a positive number, close to 1},
combines the fluid-like, $\displaystyle{\mathcal F^{s,f} = {\gamma_s^f}\lb\alpha^f_s\rb^3 Re_p^{s,f}/180}$, and solid-like, $\displaystyle{\mathcal G^{s,f}= \alpha_f^{M^{s,f} -1}}$, drag contributions between solid and fluid components in three-phase mass flows; $\mathcal U_T^{s,f}$ is the terminal velocity of a particle falling through the fluid, $\jmath = 1 \, \mbox{or}\, 2$ is selected according to whether linear or quadratic drag coefficients are {used, and} $M^{s,f} = M^{s,f}\lb Re_p^{s,f}\rb$ depends on the particle Reynolds number $Re_p^{s,f} = \rho_fd_s~\mathcal U_T^{s,f}/\eta_f$ (Richardson and Zaki, 1954; Pitman and Le, 2005). Furthermore, $d_s$ is particle diameter, $\gamma_s^f = \rho_f/\rho_s$ is the fluid to solid density ratio, and $\alpha^f_s = \alpha_f/\alpha_s$ is the fluid to solid fraction ratio.
\\[3mm]
$\displaystyle{{\mathcal S}_{\cal P}^{s,f} = \lb \frac{\mathcal P^{s,f}}{\alpha_s} + \frac{1 - {\mathcal P^{s,f}}}{\alpha_f}\rb {\mathcal K^{s,f}}}$ in (\ref{Drags_s,f}) is called the smoothing
function, where {${\mathcal K^{s,f}} = |\alpha_s{\mathbf u}_s + \alpha_f {\mathbf u}_f|$} is
determined by the mixture mass flux per unit mixture density, typically ${\mathcal K^{s,f}} = 10$ ms$^{-1}$ (Pudasaini, 2020). The emergence of $\mathcal S_{\cal P}^{s,f}$ in (\ref{Drags_s,f}) is crucial for the broad structure of the generalized drag that removes the singularity {from the existing drag coefficients (Pudasaini, 2020)}. With this, (\ref{Drags_s,f}) is called the enhanced generalized drag in mixture mass flows. This fully describes the drag for any values of the solid volume fraction $\alpha_s$. Similar discussions hold for the other drags ${C}_{DG}^{s,fs}$ and ${C}_{DG}^{fs,f}$.
\\[3mm]
{\bf B. The virtual mass} induced mass and momentum enhancements for the solid-phase due to fluid and the fine-solid {are denoted by $u_s^{vm}$ and $uu_s^{vm}$, $uv_s^{vm}$, and} are written as:
\begin{linenomath*}
\begin{subequations}\label{Virtual_mass_enhancements_s}
\begin{align}
& u_{s}^{vm} = \gamma_s^f {\mathcal C}^{s,f}\lb u_f -u_s \rb + \gamma_s^{fs} {\mathcal C}^{s,fs}\lb u_{fs} -u_{s} \rb ,\\[3mm]
& uu_{s}^{vm}= \gamma_s^f {\mathcal C}^{s,f}\lb u_f^2 -u_s^2 \rb + \gamma_s^{fs} {\mathcal C}^{s,fs}\lb u_{fs}^2 -u_{s}^2 \rb,\\[3mm]
& uv_{s}^{vm} = \gamma_s^f {\mathcal C}^{s,f}\lb u_{f}v_{f} -u_sv_s \rb + \gamma_s^{fs} {\mathcal C}^{s,fs}\lb u_{fs}v_{fs} -u_{s}v_{s}\rb.
\end{align}
\end{subequations}
\end{linenomath*}
The
virtual mass force {coefficient} ${\mathcal C}^{s,f}$ in (\ref{Virtual_mass_enhancements_s}) is given by (Pudasaini, 2019):
\begin{linenomath*}
\begin{eqnarray}
\displaystyle{ {\mathcal C}^{s,f} = \frac{\mathcal N_{vm}^0({\ell} + \alpha_s^n) - 1}{\alpha_s/\alpha_f + \gamma_s^f},
}
\label{VM_M}
\end{eqnarray}
\end{linenomath*}
where ${\mathcal N_{vm}}$ is the virtual mass number, and $\ell$ and $n$ are some numerical parameters.
This model covers any distribution of the dispersive phase (dilute to dense distribution of the solid particles) that evolves automatically as a function of solid volume fraction. The physically most relevant values for the parameters
can be: $\mathcal N_{vm}^0 = 10$, $\ell = 0.12$ and $n = 1$ (Pudasaini, 2019). The other virtual mass force coefficients ${\mathcal C^{s,fs}} $ and ${\mathcal C^{fs,f}}$ can be constructed from (\ref{VM_M}). Similarly, the virtual mass force induced mass and momentum enhancements for the fine-solid and fluid phases are given by:
\begin{linenomath*}
\begin{subequations}\label{Virtual_mass_enhancements_fs1}
\begin{align}
& u_{fs}^{vm} = \gamma_{fs}^f {\mathcal C}^{fs,f}\lb u_f - u_{fs}\rb - \alpha^s_{fs}{\mathcal C}^{s,fs}\lb u_{fs} - u_s\rb,\\[3mm]
& uu_{fs}^{vm}= \gamma_{fs}^f {\mathcal C}^{fs,f}\lb u_f^2 - u_{fs}^2\rb - \alpha^s_{fs}{\mathcal C}^{s,fs}\lb u_{fs}^2 - u_s^2\rb,\\[3mm]
& uv_{fs}^{vm} = \gamma_{fs}^f {\mathcal C}^{fs,f}\lb u_fv_f - u_{fs}v_{fs}\rb - \alpha^s_{fs}{\mathcal C}^{s,fs}\lb u_{fs}v_{fs} - u_sv_s\rb,
\end{align}
\end{subequations}
\end{linenomath*}
and
\begin{linenomath*}
\begin{subequations}\label{Virtual_mass_enhancements_f1}
\begin{align}
& u_{f}^{vm} = \alpha^s_{f} {\mathcal C}^{s,f}\lb u_f -u_s \rb + \alpha^{fs}_{f} {\mathcal C}^{fs,f}\lb u_{f} -u_{fs} \rb,\\[3mm]
& uu_{f}^{vm}= \alpha^s_{f} {\mathcal C}^{s,f}\lb u_f^2 -u_s^2 \rb + \alpha^{fs}_{f} {\mathcal C}^{fs,f}\lb u_{f}^2 -u_{fs}^2 \rb ,\\[3mm]
& uv_{f}^{vm} = \alpha^s_{f} {\mathcal C}^{s,f}\lb u_fv_f -u_s v_s \rb + \alpha^{fs}_{f} {\mathcal C}^{fs,f}\lb u_{f} v_f -u_{fs} v_{fs}\rb,
\end{align}
\end{subequations}
\end{linenomath*}
respectively, where, $\alpha^s_{fs} = \alpha_s/\alpha_{fs}, \alpha^s_{f} = \alpha_s/\alpha_{f}$ and $\alpha^{fs}_{f} = \alpha_{fs}/\alpha_{f}$ are the fraction ratios.
By consistently replacing $u$ by $v$ in (\ref{Virtual_mass_enhancements_s})-(\ref{Virtual_mass_enhancements_f1}), we obtain the virtual mass induced mass and momentum enhancements in the $y$-direction.
\\[3mm]
{\bf C. The $x$-directional fluid-type basal shear stresses in the $xz$-plane} are given, either by the no-slip condition (for both the fluid, and fine-solid):
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}\label{basal_shear_stress_u}
\displaystyle{
\left[\frac{\partial u_{f}}{\partial z}\right]_b = \chi_{u_{f}}\frac{u_{f}}{h}, \,\,\,
\left[\frac{\partial u_{fs}}{\partial z}\right]_b = \chi_{u_{fs}}\frac{u_{fs}}{h}
},
\end{array}
\end{eqnarray}
\end{linenomath*}
or by the no-slip condition for fluid, and the Coulomb-slip condition for fine-solid:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}\label{basal_shear_stress_uu}
\displaystyle{
\left[\frac{\partial u_{f}}{\partial z}\right]_b = \chi_{u_{f}}\frac{u_{f}}{h}, \,\,\,
\left[\frac{\partial u_{fs}}{\partial z}\right]_b = \frac{C_{u_{fs}}^F}{\nu_{fs}^e}p_{fs} +2 C_{u_{fs}}^F \frac{\partial u_{fs}}{\partial x},}
\end{array}
\end{eqnarray}
\end{linenomath*}
with the Coulomb friction coefficient $C_{u_{fs}}^F = - u_{fs}/|{\bf u}_{fs}|\tan\delta_{fs}$, where $\delta_{fs}$ is the basal friction angle for the fine-solid.
The parameters $\chi_{u_f}$ and $\chi_{u_{fs}}$ in (\ref{basal_shear_stress_u}) and (\ref{basal_shear_stress_uu}) model the possible velocity distributions of the respective phases in the $xz$-plane normal to the sliding surface.
\\[3mm]
{\bf D. The viscous stresses} associated with $\nu^e_{fs}$ and $\nu^e_{f}$ in (\ref{Source_x_fs})-(\ref{Source_x_f}) are related to the Newtonian-type viscous stresses. They include pressure, rate, yield strength and friction, see below.
\\[3mm]
{\bf E. The effective fluid and fine-solid kinematic viscosities} are given by:
\begin{linenomath*}
\begin{equation}\label{Viscosities}
\nu_f^e
= \nu_f + \frac{\tau_{y_f}}{||{\bf D}_f||}\left [ 1-\exp\lb-r_y||{\bf D}_f||\rb\right ],\,\,\,
\nu_{fs}^e
= \nu_{fs} + \frac{\tau_{y_{fs}}}{||{\bf D}_{fs}||}\left [ 1-\exp\lb-r_y||{\bf D}_{fs}||\rb\right ],
\end{equation}
\end{linenomath*}
where $\tau_{y_f}$ and $\tau_{y_{fs}}$ are the {corresponding} yield stresses, $r_y$ are the parameters for regularization, and $\tau_{y_{fs}} = \sin\phi_{fs} p_{fs}$,
and, {${\bf D}_f$ is the deviatoric strain-rate tensor for fluid}.
In the viscosities (\ref{Viscosities}), {the} depth-averaged norm of ${\bf D}_f$ {is obtained as}:
\begin{linenomath*}
\begin{eqnarray}
\begin{array}{lll}
\displaystyle{{||\bf D}_f|| = {\left |4\frac{\partial u_f}{\partial x}\frac{\partial v_f}{\partial y}
- \lb \frac{\partial u_f}{\partial y} + \frac{\partial v_f}{\partial x}\rb^2
- \lb \left [ \frac{\partial u_f}{\partial z}\right ]_b \rb^2
- \lb \left [ \frac{\partial v_f}{\partial z}\right ]_b \rb^2
\right |}^{1/2}},
\label{Norm_D_f}
\end{array}
\end{eqnarray}
\end{linenomath*}
$||{\bf D}_f||$ is given by the second invariant ($I\!I_{{\bf D}_f}$) of the deviatoric strain-rate tensor for fluid: $||{\bf D}_f|| = \sqrt{I\!I_{{\bf D}_f}}$ with, $I\!I_{{\bf D}_f} = \displaystyle{\frac{1}{2}\left [ \text{tr}\lb {\bf D}_f\rb^2 - \text{tr}\lb {\bf D}_f^2\rb \right ]} $.
{The norm of the deviatoric strain-rate tensor for fine-solid, ${\bf D}_{fs}$, is obtained similarly.}
\\[3mm]
{\bf Flow and No-flow Regions:}
The yield criteria help to precisely distinguish the flow and no-flow regions and depend on the rate of deformation and the material strengths for both the fine-solid and fluid phases (Prager and Drucker, 1952; Domnik et al., 2013). Both the fine-solid and fluid phases yield plastically if the measures of the deviatoric stress tensors overcome the strengths of the materials. See, Pudasaini and Mergili (2019) for more details.
\\[3mm]
{\bf F. The $x$-directional enhanced non-Newtonian viscous stress} contribution (denoted by nN) for fine-solid due to the non-uniform distribution of the solid particles in the fine-solid is given by:
{\footnotesize
\begin{linenomath*}
\begin{align}\label{Non-Newtonian_fs}
\tau_{nN}^{{fs}^x} = \displaystyle{ \frac{\mathcal A^{fs,s}}{\alpha_{fs}}\left \{ 2\frac{\partial }{\partial x}\lb \nu_{fs}^e\frac{\partial \alpha_s}{\partial x}\lb u_{fs} - u_s\rb\rb
+ \frac{\partial }{\partial y}\lb \nu_{fs}^e\lb\frac{\partial \alpha_s}{\partial x}\lb v_{fs} -v_s\rb + \frac{\partial \alpha_s}{\partial y}\lb u_{fs} - u_s\rb\rb\rb\right\}}
-\displaystyle{ \frac{\mathcal A^{fs,s}}{\alpha_{fs}}\frac{\xi_s\alpha_s\nu_{fs}^e\lb u_{fs} -u_s\rb}{h^2}}.
\end{align}
\end{linenomath*}
}
Similarly, the enhanced non-Newtonian viscous stress contribution for fluid due to the non-uniform distribution of the fine-solid and solid particles in the fluid is given by:
{\footnotesize
\begin{linenomath*}
\begin{align}\label{Non-Newtonian_f}
\tau_{nN}^{f^x} = \displaystyle{ \frac{\mathcal A^{f,s}}{\alpha_f}\left \{ 2\frac{\partial }{\partial x}\lb \nu_f^e\frac{\partial \alpha_s}{\partial x}\lb u_f - u_s\rb\rb
+ \frac{\partial }{\partial y}\lb \nu_f^e\lb\frac{\partial \alpha_s}{\partial x}\lb v_f - v_s\rb + \frac{\partial \alpha_s}{\partial y}\lb u_f - u_s\rb\rb\rb\right\}}
-\frac{\mathcal A^{f,s}}{\alpha_f}\frac{\xi_s\alpha_s\nu_f^e\lb u_f -u_s\rb}{h^2}\nonumber\\[1mm]
+\displaystyle{ \frac{\mathcal A^{f,fs}}{\alpha_f}\left \{ 2\frac{\partial }{\partial x}\lb \nu_f^e\frac{\partial \alpha_{fs}}{\partial x}\lb u_f - u_{fs}\rb\rb
+ \frac{\partial }{\partial y}\lb \nu_f^e\lb\frac{\partial \alpha_{fs}}{\partial x}\lb v_f -v_{fs}\rb + \frac{\partial \alpha_{fs}}{\partial y}\lb u_f - u_{fs}\rb\rb\rb\right\}
-\frac{\mathcal A^{f,fs}}{\alpha_f}\frac{\xi_{fs}\alpha_{fs}\nu_f^e\lb u_f -u_{fs}\rb}{h^2}}.
\end{align}
\end{linenomath*}}
{\small
| 5625c89293e15a0593fc7a53dd076fd3876ecade | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In recent years, recommendation algorithms on social platforms have greatly enhanced confirmation bias by showing users content that is the most susceptible to match their interests --- the so-called \emph{filter bubble} effect \citep{pariser}. As a consequence, more and more isolated, tightly clustered online communities of similar-minded individuals, often referred to as \emph{echo chambers}, have arisen in various domains such as politics \citep{cota2019,delvicario2017,garimella2018}, healthcare \citep{allington2020,health_polarisation,monsted2022} or science \citep{williams2015}. Because of the so-called \emph{backfire effect}, presenting these users with opposing information might have the adverse effect of reinforcing their prior beliefs \citep{bail2018,schaewitz2020}. Finding ways to prevent such polarisation of opinion is a great challenge in the actual world. This paper is a step in this direction, as we propose ways of maximising the diversity of beliefs as well as the exposure to adverse views in a social group.
To this end we rely on the well-known voter model, in which each user holds one of two possible opinions (e.g.\ liberal of conservative, pro or anti-abortion) and updates it randomly under the distribution of others' beliefs. Independently introduced by \cite{clifford_sudbury} and \cite{holley1975} in the context of particles interaction, this model has since been used to describe in a simple and intuitive manner social dynamics where people are divided between two parties and form their opinion by observing that of others around them. We assume some of the users are stubborn and never change opinion. We call them \emph{zealots} as in \cite{mobilia2003,mobilia2007}. They can represent lobbyists, politicians or activists for example. Long time dynamics and limiting behaviour of such processes have been subject to several studies \citep{mobilia2007,mukhopadhyay2020,binary_opinion}.
To achieve our goal we propose equilibrium formulas for both the opinion diversity $\sigma$ and density of active links $\rho$, which is the proportion of connections that join opposite-minded users. The former is based on earlier results from \cite{masuda2015}. The latter uses a mean-field approximations and we show that it performs well when compared to numerical simulations. We then study the problems of maximising these quantities by turning \emph{free} (\emph{i.e.}\ non-zealous) users into zealots under the presence of a backfire effect. This effect we model by assuming that any increase in the number of zealots entails the \emph{radicalisation} of some non-zealous users, turning them into zealots with the opposite opinion. We provide exact solutions in the specific case of a complete, unweighted network for both the problems of $\sigma$ and maximising $\rho$. For $\sigma$ we also propose a method to optimise it in general networks.
Finally we apply our findings on a real-life dataset. Namely, we study the evolution of the composition of the US House of Representatives since 1947. We assimilate it to a realisation of the voter model and estimate the corresponding quantity of zealots based on empirical values of the equilibrium metrics $\sigma$ and $\rho$. We then solve our maximisation problems in this case and find that maximising $\rho$ by acting on Democrat zealots can help increase both $\rho$ and $\sigma$.
All code used is available online\footnote{\url{https://github.com/antoinevendeville/howopinionscrystallise}}.
\section{Related Literature}
Perhaps the earliest milestone in the study of opinion dynamics are the works from \cite{french} and \cite{degroot} who studied how a society of individuals may or may not come to an agreement on some given topic. Assuming the society is connected and people repeatedly update their belief by taking weighted averages of those of their neighbours, they showed that consensus is reached. That is, everyone eventually agrees. Various other models have been developed since, to tackle the question of under which circumstances and how fast a population is able to reach consensus. Amongst others, \cite{friedkin_johnsen} introduce immutable innate preferences, \cite{axelrod} studies the effect of homophily, \cite{word_of_mouth} assume individuals are perfectly rational and \cite{nbsl} account for the influence of external events.
The voter model was introduced independently by \cite{clifford_sudbury} and \cite{holley1975} in the context of particles interaction. They proved that consensus is reached on the infinite $\mathbb{Z}^d$ lattice. Several works have since looked at different network topologies, wondering whether consensus is reached, on which opinion and at what speed. Complete graphs \citep{yehuda2002,sood2008,perron2009,yildiz2010}, Erdös-Rényi random graphs \citep{sood2008,yildiz2010}, scale-free random graphs \citep{sood2008,fernley2019}, and other various structures \citep{sood2008,yildiz2010} have been addressed. Variants where nodes deterministically update to the most common opinion amongst their neighbours have also been studied \citep{chen2005,mossel2013}.
An interesting case to consider is the one where zealots -- \emph{i.e.}\ stubborn agents who always keep the same opinion, are present in the graph. Such agents may for example represent lobbyists, politicians or activists, \emph{i.e.}\ entities looking to lead rather than follow and who will not easily change side. One of those placed within the network can singlehandedly change the outcome of the process \citep{mobilia2003,sood2008}. If several of them are present on both sides, consensus is usually not reachable and instead opinions converge to a steady-state in which they fluctuate indefinitely \citep{mobilia2007,binary_opinion}.
Recently, \cite{mukhopadhyay2020} considered zealots with different degrees of zealotry and proved that time to reach consensus grows linearly with their number. They also showed that if one opinion is initially preferred --- \emph{i.e.}\ agents holding that opinion have a lesser probability of changing their mind --- consensus is reached on the preferred opinion with a probability that converges to 1 as the network size increases. \cite{klamser2017} studied the impact of zealots on a dynamically evolving graph, and showed that the two main factors shaping their influence are their degrees and the dynamical rewiring probabilities.
With the increasing importance of social networks in the political debate and information diffusion, there has been a recent surge in research aiming at controlling opinions, often with the goal to reduce polarisation. With the Friedkin-Johnsen model, \cite{goyal2019} provide algorithms for selecting an optimal sets of stubborn nodes in order to push opinions in a chosen direction. \cite{yi2019disagreement} formulate different constrained optimisation problems under the French-Degroot and the Friedkin-Johnsen models. They provide solutions in the form of optimal graph construction methods.
Still within the Friedkin-Johnsen paradigm, \cite{chitra2020} prove that dynamically nudging edge weights in the user graph can reduce polarisation while preserving relevance of the content shown by the recommendation algorithm. \cite{garimella2017bis} propose a method to reduce polarisation through addition of edges in the network. The focus is put on which nodes to connect in order to get the best reduction in polarisation, while being sure that the edge is ``accepted'' --- as extreme recommendations might not work because of the backfire effect. Finally, \cite{cen2020} propose a data-driven procedure to moderate the gap between opinions influenced by a neutral or a personalised newsfeed. Importantly, they show that this can be done even without knowledge of the process through which opinions are derived from the newsfeed.
Of particular interest to us, \citep{binary_opinion,masuda2015,moreno2021} study the voter model and propose strategies to find optimal sets of zealots in order to push opinions in a chosen direction. This work places itself in a similar vein but the objective is different, as we are trying to adjust the balance between both opinions rather than promoting one of the two.
\paragraph{Our contribution}
In a previous work \citep{vendeville2022} we studied the voter model with zealots in connected graphs with arbitrary degree distribution. Extending a result from the literature, we proved that the expected average opinion $\bar{x}^*$ of the population at equilibrium is given by the proportion of opinion 1 amongst zealots. Furthermore we solved the problem of controlling $\bar{x}^*$ via injection of zealots in the presence of a backfire effect.
In the present paper, we turn ourselves to the case where the network is weighted, directed and not necessarily connected. The vector $x^*$ of individual average opinions at equilibrium is then given by the solution of a linear system \cite[eq.~(4)]{masuda2015}. We adapt the problem of controlling $\bar{x}^*$ under backfire effect to that of controlling diversity, that we define as a function of $\bar{x}^*$. We show that it can be solved efficiently by gradient descent. This approach however does not guarantee the existence of a dialogue between users, as even with $\bar{x}^*\approx 1/2$ the network might clusterise into hermetic echo chambers with opposite opinions. Thus we suggest a novel, alternative approach for the diversification of opinion in social networks. Instead of controlling the average opinion, we propose to control the density of active links -- \emph{i.e.}\ the proportion of edges that connect users with different opinions.
With the profusion of theoretical works and models on opinion dynamics in recent years, the need for real data validation has got more and more pressing. An attempt to fit the voter model with election results in the UK and in the US was the object of a previous publication from us \citep{vendeville2020}. In the present paper we illustrate our findings as we apply the developed methods to the evolving network of the House of Representatives in the United States. With the gradual disappearance of independent members and the fading of cross-party agreement, it has become a prime example of a polarised network divided in two antagonistic camps \citep{andris2015}. We find that the network exhibits high levels of opinion diversity but lower levels of active links density. In this context we solve the optimisation problems developed in the theoretical sections, providing optimal numbers of zealots that maximise either $\sigma$ or $\rho$. We find that both can be increased in some cases.
\section{The Voter Model with Zealots} \label{voter_model}
In the traditional voter model, users are placed on the $\mathbb{Z}^d$ lattice and hold individual opinions in $\{0,1\}$. Given an initial distribution of opinion, each user updates their opinion at the times of an independent Poisson process of parameter 1 by copying a neighbour chosen uniformly at random. Letting $x_i(t)$ denote the opinion of user $i$ at time $t$, we say that consensus is reached if almost surely all users eventually agree, \emph{i.e.}\ if
\begin{equation} \label{consensus}
\forall i,j, \quad \mathbb{P} \left(x_i(t) = x_j(t) \right) \underset{t \rightarrow \infty}{\longrightarrow} 1.
\end{equation}
On any finite connected network, consensus is reached \citep{aldous_fill_2014}. Intuitively, no matter the current number of opinion-0 and opinion-1 users, there exists a succession of individual opinion changes with strictly positive probability that results in everyone holding the same opinion.
It is might however seem unrealistic to imagine that all people in a group are willing to change opinions. An interesting extension of the traditional voter model is to include stubborn agents who never change their opinions, often referred to as \emph{zealots} \citep{mobilia2007,binary_opinion}. They form an inflexible core of partisans who bear great power of persuasion over the population. If all zealots defend the same opinion then via similar arguments as for eq.~(\ref{consensus}) this opinion is eventually adopted by all. When both camps count such agents within their ranks however, there always exists a strictly positive number of users with each opinion. This prevents consensus and instead the system reaches state of equilibrium in which it fluctuates indefinitely.
\paragraph{Framework}
Although we will consider a complete and unweighted user graph in our application, most of the analysis is presented in the general case of a directed, weighted network. We assume there are $N$ users among which $z$ are zealots. The remaining $F:=N-z$ users are referred to as \emph{free}. The set of free users is denoted by $\mathcal{F}$, the set of zealots with opinion 0 by $\mathcal{Z}_0$, the set of zealots with opinion 1 by $\mathcal{Z}_1$ and the set of all zealots by $\mathcal{Z}:=\mathcal{Z}_0\cup\mathcal{Z}_1$.
For any pair $(i,j)$ of users we let $w_{ij}\ge 0$ be the weight of the directed edge $j\rightarrow i$, representing the power of influence that $j$ has over $i$. If $i\in\mathcal{Z}$ we set $w_{ij}=0$ for all $j$. We do not assume uniform choice anymore and when updating their opinion, $i$ will copy $j$ with probability proportional to $w_{ij}$. We assume $w_{ii}$ to be zero, meaning users cannot choose to copy themselves -- this assumption may be relaxed in the future and we expect the results presented here to hold.
Opinions thus evolve as follows. Assume $i\in\mathcal{F}$ updates their opinion at time $t$ when the vector of opinions is $x(t)$. Then $i$ will adopt opinion 1 with probability $d_i^{-1}\sum_{j=1}^N w_{ij}x_j(t)$ where $d_i$ is the total influence exerted on them, defined by $d_i = \sum_{j=1}^N w_{ij}$. This quantity can be seen as the in-degree of node $i$. Zealots do not update their opinions and receive no external influence, thus they have in-degree 0.
Finally we let $z_0$ and $z_1$ be the $F$-dimensional vectors of zealot influence over free users, where $z_{0,i}=\sum_{j\in\mathcal{Z}_0} w_{ij}$ is the total influence exerted by all zealots with opinion 0 onto user $i\in\mathcal{F}$. The definition of $z_1$ is analog. The in-degree of a free node $i$ can then be written as $d_i = \sum_{j\in\mathcal{F}} w_{ij} +z_{0,i} +z_{1,i}$.
\section{Control of Opinion Diversity} \label{opinion_diversity_section}
We define the average diversity of opinion at equilibrium by
\begin{equation}
\sigma = 4\bar{x}^*(1-\bar{x}^*)
\end{equation}
where $\bar{x}^*$ is the average opinion over all users at equilibrium, \emph{i.e.}\ the expected result when punctually observing the opinion of a random node. It is also the average share of opinion 1 within the network and often referred to as \emph{magnetisation} in the literature.
$\sigma$ is the variance of the Bernoulli distribution of parameter $\bar{x}^*$, scaled by 4 so that it ranges in $[0,1]$. It describes the diversity of the system in that it is maximal when users are equally divided between both opinions ($\bar{x}^*=1/2$), and minimal when only one opinion is represented ($\bar{x}^*=0$ or $1$).
\subsection{Maximisation in General Networks}
Let $L=(L_{ij})_{i,j\in\mathcal{F}}$ be the Laplacian of the \emph{free} graph, \emph{i.e.}\ the $F\times F$ matrix with elements $L_{ij} = \delta_{ij}\sum_{k\in\mathcal{F}} w_{ik} - (1-\delta_{ij}) w_{ij}$ where $\delta_{ij}$ is the Kronecker delta. \cite{masuda2015} showed that the average opinion amongst free users is:
\begin{equation}
\bar{x}^*_f=\frac{1}{F}\mathbf{1}^\top[L + \text{diag}(z_0+z_1)]^{-1}z_1,
\end{equation}
where $\mathbf{1}$ is the $N$-dimensional vector filled with ones. The $i^{\text{th}}$ entry of the vector $x_f^*:=[L + \text{diag}(z_0+z_1)]^{-1}z_1$ is the average opinion of $i$ at equilibrium, given by
\begin{equation} \label{xfi}
x_{f,i}^* = \frac{\sum_{j\in\mathcal{F}}w_{ij}x_{f,j}^*+z_{1,i}}{d_i}.
\end{equation}
Finally we have
\begin{equation}
\bar{x}^* = \frac{F \bar{x}^*_f + \vert\mathcal{Z}_1\vert}{N}.
\end{equation}
Now consider a network where the set $\mathcal{Z}_0$ of 0-zealots and their influence vector $z_0$ is fixed. Given a predetermined quantity $\mathcal{Z}_1$ of 1-zealots, how should we set the values of $z_1$ to maximise the opinion diversity at equilibrium? Formally, we seek to solve
\begin{align} \label{P}
\underset{z_1\ge0}{\text{argmax}} \quad &\sigma. \tag{P}
\end{align}
Recall that the objective is function of $\bar{x}^*$ which is itself function of $z_1$. Because $\bar{x}^*$ is increasing with $\Vert z_1\Vert$ and equals zero when $\Vert z_1\Vert=0$, there exists at least one optimal vector $z_1^\star$ for which $\bar{x}^*=1/2$ and thus (\ref{P1}) is solved. This optimal vector can be found efficiently using gradient ascent on $\sigma$, as \cite{moreno2021} show that $\bar{x}^*_f$ (and thus $\bar{x}^*$) is concave with respect to $z_1$.
\subsection{Maximisation in Complete Networks}
In our application, we will consider a complete, unweighted network: $w_{ij}=\mathds{1}_{i\notin\mathcal{Z}}$ for all $i$. In that case all entries of $z_0$ are equal to the same value, which is the amount of 0-zealots nodes within the graph. For the sake of simplicity we denote this unique value by $z_0$. We proceed similarly for $z_1$. In that case it is known \citep{mobilia2007,masuda2015} that
\begin{equation} \label{xbars_complete}
\bar{x}^* = \bar{x}^*_f = x_{f,1}^* = \ldots = x_{f,{F}}^* = \frac{z_1}{z_0+z_1}.
\end{equation}
In a previous work \citep{vendeville2022} we proved that this result also holds on expectation for any connected, unweighted graph where the position of zealots is drawn uniformly at random. Hence the following.
\begin{theorem} \label{rho_complete}
In a complete unweighted user graph with $z_0$ zealots with opinion 0 and $z_1$ zealots with opinion 1,
\begin{equation}
\sigma = \frac{4z_0z_1}{(z_0+z_1)^2}.
\end{equation}
This quantity is trivially maximal when $z_0=z_1$.
\end{theorem}
In the same paper we studied the following problem: given a quantity $z_0$ of 0-zealots and a target diversity $\lambda$, what is the optimal number $z_1^\star$ of free users that should be turned into 1-zealots in order for $\bar{x}^*$ to be as close to $\lambda$ as possible? This is a generalisation of the diversity maximisation problem presented above which corresponds to the specific case $\lambda=1/2$.
Numerous empirical studies have found that rather incentivising a change in opinion, presenting certain people with opposing views might actually entrench them even deeper in their beliefs. This is often referred to as the \emph{backfire effect}. To account for this phenomenon, we assumed that the creation of $z_1$ zealots with opinion 1 will \emph{radicalise} a quantity $\alpha z_1$ of free users, who will then become 0-zealots. Thus necessarily $z_0+(1+\alpha)z_1\le N$ and the constraint $z_1\le(N-z_0)/(1+\alpha)$. The real parameter $\alpha \in [0,1)$ quantifies the intensity of the backfire effect.
We quickly summarise our findings as they will be useful to us. We also take this opportunity to correct a small mistake that was found in the paper.
\begin{theorem}
Assume there are $z_0$ zealots with opinion 0 into the system, and there exists a backfire effect of intensity $\alpha$. Set $z_1^{\textnormal{max}}:=(N-z_0)/(1+\alpha)$. After turning $z_1$ free users into zealots with opinion 1, $z_0$ is updated to $z_0+\alpha z_1$ and the average equilibrium opinion is
\begin{equation} \label{xbars_complete_formula}
\bar{x}^* = \frac{z_1}{z_0+(1+\alpha)z_1}.
\end{equation}
The solution to the problem
\begin{align} \label{P1}
\underset{0\le z_1\le z_1^\textnormal{max}}{\textnormal{argmin}} \quad &(\bar{x}^*-\lambda)^2 \tag{P1}
\end{align}
is given by
\begin{equation} \label{optim_backfire}
\begin{cases}
z_1^\star = \textnormal{min} \left(z_1^{\textnormal{max}}, \lambda z_0d^{-1} \right) &\text{ if } d>0,\\
z_1^\star = z_1^{\textnormal{max}} &\text{ if } d\leq0,
\end{cases}
\end{equation}
where $d:=1-\lambda-\alpha \lambda$. Discarding the constraint $z_1\le z_1^\textnormal{max}$ results in an unbounded problem if $d\le 0$ and $z_1^\star=\lambda z_0d^{-1}$ if $d>0$.
\end{theorem}
In our previous work we had mistakenly claimed that $z_0$ was updated to $z_0+\alpha z_0z_1$, leading to $\bar{x}^*=z_1/(z_1+(1+\alpha z_1)z_0)$ and $d=1-\lambda-\alpha \lambda z_0$. All results presented then do still hold qualitatively. Finally note that in general $z_1^\star$ will not be an integer so that in practical cases it would need to be rounded.
In the present context we restrict ourselves to $\lambda=1/2$. The objective function $(\bar{x}^*-1/2)^2$ is equal to $\frac{1}{4}-\sigma$ and the condition $d>0$ becomes $\alpha<1$ which is true by definition. Hence the following theorem.
\begin{theorem}
The problem of maximising diversity in a complete unweighted graph with $z_0$ 0-zealots and backfire effect $\alpha$ is formally written as
\begin{equation}\label{P2}\tag{P2}
\begin{aligned}
\underset{0\le z_1\le z_1^\textnormal{max}}{\textnormal{argmax}} \quad &\sigma_{z_0,\alpha}(z_1) \\
\textnormal{s.t.} \quad &\sigma_{z_0,\alpha}(z_1) = \frac{4(z_0+\alpha z_1)z_1}{(z_0+(1+\alpha)z_1)^2}
\end{aligned}
\end{equation}
where $z_1^{\textnormal{max}}=(N-z_0)/(1+\alpha)$. Its solution is given by
\begin{equation}
z_1^\star = \textnormal{min} \left(z_1^{\textnormal{max}}, \frac{z_0}{1-\alpha} \right).
\end{equation}
\end{theorem}
\section{Density of Active Links at Equilibrium} \label{active_links_section}
Because maximising diversity does not guarantee that the network will not clusterise into echo chambers, we are also interested in maximising the proportion of \emph{active links} at equilibrium. A link is said to be active if it joins two users with opposite opinions. We derive a mean-field approximation for this quantity that shows a tight fit with empirical averages obtained via numerical simulations. Let us first precise what we mean by \emph{active links} and their average density.
We denote by $\mathcal{E}'$ the set of all edges present in the graph that join two users, one of them at least being free. We write $(i,j)$ to designate the edge $j\rightarrow i$ pointing outwards from $j$ and towards $i$. Because the graph is oriented, $(i,j)$ and $(j,i)$ are two separate objects and one might exist without the other. Moreover when both are present in the graph we do not necessarily have $w_{ij}=w_{ji}$.
\begin{definition}[Active link]
At any time $t$ the directed link $(i,j)\in\mathcal{E}'$ is said to be \emph{active} if $x_i(t)\neq x_j(t)$, and inactive otherwise.
\end{definition}
\begin{definition}[Average density of active links]
Let $q_{ij}$ be the equilibrium probability of the event $\{x_i\neq x_j\}$. We define the average density of active links at equilibrium by
\begin{equation} \label{rho_def}
\rho = \frac{1}{\vert\mathcal{E}'\vert} \sum_{(i,j)\in\mathcal{E}'} q_{ij}
\end{equation}
and its weighted version by
\begin{equation} \label{rhow_def}
\rho_w = \frac{\sum_{(i,j)\in\mathcal{E}'} w_{ij}q_{ij}}{\sum_{(i,j)\in\mathcal{E}'} w_{ij}}.
\end{equation}
\end{definition}
In the weighted case, heavier edges count more towards the average and lighter ones count less. Note that $q$ is a non-oriented metric in that $q_{ij}=q_{ji}$. When $i$ and $j$ are both free users, $q_{ij}$ will be counted twice in the sum above if both $w_{ij}$ and $w_{ji}$ are positive, once if only one of them is and not counted at all if they are not connected with each other. Because zealots receive no influence from others, if $i$ is in $\mathcal{Z}$ then $w_{ij}=0$ and $q_{ij}$ is counted once if $w_{ji}>0$, not counted otherwise. If $j$ is a zealot as well then $q_{ij}$ is not counted in the sum.
The following theorem is the main theoretical contribution of this paper.
\begin{theorem} \label{rhoij_theorem}
A mean-field approximation for the values $q_{ij}$ is given by the solution of the following linear system:
\begin{align} \label{rhoij}
q_{ij}(d_i+d_j) & -\sum_{k\in\mathcal{F}\backslash\{i,j\}} (w_{ik}q_{jk}+w_{jk}q_{ik})
= \tilde{z}_j x^*_i + \tilde{z}_i x^*_j + z_{1,i} + z_{1,j},
\end{align}
where $(i,j)$ describes $\mathcal{E}'$, $\tilde{z}_k := z_{0,k}-z_{1,k}$, $d_k=\sum_{l=1}^N w_{kl}$ is the in-degree of node $k$, $x_k^*:=x_{f,k}^*$ for $k\in\mathcal{F}$, $x_k^*:=0$ for $k\in\mathcal{Z}_0$ and $x_k^*:=1$ for $k\in\mathcal{Z}_1$.
\end{theorem}
The proof can be found in \Cref{proof_section}. Consistent with intuition, if $j\in\mathcal{Z}_0$ (resp.\ $j\in\mathcal{Z}_1$) then the above yields $q_{ij}=\bar{x}^*_{f,i}$ (resp.\ $q_{ij}=1-\bar{x}^*_{f,i}$), which is simply the probability for $i$ to hold opinion 1 (resp.\ opinion 0).
\subsection{Numerical Validation}
We validate the above through a series of numerical simulations. Let us place ourselves in an Erdös-Rényi random graph with $N=100$ users, $\vert\mathcal{Z}_0\vert=23$ zealots with opinion 0 and $\vert\mathcal{Z}_1\vert=18$ zealots with opinion 1. The graph is directed and we set its density to $0.1$ so that about $10\%$ of all possible edges are present. Each edge is then attributed a weight generated uniformly at random between 0 and 1.
We perform a single simulation of the voter model on this graph for 50,000 time units. The empirical density of active links $\hat\rho$ is computed every 100 updates, starting once 10,000 time units have passed to ensure that the system has had time to stabilise. In \Cref{rho_simu_vs_theo} \emph{(top left)} we plot $\hat\rho$ over the last 1,000 time units against $\rho$. Averaging $\hat\rho$ over time yields our final empirical estimate. We proceed similarly for $\rho_w$ \emph{(top right)} and do it all as well in a Barabasi-Albert graph with weights generated under an exponential distribution of parameter 1 \emph{(bottom left and right)}. This graph has density $\approx 0.1$ as well.
In all cases the theoretical values $\rho$ and $\rho_w$ are roughly the same. Thus for the same density of edges, the topology of the graph does not seem to play an very important role here. We obtain rather small errors between theory and simulation, in the order of $10^{-4}$ for the Erdös-Rényi graph and $10^{-3}$ for the Barabasi-Albert graph. Other experiments with different quantities of zealots, weights distribution and graph topologies have shown similarly small errors as well, confirming that our mean-field approximation of $q$ performs well in practice. The error for the Barabasi-Albert network being higher is not too surprising, as the second graph is inherently less regular and the variance in the weights is higher (exponential distribution of parameter 1 against uniform distribution over $[0,1]$). This can also be seen in the oscillations of $\hat\rho$ and $\hat\rho_w$ over time, which demonstrate more variability by spanning a larger range in this case.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=\textwidth]{ER_weightedFalse_plot}
\end{subfigure}~
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=\textwidth]{ER_weightedTrue_plot}
\end{subfigure}\\
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=\textwidth]{BA_weightedFalse_plot}
\end{subfigure}~
\begin{subfigure}[b]{.5\textwidth}
\includegraphics[width=\textwidth]{BA_weightedTrue_plot}
\end{subfigure}
\caption{Verifying \Cref{rhoij_theorem} in simulation. $N=100$ users, simulation time 50,000. Last 1,000 time units are plotted. \textbf{Top:} Erdös-Rényi graph with random uniform weights. \textbf{Bottom:} Barabasi-Albert graph with random exponential weights. \textbf{Left:} theoretical density of active links as per eq.~(\ref{rho_def}) (dotted red lines) and empirical value over time for a single simulation (blue oscillations). \textbf{Right:} same with weighted density of active links (\ref{rhow_def}). In each plot we also indicate the average over the whole simulation $(\hat\rho$ or $\hat\rho_w)$ as well as the theoretical value $(\rho$ or $\rho_w)$, both rounded to $10^{-4}$.}
\label{rho_simu_vs_theo}
\end{figure}
\subsection{Maximisation in Complete Networks}
Now consider a complete unweighted graph. Again we let $z_0$ denote the amount of 0-zealots and $z_1$ the amount of 1-zealots in the network. Becomes of the asymmetry between free users and zealots it is convenient to still assume edges to be directed. Then the following holds.
\begin{theorem} \label{rho_complete}
In a complete unweighted user graph with $z_0$ zealots with opinion 0 and $z_1$ zealots with opinion 1,
\begin{equation} \label{rhoc_eq}
\rho = \frac{2z_0z_1(N-z_0-z_1)}{(N-1)(z_0+z_1)(z_0+z_1+1)}.
\end{equation}
\end{theorem}
The proof can be found in \Cref{proof_section}. Now we are interested in finding the optimal number $z_1^\star$ of free users that should be turned into 1-zealots in order to maximise $\rho$ (which is equal to $\rho_w$ as the graph is weighted). Because of the backfire effect $\alpha$, creating $z_1$ zealots with opinion 1 with entail a change in the number of 0-zealots, from $z_0$ to $z_0+\alpha z_1$.
\begin{theorem} \label{optim_rho_complete_theo}
The problem of maximising the density of active links in a complete unweighted graph with $z_0$ 0-zealots and backfire effect $\alpha$ is formally written as
\begin{equation} \label{P3} \tag{P3}
\begin{aligned}
\underset{0\le z_1\lez_1^{\textnormal{max}}}{\textnormal{argmax}} \quad &\rho_{z_0,\alpha}(z_1) \\
\textnormal{s.t.} \quad &\rho_{z_0,\alpha}(z_1) = \frac{2(z_0+\alpha z_1)z_1}{(z_0+(1+\alpha)z_1)(z_0+(1+\alpha)z_1+1)}
\end{aligned}
\end{equation}
where $z_1^{\textnormal{max}}=(N-z_0)/(1+\alpha)$. It has at least one solution which is either $z_1^{\textnormal{max}}$ or a real positive root of the derivative.
\end{theorem}
This results directly from the fact that the objective function is continue, one-dimensional, and that the optimum cannot be reached for $z_1=0$ as
\begin{equation}
\rho_{z_0,\alpha}(z_1) >0 = \rho_{z_0,\alpha}(0)
\end{equation}
for all $z_1>0$.
\section{Application to US Congress Data}
We evaluate our results on real-life data from American politics. The Voteview dataset \citep{voteview} contains very detailed information about the United States congress since its inception in 1789. Members of each Senate and House of Representatives are listed with their affiliations and what they voted in each rollcall. We discard voting data and simply focus on the composition of the House of Representatives since 1947. During this time, the proportion represented by Democrats and Republicans therein is always superior to 99.5\%, which justifies a binary approach such as the voter model.
Let $D_k,R_k$ be the respective amounts of Democrat and Republican representatives in the House during the $k^\text{th}$ congress. In 1947 started the $80^\text{th}$ congress and in 2021 the $117^\text{th}$ one so that $k$ would range in $\{80,\ldots,117\}$. For the sake of simplicity however we shift the indices by 79 and let $k\in\{1,\ldots,38\}$. We discard members of other parties from our analysis---they never represent more that 0.5\% of the House. We find ourselves with two vectors $D,R$ of length $K=38$ and assume they correspond to punctual observations of a single realisation of the voter model with zealots on a complete, unweighted graph.
Because members of the House change between congresses, users cannot represent \emph{persons} here. Rather, they represent \emph{seats} of the House, and their \emph{opinion} is the \emph{party} to which the representative occupying it is affiliated. In an effort for clarity and consistency we employ the traditional words of users and opinions, but it is important to keep in mind what they precisely mean here.
In each congress there is a small number of non-voting delegates. They are included in our analysis but their exact number may vary so that the total number of seats $N_k$ is not always the same. In the congresses considered here, and with non-Democrat, non-Republican members discarded, $N_k$ varies between 438 (3 non-voting delegates) and 453 (18 non-voting delegates).
\subsection{Parameters Estimation}
The quantity of zealots on each side is unkown to us. They represent the quantity of seats that are ``locked'' by each party and we denote these by $z_D$ (Democrat zealots) and $z_R$ (Republican zealots). To infer their number, one could for example consider as zealots users who most often agree with others from their party. This would require however the choice of aheuristic threshold to define what counts as ``most''. Moreover, we do not have the vote of every member for each rollcall, entailing the addition of uncertainty and potential errors. Rather, we choose to infer $z_D$ and $z_R$ from the two equilibrium metrics presented in sections \ref{opinion_diversity_section} and \ref{active_links_section}: $\sigma$ and $\rho$.
We propose the following estimate $\hat z=(\hat z_D, \hat z_R)$ for $(z_D,z_R)$:
\begin{equation} \label{zhat} \tag{Q}
\begin{aligned}
\hat z = \underset{(z_D,z_R)\in Z}{\text{argmin}} \quad &\epsilon \\
\text{s.t.} \quad &\epsilon=\frac{\vert\hat\sigma-\sigma\vert + \vert\hat\rho-\rho\vert}{2},
\end{aligned}
\end{equation}
where $\hat\sigma$ is an empirical estimate of $\sigma$ and $\hat\rho$ is an empirical estimate of $\rho$. Thus $\hat z$ is a minimiser of the mean distance between theoretical and empirical values of our metrics. The set $Z:=\{1,\ldots,D_\text{min}\}\times\{1,\ldots,R_\text{min}\}$ constrains the number of zealots in each party to never be higher than the quantity of members this party has in the House.
The empirical diversity of opinions is directly derived as
\begin{equation}
\hat\sigma := \frac{4}{K}\sum_{k=1}^K \frac{D_kR_k}{(D_k+R_k)^2}
\end{equation}
where $K=38$ is the total number of observations. The empirical estimation of $\rho$ is a bit trickier, as links are considered differently whether they join two free users, two zealots or a free user and a zealot. Hence we need be aware of what nodes are free and which ones are zealous. The empirical density of active links is given by
\begin{equation}
\hat\rho := \frac{1}{K}\sum_{k=1}^K \frac{2D_kR_k-D_kz_R-R_kz_D}{N_k(N_k-1)}.
\end{equation}
$N_k$ is the number of nodes in the $k^{th}$ observation, so that the denominator is the total number of links in the corresponding graph. The numerator of the fraction was simplified from
\begin{equation}
2(D_k-z_D)(R_k-z_R)+(D_k-z_D)z_R+(R_k-z_R)z_D
\end{equation}
where the first term is the number of links between all free nodes, the second term the number of links between free Democrats and zealous Republicans, and the third term the number of links between free Republicans and zealous Democrats.
The cardinal of $Z$ is small enough that (\ref{zhat}) can be solved by performing an exhaustive search over the whole set. We find:
\begin{align*}
(D_\text{min},R_\text{min}) &= (190,143), \\
(\hat z_D, \hat z_R) &= (89, 63), \\
(\hat\sigma,\hat\rho) &\simeq (0.97, 0.32), \\
\epsilon &\simeq 3.8\cdot 10^{-5}.
\end{align*}
The small error $\epsilon$ guarantees that the optimisation was efffective. The diversity is close to the theoretical optimal, indicating that the number of members from both parties is balanced over time. The majority switches back and forth between Democrats and Republicans but neither seem to truly have an upper hand over time. The density of active links is fairly high but could be better, as it pales in comparison to the values we obtained during simulations on synthetic networks (\Cref{rho_simu_vs_theo}).
\subsection{Optimising Diversity and Activity}
In this section, we investigate how $\sigma$ and $\rho$ could be optimised by changing the number of zealots. To this extent, we solve problems (\ref{P2}) and (\ref{P3}) for intensities of the backfire effect $\alpha$ spanning the whole range $[0,1)$. In a first time we consider that Democrats correspond to opinion 0, Republicans to opinion 1 and we optimise on $z_D$ keeping $z_R=\hat z_R$ fixed. Doing this we obtain a optimal number $z_D^\star$ of zealous Democrats when the number of zealous Republicans is given by the one inferred from the data. In a second time we do the opposite and optimise on $z_R$ keeping $z_D=\hat z_D$ fixed, obtaining a maximiser $z_R^\star$.
The two plots on the left part of \Cref{data_optim} describe the solution of (\ref{P2}) function of $\alpha$. The upper plot contains the values of the maximiser $z_D^\star$, its upper-bound $z_D^\text{max}$ and the empirical number of zealots $\hat z_D$ inferred from the data. Similarly for the Republican party with $z_R^\star$, $z_R^\text{max}$ and $\hat z_R$. The bottom plot contains the optimal values $\sigma(z_D^\star)$ and $\sigma(z_R^\star)$ of the objective, when optimising respectively on $z_D$ and on $z_R$. Its empirical value inferred from data $\hat\sigma$ is also represented. In addition we also show $\rho(z_D^\star)$, $\rho(z_R^\star)$ and $\hat\rho$. This is to assess what effect optimising the diversity has on the active link density.
We observe that for values of $\alpha$ up to about $0.7$, $\sigma$ can be fully optimised to its maximum value 1 whether we act on Democrat or Republican zealots. Even for the highest values of $\alpha$ it is very close to the empirical value. The impact of optimising $\sigma$ on $\rho$ is rather negative, as for almost all $\alpha$ the resulting density of active links is lower than its empirical value. Looking at the upper plot, we see that the maximisers are always above empirical values, meaning zealots would have to be added in order to improve diversity. Starting from $\alpha\simeq 0.6$ (Republicans) and $\alpha\simeq 0.7$ (Democrats), the maximiser is equal to its maximum possible value so that the system is saturated, with all nodes being zealots.
We now turn ourselves to the problem of maximising the density of active links (\ref{P3}). Its solution is described by the two plots on the right part of \Cref{data_optim}. Unlike before, here the optimal number of zealots decreases with $\alpha$, is constantly lower than its empirical estimate and never approaches its upper bound. Morevoer this time, optimising $\rho$ can also entail an improvement on $\sigma$. This happens when the number of Democrat zealots is acted upon, and for $\alpha\le 0.5$ roughly.
Finally, we remark that acting upon Democratic zealots is always more effective than upon Republican ones. This might stem from the fact that the former are in superior number from our empirical estimation, leaving more room to act efficiently on the objective functions.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{\textwidth}
\includegraphics[width=\textwidth]{data_argoptim}
\end{subfigure}\\
\vspace{.2cm}
\begin{subfigure}[b]{\textwidth}
\includegraphics[width=\textwidth]{data_optim}
\end{subfigure}
\caption{Optimal opinion diversity \textbf{(left)} and active links density \textbf{(right)} function of the backfire effect $\alpha$. \textbf{Top:} maximisers compared with empirical estimates and maximum possible values. \textbf{Bottom:} objective values and impact on the other metric, compared with empiriucal estimates.}
\label{data_optim}
\end{figure}
\section{Proofs} \label{proof_section}
\begin{proof}[Proof of \Cref{rhoij_theorem}.]
Let $\lambda_{ij}$ be the average rate at which user $i$ adopts the same opinion as $j$ while in equilibrium. Remember that each user updates their opinion at the times of an exponential clock of parameter 1. There are four different events that lead to $i$ adopting $j$'s opinion, described below with the associated frequency rates.
\begin{itemize}
\item $i$ may copy $j$ directly, which happens at rate $d_i^{-1}w_{ij}$, or
\item $i$ may copy a third free user $k$ holding the same opinion as $j$, which happens at rate $d_i^{-1}\sum_{k\in\mathcal{F}\backslash\{i,j\}} w_{ik}(1-q_{jk})$, or
\item $i$ may copy a 1-zealot while $j$ has opinion 1, which happens at rate $d_i^{-1}z_{1,i}x^*_j$,
\item $i$ may copy a 0-zealot while $j$ has opinion 0, which happens at rate $d_i^{-1}z_{0,i}(1-x^*_j)$.
\end{itemize}
By using $q_{jk}$ and $x^*_j$ we made the mean-field assumption that $i$ interacts with the average system at equilibrium rather than with its exact state. Through comparison with simulations we will show that this approximation performs well numerically. Putting it all together,
\begin{align}
\lambda_{ij} &= d_i^{-1} \left(w_{ij} + \sum_{k\in\mathcal{F}\backslash\{i,j\}} w_{ik}(1-q_{jk}) + z_{1,i}x^*_j + z_{0,i}(1-x^*_j)\right).
\intertext{Via an analogous reasoning, at equilibrium $i$ adopts the opinion opposite of $j$'s with rate}
\mu_{ij} &= d_i^{-1} \left(\sum_{k\in\mathcal{F}\backslash\{i,j\}} w_{ik}q_{jk} + z_{1,i}(1-x^*_j) + z_{0,i}x^*_j \right).
\end{align}
We obtain $\lambda_{ji}$ and $\mu_{ji}$ in a similar fashion. The discrete quantity $\mathds{1}_{x_i\neq x_j}$ describes a continuous-time Markov chain with two states 0 and 1, transitioning from 0 to 1 with rate $\mu_{ij}+\mu_{ji}$ and from 1 to 0 with rate $\lambda_{ij}+\lambda_{ji}$. The stationary probability of state 1 is exactly $q_{ij}$, so that
\begin{equation}
q_{ij} = \frac{\mu_{ij}+\mu_{ji}}{\lambda_{ij}+\mu_{ij}+\lambda_{ji}+\mu_{ji}}.
\end{equation}
After simplifications we obtain eq.~(\ref{rhoij}).
\end{proof}
\begin{proof}[Proof of \Cref{rho_complete}.]
Let $(i,j)\in\mathcal{E}'$. First if $i\in\mathcal{F}$ is free and $j\in\mathcal{Z}_0$, then as discussed before eq.~(\ref{rhoij}) immediately yields $q_{ij}=\bar{x}^*$. If $j\in\mathcal{Z}_1$ then $q_{ij}=1-\bar{x}^*$ and in both cases, $q_{ji}=0$.
Now assume $i,j\in\mathcal{F}$. Because all free nodes are topologically equivalent, they share the same value $q_f$ for $q$---just as they have the same average opinion $\bar{x}^*=z_1/(z_0+z_1)$, \emph{c.f.}\ (\ref{xbars_complete}). Replacing edge weights with 1, in-degrees with $N-1$ and zealots influence $z_{0,k},z_{1,k}$ with $z_0,z_1$, equation~(\ref{rhoij}) becomes
\begin{equation}
2q_f(N-1)-(F-2)2q_f = 2(z_0-z_1)\bar{x}^* + 2z_1
\end{equation}
and after simplifications
\begin{equation}
q_f = \frac{2z_0z_1}{(z_0+z_1)(z_0+z_1+1)}.
\end{equation}
Because there are $F(F-1)$ directed edges between zealots, $Fz_0$ edges between free users and 0-zealots and $Fz_1$ edges between free users and 1-zealots, the total \textbf{number} (and not density) of active links is given by
\begin{align}
n_{\text{active}} &= F(F-1)q_f + Fz_0\bar{x}^* + Fz_1(1-\bar{x}^*).
\intertext{Replacing $F$ by $N-z_0-z_1$ and $q_f,\bar{x}^*$ by their respective values we find}
n_{\text{active}} &= \frac{2z_0z_1N(N-z_0-z_1)}{(z_0+z_1)(z_0+z_1+1)}.
\end{align}
Finally there are $N(N-1)$ directed edges in the complete graph so that we immediately obtain (\ref{rhoc_eq}) via $\rho=n_{\text{active}}/(N(N-1))$.
\end{proof
\section{Conclusion and Future Work} \label{futurework}
In this paper we analysed the voter model with zealots on directed, weighted networks. We proposed formulas for the opinion diversity ($\sigma$) as well as for the density of active links ($\rho$) at equilibrium. The latter relied on a mean-field approximation that we showed performs well against numerical simulations. For both metrics we studied the problem of maximising it by turning free (\emph{i.e.}\ non-zealous) users into zealots in the presence of a backfire effect. We provided explicit solutions for the specific case of a complete unweighted network, and for opinion diversity we also exposed how it could be maximised in the general directed, weighted case.
As an example application, we applied our findings to a dataset detailing the evolution of members in the US House of Representatives since 1947. Assuming the data was a realisation of the voter model with zealots, we estimated the number of zealots by minimising the distance between empirical and theoretical values of the equilibrium metrics $\sigma$ and $\rho$. The opinion diversity was found to be almost maximal, indicating a balanced mix of Democrats and Republicans. We then used the optimisation problems exposed in the theoretical sections to find optimal quantities of zealots users maximising $\sigma$ and $\rho$. Of note, we found that maximising $\rho$ by acting on Democrat zealots can help increase both $\rho$ and $\sigma$.
There are many open leads for further investigation. First, we considered multiple congresses at once. It could be interesting to subdivise in several windows, separated by impactful historical moments (fuel crisis in the seventies, end of the USSR in the early nineties, etc.). There might be patterns inherent to specific periods that are not apparent in our analysis. Data from online social networks could provide interesting examples of polarised systems.
On the theoretical side, an efficient algorithm for the optimisation of $\rho$ on directed, weighted networks could help study more refined data. In the case of social media data, it could also be a good idea to optimise not on the number of zealots but on the edge weights. This would mean standing from the point of view of a platform administrator, trying to update its recommendation algorithm in order to improve opinion diversity or active links density. Such approaches have been tried in other models of opinion dynamics \citep{chitra2020,santos2021}. Finally because we are studying polarised systems, incorporating signed edges in the model could also yield more informative results \citep{keuchenius2021}.
\justify{
\subsubsection*{Data Availability}
The data used in the application is taken from \cite{voteview}. All code used and simulation data are available online at \url{https://github.com/antoinevendeville/howopinionscrystallise}.}
\justify{
\subsubsection*{Acknowledgements}
The authors have no competing interests to declare. This project was funded by the UK EPSRC grant EP/S022503/1 that supports the Centre for Doctoral Training in Cybersecurity delivered by UCL's Departments of Computer Science, Security and Crime Science, and Science, Technology, Engineering and Public Policy.}
\bibliographystyle{abbrvnat
| f0ed851e5b893de407c60c9771358c35f42a29fd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\begin{figure}[t]
\includegraphics[width=\columnwidth]{OptiTrap_teaser.png}
\caption{While previously, shapes demonstrated on levitation displays were limited to simple shapes with almost constant curvature \cite{Fushimi19,Hirayama19,Plasencia20}, our approach allows to render generic complex paths. Shapes that have not been demonstrated before include sharp edges as with the heart (B) and dolphin (D), as well as significant changes of the curvature, as with the cat (A) and flower (C) (see Supplementary Video).
Photos of the physical particle are made with an exposure time ranging from 0.2 to 1~s, such that each photo shows multiple periods of the orbit.}
\label{fig:teaser}
\Description{Four photos labelled A, B, C and D show different graphics generated with the levitation display. Image A shows an outline of a 2.4 centimetres wide cat, rendered at 5 Hertz. Image B shows an outline of a 5.4 centimetres wide heart, rendered at 10 Hertz. Image C shows an outline of a 5.6 centimetres wide three-petal flower, rendered at 10 Hertz. Image D shows an outline of a 4.7 centimetres wide dolphin, rendered at 3.3 Hertz. A human index finger is petting the dolphin. The volumetric dolphin graphics and the index finger are of similar size.}
\end{figure}
\begin{figure*}[!tb]
\includegraphics[width=2\columnwidth]{OptiTrap_flowchart.png}
\caption{\emph{OptiTrap} is an automated method to compute trap trajectories to reveal generic mid-air shapes on levitation displays. The method accepts a reference path (e.g., the shape of a heart) without any timing information as an input (A). Our approach considers the capabilities of the device and its trap-particle dynamics and combines these with a path following approach (B). The approach produces feasible trap trajectories, describing when and where the traps must be created (C). The resulting particle motion can be presented on the actual device, yielding feasible paths and supporting complex objects with sharp edges and/or significant changes of the curvature (D).}
\label{fig:teaser_diagram}
\Description{A flowchart with four elements sequentially connected with flow links. The starting state A is a reference path q of theta, which captures the positions of a reference shape of a heart. A flows to B, the OptiTrap algorithm, composed of the trap-particle dynamics and optimal path following components. The algorithm outputs a trap trajectory u of t, in state C, which contains both the trap positions and timings. An image shows that the trap trajectories do not overlap with the shape positions that were fed to the algorithm in state A. Finally, the ending state D is the particle motion. A photo of the rendered heart shape on the levitation display is shown above a stretched out hand, indicating that the shape is three dimensional and rendered in mid-air.}
\end{figure*}
Acoustic levitation has recently demonstrated the creation of volumetric content in mid-air by exploiting the Persistence of Vision (PoV) effect.
This is achieved by using acoustic traps to rapidly move single~\cite{Hirayama19} or multiple~\cite{Plasencia20} particles along a periodic reference path, revealing a shape within $0.1s$ (i.e., the integration interval of human eyes~\cite{Bowen74}).
However, the way to define such PoV content remains unsolved.
That is, there are no automated approaches to compute the \emph{trap trajectories}, i.e., the positioning and timing of the acoustic traps that will allow us to reveal the desired shape.
Currently, content creators can only rely on trial and error to find physically feasible trap trajectories resembling their intended target shapes, resulting in a time consuming and challenging process. For instance, the creator will need to define the timing of the path (i.e., not only \emph{where} the particle must be, but also \emph{when}).
Such timing is not trivial and will affect the overall rendering time and the accelerations applied to the particle, which must be within the capabilities of the levitator.
The way forces distribute around the acoustic trap will also need to be considered to decide where traps must be located.
Considering these challenges is crucial to design feasible trap trajectories, as a single infeasible point along the trajectory typically results in the particle being ejected from the levitator (e.g., approaching a sharp corner too quickly). Thus, while it has been theorised that linear PoV paths of up to 4m and peak speeds of 17m/s are possible~\cite{FushimiLimits}, actual content demonstrated to date has been limited to simple shapes with almost constant curvature along the path and much lower speeds (e.g., 0.72m/s in~\cite{Ochiai14}; 1.2m/s in~\cite{Hirayama19}; 2.25m/s in \cite{Plasencia20}, combining six particles).
In this paper, we present \emph{OptiTrap}, the first structured numerical approach to compute trap trajectories for acoustic levitation displays.
\emph{OptiTrap} automates the definition of levitated PoV content, computing physically feasible and nearly time-optimal trap trajectories given only a reference path, i.e., the desired shape.
As shown in Figure~\ref{fig:teaser}, this allows for larger and more complex shapes than previously demonstrated, as well as shapes featuring significant changes in curvature and/or sharp corners.
Our approach is summarised in Figure~\ref{fig:teaser_diagram}.
\emph{OptiTrap} assumes only a generic reference path $\textbf{q}(\theta)$ as an input, with no temporal information (see~Figure~\ref{fig:teaser_diagram}(A)). \emph{OptiTrap} formulates this as a path following problem (see~Figure~\ref{fig:teaser_diagram}(B)), computing the optimum timing in which a particle can traverse such path. Our formulation considers the \emph{Trap-Particle Dynamics} of the system, using a 3D model of acoustic forces around a trap. This results in a non-invertible model, which cannot exploit differential flatness~\cite{Fliess95a}, on which most path following approaches rely. We instead provide a coupling stage for the dynamics of our system and numerically invert the system. This approach produces a trap trajectory $\textbf{u}(t)$ (Figure~\ref{fig:teaser_diagram}(C)) that results in the intended and nearly time-optimal particle motion. That is, $\textbf{u}(t)$ defines the positions and timing of the traps that cause the particle to reveal the target shape $\textbf{q}(\theta)$, according to the capabilities of the actual device, cf.~Figure~\ref{fig:teaser_diagram}(D).
In summary, we contribute the first structured numerical approach to compute physically feasible and nearly time-optimal trap trajectories for levitation displays, given only a reference path that the particle should follow. As a core contribution, we provide a theoretical formulation of the problem in terms of path following approaches, allowing optimum timing and device properties (e.g., trap-particle dynamics) to be jointly considered. All the details of the approach we propose are provided in Section~\ref{sec:opt}.
We illustrate the potential of our approach by rendering shapes featuring straight lines, sharp corners, and complex shapes, such as those in Figure~\ref{fig:teaser}. We then provide an experimental validation of our approach, demonstrating increases of up to 563\% in the size of rendered objects, up to 150\% in the rendering frequency, and improvements in accuracy (e.g., shape revealed more accurately). While the baseline shapes we compare against require trial and error to determine optimal sizes or frequencies, our approach always yields feasible paths (i.e., working in at least 9 out of 10 attempts) and makes consistent use of accelerations very close to (but not exceeding) the maximum achievable by the device, independently of the target shape.
These features allow \emph{OptiTrap} to render complex objects involving sharp edges and significant changes in curvature that have never been demonstrated before. Even more importantly, it provides a tool to systematically explore the range of contents that levitation displays can create, as a key step to exploit their potential.
\section{Background and Challenges}
\label{sec:prob_stat}
\begin{figure*}[!tb]
\includegraphics[width=2\columnwidth]{OptiTrap_sampling_strategy.png}
\caption{ Three different path timings for the cardioid shape with fixed traversal time: (A) equidistant sampling of the arc length, (B) equidistant sampling of the path parameter~$\theta$, and (C) \emph{OptiTrap}.
The corresponding acceleration magnitudes are depicted in (D).
Strategy (A) is physically infeasible due to infinite required acceleration at the corner ($\theta=\pi$).
Strategy (B) does not share this problem, but requires careful parametrization of the path, while \emph{OptiTrap} (C) yields the timing automatically.}
\label{fig:timings}
\Description{The figure consists of four subplots. The first three plots show a cardioid shape in two dimentional space. The x-axis denotes the horizontal spatial component from -4 to 4 centimetres and the y-axis the vertical from 8 to 16 centimetres, both at increments of two centimetres. Subplot D shows the acceleration magnitude resulting from the timing strategies from A to C, on a scale from 0 to 1000 metres per seconds squared as a function of the path parameter theta, going from 0 to two pi. The most notable difference in the timing strategies occurs at the sharp corner of the cardioid shape, where the equidistant arc length strategy results in infinite acceleration. }
\end{figure*}
Our goal is is to minimise complexity for content creators. Thus, given a reference path (i.e., a shape defined by the content creator), our approach must produce physically feasible trap trajectories, which accurately reveal the desired path, while making optimal use of the capabilities of the device.
We formalise our content as a reference path $Q$, provided by the content creator without any preassigned time information.
It is an explicitly parametrized curve
\begin{equation} \label{eq:path}
Q:=\{\xi \in \mathbb{R}^3 \mid \theta \in [\theta_{0},\theta_{f}] \mapsto \textbf{q}(\theta) \},
\end{equation}
where $\textbf{q}\colon\mathbb{R} \rightarrow \mathbb{R}^3$, as later justified in Section~\ref{ssec:OPT}, must be a twice continuously differentiable function with respect to~$\theta$, which is called the path parameter.
Increasing values of~$\theta$ denote forward movement along the path~$Q$.
The starting point on the path is $\textbf{q}(\theta_0)$, while $\textbf{q}(\theta_f)$ marks the end of the path.
For periodic paths, such as a particle cyclically revealing a PoV shape, $\textbf{q}(\theta_0) = \textbf{q}(\theta_f)$ and $\dot{\textbf{q}}(\theta_0) = \dot{\textbf{q}}(\theta_f)$ hold.
Please note that while splines would provide a generic solution to this parametrization (i.e., twice continuously differentiable fitting the points in $Q$), any other parametric functions can be used.
For example, taking $\theta\in[0, 2\pi]$ and $r>0$, a cardioid as in Figure~\ref{fig:timings}, can be described by:
\begin{equation}\label{eq:cardioid}
\textbf{q}(\theta)=(0, r\sin(\theta)(1+\cos(\theta)), -r\cos(\theta)(1+\cos(\theta))+r)^\top.
\end{equation}
Whatever the approach used, revealing such reference paths typically involves rapid particle movements, as PoV content must be revealed within about $0.1s$~\cite{Bowen74}) and must consider the feasibility of the path (i.e., the trap-particle dynamics, the topology of the trap, and the capabilities of the levitator). This results in the following two main challenges for \emph{OptiTrap}: 1) determining the optimal timing for the particle revealing the path; and 2) computing the trap positions generating the required forces on the particle.
\subsection{Challenge 1: Determining an Optimal Path Timing}
The reference path $Q$ defines the geometry but not the timing (i.e., it describes \emph{where} the particle needs to be, but not \emph{when} it needs to be there). Our approach must compute such a timing, and the strategy followed will have important implications on the velocity and accelerations applied to the particle and, hence, on the physical feasibility of the content.
Different timing strategies and their effects on the feasibility can be illustrated using the cardioid example from~\eqref{eq:cardioid}. Figure~\ref{fig:timings} depicts various timing strategies applied to this shape, all of them traversing the same path in the same overall time.
A straight-forward approach would be to sample equidistantly along the path, as shown in Figure~\ref{fig:timings}(A). This works well for lines and circles, where a constant speed can be maintained, but fails at sudden changes in curvature due to uncontrolled accelerations (see the acceleration at the corner of cardioid in Figure~\ref{fig:timings}(D), at $\theta=\pi$).
The second example in Figure~\ref{fig:timings}(B) shows a simple equidistant sampling on the path parameter $\theta$. As shown in Figure~\ref{fig:timings}(D), this results in areas of strong curvature naturally involving low accelerations, and can work particularly well for low-frequency path parametrizations in terms of sinusoidals (e.g., \cite{FushimiLimits} showed such sinusoidal timing along straight paths, to theorise optimum/optimistic content sizes). In any case, the quality of the result obtained by this approach will vary depending on the specific shape and parametrization used.
As an alternative, the third example shows the timing produced by our approach, based on the shape and capabilities of the device (i.e., forces it can produce in each direction). Please note how this timing reduces the maximum accelerations required (i.e., retains them below the limits of the device), by allowing the particle to travel faster in parts that were unnecessarily slow (e.g., Figure~\ref{fig:timings}(D), at $\theta=\pi$). For the same maximum accelerations (i.e., the same device), this timing strategy could produce larger shapes or render them in shorter times, for better refresh rates. This illustrates the impact of a careful timing strategy when revealing any given reference path.
We address this challenge in Section~\ref{ssec:OPT}.
\subsection{Challenge 2: Computation of Trap Positions}
In addition to the timing, the approach must also compute the location of the traps. Previous approaches \cite{Hirayama19, Plasencia20, FushimiLimits} placed the traps along the shape to be presented, under the implied assumption that the particle would remain in the centre of such trap.
However, the location where the traps must be created almost never matches the location of the particle. Acoustic traps feature (almost) null forces at the centre of the trap and high restorative forces around them \cite{Marzo15}. As such, the only way a trap can accelerate a particle is by having it placed at a distance from the centre of the trap. Such trap-particle distances were measured by \cite{Hirayama19} to assess performance achieved during their speed tests. Distortions related to rendering fast moving PoV shapes were also shown in \cite{Fushimi19}. However, none of them considered or corrected for such displacements.
Please note the specific displacement between the trap and the particle will depend on several factors, such as the particle acceleration required at each point along the shape, as well as the trap topology (i.e., how forces distribute around it). No assumptions can be made that the traps will remain constrained to positions along or tangential to the target shape.
We describe how our approach solves this challenge in Section~\ref{ssec:extract-trap-trajectory}.
\section{Related Work}
\emph{OptiTrap} addresses the challenges above by drawing from advances in the fields of acoustic levitation and control theory, which we review in this section.
\subsection{Acoustic Levitation}
Single frequency sound-waves were first observed to trap dust particles in the lobes of a standing wave more than 150 years ago~\cite{Stevens1899}.
This has been used to create mid-air displays with particles acting as 3D voxels~\cite{Ochiai14,Omirou15,Omirou16, Sahoo16}, but such standing waves do not allow control of individual particles.
Other approaches have included Bessel beams~\cite{Norasikin19}, self-bending beams~\cite{Norasikin18}, boundary holograms~\cite{Inoue19} or near-field levitation~\cite{NearFieldLevitation}.
However, most display approaches have relied on the generic levitation framework proposed by~\cite{Marzo15}, combining a focus pattern and a levitation signature.
Although several trap topologies (i.e., twin traps, vortex traps or bottle beams) and layouts (i.e., one-sided, two-sided, v-shape) are possible, most displays proposed have adopted a top-bottom levitation setup and twin traps, as these result in highest vertical trapping forces.
This has allowed individually controllable particles/voxels~\cite{Marzo19} or even particles attached to other props and projection surfaces for richer types of content~\cite{Morales19, FenderArticulev}.
The use of single~\cite{Fushimi19,Hirayama19} or multiple~\cite{Plasencia20} fast moving particles have allowed for dynamic and free-form volumetric content, but is still limited to small sizes and simple vector graphics~\cite{FushimiLimits}.
Several practical aspects have been explored around such displays, such as selection~\cite{Freeman18} and manipulation techniques~\cite{Bachynskyi18}, content detection and initialisation~\cite{FenderArticulev} or collision avoidance~\cite{Reynal20}.
However, no efforts have been made towards optimising content considering the capabilities (i.e., the dynamics) of such displays, particularly for challenging content such as the one created by PoV high-speed particles.
\cite{Hirayama19} showed sound-fields must be updated at very high rates for the trap-particle system to engage in high accelerations, identifying optimum control for rates above 10kHz.
However, they only provided a few guidelines (e.g., maximum speed in corners, maximum horizontal/vertical accelerations) to guide the definition of the PoV content.
\cite{FushimiLimits} provided a theoretical exploration of this topic, looking at maximum achievable speeds and content sizes, according to the particle sizes and sound frequency used.
While the dynamics of the system were considered, these were extremely simplified, using a model of acoustic trap forces and system dynamics that are only applicable for oscillating recti-linear trajectories along the vertical axis of the levitator.
\cite{Paneva20} proposed a generic system simulating the dynamics of a top-bottom levitation system.
This operates as a forward model simulating the behaviour of the particle given a specific path for the traps, but unlike our approach it does not address the inverse problem.
Thus, \emph{OptiTrap} is the first algorithm allowing the definition of PoV content of generic shapes, starting only from a geometric definition (i.e., shape to present, no timing information) and optimising it according to the capabilities of the device and the dynamics of the trap-particle system.
\subsection{Path Following and Optimal Control} \label{sec:PFopt}
The particle in the trap constitutes a dynamical system, which can be controlled by setting the location of the trap.
As such, making the particle traverse the reference path can be seen as an optimal control problem (OCP).
In the context of computer graphics, optimal control approaches have been studied particularly in the areas of physics-based character animation~\cite{geijtenbeek12animation} and aerial videography~\cite{nageli2017real}.
Such OCPs can be considered as function-space variants of nonlinear programs, whereby nonlinear dynamics are considered as equality constraints. In engineering applications, OCPs are frequently solved via direct discretization~\cite{Stryk93,Bock84}, which leads to finite-dimensional nonlinear programs.
In physics-based character animation, such direct solution methods
are known as spacetime constraints~\cite{witkin88spacetime, rose96spacetime}.
If framed as \emph{a levitated particle along a given geometric reference path (i.e., the PoV shape)}, the problem can be seen as an instance of a path following problem~\cite{Faulwasser12}.
Such problems also occur in aerial videography, when a drone is to fly along a reference path~\cite{nageli2017real, roberts16generating}.
In order to yield a physically feasible trajectory, it is necessary to adjust the timing along that trajectory. This can be done by computing feed-forward input signals~\cite{Faulwasser12, roberts16generating} or, if the system dynamics and computational resources allow, using closed-loop solutions, such as Model Predictive Control in~\cite{nageli2017real}.
However, optimal control of acoustically levitated particles cannot be approached using such techniques. All of the above approaches require {\em differential flatness}~\cite{Fliess95a}.
That is, the underlying system dynamics must be invertible, allowing for the problem to be solved by projecting the dynamics of the moving object onto the path manifold~\cite{Nielsen08a}.
This inversion approach also enables formulating general path following problems in the language of optimal control, encoding desired objectives (e.g., minimum time, see~\cite{Shin85a, Verscheure09b, ifat:faulwasser14b}).
As discussed later in Section~\ref{ssec:model-acoustic-forces}, the distribution of forces around the acoustic trap (i.e., our system dynamics) are not invertible, requiring approaches that have not been widely developed in the literature. We do this by coupling the non-invertible particle dynamics into a virtual system, solvable with a conventional path following approach. We then use these coupling parameters and our model of particle dynamics to solve for the location of the traps.
\section{Optimal Control for Levitation Displays}\label{sec:opt}
This section provides a description of our \emph{OptiTrap} approach, which automates the definition of levitated content, computing physically feasible and nearly time-optimal trap trajectories using only a reference path as an input.
In Section~\ref{ssec:model-acoustic-forces}, we first describe our specific hardware setup and the general mathematical framework, then we consider existing models of the trap-particle dynamics (Subsection~\ref{sssec:exsting-models}), before introducing our proposed model (Subsection~\ref{sssec:our-model}).
Next, we describe the two stages in our algorithm, which match the challenges identified in Section~\ref{sec:prob_stat}.
That is, Section~\ref{ssec:OPT} describes the computation of optimum timing, approached as an optimal path following problem, while Section~\ref{ssec:extract-trap-trajectory} explains how to compute the trap locations.
Please note that Section~\ref{sec:opt} focuses on the general case of presenting levitated content, that is, particles cyclically traversing a reference path (shape) as to reveal it.
Other cases, such as a particle accelerating from rest as to reach the initial state (i.e., initial position and speed) required to render the content can be easily derived from the general case presented here, and are detailed in the Supplementary Material S1.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{OptiTrap_setup.png}
\caption{Overview of the components in our setup. We used two opposed arrays of transducers at a distance of 23.9 cm and an OptiTrack system to track the position of levitated particles in real time. }
\label{fig:setup}
\Description{A photo of equipment. Two square array boards of transducers with dimension 16.8 centimetres are aligned to exactly face each other at a vertical distance of 23.9 centimetres. Two optical motion cameras mounted on the side, are faced towards the volume between the boards.}
\end{figure}
\subsection{Modelling the Trap-Particle Dynamics}\label{ssec:model-acoustic-forces}
We start by describing the model of the trap-particle dynamics used for our specific setup, shown in Figure \ref{fig:setup}. This setup uses two opposed arrays of 16×16 transducers controlled by an FPGA and an OptiTrack tracking system (Prime 13 motion capture system at a frequency of 240Hz). The design of the arrays is a reproduction of the setup in~\cite{Morales21}, modified to operate at 20Vpp and higher update rates of 10kHz. The device generates a single twin-trap using the method described in \cite{Hirayama19}, allowing for vertical and horizontal forces of $4.2\cdot10^{-5}$N and $2.1\cdot10^{-5}$N, respectively, experimentally computed using the linear speed tests in \cite{Hirayama19} (i.e., 10cm paths, binary search with 9 out 10 success ratios, with particle mass $m\approx 0.7\cdot10^{-7}$ kg).
Note the substantial difference between the maximum forces in the vertical and horizontal directions. Our approach will need to remain aware of direction as this determines maximum accelerations.
In general, the trap-particle dynamics of such system can be described by simple Newtonian mechanics, i.e.,
\begin{equation}\label{eq:system}
m\ddot{\textbf{p}}(t) = F(\textbf{p}(t),\dot{\textbf{p}}(t),\textbf{u}(t)).
\end{equation}
Here, $\textbf{p}(t)=(p_x, p_y, p_z)^\top \in \mathbb{R}^3$ represents the particle position in Cartesian $(x,y,z)$ coordinates at time $t\in\mathbb{R}_0^+$, and $\dot{\textbf{p}}(t)$ and $\ddot{\textbf{p}}(t)$ are the velocity and acceleration of the particle, respectively. The force acting on the particle is mostly driven by the acoustic radiation forces and, as such, drag and gravitational forces can be neglected \cite{Hirayama19}. Therefore, the net force acting on the particle will depend only on $\textbf{p}(t)$ and on the position of the acoustic trap at time $t$ denoted by $\textbf{u}(t)=(u_x(t),u_y(t),u_z(t))^\top\in\mathbb{R}^3$.
As a result, our approach requires an accurate and ideally invertible model of the acoustic forces delivered by an acoustic trap. That is, we seek an accurate model predicting forces at any point around a trap only in terms of $\textbf{u}(t)$ and $\textbf{p}(t)$.
Moreover, an invertible model would allow us to analytically determine where to place the trap as to produce a specific force on the particle given the particle location.
For spherical particles considerably smaller than the acoustic wavelength and operating in the far-field regime, such as those used by our device, the acoustic forces exerted can be modelled by the gradient of the Gor'kov potential~\cite{Bruus2012}. This is a generic model, suitable to model acoustic forces resulting from any combination of transducer locations and transducer activations, but it also depends on all these parameters, making it inadequate for our approach.
Our case, using a top-bottom setup and vertical twin traps is much more specific. This allows for simplified analytical models with forces depending only on the relative position of the trap and the particle, which we compare to forces as predicted from the Gor'kov potential in Figure~\ref{fig:modelForces}.
\subsubsection{Existing Models of the Trap-Particle Dynamics}\label{sssec:exsting-models}
\begin{figure}[!tb]
\includegraphics[width=\columnwidth]{OptiTrap_F_max.png}
\caption{Analytical models of horizontal (A) and vertical (B) acoustic forces around a trap. The plots on the left show forces along the main axes X and Z, and analytical models provide a good fit. The plots on the right show horizontal (A) and vertical (B) forces across 2D slices through the trap centre along the X and Z axes. Spring and Sinusoidal models fail to predict forces outside the main axes, while our proposed model provides accurate reconstruction within the region of interest (highlight).}
\label{fig:modelForces}
\Description{Two subplots A and B showing the acoustic force on the y-axis going from F maximum to F minimum, as a function of the displacement between the acoustic trap and the particle, ranging from -0.01 to 0.01 metres. }
\end{figure}
Simple \emph{spring} models have been used extensively \cite{Paneva20, Fushimi18Nonlin, FushimiLimits}, modelling trapping forces according to a \emph{stiffness} parameter $\mathcal{K}_i$, that is, with forces being proportional to the distance of the particle to the centre of the trap:
\begin{equation*}
F_i(\textbf{p},\textbf{u}) := \mathcal{K}_i \cdot |u_i-p_i|, \quad i \in \{x,y,z\}.
\end{equation*}
Such models are usually refined by providing specific stiffness values for each dimension, but they are only suitable for particles remaining in close proximity to the centre of the trap, i.e., the linear region near the centre of the trap. As a result, \emph{spring} models are only accurate for systems moving particles slowly, requiring low acceleration and forces (so that particles remain within the linear region).
\emph{Sinusoidal} models have been proposed as an alternative~\cite{Fushimi18Nonlin,FushimiLimits}, providing accurate fitting from the centre of the trap to the peaks of the force distribution in Figure~\ref{fig:modelForces}, according to the peak trapping force $\mathcal{A}_i$ and characteristic frequency $\mathcal{V}_i$:
\begin{equation*}
F_i(\textbf{p},\textbf{u}) := \mathcal{A}_i \cdot \sin(\mathcal{V}_i \cdot (u_i-p_i)), \quad i \in \{x,y,z\},
\end{equation*}
However, both of these models (i.e., \emph{spring} and \emph{sinusoidal}) are only suitable for particles placed along one of the main axes of the acoustic trap, not for particles arbitrarily placed at any point around it.
This is illustrated on the right of Figure~\ref{fig:modelForces}, which provides an overview of how forces distribute on a horizontal and vertical 2D plane around a trap, according to each model (i.e., Gor'kov, \emph{spring}, \emph{sinusoidal}, and \emph{Ours}, detailed in the next subsection).
Please note how the three previous models show good matching along the horizontal and vertical axes (represented as blue and red lines).
However, the \emph{spring} and \emph{sinusoidal} models are of one-dimensional nature (e.g., force $F_x$ only depends on X distance $(u_x-p_x)$) and become inaccurate at points deviating from the main axes.
\subsubsection{Our Model} \label{sssec:our-model}
The complexity of the force distribution modelled by the Gor'kov potential increases as the distance to the centre of the trap increases.
However, we only need to derive a model allowing us to predict where to place the trap to produce a specific force on the particle.
This allows us to limit our considerations to the region corresponding to the peaks designated by $\mathcal{A}_r$ and $\mathcal{A}_z$ in Figure~\ref{fig:modelForces}.
For points within this region, forces around twin traps distribute in a mostly \emph{axis-symmetric} fashion, which can be approximated as:
\begin{subequations}\label{eq:FrFz}
\begin{align}
F_r(\textbf{p},\textbf{u}) :=\ & \mathcal{A}_r \cdot \cos\left(\mathcal{V}_z \cdot {(u_z-p_z)}\right) \cdot \\
&\sin\left(\mathcal{V}_{xr} \cdot {\sqrt{(u_x-p_x)^2+(u_y-p_y)^2}}\right), \notag \\
F_z(\textbf{p},\textbf{u}) :=\ & \mathcal{A}_z \cdot \sin\left(\mathcal{V}_z \cdot (u_z-p_z)\right) \cdot \\
& \cos\left(\mathcal{V}_{zr} \cdot \sqrt{(u_x-p_x)^2+(u_y-p_y)^2}\right), \notag
\end{align}
\end{subequations}
where $\mathcal{A}_r, \mathcal{A}_z$ represent peak trapping forces along the radial and vertical directions of the trap, respectively.
$\mathcal{V}_z$, $\mathcal{V}_{xr}$, $\mathcal{V}_{zr}$ represent characteristic frequencies of the sinusoidals describing how the forces evolve around the trap.
The resulting forces can be converted into acoustic forces in 3D space from these cylindrical coordinates, with azimuth $\phi = \arctan ((u_y-p_y)/(u_x-p_x))$:
\begin{equation} \label{eq:Facou}
F(\textbf{p},\textbf{u}) = \begin{pmatrix}
F_x(\textbf{p},\textbf{u}) \\
F_y(\textbf{p},\textbf{u}) \\
F_z(\textbf{p},\textbf{u})
\end{pmatrix} := \begin{pmatrix}
F_r(\textbf{p},\textbf{u}) \cos\phi \\
F_r(\textbf{p},\textbf{u}) \sin\phi \\
F_z(\textbf{p},\textbf{u})
\end{pmatrix}.
\end{equation}
We validated our model by comparing its accuracy against the forces predicted by the gradient of the Gor'kov potential. More specifically, we simulated 729 single traps homogeneously distributed across the working volume of our levitator (i.e., 8x8x8cm, in line with \cite{Hirayama19, Fushimi19}), testing 400 points around each trap for a total of 400x729 force estimations.
\begin{table}[t]
\caption{Fit parameters for the Spring, Sinusoidal and Axis-symmetric models and the respective average relative errors when compared to Gor'kov.}
\centering
\begin{tabular}{ |c|c|c|c|c|c|c|}
\hline
Model & \multicolumn{2}{|c|}{Spring}& \multicolumn{2}{|c|}{Sinusoidal}& \multicolumn{2}{|c|}{Axis-symmetric}
\\
\hline
\multirow{6}{*}{Fit}&$\mathcal{K}_x$&-0.0071&$\mathcal{A}_x$ &0.00009 &$\mathcal{A}_r$ & 0.0004636\\
&$\mathcal{K}_y$ & -0.0071&$\mathcal{A}_y$ &0.00009 &$\mathcal{A}_z$ & 0.0002758\\
&$\mathcal{K}_z$&-0.94&$\mathcal{A}_z$ &-0.0019 &$\mathcal{V}_z$ & 1307.83\\
& & &$\mathcal{V}_x$ &-68.92 &$\mathcal{V}_{xr}$ & -476.49\\
&&&$\mathcal{V}_y$ &-68.92 &$\mathcal{V}_{zr}$ & 287.87\\
& & &$\mathcal{V}_z$ &1307.83 &&\\
\hline
Error & \multicolumn{2}{|c|}{63.3\%}& \multicolumn{2}{|c|}{35.9\%}& \multicolumn{2}{|c|}{4.3\%}
\\
\hline
\end{tabular}
\label{table:fit_parameters}
\end{table}
Table~\ref{table:fit_parameters} summarises the error distribution achieved by these three models when compared to Gor'kov, showing an average relative error as low as 4\% for our model and much poorer fitting for the other two models. Full details and raw data used in this validation can be found in the Supplementary Material S2.
As a final summary, this results in a model for our trap-particle dynamics that only depends on $\textbf{p}$ and $\textbf{u}$ and which fits the definition of a second-order ordinary differential equation:
\begin{equation}\label{eq:system_acoustic}
m\ddot{\textbf{p}}(t) = F(\textbf{p}(t),\textbf{u}(t)).
\end{equation}
While the resulting model is accurate (4\% relative error), we note that it is not invertible, which will complicate the formulation of our solution approach. For instance, for a particle at $\textbf{p}=(0,0,0)$, force $F(\textbf{p},\textbf{u}) = (0,0, \mathcal{A}_z \cdot \cos\left(\mathcal{V}_{zr} \cdot x \right))$ can be obtained with either $\textbf{u}= (x, 0, \pi/(2\mathcal{V}_z))$ or $\textbf{u}= (-x, 0, \pi/(2\mathcal{V}_z))$.
\subsection{Open-Loop Optimal Path Following} \label{ssec:OPT}
This section computes the timing for the particle, so that it moves along the given reference path $Q$ from~\eqref{eq:path} in minimum time and according to the dynamics of the system. We approach this as an open-loop (or feed-forward) path following problem, as typically done in robotics~\cite{Faulwasser12}. That is, we design an optimal control problem that computes the timing $t \mapsto\theta(t)$ as to keep the levitated particle on the prescribed path, to traverse the path in optimum (minimum) time while considering the trap-particle dynamics (i.e., minimum \emph{feasible} time).
Our non-invertible model of particle dynamics calls for a more complex treatment of the problem, which we split in two parts.
In the first part we derive a virtual system for the timing law, which we describe as a system of first-order differential equations, solvable with traditional methods.
In the second part we couple the timing law with
our non-invertible dynamics, introducing auxiliary variables that will later enable pseudo-inversion (i.e., to compute trap placement for a given force) as described in Section~\ref{ssec:extract-trap-trajectory}.
\subsubsection{Error Dynamics and the Timing Law.}\label{sssec:OPT_timing}
The requirement that the particle follows the path $Q$ exactly and for all times means that the deviation from the path equals zero for all $t\in\mathbb{R}^+_0$, i.e.,
\[
\textbf{e}(t) := \textbf{p}(t) -\textbf{q}(\theta(t)) \equiv 0.
\]
If the path deviation $\textbf{e}(t)$ is $0$ during the whole interval $[t_0, t_1]$, this implies that the time derivatives of $\textbf{e}(t)$ also have to vanish on $(t_0, t_1)$.
Considering this (i.e., $\dot e(t)\equiv 0$ and $\ddot e(t)\equiv 0$) results in:
\begin{subequations}\label{eq:dpath_para}
\begin{align}
\textbf{p}(t) &=\textbf{q}(\theta(t)), \label{eq:dpath_dpara0} \\
\dot{\textbf{p}}(t)&=\dot{\textbf{q}}(\theta(t))=\frac{\partial \textbf{q}}{\partial \theta}\dot{\theta}(t),\\
\ddot{\textbf{p}}(t)&= \ddot{\textbf{q}}(\theta(t))=\frac{\partial^2 \textbf{q}}{\partial \theta^2}\dot{\theta}(t)^2+\frac{\partial \textbf{q}}{\partial \theta}\ddot{\theta}(t). \label{eq:dpath_dpara_2}
\end{align}
\end{subequations}
Thus, for particles on the path~$Q$, system dynamics of the form~\eqref{eq:system_acoustic}, and provided we are able to express $\textbf{u}$ as a function of $\textbf{p}$ and $\ddot{\textbf{p}}$ (i.e., the system inversion described in Section~\ref{sec:PFopt}), the
position, the velocity, and the acceleration of the particle can be expressed via $\theta, \dot\theta, \ddot\theta$, respectively.
For details, a formal derivation, and tutorial introductions we refer to~\cite{Faulwasser12,epfl:faulwasser15c}.
Observe that in~\eqref{eq:dpath_para} the partial derivatives $\frac{\partial^2 \textbf{q}}{\partial \theta^2}$ and $\frac{\partial \textbf{q}}{\partial \theta}$, as well as $\dot\theta$ and~$\ddot\theta$ appear.
This leads to two additional observations:
First, the parametrisation $\textbf{q}$ from~\eqref{eq:path} should be at least twice continuously differentiable with respect to $\theta$, so that the partial derivatives are well-defined.
This justifies our constraint in Section~\ref{sec:prob_stat}.
Please note this does not rule out corners in the path~$Q$, as shown by the parametrisation~\eqref{eq:cardioid} of the cardioid.
Second, the time evolution of $\theta$ should be continuously differentiable, as otherwise large jumps can occur in the acceleration.
To avoid said jumps, we
generate the timing $t\mapsto \theta(t)$ via the double integrator
\begin{equation}\label{eq:timingLaw}
\ddot\theta(t) = v(t),
\end{equation}
where $v(t) \in \mathbb{R}$ is a computational degree of freedom, used to control the progress of the particle along~$Q$.
The function $v(t)$ enables us to later cast the computation of time-optimal motions along reference paths as an OCP, see~\eqref{eq:OCP2}.
The in-homogeneous second-order ordinary differential equation~\eqref{eq:timingLaw} takes care of jumps, but needs to be augmented by conditions on~$\theta$ and~$\dot{\theta}$ at initial time $t=0$ and final time $t=T$. This is done to represent the periodic nature of our content (i.e., the particle reveals the same path many times per second):
\begin{equation}\label{eq:timingLaw_bc}
\theta(0)=\theta_0, \quad \theta(T)=\theta_f, \quad \dot{\theta}(0)=\dot{\theta}(T),
\end{equation}
where $\theta_0$ and $\theta_f$ are taken from~\eqref{eq:path} and the total time~$T$ will be optimally determined by the OCP. The first two equations in~\eqref{eq:timingLaw_bc} ensure that the path~$Q$ is fully traversed, while the third one ensures the speeds at the beginning and end of the path match.
With these additional considerations, we define the (state) vector $\textbf{z}(t):=(\theta(t), \dot{\theta}(t))^\top$, which finally allows us rewrite our second-order differential equations in~\eqref{eq:timingLaw} as a system of first-order differential equations:
\begin{equation} \label{eq:dz_dt}
\dot{\textbf{z}}(t)=
\begin{pmatrix}
0&1\\
0&0
\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\
1
\end{pmatrix}v(t), \quad \textbf{z}(0)=\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T.
\end{equation}
Please note that \eqref{eq:dz_dt} is an equivalent \emph{virtual} system to~\eqref{eq:timingLaw} and~\eqref{eq:timingLaw_bc}, solvable using standard Runge-Kutta methods~\cite{butcher2016numerical}. The system~\eqref{eq:dz_dt} is suitable to generate the timing law, but the particle dynamics~\eqref{eq:system_acoustic} still need to be included, as described next.
\subsubsection{Coupling of non-invertible particle dynamics:}\label{sssec:OPT_trap_placement}
To include the particle dynamics~\eqref{eq:system_acoustic} in the computation of the timing, we need to couple them with the system~\eqref{eq:dz_dt}.
The typical approach is to rewrite~\eqref{eq:system_acoustic} as:
\begin{equation}\label{eq:M_implicit}
M(\ddot{\textbf{p}}(t), {\textbf{p}}(t), {\textbf{u}}(t)):= m\ddot{\textbf{p}}(t) - F(\textbf{p}(t),\textbf{u}(t)) = 0
\end{equation}
and make sure that this equation locally admits an inverse function:
\begin{equation} \label{eq:invModel}
{\textbf{u}}(t) = M^{-1}(\ddot{\textbf{p}}(t), {\textbf{p}}(t)),
\end{equation}
This would allow us to easily compute the location of our traps.
However, such inversion is not straightforward for our model, and to the best of our knowledge, standard solution methods do not exist for such cases.
We deal with this challenge by introducing constraints related to our particle dynamics.
These will allow us to compute feasible timings while delaying the computation of the exact trap location to a later stage in the process.
More specifically, we introduce auxiliary variables~$\zeta_1,...,\zeta_6$ for each trigonometric term in the force~$F$, and we replace $p_i = q_i(\theta)$ for $i \in \{x,y,z\}$ (i.e., particle position exactly matches our reference path):
\begin{subequations}\label{eq:zeta}
\begin{align}
\zeta_1 &= \sin\left(\mathcal{V}_{xr} {\sqrt{(u_x-q_x(\theta))^2+(u_y-q_y(\theta))^2}}\right),\\
\zeta_2 &= \cos\left(\mathcal{V}_z \cdot {(u_z-q_z(\theta)}\right),\\
\zeta_3 &= \sin\left(\mathcal{V}_z \cdot {(u_z-q_z(\theta)}\right), \\
\zeta_4 &= \cos\left(\mathcal{V}_{zr} {\sqrt{(u_x-q_x(\theta))^2+(u_y-q_y(\theta))^2}}\right),\\
\zeta_5 &= \sin\phi,\\
\zeta_6 &= \cos\phi.
\end{align}
\end{subequations}
With this, we are now able to formally express the force~\eqref{eq:Facou} in terms of $\zeta:=(\zeta_1,...,\zeta_6)$, i.e.,
\begin{equation}
\tilde{F}(\zeta) := \begin{pmatrix}
\mathcal{A}_r \zeta_1\zeta_2 \zeta_6 \\
\mathcal{A}_r \zeta_1\zeta_2 \zeta_5\\
\mathcal{A}_z \zeta_4\zeta_3
\end{pmatrix}.
\end{equation}
Using these auxiliary variables, we can now couple the trap-particle dynamics~\eqref{eq:system_acoustic} with the virtual system~\eqref{eq:dz_dt}.
To this end, similar to~\eqref{eq:M_implicit}, we define the following constraint along path~$Q$:
\begin{equation}
\widetilde M(\theta(t), \dot\theta(t), v(t), \zeta(t)) := m\ddot{\textbf{q}}(\theta(t)) - \tilde{F}({\zeta}(t)) = 0.
\end{equation}
Observe that i) we need~$\dot{\theta}(t)$ due to~\eqref{eq:dpath_dpara_2}; and ii) $\ddot{\theta}(t)$ can be replaced by~$v(t)$ due to~\eqref{eq:timingLaw}.
Finally, combining this with the prior first-order differential system allows us to conceptually formulate the problem of computing a minimum-time motion along the path~$Q$ as:
\begin{equation}\label{eq:OCP2}
\begin{aligned}
\min_{v,T, \zeta}&~T + \gamma\int_0^T v(t)^2 \mathrm{d}t \\
\text{subject to}& \\
\dot{\textbf{z}}(t)&= \begin{pmatrix}0&1\\0&0\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\1\end{pmatrix}v(t),
\quad \textbf{z}(0) =\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T, \\
0&=\widetilde M(\textbf{z}(t), v(t), \zeta(t)), \\
\zeta(t)&\in [-1,1]^6.
\end{aligned}
\end{equation}
The constraint on $\zeta(t)$ is added since the trigonometric structure of~\eqref{eq:zeta} is not directly encoded in the OCP, while $\gamma\geq 0$ is a regularisation parameter.
\begin{figure}[tb]
\includegraphics[width=1\columnwidth]{OptiTrap_gamma_comparison.png}
\caption{Effect of the regularisation parameter $\gamma$ on the position of the optimised traps.
}
\label{fig:gamma_comparison}
\Description{Three subplots A, B and C show a cardioid shape in two dimensional space. The x-axis shows the horizontal spatial component from -5 to 5 centimetres, at steps of 5, while the y-axis shows the vertical from 8 to 16 centimetres, at steps of 4. Subplot A shows the computed trap positions when the regularisation parameter gamma is low. Sudden jumps at the highest and lowest vertical position, as well as at the sharp edge of the shape are highlighted.}
\end{figure}
The case~$\gamma=0$ corresponds to strictly minimising time, which typically leads to an aggressive use of the forces around the trap. The example in Figure \ref{fig:gamma_comparison}(A) shows the trap accelerating the particle before arriving at the corner, then applying aggressive deceleration before it reaches the corner.
At the top and bottom of the shape, again the particle is accelerated strongly, then it is suddenly decelerated by the traps placed behind it, in order to move around the curve.
This would be the optimum solution in an ideal case (i.e., a device working \emph{exactly} as per our model), but inaccuracies in the real device can make them unstable. Moreover, we observed the overall reductions in rendering time to usually be quite small.
The regularisation ($\gamma>0$) enables \emph{nearly} time-optimal solutions. More specifically, this is a strictly convex regularisation that penalises high magnitudes of the virtual input $v = \ddot{\theta}$. This is most closely related to the accelerations applied to the particle, hence avoiding aggressive acceleration/deceleration as shown in Figures~\ref{fig:gamma_comparison}(B) and (C). Several heuristics could be proposed to automate selection of a suitable value for $\gamma$ (e.g., use smallest $\gamma$ ensuring that the dot product of $\dot{\textbf{p}}(t)$ and $\dot{\textbf{u}}(t)$ remains always positive), but this step is considered beyond the scope of the current paper.
Note that the OCP in~\eqref{eq:OCP2} yields the (nearly) optimal timing along the path~$Q$, respecting the particle dynamics and hence solving \emph{Challenge 1}.
More\-over, it yields the required forces through $\zeta(t)$.
However, it does not give the trap trajectory~$\textbf{u}(t)$ that generates these forces.
To obtain the trap trajectory and thus solve \emph{Challenge~2}, we refine~\eqref{eq:OCP2} in the following section.
\subsection{Computing the Trap Trajectory}\label{ssec:extract-trap-trajectory}
The second stage in our approach deals with \emph{Challenge 2}, computing the trap trajectory~$\textbf{u}(t)$ based on the solution of~\eqref{eq:OCP2}. In theory, we would use the values for $\zeta_i(t)$, $i=1,...,6$ and $q_j(\theta(t))$, $j\in\{x,y,z\}$ to determine the particle position and force required at each moment in time, obtaining the required trap position~$\textbf{u}(t)$ by solving~\eqref{eq:zeta}.
In practice, solving~$\textbf{u}(t)$ from~\eqref{eq:zeta} numerically is not trivial, particularly for values of $\zeta_i$ close to $\pm 1$, where numerical instabilities could occur, particularly for $\zeta_1$ and $\zeta_4$.
We attenuate these difficulties by providing further structure to the OCP~\eqref{eq:OCP2}, specifically regarding~$\zeta$:
\begin{equation}\label{eq:zeta_trig_pythagoras}
\zeta_2^2+\zeta_3^2 = 1, \quad \zeta_5^2+\zeta_6^2 = 1.
\end{equation}
We also constrain the solvability of~\eqref{eq:zeta} by using a constant back-off $\varepsilon\in\interval[open]{0}{1}$ in order to avoid numerical instabilities:
\begin{equation}\label{eq:consistConstraints}
-1+\varepsilon \leq \zeta_i \leq 1-\varepsilon, \quad i=1,...,6.
\end{equation}
For the sake of compact notation, we summarise the above constraints in the following set notation
\begin{equation}
\mathcal{Z} := \left\{\zeta \in \mathbb{R}^6 ~ | ~\eqref{eq:zeta_trig_pythagoras} \text{ and } \eqref{eq:consistConstraints}
\text{ are satisfied}\right\}.
\end{equation}
The final OCP is then given by
\begin{equation}\label{eq:OCP3}
\begin{aligned}
\min_{v,T, \zeta}&~T + \gamma\int_0^T v(t)^2 \mathrm{d}t \\
\text{subject to}& \\
\dot{\textbf{z}}(t)&= \begin{pmatrix}0&1\\0&0\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\1\end{pmatrix}v(t),
\quad \textbf{z}(0) =\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T, \\
0&=\widetilde M(\textbf{z}(t), v(t), \zeta(t)), \\
\zeta(t)&\in \mathcal{Z}.
\end{aligned}
\end{equation}
Finally, given $\zeta(t) \in \mathcal{Z}$ and $\theta(t)$ from the solution of~\eqref{eq:OCP3}, we numerically solve~\eqref{eq:zeta} for $u_i$, $i\in\{x,y,z\}$, using Powell's dog leg method~\cite{Mangasarian94}. Please note that higher values of $\varepsilon$ will limit the magnitudes of the forces exploited by our approach. A simple solution is to perform a linear search over $\varepsilon$, only increasing its value and recomputing the solution if Powell's method fails to converge to a feasible trap location.
For details on discretising the OCP~\eqref{eq:OCP3}, we refer to the Supplementary Material S3.
\section{Evaluation}
\label{sec:eval}
This section evaluates the capability of our algorithm to support content creation by generating physically feasible trap trajectories given only a reference path (i.e., a shape) as an input. We select a range of shapes and analyse the effects each step in our approach has on the final achievable size, rendering frequency, and reconstruction error, for each shape.
We also inspect how these steps influence the presence of visual distortions in the end result.
Finally, we analyse the resulting acceleration profiles.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{OptiTrap_evaluation_shapes.png}
\caption{Test shapes used for evaluation, as rendered by our \emph{OptiTrap} approach. The circle (A) provides a trivial timing case. The cardioid (B) and squircle (C) represent simple shapes featuring sharp corners and straight lines. The fish (D) is used as an example of composite shape featuring corners, straight lines, and curves.}
\label{fig:EvalShapes}
\Description{The image consists of four photos A, B, C and D. The photos show illuminated outlines of the different shapes rendered by the levitation interface in mid-air.}
\end{figure}
\subsection{Test Shapes}
\label{ssec:eval_test_shapes}
We used four test shapes for the evaluation: a circle, a cardioid, a squircle, and a fish, all shown in Figure~\ref{fig:EvalShapes}.
The circle is selected as a trivial case in terms of timing. It can be optimally sampled by using a constant angular speed (i.e., constant acceleration), and is there to test whether our solution converges towards such optimum solutions. The cardioid and squircle represent simple shapes featuring sharp corners (cardioid) and straight lines (squircle), both of them challenging features. Finally, the fish is selected as an example composite shape, featuring all elements (i.e., corners, straight lines, and curves).
All of these shapes can be parameterised by low-frequency sinusoids (see Supplementary Material S4), ensuring the path parameter~$\theta$ progresses slowly in areas of high curvature.
This provides a good starting point for our baseline comparison that we explain in Section~\ref{ssec:eval_conditions}.
While our approach works for content in 3D, we perform the evaluation on shapes in a 2D plane of the 3D space to facilitate the analysis of accelerations in Section~\ref{ssec:eval_accelerations}, which we will perform in terms of horizontal and vertical accelerations (i.e., the independent factors given our axis-symmetric model of forces), allowing us to validate \emph{OptiTrap}'s awareness of particle directions and how close it can get to the maximum accelerations allowed by the dynamics of acoustic traps, as discussed in Section~\ref{ssec:model-acoustic-forces}.
Notice in Figure~\ref{fig:EvalShapes}, that due to the timing differences, different shape elements exhibit different levels of brightness.
To obtain homogeneous brightness along the shape, please refer to the solution proposed by \cite{Hirayama19}, where the particle illumination is adjusted to the particle speed.
\subsection{Conditions Compared}
\label{ssec:eval_conditions}
We compare three approaches to rendering levitated shapes, where each approach subsequently addresses one of the challenges introduced in Section~\ref{sec:prob_stat}.
This will help us assess the impact that each challenge has on the final results obtained.
The first condition is a straight-forward \emph{Baseline}, with homogeneous sampling of the path parameter, which matches the example strategy shown in Figure~\ref{fig:timings}(B), and placing traps where the particle should be. The \emph{Baseline} still does not address any of the challenges identified (i.e., optimum timing or trap placement), but it matches approaches used in previous works \cite{Hirayama19, Plasencia20, FushimiLimits} and will illustrate their dependence on the specific shape and initial parametrisation used.
The second condition, \emph{OCP\_Timing}, makes use of our \emph{OptiTrap} approach to compute optimum and feasible timing (see Subsection~\ref{sssec:OPT_timing}), but still ignores \emph{Challenge 2}, assuming that particles match trap location (i.e., it skips Subsection~\ref{sssec:OPT_trap_placement}). The third condition is the full \emph{OptiTrap} approach, which considers feasibility and trap-particle dynamics and deals with both the timing and trap placement challenges.
These conditions are used in a range of comparisons involving size, frequency, reconstruction error as well as comparative analysis of the effects of each condition on the resulting particle motion, detailed in the following sections.
\subsection{Maximum Achievable Sizes}
\label{ssec:eval_sizes}
This section focuses on the maximum sizes that can be achieved for each test shape and condition, while retaining an overall rendering time of 100ms\footnote{The circle rendered at 100ms would exceed the size of our device's working volume. Thus, this shape was rendered at 67ms.} (i.e., PoV threshold). For each of these cases, we report the maximum achievable size and reliability, which were determined as follows.
Maximum size is reported in terms of shape width and in terms of meters of \emph{content per second} rendered \cite{Plasencia20}, and their determination was connected to the feasibility of the trap trajectories, particularly for the \emph{Baseline} condition.
\begin{table*}
\caption{Maximum shape width and meters of content per second achieved with the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach at constant rendering frequency. Successful trials out of 10.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|c|c|c|c|c|}
\cline{3-12}
\multicolumn{2}{c}{}&\multicolumn{3}{|c}{\emph{Baseline}}& \multicolumn{3}{|c}{\emph{OCP\_Timing}}&\multicolumn{3}{|c|}{\emph{OptiTrap}}&\multicolumn{1}{c|}{Increase }\\
\cline{1-11}
\multicolumn{1}{|l|}{Shape} &Freq.& Width & Content per & Success & Width & Content per& Success & Width & Content per& Success & in Width w.r.t.\\
\multicolumn{1}{|l|}{} & (Hz) & (cm) & Second (m)& Rate & (cm) & Second (m)& Rate& (cm) & Second (m)& Rate &the \emph{Baseline}\\
\hline
\multicolumn{1}{|l|}{Circle} & 15& 7.00& 3.30 & 10/10&7.00 & 3.30 & 10/10 &7.00 & 3.30 & 10/10 & 0\% \\
\hline
\multicolumn{1}{|l|}{Cardioid} &10 & 8.05 & 2.48& 10/10& 9.09& 2.80& 10/10 & 9.09 &2.80 & 10/10 &12.9\% \\
\hline
\multicolumn{1}{|l|}{Squircle} &10& 0.80& 0.29 & 9/10&5.30& 1.90 & 10/10 & 5.30& 1.90 & 10/10 & 562.5\%\\
\hline
\multicolumn{1}{|l|}{Fish} &10& 7.77& 2.42 & 10/10&8.76 & 2.75 &10/10 &8.76& 2.75 & 10/10 & 12.7\% \\
\hline
\end{tabular}
\label{table:shape_sizes_results}
\end{table*}
More specifically, we determined maximum feasible sizes for the \emph{Baseline} condition by conducting an iterative search. That is, we increased the size at each step and tested the reliability of the resulting shape. We considered the shape feasible if it could be successfully rendered at least 9 out of 10 times in the actual levitator. On a success, we would increase size by increasing the shape width by half a centimetre. On a failure, we would perform a binary search between the smallest size failure and the largest size success, stopping after two consecutive failures and reporting the largest successful size. This illustrates the kind of trial and error that a content designer would need to go through without our approach and it was the most time consuming part of this evaluation.
For the other two conditions (i.e., \emph{OCP\_Timing} and \emph{OptiTrap}), maximum sizes could be determined without needing to validate their feasibility in the final device. Again, we iteratively increased the shape size using the same search criteria as before (i.e., 5mm increases, binary search). We assumed any result provided would be feasible (which is a reasonable assumption; see below) and checked the resulting total rendering time, increasing the size if the time was still less than 100ms. At each step, a linear search was used, iteratively increasing $\varepsilon$, until feasible trap locations could be found (i.e., equation~\eqref{eq:zeta} could be solved without numerical instabilities). Adjusting the value of the regularisation parameter $\gamma$ was done by visually inspecting the resulting trap trajectories, until no discontinuities could be observed (see Figure \ref{fig:gamma_comparison}). Please note that this is a simple and quick task, which, unlike the \emph{Baseline} condition, does not involve actual testing on the levitation device and could even be automated. All solutions provided can be assumed feasible, and the designer only needs to choose the one that better fits their needs.
Once the maximum achievable sizes were determined, these were tested in the actual device. We determined the feasibility of each shape and condition by conducting ten tests and reporting the number of cases where the particle succeeded to reveal the shape. This included the particle accelerating from rest, traversing/revealing the shape for 6 seconds and returning to rest.
Table \ref{table:shape_sizes_results} summarises the results achieved for each test shape and condition. First of all, it is worth noting that all trials were successful for \emph{OptiTrap} and \emph{OCP\_Timing} approaches, confirming our assumption that the resulting paths are indeed feasible and underpinning \emph{OptiTrap}'s ability to avoid trial and error on the actual device during content creation.
The results and relative improvements in terms of size vary according to the particular condition and shape considered. For instance, it is interesting to see that all conditions yield similar final sizes for the circle, showing that \emph{OptiTrap} (and \emph{OCP\_Timing}) indeed converge towards optimum solutions.
More complex shapes where the \emph{Baseline} parametrisation is not optimal (i.e., cardioid, squircle, and fish), show increases in size when using \emph{OCP\_Timing} and \emph{OptiTrap}. More interestingly, the increases in size vary greatly between shapes, showing increases of around $12\%$ for the fish and cardioid, and up to $562\%$ for the squircle. This is the result of the explicit parametrisation used, with reduced speeds at corners in the fish and cardioid cases, but not in the case of the squircle.
It is also interesting to see that \emph{OCP\_Timing} and \emph{OptiTrap} maintain high values of \emph{content per second} rendered, independently of the shape. That is, while the performance of the \emph{Baseline} approach is heavily determined by the specific shape (and parametrisation) used, the OCP-based approaches (\emph{OCP\_Timing} and \emph{OptiTrap}) yield results with consistent \emph{content per second}, determined by the capabilities of the device (but not so much by the specific shape). Finally, please note that no changes in terms of maximum size can be observed between \emph{OCP\_Timing} and \emph{OptiTrap}.
\subsection{Maximum Rendering Frequencies}
\label{ssec:eval_frequencies}
\begin{table}[b]
\caption{Maximum rendering frequencies achieved with the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach, for shapes of equal size. Successful trials out of 10.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|}
\cline{2-7}
&\multicolumn{2}{c|}{\emph{Baseline}}&\multicolumn{2}{c}{\emph{OCP\_Timing}}&\multicolumn{2}{|c|}{\emph{OptiTrap}}\\
\cline{1-7}
\multicolumn{1}{|l|}{Shape}& Freq. & Success & Freq. & Success &Freq. & Success \\
\multicolumn{1}{|l|}{} & (Hz)& Rate& (Hz)& Rate& (Hz)& Rate \\
\hline
\multicolumn{1}{|l|}{Circle} &15 & 10/10 & 15 &10/10 & 15 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Cardioid}& 8 & 10/10 & 10 &10/10 &10 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Squircle} & 4 & 10/10 & 10 &10/10 &10 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Fish} & 9 & 9/10& 10 &10/10 &10 & 10/10\\
\hline
\end{tabular}
\label{table:shape_frequency_results}
\end{table}
In this second evaluation, we assessed the effect of the timing strategy on the maximum achievable rendering frequencies. To do this, we selected the maximum achievable sizes obtained in the prior evaluation for each shape. We then reproduced such sizes with the \emph{Baseline} approach, searching for the maximum frequency at which this approach could reliably render the shape (using the same searching and acceptance criteria as above).
That is, while we knew \emph{OCP\_Timing} and \emph{OptiTrap} could provide 10Hz for these shapes and sizes, we wanted to determine the maximum frequency at which the \emph{Baseline} would render them, as to characterise the benefits provided by optimising the timing.
The results are summarised in Table \ref{table:shape_frequency_results}. As expected, no changes are produced for the circle. However, there is a 25\% increase for the cardioid, 150\% for the squircle, and 11\% increase for the fish using the \emph{OptiTrap} method. Again, please note that no differences can be observed in terms of maximum frequency between \emph{OCP\_Timing} and \emph{OptiTrap}.
\subsection{Reconstruction Accuracy}
\label{ssec:eval_accuracy}
To evaluate the reconstruction accuracy, we recorded each of the trials in our maximum size evaluations (see Section~\ref{ssec:eval_sizes}) using an OptiTrack camera system. We then computed the Root Mean Squared Error (RMSE) between the recorded data and the intended target shape, taking 2s of shape rendering into account (exclusively during the cyclic part, ignoring ramp-up/ramp-down). For a fair comparison across trials, we normalise the RMSE with respect to the traversed paths, by dividing the RMSE by the total path length of each individual test shape. The results are summarised in Table~\ref{table:RMSE_results}.
\begin{table}[b]
\caption{RMSE and Path-normalised (PN) RMSE with respect to the total path length of each test shape, for the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach, at constant rendering frequency.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|}
\cline{2-7}
&\multicolumn{2}{|c}{\emph{Baseline}}&
\multicolumn{2}{|c}{\emph{OCP\_Timing}}&
\multicolumn{2}{|c|}{\emph{OptiTrap}}\\
\hline
\multicolumn{1}{|l|}{Shape} & RMSE & PN & RMSE & PN & RMSE & PN\\
\multicolumn{1}{|l|}{} & (cm) & RMSE & (cm) & RMSE & (cm) & RMSE\\
\hline
\multicolumn{1}{|l|}{Circle} & 0.196 & 0.893 & 0.146 & 0.666& 0.114 & 0.521\\
\hline
\multicolumn{1}{|l|}{Cardioid} & 0.142 & 0.575 & 0.153 & 0.545 & 0.120 & 0.429\\
\hline
\multicolumn{1}{|l|}{Squircle} & 0.056 & 1.956& 0.0824 &0.434 & 0.080 & 0.421\\
\hline
\multicolumn{1}{|l|}{Fish} & 0.111 &0.458 & 0.147&0.536 &0.0581 & 0.212\\
\hline
\end{tabular}
\label{table:RMSE_results}
\end{table}
On average, both \emph{OCP\_Timing} and \emph{OptiTrap} provide better results in terms of accuracy, when compared to the \emph{Baseline}.
It is particularly worth noting that the accuracy results, in terms of the raw RMSE for \emph{OptiTrap} and \emph{OCP\_Timing}, are better for the cardioid when compared to the \emph{Baseline}, even though a larger shape is being rendered.
While the raw RMSE of \emph{OCP\_Timing} and \emph{OptiTrap} is slightly larger for the squircle, we need to take into account that the size rendered by \emph{OCP\_Timing} and \emph{OptiTrap} is almost 6 times larger.
Comparing accuracy in terms of normalised RMSE shows consistent increases of accuracy for \emph{OptiTrap} compared to the \emph{Baseline}, with overall decreases in the RMSE of $41.7\%$ (circle), $25.4\%$ (cardioid), $78.5\%$ (squircle), and $53.8\%$ (fish).
\emph{OCP\_Timing} shows a decrease of $25.4\%$, $5.14\%$, and $77.8\%$ of the normalised RMSE for the circle, cardioid, and squircle, with the exception of the fish, where the normalised RMSE increased by $17.1\%$, when compared to the \emph{Baseline}.
Comparing the reconstruction accuracy of \emph{OCP\_Timing} and \emph{OptiTrap} highlights the relevance of considering trap dynamics to determine trap locations (i.e., Section~\ref{ssec:extract-trap-trajectory}). \emph{OptiTrap} consistently provides smaller RMSE than \emph{OCP\_Timing},
as a result of considering and accounting for the trap-to-particle displacements required to apply specific accelerations.
We obtain a $21.8\%$, $21.3\%$, $2.9\%$ and $60.5\%$ decrease in the normalised RMSE, for the circle, cardioid, squircle, and fish, respectively.
Such differences are visually illustrated in Figure \ref{fig:eval_trap_location}, showing the effects on the cardioid and fish, for the \emph{OCP\_Timing} and \emph{OptiTrap} approaches. It is worth noting how placing the traps along the reference path results in the cardioid being horizontally stretched, as the particle needs to retain larger distances to the trap to keep the required acceleration (horizontal forces are weaker than vertical ones). This also results in overshooting of the particle at corner locations, which can be easily observed at the corner of the cardioid and in the fins of the fish.
For completeness, the visual comparison for the remaining two test shapes is provided in the Supplementary Material S5.
\begin{figure}[tb]
\includegraphics[width=\columnwidth]{OptiTrap_distortions.png}
\caption{Visual comparison between shapes rendered with traps located according to particle dynamics using \emph{OptiTrap} (left) and traps placed along the reference path using \emph{OCP\_Timing} (right) for the cardioid (top) and the fish (bottom) test shapes. Please note undesired increases in size and error in sharp features, such as corners.}
\label{fig:eval_trap_location}
\Description{The image consists of a two-by-two photo matrix. There are two photos of a rendered cardioid in the first row, and fish in the second row. The cardioid and the fish in the first column are rendered using the OptiTrap algorithm, and in the second column the two shapes are rendered using OCP_Timing. The levitated graphics in the second column are a bit larger in size and less precise than those in the first column. }
\end{figure}
\subsection{Analysis of Acceleration profiles}
\label{ssec:eval_accelerations}
\begin{figure}[b]
\includegraphics[width=1\columnwidth]{OptiTrap_Acc_Speed_Fish.png}
\caption{Particle acceleration and speed for the fish rendered at 10~Hz using \emph{OptiTrap} (solid black) and the \emph{Baseline} (dash-dotted purple). While the speed profiles (D) of both trap trajectories are similar, the trap trajectory generated by our approach is feasible, whereas the one generated by the \emph{Baseline} is not. The main reason is that the \emph{Baseline} exceeds the feasible horizontal acceleration, while our approach caps it to feasible values (B). Our approach compensates by using higher (available) vertical acceleration (C). Note that the times where the \emph{Baseline} applies a higher total acceleration (A) than our approach are those where our approach respects the constraints on the feasible horizontal acceleration (B).}
\label{fig:force_plots}
\Description{The figure consists of four subplots. Subplot A shows the total acceleration in metres per seconds squared from 0 to 400, in steps of 200.
Subplot B shows the horizontal acceleration in metres per seconds squared from -200 to 400, in steps of 200.
Subplot C shows the vertical acceleration in metres per seconds squared from -500 to 500, in steps of 500.
Subplot D shows the total speed in metres per seconds from 0 to 4, in steps of 2. All quantities are plotted as functions of time, going from 0 to 100 milliseconds, at steps of 20.}
\end{figure}
Finally, we examined the acceleration profiles produced by \emph{OptiTrap} and how these differ from the pre-determined acceleration profiles of the \emph{Baseline}. Please note that we only compare and discuss \emph{OptiTrap} and \emph{Baseline} in Figure \ref{fig:force_plots}, and only for the fish shape. \emph{OCP\_Timing} is not included, as it results in similar acceleration profiles as \emph{OptiTrap}. Only the fish is discussed, as it already allows us to describe the key observations that can be derived from our analysis.
For completeness, the acceleration profiles for all remaining test shapes are included in the Supplementary Material S6.
As introduced above, Figure \ref{fig:force_plots} shows the particle accelerations and speeds of a particle revealing a fish shape of maximum size (i.e., a width of 8.76cm), rendered over 100ms using the \emph{OptiTrap} and \emph{Baseline} approaches. It is worth noting that while \emph{OptiTrap} succeeded in rendering this shape, \emph{Baseline} did not (the maximum width for \emph{Baseline} was 7.77cm).
A first interesting observation is that although \emph{OptiTrap} provides lower total accelerations than the \emph{Baseline} during some parts of the path (see Figure \ref{fig:force_plots}(A)), it still manages to reveal the shape in the same time.
The key observation here is that the regions where the total acceleration is lower for \emph{OptiTrap} match with the parts of the path where the horizontal acceleration is very close to its maximum; for example, note the flat regions in Figure~\ref{fig:force_plots}(B) of around $\pm 300 m/{s}^2$). This is an example of \emph{OptiTrap}'s awareness of the dynamics and capabilities of the actual device, limiting the acceleration applied as to retain the feasibility of the path.
Second, it is worth noting that maximum horizontal accelerations are significantly smaller than vertical accelerations. As such, it does make sense for horizontal displacements to become the limiting factor. In any case, neither acceleration exceeds its respective maximum value. The \emph{OptiTrap} approach recovers any missing time by better exploiting areas where acceleration is unnecessarily small. It is also worth noting that the value of the horizontal and vertical accelerations are not simply being capped to a maximum acceleration value per direction (e.g., simultaneously maxing out at $\pm 300 m/{s}^2$ and $\pm 600 m/{s}^2$ in the horizontal and vertical directions), as such cases are not feasible according to the dynamics of our acoustic traps.
Third, it is interesting to note that the final acceleration profile is complex, retaining little resemblance with the initial parametrisation used. This is not exclusive for this shape and can be observed even in relatively simple shapes, such as the cardioid in Figure~\ref{fig:timings}(D) or the remaining shapes, provided in the Supplementary Material S6.
The complexity of these profiles is even more striking if we look at the final speeds resulting from both approaches (see Figure \ref{fig:force_plots}(D)). Even if the acceleration profiles of \emph{Baseline} and \emph{OptiTrap} are very different, their velocity profiles show only relatively subtle differences. But even if these differences are subtle, they mark the difference between a feasible path (i.e., \emph{OptiTrap}) and a failed one (i.e., \emph{Baseline}). These complex yet subtle differences also illustrate how it is simply not sensible to expect content designers to deal with such complexity, and how \emph{OptiTrap} is a necessary tool to enable effective exploitation of PoV levitated content.
\section{Discussion}
This paper presented \emph{OptiTrap}, an automated approach for optimising timings and trap placements, as to achieve feasible target shapes. We believe this is a particularly relevant step for the adoption of levitation PoV displays, as it allows the content creator to focus on the shapes to present, with feasible solutions being computed automatically, while making effective usage of the capabilities of the device. As such, we hope \emph{OptiTrap} to become an instrumental tool in helping explore the actual potential of these displays.
However, \emph{OptiTrap} is far from a complete content creation tool. Such a tool should consider the artist's workflows and practices. Similarly, testing and identifying most useful heuristics to tune our approach (e.g., the regularisation) or visualisation tools identifying tricky parts of the shapes (i.e., requiring high accelerations) should be included as a part of this process. Our goal is simply to provide the base approach enabling this kind of tools.
Even this base approach can be extended in a variety of ways.
Our levitator prototype is built from off-the-shelf hardware, and is still subject to inaccuracies that result in distortions in the sound-fields generated \cite{Fushimi19}. As such, more accurate levitation hardware or, alternatively, a model reflecting the dynamics of the system in a more accurate manner would be the most obvious pathway to improve our approach. It is worth noting that both factors should be advanced jointly. A more accurate model could also be less numerically stable, potentially leading to worse results if the hardware is not accurate enough.
The high update rates of 10kHz required by the levitator and the millisecond delays introduced by optical tracking systems indicate that closed-loop approaches can be both promising and challenging avenues to explore.
Assuming a tracking device synchronised with the levitation device (as to map current trap locations with real particle positions), \emph{OptiTrap} could be combined with learning-based approaches. These learning-based approaches will require example paths (i.e., with initial timing and trap placement), and their achievable complexity and convergence will be limited by the examples provided.
In such cases, \emph{OptiTrap} can be used to always generate feasible initial trajectories (i.e., to avoid system restarts on failure).
Thus learning-based approaches could be used to further refine \emph{OptiTrap} beyond our current model, as to account for device inaccuracies such as those discussed by~\cite{Fushimi19}.
However, higher gains can be obtained from more radical changes in the approach. For instance, our approach optimises the timing and trap placement, but it does not modify the target shapes. As illustrated in Figure~\ref{fig:force_plots}, slight changes in speed have significant effects on the acceleration profiles (and feasibility) of the shapes. This is even more prominent for position, where small changes can heavily influence the acceleration and feasibility of target shapes. As such, approaches exploiting subtle modifications to the shape could lead to significant gains in rendering performance.
Another interesting possibility would be extending our approach to use several particles. As shown in \cite{Plasencia20}, while the use of several particles does not increase the overall power that can be leveraged, it does allow for increased flexibility. That is, the intensity/forces of each trap can be individually and dynamically adjusted, as to match the needs of the region of the path that each particle is revealing. Also, particles can each be rendering specific independent features, so the content is not limited to a single connected path, and the particles do not waste time/accelerations traversing parts of the path that will not be illuminated (i.e., visible).
This approach, however, entails significant challenges. The first obvious challenge is the reliability of the intensity control of the traps. \cite{Plasencia20} demonstrate accurate control of the stiffness at the centre of the traps, but the effects of multiple (interfering) traps in each trap's topology (i.e., how forces distribute around the trap) is yet to be studied. A second challenge is that each particle is not forced to traverse the path at the same speed/rates, with such independent timing progression becoming an additional degree of freedom to account for.
Finally, further extensions to our work can come from its application to domains other than PoV displays. An obvious next step would be to adapt \emph{OptiTrap} to photophoretic displays, which trap particles using optical traps instead of acoustic traps \cite{Smalley18,Kumagai21}. This would involve including a model of the dynamics of such optical traps, but it can also involve further challenges, such as modelling the response times of galvanometers and LC panels involved in creating the trap.
Our approach can also be adapted to applications requiring objects to be transported quickly and accurately. For example, contactless transportation of matter has a wealth of applications in areas such as the study of physical phenomena, biochemical processes, materials processing, or pharmaceutics \cite{Foresti12549}. Our method can help solving such problems by directly computing a rest-to-rest solution of the matter to be transported, given an estimation of the mass of that matter, the acoustic force acting on the matter, and a path from start to target position.
\section{Conclusion}
In this paper we proposed the first structured numerical approach to compute trap trajectories for acoustic levitation displays.
\emph{OptiTrap} automatically computes physically feasible and nearly time-optimal trap trajectories to reveal generic mid-air shapes, given only a reference path.
Building on a novel multi-dimensional approximation of the acoustic forces around the trap, we formulate and show how to solve a non-linear path following problem without requiring or exploiting differential flatness of the system dynamics.
We demonstrate increases of up to 563\% in size and up to 150\% in frequency for several shapes. Additionally, we obtain better reconstruction accuracy with up to a 79\% decrease in the path-normalised RMSE.
While previously, feasible trap trajectories needed to be tuned manually for each shape and levitator, our approach requires calibration of each individual levitator just once.
We are confident that the ideas in this paper could form the basis for future content authoring tools for acoustic levitation displays and bring them a key step closer to real-world applications.
\begin{acks}
This research has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement \#737087 (Levitate) and from the AHRC UK-China Research-Industry Creative Partnerships (AH/T01136X/2).
\end{acks}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
\begin{figure}[t]
\includegraphics[width=\columnwidth]{OptiTrap_teaser.png}
\caption{While previously, shapes demonstrated on levitation displays were limited to simple shapes with almost constant curvature \cite{Fushimi19,Hirayama19,Plasencia20}, our approach allows to render generic complex paths. Shapes that have not been demonstrated before include sharp edges as with the heart (B) and dolphin (D), as well as significant changes of the curvature, as with the cat (A) and flower (C) (see Supplementary Video).
Photos of the physical particle are made with an exposure time ranging from 0.2 to 1~s, such that each photo shows multiple periods of the orbit.}
\label{fig:teaser}
\Description{Four photos labelled A, B, C and D show different graphics generated with the levitation display. Image A shows an outline of a 2.4 centimetres wide cat, rendered at 5 Hertz. Image B shows an outline of a 5.4 centimetres wide heart, rendered at 10 Hertz. Image C shows an outline of a 5.6 centimetres wide three-petal flower, rendered at 10 Hertz. Image D shows an outline of a 4.7 centimetres wide dolphin, rendered at 3.3 Hertz. A human index finger is petting the dolphin. The volumetric dolphin graphics and the index finger are of similar size.}
\end{figure}
\begin{figure*}[!tb]
\includegraphics[width=2\columnwidth]{OptiTrap_flowchart.png}
\caption{\emph{OptiTrap} is an automated method to compute trap trajectories to reveal generic mid-air shapes on levitation displays. The method accepts a reference path (e.g., the shape of a heart) without any timing information as an input (A). Our approach considers the capabilities of the device and its trap-particle dynamics and combines these with a path following approach (B). The approach produces feasible trap trajectories, describing when and where the traps must be created (C). The resulting particle motion can be presented on the actual device, yielding feasible paths and supporting complex objects with sharp edges and/or significant changes of the curvature (D).}
\label{fig:teaser_diagram}
\Description{A flowchart with four elements sequentially connected with flow links. The starting state A is a reference path q of theta, which captures the positions of a reference shape of a heart. A flows to B, the OptiTrap algorithm, composed of the trap-particle dynamics and optimal path following components. The algorithm outputs a trap trajectory u of t, in state C, which contains both the trap positions and timings. An image shows that the trap trajectories do not overlap with the shape positions that were fed to the algorithm in state A. Finally, the ending state D is the particle motion. A photo of the rendered heart shape on the levitation display is shown above a stretched out hand, indicating that the shape is three dimensional and rendered in mid-air.}
\end{figure*}
Acoustic levitation has recently demonstrated the creation of volumetric content in mid-air by exploiting the Persistence of Vision (PoV) effect.
This is achieved by using acoustic traps to rapidly move single~\cite{Hirayama19} or multiple~\cite{Plasencia20} particles along a periodic reference path, revealing a shape within $0.1s$ (i.e., the integration interval of human eyes~\cite{Bowen74}).
However, the way to define such PoV content remains unsolved.
That is, there are no automated approaches to compute the \emph{trap trajectories}, i.e., the positioning and timing of the acoustic traps that will allow us to reveal the desired shape.
Currently, content creators can only rely on trial and error to find physically feasible trap trajectories resembling their intended target shapes, resulting in a time consuming and challenging process. For instance, the creator will need to define the timing of the path (i.e., not only \emph{where} the particle must be, but also \emph{when}).
Such timing is not trivial and will affect the overall rendering time and the accelerations applied to the particle, which must be within the capabilities of the levitator.
The way forces distribute around the acoustic trap will also need to be considered to decide where traps must be located.
Considering these challenges is crucial to design feasible trap trajectories, as a single infeasible point along the trajectory typically results in the particle being ejected from the levitator (e.g., approaching a sharp corner too quickly). Thus, while it has been theorised that linear PoV paths of up to 4m and peak speeds of 17m/s are possible~\cite{FushimiLimits}, actual content demonstrated to date has been limited to simple shapes with almost constant curvature along the path and much lower speeds (e.g., 0.72m/s in~\cite{Ochiai14}; 1.2m/s in~\cite{Hirayama19}; 2.25m/s in \cite{Plasencia20}, combining six particles).
In this paper, we present \emph{OptiTrap}, the first structured numerical approach to compute trap trajectories for acoustic levitation displays.
\emph{OptiTrap} automates the definition of levitated PoV content, computing physically feasible and nearly time-optimal trap trajectories given only a reference path, i.e., the desired shape.
As shown in Figure~\ref{fig:teaser}, this allows for larger and more complex shapes than previously demonstrated, as well as shapes featuring significant changes in curvature and/or sharp corners.
Our approach is summarised in Figure~\ref{fig:teaser_diagram}.
\emph{OptiTrap} assumes only a generic reference path $\textbf{q}(\theta)$ as an input, with no temporal information (see~Figure~\ref{fig:teaser_diagram}(A)). \emph{OptiTrap} formulates this as a path following problem (see~Figure~\ref{fig:teaser_diagram}(B)), computing the optimum timing in which a particle can traverse such path. Our formulation considers the \emph{Trap-Particle Dynamics} of the system, using a 3D model of acoustic forces around a trap. This results in a non-invertible model, which cannot exploit differential flatness~\cite{Fliess95a}, on which most path following approaches rely. We instead provide a coupling stage for the dynamics of our system and numerically invert the system. This approach produces a trap trajectory $\textbf{u}(t)$ (Figure~\ref{fig:teaser_diagram}(C)) that results in the intended and nearly time-optimal particle motion. That is, $\textbf{u}(t)$ defines the positions and timing of the traps that cause the particle to reveal the target shape $\textbf{q}(\theta)$, according to the capabilities of the actual device, cf.~Figure~\ref{fig:teaser_diagram}(D).
In summary, we contribute the first structured numerical approach to compute physically feasible and nearly time-optimal trap trajectories for levitation displays, given only a reference path that the particle should follow. As a core contribution, we provide a theoretical formulation of the problem in terms of path following approaches, allowing optimum timing and device properties (e.g., trap-particle dynamics) to be jointly considered. All the details of the approach we propose are provided in Section~\ref{sec:opt}.
We illustrate the potential of our approach by rendering shapes featuring straight lines, sharp corners, and complex shapes, such as those in Figure~\ref{fig:teaser}. We then provide an experimental validation of our approach, demonstrating increases of up to 563\% in the size of rendered objects, up to 150\% in the rendering frequency, and improvements in accuracy (e.g., shape revealed more accurately). While the baseline shapes we compare against require trial and error to determine optimal sizes or frequencies, our approach always yields feasible paths (i.e., working in at least 9 out of 10 attempts) and makes consistent use of accelerations very close to (but not exceeding) the maximum achievable by the device, independently of the target shape.
These features allow \emph{OptiTrap} to render complex objects involving sharp edges and significant changes in curvature that have never been demonstrated before. Even more importantly, it provides a tool to systematically explore the range of contents that levitation displays can create, as a key step to exploit their potential.
\section{Background and Challenges}
\label{sec:prob_stat}
\begin{figure*}[!tb]
\includegraphics[width=2\columnwidth]{OptiTrap_sampling_strategy.png}
\caption{ Three different path timings for the cardioid shape with fixed traversal time: (A) equidistant sampling of the arc length, (B) equidistant sampling of the path parameter~$\theta$, and (C) \emph{OptiTrap}.
The corresponding acceleration magnitudes are depicted in (D).
Strategy (A) is physically infeasible due to infinite required acceleration at the corner ($\theta=\pi$).
Strategy (B) does not share this problem, but requires careful parametrization of the path, while \emph{OptiTrap} (C) yields the timing automatically.}
\label{fig:timings}
\Description{The figure consists of four subplots. The first three plots show a cardioid shape in two dimentional space. The x-axis denotes the horizontal spatial component from -4 to 4 centimetres and the y-axis the vertical from 8 to 16 centimetres, both at increments of two centimetres. Subplot D shows the acceleration magnitude resulting from the timing strategies from A to C, on a scale from 0 to 1000 metres per seconds squared as a function of the path parameter theta, going from 0 to two pi. The most notable difference in the timing strategies occurs at the sharp corner of the cardioid shape, where the equidistant arc length strategy results in infinite acceleration. }
\end{figure*}
Our goal is is to minimise complexity for content creators. Thus, given a reference path (i.e., a shape defined by the content creator), our approach must produce physically feasible trap trajectories, which accurately reveal the desired path, while making optimal use of the capabilities of the device.
We formalise our content as a reference path $Q$, provided by the content creator without any preassigned time information.
It is an explicitly parametrized curve
\begin{equation} \label{eq:path}
Q:=\{\xi \in \mathbb{R}^3 \mid \theta \in [\theta_{0},\theta_{f}] \mapsto \textbf{q}(\theta) \},
\end{equation}
where $\textbf{q}\colon\mathbb{R} \rightarrow \mathbb{R}^3$, as later justified in Section~\ref{ssec:OPT}, must be a twice continuously differentiable function with respect to~$\theta$, which is called the path parameter.
Increasing values of~$\theta$ denote forward movement along the path~$Q$.
The starting point on the path is $\textbf{q}(\theta_0)$, while $\textbf{q}(\theta_f)$ marks the end of the path.
For periodic paths, such as a particle cyclically revealing a PoV shape, $\textbf{q}(\theta_0) = \textbf{q}(\theta_f)$ and $\dot{\textbf{q}}(\theta_0) = \dot{\textbf{q}}(\theta_f)$ hold.
Please note that while splines would provide a generic solution to this parametrization (i.e., twice continuously differentiable fitting the points in $Q$), any other parametric functions can be used.
For example, taking $\theta\in[0, 2\pi]$ and $r>0$, a cardioid as in Figure~\ref{fig:timings}, can be described by:
\begin{equation}\label{eq:cardioid}
\textbf{q}(\theta)=(0, r\sin(\theta)(1+\cos(\theta)), -r\cos(\theta)(1+\cos(\theta))+r)^\top.
\end{equation}
Whatever the approach used, revealing such reference paths typically involves rapid particle movements, as PoV content must be revealed within about $0.1s$~\cite{Bowen74}) and must consider the feasibility of the path (i.e., the trap-particle dynamics, the topology of the trap, and the capabilities of the levitator). This results in the following two main challenges for \emph{OptiTrap}: 1) determining the optimal timing for the particle revealing the path; and 2) computing the trap positions generating the required forces on the particle.
\subsection{Challenge 1: Determining an Optimal Path Timing}
The reference path $Q$ defines the geometry but not the timing (i.e., it describes \emph{where} the particle needs to be, but not \emph{when} it needs to be there). Our approach must compute such a timing, and the strategy followed will have important implications on the velocity and accelerations applied to the particle and, hence, on the physical feasibility of the content.
Different timing strategies and their effects on the feasibility can be illustrated using the cardioid example from~\eqref{eq:cardioid}. Figure~\ref{fig:timings} depicts various timing strategies applied to this shape, all of them traversing the same path in the same overall time.
A straight-forward approach would be to sample equidistantly along the path, as shown in Figure~\ref{fig:timings}(A). This works well for lines and circles, where a constant speed can be maintained, but fails at sudden changes in curvature due to uncontrolled accelerations (see the acceleration at the corner of cardioid in Figure~\ref{fig:timings}(D), at $\theta=\pi$).
The second example in Figure~\ref{fig:timings}(B) shows a simple equidistant sampling on the path parameter $\theta$. As shown in Figure~\ref{fig:timings}(D), this results in areas of strong curvature naturally involving low accelerations, and can work particularly well for low-frequency path parametrizations in terms of sinusoidals (e.g., \cite{FushimiLimits} showed such sinusoidal timing along straight paths, to theorise optimum/optimistic content sizes). In any case, the quality of the result obtained by this approach will vary depending on the specific shape and parametrization used.
As an alternative, the third example shows the timing produced by our approach, based on the shape and capabilities of the device (i.e., forces it can produce in each direction). Please note how this timing reduces the maximum accelerations required (i.e., retains them below the limits of the device), by allowing the particle to travel faster in parts that were unnecessarily slow (e.g., Figure~\ref{fig:timings}(D), at $\theta=\pi$). For the same maximum accelerations (i.e., the same device), this timing strategy could produce larger shapes or render them in shorter times, for better refresh rates. This illustrates the impact of a careful timing strategy when revealing any given reference path.
We address this challenge in Section~\ref{ssec:OPT}.
\subsection{Challenge 2: Computation of Trap Positions}
In addition to the timing, the approach must also compute the location of the traps. Previous approaches \cite{Hirayama19, Plasencia20, FushimiLimits} placed the traps along the shape to be presented, under the implied assumption that the particle would remain in the centre of such trap.
However, the location where the traps must be created almost never matches the location of the particle. Acoustic traps feature (almost) null forces at the centre of the trap and high restorative forces around them \cite{Marzo15}. As such, the only way a trap can accelerate a particle is by having it placed at a distance from the centre of the trap. Such trap-particle distances were measured by \cite{Hirayama19} to assess performance achieved during their speed tests. Distortions related to rendering fast moving PoV shapes were also shown in \cite{Fushimi19}. However, none of them considered or corrected for such displacements.
Please note the specific displacement between the trap and the particle will depend on several factors, such as the particle acceleration required at each point along the shape, as well as the trap topology (i.e., how forces distribute around it). No assumptions can be made that the traps will remain constrained to positions along or tangential to the target shape.
We describe how our approach solves this challenge in Section~\ref{ssec:extract-trap-trajectory}.
\section{Related Work}
\emph{OptiTrap} addresses the challenges above by drawing from advances in the fields of acoustic levitation and control theory, which we review in this section.
\subsection{Acoustic Levitation}
Single frequency sound-waves were first observed to trap dust particles in the lobes of a standing wave more than 150 years ago~\cite{Stevens1899}.
This has been used to create mid-air displays with particles acting as 3D voxels~\cite{Ochiai14,Omirou15,Omirou16, Sahoo16}, but such standing waves do not allow control of individual particles.
Other approaches have included Bessel beams~\cite{Norasikin19}, self-bending beams~\cite{Norasikin18}, boundary holograms~\cite{Inoue19} or near-field levitation~\cite{NearFieldLevitation}.
However, most display approaches have relied on the generic levitation framework proposed by~\cite{Marzo15}, combining a focus pattern and a levitation signature.
Although several trap topologies (i.e., twin traps, vortex traps or bottle beams) and layouts (i.e., one-sided, two-sided, v-shape) are possible, most displays proposed have adopted a top-bottom levitation setup and twin traps, as these result in highest vertical trapping forces.
This has allowed individually controllable particles/voxels~\cite{Marzo19} or even particles attached to other props and projection surfaces for richer types of content~\cite{Morales19, FenderArticulev}.
The use of single~\cite{Fushimi19,Hirayama19} or multiple~\cite{Plasencia20} fast moving particles have allowed for dynamic and free-form volumetric content, but is still limited to small sizes and simple vector graphics~\cite{FushimiLimits}.
Several practical aspects have been explored around such displays, such as selection~\cite{Freeman18} and manipulation techniques~\cite{Bachynskyi18}, content detection and initialisation~\cite{FenderArticulev} or collision avoidance~\cite{Reynal20}.
However, no efforts have been made towards optimising content considering the capabilities (i.e., the dynamics) of such displays, particularly for challenging content such as the one created by PoV high-speed particles.
\cite{Hirayama19} showed sound-fields must be updated at very high rates for the trap-particle system to engage in high accelerations, identifying optimum control for rates above 10kHz.
However, they only provided a few guidelines (e.g., maximum speed in corners, maximum horizontal/vertical accelerations) to guide the definition of the PoV content.
\cite{FushimiLimits} provided a theoretical exploration of this topic, looking at maximum achievable speeds and content sizes, according to the particle sizes and sound frequency used.
While the dynamics of the system were considered, these were extremely simplified, using a model of acoustic trap forces and system dynamics that are only applicable for oscillating recti-linear trajectories along the vertical axis of the levitator.
\cite{Paneva20} proposed a generic system simulating the dynamics of a top-bottom levitation system.
This operates as a forward model simulating the behaviour of the particle given a specific path for the traps, but unlike our approach it does not address the inverse problem.
Thus, \emph{OptiTrap} is the first algorithm allowing the definition of PoV content of generic shapes, starting only from a geometric definition (i.e., shape to present, no timing information) and optimising it according to the capabilities of the device and the dynamics of the trap-particle system.
\subsection{Path Following and Optimal Control} \label{sec:PFopt}
The particle in the trap constitutes a dynamical system, which can be controlled by setting the location of the trap.
As such, making the particle traverse the reference path can be seen as an optimal control problem (OCP).
In the context of computer graphics, optimal control approaches have been studied particularly in the areas of physics-based character animation~\cite{geijtenbeek12animation} and aerial videography~\cite{nageli2017real}.
Such OCPs can be considered as function-space variants of nonlinear programs, whereby nonlinear dynamics are considered as equality constraints. In engineering applications, OCPs are frequently solved via direct discretization~\cite{Stryk93,Bock84}, which leads to finite-dimensional nonlinear programs.
In physics-based character animation, such direct solution methods
are known as spacetime constraints~\cite{witkin88spacetime, rose96spacetime}.
If framed as \emph{a levitated particle along a given geometric reference path (i.e., the PoV shape)}, the problem can be seen as an instance of a path following problem~\cite{Faulwasser12}.
Such problems also occur in aerial videography, when a drone is to fly along a reference path~\cite{nageli2017real, roberts16generating}.
In order to yield a physically feasible trajectory, it is necessary to adjust the timing along that trajectory. This can be done by computing feed-forward input signals~\cite{Faulwasser12, roberts16generating} or, if the system dynamics and computational resources allow, using closed-loop solutions, such as Model Predictive Control in~\cite{nageli2017real}.
However, optimal control of acoustically levitated particles cannot be approached using such techniques. All of the above approaches require {\em differential flatness}~\cite{Fliess95a}.
That is, the underlying system dynamics must be invertible, allowing for the problem to be solved by projecting the dynamics of the moving object onto the path manifold~\cite{Nielsen08a}.
This inversion approach also enables formulating general path following problems in the language of optimal control, encoding desired objectives (e.g., minimum time, see~\cite{Shin85a, Verscheure09b, ifat:faulwasser14b}).
As discussed later in Section~\ref{ssec:model-acoustic-forces}, the distribution of forces around the acoustic trap (i.e., our system dynamics) are not invertible, requiring approaches that have not been widely developed in the literature. We do this by coupling the non-invertible particle dynamics into a virtual system, solvable with a conventional path following approach. We then use these coupling parameters and our model of particle dynamics to solve for the location of the traps.
\section{Optimal Control for Levitation Displays}\label{sec:opt}
This section provides a description of our \emph{OptiTrap} approach, which automates the definition of levitated content, computing physically feasible and nearly time-optimal trap trajectories using only a reference path as an input.
In Section~\ref{ssec:model-acoustic-forces}, we first describe our specific hardware setup and the general mathematical framework, then we consider existing models of the trap-particle dynamics (Subsection~\ref{sssec:exsting-models}), before introducing our proposed model (Subsection~\ref{sssec:our-model}).
Next, we describe the two stages in our algorithm, which match the challenges identified in Section~\ref{sec:prob_stat}.
That is, Section~\ref{ssec:OPT} describes the computation of optimum timing, approached as an optimal path following problem, while Section~\ref{ssec:extract-trap-trajectory} explains how to compute the trap locations.
Please note that Section~\ref{sec:opt} focuses on the general case of presenting levitated content, that is, particles cyclically traversing a reference path (shape) as to reveal it.
Other cases, such as a particle accelerating from rest as to reach the initial state (i.e., initial position and speed) required to render the content can be easily derived from the general case presented here, and are detailed in the Supplementary Material S1.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{OptiTrap_setup.png}
\caption{Overview of the components in our setup. We used two opposed arrays of transducers at a distance of 23.9 cm and an OptiTrack system to track the position of levitated particles in real time. }
\label{fig:setup}
\Description{A photo of equipment. Two square array boards of transducers with dimension 16.8 centimetres are aligned to exactly face each other at a vertical distance of 23.9 centimetres. Two optical motion cameras mounted on the side, are faced towards the volume between the boards.}
\end{figure}
\subsection{Modelling the Trap-Particle Dynamics}\label{ssec:model-acoustic-forces}
We start by describing the model of the trap-particle dynamics used for our specific setup, shown in Figure \ref{fig:setup}. This setup uses two opposed arrays of 16×16 transducers controlled by an FPGA and an OptiTrack tracking system (Prime 13 motion capture system at a frequency of 240Hz). The design of the arrays is a reproduction of the setup in~\cite{Morales21}, modified to operate at 20Vpp and higher update rates of 10kHz. The device generates a single twin-trap using the method described in \cite{Hirayama19}, allowing for vertical and horizontal forces of $4.2\cdot10^{-5}$N and $2.1\cdot10^{-5}$N, respectively, experimentally computed using the linear speed tests in \cite{Hirayama19} (i.e., 10cm paths, binary search with 9 out 10 success ratios, with particle mass $m\approx 0.7\cdot10^{-7}$ kg).
Note the substantial difference between the maximum forces in the vertical and horizontal directions. Our approach will need to remain aware of direction as this determines maximum accelerations.
In general, the trap-particle dynamics of such system can be described by simple Newtonian mechanics, i.e.,
\begin{equation}\label{eq:system}
m\ddot{\textbf{p}}(t) = F(\textbf{p}(t),\dot{\textbf{p}}(t),\textbf{u}(t)).
\end{equation}
Here, $\textbf{p}(t)=(p_x, p_y, p_z)^\top \in \mathbb{R}^3$ represents the particle position in Cartesian $(x,y,z)$ coordinates at time $t\in\mathbb{R}_0^+$, and $\dot{\textbf{p}}(t)$ and $\ddot{\textbf{p}}(t)$ are the velocity and acceleration of the particle, respectively. The force acting on the particle is mostly driven by the acoustic radiation forces and, as such, drag and gravitational forces can be neglected \cite{Hirayama19}. Therefore, the net force acting on the particle will depend only on $\textbf{p}(t)$ and on the position of the acoustic trap at time $t$ denoted by $\textbf{u}(t)=(u_x(t),u_y(t),u_z(t))^\top\in\mathbb{R}^3$.
As a result, our approach requires an accurate and ideally invertible model of the acoustic forces delivered by an acoustic trap. That is, we seek an accurate model predicting forces at any point around a trap only in terms of $\textbf{u}(t)$ and $\textbf{p}(t)$.
Moreover, an invertible model would allow us to analytically determine where to place the trap as to produce a specific force on the particle given the particle location.
For spherical particles considerably smaller than the acoustic wavelength and operating in the far-field regime, such as those used by our device, the acoustic forces exerted can be modelled by the gradient of the Gor'kov potential~\cite{Bruus2012}. This is a generic model, suitable to model acoustic forces resulting from any combination of transducer locations and transducer activations, but it also depends on all these parameters, making it inadequate for our approach.
Our case, using a top-bottom setup and vertical twin traps is much more specific. This allows for simplified analytical models with forces depending only on the relative position of the trap and the particle, which we compare to forces as predicted from the Gor'kov potential in Figure~\ref{fig:modelForces}.
\subsubsection{Existing Models of the Trap-Particle Dynamics}\label{sssec:exsting-models}
\begin{figure}[!tb]
\includegraphics[width=\columnwidth]{OptiTrap_F_max.png}
\caption{Analytical models of horizontal (A) and vertical (B) acoustic forces around a trap. The plots on the left show forces along the main axes X and Z, and analytical models provide a good fit. The plots on the right show horizontal (A) and vertical (B) forces across 2D slices through the trap centre along the X and Z axes. Spring and Sinusoidal models fail to predict forces outside the main axes, while our proposed model provides accurate reconstruction within the region of interest (highlight).}
\label{fig:modelForces}
\Description{Two subplots A and B showing the acoustic force on the y-axis going from F maximum to F minimum, as a function of the displacement between the acoustic trap and the particle, ranging from -0.01 to 0.01 metres. }
\end{figure}
Simple \emph{spring} models have been used extensively \cite{Paneva20, Fushimi18Nonlin, FushimiLimits}, modelling trapping forces according to a \emph{stiffness} parameter $\mathcal{K}_i$, that is, with forces being proportional to the distance of the particle to the centre of the trap:
\begin{equation*}
F_i(\textbf{p},\textbf{u}) := \mathcal{K}_i \cdot |u_i-p_i|, \quad i \in \{x,y,z\}.
\end{equation*}
Such models are usually refined by providing specific stiffness values for each dimension, but they are only suitable for particles remaining in close proximity to the centre of the trap, i.e., the linear region near the centre of the trap. As a result, \emph{spring} models are only accurate for systems moving particles slowly, requiring low acceleration and forces (so that particles remain within the linear region).
\emph{Sinusoidal} models have been proposed as an alternative~\cite{Fushimi18Nonlin,FushimiLimits}, providing accurate fitting from the centre of the trap to the peaks of the force distribution in Figure~\ref{fig:modelForces}, according to the peak trapping force $\mathcal{A}_i$ and characteristic frequency $\mathcal{V}_i$:
\begin{equation*}
F_i(\textbf{p},\textbf{u}) := \mathcal{A}_i \cdot \sin(\mathcal{V}_i \cdot (u_i-p_i)), \quad i \in \{x,y,z\},
\end{equation*}
However, both of these models (i.e., \emph{spring} and \emph{sinusoidal}) are only suitable for particles placed along one of the main axes of the acoustic trap, not for particles arbitrarily placed at any point around it.
This is illustrated on the right of Figure~\ref{fig:modelForces}, which provides an overview of how forces distribute on a horizontal and vertical 2D plane around a trap, according to each model (i.e., Gor'kov, \emph{spring}, \emph{sinusoidal}, and \emph{Ours}, detailed in the next subsection).
Please note how the three previous models show good matching along the horizontal and vertical axes (represented as blue and red lines).
However, the \emph{spring} and \emph{sinusoidal} models are of one-dimensional nature (e.g., force $F_x$ only depends on X distance $(u_x-p_x)$) and become inaccurate at points deviating from the main axes.
\subsubsection{Our Model} \label{sssec:our-model}
The complexity of the force distribution modelled by the Gor'kov potential increases as the distance to the centre of the trap increases.
However, we only need to derive a model allowing us to predict where to place the trap to produce a specific force on the particle.
This allows us to limit our considerations to the region corresponding to the peaks designated by $\mathcal{A}_r$ and $\mathcal{A}_z$ in Figure~\ref{fig:modelForces}.
For points within this region, forces around twin traps distribute in a mostly \emph{axis-symmetric} fashion, which can be approximated as:
\begin{subequations}\label{eq:FrFz}
\begin{align}
F_r(\textbf{p},\textbf{u}) :=\ & \mathcal{A}_r \cdot \cos\left(\mathcal{V}_z \cdot {(u_z-p_z)}\right) \cdot \\
&\sin\left(\mathcal{V}_{xr} \cdot {\sqrt{(u_x-p_x)^2+(u_y-p_y)^2}}\right), \notag \\
F_z(\textbf{p},\textbf{u}) :=\ & \mathcal{A}_z \cdot \sin\left(\mathcal{V}_z \cdot (u_z-p_z)\right) \cdot \\
& \cos\left(\mathcal{V}_{zr} \cdot \sqrt{(u_x-p_x)^2+(u_y-p_y)^2}\right), \notag
\end{align}
\end{subequations}
where $\mathcal{A}_r, \mathcal{A}_z$ represent peak trapping forces along the radial and vertical directions of the trap, respectively.
$\mathcal{V}_z$, $\mathcal{V}_{xr}$, $\mathcal{V}_{zr}$ represent characteristic frequencies of the sinusoidals describing how the forces evolve around the trap.
The resulting forces can be converted into acoustic forces in 3D space from these cylindrical coordinates, with azimuth $\phi = \arctan ((u_y-p_y)/(u_x-p_x))$:
\begin{equation} \label{eq:Facou}
F(\textbf{p},\textbf{u}) = \begin{pmatrix}
F_x(\textbf{p},\textbf{u}) \\
F_y(\textbf{p},\textbf{u}) \\
F_z(\textbf{p},\textbf{u})
\end{pmatrix} := \begin{pmatrix}
F_r(\textbf{p},\textbf{u}) \cos\phi \\
F_r(\textbf{p},\textbf{u}) \sin\phi \\
F_z(\textbf{p},\textbf{u})
\end{pmatrix}.
\end{equation}
We validated our model by comparing its accuracy against the forces predicted by the gradient of the Gor'kov potential. More specifically, we simulated 729 single traps homogeneously distributed across the working volume of our levitator (i.e., 8x8x8cm, in line with \cite{Hirayama19, Fushimi19}), testing 400 points around each trap for a total of 400x729 force estimations.
\begin{table}[t]
\caption{Fit parameters for the Spring, Sinusoidal and Axis-symmetric models and the respective average relative errors when compared to Gor'kov.}
\centering
\begin{tabular}{ |c|c|c|c|c|c|c|}
\hline
Model & \multicolumn{2}{|c|}{Spring}& \multicolumn{2}{|c|}{Sinusoidal}& \multicolumn{2}{|c|}{Axis-symmetric}
\\
\hline
\multirow{6}{*}{Fit}&$\mathcal{K}_x$&-0.0071&$\mathcal{A}_x$ &0.00009 &$\mathcal{A}_r$ & 0.0004636\\
&$\mathcal{K}_y$ & -0.0071&$\mathcal{A}_y$ &0.00009 &$\mathcal{A}_z$ & 0.0002758\\
&$\mathcal{K}_z$&-0.94&$\mathcal{A}_z$ &-0.0019 &$\mathcal{V}_z$ & 1307.83\\
& & &$\mathcal{V}_x$ &-68.92 &$\mathcal{V}_{xr}$ & -476.49\\
&&&$\mathcal{V}_y$ &-68.92 &$\mathcal{V}_{zr}$ & 287.87\\
& & &$\mathcal{V}_z$ &1307.83 &&\\
\hline
Error & \multicolumn{2}{|c|}{63.3\%}& \multicolumn{2}{|c|}{35.9\%}& \multicolumn{2}{|c|}{4.3\%}
\\
\hline
\end{tabular}
\label{table:fit_parameters}
\end{table}
Table~\ref{table:fit_parameters} summarises the error distribution achieved by these three models when compared to Gor'kov, showing an average relative error as low as 4\% for our model and much poorer fitting for the other two models. Full details and raw data used in this validation can be found in the Supplementary Material S2.
As a final summary, this results in a model for our trap-particle dynamics that only depends on $\textbf{p}$ and $\textbf{u}$ and which fits the definition of a second-order ordinary differential equation:
\begin{equation}\label{eq:system_acoustic}
m\ddot{\textbf{p}}(t) = F(\textbf{p}(t),\textbf{u}(t)).
\end{equation}
While the resulting model is accurate (4\% relative error), we note that it is not invertible, which will complicate the formulation of our solution approach. For instance, for a particle at $\textbf{p}=(0,0,0)$, force $F(\textbf{p},\textbf{u}) = (0,0, \mathcal{A}_z \cdot \cos\left(\mathcal{V}_{zr} \cdot x \right))$ can be obtained with either $\textbf{u}= (x, 0, \pi/(2\mathcal{V}_z))$ or $\textbf{u}= (-x, 0, \pi/(2\mathcal{V}_z))$.
\subsection{Open-Loop Optimal Path Following} \label{ssec:OPT}
This section computes the timing for the particle, so that it moves along the given reference path $Q$ from~\eqref{eq:path} in minimum time and according to the dynamics of the system. We approach this as an open-loop (or feed-forward) path following problem, as typically done in robotics~\cite{Faulwasser12}. That is, we design an optimal control problem that computes the timing $t \mapsto\theta(t)$ as to keep the levitated particle on the prescribed path, to traverse the path in optimum (minimum) time while considering the trap-particle dynamics (i.e., minimum \emph{feasible} time).
Our non-invertible model of particle dynamics calls for a more complex treatment of the problem, which we split in two parts.
In the first part we derive a virtual system for the timing law, which we describe as a system of first-order differential equations, solvable with traditional methods.
In the second part we couple the timing law with
our non-invertible dynamics, introducing auxiliary variables that will later enable pseudo-inversion (i.e., to compute trap placement for a given force) as described in Section~\ref{ssec:extract-trap-trajectory}.
\subsubsection{Error Dynamics and the Timing Law.}\label{sssec:OPT_timing}
The requirement that the particle follows the path $Q$ exactly and for all times means that the deviation from the path equals zero for all $t\in\mathbb{R}^+_0$, i.e.,
\[
\textbf{e}(t) := \textbf{p}(t) -\textbf{q}(\theta(t)) \equiv 0.
\]
If the path deviation $\textbf{e}(t)$ is $0$ during the whole interval $[t_0, t_1]$, this implies that the time derivatives of $\textbf{e}(t)$ also have to vanish on $(t_0, t_1)$.
Considering this (i.e., $\dot e(t)\equiv 0$ and $\ddot e(t)\equiv 0$) results in:
\begin{subequations}\label{eq:dpath_para}
\begin{align}
\textbf{p}(t) &=\textbf{q}(\theta(t)), \label{eq:dpath_dpara0} \\
\dot{\textbf{p}}(t)&=\dot{\textbf{q}}(\theta(t))=\frac{\partial \textbf{q}}{\partial \theta}\dot{\theta}(t),\\
\ddot{\textbf{p}}(t)&= \ddot{\textbf{q}}(\theta(t))=\frac{\partial^2 \textbf{q}}{\partial \theta^2}\dot{\theta}(t)^2+\frac{\partial \textbf{q}}{\partial \theta}\ddot{\theta}(t). \label{eq:dpath_dpara_2}
\end{align}
\end{subequations}
Thus, for particles on the path~$Q$, system dynamics of the form~\eqref{eq:system_acoustic}, and provided we are able to express $\textbf{u}$ as a function of $\textbf{p}$ and $\ddot{\textbf{p}}$ (i.e., the system inversion described in Section~\ref{sec:PFopt}), the
position, the velocity, and the acceleration of the particle can be expressed via $\theta, \dot\theta, \ddot\theta$, respectively.
For details, a formal derivation, and tutorial introductions we refer to~\cite{Faulwasser12,epfl:faulwasser15c}.
Observe that in~\eqref{eq:dpath_para} the partial derivatives $\frac{\partial^2 \textbf{q}}{\partial \theta^2}$ and $\frac{\partial \textbf{q}}{\partial \theta}$, as well as $\dot\theta$ and~$\ddot\theta$ appear.
This leads to two additional observations:
First, the parametrisation $\textbf{q}$ from~\eqref{eq:path} should be at least twice continuously differentiable with respect to $\theta$, so that the partial derivatives are well-defined.
This justifies our constraint in Section~\ref{sec:prob_stat}.
Please note this does not rule out corners in the path~$Q$, as shown by the parametrisation~\eqref{eq:cardioid} of the cardioid.
Second, the time evolution of $\theta$ should be continuously differentiable, as otherwise large jumps can occur in the acceleration.
To avoid said jumps, we
generate the timing $t\mapsto \theta(t)$ via the double integrator
\begin{equation}\label{eq:timingLaw}
\ddot\theta(t) = v(t),
\end{equation}
where $v(t) \in \mathbb{R}$ is a computational degree of freedom, used to control the progress of the particle along~$Q$.
The function $v(t)$ enables us to later cast the computation of time-optimal motions along reference paths as an OCP, see~\eqref{eq:OCP2}.
The in-homogeneous second-order ordinary differential equation~\eqref{eq:timingLaw} takes care of jumps, but needs to be augmented by conditions on~$\theta$ and~$\dot{\theta}$ at initial time $t=0$ and final time $t=T$. This is done to represent the periodic nature of our content (i.e., the particle reveals the same path many times per second):
\begin{equation}\label{eq:timingLaw_bc}
\theta(0)=\theta_0, \quad \theta(T)=\theta_f, \quad \dot{\theta}(0)=\dot{\theta}(T),
\end{equation}
where $\theta_0$ and $\theta_f$ are taken from~\eqref{eq:path} and the total time~$T$ will be optimally determined by the OCP. The first two equations in~\eqref{eq:timingLaw_bc} ensure that the path~$Q$ is fully traversed, while the third one ensures the speeds at the beginning and end of the path match.
With these additional considerations, we define the (state) vector $\textbf{z}(t):=(\theta(t), \dot{\theta}(t))^\top$, which finally allows us rewrite our second-order differential equations in~\eqref{eq:timingLaw} as a system of first-order differential equations:
\begin{equation} \label{eq:dz_dt}
\dot{\textbf{z}}(t)=
\begin{pmatrix}
0&1\\
0&0
\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\
1
\end{pmatrix}v(t), \quad \textbf{z}(0)=\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T.
\end{equation}
Please note that \eqref{eq:dz_dt} is an equivalent \emph{virtual} system to~\eqref{eq:timingLaw} and~\eqref{eq:timingLaw_bc}, solvable using standard Runge-Kutta methods~\cite{butcher2016numerical}. The system~\eqref{eq:dz_dt} is suitable to generate the timing law, but the particle dynamics~\eqref{eq:system_acoustic} still need to be included, as described next.
\subsubsection{Coupling of non-invertible particle dynamics:}\label{sssec:OPT_trap_placement}
To include the particle dynamics~\eqref{eq:system_acoustic} in the computation of the timing, we need to couple them with the system~\eqref{eq:dz_dt}.
The typical approach is to rewrite~\eqref{eq:system_acoustic} as:
\begin{equation}\label{eq:M_implicit}
M(\ddot{\textbf{p}}(t), {\textbf{p}}(t), {\textbf{u}}(t)):= m\ddot{\textbf{p}}(t) - F(\textbf{p}(t),\textbf{u}(t)) = 0
\end{equation}
and make sure that this equation locally admits an inverse function:
\begin{equation} \label{eq:invModel}
{\textbf{u}}(t) = M^{-1}(\ddot{\textbf{p}}(t), {\textbf{p}}(t)),
\end{equation}
This would allow us to easily compute the location of our traps.
However, such inversion is not straightforward for our model, and to the best of our knowledge, standard solution methods do not exist for such cases.
We deal with this challenge by introducing constraints related to our particle dynamics.
These will allow us to compute feasible timings while delaying the computation of the exact trap location to a later stage in the process.
More specifically, we introduce auxiliary variables~$\zeta_1,...,\zeta_6$ for each trigonometric term in the force~$F$, and we replace $p_i = q_i(\theta)$ for $i \in \{x,y,z\}$ (i.e., particle position exactly matches our reference path):
\begin{subequations}\label{eq:zeta}
\begin{align}
\zeta_1 &= \sin\left(\mathcal{V}_{xr} {\sqrt{(u_x-q_x(\theta))^2+(u_y-q_y(\theta))^2}}\right),\\
\zeta_2 &= \cos\left(\mathcal{V}_z \cdot {(u_z-q_z(\theta)}\right),\\
\zeta_3 &= \sin\left(\mathcal{V}_z \cdot {(u_z-q_z(\theta)}\right), \\
\zeta_4 &= \cos\left(\mathcal{V}_{zr} {\sqrt{(u_x-q_x(\theta))^2+(u_y-q_y(\theta))^2}}\right),\\
\zeta_5 &= \sin\phi,\\
\zeta_6 &= \cos\phi.
\end{align}
\end{subequations}
With this, we are now able to formally express the force~\eqref{eq:Facou} in terms of $\zeta:=(\zeta_1,...,\zeta_6)$, i.e.,
\begin{equation}
\tilde{F}(\zeta) := \begin{pmatrix}
\mathcal{A}_r \zeta_1\zeta_2 \zeta_6 \\
\mathcal{A}_r \zeta_1\zeta_2 \zeta_5\\
\mathcal{A}_z \zeta_4\zeta_3
\end{pmatrix}.
\end{equation}
Using these auxiliary variables, we can now couple the trap-particle dynamics~\eqref{eq:system_acoustic} with the virtual system~\eqref{eq:dz_dt}.
To this end, similar to~\eqref{eq:M_implicit}, we define the following constraint along path~$Q$:
\begin{equation}
\widetilde M(\theta(t), \dot\theta(t), v(t), \zeta(t)) := m\ddot{\textbf{q}}(\theta(t)) - \tilde{F}({\zeta}(t)) = 0.
\end{equation}
Observe that i) we need~$\dot{\theta}(t)$ due to~\eqref{eq:dpath_dpara_2}; and ii) $\ddot{\theta}(t)$ can be replaced by~$v(t)$ due to~\eqref{eq:timingLaw}.
Finally, combining this with the prior first-order differential system allows us to conceptually formulate the problem of computing a minimum-time motion along the path~$Q$ as:
\begin{equation}\label{eq:OCP2}
\begin{aligned}
\min_{v,T, \zeta}&~T + \gamma\int_0^T v(t)^2 \mathrm{d}t \\
\text{subject to}& \\
\dot{\textbf{z}}(t)&= \begin{pmatrix}0&1\\0&0\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\1\end{pmatrix}v(t),
\quad \textbf{z}(0) =\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T, \\
0&=\widetilde M(\textbf{z}(t), v(t), \zeta(t)), \\
\zeta(t)&\in [-1,1]^6.
\end{aligned}
\end{equation}
The constraint on $\zeta(t)$ is added since the trigonometric structure of~\eqref{eq:zeta} is not directly encoded in the OCP, while $\gamma\geq 0$ is a regularisation parameter.
\begin{figure}[tb]
\includegraphics[width=1\columnwidth]{OptiTrap_gamma_comparison.png}
\caption{Effect of the regularisation parameter $\gamma$ on the position of the optimised traps.
}
\label{fig:gamma_comparison}
\Description{Three subplots A, B and C show a cardioid shape in two dimensional space. The x-axis shows the horizontal spatial component from -5 to 5 centimetres, at steps of 5, while the y-axis shows the vertical from 8 to 16 centimetres, at steps of 4. Subplot A shows the computed trap positions when the regularisation parameter gamma is low. Sudden jumps at the highest and lowest vertical position, as well as at the sharp edge of the shape are highlighted.}
\end{figure}
The case~$\gamma=0$ corresponds to strictly minimising time, which typically leads to an aggressive use of the forces around the trap. The example in Figure \ref{fig:gamma_comparison}(A) shows the trap accelerating the particle before arriving at the corner, then applying aggressive deceleration before it reaches the corner.
At the top and bottom of the shape, again the particle is accelerated strongly, then it is suddenly decelerated by the traps placed behind it, in order to move around the curve.
This would be the optimum solution in an ideal case (i.e., a device working \emph{exactly} as per our model), but inaccuracies in the real device can make them unstable. Moreover, we observed the overall reductions in rendering time to usually be quite small.
The regularisation ($\gamma>0$) enables \emph{nearly} time-optimal solutions. More specifically, this is a strictly convex regularisation that penalises high magnitudes of the virtual input $v = \ddot{\theta}$. This is most closely related to the accelerations applied to the particle, hence avoiding aggressive acceleration/deceleration as shown in Figures~\ref{fig:gamma_comparison}(B) and (C). Several heuristics could be proposed to automate selection of a suitable value for $\gamma$ (e.g., use smallest $\gamma$ ensuring that the dot product of $\dot{\textbf{p}}(t)$ and $\dot{\textbf{u}}(t)$ remains always positive), but this step is considered beyond the scope of the current paper.
Note that the OCP in~\eqref{eq:OCP2} yields the (nearly) optimal timing along the path~$Q$, respecting the particle dynamics and hence solving \emph{Challenge 1}.
More\-over, it yields the required forces through $\zeta(t)$.
However, it does not give the trap trajectory~$\textbf{u}(t)$ that generates these forces.
To obtain the trap trajectory and thus solve \emph{Challenge~2}, we refine~\eqref{eq:OCP2} in the following section.
\subsection{Computing the Trap Trajectory}\label{ssec:extract-trap-trajectory}
The second stage in our approach deals with \emph{Challenge 2}, computing the trap trajectory~$\textbf{u}(t)$ based on the solution of~\eqref{eq:OCP2}. In theory, we would use the values for $\zeta_i(t)$, $i=1,...,6$ and $q_j(\theta(t))$, $j\in\{x,y,z\}$ to determine the particle position and force required at each moment in time, obtaining the required trap position~$\textbf{u}(t)$ by solving~\eqref{eq:zeta}.
In practice, solving~$\textbf{u}(t)$ from~\eqref{eq:zeta} numerically is not trivial, particularly for values of $\zeta_i$ close to $\pm 1$, where numerical instabilities could occur, particularly for $\zeta_1$ and $\zeta_4$.
We attenuate these difficulties by providing further structure to the OCP~\eqref{eq:OCP2}, specifically regarding~$\zeta$:
\begin{equation}\label{eq:zeta_trig_pythagoras}
\zeta_2^2+\zeta_3^2 = 1, \quad \zeta_5^2+\zeta_6^2 = 1.
\end{equation}
We also constrain the solvability of~\eqref{eq:zeta} by using a constant back-off $\varepsilon\in\interval[open]{0}{1}$ in order to avoid numerical instabilities:
\begin{equation}\label{eq:consistConstraints}
-1+\varepsilon \leq \zeta_i \leq 1-\varepsilon, \quad i=1,...,6.
\end{equation}
For the sake of compact notation, we summarise the above constraints in the following set notation
\begin{equation}
\mathcal{Z} := \left\{\zeta \in \mathbb{R}^6 ~ | ~\eqref{eq:zeta_trig_pythagoras} \text{ and } \eqref{eq:consistConstraints}
\text{ are satisfied}\right\}.
\end{equation}
The final OCP is then given by
\begin{equation}\label{eq:OCP3}
\begin{aligned}
\min_{v,T, \zeta}&~T + \gamma\int_0^T v(t)^2 \mathrm{d}t \\
\text{subject to}& \\
\dot{\textbf{z}}(t)&= \begin{pmatrix}0&1\\0&0\end{pmatrix}\textbf{z}(t)+\begin{pmatrix}
0\\1\end{pmatrix}v(t),
\quad \textbf{z}(0) =\textbf{z}_0, ~\textbf{z}(T) = \textbf{z}_T, \\
0&=\widetilde M(\textbf{z}(t), v(t), \zeta(t)), \\
\zeta(t)&\in \mathcal{Z}.
\end{aligned}
\end{equation}
Finally, given $\zeta(t) \in \mathcal{Z}$ and $\theta(t)$ from the solution of~\eqref{eq:OCP3}, we numerically solve~\eqref{eq:zeta} for $u_i$, $i\in\{x,y,z\}$, using Powell's dog leg method~\cite{Mangasarian94}. Please note that higher values of $\varepsilon$ will limit the magnitudes of the forces exploited by our approach. A simple solution is to perform a linear search over $\varepsilon$, only increasing its value and recomputing the solution if Powell's method fails to converge to a feasible trap location.
For details on discretising the OCP~\eqref{eq:OCP3}, we refer to the Supplementary Material S3.
\section{Evaluation}
\label{sec:eval}
This section evaluates the capability of our algorithm to support content creation by generating physically feasible trap trajectories given only a reference path (i.e., a shape) as an input. We select a range of shapes and analyse the effects each step in our approach has on the final achievable size, rendering frequency, and reconstruction error, for each shape.
We also inspect how these steps influence the presence of visual distortions in the end result.
Finally, we analyse the resulting acceleration profiles.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{OptiTrap_evaluation_shapes.png}
\caption{Test shapes used for evaluation, as rendered by our \emph{OptiTrap} approach. The circle (A) provides a trivial timing case. The cardioid (B) and squircle (C) represent simple shapes featuring sharp corners and straight lines. The fish (D) is used as an example of composite shape featuring corners, straight lines, and curves.}
\label{fig:EvalShapes}
\Description{The image consists of four photos A, B, C and D. The photos show illuminated outlines of the different shapes rendered by the levitation interface in mid-air.}
\end{figure}
\subsection{Test Shapes}
\label{ssec:eval_test_shapes}
We used four test shapes for the evaluation: a circle, a cardioid, a squircle, and a fish, all shown in Figure~\ref{fig:EvalShapes}.
The circle is selected as a trivial case in terms of timing. It can be optimally sampled by using a constant angular speed (i.e., constant acceleration), and is there to test whether our solution converges towards such optimum solutions. The cardioid and squircle represent simple shapes featuring sharp corners (cardioid) and straight lines (squircle), both of them challenging features. Finally, the fish is selected as an example composite shape, featuring all elements (i.e., corners, straight lines, and curves).
All of these shapes can be parameterised by low-frequency sinusoids (see Supplementary Material S4), ensuring the path parameter~$\theta$ progresses slowly in areas of high curvature.
This provides a good starting point for our baseline comparison that we explain in Section~\ref{ssec:eval_conditions}.
While our approach works for content in 3D, we perform the evaluation on shapes in a 2D plane of the 3D space to facilitate the analysis of accelerations in Section~\ref{ssec:eval_accelerations}, which we will perform in terms of horizontal and vertical accelerations (i.e., the independent factors given our axis-symmetric model of forces), allowing us to validate \emph{OptiTrap}'s awareness of particle directions and how close it can get to the maximum accelerations allowed by the dynamics of acoustic traps, as discussed in Section~\ref{ssec:model-acoustic-forces}.
Notice in Figure~\ref{fig:EvalShapes}, that due to the timing differences, different shape elements exhibit different levels of brightness.
To obtain homogeneous brightness along the shape, please refer to the solution proposed by \cite{Hirayama19}, where the particle illumination is adjusted to the particle speed.
\subsection{Conditions Compared}
\label{ssec:eval_conditions}
We compare three approaches to rendering levitated shapes, where each approach subsequently addresses one of the challenges introduced in Section~\ref{sec:prob_stat}.
This will help us assess the impact that each challenge has on the final results obtained.
The first condition is a straight-forward \emph{Baseline}, with homogeneous sampling of the path parameter, which matches the example strategy shown in Figure~\ref{fig:timings}(B), and placing traps where the particle should be. The \emph{Baseline} still does not address any of the challenges identified (i.e., optimum timing or trap placement), but it matches approaches used in previous works \cite{Hirayama19, Plasencia20, FushimiLimits} and will illustrate their dependence on the specific shape and initial parametrisation used.
The second condition, \emph{OCP\_Timing}, makes use of our \emph{OptiTrap} approach to compute optimum and feasible timing (see Subsection~\ref{sssec:OPT_timing}), but still ignores \emph{Challenge 2}, assuming that particles match trap location (i.e., it skips Subsection~\ref{sssec:OPT_trap_placement}). The third condition is the full \emph{OptiTrap} approach, which considers feasibility and trap-particle dynamics and deals with both the timing and trap placement challenges.
These conditions are used in a range of comparisons involving size, frequency, reconstruction error as well as comparative analysis of the effects of each condition on the resulting particle motion, detailed in the following sections.
\subsection{Maximum Achievable Sizes}
\label{ssec:eval_sizes}
This section focuses on the maximum sizes that can be achieved for each test shape and condition, while retaining an overall rendering time of 100ms\footnote{The circle rendered at 100ms would exceed the size of our device's working volume. Thus, this shape was rendered at 67ms.} (i.e., PoV threshold). For each of these cases, we report the maximum achievable size and reliability, which were determined as follows.
Maximum size is reported in terms of shape width and in terms of meters of \emph{content per second} rendered \cite{Plasencia20}, and their determination was connected to the feasibility of the trap trajectories, particularly for the \emph{Baseline} condition.
\begin{table*}
\caption{Maximum shape width and meters of content per second achieved with the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach at constant rendering frequency. Successful trials out of 10.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|c|c|c|c|c|}
\cline{3-12}
\multicolumn{2}{c}{}&\multicolumn{3}{|c}{\emph{Baseline}}& \multicolumn{3}{|c}{\emph{OCP\_Timing}}&\multicolumn{3}{|c|}{\emph{OptiTrap}}&\multicolumn{1}{c|}{Increase }\\
\cline{1-11}
\multicolumn{1}{|l|}{Shape} &Freq.& Width & Content per & Success & Width & Content per& Success & Width & Content per& Success & in Width w.r.t.\\
\multicolumn{1}{|l|}{} & (Hz) & (cm) & Second (m)& Rate & (cm) & Second (m)& Rate& (cm) & Second (m)& Rate &the \emph{Baseline}\\
\hline
\multicolumn{1}{|l|}{Circle} & 15& 7.00& 3.30 & 10/10&7.00 & 3.30 & 10/10 &7.00 & 3.30 & 10/10 & 0\% \\
\hline
\multicolumn{1}{|l|}{Cardioid} &10 & 8.05 & 2.48& 10/10& 9.09& 2.80& 10/10 & 9.09 &2.80 & 10/10 &12.9\% \\
\hline
\multicolumn{1}{|l|}{Squircle} &10& 0.80& 0.29 & 9/10&5.30& 1.90 & 10/10 & 5.30& 1.90 & 10/10 & 562.5\%\\
\hline
\multicolumn{1}{|l|}{Fish} &10& 7.77& 2.42 & 10/10&8.76 & 2.75 &10/10 &8.76& 2.75 & 10/10 & 12.7\% \\
\hline
\end{tabular}
\label{table:shape_sizes_results}
\end{table*}
More specifically, we determined maximum feasible sizes for the \emph{Baseline} condition by conducting an iterative search. That is, we increased the size at each step and tested the reliability of the resulting shape. We considered the shape feasible if it could be successfully rendered at least 9 out of 10 times in the actual levitator. On a success, we would increase size by increasing the shape width by half a centimetre. On a failure, we would perform a binary search between the smallest size failure and the largest size success, stopping after two consecutive failures and reporting the largest successful size. This illustrates the kind of trial and error that a content designer would need to go through without our approach and it was the most time consuming part of this evaluation.
For the other two conditions (i.e., \emph{OCP\_Timing} and \emph{OptiTrap}), maximum sizes could be determined without needing to validate their feasibility in the final device. Again, we iteratively increased the shape size using the same search criteria as before (i.e., 5mm increases, binary search). We assumed any result provided would be feasible (which is a reasonable assumption; see below) and checked the resulting total rendering time, increasing the size if the time was still less than 100ms. At each step, a linear search was used, iteratively increasing $\varepsilon$, until feasible trap locations could be found (i.e., equation~\eqref{eq:zeta} could be solved without numerical instabilities). Adjusting the value of the regularisation parameter $\gamma$ was done by visually inspecting the resulting trap trajectories, until no discontinuities could be observed (see Figure \ref{fig:gamma_comparison}). Please note that this is a simple and quick task, which, unlike the \emph{Baseline} condition, does not involve actual testing on the levitation device and could even be automated. All solutions provided can be assumed feasible, and the designer only needs to choose the one that better fits their needs.
Once the maximum achievable sizes were determined, these were tested in the actual device. We determined the feasibility of each shape and condition by conducting ten tests and reporting the number of cases where the particle succeeded to reveal the shape. This included the particle accelerating from rest, traversing/revealing the shape for 6 seconds and returning to rest.
Table \ref{table:shape_sizes_results} summarises the results achieved for each test shape and condition. First of all, it is worth noting that all trials were successful for \emph{OptiTrap} and \emph{OCP\_Timing} approaches, confirming our assumption that the resulting paths are indeed feasible and underpinning \emph{OptiTrap}'s ability to avoid trial and error on the actual device during content creation.
The results and relative improvements in terms of size vary according to the particular condition and shape considered. For instance, it is interesting to see that all conditions yield similar final sizes for the circle, showing that \emph{OptiTrap} (and \emph{OCP\_Timing}) indeed converge towards optimum solutions.
More complex shapes where the \emph{Baseline} parametrisation is not optimal (i.e., cardioid, squircle, and fish), show increases in size when using \emph{OCP\_Timing} and \emph{OptiTrap}. More interestingly, the increases in size vary greatly between shapes, showing increases of around $12\%$ for the fish and cardioid, and up to $562\%$ for the squircle. This is the result of the explicit parametrisation used, with reduced speeds at corners in the fish and cardioid cases, but not in the case of the squircle.
It is also interesting to see that \emph{OCP\_Timing} and \emph{OptiTrap} maintain high values of \emph{content per second} rendered, independently of the shape. That is, while the performance of the \emph{Baseline} approach is heavily determined by the specific shape (and parametrisation) used, the OCP-based approaches (\emph{OCP\_Timing} and \emph{OptiTrap}) yield results with consistent \emph{content per second}, determined by the capabilities of the device (but not so much by the specific shape). Finally, please note that no changes in terms of maximum size can be observed between \emph{OCP\_Timing} and \emph{OptiTrap}.
\subsection{Maximum Rendering Frequencies}
\label{ssec:eval_frequencies}
\begin{table}[b]
\caption{Maximum rendering frequencies achieved with the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach, for shapes of equal size. Successful trials out of 10.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|}
\cline{2-7}
&\multicolumn{2}{c|}{\emph{Baseline}}&\multicolumn{2}{c}{\emph{OCP\_Timing}}&\multicolumn{2}{|c|}{\emph{OptiTrap}}\\
\cline{1-7}
\multicolumn{1}{|l|}{Shape}& Freq. & Success & Freq. & Success &Freq. & Success \\
\multicolumn{1}{|l|}{} & (Hz)& Rate& (Hz)& Rate& (Hz)& Rate \\
\hline
\multicolumn{1}{|l|}{Circle} &15 & 10/10 & 15 &10/10 & 15 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Cardioid}& 8 & 10/10 & 10 &10/10 &10 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Squircle} & 4 & 10/10 & 10 &10/10 &10 & 10/10 \\
\hline
\multicolumn{1}{|l|}{Fish} & 9 & 9/10& 10 &10/10 &10 & 10/10\\
\hline
\end{tabular}
\label{table:shape_frequency_results}
\end{table}
In this second evaluation, we assessed the effect of the timing strategy on the maximum achievable rendering frequencies. To do this, we selected the maximum achievable sizes obtained in the prior evaluation for each shape. We then reproduced such sizes with the \emph{Baseline} approach, searching for the maximum frequency at which this approach could reliably render the shape (using the same searching and acceptance criteria as above).
That is, while we knew \emph{OCP\_Timing} and \emph{OptiTrap} could provide 10Hz for these shapes and sizes, we wanted to determine the maximum frequency at which the \emph{Baseline} would render them, as to characterise the benefits provided by optimising the timing.
The results are summarised in Table \ref{table:shape_frequency_results}. As expected, no changes are produced for the circle. However, there is a 25\% increase for the cardioid, 150\% for the squircle, and 11\% increase for the fish using the \emph{OptiTrap} method. Again, please note that no differences can be observed in terms of maximum frequency between \emph{OCP\_Timing} and \emph{OptiTrap}.
\subsection{Reconstruction Accuracy}
\label{ssec:eval_accuracy}
To evaluate the reconstruction accuracy, we recorded each of the trials in our maximum size evaluations (see Section~\ref{ssec:eval_sizes}) using an OptiTrack camera system. We then computed the Root Mean Squared Error (RMSE) between the recorded data and the intended target shape, taking 2s of shape rendering into account (exclusively during the cyclic part, ignoring ramp-up/ramp-down). For a fair comparison across trials, we normalise the RMSE with respect to the traversed paths, by dividing the RMSE by the total path length of each individual test shape. The results are summarised in Table~\ref{table:RMSE_results}.
\begin{table}[b]
\caption{RMSE and Path-normalised (PN) RMSE with respect to the total path length of each test shape, for the \emph{Baseline}, \emph{OCP\_Timing}, and \emph{OptiTrap} approach, at constant rendering frequency.}
\centering
\begin{tabular}{ c|c|c|c|c|c|c|}
\cline{2-7}
&\multicolumn{2}{|c}{\emph{Baseline}}&
\multicolumn{2}{|c}{\emph{OCP\_Timing}}&
\multicolumn{2}{|c|}{\emph{OptiTrap}}\\
\hline
\multicolumn{1}{|l|}{Shape} & RMSE & PN & RMSE & PN & RMSE & PN\\
\multicolumn{1}{|l|}{} & (cm) & RMSE & (cm) & RMSE & (cm) & RMSE\\
\hline
\multicolumn{1}{|l|}{Circle} & 0.196 & 0.893 & 0.146 & 0.666& 0.114 & 0.521\\
\hline
\multicolumn{1}{|l|}{Cardioid} & 0.142 & 0.575 & 0.153 & 0.545 & 0.120 & 0.429\\
\hline
\multicolumn{1}{|l|}{Squircle} & 0.056 & 1.956& 0.0824 &0.434 & 0.080 & 0.421\\
\hline
\multicolumn{1}{|l|}{Fish} & 0.111 &0.458 & 0.147&0.536 &0.0581 & 0.212\\
\hline
\end{tabular}
\label{table:RMSE_results}
\end{table}
On average, both \emph{OCP\_Timing} and \emph{OptiTrap} provide better results in terms of accuracy, when compared to the \emph{Baseline}.
It is particularly worth noting that the accuracy results, in terms of the raw RMSE for \emph{OptiTrap} and \emph{OCP\_Timing}, are better for the cardioid when compared to the \emph{Baseline}, even though a larger shape is being rendered.
While the raw RMSE of \emph{OCP\_Timing} and \emph{OptiTrap} is slightly larger for the squircle, we need to take into account that the size rendered by \emph{OCP\_Timing} and \emph{OptiTrap} is almost 6 times larger.
Comparing accuracy in terms of normalised RMSE shows consistent increases of accuracy for \emph{OptiTrap} compared to the \emph{Baseline}, with overall decreases in the RMSE of $41.7\%$ (circle), $25.4\%$ (cardioid), $78.5\%$ (squircle), and $53.8\%$ (fish).
\emph{OCP\_Timing} shows a decrease of $25.4\%$, $5.14\%$, and $77.8\%$ of the normalised RMSE for the circle, cardioid, and squircle, with the exception of the fish, where the normalised RMSE increased by $17.1\%$, when compared to the \emph{Baseline}.
Comparing the reconstruction accuracy of \emph{OCP\_Timing} and \emph{OptiTrap} highlights the relevance of considering trap dynamics to determine trap locations (i.e., Section~\ref{ssec:extract-trap-trajectory}). \emph{OptiTrap} consistently provides smaller RMSE than \emph{OCP\_Timing},
as a result of considering and accounting for the trap-to-particle displacements required to apply specific accelerations.
We obtain a $21.8\%$, $21.3\%$, $2.9\%$ and $60.5\%$ decrease in the normalised RMSE, for the circle, cardioid, squircle, and fish, respectively.
Such differences are visually illustrated in Figure \ref{fig:eval_trap_location}, showing the effects on the cardioid and fish, for the \emph{OCP\_Timing} and \emph{OptiTrap} approaches. It is worth noting how placing the traps along the reference path results in the cardioid being horizontally stretched, as the particle needs to retain larger distances to the trap to keep the required acceleration (horizontal forces are weaker than vertical ones). This also results in overshooting of the particle at corner locations, which can be easily observed at the corner of the cardioid and in the fins of the fish.
For completeness, the visual comparison for the remaining two test shapes is provided in the Supplementary Material S5.
\begin{figure}[tb]
\includegraphics[width=\columnwidth]{OptiTrap_distortions.png}
\caption{Visual comparison between shapes rendered with traps located according to particle dynamics using \emph{OptiTrap} (left) and traps placed along the reference path using \emph{OCP\_Timing} (right) for the cardioid (top) and the fish (bottom) test shapes. Please note undesired increases in size and error in sharp features, such as corners.}
\label{fig:eval_trap_location}
\Description{The image consists of a two-by-two photo matrix. There are two photos of a rendered cardioid in the first row, and fish in the second row. The cardioid and the fish in the first column are rendered using the OptiTrap algorithm, and in the second column the two shapes are rendered using OCP_Timing. The levitated graphics in the second column are a bit larger in size and less precise than those in the first column. }
\end{figure}
\subsection{Analysis of Acceleration profiles}
\label{ssec:eval_accelerations}
\begin{figure}[b]
\includegraphics[width=1\columnwidth]{OptiTrap_Acc_Speed_Fish.png}
\caption{Particle acceleration and speed for the fish rendered at 10~Hz using \emph{OptiTrap} (solid black) and the \emph{Baseline} (dash-dotted purple). While the speed profiles (D) of both trap trajectories are similar, the trap trajectory generated by our approach is feasible, whereas the one generated by the \emph{Baseline} is not. The main reason is that the \emph{Baseline} exceeds the feasible horizontal acceleration, while our approach caps it to feasible values (B). Our approach compensates by using higher (available) vertical acceleration (C). Note that the times where the \emph{Baseline} applies a higher total acceleration (A) than our approach are those where our approach respects the constraints on the feasible horizontal acceleration (B).}
\label{fig:force_plots}
\Description{The figure consists of four subplots. Subplot A shows the total acceleration in metres per seconds squared from 0 to 400, in steps of 200.
Subplot B shows the horizontal acceleration in metres per seconds squared from -200 to 400, in steps of 200.
Subplot C shows the vertical acceleration in metres per seconds squared from -500 to 500, in steps of 500.
Subplot D shows the total speed in metres per seconds from 0 to 4, in steps of 2. All quantities are plotted as functions of time, going from 0 to 100 milliseconds, at steps of 20.}
\end{figure}
Finally, we examined the acceleration profiles produced by \emph{OptiTrap} and how these differ from the pre-determined acceleration profiles of the \emph{Baseline}. Please note that we only compare and discuss \emph{OptiTrap} and \emph{Baseline} in Figure \ref{fig:force_plots}, and only for the fish shape. \emph{OCP\_Timing} is not included, as it results in similar acceleration profiles as \emph{OptiTrap}. Only the fish is discussed, as it already allows us to describe the key observations that can be derived from our analysis.
For completeness, the acceleration profiles for all remaining test shapes are included in the Supplementary Material S6.
As introduced above, Figure \ref{fig:force_plots} shows the particle accelerations and speeds of a particle revealing a fish shape of maximum size (i.e., a width of 8.76cm), rendered over 100ms using the \emph{OptiTrap} and \emph{Baseline} approaches. It is worth noting that while \emph{OptiTrap} succeeded in rendering this shape, \emph{Baseline} did not (the maximum width for \emph{Baseline} was 7.77cm).
A first interesting observation is that although \emph{OptiTrap} provides lower total accelerations than the \emph{Baseline} during some parts of the path (see Figure \ref{fig:force_plots}(A)), it still manages to reveal the shape in the same time.
The key observation here is that the regions where the total acceleration is lower for \emph{OptiTrap} match with the parts of the path where the horizontal acceleration is very close to its maximum; for example, note the flat regions in Figure~\ref{fig:force_plots}(B) of around $\pm 300 m/{s}^2$). This is an example of \emph{OptiTrap}'s awareness of the dynamics and capabilities of the actual device, limiting the acceleration applied as to retain the feasibility of the path.
Second, it is worth noting that maximum horizontal accelerations are significantly smaller than vertical accelerations. As such, it does make sense for horizontal displacements to become the limiting factor. In any case, neither acceleration exceeds its respective maximum value. The \emph{OptiTrap} approach recovers any missing time by better exploiting areas where acceleration is unnecessarily small. It is also worth noting that the value of the horizontal and vertical accelerations are not simply being capped to a maximum acceleration value per direction (e.g., simultaneously maxing out at $\pm 300 m/{s}^2$ and $\pm 600 m/{s}^2$ in the horizontal and vertical directions), as such cases are not feasible according to the dynamics of our acoustic traps.
Third, it is interesting to note that the final acceleration profile is complex, retaining little resemblance with the initial parametrisation used. This is not exclusive for this shape and can be observed even in relatively simple shapes, such as the cardioid in Figure~\ref{fig:timings}(D) or the remaining shapes, provided in the Supplementary Material S6.
The complexity of these profiles is even more striking if we look at the final speeds resulting from both approaches (see Figure \ref{fig:force_plots}(D)). Even if the acceleration profiles of \emph{Baseline} and \emph{OptiTrap} are very different, their velocity profiles show only relatively subtle differences. But even if these differences are subtle, they mark the difference between a feasible path (i.e., \emph{OptiTrap}) and a failed one (i.e., \emph{Baseline}). These complex yet subtle differences also illustrate how it is simply not sensible to expect content designers to deal with such complexity, and how \emph{OptiTrap} is a necessary tool to enable effective exploitation of PoV levitated content.
\section{Discussion}
This paper presented \emph{OptiTrap}, an automated approach for optimising timings and trap placements, as to achieve feasible target shapes. We believe this is a particularly relevant step for the adoption of levitation PoV displays, as it allows the content creator to focus on the shapes to present, with feasible solutions being computed automatically, while making effective usage of the capabilities of the device. As such, we hope \emph{OptiTrap} to become an instrumental tool in helping explore the actual potential of these displays.
However, \emph{OptiTrap} is far from a complete content creation tool. Such a tool should consider the artist's workflows and practices. Similarly, testing and identifying most useful heuristics to tune our approach (e.g., the regularisation) or visualisation tools identifying tricky parts of the shapes (i.e., requiring high accelerations) should be included as a part of this process. Our goal is simply to provide the base approach enabling this kind of tools.
Even this base approach can be extended in a variety of ways.
Our levitator prototype is built from off-the-shelf hardware, and is still subject to inaccuracies that result in distortions in the sound-fields generated \cite{Fushimi19}. As such, more accurate levitation hardware or, alternatively, a model reflecting the dynamics of the system in a more accurate manner would be the most obvious pathway to improve our approach. It is worth noting that both factors should be advanced jointly. A more accurate model could also be less numerically stable, potentially leading to worse results if the hardware is not accurate enough.
The high update rates of 10kHz required by the levitator and the millisecond delays introduced by optical tracking systems indicate that closed-loop approaches can be both promising and challenging avenues to explore.
Assuming a tracking device synchronised with the levitation device (as to map current trap locations with real particle positions), \emph{OptiTrap} could be combined with learning-based approaches. These learning-based approaches will require example paths (i.e., with initial timing and trap placement), and their achievable complexity and convergence will be limited by the examples provided.
In such cases, \emph{OptiTrap} can be used to always generate feasible initial trajectories (i.e., to avoid system restarts on failure).
Thus learning-based approaches could be used to further refine \emph{OptiTrap} beyond our current model, as to account for device inaccuracies such as those discussed by~\cite{Fushimi19}.
However, higher gains can be obtained from more radical changes in the approach. For instance, our approach optimises the timing and trap placement, but it does not modify the target shapes. As illustrated in Figure~\ref{fig:force_plots}, slight changes in speed have significant effects on the acceleration profiles (and feasibility) of the shapes. This is even more prominent for position, where small changes can heavily influence the acceleration and feasibility of target shapes. As such, approaches exploiting subtle modifications to the shape could lead to significant gains in rendering performance.
Another interesting possibility would be extending our approach to use several particles. As shown in \cite{Plasencia20}, while the use of several particles does not increase the overall power that can be leveraged, it does allow for increased flexibility. That is, the intensity/forces of each trap can be individually and dynamically adjusted, as to match the needs of the region of the path that each particle is revealing. Also, particles can each be rendering specific independent features, so the content is not limited to a single connected path, and the particles do not waste time/accelerations traversing parts of the path that will not be illuminated (i.e., visible).
This approach, however, entails significant challenges. The first obvious challenge is the reliability of the intensity control of the traps. \cite{Plasencia20} demonstrate accurate control of the stiffness at the centre of the traps, but the effects of multiple (interfering) traps in each trap's topology (i.e., how forces distribute around the trap) is yet to be studied. A second challenge is that each particle is not forced to traverse the path at the same speed/rates, with such independent timing progression becoming an additional degree of freedom to account for.
Finally, further extensions to our work can come from its application to domains other than PoV displays. An obvious next step would be to adapt \emph{OptiTrap} to photophoretic displays, which trap particles using optical traps instead of acoustic traps \cite{Smalley18,Kumagai21}. This would involve including a model of the dynamics of such optical traps, but it can also involve further challenges, such as modelling the response times of galvanometers and LC panels involved in creating the trap.
Our approach can also be adapted to applications requiring objects to be transported quickly and accurately. For example, contactless transportation of matter has a wealth of applications in areas such as the study of physical phenomena, biochemical processes, materials processing, or pharmaceutics \cite{Foresti12549}. Our method can help solving such problems by directly computing a rest-to-rest solution of the matter to be transported, given an estimation of the mass of that matter, the acoustic force acting on the matter, and a path from start to target position.
\section{Conclusion}
In this paper we proposed the first structured numerical approach to compute trap trajectories for acoustic levitation displays.
\emph{OptiTrap} automatically computes physically feasible and nearly time-optimal trap trajectories to reveal generic mid-air shapes, given only a reference path.
Building on a novel multi-dimensional approximation of the acoustic forces around the trap, we formulate and show how to solve a non-linear path following problem without requiring or exploiting differential flatness of the system dynamics.
We demonstrate increases of up to 563\% in size and up to 150\% in frequency for several shapes. Additionally, we obtain better reconstruction accuracy with up to a 79\% decrease in the path-normalised RMSE.
While previously, feasible trap trajectories needed to be tuned manually for each shape and levitator, our approach requires calibration of each individual levitator just once.
We are confident that the ideas in this paper could form the basis for future content authoring tools for acoustic levitation displays and bring them a key step closer to real-world applications.
\begin{acks}
This research has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement \#737087 (Levitate) and from the AHRC UK-China Research-Industry Creative Partnerships (AH/T01136X/2).
\end{acks}
\bibliographystyle{ACM-Reference-Format}
| e22b06c14f5fdc9c20997f1273a1abb6a7428e10 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction\label{sec:1}}
In this section we provide
the background motivating us to examine the present problem; then
we sketch the main results. At the end we give an overview of existing
literature concerning PDEs on domains with small resonators.
\subsection{Motivations}
In what follows, if $\Omega$ is a domain in $\mathbb{R}^n$, we denote by $-\Delta_\Omega^D$ and $-\Delta_\Omega^N$ the Dirichlet and the
Neumann Laplacians on $\Omega$, respectively.
Also, if ${\mathcal{A}}$ is a self-adjoint operator with purely discrete spectrum bounded from below and accumulating at $\infty$, we denote by $\lambda_k({\mathcal{A}})$ its $k$th eigenvalue; as usual the eigenvalues are arranged in the ascending order and repeated according to their multiplicities.
\smallskip
It is known that the spectrum of
the Dirichlet Laplacian is stable under
small perturbations of a domain (see, e.g., \cite[Theorem~1.5]{RT75} for a precise statement).
The situation becomes more subtle for the Laplacian with the Neumann or mixed boundary conditions.
The example below demonstrating this goes back to the Courant-Hilbert
monograph \cite{CH53}; later it was elaborated in more details by Arrieta, Hale and Han in \cite{AHH91}.
Let $\Omega$ be a bounded domain in $\mathbb{R}^n$, $n\ge 2$. We perturb it to a domain $\Omega_{\varepsilon}$ by
attaching a small \emph{resonator} (in Remark~\ref{rem:reso} we will explain why such a name is reasonable) --
the union of a bounded open set $B_{\varepsilon}$ of the diameter $\varepsilon$ and
a narrow passage $P_{\varepsilon}$ of the length $\varepsilon$ and the cross-section diameter
$\mathcal{O}(\varepsilon^\alpha)$; see Figure~\ref{fig-CH}.
Since $\Omega_{\varepsilon}$ and $\Omega$ differs only in a $\mathcal{O}(\varepsilon)$-neighbourhood
of some point on $\partial\Omega$, the first naive guess is that the $k$th eigenvalues
of the Laplacians on $\Omega_{\varepsilon}$ and $\Omega$ are close as $\varepsilon\ll 1$.
This is indeed true for the Dirichlet case:
it is not hard to show\footnote{One can prove
\eqref{lambdaD:conv} using \cite[Theorem~{1.5}]{RT75} and the fact that
$\Omega_{\varepsilon}$ converges metrically to $\Omega$.}
that
\begin{gather}\label{lambdaD:conv}
\forall k\in\mathbb{N}:\quad\lambda_k(-\Delta_{\Omega_{\varepsilon}}^D)\to \lambda_k(-\Delta_{\Omega}^D)\text{ as }\varepsilon\to 0,
\end{gather}
while, for the Neumann eigenvalues this property fails:
one has $\lambda_1(-\Delta_{\Omega_{\varepsilon}}^N)=0$, and for the next eigenvalues
the following result holds if $\alpha>\frac{n+1}{n-1}$ \cite[Theorems~4.1,4.2]{AHH91}:
$$\forall k\in\mathbb{N}\setminus\{1\}:\quad \lambda_k(-\Delta_{\Omega_{\varepsilon}}^N)\to\lambda_{k-1}(-\Delta_{\Omega}^N)\text{ as }\varepsilon\to 0,$$
in particular, $\lambda_2(-\Delta_{\Omega_{\varepsilon}}^N)\to 0$ as $\varepsilon\to 0$.
\begin{figure}[t]
\begin{center}
\begin{picture}(200,105)
\includegraphics[height=35mm]{fig-CH.pdf}
\put(-150,60){$\Omega$}
\put(-35,5){$P_{\varepsilon}$}
\put(-13,71){$B_{\varepsilon}$}
\end{picture}
\end{center}
\caption{Attached resonator}\label{fig-CH}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{picture}(150,100)
\includegraphics[height=35mm]{fig-Sch.pdf}
\put(-25,30){$\Omega_{\varepsilon}^{\rm out}$}
\put(-68,78){$P_{\varepsilon}$}
\put(-143,60){$S_{\varepsilon}$}
\put(-95,45){$B_{\varepsilon}$}
\end{picture}
\end{center}
\caption{Inserted resonator}\label{fig-Sch}
\end{figure}
Instead of attaching a resonator, one can also insert it \emph{inside} $\Omega$.
Such a problem was studied by Schweizer in \cite{Sch15}.
He considered $\Omega_{\varepsilon}=\Omega\setminus S_{\varepsilon}$,
where $\Omega$ is a bounded domain in $\mathbb{R}^n$, and $S_{\varepsilon}$ is a thin layer separating
the small set $B_{\varepsilon}\subset\Omega$ and the outer domain $\Omega_{\varepsilon}^{\rm out}$;
$B_{\varepsilon}$ and $\Omega_{\varepsilon}^{\rm out}$ are connected through a narrow passage $P_{\varepsilon}$. This domain is depicted on
Figure~\ref{fig-Sch}.
Denote by $L_{\varepsilon}$ and $A_{\varepsilon}$
the length and the cross-section area of the passage $P_{\varepsilon}$, respectively,
and by $V_{\varepsilon}$ the volume of $B_{\varepsilon}$.
Let ${\mathcal{A}}_{\varepsilon}$ be the Laplacian on $\Omega_{\varepsilon}$ subject to the Neumann conditions
on $\partial S_{\varepsilon}$ and the Dirichlet conditions on $\partial\Omega$.
It was shown in \cite{Sch15} that, if $B_{\varepsilon}$ and $P_{\varepsilon}$ are appropriately scaled,
one has
$$
\lambda_1({\mathcal{A}}_{\varepsilon}) \to\gamma\coloneqq\lim_{\varepsilon\to 0}\frac{A_{\varepsilon}}{L_{\varepsilon} V_{\varepsilon}},\quad
\lambda_k({\mathcal{A}}_{\varepsilon})\to\lambda_{k-1}(-\Delta_{\Omega}^D),\ k\in\mathbb{N}\setminus\{1\}\
\text{ as }\varepsilon\to 0;
$$
the result is obtained under the assumption
that $\Omega$ is small enough in order to get
\begin{gather}
\label{ga:restr}
\gamma<\lambda_1(-\Delta_\Omega^D).
\end{gather}
The proof in \cite{Sch15} relies on variational methods.
An important ingredient is a well-known result
for compact self-adjoint operators
claiming that the existence of approximate eigenfunctions implies the existence of nearby eigenvalues;
in \cite{Sch15} this result is applied for the resolvents of ${\mathcal{A}}_{\varepsilon}$ and $-\Delta_{\Omega}^D$.
\smallskip
\emph{Our goal is to extend and complement the results obtained in
\cite{Sch15} to resonators separated from the outer domain by a closed surface with a tiny windows.}
In the two next subsections we sketch the main results of this work.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.28\textwidth]{fig-Helm-small}
\end{center}
\caption{Brass spherical Helmholtz resonator from around 1890-1900.
\href{https://commons.wikimedia.org/wiki/File:Helmholtz_resonator.jpg}{Photograph} by \href{https://en.wikipedia.org/wiki/User:Brian0918?rdfrom=commons:User:Brian0918}{brian0918}.
License: \href{https://creativecommons.org/licenses/by-sa/2.5/deed.en}{CC BY-SA 2.5}.}
\label{fig-Helm}
\end{figure}
\begin{remark}\label{rem:reso}
The reason for using the name ``resonator'' for these domain
perturbations is that their geometry resembles the well-known
\emph{Helmholtz resonators} -- acoustic devices
consisting of a resonator volume being connected by a thin channel to the outer space (Figure~\ref{fig-Helm}). The Helmholtz resonator is characterized by a resonance frequency $\omega_{\rm res}$ which can be calculated by
a widely known asymptotic formula
$$\omega_{\rm res} \approx C\sqrt\frac{A}{LV}.$$
Here $C$ is the speed of sound, $V$ is the resonator volume,
$L$ is the channel length and $A$ is the channel cross section area.
Apparently, Schweizer's work \cite{Sch15} provides the first mathematically rigorous derivation of the resonator frequency formula.
\end{remark}
\subsection{Sketch of the main convergence results}
Let $\Omega$ be a (not necessary bounded) domain in $\mathbb{R}^n$, $n\ge 2$. Let $m\in\mathbb{N}$.
Let $\varepsilon$ and $d_{k,\varepsilon}$, $k=1,\dots,m$ be small positive parameters such that $d_{k,\varepsilon}=o(\varepsilon)$ as $\varepsilon\to 0$.
We introduce the sets
\begin{gather*}
S_{k,\varepsilon}=\partial B_{k,\varepsilon}\setminus D_{k,\varepsilon}.
\end{gather*}
Here $B_{k,\varepsilon}$ (``resonator'') is a subset of $\Omega$ such that $B_{k,\varepsilon}\cong\varepsilon B_k$, where
$B_k\subset\mathbb{R}^n$ is a bounded domain,
$D_{k,\varepsilon}$ (``window'') is a subset of $\partial B_{k,\varepsilon}$ such that
$D_{k,\varepsilon}\cong d_{k,\varepsilon} D_k$, where $D_k$ is a bounded set on an $(n-1)$-dimensional hyperplane. One has few other technical assumptions on the geometry of the above sets, whose description we postpone to Section~\ref{sec:2}. Removing the sets $S_{k,\varepsilon}$ from $\Omega$ we get the domain (Figure~\ref{fig-main})
$$\displaystyle\Omega_{\varepsilon}=\Omega\setminus\left(\bigcup\limits_{k=1}^m S_{k,\varepsilon}\right).$$
\begin{figure}[h]
\begin{center}
\begin{picture}(180,125)
\includegraphics{fig-main.pdf}
\put(-83,83){${S_{k,\varepsilon}}$}
\put(-105,30){${B_{k,\varepsilon}}$}
\put(-135,106){${D_{k,\varepsilon}}$}
\end{picture}
\end{center}
\caption{The domain $\Omega_{\varepsilon}$; here $m=2$}\label{fig-main}
\end{figure}
We denote by ${\mathcal{A}}_{\varepsilon}$ the Laplacian on $\Omega_{\varepsilon}$ subject to the
Dirichlet conditions on $\partial\Omega$ and the Neumann conditions on $S_{k,\varepsilon}$.
Assume that for each $k\in\{1,\dots,m\}$ the limit
$$\gamma_k\coloneqq \lim_{\varepsilon\to 0}\frac{\mathrm{cap}(D_{k,\varepsilon})}{4|B_{k,\varepsilon}| }$$
exists and is finite; here $\mathrm{cap}(D_{k,\varepsilon})$ and $|B_{k,\varepsilon}|$ stand for the capacity of $D_{k,\varepsilon}$ and
the volume of $B_{k,\varepsilon}$, respectively. Further, in the Hilbert space $\L(\Omega)\oplus\mathbb{C}^m$ we introduce the operator
${\mathcal{A}}$ via
$$
{\mathcal{A}}=(-\Delta_{\Omega}^D)\oplus\Gamma,
$$
where
$
\Gamma:\mathbb{C}^m\to\mathbb{C}^m$ is a diagonal matrix with the numbers $\gamma_k,\ k\in\mathbb{M}$,
standing on the main diagonal.
Evidently, one has
$$\sigma({\mathcal{A}})=\sigma(-\Delta_\Omega^D)\cup\{\gamma_k,\ k=1,\dots,m\}.$$
Our main result reads as follows.
\begin{theorem*}
The spectrum of ${\mathcal{A}}_{\varepsilon}$ converges to the spectrum of ${\mathcal{A}}$
in the Hausdorff sense as $\varepsilon\to 0$, i.e.
\begin{itemize}
\item
$\forall \lambda\in \sigma({\mathcal{A}})$ there exists a family $(\lambda_{\varepsilon})_{\varepsilon>0}$ with
$\lambda_{\varepsilon}\in \sigma({\mathcal{A}}_{\varepsilon})$ such that $\lambda_{\varepsilon}\to \lambda$ as $\varepsilon\to 0$, and
\item
$\forall \lambda\in\mathbb{R}\setminus \sigma({\mathcal{A}})$ there exists $\delta>0$
such that $\sigma({\mathcal{A}}_{\varepsilon})\cap (\lambda-\delta,\lambda+\delta)=\varnothing$ for small enough $\varepsilon$.
\end{itemize}
\end{theorem*}
We derive an estimate on the rate of this convergence with respect to the
Hausdorff metrics (see~\eqref{th1:est}).
If the domain $\Omega$ is bounded, we get
\begin{gather*}
\forall k\in\mathbb{N}:\ \lambda_k({\mathcal{A}}_{\varepsilon})\to\lambda_k({\mathcal{A}})\text{ as }\varepsilon\to 0.
\end{gather*}
We also demonstrate the convergence of eigenfunctions.
The precise results are formulated in Section~\ref{sec:2} -- see Theorem~\ref{th1}, \ref{th2} and Corollary~\ref{coro:bounded}.
Note that, unlike \cite{Sch15} (cf.~\eqref{ga:restr}),
we put no restrictions on the location of the numbers $\gamma_k$.
Also note that the type of the boundary conditions on the external boundary of $\Omega_{\varepsilon}$ (i.e., on $\partial\Omega$) is not essential in our analysis, see Remark~\ref{rem:Neumann} at the end of Section~\ref{sec:2} for more details.
\subsection{Application: a waveguide with prescribed eigenvalues}
It is an interesting and long-standing topic of spectral theory --
the design of domains with
prescribed spectral properties.
For example, in the celebrated paper \cite{CdV87} Y.~Colin de Verdi\'ere
constructed a bounded domain $\Omega$ such that the
first $m$ non-zero eigenvalues of the Neumann Laplacian on $\Omega$ coincide with $m$ predefined pairwise distinct positive numbers.
In \cite{HSS91} Hempel, Seco and Simon
constructed a bounded domain such that the essential spectrum of the
Neumann Laplacian on $\Omega$ coincides with the predefined closed set $S$;
in the case
$0\in S$ the domain $\Omega$ looks like a chain of ``rooms'' and ``passages''.
This result was further elaborated by Hempel, Kriecherbauer and Plankensteiner in \cite{HKP97}, where also
a prescribed bounded part of the discrete spectrum was realized; the
designed domain has a ``comb'' structure.
We refer to the recent overview \cite{BK19}
for more references on the construction of Laplacians (and other self-adjoint operators) with the predefined
(or partly predefined) spectrum.
We are aimed to contribute to the above topic taking advantage of the convergence results we sketched in the previous subsection. These results suggest that by inserting small resonators one can create new eigenvalues having nothing in common with the eigenvalues of the Laplacian on the unperturbed domain. Moreover, it is easy to see that by a proper choice of the parameters $d_{k,\varepsilon}$ characterizing the windows sizes, one can make these new eigenvalue convergent ($\varepsilon\to 0$) to predefined numbers.
It turns out that one can achieve even more -- \emph{the precise
coincidence of these new eigenvalues with prescribed numbers} (for fixed small enough $\varepsilon$). To fix the ideas, we discuss an example demonstrating how to do this.
\begin{figure}[h]
\begin{center}
\begin{picture}(280,50)
\includegraphics[width=0.7\linewidth]{fig-waveguide.pdf}
\put(-55,15){$\Omega^+$}
\put(-220,15){$\Omega^-$}
\put(-146,40){$\widetilde\Omega_{\varepsilon}$}
\end{picture}
\end{center}
\caption{Waveguide $\Omega_{\varepsilon}$}
\label{fig-waveguide}
\end{figure}
Let the unperturbed domain be the straight unbounded tube (waveguide).
It is known that the spectrum of the Dirichlet Laplacian on such a domain
coincides with $[\Lambda',\infty)$, where $\Lambda'>0$ is the smallest eigenvalue of the Dirichlet Laplacian on the tube cross-section. We perturb $\Omega$ by narrowing the tube in some bounded part
and then by inserting $m$ resonators within this narrowed part (see Figure~\ref{fig-waveguide} -- here $\widetilde\Omega_{\varepsilon}$ is a narrowed part with resonators, and $\Omega^\pm$ are semi-infinite straight tubes).
Such a perturbation does not change the essential spectrum, but may produce
discrete eigenvalues below $\Lambda'$. \emph{We prove that the resonators actually can be chosen in such way that these eigenvalues do coincide with prescribed numbers.} The role of the narrowing is to guaranteed that only $m$ eigenvalues appear below $\Lambda'$, and no further eigenvalues emerge in the vicinity of $\Lambda'$.
The precise result is formulated in Section~\ref{sec:5} -- see Theorem~\ref{th:exact}.
The proof is based on the multi-dimensional version
of the intermediate value theorem established in \cite{HKP97}.
One of the key-points of the proof is the monotonicity of
eigenvalues of ${\mathcal{A}}_{\varepsilon}$ with respect to the parameters $d_{k,\varepsilon}$ characterizing
the sizes of the windows $D_{k,\varepsilon}$.
This is an advantage of the resonators we treat in this work
comparing to the resonators as on Figures~\ref{fig-CH} and \ref{fig-Sch}, for which the monotonicity
of eigenvalues (with respect to any of the involved geometrical parameters) is not at all obvious.
\subsection{Methods}
As it was already mentioned, we do not assume the boundedness of $\Omega$,
whence the resolvents of ${\mathcal{A}}_{\varepsilon}$ and ${\mathcal{A}}$ are non-compact operators in general.
Therefore, we cannot rely on the methods used in \cite{Sch15}.
Instead, our proofs utilize the abstract results for studying the convergence of operators
in varying Hilbert spaces developed in \cite{P06,P12,KP21}.
We recall them in Section~\ref{sec:3}.
The proof of the main results is given is Section~\ref{sec:4}.
Our main task is to show that ${\mathcal{A}}_{\varepsilon}$ converges to ${\mathcal{A}}$ as $\varepsilon\to 0$ in a generalized norm-resolvent sense
(see Definition~\ref{def:gnrc}), which requires to construct suitable identification operators between
the Hilbert spaces $\L(\Omega_{\varepsilon})$ and $\L(\Omega)\oplus\mathbb{C}^m$, as well as between the energetic spaces associated with the quadratic form generating operators ${\mathcal{A}}_{\varepsilon}$ and ${\mathcal{A}}$.
\subsection{Further literature on domains with resonators}
At the end of this section we give a brief overview of existing
literature concerning differential equations on domains
perturbed via small resonators. Note that in some other works the authors, instead of ``resonators'',
use the names ``traps'' or ``accumulators''.
The interest in PDEs in domains with small resonators
arose in the early 90th in connection with homogenization
theory. It was observed that homogenizing Neumann problems in a domain perforated by a lot of
small resonators one arrives on peculiar effective problems on an unperturbed domain.
The first result was obtained by Khruslov in \cite{Khru89}, where
the linear heat equation was considered. The
effective equation in \cite{Khru89} contains a non-local in time term (model with memory).
In the subsequent papers \cite{BCP99,BP97,Pa92,Ko95} non-linear PDEs were examined.
For more details we refer to the monograph \cite{MK06}.
One can also address another kind of homogenization problems,
when resonators are not distributed inside a domain,
but are attached along (a part of) its boundary.
We considered such a problem in \cite{CK15}, where homogenization
of the Neumann spectral problem for the (weighted) Laplacian was studied.
The obtained effective problem contains spectral parameter
both in the equation and the boundary conditions.
The remarkable applications of models with resonators is that they can be used to design materials with astonishing non-standard properties -- the so-called \emph{meta-materials}. Meta-materials consist
of small components, and,
even though the single component demonstrates ``standard'' behaviour, the meta-material behaves effectively in a way that is not known from ordinary materials.
Lamacz and Schweizer \cite{LS17} studied
acoustic properties of a medium partly filled by small resonators as on Figure~\ref{fig-Sch};
the mathematical model is the Helmholtz equation $-\Delta u_{\varepsilon}=\omega^2 u_{\varepsilon}$ subject to the Neumann conditions on the
resonators boundary (sound hard walls).
The obtained effective equation reads $-\nabla\cdot(A^*\nabla u)=\omega^2\Lambda(\omega) u$ with $A^*\gg 0$ and
a frequency dependent coefficient $\Lambda(\omega)$ (dispersive medium). This coefficient can have negative values for some ranges of the frequency $\omega$, and it also blows up when $\omega$ approaches the resonance frequency of a single resonator (cf.~\cite{Sch15}).
Of course, Helmholtz-type resonators are not the only possible components for the design of meta-materials. One can also use, for example,
wire structures (see, e.g., \cite{FB97}),
split rings structures (see, e.g., \cite{BS10}) or their mix.
We refer to the overview \cite{Sch17} for more details and further references on this topic.
Finally, resonators constitute a good tool for opening of gaps in the spectrum of periodic differential operators.
Recall that the spectrum of periodic self-adjoint differential operators
has the form of a locally finite union of compact intervals (\emph{bands}).
An open interval on $\mathbb{R}$ is called a \emph{gap}
if it has an empty intersection with the spectrum, but its endpoints belong to it.
The band structure of the spectrum suggests that gaps may exist in principle, but, in general, the presence of gaps is not guaranteed: two spectral bands may overlap, and then the corresponding gap disappears.
Existence of spectral gaps are of primary interest because of various applications, for example in physics of photonic crystals -- see, e.g., \cite{DLPSW11} for more details.
It was proven in \cite{Kh14,KK15} that the Neumann Laplacian
on $\mathbb{R}^n$ with $m$ families of periodically distributed resonators has at least $m$ spectral gaps;
their location and lengths of can be controlled by
a suitable choice of the resonators sizes.
Close results for waveguide-like domains with periodically attached
resonators were obtained in \cite{CK17}.
\section{Setting of the problem and main results\label{sec:2}}
Let $n\in\mathbb{N}\setminus\{1\}$. In the following,
$x'=(x^1,\dots,x^{n-1})$ and $x=(x',x^n)$ stand for the Cartesian coordinates in $\mathbb{R}^{n-1}$ and $\mathbb{R}^{n}$, respectively.
For $r>0$ and $z\in\mathbb{R}^n$ we denote
$$\mathscr{B}(r,z)\coloneqq \left\{x\in\mathbb{R}^n:\ |x-z|<r\right\}.$$
Let $\Omega$ be a (not necessary bounded) domain in $\mathbb{R}^n$, $n\ge 2$.
Let $m\in\mathbb{N}$. We set $$\mathbb{M}\coloneqq\{1,\dots,m\}.$$
Let
$B_k$, ${k\in\mathbb{M}}$ be bounded Lipschitz domains in $\mathbb{R}^{n}$.
We assume that there exist positive numbers $\rho_k$, ${k\in\mathbb{M}}$ such that
\begin{gather}
\label{DB1}
\forall k\in\mathbb{M}:\quad \mathscr{B}(\rho_k,0)\cap B_k = \mathscr{B}(\rho_k,0)\cap
\left\{x=(x',x^n)\in\mathbb{R}^n:\ x^n<0\right\},
\end{gather}
whence, in particular, the boundary of $B_k$ is flat in the $\rho_k$-neighbourhood of the origin.
Further, let $\widetilde D_k$, ${k\in\mathbb{M}}$ be bounded Lipschitz domains in $\mathbb{R}^{n-1}$, and
\begin{gather}\label{Dk}
D_k\coloneqq \left\{x=(x',x^n)\in\mathbb{R}^n:\ x' \in \widetilde D_k,\ x^n=0\right\}.
\end{gather}
We assume that
\begin{gather}
\label{DB2}
\begin{array}{l}
\text{the smallest ball containing $D_k$ has center at the origin,}\\
\text{and its radius $\ell_k$ satisfies }
\ell_k< \rho_k.
\end{array}
\end{gather}
In particular, it follows from \eqref{DB1}, \eqref{DB2} that ${D_k}\subset\partial B_k$.
Let $\varepsilon>0$ be a small parameter. Let $d_{k,\varepsilon}$, ${k\in\mathbb{M}}$ be positive numbers satisfying
\begin{gather}\label{deps}
d_{k,\varepsilon}<\varepsilon
\end{gather}
Below we specify $d_{k,\varepsilon}$ more precisely, see~\eqref{gamma:2}. Let $z_k$, ${k\in\mathbb{M}}$ be pairwise distinct points in $\Omega$. For $k\in\mathbb{M}$ we define
\begin{gather}\label{DBe}
B_{k,\varepsilon}\coloneqq\varepsilon B_k+z_k,\quad
D_{k,\varepsilon}\coloneqq d_{k,\varepsilon} D_k+z_k.
\end{gather}
Due to \eqref{DB1}, \eqref{DB2}, \eqref{deps}
we have ${D_{k,\varepsilon}}\subset \partial B_{k,\varepsilon}$, and
one has (see also Figure~\ref{fig-circles} in Section~\ref{sec:4}):
\begin{gather}\label{DBeps}
\mathscr{B}(\rho_k\varepsilon,z_k)\cap B_{k,\varepsilon} = \mathscr{B}(\rho_k\varepsilon,z_k)\cap
\left\{x=(x',x^n)\in\mathbb{R}^n:\ x^n<z_k^n\right\},
\end{gather}
where $z_k^n$ stands for the $x^n$-\,coordinate of $z_k$.
We assume that $\varepsilon$ is sufficiently small so that
\begin{gather*}
\overline{B_{k,\varepsilon}}\subset\Omega \quad\text{and}\quad
\overline{B_{k,\varepsilon}}\cap \overline{B_{l,\varepsilon}}=\emptyset\text{ if }k\not=l.
\end{gather*}
Finally, we define the domain $\Omega_{\varepsilon}$ (see Figure~\ref{fig-main}) by
\begin{gather}\label{Omega:e}
\Omega_{\varepsilon}\coloneqq \Omega\setminus\left(\bigcup\limits_{k\in\mathbb{M}} S_{k,\varepsilon}\right),\text{ where }S_{k,\varepsilon}\coloneqq \partial B_{k,\varepsilon}\setminus D_{k,\varepsilon}.
\end{gather}
Now, we impose extra conditions on the sizes of the windows $D_{k,\varepsilon}$.
We assume that for each $k\in\mathbb{M}$ the following limit
exists and is finite:
\begin{gather}\label{gamma:1}
\gamma_k\coloneqq\lim\limits_{\varepsilon\to 0}\gamma_{k,\varepsilon},\quad
\text{where }\gamma_{k,\varepsilon}\coloneqq \frac{\mathrm{cap}( {D_{k,\varepsilon}})}{4|B_{k,\varepsilon}|}.
\end{gather}
Here $|B_{k,\varepsilon}|$ stands for the volume of the domain $B_{k,\varepsilon}$, and
$\mathrm{cap}({D_{k,\varepsilon}})$ stands for the capacity of the set $D_{k,\varepsilon}$;
the latter one is defined via
\begin{gather}\label{capty}
\mathrm{cap}(D_{k,\varepsilon})=
\begin{cases}
\|\nabla H_{k,\varepsilon}\|_{\L(\mathbb{R}^n)}^2,&n\ge 3,\\
\|\nabla H_{k,\varepsilon}\|_{\L(\mathscr{B}(1,z_k))}^2,&n=2,
\end{cases}
\end{gather}
where $H_{k,\varepsilon}(x)$ is the solution to the problem
\begin{gather}\label{BVP:n3}
\begin{cases}
\Delta H(x)=0,&x\in\mathbb{R}^n\setminus\overline{D_{k,\varepsilon}},\\
H=1,&x\in\partial D_{k,\varepsilon} =\overline{D_{k,\varepsilon}},\\
H\to 0,&|x|\to\infty,
\end{cases}
\end{gather}
if $n\ge 3$, or to the problem
\begin{gather}\label{BVP:n2}
\begin{cases}
\Delta H(x)=0,&x\in \mathscr{B}(1,z_k)\setminus\overline{D_{k,\varepsilon}},\\
H=1,&x\in\partial D_{k,\varepsilon} =\overline{D_{k,\varepsilon}},\\
H= 0,&x\in\partial \mathscr{B}( 1,z_k),
\end{cases}
\end{gather}
if $n=2$; here we assume that $\varepsilon$ is sufficiently small in order to have
$\overline{D_{k,\varepsilon}}\subset \mathscr{B}(1,z_k)$. Note that the capacity of $D_{k,\varepsilon}$ is positive despite its Lebesgue measure is zero.
One has the following asymptotics in the two-dimensional case
\cite[Lemma~3.3]{CDG02}:
\begin{gather}\label{cap:rescaling1}
\mathrm{cap}( {D_{k,\varepsilon}})=2\pi|\ln d_{k,\varepsilon}|^{-1}(1+\alpha(\varepsilon))\text{ with } \lim_{\varepsilon\to 0}\alpha(\varepsilon)= 0 ,\ n=2 ,
\end{gather}
while in higher dimensions, by using simple re-scaling arguments, we get
\begin{gather}\label{cap:rescaling2}
\mathrm{cap}( {D_{k,\varepsilon}})=(d_{k,\varepsilon})^{n-2}\mathrm{cap}( {D_k}),\ n\geq 3.
\end{gather}
It follows from \eqref{gamma:1}, \eqref{cap:rescaling1}, \eqref{cap:rescaling2} that
\begin{gather}\label{gamma:2}
d_{k,\varepsilon}= C_{\varepsilon}\varepsilon^{\frac{n}{n-2}}\text{ if }n\ge 3
\qquad\text{and}\qquad
|\ln d_{k,\varepsilon}|^{-1}= C_{\varepsilon}\varepsilon^2\text{ if }n= 2,
\end{gather}
where $C_{\varepsilon}=\mathcal{O}(1)$ as $\varepsilon\to 0$.
To define the Laplace operator on $\Omega_{\varepsilon}$ subject to the Neumann boundary conditions on $\cup_{k\in\mathbb{M}}S_{k,\varepsilon}$ and the Dirichlet boundary conditions on $\partial\Omega$, we need an appropriate Sobolev space, which we denote $\H_{0,\partial \Omega}^1(\Omega_{\varepsilon})$.
If $\partial\Omega$ is sufficiently regular, we defined it simply by
$$\H_{0,\partial \Omega}^1(\Omega_{\varepsilon})=\left\{u\in \H^1(\Omega_{\varepsilon}):\ u\hspace{-3pt}\restriction_{\partial\Omega}=0\right\},$$
where $u\hspace{-3pt}\restriction_{\partial\Omega}$ is understood in the sense of traces.
In the general case, following \cite{Hemp06}, we define it as follows.
Let $\Sigma'$ and $\Sigma''$ be bounded smooth domains in $\mathbb{R}^n$ satisfying
$$
\cup_{k\in\mathbb{M}}S_{k,\varepsilon}\subset\joinrel\subset\Sigma''\subset\joinrel\subset\Sigma'\subset\joinrel\subset\Omega
$$
($D\subset\joinrel\subset G$ means $\overline{D}\subset G$).
Let $\psi\in C^\infty(\mathbb{R}^n)$ with $\psi\hspace{-3pt}\restriction_{\mathbb{R}^n\setminus\Sigma'}=1$,
$\psi\hspace{-3pt}\restriction_{\Sigma''}=0$.
We define
$$ \H_{0,\partial \Omega}^1(\Omega_{\varepsilon})\coloneqq \left\{u\in \H^1(\Omega_{\varepsilon}):\ \psi u\in \H^1_0(\Omega)\right\}.$$
It is easy to show that the space $\H_{0,\partial \Omega}^1(\Omega_{\varepsilon})$ is independent of the choice of
$\Sigma'$, $\Sigma''$, $\psi$.
Now, we can introduce the operator ${\mathcal{A}}_{\varepsilon}$.
In the Hilbert space $\L(\Omega_{\varepsilon})$ we define the sesquilinear form $\a_{\varepsilon}$ by
\begin{equation}\label{ae}
\a_{\varepsilon}[u,v]=\int_{\Omega_{\varepsilon}}\nabla u\cdot\overline{\nabla v}\,\d x,\qquad
\mathrm{dom}(\a_{\varepsilon})=\H_{0,\partial \Omega}^1(\Omega_{\varepsilon}).
\end{equation}
The form $\a_{\varepsilon}$ is densely defined in $\L(\Omega)$, nonnegative, and closed.
By the first representation
theorem (see, e.g. \cite[Chapter 6, Theorem 2.1]{Ka66}) there is a unique nonnegative
self-adjoint operator ${\mathcal{A}}_{\varepsilon}$ in $\L(\Omega)$ associated with $\a_{\varepsilon}$, i.e.,
$\mathrm{dom}({\mathcal{A}}_{\varepsilon})\subset\mathrm{dom}(\a_{\varepsilon})$ and
\begin{equation*}
({\mathcal{A}}_{\varepsilon} u, v)_{\L(\Omega_{\varepsilon})}=\a_{\varepsilon}[u,v],\quad \forall u\in \mathrm{dom}({\mathcal{A}}_{\varepsilon}),\,\,v\in\mathrm{dom}(\a_{\varepsilon}).
\end{equation*}
Further, we define the limiting operator ${\mathcal{A}}$.
In the following, for $f\in\L(\Omega)\oplus \mathbb{C}^m$ we denote its
$\L(\Omega)$ and $\mathbb{C}^m$ components by $f_0$ and $(f_k)_{k\in\mathbb{M}}$, respectively.
We define the sesquilinear form $\a$ in $\L(\Omega)\oplus \mathbb{C}^m$ by
\begin{gather}\label{a}
\displaystyle\a[f,g]=\int_{\Omega}\nabla f_0\cdot\overline{\nabla g_0}\,\d x +
\sum\limits_{k\in\mathbb{M}}\gamma_k f_k\overline{g_k},
\quad\mathrm{dom}(\a)= \H^1_0(\Omega)\times \mathbb{C}^m.
\end{gather}
This form is densely defined in $\L(\Omega)\oplus \mathbb{C}^m$, nonnegative, and closed.
We denote by ${\mathcal{A}}$ the nonnegative
self-adjoint operator in $\L(\Omega)\oplus \mathbb{C}^m$ associated with $\a$.
It is easy to see that
\begin{gather}
\label{A:oplus}
{\mathcal{A}}=(-\Delta^D_\Omega)\oplus \Gamma\text{\quad and\quad}\sigma({\mathcal{A}})=\sigma(-\Delta_\Omega^D)\cup \{\gamma_k,\, k\in\mathbb{M}\},
\end{gather}
where
$\Gamma:\mathbb{C}^m\to\mathbb{C}^m$ acts on $f=(f_k)_{k=1}^m\in\mathbb{C}^m$ by
$$\displaystyle(\Gamma f)_k= \gamma_k f_k,\ k\in\mathbb{M}.$$
Our aim is to show that $\sigma({\mathcal{A}}_{\varepsilon})$ converges to $\sigma({\mathcal{A}})$
in a suitable sense as $\varepsilon\to 0$.
\smallskip
Recall that
for closed sets $X,Y\subset\mathbb{R}$ the \emph{Hausdorff distance} $d_H (X,Y)$ is given by
\begin{gather*}
d_H (X,Y)\coloneqq\max\left\{\sup_{x\in X} \inf_{y\in Y}|x-y|;\,\sup_{y\in Y} \inf_{x\in X}|y-x|\right\}.
\end{gather*}
The notion of convergence provided by this metric is too restrictive for our purposes, since
the convergence of $\sigma({\mathcal{A}}_{\varepsilon})$ to $\sigma({\mathcal{A}})$ in the metric $d_H(\cdot,\cdot)$ would mean that $\sigma({\mathcal{A}}_{\varepsilon})$ and $\sigma({\mathcal{A}})$ look nearly the same \emph{uniformly in the whole of $[0,\infty)$} -- a situation which is usually not guaranteed even if ${\mathcal{A}}_{\varepsilon}$ converges to ${\mathcal{A}}$
in (a kind of) the norm resolvent topology.
Therefore, it is convenient to introduce the new metric $\widetilde{d}_H(\cdot,\cdot)$ via
\begin{gather*}
\widetilde{d}_H(X,Y)\coloneqq d_H( \overline{(1+X)^{-1}}, \overline{(1+Y)^{-1}}),\ X,Y\subset[0,\infty),
\end{gather*}
where $(1+X)^{-1}\coloneqq {\{(1+x)^{-1}:\ x\in X\}}$, $(1+Y)^{-1}\coloneqq {\{(1+y)^{-1}:\ y\in Y\}}$.
With respect to this metric two spectra can be close even if they differ significantly at high energies.
To clarify the statement $\widetilde{d}_H(X_{\varepsilon},X)\to 0$ we introduce the following definition.
\begin{definition}
Let $X_{\varepsilon},X\in\mathbb{R}$. One says that
\begin{itemize}
\item \emph{$X_{\varepsilon}$ converges from inside to $X$} (we denote $X_{\varepsilon}\nearrow X$)
if for any $ x\in X$ there exists a family $(x_{\varepsilon})_{\varepsilon>0}$ with
$x_{\varepsilon}\in X_{\varepsilon}$ such that $x_{\varepsilon}\to x$ as $\varepsilon\to 0$;
\item \emph{$X_{\varepsilon}$ converges from outside to $X$} (we denote $X_{\varepsilon}\searrow X$)
if for any $x\in\mathbb{R}\setminus X$ there exist $\delta>0$
such that $X_{\varepsilon}\cap (x-\delta,x+\delta)=\varnothing$ for sufficiently small $\varepsilon$.
\item \emph{$X_{\varepsilon}$ converges to $X$} (we denote $X_{\varepsilon}\to X$)
if simultaneously $X_{\varepsilon}\nearrow X$ and $X_{\varepsilon}\searrow X$.
\end{itemize}
\end{definition}
\begin{remark}\label{rem:equiv}
It is easy to see that $X_{\varepsilon}\searrow X$ iff
for any sequence $(x_{\varepsilon_k})_{k\in\mathbb{N}}$ with $x_{\varepsilon_k}\in X_{\varepsilon_k}$ converging to some $x\in\mathbb{R}$ as $\varepsilon_k\to 0$, one has $x\in X$.
\end{remark}
\begin{lemma}[{\cite[Lemma~A.2]{HN99}}] \label{lemma:HN}
Let $(X_{\varepsilon})_{\varepsilon>0}$ be a family of closed sets, $X_{\varepsilon}\subset[0,\infty)$.
Let $X\subset [0,\infty)$ be a closed set. Then
$
\widetilde{d}_H(X_{\varepsilon},X)\to 0$ if and only if $X_{\varepsilon}\to X$.
\end{lemma}
We are now in position to formulate the main convergence results.
In what follows, by $C,C_1,C_2,\dots$ we denote generic positive constants being independent of $\varepsilon$.
\begin{theorem}\label{th1}
One has
$\sigma({\mathcal{A}}_{\varepsilon})\to \sigma({\mathcal{A}})$ as $\varepsilon\to 0$, and the estimate
\begin{gather} \label{th1:est}
\widetilde{d}_H\left(\sigma({\mathcal{A}}_{\varepsilon}),\sigma({\mathcal{A}})\right)\leq
\begin{cases}
\displaystyle C \sum_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+C\varepsilon ,&n\ge 3,\\
\displaystyle C \sum_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+C\varepsilon|\ln\varepsilon|^{3/2} ,&n=2,
\end{cases}
\end{gather}
holds for sufficiently small $\varepsilon$.
\end{theorem}
As usual, we denote by
$\sigma_{{\rm ess}}(\cdot)$ and $\sigma_{{\rm disc}}(\cdot)$ the essential and the discrete spectra of a self-adjoint operator.
Using standard methods of perturbation theory, one infers that inserting
finitely many compact Lipschitz surfaces $S_{k,\varepsilon}$ inside $\Omega$ and posing the Neumann boundary conditions on them, one does not change the essential spectrum of the Laplacian, i.e.
\begin{gather}\label{essAA}
\sigma_{\rm ess}({\mathcal{A}}_{\varepsilon})=\sigma_{\rm ess}(-\Delta_\Omega^D)=\sigma_{\rm ess}({\mathcal{A}})
\end{gather}
(the second equality in \eqref{essAA} follows from \eqref{A:oplus}). For the discrete spectrum of
${\mathcal{A}}_{\varepsilon}$ we have the following result.
\begin{theorem}\label{th2}
One has
$
\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\nearrow \sigma_{\rm disc}({\mathcal{A}})\text{ as }\varepsilon\to 0.$
The multiplicity is preserved:
if $\lambda\in\sigma_{\rm disc}({\mathcal{A}})$ is of multiplicity $\mu$ and
$[\lambda-L,\lambda+L]\cap\sigma({\mathcal{A}})=\{\lambda\}$ with $L>0$, then for sufficiently small $\varepsilon$
the spectrum of ${\mathcal{A}}_{\varepsilon}$ in $[\lambda-L,\lambda+L]$ is purely discrete and
the total multiplicity of the eigenvalues of ${\mathcal{A}}_{\varepsilon}$ contained in $[\lambda-L,\lambda+L]$ equals $\mu$.
If, in addition, $\mu=1$ (i.e. the eigenvalue $\lambda$ is simple), and $\psi=(\psi_0,\psi_1,\dots,\psi_m)$ with $\psi_0\in\L(\Omega)$ and $(\psi_k)_{k\in\mathbb{M}}\in \mathbb{C}^m$ is the corresponding normalized in ${\L(\Omega)\oplus\mathbb{C}^m}$ eigenfunction,
then there exists a sequence of normalized in $\L(\Omega_{\varepsilon})$ eigenfunctions $\psi_{\varepsilon}$ of ${\mathcal{A}}_{\varepsilon}$ such that
\begin{gather*}
\|\psi_{\varepsilon} - \psi_0\|^2_{\L(\Omega_{\varepsilon}\setminus \overline{\cup_{k\in\mathbb{M}}B_{k,\varepsilon}})}+
\sum\limits_{k\in\mathbb{M}}\|\psi_{\varepsilon} - |B_{k,\varepsilon}|^{-1/2}\psi_k\|^2_{\L(B_{k,\varepsilon})} \to 0.
\end{gather*}
\end{theorem}
\begin{remark}Let $\lambda$ be a simple eigenvalue of ${\mathcal{A}}$, and let $\lambda_{\varepsilon}\in\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})$ with $\lambda_{\varepsilon}\to\lambda$ as $\varepsilon\to 0$. The second part of
Theorem~\ref{th2} asserts that the normalized eigenfunction, which corresponds to $\lambda_{\varepsilon}$, concentrates on ${\L(\Omega_{\varepsilon}\setminus \overline{\cup_{k\in\mathbb{M}}B_{k,\varepsilon}})}$
if $\lambda$ is an eigenvalue of $-\Delta_\Omega^D$ and concentrates on $B_{k,\varepsilon}$ if $\lambda=\gamma_k$ for some $k\in\mathbb{M}$.
\end{remark}
\begin{remark}
If $\Omega$ is a bounded domain, the spectra of ${\mathcal{A}}_{\varepsilon}$ and ${\mathcal{A}}$ are purely discrete;
hence, due to Theorem~\ref{th1}, additionally to $
\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\nearrow \sigma_{\rm disc}({\mathcal{A}})$ we also have
$\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\searrow\sigma_{\rm disc}({\mathcal{A}})$ as $\varepsilon\to 0$.
However, if $\Omega$ is unbounded,
the latter property does not necessary hold true: it may happen that
there exists a sequence $(\lambda_{\varepsilon_k})_{k\in\mathbb{N}}$ with $\lambda_{\varepsilon_k}\in\sigma_{{\rm disc}}({\mathcal{A}}_{\varepsilon_k})$ converging to $\lambda_0\in\sigma_{{\rm ess}}({\mathcal{A}})$ as $\varepsilon_k\to 0$.
In Subsection~\ref{subsec:4:1} we present an example demonstrating this.
\end{remark}
From Theorems~\ref{th1} and \ref{th2} one can easily get the following corollary.
\begin{corollary}\label{coro:bounded}
Let $\Omega$ be a bounded domain (consequently, the spectra of ${\mathcal{A}}_{\varepsilon}$ and ${\mathcal{A}}$ are purely discrete).
Then $\forall k\in\mathbb{N}$ one has $\lambda_k({\mathcal{A}}_{\varepsilon})\to\lambda_k({\mathcal{A}})$ as $\varepsilon\to 0$.
\end{corollary}
\begin{remark}\label{rem:Neumann}
The choice of the boundary conditions on the external boundary of $\Omega_{\varepsilon}$ (i.e., on $\partial\Omega$) plays no essential role in our analysis.
If one prescribes the Neumann, Robin or mixed boundary conditions, all convergence results (Theorems~\ref{th1}, \ref{th2} and Corollary~\ref{coro:bounded}) remains the same, but
with $-\Delta_\Omega^D$ being replaced by the Laplacian on $\Omega$ subject to the Neumann, Robin or mixed boundary conditions, respectively. Up to a few simple technical details the proofs are similar to the Dirichlet case, and to simplify the presentation we omit them.
\end{remark}
\section{Abstract tools\label{sec:3}}
In this section we present an abstract scheme for studying the convergence of operators
in varying Hilbert spaces proposed by Post in \cite{P06}, and further elaborated in the monograph \cite{P12}.
Initially, this scheme was applied to study convergence of the Laplace-Beltrami operator on ``fat'' graphs;
later it also has shown to be effective to investigate resolvent and spectral convergence
in domains with holes \cite{KPl21,KP18,AP21}.
Let $(\mathcal{H}_{\varepsilon})_{\varepsilon>0}$ be a family of Hilbert spaces,
$({\mathcal{A}}_{\varepsilon})_{\varepsilon>0}$ be a family of non-negative, self-adjoint, unbounded operators in $\mathcal{H}_{\varepsilon}$, $(\a_{\varepsilon})_{\varepsilon>0}$ be a family of associated
sesquilinear forms. Also, let $\mathcal{H}$ be a Hilbert space, ${\mathcal{A}}_{\varepsilon}$ be a non-negative, self-adjoint, unbounded operator in $\mathcal{H}$, and $\a$ be the associated
sesquilinear form.
Along with $\mathcal{H}_{\varepsilon}$ and $\mathcal{H}$ we define the Hilbert spaces $\mathcal{H}^1_{\varepsilon}$ and $\mathcal{H}^1$ (energetic space associated with the forms $\a_{\varepsilon}$ and $\a$) via
\begin{equation}\label{scale:1}
\begin{array}{ll}
\mathcal{H}^1_{\varepsilon}=\mathrm{dom}(\a_{\varepsilon}),&
\|u\|_{\mathcal{H}^1_{\varepsilon}}^2=\a_{\varepsilon}[u,u]+\|u\|^2_{\mathcal{H}_{\varepsilon}},\\[2mm]
\mathcal{H}^1=\mathrm{dom}(\a),&
\|f\|_{\mathcal{H}^1}^2=\a[f,f]+\|f\|^2_{\mathcal{H}}.
\end{array}
\end{equation}
The definition below generalizes the standard notion of norm resolvent convergence
to the setting of varying Hilbert spaces.
\begin{definition}[{\cite{P12}}]\label{def:gnrc}
We say that ${\mathcal{A}}_{\varepsilon}$ converges to ${\mathcal{A}}$ in the generalized norm-resolvent sense
(${\mathcal{A}}_{\varepsilon}\overset{g.n.r.c}\to {\mathcal{A}}$) as $\varepsilon\to 0$ if there is a sequence of $(\delta_{\varepsilon})_{\varepsilon>0}$ with $\delta_{\varepsilon}\to 0$
and linear bounded operators
$\J_{\varepsilon} \colon \mathcal{H}\to \mathcal{H}_{\varepsilon}$, ${\widetilde\J_{\varepsilon} }\colon {\mathcal{H}_{\varepsilon}}\to \mathcal{H},$
satisfying the conditions
\begin{align}
\label{C1}
|(u,\J_{\varepsilon} f)_{\mathcal{H}_{\varepsilon}} - (\widetilde\J_{\varepsilon} u,f)_{\mathcal{H}}| \leq \delta_{\varepsilon}\|f\|_{\mathcal{H}}\|u\|_{\mathcal{H}_{\varepsilon}},&\quad \forall f\in\mathcal{H},\, u\in\mathcal{H}_{\varepsilon},
\\
\label{C2}
\|f-\widetilde\J_{\varepsilon} \J_{\varepsilon} f \|_{\mathcal{H}} \leq \delta_{\varepsilon}\|f\|_{\mathcal{H}^1},&\quad \forall f\in\mathcal{H}^1,
\\ \label{C3}
\|u-\J_{\varepsilon}\widetilde \J_{\varepsilon} u \|_{\mathcal{H}_{\varepsilon}}\leq \delta_{\varepsilon}\|u\|_{\mathcal{H}_{\varepsilon}^1},&\quad \forall u\in \mathcal{H}^1_{\varepsilon},
\\
\label{C4}
\|\J_{\varepsilon} f \|_{\mathcal{H}_{\varepsilon}} \leq 2\|f\|_{\mathcal{H}},&\quad \forall f\in \mathcal{H},
\\
\label{C5}
\|({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}\J_{\varepsilon} f - \J_{\varepsilon} ({\mathcal{A}}+\mathrm{I})^{-1} f \|_{\mathcal{H}_{\varepsilon}} \leq \delta_{\varepsilon}\|f\|_{\mathcal{H}},&\quad\forall f\in \mathcal{H}.
\end{align}
\end{definition}
\begin{remark}
For the sake of clarity, we formulate Definition~\ref{def:gnrc} in a slightly different form than the one in \cite{P12}. For example, \eqref{C1} is written in \cite{P12} in the following equivalent form:
$$\|\J_{\varepsilon}-(\widetilde\J_{\varepsilon})^*\|_{\mathcal{H}\to\mathcal{H}_{\varepsilon}}\leq \delta_{\varepsilon}.$$
In fact, the above definition is a mix of Definitions~4.1.1, 4.2.1, 4.2.3, 4.2.6 from \cite{P12}.
\end{remark}
Classical perturbation theory yields that the norm resolvent convergence of self-adjoint operators in a \textit{fixed} Hilbert space implies the convergence of their spectra.
The theorem below extends this result to the setting of varying Hilbert spaces.
\begin{theorem}[{\cite[Theorems~4.3.3 \& 4.3.5]{P12}}]
\label{thA1}
Let ${\mathcal{A}}_{\varepsilon}\overset{g.n.r.c}\to {\mathcal{A}}$ as $\varepsilon\to 0$.
Then
$$
\sigma({\mathcal{A}}_{\varepsilon})\to \sigma({\mathcal{A}}),\quad
\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\nearrow\sigma_{\rm disc}({\mathcal{A}})
\quad\text{as}\quad\varepsilon\to 0.$$
The multiplicity is preserved: if $\lambda\in\sigma_{\rm disc}({\mathcal{A}})$ is of multiplicity $\mu$ and
$[\lambda-L,\lambda+L]\cap\sigma({\mathcal{A}})=\{\lambda\}$ with $L>0$, then
for sufficiently small $\varepsilon$ the spectrum of
${\mathcal{A}}_{\varepsilon}$ in $[\lambda-L,\lambda+L]$ is purely discrete
and
the total multiplicity of the eigenvalues of ${\mathcal{A}}_{\varepsilon}$ contained in $[\lambda-L,\lambda+L]$ equals $\mu$.
If, in addition, $\mu=1$ (i.e. the eigenvalue $\lambda$ is simple), and $\psi$ is the corresponding normalized in $\mathcal{H}$ eigenfunction,
then there exists a sequence of normalized in $\mathcal{H}_{\varepsilon}$ eigenfunctions $\psi_{\varepsilon}$ of ${\mathcal{A}}_{\varepsilon}$ such that $$\|\psi_{\varepsilon} - \J_{\varepsilon} \psi \|_{\mathcal{H}_{\varepsilon}}\to 0\text{ as }\varepsilon\to 0.$$
\end{theorem}
For two bounded normal operators ${\mathcal{R}}_{\varepsilon}$ and ${\mathcal{R}}$ in a \textit{fixed} Hilbert space $\mathcal{H}$
one has the following estimate \cite[Lemma~A.1]{HN99}:
$$d_{H}(\sigma({\mathcal{R}}_{\varepsilon}),\sigma({\mathcal{R}}))\leq \|{\mathcal{R}}_{\varepsilon}-{\mathcal{R}}\|_{\mathcal{H}\to\mathcal{H}}.$$
Applying it for the resolvents ${\mathcal{R}}_{\varepsilon}\coloneqq ({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}$ and ${\mathcal{R}} \coloneqq ({\mathcal{A}}+\mathrm{I})^{-1}$ of non-negative self-adjoint operators ${\mathcal{A}}_{\varepsilon}$, ${\mathcal{A}}$ acting in $\mathcal{H}$ and taking into account that by spectral mapping theorem $\widetilde{d}_{H}(\sigma({\mathcal{A}}_{\varepsilon}),\sigma({\mathcal{A}}))=d_{H}(\sigma({\mathcal{R}}_{\varepsilon}),\sigma({\mathcal{R}}))$, we get
$$\widetilde d_{H}(\sigma({\mathcal{A}}_{\varepsilon}),\sigma({\mathcal{A}}))\leq \|({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}-({\mathcal{A}}+\mathrm{I})^{-1}\|_{\mathcal{H}\to\mathcal{H}}.$$
The theorem below is an analogue of this result to the setting of varying spaces.
\begin{theorem}[{\cite[Theorem~3.4]{KP21}}]\label{thA2}
Let
$\J_{\varepsilon} \colon\mathcal{H}\to {\mathcal{H}_{\varepsilon}}$, $\widetilde\J_{\varepsilon} \colon\mathcal{H}_{\varepsilon}\to {\mathcal{H}} $ be linear bounded operators satisfying
\begin{align*}
\|({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}\J_{\varepsilon} - \J_{\varepsilon} ({\mathcal{A}}+\mathrm{I})^{-1} \|_{\mathcal{H}\to \mathcal{H}_{\varepsilon}}\leq \rho_{\varepsilon},
\\
\|\widetilde\J_{\varepsilon} ({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1} - ({\mathcal{A}}+\mathrm{I})^{-1}\widetilde\J_{\varepsilon} \|_{\mathcal{H}_{\varepsilon}\to \mathcal{H}}\leq \widetilde\rho_{\varepsilon} ,
\end{align*}
and, moreover,
\begin{align}
\label{thA2:3}
\|f\|^2_{ \mathcal{H}}&\leq \mu_{\varepsilon} \|\J_{\varepsilon} f\|^2_{\mathcal{H}_{\varepsilon}}+\nu_{\varepsilon} \, \a[f,f],\quad \forall f\in \mathrm{dom}( \a),\\
\label{thA2:4}
\|u\|^2_{\mathcal{H}_{\varepsilon}}&\leq \widetilde\mu_{\varepsilon} \|\widetilde\J_{\varepsilon} u\|^2_{\mathcal{H}}+\widetilde\nu_{\varepsilon} \,
\a_{\varepsilon}[u,u],\quad \forall u\in \mathrm{dom}( \a_{\varepsilon})
\end{align}
for some positive constants $\rho_{\varepsilon} ,\,\mu_{\varepsilon} ,\,\nu_{\varepsilon} ,\, \widetilde\rho_{\varepsilon} ,\,\widetilde\mu_{\varepsilon} ,\,\widetilde\nu_{\varepsilon} $.
Then one has
\begin{gather}\label{thA2:est}
\widetilde{d}_H\left(\sigma({\mathcal{A}}_{\varepsilon}),\,\sigma({\mathcal{A}})\right)\leq
\max
\left\{
{\nu_{\varepsilon}\over 2}+\sqrt{{\nu_{\varepsilon}^2\over 4}+\rho_{\varepsilon}^2\mu_{\varepsilon}};\,
{\widetilde\nu_{\varepsilon}\over 2}+\sqrt{{\widetilde\nu_{\varepsilon}^2\over 4}+\widetilde\rho_{\varepsilon}^2\widetilde\mu_{\varepsilon}}
\right\}.
\end{gather}
\end{theorem}
\begin{remark}
\label{rem:tau}
In fact, in \cite{KP21} the obtained estimate reads
\begin{gather}\label{rem:tau:est}
\widetilde{d}_H\left(\sigma({\mathcal{A}}_{\varepsilon}),\,\sigma({\mathcal{A}})\right)
\leq
\max\left\{
\rho_{\varepsilon}\sqrt{\mu_{\varepsilon}\over \tau};\,{\nu_{\varepsilon}\over 1-\tau };\,
\widetilde\rho_{\varepsilon}\sqrt{\widetilde\mu_{\varepsilon}\over \widetilde\tau};\,{\widetilde\nu_{\varepsilon}\over 1-\widetilde\tau }
\right\},\
\forall\tau,\widetilde\tau\in (0,1).
\end{gather}
Minimizing
the right-hand-side of \eqref{rem:tau:est} over $\tau,\widetilde\tau\in (0,1)$,
one gets the estimate \eqref{thA2:est}.
\end{remark}
\begin{remark}
We established slightly weaker version of Lemma~\ref{thA2} in \cite{CK19}.
\end{remark}
\begin{proposition}[{\cite[Proposition~3.8]{KP21}}]
\label{prop:2/3}
Let the conditions \eqref{C1} and \eqref{C2} (respectively, \eqref{C1} and \eqref{C3})
are fulfilled with $\delta_{\varepsilon}<2/3$. Then
the estimate \eqref{thA2:3} (respectively, estimate \eqref{thA2:4}) is valid with
$$
\mu_{\varepsilon}=1+{4\delta_{\varepsilon}\over 2-3\delta_{\varepsilon}},\
\nu_{\varepsilon}={\delta_{\varepsilon}\over 2-3\delta_{\varepsilon}}\qquad
\text{(respectively, }\widetilde\mu_{\varepsilon}=1+{4\delta_{\varepsilon}\over 2-3\delta_{\varepsilon}},\
\widetilde\nu_{\varepsilon}={\delta_{\varepsilon}\over 2-3\delta_{\varepsilon}}\text{)}.
$$
\end{proposition}
The last theorem gives a useful tool to establish the crucial condition \eqref{C5} in the definition of the generalized norm resolvent convergence. It is well-known (see, e.g., \cite[Theorem~VI.3.6]{Ka66} or \cite[Theorem~VIII.25]{RS72}) that convergence of sesquilinear forms with \emph{common domain} implies norm resolvent convergence of the associated operators. The theorem below is as a generalization of this fact to the setting of varying spaces.
To formulate it we need to define one more Hilbert space $\mathcal{H}^2$ via
\begin{gather} \label{scale:2}
\mathcal{H}^2\coloneqq\mathrm{dom}({\mathcal{A}}),\quad
\|f\|_{\mathcal{H}^2}\coloneqq\|{({\mathcal{A}}+\mathrm{I}) f}\|_{\mathcal{H}}.
\end{gather}
\begin{theorem}[{\cite[Theorem~A.5]{P06}}]
\label{thA3}
Let $\J_{\varepsilon} \colon \mathcal{H}\to \mathcal{H}_{\varepsilon}$, ${\widetilde\J_{\varepsilon} }\colon {\mathcal{H}_{\varepsilon}}\to \mathcal{H}$ be linear bounded operators satisfying \eqref{C1}. Furthermore, let
$\J_{\varepsilon}^1 \colon \mathcal{H}^1\to \mathcal{H}_{\varepsilon}^1$, ${\widetilde\J_{\varepsilon}^1 }\colon {\mathcal{H}^1_{\varepsilon}}\to \mathcal{H}^1$
be linear operators satisfying
\begin{gather}
\label{thA3:1}
\|\J_{\varepsilon}^1 f-\J_{\varepsilon} f\|_{\mathcal{H}_{\varepsilon}}\leq \delta_{\varepsilon}\|f\|_{\mathcal{H}^1 },
\quad \forall f\in \mathcal{H}^1,
\\
\label{thA3:2}
\|\widetilde\J_{\varepsilon}^1 u - \widetilde\J_{\varepsilon} u \|_{\mathcal{H}}\leq \delta_{\varepsilon}\|u\|_{ \mathcal{H}^1_{\varepsilon}},
\quad \forall u\in\mathcal{H}_{\varepsilon}^1,
\\\label{thA3:3}
|\a_{\varepsilon}[u,\J^1_{\varepsilon} f]-\a[\widetilde\J^{1}_{\varepsilon} u,f] | \leq
\delta_{\varepsilon}\|f\|_{\mathcal{H}^2 }\|u\|_{\mathcal{H}^1_{\varepsilon}},\quad \forall f\in \mathcal{H}^2 ,\ u\in \mathcal{H}^1_{\varepsilon} .
\end{gather}
Then
\begin{align}\label{thA3:result}
\|({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}\J_{\varepsilon} -\J_{\varepsilon} ({\mathcal{A}}+\mathrm{I})^{-1}\|_{\mathcal{H}\to\mathcal{H}_{\varepsilon}}\leq 4\delta_{\varepsilon}.
\end{align}
\end{theorem}
\begin{remark}
Tracing the proof of \cite[Theorem~A.5]{P06} one observes that
the estimate \eqref{thA3:result}
remains valid if \eqref{thA3:3} is substituted
by the weaker condition
\begin{gather*}
\left|\a_{\varepsilon}[u,\J^1_{\varepsilon} f]-\a[\widetilde\J^{1}_{\varepsilon} u,f] \right|\leq
\delta_{\varepsilon}\|f\|_{\mathcal{H}^2 }\|u\|_{\mathcal{H}^2_{\varepsilon}},\quad
\forall f\in \mathcal{H}^2 ,\ u\in \mathcal{H}^2_{\varepsilon},
\end{gather*}
where $\mathcal{H}^2_{\varepsilon}\coloneqq\mathrm{dom}({\mathcal{A}}_{\varepsilon})$, $\|u\|_{\mathcal{H}^2_{\varepsilon}}\coloneqq\|{({\mathcal{A}}_{\varepsilon}+\mathrm{I}) f}\|_{\mathcal{H}_{\varepsilon}}$.
Nevertheless, in most of the applications one is able to establish stronger condition~\eqref{thA3:3}.
\end{remark}
\section{Proof of the main results\label{sec:4}}
Recall that $\mathscr{B}(\ell_k ,0)$ is the smallest ball containing
$D_k$,
$( \rho_k)_{k\in\mathbb{M}}$ are positive numbers satisfying \eqref{DB1}, \eqref{DB2},
$(z_k)_{k\in\mathbb{M}}$ are points
standing at the definitions of $B_{k,\varepsilon}$ and $D_{k,\varepsilon}$. Evidently,
\begin{gather}\label{elleps}
\mathscr{B}(\ell_k d_{k,\varepsilon},z_k)\text{ is the smallest ball containing }D_{k,\varepsilon}.
\end{gather}
Further, we denote
\begin{align*}
R_k&\coloneqq {1\over 2}\min\left\{\mathrm{dist}(z_k,\cup_{l\not=k}\{z_l\}); \mathrm{dist}(z_k,\partial\Omega) \right\},\\
\tau_k&\coloneqq \inf\left\{\tau>0:\ B_k\cup\mathscr{B}(\rho_k,0)\subset \mathscr{B}(\tau,0)\right\}\\
&\,=
\inf\left\{\tau>0:\ B_{k,\varepsilon}\cup\mathscr{B}(\rho_k\varepsilon,z_k)\subset \mathscr{B}(\tau\varepsilon,z_k)\right\}.
\end{align*}
It is easy to see that
\begin{equation}\label{BRk:prop}
\overline{\mathscr{B}({R_k },z_k)}\subset\Omega\quad\text{and}\quad
{\mathscr{B}({R_k },z_k)}\cap{\mathscr{B}({R_l },z_l)}=\emptyset\text{ if } k\not=l.
\end{equation}
We also introduce the sets (below $z_k^n$ stands for the $x^n$-\,coordinate of $z_k$)
\begin{align}
\label{Bke+}
B^+_{k,\varepsilon}&\coloneqq \mathscr{B}(\rho_k\varepsilon,z_k)\cap\left\{x=(x',x^n)\in\mathbb{R}^n:\ x^n>z_k^n\right\},
\\\label{Bke-}
B^-_{k,\varepsilon}&\coloneqq \mathscr{B}(\rho_k\varepsilon,z_k)\cap\left\{x=(x',x^n)\in\mathbb{R}^n:\ x^n<z_k^n\right\}.
\end{align}
In the following
we assume that $\varepsilon\in (0,\varepsilon_0]$
with $\varepsilon_0>0$ satisfying
\begin{gather}
\label{tauR}
\varepsilon_0<\min\left\{1, \min_{k\in\mathbb{M}}{R_k\over\tau_k}\right\},
\\
\label{CC}
\sup_{\varepsilon\in(0,\varepsilon_0]}C_{\varepsilon}<\infty,
\\
\label{deps<1}
\sup_{\varepsilon\in(0,\varepsilon_0]} {d_{\varepsilon}\over \varepsilon}<1,
\end{gather}
where $C_{\varepsilon}=\mathcal{O}(1)$ is a constant standing in \eqref{gamma:2};
the existence of $\varepsilon_0$ satisfying \eqref{deps<1} is guaranteed by \eqref{gamma:2}.
We also denote
\begin{gather}\label{kappa}
\widetilde\tau_k\coloneqq{\tau_k\varepsilon_0+R_k\over 2\varepsilon_0}.
\end{gather}
It follows from the definitions of $\ell_k$, $\rho_k$, $\tau_k$, $\widetilde\tau_k$, $R_k$,
and \eqref{DB2}, \eqref{deps}, \eqref{tauR} that
\begin{gather}\label{all:paramaters}
\forall k\in\mathbb{M}\ \forall\varepsilon\in (0,\varepsilon_0]:\quad
\ell_k d_{k,\varepsilon}<\ell_k\varepsilon<\rho_k\varepsilon\le\tau_k\varepsilon<\widetilde\tau_k\varepsilon <R_k.
\end{gather}
Note that $\rho_k=\tau_k$ only if $B_{k,\varepsilon}=B_{k,\varepsilon}^-$.
The inequalities \eqref{all:paramaters} (together with the properties \eqref{DBeps}, \eqref{elleps}) are clarified on Figure~\ref{fig-circles}.
\begin{figure}[h]
\begin{center}
\begin{picture}(250,250)
\scalebox{0.65}{\includegraphics{fig-circles.pdf}}
\put(-60,243){$_{R_k}$}
\put(-78,220){$_{\widetilde\tau_k\varepsilon}$}
\put(-85,195){$_{\tau_k\varepsilon}$}
\put(-100,168){$_{\rho_k\varepsilon}$}
\put(-108,148){$_{\ell_k\varepsilon}$}
\put(-115,135){$_{\ell_k d_{k,\varepsilon}}$}
\end{picture}
\end{center}
\caption{The domain $B_{k,\varepsilon}$ (dark gray color). The part of $\partial B_{k,\varepsilon}$ corresponding to $S_{k,\varepsilon}$ is drawn via solid bold line, the part corresponding to $D_{k,\varepsilon}$ is drawn via dotted line.
The light gray half disc of the radius $\rho_k$ corresponds to $B_{k,\varepsilon}^+$;
the corresponding lower half disc (its boundary is drawn via solid line) corresponds to $B_{k,\varepsilon}^-$. All circles has their center at $z_k$.}
\label{fig-circles}
\end{figure}
We denote
$$\mathcal{H}_{\varepsilon}\coloneqq \L(\Omega_{\varepsilon}),\qquad\mathcal{H}\coloneqq \L(\Omega)\oplus \mathbb{C}^m.$$
Recall that ${\mathcal{A}}_{\varepsilon}$ and ${\mathcal{A}}$ are the self-adjoint non-negative operators
in $\mathcal{H}_{\varepsilon}$ and $\mathcal{H}$, respectively, being
associated with the forms $\a_{\varepsilon}$ and $\a$ defined by
\eqref{ae} and \eqref{a}. We introduce the spaces $\mathcal{H}_{\varepsilon}^1$, $\mathcal{H}^1$, $\mathcal{H}^2$ as in \eqref{scale:1} and \eqref{scale:2}, i.e.
\begin{gather*}
\begin{array}{ll}
\mathcal{H}^1_{\varepsilon}=\H^1_{0,\partial\Omega}(\Omega_{\varepsilon}),
&
\displaystyle\|u\|^2_{\mathcal{H}^1_{\varepsilon}}=\|u\|^2_{\H^1(\Omega_{\varepsilon})},\\[1.5ex]
\mathcal{H}^1=\H_0^1(\Omega)\times\mathbb{C}^m,
&
\displaystyle\|f\|^2_{\mathcal{H}^1}=\|f_0\|^2_{\H^1(\Omega)}+
\sum_{k\in\mathbb{M}}(\gamma_k+1)|f_k|^2,
\\[1ex]
\mathcal{H}^2=\mathrm{dom}(\Delta_\Omega^D)\times \mathbb{C}^m,
&\displaystyle
\|f\|_{\mathcal{H}^2}^2=
\|-\Delta f_0+f_0\|^2_{\L(\Omega)}+\sum_{k\in\mathbb{M}}(\gamma_k+1)^2|f_k|^2.
\end{array}
\end{gather*}
Hereinafter, $f_0$ and $(f_k)_{k\in\mathbb{M}}$ stand, respectively, for
$\L(\Omega)$ and $\mathbb{C}^m$ components of $f\in\mathcal{H}$.
Our goal is to construct suitable operators
$$\J_{\varepsilon} : \mathcal{H} \to \mathcal{H}_{\varepsilon},\quad
\widetilde\J_{\varepsilon} : \mathcal{H}_{\varepsilon} \to \mathcal{H}, \quad
\J_{\varepsilon}^1 : \mathcal{H}^1 \to \mathcal{H}^1_{\varepsilon}, \quad
\widetilde\J_{\varepsilon}^1: \mathcal{H}_{\varepsilon}^1\to \mathcal{H}^1,$$
satisfying the assumptions of abstract Theorems~\ref{thA1}, \ref{thA2}, \ref{thA3}.
\smallskip
The following lemma will be frequently used further.
We denote
\begin{gather*}
\eta_{\varepsilon}\coloneqq
\begin{cases}
\varepsilon^2,& n\ge 3
\\
\varepsilon^2|\ln\varepsilon|, & n=2 .
\end{cases}
\end{gather*}
Recall that ${B_{k,\varepsilon}\cup B_{k,\varepsilon}^+ }\subset \mathscr{B}(R_k,z_k)$ (see \eqref{all:paramaters} and Figure~\ref{fig-circles}).
\begin{lemma}\label{lm:Best}
One has
\begin{gather}\label{lm:Best:est}
\forall u\in \H^1(\mathscr{B}(R_k,z_k)):\quad
\|u\|^2_{\L(B_{k,\varepsilon}\cup B_{k,\varepsilon}^+)}\leq C\eta_{\varepsilon}\|u\|^2_{\H^1(\mathscr{B}(R_k,z_k))}
\end{gather}
\end{lemma}
\begin{proof}
Evidently, it is enough to prove \eqref{lm:Best:est} for $u\in \mathsf{C}^\infty(\overline{ \mathscr{B}(R_k,z_k) })$.
One has:
\begin{align}\label{lm:Best:est1}
\|u\|_{\L(B_{k,\varepsilon}\cup B_{k,\varepsilon}^+)}^2&\leq
\|u\|_{\L(\mathscr{B}(\tau_k \varepsilon,z_k))}^2\\&\leq
C\left(\varepsilon\|u\|^2_{\L(\partial(\mathscr{B}(\tau_k \varepsilon,z_k)))}+\varepsilon^2\|\nabla u\|^2_{\L(\mathscr{B}(\tau_k \varepsilon,z_k))}\right), \notag
\end{align}
where the first inequality is due to
$B_{k,\varepsilon}\cup B_{k,\varepsilon}^+ \subset\mathscr{B}(\tau_k \varepsilon,z_k)$ (see~\eqref{all:paramaters} and Figure~\ref{fig-circles}), while
the second inequality is deduced from the estimate
\begin{gather*}
\|v\|_{\L(\mathscr{B}(\tau_k,0))}^2\leq
(\lambda_1(-\Delta^R_{\mathscr{B}(\tau_k,0)}))^{-1}\left(\|v\|^2_{\L(\partial(\mathscr{B}(\tau_k,0)))}+\|\nabla v\|^2_{\L(\mathscr{B}(\tau_k ,0))}\right)\label{Robin}
\end{gather*}
via the coordinate transformation $ y= (x-z_k)\varepsilon^{-1}$, where $y\in \mathscr{B}(\tau ,0)$, $x\in \mathscr{B}(\tau\varepsilon ,z_k)$; here $\Delta^R_{\mathscr{B}(\tau_k,0)}$
is the Laplacian on $\mathscr{B}(\tau_k,0)$ subject to the Robin boundary conditions
${\partial_n u} +u=0$.
We introduce the spherical coordinate system $(r,\phi)$ in $\overline{\mathscr{B}(R_k,z_k)\setminus \mathscr{B}(\tau_k\varepsilon ,z_k)}$.
Here $r\in [\tau_k\varepsilon,R_k]$ stands for the distance to $z_k$,
$\phi=(\phi_1,{\dots},\phi_{n-1})$ are the angular coordinates, where $\phi_j\in [0,\pi]$ for $j=1,\dots,n-2$,
$\phi_{n-1}\in [0,2\pi)$.
One has
$$
u(\tau_k\varepsilon,\phi)=u(r,\phi)-\int_{\tau_k\varepsilon}^{r}{\partial
u(\tau,\phi)\over\partial\tau}\d\tau,
$$
whence
\begin{align} \label{lm:Best:est2}
|u(\tau_k\varepsilon,\phi)|^2&\le 2|u(r,\phi)|^2+2\left|\int_{\tau_k\varepsilon}^{r}{\partial
u(\tau,\phi)\over\partial\tau}\d\tau\right|^2\\
\notag
&\leq
2|u(r,\phi)|^2+
2M_{\varepsilon} \int_{\tau_k\varepsilon}^{R_k}\left|{\partial
u(\tau,\phi)\over\partial\tau}\right|^2\tau^{n-1}\d\tau,
\end{align}
where $M_{\varepsilon}\coloneqq \displaystyle\int_{\tau_k\varepsilon}^{R_k}\tau^{1-n}\d\tau$.
We denote $N_{\varepsilon}\coloneqq \displaystyle \int_{\tau_k\varepsilon}^{R_k} r^{n-1}\d r $.
Multiplying \eqref{lm:Best:est2} by $(N_{\varepsilon})^{-1}(\tau_k\varepsilon)^{n-1}r^{n-1}\prod_{j=1}^{n-2}\left(\sin\phi_j\right)^{n-1-j}$, and then integrating
over $r\in (\tau_k\varepsilon,R_k)$, $\phi_j\in (0,\pi)$, $j=1,\dots,n-2$,
$\phi_{n-1}\in (0,2\pi)$,
we get
\begin{equation}\label{lm:Best:est3}
\|u\|^2_{\L(\partial(\mathscr{B}(\tau_k\varepsilon,z_k)))}
\leq 2(\tau_k\varepsilon)^{n-1}\left((N_{\varepsilon})^{-1}\|u\|^2_{\L(T_{k,\varepsilon})}
+
M_{\varepsilon}\|\nabla u\|^2_{\L(T_{k,\varepsilon})}\right).
\end{equation}
where $T_{k,\varepsilon}\coloneqq \mathscr{B}(R_k,z_k)\setminus \overline{\mathscr{B}(\tau_k\varepsilon ,z_k)}$.
One has
\begin{gather}\label{MN:eps}
M_{\varepsilon}\leq C\varepsilon^{ -n}\eta_{\varepsilon},\quad N_{\varepsilon}\geq C>0
\end{gather}
(the first inequality in \eqref{MN:eps} is obtained via straightforward calculations,
while the second follows from $\tau_k\varepsilon\leq \tau_k\varepsilon_0<R_k$).
Using \eqref{MN:eps},
we can extend \eqref{lm:Best:est3} as follows,
\begin{gather} \label{lm:Best:est4}
\|u\|^2_{\L(\partial(\mathscr{B}(\tau_k\varepsilon,z_k)))}
\leq C\left(\varepsilon^{n-1}\|u\|^2_{\L(T_{k,\varepsilon})}
+
\varepsilon^{-1}\eta_{\varepsilon}\|\nabla u\|^2_{\L(T_{k,\varepsilon})}\right).
\end{gather}
Combining \eqref{lm:Best:est1} and \eqref{lm:Best:est4}, and taking into account that
\begin{gather*}
\varepsilon^n\leq C\varepsilon^2\text{ with }C=\varepsilon_0^{n-2},\quad
\varepsilon^2=\eta_{\varepsilon}\text{ if }n\ge 3,\quad
\varepsilon^2\leq C \eta_{\varepsilon}\text{ with }C=|\ln\varepsilon_0|^{-1}\text{ if }n=2,
\end{gather*}
we arrive at the required estimate \eqref{lm:Best:est}. The lemma is proven.
\end{proof}
\subsection{Operators $\J_{\varepsilon}$ and $\widetilde\J_{\varepsilon}$ and their properties}
We denote
\begin{gather*}
\Omega_{\varepsilon}^{\rm out}\coloneqq \Omega\setminus\overline{\cup_{k\in\mathbb{M}}B_{k,\varepsilon}},
\end{gather*}
i.e. $\Omega=\Omega_{\varepsilon}^{\rm out}\cup\left(\cup_{k\in\mathbb{M}}B_{k,\varepsilon}\right)\cup\left(\cup_{k\in\mathbb{M}}D_{k,\varepsilon}\right)$.
We define the operator $\J_{\varepsilon}:\mathcal{H}\to\mathcal{H}_{\varepsilon}$ via
\begin{gather*}
(\J_{\varepsilon} f)(x)\coloneqq
\begin{cases}
f_0(x),&x\in \Omega_{\varepsilon}^{\rm out},
\\
|B_{k,\varepsilon}|^{-1/2}f_k,&x\in B_{k,\varepsilon},\ k\in\mathbb{M},\\
0,&x\in D_{k,\varepsilon},\ k\in\mathbb{M},
\end{cases}
\end{gather*}
where $f=(f_0,f_1,\dots,f_m)$ with $f_0\in\L(\Omega)$, $(f_k)_{k\in\mathbb{M}}\in\mathbb{C}^m$.
Since the measure of $D_{k,\varepsilon}$ is zero, it makes no odds how to define $\J_{\varepsilon} f$ at $D_{k,\varepsilon}$ -- one can choose any other value instead of $0$.
Further, we introduce the operator $\widetilde\J_{\varepsilon}:\mathcal{H}_{\varepsilon}\to\mathcal{H}$ acting on $u\in\L(\Omega_{\varepsilon})$ as follows:
\begin{gather}\label{wtJ}
(\widetilde\J_{\varepsilon} u)_0\coloneqq
\begin{cases}
u(x),&x\in \Omega_{\varepsilon}^{\rm out},
\\
0,&x\in \overline{\cup_{k\in \mathbb{M}}B_{k,\varepsilon}},
\end{cases}\qquad
(\widetilde\J_{\varepsilon} u)_k\coloneqq |B_{k,\varepsilon}|^{1/2}\langle u \rangle_{B_{k,\varepsilon}},\ k\in\mathbb{M},
\end{gather}
where $\langle u \rangle_{B_{k,\varepsilon}}$ is the mean value of $u$ in $B_{k,\varepsilon}$, i.e.
$$\langle u \rangle_{B_{k,\varepsilon}}\coloneqq |B_{k,\varepsilon}|^{-1}\int_{B_{k,\varepsilon}}u(x)\d x .$$
The above defined operators satisfy the following properties.
\begin{lemma}\label{lm:JJ:1}
One has
\begin{align}
\label{JJ:1}
(u,\J_{\varepsilon} f)_{\mathcal{H}_{\varepsilon}}-(\widetilde\J_{\varepsilon} u,f)_{\mathcal{H}}=0,&\quad\forall f\in\mathcal{H},\ u\in\mathcal{H}_{\varepsilon},
\\
\label{JJ:2}
\|\J_{\varepsilon} f\|_{\mathcal{H}_{\varepsilon}}\leq \|f\|_{\mathcal{H}},&\quad\forall f\in\mathcal{H},
\\
\label{JJ:3}
\|f-\widetilde\J_{\varepsilon}\J_{\varepsilon} f\|_{\mathcal{H}}\leq C \eta_{\varepsilon} ^{1/2}\|f\|_{\mathcal{H}},&\quad\forall f\in\mathcal{H},
\\
\label{JJ:4}
\|u-\J_{\varepsilon}\widetilde\J_{\varepsilon} u\|_{\mathcal{H}_{\varepsilon}}\leq C{ \varepsilon}\|u\|_{\mathcal{H}_{\varepsilon}},&\quad\forall u\in\mathcal{H}_{\varepsilon}.
\end{align}
\end{lemma}
\begin{proof}
Properties \eqref{JJ:1}--\eqref{JJ:2} follow immediately from the definition of
$\J_{\varepsilon}$, $\widetilde\J_{\varepsilon}$.
Using Lemma~\ref{lm:Best}, we obtain the estimate \eqref{JJ:3}:
\begin{align*}
\|f-\widetilde\J_{\varepsilon}\J_{\varepsilon} f\|^2_{\mathcal{H}}&=
\sum\limits_{k\in\mathbb{M}}\|f_0\|^2_{\L(B_{k,\varepsilon})}
\leq C \eta_{\varepsilon}\sum\limits_{k\in\mathbb{M}}\|f_0\|^2_{\H^1(\mathscr{B}(R_k,z_k))}\\&
\leq C \eta_{\varepsilon} \|f_0\|^2_{\H^1(\Omega)}
\leq C \eta_{\varepsilon} \|f\|_{\mathcal{H}^1}^2.
\end{align*}
Finally, we prove \eqref{JJ:4}.
For each $u\in\H^1(B_{k,\varepsilon})$
one has the Poincar\'{e} inequality
\begin{gather}\label{Poincare}
\|u-\langle u \rangle_{B_{k,\varepsilon}}\|^2\leq {1\over \lambda_2(-\Delta_{B_{k,\varepsilon}}^N)}\|\nabla u\|^2_{\L(B_{k,\varepsilon})}=
{\varepsilon^2\over \lambda_2(-\Delta_{B_k}^N)}\|\nabla u\|^2_{\L(B_{k,\varepsilon})}
\end{gather}
(the last equality follows from $B_{k,\varepsilon}\cong\varepsilon B_k$).
Using \eqref{Poincare}, we get the estimate \eqref{JJ:4}:
\begin{gather*}
\|u-\J_{\varepsilon}\widetilde\J_{\varepsilon} u\|^2_{\mathcal{H}_{\varepsilon}}=
\sum_{k\in\mathbb{M}}\|u- \langle u \rangle_{B_{k,\varepsilon}}\|^2_{\L(B_{k,\varepsilon})}\leq
C\varepsilon^2\sum_{k\in\mathbb{M}}\|\nabla u\|^2_{\L(B_{k,\varepsilon})}\leq C\varepsilon^2 \|u\|_{\mathcal{H}_{\varepsilon}^1}^2.
\end{gather*}
The lemma is proven.
\end{proof}
The lemma below follows immediately
from \eqref{JJ:1}, \eqref{JJ:3}, \eqref{JJ:4} by using
Proposition~\ref{prop:2/3}.
\begin{lemma}
For sufficiently small $\varepsilon$ one has
\begin{align}
\label{JJ:3+}
\|f\|^2_{\mathcal{H}}\leq C(1+\eta_{\varepsilon}^{1/2})\|\J_{\varepsilon} f\|^2_{\mathcal{H}_{\varepsilon}} + C\eta_{\varepsilon}^{1/2}\,\a[f,f],&\quad\forall f\in\mathcal{H}^1,
\\
\label{JJ:4+}
\|u\|^2_{\mathcal{H}_{\varepsilon}}\leq C(1+\varepsilon)\|\widetilde \J_{\varepsilon} u\|^2_{\mathcal{H}} + C\varepsilon\,\a_{\varepsilon}[u,u],&\quad\forall u\in\mathcal{H}_{\varepsilon}^1.
\end{align}
\end{lemma}
\subsection{Operators $\J^1_{\varepsilon}$ and $\widetilde\J^1_{\varepsilon}$ and their properties}
To construct the operator $\widetilde\J^1_{\varepsilon}$, one needs the following auxiliary result.
\begin{lemma}[{\cite[Lemma~2.6]{ACDP92}}]\label{lm:ext}
Let $Y$ be a bounded domain in $\mathbb{R}^n$.
Let $Y^* $ be a domain in $\mathbb{R}^n$ having Lipschitz boundary
at each point of $\partial Y^*\cap\overline{Y}$.
Then there exists a linear operator
$P: \H^1(Y^*)\to \H^1(Y)$ such that for each $u\in \H^1(Y^*)$
one has
\begin{gather}\label{Pext}
\begin{array}{l}
P u=u\ \text{ a.e. in }Y\cap Y^*,\\[2mm]
\|P u \|_{\L(Y )}\leq C_1\| u\|_{\L(Y^*)},\quad
\|\nabla(P u )\|_{\L(Y )}\leq C_2\|\nabla u\|_{\L(Y^*)},
\end{array}
\end{gather}
where the constants $C_1,C_2$ depend only on $Y$ and $Y^*$.
\end{lemma}
Owing to the fact that in Lemma~\ref{lm:ext} one has separate estimates
for $u$ and $\nabla u$, we immediately get the following corollary.
\begin{corollary}\label{coro:ext}
Let $Y,Y^*\subset\mathbb{R}^n$ be as in Lemma~\ref{lm:ext}, and $P: \H^1(Y^*)\to \H^1(Y)$ be the operator satisfying \eqref{Pext}.
Let $Y_{\varepsilon}\coloneqq\varepsilon Y+z$, $Y^*_{\varepsilon}\coloneqq \varepsilon Y^*+z$, where $\varepsilon>0$, $z\in\mathbb{R}^n$.
We define the operator
$$
P_{\varepsilon}: \H^1(Y_{\varepsilon}^*)\to \H^1(Y_{\varepsilon}),\quad
(P_{\varepsilon} u)(x)= (P v_{\varepsilon})((x-z)\varepsilon^{-1}),\ x\in Y_{\varepsilon},
$$
where $v_{\varepsilon}(y)= u(y\varepsilon+z)$, $y\in Y^*$.
Then for each $ u\in \H^1(Y^*_{\varepsilon})$ one has
\begin{gather*}
\begin{array}{l}
P_{\varepsilon} u=u\ \text{ a.e. in }Y_{\varepsilon}\cap Y_{\varepsilon}^*,\\[2mm]
\|P_{\varepsilon} u \|_{\L(Y_{\varepsilon} )}\leq C_1\| u\|_{\L(Y_{\varepsilon}^*)},\quad
\|\nabla(P_{\varepsilon} u )\|_{\L(Y_{\varepsilon} )}\leq C_2\|\nabla u\|_{\L(Y^*_{\varepsilon})},
\end{array}
\end{gather*}
where the constants $C_1,C_2$ are the same constants as in \eqref{Pext}.
\end{corollary}
One has $B_{k,\varepsilon}=\varepsilon B_k+z_k$ and $\mathscr{B}(\widetilde\tau_k \varepsilon,z_k)=\varepsilon\mathscr{B}(\widetilde\tau_k,0)+z_k$
(recall that $\widetilde\tau_k$ is defined by \eqref{kappa}).
Since $B_{k,\varepsilon}\subset\mathscr{B}(\tau_k,z_k)$ and $\tau_k<\widetilde\tau_k$, we have $\overline{B_{k,\varepsilon}}\subset
\mathscr{B}(\widetilde\tau_k \varepsilon,z_k)$ (that is, $\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\setminus \overline{B_{k,\varepsilon}}$ is an open connected set with Lipschitz boundary).
By Corollary~\ref{coro:ext} there exists
a linear operator
$$P_{k,\varepsilon}:\H^1(\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\setminus \overline{B_{k,\varepsilon}})\to \H^1(\mathscr{B}(\widetilde\tau_k\varepsilon,z_k))$$
such
that
\begin{gather}\label{Pk:ext}
\begin{array}{l}
P_{k,\varepsilon} u=u\ \text{ a.e. in }\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\setminus \overline{B_{k,\varepsilon}},\\[2mm]
\|P_{k,\varepsilon} u \|_{\H^1(\mathscr{B}(\widetilde\tau_k\varepsilon,z_k))}\leq C_k\| u\|_{\H^1(\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\setminus \overline{B_{k,\varepsilon}})}.
\end{array}
\end{gather}
Finally, we define the operator $\widetilde\J_{\varepsilon}^1:\mathcal{H}_{\varepsilon}^1\to\mathcal{H}^1 $ by
\begin{gather}\label{wtJ1}
\begin{array}{l}
(\widetilde\J_{\varepsilon}^1 u)_0(x)\coloneqq
\begin{cases}
(P_{k,\varepsilon}(u\hspace{-3pt}\restriction_{\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\setminus \overline{B_{k,\varepsilon}}}))(x),&x\in \mathscr{B}(\widetilde\tau_k\varepsilon,z_k),\ k\in\mathbb{M},\\
u(x),&x\in\Omega\setminus\left(\cup_{k\in\mathbb{M}}\mathscr{B}(\widetilde\tau_k\varepsilon,z_k)\right),
\end{cases} \\[5mm]
(\widetilde\J_{\varepsilon}^1 u)_k\coloneqq (\widetilde\J_{\varepsilon} u)_k,\ k\in\mathbb{M}.
\end{array}
\end{gather}
It follows easily from \eqref{Pk:ext}, \eqref{BRk:prop} and the last inequality in \eqref{all:paramaters} that
$(\widetilde\J_{\varepsilon}^1 u)_0 \in\H^1_0(\Omega)$; thus \eqref{wtJ1} provides a well-defined linear operator $\widetilde\J_{\varepsilon}^1:\mathcal{H}^1_{\varepsilon}\to\mathcal{H}^1$ such that
\begin{gather}
\label{E}
(\widetilde\J_{\varepsilon}^1 u)_0 = u\ \text{ a.e. on }\Omega_{\varepsilon}^{\rm out},\quad
\|(\widetilde\J_{\varepsilon}^1 u)_0 \|_{\H^1(\Omega)}\leq C\| u\|_{\H^1(\Omega_{\varepsilon}^{\rm out})}.
\end{gather}
\begin{lemma}
One has
\begin{align}
\label{JJ:6}
\|\widetilde\J^1_{\varepsilon} u - \widetilde\J_{\varepsilon} u\|_{\mathcal{H} }\leq C\eta_{\varepsilon}\|u\|_{\mathcal{H}^1_{\varepsilon}},&\quad\forall u\in\mathcal{H}^1_{\varepsilon}.
\end{align}
\end{lemma}
\begin{proof}
Let $u\in \mathcal{H}^1_{\varepsilon}$. Using Lemma~\ref{lm:Best} and \eqref{E} we get:
\begin{align*}
\|\widetilde\J^1_{\varepsilon} u - \widetilde\J_{\varepsilon} u\|_{\mathcal{H} }^2&=
\sum\limits_{k\in\mathbb{M}}\|(\widetilde\J^1_{\varepsilon} u)_0\|^2_{\L(B_{k,\varepsilon})}
\leq
C \eta_{\varepsilon}\sum\limits_{k\in\mathbb{M}}\|(\widetilde\J^1_{\varepsilon} u)_0 \|^2_{\H^1(\mathscr{B}(R_k,z_k))}
\\&\leq
C \eta_{\varepsilon} \|(\widetilde\J^1_{\varepsilon} u)_0 \|^2_{\H^1(\Omega)}
\leq
C_1 \eta_{\varepsilon}\|u\|^2_{\H^1(\Omega_{\varepsilon}^{\rm out})}
\leq
C_1 \eta_{\varepsilon}\|u\|^2_{\mathcal{H}_{\varepsilon}^1}.
\end{align*}
The lemma is proven.
\end{proof}
Now, we introduce the most tricky operator $\J_{\varepsilon}^1:\mathcal{H}^1\to\mathcal{H}_{\varepsilon}^1$.
Recall that $H_{k,\varepsilon}(x)$ is a solution to the problem
\eqref{BVP:n3} if $n\ge 3$ or to the problem
\eqref{BVP:n2} if $n= 2$; in the latter case we extend $H_{k,\varepsilon}(x)$ by $0$ to the whole $\mathbb{R}^2$.
Recall that the sets $B_{k,\varepsilon}^\pm$ are defined by \eqref{Bke+}--\eqref{Bke-}.
For $f=(f_0,f_1,\dots,f_m)\in\mathcal{H}^1$ with $f_0\in\H^1_0(\Omega)$ and $ f_k \in\mathbb{C}$, ${k\in\mathbb{M}}$
we set
\begin{gather}
\label{J1}
(\J_{\varepsilon}^1 f)(x)
\coloneqq
\begin{cases}\displaystyle
f_0(x)-\sum\limits_{k\in\mathbb{M}}\widehat\phi_{k,\varepsilon}(x)f_0(x)\\
\displaystyle +{1\over 2}H_{k,\varepsilon}(x)\phi_{k,\varepsilon}(x)
|B_{k,\varepsilon}|^{-1/2}f_k,&x\in\Omega_{\varepsilon}^{\rm out},\\\displaystyle
\left(1-{1\over 2} H_{k,\varepsilon}(x)\phi_{k,\varepsilon}(x)\right)|B_{k,\varepsilon}|^{-1/2}f_k,&x\in B_{k,\varepsilon},\, k\in\mathbb{M},\\\displaystyle
{1\over 2}|B_{k,\varepsilon}|^{-1/2}f_k,&x\in D_{k,\varepsilon},\, k\in\mathbb{M}.
\end{cases}
\end{gather}
Here $\widehat\phi_{k,\varepsilon}$ and $\phi_{k,\varepsilon}$ are cut-off functions defined as follows:
\begin{itemize}
\item $\widehat\phi_{k,\varepsilon}(x)=\begin{cases}
1,&|x-z_k|\le \ell_k d_{k,\varepsilon}, \\
\displaystyle\frac
{\mathcal{G}(|x-z_k|)-\mathcal{G}(\ell_k\varepsilon)}
{\mathcal{G}(\ell_k d_{k,\varepsilon})-\mathcal{G}(\ell_k\varepsilon)},
&\ell_k d_{k,\varepsilon}< |x-z_k|<\ell_k\varepsilon, \\
0,&|x-z_k|\ge \ell_k\varepsilon,
\end{cases}
$ \\[3pt]
where the function $\mathcal{G}:(0,\infty)\to\mathbb{R}$ is given by
$$
\mathcal{G}(t)\coloneqq
\begin{cases}
t^{2-n},&n\ge 3,\\
\ln t,& n=2,
\end{cases}
$$
\item
$\phi_{k,\varepsilon}(x)=\phi\left({|x-z_k|\over \varepsilon}\right)$, where $\phi:(0,\infty)\to\mathbb{R}$ is a smooth function
satisfying
\begin{gather*}
\phi(t)=1\text{ if }t\le \ell_k,\qquad \phi(t)=0\text{ if }t\ge \rho_k.
\end{gather*}
\end{itemize}
It is easy to see that
$(\J_{\varepsilon}^1 f)\hspace{-3pt}\restriction_{\Omega_{\varepsilon}^{\rm out}}\in\H^1(\Omega_{\varepsilon}^{\rm out})$ and
$(\J_{\varepsilon}^1 f)\hspace{-3pt}\restriction_{B_{k,\varepsilon}}\in \H^1(B_{k,\varepsilon})$ for each $k\in\mathbb{M}$.
Moreover, the traces of
$(\J_{\varepsilon}^1 f)\hspace{-3pt}\restriction_{B_{k,\varepsilon}}$ and $(\J_{\varepsilon}^1 f)\hspace{-3pt}\restriction_{\Omega_{\varepsilon}^{\rm out}}$
on $D_{k,\varepsilon}$ coincide -- they both equal to the constant ${1\over 2}|B_{k,\varepsilon}|^{-1/2}f_k$.
Hence $\J_{\varepsilon}^1 f\in\H^1(\Omega_{\varepsilon})$. Finally, since $\J_{\varepsilon}^1 f=f_0$ on
$\Omega\setminus \overline{\cup_{k\in\mathbb{M}}\mathscr{B}(R_k,z_k)}$, one gets $\J_{\varepsilon}^1 f\in \H^1_{0,\partial\Omega_{\varepsilon}}(\Omega_{\varepsilon})$.
Thus \eqref{J1} provides a well-defined linear operator
from $\mathcal{H}^1$ to $\mathcal{H}_{\varepsilon}^1$.
To proceed further,
we need pointwise estimates on the function $H_{k,\varepsilon}$.
Recall that $\mathscr{B}(\ell_k d_{k,\varepsilon},z_k)$ is the smallest ball containing $D_{k,\varepsilon}$.
\begin{lemma}[{\cite[Lemma~2.4]{MK06}}]
\label{lemma:Hest}
Let $x\in\mathbb{R}^n\setminus\overline{\mathscr{B}(\ell_k d_{k,\varepsilon},z_k)}$ with
\begin{gather}\label{C0}
\rho(x)\geq C_0 d_{k,\varepsilon}\text{ if }n\geq 3\text{\quad and\quad }
\rho(x)\geq \exp(-C_0 {|\ln d_{k,\varepsilon}|^{1/2}})\text{ if n=2},
\end{gather}
where $\rho(x)=|x-z_k|-\ell_k d_{k,\varepsilon}$ is the distance from $x$ to $\mathscr{B}(\ell_k d_{k,\varepsilon},z_k)$, $C_0>0$ is a constant.
Then
\begin{gather}\label{H:estimates}
\begin{array}{lll}
|H_{k,\varepsilon}(x)|\leq
C\dfrac{ (d_{k,\varepsilon})^{n-2}}{(\rho(x))^{n-2}},&
|\nabla H_{k,\varepsilon}(x)|\leq
C\dfrac{ (d_{k,\varepsilon})^{n-2}}{(\rho(x))^{n-1}}&\text{as }n\geq 3,\\[2ex]
|H_{k,\varepsilon}(x)|\leq
C\dfrac{ |\ln d_{k,\varepsilon}|^{-1}}{|\ln \rho(x) |^{-1} },&
|\nabla H_{k,\varepsilon}(x)|\leq
C\dfrac{ |\ln d_{k,\varepsilon}|^{-1} }{ \rho(x) }&\text{as }n=2,
\end{array}
\end{gather}
where $C>0$ is independent on $d_{k,\varepsilon}$ and $\varepsilon$ (but may depend on $C_0$).
\end{lemma}
\begin{remark}\label{rem:Hest}
Let $x\in \mathbb{R}^n$ such that $|x-z_k|\ge \ell_k\varepsilon$; then $x$ satisfies \eqref{C0} with
$$C_0=
\begin{cases}
\ell_k(\kappa^{-1}-1),&n\ge 3,\\
{\sup}_{\varepsilon\in (0,\varepsilon_0]}\big(C_{\varepsilon}^{1/2}\varepsilon|\ln (\ell_k\varepsilon(1-\kappa))|\big),&n=2,
\end{cases}
$$
where $\kappa\coloneqq \sup_{\varepsilon\in(0,\varepsilon_0]} {d_{\varepsilon} \varepsilon^{-1}}<1$ (see \eqref{deps<1}), $C_{\varepsilon}>0$ is a constant standing in \eqref{gamma:2} and satisfying $\sup_{\varepsilon\in (0,\varepsilon_0]}C_{\varepsilon}<\infty$ (see \eqref{CC}).
\end{remark}
\begin{lemma}\label{lm:JJ:5}
One has
\begin{align}
\label{JJ:5}
\|\J^1_{\varepsilon} f - \J_{\varepsilon} f\|_{\mathcal{H}_{\varepsilon}}\leq C\eta_{\varepsilon}^{1/2}\|f\|_{\mathcal{H}^1},&\quad\forall f\in\mathcal{H}^1.
\end{align}
\end{lemma}
\begin{proof}
Let $f=(f_0,f_1,\dots,f_m)\in\mathcal{H}^1$ with $f_0\in \H^1_0(\Omega)$, $f_k\in\mathbb{C}$, $k\in\mathbb{M}$.
One has
\begin{gather*}
\J^1_{\varepsilon} f - \J_{\varepsilon} f=
\begin{cases}\displaystyle
\sum\limits_{k\in\mathbb{M}}\left( -\widehat\phi_{k,\varepsilon}(x)f_0(x) +{1\over 2}H_{k,\varepsilon}(x)\phi_{k,\varepsilon}(x)
|B_{k,\varepsilon}|^{-1/2}f_k\right),&x\in\Omega_{\varepsilon}^{\rm out},\\\displaystyle
- {1\over 2} H_{k,\varepsilon}(x)\phi_{k,\varepsilon}(x) |B_{k,\varepsilon}|^{-1/2}f_k,&x\in B_{k,\varepsilon},\\\displaystyle
{1\over 2}|B_{k,\varepsilon}|^{-1/2}f_k,&x\in D_{k,\varepsilon},
\end{cases}
\end{gather*}
whence, using $|\widehat\phi_{k,\varepsilon}|\leq C$ and $|B_{k,\varepsilon}|=\varepsilon^n|B_k|$, we infer the estimate
\begin{gather}\label{JminusJ:2}
\|\J^1_{\varepsilon} f - \J_{\varepsilon} f \|^2_{\mathcal{H}_{\varepsilon}}\leq
C\sum_{k\in\mathbb{M}}\left[\|f_0\|^2_{\L( B^+_{k,\varepsilon})}
+\varepsilon^{-n}|f_k|^2 \|H_{k,\varepsilon}\phi_{k,\varepsilon} \|^2_{\L( \mathscr{B}(\rho_k \varepsilon,z_k) )} \right].
\end{gather}
The first term in the right-hand-side of \eqref{JminusJ:2}
is estimated via Lemma~\ref{lm:Best}:
\begin{gather}\label{poincare:f0}
\sum\limits_{k\in\mathbb{M}}\|f_0\|^2_{\L( B^+_{k,\varepsilon})}
\leq
C \eta_{\varepsilon}\sum\limits_{k\in\mathbb{M}}\| f_0\|^2_{\H^1(\mathscr{B}(R_k,z_k))}
\leq
C \eta_{\varepsilon} \|f_0\|^2_{\H^1(\Omega)}.
\end{gather}
Now, we estimate the second term.
One has the following Poincar\'{e} inequality:
\begin{gather}\label{Poincare:0}
\forall v\in\H_0^1(\mathscr{B}(\rho_k \varepsilon,z_k)):\quad
\|v\|^2_{\L(\mathscr{B}(\rho_k \varepsilon,z_k))}\leq {\varepsilon^2\over \lambda_1(-\Delta_{\mathscr{B}(\rho_k,0)}^D)} \|\nabla u\|^2_{\L(\mathscr{B}(\rho_k \varepsilon,z_k))}.
\end{gather}
Since $\mathrm{supp}(\phi_{k,\varepsilon}) \subset\overline{\mathscr{B}(\rho_k \varepsilon,z_k)}$,
one has $H_{k,\varepsilon} \phi_{k,\varepsilon}\in \H^1_0(\mathscr{B}(\rho_k \varepsilon,z_k))$.
Applying \eqref{Poincare:0} for $v\coloneqq H_{k,\varepsilon} \phi_{k,\varepsilon}$, and then using
$|\phi_{k,\varepsilon}|\leq C$, $|\nabla\phi_{k,\varepsilon}|\leq C\varepsilon^{-1}$ and
\eqref{capty},
we obtain
\begin{align}\label{Hphi:est}
\|H_{k,\varepsilon} \phi_{k,\varepsilon}\|^2_{\L( \mathscr{B}(\rho_k \varepsilon,z_k) )}
&\leq C\varepsilon^2\|\nabla( H_{k,\varepsilon} \phi_{k,\varepsilon} )\|^2_{\L( \mathscr{B}(\rho_k \varepsilon,z_k) )}
\\\notag
&\leq
C_1\left(\varepsilon^2 \mathrm{cap}( {D_{k,\varepsilon}})+
\| H_{k,\varepsilon}\|^2_{\L(\mathrm{supp}(\nabla\phi_{k,\varepsilon}))}\right).
\end{align}
By virtue of \eqref{cap:rescaling1}--\eqref{gamma:2}, \eqref{CC} we have
\begin{gather}\label{cap:est}
\mathrm{cap}( {D_{k,\varepsilon}})\leq C\varepsilon^n.
\end{gather}
One has $\mathrm{supp}(\nabla\phi_{k,\varepsilon})\subset \mathbb{R}^n\setminus\mathscr{B}(\ell_k\varepsilon,z_k)$, whence
(see Remark~\ref{rem:Hest}) the estimates \eqref{H:estimates} fulfill for
$x\in \mathrm{supp}(\nabla\phi_{k,\varepsilon})$. Using these estimates, \eqref{gamma:2}, \eqref{CC}, and $|\mathrm{supp}(\nabla\phi_{k,\varepsilon})|\leq C\varepsilon^n$, we get
\begin{gather}\label{HL2}
\|H_{k,\varepsilon}\|^2_{\L(\mathrm{supp}(\nabla\phi_{k,\varepsilon}))}
\leq
C_1\varepsilon^{4-n}\cdot\left.
\begin{cases}
{(d_{k,\varepsilon})^{2n-4}},& n\ge 3
\\
{|\ln\varepsilon|^2}\cdot {|\ln d_{k,\varepsilon}|^{-2}}, & n=2
\end{cases}\right\}\leq
C_2\varepsilon^{n }\eta_{\varepsilon}^2.
\end{gather}
The estimate \eqref{JJ:5} follows from \eqref{JminusJ:2}, \eqref{poincare:f0},
\eqref{Hphi:est}--\eqref{HL2} (in the case $n=2$ we also use $\varepsilon^2\leq C \eta_{\varepsilon}$, $C=|\ln\varepsilon_0|^{-1}$ and $\eta_{\varepsilon}<\varepsilon$).
The lemma is proven.
\end{proof}
Before to proceed to the last lemma we give several several further properties of the function $H_{k,\varepsilon}$.
Standard elliptic regularity theory yields
\begin{gather}
\label{H:regul}
H_{k,\varepsilon}\in \mathsf{C}^\infty(\mathbb{R}^n\setminus\overline{D_{k,\varepsilon}})\text{ if }n\geq 3
\quad\text{and}\quad
H_{k,\varepsilon}\in \mathsf{C}^\infty(\mathscr{B}(1,z_k)\setminus\overline{D_{k,\varepsilon}})\text{ if }n=2.
\end{gather}
Evidently, $H_{k,\varepsilon}$ is symmetric
with respect to the hyperplane
\begin{align}
\label{Gak}
\S_k\coloneqq \left\{x=(x',x^n)\in\mathbb{R}^n:\ x^n=z_k^n\right\},
\end{align}
i.e.
\begin{gather}
\label{H:symm}
H_{k,\varepsilon}(x',z_k^n+\tau)=H_{k,\varepsilon}(x',z_k^n-\tau),\ \forall \tau>0.
\end{gather}
From \eqref{H:regul}, \eqref{H:symm} we deduce
\begin{gather}\label{H:n1}
\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k+0}=\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}=0\text{ on } \S_k\setminus \overline{D_{k,\varepsilon}},\\
\label{H:n2}
\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k+0}=-\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}\text{ on } D_{k,\varepsilon}.
\end{gather}
Finally, using \eqref{BVP:n3}--\eqref{BVP:n2}
and the asymptotics (see Lemma~\ref{lemma:Hest})
\begin{gather}\label{decay}
| H_{k,\varepsilon}(x)|=\mathcal{O}(|x|^{2-n}),\quad
|\nabla H_{k,\varepsilon}(x)|=\mathcal{O}(|x|^{1-n})\quad\text{if }n\ge 3,
\end{gather}
we obtain
the Green's identity
\begin{gather}\label{Green}
\int_{D_{k,\varepsilon}}\left(\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}-
\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k+0}\right)\d x'=
\begin{cases}\displaystyle
\|\nabla H_{k,\varepsilon}\|^2_{\L(\mathbb{R}^n)},&n\ge 3,\\\displaystyle
\|\nabla H_{k,\varepsilon}\|^2_{\L(\mathscr{B}(1,z_k)},&n= 2,\\
\end{cases}
\end{gather}
where $\d x'$ stands for the area measure on $\partial D_{k,\varepsilon}$ (the decaying property \eqref{Green} is required
to guarantee that $\int_{\partial \mathscr{B}(R,0)}H_{k,\varepsilon}{\partial_n H_{k,\varepsilon} }\d s\to 0$
as $R\to \infty$).
From \eqref{capty}, \eqref{H:n2}, \eqref{Green} we infer
\begin{gather}\label{cap:D}
\mathrm{cap}( {D_{k,\varepsilon}})=2\int_{D_{k,\varepsilon}}
\left.{\partial H_{k,\varepsilon}\over \partial x^n}\right|_{x^n=z_k^n-0}\d x'.
\end{gather}
\begin{lemma}
One has $\forall f\in \mathcal{H}^2 ,\ u\in \mathcal{H}^1_{\varepsilon}$:
\begin{align}
\label{JJ:7}
\left|\a_{\varepsilon}[u,\J^1_{\varepsilon} f]-\a[\widetilde\J^{1}_{\varepsilon} u,f] \right| \leq
C\left( \sum\limits_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+(\eta_{\varepsilon})^{3/2}\varepsilon^{-2}\right)\|f\|_{\mathcal{H}^2 }\|u\|_{\mathcal{H}^1_{\varepsilon}}.
\end{align}
\end{lemma}
\begin{proof}
Let $f\in \mathcal{H}^2$ and $u\in \mathcal{H}^1_{\varepsilon}$. We have
\begin{gather}\notag
\a_{\varepsilon}[u,\J^1_{\varepsilon} f]-\a[\widetilde\J^{1}_{\varepsilon} u,f] = \sum\limits_{k\in\mathbb{M}}\left(P_{\varepsilon}^k+Q_{\varepsilon}^k+R_{\varepsilon}^k\right),
\end{gather}
where
\begin{align*}
P_{\varepsilon}^k=&\,
-\sum\limits_{k\in\mathbb{M}} \int_{ B^+_{k,\varepsilon}}\nabla u\cdot \nabla\left(\widehat\phi_{k,\varepsilon} \overline{f_0}\right)\d x,
\\
Q_{\varepsilon}^k=&\,-\sum\limits_{k\in\mathbb{M}} \int_{B_{k,\varepsilon}}\nabla (\widetilde\J^{1}_{\varepsilon} u)\cdot \nabla \overline{f_0}\d x,
\\
R_{\varepsilon}^k=&\,
\sum\limits_{k\in\mathbb{M}}
\left({1\over 2}\int_{ B^+_{k,\varepsilon}}\nabla u\cdot \nabla\left(H_{k,\varepsilon}\phi_{k,\varepsilon}
\right) \d x\right. \\
&\left.- {1\over 2}
\int_{B_{k,\varepsilon}^-}
\nabla u\cdot
\nabla \left(H_{k,\varepsilon}\phi_{k,\varepsilon} \right)\d x
-\gamma_k |B_{k,\varepsilon}| \langle u \rangle_{B_{k,\varepsilon}}\right) |B_{k,\varepsilon}|^{-1/2} \overline{f_k }.
\end{align*}
\paragraph{\it Estimate of $P_{\varepsilon}^k$}
Using standard theory of elliptic PDEs (see, e.g., \cite{Ev98}), we infer
$\mathrm{dom}(\Delta_\Omega^D)\xhookrightarrow{}\H^2(\Sigma')$
for any bounded domain $\Sigma'\subset\joinrel\subset\Omega$. In particular,
there exists $C>0$ such that
\begin{gather}\label{h}
\forall h\in\mathrm{dom}(\Delta_\Omega^D):\quad
\|h\|_{\H^2(\mathscr{B}(R_k,z_k))}\leq C\|-\Delta h + h\|_{\L(\Omega)}.
\end{gather}
One has
\begin{gather}\label{P}
P_{\varepsilon}^{k}=
-\sum\limits_{k\in\mathbb{M}} \underbrace{\int_{ B^+_{k,\varepsilon}}(\nabla u\cdot \nabla \overline{f_0})\widehat\phi_{k,\varepsilon}\d x}_{P_{\varepsilon}^{k,1}\coloneqq}
-\sum\limits_{k\in\mathbb{M}} \underbrace{\int_{ B^+_{k,\varepsilon}}(\nabla u\cdot \nabla \widehat\phi_{k,\varepsilon})
\overline{f_0}\d x}_{P_{\varepsilon}^{k,2}\coloneqq}.
\end{gather}
Using Lemma~\ref{lm:Best} and \eqref{h}, and taking into account that $|\widehat\phi_{k,\varepsilon}|\leq C$, we estimate
the first integral in \eqref{P} as follows:
\begin{align}\label{P1}
|P_{\varepsilon}^{k,1}|&\leq \|\nabla u\|_{\L( B^+_{k,\varepsilon})}
\|\nabla f_0\|_{\L( B^+_{k,\varepsilon})}
\leq
C\eta_{\varepsilon}^{1/2}\|\nabla u\|_{\L(\Omega_{\varepsilon})}
\|\nabla f_0\|_{\H^1(\mathscr{B}(R_k,z_k))} \\
\notag
&\leq C\eta_{\varepsilon}^{1/2}\|\nabla u\|_{\L(\Omega_{\varepsilon})}\|f_0\|_{\H^2(\mathscr{B}(R_k,z_k))}\\
&\leq C_1\eta_{\varepsilon}^{1/2}\|\nabla u\|_{\L(\Omega)}\|-\Delta f_0 + f_0\|_{\L(\Omega)}
\notag
\leq C_1\eta_{\varepsilon}^{1/2}\|u\|_{\mathcal{H}_{\varepsilon}^1}\|f\|_{\mathcal{H}^2}.
\end{align}
To estimate the second integral in \eqref{P} we use the generalised H\"older's inequality:
\begin{gather}\label{P2:start}
|P_{\varepsilon}^{k,2}|\leq
\|\nabla u\|_{\L( B^+_{k,\varepsilon})} \|f_0\|_{\mathsf{L}^p( B^+_{k,\varepsilon})}
\|\nabla \widehat \phi_{k,\varepsilon}\|_{\mathsf{L}^q( B^+_{k,\varepsilon})},
\end{gather}
where $p,q\in [2,\infty]$ with ${1\over p}+{1\over q}={1\over 2}$ are chosen as follows:
\begin{gather}\label{pq}
\begin{array}{lll}
p={2n^2\over (n-2)^2},& q={n^2\over 2(n-1)}&\text{if }n\ge 3,
\\
p=\infty,& q=2&\text{if }n=2.
\end{array}
\end{gather}
{The Sobolev embedding theorem \cite[Theorem~5.4]{Ad75} asserts that the space $\mathsf{L}^p(\mathscr{B}(R_k,z_k))$
is embedded continuously into the space $\H^2(\mathscr{B}(R_k,z_k))$ provided $p$ satisfies
$1\leq p\leq \frac{2n}{n-4}$ if $n\geq 5$,
$1\leq p<\infty$ if $n=4$, and
$1\le p\le \infty$ if $n=2,3$; it is easy to check that
$p$ defined by \eqref{pq} falls within the above restrictions.}
Using this and \eqref{h}, we get
\begin{align} \label{sobolev}
\|f_0\|_{\mathsf{L}^p( B^+_{k,\varepsilon})}&\le
\|f_0\|_{\mathsf{L}^p(\mathscr{B}(R_k,z_k))}\le
C\|f_0\|_{\mathsf{H}^2(\mathscr{B}(R_k,z_k))}
\\&\leq
C_1\|-\Delta f_0+f_0\|_{\L(\Omega)}\leq
C_1\|f\|_{\mathcal{H}^2}.\notag
\end{align}
Furthermore, via direct calculations we obtain
\begin{gather}\label{phi:est}
\|\nabla \widehat \phi_{k,\varepsilon}\|_{\mathsf{L}^q( B^+_{k,\varepsilon})}\leq
C\left.\begin{cases}
(d_{\varepsilon})^{1-2/n},&n\ge 3
\\
|\ln d_{\varepsilon}|^{-1/2},&n=2
\end{cases}\right\}\leq C_1\varepsilon
\end{gather}
(the last inequality is valid by virtue of \eqref{gamma:2}, \eqref{CC}).
It follows from \eqref{P2:start}, \eqref{sobolev}, \eqref{phi:est}
that
\begin{gather}\label{P2}
|P_{\varepsilon}^{k,2}|\leq
C\varepsilon\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}^2}\leq
C_1\eta_{\varepsilon}^{1/2}\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}^2}.
\end{gather}
Combining \eqref{P1} and \eqref{P2}, we finally arrive at the estimate
\begin{gather}\label{P:final}
|P_{\varepsilon}^{k}|\leq
C\eta_{\varepsilon}^{1/2}\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}^2}.
\end{gather}
\medskip
\paragraph{\it Estimate of $Q_{\varepsilon}^k$}
One has:
\begin{align}\label{Q:first}
|Q_{\varepsilon}^k|^2&\leq
{\sum\limits_{k\in\mathbb{M}}\|\nabla (\widetilde J^1_{\varepsilon} u)\|^2_{\L(B_{k,\varepsilon})}}
{\sum\limits_{k\in\mathbb{M}}\|\nabla f_0\|^2_{\L(B_{k,\varepsilon})}}
\\ &\leq
\|\widetilde J^1_{\varepsilon} u\|^2_{\H^1(\Omega)}
{\sum\limits_{k\in\mathbb{M}}\|\nabla f_0\|^2_{\L(B_{k,\varepsilon})}}.\notag
\end{align}
Due to \eqref{E}, we have
\begin{gather}\label{Q1}
\|\widetilde\J_{\varepsilon}^1 u\|^2_{\H^1(\Omega)}\leq
C\| u\|^2_{\H^1(\Omega_{\varepsilon}^{\rm out})}\leq
C\|u\|^2_{\mathcal{H}^1_{\varepsilon}}.
\end{gather}
Lemma~\ref{lm:Best} and \eqref{h} yield
\begin{align}\label{Q2}
\|\nabla f_0\|_{\L( B_{k,\varepsilon})}^2
& \leq
C \eta_{\varepsilon}\|\nabla f_0\|^2_{\H^1(\mathscr{B}(R_k,z_k))}
\leq C \eta_{\varepsilon} \|f_0\|_{\H^2(\mathscr{B}(R_k,z_k))}\\ \notag
&\leq C_1 \eta_{\varepsilon} \|-\Delta f_0 + f_0\|^2_{\L(\Omega)}
\leq C_1 \eta_{\varepsilon} \|f\|^2_{\mathcal{H}^2}.
\end{align}
Combining \eqref{Q:first}--\eqref{Q2} we get the estimate
\begin{gather}\label{Q:final}
|Q_{\varepsilon}^k|\leq C \eta_{\varepsilon}^{1/2}\|u\|_{\mathcal{H}_{\varepsilon}^1}\|f\|_{\mathcal{H}^2}.
\end{gather}
\paragraph{\it Estimate of $R_{\varepsilon}^k$}
Recall that the hyperplane $\S_k$ is defined by \eqref{Gak}.
The cut-off function $\phi_{k,\varepsilon}=\phi(|\cdot-z_k|\varepsilon^{-1})\in C^\infty(\mathbb{R}^n)$ has the following properties:
\begin{gather}\label{phi:n}
\begin{array}{c}
\displaystyle{\partial \phi_{k,\varepsilon}\over\partial x^n}=0\text{ on } \S_k,\\[2mm]
\mathrm{supp}(\phi_{k,\varepsilon})\subset \overline{\mathscr{B}(\rho_k\varepsilon,z_k)},\quad
\phi_{k,\varepsilon}=1\text{ in a neighbourhood of }D_{k,\varepsilon}.
\end{array}
\end{gather}
Then, using \eqref{H:n1} and \eqref{phi:n},
we obtain via partial integration:
\begin{multline}\label{byparts1}
\int_{ B^+_{k,\varepsilon}}\nabla u\cdot \nabla\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x\\
=
-\int_{ B^+_{k,\varepsilon}} u\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x -
\int_{D_{k,\varepsilon}}
u\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k+0}\d x'.
\end{multline}
Similarly, using \eqref{H:n1} and \eqref{phi:n}, we get
\begin{align}\label{byparts2}
&\int_{ B^-_{k,\varepsilon}}\nabla u\cdot \nabla\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x
\\\notag
=&
-\int_{ B^-_{k,\varepsilon}} u\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x +
\int_{D_{k,\varepsilon}}
u\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}\d x'
\\\notag
=&-\int_{ B^-_{k,\varepsilon}} (u-\langle u\rangle_{B_{k,\varepsilon}})\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x
\\\notag
-&
\langle u\rangle_{B_{k,\varepsilon}}\int_{ B^-_{k,\varepsilon}} \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x
+
\int_{D_{k,\varepsilon}}
u\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}\d x'
\\\notag
=&
-\int_{ B^+_{k,\varepsilon}} (u-\langle u\rangle_{B_{k,\varepsilon}})\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x
\\&-
\langle u\rangle_{B_{k,\varepsilon}}\int_{D_{k,\varepsilon}}
\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}\d x'
+
\int_{D_{k,\varepsilon}}
u\left.{\partial H_{k,\varepsilon}\over\partial x^n}\right|_{x^n=z^n_k-0}\d x'.\notag
\end{align}
Combining \eqref{byparts1} and \eqref{byparts2} and taking into account \eqref{H:n2} and \eqref{cap:D},
we get the equality
\small
\begin{multline*}
R_{\varepsilon}^k=
\sum\limits_{k\in\mathbb{M}}
\underbrace{\left({1\over 4}\mathrm{cap}( {D_{k,\varepsilon}})
-\gamma_k |B_{k,\varepsilon}|\right) \langle u \rangle_{B_{k,\varepsilon}} |B_{k,\varepsilon}|^{-1/2} \overline{f_k }}_{R_{\varepsilon}^{k,1}\coloneqq}\\
+{1\over 2}\sum\limits_{k\in\mathbb{M}}
\underbrace{\left(-\int_{ B^+_{k,\varepsilon}} u\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x
+\int_{ B_{k,\varepsilon}^-} (u-\langle u\rangle_{B_{k,\varepsilon}})\cdot \Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right) \d x \right)|B_{k,\varepsilon}|^{-1/2} \overline{f_k}}_{R_{\varepsilon}^{k,2}\coloneqq}
\end{multline*}\normalsize
Due to \eqref{gamma:1} and \eqref{wtJ}, one has
$R_{\varepsilon}^{k,1} = \left(\gamma_{k,\varepsilon}
-\gamma_k \right) (\widetilde \J_{\varepsilon} u)_k \overline{f_k }$, whence, taking into account
that $\|\widetilde\J_{\varepsilon} u\|_{\mathcal{H}}\leq \|u\|_{\mathcal{H}_{\varepsilon}}$ (this follows immediately from \eqref{JJ:1}--\eqref{JJ:2}),
we obtain
\begin{align}\label{R1:final}
|R_{\varepsilon}^{k,1}|\leq
C|\gamma_{k,\varepsilon}-\gamma_k| \cdot \|\widetilde \J_{\varepsilon} u\|_{\mathcal{H}}\cdot\|f\|_{\mathcal{H}}
&\leq
C_1|\gamma_{k,\varepsilon}-\gamma_k| \cdot \|u\|_{\mathcal{H}_{\varepsilon}}\cdot\|f\|_{\mathcal{H}}
\\ \notag
&\leq
C_1|\gamma_{k,\varepsilon}-\gamma_k| \cdot \|u\|_{\mathcal{H}_{\varepsilon}^1}\cdot\|f\|_{\mathcal{H}^2}.
\end{align}
It remains to estimate the term $R_{\varepsilon}^{k,2}$.
One has, taking into account $B_{k,\varepsilon}^-\subset B_{k,\varepsilon}$:
\begin{align*}
|R_{\varepsilon}^{k,2}|^2&\leq
2\left(\|u\|_{\L( B^+_{k,\varepsilon})}^2 +
\|u-\langle u\rangle_{B_{k,\varepsilon}}\|_{\L(B_{k,\varepsilon})}^2\right)\times\\
&\times
\|\Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right)\|_{\L(\mathscr{B}(\rho_k\varepsilon,z_k))}^2
|B_{k,\varepsilon}|^{-1} |f_k|^2
\end{align*}
Applying Lemma~\ref{lm:Best} and \eqref{E}, we get
\begin{align}
\|u\|_{\L( B^+_{k,\varepsilon})}^2=
\|\widetilde\J_{\varepsilon}^1 u\|_{\L( B^+_{k,\varepsilon})}^2
\leq
C \eta_{\varepsilon} \|\widetilde\J_{\varepsilon}^1 u\|^2_{\H^1(\mathscr{B}(R_k,z_k))}\leq
C_1 \eta_{\varepsilon}\|u\|_{\mathcal{H}_{\varepsilon}^1}^2.
\label{u:whB}
\end{align}
Further, since
$\Delta H_{k,\varepsilon}=0$, one has
\begin{gather}\label{Leibniz}
\Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right)=
2\nabla H_{k,\varepsilon}\cdot \nabla \phi_{k,\varepsilon}+
H_{k,\varepsilon}\Delta \phi_{k,\varepsilon}.
\end{gather}
One has
$\mathrm{supp}(\nabla\phi_{k,\varepsilon})\cup\mathrm{supp}(\Delta\phi_{k,\varepsilon})\subset \mathbb{R}^n\setminus\mathscr{B}(\ell_k\varepsilon,z_k)$, whence (see~Remark~\ref{rem:Hest})
the estimates \eqref{H:estimates} hold
for $x\in \mathrm{supp}(\nabla\phi_{k,\varepsilon})\cup\mathrm{supp}(\Delta\phi_{k,\varepsilon})$. Using them, \eqref{gamma:2} and \eqref{CC}, we get
\begin{gather}\label{HgradH:est}
x\in\mathrm{supp}(\nabla\phi_{k,\varepsilon})\cup\mathrm{supp}(\Delta\phi_{k,\varepsilon}):\quad
|(\nabla H_{k,\varepsilon})(x)|\leq C\varepsilon,\
|H_{k,\varepsilon}(x)|\leq
C\eta_{\varepsilon}.
\end{gather}
It follows from \eqref{Leibniz}, \eqref{HgradH:est},
$|\nabla\phi_{k,\varepsilon}|\leq C\varepsilon^{-1}$, $|\Delta\phi_{k,\varepsilon}|\leq C\varepsilon^{-2}$,
$|\mathscr{B}(\rho_k\varepsilon,z_k)|\leq C\varepsilon^n$ that
\begin{gather}
\label{DeltaHphi:est}
\|\Delta\left(H_{k,\varepsilon}\phi_{k,\varepsilon}\right)\|_{\L(\mathscr{B}(\rho_k\varepsilon,z_k))}^2
\leq C\varepsilon^{n-4} \eta_{\varepsilon}^2.
\end{gather}
Combining \eqref{Poincare}, \eqref{u:whB}, \eqref{DeltaHphi:est} and taking into account that
$|B_{k,\varepsilon}|= C\varepsilon^n$, we get
\begin{gather}\label{R2:final}
|R_{\varepsilon}^{k,2}|\leq C\eta_{\varepsilon}^{3/2}\varepsilon^{-2}\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}}\leq C\eta_{\varepsilon}^{3/2}\varepsilon^{-2}\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}^2}.
\end{gather}
Finally, from \eqref{R1:final} and \eqref{R2:final}, we conclude
\begin{gather}\label{R:final}
|R_{\varepsilon}^{k}|\leq C\left(\sum\limits_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+\eta_{\varepsilon}^{3/2}\varepsilon^{-2}\right)\|u\|_{\mathcal{H}^1_{\varepsilon}}\|f\|_{\mathcal{H}^2}.
\end{gather}
The required estimate \eqref{JJ:7} follows from \eqref{P:final}, \eqref{Q:final}, \eqref{R:final}
(in the case $n=2$ we also use $\eta_{\varepsilon}^{1/2}\leq C \eta_{\varepsilon}^{3/2}\varepsilon^{-2}$).
The lemma is proven.
\end{proof}
\subsection{End of proofs of Theorems~\ref{th1} and \ref{th2}}
It follows from \eqref{JJ:1}, \eqref{JJ:6}, \eqref{JJ:5}, \eqref{JJ:7}
that conditions \eqref{C1}, \eqref{thA3:1}--\eqref{thA3:3} hold with
$$\delta_{\varepsilon}\coloneqq C\left(\sum\limits_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+(\eta_{\varepsilon})^{3/2}\varepsilon^{-2}\right)=\begin{cases}
\displaystyle C \sum_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+C\varepsilon ,&n\ge 3,\\
\displaystyle C \sum_{k\in\mathbb{M}}|\gamma_{k,\varepsilon}-\gamma_k|+C\varepsilon|\ln\varepsilon|^{3/2} ,&n=2.
\end{cases}$$
(in the case $n=2$ we also use
$\varepsilon\leq C \eta_{\varepsilon}^{3/2}\varepsilon^{-2}$,
$\eta_{\varepsilon}^{1/2}\leq C \eta_{\varepsilon}^{3/2}\varepsilon^{-2}$).
Hence, using Theorem~\ref{thA3} and taking into account that $\widetilde\J_{\varepsilon} = (\J_{\varepsilon})^*$ (see \eqref{JJ:1}), we arrive at
\begin{gather}\label{ResRes:1}
\|({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1}\J_{\varepsilon} -\J_{\varepsilon} ({\mathcal{A}}+\mathrm{I})^{-1}\|_{\mathcal{H}\to\mathcal{H}_{\varepsilon}}=\|\widetilde\J_{\varepsilon}({\mathcal{A}}_{\varepsilon}+\mathrm{I})^{-1} - ({\mathcal{A}}+\mathrm{I})^{-1} \widetilde\J_{\varepsilon}\|_{\mathcal{H}\to\mathcal{H}_{\varepsilon}}\leq 4\delta_{\varepsilon}.
\end{gather}
Futhermore, it follows from \eqref{JJ:1}--\eqref{JJ:4}, \eqref{ResRes:1} that
\begin{gather}\label{gnrc}
{\mathcal{A}}_{\varepsilon}\overset{g.n.r.c.}\to {\mathcal{A}}\text{ as }\varepsilon\to 0.
\end{gather}
Now, by virtue of Theorem~\ref{thA2}, the estimates \eqref{JJ:3+}, \eqref{JJ:4+},
\eqref{ResRes:1} imply the statement of Theorem~\ref{th1}, and,
by virtue of Theorem~\ref{thA1}, the property \eqref{gnrc} and the equality
$$
\|\psi_{\varepsilon} - \J_{\varepsilon} \psi \|_{\mathcal{H}_{\varepsilon}}^2=
\|\psi_{\varepsilon} - \psi_0\|^2_{\L(\Omega_{\varepsilon}\setminus \overline{\cup_{k\in\mathbb{M}}B_{k,\varepsilon}})}+
\sum\limits_{k\in\mathbb{M}}\|\psi_{\varepsilon} - |B_{k,\varepsilon}|^{-1/2}\psi_k\|^2_{\L(B_{k,\varepsilon})}$$
yield the statements of Theorems~\ref{th2}.
\subsection{Absence of outside convergence of discrete spectrum\label{subsec:4:1}}
In this subsection we present an example showing that
the property $\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\searrow\sigma_{\rm disc}({\mathcal{A}})$ does not always hold true.
Let $\Omega=\Omega'\times\mathbb{R}$ be a tubular domain with a bounded Lipschitz
cross-section $\Omega'\subset \mathbb{R}^{n-1}$. To simplify presentation, we restrict ourselves to the case $n\ge 3$. The spectrum of the Dirichlet Laplacian
$-\Delta_\Omega^D$ on $\Omega$ is purely essential, namely
\begin{gather}\label{Delta:spec}
\sigma(-\Delta_\Omega^D)=\sigma_{\rm ess}(-\Delta_\Omega^D)=[\Lambda',\infty),
\end{gather}
where $\Lambda'\coloneqq \lambda_1(-\Delta^D_{\Omega'})>0$.
As before, we define the domain $\Omega_{\varepsilon}$ by \eqref{Omega:e}, \eqref{DBe}, \eqref{Dk}
with $m=1$ (i.e., only one resonator is inserted), arbitrary sets $B_1\subset\mathbb{R}^n$ and $\widetilde D_1\subset\mathbb{R}^{n-1}$ such that \eqref{DB1}, \eqref{DB2} hold,
arbitrary $z_1\in\Omega$, and $d_{1,\varepsilon}$ being chosen as follows:
\begin{gather*}
d_{1,\varepsilon}=\left({4(\Lambda'-\varepsilon^{1/2})|B_1|\over \mathrm{cap}(D_1)}\right)^{1\over n-2}\varepsilon^{n\over n-2}.
\end{gather*}
With such a choice of $d_{1,\varepsilon}$ we obtain, using \eqref{cap:rescaling2}:
\begin{gather}\notag
\gamma_{1,\varepsilon}=\Lambda'-\varepsilon^{1/2},\\\label{gaga2}
\text{consequently, }\gamma_1=\Lambda'.
\end{gather}
Let ${\mathcal{A}}$ be the corresponding limiting operator, see~\eqref{A:oplus}.
From \eqref{A:oplus}, \eqref{Delta:spec}, \eqref{gaga2} we infer
\begin{gather}
\label{A:ess}
\sigma_{\rm ess}({\mathcal{A}})=[\Lambda',\infty),\quad
\sigma_{\rm disc}({\mathcal{A}})=\emptyset.
\end{gather}
Now, we look closely at the spectrum of ${\mathcal{A}}_{\varepsilon}$.
From \eqref{essAA} and \eqref{A:ess} we deduce
\begin{gather}
\label{Ae:ess}
\sigma_{\rm ess}({\mathcal{A}}_{\varepsilon})=[\Lambda',\infty).
\end{gather}
Furthermore, one can find such a function
$v_{\varepsilon}\in \mathrm{dom}(\a_{\varepsilon})$ that
\begin{gather}\label{ve:RelQ}
{\a_{\varepsilon}[v_{\varepsilon},v_{\varepsilon}]\over \|v_{\varepsilon}\|^2_{\L(\Omega_{\varepsilon})}}<
\Lambda'\text{ for sufficiently small $\varepsilon$}.
\end{gather}
For convenience of presentation, we postpone
the construction of $v_{\varepsilon}$ and the justification of \eqref{ve:RelQ}
to the end of this subsection.
Now, using \eqref{Ae:ess}, \eqref{ve:RelQ} and
the min-max principle (see, e.g., \cite[Section~4.5]{D95}), we conclude that
\begin{gather}
\label{Ae:disc}
[0,\Lambda')\cap\sigma_{{\rm disc}}({\mathcal{A}}_{\varepsilon})\not=\emptyset \text{ for sufficiently small $\varepsilon$.}
\end{gather}
From \eqref{Ae:disc} we immediately infer that
there exists a sequence $(\lambda_{\varepsilon_k})_{k\in\mathbb{N}}$ with $\lambda_{\varepsilon_k}\in\sigma_{{\rm disc}}({\mathcal{A}}_{\varepsilon_k})$ converging to some $\lambda_0\in [0,\Lambda']$ as $\varepsilon_k\to 0$. By Theorem~\ref{th1} $\lambda_0\in\sigma({\mathcal{A}})$. On the other hand $\sigma({\mathcal{A}})\cap [0,\Lambda']=\{\Lambda'\}$, whence $\lambda_0=\Lambda'\in\sigma_{{\rm ess}}({\mathcal{A}})$,
which means
(cf.~Remark~\ref{rem:equiv}) that the property
$\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon})\searrow\sigma_{\rm disc}({\mathcal{A}})$ does not hold true.
It remains to construct the function $v_{\varepsilon}$ satisfying \eqref{ve:RelQ}.
We choose it as follows:
\begin{align*}
v_{\varepsilon}=
\begin{cases}\displaystyle
{1\over 2}H_{1,\varepsilon}(x)\phi_{1,\varepsilon}(x)
|B_{1,\varepsilon}|^{-1/2},&x\in\Omega_{\varepsilon}^{\rm out},\\\displaystyle
\left(1-{1\over 2} H_{1,\varepsilon}(x)\phi_{1,\varepsilon}(x)\right)|B_{1,\varepsilon}|^{-1/2},&x\in B_{1,\varepsilon},\\\displaystyle
{1\over 2}|B_{1,\varepsilon}|^{-1/2},&x\in D_{1,\varepsilon},
\end{cases}
\end{align*}
i.e. $v_{\varepsilon}=\J_{\varepsilon}^1 g$, where
$g=(0,1)\in\mathcal{H}^1=\H^1_0(\Omega)\times\mathbb{C}$.
It is easy to see that
\begin{align*}
\a_{\varepsilon}[v_{\varepsilon},v_{\varepsilon}]
={1\over 4}|B_{1,\varepsilon}|^{-1}
\|\nabla (H_{1,\varepsilon}\phi_{1,\varepsilon})\|^2_{\L(\mathbb{R}^n )}
={1\over 4}|B_{1,\varepsilon}|^{-1} \left(
\mathrm{cap}(D_{1,\varepsilon})+U_{\varepsilon,1}+U_{\varepsilon,2}\right),
\end{align*}
where
$$U_{\varepsilon,1}\coloneqq\|\nabla (H_{1,\varepsilon}(\phi_{1,\varepsilon}-1))\|^2_{\L(\mathbb{R}^n )},\quad U_{\varepsilon,2}\coloneqq 2(\nabla H_{1,\varepsilon},\nabla (H_{1,\varepsilon}(\phi_{1,\varepsilon}-1)))_{\L(\mathbb{R}^n)}.$$
Taking into account that
$|\phi_{1,\varepsilon}|\leq C$ and $|\nabla\phi_{1,\varepsilon}|\leq C\varepsilon^{-1}$,
we estimate $U_{1,\varepsilon}$ as follows:
\begin{align}\label{inter:1}
\|\nabla (H_{1,\varepsilon}(\phi_{1,\varepsilon}-1))\|^2_{\L(\mathbb{R}^n )}\leq
C\left(\|\nabla H_{1,\varepsilon}\|^2_{\L(\mathrm{supp}(\phi_{1,\varepsilon}-1))}+
\varepsilon^{-2}\|H_{1,\varepsilon}\|^2_{\L(\mathrm{supp}(\nabla\phi_{1,\varepsilon}))}\right).
\end{align}
One has (see~\eqref{HL2}; recall that we assume $n\ge 3$):
\begin{gather}\label{HL2+}
\|H_{k,\varepsilon}\|^2_{\L(\mathrm{supp}(\nabla\phi_{k,\varepsilon}))}
\leq C_1\varepsilon^{n}\eta^2_{\varepsilon}=C_1\varepsilon^{n+4}.
\end{gather}
Similarly, using Lemma~\ref{lemma:Hest} (note that $\mathrm{supp}(\phi_{1,\varepsilon}-1)\subset\mathbb{R}^n\setminus\mathscr{B}(\ell_1\varepsilon,z_1)$, whence the estimates \eqref{H:estimates} are valid due to Remark~\ref{rem:Hest}), we get
\begin{gather}\label{nablaHL2}
\|\nabla H_{1,\varepsilon}\|^2_{\L(\mathrm{supp}(\phi_{1,\varepsilon}-1))}
\leq
C\varepsilon^{n+2}.
\end{gather}
Combining \eqref{inter:1}--\eqref{nablaHL2}, we obtain
\begin{gather}\label{U1}
U_{1,\varepsilon}
\leq
C\varepsilon^{n+2}.
\end{gather}
Furthermore, using \eqref{nablaHL2} and \eqref{U1}, we arrive at the estimate for $U_{2,\varepsilon}$:
\begin{gather}\label{U2}
|U_{2,\varepsilon}|\leq
2\|\nabla H_{1,\varepsilon}\|_{\L(\mathrm{supp}(\phi_{1,\varepsilon}-1))}
\sqrt{U_{1,\varepsilon}} \leq
C\varepsilon^{n+2}.
\end{gather}
Combining \eqref{U1} and \eqref{U2}, we infer
\begin{gather}\label{ae:asy1}
\a_{\varepsilon}[v_{\varepsilon},v_{\varepsilon}]=
{1\over 4}|B_{1,\varepsilon}|^{-1}\left(\mathrm{cap}( {D_{1,\varepsilon}})+\mathcal{O}(\varepsilon^{n+2}) \right)=
\gamma_{1,\varepsilon}+\mathcal{O}(\varepsilon^2)=
\Lambda'-\varepsilon^{1/2}+\mathcal{O}(\varepsilon^2).
\end{gather}
Similarly, we derive the asymptotics
\begin{gather}\label{ae:asy2}
\|v_{\varepsilon}\|^2_{\mathcal{H}_{\varepsilon}}=
1+\mathcal{O}(\varepsilon).
\end{gather}
The required estimate \eqref{ve:RelQ} follows from \eqref{ae:asy1}--\eqref{ae:asy2}.
\section{Waveguide with prescribed eigenvalues\label{sec:5}}
The spectrum of the Dirichlet Laplacian on the straight tubular domain
coincides with $[\Lambda',\infty)$, where $\Lambda'>0$ is the smallest eigenvalue of the Dirichlet Laplacian on the tube cross-section. Let us perturb this tube by narrowing
its bounded part and then by inserting $m$ resonators within this narrowed part.
Such a perturbation does not change the essential spectrum, but may produce
discrete eigenvalues below $\Lambda'$. Our goal is to show that these eigenvalues can be made coinciding with prescribed numbers via a suitable choice of the parameters $d_{k,\varepsilon}$.
Below we formulate the problem and the result precisely.
Let $\Omega'$ and $\Omega''$ be bounded Lipschitz domains in $\mathbb{R}^{n-1}$ ($n\ge 2$) such that $\overline{\Omega''}\subset\Omega'$.
Let $L>0$. We introduce the following domains in $\mathbb{R}^n$:
$$
\widetilde\Omega\coloneqq \Omega''\times (-L,L),\quad
\Omega^+\coloneqq \Omega'\times (L,\infty),\quad
\Omega^-\coloneqq \Omega'\times(-\infty,-L),\quad
S^\pm=\Omega''\times\{\pm L\}.
$$
Further, let $S_{k,\varepsilon}\subset \widetilde\Omega$, $k\in\mathbb{M}=\{1,\dots,m\}$
be the sets we introduced in Section~\ref{sec:2}, namely
\begin{gather*}
S_{k,\varepsilon}=(\partial(\underbrace{\varepsilon B_k + z_k}_{B_{k,\varepsilon}}))\setminus (\underbrace{d_{k,\varepsilon} D_k+z_k}_{D_{k,\varepsilon}}).
\end{gather*}
Here $\varepsilon>0$,
$(z_k)_{k\in\mathbb{M}}$ are pairwise distinct points in $\widetilde\Omega$,
$(B_k)_{k\in\mathbb{M}}$ and $(D_k)_{k\in\mathbb{M}}$
are sets in $\mathbb{R}^n$ satisfying \eqref{DB1}--\eqref{DB2}, finally, the numbers $(d_{k,\varepsilon})_{k\in\mathbb{M}}$ are specified as follows:
\begin{gather}
d_{k,\varepsilon}=
\begin{cases}
d_k\varepsilon^{n\over n-2},&n\ge 3,
\\
\exp\left(-{1\over d_k\varepsilon^2}\right),&n=2,
\end{cases}\label{dke}
\end{gather}
where $d_k$ are positive constants. The parameter $\varepsilon$ is supposed to be sufficiently small
in order to have
$\overline{B_{k,\varepsilon}}\subset \widetilde\Omega $ and $\overline{B_{k,\varepsilon}}\cap \overline{B_{l,\varepsilon}}=\emptyset$ if $k\not=l$.
We set
$$\widetilde\Omega_{\varepsilon} \coloneqq \widetilde\Omega\setminus \left(\bigcup\limits_{k\in\mathbb{M}} S_{k,\varepsilon}\right).$$
Finally, we connect $\widetilde\Omega_{\varepsilon}$ (via $S^\pm$) with $\Omega^-$ and $\Omega^+$, and
arrive at the domain (see Figure~\ref{fig-waveguide})
$$
\Omega_{\varepsilon}\coloneqq \Omega^-\cup S^-\cup\widetilde\Omega_{\varepsilon}\cup S^+\cup\Omega^+
$$
We also introduce the set
$$
\Omega\coloneqq \Omega^-\cup S^-\cup\widetilde\Omega\cup S^+\cup\Omega^+.
$$
As before, ${\mathcal{A}}_{\varepsilon}$ stands for the operator acting in $\L(\Omega_{\varepsilon})$ and
being associated with the sesquilinear form \eqref{ae}, i.e. ${\mathcal{A}}_{\varepsilon}$ is the Laplace operator on $\Omega_{\varepsilon}$ subject to the Neumann boundary conditions on $\cup_{k\in\mathbb{M}}S_{k,\varepsilon}$ and the Dirichlet boundary conditions on $\partial\Omega$.
Using standard methods of perturbation theory one can easily demonstrate that the essential spectrum of ${\mathcal{A}}_{\varepsilon}$ coincides with the essential spectrum of
the Dirichlet Laplacian $-\Delta^D_{\Omega'\times\mathbb{R}}$ on the unperturbed waveguide $\Omega'\times\mathbb{R}$,
i.e.
\begin{gather}\label{ess:ess:1}
\sigma_{\rm ess}({\mathcal{A}}_{\varepsilon})=\sigma_{\rm ess}(-\Delta_{\Omega'\times\mathbb{R}}^D)=[\Lambda',\infty),
\text{ where } \Lambda'\coloneqq \lambda_1(-\Delta_{\Omega'}^D).
\end{gather}
Thus the discrete spectrum of ${\mathcal{A}}_{\varepsilon}$ (if any) belongs to $[0,\Lambda')$.
The theorem below asserts that, under a suitable choice of the constants $d_k$ in \eqref{dke},
$\sigma_{{\rm disc}}({\mathcal{A}}_{\varepsilon})$ has $m$ simple eigenvalues, which coincide with prescribed numbers.
The role of the local narrowing is to guaranteed that
there are no further eigenvalues in the vicinity of $\Lambda'$.
Note that solely the local narrowing will not change the spectrum. Indeed,
similarly to \eqref{ess:ess:1}, we get
\begin{gather*}
\sigma_{\rm ess}(-\Delta_{\Omega}^D)=\sigma_{\rm ess}(-\Delta_{\Omega'\times\mathbb{R}}^D)=[\Lambda',\infty).
\end{gather*}
Moreover, $-\Delta_{\Omega}^D$ has no eigenvalues below $\Lambda'$
(this follows easily from
$\sigma_{{\rm disc}}(-\Delta_{\Omega'\times\mathbb{R}}^D)=\varnothing$,
$\Omega\subset\Omega'\times\mathbb{R}$
and the min-max principle).
To formulate the result we introduce the functions $\mathcal{F}_k:\mathbb{R}_+\to\mathbb{R}_+$, $k\in\mathbb{M}$, via
\begin{gather*}
\mathcal{F}_k(t)=
\begin{cases}
\displaystyle\left(t{4|B_k|\over \mathrm{cap}(D_k)}\right)^{1\over n-2} ,&n\ge 3,\\
\displaystyle t{2|B_k| \over \pi},&n=2.
\end{cases}
\end{gather*}
In the following, in order to emphasize the dependence of ${\mathcal{A}}_{\varepsilon}$ (and its spectrum) on $ d_1,\dots,d_m$,
we will use the notation ${\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$ instead of ${\mathcal{A}}_{\varepsilon}$.
\begin{theorem}\label{th:exact}
Let $\widetilde\gamma_k$, $k\in\mathbb{M}$ be arbitrary numbers satisfying
\begin{gather*}
0<\widetilde\gamma_1<\widetilde\gamma_2<\dots<\widetilde\gamma_m<\Lambda'.
\end{gather*}
Then for any $\delta>0$ there exist $\varepsilon=\varepsilon(\delta)>0$ and $\widetilde d_k\in (\mathcal{F}_k(\widetilde\gamma_k)-\delta,\mathcal{F}_k(\widetilde\gamma_k)+\delta)$, $k\in\mathbb{M}$
such that
$$
\sigma_{{\rm disc}}({\mathcal{A}}_{\varepsilon}^{\widetilde d_1,\dots,\widetilde d_m})=\cup_{k\in\mathbb{M}}\{\widetilde\gamma_k \},
$$
and
the eigenvalues $\widetilde\gamma_k$ are simple.
\end{theorem}
\begin{remark}\label{rem:almost}
The above theorem asserts that $\widetilde d_k$
can be indicated ``almost explicitly'': they belong to the $\delta$-neighbourhoods of
the numbers $\mathcal{F}_k(\widetilde\gamma_k)$, where $\delta>0$ can be chosen arbitrary small.
Of course, the smaller $\delta$ is chosen, the smaller $\varepsilon(\delta)$ will be.
\end{remark}
\begin{proof}[Proof of Theorem~\ref{th:exact}]
Let us fix $\delta>0$.
Let us also six sufficiently small $\eta>0$ satisfying
\begin{gather}\label{eta:smallness:1}
\forall k\in\mathbb{M}:\quad [\widetilde\gamma_k-2\eta,\widetilde\gamma_k+2\eta]\subset (0,\Lambda'),\\[2mm]
\label{eta:smallness:2}
\forall k,l\in\mathbb{M},\ k\not=l:\quad
[\widetilde\gamma_k-2\eta,\widetilde\gamma_k+2\eta]\cap[\widetilde\gamma_l-2\eta,\widetilde\gamma_l+2\eta]=\emptyset,\\[2mm]
\label{ddist}
\max_{k\in\mathbb{M}}(\mathcal{F}_k(\widetilde\gamma_k+\eta)-\mathcal{F}_k(\widetilde\gamma_k-\eta))<\delta
\end{gather}
(the last property can be achieved since $\mathcal{F}_k$, $k\in\mathbb{M}$ are continuous functions).
We denote
$$
d_k^\pm\coloneqq \mathcal{F}_k(\widetilde\gamma_k\pm\eta).
$$
Due to a strict
monotonicity of $\mathcal{F}_k$,
we have
$0< d_k^-<d_k^+$. Finally, we introduce the set
$$
\mathcal{D}\coloneqq \prod\limits_{k=1}^m [d^-_k, d^+_k]\subset\mathbb{R}^m.
$$
By virtue of Theorems~\ref{th1}--\ref{th2} one has for each $(d_1,\dots,d_m)\in\mathcal{D}$:
\begin{gather}\label{recall}
\begin{array}{l}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\to \sigma({\mathcal{A}}^{d_1,\dots,d_m}),\\[2mm]
\sigma_{\rm disc}({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\nearrow \sigma_{\rm disc}({\mathcal{A}}^{d_1,\dots,d_m})
\text{ with the multiplicity being preserved.}
\end{array}
\end{gather}
Here ${\mathcal{A}}^{d_1,\dots,d_m}$ is an operator in $\L(\Omega)\oplus\mathbb{C}^m$
given by
$${\mathcal{A}}^{d_1,\dots,d_m}=(-\Delta_{\Omega}^D)\oplus\mathrm{diag}\{\gamma_1^{d_1},\dots,\gamma_m^{d_m}\},$$
with
$\gamma_k^{d_k}=\lim_{\varepsilon\to 0}\mathrm{cap}(D_{k,\varepsilon})|B_{k,\varepsilon}|^{-1}$. It follows from
\eqref{dke} and \eqref{cap:rescaling1}--\eqref{cap:rescaling2} that
$$
\gamma_k^{d_k}=\mathcal{F}^*_k(d_k),
$$
where $\mathcal{F}^*_k:\mathbb{R}_+\to\mathbb{R}_+$ is a strictly increasing function defined by
$$
\mathcal{F}^*_k(s)=
\begin{cases}
\displaystyle s^{n-2}{\mathrm{cap}(D_k)\over 4|B_k|},&n\ge 3,\\
\displaystyle s {\pi\over 2|B_k|},&n=2.
\end{cases}
$$
Observe that
$\mathcal{F}_k(\mathcal{F}^*_k(s))=s$ and $\mathcal{F}^*_k(\mathcal{F}_k(t))=t$;
we have
\begin{gather}\label{gagaga}
\begin{array}{c}
\gamma_k^{d_k^-}=\widetilde\gamma_k-\eta< \widetilde\gamma_k< \widetilde\gamma_k+\eta=\gamma_k^{d_k^+},\quad
\gamma_k^{d_k}\in [\widetilde\gamma_k-\eta,\widetilde\gamma_k+\eta]\text{ as }d_k\in [d_k^-,d_k^+].
\end{array}
\end{gather}
Let us fix some $\widehat\Lambda\in(\gamma_m+2\eta,\Lambda')$.
\begin{lemma}\label{lm:spec:struct}
There exists $\varepsilon'>0$ such that
for any $\varepsilon\in(0,\varepsilon']$ and $(d_1,\dots,d_m)\in\mathcal{D}$ the spectrum of
${\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$ has the following structure within $ [0,\widehat\Lambda]$:
\begin{gather}
\label{spec:d}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap [0,\widehat\Lambda]=\bigcup\limits_{k\in\mathbb{M}}\{\gamma_{k,\varepsilon}^{d_1,\dots,d_m}\},
\end{gather}
where the numbers $\gamma_{k,\varepsilon}^{d_1,\dots,d_m}$ are simple eigenvalues satisfying
\begin{gather}\label{3/2}
\widetilde\gamma_{k}-{3\eta\over 2}\leq \gamma_{k,\varepsilon}^{d_1,\dots,d_m}\leq \widetilde\gamma_{k}+{3\eta\over 2}.
\end{gather}
Moreover, for each $k\in\mathbb{M}$ one has
\begin{gather}
\label{Hempel:ineq}
\gamma_{k,\varepsilon}^{d_1^+,d_2^+,\dots,d_{k-1}^+,d_k^-,d_{k+1}^+,\dots,d_m^+}
<\widetilde\gamma_k<
\gamma_{k,\varepsilon}^{d_1^-,d_2^-,\dots,d_{k-1}^-,d_k^+,d_{k+1}^-,\dots,d_m^-}.
\end{gather}
\end{lemma}
\begin{proof}
Due to \eqref{eta:smallness:1}, \eqref{eta:smallness:2}, \eqref{gagaga},
the intervals $[\gamma_k^{d_k^-}-{\eta\over 2}, \gamma_{k}^{d_k^-}+{\eta\over 2}]$, $k\in\mathbb{M}$ are pairwise disjoint and belong to $(0,\widehat\Lambda)$.
Hence, by virtue of
\eqref{ess:ess:1} and
\eqref{recall}, for sufficiently small $\varepsilon$ the spectrum of ${\mathcal{A}}^{d_1^-,\dots,d_m^-}$
withing the interval $[0,\widehat\Lambda]$ consists of $m$ simple eigenvalues
so that in each interval $[\gamma_k^{d_k^-}-{\eta\over 2}, \gamma_{k}^{d_k^-}+{\eta\over 2}]$ one has precisely one eigenvalue:
there exist $\varepsilon_->0$ such that
\begin{gather}\label{spec-}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1^-,\dots,d_m^-})\cap[0,\widehat\Lambda]=\bigcup\limits_{k\in\mathbb{M}}\{\gamma_{k,\varepsilon}^-\}
\text{ for } \varepsilon\in (0,\varepsilon_-],
\end{gather}
where the numbers $\gamma_{k,\varepsilon}^-$ are simple eigenvalues satisfying
\begin{gather}
\gamma_{k,\varepsilon}^-\in
[\gamma_k^{d_k^-}-{\eta\over 2}, \gamma_{k}^{d_k^-}+{\eta\over 2}]=
[\widetilde\gamma_k-{3\eta\over 2}, \widetilde\gamma_{k}-{\eta\over 2}].
\end{gather}
Similarly, there exist $\varepsilon_+>0$ such that
\begin{gather}\label{spec+}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1^+,\dots,d_m^+})\cap [0,\widehat\Lambda]=\bigcup\limits_{k\in\mathbb{M}}\{\gamma_{k,\varepsilon}^+\}
\text{ for } \varepsilon\in (0,\varepsilon_+],
\end{gather}
where the numbers $\gamma_{k,\varepsilon}^+$ are simple eigenvalues satisfying
\begin{gather}
\gamma_{k,\varepsilon}^+\in [ \gamma_k^{d_k^+}-{\eta\over 2}, \gamma_{k}^{d_k^+}+{\eta\over 2}]=
[\widetilde\gamma_k+{\eta\over 2}, \widetilde\gamma_{k}+{3\eta\over 2}].
\end{gather}
Let $\a_{\varepsilon}^{d_1,\dots,d_m}$ be the sesquilinear form associated
with ${\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$.
It is easy to see that for any
$(d^1_1,\dots,d^1_m)\in\mathcal{D}$,
$(d^2_1,\dots,d^2_m)\in\mathcal{D}$ satisfying $d_k^1\leq d_k^2$, one has
\begin{gather}\label{a:monoton}
\begin{array}{c}
\mathrm{dom}(\a_{\varepsilon}^{d^1_1,\dots,d^1_m})\supset \mathrm{dom}(\a_{\varepsilon}^{d_1^2,\dots,d_m^2}),\\[2mm]
\forall u\in\mathrm{dom}(\a_{\varepsilon}^{d^2_1,\dots,d^2_m}): \ \a_{\varepsilon}^{d_1^1,\dots,d_m^1}[u,u]=
\a_{\varepsilon}^{d_1^2,\dots,d_m^2}[u,u].
\end{array}
\end{gather}
Then, using the min-max principle, we conclude from \eqref{spec-}--\eqref{a:monoton} that
\eqref{spec:d}--\eqref{3/2} hold for $\varepsilon\le \min\{\varepsilon_-,\varepsilon_+\}$; moreover,
by Theorem~\ref{th1},
$
\gamma_{k,\varepsilon}^{d_1,\dots,d_m}\to \gamma_{k}^{d_k}\text{ as }\varepsilon\to 0$.
In particular, one has
\begin{gather}\label{Hempel:lim}
\lim_{\varepsilon\to 0}\gamma_{k,\varepsilon}^{d_1^+,d_2^+,\dots,d_{k-1}^+,d_k^-,d_{k+1}^+,\dots,d_m^+}=\gamma_k^{d_k^-},
\quad
\lim_{\varepsilon\to 0}\gamma_{k,\varepsilon}^{d_1^-,d_2^-,\dots,d_{k-1}^-,d_k^+,d_{k+1}^-,\dots,d_m^-}=\gamma_k^{d_k^+}.
\end{gather}
Due to \eqref{gagaga} and \eqref{Hempel:lim} there exists $\varepsilon'\le \min\{\varepsilon_-,\varepsilon_+\}$
such that
\eqref{Hempel:ineq} holds
for $\varepsilon\in (0,\varepsilon']$.
The lemma is proven.
\end{proof}
\begin{lemma}\label{lm:no:spec}
There exists $\widehat\varepsilon\in (0,\varepsilon']$ such that $(d_1,\dots,d_m)\in\mathcal{D}$ and
for any $\varepsilon\in(0,\widehat\varepsilon]$ one has
\begin{gather}\label{nospectrum}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap (\widehat\Lambda,\Lambda')=\emptyset.
\end{gather}
\end{lemma}
\begin{proof}
For the proof we use bracketing technique.
We represent $\L(\Omega_{\varepsilon})$
as a direct sum
$$\L(\Omega_{\varepsilon})=\L(\Omega^-)\oplus \L(\widetilde\Omega_{\varepsilon})\oplus\L(\Omega^+).$$
With respect to this space decomposition we introduce the operator
$$\widehat{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}={\mathcal{A}}^-\oplus\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}\oplus{\mathcal{A}}^+,$$
where
\begin{itemize}
\item the operator $\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$ is associated with the sesquilinear form
$\widetilde\a_{\varepsilon} $ given by\\
$\displaystyle\widetilde\a_{\varepsilon} [u,v]=\int_{\widetilde\Omega_{\varepsilon}}\nabla u\cdot\overline{\nabla v}\d x,\ \mathrm{dom}(\widetilde\a_{\varepsilon} )=\{u\in \H^1(\widetilde\Omega_{\varepsilon}):\ u=0\text{ on }\partial\Omega''\times(-L,L)\}.$\\
In other words, $\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$ is the Laplacian on $\widetilde\Omega_{\varepsilon}$
subject to the Dirichlet boundary conditions on $\partial\Omega''\times (-L,L)$ and the Neumann boundary conditions on $S^\pm$.
\item the operator ${\mathcal{A}}^-$ (respectively, ${\mathcal{A}}^+$) in the Laplace operator on $\Omega^-$ (respectively, on $\Omega^+$) subject to the
Neumann boundary conditions on
$S^-$ (respectively, on $S^+$), and the Dirichlet boundary conditions on
$(\partial\Omega^-)\setminus S^-$ (respectively, on $(\partial\Omega^+)\setminus S^+$).
\end{itemize}
We denote by $\widetilde{\mathcal{A}}$ the Laplacian on $\widetilde\Omega$
subject to the
Neumann boundary conditions on
$S^\pm$, and the Dirichlet boundary conditions on
$\partial\Omega''\times (-L,L)$.
It is easy to see that
\begin{gather}\label{inf''}
\inf\sigma(\widetilde{\mathcal{A}})=\Lambda''\coloneqq\lambda_1(-\Delta_{\Omega''}^D).
\end{gather}
Also, since $\overline{\Omega''}\subset\Omega'$, we get (see, e.g, \cite[Subsection~1.3.2]{He06})
\begin{gather}\label{LaLa}
\Lambda'<\Lambda''.
\end{gather}
By virtue of Theorems~\ref{th1}--\ref{th2} (taking into account Remark~\ref{rem:Neumann}) we have
\begin{gather}\label{recall:wt}
\begin{array}{l}
\sigma(\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\to \sigma(\widetilde{\mathcal{A}}^{d_1,\dots,d_m}),
\\[2mm]
\sigma_{\rm disc}(\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\nearrow \sigma_{\rm disc}(\widetilde{\mathcal{A}}^{d_1,\dots,d_m})
\text{ with the multiplicity being preserved},
\end{array}
\end{gather}
where $\widetilde{\mathcal{A}}^{d_1,\dots,d_m}=\widetilde{\mathcal{A}}\oplus\mathrm{diag}\{\gamma_1^{d_1},\dots,\gamma_m^{d_m}\}.$
Combining \eqref{inf''}--\eqref{recall:wt} and using the same arguments as in the proof of \eqref{spec:d},
we conclude there exists $\widetilde\varepsilon>0$ such that
$ \sigma(\widetilde{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap[0,\Lambda')$
consists of precisely $m$ simple eigenvalues
provided $\varepsilon\in (0,\widetilde\varepsilon]$ and $(d_1,\dots,d_m)\in\mathcal{D}$.
Hence, since $\sigma({\mathcal{A}}^\pm)=[\Lambda',\infty)$, we infer
\begin{gather}\label{nospectrum:wh}
\sigma(\widehat{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap[0,\Lambda')\text{ consists of }m\text{ simple eigenvalues}
\end{gather}
provided $\varepsilon\in(0,\widetilde\varepsilon]$ and $(d_1,\dots,d_m)\in\mathcal{D}$.
Finally, we observe that $\widehat{\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}\leq {\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m}$ in the form sense. Using this fact and the min-max principle, we conclude from \eqref{nospectrum:wh} that
\begin{multline*}
\sigma({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap[0,\Lambda')\text{ consists of \emph{at most} }m\text{ eigenvalues} \\\text{(with multiplicities taken into account)}
\end{multline*}
for $\varepsilon\in(0,\widetilde\varepsilon]$, $(d_1,\dots,d_m)\subset\mathcal{D}$.
On the other hand (see \eqref{spec:d}) $\sigma({\mathcal{A}}_{\varepsilon}^{d_1,\dots,d_m})\cap[0,\widetilde\Lambda]$ consists of
precisely $m$ simple eigenvalues for $\varepsilon\in(0,\varepsilon']$.
Hence \eqref{nospectrum} holds with $\widehat\varepsilon\coloneqq\min\{\varepsilon',\widetilde\varepsilon\}$.
\end{proof}
We need the following multi-dimensional version
of the intermediate value theorem.
\begin{lemma}[{\cite[Lemma~3.5]{HKP97}}]\label{lemma-hempel}
Let $\mathcal{D}=\Pi_{k=1}^n[a_k, b_k]$ with $a_k < b_k$, $k=1,\dots,n$, assume that
$f:\mathcal{D}\to\mathbb{R}^n$ is continuous and each coordinate function $f_k$ of $f$ is
monotonically increasing in each of its arguments. If
$F_k^-<F_k^+$, $k=1,\dots,n$, where
\begin{gather*}
F_k^-\coloneqq f_k(b_1,b_2,\dots,b_{k-1},a_k,b_{k+1},\dots,b_n),\\
F_k^+\coloneqq f_k(a_1,a_2,\dots,a_{k-1},b_k,a_{k+1},\dots,b_n),
\end{gather*}
then for any $F\in\Pi_{k=1}^n[F_k^-,F_k^+]$
there exists $x\in\mathcal{D}$ such that $f(x)=F$.
\end{lemma}
Now, we are ready to finish the proof of Theorem~\ref{th:exact}.
Let us \textit{fix} $\varepsilon\in (0,\widehat\varepsilon]$; with this $\varepsilon$ the properties \eqref{spec:d}--\eqref{Hempel:ineq}, \eqref{nospectrum} hold true.
We introduce $f=(f_1,\dots,f_m):\mathcal{D}\to\mathbb{R}^m$
by
$$
f_k(d_1,\dots,d_m)= \gamma_{k,\varepsilon}^{d_1,\dots,d_m}.
$$
By the min-max principle, each $f_k$ is
monotonically increasing in each of its arguments.
Moreover, each $f_k$ depends continuously on $(d_1,\dots,d_m)\in\mathcal{D}$
(for $n=2$ the proof can be found in \cite[Theorem~A.1]{BK19}, in the general case the proof is similar).
Finally, due to \eqref{Hempel:ineq}, one has for each $k\in\mathbb{M}$:
\begin{gather*}
f_k( d_1^+,\dots, d_{k-1}^+, d_k^-, d_{k+1}^+,\dots,\dots, d_m^+)<
\widetilde\gamma_k<
f_k( d_1^-,\dots, d_{k-1}^-, d_k^+, d_{k+1}^-,\dots,\dots,d_m^-)
\end{gather*}
Then, by Lemma~\ref{lemma-hempel}, there exists $(\widetilde d_1,\dots,\widetilde d_m)\in\mathcal{D}$ such
that $f_k(\widetilde d_1,\dots,\widetilde d_m)=\widetilde\gamma_k$, $\forall k\in\mathbb{M}$. Thus, taking into account \eqref{ess:ess:1} and \eqref{nospectrum}, we arrive at
$$\sigma_{{\rm disc}}( {\mathcal{A}}_{\varepsilon}^{\widetilde d_1,\dots,\widetilde d_m})=\cup_{k\in\mathbb{M}}\{\widetilde\gamma_k\}.$$
Evidently, all eigenvalues $\widetilde\gamma_k$ are simple.
Finally, using \eqref{ddist} and taking into account that $\mathcal{F}_k(\widetilde\gamma_k)\in [\mathcal{F}_k(\widetilde\gamma_k-\eta),\mathcal{F}_k(\widetilde\gamma_k+\eta)]=[d_k^-,d_k^+]$,
we get
\begin{gather*}
|\widetilde d_k - \mathcal{F}_k(\widetilde\gamma_k)|\leq d_k^+-d_k^-=\mathcal{F}_k(\widetilde\gamma_k+\eta)-\mathcal{F}_k(\widetilde\gamma_k-\eta)<\delta
\end{gather*}
Theorem~\ref{th:exact} is proven.
\end{proof}
\section*{Acknowledgements}
The work of A.K. is supported by the Czech Science Foundation (GA\v{C}R) through the project 22-18739S.
\bibliographystyle{siamplain}
| 7c411366ed7eea734df33215e4b2706e17155898 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Constructing open source software (OSS) depends on contributors with a diverse set of skills. Each skill is vital to OSS: problem-solving skills allow software developers to build new features to address issues; organizational skills help software maintainers manage the moving parts of an OSS project; and communication skills enable writers to generate clear, concise documentation and facilitate collaboration. Unlike in software engineering, where programmers primarily write code, many OSS contributors provide equally valuable non-code contributions \cite{allcontributors2022, trinkenreich2020hidden}. While OSS-related skills include a subset of software engineering skills, contributors also work in contexts unique to OSS (e.g., wrangling contributors, identifying funding, consistently collaborating in a distributed form).
Despite the importance of skills related to OSS development, to our knowledge, there are no tools that currently exist which detect such skills.
Related work in software engineering has developed techniques to detect specific software engineering skills, for example, Java programming skills \cite{bergersen2014construction} and general programming experience \cite{siegmund2014measuring}.
Meanwhile, other work detects programming-related skills by pulling data from version control systems such as GitHub\xspace \cite{papoutsoglou2019extracting, greene2016cvexplorer, montandon2019identifying, mockus2002icse, montandon2021mining, hauff2015matching}. Montandon et al. showed that this data could help identify experts in OSS communities \cite{montandon2019identifying} and predict technical roles of GitHub\xspace users \cite{montandon2021mining}. Other work has demonstrated that GitHub\xspace data can be used to extract skills for job recommendations \cite{greene2016cvexplorer, hauff2015matching}. %
These studies have significantly advanced the field, but focus largely on a single topic: mining technical programming skills. Thus, a gap still remains in mining OSS expertise, which encapsulates both soft and hard skills \cite{vadlamani2020studying} across many roles aside from software engineering. Rather than simply focusing on technical software development skills, our work focuses on mining GitHub\xspace data to detect both soft skills and hard skills, which are vital to OSS development.
In this paper, we introduce a method to detect OSS skills and implement it in a tool called \ADDED{\toolname (\textbf{D}etect\textbf{I}ng \textbf{SK}ills in \textbf{O}SS)}, with promising results. Our approach relies on identifying accurate \emph{signals}, which are measurable activities or cues associated with a skill used to identify the presence of having that skill. For example, having proficiency in a programming language could be measured using a certain number of lines of code written in that language \cite{bergersen2014construction}. The notion of signals follows prior work, such as Marlow et al., who identified cues that GitHub\xspace users utilized to judge a contributor's coding ability, project-relevant skills, and personality \cite{marlow2013activity}.
We discuss how we identify relevant signals (Section \ref{sec:SkillsSignalsModel}) and outline \toolname's features (Section \ref{sec:ImplementingTool}). We then present an evaluation of \toolname and its results (Section \ref{sec:Evaluation}). Finally, we discuss the implications (Section \ref{sec:Discussion}) and next steps for this research (Section \ref{sec:FuturePlans}).
\section{Identifying Signals for Skills}
\label{sec:SkillsSignalsModel}
\input{tables/skills-signals-table}
To develop \toolname, OSS skills and signals need to be identified. The first author extracted relevant \textbf{OSS skills} by reading prior literature on software engineering expertise \cite{baltes2018towards, ahmed2012evaluating, li2020distinguishes, papoutsoglou2017mining} and social factors of OSS \cite{dias2021makes, gousios2015work, marlow2013activity, steinmacher2015social, tsay2014influence}. These papers were selected to include a broad set of roles and contribution types. The first author read each manuscript and identified key skills. %
Next, we performed additional literature review to identify \textbf{signals} for each skill. In this work, \emph{signals are expressed as true or false statements} (i.e., the signal is either present or not). Potential signals were elicited through the nine papers to identify OSS skills \cite{dias2021makes, tsay2014influence, gousios2015work, steinmacher2015social, baltes2018towards, ahmed2012evaluating, li2020distinguishes, marlow2013activity, papoutsoglou2017mining} and relevant literature on mining OSS activity \cite{bergersen2014construction, hata2021github, chouchen2021anti, chatterjee2021aid, zhou2012make, lee2017one, eluri2021predicting, bao2019large, raman2020stress}. We read through each manuscript, identified potential signals for as many OSS skills as possible, and recorded the source of each signal. We finalized the signals from literature by converting each one as a true or false statement by defining thresholds (e.g., an activity happens at least $N$ times, an activity occurs with a frequency at least at the $N$th percentile). We then generated signals which seemed viable to compute and were representative of the skill.
We designed our own signals for \skill{Is familiar with OSS practices} because it was cited often in literature \cite{ahmed2012evaluating, gousios2015work, steinmacher2015social}, but prior work did not define clear signals for this skill.
During this process, authors consulted three OSS community experts who made significant contributions to OSS and regularly work with OSS stakeholders for their profession. They provided insight on important skills and signals. Based on the experts' feedback and the frequency of the skills cited in literature, we reduced the final model to four skills, which is shown in Table \ref{tab:SkillsSignalsTable}.
\section{Automatically Detecting Skills}
\label{sec:ImplementingTool}
After developing a model of OSS skills and their associated signals (see Section \ref{sec:SkillsSignalsModel}), the first author implemented \toolname in Python to detect a contributor's OSS skills from the model based on GitHub\xspace user data. We use the GitHub\xspace GraphQL API \cite{github2021graphql}, GitHub\xspace REST API \cite{github2021rest}, and GHTorrent \cite{gousios2012ghtorrent} as sources of GitHub\xspace data.
Our tool rates skills on a zero to five scale, where zero represents no proficiency and five represents a high level of proficiency. %
We use the function $rate(N, M)$ to determine skill level, where $N$ represents the number of signals present and $M$ represents the total number of signals a skill has. This function weights each signal equally.
\[
rate(N, M) =
\begin{cases}
0 & \tfrac{N}{M} = 0 \\
1 & 0 < \tfrac{N}{M} \leq 0.2 \\
2 & 0.2 < \tfrac{N}{M} \leq 0.4 \\
3 & 0.4 < \tfrac{N}{M} \leq 0.6 \\
4 & 0.6 < \tfrac{N}{M} \leq 0.8 \\
5 & 0.8 < \tfrac{N}{M} \leq 1.0 \\
\end{cases}
\]
\subsubsection*{Selecting programming languages}
Programming languages that the tool detected were selected based on if they were present in both the 2020 State of the Octoverse \cite{github2020state} and the 2020 Stack Overflow Developer Survey \cite{stackoverflow2020developer}. This resulted in 9 programming languages: C, C\#, Java, JavaScript, PHP, Python, Ruby, Shell, and TypeScript.
\subsubsection*{\ADDEDZ{Excluding third-party libraries as} user contributions}
Contributors often included code from third-party libraries in their commits, which has also been shown in prior work \cite{lopes2017dejavu}. \ADDED{Code from third-party libraries thus should be excluded from contributors' mined contributions.} To that end, we identify popular package managers for each programming language from a Wikipedia article on popular package managers \cite{wikipedia2021list}. We compile a list of installation folder names across all the package managers. Next, we analyze the top folder names for each programming language. We use a list of OSS for social good (OSS4SG) projects (i.e., OSS projects which address a societal issue and target a specific community) from Huang et al. \cite{huang2021leaving} We download the contents of each repository in this list and record its file tree. From this, we identify files written in the language based on file extensions and extract the path from the repository folder. We then retrieve the top 50 folder names associated with each programming language. For each language, we exclude the entire file tree under a folder name from the analysis when: 1) the folder name corresponds to an installation location of the programming language's package manager(s), or 2) the folder name is in the top 50 folder names for the programming language and the list of installation folder names across all package managers.
\subsubsection{Computing distributions}
Some signals rely on the user activity being at a certain percentile and thus depends on an underlying distribution to compare to. To compute this, we generate a list of OSS contributors by recording the top 30 contributors \ADDEDZ{per project} from Huang et al.'s list of 437 OSS4SG projects~\cite{huang2021leaving}. We randomly select 500 GitHub\xspace users from this list. Each distribution is computed based on these 500 users' relevant activity for the distribution.
\section{Evaluation}
\label{sec:Evaluation}
\subsection{Design}
We designed a two-part survey to validate \toolname (see Section \ref{sec:ImplementingTool}):
\begin{enumerate}
\item a Qualtrics survey where participants submitted anonymized responses and
\item a Microsoft Forms survey where participants submitted personal identifying information (PII). \ADDEDZ{This survey was displayed after the completion of the Qualtrics survey.}
\end{enumerate} \ADDEDZ{The survey was implemented in two parts so PII was linked only to a small number of questions.} Participants could choose to only take the Qualtrics survey \ADDEDZ{and not provide any PII}.
Topics in the Qualtrics survey included the importance of our tool's OSS skills and the participants' willingness to display their skills and ratings on GitHub\xspace. Topics in the Microsoft Forms survey included GitHub\xspace usernames and self-assessments of the skills from \toolname.
Some survey questions are shown in Figure~\ref{fig:survey} and the full survey instrument is available as supplemental material~\cite{supplemental-materials}.
We sent the survey to a subset of contributors who authored commits, opened issues and pull requests, or commented on others’ issues and pull requests from Huang et al.'s list of 1,079 OSS and OSS4SG projects \cite{huang2021leaving}. \ADDED{Our survey was sent to a total of 9,095 OSS contributors with a response rate of 5\%.} The Qualtrics survey received 455 responses. The Microsoft Forms survey received 386 responses, resulting in 316 valid usernames with public code contributions from merged pull requests.
After completing the survey, participants could join a sweepstakes to win one of four \$100 Amazon.com gift cards.
\ADDEDZ{The GitHub\xspace usernames allowed us to compare the skill self-assessments with the skills scores computed by \toolname.}
To evaluate \toolname, we focus on its precision---that is, when the participant is confident they have a skill, does \toolname agree? We used two measurements of precision: 1) the precision of users who had self-evaluated skill levels greater than 0 (i.e., detecting the presence of a skill) and 2) the precision of users who had self-evaluated skill levels greater than 3 (i.e., detecting moderate to high skill proficiency).
For the analysis of the survey, we only look at close-ended questions and use standard statistical analysis techniques. We report percentages on how frequently participants agreed or strongly agreed with a statement and how frequently participants said a skill was important or very important. This follows Kitchenham's and Pfleeger's best practices to analyze survey data~\cite{survey-guidelines}.
\begin{figure}
\begin{tcolorbox}[left=-14pt,right=2pt,top=2pt,bottom=2pt]
{\sc \hspace{16pt}Part 1: Qualtrics Survey Questions (anonymous)}
\begin{itemize}
\item \questiontext{Q16}
\item \questiontext{Q34}
\item \questiontext{Q35}
\end{itemize}
\end{tcolorbox}
\vspace{0.5\baselineskip}
\begin{tcolorbox}[left=-14pt,right=2pt,top=2pt,bottom=2pt]
{\sc \hspace{16pt}Part 2: Microsoft Forms Survey Questions}
\begin{itemize}
\item \ADDEDZ{\questiontext{MF5}}
\item \questiontext{MF6}
\item \questiontext{MF7}
\vspace{0.5\baselineskip}
\end{itemize}
\vspace{-0.5\baselineskip}
\end{tcolorbox}
\caption{A subset of the survey questions. The complete survey instrument is in the supplemental materials~\cite{supplemental-materials}. }
\label{fig:survey}
\end{figure}
\subsection{Preliminary Results}
\input{tables/skills-importance}
\input{tables/merged}
\subsubsection{Skill importance} All the detected skills were important to participants; the results are shown in Table \ref{tab:SkillsImportanceTable}. A majority of participants found the soft skills (\skill{Teaches others to be involved in the OSS project}, \skill{Shows commitment towards the OSS project}) to be important. Notably, participants found soft skills more important than hard skills. \skill{Is familiar with OSS practices} was also rated as important by a majority of participants, but was less important than the soft skills.
\subsubsection{Displaying skills} All detected skills would be displayed by a majority of participants (see Table \ref{tab:MergedTable}). Participants were most willing to share \skill{Is familiar with OSS practices}, \skill{JavaScript}, and \skill{Python}. \skill{PHP} was the least popular skill to share. Overall, there was variation in how willing participants were to share certain skills. In the free response, some participants were excited by sharing skills and suggested new languages to detect (e.g., Golang, Rust) or expressed encouragement. Others expressed concerns about the impact of \toolname, questioning whether it should be deployed.
\subsubsection{Tool accuracy} \toolname's performance is displayed in Table \ref{tab:MergedTable}. We find that \toolname identifies the presence of OSS skills with impressive performance, with precision scores ranging from 77\% (\skill{Ruby}) to 97\% (\skill{Is familiar with OSS practices}, \skill{Teaches others to be involved in the OSS project}). However, the overall performance of \toolname drops while detecting moderately skilled to expert users. It is also notable that the rating distributions for \skill{Teaches others to be involved in the OSS project}, \skill{Shows commitment towards the OSS project}, and \skill{Is familiar with OSS practices} are skewed positively in the survey.
\section{Discussion}
\label{sec:Discussion}
Our evaluation shows encouraging results for \toolname. Participants are most excited to display programming language-related skills, which our tool detects with reasonable confidence. However, the high importance placed on soft skills by our participants should not be overlooked. This is especially relevant with emerging neural models that generate code with high quality, such as GitHub\xspace Copilot \cite{github2021copilot}. Given its importance, future versions of \toolname could be enhanced to more accurately detect soft skills.
\subsubsection*{Potential applications.} \ADDED{Our results indicate a potential future for skill detection approaches within OSS development and software engineering.} While skills underlie these activities, they are not widely supported in tooling. Skills detection methods such as ours could be applied in practice to transform existing experiences. For example, \toolname could assist project maintainers with OSS team formation by supporting a search engine for potential collaborators based on skills or automatically recommending contributors for particular OSS roles. Furthermore, \toolname's skill ratings could be used to identify a contributor's potential areas of growth or recommend mentors with the expertise for the contributor's professional goals. Skills could also be displayed publicly to GitHub\xspace profiles---users could self-select skills they were proficient in, while the platform could display a verified badge for detected skills.
\subsubsection*{Limitations.} One limitation of our evaluation is that self-rated skills is a biased measure, as participants may systematically overestimate or underestimate their skills. Additionally, limitations in describing skills in a survey may cause mismatched expectations of a skill. These may contribute to lower evaluation scores. For example, the signals for \skill{Is familiar with OSS practices} was designed to be simple to compute and beginner friendly, but participants rated themselves more harshly on the skill. Thus, when designing experiences with automatic skills detection, transparency in how the skill rating is computed is paramount and should be communicated clearly.
\section{Future Plans}
\label{sec:FuturePlans}
Our preliminary results indicate that there are some promising directions for \toolname. However, additional steps are required to improve upon our current approach, which we outline below.
\smallskip
\emph{Identifying additional signals from practitioners.} We hope to interview OSS contributors to understand how they evaluate their peers' expertise for the skills our tool detects. This could generate new signals to improve the accuracy of our tool. 258 participants have agreed to be interviewed for this work.
\smallskip
\emph{Defining weights of the signals.} In practice, each signal is not equally weighted in predicting a contributor's skill. Since our current approach does not support this, one next step could use a linear regression model to examine the relationship between signals and participants' self-evaluated skill level and then use the resulting model weights to weigh each signal in our tool.
\smallskip
\emph{Evaluating the tool with other measures of skill.} We hope to run an additional evaluation using data sources less impacted by personal bias. For example, soft skills may be evaluated using peer evaluations, while hard skills may be evaluated using skill tests. Previous work has administered skill tests to determine Java expertise \cite{bergersen2014construction}.
\section{Conclusion}
We present \toolname, a tool to detect OSS-related skills which identifies \emph{signals} (i.e., measurable activities or cues associated with the skill) and then computes them from GitHub\xspace data. \toolname detects the following skills: \skill{Teaches others to be involved in the OSS project}, \skill{Shows commitment towards the OSS project}, \skill{Has knowledge in specific programming languages}, and \skill{Is familiar with OSS practices}. We demonstrate this approach yields positive results, as \toolname detects the presence of OSS skills with precision scores between 77\% to 97\%. Additionally, a near majority of participants find the detected skills important to OSS. We expect to improve the tool and perform more rigorous evaluations in the future. Future work could design tools to augment existing OSS experiences or improve upon our current approach. Our supplemental materials are publicly available at \cite{supplemental-materials}.
\begin{acks}
We thank our survey participants for their insight and Christian Bird, Mala Kumar, Victor Grau Serrat, and Lucy Harris for their feedback. Jenny T. Liang conducted this work for an internship at Microsoft Research's Software Analysis and Intelligence Group.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
| ddcb836a9001922b2354301baface29b6ec46357 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Randomized experimentation is a fundamental tool for obtaining counterfactual estimates.
The efficacy of randomization comes from a very simple intuition--by randomly assigning treatment status dependence between observed (and unobserved) treatment and pre-treatment covariates necessarily tends to zero as a function of the number of units.
In the context of experimentation, this independence condition on observed covariates, commonly known as balance~\citep{imai2008misunderstandings, imai2014covariate}, reduces the variance of estimates of the average treatment effect~\citep{greevy2004optimal, higgins2016improving, kallus2018optimal, li2018asymptotic, harshaw2020balancing}.
Under the appropriate conditions such corrections can result in large increases in effective sample size, allowing for the detection of the small effects which are commonplace in many large scale studies and industrial applications~\citep{dimmery2019shrinkage, azevedo2020b}.
These contexts rely heavily on experimentation for decision-making, so reduced variance directly translates into more reliable decisions~\citep{kohavi2012trustworthy}.
Traditional experimental design like blocking~\citep{greevy2004optimal, higgins2016improving} or even the novel Gram-Schmidt Walk design~\citep{harshaw2020balancing} require more than one pass over the sample and their sample complexity is greater than $\mathcal{O}(n)$.
Even algorithms which admit sequential assignment such as \citet{moore2017blocking} suffer from the fact that the algorithm is not linear time and, thus, respondents late in the experiment may take substantially longer to receive a treatment assignment~\citep{cavaille2018implementing}.
Our work is motivated by this setting to provide a linear-time, single-pass (i.e. sequential) algorithm for balancing experimental design.
Our focus is on linear measures of balance (often of particular interest to applied researchers).
This provides a new avenue through which experimenters can ensure that their experiments optimize the information they gain from costly samples.
We start by presenting four desiderata for effective, practical online experimental design.
First, a method must be computationally efficient.
In a review of existing methods for online assignment in the case of survey experiments, \citet{cavaille2018implementing} finds that existing methods become slow to unusable as increasingly more respondents are included in a study.
This resulting speed is fundamentally incompatible with effective administration of an experiment.
In short, high latency will cause disproportionately large dropoff in an experiment, which may completely nullify the gains from using more sophisticated experimental design.
Any algorithm with greater than linear time complexity exacerbate this problem: higher latency for later subjects than earlier subjects will tend to cause non-random sample attrition, as later subjects (who may be different than earlier respondents) will be more likely to drop out.
Second, an experimental design must reduce covariate imbalance to be effective.
This is the entire justification for using methods more sophisticated than Bernoulli randomization, so if the design is not able to improve on balance, then there will be no subsequent reduction in variance and therefore no compelling reason to use it.
Third, the design must incorporate randomization.
\citet{harshaw2020balancing} provides an extensive discussion on the inherent tradeoff between robustness and balance within experimental design.
A design which solely optimizes for balance will tend to operate on a knife-edge of accidental bias~\citep{efron1971forcing}: the potential bias from an adversarially chosen confounder.
With higher accidental bias than Bernoulli randomization, if units do not arrive precisely i.i.d., then the entire design may be compromised.
Given that experimental settings are prized specifically for their unbiasedness, this could completely undermine any gains from improved balance.
Strong theoretical guarantees on robustness are extremely important in this setting in order to ensure that inferences rest squarely on the design.
If assumptions about sampling procedures or data generating processes are necessary to ensure the reliability of estimation, then inferences are not based solely on properties of the experiment, but rather on factors outside the control of the experimenter~\citep{aronow2021nonparametric}.
Fourth, units which do not show up in the sample should not be included in the balancing.
An offline algorithm used in an online environment would fail this condition, because to use it would require to balance the entire population of units \emph{expected} to show up to an experiment.
Units within that population who fail to show would nevertheless be assigned a treatment.
Depending on the distribution of units who actually show up in the sample, there are no longer any guarantees that balance will obtain.
Our approach satisfies all four of these desiderata.
We propose an online method for covariate balancing requiring linear time and space which provably provides variance reduction.
Building upon recent work on discrepancy minimization---the self-balancing walk~\citep{alweiss2020discrepancy} and kernel thinning~\citep{dwivedi2021kernel}---we provide an algorithm whose $L_2$-imbalance matches the best known \emph{online} discrepancy algorithm.
Where $\delta$ is a failure probability, performing this optimization online results in a $\log(n/\delta)$ cost in convergence of the average treatment effect over the offline algorithm of discrepancy minimization by~\citet{harshaw2020balancing}.
A conjectured lower bound for online discrepancy minimization would further reduce this cost by a square root.
We also provide an extension to multiple treatments and non-uniform treatment probabilities.
The rest of the paper is organized as follows.
Section~\ref{sec:related} with a discussion of related work to position our contribution in the literature.
Section~\ref{sec:problem} defines notation and formally introduces the problem.
Section~\ref{sec:method} provides our proposed algorithms and methods.
Section~\ref{sec:simulations} provides a detailed simulation study of the behavior of the proposed algorithms.
\section{Related Work}
\label{sec:related}
There are two common approaches for achieving improved covariate balance in experiments.
The first, and most common especially within industrial settings, approach is to perform a post-hoc regression adjustment which includes pre-treatment covariates~\citep{deng2013improving, lin2013agnostic}.
The second approach is to consider covariate balance during the design phase of the experiment, i.e., explicitly optimizing treatment assignment in order to minimize imbalance between treatment groups~\citep{greevy2004optimal, higgins2016improving, kallus2018optimal}.
Post-hoc stratification may be seen as asymptotically equivalent in terms of variance reduction to its analogous pre-stratified design as shown by \citet{miratrix2013adjusting}.
\posscite{miratrix2013adjusting} analysis is limited by two factors: it assumes a fixed number of stratification cells (that do not grow with sample size) and it is conditioned on the post-stratification estimator being defined (e.g. treatment and control units within each stratification cell).
These limitations may weaken it's asymptotic equivalence argument.
A key limitation of post-hoc adjustment approaches is that the desire for simplicity and scalability implies that practitioners typically adjust for only a linear function of the pre-treatment covariates.
Indeed, in the common ``CUPED'' approach, adjustment is performed solely on a linear function of a single pre-treatment outcome measurement~\citep{deng2013improving}.
Second, many common approaches for constructing stratification cells (i.e. clustering algorithms) may be computationally infeasible in practice for industrial applications when the number of simultaneous experiments and the number of outcome variables of interest are large.
Third, without sample splitting (or when naively applied) advanced machine-learning based methods for adjustment may slip in assumptions of correct-specification of the outcome model, or have confidence intervals with poor coverage properties.
While cross-fitting may ameliorate some of these problems in larger samples, sample splitting may prove too high a cost when sample sizes are low.
A key shortcoming of design-based covariate balance is the lack of computationally efficient algorithms which provide theoretical guarantees over worst case behavior.
Blocking~\citep{fisher1935design} partitions variables into non-overlapping sets and performs complete randomization within each partition~("blocks").
\citet{higgins2016improving} introduced a computationally approximation of blocking which runs in $\mathcal{O}(n\log(n))$ time.
\citet{kallus2018optimal, bertsimas2015power} propose an optimization based approach, \citet{kallus2018optimal} additionally considers a partially random approach using semi-definite programming.
\citet{zhou2018sequential} provide a method combining \emph{batch-based} sequential experimentation with rerandomization to achieve balance, but which is not computationally feasible in moderate to large sample sizes.
Perhaps the closest to the current work is \citet{harshaw2020balancing} which propose a balancing design using the Gram-Schmidt walk, an offline method for (linear) discrepancy minimization.
Current state of the art for balancing treatment assignment requires polynomial running time and generally requires knowing all of the covariate vectors prior to determining assignment~\citep{higgins2016improving, harshaw2020balancing,arbour2021efficient}.
As we discuss in section \ref{sec:problem}, this is a non-starter for online treatment assignment.
In this setting, subjects must be allocated as they arrive; it does no good to know how you \emph{should have} assigned a user at the end of the experiment; you need to know when that subject arrives.
It's crucial that when a subject in an experiment arrives they be swiftly allocated to a unit.
Especially in an online environment, high latency will lead to attrition, which may counteract any potential gains from greater efficiency.
Moreover, the users who attrit may be the very subjects of interest~\citep{munger2021accessibility}.
By inducing differential attrition based on patience, the sample in the experiment may differ greatly from the population of interest on unobserved characteristics that make it difficult to extrapolate to a population-level effect~\citep{egami2020elements}.
There is also a variety of methods aimed at sequential, online assignment in experiments.
The seminal work in this literature is \citet{efron1971forcing} which introduced an online variant of complete randomization which aims to ensure that a pre-specified marginal treatment probability is met without introducing too much accidental bias.
\citet{smith1984sequential} provides a generalization of the \citet{efron1971forcing} approach which extends gracefully to multiple treatments.
There are a variety of online balanced coin designs which seek to reduce covariate imbalance~\citep[e.g.][]{baldi2011covariate, moore2017blocking}.
\citet{moore2017blocking} is based around Mahalanobis distance.
As such, it has polynomial time-complexity \emph{at each arrival time}.
In addition to inefficiency, the theoretical worst-case behavior of this algorithm has not been resolved, even in the stochastic setting.
Theoretical guarantees of this sort are paramount in the design setting, as practitioners need to know the credibility of their inferences (and how they may differ from simple Bernoulli randomization).
\section{Background and Problem Setting}
\label{sec:problem}
We first fix notation.
Random variables will be denoted in upper case, with sets in bold.
The problem setting, which we refer to as experimental treatment allocation, is as follows.
We assume that we observe $1, \dots, n$ i.i.d. observations of $\boldsymbol{X} \in \mathbb{R}^{n \times d}$: the covariates\footnote{We assume linear feature maps throughout. We note that nonlinearities can be handled with the same guarantees following \citet{dwivedi2021kernel} at the cost of additional computational complexity.}.
The experimenter is asked to assign a treatment assignment, $A \in \{1, -1\}$ (we will later loosen this to multiple discrete treatment values).
We will refer to the assignments of $A$ as treatment and control, respectively.
Each unit is imbued with \emph{potential outcomes} for each treatment, the value of the outcome if that unit had been assigned to the given group: $y(1)$ for treatment and $y(-1)$ for control.
After assignment we observe only the potential outcome corresponding to the realized treatment assignment, $Y$.
We assume that the outcomes are not available until the conclusion of the experiment.
At the end of the experiment we are interested in measuring the sample average treatment effect~(SATE) between any two treatments, $k$ and $k'$ with the difference in means estimator:
\begin{align}
\label{eq:SATE}
\hat{\tau}_{kk'} = \frac{1}{n}\sum_i^n \frac{A_i}{p(A_i)} Y_i
\end{align}
where $p(A_i)$ denotes the probability of assigning treatment $A_i$ to instance $i$.
Note that this is simply the difference-in-means rather than the more general Horwitz-Thompson estimator~\citep{horwitz1952generalization}, as the treatment probability is marginal rather than conditional.
More sophisticated estimators are usable in this setting~\citep[e.g.][]{tsiatis2008,aronow2013}, but we will focus on the simplest as we optimize design as is commonplace for studying design~\citep{kallus2017balanced, harshaw2020balancing}.
If propensity scores are constant, the estimator of the SATE given by equation~\ref{eq:SATE} will be unbiased and consistent for its oracle counterpart,
\begin{align}
\tau_{kk'} = \frac{1}{n}\sum_i^n y_i(k) - y_i(k'),
\end{align}
the difference of potential outcomes of the $k$ and $k'$ treatments.
This SATE is our estimand of interest, as estimated by equation~\ref{eq:SATE}.
We will maintain the following assumptions:
\begin{assumption}[Consistency] $Y_i = y_i(k)$ if $A_i = k \quad \forall i, k$.
\end{assumption}
The problem of experimental allocation is to observe covariate vectors and assign $A$ to units so as to achieve desirable properties of the SATE (for instance, to minimize variance).
In the most general setting where no assumptions are placed the relationship between the covariates and outcome complete randomization---randomly drawing assignments without respect to background covariates---is known to be minimax optimal~\citep{kallus2018optimal}.
\subsection{Robustness in sequential design}
Experiments are prized for their ability to provide unbiased estimates of causal effects with relatively mild assumptions.
These assumptions, in fact, typically flow from the \emph{design} of the experiment rather than more difficult assumptions about the data used in the course of analysis~\citep{sekhon2009opiates,aronow2021nonparametric}.
In the study of vector balancing, there are three main sampling schemes of interest, listed in order of how adversarial they are: \begin{enumerate*}
\item Stochastic arrivals.
\item Oblivious adversarial arrivals.
\item Fully adversarial arrivals.
\end{enumerate*}
In the stochastic arrivals setting, units are sampled i.i.d. from some fixed (possibly infinite) population.
As such, a given covariate vector is just as likely to arrive early in the sequence as late.
In the oblivious adversarial setting, the adversary knows the process which will be used to assign units to groups, but cannot condition specifically on those assignments in making its decisions about the order of arrival of units.
This implies the following conditional independence:
\begin{assumption}[Oblivious Adversary]
$$
\boldsymbol{w}_{i-1} \indep \boldsymbol{x}_i | \mathcal{H}_{i-1}
$$
where $\mathcal{H}_i$ is the history of $\boldsymbol{x}$ up to and including $\boldsymbol{x}_i$.
\end{assumption}
This stands in contrast to the fully adversarial case, in which the adversary is able to condition its decision in each period on the full set of past assignments in addition to the history of the covariate vector.
We will focus on the oblivious adversary in this paper.
\section{Online Assignment of Treatments}
\label{sec:method}
\textbf{Weighted Online Discrepancy Minimization:} We consider a variant of the online Koml\'os problem~\citep{spencer1977balancing}, where vectors $\boldsymbol{x}_1, \ldots \boldsymbol{x}_n$ arrive one by one and must be immediately assigned a weighted sign of either $-2q$ or $2(1-q)$, for $0 < q < 1$, such that the weighed discrepancy $\left\| \sum_{i=1}^n \eta_i \boldsymbol{x}_i \right\|_{\infty}$, where $\eta_i$ is the weighted sign given to $\boldsymbol{x}_i$, is minimized.
Notice that when $q = 1/2$, the signs become $\pm 1$.
Algorithm \ref{alg:weighted}, takes $i=1,\dots,n$ unit vectors in sequentially and assigns them to a treatment and control,
represented by the value of $\eta_i$.
The procedure, an extension of recent work in online discrepancy minimization~\citep{alweiss2020discrepancy, dwivedi2021kernel}, assigns treatment with probability proportional to the inner product between a running sum of the signed prior observations.
The algorithm and analysis differs from prior work for discrepancy in two aspects which are necessary for use in experimentation. One is to give a ridge regression guarantee, we need to characterize the random vectors output by the algorithm in terms of the projection matrix $\boldsymbol{P}.$ \citet{harshaw2020balancing} do that for an offline discrepancy minimization algorithm, and here we do it for an online version. The other way we differ is that our algorithm is a slight generalization of the one in \citet{dwivedi2021kernel} in that we have a parameter $q$. This allows for the case of imbalanced treatment assignments. A straightforward adoption of the analysis in \citet{dwivedi2021kernel} to this case results in a worse dependence on $1/q.$ Therefore, we derive a sub-exponential concentration bound and get a better dependence on $1/q.$
\begin{algorithm}
\label{alg:main}
\caption{takes each input vector ${\boldsymbol{x}}_i$ and assigns it $\{-2q, 2(1-q)\}$ signs online to maintain low weighted discrepancy with probability $1-\delta$.}
\begin{algorithmic}
\STATE {\bfseries Input: } {${\boldsymbol{x}}, q$}
$c \gets \min(1/q, 9.3) \log(2n/\delta)$ \\
\FOR {$i$ from 1 to $n$}
\IF {$|\boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i| > c$}
\STATE ${\boldsymbol{w}_{i}} \gets {\boldsymbol{w}_{i-1}} -
2 q \frac{\boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i}{c} \boldsymbol{x}_i $
\ELSE
\STATE $\eta_i \gets \begin{cases}
2(1-q), & \text{ w.p. } q(1 - \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c)\\
-2q,& \text{ w.p. }
1- q(1 - \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c)
\end{cases}
$
\STATE ${\boldsymbol{w}_{i}} \gets {\boldsymbol{w}_{i-1}} + \eta_i {\boldsymbol{x}_i}$
\ENDIF
\ENDFOR
\STATE {\bfseries Output: } {$\boldsymbol{\eta}, \boldsymbol{w}$}
\end{algorithmic}
\label{alg:weighted}
\end{algorithm}
{\bf Notation: }Let $\boldsymbol{P}_i, \, i\in [n]$ be orthogonal projection matrices onto the span of $\{\boldsymbol{x}_1,..,\boldsymbol{x}_i\},$ that is,
\[\boldsymbol{P}_i = \boldsymbol{X}_i^{\top} \br{\boldsymbol{X}_i \boldsymbol{X}_i^{\top}}^{-1} \boldsymbol{X}_i ,\]
where $\boldsymbol{X}_i$ is the $i \times d$ submatrix of $\boldsymbol{X}$ corresponding to covariates $\{\boldsymbol{x}_1,..,\boldsymbol{x}_i\}.$ Let $A = 0.5803, B = 0.4310$ and $ \alpha = 2/B $.
\begin{definition}
[Sub-Gaussian]
A mean zero random variable $X$ is sub-Gaussian with parameter $\sigma$ if for all $\lambda \in \mathbb{R}$,
$
\mathrm{E} [ \exp\br{ \lambda X } ] \leq \exp \br{ \frac{\lambda^2 \sigma^2}{2}}.
$
A mean zero random vector $\boldsymbol{w}$ is $(\sigma, \boldsymbol{P})$ sub-Gaussian if for all unit vectors $\boldsymbol{u}$ and $\lambda \in \mathbb{R}$,
$
\mathrm{E} [ \exp\br{ \lambda \boldsymbol{w}^{\top} \boldsymbol{u} } ] \leq \exp \br{ \frac{\lambda^2 \sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u} }{2}}.
$
In particular, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $\sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}$ sub-Gaussian.
\end{definition}
\begin{definition}
[Sub-exponential]
A mean zero random variable $X$ is $(\nu, \alpha)$ sub-exponential if for all $|\lambda| \leq \frac{1}{\alpha}$,
$
\mathrm{E} [ \exp\br{ \lambda X } ] \leq \exp \br{ \frac{\lambda^2 \nu^2}{2}}.
$
A mean zero random vector $\boldsymbol{w}$ is $(\nu, \alpha, \boldsymbol{P})$ sub-exponential if for all unit vectors $\boldsymbol{u}$ and $|\lambda| \leq \frac{1}{\alpha \sqrt{ \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}$,
$
\mathrm{E} [ \exp\br{ \lambda \boldsymbol{w}^{\top} \boldsymbol{u} } ] \leq \exp \br{ \frac{\lambda^2 \nu^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u} }{2}}.
$
In particular, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $(\nu \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}, \alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}})$ sub-exponential.
\end{definition}
\begin{restatable}[Main]{theorem}{main}
\label{thm:main}
Let $\boldsymbol{w}_1, ... \boldsymbol{w}_n$ be as in Algorithm~\ref{alg:main}. Then \begin{enumerate}
\item $\boldsymbol{w}_i$ is mean zero $\br{\sqrt{c/2q}, P_i}$ sub-Gaussian.
\item $\boldsymbol{w}_i$ is mean zero $\br{\sqrt{8Ac}, \alpha, P_i}$ sub-exponential.
\item With probability $1-\delta,$ for all $i,$ $|\boldsymbol{w}_i^T \boldsymbol{x}_i| \leq c .$
\end{enumerate}
\end{restatable}
Note that $\eta_i$ is defined only when $|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i| \leq c$. Therefore, $\boldsymbol{\eta}$ is defined with probability at least $1-\delta.$
\begin{remark}
If $\boldsymbol{w}_i$ is a is a mean zero $\br{\sigma, \boldsymbol{P}_i}$ sub-Gaussian random vector, then $\mathrm{Cov}(\boldsymbol{w}_i) \leq \sigma^2 \boldsymbol{P}_i.$ Similarly, we have that if $\boldsymbol{w}_i$ is a is a mean zero $\br{\nu, \alpha, P_i}$ sub-exponential random vector, then $\mathrm{Cov}(\boldsymbol{w}_i) \leq \frac{3}{2} \nu^2 \boldsymbol{P}_i$.
A proof is given in Lemma~\ref{lem:vector-var}.
\end{remark}
We will use Threorem~\ref{thm:main} to derive results on the average treatment effect using the framework developed by \cite{harshaw2020balancing}.
Let $z_i = \eta_i + 2q - 1 \in \{-1, 1\}.$ We will show that $\mathrm{E}[\eta_i] = 0,$ and so we have $\mathrm{E}[z_i] = 2q-1$ and $\boldsymbol{z} - \mathrm{E}[\boldsymbol{z}] = \boldsymbol{\eta}.$ Let $ \boldsymbol{\mu} = \frac{\boldsymbol{Y}(1)}{4 q} +
\frac{\boldsymbol{Y}(0)}{4(1-q)}.$
\citet{harshaw2020balancing} give a linear algberaic expression for the error of HT-estimators in terms of $\boldsymbol{\mu}$. In particular, they show
\begin{lemma}[Lemma A2 and Corollary A1 in \cite{harshaw2020balancing}]
\label{lem:ate-err}
\[
\hat{\tau} - \tau = \frac{2}{n} \br{\boldsymbol{z} - \mathrm{E}[\boldsymbol{z}]}^{\top} \boldsymbol{\mu} = \frac{2}{n} \boldsymbol{\eta}^{\top}\boldsymbol{\mu}
\]
and hence,
\[ \mathrm{Var}(\hat{\tau}) = \mathrm{E} \brs{ (\hat{\tau} - \tau)^2 } = \frac{4}{n^2} \boldsymbol{\mu}^{\top} \mathrm{Cov}(\boldsymbol{z}) \boldsymbol{\mu}.
\]
\end{lemma}
Theorem~\ref{thm:main} immediately implies that assignments generated by Algorithm~\ref{alg:main} are well balanced. But optimizing just for balance can lead to accidental bias~\citep{efron1971forcing}. We have from Lemma~\ref{lem:ate-err} that $\text{Var}(\hat{\tau}) = \frac{4}{n^2} \lambda_{\max}\br{\mathrm{Cov}(\boldsymbol{z})} \|\boldsymbol{\mu}\|^2 $ in the worst case when $\boldsymbol{\mu}$ is along the top eigenvector of $\mathrm{Cov}(\boldsymbol{z})$. %
Therefore, to control accidental bias we need to make sure $\lambda_{\max}\br{\mathrm{Cov}(\boldsymbol{z})}$ is not high.
We achieve this by augmenting the original covariates $\boldsymbol{x}_i$ by $\sqrt{\phi} \boldsymbol{e}_i$ to get
$\left[ \begin{array}{c}
\sqrt{\phi} \boldsymbol{e}_i \\
\sqrt{1-\phi} \boldsymbol{x}_i
\end{array}
\right]
,$
where $\phi \in [0, 1]$ is a parameter which controls the extent of the covariate balance, and $\boldsymbol{e}_i$ is a basis vector in dimension $n$.
By a simple calculation, we can see that running Algorithm~\ref{alg:main} on $\boldsymbol{x}_i$ augmented with $\sqrt{\phi} \boldsymbol{e}_i$ is equivalent to running it with $\boldsymbol{x}_i$ and replacing $\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i$ by $\sqrt{1- \phi} \boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i$ everywhere. Therefore, we don't have to explicitly augment the covariates in the algorithm.
We note that with augmented covariates, $\boldsymbol{w}_n = \left[\begin{array}{c}
\sqrt{\phi} \boldsymbol{\eta} \\
\sqrt{1-\phi} \boldsymbol{X}^{\top} \boldsymbol{\eta}
\end{array}\right]$ is a sub-Gaussian or a sub-exponential random vector as in Theorem~\ref{thm:main}.
{\small Let $\phi \in (0,1)$ be fixed, and $\boldsymbol{Q} = \br{\phi \boldsymbol{I} + (1-\phi)\boldsymbol{X} \boldsymbol{X}^{\top}}^{-1}.$}
\begin{proposition}
\label{prop:z-dist}
When Algorithm~\ref{alg:main} is run with augmented covariates as described above, then
$\boldsymbol{\eta} = \boldsymbol{z} - \mathrm{E}[\boldsymbol{z}]$ is a mean zero $(\sqrt{c/2q}, \boldsymbol{Q})$ sub-Gaussian random vector and also, $\boldsymbol{\eta}$ is a mean zero $(\sqrt{8Ac},\alpha, \boldsymbol{Q})$ sub-exponential random vector.
\end{proposition}
Proposition~\ref{prop:z-dist} shows that $\mathrm{E}[\boldsymbol{\eta}] = 0$ for all $i.$
\begin{proposition}[Unbiasedness]
When Algorithm~\ref{alg:main} is run with augmented covariates, we have with probability at least $1-\delta$, $\mathrm{E} \brs{\sum_i \boldsymbol{x}_i \eta_i} = 0.$
\end{proposition}
For rest of the section, we let $\sigma^2 = c/2q$ if $c = \log(2n/\delta)/q$ and $\sigma^2 = 12Ac$ if $c = 9.3 \log(2n/\delta).$
\begin{proposition}[Eigenvalues of Treatment Covariance]
\label{prop:spectral-bound}
With probability at least $1-\delta$, Algorithm~\ref{alg:main} produces $\boldsymbol{\eta}$ satisfying
$
\mathrm{Cov}(\boldsymbol{z}) = \mathrm{Cov}(\boldsymbol{\eta}) \preceq \sigma^2 \boldsymbol{Q}.
$
\end{proposition}
\subsection{Balance}
\label{sec:balance}
\begin{proposition}
\label{prop:balance}
Let $\boldsymbol{w} = \sum_i \eta_i \boldsymbol{x}_i$. With probability at least $1-\delta$,
{
\small
\[
\left \| \boldsymbol{w} \right \|_{2} \leq \sqrt{d} \left \| \boldsymbol{w} \right \|_{\infty} \leq
\min \left( \frac{1}{q}, 9.3 \right) \sqrt{\frac{ {d \log (4d/\delta)\log(4n/\delta)} }{2 (1-\phi) \phi}}.
\]
}
\end{proposition}
\subsection{Error bounds}
\begin{proposition}[Concentration of ATE]
\label{prop:ate-concentration}
Algorithm~\ref{alg:main} when run with augmented covariates, generates a random assignment $\boldsymbol{z}$ such that
\[
|\hat{\tau} - \tau| = \frac{2}{n} | \boldsymbol{\eta}^{\top} \boldsymbol{\mu}| \leq \frac{2c}{n}
\sqrt{\boldsymbol{\mu}^{\top} \boldsymbol{Q} \boldsymbol{\mu}}.
\]
with probability $1-\delta.$
\end{proposition}
Lemma A10 in~\cite{harshaw2020balancing} shows that $\boldsymbol{\mu}^{\top} \boldsymbol{Q} \boldsymbol{\mu} = \min _{\boldsymbol{\beta} \in \mathbb{R}^{d}}\left[\frac{1}{\phi }\|\boldsymbol{\mu}-\boldsymbol{X} \boldsymbol{\beta}\|^{2}+\frac{\|\boldsymbol{\beta}\|^{2}}{(1-\phi) } \right].$
\begin{proposition}
\label{prop:ate-worstcase}
The worst-case mean squared error of the online balancing walk design is upper bounded by %
\begin{align*}
&\mathrm{E}\left[(\widehat{\tau}-\tau)^{2}\right] \leq \frac{4 \sigma^2}{\phi n^2} \sum_{i=1}^{n} \mu_{i}^{2}\\ %
\end{align*}
where $\phi \in(0,1]$ with probability $1 - \delta$.
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{prop:ate-worstcase}]
This follows from Lemma~\ref{lem:ate-err} and Proposition~\ref{prop:spectral-bound}. We note that $\boldsymbol{Q} \preceq \frac{\sigma^2}{\phi} \boldsymbol{I}.$
\end{proof}
\begin{proposition}[Ridge Connection]
\label{prop:ate-ridge}
The worst-case mean squared error of the online balancing walk design is upper bounded by an implicit ridge regression estimator with regularization proportional to $\phi$. That is, %
\begin{align*}
&\mathrm{E} \left[(\widehat{\tau}-\tau)^{2}\right] \leq \frac{4 \sigma^2 L}{n} \\
&\text { where } \quad L=\min _{\boldsymbol{\beta} \in \mathbb{R}^{d}}\left[\frac{1}{\phi n}\|\boldsymbol{\mu}-\boldsymbol{X} \boldsymbol{\beta}\|^{2}+\frac{\|\boldsymbol{\beta}\|^{2}}{(1-\phi) n}\right]
\end{align*}
with probability $1 - \delta$.
\end{proposition}
\subsection{Algorithm with Restart}
\label{sec:restart}
We saw earlier that $\boldsymbol{\eta}, \boldsymbol{z}$ are defined only with probability $1-\delta.$ This is because with our choice of $c$, only with probability $1-\delta$ we have for all $i, |\boldsymbol{w}_i^{\top} \boldsymbol{x}_i| \leq c.$ This means our treatment assignment fails with probability $\delta.$ There is a simple way to make sure that the algorithm never fails and have same error bound guarantees with a slightly worse constant.
This is achieved by slightly modifying Algorithm~\ref{alg:main} so that whenever we have $|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i| > c$ for a particular $i,$ we start a new instance of the algorithm for covariates $\boldsymbol{x}_{i+1}, ...\boldsymbol{x}_n.$ This is equivalent to setting $\boldsymbol{w}_{i-1} = 0$ and continuing with the algorithm.
Since for any treat assignment procedure $\mathrm{E} \br{\hat{\tau}-\tau}^2$ just depends on $\mathrm{Cov}(\boldsymbol{z})$, and Algorithm~\ref{alg:main} fails with probability $\geq \delta,$ we can show that
\begin{align*}
\mathrm{Cov}(\boldsymbol{z}) &\preceq (1-\delta)\boldsymbol{Q} + \delta \boldsymbol{Q} + \delta^2 \boldsymbol{Q} + ...\\
&\preceq 2 \boldsymbol{Q} \text{ when } \delta \leq 1/2.
\end{align*}
Therefore, for the modified algorithm, we will have error guarantees as in Propositions~\ref{prop:ate-worstcase} and \ref{prop:ate-ridge} (but worse by at most a factor of $2$) and with probability one.
\subsection{Multiple Treatments}
In this section, we consider an online multi-treatment setting, where each vector is assigned to a group in $M = \{m_1, m_2, \ldots, m_k\}$ immediately on arrival. For each $1 \leq i\leq k$, group $m_i$ is associated with a weight $\alpha_i$.
The goal is to minimize the multi-treatment discrepancy:
\begin{align*}
&\max_{m_1, m_2 \in M} 2 \left\| \frac{\boldsymbol{s}(m_1)/{\alpha_1} - \boldsymbol{s}(m_2)/{\alpha_2}}{ {1}/{\alpha_1} + {1}/{\alpha_2}} \right\|_{\infty}
\end{align*}
where $\boldsymbol{s}(m)$ is the sum of all vectors assigned to treatment $m$.
Notice that by setting $\alpha_1 = \frac{1}{1-q}$ and $\alpha_2 = \frac{1}{q}$,
we can recover the definition given for the weighted discrepancy between two treatments $m_1$ and $m_2$.
Our algorithm can leverage any oracle (we call it \texttt{BinaryBalance} in Algorithm~\ref{alg:multi}) that minimizes the weighted discrepancy for two treatments. Our results are obtained by using Algorithm~\ref{alg:weighted}.
We first build a binary tree where each leaf of the tree corresponds to one of the $k$ treatments in $M$. Let $h$ be the smallest integer such that $2^{h} \geq k$. We start with a complete binary tree of height $h$, and then remove $2^h - k$ leaves from the tree such that no two siblings are removed. Note that this is possible by the definition of $h$. We further contract each internal node with only one child to its child. This process does not change the number of leaves in the tree. Let $T$ be the obtained tree. By construction, all internal nodes of $T$ have 2 children and $T$ has exactly $k$ leaves. We then associate each leaf of $T$ with a treatment in $M$. For each vector assigned to treatment $m_i$, we also say that it is assigned to the leaf corresponding to $m_i$,
$\forall 1 \leq i \leq k$.
For each node $v \in T$, denote by $\boldsymbol{s}(v)$ the sum of all vectors assigned to leaves under $v$.
In addition, let $\alpha(v)$ be the sum of all weights assigned to leaves under $v$.
For each internal node $v$ of $T$, the weighted discrepancy vector at $v$ is defined as:
\begin{align*}
\boldsymbol{w}(v) &= \frac{\alpha(v_r)}{\alpha(v_l) + \alpha(v_r)} \boldsymbol{s}(v_l) - \frac{\alpha(v_l)}{\alpha(v_l) + \alpha(v_r)} \boldsymbol{s}(v_r) \\
&= \frac{{\boldsymbol{s}(v_l)}/{\alpha(v_l)} - {\boldsymbol{s}(v_r)}/{\alpha(v_r)}}{{ 1}/{\alpha(v_l)} + { 1}/{\alpha(v_r)}},
\end{align*}
where $v_l$ and $v_r$ are the left and right child of $v$ respectively.
For each internal node $v$ in $T$, we maintain an independent run of a two-treatment algorithm that minimizes $\|\boldsymbol{w}(v)\|_{\infty}$. At a high level, we minimize the weighted discrepancies at all internal nodes simultaneously. When a new vector $\boldsymbol{x}$ arrives, we first feed it to the algorithm at root $r$. If the result is $+$, we continue with the left sub-tree of $r$. Otherwise, we go to the right sub-tree. We continue in that manner until we reach a leaf $l$. $\boldsymbol{x}$ will then be assigned to $l$ (and the treatment associated with $l$).
\begin{restatable}{theorem}{multi}
\label{thm:multi}
Let \texttt{BinaryBalance} be Algorithm~\ref{alg:weighted}. Then Algorithm~\ref{alg:multi} obtains
$O\left(\log k \sqrt{\frac{ {(1-\phi)d \log (dk/\delta)\log(nk/\delta)} }{\phi}} \right)$
multi-treatment discrepancy with probability $1-\delta$.
\end{restatable}
\begin{algorithm}[h]
\caption{\texttt{KGroupBalance} takes each input vector ${x}_i$ and assigns it to one of the groups online to maintain low discrepancy with probability $1-\delta$.}
\begin{algorithmic}
\STATE {\bfseries Input:} ${\boldsymbol{x}}, k, \alpha$.
\STATE $h$ $\gets$ smallest integer such that $2^h \geq k$.
\STATE $T \gets$ complete binary tree with height $h$. Remove $2^h - k$ leaves from $T$ such that no two siblings are removed. Associate each treatment to a leaf of $T$. Contract each internal node in $T$ with one child to its child.
\FOR{node $v$ in $T$}
\STATE $\alpha(v) \gets$ sum of all weights of groups associating with leaves under $v$.
\ENDFOR
\FOR{internal node $v$ of $T$}
\STATE Instantiate \texttt{BinaryBalance}($v$) $\gets$ oracle for weighted discrepancy problem at $v$ with weighted signs $\frac{\alpha(v_r)}{\alpha(v_l) + \alpha(v_r)}$
and $-\frac{\alpha(v_l)}{\alpha(v_l) + \alpha(v_r)}$.
\ENDFOR
\FOR{i from 1 to n}
\STATE $v \gets $ root of $T$.
\FOR{$v$ is an internal node of $T$}
\STATE Feed $\boldsymbol{x}_i$ to \texttt{BinaryBalance}($v$)
\STATE $v \gets $ one of the children of $v$ according to the assignment of \texttt{BinaryBalance}($v$) on input $\boldsymbol{x}_i$
\ENDFOR
\STATE Assign $x_i$ to the group corresponding to $v$.
\ENDFOR
\end{algorithmic}
\label{alg:multi}
\end{algorithm}
\begin{figure}
\centering
\includegraphics[width=0.375\textwidth]{figures/time-1.pdf}
\caption{Time to design. All timings performed on a \texttt{ml.r5.2xlarge} instance of Amazon SageMaker. The y-axis is scaled by the square root for easier visualization.}
\label{fig:time}
\end{figure}
\section{Experiments}
\label{sec:simulations}
In this section, we provide simulation evidence on the efficacy of our proposed methods.
In particular, we use a wide variety of data generating processes, many of which do not assume that units arrive i.i.d. as is often standard in simulation settings for this problem.
Subjects are unlikely to arrive truly i.i.d. in the real world.
Earlier arrivals will typically be more active than late-arriving units, for example.
All data generating processes used in simulations are shown in Table~\ref{tab:dgps}.
If not otherwise specified, the sample size is $1000$ subjects, the number of groups is two and the marginal probability of treatment is $\frac{1}{2}$.
Methods compared in the simulations are simple randomization (Bernoulli coin flips), complete randomization (fixed-margins randomization), the biased coin designs of \citet{efron1971forcing} and of \citet{smith1984sequential}, QuickBlock of \citet{higgins2016improving}, and \citet{alweiss2020discrepancy}.
These are compared to our proposed methods which are generalized versions of the discrepancy minimization procedure of \citet{dwivedi2021kernel}.
We provide three versions of our proposed algorithms, called BWD for "Balancing Walk Design".
The most basic version (BWDRandom) reverts to simple random assignment for all remaining periods once $|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i| > c$.
Our preferred approach restarts the algorithm in this case as described in section~\ref{sec:restart}.
We examine this with two levels of robustness, $\phi$: 0 (purely balancing) and 0.5 (a uniform mix between randomization and balancing): BWD(0) and BWD(0.5) respectively. For further details, see section~\ref{sec:robust}.
In these comparisons, BWD need not out-perform all methods in all data-generating processes.
QuickBlock, for instance, is a fully off-line method, so comparable performance by an online method is noteworthy.
In general, Alweiss and BWD will be most effective when the true relationship between covariates and outcome is linear, since they seek linear balance.
All plots incorporate 95\% confidence intervals.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/mse-1.pdf}
\caption{MSE. BWD provides effective variance reduction across a wide array of simulation environments.}
\label{fig:mse}
\end{figure}
\subsection{Binary treatment}
\paragraph{Timing.} Our proposed method (BWD) is highly efficient, scaling substantially better than other balancing methods.
This analysis of runtime directly compares online methods to a widely used offline balancing method (QuickBlock).
It's important to note that the QuickBlock algorithm cannot be used in the online setting, even if its runtime did not make that prohibitive.
While QuickBlock is $\mathcal{O}(n \log n)$, the proposed online balancing methods are all linear-time.
Given QuickBlock's runtime, the following simulations only include it for comparisons up to sample sizes of $10^4$.
\paragraph{MSE.}
Next, we demonstrate in Figure~\ref{fig:mse} how imbalance minimization translates to improved estimation of causal effects, measured by the mean squared-error of our estimate of the average treatment effect.
We normalize this graph based on $n$, the rate of convergence of the difference-in-means estimator under simple randomization.
The results depend strongly on the true nature of the data-generating process.
In short, on non-linear data generating processes, offline blocking performs better than anything else, but in many settings BWD converges to similar error rates as QuickBlock.
On linear or near-linear data-generating processes, our proposed algorithms perform very strongly, outperforming QuickBlock even in small sample-sizes.
When there is a break from the purely i.i.d. stochastic setting (such as \texttt{LinearDriftDGP} and \texttt{LinearSeasonDGP}), BWD behaves well, as expected.
\subsection{Multiple Treatments}
\paragraph{MISE.}
BWD gracefully extends to the multiple-treatment setting, which we demonstrate in Figure~\ref{fig:mse-multi}.
This chart measures the mean integrated squared-error of our estimates of the ATEs (relative to a single control group).
Figure~\ref{fig:mse-multi} shows the results.
BWD consistently outperforms existing online assignment methods by substantial margins.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/multi_mse-1.pdf}
\caption{MISE. BWD effective reduces variance no matter the number of treatments.}
\label{fig:mse-multi}
\end{figure}
An array of additional simulation results may be found in Appendix~\ref{app:sims}.
\section{Conclusion}
Experiments are a crucial part of how humans learn about the world and make decisions.
This paper is aimed at providing a way to more effectively run experiments in the online setting.
Practitioners must commonly operate their experiments in this environment, but due to the lack of suitable options for design fall back to simple randomization as the assignment mechanism.
In this paper, we have shown how the Balancing Walk Design can be an effective tool in this setting.
It is efficient, effective at reducing imbalance (and, therefore, the variance of resulting causal estimates), robust and it is fully suited to the particularities of online treatment assignment.
Simulations have shown it to work well across a range of diverse settings.
The Balancing Walk Design can improve the practice of large-scale online experimentation.
\begin{appendix}
\setcounter{figure}{0}
\setcounter{table}{0}
\renewcommand\thefigure{A\arabic{figure}}
\renewcommand\thetable{A\arabic{table}}
\section{Proofs}
\subsection{Proof of Theorem~\ref{thm:main}}
\begin{restatable}[Loewner Order]{proposition}{loewner}
\label{prop:loewner}
Let $\boldsymbol{M} \succeq 0, C \geq 1, L \geq 0 .$ If $$\boldsymbol{M} \preceq CL\boldsymbol{P} \coloneqq CL\boldsymbol{B}^\top\left(\boldsymbol{B}^\top\boldsymbol{B}\right)^{-1}\boldsymbol{B}$$ then for any vector $\boldsymbol{v} \in \mathbb{R}^{n}$ with $\|\boldsymbol{v}\|_{2} \leq 1$
$$
\boldsymbol{M}^{\prime}=\left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right) \boldsymbol{M}\left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right)+L \boldsymbol{v} \boldsymbol{v}^{\top}
$$
satisfies $$0 \preceq \boldsymbol{M}^{\prime} \preceq CL\boldsymbol{P}' \coloneqq CL\boldsymbol{P} + CL\frac{\left(\boldsymbol{I} - \boldsymbol{P}\right)\boldsymbol{v} \boldsymbol{v}^\top\left(\boldsymbol{I} - \boldsymbol{P}\right)}{\boldsymbol{v}^\top\left(\boldsymbol{I} - \boldsymbol{P}\right)\boldsymbol{v}}.$$
\end{restatable}
\begin{proof}
By definition of $\boldsymbol{M}'$ and the assumption $\boldsymbol{M} \preceq cL\boldsymbol{P},$ we have
\[
\boldsymbol{M}^{\prime} \preceq CL \left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right) \boldsymbol{P}\left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right)+L \boldsymbol{v} \boldsymbol{v}^{\top}.
\]
Therefore, it is sufficient to prove
\[
\left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right) \boldsymbol{P}\left(\boldsymbol{I}-C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top}\right)+ C^{-1} \boldsymbol{v} \boldsymbol{v}^{\top} \preceq
\boldsymbol{P} + \frac{\left(\boldsymbol{I} - \boldsymbol{P}\right)\boldsymbol{v}\vv^\top\left(\boldsymbol{I} - \boldsymbol{P}\right)}{\boldsymbol{v}^\top\left(\boldsymbol{I} - \boldsymbol{P}\right)\boldsymbol{v}}.
\]
Also note that, since $\|\boldsymbol{v}\| \leq 1$ and $C \geq 1,$ we can absorb $C$ into $\boldsymbol{v}$ (by taking $\boldsymbol{v} \coloneqq \boldsymbol{v}/\sqrt{C}$) and therefore without loss of generality assume that $C=1$ and $\|\boldsymbol{v}\| \leq 1.$
Define:
\begin{align*}
&\boldsymbol{P}_x = \boldsymbol{B}^\top\left(\boldsymbol{B}^\top\boldsymbol{B}\right)^{-1}\boldsymbol{B}\quad
&\boldsymbol{Q}_x \coloneqq \left(\boldsymbol{I} - \boldsymbol{P}\right)\quad
&\boldsymbol{P}_v = \boldsymbol{v}\vv^\top\quad
&\boldsymbol{Q}_v = \left(\boldsymbol{I} - \boldsymbol{v}\vv^\top\right)\\
&\boldsymbol{a} \coloneqq \boldsymbol{P}_x \boldsymbol{v}\quad
&\boldsymbol{b} \coloneqq \boldsymbol{Q}_x \boldsymbol{v}\quad
&\alpha \coloneqq \|\boldsymbol{a}\|^2\quad
&\beta \coloneqq \|\boldsymbol{b}\|^2
\end{align*}
We want to show
\begin{align*}
\boldsymbol{Q}_v\boldsymbol{P}_x\boldsymbol{Q}_v + \boldsymbol{P}_v \preccurlyeq
\boldsymbol{P}_x + \frac{\boldsymbol{Q}_x\boldsymbol{P}_v\boldsymbol{Q}_x}{\beta}
\end{align*}
Beginning by rewriting the LHS
\begin{align*}
&\boldsymbol{Q}_v\boldsymbol{P}_x\boldsymbol{Q}_v + \boldsymbol{P}_v = (\boldsymbol{I} - \boldsymbol{P}_v)\boldsymbol{P}_x(\boldsymbol{I} - \boldsymbol{P}_v) + \boldsymbol{P}_v\\
&=\boldsymbol{P}_x - \boldsymbol{P}_v\boldsymbol{P}_x - \boldsymbol{P}_x\boldsymbol{P}_v + \boldsymbol{P}_v\boldsymbol{P}_x\boldsymbol{P}_v + \boldsymbol{P}_v\\
&=\boldsymbol{P}_x - \boldsymbol{v}\boldsymbol{a}^\top - \boldsymbol{a}\boldsymbol{v}^\top + \boldsymbol{v}\vv^\top\alpha + \boldsymbol{P}_v\\
&=\boldsymbol{P}_x - (\boldsymbol{a} + \boldsymbol{b})\boldsymbol{a}^\top - \boldsymbol{a}(\boldsymbol{a} + \boldsymbol{b})^\top + (\boldsymbol{a} + \boldsymbol{b})(\boldsymbol{a} + \boldsymbol{b})^\top\alpha + \boldsymbol{P}_v
\end{align*}
Expanding $\boldsymbol{P}_v$ as $\boldsymbol{P}_v = \left(\boldsymbol{a} + \boldsymbol{b}\right)\left(\boldsymbol{b} + \boldsymbol{a}\right)^\top = \boldsymbol{a}\aaa^\top + \boldsymbol{b}\bb^\top + \boldsymbol{a}\boldsymbol{b}^\top + \boldsymbol{b}\boldsymbol{a}^\top$ gives
\begin{align*}
&\boldsymbol{P}_x - \left(\boldsymbol{a} + \boldsymbol{b}\right)\boldsymbol{a}^\top - \boldsymbol{a}\left(\boldsymbol{a} + \boldsymbol{b}\right)^\top + (1 + \alpha)\left(\boldsymbol{a}\aaa^\top + \boldsymbol{b}\bb^\top + \boldsymbol{a}\boldsymbol{b}^\top + \boldsymbol{b}\boldsymbol{a}^\top\right)\\
= &\boldsymbol{P}_x + (\alpha - 1)\boldsymbol{a}\aaa^\top + \alpha\left(\boldsymbol{a}\boldsymbol{b}^\top + \boldsymbol{b}\boldsymbol{a}^\top\right) + (1 + \alpha)\boldsymbol{b}\bb^\top
\end{align*}
Since $\| \boldsymbol{v}\|^2 \leq 1$, we have $\alpha + \beta \leq 1.$ Now considering the difference of the LHS from the RHS after multiplying both sides by $\beta$ we arrive at
\begin{align*} \beta (\text{RHS} - \text{LHS}) =
&\boldsymbol{b}\bb^\top(\underbrace{1 - \beta(1 + \alpha)}_{1 - \beta - \beta\alpha \geq \alpha(1 - \beta) \geq \alpha^2}) + \beta(1 - \alpha)\boldsymbol{a}\aaa^\top - \alpha\beta(\boldsymbol{a}\boldsymbol{b}^\top + \boldsymbol{b}\boldsymbol{a}^\top)\\
\succcurlyeq &\alpha^2\boldsymbol{b}\bb^\top + \beta^2\boldsymbol{a}\aaa^\top - \alpha\beta(\boldsymbol{a}\boldsymbol{b}^\top + \boldsymbol{b}\boldsymbol{a}^\top)\\
= &\left(\alpha \boldsymbol{b} - \beta \boldsymbol{a}\right)\left(\alpha \boldsymbol{b} - \beta \boldsymbol{a}\right)^\top \succcurlyeq 0.
\end{align*}
\end{proof}
\begin{lemma}
When $|\boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i| < c$, we have
$$ \mathrm{E} \left[ \eta_i \right] = - 2 q \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c .$$
\end{lemma}
\begin{proof}
\begin{align*}
\mathrm{E} \left[ \eta_i \right] &= 2(1-q) \cdot (q(1 - \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c)) +
(-2q) \cdot (1- q(1 - \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c)) \\
&= 0 + 2(1-q)(- q \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c) +
(-2q)(q \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c) \\
&= - 2 q \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c.
\end{align*}
\end{proof}
For all $i$ and $\boldsymbol{u} \in \mathbb{R}^d$, we have
\begin{align*}
\langle \boldsymbol{w}_i, \boldsymbol{u} \rangle &= \left \langle \boldsymbol{w}_{i-1},
\boldsymbol{u} - 2 q \frac{\boldsymbol{x}_i^{\top} \boldsymbol{u}}{c} \boldsymbol{x}_i \right \rangle + \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}\\
& = \langle \boldsymbol{w}_i, \left( \boldsymbol{I} - \frac{2q}{c} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \right) \boldsymbol{u} \rangle + \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u} \\
& = \langle \boldsymbol{w}_i, \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}
\rangle + \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u},
\end{align*}
where
$\epsilon_i =0$ if $|\boldsymbol{w}_{i-1}^{\top}\boldsymbol{x}_i| > c$ and $\epsilon_i = \eta_i - E [\eta_i]$ otherwise and $\boldsymbol{Q}_{\boldsymbol{x}_i, c} = \left( \boldsymbol{I} - \frac{2q}{c} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \right).$
Consider the case when $|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_i| \leq c$ and $\epsilon_i = \eta_i - E [\eta_i]$. Let $\tilde{q} = q(1 - \boldsymbol{w}_{i-1}^{\top} {\boldsymbol{x}}_i / c).$ Note that $0 \leq \tilde{q} \leq 2q.$ We have
$$\epsilon_i \gets \begin{cases}
2(1 - \tilde{q}) & \text{ with probability } \tilde{q},\\
-2\tilde{q}& \text{ with probability }
1- \tilde{q}.
\end{cases}
$$
\begin{definition}
[Sub-Gaussian]
A mean zero random variable $X$ is sub-Gaussian with parameter $\sigma$ if for all $\lambda \in \mathbb{R}$,
\[
\mathrm{E} [ \exp\br{ \lambda X } ] \leq \exp \br{ \frac{\lambda^2 \sigma^2}{2}}.
\]
A mean zero random vector $\boldsymbol{w}$ is $(\sigma, \boldsymbol{P})$ sub-Gaussian if for all unit vectors $\boldsymbol{u}$ and $\lambda \in \mathbb{R}$,
\[
\mathrm{E} [ \exp\br{ \lambda \boldsymbol{w}^{\top} \boldsymbol{u} } ] \leq \exp \br{ \frac{\lambda^2 \sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u} }{2}}.
\]
In particular, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $\sigma'$ sub-Gaussian, where $\sigma'^2 = \sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}$.
\end{definition}
\begin{definition}
[Sub-exponential]
A mean zero random variable $X$ is $(\nu, \alpha)$ sub-exponential if for all $|\lambda| \leq \frac{1}{\alpha}$,
\[
\mathrm{E} [ \exp\br{ \lambda X } ] \leq \exp \br{ \frac{\lambda^2 \nu^2}{2}}.
\]
A mean zero random vector $\boldsymbol{w}$ is $(\nu, \alpha, \boldsymbol{P})$ sub-exponential if for all unit vectors $\boldsymbol{u}$ and $|\lambda| \leq \frac{1}{\alpha \sqrt{ \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}$,
\[
\mathrm{E} [ \exp\br{ \lambda \boldsymbol{w}^{\top} \boldsymbol{u} } ] \leq \exp \br{ \frac{\lambda^2 \nu^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u} }{2}}.
\]
In particular, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $(\nu', \alpha')$ sub-exponential, where $\nu'^2 = \nu^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}$ and $\alpha' = \alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}$.
\end{definition}
The following concentration bounds for sub-Gaussian and sub-exponential vectors are obtained from standard bounds for scalar sub-Gaussian and sub-exponential random variables \cite{wainwright_2019} by scaling $\sigma$, $\nu$ and $\alpha$ by appropriate factors.
\begin{lemma}[Sub-Gaussian Concentration]
\label{lem:subgaussian-bound}
If a random vector $\boldsymbol{w}$ is $(\sigma, \boldsymbol{P})$ sub-Gaussian, then for all unit vectors $\boldsymbol{u}$,
$$
P\br{|\boldsymbol{w}^{\top}\boldsymbol{u}| \geq t} \leq
2 \exp \br{-\frac{t^2}{2\sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}.
$$
\end{lemma}
\begin{lemma}[Sub-exponential Concentration]
\label{lem:subexp-bound}
If a random vector $\boldsymbol{w}$ is $(\nu, \alpha, \boldsymbol{P})$ sub-exponential, then for all unit vectors $\boldsymbol{u}$,
$$
P\br{|\boldsymbol{w}^{\top}\boldsymbol{u}| \geq t} \leq
\begin{cases}
2\exp \br{-\frac{t^2}{2 \nu^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}} & \text{ if } 0 \leq t \leq \frac{\nu^2 \sqrt{ \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}{\alpha} \\
2\exp \br{- \frac{t}{2 \alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}} & \text{ if } t > \frac{\nu^2 \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}}{\alpha}.
\end{cases}
$$
\end{lemma}
\begin{lemma}
\label{lem:subexp-var}
Suppose $X$ is a $(\nu, \alpha)$ sub-exponential random variable with $\frac{2}{\nu^2} \leq \frac{1}{\alpha^2}$. Then
\[
\mathrm{Var}(X) \leq \frac{3}{2}\nu^2.
\]
\end{lemma}
\begin{proof}
We will use the inequality $x^2 \leq C \br{e^x + e^{-x}}, \, \forall x$ and $C = 1.5/e.$ This gives
\begin{align*}
\text{Var}(\lambda X) & \leq C \mathrm{E} \br{e^{\lambda X} + e^{- \lambda X}} \\
& \leq 2C\exp(0.5 \lambda^2 \nu^2)
, \text{ if } |\lambda| \leq \frac{1}{\alpha}.
\end{align*}
We therefore have
\begin{align*}
\text{Var}(X) &\leq C \nu^2 \frac{\exp(0.5 \lambda^2 \nu^2)}{0.5 \lambda^2 \nu^2} \\
&= eC \nu^2 \text{ when } 0.5 \lambda^2 \nu^2 = 1.
\end{align*}
We get the result with $C = 1.5/e.$
\end{proof}
\begin{lemma}
\label{lem:vector-var}
If $\boldsymbol{w}$ is a is a mean zero $\br{\sigma, \boldsymbol{P}}$ sub-Gaussian random vector, then $\mathrm{Cov}(\boldsymbol{w}) \preceq \sigma^2 \boldsymbol{P}$. If $\boldsymbol{w}$ is a is a mean zero $\br{\nu, \alpha, \boldsymbol{P}}$ sub-exponential random vector with $\frac{2}{\nu^2} \leq \frac{1}{\alpha^2}$, $\mathrm{Cov}(\boldsymbol{w}) \preceq \frac{3}{2} \nu^2 \boldsymbol{P}$.
\end{lemma}
\begin{proof}
For all unit vector $\boldsymbol{u}$, we have
\[
\boldsymbol{u}^{\top} \mathrm{Cov}(\boldsymbol{w}) \boldsymbol{u} = \boldsymbol{u}^{\top} \mathrm{E}(\boldsymbol{w} \boldsymbol{w}^{\top}) \boldsymbol{u} = \mathrm{E}[(\boldsymbol{w}^{\top} \boldsymbol{u} )^2] = \mathrm{Var}(\boldsymbol{w}^{\top} \boldsymbol{u}).
\]
By definition, if $\boldsymbol{w}$ is a is a mean zero $\br{\sigma, \boldsymbol{P}}$ sub-Gaussian random vector, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $\sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}$ sub-Gaussian. Therefore,
\[ \boldsymbol{u}^{\top} \mathrm{Cov}(\boldsymbol{w}) \boldsymbol{u} = \mathrm{Var}(\boldsymbol{w}^{\top} \boldsymbol{u}) \leq \sigma^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}
\]
as desired.
Similarly, if $\boldsymbol{w}$ is a is a mean zero $\br{\nu, \alpha, \boldsymbol{P}}$ sub-exponential random vector, $\boldsymbol{w}^{\top} \boldsymbol{u}$ is $(\nu \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}}, \alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}})$ sub-exponential. By Lemma~\ref{lem:subexp-var},
\[ \boldsymbol{u}^{\top} \mathrm{Cov}(\boldsymbol{w}) \boldsymbol{u} = \mathrm{Var}(\boldsymbol{w}^{\top} \boldsymbol{u}) \leq \frac{3}{2} \nu^2 \boldsymbol{u}^{\top} \boldsymbol{P} \boldsymbol{u}.
\]
\end{proof}
\begin{lemma}
\label{lem:epsilon}
For all $i \in [n]$
\begin{enumerate}
\item $\epsilon_i$ is sub-Gaussian with $\sigma = 1$, and
\item $\epsilon_i$ is $(4\sqrt{Aq}, 2/B)$ sub-exponential for any $A,B > 0$ satisfying $e^x < 1 + x + Ax^2$ for $x < B$.
\end{enumerate}
\end{lemma}
\begin{proof}
For the first claim, note that a random variable bounded in $[a, b]$ is sub-Gaussian with $\sigma^2 = \frac{(b-a)^2}{4}$.
To prove second claim, we have
\begin{align*}
\exp(\lambda \epsilon_i) &= \tilde{q} \exp(2 \lambda(1-\tilde{q})) + (1-\tilde{q}) \exp(-2 \lambda \tilde{q}) \\
&= \exp(-2 \lambda \tilde{q}) (1 + \tilde{q}(\exp(2 \lambda)-1) ) \\
& < \exp \br{-2 \lambda \tilde{q}} \exp \br{\tilde{q}(\exp(2 \lambda)-1) )} \\
& \leq \exp \br{\tilde{q} A (2 \lambda)^2},
\end{align*}
for $2\lambda < B.$ The last step follows from $e^x \leq 1+ x + Ax^2$ for $x <B.$
Recall that $\tilde{q} < 2q$, we have
$$
\exp(\lambda \epsilon_i) \leq \exp \br{ \frac{16qA\lambda^2}{2}}
$$
for $\lambda < B/2$ as desired.
\end{proof}
Let $\boldsymbol{P}_i, \, i\in [n]$ be orthogonal projection matrices onto the span of $\{\boldsymbol{x}_1,..,\boldsymbol{x}_i\},$ that is,
\[\boldsymbol{P}_i := \boldsymbol{P}_{i-1} + \frac{(\boldsymbol{I} - \boldsymbol{P}_{i-1}) \boldsymbol{x}_i \boldsymbol{x}_i^{\top} (\boldsymbol{I} - \boldsymbol{P}_{i-1})}{\|(\boldsymbol{I} - \boldsymbol{P}_{i-1})\boldsymbol{x}_i \|^2},
\]
with $\boldsymbol{P}_0 = 0.$
\begin{lemma}
\label{lem:main}
Suppose $A,B > 0$ satisfy $e^x < 1 + x + Ax^2$ for all $x < B$. Let $\sigma^2 :=c/2q$, $\nu^2 := 8Ac$ and $\alpha := 2/B$. Then
\begin{enumerate}
\item If $\boldsymbol{w}_{i-1}$ is $(\sigma, \boldsymbol{P}_{i-1})$ sub-Gaussian, $\boldsymbol{w}_i$ is $(\sigma, \boldsymbol{P}_i)$ sub-Gaussian.
\item If $\boldsymbol{w}_{i-1}$ is $(\nu, \alpha, \boldsymbol{P}_{i-1})$ sub-exponential, $\boldsymbol{w}_i$ is $(\nu, \alpha, \boldsymbol{P}_i)$ sub-exponential.
\end{enumerate}
\end{lemma}
\begin{proof}
We have
\begin{align*}
\mathrm{E} \left[ \exp( \lambda \boldsymbol{w}_i^{\top} \boldsymbol{u}) \right] &= \mathrm{E} \brs{ \mathrm{E} \brs{
\exp( \lambda \boldsymbol{w}_i^{\top} \boldsymbol{u}) \big| \boldsymbol{w}_{i-1}
} } \\
&= \mathrm{E} \brs{ \mathrm{E} \brs{ e^{ \lambda
\left \langle \boldsymbol{w}_{i-1},
\boldsymbol{u} - 2 q \frac{\boldsymbol{x}_i^{\top} \boldsymbol{u}}{c} \boldsymbol{x}_i \right \rangle + \lambda \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}} \big| \boldsymbol{w}_{i-1}}} \\
& = \mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \mathrm{E} \brs{ e^{\lambda \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}} \big| \boldsymbol{w}_{i-1}}}.
\end{align*}
First, suppose that $\boldsymbol{w}_{i-1}$ is $(\sigma, \boldsymbol{P}_{i-1})$ sub-Gaussian, we will prove that $\boldsymbol{w}_i$ is $(\sigma, \boldsymbol{P}_i)$ sub-Gaussian. By Lemma~\ref{lem:epsilon}, $\epsilon_i$ is 1-sub-Gaussian. Therefore,
\begin{align*}
\mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \mathrm{E} \brs{ e^{\lambda \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}} | \boldsymbol{w}_{i-1}}}
& \leq \mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \cdot e^{\frac{1}{2} \lambda^2 \|\boldsymbol{x}_i^{\top} \boldsymbol{u}\|^2 }} \\
& \leq e^{ \frac{ \lambda^2 \sigma^2}{2} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}^{\top} \boldsymbol{P}_{i-1} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}
+\frac{\lambda^2 }{2} (\boldsymbol{u}^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u}) }.
\end{align*}
We now consider the exponent (divided by the common factor $\lambda^2 $). It is sufficient to show
\begin{align*}
\frac{\sigma^2}{2} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}^{\top} \boldsymbol{P}_{i-1} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}
+ \frac{1}{2}u^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u} \leq \frac{\sigma^2}{2} \boldsymbol{u}^{\top} \boldsymbol{P}_i \boldsymbol{u} \, \, \forall \boldsymbol{u}\\
\iff {\sigma^2} \boldsymbol{Q}_{\boldsymbol{x}_i, c} ^{\top} \boldsymbol{P}_{i-1} \boldsymbol{Q}_{\boldsymbol{x}_i, c}
+ \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \preceq {\sigma^2} \boldsymbol{P}_i.
\end{align*}
Using Proposition~\ref{prop:loewner} (with $L \leftarrow 1, C \leftarrow c/2q$) and assuming $ \sigma^2 = c/2q \geq 1$, we have
\begin{align*}
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{P}_{i-1} \boldsymbol{Q}_{\boldsymbol{x}_i, c} \sigma^2
+ \boldsymbol{P}_{\boldsymbol{x}_i} &= \frac{c}{2q} \left( \boldsymbol{I} - \frac{2q}{c} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \right) \boldsymbol{P}_{i-1} \left( \boldsymbol{I} - \frac{2q}{c} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \right) + \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \\
&\preceq \frac{c}{2q} \boldsymbol{P}_i.
\end{align*}
Therefore, $\boldsymbol{w}_i$ is a $(c/2q, \boldsymbol{P}_i) $ sub-Gaussian random vector.
Now suppose that $\boldsymbol{w}_{i-1}$ is $(\nu, \alpha, \boldsymbol{P}_{i-1})$ sub-exponential, we will prove that $\boldsymbol{w}_i$ is $(\nu, \alpha, \boldsymbol{P}_{i})$ sub-exponential. Again, by Lemma~\ref{lem:epsilon}, $\epsilon_i$ is $(4\sqrt{Aq}, 2/B)$ sub-exponential. Therefore,
\begin{align*}
\mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \mathrm{E} \brs{ e^{\lambda \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}} | \boldsymbol{w}_{i-1}}}
& \leq \mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \cdot e^{8Aq \lambda^2 \|\boldsymbol{x}_i^{\top} \boldsymbol{u}\|^2 }}
\end{align*}
for $|\lambda| < \frac{2}{B|\boldsymbol{x}_i^{\top} \boldsymbol{u}|} = \frac{1}{\alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u}}}$.
Since $\boldsymbol{w}_{i-1}$ is $(\nu, \alpha, \boldsymbol{P}_{i-1})$ sub-exponential,
\[
\mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} }
\leq
e^{ \frac{ \lambda^2 \nu^2}{2} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}^{\top} \boldsymbol{P}_{i-1} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}}
\]
for $|\lambda| < \frac{1}{\alpha \sqrt{ \boldsymbol{u}^{\top} \boldsymbol{P}_{i-1} \boldsymbol{u}}}$.
Note that $\boldsymbol{u}^{\top} \boldsymbol{P}_{i} \boldsymbol{u}$ is greater than both $\boldsymbol{u}^{\top} \boldsymbol{P}_{i-1} \boldsymbol{u}$ and $\boldsymbol{u}^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u}$. We have
\begin{align*}
\mathrm{E} \brs{ e^{ \lambda
\langle \boldsymbol{w}_{i-1},
\boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u} \rangle} \mathrm{E} \brs{ e^{\lambda \epsilon_i \boldsymbol{x}_i^{\top} \boldsymbol{u}} | \boldsymbol{w}_{i-1}}}
\leq
e^{ \frac{ \lambda^2 \nu^2}{2} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}^{\top} \boldsymbol{P}_{i-1} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}
+\ {8Aq\lambda^2 } (\boldsymbol{u}^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u}) }
\end{align*}
for all $|\lambda| < \frac{1}{\alpha \sqrt{\boldsymbol{u}^{\top} \boldsymbol{P}_{i} \boldsymbol{u}}}$.
Hence, it is sufficient to show
\begin{align*}
\frac{\nu^2}{2} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}^{\top} \boldsymbol{P}_{i-1} \br{ \boldsymbol{Q}_{\boldsymbol{x}_i, c} \boldsymbol{u}}
+ 8Aq \boldsymbol{u}^{\top} \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \boldsymbol{u} \leq \frac{\nu^2}{2} \boldsymbol{u}^{\top} \boldsymbol{P}_i \boldsymbol{u} \, \, \forall \boldsymbol{u}\\
\iff {\nu^2} \boldsymbol{Q}_{\boldsymbol{x}_i, c} ^{\top} \boldsymbol{P}_{i-1} \boldsymbol{Q}_{\boldsymbol{x}_i, c}
+ 16Aq \boldsymbol{x}_i \boldsymbol{x}_i^{\top} \preceq {\nu^2} \boldsymbol{P}_i.
\end{align*}
This follows from Proposition~\ref{prop:loewner} by substituting $L \leftarrow 16Aq$ and $C \leftarrow c/2q$ and noting that $\nu^2 = 8Ac = Lc$.
\end{proof}
\begin{lemma}
\label{lem:success-prob}
If $c = \min(1/q, 9.3) \log(2n/\delta)$ then with probability at least $1-\delta$, we have
\[
|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_{i}| < c \text{ for all } i \in [n].
\]
\end{lemma}
\begin{proof}
By definition, $c$ is either equal to $\log(2n/\delta)/q$ or $9.3 \log(2n/\delta)$. We consider these two cases.
First suppose $c = \log(2n/\delta)/q$. With $\sigma^2 = c/2q$, we have
$$
c = \log(2n/\delta)/q = \sigma \sqrt{2 \log(2 n/ \delta)}.
$$
By Lemma~\ref{lem:subgaussian-bound},
\begin{align*}
P\br{|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_{i}| > c}
&\leq P\br{|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_{i}| > c \sqrt{\boldsymbol{x}_i^{\top} \boldsymbol{P}_{i-1} \boldsymbol{x}_i}} \\
&\leq 2 \exp \br{ -\frac{c^2}{2 \sigma^2 }} \leq \delta/n.
\end{align*}
The result then follows by a union bound over $i \in [n].$
Now suppose $c = 9.3 \log(2n/\delta)$. Note that $A = 0.5803$ and $B = 0.4310$ satisfy $e^x < 1 + x + Ax^2$ for $x < B$. Let $\nu^2 =8Ac$ and $\alpha = 2/B$. We have
$$
\frac{\nu^2}{\alpha} = \frac{8Ac }{2/B} = {4ABc} > c .
$$
Therefore, by Lemma~\ref{lem:subexp-bound},
\begin{align*}
P\br{|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_{i}| > c}
&\leq P\br{|\boldsymbol{w}_{i-1}^{\top} \boldsymbol{x}_{i}| > c \sqrt{\boldsymbol{x}_i^{\top} \boldsymbol{P}_{i-1} \boldsymbol{x}_i}} \\
&\leq 2 \exp \br{ -\frac{c^2}{2 \nu^2 }}
= 2 \exp \br{-\frac{c}{16A}} \\
&< 2 \exp \br{-\frac{c}{9.3}} \leq \delta/n.
\end{align*}
Again, the result follows by union bounding over $i \in [n].$
\end{proof}
Lemma~\ref{lem:success-prob} and Lemma~\ref{lem:main} together prove Theorem~\ref{thm:main}.
\subsection{Proof of Theorem~\ref{thm:multi}}
\begin{proof}[Proof of Theorem~\ref{thm:multi}]
Let $D(\delta)$ be the discrepancy obtained by \texttt{BinaryBalance} as a function of the failure probability $\delta$. We will show that $(2 \log k) D(\delta/k)$ is the corresponding discrepancy obtained by Algorithm~\ref{alg:multi}.
Theorem~\ref{thm:multi} will then follow from Proposition~\ref{prop:balance}.
Notice that with probability $\delta/k$, each run of \texttt{BinaryBalance} at an internal node of $T$ has discrepancy $D(\delta/k)$. By union bounding over $O(k)$ internal nodes, we have that with probability $1-\delta$, all of the discrepancies are bounded by $D(\delta/k)$.
Assume all the discrepancies at the internal nodes in $T$ are bounded, we show how to bound the discrepancy between any two treatments. Let $l$ and $l'$ be any two leaves in $T$. The goal is to show that
$$\left\| \frac{\alpha(l')}{\alpha(l') + \alpha(l)}s(l) - \frac{\alpha(l)}{\alpha(l') + \alpha(l)}s(l')\right\|_\infty$$
is small.
First we relate $s(v)$ to $s(v_l)$ and $s(v_r)$ where $v_l, v_r$ are the left and right children of $v$. By definition,
\[ w(v) = \frac{\alpha(v_r)}{\alpha(v_l) + \alpha(v_r)} s(v_l) - \frac{\alpha(v_l)}{\alpha(v_l) + \alpha(v_r)} s(v_r) \]
and
\[s(v) = s(v_l) + s(v_r).\]
Therefore,
\[w(v) = s(v_l) - \frac{\alpha(v_l)}{\alpha(v_l) + \alpha(v_r)} s(v),\]
and
\[- w(v) = s(v_r) - \frac{\alpha(v_r)}{\alpha(v_l) + \alpha(v_r)} s(v).\]
Hence, both
\[ \left\|s(v_l) - \frac{\alpha(v_l)}{\alpha(v)} s(v) \right\|_{\infty}
\hspace{0.5cm} \text{and} \hspace{0.5cm}
\left\|s(v_r) - \frac{\alpha(v_r)}{\alpha(v)} s(v) \right\|_{\infty} \]
are bounded by $D(\delta/k)$.
Now consider $v_1, v_2$ and $v_3$ in $T$ such that $v_1$ is a child of $v_2$ and $v_2$ is a child of $v_3$. We have, by triangle inequality,
\[ \left\|s(v_1) - \frac{\alpha(v_1)}{\alpha(v_3)} s(v_3) \right\|_{\infty} \leq \left\|s(v_1) - \frac{\alpha(v_1)}{\alpha(v_2)} s(v_2) \right\|_{\infty} + \frac{\alpha(v_1)}{\alpha(v_2)}\left\|s(v_2) - \frac{\alpha(v_2)}{\alpha(v_3)} s(v_3) \right\|_{\infty} \leq \left(1+ \frac{\alpha(v_1)}{\alpha(v_2)}\right) D(\delta/k). \]
Let $l$ be a leaf in $T$ and let $l,v_1,v_2 \ldots r$ be the path from $l$ to the root $r$. Repeatedly applying the above relation along the path gives
\begin{equation}
\label{eqn:sum-path}
\left\|s(l) - \frac{\alpha(l)}{\alpha(r)} s(r) \right\|_{\infty} \leq
\left( 1 + \frac{\alpha(l)}{\alpha(v_1)} + \frac{\alpha(l)}{\alpha(v_2)} \ldots + \frac{\alpha(l)}{\alpha(r)} \right) D(\delta/k).
\end{equation}
Since there are at most $\log k$ nodes in the path from $l$ to $r$,
\[ \left\|s(l) - \frac{\alpha(l)}{\alpha(r)} s(r) \right\|_{\infty} \leq (\log k) D(\delta/k).\]
Finally, for any two leaves $l$ and $l'$,
\[ \left\|\frac{s(l)/\alpha(l) - s(l') /\alpha(l')}{1/\alpha(l) + 1/\alpha(l') }\right\|_{\infty} \leq \left\|\frac{s(l)/\alpha(l) - s(r) /\alpha(r)}{1/\alpha(l) + 1/\alpha(l') }\right\|_{\infty} + \left\|\frac{s(l')/\alpha(l') - s(r) /\alpha(r)}{1/\alpha(l) + 1/\alpha(l') }\right\|_{\infty} \leq (2 \log k) D(\delta/k).\]
\end{proof}
\begin{remark}
If all weights are uniform, the summation in (\ref{eqn:sum-path}) becomes a geometric series and can be bounded by a constant. Therefore, we can remove the factor $\log k$ in Theorem~\ref{thm:multi}.
\end{remark}
\subsection{Other Proofs}
\begin{proof}[Proof of Proposition~\ref{prop:z-dist}]
Let $\boldsymbol{B}=\left[\begin{array}{c}
\sqrt{\phi} \boldsymbol{I} \\
\sqrt{1-\phi} \boldsymbol{X}^{\top}
\end{array}\right]$.
We have from Threorem~\ref{thm:main} that $\boldsymbol{w}_n = \boldsymbol{B} \boldsymbol{\eta} = \left[\begin{array}{c}
\sqrt{\phi} \boldsymbol{\eta} \\
\sqrt{1-\phi} \boldsymbol{X}^{\top} \boldsymbol{\eta}
\end{array}\right] $ is a mean zero $(c/2q, \boldsymbol{P})$ sub-Gaussian random vector, where
\begin{align*}
\boldsymbol{P}
& = \boldsymbol{B} (\boldsymbol{B}^{\top} \boldsymbol{B})^{-1} \boldsymbol{B}^{\top} \\
& =
\left[
\begin{array}{cc}
\phi \boldsymbol{Q} & *\\
* & *
\end{array}
\right].
\end{align*}
Therefore, by sub-Gaussianity of $\boldsymbol{w}_n,$ for any vector $\boldsymbol{u},$ we have
\[
\mathrm{E} \brs{ \exp\br{\sqrt{\phi}\boldsymbol{\eta}^{\top} \boldsymbol{u} }} \leq
\exp \br{ \frac{c}{4q} \boldsymbol{u}^{\top} \br{\phi \boldsymbol{Q}} \boldsymbol{u}.}
\]
We therefore have the sub-Gaussian claim. The sub-exponential result follows similarly.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:spectral-bound}]
Suppose $c = \log(2n/\delta)/q.$ We have from Proposition~\ref{prop:z-dist} that $\boldsymbol{z}$ is a $(\sqrt{c/2q}, \boldsymbol{Q})$ sub-Gaussian vector.
This implies that
$\mathrm{Cov}(\boldsymbol{z}) \preceq \frac{c}{2q} \boldsymbol{Q}.$
When $c= 9.3 \log(2n/\delta),$ we have from Proposition~\ref{prop:z-dist} that $\boldsymbol{\eta}$ is a $(\sqrt{8Ac}, \alpha, \boldsymbol{Q})$ sub-exponential vector. Now, Lemma~\ref{lem:subexp-var} gives that
\begin{align*}
\mathrm{Cov}(\boldsymbol{z}) \preceq \frac{3}{2} 8Ac \boldsymbol{Q}
= 12 A c \boldsymbol{Q}.
\end{align*}
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:balance}]
When $c = \frac{\log(4n/\delta)}{q}$ we have that with probability at least $1-\delta/2,$ $\boldsymbol{B} \boldsymbol{\eta}$ is a $\br{\sqrt{c/2q}, \boldsymbol{Q}}$ sub-Gaussian vector. Since $\boldsymbol{B} \boldsymbol{\eta} = \left[\begin{array}{c}
\sqrt{\phi} \boldsymbol{\eta} \\
\sqrt{1-\phi} \boldsymbol{X}^{\top} \boldsymbol{\eta}
\end{array}\right]$ and $\boldsymbol{Q} \preceq \phi \boldsymbol{I},$ we have
$\sqrt{1-\phi} \sum_i \eta_i \boldsymbol{x}_i = \sqrt{1-\phi} \boldsymbol{X}^{\top}\boldsymbol{\eta}$ is a $\br{\sqrt{1/\phi}\sqrt{c/2q}, \boldsymbol{I} } $ sub-Gaussian random vector. By sub-Gaussian concentration, we have with probability at least $1-\delta/2d$, $|\br{\boldsymbol{X}^{\top} \boldsymbol{\eta}}^{\top} \boldsymbol{e}_i| \leq \br{\sqrt{(c/2\phi (1-\phi) q}} \sqrt{4d/\delta}. $ The result follows by a union bound over $\boldsymbol{e}_1,...,\boldsymbol{e}_d$ and $\|\boldsymbol{X}^{\top}\boldsymbol{\eta} \|_{2} \leq \sqrt{d}\|\boldsymbol{X}^{\top}\boldsymbol{\eta} \|_{\infty}. $
When $c = 9.3 \log(4n /\delta),$ then with probability $1-\delta,$ $ \left[\begin{array}{c}
\sqrt{\phi} \boldsymbol{\eta} \\
\sqrt{1-\phi} \boldsymbol{X}^{\top} \boldsymbol{\eta}
\end{array}\right]$ is a $(\sqrt{8Ac}, \alpha, \boldsymbol{Q})$ random vector. Like before, this implies $\sqrt{1-\phi} \boldsymbol{X}^{\top} \boldsymbol{\eta}$ is a $(\sqrt{8Ac}, \alpha, \phi \boldsymbol{I})$ random vector. By sub-exponential concentration, we have
\begin{align*}
P \br{ \left|\sqrt{1-\phi} \br{\boldsymbol{X}^{\top} \boldsymbol{\eta}}^{\top} \boldsymbol{e}_i \right| \geq t}
\leq \left\{ \begin{array}{cc}
2 \exp \br{-t^2 \phi/2 \nu^2} & \text{ when } t \leq \frac{\nu^2}{\alpha \sqrt{\phi}} \\
2 \exp \br{ -t/2 \alpha} & \text{ otherwise }
\end{array} \right .
\end{align*}
Setting $t = \sqrt{ 2 \log (4d \delta) (8Ac)} \leq c \leq \nu^2/\alpha, $ (when $n \geq d$), we get that with probability at least $1-\delta/2d,$ we have
\[
|\sqrt{1-\phi} \br{\boldsymbol{X}^{\top} \boldsymbol{\eta}}^{\top} \boldsymbol{e}_i| \leq 9.3 \sqrt{ \frac{ \log (4d /\delta) \log (4n/\delta)}{\phi}.}
\]
The rest follows by a union bound.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:ate-concentration}.]
First consider the case when $c = \log(2n/\delta)/q = \sigma \sqrt{2 \log(2 n/ \delta)}$. By Proposition~\ref{prop:z-dist}, $\boldsymbol{\eta}$ is $(\sigma, \boldsymbol{Q})$ sub-Gaussian with $\sigma = \sqrt{c/2q}$. From Lemma~\ref{lem:subgaussian-bound}, we have
\begin{align*}
P(|\boldsymbol{\eta}^{\top}\boldsymbol{\mu}| > c\sqrt{\boldsymbol{\mu}^T\boldsymbol{Q}\boldsymbol{\mu}}) \leq 2 \exp \br{ - \frac{c^2}{2\sigma^2}} = \delta / n.
\end{align*}
Now consider $c = 9.3 \log(2n/ \delta)$. By Proposition~\ref{prop:z-dist}, $\boldsymbol{\eta}$ is $(\nu,\alpha, \boldsymbol{Q})$ sub-exponential with $\nu = \sqrt{8Ac}$. Note that
$$
{\nu^2}/{\alpha} = {8Ac } /{(2/B)} = {4ABc} > c .
$$
From Lemma~\ref{lem:subexp-bound}, we have
\begin{align*}
P(|\boldsymbol{\eta}^{\top}\boldsymbol{\mu}| > c\sqrt{\boldsymbol{\mu}^T\boldsymbol{Q}\boldsymbol{\mu}}) \leq 2 \exp \br{ - \frac{c^2}{2\nu^2}} \\
= 2 \exp \br{ - \frac{c}{16A}}\leq \delta / n.
\end{align*}
The result then follows by a union bound.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:ate-ridge}.]
This follows from Proposition~\ref{prop:spectral-bound} and the proof of Theorem 3 in \cite{harshaw2020balancing}.
\end{proof}
\subsection{Robustness}
\label{sec:robust}
Proposition~\ref{prop:spectral-bound} immediately gives a bound on
$\lambda_{\max}(\mathrm{Cov}(\boldsymbol{z}))$ and hence bounds accidental bias.
\begin{remark}[Accidental Bias]
With probability $\geq 1-\delta$ the maximum eigenvalue of $\mathrm{Cov}(\boldsymbol{z})$ satisfies
\[
\lambda_{\max}\br{\mathrm{Cov}(\boldsymbol{z})} \leq \frac{\sigma^2}{\phi}.
\]
\end{remark}
\section{Simulations}
\label{app:sims}
\subsection{Description}
\begin{table*}[h!]
\centering
\begin{tabular}{rccc}
\toprule
\textbf{DGP Name} & $\mathbf{X}$ & $\mathbf{y(0)}$ & $\mathbf{y(a)\; \mathrm{\bf s.t.}\; a \neq 0}$\\
\midrule
\textbf{QuickBlockDGP} & $X_{i,k} \sim \mathcal{U}(0,10), \forall k \in \{1,2\}$ & $\prod_{k=1}^2 X_{k} + \epsilon$ & 1 + y(0) \\
\midrule
\textbf{LinearDGP} & $X_{i,k} = \epsilon_{k}, k \in \{1,\dots,4\}$& $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(1)}$\\
\midrule
\textbf{LinearDriftDGP} & $X_{i,k} = \frac{i}{N} + \epsilon_{k}, k \in \{1,\dots,4\}$& $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(1)}$\\
\midrule
\textbf{LinearSeasonDGP} & $X_{i,k} = \sin(2\pi\frac{i}{N}) + \epsilon_{k}, k \in \{1,\dots,4\}$& $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mathbf{X} \beta + \frac{1}{10}\epsilon_{y(1)}$\\
\midrule
\textbf{QuadraticDGP} & $X_{i,k} = 2\beta_k - 1, k \in \{1, 2\}$& $\mu_0 + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mu_0 + \frac{1}{10}\epsilon_{y(1)}$\\
&$\mu_0 = X_1 - X_2 + X_1^2 + X_2^2 - 2 X_1 X_2$&& \\
\midrule
\textbf{CubicDGP} & $X_{i, k} = 2\beta_k - 1, k \in \{1, 2\}$& $\mu_0 + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mu_0 + \frac{1}{10}\epsilon_{y(1)}$\\
&$\mu_0 = X_1 - X_2 + X_1^2 + X_2^2 - 2 X_1 X_2$&& \\
&$+ X_1^3 -X_2^3 - 3 X_1^2 X_2 + 3X_1 X_2^2$&&\\
\midrule
\textbf{SinusoidalDGP} & $X_{i,k} = 2\beta_k - 1, k \in \{1, 2\}$& $\mu_0 + \frac{1}{10}\epsilon_{y(0)}$ & 1 + $\mu_0 + \frac{1}{10}\epsilon_{y(1)}$\\
&$\mu_0 = \sin\left(\frac{\pi}{3} + \frac{\pi X_1}{3} - \frac{2 \pi X_2}{3}\right)$&& \\
&$-6 \sin(\frac{\pi X_1}{3} + \frac{\pi X_2}{4}) +6 \sin(\frac{\pi X_1}{3} + \frac{\pi X_2}{6})$&&\\
\bottomrule
\end{tabular}
\caption{Data generating processes used in simulations. All $\epsilon$s indicate a standard normal variate and all $\beta$s indicate a standard uniform variate. $i$ indicates a unit's index. Covariate vectors are row-normalized to unit norm, except for the QuickBlock simulation which just normalized relative to the maximum row norm.}
\label{tab:dgps}
\end{table*}
\clearpage
\subsection{Binary Treatments}
\paragraph{Bias.} Figure~\ref{fig:bias} shows that none of the examined methods are biased (but that does not imply that they are robust \citep{efron1971forcing, harshaw2020balancing}).
\begin{figure}
\centering
\includegraphics[width=0.475\textwidth]{figures/bias-1.pdf}
\caption{Bias}
\label{fig:bias}
\end{figure}
\paragraph{Imbalance.}
To measure linear imbalance, we calculate the $L_2$ norm of the difference in covariate means.
While this is far from the only measure of imbalance, it serves as an effective metric to demonstrate how well various metrics serve to eliminate linear imbalances, a common diagnostic used my experimenters.
We further normalize across sample-size by multiplying by the sample size.
Since all methods are unbiased, this has the effect of showing parametric convergence rates as a flat line in the graph.
Unsurprisingly, methods which directly optimize for linear imbalance perform very well in Figure~\ref{fig:imba}.
Our proposed algorithm has better finite sample imbalance minimization than does the algorithm of \citep{alweiss2020discrepancy} due to the finite sample improvements of \citep{dwivedi2021kernel}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/imba-1.pdf}
\caption{Imbalance. BWD is highly effective at eliminating linear imbalance between groups.}
\label{fig:imba}
\end{figure}
\subsection{Non-uniform assignment}
\paragraph{MSE.}
Figure~\ref{fig:mse-q} shows the resulting mean squared-error attained by methods which support marginal probabilities not equal to one-half.
All methods perform well, with DM performing effectively on nearly linear processes, and QuickBlock performing slightly better when the true process is highly non-linear.
\begin{figure}
\centering
\includegraphics[width=0.475\textwidth]{figures/q-mse-1.pdf}
\caption{Marginal probability of treatment}
\label{fig:mse-q}
\end{figure}
\paragraph{Marginal probability.}
The evaluation in Figure~\ref{fig:marginal-q} examines how closely each method hews to the desired marginal probability of treatment.
All methods do a good job of ensuring the appropriate marginal distribution.
\begin{figure}
\centering
\includegraphics[width=0.475\textwidth]{figures/q-marginal-1.pdf}
\caption{Marginal probability of treatment}
\label{fig:marginal-q}
\end{figure}
\subsection{Multiple Treatments}
\paragraph{Entropy.}
To ensure that treatment is being assigned with the correct marginal probability, we can measure the normalized entropy of the empirical marginal treatment probabilities.
If the values are perfectly uniform, then the value will be exactly one.
As the normalized entropy decreases, the marginal probabilities are more uneven, indicating a failure to match the desired marginal distribution.
BWD performs very similarly to complete randomization (which almost exactly matches the desired marginal probabilities), slightly out-performing \citep{smith1984sequential}.
Note that while \citet{smith1984sequential} only seeks to optimize the marginal probability of treatment for each unit, BWD additionally balances covariates.
\begin{figure}
\centering
\includegraphics[width=0.475\textwidth]{figures/multi-marginal-entropy-1.pdf}
\caption{Entropy}
\label{fig:entropy}
\end{figure}
\end{appendix}
\bibliographystyle{abbrvnat}
| a828d720bcfad6ad737333a2ed13dc8ffd3eea53 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Autophoretic colloids that isotropically emit or absorb solute molecules at their interface are prototypical examples of synthetic active matter. At low levels of chemical activity (quantified by a suitably small intrinsic P\'{e}clet number, \Pen), diffusion is dominant enough to homogenize any perturbations to the solute distribution around such a particle, thereby rendering the concentration profile isotropic, and the particle stationary. Beyond a threshold P\'{e}clet number, however, small perturbations to the concentration field result in the breaking of the fore-aft symmetry of the concentration profile, and the particle may undergo self-propulsion along a rectilinear path, or, at sufficiently large $\Pen$, execute meandering, helical or chaotic motion. The spontaneous self-propulsion of autophoretic particles and droplets has been predicted theoretically~\citep{Michelin2013,Hu2019}, as well as being observed in experiments~\citep{Izri2014,Maass2016,Hokmabad2021,Suda2021} and numerical simulations~\citep{Michelin2013,Morozov2019}. While the steady self-propulsion of autophoretic particles is well-characterised, dynamics of autophoresis in the chaotic regime, and the transition from steady to chaotic self-propulsion, is relatively less so.
\citet{Michelin2013} established from numerical computations of rigid spherical autophoretic particles that a transition from quiescence to steady self-propulsion occurs at $\Pen=4$. They observed a maximum in the swimming speed at $\Pen\approx9$, after which the speed decreases up to $\Pen=20$, which was the largest value of $\Pen$ they studied.~\citet{Morozov2019} performed simulations on autophoretic particles with a tuneable parameter $m$ that dictates the relative importance of diffusiophoretic effects to Marangoni flow, with $m=0$ corresponding to an active droplet that moves solely under the effect of Marangoni stresses, and $m\to\infty$ corresponding to a rigid particle moving solely due to diffusioosmotic slip. A larger range of P\'{e}clet numbers is explored in their work, and the transition of a particle (with $m=2$) from its quiescent state to rectilinear self-propulsion, and ultimately chaotic motion is predicted as $\Pen$ is increased. Their calculations assume that the solute concentration and flow field around the particle remain axially symmetric about its initial direction of translation; hence, the particle undergoes unsteady, rectilinear translation. Interestingly, ~\citet{Morozov2019} observe a ballistic mean square displacement (MSD) scaling in the chaotic regime: that is, they do not observe a transition to non-ballistic motion associated with the chaotic dynamics. They do, however, comment that much longer computations than they conducted might be needed to fully characterize the chaotic regime. They also note that their computational approach, which employs a truncated spectral expansion of the concentration field, converges slowly with the number of Legendre modes for large $\Pen$, rendering computations expensive in this region of parameter space. Computations of autophoretic circular disks by ~\citet{Hu2019}, in which the particle is allowed to move in two dimensions, show that: (a) the transition to chaos occurs via an intermittency scenario~\citep{Pomeau1980}; and (b) the chaotic regime is characterized by a change in the scaling of the mean MSD of the particle as a function the lag time $\tau$, from a short-time ballistic profile ($\sim\tau^2$), to normal diffusion ($\sim\tau$) at longer times. Experiments on autophoretic droplets confined between walls in a Hele-Shaw type configuration~\citep{Hokmabad2021} show that an increase in the P\'{e}clet number results in a transition from ballistic motion to a two-dimensional self-avoiding walk ($\sim\tau^{3/2}$) at long times, as the droplets tend to avoid the chemical trail that they emit. More broadly, it is known that chaotic maps can lead to either normal~\citep{Geisel1982} or anomalous diffusion~\citep{Geisel1984,Sato2018}. Delay-differential equations, used to model phenomena such as blood cell production and the transmission of light from optical ring cavities, have also been shown to exhibit chaotic diffusion~\citep{Ikeda1980,Albers2022}. Given the differing theoretical and experimental results mentioned above, it is not evident what the MSD scaling of autophoretic particles in the chaotic regime is.
In this paper, using a spectral element based algorithm to simultaneously solve the unsteady advection-diffusion equation governing the concentration profile of the solute, and the quasi-steady Stokes equations governing the velocity field, we aim to chart the transition of a rigid autophoretic particle from steady self-propulsion to chaotic motion. Our computations are performed on an axisymmetric model, so the motion is restricted to unsteady rectilinear translation. A primary question we will address is the nature of the chaotic motion, i.e., does the active particle execute ballistic motion as predicted by~\citet{Morozov2019}, normal diffusion as seen in~\citet{Hu2019}, or super-diffusion as observed by~\citet{Hokmabad2021}? In fact, our computations indicate that none of these outcomes are attained. Instead, we find that the chaotic dynamics are characterised by long-time subdiffusion, where the MSD scales sublinearly with $\tau$. Notably, the scaling exponent is nearly identical for all values of $\Pen$ studied.
In~\cref{sec:prob_form}, we specify the governing equations for a neutrally buoyant, rigid, spherical autophoretic particle. The numerical details for the solution of the coupled system of nonlinear partial differential equations governing the transient evolution of the concentration field and the quasi-steady velocity field about the rigid particle are provided in~\cref{sec:num_details}, along with the details of the MSD and the velocity autocorrelation (VAC) calculations. We present and discuss the results in~\cref{sec:res_disc}, followed by concluding remarks in~\cref{sec:conclusion}.
\section {Problem formulation}
\label{sec:prob_form}
We consider a rigid, spherical autophoretic particle of radius $a^{*}$, freely suspended in an incompressible Newtonian fluid of viscosity $\eta^{*}$, whose flow obeys the Stokes equations. Above and henceforth, dimensional variables are marked with an asterisk as superscript. There is a constant flux $\mathcal{A}^{*}$ of solute at the particle surface, which is positive (negative) when the solute is being emitted (absorbed) at the interface. Far away from the particle, the uniform solute concentration is $C^{*}_{\infty}$, and the difference between the local concentration and its far-field value is denoted by $c^{*}=C^{*}-C^{*}_{\infty}$. The solute molecules, of diffusivity $D^{*}$, interact with the active particle via a short-ranged potential whose characteristic length is $b^{*}$, such that $b^{*}\ll a^{*}$. These interactions give rise to a tangential slip velocity along the particle surface~\citep{Anderson1989}, whose magnitude is set by the concentration gradient of the solute at the interface, and the mobility parameter, $\mathcal{M}^{*}=\pm {k_B^{*}T^{*}b^{*2}}/{\eta^{*}}$, where $k^{*}_{B}$ is Boltzmann's constant and $T^{*}$ is the absolute temperature. Attractive (repulsive) interactions between the solute and the active particle are described by a negative (positive) $\mathcal{M}^{*}$~\citep{Michelin2014}. The concentration profile of the solute, therefore, evolves due to diffusion and advection by fluid flow.
Following~\citet{Michelin2013}, length, time, fluid velocity, pressure, and concentration are scaled by $a^{*}$, ${a^{*}D^{*}}/{|\mathcal{A}^{*}\mathcal{M}^{*}|}$, ${|\mathcal{A}^{*}\mathcal{M}^{*}|}/{D^{*}}$, ${\eta^{*}|\mathcal{A}^{*}\mathcal{M}^{*}|}/{a^{*}D^{*}}$, and ${a^{*}|\mathcal{A}^{*}|}/{D^{*}}$, respectively. In the subsequent discussion, variables without asterisks are the dimensionless equivalent of their dimensional counterparts. A key dimensionless parameter is the intrinsic P\'{e}clet number,
\begin{equation}
\Pen=\dfrac{a^{*}|\mathcal{A}^{*}\mathcal{M}^{*}|}{D^{*2}},
\end{equation}
which quantifies the relative importance of the solute advection with respect to its diffusion, and is a measure of the chemical activity of the particle. It is convenient to introduce the scaled flux and mobility parameters,
\begin{equation}
A=\dfrac{\mathcal{A}^{*}}{|\mathcal{A}^{*}|},\,M=\dfrac{\mathcal{M}^{*}}{|\mathcal{M}^{*}|}.
\end{equation}
The concentration field is governed by the unsteady advection-diffusion equation,
\begin{equation}\label{eq:adv_diff}
\Pen\left(\dfrac{\partial c}{\partial t}+\boldsymbol{v}\bcdot\bnabla c\right)=\nabla^2 c,
\end{equation}
where $t$ is time and $\boldsymbol{v}$ denotes the velocity field. Eq.~\ref{eq:adv_diff} is subject to the two boundary conditions of: (i) constant flux of the solute at the surface of the active particle
\begin{equation}\label{eq:const_flux}
\dfrac{\partial c}{\partial r}=-A\quad\text{at}\quad r=1,
\end{equation}
and (ii) an attenuation condition far away from the sphere
\begin{equation}\label{eq:atten}
c\to0\quad\text{as}\quad r\to\infty.
\end{equation}
We employ a cylindrical coordinate system $(z,\rho,\phi)$ with its origin attached to the particle centre. The $z$-axis represents the axis of symmetry along which the particle motion is constrained. The perpendicular distance from the $z$-axis is measured by the $\rho$-coordinate, and $\phi$ denotes the azimuthal angle of rotation about the $z$-axis. This frame of reference is non-inertial since the particle is accelerating during its unsteady translation. However, this choice does not affect the advection-diffusion equation or the Stokes equations, since the motion occurs at zero Reynolds number. The particle surface is denoted by $r=1$ where $r=\sqrt{\rho^2+z^2}$. The axisymmetric flow around the particle is represented as
\begin{equation}
\boldsymbol{v}={v}_{z}(\rho,z)\boldsymbol{e}_{z}+{v}_{\rho}(\rho,z)\boldsymbol{e}_{\rho},
\end{equation}
where $\boldsymbol{e}_{z}$ and $\boldsymbol{e}_{\rho}$ are unit vectors along the $z$ and $\rho$ axis, respectively. It is useful to define the polar angle $\theta\equiv\arctan(\rho/z)$ measured anticlockwise such that $\theta=0$ lies on the positive $z$-axis.
The velocity field, $\boldsymbol{v}$, in (\ref{eq:adv_diff}) is governed by the incompressibility criterion and the Stokes equation,
\begin{equation}
\bnabla\bcdot\boldsymbol{v}=0,\nabla^2\boldsymbol{v}=\bnabla p,
\end{equation}
where $p$ is the dynamic pressure, and subject to the following boundary conditions
\begin{equation}\label{eq:slip_con}
\boldsymbol{v}\equiv M\bnabla_{\text{s}}c=v_{\text{s}}\left(\cos\theta\boldsymbol{e}_{\rho}-\sin\theta\boldsymbol{e}_{z}\right)\quad\text{at}\quad r=1,
\end{equation}
where
\begin{equation}
v_{\text{s}}=M\left(\cos\theta\dfrac{\partial c}{\partial \rho}-\sin\theta\dfrac{\partial c}{\partial z}\right),
\end{equation}
and
\begin{equation}\label{eq:far_field_vel}
\boldsymbol{v}\to-U\boldsymbol{e}_{z}\quad\text{as}\quad r\to\infty.
\end{equation}
The far-field speed $U$ in eq.~(\ref{eq:far_field_vel}) is unknown \textit{a priori}, and is determined iteratively by requiring that the total hydrodynamic force on the particle in the $z$-direction is zero at all times. The presumed axisymmetry of the problem permits us to rewrite the velocity field in terms of the Stokes streamfunction $\psi$,
\begin{equation}\label{eq:vel_def}
\boldsymbol{v}=\dfrac{1}{\rho}\left(\boldsymbol{e}_{z}\dfrac{\partial \psi}{\partial \rho}-\boldsymbol{e}_{\rho}\dfrac{\partial \psi}{\partial z}\right).
\end{equation}
Taking the curl of the Stokes equations and introducing the vorticity vector $\boldsymbol{\omega}=\bnabla\times\boldsymbol{v}$ eliminates the pressure from the governing equations, resulting in the following system of coupled partial differential equations,
\begin{equation}\label{eq:vort_def}
\omega\rho+E^2\psi=0,
\end{equation}
\begin{equation}\label{eq:vort_transport}
\nabla^2 \omega-\dfrac{\omega}{\rho^2}=0,
\end{equation}
where $\omega$ is the $\phi$-component of the vorticity $\boldsymbol{\omega}$ about the $z$-axis, with the other components ($\rho$ and $z$) of the vorticity vector vanishing due to symmetry, and the operators
\begin{equation}\label{eq:grad_op_def}
\nabla^2=\dfrac{1}{\rho}\dfrac{\partial}{\partial \rho}\left(\rho\dfrac{\partial}{\partial \rho}\right)+\dfrac{\partial^2}{\partial z^2},\,E^2=\nabla^2-\dfrac{2}{\rho}\dfrac{\partial}{\partial \rho}.
\end{equation}
Equations~(\ref{eq:vort_def}) and ~(\ref{eq:vort_transport}) must be solved subject to the following boundary conditions
\begin{equation}\label{eq:dn_grad_sfn_near}
\boldsymbol{n}\bcdot\bnabla\psi=-\rho v_{\text{s}}\quad\text{at}\quad r=1,
\end{equation}
\begin{equation}\label{eq:dn_grad_sfn_far}
\boldsymbol{n}\bcdot\bnabla\psi\to-\rho U\sin\theta\quad\text{as}\quad r\to\infty,
\end{equation}
\begin{equation}\label{eq:dn_grad_vort_far}
\boldsymbol{n}\bcdot\bnabla\omega\to0\quad\text{as}\quad r\to\infty,
\end{equation}
\begin{equation}
\psi=0\quad\text{at}\quad r=1,
\end{equation}
with $\boldsymbol{n}=\cos\theta\boldsymbol{e}_{z}+\sin\theta\boldsymbol{e}_{\rho}$ being the unit normal to the particle surface pointing into the fluid;
and the symmetry condition
\begin{equation}\label{eq:symaxis}
\psi=\omega=0\quad\text{along}\quad \rho=0.
\end{equation}
There exists a trivial solution to the system of equations given by eqs.~(\ref{eq:adv_diff}),~(\ref{eq:vort_def}),~(\ref{eq:vort_transport}), which is the isotropic concentration profile $c=1/r$, corresponding to the quiescent state of zero phoretic velocity and no fluid motion at all times. Beyond $\Pen=4$, however, the quiescent state becomes unstable with respect to dipolar perturbations in the concentration field~\citep{Michelin2013,Morozov2019}, and the autophoretic particle sets into motion. Further types of perturbation (such as quadrupolar) are unstable at larger values of the P\'{e}clet number as the higher hydrodynamic modes are excited. We therefore supply an asymmetrical concentration field as the initial condition to eq.~(\ref{eq:adv_diff}), as follows
\begin{equation}\label{eq:init_cond}
c(r,t=0)=\dfrac{1}{r}-\delta_{\text{per}}\left(\dfrac{\cos\theta}{r^2}\right),
\end{equation}
with $|\delta_{\text{per}}|<1$.
\section {Numerical solution methodology}
\label{sec:num_details}
As the first step in the numerical solution process, the initial condition given in~(\ref{eq:init_cond}) is used to specify the slip boundary condition~(\ref{eq:slip_con}) for the flow problem. The self-propulsion speed $U(t)$ at any instant of time $t$ must satisfy the requirement that total hydrodynamic drag force on the particle in the $z$-direction, $F_{z}$, given by~\citep{Khair2014}
\begin{equation}\label{eq:drag_calc}
F_{z}\equiv\pi\int_{0}^{\pi}\left[\dfrac{\partial(\omega r)}{\partial r}-2\omega\right]_{r=1}\sin^2\theta\,d\theta
\end{equation}
vanishes at $t$. The self-propulsion speed is evaluated iteratively using a secant method, as described in~\citet{Chisholm2016}. Given two initial guesses $U^{\left<j\right>}$ and $U^{\left<j-1\right>}$ at time $t$, where $j$ denotes the iteration number, the hydrodynamic drag at the two values of the self-propulsion speed are evaluated. Using linear interpolation, an improved estimate for $U$ is obtained as: $U^{\left<j+1\right>}=\left(U^{\left<j\right>}F_z^{\left<j-1\right>}-U^{\left<j-1\right>}F_z^{\left<j\right>}\right)/\left(F_z^{\left<j-1\right>}-F_z^{\left<j\right>}\right)$. The procedure is terminated when the magnitude of the difference in the computed speeds between successive iterations $|U^{\left<j\right>}-U^{\left<j-1\right>}|$ is reduced below $10^{-5}$. The converged solution for the flowfield $\left(\psi,\omega\right)$ at each time instant is then used to solve the advection-diffusion equation for the concentration field.
Numerical computations require the stipulation of a finite outer boundary, and a value of $R_{\text{o}}=100$ is chosen as the radius of the spherical shell on which the far-field boundary conditions are specified. The computational domain is discretized into quadrilateral spectral elements of order $N_{\text{o}}=8$ using the GMSH software package~\citep{Remacle2009}, and the system of equations given by~(\ref{eq:adv_diff}),~(\ref{eq:vort_def}), and~(\ref{eq:vort_transport}) is solved using a custom code based on the spectral element method~\citep{KarniadakisSherwin2005}. A detailed account of the solution methodology and its validation is provided in ~\citet{Chisholm2016} and ~\citet{Khair2018}. We only briefly recap the salient features here.
The shape functions are defined as tensor products of 1D Lagrange polynomials of order $N_{\text{o}}$, supported at $N_{\text{o}}+1$ Gauss-Lobatto quadrature points over the standard region $[-1,1]^2$. The time-derivative in eq.~(\ref{eq:adv_diff}) is discretized using the backward Euler algorithm, as follows
\begin{equation}\label{eq:euler_def}
\dfrac{\partial c}{\partial t}\approx \dfrac{c^{(i+1)}-c^{(i)}}{\Delta t},
\end{equation}
where $c^{(i)}$ denotes the value of the concentration at the discrete time $t^{(i)}$, and $\Delta t\equiv t^{(i+1)}-t^{(i)}$ is the width of the discrete timestep. The concentration variable is taken to be the arithmetic mean of its values at the current and previous timestep, that is,
\begin{equation}\label{eq:crank_nicol}
c\equiv\dfrac{c^{(i+1)}+c^{(i)}}{2}.
\end{equation}
Equations~(\ref{eq:euler_def}) and ~(\ref{eq:crank_nicol}) together represent the implicit Crank-Nicolson scheme~\citep{Tanaka1994} that is second-order accurate in the timestep width, $\Delta t$, and also unconditionally stable~\citep{DoneaHuerta2003}. A value of $\Delta t=1.0$ is used for simulations with $\Pen<20$, while a timestep width of 0.1 is used at higher values of the P\'{e}clet number.
The spatial integral over each element is computed using the Gauss-Lobatto quadrature rule, resulting in a system of algebraic equations which is solved iteratively using a Newton-Raphson scheme. The discretised version of eq.~(\ref{eq:adv_diff}), being linear (when the velocity field is prescribed), converges in one step, while the iterative procedure for the solution of the nonlinear system [eqs.~(\ref{eq:vort_def}) and~(\ref{eq:vort_transport})] is terminated when the $L^2$-norm of the difference between the solutions in successive steps, computed over all discretisation points, drops below $10^{-6}$.
Previous investigations~\citep{Michelin2013,Hu2019} on autophoresis have established that for particles with oppositely signed $A$ and $M$, perturbations to the concentration field vanish in the long-time limit, where the particle remains stationary. We therefore concern ourselves with the nontrivial case of similarly signed $A$ and $M$ values, which result in the self-propulsion of the active particle, and we pick $A=M=1$ without loss of generality.
At $\Pen\leq10$ the solver is supplied an initial condition of the form given by eq.~(\ref{eq:init_cond}), with the choice of $\delta_{\text{per}}=0.1$. It is observed that different values of $\delta_{\text{per}}$ in this regime result in the same long-time prediction for the steady phoretic velocity. At $\Pen>10$ the concentration profile computed at a previous (lower) value of the P\'{e}clet number is used as the initial condition for the subsequent simulation at the next higher \Pen.
The axisymmetry of the flow field restricts motion of the particle along the $z$ axis. From the computed time-series of the swimming speed, the particle position at a discrete time $i+1$ is evaluated as
\begin{equation}\label{eq:z_pos}
z({i+1})=z({i})+U(i)\Delta t;\,\,z(0)=0.
\end{equation}
In the time interval $[0,t_{\text{sim}}]$, a total of $N\equiv t_{\text{sim}}/\Delta t$ data points are recorded. The MSD and VAC, $C_v$, at the $n^{\text{th}}$ time interval are evaluated as follows
\begin{equation}\label{eq:msd_def}
\text{MSD}(n)=\dfrac{1}{N-n}\sum_{i=0}^{N-n}\left[z(i+n)-z(i)\right]^2,
\end{equation}
\begin{equation}\label{eq:vac_def}
C_{v}(n)=\dfrac{1}{N-n}\sum_{i=0}^{N-n}\left[\boldsymbol{e}_{u}(i+n)\bcdot\boldsymbol{e}_{u}(i)\right].
\end{equation}
where $\boldsymbol{e}_{u}(i)\equiv U(i)/|U(i)|$ is the unit direction vector of the self-propulsion speed at discrete time $i$, and can take the value of $+1$ or $-1$. Equations~(\ref{eq:msd_def}) and ~(\ref{eq:vac_def}) represent calculations of the time average over a single trajectory. The lag-time is defined as $\tau\equiv n\Delta t$ and used in reporting the MSD and VAC results. For $\Pen\geq50$, the total simulation time is at least $t_{\text{sim}}=5\times10^3$, while that for $\Pen=75$ and $\Pen=80$ are $t_{\text{sim}}=1\times10^4$ and $t_{\text{sim}}=2\times10^4$, respectively. Data points corresponding to at least the first $1\times 10^4$ timesteps in these runs are discarded prior to the calculation of the MSD and VAC, to remove the effects of transients.
\section {Results and discussion}
\label{sec:res_disc}
\begin{figure}[h]
\centerline{\includegraphics[width=4.8in,height=!]{fig1}}
\caption{ (Colour online) Steady self-propulsion speed of an autophoretic particle at various P\'{e}clet numbers across the quiescent, steady, and stirring regimes. The unfilled diamonds (\markerfive) represent results obtained by~\citet{Michelin2013}, and the unfilled circles are from the present work.}
\label{fig:steady_swim}
\end{figure}
\begin{figure*}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=2.5in, scale=0.8]{fig2a}&
\includegraphics[width=2in, scale=0.8]{fig2b}\\[5pt]
(a)&(b)\\
\includegraphics[width=2.5in, scale=0.8]{fig2c}&
\includegraphics[width=2in, scale=0.8]{fig2d}\\[5pt]
(c)&(d)
\end{tabular}
\end{center}
\caption{(Colour online) Steady-state concentration profile and streamlines of the flow around the autophoretic particle at $\Pen=20$ [(a),(b)] and (b) $\Pen=30$ [(c),(d)]. The flow and concentration fields are symmetric about the horizontal axis; therefore only half the particle is shown. The colour bar indicates the value of the solute concentration $c$.}
\label{fig:conc_sfn_map}
\end{figure*}
We analyze the dynamics of the autophoretic particle as a function of the P\'{e}clet number.~\citet{Michelin2013} have shown that the particle remains stationary until $\Pen=4$. This range, $0\leq\Pen\leq4$, may be termed as the quiescent regime. As the P\'{e}clet number is increased beyond 4, the fore-aft symmetry in the concentration profile is broken, and the particle executes steady rectilinear self-propulsion. In fig.~\ref{fig:steady_swim}, the steady self-propulsion speed $U$ from our computations is plotted as a function of \Pen, and it is seen that $U$ attains a maximum at $\Pen\approx9$, and decreases smoothly up to a value of $\Pen\approx26$. This range, $4<\Pen\leq26$, is classified as the steady self-propulsion regime. Furthermore, the good agreement of our results with data from~\citet{Michelin2013} establishes the validity of the numerical procedure used in the present work.
\begin{figure}
\centerline{\includegraphics[width=5in,height=!]{fig3}}
\caption{ (Colour online) Time evolution of the self-propulsion speed at (a) $\Pen=10$, (b) $\Pen=40$, (c) $\Pen=52$, (d) $\Pen=53$, (e) $\Pen=55$, (f) $\Pen=60$.}
\label{fig:u_vs_time}
\end{figure}
At $Pe\approx27$, however, there is another qualitative change in the motion of the active particle, as it experiences a drop of almost three orders of magnitude in its self-propulsion speed, becoming practically stationary and entering what we refer to as a stirring regime. The seemingly abrupt transition in particle motion may be understood by considering the relative growth rates of the various angular modes near the P\'{e}clet number at which the transition is observed. In fig.~1 of ~\citet{Michelin2013}, analytical solutions to the growth rates of the various unstable modes are plotted as a function of the P\'{e}clet number. Near $Pe\approx30$, the growth rates of the higher order modes are seen to outweigh the dipolar self-propulsion mode, and this could be a reason for the qualitative change in the dynamics of the particle from the swimming to the stirring regime. The transition may also be examined by a comparison of the concentration profile of the solute cloud, and the streamlines of the flow profile, at two different values of $\Pen$, as illustrated in fig.~\ref{fig:conc_sfn_map}. The solute distribution around the autophoretic particle is fore-aft asymmetric at $\Pen=20$, but is nearly fore-aft symmetric at $\Pen=30$ with a maximum concentration at the front and rear stagnation points. Furthermore, the streamlines in fig.~\ref{fig:conc_sfn_map}~(d) indicate a recirculation of the fluid around the autophoretic particle, where fluid is brought in along the polar axis and expelled at the equator. This flow pattern is consistent with the dominance of a quadrupolar disturbance to concentration profile. This region of parameter space $27\leq \Pen\leq 50$ is therefore analogous to the symmetric extensile pumping regime identified by~\citet{Morozov2019}, wherein the activity of the particle results in the stirring of the fluid around it, without resulting in its translation.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=5.5in,height=!]{fig4a}\\
(a)\\
\includegraphics[width=5.5in,height=!]{fig4b}\\[5pt]
(b)\\
\end{tabular}
\begin{tabular}{cc}
\includegraphics[width=3.2in, scale=0.8]{fig4c}&
\includegraphics[width=2.5in, scale=0.6]{fig4d}\\[5pt]
(c)&(d)
\end{tabular}
\end{center}
\caption{(Colour online) Time evolution of (a) the self-propulsion speed, and (b) displacement along $z-$ axis, for an autophoretic particle at $\Pen=75$. The complete simulation output in the window $t\in\left(3000,8000\right)$ is saved at intervals of fifty dimensionless time units. The time-averaged concentration profile and streamlines of the flow around the autophoretic particle in this window are shown in (c) and (d), respectively.}
\label{fig:pe75_analysis}
\end{figure}
In figure~\ref{fig:u_vs_time}, the instantaneous self-propulsion speed of the autophoretic particle is plotted as a function of time for various values of $\Pen$. The swimming speed in the steady regime, as seen from fig.~\ref{fig:u_vs_time}~(a), settles to a constant value of $\textit{O}(10^{-1})$ following an initial transient, while that in the stirring regime (shown in fig.~\ref{fig:u_vs_time}~(b)) is orders of magnitude lower. Beyond $\Pen=51$, an onset of oscillations in the swimming speed is observed. While these oscillations are transient and vanish at longer times for $\Pen=52$, they become more persistent at $\Pen=53$, as seen from figures~\ref{fig:u_vs_time}~(c) and (d), respectively. With a further increase in $\Pen$, the particle begins to move back and forth along the $z$-axis, the time scale for the reversal in self-propulsion direction decreases and the magnitude of the instantaneous self-propulsion speed increases, as seen from figures~\ref{fig:u_vs_time}~(e) and (f). The region $\Pen\geq55$ may be termed as the chaotic regime, characterized by short bursts of self-propulsion in arbitrary directions (along the $\pm z$-axis) and sharp changes in both the magnitude and direction of the self-propulsion speed.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=5.5in,height=!]{fig5a}\\
(a)\\
\includegraphics[width=5.5in,height=!]{fig5b}\\
(b)
\end{tabular}
\end{center}
\caption{(Colour online) Time evolution of (a) the self-propulsion speed, and (b) displacement along $z-$ axis, for an autophoretic particle at $\Pen=80$.}
\label{fig:pe80_analysis}
\end{figure}
The dynamics in the chaotic regime is explored further by considering a representative case of $\Pen=75$. The time series of instantaneous self-propulsion speed at $\Pen=75$ is shown in figure~\ref{fig:pe75_analysis}~(a). Following an initial transient that lasts till $t\approx3000$, the velocity settles into a characteristic pattern, with intermittent bursts of chaos that interrupt nearly regular oscillations of slowly varying amplitude~\citep{bpv1984}. The instantaneous position of the particle along the $z-$axis, evaluated using eq.~(\ref{eq:z_pos}), is plotted in fig~\ref{fig:pe75_analysis}~(b) as a function of time. After a transient period, the particle is seen to oscillate about a mean position of $z=-1$. The time-averaged concentration profile and streamlines of flow around the autophoretic particle, evaluated in the window $t\in\left(3000,8000\right)$, are plotted in figs.~\ref{fig:pe75_analysis}~(c) and (d), respectively. The concentration map is largely fore-aft symmetric, and this is consistent with the limited (net) mobility of the particle seen in (b). The pattern of the streamlines is qualitatively similar to that observed in the stirring regime (\textit{cf}. fig.~\ref{fig:conc_sfn_map}~(d)). A similar behaviour is observed at $\Pen=80$, following an initial transient period that lasts till $t\approx1500$. The intermittent bursts of chaos are more frequent, however, as observed in figure~\ref{fig:pe80_analysis}~(a). Due to the more frequent injections of this `turbulent' motion, the particle travels a larger distance than at $\Pen=75$, as seen from figure~\ref{fig:pe80_analysis}~(b). While the velocity and displacement time series of the autophoretic particle in the chaotic regime vary markedly depending upon the P\'{e}clet number, an analysis of the particle's mean square displacement and velocity autocorrelation reveals certain unifying features, as discussed below.
\begin{figure}[t]
\centerline{\includegraphics[width=5.0in,height=!]{fig6}}
\caption{ (Colour online) Mean square displacement of an active particle at various P\'{e}clet numbers.}
\label{fig:msd_mult}
\end{figure}
\begin{figure}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=3.1in,height=!]{fig7a}&
\includegraphics[width=3.1in,height=!]{fig7b}\\[5pt]
(a)&(b)\\
\end{tabular}
\end{center}
\caption{(Colour online) Velocity autocorrelation of an active particle (a) at $\Pen=80$, and (b) over a range of P\'{e}clet numbers.}
\label{fig:vac_comparison}
\end{figure}
In figure~\ref{fig:msd_mult}, the MSD of the particle is plotted as a function of the lag time for a range of P\'{e}clet numbers. At $\Pen=5$, corresponding to steady self-propulsion, the MSD grows as $\sim \tau^2$, since the particle moves rectilinearly at a constant speed. P\'{e}clet numbers in the range $27\leq\Pen<50$ are not examined since the particle is in a state of dynamic arrest, with only the fluid around it undergoing symmetric pumping, or stirring. In the chaotic regime, the MSD initially grows ballistically before transitioning to subdiffusive motion ($\sim \tau^{0.8}$). The timescale for the crossover in scaling is pushed to smaller values as $\Pen$ is increased. The subdiffusive exponent of $0.8$ is observed over a range of P\'{e}clet numbers between $55\leq\Pen\leq 80$, and thereby appears to be a universal signature of chaotic dynamics in the axisymmetric motion of autophoretic particles.
Anomalous diffusion (that is, MSD $\sim\tau^{\alpha}$ with $\alpha\neq1$) is observed in several contexts, including: the dynamics of colloidal tracers in entangled actin filament networks~\citep{Wong2004}; the motion of RNA tracers and DNA loci in living cells~\citep{Golding2006,Weber2010a}, and that of bacteriophages in mucus layers~\citep{Barr2015}; and the self-diffusion of sticky polymers in associated networks~\citep{MahmadRasid2021}. The two most commonly attributed reasons for this phenomenon are fractional Brownian motion or continuous time random walk~\citep{Bouchaud1990,Burov2011,Weber2010}. While a particle subject to fractional Brownian motion evolves its position in time due to a Gaussian noise with a long-range correlation, that undergoing a continuous time random walk process takes steps of random length, with the waiting time between jumps also sampled from a probability distribution. Neither of these two stochastic models, however, directly describe the subdiffusive motion in our computations of an autophoretic particle, which is deterministic and free from external noise. ~\citet{Geisel1984} have shown that discrete maps of the form $x_{t+1}=x_{t}+hx^{\mu}_{t},\,\text{where}\,h>0,\,\text{and}\,\mu>1$, result in chaotic dynamics, and can exhibit subdiffusive behavior upon tuning the $\mu$ parameter. This occurrence of subdiffusion in the chaotic dynamics of a deterministic system is in line with our computations.
In fig~\ref{fig:vac_comparison}~(a), the velocity autocorrelation for an autophoretic particle with $\Pen=80$ is plotted as a function of time. Rapid changes in the direction of self-propulsion result in a decorrelation of the signal, which crosses zero at $\tau\approx15$. This is followed by the appearance of negative dips and oscillations in the VAC: these features have also been observed experimentally by~\citet{Suda2021} and~\citet{Hokmabad2021}, as discussed below in detail. The variation in the VAC with $\Pen$ is illustrated in fig~\ref{fig:vac_comparison}~(b), where the data for $\Pen\geq55$ have been shown only till the first instance of their crossing zero, for clarity. A decrease in the
P\'{e}clet number from $\Pen=80$ results in less rapid changes in the particle motion, which is reflected in the longer time required for the velocity to decorrelate. Finally, at $\Pen=5$, the VAC is constant in time, indicating a complete correlation in the velocity, due to the persistent unidirectional translation of the particle. The timescale for the loss in correlation with a variation in $\Pen$ corresponds approximately to the time at which a crossover in the MSD scaling in figure~\ref{fig:msd_mult} is observed. Such a decrease in the correlation time with an increase in $\Pen$ has also been reported by~\citet{Chen2021}, who performed 3D simulations of a phoretic particle using the immersed boundary method. That study, however, does not report a zero crossing in the VAC for the range of lag times investigated.
We note a qualitative similarity between the trends in figures~\ref{fig:msd_mult} and ~\ref{fig:vac_comparison}, and the experimental findings on active droplets suspended in an aqueous surfactant solution within a Hele-Shaw cell~\citep{Hokmabad2021}. In their experiments, the droplets undergo self-propulsion due to a Marangoni stress imbalance at the interface, driven by a reaction between the droplet contents and the surfactant, resulting in the generation of oil-filled micelles. The active droplets undergo a transition from ballistic motion to anomalous diffusion with an increase in the P\'{e}clet number, which is also accompanied by a faster decay in the velocity autocorrelation. The droplets are found to avoid the chemical trail that they generate, and undergo a two-dimensional self avoiding walk, with MSD $\sim\tau^{3/2}$. In our model, the particle is restricted to motion along a line, where the only route for trail avoidance is a reversal in the direction of self-propulsion. The long-time MSD, therefore, grows less rapidly than the experimental system where the particle may move in two dimensions to avoid the self-generated chemical gradient.
\section {Conclusions}
\label{sec:conclusion}
Using a spectral element method to solve for the velocity and concentration fields around a rigid, spherical autophoretic particle in axisymmetric translation, we have demonstrated that the transition to chaotic dynamics proceeds through quiescent, steady, and stirring regimes, as the chemical activity ($\Pen$) of the particle is increased. The motion of the particle in the chaotic regime is analyzed using the mean square displacement and the velocity autocorrelation, and the dynamics is observed to be subdiffusive in the long-time limit (that is, MSD$\sim\tau^{\alpha}$, with $\alpha<1$) and not ballistic. Our findings thereby answer a question posed by ~\citet{Morozov2019} on whether axisymmetric computations of autophoretic particles could give rise to non-ballistic long-time dynamics. In fact, the value of $\alpha$ is found to be universal in the chaotic regime ($55\leq Pe\leq80$), with higher values of the P\'{e}clet number resulting in a faster decay of the velocity autocorrelation.
It would be natural to extend our analysis to particles with a finite value of the $m$ parameter, that is, where both diffusiophoresis and Marangoni flow contribute to self-propulsion. Here, one could examine if the scaling exponent of the MSD is dependent upon the value of $m$. This in relation to the findings by~\citet{Morozov2019}, which suggest that some amount of diffusiophoresis is needed in order to induce the transition to chaos, thereby implying that only steady translation is observed for purely Marangoni propulsion, $m=0$.
The simulations in the present work are restricted to the axisymmetric case, where the particle can move along a single dimension. Active droplets swimming in 1D have been studied experimentally~\citep{DeBlois2021}, albeit in the presence of a confining geometry.~\citet{Picella2022} have recently investigated the motion of autophoretic particles under confinement in a tube using fully 3D simulations, and observe rectilinear steady self-propulsion up to $Pe<15$. Should a transition to chaotic dynamics occur at larger $\Pen$, it may be that the effect of the confinement in rendering the propulsion effectively one-dimensional leads to subdiffusion as predicted here. Experiments of \citet{Hokmabad2021} on active droplets in a Hele-Shaw geometry reveal that the droplets undergo a self avoiding walk ($\alpha=1.5$) at sufficiently large $\Pen$, while simulations of 2D active circular disks by~\citet{Hu2019} predict normal diffusion ($\alpha=1.0$). The present results suggest that the values of these scaling exponents could be universal within the regime of chaotic self-propulsion. Of course, it would be interesting to determine the scaling exponent for an autophoretic particle free to move in three dimensions, although such computations would be numerically expensive.
\noindent\textbf{Acknowledgements.} R.K. thanks Nicholas Chisholm for detailed discussions regarding the numerical solver.
\noindent\textbf{Funding.} We gratefully acknowledge the support of the Charles E. Kaufmann Foundation of the Pittsburgh Foundation.
\noindent\textbf{Declaration of interests.} The authors report no conflict of interest.
| 9f66646fe8f29a00fcbbfd2edb642ad84bd4a516 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The most common approach in neural lossy video compression is to train a VAE-like network to minimize the expected rate-distortion (R-D) objective $D + \beta R$. Although this approach has proven to be very successful~\cite{habibian2019video, minnen2018joint, ssf, elfvc, lin2020mlvc}, a model trained to minimize expected R-D uniformly over all pixels may allocate too few bits to salient regions of a specific video.
This clashes with the model of the human visual system, which is space-variant and has the highest spatial resolution
at the the foveation point~\cite{wandell1995foundations,itti2004automatic}.
Exploiting this phenomenon by encoding regions-of-interest (ROI), e.g., based on saliency, objects-of-interest, or visual foveation, with higher fidelity can significantly contribute to the subjective quality under a low bitrate regime.
\input{figures/spotlight}
The key idea of \emph{ROI-based coders}~\cite{sikora1995shape, li2000shape, chen2006dynamic, han2008object, li2018learning, cai2020endtoend, xia2020object} is
to allocate different bitrate budgets for objects or regions-of-interest (ROI) and allow non-uniform reconstruction qualities, targeting use-cases like teleconferencing.
Traditional codecs like JPEG2000~\cite{skodras2001jpeg} and MPEG-4~\cite{vetro1999mpeg4}, which support ROI-based coding, were used as basis to build object-based coding methods~\cite{han2008object, chen2006dynamic}.
However, they lacked adoption
due to their high complexity and inherent incapacity to model arbitrary shapes due to their block-based approach. More recently, some works have developed ROI-based neural \emph{image} codecs. Some implicitly learn the ROI~\cite{li2018learning, cai2020endtoend} using a sub-network in the encoder, while others~\cite{xia2020object} cascade ROI retrieval using an external algorithm and use it for ROI-coding. Both approaches typically weigh the R-D objective spatially and~/~or mask the latent before quantization to reduce its entropy~\cite{habibian2019video, cai2020endtoend, xia2020object}.
Nevertheless, existing neural ROI-based codecs~\cite{li2018learning, cai2020endtoend} have the following limitations:
(1) they only work for images, (2) they use intricate masking schemes to control the rate spatially which do not exploit the Gaussian prior structure of the latent and
(3) the encoding operations are tightly coupled with ROI prediction, which makes it hard for the codecs to be adapted to different ROI requirements.
In this paper we present two methods for ROI-based neural video compression \ie, using ROI information to allow non-uniform reconstructed quality in the video frames: an implicit method and a latent-scaling based explicit method.
To the best of our knowledge, this is the first work applying neuralROI-based coding to video.
We propose two solutions and demonstrate their effectiveness on top of Scale-Scale Flow~\cite{ssf}, a popular neural video codec architecture. Even though the solutions are described in the context of SSF, they are in principle compatible with any solution that uses factorized hyperprior autoencoder (AE) based architectures, which encompasses a wide collection of state-of-the-art codecs
like ELF-VC~\cite{elfvc}, M-LVC~\cite{lin2020mlvc} or B-EPIC~\cite{pourreza2021extending}.
The first implicit model has access to the ROI mask and is trained with an ROI-aware loss, where the distortion of the background is de-emphasized. Secondly, the latent-scaling model extends the implicit model by exploiting a recent technique originally developed for variable rate coding~\cite{cui2020g,chen2020variable,chen2019neural,plonq}.
We extend it
by designing an auxiliary AE that takes as input the ROI map, and outputs a gain tensor with the same dimension as the latent variables of the main AEs,
allowing explicit control of the quantization binwidth for different spatial regions.
This can be seen as the continuous equivalent of the masking scheme used in conjunction with scalar quantization~\cite{li2018learning, mentzer2018conditional, cai2020endtoend}.
In this work we assume the availability of the ROI and decouple the ROI retrieval from coding as in Xia \etal~\cite{xia2020object} as it
allows for a generic ROI-based codec, which can be adapted at inference time to unseen use-cases. In contrast, methods which either use the semantic segmentation classes~\cite{agustsson2019extreme, akbari2019dsslic} or implicitly learn the ROI~\cite{li2018learning, cai2020endtoend} hardly adapt to out-of-distribution data without retraining.
We show that
(i) our methods outperform all our baselines on the DAVIS dataset~\cite{davis} in terms of R-D performance, as measured in PSNR in the ROI (Fig.~\ref{fig:spotlight}).
Moreover, (ii) they generalize to any arbitrary ROI which can be specified by the user at inference time. In addition,
(iii) one does not require obtaining expensive pixel-dense annotations during training, as synthetic ROI can be used, with little to no degradation in performance.
\section{Related work}
\begin{figure*}[t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{ccc}
\includegraphics[height=7cm]{images/sssf_concat_v2.pdf}&&
\includegraphics[height=7cm]{images/sssf_latent_scaling_v2.pdf}\\
(a) Implicit ROI Scale-Space Flow && (b) Latent-scaling ROI Scale-Space Flow
\end{tabular}}
\caption{Illustration of the proposed ROI-based neural video compression models.}
\label{fig:models}
\end{figure*}
\paragraph{Neural video compression.} Compressing videos with neural networks has been an active field of research in recent years~\cite{wu2018video, habibian2019video, dvc, rippel2019lvc, golinski2020frae, lin2020mlvc, ssf, elfvc, pourreza2021extending}. While varying in their choice of architecture and quantization strategy, neural video codecs generally follow the DVC~\cite{dvc} framework where an I-frame codec compresses the first frame in a Group-of-Picture (GoP) independently, and a P-frame codec uses motion estimation
(most commonly in the form of optical flow)
and a residual network to model the subsequent P-frames using previous reconstructions. Recently, Agustsson \etal~\cite{ssf} proposed to use a scale-space flow which addresses uncertainties in motion estimation via interpolation through a Gaussian pyramid. This allows blurring of the warped frames in regions where optical flow prediction is uncertain or ill-posed, like chaotic motions and obstructed objects.
While our work is based on SSF, it potentially translates to any method~\cite{elfvc, pourreza2021extending} using a mean-scale hyperprior~\cite{minnen2018joint} autoencoder as base quantization and entropy modeling scheme
, as it exploits the recent body of work on latent-scaling~\cite{chen2019neural, chen2020variable, cui2020g, plonq}.
Latent-scaling~\cite{chen2019neural, chen2020variable, cui2020g, plonq} was originally introduced to allow variable bitrate coding by essentially learning to adjust the quantization step size of the latents.
It is the continuous equivalent of mask-based methods like~\cite{li2018learning, mentzer2018conditional} which were used with scalar quantization with a fixed size of learnable codebook.
After training, the interpolation of the learned scaling parameters can lead to continuous variable bitrate performance.
\paragraph{Object-Based non-uniform coding.} The human visual system is space-variant and the spatial acuity is highest at the foveation point~\cite{wandell1995foundations, itti2004automatic}. Exploiting this phenomenon, by encoding objects at the foveation point in the scene with better quality, can significantly contribute to the subjective quality. Spatially non-uniform image encoding has been investigated to allocate different bitrate budgets for objects or regions of interest with non-uniform reconstructed qualities.
The literature can be broadly separated into two categories: (a) automatic ROI retrieval and (b) using the retrieved ROI to improve coding. Most traditional block-based coding methods~\cite{han2006image, han2008object, chen2006dynamic} fall under the first category. They exploit non-uniform coding capabilities of standard image codecs like JPEG2000~\cite{skodras2001jpeg} and video codecs like MPEG-4~\cite{vetro1999mpeg4}. These are limited in their capabilities due to their block-based approach to compression, which hinders the encoding of arbitrarily shaped objects and does not allow for pixel-level bit allocation optimisation~\cite{sikora1995shape, li2000shape}.
In contrast, recent work in neural image coding combine (a) and (b) and exploit pixel-level ROI~\cite{li2018learning, cai2020endtoend, duan2019contentaware, xia2020object} and semantic segmentation~\cite{agustsson2019extreme, akbari2019dsslic, duan2020jpad} information to improve perceptual quality in the low bitrate regime. Semantic segmentation restricts the codec to be used with the finite number of classes (car, road, tree) used during training, while ROI-based approaches are more generic, essentially only distinguishing foreground from background. Li \etal~\cite{li2018learning} and Cai \etal~\cite{cai2020endtoend} both learn the ROI implicitly by spatially masking out the latents before scalar quantization. Xia \etal~\cite{xia2020object} use the down-scaled output of the DeepLab~\cite{chen2018deeplab} segmentation network to mask out foreground from background, and send each stream to a separate hyper-codec for quantization.
Similar to them, our work focuses on (ii) \ie, how to use the ROI effectively for non-uniform coding of object at low bit-rate, but it extends neural ROI-coding to video.
In doing so, we focus on the coding perspective, and rely on groundtruth labels in segmentation datasets to derive ROIs.
We however remark that our codec can be deployed in pair with automatic ROI extractors~\cite{zhao2017pspnet, chen2018deeplab, le2018video, Wang_2018_CVPR, lai2019video, Wang_2019_revisitingVS, wang2021hrnet}.
\section{ROI-based neural video compression} \label{sec:methodology}
In this section we first present the neural video codec we use as backbone for our work, Scale Space Flow (SSF)~\cite{ssf}. Next, we extend SSF to be an ROI-based codec by proposing two models: the \emph{Implicit} and \emph{Latent-scaling} ROI SSF. Lastly, we will describe the optimization for SSF and the ROI-aware methods.
We define a video frame $x_i \in \mathds{R}^{H \times W\times 3}$ at time step $i$, where $H$ and $W$ represent its height and width respectively. Then a video sequence is denoted as $\mathbf{x}= \{ x_0, x_1, \hdots, x_{T} \}$, with $T+1$ frames. The sequence of binary ROI masks corresponding to the video sequence is defined as $\mathbf{s} = \{s_0, s_1, \hdots, s_{T}\}$, where $s_i \in \{0, 1\}^{H \times W}$.
The neural video codec SSF consists of an I-frame codec and a P-frame codec. The I-frame codec is a mean-scale hyperprior AE~\cite{minnen2018joint} which encodes individual frames $x_0$ independently to produce a reconstruction $\hat{x}_0$. The P-frame codec is comprised of two hyperprior AEs. The first, the \emph{P-frame flow hyperprior AE}, estimates a scale-space flow $g_i$ from the previous reconstruction $\hat{x}_{i-1}$ and current frame $x_i$, which is used to warp the previous reconstruction into $\bar{x}_i$. The second hyperprior AE, \emph{the P-frame residual hyperprior AE}, encodes the residual $r_i = \bar{x}_i - x_i$. The final reconstruction $\hat{x}_i$ is obtained by adding the warped prediction $\bar{x}_i$ and the estimated residual $\hat{r}_i$. The latent codes of each hyperprior AE are denoted by $z_0$, $w_i$ and $v_i$ and are rounded to integer values then entropy coded using the prior parameters estimated by their respective hyper-decoder. We omit hyper latent codes for ease of exposition, and we refer to~\cite{ssf} for further details.
\paragraph{Implicit ROI Scale-Space Flow} \label{sec:implicit_ssf}
An immediate extension to SSF to make it ROI-aware is to concatenate the current ROI mask $s_i$ as an additional input to each of the three hyperpriors, see Fig.~\ref{fig:models}a. Note that the ROI mask is not fed to the decoder, meaning we expect the encoders to implicitly store the relevant ROI information inside the existing latent codes. Since the decoder does not require the ROI mask, we do not need to transmit a representation of the mask itself. Albeit simple, we show the effectiveness of this approach when paired with an ROI-aware loss in Sec.~\ref{sec:results}.
\input{figures/latent_scaling}
\paragraph{Latent-scaling ROI Scale-Space Flow}\label{sec:latent_scaling_ssf}
Inspired by methods like~\cite{li2018learning, cai2020endtoend} which introduce a mechanism to explicitly control the spatial bit allocation, we adapted a recent technique called latent-scaling~\cite{chen2020variable, cui2020g}. Albeit similar in its motivation, it differs from the masking approach of~\cite{cui2020g} by exploiting the Gaussian prior structure of mean-scale hyperprior AE.
The key idea is to apply a scaling factor to the latent which changes the quantization step size, leading to different trade-offs between rate and distortion in ROI and non-ROI areas.
For ease of exposition, we will describe in the next paragraphs latent-scaling for the I-frame hyperprior AE, but the same method is applied to the P-frame residual hyperprior AE. We do not apply it to the P-frame flow hyperprior AE as initial studies showed the flow code $w_i$ only accounts for a small fraction of the total rate.
For the similar reasons, we only apply latent-scaling latents, leaving hyper-latents, which are cheap to encode, unaffected.
We introduce a new hyperprior-like network called \emph{gain hyperprior AE} (see leftmost autoencoder in Fig. \ref{fig:models}b). This network encodes the ROI mask $s_0$ into a latent code $z^s_0$, that is decoded to a gain variable $h_0$ which shares the same dimensions as the latent variable $z_0$, both spatially and channel-wise \footnote{previous latent-scaling~\cite{chen2020variable, cui2020g, plonq} work only use channel-wise gain}.
We scale the latent $z_0$ with the inverse of the estimated spatial gain variable $h_0$,
where we restrict $h_0 \geq 1$.
Such a procedure is akin to making the quantization range larger, depending on the value of $h_0$.
We further denote the mean $\mu$ and scale $\sigma$ as the prior parameters estimated by the I-frame hyper-decoder. In the quantization step, we choose to center the scaled latent $z_0 \oslash h_0$ by its prior mean $\mu \oslash h_0$, where $\oslash$ is a elementwise division. Next, we apply the rounding operator $\round{\cdot}$ on $(z_0 - \mu) \oslash h_0$ such that the estimated mean $\mu$ learned by the hyper-encoder is on the grid, and then add the offset $\mu \oslash h_0$ back. The dequantized latent $\hat{z}_0(h_0)$ is obtained by multiplying by $h_0$ after the quantization block. More precisely:
\begin{equation}
\hat{z}_0(h_0) = \round{(z_0 - \mu) \oslash h_0} \odot h_0 + \mu,
\end{equation}
where $\odot$ denotes elementwise multiplication.
After the dequantized latent $\hat{z}_0(h_0)$ is obtained, it is passed to the decoder to obtain reconstructed frame $\hat{x}_0$.
The whole procedure is illustrated in Fig.~\ref{fig:latent_scaling}a.
For rate computation and entropy coding, we use the modified probability $\mathbb{P}$ of $\hat{z}_0(h_0)$ as follows:
\begin{align}
\mathbb{P}\left(\hat{z}_0(h_0)\right) &= \int_{\hat{z}_0(h_0) - h_0/2}^{\hat{z}_0(h_0) + h_0/2} \mathcal{N}(x - \mu|0, \sigma) dx
\label{eq:change_grid} \\ \label{eq:change_prior} =& \int_{\hat{z}_0(h_0)/ h_0 - 1/2}^{\hat{z}_0(h_0)/ h_0 + 1/2} \mathcal{N}\left(x - \frac{\mu}{h_0}{\Big|}0, \frac{\sigma}{h_0}\right) dx
\end{align}
As shown in Fig.~\ref{fig:latent_scaling}b (rightmost plot) and Eq.~\eqref{eq:change_grid}, latent-scaling can be interpreted as effectively changing the quantization grid ~/~ binwidth. In practice, for entropy coding we do not change the quantization grid and round to the integer grid and scale the prior appropriately, as in Fig.~\ref{fig:latent_scaling}a and b (middle plot) and Eq.~\eqref{eq:change_prior}. As stated above, the same procedure is applied to the P-frame residual latent code $v_t$, as shown in Fig. \ref{fig:models}b.
\paragraph{ROI-aware R-D Loss}
We modify the regular R-D loss from SSF to take into account the ROI mask. We sum the rate and distortion for all $T$ frames in the video sequence $\mathbf{x}$ with corresponding ROI masks $\mathbf{s}$:
\begin{equation}
\label{eq:rd_loss}
\mathcal{L} = \beta \mathcal{L}_{R} + \sum_{i=0}^{T} \mathcal{L}_{D, i},
\end{equation}
where $\beta$ is rate-distortion trade-off variable. $\mathcal{L}_D$ represents the distortion loss which is a modified mean squared error (MSE) involving the binary ROI mask:
\begin{align} \label{eq:bin_mse}
\mathcal{L}_{D, i} = \operatorname{mean}\left(s_i \odot \epsilon_i + \frac{1}{\gamma} \cdot (1 - s_i) \odot \epsilon_i\right),
\end{align}
where $\operatorname{mean}$ computes the average over the image dimensions, $\gamma$ is a penalty hyperparameter for the non-ROI, $ \epsilon_i = (x_i - \hat{x}_i)^2$ is the squared error
and $s_i$ is broadcasted over the channel dimension. Note that the distortion loss of the original SSF corresponds to the special case where $s_i$ equals one everywhere.
Further, the rate loss $\mathcal{L}_R$ is computed with the estimated cross-entropy $\mathcal{H}(\cdot)$ by the hyperprior of each latent variable present in the model. For the implicit ROI SSF the rate loss $\mathcal{L}_{I,R}$ is equal to:
\begin{equation}
\label{eq:rate_implicit}
\mathcal{L}_{I,R} = \mathcal{H}(z_0) + \sum_{i=1}^{T} \left[ \mathcal{H}(v_{i}) + \mathcal{H}(w_{i}) \right].
\end{equation}
The rate loss $\mathcal{L}_{LS,R}$ of the latent-scaling ROI SSF also includes latent variables $z^s_0$ for the latent scaling of the I-frame hyperprior AE and $v^s_i$ for the latent scaling of the P-frame residual hyperprior AE.
As such, it is given by:
\begin{equation}
\label{eq:rate_ls}
\begin{split}
\mathcal{L}_{LS,R} =
& \mathcal{H}(z_0^s) +
\mathcal{H}(z_0) \\
& +
\sum_{i=1}^{T} \left[ \mathcal{H}(v_{i}^s) + \mathcal{H}(v_{i}) + \mathcal{H}(w_{i}) \right].
\end{split}
\end{equation}
In practice we found that the two extra rate contributions from the ROI masks $\mathcal{H}(z^s_0)$ and $\mathcal{H}(v^s_i)$ are only a small fraction compared to the standard rate components $\mathcal{H}(z_0)$ and $\mathcal{H}(v_i)$ of the model.
Please note that in both Eq.~\ref{eq:rate_implicit} and~\ref{eq:rate_ls} we omit the rate of the hyper latent codes to avoid notational clutter.
\section{Experimental setup}
\label{sec:experiments}
\paragraph{Datasets}
Typically, video compression algorithms are evaluated on datasets where the source is uncompressed, such as HEVC~\cite{HEVC_dataset}, UVG~\cite{UVG} and Xiph~\cite{Xiph} sequences.
However, these datasets are not immediately suitable to evaluate ROI-based methods, as they lack segmentation annotations: as such, a choice for ROI-mask to employ for foreground and background coding is not trivial.
Therefore, we use publicly available video segmentation datasets, namely DAVIS~\cite{davis} and Cityscapes~\cite{cityscapes}, both for training and evaluation.
DAVIS is composed of 90 diverse and short video sequences representing anything from a motocross jumping to a goat walking in the mountain.
Every video is labeled with pixel-wise annotations of objects of interest.
We use 60 sequences for training and 30 for validation, comprising 4209 and 1999 frames respectively.
Frames are shared in JPEG compressed format and 90\% have a resolution of 1080p or above, with a minimum resolution of 720p. Cityscapes is composed of 2120 video sequences from dashcam of vehicles driving around German cities. 1885 sequences are used for training and 235 for validation, or 89248 and 15000 frames respectively.
As groundtruth segmentation labels are provided only at 1 fps, we extract semantic labels automatically for every frame by running the state of the art model in~\cite{nvidia_segm}.
All frames are of 2048x1024 resolution and although being shared as PNG they were extracted from compressed videos. Consequently, we resize frames from both DAVIS and Cityscapes to 720p to reduce compression artefacts using Pillow~\cite{clark2015pillow}.
\paragraph{ROI construction}
\label{sec:semantic_instances}
In the DAVIS dataset, a single frame can be labeled with multiple segmentation instances, referring to different objects.
To obtain a binary ROI mask, we select all labeled instances, and note them as ROI (1), while the unlabeled instances are notated as non-ROI (0).
Cityscapes provides a categorization of every pixel into one of 19 classes. We select the following classes as classes of interest: "\textit{vehicle}", "\textit{road}", "\textit{pedestrian}", "\textit{bicycle}", "\textit{motorcycle}" and label all pixels from these classes as ROI (1) while the remaining classes are labeled as non-ROI (0).
We refer the reader to App.~\ref{apd:roi_creation} for further details on ROI construction.
As an alternative to ground-truth ROI masks, we also train with synthetic masks
generated using Perlin noise~\cite{perlin_noise} (see Sec.~\ref{sec:results_synthetic_roi}).
The masks contain blobs that move over time to cover each of the video frames. This makes the ROI masks more realistic than independently sampled noise. Although Perlin noise is typically a continuous value in $[0,1]$, we binarize the noise with a threshold of $0.5$.
\paragraph{Training details}
We optimize
all methods but SSF
with the ROI-aware MSE as distortion metric, see Eq.~\eqref{eq:bin_mse}, and use $\gamma=30$ as penalty for the non-ROI areas (see App.~\ref{apd:additional} for results with $\gamma=10$).
Following the training scheme from~\cite{ssf, pourreza2021extending}, all models are warm-started from an SSF pre-trained on the Vimeo-90k dataset~\cite{vimeo90k} for 1M steps, then fine-tuned on the dataset of interest for 300K steps. We trained all models at various rate-distortion tradeoffs
with $\beta = 2^\alpha \times 10^{-4} : \alpha \in {0,1,...,7}$.
We use Adam optimizer with a learning rate of $10^{-4}$ with batch size 8. Each example in the batch is comprised of 3 frames (IPP), randomly cropped to $256 \times 384$.
The models take about 3 days to train on a single NVIDIA V100 GPU.
\paragraph{Evaluation method} We report video quality in terms of PSNR in ROI and non-ROI, where both are first calculated per-frame in the RGB color space, then averaged over all the frames of each video, and finally averaged over all the videos of a dataset. The results we report are based on
GoP size of 12 for consistency with other neural compression works~\cite{ssf, pourreza2021extending, dvc, lin2020mlvc}.
\paragraph{Compared methods}
We compare our method to our re-implementation of SSF
and two ROI-based baselines. The first, dubbed \emph{ROI-aware loss}, is SSF trained with our ROI-aware loss as described in Eq.~\eqref{eq:rd_loss}. While the codec is blind to the ROI, it is expected to implicitly learn it through the training objective, in a similar fashion as the semantic models in Habibian~\etal~\cite{habibian2019video}. The second method, dubbed \emph{OBIC SSF}, is based on a recent ROI-based neural image codec called LearntOBIC.
We use the same hyperprior structure where a shared codec produces a latent, which is then masked using the bilinear downscaled ROI mask and split between foreground (ROI) and background (non-ROI). Each masked latent is passed to a different hyper-codec for prior parameter estimation, before being re-combined by a simple addition. For similar reasons to Latent-Scaling ROI SSF, we only modify the I-frame and P-frame residual hyperprior AEs in this way. To enable a fair comparison, we train this architecture using our ROI-aware loss, which is slightly different from the formulation in Xia~\etal (where the rate loss of the foreground is de-emphasized).
\section{Results}
\label{sec:results}
In this section we investigate the R-D performances of the proposed models, assessing their capability to properly achieve ROI-based video compression.
We further investigate the potential for generalization to a new dataset and hence unseen ROI distribution.
Finally, we explore training on synthetically generated ROI masks and analyze the adaptability of our ROI-based codecs to different ROI instance during evaluation.
\paragraph{ROI-based coding}
In Fig.~\ref{fig:semantic_models}, we report the RD-plots of Implicit ROI SSF and Latent-scaling ROI SSF.
We compare our proposed models to the described ROI-aware loss and OBIC SSF baselines, as well as to a plain SSF model that does not involve any ROI-based compression.
For all compared models, solid lines and dashed lines correspond to RD curves in ROI and non-ROI regions respectively.
The figure shows several insights.
First, the plain SSF shows better compression results on non-ROI regions, that are seemingly easier to compress than ROI areas on DAVIS.
This result - that we hypothesize is due to the high degree of motion affecting foreground objects on the dataset - underlines that such a codec might be suboptimal, in that a typical user is likely to attend mainly ROI regions.
The ROI-based baselines we consider, namely ROI-aware loss and OBIC SSF, succeed in delivering a better tradeoff for foreground regions.
Overall, their performances seem comparable across the rate spectrum.
Interestingly, the separate hyperprior models envisioned by OBIC SSF for foreground and background barely outperforms a simple ROI-aware loss in our experiments.
Finally, the figure clearly shows the superiority of the proposed implicit and latent-scaling ROI SSF.
Indeed, their RD-curves performs on par with the mentioned baselines on background regions, while achieving a superior tradeoff for ROI regions.
In this respect, our latent-scaling based model seems to slightly outperform our implicit model in ROI areas, especially at higher bpps ($>0.1$).
\input{figures/psnr_bit_allocation}
Furthermore, we investigate the behavior of the proposed Latent Scaling ROI SSF codec in terms of spatial bit allocation and reconstruction quality.
Fig.~\ref{fig:psnr_bit_allocation} we show, on a reference validation frame from DAVIS, its pixel-wise bpp and PSNR as compared to the ones achieved by SSF.
For SSF, we observe that bit allocation and reconstruction quality are uniformly distributed over the image.
Differently, Latent Scaling ROI SSF model successfully focus both bpp and PSNR on the ROI.
Moreover, it is worth noting how, despite the fact latent scaling operates at the reduced resolution of the latents (resulting in block-wise bpp allocation), the PSNR of the reconstructed frame properly aligns with the ROI at pixel-level.
In Fig.~\ref{fig:qualitative} we shows a few qualitative compression results of our model, compared to SSF. More qualitative results can be found in App.~\ref{apd:qualitative}.
\input{figures/qualitative}
\input{figures/cityscapes_plot}
\input{figures/perlin_noise_plot}
\paragraph{Generalization}
We investigate the generalization capability of our proposed latent-scaling ROI SSF model to different data and regions of interest.
To do so, we evaluate a model trained on DAVIS on a separate dataset, Cityscapes.
We expect (at least) two main sources of generalization gap.
First, the videos in the two datasets depict very different content (\textit{data gap}), and differences in the acquisition settings may generate discrepancies in low-level image statistics and global motion\footnote{for instance, in Cityscapes the motion is dominated by the ego-motion of the camera, which is car-mounted.}.
Moreover, the ROI specification described in Sec.~\ref{sec:experiments} might impact training (\textit{ROI gap}).
Indeed, in DAVIS we consider any labeled instance as foreground, resulting in ROIs that belong to different semantic classes, and are typically affected by motion.
Somewhat differently, for Cityscapes we mark pixels of specific classes as ROI, that are therefore consistent among training examples.
To monitor both effects, we plot in Fig.~\ref{fig:cityscapes_plot} the RD curves of our latent scaling model and plain SSF, trained either on DAVIS or on Cityscapes, and evaluated on Cityscapes.
By monitoring the gap between the SSF model (blue lines) trained on DAVIS and the one trained on Cityscapes, we appreciate the former performs slightly worse than the latter, both for ROI and non-ROI areas.
This gives a sense of the severity of the data gap alone, as no ROI was employed during training whatsoever.
In order to assess the effect of the ROI gap, we examine the margin between the two trainings of Latent Scaling ROI SSF (pink lines).
Interestingly, we observe a similar edge as the one observed for plain SSF.
The fact that the performance gap does not increase significantly suggests that most of the discrepancy is still due to the data gap, and that our codec is barely susceptible to the nature of ROIs used during training.
We will follow up on this intuition in the next paragraph.
Finally, we observe that, when evaluated on Cityscapes ROI areas, the ROI-based model trained on DAVIS outperforms the SSF model.
This might suggest that, when interested in ROI-based compression on a target dataset, our codec trained on a different dataset might still be a better choice than its non ROI-based counterpart, even when the latter is trained on the target dataset itself.
\paragraph{Synthetic ROI masks}\label{sec:results_synthetic_roi}
In order to further investigate the sensitivity of our latent scaling based codec to the nature of ROIs used during training, we carry out an experiment where we train it using synthetically generated masks.
Specifically, we rely on the DAVIS dataset and we generate the ROI for every training clip randomly, by taking advantage of perlin noise~\cite{perlin_noise}, as specified in Sec.~\ref{sec:experiments}.
The resulting masks are temporally smooth, but do not correlate with the video itself.
In Fig.~\ref{fig:perlin_noise_plot} we plot the performance of such a model (in purple) against a model trained on regular semantic masks, obtained by manual annotation (in pink).
We emphasize both models are tested, on the validation set, on regular semantic masks of ROI objects.
Interestingly, the results show that a gap exists between the two models, but it is almost negligible.
This result confirms the intuition that our model is minimally affected by the nature of training ROIs.
Such a property is appealing, as it enables the codec to be trained even in the absence of precise pixel-level ROI masks - that are typically expensive to obtain - while still yielding competitive results on realistic use cases.
In App.~\ref{apd:perlin_noise} we repeat the experiment for the Implicit ROI SSF and for a different value of $\gamma$, reaching similar conclusions.
\input{figures/custom_foreground_barplots_dogs}
\paragraph{Inference time ROI selection}
We now evaluate the capability of our model to adapt to different ROI specifications in front of the same video to be compressed.
We remark that this trait is appealing as it would elect our model as general purpose, as the same trained model could be deployed for ROI-based compression in disparate use cases.
We also notice how this feature lacks in current works for neural codecs~\cite{habibian2019video,agustsson2019extreme}, as they typically commit to specific semantic classes during optimization and are trained such that their encoder would implicitly recognize and favor important regions.
On the contrary, our model is explicitly fed with a mask specifying the desired (non-)ROI areas, allowing to compress the same video differently, depending on the desired ROI specifications.
We select the \texttt{dogs-jump} sequence from DAVIS validation set, as it is labeled with more than one instance.
Instead of merging all instances into a single ROI mask (as we do in all other experiments), we compress the video multiple times, by considering different instances as ROI in different runs.
We
consistently monitor PSNR on all instances, and observe it is consistently higher in the region considered as ROI.
We represent these results color-coded in the barplots in Fig.~\ref{fig:custom_foreground_barplots_dogs}. This clearly shows that our codec can be used, at approximately the same bitrate, to improve reconstruction quality in any ROI of choice.
A qualitative representation of such a feature is represented in Fig.~\ref{fig:custom_foreground_qualitative}.
\input{figures/custom_foreground_qualitative}
\paragraph{Limitations}
Our main motivation for the Latent Scaling ROI Scale-Space Flow was to allow for inference-time single model multirate behavior for the largest rate model, without the need to re-train or to adapt the training scheme like in~\cite{cui2020g,elfvc} (similar to what was demonstrated in~\cite{plonq} for image compression).
This would make our ROI codec more practical to deploy by drastically reducing the number of parameters and allowing fine-grain control of the rate.
In our experiments in App.~\ref{apd:additional:multirate}, latent-scaling does indeed allow reducing the number of trained models required to cover the bitrate region of interest (0.05 to 0.5 bpp) from 8 to 3, whilst also improving the distortion in the high bpp region.
However, it does not allow for a fully multirate model (\textit{i.e.} a single model covering the whole rate spectrum), and it comes with an increase in implementation complexity with minor performance benefits over the simpler implicit ROI approach.
In addition, visual assessments highlighted how, in their current implementation, both ROI-based models can sometimes produce sharp quality transitions between ROI and non-ROI regions. The problem would probably be exacerbated if the ROI masks suffered both in terms of quality and in temporal consistency. Both of these issues may be overcome by using smooth masks during training and/or inference.
Finally, a user study would benefit the evaluation of subjective quality of the compressed videos and further confirm that higher fidelity in the ROI at the cost of fidelity in the non-ROI can lead to a net boost in perceptual quality.
\paragraph{Societal impact}
Concerning societal impact, we do not see immediate harmful applications of our method that might negatively affect any public. Note that because the ROI codecs depend on an ROI retrieval algorithm, the methods may suffer from (and potentially amplify) its biases and shortcomings.
\section{Conclusions}
We introduced two methods for ROI-based neural video compression.
We showed that our methods outperform all our baselines in terms of Rate-Distortion (R-D) performance in the ROI. Moreover, they can generalize to different datasets and to any arbitrary ROI at inference time. Finally, they do not require expensive pixel-level annotations during training, as synthetic ROI masks can be used with little to no degradation in performance.
\section{ROI creation}
\label{apd:roi_creation}
\input{figures/appendix_roi_creation}
In Sec~\ref{sec:experiments} we explained how we created binary ROI mask from ground-truth annotations. In Fig.~\ref{fig:roi_creation} we show visual examples of this process for the DAVIS (top) and Cityscapes (bottom) datasets.
\section{Additional results} \label{apd:additional}
\subsection{Quantitative results} \label{apd:gammas}
During our research we tested two different penalty term $\gamma$ for non-ROI distortion, as defined in Eq.~\ref{eq:bin_mse}, namely $\gamma=\{10, 30\}$. In Sec.~\ref{sec:results}, all results are shared with $\gamma=30$ for ease of exposition.
In this section we provide additional results with $\gamma=10$. We allow side-by-side comparison for all experiments of Sec.~\ref{sec:results} for each penalty $\gamma$. Finally, we provide an additional multirate analysis.
\paragraph{ROI-based coding}
In Fig.~\ref{fig:apd_a_all_models} we show all ROI-based models trained with $\gamma=\{10, 30\}$ on DAVIS and evaluated on DAVIS val, with SSF as reference.
As expected from our loss formulation, a smaller penalty $\gamma$ results in a smaller performance gap between ROI and non-ROI across all ROI-based methods.
Interestingly, both the ROI-aware loss and OBIC SSF baselines which are blind to the ROI mask seem to only allow higher PSNR in the ROI than in the non-ROI at low bitrate, namely $\leq 0.15$ bpp.
For $\gamma=30$, the ROI PSNR is consistently better than non-ROI PSNR across the entire rate spectrum.
The two methods may perform similarly as they are both blind to the ROI mask, \ie the encoding operation does not get the ROI mask as input, although OBIC SSF foreground and background hypercodecs do get ROI information as their input is the ROI masked latent. We hypothesize that it may be insufficient for the hyper-codec network to implicitly learn to scale the prior parameters, and does not allow the encoder to scale the latent.
\input{figures/appendix_B1_a}
\paragraph{Generalization}
In Fig.~\ref{fig:apd_c_cityscapes} we show the SSF and latent-scaling ROI SSF models trained on either DAVIS or Cityscapes and evaluated on Cityscapes val for both values of $\gamma = \{10, 30\}$. As expected from our loss formulation, for $\gamma=10$ latent-scaling ROI SSF exhibits a smaller gap between ROI PSNR and non-ROI PSNR than with $\gamma=30$. Yet, irrespective of $\gamma$, the same observation can be made: the ROI PSNR of latent-scaling ROI SSF trained on DAVIS is higher than SSF trained on Cityscapes. This indicates that when interested in ROI-based compression on a target dataset, our codec trained on a different dataset might still be a better choice than its non ROI-based counterpart, even when the latter is trained on the target dataset itself.
\input{figures/appendix_B1_c}
\paragraph{Synthetic ROI masks} \label{apd:perlin_noise}
In Fig.~\ref{fig:apd_perlin_noise1} we show the effect of using synthetic ROI mask during training instead of ground-truth annotations, for $\gamma=\{10, 30\}$.
In addition to the experiment in the main text, we not only show latent-scaling ROI SSF but also implicit ROI SSF.
We find that for each of our proposed models, training with synthetically generated masks results only in a minor performance drop, albeit slightly larger for the implicit model. Since the performance of our proposed ROI-based models seem to be minimally affected by the type of ROI masks used during training, one could train them without requiring expensive pixel-wise annotations. This allows training on a target dataset of interest which may be different from dataset with available annotations like DAVIS. Consider, for instance, cartoons instead of natural videos.
\input{figures/appendix_B2}
\paragraph{Inference time ROI selection} In Fig.~\ref{fig:custom_foreground_barplots} we show result of swapping the ROI mask between each available instance in two DAVIS val sequences, instead of merging all instances into a single ROI mask as we do in all other experiments. We show that in each case, the instance considered as ROI sees a boost of 5dB or more in PSNR. This showcases the versatility of our ROI-based methods, as the same trained model could be deployed for ROI-based compression in disparate use cases.
Note this experiment is akin to the one reported by Fig.~\ref{fig:custom_foreground_barplots_dogs}, but validated on two more sequences from the DAVIS validation set (``\texttt{pigs}'' and ``\texttt{gold-fish}'').
\input{figures/custom_foreground_barplots}
\paragraph{Multirate capabilities}\label{apd:additional:multirate}
We experimented with the "naive" latent-scaling technique described in Lu \etal~\cite{plonq}. With the use of a gain amplifier $\mathrm{ga}$, it allows navigating different R-D tradeoffs with a single trained model during evaluation. The gain variable $h$ output by the gain hyperprior AE is transformed using
\begin{equation}
\label{eq:gain_amplifier}
\tilde{h} = (h-1) \cdot \mathrm{ga} + 1
\end{equation}
before being used to scale the prior parameters and latent code, see Sec~\ref{sec:methodology} for details.
Note that the higher the $\mathrm{ga}$ value, the coarser the quantization grid becomes, which in return is cheaper to encode.
In Fig.~\ref{fig:apd_b_ls} we show the latent-scaling ROI SSF for different rate regularization coefficients $\beta$ with gain amplifier $ga=1$ in pink. In addition
we select three trained models ($\beta=\{0.0001, 0.0008, 0.0064\}$) and sweep the gain amplifier $\mathrm{ga} \in \{1,2,4,8,16,32,64\}$; such curves are represented in red, purple and brown, and marked as ``MR'' (multirate) in the plot.
The figure shows how, in general, the multirate curves can follow the baseline curve for several values of the gain amplifier, before falling below it.
This allows to cover the target bpp range with 3 trained models instead of the 8 originally achieved by separate trainings.
More specifically, for high bpps ($\beta=0.0001$) we observe favorable performance for low values of the gain amplifier, with a severe drop as $ga$ increases.
We however appreciate that for higher compression rates ($\beta=\{0.0008, 0.0064\}$) the MR curves closely resemble the one achieved by separate trainings.
This shows promise for training a single model to support multiple bitrate by following training schemes as proposed in Cui \etal~\cite{cui2020g}.
\input{figures/appendix_B1_b}
\subsection{Qualitative results} \label{apd:qualitative}
In this section, we provide additional visual results for several variants of the proposed ROI-based methods.
\paragraph{Different background penalty}
In Fig.~\ref{fig:qualitative_gammas} we report for frames from the DAVIS validation set the ROI-based encodings achieved by Implicit ROI-SSF and Latent Scaling ROI-SSF at different values of the background penalty $\gamma$ (Eq.~\ref{eq:bin_mse}).
Such an hyperparameter controls to which extent background distortion can be de-emphasized to achieve (under rate constraints) a better quality in ROI regions.
\paragraph{Training on synthetic ROI masks}
As validated in Fig.~\ref{fig:perlin_noise_plot} and Fig.~\ref{fig:apd_perlin_noise1}, our models can be trained even in the absence of pixel-level ROI masks, as synthetically generated ones can be used instead, with similar validation performances.
In Fig.~\ref{fig:qualitative_random_maps} we report some examples of encodings for comparable models, when trained either on synthetic or groundtruth masks.
The visual quality of the resulting encoded frames appears comparable, confirming quantitative measurements.
\input{figures/appendix_gammas}
\input{figures/appendix_random_maps}
\subsection{Runtime performance} \label{apd:complexity}
In table \ref{table:timings}, we benchmarked the runtime of SSF and Latent-Scaling ROI SSF on HD 720p inputs on an NVIDIA Tesla V100 and Intel CPU E5-1620 v4 @ 3.50GHz. We show timings in frames-per-second (FPS) for encode and decode operations: neural-network execution only, and together with entropy coding on CPU including data transfer, for I-frame and P-frame codec separately.
\bgroup
\begin{table*}[tbh]
\centering
\begin{tabular}{lccccccccc}
\toprule
& & \multicolumn{2}{c}{Encode} & \multicolumn{2}{c}{Decode} & \multicolumn{2}{c}{Encode (no EC)} & \multicolumn{2}{c}{Decode (no EC)} \\
& & I-frame & P-frame & I-frame & P-frame & I-frame & P-frame & I-frame & P-frame \\ \midrule
SSF & FPS & 3.5 & 1.7 & 3.8 & 1.8 & 378 & 192 & 682 & 340 \\ \midrule
\multirow{2}{*}{LS ROI SSF} & FPS & 2.9 & 1.5 & 3.2 & 1.7 & 247 & 156 & 410 & 259 \\
& FPS drop & -17\% & -12\% & -16\% & -6\% & -35\% & -19\% & -40\% & -24\%
\\\bottomrule
\end{tabular}
\caption{Comparison of runtime (FPS) for 720p inputs of SSF and LS ROI SSF I/P-frame codecs on NVIDIA V100.
}
\label{table:timings}
\end{table*}
\egroup
\section{Architecture Details}
We use the same SSF architecture as described in Pourreza~\etal~\cite{pourreza2021extending}, Appendix A.1, except we share the hyper decoder for mean and scale, and last layer outputs twice as many channels. Our gain hyperprior autoencoder follows a similar architecture, except for the codec decoder which does not upsample and replaces transpose convolutions with regular convolutions with stride 1, see details in Fig.~\ref{fig:apb_gain_hyperprior_ae_codec} for the codec and Fig~\ref{fig:apb_gain_hyperprior_ae_hyper_codec} for the hyper-codec.
\input{figures/appendix_C} | ce8dcfdac01c0a2cefb1782ca7320e293c947503 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}
The Brazilian Supreme Court (\textit{Supremo Tribunal Federal} (STF) in Portuguese) is the highest court of law in Brazil. It is primarily responsible for guarding the rights in the Brazilian Constitution.
There are eleven justices, nominated by the President and confirmed by the Senate.
Among its multiple attributions, STF can provide a final judgment on appeals from other courts and a judicial review on some norms created in the country.
In 2004, Brazil had more than 100 million pending cases in the Judiciary\cite{cnjpendingcases}.
To reduce those numbers and avoid conflicting decisions, the National Congress amended the Constitution that same year, allowing STF to create binding precedents (Súmulas Vinculantes in Portuguese).
These precedents consolidate understandings on judicial issues that both Executive and lower Judiciary branches must follow.
From 2004 to 2020, the Court ruled 58 of those precedents, covering some of the most critical discussions from crime to tax law\cite{stfsumulas}. For example, binding precedent 37 addresses the issue of salaries of public servants and has the following statement:
\textit{``It is not up to the Judiciary, which has no legislative function, to increase the salaries of public servants, on the grounds of isonomy.''}
At first, the intention was to end the controversy. Still, the STF ends up later resolving thousands of cases that present a divergence on the application of a binding precedent (BP).
Consequently, STF's justices regularly cite these binding precedents in their decisions.
We consider a decision as a justice's ruling on a particular legal case, written and published on a legal document afterward. The terms ``decision'' and ``legal document'' are used interchangeably in this paper.
Given the large number of decisions in STF\,---\,more than 1 million from 2011 to 2020 \cite{stfnumberrulings}\,---\,, lawyers and other judicial experts face difficulties finding citations to BPs and analyzing them.
Such difficulties arise from the lack of a computational tool to help the experts to find, for instance, decisions that cite a particular BP of interest.
Moreover, it is usual to have decisions with hundreds of pages, making it hard for lawyers and other experts to find parts of interest where justices cite, quote, or implicitly mention a BP.
We worked alongside a domain expert to address these challenges to develop a web-based visual analytic tool named LegalVis.
We designed this system to explore and analyze the texts of STF's legal documents, particularly their explicit or implicit relationships with binding precedents.
In contrast to an explicit citation to a BP, which can be easily found by searching ``Binding Precedent \#X'' (in free translation) in the document's content, finding potential citations to a BP is not a straightforward task.
To tackle this issue, LegalVis relies on a machine learning model to identify potential citations, building upon an interpretability mechanism to provide reliable explanations for the model's decisions.
This system also provides mechanisms for quickly identifying documents and BPs of interest through multiple linked views and interactive tools.
These features permit easy analysis of the content of the documents and, at the same time, allow us to interpret the relationships that exist between similar documents and their relation with the BPs.
We demonstrate the potential of LegalVis through two usage scenarios that show insightful and relevant findings.
Our usage scenarios were validated by six domain experts not involved in the system's development, who also outlined the usefulness and potential of LegalVis.
Both components in LegalVis, the potential citation identification model and the visual analytics tool, could be applied to any court within the judicial system. We have chosen to deal with STF's decisions due to their relevance in the national context.
Our main contributions are:
\begin{itemize}
\item A pipeline for the identification of potential (\emph{i.e.}\xspace, non-explicit) citations that relies on (i) a machine learning model and (ii) an interpretability mechanism that provides reliable explanations for the model's decisions. Properly handling legal data, especially non-English data, is a sensitive task by its nature. Our pipeline and obtained results benefit this application domain and may inspire other applications.
\item LegalVis, a visual analytics system that assists lawyers and other judicial experts in exploring legal documents and binding precedents.
\item Two usage scenarios, validated by six domain experts not involved in the system's development, showing relevant judicial findings concerning STF's decisions.
\end{itemize}
\section{Conclusion}
In this paper, we presented LegalVis, a web-based visual analytic system designed to assist lawyers and other judicial experts in analyzing legal documents that cite or could potentially cite binding precedents. LegalVis first identifies potential citations by implementing a simple yet powerful machine learning model based on classification. Then, an interpretability mechanism also incorporated into the system provides reliable explanations for the model's decisions. Finally, all this information becomes accessible through the three interactive and linked views that compose the system. Qualitative and quantitative analyses validated the performance of the proposed model and two usage scenarios demonstrated the usefulness and effectiveness of LegalVis system. Both scenarios were validated by six domain experts not involved in the system's development, who also reported positive feedback.
\section{Related Work}
In this paper, we are interested in identifying and analyzing decisions somehow related to binding precedents, either by explicitly citing or implicitly mentioning them. In this context, a crucial step is to compute and visualize the similarity between the binding precedent and parts of a legal document. In the following, we discuss related research under two perspectives: visualization of similarities between parts of documents and works related to legal documents analysis.
\subsection{Visualizing Similarities between Parts of Documents}
Several text visualization techniques have been proposed for many tasks, domains, and data types. We point the reader to the surveys by Alharbi and Laramee~\cite{alharbi2018sos} and by Kucher and Kerren~\cite{kucher2015text} for a broader understanding of this topic.
One particular category of techniques focuses on visualizing document-level similarity, which is generally achieved by representing documents as points in a 2D or 3D visualization plane and positioning them accordingly~\cite{cao2016overview, alsakran2011streamit, van2008visualizing}.
Rather than just knowing whether two documents are similar to each other, one may be interested in exploring similarity at a fragment-level, \emph{i.e.}\xspace, by visually comparing parts (or fragments) of documents (\emph{e.g.}\xspace, paragraphs or sentences) to a reference text and highlighting their similarities.
A high similarity between the two parts is expected when there is text repetition, which is one of the text reuse possibilities~\cite{surveyTextAlignment}. Still, even parts that do not share many words may be similar to each other depending on the underlying semantics~\cite{use}. Either way, a visual analysis of these similar parts' contents is helpful in various contexts, as in tasks related to technical writing~\cite{soto2015similarity} and plagiarism identification~\cite{potthast2013overview,riehmann2015visual}. Examples include \textit{PicaPica}~\cite{riehmann2015visual}, a visual analytic tool that highlights differences and commonalities between two documents to help users in detecting plagiarism, and the \textit{Text Re-use Browser}~\cite{janicke2014visualizations}, which highlights pairs of similar sentences of the Bible via a dot plot matrix.
Different visualization techniques support the comparison between documents at a fragment-level, from
grid-based heatmaps~\cite{janicke2014visualizations, abdul2017constructive} to side-by-side views~\cite{riehmann2015visual}. They usually fall into one of two basic categories depending on the level of details on the raw text they provide~\cite{janicke2017visual}: while \emph{close reading} techniques allow the analysis of the text itself (words, phrases, ideas, and so on\,---\,see, \emph{e.g.}\xspace, \textit{Text Re-use Reader}~\cite{janicke2014visualizations}), \emph{distant reading} techniques\footnote{\emph{Close reading} and \emph{distant reading} are terms firstly created to categorize visualization techniques applied to digital humanities~\cite{cheema2016annotatevis}.} generate abstract representations that summarize the texts' information (see, \emph{e.g.}\xspace, \textit{Compare Cloud}~\cite{diakopoulos2015compare} and ViTA~\cite{ abdul2017constructive}). As stated by Jänicke et al.~\cite{janicke2017visual}, hybrid strategies that guide users from distant to close reading (top-down approach) successfully apply the well-established visual information-seeking mantra \emph{``overview first, zoom and filter, then details-on-demand"}~\cite{visualmantra}. Visual analytic tools usually rely on linked views to accomplish hybrid visualizations.
One such example is the system proposed by Kiesel et al. to compare argument structures in essays~\cite{kiesel2020visual}.
Our approach computes syntactic and semantic similarities (for explicit and potential citations, respectively) between parts of a document and the binding precedent being cited. To show such similarities intuitively and effectively, we rely on a hybrid approach that allows both distant and close readings in a top-down manner. Contrary to the mentioned systems, LegalVis is designed to handle Brazilian legal documents and particularities. We discuss studies related to the analysis of legal documents in the following section.
Finally, different visual encoding strategies may be employed to optimize the analysis of similar parts' contents when in close reading. One such choice is the use of colors (in the background or font color) to indicate common words or shared ideas~\cite{riehmann2015visual}. Other possibilities include changes in the font size, connections, and glyphs~\cite{janicke2017visual}. LegalVis employs the first option.
\subsection{Legal Documents Analysis}
In common law countries, such as the United States and Australia, lawyers and other judicial experts often search for precedents, looking at how old decisions could help solve open cases~\cite{zhang2007semantics}. In this context, automated text processing becomes a valuable tool that allows these professionals to search and analyze legal documents~\cite{barros2018case,legal_document_review}, especially when the task of interest involves finding relevant precedents or similar decisions~\cite{correia2019exploratory, mandal2017measuring}.
Given the many benefits of visual analytics, some authors have created visualization tools to help users explore such a high amount of legal data. An example is \textit{Knowlex}~\cite{lettieri2017legal}, a visual analytic system that links and exhibits Italian legal documents from different sources related to the user's law of interest. By employing citation networks, \textit{EUCaseNet}~\cite{lettieri2016computational} allows one to visually explore citations involving judgments from the European Court of Justice through network structural properties that will enable, \emph{e.g.}\xspace, recognizing relevant precedents and judges' behaviors. Visualizations applied to other countries' legal data also exist, for example, in Netherlands~\cite{wyner2017answering} and Portugal~\cite{carvalho2018transforming}.
Despite advances in the development of visual frameworks to assist in legal documents analysis, users are still challenged in exploring and visualizing legal data in Brazil. The Brazilian courts' websites and other official platforms usually exhibit copious pages of results, each containing several and long decisions\cite{barros2018case}. A few efforts have been made to enhance the visualization of these data. One of them is the \textit{Supremo 2.0} system~\cite{chada2015visualizing}, a tool designed to visualize STF's quantitative and multidimensional data. By working with a dataset similar to ours, Gomez-Nieto et al.~\cite{gomez2015understanding} proposed a visual analytic tool that allows exploring case-related aggregated information from STF through different representations (\emph{e.g.}\xspace, stacked graphs, treemaps, and heatmaps).
Our approach addresses a challenge not pursued by any other tool, even though some of them have characteristics in common with ours, such as linked views or the use of STF's data. LegalVis's primary goal is to provide a compelling exploration of legal documents that explicitly or potentially cite binding precedents. This involves an effective visualization and suitable machine learning and interpretability methods to identify and explain potential (non-explicit) citations. Not even those studies that also rely on STF's data~\cite{gomez2015understanding,chada2015visualizing} consider this type of citation information in their work.
\myparagraph{Other tools and search engines.} Various platforms developed search engines for legal documents and jurisprudence.
In Brazil, \textit{Buscador Dizer o Direito}~\cite{dizerDireito} presents commented documents, although the number of commentaries may be considered small, and \textit{OAB Juris~}\cite{oabjuris} uses artificial intelligence algorithms to present more relevant results. Other Brazilian platforms include \textit{JusBrasil}~\cite{jusbrasil} and \textit{LexML}~\cite{lexml}.
In the Netherlands, \textit{Bluetick}~\cite{bluetick} offers suggestions and similarity concepts during the search.
In general, however, these platforms do not offer efficient visualization or natural language processing (NLP) algorithms that could improve results and time during a user's search.
Some other tools are more elaborate, offering solutions based on analytics, visualization, machine learning, digitalization, and process automation. In Brazil, \textit{Finch Platform}~\cite{finchplatform} and \textit{Legal One}~\cite{legalone} are some of the most known tools, and, in United States, there are \textit{Lexis}~\cite{lexis}, \textit{Westlaw}~\cite{westlaw}, and \textit{Casetext}~\cite{casetext}. The last one, in particular, allows the use of Transformer-based models~\cite{transformer} to search for case laws that are similar to given sentences. As stated, none of these tools and search engines meets LegalVis's objectives.
\section{System Overview}
After several weekly meetings with a domain expert, we identified the challenges behind exploring legal documents and binding precedents.
We characterize a set of questions/requirements to be addressed by the analytical tool.
We present the system's requirements and the concrete visualization tasks that guided our visual design throughout this section.
We also provide an overview of the system's workflow.
\figWorkflow
\subsection{System Requirements}
After elucidating the domain experts' needs, we came up with several questions grouped into three categories.
The first category comprises questions related to the identification of decisions that cite or could potentially cite a binding precedent:
\myparagraph{Q1.1.} Which decisions in the STF (overall, per Justice rapporteur, and type of decision) explicitly cite a specific binding precedent?
\myparagraph{Q1.2.} Which decisions in the STF (overall, per Justice rapporteur, and type of decision) could potentially cite a specific binding precedent? In other words, which decisions end up citing a particular binding precedent in a non-explicit manner?
\myparagraph{Q1.3.} Are there decisions related to a to-be-created binding precedent?
\vspace{0.2cm}
The second category of questions are related to the filtering and analysis of relevant decisions:
\myparagraph{Q2.1} Which decisions cite a specific binding precedent on a given date?
\myparagraph{Q2.2.} Given a date, are there similar decisions that cite the same binding precedent?
\myparagraph{Q2.3.} Which are the most relevant decisions among those that cite a specific binding precedent on a given date regarding the similarity between binding precedent and decision?
Finally, the third category of questions addresses issues related to the similarity between parts of decisions and BPs and the relevance of those parts:
\myparagraph{Q3.1.} How similar is each part of a decision to the corresponding binding precedent?
\myparagraph{Q3.2.} Which parts of a decision cite or are likely to cite a specific binding precedent?
\myparagraph{Q3.3.} Given many decisions, can one quickly access multiple decisions of interest and their relevant parts when analyzing different dates and binding precedents?
\subsection{Design Tasks}
Based on the requirements described above, we raised concrete visualization tasks to guide the LegalVis system's development, following the mantra \emph{``overview first, zoom and filter, then details-on-demand"}~\cite{visualmantra}.
\myparagraph{T1 - Identification of explicit and potential citations:} The system should exhibit decisions that explicitly cite a specific binding precedent (Q1.1) and identify/exhibit those that could potentially mention it (non-explicit citation) (Q1.2, Q1.3).
\myparagraph{T2 - Overview of decisions and binding precedents:} The system should provide an overview of all the decisions and binding precedents, highlighting the chronological order they appear (Q1.3).
\myparagraph{T3 - Filtering and selecting decisions:} The system should allow filtering decisions by Justice rapporteur, by decision type, and by type of citation (explicit or potential) (Q1.1, Q1.2, Q1.3). It should also allow selecting decisions based on the date and binding precedent (Q2.1).
\myparagraph{T4 - Grouping and ordering decisions:} The system should identify and group similar decisions that cite the same binding precedent on a given date (Q2.2) and order them in the layout according to the similarity between binding precedent and parts of a decision (Q2.3).
\myparagraph{T5 - Highlight in decisions' relevant parts:} The system should quickly identify the most similar paragraphs/sentences to the binding precedent (Q3.1). It should also highlight those parts that (either explicitly or potentially) cite that precedent (Q3.2).
\myparagraph{T6 - Document browsing history:} The system must provide means to access decisions already analyzed quickly, regardless of the date and binding precedent (Q3.3).
\subsection{Workflow}
After mapping the requirements into design tasks, we define the system's workflow (see \figref{fig:workflow}).
In the first stage, \textbf{data collection (a)}, we consult the digitized decisions stored in a database.
They already contain the explicit citations that were found using regular expressions.
However, it is necessary to query and label the decisions and binding precedents that we will use in the following stages (details in \secref{sec:data_set}).
In the second stage, \textbf{identification of potential citations (b)}, we use the collected data to train a machine learning model that characterizes a citation to a particular binding precedent.
After that, the procedure infers potential citations and explains the reason behind the decision by employing an interpretability method (details in \secref{sec:potentialCitation}).
Finally, in the third stage, \textbf{visualization (c)}, users explore and visualize documents that explicitly cite or that can potentially be citing binding precedents.
LegalVis comprises three linked view components:
Global View, which presents an overview of the data under a temporal perspective;
Paragraph Similarities View, which allows for filtering and grouping relevant documents; and
Document Reader, which shows a document's content and points out which parts of the document are likely to mention a binding precedent (details in \secref{sec:visual}).
\section{Dataset}
\label{sec:data_set}
\figCitations
When an STF justice decides a case, this decision is consolidated in a document.
Each document contains the decision's text, publication date, the justice that conducted the process (justice rapporteur), and the document type (\emph{e.g.}\xspace, Complaint or Extraordinary Appeal).
In this work, we are interested in STF decisions that are somehow related to a binding precedent, explicitly or \emph{potentially} citing it.
In total, we have 58 binding precedents, idealized as a mechanism to create a consolidated understanding among the STF's justices.
We collected our dataset from a partnership with the \emph{Supremo em Números} (``STF in Numbers'', in free translation) project~\cite{falcao2013relatorio}\,---\,a project that seeks to assess legal and computer knowledge to produce unprecedented data on the Supreme Court.
This dataset contains more than 2,500,000 documents since 1988 and metadata such as the number of the BP being cited (if that is the case) and document type.
Considering only decisions that explicitly cite at least one of the 58 BPs, there are 38,364 documents, totaling 41,031 citations (some documents may mention more than one BP).
The number of citations per BP does not follow a uniform distribution, as we can see in \figref{fig:citations_bp}.
As requested by the domain experts, the name of the justice rapporteur is essential for the analysis. %
However, this information was not in the initial dataset, so we extracted it using regular expressions in the document's content.
As part of the data cleaning process, we had some difficulties: (i) documents with different titles but same content (we refer to them as \emph{duplicated} documents), and (ii) documents in which it is not trivial to extract out the justice name.
However, these problems did not impact our work's development, and we give more details in~\secref{sec:limitations}.
In this work, we decided to deal with documents that explicitly cite at least one of the ten most-cited BPs (\emph{i.e.}\xspace, 3, 4, 10, 11, 14, 17, 20, 26, 33, and 37).
We made this decision to have a representative sample of documents and citations. In fact, after ignoring the ``duplicated" documents and considering only documents that cite just one BP each, we gathered 29,743 documents and 31,070 citations related to these 10 BPs, which corresponds to 77.5\% and 75.7\%, respectively, compared to all documents and citations from the initial dataset.
Besides these 29,743 documents, we also consider other 30 thousand documents, without explicit citations, that we will use to identify potential citations (details in~\secref{sec:potentialCitation}).
Finally, our documents generally have less than four thousand words (we have a word count histogram in Appendix B).
\section{Identification of Potential Citations}
\label{sec:potentialCitation}
In the context of citations to binding precedents, relevant questions may arise.
For instance, \emph{``are there documents that potentially cite a BP?''} (\emph{i.e.}\xspace, the document should have cited the BP but it has not), or \emph{``are the document and the BP related enough?''}
These questions can help understand how STF works and if the BPs are correctly applied.
In this section, we investigate the possibility of potential citations and how to find them.
We describe our modeling of identification of potential citations as a classification task, and we discuss how to use classifiers to infer the citations.
Our pipeline for identifying potential citations (\figref{fig:potential_pipeline}) is composed of two main steps: (a) the learning process, when the models learn what makes a citation, and (b) the potential citation inference, when we use the models to identify a potential citation and an interpretability technique to explain this decision.
We give more details about these stages in the following.
\subsection{Learning Process}
\label{sec:learning_process}
Suppose there is a function $d_{X} \colon U \to \mathbb{R}$ that takes a document $D$ from our corpus $U$ and returns a score of citing the binding precedent $X$. When a document has a ``good'' score through $d_X$, we can assign it to a potential citation group. Motivated by this reasoning, the potential citation identification process starts searching for such a function $d_X$.
There are many ways to model the problem above. Two possible approaches are to consider the score $d_X$ as the distance between the document $D$ and the BP $X$ in some latent space or compute the probability of $D$ citing the BP $X$.
With these approaches, when the distance between $D$ and BP $X$ is small, or $D$ citing $X$'s probability is high, we say that we found a potential citation.
The distance between $D$ and BP $X$ can be computed by embedding $D$ and BP $X$ in a high dimensional vector space (\emph{e.g.}\xspace, Doc2vec~\cite{doc2vec}) and then calculating the distance in this space.
In the second case, we could compute the probability of a document $D$ citing BP $X$ by modeling the problem as a classification task.
If we fit a classification model in documents that cite the BP $X$, the model can learn what makes a document cite the precedent.
Therefore, we could employ the trained classifier to verify whether a new document cites the BP $X$, using the model's probability as a confidence level for this assignment.
In this work, we rely on the classification approach to identify potential citations to BPs.
The distance-based approach is similar to clustering with centers in the BPs, so the approach is unsupervised by nature.
On the other hand, the classification is supervised and leverages the labels of the other documents. It is also easier to identify a potential citation confidence threshold in terms of probabilities than distances (\emph{e.g.}\xspace, documents with model's probability greater than $t_c$ are assigned to a potential citation).
Moreover, it is more straightforward to apply interpretability techniques to classification models' decisions (\emph{e.g.}\xspace, using Lime \cite{lime} or Anchors \cite{ribeiro2018anchors}) instead of trying to interpret citation inference using vector representations.
\myparagraph{Data Pre-processing:}
Before training our models, we need to process our data and create a balanced sample dataset, further split into training, validation, and test datasets.
We generate the sample dataset by choosing a random sample of documents that cite one of the ten most-cited BPs.
As described in~\secref{sec:data_set}, we do not consider ``duplicated'' documents or documents citing more than one BP.
Avoiding duplicated documents is necessary to prevent the classifier from overfitting the data.
Having documents citing only one BP ensures they only have one label, making the classification task more straightforward.
We chose only to consider the ten most-cited precedents. The reasons for this decision are twofold.
First, the number of documents associated with less frequent BPs is not enough to train the models, demanding extra effort to remedy this situation, such as using a data augmentation scheme.
Second, the reduced number of documents associated with the less frequent BPs can quickly be inspected manually, so there is no need to undertake efforts to develop sophisticated resources to analyze them.
In total, our sample dataset contains 6,730 documents balanced among the ten classes.
Given that most documents explicitly contain the text ``Binding Precedent \#X'' (in free translation), we remove these citations using regular expressions to avoid hints during the classification process.
The documents also underwent a standard NLP text pre-processing, which included converting the text to lowercase, \emph{tokenization}, removing punctuation and stop words, and \emph{lemmatization}.
\myparagraph{Text Embedding:}
As mentioned above, we use a classification model to assess the probability of a document citing a BP.
To assess different classifiers' performance, we rely on text embedding methods to generate a vector representation of the documents.
We test different embeddings to evaluate their performance: TF-IDF~\cite{Robertson2004}, Doc2vec~\cite{doc2vec}, Universal Sentence Encoder (USE)~\cite{use}, and Longformer~\cite{beltagy2020longformer}.
In TF-IDF's particular case, we have also applied a dimensionality reduction procedure to map the documents to a 50-dimensional space using Truncated Singular Value Decomposition (SVD)~\cite{truncated_svd}.
We call the final representation Truncated TF-IDF.
The reason to consider Longformer, instead of a more popular language model, \emph{e.g.}\xspace BERT \cite{bert}, is that Longformer overcomes BERT's limitation of 512 tokens.
Note that USE and Longformer do not need a pre-processed text (as described in the previous paragraph) because both have their tokenization mechanism.
\figPipeline
\myparagraph{Classification:}
Support Vector Machines (SVM) are complex enough for real-world classification problems and simple enough to be analyzed mathematically~\cite{svm}.
Our study considers SVM with linear and Radial Basis Function (RBF) kernels. Moreover, given that Longformer~\cite{beltagy2020longformer} has a linear classification neural network layer plugged into it, we also fine-tuned it to assess its classification performance in our context.
The described classifiers are used to search for potential citations in unlabeled data, assigning probabilities to each document to belong to each class.
To get probabilities from an SVM, we need to make calibration: using labeled data, create a map from the classifier's output (SVM scores) to a probability estimate for each class, which sum up to 1.
To create this map, we use the calibration method from Platt et al.~\cite{platt1999probabilistic} for SVM with linear kernel and the Wu et al.~\cite{wu2004probability} method for RBF kernel.
Platt et al.'s method, precisely, does not support the multiclass case, so we calibrate each class in a ``one-vs-rest'' approach and normalize the results in the end to sum up to 1.
This map will receive the SVM scores from a data instance and output a probability estimate for each class.
\subsection{Potential Citation Inference}
\label{sec:pot-ite-inf}
This section describes how we use the classifiers to identify potential citations and understand the models' decisions (see~\figref{fig:potential_pipeline}~(b)).
\myparagraph{Citation Inference:}
The association between a text embedding technique and a classifier gives us what we call a \emph{model}.
A model receives the raw text and returns the classifier's probabilities, which we interpret as the document's probability to cite each precedent.
We consider a potential citation when the citation probability is greater or equal to a threshold $t_c \in [0,1]$ (chosen by the user).
\myparagraph{Interpretability:}
A concern when dealing with machine learning models is their interpretability.
For instance, if Truncated TF-IDF combined with SVM points out that a document potentially cites BP 10 because of high returned probability, how can we understand the reasons behind this decision? Generally speaking, \emph{how can we trust this model?}
Understanding why a model is taking a particular decision is of paramount importance~\cite{lime}, especially in sensitive scenarios like a legal document analysis.
Consider document $X$ and its embedding vector $\mathbf{x} \in \mathbb{R}^d$ (\emph{e.g.}\xspace, TF-IDF vector).
This document's probability of belonging to class $C$ is given by $f_C(\mathbf{x})$, with $f_C \colon \mathbb{R}^d \to [0,1]$ a model's returned probability for class $C$.
Our particular interest is to know the importance of each sentence from document $X$ to the given probability, \emph{i.e.}\xspace, if each sentence has positive, negative, or neutral importance, and the magnitude of this importance.
After preliminary tests have discarded the employment of the \emph{leave-one-out feature importance} (LOO)~\cite{loo} for interpretability due to its high sensitivity, we chose the \emph{Local Interpretable Model-agnostic Explanations} (Lime)~\cite{lime} method to tackle this issue.
The intuition is that, by removing a specific sentence and obtaining the variation of probability $\Delta f_C$, we can measure the importance of this sentence for the prediction $f_C(\mathbf{x})$. More formally, we randomly remove some sentences from document $X$, vectorize this new document to $\mathbf{z}$, and add it to set $Z$. Doing this many times, we have a collection $Z$ of vectors around $\mathbf{x}$ in high dimensional space.
We approximate $f_C(\mathbf{z})$ using a linear model $g(\mathbf{z})$, minimizing the approximation error $\mathcal{L}(f_C, g, \mathbf{z})$ in $\mathbf{z} \in Z$, but also constraining the complexity $\Omega(g)$ of $g$.
This task can be interpreted as a weighted linear regression with regularization. We end with a model $g(\mathbf{z})$ that is linear over the sentences (\emph{i.e.}\xspace, the presence or absence of a sentence), where the linear model's coefficients can be interpreted as the importance score of each sentence to the final prediction $f_C(\mathbf{x})$.
By default, Lime works with words, and there is also the possibility of interpreting entire paragraphs.
A word division brings some advantages, such as detailing.
Still, it does not immediately assess each sentence's importance to the decision (\emph{e.g.}\xspace, sentences similar to the BP text).
A paragraph division is also interesting for visualization purposes (see, for example, \secref{sec:document_reader}), but it merges various sentences into one, hampering the precise identification of relevant parts.
Therefore, our choice of working with sentences comprises the ``best of both worlds''. In our experiments, Lime proved to be very robust and reliable when dealing with sentences.
\subsection{Modeling Results}
\label{sec:modeling_results}
The modeling described in \figref{fig:potential_pipeline} (\emph{i.e.}\xspace, fitting classifiers to data, using the classifiers to infer citations, and interpreting these inferences) needs to be validated.
In this section, we evaluate the modeling process quantitatively and qualitatively.
\figConfusion
\myparagraph{Quantitative Analysis.}
We fit the classifiers with the text embedding to a training set corresponding to 80\% of our labeled, balanced sample dataset (randomly chosen, preserving balancing). Before applying SVM, each text embedding coordinate was standardized, improving SVM classification performance~\cite{standardizeSVM}. Then, we performed a grid search and cross-validation to optimize SVM parameters (linear and RBF). We also fine-tuned the Longformer model for four epochs, with a batch size of 12 and a linear decrease of the learning rate. In general, all models performed well in the validation set (10\% of the sample dataset, preserving balancing), especially TF-IDF-based models and Longformer, as depicted in \tabref{tab:accuracies}. In particular, Truncated TF-IDF with linear SVM has an excellent performance when predicting the correct BP on this set, as shown in the confusion matrix comparing the true with the predicted BPs from \figref{fig:confusion_matrix}.
Recall that the explicit mentions of the BPs were removed from the documents in the pre-processing step.
\begin{table}[t!]
\color{black}
\footnotesize
\centering
\caption{Evaluation of models' performance on validation portion of the sample dataset. Precision, recall, and F1 score are the average of each class' metrics weighted by number of class instances.}
\label{tab:accuracies}
\begin{tabular}{cc|cccc}
\textbf{Embedding} & \textbf{Classifier} & \textbf{Acc.} & \textbf{Precision} & \textbf{Recall} & \textbf{F1 score} \\ \hline
\multirow{2}{*}{Trunc. TF-IDF} & Linear & 0.94 & 0.94 & 0.94 & 0.94 \\
& RBF & 0.94 & 0.94 & 0.94 & 0.94 \\ \hline
\multirow{2}{*}{TF-IDF} & Linear & 0.94 & 0.94 & 0.94 & 0.94 \\
& RBF & 0.93 & 0.93 & 0.93 & 0.93 \\ \hline
\multicolumn{2}{c|}{Longformer} & 0.93 & 0.93 & 0.93 & 0.93 \\ \hline
\multirow{2}{*}{Doc2vec} & Linear & 0.84 & 0.85 & 0.84 & 0.84 \\
& RBF & 0.88 & 0.88 & 0.88 & 0.88 \\ \hline
\multirow{2}{*}{USE} & Linear & 0.85 & 0.86 & 0.85 & 0.85 \\
& RBF & 0.86 & 0.87 & 0.86 & 0.86 \\ \hline
\end{tabular}
\end{table}
\myparagraph{Qualitative Analysis.}
The models' quantitative results support the idea that searching for potential citations can be solved as a classification task.
Moreover, TF-IDF combined with SVM and Longformer presented pretty good results.
However, Longformer tends to assign probability close to 1 to a particular class and the remaining probabilities close to zero. This fact makes it hard to interpret the probabilities returned by Longformer as a confidence level.
In contrast, Truncated TF-IDF with SVM better distributes probabilities between 0 and 1.
This fact added to other properties\,---\,ease of implementation, low dimensionality, and validation performance\,---\,rendered Truncated TF-IDF with linear SVM the better choice in our context.
\myparagraph{Chosen model on test data.}
The chosen model (Truncated TF-IDF with linear SVM), which had good performance on validation data (\autoref{tab:accuracies}), also performed well on test data (10\% of a sample dataset, balanced), achieving 0.96 in all metrics of \autoref{tab:accuracies}.
\myparagraph{Identification Analysis.}
To search for potential citations, we run our model with a document collection containing 30,000 documents that do not cite any BP (\secref{sec:data_set}). For sanity check, we also run our model in documents that explicitly cite the ten chosen BPs.
Of course, when the model assigns a potential citation that is the same as the explicit citation, we continue calling it explicit citation.
The goal is to rely on the classification model to find potential citations as one of the ten chosen BPs.
The larger the user-defined threshold $t_c$ (\secref{sec:pot-ite-inf}, Citation Inference), the smaller the number of documents with potential citations, as presented in~\tabref{tab:potential_citations}.
\figVisualComp
\begin{table}[t!]
\footnotesize
\centering
\caption{Number of documents with potential citations of each BP found by Truncated TF-IDF, for three different $t_c$ values.}
\label{tab:potential_citations}
\begin{tabularx}{\columnwidth}{XXXXXXXXXXX}
\multirow{2}{*}{\textbf{$t_c$}} & \multicolumn{10}{c}{\textbf{BP}} \\
& \textbf{3} & \textbf{4} & \textbf{10} & \textbf{11} & \textbf{14} & \textbf{17} & \textbf{20} & \textbf{26} & \textbf{33} & \textbf{37}\\ \hline
\textbf{0.99} & 4 & 73 & 1 & 1 & 2 & 7 & 4 & 5 & 50 & 1 \\ \hline
\textbf{0.95} & 45 & 237 & 65 & 8 & 78 & 184 & 59 & 118 & 344 & 42\\ \hline
\textbf{0.90} & 124 & 374 & 476 & 85 & 212 & 524 & 88 & 368 & 469 & 119\\ \hline
\end{tabularx}
\end{table}
We use Lime (\secref{sec:pot-ite-inf}) to understand why the model assigned a potential citation to a document.
The local model fitted by Lime is a linear classifier. We can interpret this model's weights as the importance of each sentence to the model's decision, \emph{i.e.}\xspace, the potential citation assignment.
These weights were mapped into a color range in the document background to visually indicate which sentences are more important.
The validation of this assignment's interpretability, so as the potential citation assignment, is established in the usage scenarios discussed in~\secref{sec:case_studies}.
\section{LegalVis}
\label{sec:visual}
In this section we describe the visual components that compose LegalVis and also provide implementation details. As shown in~\figref{fig:visualComponents}, the system is made up of three main components, namely \textit{Global View}, \textit{Paragraph Similarities View}, and \textit{Document Reader}, that address the proposed design tasks as depicted in~\tabref{tab:views_tasks}.
\begin{table}[t!]
\footnotesize
\centering
\caption{Visual components of LegalVis and the addressed tasks.}
\label{tab:views_tasks}
\begin{tabular}{lllllllllll}
\multirow{2}{*}{} & \multicolumn{6}{c}{\textbf{Design Task}} \\
& \textbf{T1} & \textbf{T2} & \textbf{T3} & \textbf{T4} & \textbf{T5} & \textbf{T6} \\ \hline
\textbf{Global View} & \checkmark & \checkmark & \checkmark & & & \\ \hline
\textbf{Parag. Similarities} & \checkmark & & & \checkmark & \checkmark & \\ \hline
\textbf{Document Reader} & \checkmark & & & & \checkmark & \checkmark
\end{tabular}
\end{table}
\subsection{Global View}
\textit{Global View} (\figref{fig:visualComponents}~(a)) is the main and first opened panel. This view is responsible for showing an overview of the dataset under a temporal perspective (T2) and providing interactive tools that guide the user in searching for documents of interest (T3). Some documents in the dataset do not have valid date information, as will be discussed in~\secref{sec:limitations}. Since Global View depends on this attribute, these documents were potentially used to train the classification model (\secref{sec:potentialCitation}) but ignored in the visualization.
In this view, the $x$-axis represents the time a document was published (monthly resolution), and the $y$-axis represents the BP. The red pins (\img{pin.png}) refer to the publication date of the corresponding BPs. For each one, a vertical bar on a particular date indicates the existence of documents published on that date that cite such BP (T1). The height
of each bar reflects the number of documents published on that date, and its color is defined such that (i) blue bars (without borders) indicate
that every document in that month cites the BP explicitly, (ii) orange bars (without borders) suggest that every document potentially (rather than explicitly) cites the BP, and (iii) blue bars with orange edges indicate the presence of both explicit and potential citations.
For better understanding and exploration of regions of interest, users can \textit{zoom~in}, \textit{zoom~out} and \textit{pan}. Other interactions include filtering documents by
(i) the type of citation (explicit and/or potential); (ii) Justice rapporteur; (iii) document type (\emph{e.g.}\xspace, complaints or appeals); and
(iv) potential citation confidence (threshold $t_c$). Finally, a tooltip showing the number of the BP, the date, and the number of documents (total, with explicit citations, and with potential citations) appears when a bar has hovered over. The publication date of the BP is also shown alongside the red pin.
We considered other design choices for \textit{Global View}, for example, fixed-size squares with the number of documents mapped by a color-scale (a strategy similar to the \textit{Temporal Activity Map}~\cite{TAM}) and varying-size circles instead of vertical bars. Since color coding was already considered the more suitable strategy to represent explicit/potential citations in this view (and in \textit{Paragraph Similarities View}, as we show in the next section), and given the high level of overlaps and cluttering obtained with varying-size circles, mapping the number of documents through bars with varying heights was preferred.
\subsection{Paragraph similarities View}
Once the user has found a set of documents (\emph{i.e.}\xspace, a bar) of interest in \textit{Global View}, he/she can select such a set by clicking on the bar. The user is then
redirected to the \textit{Paragraph Similarities View} (\figref{fig:visualComponents}~(b)), which presents in the $y$-axis all documents from the selected bar, that is,
documents that cite a particular BP on a given date.
Each document is divided into paragraphs, indicated by the horizontal stack of bars. The bar's size means the size of the paragraph, and the color intensity reflects the similarity between the corresponding paragraph and the BP text; the darker the color, the greater the similarity, and the more common parts exist between the BP and the paragraph (T5). Similarly to Global View, the color of each stacked bar stands for explicit (blue) or potential (orange) citation (T1).
The similarity is given by the angular distance~\cite{use} between two Truncated-TF-IDF vectors, but other embeddings and similarities (\emph{e.g.}\xspace, the raw cosine similarity) could be adopted.
To guide users further exploring the set of documents, we group them into clusters (T4 -- details below).
Inside each cluster, the documents are positioned in descending order of similarity between the document and the BP,
which is defined as the maximum similarity between its paragraphs and the BP (T4).
Showing documents in descending order of similarity is helpful because the user can promptly identify the top-$k$ most similar, and therefore most interesting, documents. Furthermore, to quickly assess the document distribution w.r.t. the similarity values, we also show an interactive bar chart above each color bar.
\myparagraph{Document Clustering.} Since a binding precedent may cover decisions related to different subjects, the Brazilian Supreme Court website provides, for each BP, some clusters of decisions created according to their subjects. There are only a few clusters for each BP, each containing a few labeled documents (for instance, there are 8 clusters for BP 4, each containing two documents on average). We employed NLP text pre-processing and applied topic
modeling strategies to cluster documents according to their relevant words to take advantage of this limited but useful ground truth. We have tested different and well-established topic extraction methods, including LDA~\cite{blei2003latent}, NMF~\cite{nmf}, SNMF~\cite{snmf}, and PSMF~\cite{psmf}.
The NMF (Frobenius norm) method presented the best results, so we adopted it as the default method.
In the Paragraph Similarities View, users are free to choose the desired number of clusters/topics.
A word cloud with the top-10 most relevant keywords is shown alongside each cluster to optimize the task of finding clusters of interest. Buttons \img{next.png} and \img{previous.png} allow one to navigate between the detected clusters.
\subsection{Document Reader}
\label{sec:document_reader}
After finding a decision that seems interesting, the user selects one of its paragraphs by clicking on it. At this moment, he/she is redirected to \textit{Document Reader} (\figref{fig:visualComponents}~(c)), a view that enables the analysis of both the document's content and the BP text. In this view, the colored bars from the Paragraph Similarities View are also visible alongside the paragraphs (\figref{fig:visualComponents}~(c) and \figref{fig:show_text_similarity}).
When a paragraph is hovered over, its content and the BP text are shown as illustrated in~\figref{fig:show_text_similarity}, with the common parts highlighted or not depending on the user's needs (T5 --- flag ``\textit{Show text similarity}''). In the example shown in the figure, one may notice that the text of the paragraph associated with the darkest blue bar is contained in the BP text (see yellow highlights). To analyze any other paragraph of the current document, one may hover over its text without revisiting Paragraph Similarities View.
Document Reader also incorporates Lime when the opened document refers to a potential citation (T1, T5). In this case, instead of just showing the document's raw text (as in \figref{fig:show_text_similarity}), the parts of the document's content are highlighted according to their influence (negative, neutral, or positive) in the potential citation identification process (see~\figref{fig:visualComponents}~(c)).
\figShowTextSimilarity
Since one may analyze many documents published on different dates and citing different BPs, Document Reader also keeps track of recently opened documents (T6),
allowing the user to quickly revisit them through the document browsing history (\figref{fig:show_text_similarity}).
In practice, it is not uncommon to find complex cases in which a single justice writes his or her decision in a document with more than 100 pages.
Document Reader minimizes experts' reading time by tracking recently opened documents of interest and pointing out exactly which paragraphs of the document are likely to mention the binding precedent, even if not explicitly. These characteristics increase the time efficiency of those who search for specific citations in long documents.
\subsection{Implementation Details}
\label{sec:implementation_details}
Our system adopts a client-server architecture. From the client perspective (front-end), all three visual components were implemented using the D3 library~\cite{d3js}. The server-side (back-end) was implemented in Python and employs some well-established libraries and tools, such as Flask~\cite{flask}, Scikit-learn~\cite{scikit-learn}, NLTK~\cite{nltk}, ElasticSearch~\cite{elastic}, and others.
For storing and retrieving documents' and binding precedents' information, LegalVis relies on a MySQL database.
\section{Usage Scenarios}
\label{sec:case_studies}
In this section, we discuss two usage scenarios that demonstrate the capabilities of LegalVis in exploring legal documents related to binding precedents. The first usage scenario is related to tasks T1, T2, T5 and highlights the system's effectiveness in identifying potential citations under three different situations. The second usage scenario addresses T2, T3, T4, and T6 by showing an exploratory analysis involving related decisions reported by two particular justice rapporteurs. Unless explicitly
stated otherwise, all analyses adopt $t_c = 0.95$. To better understand, we also translated those parts of documents and binding precedents crucial for the analyses to English. The original text contents (in Portuguese) are available in Appendix A.
\subsection{Exploring Potential Citations}
As observed in the explicit citations, potential citations are expected to exist in documents published after the publication of the binding precedent they refer to.
There are, though, documents with potential citations to a particular BP whose publishing date is before the creation date of that BP.
Recall that a binding precedent represents a consolidated understanding among the Supreme Court's justices. Thus a series of decisions may have existed
related to this underlying judicial issue that led to the precedent's creation. In this usage scenario, we explore three particular
potential citations (\figref{fig:caseStudy1}), two of them identified in documents prior to the binding precedent publication (\figref{fig:caseStudy1},~left~and~middle).
The first potential citation involves a document published in Feb. 2013 containing an instrument appeal (\textit{``agravo de instrumento''}, in free translation). According to the system, this document potentially cites BP~37, even though this BP was published in Oct. 2014 (\figref{fig:caseStudy1}~(a)). When analyzing how similar the paragraphs of this document are to BP~37 in the Paragraph Similarities View, we see that one of them is very close to it (the darkest orange bar in~\figref{fig:caseStudy1}~(b)). By looking at the content of this presumable relevant paragraph in the Document Reader (\figref{fig:caseStudy1}~(c)), one may notice that Lime considered a long part of it as of great importance in the citation identification process (see the darkest green background on the text). Recall that Document Reader also allows users to compare the text of the corresponding binding precedent with any selected paragraph, with the possibility of highlighting their common parts. By using this feature (\figref{fig:caseStudy1}~(d)), we can easily perceive that this paragraph contains the entire BP text, with a single modification (``on the'' \textit{vs} ``on''). At this point, a relevant question emerges: \textit{How can a document have the text of a BP before its creation?} In fact, this document does not quote BP 37; instead, it quotes Precedent 339 (\textit{Súmula 339}), a precedent that was eventually converted into BP 37 \cite{sv37} (see explicit mention to Precedent 339 just before the highlighted parts in~\figref{fig:caseStudy1}~(d)).
A potential citation with the explicit BP text is an exception, especially for documents published before the creation of the BP. The expected behavior is that documents and BPs involved in potential citations have the same implicit reasoning. Our second analysis (\figref{fig:caseStudy1}~(middle)) illustrates this case by considering an instrument appeal, published in Nov. 2004 (\figref{fig:caseStudy1}~(e)), that potentially cites BP 4 (created in April 2008). After selecting one of its paragraphs in the
Paragraph Similarities View (\figref{fig:caseStudy1}~(f)), we can compare its text with the BP text in the Document Reader.
When analyzing the BP text and also two parts that were considered relevant by Lime in two adjacent paragraphs (the selected one and a neighbor), we see that both the document and the BP refer to the prohibition of using the minimum wage as an indexing factor (\figref{fig:caseStudy1}~(g)), which supports the claim of a valid potential citation; they however describe this subject in different ways.
The two cases described so far demonstrate the usefulness of LegalVis for identifying and analyzing documents published before creating the binding precedent they are somehow related to. Identifying and exploring these documents help users to understand why the Supreme Court creates binding precedents, one of the key needs pointed out by judicial experts that LegalVis successfully addresses through tasks T1 (identification of potential citations), T2 (overview of documents under a temporal perspective), and T5 (highlight in relevant parts of the documents).
In practice, a justice may refer to a binding precedent X as ``binding entry \#X of the precedent" (\textit{``verbete vinculante no X da súmula''}, in free translation). This latter terminology was not considered when labeling the documents, so (i) documents adopting it were marked as not having explicit citations (\emph{i.e.}\xspace, unlabeled documents), and (ii) it can be used as ground-truth to validate the identification process. To study this case, we rely on a complaint (\textit{``reclamação''}) that potentially cites BP 10 when adopting $t_c = 0.91$ (\figref{fig:caseStudy1}~(right)). This document was selected among all documents that (explicitly or potentially) cite BP 10 in Sep. 2013 (\figref{fig:caseStudy1}~(h-i)). When comparing the text of the BP with one of the paragraphs marked as of great importance by Lime (\figref{fig:caseStudy1}~(j)), we can notice that both refer to some violation of the so-called plenary reserve clause. Still, while the BP defines the rule, the document brings a real-world and related situation. Although this common subject suggests a coherent and valid potential citation, we can validate it by observing an explicit mention to ``Binding Entry \#10 of the Precedent of the Supreme Court" in another document's paragraph (\figref{fig:caseStudy1}~(j)). This example supports once more the usefulness of LegalVis in identifying and exploring potential citations and demonstrates the potential of the system to optimize and validate document labeling.
\subsection{Related Decisions}
Users may explore LegalVis to search for patterns involving decisions and/or BPs. As shown in~\figref{fig:caseStudy2}, this usage scenario presents an exploratory analysis focusing on related decisions reported by two Supreme Court's justices, Teori Zavascki and Edson Fachin.
\figCaseTwo
BP 14 covers a criminal law theme. It grants lawyers the right to access every documented evidence already collected during an investigation that could help them prove their client's innocence. It was applied, for example, at the \textit{Lava Jato} case, a criminal investigation started in 2014 in Brazil on alleged irregularities involving several individuals and companies, among them Petrobras, the largest state-owned company in Brazil \cite{lavajato}.
Among the 64 documents of type ``petition'' (\textit{``petição''}, in Portuguese) that cite or are likely to cite BP 14, 33 are associated with Teori Zavascki and 21 to Edson Fachin.
Likewise, among the 67 documents of type ``investigation'' (\textit{``inquérito''}) related to this BP, Zavascki is the rapporteur of 26, and Fachin has 18, followed by other justices. Except for a single document, Fachin only started dealing with documents of these types after Zavascki stopped, as illustrated by \figref{fig:caseStudy2}~(a) for documents ``investigation''.
A brief background is required to understand this behavior: Teori Zavascki was the Supreme Court's justice rapporteur for the Lava Jato operation cases. Several of Zavascki's decisions, therefore, were related to Lava Jato. He passed away in Jan. 2017 due to an airplane crash, and Fachin was appointed as a new rapporteur for the Lava Jato's cases from that moment on \cite{lavajatonovorelatorstf}.
In this usage scenario, rather than validating potential citation identification, we want to find out if there is a relation between decisions from Zavascki and Fachin connected to Lava Jato. We begin our exploratory analysis by selecting in Global View all documents associated with Edson Fachin for BP~14 in Sep. 2017. The five retrieved documents are shown in Paragraph Similarities View as a single cluster. After asking for two clusters using topic modeling, the documents are clustered as shown in~\figref{fig:caseStudy2}~(b), \emph{i.e.}\xspace, a cluster ``Topic~1'' containing two complaints (documents of type ``Rcl'') that explicitly cite BP 14 and a cluster ``Topic~2'', containing three investigations (documents of type ``Inq'') that potentially cite BP 14. We also see terms related to investigation (\emph{e.g.}\xspace, ``investigar'' and ``apurar'') and ``petrobras'' in the word-cloud relative to Topic~2 (\figref{fig:caseStudy2}~(c)), which is interesting for our search. It is worth noting that the quality of this clustering supports the suitability of topic modeling for separating the documents. As expected, documents of different types tend to be characterized by different words, and consequently, separated into different clusters.
By exploring Topic 2's last document (``Inq 4413'') with Document Reader, we can identify three parts in the text that are relevant for our analysis (\figref{fig:caseStudy2}~(d)). By observing these parts, we can finally establish the relationship we were looking for: Fachin explicitly agrees with Zavascki's decision at ``Investigation 4,231'' (\figref{fig:caseStudy2}~(d-2)). This decision, which is related to Lava Jato and Petrobras (\figref{fig:caseStudy2}~(d-3)), is unfortunately under legal secrecy and is not available for analysis. In his decision, Fachin also mentions two other investigations (ids 4,325 and 4,326 -- see \figref{fig:caseStudy2}~(d-1)), where one may find other agreements or controversies. By identifying this kind of relation between decisions or justices, experts can understand and explore the adopted arguments in new and related processes or anticipate possible outcomes of ongoing ones. The system also facilitates comparisons of decisions. If we open multiple documents in the system (\emph{e.g.}\xspace, those related to these investigations), we can promptly switch between them by using the document browsing history.
Recall that LegalVis has identified a potential citation between ``Inq 4413'' and BP 14 with confidence $t_c = 0.95$. Even though we could not establish reasons that justify this identification, the presented usage scenario showed the system's effectiveness in assisting users in finding patterns and behaviors related to particular justices and documents' types. Since LegalVis properly tackles tasks T2 and T3, finding the temporal relationship between cases reported by Zavasci and Fachin was easy in the explored scenario. Reaching the analyzed document of interest was easy, mainly because of the topic modeling clustering (T4). Finally, comparisons among different decisions are possible and uncomplicated through the document browsing history~(T6).
\section{Evaluation from Experts}
\label{sec:experts}
This section describes an evaluation with experts performed to collect detailed feedback about the presented usage scenarios, the usefulness and usability of the tool, and ideas for further improvement.
\subsection{Participants}
We recruited six domain experts not involved in the tool's development to evaluate our proposal. They work as attorneys (2), researchers (2), legal assistants (1), and trainees (1) and have from 1.5 to 22 years of experience in analyzing legal documents and/or binding precedents. The participation was voluntary and without payment.
\subsection{Evaluation Process}
The expert evaluation involved three steps. First, the participants had to watch a video presenting the pipeline to identify potential citations, the visual components of LegalVis, and the two usage scenarios described in \secref{sec:case_studies}. After that, we invited them to explore the tool through two well-defined tasks:
\myparagraph{T1:} We asked the participants to select the documents associated with Justice Ricardo Lewandowski and: (i) identify the BP least cited by him; (ii) identify the time interval in which he did not cite any BP (if any); (iii) based on the provided word cloud, give their opinion on what they think the 16 decisions associated to him and explicitly citing BP 20 on May 2014 refer to; (iv) find a decision, reported by him, that not only explicitly cites a BP (any BP) but also contains the BP's text. While we consider (i) and (ii) relatively straightforward as they can be answered through Global View with few interactions (filtering options), our goal with (iii) was to guide the participants to Paragraph Similarities View and evaluate how helpful and trustful the word cloud containing relevant keywords is. We also aimed to assess the system for exploring decisions with explicit citations (iv).
\myparagraph{T2:} We asked the participants to find any decision with a potential citation ($t_c = 0.95$) and inform which part of the decision's text they consider relevant w.r.t. the corresponding BP. Besides evaluating the system for explorations involving potential citations, we aimed to assess the quality of the results found by LegalVis on this matter.
Finally, the experts answered a set of quantitative (QT) and qualitative (QL) questions. Regarding the quantitative ones, participants had to answer, through a 5-point Likert scale, whether they agree with the following statements:
``\textit{Usage scenario 1 is relevant.}'' (QT1);
``\textit{Usage scenario 2 is relevant.}'' (QT2);
``\textit{It is easy to find decisions with explicit or potential citations throughout time.}'' (QT3);
``\textit{It is easy to filter interesting decisions for analysis.}'' (QT4);
``\textit{It is easy to identify parts of decisions related to the BP of interest.}'' (QT5);
``\textit{LegalVis is useful.}'' (QT6);
``\textit{It is easy to learn how to use LegalVis.}'' (QT7);
``\textit{LegalVis is efficient and would optimize my time.}'' (QT8);
``\textit{LegalVis is easy to use.}'' (QT9);
``\textit{LegalVis has an intuitive interface.}'' (QT10).
Besides asking for comments on QT3-QT6, and also a general comment on QT7-QT10, the qualitative questions include: ``\textit{What is your impression about the proposed pipeline for the identification of potential citations?}'' (QL1);
``\textit{What do you highlight as relevant or interesting in usage scenarios 1 (QL2) and 2 (QL3)? }'';
``\textit{Without LegalVis, what are the challenges in performing analysis similar to the ones you have made?}'' (QL4);
``\textit{What other tools that allow this type of analysis do you know?}'' (QL5);
``\textit{What are the advantages (QL6) and disadvantages (QL7) of LegalVis compared to other tools you are used to?}'';
``\textit{In your opinion, which are the most helpful visual components?}'' (QL8);
``\textit{Besides the existing visual components, which other components do you think could be incorporated in LegalVis?}'' (QL9);
``\textit{Would you like to leave a final comment?}'' (QL10). We only asked QL6 and Q7 to those who know other tools that allow similar analyses (QL5).
\subsection{Results}
Five of the six participants responded to T1(i) correctly, and all six provided similar/correct answers for T1(ii-iv). We invalidated one answer for T2 because the participant used a document with explicit citation (instead of a potential one) to answer it; four of the other participants provided answers in line with Lime.
We present in the following the experts' opinions about (i) the pipeline of identification of potential citations, (ii) relevance of the usage scenarios described in~\secref{sec:case_studies}, (iii) system's usefulness, (iv) usability, and (v) scope for further improvements.
\myparagraph{Potential citation identification (QL1, QL10):}
The experts considered the pipeline interesting, robust, and promising. One of the participants commented: \emph{``Although I don't know how the machine learning works, the proposal is exciting and, without doubt, of great value not only to lawyers but also to all careers that depend on the analysis of judicial decisions."} For another expert, \emph{``The pipeline seems excellent. The use and improvement of machine learning algorithms lead to great results"}. According to a third participant, \emph{``It enhances the search for binding precedents and similar decisions, which helps attorneys to support their petitions"}.
\myparagraph{Usage scenarios (QT1, QT2, QL2, QL3):}
The first usage scenario was considered relevant by five of the six participants (\figref{fig:likert}~(QT1)). According to their comments, they were particularly delighted because it \textit{``clearly demonstrates the applicability of the tool''} and its results show that LegalVis \textit{``enables and facilitates the identification of previous court positions that led to the creation of BPs''} in a \textit{``practical and intelligent way''}.
\figLikert
Although the overall rating for the second usage scenario is also positive (\figref{fig:likert}~(QT2)), one participant considered it not relevant. He/she stated: \textit{``This analysis involves going through several steps that seem less necessary. If the objective is to compare decisions, a simpler way would be to compare the decisions they cite and their content directly''}. It is worth noting that this usage scenario describes the comparison between decisions and demonstrates how users can find relevant decisions and justices to be compared. For another participant, this usage scenario is relevant because it shows that the system \textit{``helps to find similar decisions, which bring unity to the judiciary''}. A third expert commented: \textit{``I foresee great applicability in the field of legal arguments. Great.''}
\myparagraph{System's usefulness (QT6, QT8, QL4-QL6, QL10):}
All participants consider LegalVis a useful tool, and all but one think it is efficient and time-saving (\figref{fig:likert}~(QT6, QT8)). One of them commented: \textit{``While learning the features, I already anticipated using the tool better illustrate my classes, opinions, and scientific articles''}. According to another participant: \textit{``It will greatly help all law professionals, bringing more quality to the services provided and reducing research time''}. A third expert commented: \textit{``I hope I can use the tool for professional purposes as soon as possible.''}
According to the experts, LegalVis is useful because it addresses some critical issues related to large-scale data analysis (3 participants mentioned this issue), filtering options (2), and searches based only on keywords (3). Only two participants responded `yes' when asked whether they know similar tools. The tools they mentioned are the STF's website and a platform named JusBrasil~\cite{jusbrasil}. According to one participant, \textit{``the search offered by these websites is based only on keywords and some filters. Thus, LegalVis proves to be much more complete and useful for the better development of activities related to legal arguments. In my opinion, the LegalVis proposal is broader and more useful.''}
Another participant commented: \textit{``LegalVis seems to have a proposal that simplifies large-scale data analysis. Most jurists I know rely on search tools based on not much more than keywords. In that sense, LegalVis would be an important tool''}. A third expert highlighted one problem with keyword-based search engines: \textit{``Without LegalVis, we would have to search for keywords on the STF's website. The problem is that it is not always possible to think of good keywords to use. Sometimes even their search engine is poor.''}
Besides the advantages mentioned above, LegalVis was also considered useful because \textit{``potential citations seem to be an excellent source of argument to use in the Courts.''} This participant's comment continues: \textit{``I consider the tool very useful, not only in terms of litigation but also in the development of academic activities (scientific articles, research, classes, etc.).''}
\myparagraph{System's usability (QT3-QT5, QT7, QT9, QT10, QL8):}
Five of the six participants consider it easy to find decisions with citations throughout time (\figref{fig:likert}~(QT3)), especially those with explicit citations. They also agree that one can quickly identify parts of decisions related to the BP of interest (\figref{fig:likert}~(QT5)). Filtering interesting decisions for analysis, on the other hand, was considered a more difficult task (\figref{fig:likert}~(QT4)). About that, one participant commented: \textit{``The learning curve of the system is a little slower at first due to the complexity and a large number of utilities that can be extracted. However, I emphasize that it is possible to filter very interesting documents for analysis''}. Despite this comment about the learning curve, most experts agree that it is easy to learn how to use LegalVis (\figref{fig:likert}~(QT7)).
According to most participants, LegalVis is easy to use and has an intuitive interface (\figref{fig:likert}~(QT9, QT10)). Regarding the most helpful visual components, the experts reported different opinions: \textit{``the similarity color scale''}, \textit{``highlight in relevant excerpts''}, \textit{``timeline visualization''}, \textit{``similarity bar alongside the decision's paragraph''}, \textit{``all of them are useful''}.
\myparagraph{Scope for improvements (QT4, QT10, QL7, QL9, QL10):}
Although LegalVis shows strong capabilities to enhance the analysis of legal documents and binding precedents, it can still be improved as suggested by the domain experts.
Regarding filtering interesting decisions and new features, received suggestions include \textit{``filtering by justice and document type are interesting, but I think more important is a keyword filter, similar to the one offered by the STF's website, to make it easier to search for a specific subject.''} and \textit{``it would be nice if LegalVis could incorporate subject filtering''}. It is worth mentioning that, since we have both the decisions' texts and the capability to detect relevant keywords through topic modeling, LegalVis could incorporate these features with little effort.
Some experts felt that LegalVis could become more user-friendly and offer more filtering possibilities. One of them commented: \textit{``The system is really cool! I think, however, that it is still possible to improve the visual aspect, for example, by removing from the visualization those BPs that are not subject of the person's research''}. Another expert also suggested filtering decisions based on time intervals defined by the user. LegalVis can naturally incorporate these filtering options.
About the quality of the potential citation identification, one participant reported that \textit{``there seem to be a few cases where the potential citation is not very related to the BP''}. We will discuss ideas for improving our model in the next section.
\section{Discussion and Limitations}
\label{sec:limitations}
The presented usage scenarios and experts' evaluations show that LegalVis is a helpful tool to assist judicial experts in analyzing legal documents.
LegalVis can also be adapted to assist experts from the legal field in deciding whether or not a BP should be associated with a process under analysis, reducing the processing times.
A similar framework could also be applied to other domains like medicine and public purchases. Medical prescriptions and government purchases can be seen as legal documents while the regulatory rules as BPs, for example.
Another potential application is in scientific research, where papers to be cited can be seen as precedents for new ones.
\myparagraph{TF-IDF performance.} An important question from our results is why TF-IDF presents better performance than the new generation of text representations tools (Doc2vec, USE, and Longformer).
One hypothesis is that the presence of specific words dictates the behavior of the classifier (SVM in our case). We performed several experiments considering both a decision tree classifier over bag-of-words and a logistic regression using TF-IDF with L1 regularization in a one-vs-rest multiclass classification approach to analyze this aspect further. Both experiments showed that the presence of specific words led to high accuracy in the classification (over 85\% on validation set).
However, each classification method highlights a different set of relevant words, making it hard to establish a consensus on which words are essential for classifying each BP.
This fact reinforces the critical role of Lime, as it allows for breaking the document into sentences rather than words, facilitating the interpretation. Moreover, Lime explanations are local (per sample), enabling individual understanding of the model's decision.
The issue related to the inferior performance of Doc2vec, USE, and Longformer may also be associated with the lack of pre-trained models with Brazilian Portuguese, in particular with Brazilian Portuguese legal language, which has several particularities in the writing style and terms used throughout the documents. This key factor might be negatively impacting the models' performance.
Training models to suit the characteristics of legal data is a task we are very interested in, and the accomplishment of this task is in our near future plans.
It is essential to mention that in certain cases, TF-IDF can, indeed, outperform Doc2vec, as already reported by some previous work~\cite{7822730, Wang_2017, tfidf_melhor_alguns_datasets}.
\myparagraph{Problems with data.} The dataset we use came from a partnership with another project (recall~\secref{sec:data_set}). Unfortunately, it presents some limitations that required some adaptations in our system: (i) presence of some ``duplicated'' documents, \emph{i.e.}\xspace, documents with different titles but same raw texts that had to be ignored in the classification to avoid overfitting the data; (ii) presence of some documents without valid date attribute (marked as Jan. 1970), which were ignored in the visualization;
(iii) presence of some documents for which it was impossible to automatically extract justice information (marked as ``unknown justice'').
\myparagraph{Potential citation identification model.} Although efficient, our model is simple in the sense that it only considers the documents' raw texts. As future work, we intend to incorporate metadata (\emph{e.g.}\xspace, justice rapporteur and date) directly into the process. These metadata already exist as parts of the texts, but we hypothesize that giving more weight to them would improve the identification. We also plan to experiment with other models (\emph{e.g.}\xspace, \textit{Big Bird}~\cite{zaheer2020big}) and analyze their performance. A promising experiment would be to test with Transformer-based models pretrained using Portuguese texts (\emph{e.g.}\xspace, BERTimbau~\cite{souza2020bertimbau}), differently of Longformer, which we hypothesize is a downside in our case. Not least, the system's current version does not allow users to influence the model's decisions. We now intend to aggregate user relevance feedback to improve the performance through additional training. Such feedback would be especially relevant when dealing with more (and imbalanced) BPs or citations to multiple BPs.
\myparagraph{Number of documents and binding precedents.} There are 58 binding precedents up to this moment, and we chose to explore only the ten most cited of them. Before increasing this number, one should be aware that there are binding precedents with just a few citations, consequently impairing the classification task. To address these issues, we intend to explore data augmentation methods (\emph{e.g.}\xspace, EDA~\cite{wei2019eda}) and other unbalanced classification methods \cite{krawczyk_learning_2016}. We also plan to exhibit each class performance and each document's potential citation probability. This way, users are better informed and can influence the system's decisions through the relevance feedback mentioned.
\myparagraph{Topic Modeling.} As shown in the second usage scenario, our clustering strategy based on topic modeling is helpful to guide users to documents of interest. However, the lack of ground truth with a representative sample of decisions impaired a formal evaluation of its quality. Although validated by domain experts for particular cases, we intend to augment our ground truth and perform a more robust analysis.
\section{Original Images in Portuguese}
Through two usage scenarios (Sect.~7) validated by domain experts, our paper demonstrates the usefulness of LegalVis to assist users in exploring legal documents and binding precedents. In the first usage scenario, we focused on highlighting the system's capabilities in identifying potential citations under different situations. Here, we show in~\figref{fig:supcaseStudy1} the original texts (in Portuguese) of those parts of the binding precedents and documents explored throughout this scenario. Our second usage scenario, by its turn, showed an exploratory analysis that aimed to find patterns and behaviors related to particular justices and types of documents. To help in the analysis, we presented parts of a decision reported by justice Edson Fachin (Fig.~8~(d)). Here,~\figref{fig:supcaseStudy2} shows the original texts (in Portuguese) that compose those parts.
\figSupMatCaseTwo
\newpage
\section{Word Count Histogram}
\figref{fig:histogram} shows a word count histogram of the documents considered in our study (i.e., 29,743 documents that explicitly cite at least one of the ten most cited BPs plus 30,000 documents without
explicit citations). In general, we have less than four thousand words.
\figHistogram
\end{document}
| c08a7fe9c763a918fc968b188f0a063d7e4c3b3e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and main results}\label{Sec:Introworm}
Worm problems have a long history. The earliest known problem of this type was posed by L. Moser in \cite{Moser1966} (cf.\;also \cite{Moser1991}) more than 50 years ago:\\
\par
\begingroup
\leftskip=1cm
\noindent \textit{Moser's worm problem: Find a/the (convex) set of least area that contains a congruent copy of each arc in the plane of lenth one.}\\
\par
\endgroup
\noindent Here, the unit arcs are sometimes called worms, while the problem has been phrased in many different ways in the literature: the architect's version (find the smallest comfortable living quarters for a unit worm), the humanitarian version (find the shape of the most efficient worm blanket), the sadistic version (find the shape of the best mallet head), and so on (cf.\;\cite{Wetzel2003}). So far, despite a lot of research, only partial results are known, including the existence of such a minimum cover in the convex case (probably the first time proven in \cite{LaiPoo1986}), but its shape and area remain unknown. The best bounds presently known for its area $\mu$ are:
\begin{equation*}}\def\eeqq{\end{equation*} 0.23224\leq \mu \leq 0.27091\footnote{We round all decimal numbers up to the fifth decimal place.}\eeqq
(cf.\;\cite{KhPagSri2013} for the lower and \cite{Wang2006} for the upper bound).
Worm problems can be formulated in considerable generality (cf.\;\cite{Wetzel2003}):\\
\par
\begingroup
\leftskip=1cm
\noindent\textit{Given a collection $\mathcal{F}$ of $n$-dimensional figures $F$ and a transitive group $\mathcal{M}$ of motions $m$ on $\mathbb{R}^n$, find minimal convex target sets $K\subset \mathbb{R}^n$--minimal in the sense of having least volume, surface volume, or whatever--so that for each $F\in\mathcal{F}$ there is a motion $m \in \mathcal{M}$ with
\begin{equation*}}\def\eeqq{\end{equation*} m(F)\subseteq K.\eeqq}
\par
\endgroup
The existence of solutions to this problem can be guaranteed under certain natural hypotheses by fundamental compactness results like the Blaschke selection theorem (cf.\;\cite[\S 18]{Blaschke1916} for Blaschke's selection theorem and \cite{KellyWeiss1979} or \cite{LaiPoo1986} for its application; cf.\;also Theorem \ref{Thm:Blaschkeselection} and its application in Propositions \ref{Prop:minimumexists}, \ref{Prop:minimaxproblem1}, \ref{Prop:minimaxproblem2}, and \ref{Prop:minimaxproblem3}).
When the problem does not permit an arc to be replaced by its mirror image, then it is appropriate to consider the subgroup of orientation preserving motions. For other problems, e.g., Moser's original worm problem, orientation reversing motions are permitted. Many problems whose motion group is the group of translations have been studied in the literature (cf.\;\cite{BezCon1989}, \cite{CroftFalconerGuy1991}, \cite{Wetzel1973}).
In order to formulate the specific worm problem which is of main interest for our study, we introduce the following definition: Let $T\subset\mathbb{R}^n$ be a convex body, i.e., a compact convex set in $\mathbb{R}^n$ with nonempty interior, and $T^\circ$ its polar. Using the Minkowski functional $\mu_{T^\circ}$ with respect to $T$'s polar, we define the \textit{$\ell_T$-length} of a closed $H^1([0,\widetilde{T}],\mathbb{R}^n)$-curve\footnote{This implies that $q$ is differentiable almost everywhere with $\dot{p}\in L^2([0,\widetilde{T}],\mathbb{R}^n)$.} $q$ (from now on, for the sake of simplicity, every closed curve is assumed to fulfill this Sobolev property), $\widetilde{T}\geq 0$, by
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q):=\int_0^{\widetilde{T}} \mu_{T^\circ}(\dot{p}(t))\, \mathrm{d}t.\eeqq
The worm problem which is of main interest for our study we call the \textit{Minkowski worm problem}. Referring to the above general worm problem formulation, for this for convex body $T\subset\mathbb{R}^n$, we consider $\mathcal{F}=\mathcal{F}(T,\alpha)$ as the set of closed curves of $\ell_T$-length $\alpha > 0$, $\mathcal{M}$ as the group of translations and the minimization in the sense of having minimal volume:\\
\par
\begingroup
\leftskip=1cm
\noindent \textit{Minkowski worm problem: Let $T\subset\mathbb{R}^n$ be a convex body. Find the volume-minimizing convex bodies $K\subset\mathbb{R}^n$ that contain a translate of every closed curve of $\ell_T$-length $\alpha$.}\\
\par
\endgroup
\noindent So, in contrast to Moser's worm problem, we consider general dimension (instead of just dimension two), length-measuring with Minkowski functionals with respect to arbitrary convex bodies (instead of with respect to the Euclidean unit ball), closed curves (instead of not necessarily closed arcs), and translations (instead of congruence transformations). In other words and introducing a notation which will be useful throughout this paper: Let $cc(\mathbb{R}^n)$ be the set of closed curves in $\mathbb{R}^n$. Find the minimizers\footnote{In Proposition \ref{Prop:minimumexists}, we will prove that in fact there exists at least one minimizer.} of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)} \vol(K),\quad \eeqq
where for convex body $T\subset\mathbb{R}^n$ and $\alpha > 0$, we define
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha):=\left\{K\subset\mathbb{R}^n \text{ convex body}: L_T(\alpha)\subseteq C(K)\right\}\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha):=\left\{q \in cc(\mathbb{R}^n) : \ell_T(q)=\alpha\right\}\eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} C(K):=\left\{q \in cc(\mathbb{R}^n) : \exists k\in\mathbb{R}^n \text{ s.t. } q \subseteq k + K\right\},\eeqq
where, for the sake of simplicity, we, in general, identify $q$ with its image.
The case when the dimension is $2$, $T$ is the Euclidean unit ball in $\mathbb{R}^2$, and, without loss of generality, $\alpha =1$ is the only case (one could say: the two-dimensional Euclidean worm problem) which has been investigated so far. It is known as \textit{Wetzel's problem}. So far, the minimal volume (area) for this problem is not known, but the best bounds presently known for the minimum are $0.15544$ as lower (cf.\;\cite{Wetzel1973}, where an argument from \cite{Pal1921} is used) and $0.16526$ as upper bound (cf.\;\cite{BezCon1989}; note that in \cite{Wetzel1973} it was claimed incorrectly an upper bound of $0.159$). In comparison to that: The volumes of the obvious covers of constant width, the ball of radius $1/4$ and the Reuleaux triangle of width $1/2$, are $0.19635$ and $0.17619$, respectively. Since, by the Blaschke-Lebesgue theorem, the Reuleaux triangle is the volume-minimizing set of constant width (cf.\;\cite{Blaschke1915} and \cite{Lebesgue1914}; cf.\;\cite{Harrell2002} for a direct proof by analyzing the underlying variational problem), we can conclude that a minimizer for Wetzel's problem is not of constant width. We refer to Figure \ref{img:upperbounds} for three examples whose volumes are approaching (not achieving) the minimum (clearly, the middle and right convex bodies are not of constant width).
\begin{figure}[h!]
\centering
\def380pt{250pt}
\includegraphics[scale=.9]{images/upperbounds}
\caption[Three examples of convex bodies approaching the minimizer in Wetzel's problem]{On the left side is the Reuleaux triangle with width $\frac{1}{2}$ and volume $0.17619$, in the middle is a convex body with volume $0.17141$ which was found by Wetzel in \cite{Wetzel1973}, and on the right is a convex body, looking a bit like a church window, with base length and height equal to $\frac{1}{2}$ and volume $\frac{1}{6}\approx 0.16667$ (for both the middle and right convex body we refer to \cite{BezCon1989}). Some worms are drawn in in each case.}
\label{img:upperbounds}
\end{figure}
Although we derive some results, the primary goal of our study will not be to solve these Minkowski worm problems, rather to relate them to Viterbo's conjecture from symplectic geometry (cf.\;\cite{Viterbo2000}) which for convex bodies $C\subset\mathbb{R}^{2n}$ reads
\begin{equation*}}\def\eeqq{\end{equation*} \vol(C)\geq \frac{c_{EHZ}(C)^n}{n!}.\eeqq
For that, we recall that the EHZ-capacity of a convex body $C\subset\mathbb{R}^{2n}$ can be defined\footnote{This definition is in fact the outcome of a historically grown study of symplectic capacities. Traced back--recalling that $c_{EHZ}$ in its present form is the generalization of a symplectic capacity by K\"{u}nzle in \cite{Kuenzle1996} after applying the dual action functional introduced by Clarke in \cite{Clarke1979}--, the EHZ-capacity denotes the coincidence of the Ekeland-Hofer- and Hofer-Zehnder-capacities, originally constructed in \cite{EkHo1989} and \cite{HoferZehnder1990}, respectively.} by
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(C)=\min\{\mathbb{A}(x):x \text{ closed characteristic on }\partial C\},\eeqq
where a closed characteristic on $\partial C$ is an absolutely continuous loop in $\mathbb{R}^{2n}$ satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \begin{cases}\dot{x}(t)\in J\partial H_{C}(x(t))\quad \text{a.e}. \\ H_{C}(x(t))=\frac{1}{2}\;\forall t\in \mathbb{T}\end{cases}\eeqq
where
\begin{equation*}}\def\eeqq{\end{equation*} H_{C}(x)=\frac{1}{2}\mu_{C}(x)^2,\quad J=\begin{pmatrix} 0 & \mathbb{1} \\ -\mathbb{1} & 0 \end{pmatrix},\quad \mathbb{T}=\mathbb{R}/\widetilde{T}\mathbb{Z},\; \widetilde{T}>0.\eeqq
$\widetilde{T}$ is the period of the loop and by $\mathbb{A}$ we denote its action defined by
\begin{equation*}}\def\eeqq{\end{equation*} \mathbb{A}(x)= -\frac{1}{2}\int_0^{\widetilde{T}} \langle J\dot{x}(t),x(t)\rangle \, \mathrm{d}t.\eeqq
The first main result of this paper addresses the special case of Lagrangian products
\begin{equation*}}\def\eeqq{\end{equation*} C=K\times T\subset \mathbb{R}^n_q \times \mathbb{R}^n_p \cong \mathbb{R}^{2n},\eeqq
where $K$ and $T$ are convex bodies in $\mathbb{R}^n$. We denote by $\mathcal{C}(\mathbb{R}^n)$ the set of convex bodies in $\mathbb{R}^n$.
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:main1}
Viterbo's conjecture for convex Lagrangian products $K\times T\subset\mathbb{R}^n \times \mathbb{R}^n$
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K\times T) \geq \frac{c_{EHZ}(K\times T)^n}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n),\eeqq
is equivalent to
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=1}\;\min_{K\in A(T,1)}\; \vol(K) \geq \frac{1}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\footnote{In order not to let the entries under the maxima and minima become too long, we decided in general to write extremization problems of the form
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=1,\; T\in \mathcal{C}(\mathbb{R}^n)}\;\min_{K\in A(T,1),\, K\in \mathcal{C}(\mathbb{R}^n)}\; \vol(K)\eeqq
(here, $K\in\mathcal{C}(\mathbb{R}^n)$ is already included within $K\in A(T,1)$; however, it happens sometimes that only convex and centrally symmetric convex bodies are considered in which case the formulations of the extremization problems become as long as indicated) as
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=1}\;\min_{K\in A(T,1)}\; \vol(K),\quad K,T\in \mathcal{C}(\mathbb{R}^n).\eeqq}\label{eq:main11}\eeq
Additionally, equality cases $K^*\times T^*$ of Viterbo's conjecture satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K^*)=\vol(T^*)=1\eeqq
are composed of equality cases $(K^*,T^*)$ of \eqref{eq:main11}. Conversely, equality cases $(K^*,T^*)$ of \eqref{eq:main11} form equality cases $K^*\times T^*$ of Viterbo's conjecture.
\ethm
This yields the following corollary, which seems to be more suitable in order to approach Viterbo's conjecture as an optimization problem (cf.\;Section \ref{Sec:OptimizationProblem}).
\begin{corollary}}\def\ecor{\end{corollary}\label{Cor:optimizationmain1}
Viterbo's conjecture for convex Lagrangian products $K\times T\subset\mathbb{R}^n \times \mathbb{R}^n$
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K\times T) \geq \frac{c_{EHZ}(K\times T)^n}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n),\eeqq
is equivalent to
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=1}\;\min_{a_q\in\mathbb{R}^n}\vol\bigg(\conv\bigg\{ \bigcup_{q\in L_T(1)} (q+a_q)\bigg\}\bigg)\geq \frac{1}{n!},\quad T \in \mathcal{C}(\mathbb{R}^n),\footnote{Here, we note that $K$ has been dissolved by replacing it by an expression that extremizes over all possible $K$s. The extremizing $K$ is of the form \eqref{eq:Kstern}.}\label{eq:optimizationmain11}\eeq
where the second minimum on the left runs for every $q\in L_T(1)$ over all possible translations in $\mathbb{R}^n$. Additionally, equality cases $K^*\times T^*$ of Viterbo's conjecture satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K^*)=\vol(T^*)=1\eeqq
are composed of equality cases $T^*$ of \eqref{eq:optimizationmain11} with
\begin{equation}}\def\eeq{\end{equation} K^*=\conv\bigg\{ \bigcup_{q\in L_{T^*}(1)} (q+a_q^*) \bigg\},\label{eq:Kstern}\eeq
where $a_q^*$ are the minimizers in \eqref{eq:optimizationmain11}. Conversely, equality cases $T^*$ of \eqref{eq:optimizationmain11} with $K^*$ as in \eqref{eq:Kstern} form equality cases $K^*\times T^*$ of Viterbo's conjecture.
\ecor
In analogy to Theorem \ref{Thm:main1}, also Mahler's conjecture from convex geometry (cf.\;\cite{Mahler1939}), i.e.,
\begin{equation}}\def\eeq{\end{equation} \vol(T)\vol(T^\circ)\geq \frac{4^n}{n!},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n),\label{eq:mahlerdefinition}\eeq
where by $\mathcal{C}^{cs}(\mathbb{R}^n)$ we denote the set of all centrally symmetric convex bodies in $\mathbb{R}^n$, can be expressed as a worm problem. As shown in \cite{ArtKarOst2013}, this is due to the fact that Mahler's conjecture is a special case of Viterbo's conjecture.
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:Mahler}
Mahler's conjecture for centrally symmetric convex bodies
\begin{equation}}\def\eeq{\end{equation} \vol(T)\vol(T^\circ)\geq \frac{4^n}{n!},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n),\label{eq:Mahler0}\eeq
is equivalent to
\begin{equation}}\def\eeq{\end{equation} \min_{T\in A\left(T^\circ,\sqrt[n]{\vol(T^\circ)}\right)} \vol(T) \geq \frac{1}{n!},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\label{eq:Mahler1}\eeq
Additionally, equality cases $T^*$ of Mahler's conjecture \eqref{eq:Mahler0} satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T^*)=1\eeqq
are equality cases of \eqref{eq:Mahler1}. And conversely, equality cases $T^*$ in \eqref{eq:Mahler1} are equality cases of Mahler's conjecture \eqref{eq:Mahler0}.
\ethm
Furthermore, also \textit{systolic Minkowski billiard inequalities} within the field of billiard dynamics can be related to worm problems.
In order to state this, let us recall some relevant notions from the theory of Minkowski billiards (cf.\;\cite{KruppRudolf2022}): For convex bodies $K,T\subset\mathbb{R}^n$, we say that a closed polygonal curve\footnote{For the sake of simplicity, whenever we talk of the vertices $q_1,...,q_m$ of a closed polygonal curve, we assume that they satisfy $q_j\neq q_{j+1}$ and $q_j$ is not contained in the line segment connecting $q_{j-1}$ and $q_{j+1}$ for all $j\in\{1,...,m\}$. Furthermore, whenever we settle indices $1,...,m$, then the indices in $\mathbb{Z}$ will be considered as indices modulo $m$.\label{foot:polygonalline}} with vertices $q_1,...,q_m$, $m\geq 2$, on the boundary of $K$ is a \textit{closed weak $(K,T)$-Minkowski billiard trajectory} if for every $j\in \{1,...,m\}$, there is a $K$-supporting hyperplane $H_j$ through $q_j$ such that $q_j$ minimizes
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{T^\circ}(\widebar{q}_j-q_{j-1})+\mu_{T^\circ}(q_{j+1}-\widebar{q}_j),\eeqq
over all $\widebar{q}_j\in H_j$. We encode this closed $(K,T)$-Minkowski billiard trajectory by $(q_1,...,q_m)$. Furthermore, we say that a closed polygonal curve with vertices $q_1,...,q_m$, $m\geq 2$, on the boundary of $K$ is a \textit{closed (strong) $(K,T)$-Minkowski billiard trajectory} if there are points $p_1,...,p_m$ on $\partial T$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \begin{cases}q_{j+1}-q_j\in N_T(p_j),\\ p_{j+1}-p_j=-N_K(q_{j+1})\end{cases}\eeqq
is fulfilled for all $j\in\{1,...,m\}$. We denote by $M_{n+1}(K,T)$ the set of closed $(K,T)$-Minkowski billiard trajectories with at most $n+1$ bouncing points.
Then, for convex body $K\subset\mathbb{R}^n$, introducing $F^{cp}(K)$ as the set of all closed polygonal curves in $\mathbb{R}^n$ that cannot be translated into $\mathring{K}$, we have the following relations:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:relations}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha,c >0$. Then, the following statements are equivalent:
\begin{itemize}
\item[(i)]
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\; \min_{q\in F^{cp}(K)}\ell_T(q) \leq \alpha, \quad K\in\mathcal{C}(\mathbb{R}^n),\eeqq
\item[(ii)]
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\; c_{EHZ}(K\times T) \leq \alpha, \quad K\in\mathcal{C}(\mathbb{R}^n),\eeqq
\item[(iii)]
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\; \min_{q \in M_{n+1}(K,T)} \ell_{T}(q) \leq \alpha, \quad K\in\mathcal{C}(\mathbb{R}^n),\eeqq
\item[(iv)]
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)} \vol(K)\geq c, \quad K\in\mathcal{C}(\mathbb{R}^n),\eeqq
\item[(v)]
\begin{equation*}}\def\eeqq{\end{equation*} \min_{a_q\in\mathbb{R}^n} \vol\bigg(\conv \bigg\{ \bigcup_{q\in L_T(1)}(q+a_q) \bigg\}\bigg) \geq c, \quad K\in\mathcal{C}(\mathbb{R}^n).\eeqq
\end{itemize}
If $T$ is additionally assumed to be strictly convex, then the following systolic weak Minkowski billiard inequality can be added to the above list of equivalent expressions:
\begin{itemize}
\item[(vi)]
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\; \min_{q \text{ cl.\,weak }(K,T)\text{-Mink.\,bill.\,traj.}} \ell_{T}(q) \leq \alpha, \quad K\in\mathcal{C}(\mathbb{R}^n).\eeqq
\end{itemize}
Moreover, every equality case $(K^*,T^*)$ of any of the above inequalities is also an equality case of all the others.
\ethm
Now, we turn our attention to the general Viterbo conjecture for convex bodies in $\mathbb{R}^{2n}$. For that, we first introduce the following definitions: We denote by $\mathcal{C}^p\left(\mathbb{R}^{2n}\right)$ the set of convex polytopes in $\mathbb{R}^{2n}$. For $P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right)$, we denote by
\begin{equation*}}\def\eeqq{\end{equation*} F^{cp}_*(P)\subset F^{cp}(P)\eeqq
the set of all closed polygonal curves $q=(q_1,...,q_m)$ in $F^{cp}(P)$ for which $q_j$ and $q_{j+1}$ are on neighbouring facets $F_j$ and $F_{j+1}$ of $P$ such that there are $\lambda_j,\mu_{j+1}\geq 0$ with
\begin{equation*}}\def\eeqq{\end{equation*} q_{j+1}=q_j + \lambda_j J\nabla H_P(x_j)+\mu_{j+1}J\nabla H_P(x_{j+1}),\eeqq
where $x_j$ and $x_{j+1}$ are arbitrarily chosen interior points of $F_j$ and $F_{j+1}$, respectively. Later, we will see that the existence of such closed polygonal curves is guaranteed.
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:main2}
Viterbo's conjecture for convex polytopes in $\mathbb{R}^{2n}$
\begin{equation}}\def\eeq{\end{equation} \vol(P)\geq \frac{c_{EHZ}(P)^n}{n!},\quad P \in \mathcal{C}^p\left(\mathbb{R}^{2n}\right),\label{eq:main20}\eeq
is equivalent to
\begin{equation}}\def\eeq{\end{equation} \min_{P\in A\left(JP,\frac{1}{R_{P}}\right)} \vol(P) \geq \frac{1}{2^n n!},\quad P \in \mathcal{C}^p\left(\mathbb{R}^{2n}\right),\label{eq:main21}\eeq
where we define
\begin{equation*}}\def\eeqq{\end{equation*} R_{P}:=\frac{\min_{q\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(q)}{\min_{q\in F^{cp}(P)}\ell_{\frac{JP}{2}}(q)} \geq 1.\eeqq
Additionally, $P^*$ is an equality case of Viterbo's conjecture for convex polytopes \eqref{eq:main20} satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(P^*)=1\eeqq
if and only if $P^*$ is an equality case of \eqref{eq:main21}.
\ethm
When we look at the operator norm of the complex structure/symplectic matrix $J$ with respect to a convex body $C\subset\mathbb{R}^{2n}$ as map from
\begin{equation*}}\def\eeqq{\end{equation*} \left(\mathbb{R}^{2n},||\cdot||_{C^\circ}\right) \; \text{ to } \; \left(\mathbb{R}^{2n},||\cdot||_{C}\right)\eeqq
as it has been done in \cite{AkopKar2017} and \cite{GlusOst2015}, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} ||J||_{C^\circ \rightarrow C}=\sup_{||v||_{C^\circ}\leq 1}||Jv||_C,\eeqq
then we derive the following theorem:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:main3}
Viterbo's conjecture for convex bodies in $\mathbb{R}^{2n}$
\begin{equation}}\def\eeq{\end{equation} \vol(C) \geq \frac{c_{EHZ}(C)^n}{n!},\quad C\in \mathcal{C}\left(\mathbb{R}^{2n}\right),\label{eq:main300}\eeq
is equivalent to
\begin{equation}}\def\eeq{\end{equation} \min_{C\in A\left(C^\circ,\frac{1}{\widetilde{R}_C}\right)} \vol(C) \geq \frac{1}{n!},\quad C\in \mathcal{C}\left(\mathbb{R}^{2n}\right),\label{eq:main3}\eeq
where
\begin{equation*}}\def\eeqq{\end{equation*} \widetilde{R}_C:=\frac{c_{EHZ}(C)}{c_{EHZ}(C\times C^\circ)}\geq \frac{1}{2||J||_{C^\circ \rightarrow C}}.\eeqq
Additionally, $C^*$ is an equality case of Viterbo's conjecture for convex bodies in $\mathbb{R}^{2n}$ \eqref{eq:main300} satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(C^*)=1\eeqq
if and only if $C^*$ is an equality case of \eqref{eq:main3}.
\ethm
Finally, we turn to Wetzel's problem. For that, we keep the current state of things in mind:
\begin{theorem}}\def\ethm{\end{theorem}[Wetzel in \cite{Wetzel1973}, '73; Bezdek \& Connelly in \cite{BezCon1989}, '89]\label{Thm:WetzelBezdekConnelly}
In dimension $n=2$, we have
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A\left(B_1^2,1\right)}\vol(K)\in (0.15544,0.16526),\quad K\in\mathcal{C}(\mathbb{R}^2),\eeqq
where we denote by $B_1^2$ the Euclidean unit ball in $\mathbb{R}^2$.
\ethm
Then, as application of Theorem \ref{Thm:main1}, we transfer Viterbo's conjecture onto Wetzel's problem. This results in the following conjecture:
\begin{conjecture}}\def\econj{\end{conjecture}\label{Conj:Wetzelsproblem}
We have
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A\left(B_1^2,1\right)}\vol(K)\geq \frac{1}{2\pi} \approx 0.15915, \quad K\in\mathcal{C}(\mathbb{R}^2).\eeqq
\econj
Applying \cite[Theorem 3.12]{KruppRudolf2022} and Theorem \ref{Thm:relations}, we note that this conjecture can be equivalently expressed as \textit{systolic Euclidean billiard inequality}:
\begin{conjecture}}\def\econj{\end{conjecture}\label{Conj:WetzelsproblemBilliard}
We have
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=\frac{1}{2\pi}}\;\; \min_{q \text{ cl.\,}(K,B_1^2)\text{-Mink.\,bill.\,traj.}} \ell_{B_1^2}(q)\leq 1\eeqq
for $K\in \mathcal{C}(\mathbb{R}^2)$.
\econj
We remark that, for the configuration $(K,B_1^2)$, due to the strict convexity of $B_1^2$, the notions of weak and strong $(K,B_1^2)$-Minkowski billiards coincide and are equal to the one of billiards in the Euclidean sense.
Although much work has been done around Wetzel's problem and the systolic Euclidean billiard inequality, this shows that Viterbo's conjecture is even unsolved for the \say{trivial} configuration
\begin{equation*}}\def\eeqq{\end{equation*} K \times B_1^2\subset\mathbb{R}^2\times\mathbb{R}^2.\eeqq
On the other hand, looking at these two problems from the symplectic point of view, can help us to conceptualize them from a very different point of view.
The worm problems in Theorems \ref{Thm:Mahler}, \ref{Thm:main2} and \ref{Thm:main3} seem to be very hard to solve (as it is expected from the perspective of Mahler's/Viterbo's conjecture). On the one hand, this is a consequence of the inner dependencies within
\begin{equation*}}\def\eeqq{\end{equation*} T\in A\left(T^\circ,\sqrt[n]{\vol(T^\circ)}\right) ,\; P\in A\left(JP,\frac{1}{R_{P}}\right) \; \text{ and }\; C\in A\left(C^\circ,\frac{1}{\widetilde{R}_C}\right),\eeqq
on the other hand, the estimates, hidden in $R_P$ and $\widetilde{R}_C$, beyond specific configurations, do not seem to be so accessible. Nevertheless, beyond the inner dependencies of the second arguments of $A(\cdot,\cdot)$, perhaps it turns out to be fruitful to investigate worm problems of the following structure a little bit more in detail: For $\alpha > 0$, find
\begin{equation*}}\def\eeqq{\end{equation*} \min_{C\in A(C^\circ,\alpha)} \vol(C) \; \text{ and } \; \min_{C\in A(J^{-1}C,\alpha)}\vol(C).\eeqq
Interestingly enough, from this perspective, Viterbo's and Mahler's conjecture are very similar in structure.
Motivated by a relationship between Moser's worm problem and a version of \textit{Bellman's lost-in-a-forest problem} shown by Finch and Wetzel in \cite{FinchWetzel2004}, we further investigate whether it is possible also to relate Minkowski worm problems to versions of Bellman's lost-in-a-forest problem. And indeed, it will turn out that the relationship established in \cite{FinchWetzel2004} is somewhat similar to the relationship between Minkowski worm problems and Viterbo's conjecture for convex Lagrangian products. However, before we will elaborate on this, we will give a short introduction to Bellman's lost-in-a-forest problem and general escape problems of this type.
In 1955, Bellman stated in \cite{Bellman1956} the following research problem (cf.\;also \cite{Bellman1957} and \cite{Bellman1963}):\\
\par
\begingroup
\leftskip=1cm
\noindent \textit{We are given a region $R$ and a random point $P$ within the region. Determine the paths which (a) minimize the expected time to reach the boundary, or (b) minimize the maximum time required to reach the boundary.}\\
\par
\endgroup
\noindent This problem can be phrased as:\\
\par
\begingroup
\leftskip=1cm
\noindent \textit{A hiker is lost in a forest whose shape and dimensions are precisely known to him. What is the best path for him to follow to escape from the forest?}\\
\par
\endgroup
\noindent In other words: To solve the lost-in-a-forest problem one has to find the \textit{best} escape path--the best in terms of minimizing the maximum or expected time required to escape the forest. A third interpretation of \textit{best} has been given in \cite{CroftFalconerGuy1991}: Find the best escape path in terms of maximizing the probability of escape within a specified time period.
Bellman asked about two configurations in particular: on the one hand, the configuration in which the region is the infinite strip between two parallel lines a known distance apart, on the other hand, the configuration in which the region is a half-plane and the hiker's distance from the boundary is known. For the case when \textit{best} is understood in terms of the maximum time to escape, both of these two configurations have been studied: for the first configuration, the best path was found in \cite{Zalgaller1961} ('61), for the second, in \cite{Isbell1957} ('57) (where a complete and detailed proof was not published until it was done in \cite{Joris1980} ('80); cf.\;\cite{Finch2019} for an english translation). In each of these two cases, the shortest escape path is unique up to congruence. Apart from that, not much is known for other interpretations of \textit{best}. We refer to \cite{Ward2008} for a detailed survey on the different types, results, and some related material.
Finch and Wetzel studied in \cite{FinchWetzel2004} the case in which the \textit{best} escape path is the shortest. As already mentioned above, in this case, they could show a fundamental relation to Moser's worm problem.
Before we further elaborate on this, it is worth mentioning to note that Williams in \cite{Williams2002} has included lost-in-a-forest problems in his recent list \say{Million Buck Problems} of unsolved problems of high potential impact on mathematics. He justified the selection of these problems by mentioning that the techniques involved in their resolution will be worth at least one million dollars to mathematics.
Now, let's consider the case studied by Finch and Wetzel and take it a little more rigorously. For that, let $\gamma$ be a \textit{path} in $\mathbb{R}^2$, i.e., a continuous and rectifiable mapping of $[0,1]$ into $\mathbb{R}^2$. Let $\ell_{B_1^2}(\gamma)$ be its Euclidean length and $\{\gamma\}$ its trace $\gamma([0,1])$. We call a \textit{forest} a closed, convex region in the plane with nonempty interior. A path $\gamma$ is an \textit{escape path} for a forest $K$ if a congruent copy of it meets the boundary $\partial K$ no matter how it is placed with its initial point in $K$, i.e., for each point $P\in K$ and each Euclidean motion (translation, rotation, reflection and combinations of them) $\mu$ for which $P=\mu(\gamma(0))$ the intersection $\mu\left(\{\gamma\}\right)\cap \partial K$ is nonempty. Then, among all the escape paths for a forest $K$, there is at least one whose length is the shortest. The \textit{escape length} $\alpha$ of a forest $K$ is the length of one of these shortest escape paths for $K$. Based on these notions, Finch and Wetzel proved the following:
\begin{theorem}}\def\ethm{\end{theorem}[Theorem 3 in \cite{FinchWetzel2004}]\label{Thm:FinchWetzel}
Let $K\subset\mathbb{R}^2$ be a convex body. The escape length $\alpha^*$ of $K$ is the largest $\alpha$ for which for every path $\gamma$ with length $\leq \alpha$, there is a Euclidean motion $\mu$ such that $K$ covers $\mu\left(\{\gamma\}\right)$.
\ethm
For Finch and Wetzel, this theorem established the connection to Moser's worm problem. For that, we recall that in Moser's worm problem one tries to find a/the convex set of least area that contains a congruent copy of each arc in the plane of a certain length. Clearly, the condition of having a certain length can be replaced by the condition of having a length which is bounded from above by that certain length.
Now, translated into our setting, we can derive a similar result. For that, we first have to define a version of a lost-in-a-forest problem which is compatible with the Minkowski worm problems discussed in the previous sections.
In order to indicate the connection to Minkowski worm problems in our setting, we will call the problem the \textit{Minkowski escape problem}. We start by generalizing the problem to any dimension. So, we are considering higher dimensional \say{forests} which one aims to escape. We let $K\subset\mathbb{R}^n$ be a convex body, measure lengths by $\ell_T$, where $T\subset\mathbb{R}^n$ is a convex body, and we call $\gamma$ a \textit{closed Minkowski escape path} for $K$ if $\gamma$ is a closed curve and for each point $P\in K$ and each translation $\mu$ for which $P=\mu(\gamma(0))$ the intersection $\mu\left(\{\gamma\}\right)\cap \partial K$ is nonempty. So, in contrast to considering not necessarily closed paths, allowing the motions to be Euclidean motions and measuring the lengths in the standard Euclidean sense in the escape problem of Finch and Wetzel, we only consider closed paths, translations and measure the lengths by the metric induced by the Minkowski functional with respect to the polar of $T$. Translating this problem into \say{our (mesocosmic) reality}--therefore, requiring $n=2$ and Euclidean measurements--, we get a slightly different problem (of course there are no limits to creativity) (cf.\;Figure \ref{img:minkowskiescapeproblem}):\\
\begin{figure}[h!]
\centering
\def380pt{375pt}
\input{images/minkowskiescapeproblem.pdf_tex}
\caption[Visualization of the Minkowski escape problem]{Visualization of the Minkowski escape problem for the special case of two dimensions with Euclidean measurement. This presents two possible Minkowski escape paths which, however, are not the length-minimizing one. For this $K$, the shortest Minkowski escape path is most likely a closed polygonal cruve with two vertices.}
\label{img:minkowskiescapeproblem}
\end{figure}
\par
\begingroup
\leftskip=1cm
\noindent \textit{Two hikers walk in a forest. One of them gets injured and is in need of medical attention. The unharmed hiker would like to make the emergency call. Although he has his cell phone with him, there is only reception outside the forest. He has a map of the forest, i.e., the shape of the forest and its dimensions are known to him, and a compass to orient himself in terms of direction. Furthermore, he is able to measure the distance he has walked. However, he does not know exactly where in the forest he is. What's the best way to get out of the forest, put off the emergency call, and then get back to the injured hiker?}\\
\par
\endgroup
\noindent The fact that in our story the unharmed hiker knows the shape of the forest and has a compass to orient himself in terms of direction is due to the fact that in our Minkowski escape problem, translations are the only allowed motions. The condition of coming back to the injured hiker is a consequence of our demand to consider only closed curves.
We can prove the analogue to Theorem \ref{Thm:FinchWetzel}:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:analogueFinchWetzel}
Let $K,T\subset\mathbb{R}^n$ be convex bodies. Then, an/the $\ell_T$-minimizing closed Minkowski escape path for $K$ has $\ell_T$-length $\alpha^*$ if and only if $\alpha^*$ is the largest $\alpha$ for which
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha),\eeqq
i.e., for which for every closed path $\gamma$ of $\ell_T$-length $\leq \alpha$, there is a translation $\mu$ such that $K$ covers $\mu\left(\{\gamma\}\right)$.
\ethm
Having in mind that Minkowski escape paths for a convex body $K\subset\mathbb{R}^n$ can be understood as closed curves which cannot be translated into the interior of $K$, we can use the Minkowski billiard characterization of shortest closed polygonal curves that cannot be translated into the interior of $K$, in order to directly conclude the following corollary. Note for this line of argumentation that shortest closed curves that cannot be translated into the interior of $K$ are in fact closed polygonal curves.
\begin{corollary}}\def\ecor{\end{corollary}\label{Cor:analogue}
Let $K,T\subset\mathbb{R}^n$ be convex bodies, where $T$ is additionally assumed to be strictly convex. An/The $\ell_T$-minimizing closed $(K,T)$-Minkowski billiard trajectory has $\ell_T$-length $\alpha^*$ if and only if $\alpha^*$ is the largest $\alpha$ for which
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
\ecor
So, the unharmed hiker in our story can conceptualize his problem by searching for length-minimizing closed Euclidean billiard trajectories.
In general, the problem of minimizing over Minkowski escape problems in the sense of varying the forest while maintaining their volume in order to find the forest with minimal escape length becomes the problem of solving systolic Minkowski billiard inequalities, or equivalently, the problem of proving/investigating Viterbo's conjecture for Lagrangian products in $\mathbb{R}^n\times\mathbb{R}^n$.
This means: If the hikers want to play it safe from the outset by choosing, among forests of equal area, the one where the time needed to help an injured hiker is minimized, then it is useful for them to be familiar with symplectic geometry or billiard dynamics. Of course, they could have paid attention from the beginning to where they entered the forest from and how they designed their path. Then they do not have to solve too difficult problems.
This paper is organized as follows: In Section \ref{Sec:Preliminaries}, we start with some relevant preliminaries before, in Section \ref{Sec:Properties}, we derive properties of Minkowski worm problems and the fundamental results in order to prove Theorems \ref{Thm:main1}, \ref{Thm:Mahler}, \ref{Thm:relations}, \ref{Thm:main2}, and \ref{Thm:main3} and Corollary \ref{Cor:optimizationmain1} in Sections \ref{Sec:main1}, \ref{Sec:main2}, and \ref{Sec:main3}. In Section \ref{Sec:Wetzelproblem}, we prove that it is justified to transfer a special case of Viterbo's conjecture into one for Wetzel's problem which becomes Conjecture \ref{Conj:Wetzelsproblem}. In Section \ref{Sec:escape}, we prove Theorem \ref{Thm:FinchWetzel} as analogue to the relationship between Moser's worm problem and Bellman's lost-in-a-forest problem. Finally, in Section \ref{Sec:OptimizationProblem}, we discuss a computational approach for improving lower bounds in Minkowski worm problems, especially lower bounds for Wetzel's problem.
\section{Preliminaries}\label{Sec:Preliminaries}
We begin by collecting some results concerning the \textit{Fenchel conjugate} of a convex and continuous function $H:\mathbb{R}^n\rightarrow\mathbb{R}$, which for $x^*\in\mathbb{R}^n$ is defined by
\begin{equation*}}\def\eeqq{\end{equation*} H^*(x^*)=\sup_{x\in \mathbb{R}^n}(\langle x,x^*\rangle - H(x)).\eeqq
\begin{proposition}}\def\eprop{\end{proposition}[Proposition II.1.8 in \cite{Ekeland1990}]\label{Prop:ConsofFenchelConjugateDef}
If $H^*$ is the Fenchel conjugate of a convex and continuous function $H:\mathbb{R}^n\rightarrow\mathbb{R}$, then for $x\in\mathbb{R}^n$ we have
\begin{equation*}}\def\eeqq{\end{equation*} H(x)=\sup_{x^*\in\mathbb{R}^n}(\langle x^*,x\rangle - H^*(x^*)).\eeqq
\eprop
The \textit{subdifferential} of $H$ in $x\in\mathbb{R}^n$ is given by
\begin{equation*}}\def\eeqq{\end{equation*} \partial H(x)=\{x^*\in\mathbb{R}^n \vert H^*(x^*)=\langle x^*,x\rangle -H(x)\}\eeqq
Then, we get the \textit{Legendre recipocity formula}:
\begin{proposition}}\def\eprop{\end{proposition}[Proposition II.1.15 in \cite{Ekeland1990}]\label{Prop:Legendrereciprocityformula}
For a convex and continuous function $H:\mathbb{R}^n\rightarrow\mathbb{R}$ the Legendre reciprocity formula is given by
\begin{equation*}}\def\eeqq{\end{equation*} x^*\in\partial H(x)\;\Leftrightarrow\; H^*(x^*)+H(x)=\langle x^*,x\rangle \;\Leftrightarrow\; x\in\partial H^*(x^*),\eeqq
where $x,x^*\in\mathbb{R}^n$.
\eprop
We state the \textit{generalized Euler identity}:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:Euleridentity}
Let $H:\mathbb{R}^n\rightarrow \mathbb{R}$ be a $p$-positively homogeneous, convex and continuous function of $\mathbb{R}^n$. Then, for each $x\in \mathbb{R}^n$ the following identity holds:
\begin{equation*}}\def\eeqq{\end{equation*} \langle x^*,x\rangle=pH(x)\quad \forall x^*\in\partial H(x).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
For each $x\in \mathbb{R}^n$, since $H$ is convex and continuous, we have
\begin{equation*}}\def\eeqq{\end{equation*} \partial H(x)\neq \emptyset.\eeqq
For each
\begin{equation*}}\def\eeqq{\end{equation*} x^*\in\partial H(x)\eeqq
Proposition \ref{Prop:Legendrereciprocityformula} provides
\begin{equation}}\def\eeq{\end{equation} H^*(x^*)+H(x)=\langle x^*,x\rangle\label{eq:Eulerlem1}\eeq
and from Proposition \ref{Prop:ConsofFenchelConjugateDef}, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} H(x)=\sup_{x^*\in \mathbb{R}^n}(\langle x^*,x\rangle -H^*(x^*)),\eeqq
we get
\begin{equation}}\def\eeq{\end{equation} H(y)\geq \langle x^*,y \rangle-H^*(x^*)\quad \forall y\in \mathbb{R}^n.\label{eq:Eulerlem2}\eeq
Combining \eqref{eq:Eulerlem1} and \eqref{eq:Eulerlem2} we get
\begin{equation}}\def\eeq{\end{equation} H(y)\geq \langle x^*,y-x\rangle +H(x)\quad \forall y\in \mathbb{R}^n.\label{eq:Eulerlem3}\eeq
Now, we set
\begin{equation*}}\def\eeqq{\end{equation*} y=\lambda x \quad (\lambda >0)\eeqq
and recognize to have equality in \eqref{eq:Eulerlem3} for $\lambda \rightarrow 1$. Furthermore, we obtain by the $p$-homogeneity of $H$ for $\lambda \rightarrow 1$:
\begin{equation*}}\def\eeqq{\end{equation*} \lim_{\lambda \rightarrow 1}\frac{g(\lambda)-g(1)}{\lambda -1}H(x)= \langle x^*,x\rangle,\eeqq
where we introduced the function
\begin{equation*}}\def\eeqq{\end{equation*} g(x):=x^p.\eeqq
Because of
\begin{equation*}}\def\eeqq{\end{equation*} g'(1)=p\eeqq
we get
\begin{equation*}}\def\eeqq{\end{equation*} pH(x)=\langle x^*,x\rangle.\eeqq
\epf
Noting that for convex body $C\subset\mathbb{R}^n$
\begin{equation*}}\def\eeqq{\end{equation*} H_C=\frac{1}{2}\mu_C^2\eeqq
is a $2$-positively homogeneous, convex and continuous function, we derive the following properties:
\begin{proposition}}\def\eprop{\end{proposition}\label{Lem:ConsLegendre}
For convex body $C\subset\mathbb{R}^n$ we have
\begin{equation*}}\def\eeqq{\end{equation*} H_C^*=H_{C^\circ}.\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
For $\xi\in\mathbb{R}^{n}$ we have
{\allowdisplaybreaks\begin{align*}
\mu_{C^{\circ}}(\xi)&=\min\lbrace t\geq 0 : \xi \in tC^{\circ}\rbrace\\
&=\min \left\lbrace t\geq 0 : \xi \in t \lbrace \widehat{\xi}\in\mathbb{R}^{n}: \langle\widehat{\xi},x\rangle\leq 1 \,\forall x\in C\rbrace\right\rbrace\\
&=\min \left\lbrace t\geq 0 : \xi \in \lbrace \widehat{\xi}\in\mathbb{R}^{n}: \langle \widehat{\xi},x\rangle\leq t \,\forall x\in C\rbrace\right\rbrace\\
&= \min \lbrace t\geq 0 : \langle \xi,x \rangle \leq t \,\forall x\in C\rbrace\\
&=\max_{x\in C}\;\langle \xi,x\rangle\\
& =\max_{\mu_C(x)=1}\langle \xi,x\rangle,
\end{align*}}%
and therefore
{\allowdisplaybreaks\begin{align*}
H_C^*(\xi)&=\sup_{x\in\mathbb{R}^{n}}\left(\langle\xi,x\rangle-H_C(x)\right)\\
&=\sup_{r\geq 0}\sup_{\mu_C(x)=1}\left( \langle\xi,rx\rangle-\frac{1}{2}\mu_C(rx)^2\right)\\
&= \sup_{r\geq 0} \left( r \left(\sup_{\mu_C(x)=1}\langle\xi,x\rangle\right) - \frac{r^2}{2}\right) \\
& = \max_{r\geq 0} \left( r \left(\max_{\mu_C(x)=1}\langle\xi,x\rangle\right) - \frac{r^2}{2}\right)\\
&=\max_{r\geq 0}\left(r\mu_{C^{\circ}}(\xi)-\frac{r^2}{2}\right)\\
&=\frac{\mu_{C^{\circ}}(\xi)^2}{2}\\
&=H_{C^{\circ}}(\xi).
\end{align*}}%
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:ConsEuler}
Let $C\subset\mathbb{R}^n$ be a convex body. If $x^*\in\partial H_C(x)$ for $x\in \mathbb{R}^n$, then
\begin{equation*}}\def\eeqq{\end{equation*} H_{C^\circ}(x^*)=H_C(x).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
With Proposition \ref{Prop:Euleridentity} and the $2$-homogeneity of $H_{C^\circ}$ we can write
\begin{equation*}}\def\eeqq{\end{equation*} 2H_{C^\circ}(x^*)=\langle x',x^*\rangle,\eeqq
where
\begin{equation*}}\def\eeqq{\end{equation*} x'\in\partial H_{C^\circ}(x^*),\eeqq
which together with Propositions \ref{Prop:Legendrereciprocityformula} and \ref{Lem:ConsLegendre} and
\begin{equation*}}\def\eeqq{\end{equation*} (C^\circ)^\circ =C\eeqq
is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} x^*\in\partial H_{C^\circ}^*(x')=\partial H_C(x').\eeqq
Therefore, again using Proposition \ref{Prop:Euleridentity}, we can conclude
\begin{equation*}}\def\eeqq{\end{equation*} 2H_{C^\circ}(x^*)=\langle x',x^*\rangle=2H_{C}(x').\eeqq
In the following we show that
\begin{equation*}}\def\eeqq{\end{equation*} H_C(x')=H_C(x).\eeqq
This would prove the claim.
Again, using Propositions \ref{Prop:Legendrereciprocityformula} and \ref{Lem:ConsLegendre}, the fact
\begin{equation*}}\def\eeqq{\end{equation*} x^*\in\partial H_C(x)\eeqq
is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} x\in \partial H_C^*(x^*)=\partial H_{C^\circ}(x^*).\eeqq
All previous informations now can be summarized by the following two equations:
\begin{equation*}}\def\eeqq{\end{equation*} H_C(x)+H_{C^\circ}(x^*)=\langle x,x^*\rangle,\quad H_{C^\circ}(x^*)+H_C(x')=\langle x',x^*\rangle.\eeqq
The difference yields
\begin{equation*}}\def\eeqq{\end{equation*} H_C(x)-H_C(x')=\langle x-x',x^*\rangle,\eeqq
which implies
\begin{equation*}}\def\eeqq{\end{equation*} H_C(x')=H_C(x)-\langle x-x',x^*\rangle= H_C(x)-\langle x,x^*\rangle + \langle x',x^*\rangle.\eeqq
The conditions
\begin{equation*}}\def\eeqq{\end{equation*} x\in\partial H_{C^\circ}(x^*) \; \text{ and } \; x'\in \partial H_{C^\circ}(x^*)\eeqq
imply, applying Proposition \ref{Prop:Euleridentity},
\begin{equation*}}\def\eeqq{\end{equation*} -\langle x,x^*\rangle + \langle x',x^*\rangle=-2H_{C^\circ}(x^*)+2H_{C^\circ}(x^*)=0,\eeqq
therefore
\begin{equation*}}\def\eeqq{\end{equation*} H_C(x')=H_C(x).\eeqq
\epf
The following proposition is the generalization of \cite[Proposition 3.11]{KruppRudolf2022} from closed polygonal curves to closed curves:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:lengthpropertygeneral}
Let $T\subset\mathbb{R}^n$ be a convex body, $q\in cc(\mathbb{R}^n)$ and $\lambda >0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda q) = \ell_{\lambda T}(q) = \lambda \ell_T(q).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
From
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{T^\circ}(\lambda x)=\mu_{(\lambda T)^\circ}(x)=\lambda \mu_{T^\circ}(x),\quad x\in\mathbb{R}^n,\eeqq
(cf.\;\cite[Proposition 2.3(iii)]{KruppRudolf2022}) we conclude
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda q) =\int_0^{\widetilde{T}} \mu_{(T)^\circ}((\dot{\lambda q})(t))\,\mathrm{d}t=\int_0^{\widetilde{T}} \mu_{(\lambda T)^\circ}(\dot{q}(t))\,\mathrm{d}t = \ell_{\lambda T}(q)\eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda q)=\int_0^{\widetilde{T}} \mu_{(T)^\circ}((\dot{\lambda q})(t))\,\mathrm{d}t = \lambda \int_0^{\widetilde{T}} \mu_{(T)^\circ}((\dot{q})(t))\,\mathrm{d}t=\lambda \ell_T(q).\eeqq
\epf
We continue by recalling \cite[Theorem 1.1]{Rudolf2022} which will be useful throughout this paper:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:relationcapacity}
Let $K,T\subset\mathbb{R}^n$ be convex bodies. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(K\times T)=\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{p\in F^{cp}(T)}\ell_K(p)=\min_{q \in M_{n+1}(K,T)} \ell_T(q).\eeqq
\ethm
We note that in \cite[Theorem 1.1]{Rudolf2022} actually appear $F^{cp}_{n+1}(K)$ and $F^{cp}_{n+1}(T)$ instead of $F^{cp}(K)$ and $F^{cp}(T)$, respectively. However, for the purposes within this paper, we only need this more general formulation which is valid since there are no $\ell_T$/$\ell_K$-minimizing closed polygonal curves in $F^{cp}(K)$/$F^{cp}(T)$ with more than $n+1$ vertices and shorter $\ell_T$/$\ell_K$-length than the $\ell_T$/$\ell_K$-minimizing closed polygonal curves in $F^{cp}_{n+1}(K)$/$F^{cp}_{n+1}(T)$ (cf.\;the proof of point (i) in the proof of \cite[Theorem 2.2]{Rudolf2022}).
We collect some invariance properties of Viterbo's as well as of Mahler's conjecture:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:invViterbotranslation}
Viterbo's conjecture is invariant under translations.
\eprop
\begin{proof}}\def\epf{\end{proof}
Translations
\begin{equation*}}\def\eeqq{\end{equation*} t_a:\mathbb{R}^n\rightarrow \mathbb{R}^n, \quad \xi \mapsto \xi + a,\; a\in\mathbb{R}^n,\eeqq
are symplectomorphism because of
\begin{equation*}}\def\eeqq{\end{equation*} \mathrm{d}t_a(\xi)=\mathbb{1}\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} \mathrm{d}t_a(\xi)^TJ\mathrm{d}t_a(\xi)=J.\eeqq
Finally, we know that Viterbo's conjecture is invariant under symplectomorphisms, since symplectomorphisms preserve the volume as well as the action and therefore the EHZ-capacity.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:invViterbo}
Let $C\subset\mathbb{R}^{2n}$ and $K,T\subset\mathbb{R}^n$ be convex bodies. Then
\begin{equation*}}\def\eeqq{\end{equation*} \vol(C)\geq \frac{c_{EHZ}(C)^n}{n!} \;\Leftrightarrow\; \vol(\lambda C)\geq \frac{c_{EHZ}(\lambda C)^n}{n!}\eeqq
for $\lambda >0$, and
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K \times T)\geq \frac{c_{EHZ}( K \times T)^n}{n!} \; \Leftrightarrow \; \vol(\lambda K \times \mu T)\geq \frac{c_{EHZ}(\lambda K \times \mu T)^n}{n!}\eeqq
for $\lambda,\mu >0$. If
\begin{equation*}}\def\eeqq{\end{equation*} \Phi:\mathbb{R}^n\rightarrow\mathbb{R}^n\eeqq
is an invertible linear transformation, then
\begin{align*}
\vol(K \times T) &\geq \frac{c_{EHZ}( K \times T)^n}{n!}\\
\Leftrightarrow \; \vol\left(\Phi(K)\times \left(\Phi^T\right)^{-1}(T)\right) &\geq \frac{c_{EHZ}\left(\Phi(K)\times \left(\Phi^T\right)^{-1}(T)\right)^n}{n!}.
\end{align*}
\eprop
\begin{proof}}\def\epf{\end{proof}
We have
\begin{equation*}}\def\eeqq{\end{equation*} \vol(\lambda C)=\lambda^{2n}\vol(C)\eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(\lambda C)=\lambda^2 c_{EHZ}(C)\eeqq
due to the $2$-homogeneity of the action. Further,
\begin{equation*}}\def\eeqq{\end{equation*} \vol(\lambda K \times \mu T)= \vol(\lambda K)\vol(\mu K)=\lambda^n \mu^n \vol(K)\vol(T)=\lambda^n \mu^n \vol(K\times T)\eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(\lambda K \times \mu T) = \min_ {q\in F^{cp}(\lambda K)}\ell_{\mu T}(q)= \lambda\mu \min_{q\in F^{cp}(K)}\ell_T(q)\eeqq
due to Theorem \ref{Thm:relationcapacity} and \cite[Proposition 3.11(ii)$\&$(iv)]{KruppRudolf2022} (cf.\;also Lemma \ref{Lem:Fhomogenity}).
Furthermore,
\begin{equation*}}\def\eeqq{\end{equation*} \Phi\times \left(\Phi^T\right)^{-1}\eeqq
is a symplectomorphism, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \left(\Phi\times \left(\Phi^T\right)^{-1}\right)^T J \left(\Phi\times \left(\Phi^T\right)^{-1}\right) = J.\eeqq
Indeed, for $a,b\in\mathbb{R}^n$, we calculate
\begin{align*}
\left(\Phi\times \left(\Phi^T\right)^{-1}\right)^T J \left(\Phi\times \left(\Phi^T\right)^{-1}\right)(a,b) = & \left(\Phi\times \left(\Phi^T\right)^{-1}\right)^T J \left(\Phi(a),\left(\Phi^T\right)^{-1}(b)\right)\\
=&\left(\Phi\times \left(\Phi^T\right)^{-1}\right)^T \left(\left(\Phi^T\right)^{-1}(b),-\Phi(a)\right)\\
=&\left(\Phi^T\times \Phi^{-1}\right)\left(\left(\Phi^T\right)^{-1}(b),-\Phi(a)\right)\\
=&\left(\Phi^T\left(\left(\Phi^T\right)^{-1}(b)\right),\Phi^{-1}(-\Phi(a))\right)\\
=&(b,-a)\\
=&J(a,b),
\end{align*}
where we used the facts
\begin{equation*}}\def\eeqq{\end{equation*} \left(\Phi^T\right)^T=\Phi\; \text{ and }\; \left(\Phi^T\right)^{-1}=\left(\Phi^{-1}\right)^{T}.\eeqq
Finally, we recall that every symplectomorphism preserves the volume as well as the action and therefore the EHZ-capacity.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:invMahler}
If $T\subset\mathbb{R}^n$ is a centrally symmetric convex body and
\begin{equation*}}\def\eeqq{\end{equation*} \Phi:\mathbb{R}^n\rightarrow\mathbb{R}^n\eeqq
an invertible linear transformation, then
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T)\vol(T^\circ)\geq \frac{4^n}{n!}\;\Leftrightarrow \; \vol(\Phi(T))\vol(\Phi(T)^\circ)\geq \frac{4^n}{n!}.\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Because of
\begin{equation*}}\def\eeqq{\end{equation*} \Phi(T)^\circ=\left(\Phi^T\right)^{-1}\left(T^\circ\right)\eeqq
and the volume preservation of
\begin{equation*}}\def\eeqq{\end{equation*} \Phi\times \left(\Phi^T\right)^{-1},\eeqq
we have
\begin{align*}
\vol\left(\Phi(T)\right)\vol\left((\Phi(T))^\circ\right) =&\vol\left(\Phi(T)\times \Phi(T)^\circ\right)\\
=&\vol\left(\Phi(T)\times \left(\Phi^T\right)^{-1}(T^\circ)\right)\\
=&\vol\left(\left(\Phi\times \left(\Phi^T\right)^{-1}\right)(T\times T^\circ)\right)\\
=&\vol(T\times T^\circ)\\
=&\vol(T)\vol(T^\circ).
\end{align*}
\epf
\section{Properties of Minkowski worm problems}\label{Sec:Properties}
We begin by concluding some basic properties of the set
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha), \quad T\in\mathcal{C}(\mathbb{R}^n),\; \alpha > 0.\eeqq
We note that all of the following properties can be easily extended to the case $\alpha \geq 0$. Nevertheless, for the sake of simplicity and in order to avoid trivial case distinctions when it is not possible to divide by $\alpha$, for the following we just treat the case $\alpha >0$.
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:Ahomogenity1}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha,\lambda,\mu > 0$. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} A(\lambda T,\mu \alpha) = \frac{\mu}{\lambda}A(T,\alpha).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
We have
\begin{equation*}}\def\eeqq{\end{equation*} A(\lambda T,\mu\alpha)=\{K\in\mathcal{C}(\mathbb{R}^n) : L_{\lambda T}(\mu\alpha) \subseteq C(K)\}.\eeqq
Because of
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{\lambda T}(q)=\ell_T(\lambda q)\eeqq
(cf.\;Proposition \ref{Prop:lengthpropertygeneral}) we conclude
\begin{equation*}}\def\eeqq{\end{equation*} q\in L_{\lambda T}(\mu\alpha) \; \Leftrightarrow \; \lambda q \in L_T(\mu\alpha)\eeqq
which together with
\begin{equation*}}\def\eeqq{\end{equation*} q\in C(K)\; \Leftrightarrow \; \lambda q \in C(\lambda K)\eeqq
implies
\begin{align*}
A(\lambda T,\mu\alpha)=&\{K\in\mathcal{C}(\mathbb{R}^n):q\in L_{\lambda T}(\mu\alpha)\Rightarrow q\in C(K)\}\\
=&\{K\in\mathcal{C}(\mathbb{R}^n):\lambda q\in L_{T}(\mu\alpha)\Rightarrow \lambda q\in C(\lambda K)\}\\
\overset{(K^*=\lambda K)}{\underset{(q^*=\lambda q)}{=}}&\left\{\frac{1}{\lambda}K^*\in\mathcal{C}(\mathbb{R}^n):q^*\in L_{T}(\mu\alpha)\Rightarrow q^*\in C(K^*)\right\}\\
=&\frac{1}{\lambda}A(T,\mu\alpha).
\end{align*}
Again referring to Proposition \ref{Prop:lengthpropertygeneral} we conclude
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q)=\mu\alpha \; \Leftrightarrow\; \ell_T\left(\frac{q}{\mu}\right)=\alpha,\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} q\in L_T(\mu\alpha)\; \Leftrightarrow \; \frac{q}{\mu} \in L_T(\alpha).\eeqq
This implies
\begin{align*}
A(T,\mu\alpha) =& \{K\in\mathcal{C}(\mathbb{R}^n):q\in L_{T}(\mu\alpha)\Rightarrow q\in C(K)\}\\
=& \left\{K\in\mathcal{C}(\mathbb{R}^n):\frac{q}{\mu}\in L_{T}(\alpha)\Rightarrow \frac{q}{\mu}\in C\left(\frac{K}{\mu}\right)\right\}\\
\overset{(K^*=\frac{K}{\mu})}{\underset{(q^*=\frac{q}{\mu})}{=}}& \{\mu K^* \in\mathcal{C}(\mathbb{R}^n) :q^*\in L_T(\alpha)\Rightarrow q^*\in C(K^*)\}\\
=& \mu A(T,\alpha).
\end{align*}
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:cases1}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha_1,\alpha_2 >0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} \alpha_1 \begin{rcases}\begin{dcases} \leq \\ < \\ = \end{dcases}\end{rcases} \alpha_2 \; \Rightarrow \; A(T,\alpha_1) \begin{rcases}\begin{dcases} \supseteq \\ \supsetneq \\ = \end{dcases}\end{rcases} A(T,\alpha_2).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
We find $\mu >0$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \mu\alpha_1 = \alpha_2.\eeqq
Then, using Proposition \ref{Prop:Ahomogenity1} we have
\begin{equation}}\def\eeq{\end{equation} A(T,\alpha_2)=A(T,\mu \alpha_1) = \mu A(T,\alpha_1).\label{eq:cases10}\eeq
This implies
\begin{equation*}}\def\eeqq{\end{equation*} \alpha_1 \begin{rcases}\begin{dcases} \leq \\ < \\ =\end{dcases}\end{rcases} \alpha_2 \; \Leftrightarrow \; \mu \begin{rcases}\begin{dcases} \geq \\ > \\ =\end{dcases}\end{rcases}1\; \Leftrightarrow \; A(T,\alpha_1) \begin{rcases}\begin{dcases} \supseteq \\ \supsetneq \\ = \end{dcases}\end{rcases} A(T,\alpha_2),\eeqq
where the last equivalence follows from the following considerations: If we have \eqref{eq:cases10} with $\mu \geq 1$, then
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha_2)=\mu A(T,\alpha_1)\eeqq
means that
\begin{equation*}}\def\eeqq{\end{equation*} \frac{1}{\mu}K\in A(T,\alpha_1),\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha_1)\subseteq C\left(\frac{1}{\mu}K\right) \subseteq C(K).\eeqq
This implies
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha_1)\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha_2) \subseteq A(T,\alpha_1).\eeqq
The case $\mu > 1$ follows by considering that in this case there can be find a convex body $K^*$ with
\begin{equation*}}\def\eeqq{\end{equation*} K^*\in A(T,\alpha_1)\setminus A(T,\alpha_2).\eeqq
Indeed, for
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha_2)\eeqq
we define
\begin{equation*}}\def\eeqq{\end{equation*} \widehat{K}:=\widehat{\lambda}K,\quad \widehat{\lambda}:=\min\{0<\lambda \leq 1 : \lambda K\in A(T,\alpha_2)\}.\eeqq
Then, one has
\begin{equation*}}\def\eeqq{\end{equation*} \widehat{K}\in A(T,\alpha_2)\eeqq
and with \eqref{eq:cases10}
\begin{equation*}}\def\eeqq{\end{equation*} \frac{1}{\mu}\widehat{K}\in A(T,\alpha_1).\eeqq
With
\begin{equation*}}\def\eeqq{\end{equation*} K^*:=\frac{1}{\mu}\widehat{K}\eeqq
it follows
\begin{equation*}}\def\eeqq{\end{equation*} K^*\in A(T,\alpha_1)\setminus A(T,\alpha_2)\eeqq
by the definition of $\widehat{K}$.
\epf
For convex body $T\subset\mathbb{R}^n$ and $\alpha > 0$ we define the set
\begin{equation*}}\def\eeqq{\end{equation*} A^{\leq}(T,\alpha)=\left\{K\in \mathcal{C}(\mathbb{R}^n) : L_{T}^{\leq}(\alpha) \subseteq C(K)\right\},\eeqq
where
\begin{equation*}}\def\eeqq{\end{equation*} L_T^{\leq}(\alpha)=\left\{q\in cc(\mathbb{R}^n) : 0<\ell_T(q)\leq \alpha\right\}=\bigcup_{0 < \widetilde{\alpha} \leq \alpha} L_T\left(\widetilde{\alpha}\right).\eeqq
Then, we have the following identity:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:Aidentity}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha > 0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha)=A^{\leq}(T,\alpha).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
By definition it is clear that
\begin{equation*}}\def\eeqq{\end{equation*} A^\leq(T,\alpha) \subseteq A(T,\alpha).\eeqq
Indeed, if
\begin{equation*}}\def\eeqq{\end{equation*} K\in A^{\leq}(T,\alpha),\eeqq
then this means
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\widetilde{\alpha})\subseteq C(K),\quad \text{for all }0<\widetilde{\alpha}\leq \alpha.\eeqq
For $\widetilde{\alpha}=\alpha$ it follows
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha)\subseteq C(K)\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
Let $0<\widetilde{\alpha}\leq\alpha$. Then, it follows from Proposition \ref{Prop:cases1} that
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha) \subseteq A\left(T,\widetilde{\alpha}\right),\quad \text{ for all } 0< \widetilde{\alpha}\leq \alpha.\eeqq
This implies
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\alpha) \subseteq \bigcap_{0<\widetilde{\alpha}\leq \alpha}A\left(T,\widetilde{\alpha}\right)=A^{\leq}(T,\alpha).\eeqq
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:Aproperty11}
Let $\alpha > 0$ and $T_1,T_2\subset\mathbb{R}^n$ be two convex bodies. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} T_1 \subseteq T_2 \; \Rightarrow \; A(T_1,\alpha) \subseteq A(T_2,\alpha).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let
\begin{equation*}}\def\eeqq{\end{equation*} T_1\subseteq T_2.\eeqq
If
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T_1,\alpha),\eeqq
then it follows from Proposition \ref{Prop:Aidentity} that
\begin{equation*}}\def\eeqq{\end{equation*} L_{T_1}^\leq(\alpha)\subseteq C(K).\eeqq
Because of
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{T_1}(q)\leq \ell_{T_2}(q) \; \text{ for all }\; q\in cc(\mathbb{R}^n),\eeqq
as consequence of
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{T_1^\circ}(x)\leq \mu_{T_2^\circ}(x)\quad \forall x\in\mathbb{R}^n,\eeqq
it follows that
\begin{equation*}}\def\eeqq{\end{equation*} L_{T_2}^\leq(\alpha) \subseteq L_{T_1}^\leq(\alpha)\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} L_{T_2}^\leq(\alpha) \subseteq C(K).\eeqq
With Proposition \ref{Prop:Aidentity} this implies
\begin{equation*}}\def\eeqq{\end{equation*} K\in A\left(T_2,\alpha\right).\eeqq
Consequently, it follows
\begin{equation*}}\def\eeqq{\end{equation*} A(T_1,\alpha) \subseteq A(T_2,\alpha).\eeqq
\epf
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:easyLemma}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha > 0$. Further, let $K_1,K_2\subset\mathbb{R}^n$ be two convex bodies with
\begin{equation*}}\def\eeqq{\end{equation*} K_1 \subseteq K_2.\eeqq
Then it holds:
\begin{equation*}}\def\eeqq{\end{equation*} K_1\in A(T,\alpha) \; \Rightarrow \; K_2\in A(T,\alpha).\eeqq
\elem
\begin{proof}}\def\epf{\end{proof}
Let
\begin{equation*}}\def\eeqq{\end{equation*} K_1\in A(T,\alpha),\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha)\subseteq C(K_1).\eeqq
It obviously holds
\begin{equation*}}\def\eeqq{\end{equation*} K_1 \subseteq K_2 \; \Rightarrow \; C(K_1)\subseteq C(K_2).\eeqq
Therefore
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha) \subseteq C(K_2),\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} K_2 \in A(T,\alpha).\eeqq
\epf
For the next Lemma we recall the following: If $(M,d)$ is a metric space and $P(M)$ the set of all nonempty compact subsets of $M$, then $(P(M),d_H)$ is a metric space, where by $d_H$ we denote the \textit{Hausdorff metric} $d_H$ which for nonempty compact subsets $X,Y$ of $(M,d)$ is defined by
\begin{equation*}}\def\eeqq{\end{equation*} d_H(X,Y)=\max\left\{\sup_{x\in X}\inf_{y\in Y} d(x,y),\sup_{y\in Y} \inf_{x\in X} d(x,y)\right\}.\eeqq
Then, $(cc(\mathbb{R}^n),d_H)$ is a metric subspace of the complete metric space $(P(\mathbb{R}^n),d_H)$ which is induced by the Euclidean space $(\mathbb{R}^n,|\cdot|)$. For convex body $K\subset\mathbb{R}^n$ we consider
\begin{equation*}}\def\eeqq{\end{equation*} (F^{cc}(K),d_H) \; \text{ and } \; (C(K),d_H)\eeqq
as metric subspaces of $(cc(\mathbb{R}^n),d_H)$. We have that
\begin{equation*}}\def\eeqq{\end{equation*} F^{cc}(K)\setminus C(K) \; \text{ and } \; C(K)\eeqq
are complements of each other in $cc(\mathbb{R}^n)$.
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:C(T)closed}
Let $K\subset\mathbb{R}^n$ be a convex body. Then, $(C(K),d_H)$ is a closed metric subspace of $(cc(\mathbb{R}^n),d_H)$.
\elem
\begin{proof}}\def\epf{\end{proof}
Since $C(K)$ is a subset of $cc(\mathbb{R}^n)$, $(C(K),d_H)$ is a metric subspace of the metric space $(cc(\mathbb{R}^n),d_H)$. It remains to show that $C(K)$ is a closed subset of $cc(\mathbb{R}^n)$. For that let $(q_j)_{j\in\mathbb{N}}$ be a sequence of closed curves in $C(K)$ $d_H$-converging to the closed curve $q^*$. If
\begin{equation*}}\def\eeqq{\end{equation*} q^*\notin C(K),\eeqq
then $q^*$ cannot be translated into $K$. Using the closedness of $K$ in $\mathbb{R}^n$ this means
\begin{equation*}}\def\eeqq{\end{equation*} \min_{k\in\mathbb{R}^n} \; d_H\left(\partial \conv\{K+k,q^*\},\partial (K+k)\right)=:m>0.\eeqq
Then, we can find a $j_0\in\mathbb{N}$ such that
\begin{equation*}}\def\eeqq{\end{equation*} d_H(p_j,q^*)<m\eeqq
for all $j\geq j_0$. But this implies
\begin{equation*}}\def\eeqq{\end{equation*} \min_{k\in\mathbb{R}^n} \; d_H\left(\partial \conv\{K+k,q_j\},\partial (K+k)\right) >0\eeqq
for all $j\geq j_0$, i.e., $p_j$ cannot be translated into $K$ for all $j\geq j_0$, a contradiction to
\begin{equation*}}\def\eeqq{\end{equation*} q_j\in C(K)\quad \forall j\in\mathbb{N}.\eeqq
Therefore, it follows
\begin{equation*}}\def\eeqq{\end{equation*} q^*\in C(K),\eeqq
and consequently, $(C(K),d_H)$ is a closed metric subspace of $(cc(\mathbb{R}^n),d_H)$.
\epf
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:easyLemma2}
Let $T\subset\mathbb{R}^n$ be a convex body and $(\alpha_k)_{k\in\mathbb{N}}$ an increasing sequence of positive real numbers converging to $\alpha > 0$ for $k\rightarrow\infty$. If
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha_k)\quad \forall k\in\mathbb{N},\eeqq
then also
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
\elem
\begin{proof}}\def\epf{\end{proof}
Let
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha_k)\quad \forall k\in\mathbb{N},\eeqq
i.e.,
\begin{equation}}\def\eeq{\end{equation} L_T(\alpha_k)\subseteq C(K)\quad \forall k\in\mathbb{N}.\label{eq:easyLemma1}\eeq
This means for all $k\in\mathbb{N}$ that for all
\begin{equation*}}\def\eeqq{\end{equation*} q\in cc(\mathbb{R}^n) \; \text{ with } \; \ell_T(q)=\alpha_k\eeqq
holds
\begin{equation*}}\def\eeqq{\end{equation*} q\in C(K).\eeqq
Let us assume that
\begin{equation*}}\def\eeqq{\end{equation*} K\notin A(T,\alpha),\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha) \nsubseteq C(K).\eeqq
This means that there is a $q^*\in cc(\mathbb{R}^n)$ with
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q^*)=\alpha \; \text{ and } \; q^*\in F^{cc}(K)\setminus C(K).\eeqq
Due to Lemma \ref{Lem:C(T)closed} $(C(K),d_H)$ is a closed metric subspace of $(cc(\mathbb{R}^n),d_H)$. Since $F^{cc}(K)\setminus C(K)$ is the complement of $C(K)$ in $cc(\mathbb{R}^n)$ it follows that $F^{cc}(K)\setminus C(K)$ is an open subset of the metric space $(cc(\mathbb{R}^n),d_H)$. Consequently there is a $k_0\in \mathbb{N}$ sufficiently big such that
\begin{equation}}\def\eeq{\end{equation} q^*\frac{\alpha_k}{\alpha}\in F^{cc}(K)\setminus C(K)\quad \forall k\geq k_0.\label{eq:easyLemma2}\eeq
But with \cite[Proposition 3.11(iv)]{KruppRudolf2022} it is
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T\left(q^*\frac{\alpha_k}{\alpha}\right)=\ell_T\left(q^*\right)\frac{\alpha_k}{\alpha}=\alpha_k\quad \forall k\in\mathbb{N},\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} q^*\frac{\alpha_k}{\alpha} \in L_T(\alpha_k)\quad \forall k\in\mathbb{N},\eeqq
which produces a contradiction between \eqref{eq:easyLemma1} and \eqref{eq:easyLemma2}.
Therefore it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
\epf
The next Proposition justifies to write \say{$\min$} instead of \say{$\inf$} within the Minkowski worm problem. The main ingredient of its proof will be Blaschke's selection theorem (cf.\;\cite[\S 18]{Blaschke1916}).
\begin{theorem}}\def\ethm{\end{theorem}[Blaschke selection theorem]\label{Thm:Blaschkeselection}
Let $(C_k)_{k\in\mathbb{N}}$ be a sequence of convex bodies in $\mathbb{R}^n$ satisfying
\begin{equation*}}\def\eeqq{\end{equation*} C_k \subset B_R^n,\quad R>0,\eeqq
for all $k\in\mathbb{N}$. Then there is a subsequence $(C_{k_l})_{l\in\mathbb{N}}$ and a convex body $C$ in $\mathbb{R}^n$ such that $C_{k_l}$ $d_H$-converges to $C$ for $l\rightarrow\infty$.
\ethm
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:minimumexists}
Let $T$ be a convex body and $\alpha > 0$. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} \inf_{K\in A(T,\alpha)} \vol(K)=\min_{K\in A(T,\alpha)}\vol(K).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let $(K_k)_{k\in\mathbb{N}}$ be a minimizing sequence of
\begin{equation}}\def\eeq{\end{equation} \inf_{K\in A(T,\alpha)}\vol(K).\label{eq:minimumexists0}\eeq
Then, there is a $k_0\in\mathbb{N}$ and a sufficiently big $R>0$ such that
\begin{equation*}}\def\eeqq{\end{equation*} K_k \subset B_R^n \quad \forall k\geq k_0.\eeqq
Indeed, if this is not the case, then there is a subsequence $\left(K_{k_j}\right)_{j\in\mathbb{N}}$ such that
\begin{equation}}\def\eeq{\end{equation} R_j:=\max\{R>0 : K_{k_j} \in F(B_R^n)\} \rightarrow \infty\quad (j\rightarrow\infty).\label{eq:minimumexists1}\eeq
Guaranteeing
\begin{equation*}}\def\eeqq{\end{equation*} K_{k_j}\in A(T,\alpha)=\{K\in\mathcal{C}(\mathbb{R}^n):L_T(\alpha)\subseteq C(K)\}\quad \forall j\in\mathbb{N}\eeqq
means that
\begin{equation}}\def\eeq{\end{equation} V_j:= \vol\left(K_{k_j}\right) \rightarrow \infty \quad (j\rightarrow\infty).\label{eq:minimumexists2}\eeq
The latter follows together with \eqref{eq:minimumexists1} and the convexity of $K_{k_j}$ for all $j\in\mathbb{N}$ from the fact that due to
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha)\subseteq C\left(K_{k_j}\right)\quad \forall j\in\mathbb{N}\eeqq
there is no direction in which $K_{k_j}$ can be shrunk. But \eqref{eq:minimumexists2} is not possible since $\left(K_{k_j}\right)_{j\in\mathbb{N}}$ is a minimizing sequence of \eqref{eq:minimumexists0}.
Applying Theorem \ref{Thm:Blaschkeselection}, there is a subsequence $(K_{k_l})_{l\in\mathbb{N}}$ and a convex body $K\subset\mathbb{R}^n$ such that $K_{k_l}$ $d_H$-converges to $K$ for $l\rightarrow\infty$. It remains to show that
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
The fact that $K_{k_l}$ $d_H$-converges to $K$ for $l\rightarrow \infty$ implies that for every $\varepsilon >0$ there is an $l_0\in\mathbb{N}$ such that
\begin{equation}}\def\eeq{\end{equation} (1-\varepsilon)K \subseteq K_{k_l} \subseteq (1+\varepsilon)K\quad \forall l\geq l_0.\label{eq:minimumexists3}\eeq
Then, with
\begin{equation*}}\def\eeqq{\end{equation*} K_{k_l}\in A(T,\alpha)\quad \forall l\in\mathbb{N}\eeqq
it follows from the second inclusion in \eqref{eq:minimumexists3} together with Lemma \ref{Lem:easyLemma} that
\begin{equation*}}\def\eeqq{\end{equation*} (1+\varepsilon)K \in A(T,\alpha).\eeqq
Applying Proposition \ref{Prop:Ahomogenity1} this means
\begin{equation*}}\def\eeqq{\end{equation*} K\in A\left(T,\frac{\alpha}{1+\varepsilon}\right).\eeqq
We define the sequence
\begin{equation*}}\def\eeqq{\end{equation*} \alpha_k:=\frac{\alpha}{1+\frac{1}{k}} \quad \forall k\in\mathbb{N}.\eeqq
Then, $(\alpha_k)_{k\in\mathbb{N}}$ is an increasing sequence of positive numbers converging to $\alpha$ for $k\rightarrow \infty$ and together with the aboved mentioned ($\varepsilon >0$ can be chosen arbitrarily) we have
\begin{equation*}}\def\eeqq{\end{equation*} K\in A\left(T,\alpha_k\right)\quad \forall k\in\mathbb{N}.\eeqq
Applying Lemma \ref{Lem:easyLemma2} it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha).\eeqq
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:Ahomogenity2}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha,\lambda,\mu > 0$. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(\lambda T,\mu \alpha)}\vol(K)=\frac{\mu^n}{\lambda^n} \min_{K\in A(T,\alpha)}\vol(K).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
From
\begin{equation*}}\def\eeqq{\end{equation*} A(\lambda T,\mu\alpha)=\frac{1}{\lambda} A(T,\mu \alpha)\eeqq
(cf.\;Proposition \ref{Prop:Ahomogenity1}) it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(\lambda T,\mu\alpha) \;\Leftrightarrow\; \lambda K\in A(T,\mu \alpha)\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(\lambda T,\mu\alpha)}\vol(K)\overset{(K^*=\lambda K)}{=}\min_{K^*\in A(T,\mu\alpha)}\vol\left(\frac{K^*}{\lambda}\right)=\frac{1}{\lambda^n} \min_{K^*\in A(T,\mu\alpha)}\vol(K^*).\eeqq
From
\begin{equation*}}\def\eeqq{\end{equation*} A(T,\mu\alpha)=\mu A(T,\alpha)\eeqq
(cf.\;Proposition \ref{Prop:Ahomogenity1}) it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\mu\alpha) \;\Leftrightarrow\; \frac{K}{\mu}\in A(T,\alpha)\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\mu\alpha)}\vol(K)\overset{(K^*=\frac{K}{\mu})}{=}\min_{K^*\in A(T,\alpha)}\vol(\mu K^*)=\mu^n \min_{K^*\in A(T,\alpha)}\vol(K^*).\eeqq
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:monotony1}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha_1,\alpha_2 >0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} \alpha_1 \begin{rcases}\begin{dcases} \leq \\ < \\ = \end{dcases}\end{rcases} \alpha_2 \; \Leftrightarrow \; \min_{K\in A(T,\alpha_1)}\vol(K) \begin{rcases}\begin{dcases} \leq \\ < \\ = \end{dcases}\end{rcases} \min_{K\in A(T,\alpha_2)}\vol(K)\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
We find $\mu >0$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \mu\alpha_1 = \alpha_2.\eeqq
Then, we apply Proposition \ref{Prop:Ahomogenity2}.
\epf
Now, for convex bodies $K,T\subset\mathbb{R}^n$ we will turn our attention to the minimization problem
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
The existence of the minimum is guaranteed by Theorem \ref{Thm:relationcapacity}.
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:Fhomogenity}
Let $K,T\subset\mathbb{R}^n$ be convex bodies and $\lambda >0$. Then
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(\lambda K)}\ell_T(q)=\min_{q\in F^{cp}(K)}\ell_T(\lambda q)=\lambda\min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
\elem
\begin{proof}}\def\epf{\end{proof}
Similar to \cite[Proposition 3.11(ii)]{KruppRudolf2022} we have
\begin{equation*}}\def\eeqq{\end{equation*} q\in F^{cp}(\lambda K) \; \Leftrightarrow \; \frac{q}{\lambda}\in F^{cp}(K)\eeqq
and using \cite[Proposition 3.11(iv)]{KruppRudolf2022} therefore
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(\lambda K)}\ell_T(q)=\min_{\frac{q}{\lambda}\in F^{cp}(K)}\ell_T(q)\overset{(q^*=\frac{q}{\lambda})}{=}\min_{q^*\in F^{cp}(K)}\ell_T(\lambda q^*)=\lambda \min_{q^*\in F^{cp}(K)}\ell_T(q^*).\eeqq
\epf
In the following for convex body $T\subset\mathbb{R}^n$ and $c>0$ we consider the minimax problem\footnote{Whenever we write \begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)\eeqq the maximum is understood to consider only convex bodies $K\subset\mathbb{R}^n$. This is implicitly indicated by the fact that we defined $F^{cp}(K)$ only for convex bodies $K\subset\mathbb{R}^n$.}
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
The following Proposition guarantees the existence of its maximum:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:minimaxproblem1}
Let $T\subset\mathbb{R}^n$ be a convex body and $c>0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} \sup_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let $(K_k)_{k\in\mathbb{N}}$ be a maximizing sequence of
\begin{equation}}\def\eeq{\end{equation} \sup_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q).\label{eq:minimaxproblem10}\eeq
Then, there is a $k_0\in\mathbb{N}$ and an $R>0$ such that
\begin{equation}}\def\eeq{\end{equation} K_k \subset B_R^n\quad \forall k\geq k_0.\label{eq:minimaxproblem1}\eeq
Indeed, if this is not the case, then there is a subsequence $\left(K_{k_j}\right)_{j\in\mathbb{N}}$ such that
\begin{equation}}\def\eeq{\end{equation} R_j:=\max\{R>0 : K_{k_j} \in F(B_R^n)\}\rightarrow \infty \quad (j\rightarrow \infty).\label{eq:minimaxproblem11}\eeq
But this implies
\begin{equation}}\def\eeq{\end{equation} L_j:=\min\left\{\ell_T(q) : q\in F^{cp}(K_{k_j})\right\}\rightarrow 0 \quad (j\rightarrow\infty).\label{eq:minimaxproblem12}\eeq
This follows from the fact that for every $j\in\mathbb{N}$ we can find a
\begin{equation*}}\def\eeqq{\end{equation*} q_j\in F^{cp}(K_{k_j})\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q_j)\rightarrow 0 \quad (j\rightarrow\infty).\eeqq
The latter is a consequence of \eqref{eq:minimaxproblem11} and the constraint
\begin{equation}}\def\eeq{\end{equation} \vol\left(K_{k_j}\right)=c\quad \forall j\in\mathbb{N},\label{eq:minimaxproblem13}\eeq
i.e., due to the convexity of $K_{k_j}$ for all $j\in\mathbb{N}$ guaranteeing \eqref{eq:minimaxproblem13} there are directions from the origin in which $K_{k_j}$ has to shrink for $j\rightarrow\infty$ and which are suitable in order to construct convenient $q_j$. But \eqref{eq:minimaxproblem12} is not possible since $\left(K_{k_j}\right)_{j\in\mathbb{N}}$ is a maximizing sequence of \eqref{eq:minimaxproblem10}.
Then, we apply Theorem \ref{Thm:Blaschkeselection} and find a subsequence $(K_{k_l})_{l\in\mathbb{N}}$ and a convex body $K\subset\mathbb{R}^n$ such that $K_{k_l}$ $d_H$-converges to $K$ for $l\rightarrow\infty$. It remains to prove that
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K)=c,\eeqq
but this is an immediate consequence of the $d_H$-continuity of the volume function.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:monotony2}
Let $T\subset\mathbb{R}^n$ be a convex body. Then,
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)\eeqq
increases/decreases strictly if and only if this is the case for $c> 0$.
\eprop
\begin{proof}}\def\epf{\end{proof}
We make use of the implication
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q)\; \Rightarrow\; c_1=c_2\label{eq:monotony21}\eeq
for all $c_1,c_2>0$.
Let us verify \eqref{eq:monotony21}: We assume
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q)\label{eq:monotony22}\eeq
and without loss of generality $c_1 < c_2$. Let the pair
\begin{equation*}}\def\eeqq{\end{equation*} (K_1^*,q_1^*) \; \text{ with } \; \vol(K_1^*)=c_1 \; \text{ and } \; q_1^*\in F^{cp}(K_1^*)\eeqq
be a maximizer of the left side in \eqref{eq:monotony22}, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{q\in F^{cp}(K_1^*)}\ell_T(q)=\ell_T(q_1^*).\eeqq
With
\begin{equation*}}\def\eeqq{\end{equation*} q_1^*\in F^{cp}(K_1^*)\eeqq
similar to \cite[Proposition 3.11(ii)]{KruppRudolf2022} we have
\begin{equation*}}\def\eeqq{\end{equation*} \sqrt[n]{\frac{c_2}{c_1}}q_1^*\in F^{cp}\left(\widetilde{K}\right)\eeqq
for
\begin{equation*}}\def\eeqq{\end{equation*} \widetilde{K}:=\sqrt[n]{\frac{c_2}{c_1}}K_1^*.\eeqq
From
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K_1^*)}\ell_T(q)=\ell_T(q_1^*)\eeqq
it follows together with Lemma \ref{Lem:Fhomogenity} that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}\left(\widetilde{K}\right)}\ell_T(q)=\min_{q\in F^{cp}\left(\sqrt[n]{\frac{c_2}{c_1}}K_1^*\right)}\ell_T(q)=\sqrt[n]{\frac{c_2}{c_1}} \min_{q\in F^{cp}(K_1^*)}\ell_T(q)=\sqrt[n]{\frac{c_2}{c_1}} \ell_T(q_1^*).\eeqq
Since
\begin{equation*}}\def\eeqq{\end{equation*} \vol\left(\widetilde{K}\right)=\vol\left(\sqrt[n]{\frac{c_2}{c_1}}K_1^*\right)=\frac{c_2}{c_1}\vol(K_1^*)=c_2,\eeqq
we conclude
\begin{align*}
\max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{q\in F^{cp}(K_1^*)}\ell_T(q)&=\ell_T(q_1^*)\\
&<\sqrt[n]{\frac{c_2}{c_1}}\ell_T(q_1^*)\\
&=\min_{q\in F^{cp}\left(\widetilde{K}\right)}\ell_T(q)\\
&\leq \max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q),
\end{align*}
which is a contradiction to \eqref{eq:monotony22}. Therefore, noting that the assumption $c_1 > c_2$ would have led analogously to the same contradiction, it follows
\begin{equation*}}\def\eeqq{\end{equation*} c_1=c_2.\eeqq
We now prove the equivalence
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q) <\max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q) \; \Leftrightarrow \; c_1 <c_2\label{eq:monotony25}\eeq
for $c_1,c_2>0$.
If
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q) <\max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q),\label{eq:monotony23}\eeq
then from the first part of the proof it necessarily follows $c_1 \neq c_2$. Let us assume $c_1>c_2$. We further assume that the pair
\begin{equation*}}\def\eeqq{\end{equation*} (K_2^*,q_2^*) \; \text{ with } \; \vol(K_2^*)=c_2 \; \text{ and } \; q_2^*\in F^{cp}(K_2^*)\eeqq
is a maximizer of the right side in \eqref{eq:monotony23}, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{q\in K_2^*}\ell_T(q)=\ell_T(q_2^*).\eeqq
We define
\begin{equation*}}\def\eeqq{\end{equation*} \widehat{K}:=\sqrt[n]{\frac{c_1}{c_2}}K_2^*.\eeqq
From
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in K_2^*}\ell_T(q)=\ell_T(q_2^*)\eeqq
it follows together with Lemma \ref{Lem:Fhomogenity} that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}\left(\widehat{K}\right)}\ell_T(q)=\min_{q\in F^{cp}\left(\sqrt[n]{\frac{c_1}{c_2}}K_2^*\right)}\ell_T(q)=\sqrt[n]{\frac{c_1}{c_2}} \min_{q\in F^{cp}(K_2^*)}\ell_T(q)=\sqrt[n]{\frac{c_1}{c_2}} \ell_T(q_2^*).\eeqq
Since
\begin{equation*}}\def\eeqq{\end{equation*} \vol\left(\widehat{K}\right)=\vol\left(\sqrt[n]{\frac{c_1}{c_2}}K_2^*\right)=\frac{c_1}{c_2}\vol(K_2^*)=c_1,\eeqq
we conclude
\begin{align*}
\max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{q\in F^{cp}(K_2^*)}\ell_T(q)&=\ell_T(q_2^*)\\
&<\sqrt[n]{\frac{c_1}{c_2}}\ell_T(q_2^*)\\
&=\min_{q\in F^{cp}\left(\widehat{K}\right)}\ell_T(q)\\
&\leq \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q),
\end{align*}
which is a contradiction to \eqref{eq:monotony23}. Therefore, we conclude $c_1<c_2$.
Conversely, let $c_1<c_2$. From \eqref{eq:monotony21} we conclude
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q) \neq \max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q).\label{eq:monotony24}\eeq
If the strict inequality "$>$" holds in \eqref{eq:monotony24}, then we conclude from the above proven implication "$\Rightarrow$" in \eqref{eq:monotony25} that $c_1 > c_2$, a contradiction. Therefore, it follows
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c_1}\;\min_{q\in F^{cp}(K)}\ell_T(q) < \max_{\vol(K)=c_2}\;\min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:p^*minF}
Let $K,T\subset\mathbb{R}^n$ be convex bodies with $q^*$ as minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
Then, it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\ell_T(q^*))=A\left(T,\min_{q\in F^{cp}(K)}\ell_T(q)\right).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let $q^*$ be a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
Then, it follows
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\ell_T(q^*)) \subseteq C(K).\eeqq
Indeed, otherwise, if there is
\begin{equation*}}\def\eeqq{\end{equation*} \widetilde{q}\in L_T(\ell_T(q^*))\setminus C(K),\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\widetilde{q})=\ell_T(q^*) \; \text{ and } \; \widetilde{q}\in F^{cc}(K)\setminus C(K),\eeqq
then, due to the openess of
\begin{equation*}}\def\eeqq{\end{equation*} F^{cc}(K)\setminus C(K) \; \text{ in } \; cc(\mathbb{R}^n)\eeqq
with respect to $d_H$ (cf.\;Lemma \ref{Lem:C(T)closed}), there is a $\lambda <1$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \lambda \widetilde{q} \in F^{cc}(K)\setminus C(K).\eeqq
Then, using \cite[Proposition 3.11(iv)]{KruppRudolf2022}, we conclude
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda \widetilde{q})=\lambda \ell_T(\widetilde{q})< \ell_T(\widetilde{q})=\ell_T(q^*)=\min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
Because of the $d_H$-density of $F^{cp}(K)$ in $F^{cc}(K)$ and the $d_H$-continuity of $\ell_T$ on $F^{cc}(K)$ (cf.\;\cite[Proposition 3.11(v)]{KruppRudolf2022}--which is also valid for closed curves) then we can find a
\begin{equation*}}\def\eeqq{\end{equation*} \widehat{q}\in F^{cp}(K)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T\left(\widehat{q}\right)< \min_{q\in F^{cp}(K)}\ell_T(q),\eeqq
a contradiction.
Finally, from
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\ell_T(q^*)) \subseteq C(K)\eeqq
it follows
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\ell_T(q^*))=A\left(T,\min_{q\in F^{cp}(K)}\ell_T(q)\right).\eeqq
\epf
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:intersectionFC}
Let $K\subset\mathbb{R}^n$ be a convex body and $\lambda >1$. If
\begin{equation}}\def\eeq{\end{equation} q\in F^{cc}(K)\cap C(K),\label{eq:intersectionFC1}\eeq
then it follows that
\begin{equation}}\def\eeq{\end{equation} \lambda q \in F^{cc}(K)\setminus C(K).\label{eq:intersectionFC2}\eeq
\elem
\begin{proof}}\def\epf{\end{proof}
If we assume \eqref{eq:intersectionFC1} but \eqref{eq:intersectionFC2} does not hold. Then it follows
\begin{equation*}}\def\eeqq{\end{equation*} q,\lambda q \in C(K)\eeqq
and due to $\lambda >1$ therefore
\begin{equation*}}\def\eeqq{\end{equation*} q\in C\left(\mathring{K}\right).\eeqq
But this is a contradiction to
\begin{equation*}}\def\eeqq{\end{equation*} q\in F^{cc}(K).\eeqq
Therefore, it follows \eqref{eq:intersectionFC2}.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:later}
Let $T\subset\mathbb{R}^n$ be a convex body and $\alpha > 0$. If $K^*$ is a minimizer of
\begin{equation}}\def\eeq{\end{equation} \min_{K\in A(T,\alpha)}\vol(K),\label{eq:later}\eeq
then
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q)=\alpha.\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
If $q^*$ is a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q),\eeqq
then it follows from Proposition \ref{Prop:p^*minF} that
\begin{equation*}}\def\eeqq{\end{equation*} K^*\in A(T,\ell_T(q^*)).\eeqq
This means
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\ell_T(q^*))}\vol(K)\leq \vol(K^*)=\min_{K\in A(T,\alpha)}\vol(K).\eeqq
Proposition \ref{Prop:monotony1} implies
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q^*)\leq \alpha.\eeqq
If
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(q^*)<\alpha,\eeqq
then with Proposition \cite[Proposition 3.11(iv)]{KruppRudolf2022} there is $\lambda >1$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda q^*)=\alpha.\eeqq
Together with
\begin{equation*}}\def\eeqq{\end{equation*} F^{cp}(K^*)\subseteq F^{cc}(K^*)\eeqq
and Lemma \ref{Lem:intersectionFC} the fact
\begin{equation*}}\def\eeqq{\end{equation*} q^*\in F^{cp}(K^*)\eeqq
implies
\begin{equation*}}\def\eeqq{\end{equation*} \lambda q^*\in F^{cc}(K^*)\setminus C(K^*),\eeqq
therefore, there is no translate of $K^*$ that covers $\lambda q^*$. Consequently,
\begin{equation*}}\def\eeqq{\end{equation*} K^*\notin A(T,\ell_T(\lambda q^*))=A(T,\alpha),\eeqq
a contradiction to the fact that $K^*$ is a minimizer of \eqref{eq:later}. Therefore, it follows that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q)=\ell_T(q^*)=\alpha.\eeqq
\epf
The idea which underlies the following Theorem leads to the heart of this paper.
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:MainProperty1}
Let $T\subset\mathbb{R}^n$ be a convex body. If $K^*$ is a minimizer of
\begin{equation}}\def\eeq{\end{equation} \min_{K\in A(T,\alpha)} \vol(K)\label{eq:MoserRelation1}\eeq
for $\alpha > 0$, then $K^*$ is a maximizer of
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(K)=c}\; \min_{q\in F^{cp}(K)} \ell_T(q)\label{eq:MoserRelation2}\eeq
for
\begin{equation*}}\def\eeqq{\end{equation*} c:=\vol(K^*)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q)=\alpha.\eeqq
Conversely, if $K^*$ is a maximizer of \eqref{eq:MoserRelation2} for $c>0$, then $K^*$ is a minimizer of \eqref{eq:MoserRelation1} for
\begin{equation*}}\def\eeqq{\end{equation*} \alpha:= \min_{q\in F^{cp}(K^*)}\ell_T(q)\eeqq
and with
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K^*)=c.\eeqq
Consequently, for $\alpha, c>0$ we have the equivalence
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)}\vol(K)= c \; \Leftrightarrow \; \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)= \alpha\eeqq
and moreover
\begin{equation}}\def\eeq{\end{equation} \min_{K\in A(T,\alpha)}\vol(K)\geq c \; \Leftrightarrow \; \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)\leq \alpha.\label{eq:mainproperty11}\eeq
\ethm
\begin{proof}}\def\epf{\end{proof}
Let $K^*$ be a minimizer of \eqref{eq:MoserRelation1} for $\alpha>0$. If $K^*$ is not a maximizer of \eqref{eq:MoserRelation2} for
\begin{equation*}}\def\eeqq{\end{equation*} c=\vol(K^*),\eeqq
then there is a convex body
\begin{equation*}}\def\eeqq{\end{equation*} K^{**}\subset\mathbb{R}^n \; \text{ with } \; \vol(K^{**})=c\eeqq
and a
\begin{equation*}}\def\eeqq{\end{equation*} q^{**}\in F^{cp}(K^{**})\eeqq
such that
\begin{equation}}\def\eeq{\end{equation} \ell_T(q^{**})=\min_{q\in F^{cp}(K^{**})}\ell_T(q)>\min_{q\in F^{cp}(K^*)}\ell_T(q)=\ell_T(q^*),\label{eq:mainproperty12}\eeq
where by $q^*$ we denote a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q).\eeqq
From Proposition \ref{Prop:p^*minF} it follows
\begin{equation}}\def\eeq{\end{equation} K^{**}\in A(T,\ell_T(q^{**})),\label{eq:mainproperty13}\eeq
and further from Proposotion \ref{Prop:later} that
\begin{equation}}\def\eeq{\end{equation} \min_{q\in F^{cp}(K^*)}\ell_T(q)=\ell_T(q^*)=\alpha.\label{eq:mainproperty14}\eeq
From \eqref{eq:mainproperty12}, \eqref{eq:mainproperty13} and \eqref{eq:mainproperty14} together with Proposition \ref{Prop:monotony1} we conclude
\begin{align*}
c=\vol(K^{**})\stackrel{\eqref{eq:mainproperty13}}{\geq} \min_{K\in A(T,\ell_T(q^{**}))}\vol(K)&\stackrel{\eqref{eq:mainproperty12}}{>}\min_{K\in A(T,\ell_T(q^*))}\vol(K)\\
&\stackrel{\eqref{eq:mainproperty14}}{=}\min_{K\in A(T,\alpha)}\vol(K)\\
&=\vol(K^*)\\
&=c,
\end{align*}
a contradiction. Therefore, $K^*$ is a maximizer of \eqref{eq:MoserRelation2} for
\begin{equation*}}\def\eeqq{\end{equation*} c=\vol(K^*).\eeqq
Conversely, let $K^*$ be a maximizer of \eqref{eq:MoserRelation2} for $c>0$ with
\begin{equation*}}\def\eeqq{\end{equation*} q^*\in F^{cp}(K^*)\eeqq
such that
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\min_{q\in F^{cp}(K^*)}\ell_T(q)=\ell_T(q^*)=:\alpha.\eeqq
Then, from Proposition \ref{Prop:p^*minF} it follows that
\begin{equation*}}\def\eeqq{\end{equation*} K^*\in A(T,\alpha),\eeqq
and consequently
\begin{equation*}}\def\eeqq{\end{equation*} c=\vol(K^*)\geq \min_{K\in A(T,\alpha)} \vol(K).\eeqq
If $K^*$ is not a minimizer of \eqref{eq:MoserRelation1} for
\begin{equation*}}\def\eeqq{\end{equation*} \alpha=\ell_T(q^*),\eeqq
then there is a
\begin{equation*}}\def\eeqq{\end{equation*} K^{**}\in A(T,\alpha)\eeqq
with
\begin{equation}}\def\eeq{\end{equation} c=\vol(K^*)>\min_{K\in A(T,\alpha)}\vol(K)=\vol(K^{**}).\label{eq:mainproperty15}\eeq
Then, from Proposition \ref{Prop:later} it follows that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^{**})}\ell_T(q)=\alpha.\eeqq
This implies
\begin{align*}
\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)= \min_{q\in F^{cp}(K^*)} \ell_T(q)&=\ell_T(q^*)\\
&=\alpha\\
&= \min_{q\in F^{cp}(K^{**})}\ell_T(q)\\
&\leq \max_{\vol(K)=\vol(K^{**})}\;\min_{q\in F^{cp}(K)}\ell_T(q),
\end{align*}
which because of \eqref{eq:mainproperty15} is a contradiction to Proposition \ref{Prop:monotony2}. We conclude that $K^*$ is a minimizer of \eqref{eq:MoserRelation1} for
\begin{equation*}}\def\eeqq{\end{equation*} \alpha=\ell_T(q^*).\eeqq
From the before proven it clearly follows the equivalence
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)}\vol(K)= c \; \Leftrightarrow \; \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)= \alpha\eeqq
for $\alpha,c >0$. In order to prove \eqref{eq:mainproperty11} it remains to show
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)}\vol(K)> c \; \Leftrightarrow \; \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q) < \alpha.\eeqq
Let $K^*$ be a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)}\vol(K),\eeqq
where $c>0$ is chosen such that
\begin{equation}}\def\eeq{\end{equation} \vol(K^*)> c.\label{eq:mainproperty16}\eeq
Then we know from the above reasoning that $K^*$ is a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=\vol(K^*)}\;\min_{q\in F^{cp}(K)}\ell_T(q)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q)=\alpha.\eeqq
From \eqref{eq:mainproperty16} and Proposition \ref{Prop:monotony2} it follows
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q) < \max_{\vol(K)=\vol(K^*)}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\alpha.\eeqq
Conversely, let $K^*$ be a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q),\eeqq
where $\alpha >0$ is chosen such that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_T(q)=:\widetilde{\alpha}<\alpha.\eeqq
Then we know from the above reasoning that $K^*$ is a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{T\in A(T,\widetilde{\alpha})}\vol(K),\eeqq
and from Proposition \ref{Prop:monotony1} it follows
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,\alpha)}\vol(K)>\min_{K\in A(T,\widetilde{\alpha})}\vol(K)= \vol(K^*)=c.\eeqq
\epf
Hereinafter we will deal with the following two minimax problems\footnote{Whenever we write \begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K)\eeqq or \begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=d}\; \max_{\vol(K)=c} \; \min_{q\in F^{cp}(K)} \ell_T(q)\eeqq the minimum/maximum is understood to consider only convex bodies $T\subset\mathbb{R}^n$. This is implicitly indicated by the fact that we defined $A(\cdot,\alpha)$ and $\ell_{\cdot}(q)$ only for convex bodies $T\subset\mathbb{R}^n$.}: For $\alpha,d >0$ we will consider
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K),\eeqq
and for $c,d >0$ we will consider
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=d}\; \max_{\vol(K)=c} \; \min_{q\in F^{cp}(K)} \ell_T(q).\eeqq
It is indeed justified to write \say{$\min$} and \say{$\max$} respectively:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:minimaxproblem2}
Let $\alpha,d > 0$. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} \inf_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K) = \min_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let $(T_k)_{k\in\mathbb{N}}$ be a minimizing sequence of
\begin{equation}}\def\eeq{\end{equation} \inf_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K).\label{eq:minimaxproblem21}\eeq
Then there is an $R>0$ and a $k_0\in\mathbb{N}$ such that
\begin{equation*}}\def\eeqq{\end{equation*} T_k \subset B_R^n\quad \forall k\geq k_0.\eeqq
Indeed, if this is not the case, then there is a subsequence $\left(T_{k_j}\right)_{j\in\mathbb{N}}$ such that
\begin{equation}}\def\eeq{\end{equation} R_j:=\max\left\{R>0 : T_{k_j} \in F\left(B_R^n\right)\right\}\; \rightarrow \infty \quad (j\rightarrow \infty).\label{eq:minimaxproblem22}\eeq
This implies
\begin{align*}
V_j:&=\min\left\{\vol(K):K\in A\left(T_{k_j},\alpha\right)\right\}\\
&= \min\left\{\vol(K):K\in\mathcal{C}(\mathbb{R}^n),\; L_{T_{k_j}}(\alpha)\subseteq C(K)\right\}\\
&\rightarrow \infty \quad (j\rightarrow \infty).
\end{align*}
The latter follows from the fact that--\eqref{eq:minimaxproblem22} together with the convexity of $T_{k_j}$ for all $j\in\mathbb{N}$ and the constraint
\begin{equation*}}\def\eeqq{\end{equation*} \vol\left(T_{k_j}\right)=d\quad \forall j\in\mathbb{N}\eeqq
means that there are directions from the origin in which $T_{k_j}$ has to shrink for $j\rightarrow\infty$--for every $j\in\mathbb{N}$ we can find
\begin{equation*}}\def\eeqq{\end{equation*} q_j\in L_{T_{k_j}}(\alpha)\eeqq
($q_j$ can be constructed by using the aforementioned directions) for which
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{T_{k_j}}(q_j)=\alpha\eeqq
means
\begin{equation*}}\def\eeqq{\end{equation*} \max_{t\in [0,\widetilde{T}_j]} |q_j(t)|\rightarrow\infty \quad (j\rightarrow\infty),\eeqq
where by $\widetilde{T}_j$ we denote the period of the closed curve $q_j$, and for every convex body $K_j\subset\mathbb{R}^n$ minimizing
\begin{equation*}}\def\eeqq{\end{equation*} \min\left\{\vol(K):K\in\mathcal{C}(\mathbb{R}^n),\; L_{T_{k_j}}(\alpha)\subseteq C(K)\right\}\eeqq
means
\begin{equation*}}\def\eeqq{\end{equation*} V_j=\vol(K_j)\rightarrow \infty \quad (j\rightarrow\infty).\eeqq
But this is not possible since $(T_k)_{k\in\mathbb{N}}$ is a minimizing sequence of \eqref{eq:minimaxproblem21}.
Then, we can apply Theorem \ref{Thm:Blaschkeselection}: There is a subsequence $\left(T_{k_l}\right)_{l\in\mathbb{N}}$ and a convex body $T\subset\mathbb{R}^n$ such that $T_{k_l}$ $d_H$-converges to $T$ for $l\rightarrow\infty$. We clearly have
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T)=\vol\left(\lim_{l\rightarrow\infty}T_{k_l}\right)=\lim_{l\rightarrow\infty}\vol\left(T_{k_l}\right)=d.\eeqq
Therefore, $T$ is a minimizer of \eqref{eq:minimaxproblem21}.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:minimaxproblem3}
Let $c,d >0$. Then we have
\begin{equation}}\def\eeq{\end{equation} \sup_{\vol(T)=d}\; \max_{\vol(K)=c} \; \min_{q\in F^{cp}(K)} \ell_T(q) = \max_{\vol(T)=d}\; \max_{\vol(K)=c} \; \min_{q\in F^{cp}(K)} \ell_T(q).\label{eq:minimaxproblem30}\eeq
\eprop
\begin{proof}}\def\epf{\end{proof}
Let $\alpha >0$ and let us consider the minimax problem
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=d} \; \min_{K\in A(T,\alpha)} \vol(K).\label{eq:minimaxproblem31}\eeq
Let the pair
\begin{equation*}}\def\eeqq{\end{equation*} (K^*,T^*) \; \text{ with } \; \vol(T^*)=d \; \text{ and } \; K^*\in A(T^*,\alpha)\eeqq
be a minimizer of \eqref{eq:minimaxproblem31}, i.e., it is
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\; \min_{K\in A(T,\alpha)} \vol(K)=\min_{K\in A(T^*,\alpha)}\vol(K)=\vol(K^*)=:\widetilde{c}.\eeqq
By Theorem \ref{Thm:MainProperty1} $K^*$ is a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)=\alpha.\eeqq
Then, due to
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T^*)=d\eeqq
we clearly have
\begin{equation}}\def\eeq{\end{equation} \alpha=\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\leq \sup_{\vol(T)=d}\; \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T}(q).\label{eq:minimaxproblem32}\eeq
If this is a strict inequality, then there is a pair of convex bodies
\begin{equation*}}\def\eeqq{\end{equation*} (K^{**},T^{**}) \; \text{ with } \; \vol(T^{**})=d \; \text{ and } \; \vol(K^{**})=\widetilde{c}\eeqq
such that
\begin{equation*}}\def\eeqq{\end{equation*} \alpha < \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^{**}}(q)=\min_{q\in F^{cp}(K^{**})}\ell_{T^{**}}(q)=:\widetilde{\alpha}.\eeqq
Then, by Theorem \ref{Thm:MainProperty1} $K^{**}$ is a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T^{**},\widetilde{\alpha})}\vol(K)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T^{**},\widetilde{\alpha})}\vol(K)=\vol(K^{**})=\widetilde{c}.\eeqq
Now, $\widetilde{\alpha} >\alpha$ together with Proposition \ref{Prop:monotony1} implies
\begin{align*}
\widetilde{c}=\vol(K^{**})=\min_{K\in A(T^{**},\widetilde{\alpha})}\vol(K)&\geq \min_{\vol(T)=d}\;\min_{K\in A(T,\widetilde{\alpha})}\vol(K)\\
&>\min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)\\
&=\min_{K\in A(T^*,\alpha)}\vol(K)\\
&=\vol(K^*)\\
&=\widetilde{c},
\end{align*}
a contradiction. Therefore, it follows that the inequality in \eqref{eq:minimaxproblem32} is in fact an equality, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \sup_{\vol(T)=d}\; \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T}(q)=\alpha=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q).\eeqq
This means that the pair $(K^*,T^*)$ is a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \sup_{\vol(T)=d}\; \max_{\vol(K)=\widetilde{c}} \; \min_{q\in F^{cp}(K)} \ell_T(q).\eeqq
Since it is sufficient to prove the claim \eqref{eq:minimaxproblem30} for one $c>0$, we are done.
\epf
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:MainProperty2}
If the pair $(K^*,T^*)$ is a minimizer of
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=d}\; \min_{K\in A(T,\alpha)} \vol(K)\label{eq:minmin}\eeq
for $\alpha,d >0$, then $(K^*,T^*)$ is a maximizer of
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(T)=d}\; \max_{\vol(K)=c}\; \min_{q\in F^{cp}(K)} \ell_T(q)\label{eq:maxmaxmin}\eeq
for
\begin{equation*}}\def\eeqq{\end{equation*} c:=\vol(K^*)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)=\alpha.\eeqq
Conversely, if the pair $(K^*,T^*)$ is a maximizer of \eqref{eq:maxmaxmin} for $c,d>0$, then $(K^*,T^*)$ is a minimizer of \eqref{eq:minmin} for
\begin{equation*}}\def\eeqq{\end{equation*} \alpha:=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K^*)=c.\eeqq
Consequently, for $\alpha,c,d>0$ we have the equivalence
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)= c \; \Leftrightarrow \; \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)= \alpha\eeqq
and moreover
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)\geq c \; \Leftrightarrow \; \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)\leq \alpha.\label{eq:mainproperty21}\eeq
\ethm
\begin{proof}}\def\epf{\end{proof}
Let the pair $(K^*,T^*)$ be a minimizer of \eqref{eq:minmin} for $\alpha,d>0$, i.e., it is
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T^*)=d \; \text{ and } \; K^* \in A(T^*,\alpha)\eeqq
such that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\; \min_{K\in A(T,\alpha)} \vol(K)=\min_{K\in A(T^*,\alpha)}\vol(K)=\vol(K^*).\eeqq
Then, in the proof of Proposition \ref{Prop:minimaxproblem3} we have seen that $(K^*,T^*)$ is a maximizer of \eqref{eq:maxmaxmin} for
\begin{equation*}}\def\eeqq{\end{equation*} c:=\vol(K^*) \; \text{ with } \; \min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)=\alpha.\eeqq
Conversely, let the pair $(K^*,T^*)$ be a maximizer of \eqref{eq:maxmaxmin} for $c,d>0$, i.e., $K^*,T^*\subset\mathbb{R}^n$ are convex bodies of volume $c$ and $d$, respectively, such that
\begin{align*}
\max_{\vol(T)=d}\; \max_{\vol(K)=c}\; \min_{q\in F^{cp}(K)} \ell_T(q)&=\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\\
&=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)\\
&=:\alpha.
\end{align*}
By Theorem \ref{Thm:MainProperty1} $K^*$ minimizes
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T^*,\alpha)}\vol(K)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K^*)=c.\eeqq
Then, we clearly have
\begin{equation*}}\def\eeqq{\end{equation*} c=\vol(K^*)=\min_{K\in A(T^*,\alpha)}\vol(K)\geq \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K).\eeqq
If this is a strict inequality, then there is a pair $(K^{**},T^{**})$ with
\begin{equation*}}\def\eeqq{\end{equation*} c>\min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)=\min_{K\in A(T^{**},\alpha)}\vol(K)=\vol(K^{**})=:\widetilde{c},\eeqq
where
\begin{equation*}}\def\eeqq{\end{equation*} K^{**}\in A(T,\alpha)\eeqq
and $T^{**}\subset\mathbb{R}^n$ is a convex body of volume $d$. Then, by Theorem \ref{Thm:MainProperty1} $K^{**}$ is a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^{**}}(q)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^{**}}(q)=\min_{q\in F^{cp}(K^{**})}\ell_{T^{**}}(q)=\alpha.\eeqq
Now, $\widetilde{c}<c$ together with Proposition \ref{Prop:monotony2} implies
{\allowdisplaybreaks\begin{align*}
\alpha=\min_{q\in F^{cp}(K^{**})}\ell_{T^{**}}(q)&=\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^{**}}(q)\\
&\leq \max_{\vol(T)=d}\;\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T}(q)\\
&<\max_{\vol(T)=d}\; \max_{\vol(K)=c}\; \min_{q\in F^{cp}(K)} \ell_T(q)\\
&=\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\\
&=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)\\
&=\alpha,
\end{align*}}%
a contradiction. Therefore,
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)=c=\vol(K^*)=\min_{K\in A(T^*,\alpha)}\vol(K),\eeqq
i.e., the pair $(K^*,T^*)$ is a minimizer of \eqref{eq:minmin}.
From the before proven it clearly follows the equivalence
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)= c \;\Leftrightarrow\; \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)= \alpha.\eeqq
for $\alpha, c, d >0$.
In order to prove \eqref{eq:mainproperty21} it is sufficient to show
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)> c \;\Leftrightarrow\; \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)< \alpha.\eeqq
Let the pair $(K^*,T^*)$ be a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K),\eeqq
where $c>0$ is chosen such that
\begin{equation*}}\def\eeqq{\end{equation*} c<\min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)=\min_{K\in A(T^*,\alpha)}\vol(K)=\vol(K^*)=:\widetilde{c}.\eeqq
From above reasoning we know that $(K^*,T^*)$ maximizes \eqref{eq:maxmaxmin} (for $c$ replaced by $\widetilde{c}$), i.e., $K^*,T^*\subset\mathbb{R}^n$ are convex bodies of volume $\widetilde{c}$ and $d$, respectively, such that
\begin{align*}
\max_{\vol(T)=d}\;\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_T(q)&=\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\\
&=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)\\
&=\alpha.
\end{align*}
Now, $c<\widetilde{c}$ together with Proposition \ref{Prop:monotony2} implies
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q) < \max_{\vol(T)=d}\;\max_{\vol(K)=\widetilde{c}}\;\min_{q\in F^{cp}(K)}\ell_T(q)=\alpha.\eeqq
Conversely, let $(K^*,T^*)$ be a maximizer of
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q),\eeqq
i.e., $K^*,T^*\subset\mathbb{R}^n$ are convex bodies of volume $c$ and $d$, respectively, where $\alpha >0$ is chosen such that
\begin{align*}
\alpha > \max_{\vol(T)=d}\;\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_T(q)&=\max_{\vol(K)=c}\;\min_{q\in F^{cp}(K)}\ell_{T^*}(q)\\
&=\min_{q\in F^{cp}(K^*)}\ell_{T^*}(q)\\
&=:\widetilde{\alpha}.
\end{align*}
Then we know from above reasoning that $(K^*,T^*)$ minimizes \eqref{eq:minmin} (for $\alpha$ replaced by $\widetilde{\alpha}$), i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\widetilde{\alpha})}\vol(K)=\min_{K\in A(T^*,\widetilde{\alpha})}\vol(K)=\vol(K^*)=c.\eeqq
Now, $\alpha > \widetilde{\alpha}$ together with Proposition \ref{Prop:monotony1} implies
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=d}\;\min_{K\in A(T,\alpha)}\vol(K)>\min_{\vol(T)=d}\;\min_{K\in A(T,\widetilde{\alpha})}\vol(T)=c.\eeqq
\epf
\section{Proofs of Theorems \ref{Thm:main1}, \ref{Thm:Mahler}, \ref{Thm:relations} and Corollary \ref{Cor:optimizationmain1}}\label{Sec:main1}
In the following, we mainly make use of Theorems \ref{Thm:MainProperty1} and \ref{Thm:MainProperty2}. However, we begin by rewriting Viterbo's conjecture for convex Lagrangian products:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:rewriteViterbo}
Viterbo's conjecture for convex Lagrangian products $K\times T \subset \mathbb{R}^n\times\mathbb{R}^n$
\begin{equation*}}\def\eeqq{\end{equation*} \vol(K\times T) \geq \frac{c_{EHZ}(K\times T)^n}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n),\eeqq
is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=1}\; \max_{\vol(T)=1}\; \min_{q\in F^{cp}(K)} \ell_T(q) \leq \sqrt[n]{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Using Proposition \ref{Prop:invViterbo}, Viterbo's conjecture for convex Lagrangian products is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=1}\; \max_{\vol(T)=1}\; c_{EHZ}(K\times T) \leq \sqrt[n]{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
By Theorem \ref{Thm:relationcapacity}, this is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=1}\; \max_{\vol(T)=1}\; \min_{q\in F^{cp}(K)} \ell_T(q) \leq \sqrt[n]{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:main1}]
Using Proposition \ref{Prop:rewriteViterbo}, Viterbo's conjecture for convex Lagrangian products is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=1}\; \max_{\vol(T)=1}\; \min_{q\in F^{cp}(K)} \ell_T(q) \leq \sqrt[n]{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
After applying Theorem \ref{Thm:MainProperty2}, it is further equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=1}\;\min_{K\in A\left(T,\sqrt[n]{n!}\right)}\vol(K)\geq 1,\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
Using Proposition \ref{Prop:Ahomogenity2}, this can be written as
\begin{equation*}}\def\eeqq{\end{equation*} \min_{\vol(T)=1}\;\min_{K\in A(T,1)}\vol(K)\geq \frac{1}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\eeqq
By similar reasoning, Theorem \ref{Thm:MainProperty2} also guarantees the equivalence of the equality case of Viterbo's conjecture for convex Lagrangian products $K\times T\subset\mathbb{R}^n\times\mathbb{R}^n$
\begin{equation}}\def\eeq{\end{equation} \vol(K\times T) = \frac{c_{EHZ}(K\times T)^n}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n),\label{eq:rewriteViterbo1}\eeq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(K)=1}\; \max_{\vol(T)=1}\; \min_{q\in F^{cp}(K)} \ell_T(q) = \sqrt[n]{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n),\eeqq
and
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=1}\;\min_{K\in A(T,1)}\vol(K)= \frac{1}{n!},\quad K,T \in \mathcal{C}(\mathbb{R}^n).\label{eq:rewriteViterbo2}\eeq
Moreover, Theorem \ref{Thm:MainProperty2} guarantees the following: If $K^*\times T^*$ is a solution of \eqref{eq:rewriteViterbo1} satisfying
\begin{equation}}\def\eeq{\end{equation} \vol(K^*)=\vol(T^*)=1\label{eq:rewriteViterbo3}\eeq
(note that, applying Proposition \ref{Prop:invViterbo}, the property of being a solution of \eqref{eq:rewriteViterbo1} is invariant under scaling), then the pair $(K^*,T^*)$ is a solution of \eqref{eq:rewriteViterbo2}. And conversely, if the pair $(K^*,T^*)$ is a solution of \eqref{eq:rewriteViterbo2}, then $K^*\times T^*$ is a solution of \eqref{eq:rewriteViterbo1}.
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Corollary \ref{Cor:optimizationmain1}]
In view of the proof of Theorem \ref{Thm:main1}, for convex bodies $K,T\subset\mathbb{R}^n$, it is sufficient to prove the following equality:
\begin{equation}}\def\eeq{\end{equation} \min_{\vol(T)=1}\; \min_{K\in A(T,1)} \vol(K) = \min_{\vol(T)=1}\;\min_{a_q\in\mathbb{R}^n}\vol\bigg(\conv\bigg\{ \bigcup_{q\in L_T(1)} (q+a_q)\bigg\}\bigg).\label{eq:approach0}\eeq
But this follows from the following gradually observation: First, we notice that the volume-minimizing convex cover for a set of closed curves is, equivalently, the volume-minimizing convex hull of these closed curves. So, if we ask for lower bounds of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{K\in A(T,1)}\vol(K),\eeqq
we note that for $q_1,...,q_k\in L_T(1)$, we have
\begin{equation*}}\def\eeqq{\end{equation*} \min_{(a_1,...,a_k)\in (\mathbb{R}^n)^k}\;\vol\left(\conv\{q_1+a_1,...,q_k+a_k\}\right) \leq \min_{K\in A(T,1)}\vol(K).\eeqq
This estimate can be further improved by
\begin{equation*}}\def\eeqq{\end{equation*} \max_{q_1,...,q_k\in L_T(1)}\;\min_{(a_1,...,a_k)\in (\mathbb{R}^n)^k}\;\vol(\conv\{q_1+a_1,...,q_k+a_k\}) \leq \min_{K\in A(T,1)}\vol(K),\eeqq
so that eventually we get
\begin{equation*}}\def\eeqq{\end{equation*} \min_{a_q\in\mathbb{R}^n}\vol\bigg(\conv\bigg\{ \bigcup_{q\in L_T(1)} (q+a_q)\bigg\}\bigg)=\min_{K\in A(T,1)}\vol(K),\eeqq
where the minimum on the left runs for every $q\in L_T(1)$ over all possible translations in $\mathbb{R}^n$. Minimizing this equation over all convex bodies $T\subset\mathbb{R}^n$ of volume $1$, we get \eqref{eq:approach0}.
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:Mahler}]
Because of
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(T\times T^\circ)=4\eeqq
for all centrally symmetric convex bodies $T\subset\mathbb{R}^n$ (cf.\;\cite{ArtKarOst2013}), Mahler's conjecture for centrally symmetric convex bodies is equivalent to
\begin{equation}}\def\eeq{\end{equation} \vol(T\times T^\circ)\geq \frac{c_{EHZ}(T\times T^\circ)^n}{n!}, \quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\label{eq:mahlerpr1}\eeq
Fixing
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T)=1,\eeqq
which is without loss of generality due to Proposition \ref{Prop:invMahler}, and using Theorem \ref{Thm:relationcapacity}, \eqref{eq:mahlerpr1} is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \sqrt[n]{n! \vol(T^\circ)}\geq c_{EHZ}(T\times T^\circ)=\min_{q\in F^{cp}(T)}\ell_{T^\circ}(q),\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\eeqq
This can be written as
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=1}\;\min_{q\in F^{cp}(T)}\ell_{T^\circ}(q)\leq \sqrt[n]{n! \vol(T^\circ)},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n),\eeqq
which, by Theorem \ref{Thm:MainProperty1}, is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{T\in A\left(T^\circ,\sqrt[n]{n! \vol(T^\circ)}\right)}\vol(T)\geq 1,\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\eeqq
Applying Proposition \ref{Prop:Ahomogenity2}, we finally conclude that Mahler's conjecture for centrally symmetric convex bodies is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{T\in A\left(T^\circ,\sqrt[n]{\vol(T^\circ)}\right)}\vol(T)\geq \frac{1}{n!},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\eeqq
By similar reasoning, Theorem \ref{Thm:MainProperty1} also guarantees the equivalence of the equality case of Mahler's conjecture for centrally symmetric convex bodies $T\subset\mathbb{R}^n$
\begin{equation}}\def\eeq{\end{equation} \vol(T)\vol(T^\circ) =\frac{4^n}{n!},\label{eq:mahlerpr2}\eeq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(T)=1}\;\min_{q\in F^{cp}(T)}\ell_{T^\circ}(q)= \sqrt[n]{n! \vol(T^\circ)},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n),\eeqq
and
\begin{equation}}\def\eeq{\end{equation} \min_{T\in A(T^\circ,\sqrt[n]{\vol(T^\circ)})}\vol(T) = \frac{1}{n!},\quad T\in \mathcal{C}^{cs}(\mathbb{R}^n).\label{eq:mahlerpr3}\eeq
Moreover, Theorem \ref{Thm:MainProperty1} guarantees the following: If $T^*$ is a solution of \eqref{eq:mahlerpr2} satisfying
\begin{equation*}}\def\eeqq{\end{equation*} \vol(T^*)=1\eeqq
(note that, applying Proposition \ref{Prop:invMahler}, the property of being a solution of \eqref{eq:mahlerpr2} is invariant under scaling), then it is a solution of \eqref{eq:mahlerpr3}. And conversely, if $T^*$ is a solution of \eqref{eq:mahlerpr3}, then it is also a solution of \eqref{eq:mahlerpr2}.
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:relations}]
The equivalence of (i), (ii), and (iii) follows from Theorem \ref{Thm:relationcapacity}. The equivalence of (i) and (iv) follows from Theorem \ref{Thm:MainProperty1}. The equivalence of (iv) and (v) can be concluded as within the proof of Corollary \ref{Cor:optimizationmain1}. For the case of strictly convex $T\subset\mathbb{R}^n$, the equivalence of (i) and (vi) follows from \cite[Theorem 1.3]{KruppRudolf2022}.
The addition that every equality case $(K^*,T^*)$ of any of the inequalities is also an equality case of all the others is guaranteed by Theorem \ref{Thm:MainProperty1}.
\epf
\section{Proof of Theorem \ref{Thm:main2}}\label{Sec:main2}
We recall a sligthly rephrased version of the main result of Haim-Kislev in \cite{Haim-Kislev2019}:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:HaimKislev}
Let $P\subset\mathbb{R}^{2n}$ be a convex polytope. Then, there is an action-minimizing closed characteristic $x$ on $\partial P$ which is a closed polygonal curve consisting of finitely many segments
\begin{equation*}}\def\eeqq{\end{equation*} [x(t_j),x(t_{j+1})]\eeqq
given by
\begin{equation*}}\def\eeqq{\end{equation*} x(t_{j+1})=x(t_j)+\lambda_j J\nabla H_P(x_j),\quad \lambda_j >0,\eeqq
while $x_j\in \mathring{F}_j$, $F_j$ is a facet of $P$ and $x$ visits every facet $F_j$ at most once.
\ethm
For the proof of Theorem \ref{Thm:main2}, we need the following theorem:
\begin{theorem}}\def\ethm{\end{theorem}\label{Thm:Estimate}
If $P\subset\mathbb{R}^{2n}$ is a convex polytope, then we have
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(P)= R_P \min_{q\in F^{cp}(P)}\ell_{\frac{JP}{2}}(q)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} R_P = \frac{\min_{q\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(q)}{\min_{q\in F^{cp}(P)}\ell_{\frac{JP}{2}}(q)}\geq 1.\eeqq
\ethm
We remark that, in light of Theorem \ref{Thm:relationcapacity}, Theorem \ref{Thm:Estimate} implies the following relationship between the EHZ-capacity of $P$ and the EHZ-capacity of the Lagrangian product $P\times \frac{1}{2}JP$:
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(P)=R_P c_{EHZ}\left(P\times \frac{1}{2}JP\right).\eeqq
For the proof of Theorem \ref{Thm:Estimate}, we need the following proposition. We remark that in the proof of Theorem \ref{Thm:Estimate}, we need it only in the case of action-minimizing closed characteristics on the boundary of a polytope. However, we will state it in full generality which has relevance beyond its use in the proof of Theorem \ref{Thm:Estimate} (which we will briefly address below).
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:characteristicgeodesic}
Let $C\subset\mathbb{R}^{2n}$ be a convex body. Let $x$ be any closed characteristic on $\partial C$. Then, the action of $x$ equals its $\ell_{\frac{JC}{2}}$-length:
\begin{equation*}}\def\eeqq{\end{equation*} \mathbb{A}(x)=\ell_{\frac{JC}{2}}(x).\eeqq
\eprop
Proposition \ref{Prop:characteristicgeodesic} implies a noteworthy connection between closed characteristics and closed Finsler geodesics: Every closed characteristic on $\partial C$ can be interpreted as a Finsler geodesic with respect to the Finsler metric determined by $\mu_{2JC^\circ}$ and which is parametrized by arc length. This raises a number of questions; for example, which Finsler geodesics are closed characteristics (we note that, usually, there are more geodesics than those which, by the least action principle and Proposition \ref{Prop:characteristicgeodesic}, can be associated to closed characteristics) and whether the length-minimizing Finsler geodesics are of this kind. Following this line of thought, would lead to the question whether it is possible to deduce Viterbo's conjecture from systolic inequalities for Finsler geodesics. However, we leave these questions for further research.
\begin{proof}}\def\epf{\end{proof}[Proof of Proposition \ref{Prop:characteristicgeodesic}]
By
\begin{equation*}}\def\eeqq{\end{equation*} \dot{x}(t)\in J\partial H_C(x(t))\quad \text{a.e.},\eeqq
we conclude
\begin{equation*}}\def\eeqq{\end{equation*} \frac{1}{2}\left(\mu_{2JC^\circ}(\dot{x}(t))\right)^2=H_{2JC^\circ}(\dot{x}(t))\in H_{2JC^\circ}(J\partial H_C(x(t)))=\frac{1}{4} H_{C^\circ}(\partial H_C(x(t)))\quad \text{a.e.},\eeqq
where we used the facts
\begin{equation*}}\def\eeqq{\end{equation*} J^{-1}=-J,\quad H_C(Jx)=H_{J^{-1}C}(x) \eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} H_{\lambda C}(x)=H_{C}\left(\frac{1}{\lambda}x\right)=\frac{1}{\lambda^2}H_C(x),\quad \lambda \neq 0,\eeqq
(cf.\;\cite[Proposition 2.3(iii)]{KruppRudolf2022}). From Proposition \ref{Prop:ConsEuler}, we therefore conclude
\begin{equation*}}\def\eeqq{\end{equation*} \frac{1}{2}\left(\mu_{2JC^\circ}(\dot{x}(t))\right)^2=\frac{1}{4}H_C(x(t))=\frac{1}{8} \quad \text{a.e.}\eeqq
and consequently
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{2JC^\circ}(\dot{x}(t))=\frac{1}{2} \quad \text{a.e.}\eeqq
Considering
\begin{equation*}}\def\eeqq{\end{equation*} (2JC^\circ)^\circ = \frac{1}{2}JC\eeqq
(cf.\;\cite[Proposition 2.1]{KruppRudolf2022}), we obtain
\begin{align*}
\ell_{\frac{JC}{2}}(x)=\int_0^T \mu_{\left(\frac{JC}{2}\right)^\circ}(\dot{x}(t))\,\mathrm{d}t = \int_0^T \mu_{2JC^\circ}(\dot{x}(t))\,\mathrm{d}t=\int_0^T \frac{1}{2}\,\mathrm{d}t=\frac{T}{2}=\mathbb{A}(x),
\end{align*}
where the last equality follows from
\begin{equation*}}\def\eeqq{\end{equation*} \mathbb{A}(x)=-\frac{1}{2}\int_0^T \langle J\dot{x}(t),x(t)\rangle \, \mathrm{d}t \in \frac{1}{2}\int_0^T \langle \partial H_C(x(t)),x(t)\rangle\, \mathrm{d}t\eeqq
which by Proposition \ref{Prop:Euleridentity} and the $2$-homogeneity of $H_C$ implies
\begin{equation*}}\def\eeqq{\end{equation*} \mathbb{A}(x) =\int_0^T H_C(x(t)) \,\mathrm{d}t = \frac{T}{2}.\eeqq
\epf
Then, we come to the proof of Theorem \ref{Thm:Estimate}:
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:Estimate}]
The idea behind the proof is to associate action-minimizing closed characteristics on $\partial P$ in the sense of Theorem \ref{Thm:HaimKislev} with $\ell_{\frac{1}{2}JP}$-minimizing closed $(P,\frac{JP}{2})$-Minkowski billiard trajectories.
Let $x$ be an action-minimizing closed characteristic on $\partial P$ in the sense of Theorem \ref{Thm:HaimKislev}. Let us assume $x$ is moving on the facets of $P$ according to the order
\begin{equation*}}\def\eeqq{\end{equation*} F_1\rightarrow F_2 \rightarrow ... \rightarrow F_m \rightarrow F_1,\eeqq
while the linear flow on every facet is given by the $J$-rotated normal vector at the interior of this facet. Out of every trajectory segment
\begin{equation*}}\def\eeqq{\end{equation*} {\rm orb}(x)\cap \mathring{F}_j,\eeqq
we choose one point $q_j$ arbitrarily (on the whole requiring $q_i\neq q_j$ for $i\neq j$) and connect these points by straight lines (by maintaining the order of the corresponding facets) constructing a closed polygonal curve
\begin{equation*}}\def\eeqq{\end{equation*} q:=(q_1,...,q_m)\eeqq
within $P$ which has its vertices on $\partial P$. From Lemma \ref{Lem:Pre2} (which we provide subsequently), we derive
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{\frac{JP}{2}}(q)=\ell_{\frac{JP}{2}}(x)\eeqq
since the trajectory segment of $x$ between the two consecutive points $q_j$ and $q_{j+1}$--let us call it ${\rm orb}(x)_{q_j\rightarrow q_{j+1}}$--together with the line from $q_j$ to $q_{j+1}$ (as trajectory segment of $q$)--let us call it $[q_j,q_{j+1}]$--builds a triangle with the property that
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{2JP^\circ}\left({\rm orb}(x)_{q_j\rightarrow q_{j+1}}\right)=\mu_{2JP^\circ}([q_j,q_{j+1}]).\eeqq
We therefore conclude from Proposition \ref{Prop:characteristicgeodesic} that
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{\frac{JP}{2}}(q) = \mathbb{A}(x).\eeqq
Because of the arbitrariness of the choice of $q_j$ within ${\rm orb}(x)\cap \mathring{F}_j$, we can assign infinitely many different closed polygonal curves of the above kind to one action-minimizing closed characteristic fulfilling the demanded conditions.
Each of these closed polygonal curves $q$ is a closed $(P\times \frac{1}{2}JP)$-Minkowski billiard trajectory: This follows from the fact that $q$ fulfills
\begin{equation*}}\def\eeqq{\end{equation*} \begin{cases}q_{j+1}-q_j\in N_{\frac{1}{2}JP}(p_j),\\p_{j+1}-p_j\in -N_P(q_{j+1}),\end{cases}\eeqq
for the closed polygonal curve $p=(p_1,...,p_m)$ in $\frac{1}{2}JP$ with
\begin{equation*}}\def\eeqq{\end{equation*} p_{j-1}\in \partial \left(\frac{1}{2}JP\right)\eeqq
given as the intersection point
\begin{equation*}}\def\eeqq{\end{equation*} \frac{1}{2}J\left(\{q_{j-1}+tJ\nabla H_P(q_{j-1}):t\in\mathbb{R}\}\cap \{q_{j}+tJ\nabla H_P(q_{j}):t\in\mathbb{R}\}\right) \subset \frac{JF_{j-1}}{2} \cap \frac{JF_{j}}{2} \eeqq
for all $j\in\{2,...,m+1\}$.
Indeed, from the definition of $p$, it follows
\begin{equation}}\def\eeq{\end{equation} p_{j+1}-p_j\in -N_P(q_{j+1}) \quad \forall j\in\{1,...,m\}\label{eq:minusexplanation}\eeq
since, by construction, $p_{j+1}-p_j$ is a multiple of the outer normal vector at $P$ in $q_j$ rotated by twofold multiplication with $J$ ($J^2=-\mathbb{1}$ produces the minus sign in \eqref{eq:minusexplanation}). Since, by construction,
\begin{equation*}}\def\eeqq{\end{equation*} J^{-1}(q_j-q_{j-1})\eeqq
is in the normal cone at $P$ in the intersection point
\begin{equation*}}\def\eeqq{\end{equation*} \{q_{j-1}+tJ\nabla H_P(q_{j-1}):t\in\mathbb{R}\}\cap \{q_{j}+tJ\nabla H_P(q_{j}):t\in\mathbb{R}\} \subset F_{j-1} \cap F_{j},\eeqq
roation by $\frac{1}{2}J$ then implies that $q_j-q_{j-1}$ is in the normal cone at $\frac{1}{2}JP$ in $p_{j-1}$. This implies
\begin{equation*}}\def\eeqq{\end{equation*} q_{j}-q_{j-1}\in N_P(p_{j-1})\quad \forall j\in\{1,...,m\}.\eeqq
\begin{figure}[h!]
\centering
\def380pt{420pt}
\input{images/Billiard.pdf_tex}
\caption[Association of closed characteristics to closed Minkowski billiard trajectories]{$q=(q_1,...,q_m)$ is a closed $(P,\frac{1}{2}JP)$-Minkowski billiard trajectory with $p=(p_1,...,p_m)$ as its dual billiard trajectory in $\frac{1}{2}JP$.}
\label{img:Billiard}
\end{figure}
From \cite[Proposition 3.9]{KruppRudolf2022}, it follows that $q$ cannot be translated into $\mathring{P}$, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} q\in F^{cp}(P).\eeqq
From the construction of $q$, we moreover know
\begin{equation}}\def\eeq{\end{equation} q\in F^{cp}_*(P),\label{eq:wormoneinequality}\eeq
where we recall that $F^{cp}_*(P)$ as subset of $F^{cp}(P)$ was defined as the set of all closed polygonal curves $q=(q_1,...,q_m)$ in $F^{cp}(P)$ for which $q_j$ and $q_{j+1}$ are on neighbouring facets $F_j$ and $F_{j+1}$ of $P$ such that there are $\lambda_j,\mu_{j+1}\geq 0$ with
\begin{equation*}}\def\eeqq{\end{equation*} q_{j+1}=q_j + \lambda_j J\nabla H_P(x_j)+\mu_{j+1}J\nabla H_P(x_{j+1}),\eeqq
where $x_j$ and $x_{j+1}$ are arbitrarily chosen interior points of $F_j$ and $F_{j+1}$, respectively.
Because of \eqref{eq:wormoneinequality}, we have
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{\frac{JP}{2}}(q) \geq \min_{\widetilde{q}\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(\widetilde{q}).\eeqq
Since, by definition and the above considerations, every closed polygonal curve in $F^{cp}_*(P)$ is associated with a closed characteristic on $\partial P$, where the $\ell_{\frac{JP}{2}}$-length of the former coincides with the action of the latter, and $x$ (to which $q$ is associated) was chosen to be action-minimizing, we actually have
\begin{equation*}}\def\eeqq{\end{equation*} \ell_{\frac{JP}{2}}(q) = \min_{\widetilde{q}\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(\widetilde{q}).\eeqq
Altogether, this implies
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(P)=\mathbb{A}(x)=\ell_{\frac{JP}{2}}(x)=\ell_{\frac{JP}{2}}(q)=\min_{\widetilde{q}\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(\widetilde{q})=R_P \min_{\widetilde{q}\in F^{cp}(P)}\ell_{\frac{JP}{2}}(\widetilde{q})\eeqq
for
\begin{equation*}}\def\eeqq{\end{equation*} R_P = \frac{\min_{q\in F^{cp}_*(P)}\ell_{\frac{JP}{2}}(q)}{\min_{q\in F^{cp}(P)}\ell_{\frac{JP}{2}}(q)}\geq 1.\eeqq
\epf
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:Pre2}
Let $P\subset\mathbb{R}^{2n}$ be a convex polytope. If
\begin{equation*}}\def\eeqq{\end{equation*} y=\lambda_iJ\nabla H_P(x_i)+\lambda_j J\nabla H_P(x_j), \; \lambda_i,\lambda_j\geq 0,\eeqq
where $F_i$ and $F_j$ are neighbouring facets of $P$ with $x_i\in \mathring{F}_i$ and $x_j\in \mathring{F}_j$, then
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{2JP^{\circ}}(y)=\lambda_i\mu_{2JP^{\circ}}(J\nabla H_P(x_i))+\lambda_j\mu_{2JP^{\circ}}(J\nabla H_P(x_j))=\frac{1}{2}(\lambda_i + \lambda_j).\eeqq
\elem
\begin{proof}}\def\epf{\end{proof}
We first notice that
\begin{equation*}}\def\eeqq{\end{equation*} \nabla H_P(x_i) \; \text{ and } \; \nabla H_P(x_j)\eeqq
are neighbouring vertices of $P^\circ$, i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} t \nabla H_P(x_i)+(1-t)\nabla H_P(x_j)\in \partial P^\circ\;\;\; \forall t\in[0,1].\eeqq
Indeed, from the fact that $\nabla H_P(x_i)$ and $\nabla H_P(x_j)$ are elements of the one dimensional normal cone at $\mathring{F}_i$ and $\mathring{F}_j$, we conclude by the properties of the polar of convex polytopes (cf.\;\cite[Chapter 3.3]{GallierQuaintance2017}) that they point into the direction of two neigbouring vertices of $P^\circ$. Using Proposition \ref{Prop:ConsEuler}, we calculate
\begin{equation*}}\def\eeqq{\end{equation*} H_{P^\circ}(\nabla H_P(x_{i}))=H_P(x_{i})=\frac{1}{2}\eeqq
and
\begin{equation*}}\def\eeqq{\end{equation*} H_{P^\circ}(\nabla H_P(x_{j}))=H_P(x_{j})=\frac{1}{2}\eeqq
and conclude that $\nabla H_P(x_i)$ and $\nabla H_P(x_j)$ actually are these two neighbouring vertices of $P^\circ$.
Using for convex body $C\subset\mathbb{R}^{2n}$ and $\lambda > 0$ the properties
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{\lambda C}(x)=\frac{1}{\lambda}\mu_{C}(x) \;\text{ and }\;\mu_{J C}(Jx)=\mu_C(x),\;\; x\in\mathbb{R}^{2n},\eeqq
(cf.\;\cite[Proposition 2.3(iii)]{KruppRudolf2022}), we derive
\allowdisplaybreaks{\begin{align*}
\mu_{2JP^\circ}(y)&=\mu_{2JP^\circ}(\lambda_iJ\nabla H_P(x_i)+\lambda_j J\nabla H_P(x_j))\\
&= \mu_{2P^\circ}(\lambda_i\nabla H_P(x_i)+\lambda_j\nabla H_P(x_j))\\
&=\frac{1}{2}\left(\mu_{P^\circ}(\lambda_i\nabla H_P(x_i)+\lambda_j\nabla H_P(x_j))\right)\\
&\stackrel{(\star)}{=} \frac{1}{2}\left(\mu_{P^{\circ}}(\lambda_i\nabla H_P(x_i))+\mu_{P^{\circ}}(\lambda_j\nabla H_P(x_j))\right)\\
&= \frac{1}{2}\left(\lambda_i\mu_{P^{\circ}}(\nabla H_P(x_i))+\lambda_j\mu_{P^{\circ}}(\nabla H_P(x_j))\right)\\
&=\frac{1}{2}(\lambda_i+\lambda_j),
\end{align*}}%
where in $(\star)$ we used that, by the choice of $x_i$ and $x_j$ and the properties of polar bodies, $\nabla H_P(x_i)$ and $\nabla H_P(x_j)$ are neighbouring vertices of $P^\circ$ and, therefore, in $(\star)$, the initial term can be splitted linearly.
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:main2}]
Viterbo's conjecture for convex polytopes in $\mathbb{R}^{2n}$ can be written as
\begin{equation*}}\def\eeqq{\end{equation*} \vol(P)\geq \frac{c_{EHZ}(P)^n}{n!},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right),\eeqq
which, by Theorem \ref{Thm:Estimate}, is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \vol(P)\geq \frac{R_P^n}{2^n n!}c_{EHZ}(P\times JP)^n,\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right).\eeqq
By referring to Proposition \ref{Prop:invViterbo}, we can assume
\begin{equation*}}\def\eeqq{\end{equation*} \vol(P)=1\eeqq
without loss of generality and get
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(P\times JP)\leq \frac{2\sqrt[n]{n!}}{R_P},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right),\eeqq
which, by Theorem \ref{Thm:relationcapacity}, is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(P)=1}\; \min_{q\in F^{cp}(P)}\ell_{JP}(q) \leq \frac{2\sqrt[n]{n!}}{R_P},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right).\eeqq
By Theorem \ref{Thm:MainProperty1}, this is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{P\in A\left(JP,\frac{2\sqrt[n]{n!}}{R_P}\right)}\vol(P)\geq 1,\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right),\eeqq
and, after applying Proposition \ref{Prop:Ahomogenity2}, to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{P\in A\left(JP,\frac{1}{R_{P}}\right)}\vol(P)\geq \frac{1}{2^n n!},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right).\eeqq
By similar reasoning, Theorem \ref{Thm:MainProperty1} also guarantees the equivalence of
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(P)=1}\; \min_{q\in F^{cp}(P)}\ell_{JP}(q) = \frac{2\sqrt[n]{n!}}{R_P},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right),\label{eq:proofmain21}\eeq
and
\begin{equation}}\def\eeq{\end{equation} \min_{P\in A\left(JP,\frac{1}{R_{P}}\right)}\vol(P)= \frac{1}{2^n n!},\quad P\in\mathcal{C}^p\left(\mathbb{R}^{2n}\right).\label{eq:proofmain22}\eeq
Moreover, Theorem \ref{Thm:MainProperty1} guarantees the following: $P^*$ is a solution of \eqref{eq:proofmain21} if and only if $P^*$ is a solution of \eqref{eq:proofmain22}.
\epf
\section{Proof of Theorem \ref{Thm:main3}}\label{Sec:main3}
In order to prove Theorem \ref{Thm:main3}, we need the following propositon:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:characteristicnotranslate}
Let $C\subset\mathbb{R}^{2n}$ be a convex body and $x$ a closed characteristic on $\partial C$. Then, $x$ cannot be translated into $\mathring{C}$.
\eprop
\begin{proof}}\def\epf{\end{proof}
Let us assume that $x$ can be translated into $\mathring{C}$. Let $\widetilde{T}>0$ be the period of $x$. Because of
\begin{equation*}}\def\eeqq{\end{equation*} \dot{x}(t)\in J\partial H_C(x(t))\quad \text{a.e. on }[0,\widetilde{T}],\eeqq
there is a vector-valued function $n_C$ on $\partial C$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \dot{x}(t)=Jn_C(x(t))\quad \text{a.e. on }[0,\widetilde{T}]\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} n_C(x(t))\in \partial H_C(x(t))\eeqq
for all $t\in[0,\widetilde{T}]$ for which $\dot{x}(t)$ exists and
\begin{equation*}}\def\eeqq{\end{equation*} n_C(x(t))=0\eeqq
for all $t\in[0,\widetilde{T}]$ for which $\dot{x}(t)$ does not exist.
Then, the convex cone $U$ spanned by
\begin{equation*}}\def\eeqq{\end{equation*} n_C(x(t))\in N_C(x(t)),\;t\in [0,\widetilde{T}],\eeqq
has the property
\begin{equation*}}\def\eeqq{\end{equation*} \forall u\in U\setminus\{0\}:\;-u\notin U,\eeqq
since, otherwise, one could find points on $x$ and $C$-supporting hyperplanes through these points with the property that the intersection of the $C$-containing half-spaces bounded by these hyperplanes is nearly bounded (what would imply that $x$ cannot be translated into $\mathring{C}$). By the convexity of $U$, this implies that
\begin{equation*}}\def\eeqq{\end{equation*} \int_0^{\widetilde{T}} n_C(x(t))\, \mathrm{d}t \neq 0,\eeqq
and therefore
\begin{equation*}}\def\eeqq{\end{equation*} \int_0^{\widetilde{T}} Jn_C(x(t))\, \mathrm{d}t \neq 0.\eeqq
Since $x$ is a closed characteristic on $\partial C$, $x$ fulfills $x(0)=x(\widetilde{T})$. This implies
\begin{equation*}}\def\eeqq{\end{equation*} 0=x(\widetilde{T})-x(0)=\int_0^{\widetilde{T}} \dot{x}(t)\, \mathrm{d}t = \int_0^{\widetilde{T}} Jn_C(x(t))\, \mathrm{d}t \neq 0,\eeqq
a contradiction. Therefore, it follows that $x$ cannot be translated into $\mathring{C}$.
\epf
We now consider the operator norm of the complex structure/symplectic matrix $J$. It is given by:
\begin{equation*}}\def\eeqq{\end{equation*} ||J||_{C^\circ\rightarrow C}=\sup_{||v||_{C^\circ}\leq 1}||Jv||_{C}=\sup_{\mu_{C^\circ}(v)\leq 1} \mu_C(Jv).\eeqq
We derive the following lemma:
\begin{lemma}}\def\elem{\end{lemma}\label{Lem:Karasev}
Let $C\subset\mathbb{R}^{2n}$ be a convex body and $x$ a closed characteristic on $\partial C$ which has period $\widetilde{T}>0$. Then, we have
\begin{equation*}}\def\eeqq{\end{equation*} \mu_C(\dot{x}(t)) \leq ||J||_{C^\circ \rightarrow C}\quad \text{a.e. on }[0,\widetilde{T}].\eeqq
\elem
\begin{proof}}\def\epf{\end{proof}
Since $x$ is a closed characteristic on $\partial C$, we have
\begin{equation*}}\def\eeqq{\end{equation*} \dot{x}(t)\in J\partial H_C(x(t))\quad \text{a.e. on }[0,\widetilde{T}].\eeqq
This implies
\begin{equation*}}\def\eeqq{\end{equation*} H_{C^\circ}(-J\dot{x}(t))\in H_{C^\circ}(\partial H_C(x(t)))\quad \text{a.e. on }[0,\widetilde{T}].\eeqq
Using Proposition \ref{Prop:ConsEuler}, we conclude
\begin{equation*}}\def\eeqq{\end{equation*} H_{C^\circ}(-J\dot{x}(t))=H_C(x(t))=\frac{1}{2}\quad \text{a.e. on }[0,\widetilde{T}],\eeqq
i.e.,
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{C^\circ}(-J\dot{x}(t))=1\quad \text{a.e. on }[0,\widetilde{T}].\eeqq
Therefore, for
\begin{equation*}}\def\eeqq{\end{equation*} v(t):=-J\dot{x}(t) \quad \text{a.e. on }[0,\widetilde{T}],\eeqq
we have
\begin{equation*}}\def\eeqq{\end{equation*} \mu_{C^\circ}(v(t))=1 \; \text{ and } \; Jv(t)=\dot{x}(t) \quad \text{a.e. on }[0,\widetilde{T}]\eeqq
and consequently
\begin{equation*}}\def\eeqq{\end{equation*} \mu_C(\dot{x}(t)) \leq \sup_{\mu_{C^\circ}(v)\leq 1}\mu_C(Jv) = ||J||_{C^\circ\rightarrow C}\quad \text{a.e. on }[0,\widetilde{T}].\eeqq
\epf
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:main3}]
By Theorem \ref{Thm:relationcapacity}, we have
\begin{equation}}\def\eeq{\end{equation} c_{EHZ}(C\times C^\circ)=\min_{q\in F^{cp}(C)}\ell_{C^\circ}(q).\label{eq:main31}\eeq
Let $x$ be an action-minimizing closed characteristic on $\partial C$, i.e., $x$ fulfills
\begin{equation*}}\def\eeqq{\end{equation*} \dot{x}\in J\partial H_C(x)\quad \text{a.e.}\eeqq
and minimizes the action with
\begin{equation}}\def\eeq{\end{equation} \mathbb{A}(x)=-\frac{1}{2}\int_0^{\widetilde{T}} \langle J\dot{x}(t),x(t)\rangle\, \mathrm{d}t=\int_0^{\widetilde{T}} H_C(x)\, \mathrm{d}t=\frac{\widetilde{T}}{2},\label{eq:main33}\eeq
where we used Euler's identity (cf.\;Proposition \ref{Prop:Euleridentity}) to derive
\begin{equation*}}\def\eeqq{\end{equation*} \langle y,x(t)\rangle = H_C(x(t))\quad \forall y \in \partial H_C(x(t)).\eeqq
Then, since $x$ is in $\partial C$ and, by Proposition \ref{Prop:characteristicnotranslate}, cannot be translated into $\mathring{C}$, \eqref{eq:main31} together with
\begin{equation}}\def\eeq{\end{equation} \min_{q\in F^{cp}(C)}\ell_{C^\circ}(q)=\min_{q\in F^{cc}(C)}\ell_{C^\circ}(q)\label{eq:main34}\eeq
(cf.\;Proposition \ref{Prop:selfevident2}) implies that
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(C\times C^\circ)\leq \ell_{C^\circ}(x) = \int_0^{\widetilde{T}} \mu_C(\dot{x}(t))\, \mathrm{d}t.\eeqq
Using Lemma \ref{Lem:Karasev} and \eqref{eq:main33}, we conclude
\begin{align*}
c_{EHZ}(C\times C^\circ)\leq \int_0^{\widetilde{T}} \mu_C(\dot{x}(t))\, \mathrm{d}t &\leq \int_0^{\widetilde{T}} ||J||_{C^\circ \rightarrow C}\, \mathrm{d}t\\
& = \widetilde{T} ||J||_{C^\circ \rightarrow C}\\
& = 2 \mathbb{A}(x)||J||_{C^\circ \rightarrow C}\\
& = 2c_{EHZ}(C)||J||_{C^\circ \rightarrow C}.
\end{align*}
This implies
\begin{equation*}}\def\eeqq{\end{equation*} \widetilde{R}_C=\frac{c_{EHZ}(C)}{c_{EHZ}(C\times C^\circ)} \geq \frac{1}{2||J||_{C^\circ \rightarrow C}}.\eeqq
Therefore, Viterbo's conjecture for convex bodies in $\mathbb{R}^{2n}$ is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \vol(C)\geq \frac{c_{EHZ}(C)^n}{n!}=\frac{\widetilde{R}_C^n c_{EHZ}(C\times C^\circ)^n}{n!},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right).\eeqq
By referring to Proposition \ref{Prop:invViterbo}, we can assume
\begin{equation*}}\def\eeqq{\end{equation*} \vol(C)=1\eeqq
without loss of generality and get
\begin{equation*}}\def\eeqq{\end{equation*} c_{EHZ}(C\times C^\circ)\leq \frac{\sqrt[n]{n!}}{\widetilde{R}_C},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right),\eeqq
which, by Theorem \ref{Thm:relationcapacity}, is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\vol(C)=1}\; \min_{q\in F^{cp}(C)}\ell_{C^\circ}(q) \leq \frac{\sqrt[n]{n!}}{\widetilde{R}_C},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right).\eeqq
By Theorem \ref{Thm:MainProperty1}, this is equivalent to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{C\in A\left(C^\circ,\frac{\sqrt[n]{n!}}{\widetilde{R}_C}\right)}\vol(T)\geq 1,\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right),\eeqq
and, after applying Proposition \ref{Prop:Ahomogenity2}, to
\begin{equation*}}\def\eeqq{\end{equation*} \min_{C\in A\left(C^\circ,\frac{1}{\widetilde{R}_C}\right)}\vol(T)\geq \frac{1}{n!},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right).\eeqq
By similar reasoning, Theorem \ref{Thm:MainProperty1} also guarantees the equivalence of
\begin{equation}}\def\eeq{\end{equation} \max_{\vol(C)=1}\; \min_{q\in F^{cp}(C)}\ell_{C^\circ}(q) = \frac{\sqrt[n]{n!}}{\widetilde{R}_C},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right),\label{eq:main35}\eeq
and
\begin{equation}}\def\eeq{\end{equation} \min_{C\in A\left(C^\circ,\frac{1}{\widetilde{R}_C}\right)}\vol(T) = \frac{1}{n!},\quad C\in\mathcal{C}\left(\mathbb{R}^{2n}\right).\label{eq:main36}\eeq
Moreover, Theorem \ref{Thm:MainProperty1} guarantees the following: $C^*$ is a solution of \eqref{eq:main35} if and only if $C^*$ is a solution of \eqref{eq:main36}.
\epf
\section{Justification of Conjectures \ref{Conj:Wetzelsproblem} and \ref{Conj:WetzelsproblemBilliard}}\label{Sec:Wetzelproblem}
We transfer Viterbo's conjecture onto Wetzel's problem. For that, we define
\begin{equation*}}\def\eeqq{\end{equation*} y:=\min_{K\in A\left(B_1^2,1\right)} \vol(K)\eeqq
and let $K^*\subset\mathbb{R}^2$ be an arbitrarily chosen convex body of volume $y$. Then, applying Theorems \ref{Thm:relationcapacity} and \ref{Thm:MainProperty1}, we have
\begin{align*}
\frac{c_{EHZ}\left(K^*\times B_1^2\right)^2}{2} &\leq \max_{\vol(K)=y}\frac{c_{EHZ}\left(K\times B_1^2\right)^2}{2}\\
& = \max_{\vol(K)=y}\;\min_{q\in F^{cp}(K)}\frac{\ell_{B_1^2}(q)^2}{2}\\
& =\frac{1}{2}.
\end{align*}
Further, we have
\begin{equation*}}\def\eeqq{\end{equation*} \vol\left(K^* \times B_1^2\right)=\pi y.\eeqq
The truth of Viterbo's conjecture requires
\begin{equation*}}\def\eeqq{\end{equation*} \vol\left(K^* \times B_1^2\right)\geq \frac{c_{EHZ}\left(K^* \times B_1^2\right)^2}{2},\eeqq
i.e., $\pi y \geq \frac{1}{2}$, which means
\begin{equation*}}\def\eeqq{\end{equation*} y\geq \frac{1}{2\pi}\approx 0.15915.\eeqq
Theorem \ref{Thm:MainProperty1} also guarantees the sharpness of this estimate.
Together with Theorem \ref{Thm:relationcapacity}, this justifies the formulation of Conjectures \ref{Conj:Wetzelsproblem} and \ref{Conj:WetzelsproblemBilliard}.
\section{Proof of Theorem \ref{Thm:analogueFinchWetzel}}\label{Sec:escape}
In order to prove Theorem \ref{Thm:analogueFinchWetzel}, we start with the two following obvious observations:
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:selfevident}
Let $K\subset\mathbb{R}^n$ be a convex body. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} \{\text{closed Minkowski escape paths for $K$}\} = F^{cc}(K).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
The statement follows directly by recalling that a closed Minkowski escape path is a closed curve whose all translates intersect $\partial K$ and therefore, equivalently, cannot be translated into $\mathring{K}$.
\epf
\begin{proposition}}\def\eprop{\end{proposition}\label{Prop:selfevident2}
Let $K,T\subset\mathbb{R}^n$ be convex bodies. Then we have
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cc}(K)}\ell_T(q) = \min_{q\in F^{cp}(K)}\ell_T(q).\eeqq
\eprop
\begin{proof}}\def\epf{\end{proof}
Since
\begin{equation*}}\def\eeqq{\end{equation*} F^{cp}(K) \subset F^{cc}(K),\eeqq
it suffices to find for every closed curve $q\in F^{cc}(K)$ a closed polygonal curve $\widetilde{q}\in F^{cp}(K)$ with
\begin{equation}}\def\eeq{\end{equation} \ell_T(\widetilde{q}) \leq \ell_T(q).\label{eq:selfevident2}\eeq
If $q$ cannot be translated into $\mathring{K}$, then by the remark beyond \cite[Lemma 2.1]{KruppRudolf2020}, there are $n+1$ points on $q$ that cannot be translated into $\mathring{K}$. By connecting these points, we obtain a closed polygonal curve in $F^{cp}(K)$ which we call $\widetilde{q}$. By the subadditivity of the Minkowski functional, it follows \eqref{eq:selfevident2}.
\epf
Based on these propositions, we can prove the analogue to Theorem \ref{Thm:FinchWetzel}:
\begin{proof}}\def\epf{\end{proof}[Proof of Theorem \ref{Thm:analogueFinchWetzel}]
We first use Proposition \ref{Prop:selfevident} in order to reduce the statement of Theorem \ref{Thm:analogueFinchWetzel} to: An/The $\ell_T$-minimizing closed curve in $F^{cc}(K)$ has $\ell_T$-length $\alpha^*$ if and only if $\alpha^*$ is the largest $\alpha$ for which
\begin{equation}}\def\eeq{\end{equation} K\in A(T,\alpha).\label{eq:analogue0}\eeq
First, let us asssume that $\alpha^*$ is the $\ell_T$-length of an/the $\ell_T$-minimizing closed curve in $F^{cc}(K)$. Then, from Proposition \ref{Prop:selfevident2}, we know that there is a closed polygonal curve
\begin{equation*}}\def\eeqq{\end{equation*} q^*\in F^{cp}(K) \; \text{ with } \; \ell_T(q^*)=\alpha^*,\eeqq
i.e., $q^*$ is a minimizer of
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cp}(K)} \ell_T(q).\eeqq
Then it follows from Proposition \ref{Prop:p^*minF} that
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\ell_T(q^*))=A(T,\alpha^*).\eeqq
Let $\alpha > \alpha^*$. If
\begin{equation}}\def\eeq{\end{equation} K\in A(T,\alpha),\label{eq:analogue1}\eeq
then
\begin{equation*}}\def\eeqq{\end{equation*} L_T(\alpha)\subseteq C(K),\eeqq
i.e., every closed curve of $\ell_T$-length $\alpha$ can be covered by a translate of $K$. This implies that every closed curve of $\ell_T$-length $\lambda \alpha $, $\lambda <1$, can be covered by a translate of $\mathring{K}$. From this we conclude
\begin{equation*}}\def\eeqq{\end{equation*} q^* \notin F^{cp}(K).\eeqq
Therefore, there is no $\alpha > \alpha^*$ for which \eqref{eq:analogue1} is fulfilled, i.e., $\alpha^*$ is the largest $\alpha$ for which \eqref{eq:analogue0} holds.
Conversely, if $\alpha^*$ is the largest $\alpha$ for which \eqref{eq:analogue0} holds. Then, there is a closed curve $q^*$ with
\begin{equation}}\def\eeq{\end{equation} q^*\in F^{cc}(K)\cap C(K) \; \text{ and } \; \ell_T(q^*)=\alpha^*.\label{eq:analogue2}\eeq
Otherwise, if not, then one has
\begin{equation*}}\def\eeqq{\end{equation*} q\in C(K)\setminus F^{cc}(K)\eeqq
for all closed curves $q$ of $\ell_T$-length $\alpha^*$. This implies
\begin{equation*}}\def\eeqq{\end{equation*} q\in C(\mathring{K})\eeqq
for all closed curves $q$ of $\ell_T$-length $\alpha^*$. But then there is a $\lambda >1$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \lambda q \in C(\mathring{K})\eeqq
for all closed curves of $\ell_T$-length $\alpha^*$. But this is a contradiction to the fact that $\alpha^*$ is the largest $\alpha$ for which \eqref{eq:analogue0} holds.
Now, if
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cc}(K)}\ell_T(q)=:\widetilde{\alpha}<\alpha^*\eeqq
and $\widetilde{q}$ is a minimizer of the left side, then it follows
\begin{equation*}}\def\eeqq{\end{equation*} \widetilde{q}\in C(K)\eeqq
because, due to Proposition \ref{Prop:cases1}, with $\widetilde{\alpha}<\alpha^*$ one has
\begin{equation*}}\def\eeqq{\end{equation*} K\in A(T,\alpha^*)\subseteq A(T,\widetilde{\alpha}).\eeqq
Then, with Lemma \ref{Lem:intersectionFC}, there is a $\lambda >1$ such that
\begin{equation*}}\def\eeqq{\end{equation*} \lambda \widetilde{q}\in F^{cc}(K)\setminus C(K)\eeqq
with
\begin{equation*}}\def\eeqq{\end{equation*} \ell_T(\lambda\widetilde{q})<\alpha^*.\eeqq
But this is a contradiction to the fact that every closed curve of $\ell_T$-length $\leq \alpha^*$ can be covered by a translate of $K$. Therefore, it follows
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cc}(K)}\ell_T(q) \geq \alpha^*,\eeqq
and together with \eqref{eq:analogue2}, we conclude that
\begin{equation*}}\def\eeqq{\end{equation*} \min_{q\in F^{cc}(K)}\ell_T(q) = \alpha^*.\eeqq
\epf
The proof of Corollary \ref{Cor:analogue} follows immediately:
\begin{proof}}\def\epf{\end{proof}[Proof of Corollary \ref{Cor:analogue}]
The proof follows directly by combining Proposition \ref{Prop:selfevident2}, \cite[Theorem 3.12]{KruppRudolf2022}, and Theorem \ref{Thm:analogueFinchWetzel}.
\epf
\section[Approach for improving the lower bound in Wetzel's problem]{Computational approach for improving the lower bound in Wetzel's problem}\label{Sec:OptimizationProblem}
In this section, we aim to present a computational approach for improving the best lower bound in Wetzel's problem, which, as stated in Theorem \ref{Thm:WetzelBezdekConnelly}, is due to Wetzel himself (cf.\;\cite{Wetzel1973}). But not only that, our approach most likely also allows to find, more generally, lower bounds in Minkowski worm problems. By Theorem \ref{Thm:MainProperty1}, these lower bounds eventually translate into upper bounds for systolic Minkowski billiard inequalities as well as for Viterbo's conjecture for convex Lagrangian products.
The main idea of this approach is inspired by a series of works related to the search for area-minimizing convex hulls of closed curves in the plane which are allowed to be translated and rotated. Since the area-minimizing convex cover for a set of closed curves is, equivalently, the area-minimizing convex hull of these closed curves (note that this observation has already used within the proof of Corollary \ref{Cor:optimizationmain1}), these works treat the question of lower bounds for the following version of Moser's worm problem in which \textit{closed} arcs are considered:\\
\par
\begingroup
\leftskip=1cm
\noindent \textit{Find a/the convex set of least area that contains a congruent copy of each closed arc in the plane of length one.}\\
\par
\endgroup
\noindent In \cite{ChakerianKlamkin1971} (applying results from \cite{FaryRedei1950}), the first lower bound for the area was found considering the convex hull of a circle and a line segment. In \cite{FurediWetzel2011}, this lower bound was improved by first considering a circle and a certain rectangle and later a circle and a curvilinear rectangle. The latest improvements are due to Grechuk and Som-am who in \cite{GrechukSomam2019} considered the convex hull of a circle, an equilateral triangle and a certain rectangle, and in \cite{GrechukSomam2020} the convex hull of a circle, a certain rectangle, and a line segment. However, in order to adapt these approaches to our setting, in the details, we have to make some changes.
But let us first start with some underlying considerations (as in the proof of Corollary \ref{Cor:optimizationmain1}) in the most general case: For arbitrary convex body $T\subset\mathbb{R}^n$, we ask for lower bounds of
\begin{equation}}\def\eeq{\end{equation} \min_{K\in A(T,1)}\vol(K).\label{eq:approach1}\eeq
By referring to the above mentioned main idea, we start by noting that for
\begin{equation*}}\def\eeqq{\end{equation*} q_1,...,q_k\in L_T(1)\eeqq
we have
\begin{equation}}\def\eeq{\end{equation} \min_{(a_1,...,a_k)\in (\mathbb{R}^n)^k}\;\vol\left(\conv\{q_1+a_1,...,q_k+a_k\}\right) \leq \min_{K\in A(T,1)}\vol(K).\label{eq:approach2}\eeq
This estimate can be further improved by
\begin{equation*}}\def\eeqq{\end{equation*} \max_{q_1,...,q_k\in L_T(1)}\;\min_{(a_1,...,a_k)\in (\mathbb{R}^n)^k}\;\vol(\conv\{q_1+a_1,...,q_k+a_k\}) \leq \min_{K\in A(T,1)}\vol(K),\eeqq
so that, eventually, we get
\begin{equation*}}\def\eeqq{\end{equation*} \min_{a_q\in\mathbb{R}^n}\vol\bigg(\conv\bigg\{ \bigcup_{q\in L_T(1)} (q+a_q)\bigg\}\bigg)=\min_{K\in A(T,1)}\vol(K),\eeqq
where the minimum on the left runs for every $q\in L_T(1)$ over all possible translations in $\mathbb{R}^n$.
Let us now exemplary show how \eqref{eq:approach2} can be used to calculate lower bounds of \eqref{eq:approach1} within the setting of Wetzel's problem, i.e., $n=2$ and $T=B_1^2$.
Let $q_1$ be the boundary of $B_{\frac{1}{2\pi}}^2$,
\begin{equation*}}\def\eeqq{\end{equation*} q_2=q_2(t_1,t_2,\theta)\eeqq
the boundary of an equilateral triangle $T_{t_1,t_2,\frac{1}{3},\theta}$ with mass point $(t_1,t_2)$, side length $\frac{1}{3}$, and angle $\theta$ between one of the sides and the horizontal line, and let
\begin{equation*}}\def\eeqq{\end{equation*} q_3=q_3\left(r_1,r_2,\widehat{q}\right)\eeqq
be the boundary of a rectangle $R_{r_1,r_2,1,\widehat{q}}$ with middle point $(r_1,r_2)$, perimeter $1$, and quotient of the side lengths $\widehat{q}$.
\begin{figure}[h!]
\centering
\def380pt{380pt}
\input{images/approach_grey.pdf_tex}
\caption{Illustration of the convex hull of $B_{\frac{1}{2\pi}}^2$, $R_{r_1,r_2,1,\widehat{q}}$ and $T_{t_1,t_2,\frac{1}{3},\theta}$.}
\label{img:approach}
\end{figure}
Then, by definition, we have
\begin{equation*}}\def\eeqq{\end{equation*} q_1,\,q_2(t_1,t_2,\theta),\,q_3\left(r_1,r_2,\widehat{q}\right)\in L_{B_1^2}(1)\eeqq
for all
\begin{equation*}}\def\eeqq{\end{equation*} t_1,t_2\in\mathbb{R}, \, \theta \in \left[0,\frac{3\pi}{4}\right],\,r_1,r_2\geq 0,\,\widehat{q} >0\eeqq
and \eqref{eq:approach2} (because of $\theta\in\left[0,\frac{3\pi}{4}\right]$ and $\widehat{q}>0$, one has $k=\infty$) becomes
\begin{align*}
\max_{\theta\in[0,\frac{3\pi}{4}],\;\widehat{q}>0}\;\min_{t_1,t_2\in\mathbb{R}, \; r_1,r_2\geq 0}\; \vol &\left(\conv\left\{B_{\frac{1}{2\pi}}^2,T_{t_1,t_2,\frac{1}{3},\theta},R_{r_1,r_2,1,\widehat{q}}\right\}\right)\\
&\leq \min_{K\in A(B_1^2,1)}\vol(K).
\end{align*}
Then, one can define
\begin{equation*}}\def\eeqq{\end{equation*} f\left(t_1,t_2,r_1,r_2,\theta,\widehat{q}\right):=\vol \left(\conv\left\{B_{\frac{1}{2\pi}}^2,T_{t_1,t_2,\frac{1}{3},\theta},R_{r_1,r_2,1,\widehat{q}}\right\}\right)\eeqq
which is a convex function with respect to the first four coordinates $(t_1,t_2,r_1,r_2)$ (this can be shown similar to in \cite{GrechukSomam2019}) and compute
\begin{equation*}}\def\eeqq{\end{equation*} \max_{\theta\in[0,\frac{3\pi}{4}],\;\widehat{q}>0}\;\min_{t_1,t_2\in\mathbb{R}, \; r_1,r_2\geq 0}\; f\left(t_1,t_2,r_1,r_2,\theta,\widehat{q}\right).\eeqq
We leave it at that, starting with \eqref{eq:approach2}, gives us the ability to tackle many different Minkowski worm problems--in any dimension, for many different $T$s and by using diverse closed curves
\begin{equation*}}\def\eeqq{\end{equation*} q_1,...,q_k\in L_T(1).\eeqq
\section*{Acknowledgement}
This research is supported by the SFB/TRR 191 'Symplectic Structures in Geometry, Algebra and Dynamics', funded by the \underline{German Research Foundation}, and was carried out under the supervision of Alberto Abbondandolo (Ruhr-Universit\"at Bochum). The author is thankful to the supervisor's support.
\medskip
| c7f0756bcbaaa07e880627eead4e372967d1a855 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Surgical resection is a common procedure in the treatment of pediatric posterior fossa tumors. However, surgical resection is associated with potential serious complications~\cite{dubey2009complications}, such as postoperative cerebellar mutism~\cite{van1995transient,gelabert2001mutism,pollack1997posterior,10.1093/neuonc/noaa215.536} and resultant long-term neuropsychological dysfunction~\cite{grill2004critical,levisohn2000neuropsychological}. While brain tissue damage caused by the tumor itself may be minor, it is typically not possible to remove the tumor without impacting surrounding healthy brain tissue. Thus, surgical damage is often unavoidable. Understanding the relationship between the damaged cerebellar regions and postoperative neurological and cognitive outcomes has the potential to guide surgery and significantly affect quality of survival of patients. Therefore, a reliable localization and measure of cerebellar damage is needed and fundamental to conduct this type of investigation.
Defining the volume-of-interest (VOI) for cerebellar damage in the postoperative brain MR image is challenging, even for human experts performing this task manually. First, the damage we want to identify herein is the ``loss of normal brain tissue''. Instead of showing an abnormal intensity in the MR image, this missing tissue presents as just an empty space (cavity) filled with cerebrospinal fluid (CSF). In addition, an abnormal CSF distribution does not necessarily reflect damage, because normal brain tissues may have already been displaced by the presence of the tumor and disease complications (e.g., hydrocephalus) before the surgery. Presence of the tumor near the fourth ventricle complicates this issue, as the ventricle space is connected to the postsurgical cavity. Given these challenges, we aim to develop a fully automatic procedure to define the damage VOI. We measure the damage by comparing the spatially unbiased atlas template (SUIT) of the cerebellum~\cite{diedrichsen2006spatially} and the normalized patient's cerebellum, which distinguishes the missing brain tissue from the fourth ventricle space.
Efforts have been made in cerebellum normalization to understand structure-function relationships~\cite{grosse2021mapping,schmahmann2000mri,diedrichsen2006spatially,diedrichsen2010advances,hernandez2019representative,diedrichsen2009probabilistic}. Grosse et al.~\cite{grosse2021mapping} related cerebellar lesion to the cognitive and motor deficits in pediatric brain tumor survivors. However, their approach to identify cerebellar lesion was achieved manually. A lesion mask was manually created in both native and atlas space. Our approach is fully automatic, and is closely related to SUIT normalization, first proposed in ~\cite{diedrichsen2006spatially} by Diedrichsen. In their approach, an isolation algorithm is applied to segment the cerebellum from the whole brain, using the ICBM152 template and prior information on the brain tissue. The the cerebellum is then normalized to the SUIT atlas via the algorithm proposed in~\cite{ashburner1999nonlinear} and implemented in the SPM2 package~\cite{spm}, an early version of SPM. Diedrichen et al. later developed a
VOI-based SUIT normalization\cite{diedrichsen2011imaging}, which uses the VOI of the dentate nuclei to guide normalization. The VOI-guided approach yields more accurate normalization, but requires the dentate nuclei VOI as an additional input. The SUIT normalization software is available at~\cite{suit}, which uses SPM12 segmentation for brain tissue segmentation and DARTEL~\cite{ashburner2007fast} as the normalization algorithm in its latest implementation. Despite its success in many applications~\cite{o2008cerebellum,donchin2012cerebellar,kuper2011evidence}, the SUIT normalization technique, however, cannot be directly applied here to detect cerebellar damage, because it does not take the surgical cavity into account and damage to the dentate nuclei may be present. The cerebellum isolation algorithm is likely to fail in the first place by miss-classifying surgical cavity as gray matter, regardless of the normalization algorithm used later. In addition, given an accurate cerebellum mask, the DARTEL algorithm tends to over stretch the brain tissues to match the template, resulting in little or no damage in the atlas space. Therefore, both the brain tissue segmentation and the cerebellum normalization need to be carefully designed for postoperative images.
In this work, we propose a novel automatic approach to detect and measure cerebellar damage. This work has three major contributions: (1) We develop a brain tissue segmentation algorithm that considers the presence of surgical cavity, enlarged ventricles and other non-tissue components such as the presence of absorbable hemostat and blood products, which are commonly seen in postoperative images.
(2) We develop a customized cerebellum normalization procedure that uses anatomical information in combination with the state-of-the-art nonlinear registration algorithm, SyN~\cite{avants2008symmetric,klein2009evaluation}, implemented in the Advanced Normalization Tools (ANTs)~\cite{ants}. The normalization technique proposed here is more accurate than SUIT normalization on postoperative images, thus can also be useful for applications other than damage detection, such as functional magnetic resonance imaging (fMRI) analyses where mapping from the blood-oxygen-level-dependent (BOLD) signal to the cerebellum functional regions is required.
(3) We provide an automatic and accurate VOI estimate of cerebellar damage in the atlas space, which is vital to study structural damage and function relationship and also to perform group comparisons.
\section{Methods}
\label{sec:method}
\subsection{Overview}
Detecting surgical damage is challenging because the void region in the postoperative image does not necessarily indicate damage. This is evidenced by the following: 1) If the tumor was near the fourth ventricle before resection, the void region of the postoperative image would show both the cerebellar damage and the fourth ventricle space as one unified fluid-filled space. 2) Even for lateral tumors, the volume of the void region does not reflect the volume of damage, because the degree of deformation of normal brain tissues varies by patient. Hence, we detect surgical damage in the atlas space instead of native space. The accuracy of the damage detection then heavily relies on the accuracy of segmentation and normalization algorithms. Therefore, we specifically design a brain tissue segmentation algorithm and a cerebellum normalization approach to address this issue, which constitute two major components of the overall damage detection tool.
An overview of the proposed framework is shown in Fig.~\ref{fig:workflow}. The 3D T1 image is first preprocessed by cropping, bias correction and brain extraction. The white matter (WM) and gray matter (GM) are then segmented via a brain tissue segmentation algorithm we developed taking the surgical cavity into account (see detailed description in Section~\ref{sec:brain_tissue_seg}). The cerebellum is isolated by nonlinear registration of the ICBM template to the native space, keeping only WM and GM in the moving and fixed image. The isolated cerebellum of the patient is then normalized into the SUIT template space~\cite{diedrichsen2006spatially} using labeled (background: 0, WM: 1, GM: 2, stem: 3) fixed and moving images. Finally, the surgical damage is detected by comparing the difference between the normalized image and the SUIT template.
\begin{figure}
\includegraphics[width=\textwidth]{figures/overview.png}
\caption{An overview of the proposed framework for detecting cerebellar surgical damage.} \label{fig:workflow}
\end{figure}
\subsection{Templates and processing}
\label{sec:templates}
\subsubsection{ICBM} The ICBM template is available at UCLA brain mapping center~\cite{icbmtempalate} and is used here as our the whole brain template. Our implementation uses only WM and GM for more accurate registration to the native space. WM and GM can be obtained by thresholding the template after brain extraction using FSL BET~\cite{smith2002fast}. The cerebellum mask of ICBM is the combined region of the zones with the following labels: 58, 67, 237, 238, and 251. The probability maps of WM and GM of this atlas~\cite{icbmprob} are used in our brain tissue segmentation algorithm as prior probabilities.
\subsubsection{SUIT}
The SUIT atlas~\cite{diedrichsen2006spatially} is used as the cerebellum template for cerebellum normalization. The same brain tissue segmentation algorithm is used to generate probability maps of WM and GM and label this template (WM: 1, GM: 2). The stem VOI was created manually and labeled as 3.
The original and processed templates used in our implementation are shown in Fig~\ref{fig:templates}.
\begin{figure}
\centering
\includegraphics[width=3.5in]{figures/templates.png}
\caption{Original and processed templates used in this study. The ICBM atlas (A-E). A: original (sagittal); B: brain extracted (sagittal); C: cerebellum mask overlaid on the thresholded template (sagittal); D: white matter probability map (axial); E: gray matter probability map (axial). The SUIT atlas (F-G). F: original (sagittal); G: labeled (sagittal)} \label{fig:templates}
\end{figure}
\subsection{Input preprocessing}
\label{sec:preprocessing}
Neck and lower head are first removed using FSL robustfov\cite{robustfov} from the input postsugical 3D T1 image. The cropped image is then bias corrected using N4 bias correction~\cite{tustison2010n4itk}, and brain extracted using FSL BET~\cite{smith2002fast}. Because of the presence of the surgical cavity and potentially contaminated brain tissues, the BET mask is generally not satisfactory. Thus, a better brain mask is obtained by affine registration of the brain extracted ICBM template to the brain extracted input image, and then binarizing the warped template.
\subsection{Brain tissue segmentation}
\label{sec:brain_tissue_seg}
Patients being treated for posterior fossa brain tumors present with abnormalities that complicate standard neuroimaging anatomical analysis such as segmentation of brain tissues. In addition to the presence of a tumor, or relatively large postsurgical cavity (which may progressively contract later), abnormalities such as enlarged ventricles (i.e., hydrocephalus) are common. Therefore, commonly used brain tissue segmentation tools (e.g., SPM unified segmentation~\cite{ashburner2005unified}) designed for the healthy population are susceptible to failure on postoperative images. Here we present a brain tissue segmentation approach modified from the brain tumor segmentation algorithm based on outlier detection, proposed in ~\cite{prastawa2004brain}. As this work is aimed to detect damage on postoperative images, we assume that there is no tumor present in the image. Therefore, only three classes are modeled in our algorithm: WM, GM and other, i.e., class label $\mathbb{Y}=\{\mathrm{WM},\mathrm{GM},\mathrm{other}\}$. The ``other'' class contains primarily CSF, but may also contain blood products, hemostat, non-brain regions (due to imperfect brain extraction), and other components that are not considered as normal brain tissue.
The WM and GM probabilities are initialized as ICBM atlas priors (affine transformed into the native space). Because of the presence of surgical cavity, enlarged ventricles and other non-tissue components, probability of the ``other'' class is initialized as uniform distribution ($
\Pr(\mathrm{other})=0.5$), rather than using the CSF prior of ICBM. Then, the probability density function of each class, denoted as $p(x|Y)$, is estimated via kernel density estimation using voxels sampled according to prior probability. Because the WM and GM samples can have contaminants (due to inaccurate priors), an outlier detection approach is used to detect and remove the contaminants (key idea proposed in ~\cite{prastawa2004brain}). Outlier detection is performed by using the Minimum Covariance Determinant estimator, setting the support fraction to 0.5. Unlike ~\cite{prastawa2004brain} where T1 and T2 are taken as inputs, our implementation uses only 3D T1 as the input, as it provides good spatial resolution and good contrast between WM, GM, and others. The outlier detection in our implementation is hence performed in one dimension (T1 only) instead of two (T1 and T2 as in ~\cite{prastawa2004brain}), and not performed for the ``other'' class, since we allow this class to have contaminants. The posterior probabilities are then calculated according to Bayes’ theorem:
$$
Pr(Y|x)=\frac{p(x|Y)Pr(Y)}{p(x)},
$$
where $p(x)=\sum_{Y\in \mathbb{Y}}p(x|Y)Pr(Y).$
The posteriors are used as priors in the next iteration, and the probabilities are updated through a total of 3 iterations in our implementation. The final WM and GM masks are generated by thresholding the smoothed posteriors at 0.5. The union mask of WM and GM is denoted as $\mathrm{WM} \cup \mathrm{GM}$.
\subsection{Cerebellum isolation, normalization and damage detection}
\subsubsection{Cerebellum isolation}
The cerebellum is isolated by non-linear transformation of the ICBM atlas to the native space. To achieve a better alignment, only WM and GM regions of the moving and the fixed images are used for registration. Registration is performed using ANTs~\cite{ants}, with affine transformation as the initial step followed by nonlinear transformation using the SyN algorithm~\cite{avants2008symmetric}. The cerebellum mask of the ICBM template is then warped into the native space. The warped ICBM cerebellum mask is denoted as $\mathrm{Cb_w}$. The patient's cerebellum mask is the intersection of $\mathrm{WM} \cup \mathrm{GM}$ and $\mathrm{Cb_w}$ adjusted by morphological operations and smoothing.
\subsubsection{Cerebellum normalization}
The initial step of cerebellum normalization is affine registration of the patient's cerebellum (along with the WM and GM masks) to the SUIT template. Next, to achieve detailed matching in the nonlinear transformation, both the moving (affine transformed input image) and fixed (SUIT template) images are labeled (background: 0, WM: 1, GM: 2, stem: 3). The stem mask of the patient was generated after affine registration using the template's label information. The ANTs SyN is used to perform the final step of normalization.
\subsubsection{Cerebellar damage detection}
We define the cerebellar damage as the missing brain tissue (WM or GM) of the patient's cerebellum. Therefore, assuming a perfect normalization, the damage VOI is segmented in the SUIT atlas space by subtracting the normalized cerebellum from the template after binarization of both, followed by taking the largest connected component. Detecting damage in the atlas space has two major advantages: 1) the fourth ventricle space is not counted as damage because it has the background intensity in the SUIT template; 2) the damage VOIs can be compared (e.g., size, location) across patients.
\section{Experimental Results}
\subsection{Datasets}
We used 3D T1 scans of a prospective study (SJMB12; NCT 01878617) at St. Jude Children's Research Hospital to test our cerebellar damage detection algorithm. Specifically, we included 245 patients diagnosed with medulloblastoma (MB) and had the complete tumor resected (no remaining tumor or metastasis). Our damage detection approach was applied on one postoperative 3D T1 scan for each patient. 3D T1 images were acquired 1-368 days after the last surgery. Evaluation of performance was established on a subset of imaging (N=153) with ground truth created by human experts. To better understand the performance of the proposed approach in wider scenarios, we also created a simulated dataset by generating damages on healthy brain images. The manual evaluation procedure and the simulation design are described below.
\subsubsection{The CMS Dataset}
Manual evaluation was performed on results of 153 patients who either had postoperative cerebellar mutism syndrome (CMS) or was asymptomatic after surgery, for the purpose of studying CMS related surgical damage. Evaluation of the results is challenging, as it is extremely difficult for human experts to draw the damage VOI on the SUIT template by just looking at the 3D T1 in native space, blinded to the auto generated results (the normalized T1 and auto mask). To reduce bias, we asked two raters, an expert in cerebellar anatomy and a neuroradiologist, to evaluate the damage VOIs generated by the algorithm. During the evaluation, the raters first inspected the 3D T1 scan, and then examined the normalized image and the damage mask in the atlas space. Each damage mask was either considered accurate as is or edited manually and approved by both raters. Then, the performance of the algorithm was measured by the S$\phi$rensen-Dice Similarity Coefficient (Dice) between the auto generated mask (AM) and the approved ground truth (GT) mask. The Dice score is defined as
$$
Dice = \frac{2|GT\cap AM|}{|GT|+|AM|}.
$$
\subsubsection{Simulated Dataset}
The evaluation on the CMS dataset had some limitations. First, this dataset was relatively small, and only 13 out of 153 cases had a lateral damage. Thus, it's hard to conclude how well the algorithm does in detecting lateral damages. Second, the raters were potentially biased by the auto prediction due the natural of the evaluation process. Hence, we designed a simulation procedure to generate sufficient data that covers both ventricular and lateral damages of different sizes.
The main idea of the simulation was to generate a postoperative 3D T1 image in the native space, the normalized cerebellum in the SUIT atlas space, a damage VOI showing the missing tissue, and the transformation that maps the two spaces. To achieve this, we also used 3D T1 images of 69 healthy children (not necessarily age matched to the patients) recruited for a separate study. The healthy controls do not have a history of cancer or are a first degree relative or direct friend of the MB patients. The simulation contains the following steps: 1) Brain tissue segmentation and cerebellum isolation (described in Section~\ref{sec:method}) was performed on a 3D T1 of a healthy brain. 2) The SUIT template was registered to the isolated healthy cerebellum using ATNs SyN. The transformation of this registration was reliable because the anatomies of both the moving and fixed images were complete. The registration was also inspected to ensure quality. 3) A simulated damage VOI was created on the SUIT atlas by randomly sampling one VOI from all the 245 VOIs generated by the algorithm on the SJMB12 patients (including those that were not verified by human expects), followed by applying random deformation. 4) The simulated damage VOI was then overlaid on the native image by applying the transformation generated from step 2. 5) The simulated postoperative 3D T1 image was created by replacing the damage VOI in the native space with CSF intensities using Gaussian distribution $\mathcal{N}(\mu,10)$, followed by random deformation of the whole image. The mean of CSF $\mu$ can be set as the median intensity in the ``other'' mask from brain tissue segmentation, since the ``other'' class has no containment in healthy brains.
We generated 10 random damage VOIs for each of the 69 healthy children, which resulted in a total of 690 pairs of simulated 3D T1 and GT damage VOI. Our damage detection tool was applied on the simulated 3D T1 and the auto generated damage VOI was compared to GT using Dice score.
\subsection{Results}
\subsubsection{The CMS Dataset}
The average run time of the whole damage detection pipeline was about 40 minutes and the average memory usage was about 2 GB.
82 out of 153 damage masks were considered accurate as is, and for the remaining cases that required manual adjustment, the average time needed for editing was about 10 minutes. The mean Dice score of 153 patients was $0.892 (\pm 0.217)$. The GT mask size range from 0 (no damage) to 37210 $mm^3$, with a mean of 3455 $mm^3$. In this dataset, there were 4 patients that had no surigical damage observed by the raters. Because the algorithm always detect the missing voxels in the normalized image as damage, the Dice score for these 4 cases was 0. The masks that were approved as is had a Dice score of 1. We observed that the performance of the algorithm depends on the size of the damage. The smaller the damage, the harder to detect. This is because when the surgical cavity is small, the result can be skewed by MR artifact, abnormal anatomy not caused by resection, or misalignment during normalization, etc. The mean Dice scores of different damage size ranges are summarized in Table 1.
\begin{table}
\caption{Performance (Dice) of the cerebellar damage detection tool on patients with different damage sizes.} \label{table:dice_size}
\includegraphics[width=\textwidth]{figures/dice_size.png}
\end{table}
To better demonstrate our results, we selected one patient from each size range that had a Dice score close to the mean value (per Table~\ref{table:dice_size}). The results of the five example cases are shown in Fig.~\ref{fig:results}.
The first row shows the 3D T1 scans that the algorithm took as inputs. The second row shows the WM (red) and GM (green) segmentations from our brain tissue segmentation model. The third row shows the cerebellar mask (red) overlaid on the 3D T1 image. The fourth row shows the auto generated cerebellar damage mask (red) overlaid on the normalized cerebellum in the atlas space. The fifth row shows the GT mask (green) overlaid on the SUIT template. The results show the superior performance of our proposed approach under various scenarios (different damage sizes and locations).
\begin{figure}
\includegraphics[width=\textwidth]{figures/figure_CMS.png}
\caption{Representative examples of patients with different cerebellar damages in size and location. First row: cropped 3D T1. Second row: white matter (WM, red) and gray matter (GM, green) segmentations. Third row: cerebellum mask (red) overlaid on 3D T1. Fourth row: auto generated damage mask (AM, red) overlaid on the normalized cerebellum. Fifth row: approved ground truth (GT, green) mask overlaid on the SUIT template. Patient 1-4 had a damage near the fourth ventricle and results are shown in the sagittal plane. Patient 5 had a lateral damage and results are shown in the coronal plane.} \label{fig:results}
\end{figure}
Figure~\ref{fig:FN_FP} shows the heatmaps of false negative (FN) (Figure~\ref{fig:FN_FP}A) and false positive (FP) (Figure~\ref{fig:FN_FP}B) voxels detected by our approach. The count of FN is relatively small, with the maximum count being 7, and FN voxels appear to be randomly distributed. The FP voxels are centered near the superior cerebellar peduncles (SCPs). This is a result from misalignment in the SCP region during the normalization process. The SCPs have very thin structure and therefore are under emphasized by the registration algorithm.
\begin{figure}
\includegraphics[width=\textwidth]{figures/FN_FP.png}
\caption{The heatmap of false negative (A) and false positive (B) voxels detected by the proposed damage detection approach.} \label{fig:FN_FP}
\end{figure}
\subsubsection{Simulated Dataset}
The performance of the damage detection algorithm on the simulated data is summarized in Table~\ref{table:simulation}. The mean Dice score on ventricular damages was 0.723, higher than the score on lateral cases, which was 0.365. Similar to what was observed from the CMS dataset, the algorithm was more accurate on damages with larger size. Qualitative results are shown in Fig.~\ref{fig:simulation}. For each size range, we selected a representative example that had the closest Dice to the mean value (per Table~\ref{table:simulation}) for both ventricular (Fig.~\ref{fig:simulation}A) and lateral damages (Fig.~\ref{fig:simulation}B). The algorithm was able to accurately locate the damage, even when damage was small (first case in Fig.~\ref{fig:simulation}A and first two cases in Fig.~\ref{fig:simulation}B). Because the normalization was not perfect, a small degree of misalignment of the healthy brain tissues near the damage would dramatically lower the dice score if the missing VOI was small. These results suggest that inspection and manual editing by human expect is required to ensure the quality of damage detection. Overall, the simulation provided an unbiased evaluation of the proposed damage detection approach, which achieved reasonable normalization of the cerebellum and robust damage detection.
\begin{table}
\caption{Performance (Dice) of the cerebellar damage detection tool on simulated datasets categorized by size and location.} \label{table:simulation}
\includegraphics[width=\textwidth]{figures/table_simulation.png}
\end{table}
\begin{figure}
\includegraphics[width=\textwidth]{figures/figure_simulation.png}
\caption{Visualization of representative examples of the simulated data and damage detection results. First row: the healthy 3D T1 brain image used for simulation. Second row: the simulated postoperative 3D T1 image after applying the simulated damage VOI. Third row: the automatically detected damage mask (AM, red) on the normalized cerebellum. Fourth row: the ground truth (GT, green) simulated damage mask on the SUIT template. A) Ventricular damages. Results are shown in the saggittal plane. B) Lateral damages. Results are shown in the coronal plane.} \label{fig:simulation}
\end{figure}
\section{Discussions}
Existing cerebellum normalization tools are typically designed for healthy brain imaging data, thus tend to fail when lesions are present. As a by-product of this cerebellar damage detection approach, the normalization of cerebellum is also suitable for broader neuroimaging applications (e.g., fMRI preprocessing and analysis), and is more accurate than SUIT normalization~\cite{diedrichsen2011imaging} on postoperative images. Our normalization is more accurate because (1) it has a more accurate cerebellum mask and WM/GM segmentation to start with, (2) labeled inputs allow for easier and better optimization, and (3) the choice of state-of-the art nonlinear registration algorithm, Syn~\cite{klein2009evaluation}.
Our work has some limitations: (1) An important assumption of this approach is the absence of the tumor in the image. Hence, this method cannot handle cases with partial tumor resection or metastases. (2) Our brain tissue segmentation algorithm assumes everything with abnormal intensity (hyper or hypo) is not healthy brain tissue, which may not always be the case. For example, brain tissue might be stained by blood and therefore have an abnormal appearance in the image, but be otherwise healthy. However, this is not easy to tell from image at the time point immediately following surgery. MR scans at later time points may further inform healthy or damaged anatomy after the stain has been removed via natural processes. (3) Our damage detection algorithm is not learning-based. As a result, its performance can not be improved by collecting more accurate labels (manually edited masks). Our future work is to develop a deep learning based model to further improve the performance by taking advantage of the inputs from neuroradiologists.
\section{Conclusions}
In this paper, we present a fully automatic localization and measurement of postoperative damage in the cerebellum in the atlas space using postoperative MRI. The proposed framework has a novel brain tissue segmentation algorithms that considers surgery-related features (e.g., postsurgical cavity, absorbable hemostat, blood products) and other non-surgical features (e.g., enlarged ventricles), a more accurate cerebellum normalization procedure, and provides quantification of cerebellar damage in the atlas space. This automation greatly reduced human annotation time from hours to about 10 minutes. The normalization proposed in this work is also helpful to perform other neuroimaging analyses (e.g., fMRI analyses) where cerebellum normalization is needed, especially on postoperative images.
\bibliographystyle{splncs04}
\section{Introduction}
Surgical resection is a common procedure in the treatment of pediatric posterior fossa tumors. However, surgical resection is associated with potential serious complications~\cite{dubey2009complications}, such as postoperative cerebellar mutism~\cite{van1995transient,gelabert2001mutism,pollack1997posterior,10.1093/neuonc/noaa215.536} and resultant long-term neuropsychological dysfunction~\cite{grill2004critical,levisohn2000neuropsychological}. While brain tissue damage caused by the tumor itself may be minor, it is typically not possible to remove the tumor without impacting surrounding healthy brain tissue. Thus, surgical damage is often unavoidable. Understanding the relationship between the damaged cerebellar regions and postoperative neurological and cognitive outcomes has the potential to guide surgery and significantly affect quality of survival of patients. Therefore, a reliable localization and measure of cerebellar damage is needed and fundamental to conduct this type of investigation.
Defining the volume-of-interest (VOI) for cerebellar damage in the postoperative brain MR image is challenging, even for human experts performing this task manually. First, the damage we want to identify herein is the ``loss of normal brain tissue''. Instead of showing an abnormal intensity in the MR image, this missing tissue presents as just an empty space (cavity) filled with cerebrospinal fluid (CSF). In addition, an abnormal CSF distribution does not necessarily reflect damage, because normal brain tissues may have already been displaced by the presence of the tumor and disease complications (e.g., hydrocephalus) before the surgery. Presence of the tumor near the fourth ventricle complicates this issue, as the ventricle space is connected to the postsurgical cavity. Given these challenges, we aim to develop a fully automatic procedure to define the damage VOI. We measure the damage by comparing the spatially unbiased atlas template (SUIT) of the cerebellum~\cite{diedrichsen2006spatially} and the normalized patient's cerebellum, which distinguishes the missing brain tissue from the fourth ventricle space.
Efforts have been made in cerebellum normalization to understand structure-function relationships~\cite{grosse2021mapping,schmahmann2000mri,diedrichsen2006spatially,diedrichsen2010advances,hernandez2019representative,diedrichsen2009probabilistic}. Grosse et al.~\cite{grosse2021mapping} related cerebellar lesion to the cognitive and motor deficits in pediatric brain tumor survivors. However, their approach to identify cerebellar lesion was achieved manually. A lesion mask was manually created in both native and atlas space. Our approach is fully automatic, and is closely related to SUIT normalization, first proposed in ~\cite{diedrichsen2006spatially} by Diedrichsen. In their approach, an isolation algorithm is applied to segment the cerebellum from the whole brain, using the ICBM152 template and prior information on the brain tissue. The the cerebellum is then normalized to the SUIT atlas via the algorithm proposed in~\cite{ashburner1999nonlinear} and implemented in the SPM2 package~\cite{spm}, an early version of SPM. Diedrichen et al. later developed a
VOI-based SUIT normalization\cite{diedrichsen2011imaging}, which uses the VOI of the dentate nuclei to guide normalization. The VOI-guided approach yields more accurate normalization, but requires the dentate nuclei VOI as an additional input. The SUIT normalization software is available at~\cite{suit}, which uses SPM12 segmentation for brain tissue segmentation and DARTEL~\cite{ashburner2007fast} as the normalization algorithm in its latest implementation. Despite its success in many applications~\cite{o2008cerebellum,donchin2012cerebellar,kuper2011evidence}, the SUIT normalization technique, however, cannot be directly applied here to detect cerebellar damage, because it does not take the surgical cavity into account and damage to the dentate nuclei may be present. The cerebellum isolation algorithm is likely to fail in the first place by miss-classifying surgical cavity as gray matter, regardless of the normalization algorithm used later. In addition, given an accurate cerebellum mask, the DARTEL algorithm tends to over stretch the brain tissues to match the template, resulting in little or no damage in the atlas space. Therefore, both the brain tissue segmentation and the cerebellum normalization need to be carefully designed for postoperative images.
In this work, we propose a novel automatic approach to detect and measure cerebellar damage. This work has three major contributions: (1) We develop a brain tissue segmentation algorithm that considers the presence of surgical cavity, enlarged ventricles and other non-tissue components such as the presence of absorbable hemostat and blood products, which are commonly seen in postoperative images.
(2) We develop a customized cerebellum normalization procedure that uses anatomical information in combination with the state-of-the-art nonlinear registration algorithm, SyN~\cite{avants2008symmetric,klein2009evaluation}, implemented in the Advanced Normalization Tools (ANTs)~\cite{ants}. The normalization technique proposed here is more accurate than SUIT normalization on postoperative images, thus can also be useful for applications other than damage detection, such as functional magnetic resonance imaging (fMRI) analyses where mapping from the blood-oxygen-level-dependent (BOLD) signal to the cerebellum functional regions is required.
(3) We provide an automatic and accurate VOI estimate of cerebellar damage in the atlas space, which is vital to study structural damage and function relationship and also to perform group comparisons.
\section{Methods}
\label{sec:method}
\subsection{Overview}
Detecting surgical damage is challenging because the void region in the postoperative image does not necessarily indicate damage. This is evidenced by the following: 1) If the tumor was near the fourth ventricle before resection, the void region of the postoperative image would show both the cerebellar damage and the fourth ventricle space as one unified fluid-filled space. 2) Even for lateral tumors, the volume of the void region does not reflect the volume of damage, because the degree of deformation of normal brain tissues varies by patient. Hence, we detect surgical damage in the atlas space instead of native space. The accuracy of the damage detection then heavily relies on the accuracy of segmentation and normalization algorithms. Therefore, we specifically design a brain tissue segmentation algorithm and a cerebellum normalization approach to address this issue, which constitute two major components of the overall damage detection tool.
An overview of the proposed framework is shown in Fig.~\ref{fig:workflow}. The 3D T1 image is first preprocessed by cropping, bias correction and brain extraction. The white matter (WM) and gray matter (GM) are then segmented via a brain tissue segmentation algorithm we developed taking the surgical cavity into account (see detailed description in Section~\ref{sec:brain_tissue_seg}). The cerebellum is isolated by nonlinear registration of the ICBM template to the native space, keeping only WM and GM in the moving and fixed image. The isolated cerebellum of the patient is then normalized into the SUIT template space~\cite{diedrichsen2006spatially} using labeled (background: 0, WM: 1, GM: 2, stem: 3) fixed and moving images. Finally, the surgical damage is detected by comparing the difference between the normalized image and the SUIT template.
\begin{figure}
\includegraphics[width=\textwidth]{figures/overview.png}
\caption{An overview of the proposed framework for detecting cerebellar surgical damage.} \label{fig:workflow}
\end{figure}
\subsection{Templates and processing}
\label{sec:templates}
\subsubsection{ICBM} The ICBM template is available at UCLA brain mapping center~\cite{icbmtempalate} and is used here as our the whole brain template. Our implementation uses only WM and GM for more accurate registration to the native space. WM and GM can be obtained by thresholding the template after brain extraction using FSL BET~\cite{smith2002fast}. The cerebellum mask of ICBM is the combined region of the zones with the following labels: 58, 67, 237, 238, and 251. The probability maps of WM and GM of this atlas~\cite{icbmprob} are used in our brain tissue segmentation algorithm as prior probabilities.
\subsubsection{SUIT}
The SUIT atlas~\cite{diedrichsen2006spatially} is used as the cerebellum template for cerebellum normalization. The same brain tissue segmentation algorithm is used to generate probability maps of WM and GM and label this template (WM: 1, GM: 2). The stem VOI was created manually and labeled as 3.
The original and processed templates used in our implementation are shown in Fig~\ref{fig:templates}.
\begin{figure}
\centering
\includegraphics[width=3.5in]{figures/templates.png}
\caption{Original and processed templates used in this study. The ICBM atlas (A-E). A: original (sagittal); B: brain extracted (sagittal); C: cerebellum mask overlaid on the thresholded template (sagittal); D: white matter probability map (axial); E: gray matter probability map (axial). The SUIT atlas (F-G). F: original (sagittal); G: labeled (sagittal)} \label{fig:templates}
\end{figure}
\subsection{Input preprocessing}
\label{sec:preprocessing}
Neck and lower head are first removed using FSL robustfov\cite{robustfov} from the input postsugical 3D T1 image. The cropped image is then bias corrected using N4 bias correction~\cite{tustison2010n4itk}, and brain extracted using FSL BET~\cite{smith2002fast}. Because of the presence of the surgical cavity and potentially contaminated brain tissues, the BET mask is generally not satisfactory. Thus, a better brain mask is obtained by affine registration of the brain extracted ICBM template to the brain extracted input image, and then binarizing the warped template.
\subsection{Brain tissue segmentation}
\label{sec:brain_tissue_seg}
Patients being treated for posterior fossa brain tumors present with abnormalities that complicate standard neuroimaging anatomical analysis such as segmentation of brain tissues. In addition to the presence of a tumor, or relatively large postsurgical cavity (which may progressively contract later), abnormalities such as enlarged ventricles (i.e., hydrocephalus) are common. Therefore, commonly used brain tissue segmentation tools (e.g., SPM unified segmentation~\cite{ashburner2005unified}) designed for the healthy population are susceptible to failure on postoperative images. Here we present a brain tissue segmentation approach modified from the brain tumor segmentation algorithm based on outlier detection, proposed in ~\cite{prastawa2004brain}. As this work is aimed to detect damage on postoperative images, we assume that there is no tumor present in the image. Therefore, only three classes are modeled in our algorithm: WM, GM and other, i.e., class label $\mathbb{Y}=\{\mathrm{WM},\mathrm{GM},\mathrm{other}\}$. The ``other'' class contains primarily CSF, but may also contain blood products, hemostat, non-brain regions (due to imperfect brain extraction), and other components that are not considered as normal brain tissue.
The WM and GM probabilities are initialized as ICBM atlas priors (affine transformed into the native space). Because of the presence of surgical cavity, enlarged ventricles and other non-tissue components, probability of the ``other'' class is initialized as uniform distribution ($
\Pr(\mathrm{other})=0.5$), rather than using the CSF prior of ICBM. Then, the probability density function of each class, denoted as $p(x|Y)$, is estimated via kernel density estimation using voxels sampled according to prior probability. Because the WM and GM samples can have contaminants (due to inaccurate priors), an outlier detection approach is used to detect and remove the contaminants (key idea proposed in ~\cite{prastawa2004brain}). Outlier detection is performed by using the Minimum Covariance Determinant estimator, setting the support fraction to 0.5. Unlike ~\cite{prastawa2004brain} where T1 and T2 are taken as inputs, our implementation uses only 3D T1 as the input, as it provides good spatial resolution and good contrast between WM, GM, and others. The outlier detection in our implementation is hence performed in one dimension (T1 only) instead of two (T1 and T2 as in ~\cite{prastawa2004brain}), and not performed for the ``other'' class, since we allow this class to have contaminants. The posterior probabilities are then calculated according to Bayes’ theorem:
$$
Pr(Y|x)=\frac{p(x|Y)Pr(Y)}{p(x)},
$$
where $p(x)=\sum_{Y\in \mathbb{Y}}p(x|Y)Pr(Y).$
The posteriors are used as priors in the next iteration, and the probabilities are updated through a total of 3 iterations in our implementation. The final WM and GM masks are generated by thresholding the smoothed posteriors at 0.5. The union mask of WM and GM is denoted as $\mathrm{WM} \cup \mathrm{GM}$.
\subsection{Cerebellum isolation, normalization and damage detection}
\subsubsection{Cerebellum isolation}
The cerebellum is isolated by non-linear transformation of the ICBM atlas to the native space. To achieve a better alignment, only WM and GM regions of the moving and the fixed images are used for registration. Registration is performed using ANTs~\cite{ants}, with affine transformation as the initial step followed by nonlinear transformation using the SyN algorithm~\cite{avants2008symmetric}. The cerebellum mask of the ICBM template is then warped into the native space. The warped ICBM cerebellum mask is denoted as $\mathrm{Cb_w}$. The patient's cerebellum mask is the intersection of $\mathrm{WM} \cup \mathrm{GM}$ and $\mathrm{Cb_w}$ adjusted by morphological operations and smoothing.
\subsubsection{Cerebellum normalization}
The initial step of cerebellum normalization is affine registration of the patient's cerebellum (along with the WM and GM masks) to the SUIT template. Next, to achieve detailed matching in the nonlinear transformation, both the moving (affine transformed input image) and fixed (SUIT template) images are labeled (background: 0, WM: 1, GM: 2, stem: 3). The stem mask of the patient was generated after affine registration using the template's label information. The ANTs SyN is used to perform the final step of normalization.
\subsubsection{Cerebellar damage detection}
We define the cerebellar damage as the missing brain tissue (WM or GM) of the patient's cerebellum. Therefore, assuming a perfect normalization, the damage VOI is segmented in the SUIT atlas space by subtracting the normalized cerebellum from the template after binarization of both, followed by taking the largest connected component. Detecting damage in the atlas space has two major advantages: 1) the fourth ventricle space is not counted as damage because it has the background intensity in the SUIT template; 2) the damage VOIs can be compared (e.g., size, location) across patients.
\section{Experimental Results}
\subsection{Datasets}
We used 3D T1 scans of a prospective study (SJMB12; NCT 01878617) at St. Jude Children's Research Hospital to test our cerebellar damage detection algorithm. Specifically, we included 245 patients diagnosed with medulloblastoma (MB) and had the complete tumor resected (no remaining tumor or metastasis). Our damage detection approach was applied on one postoperative 3D T1 scan for each patient. 3D T1 images were acquired 1-368 days after the last surgery. Evaluation of performance was established on a subset of imaging (N=153) with ground truth created by human experts. To better understand the performance of the proposed approach in wider scenarios, we also created a simulated dataset by generating damages on healthy brain images. The manual evaluation procedure and the simulation design are described below.
\subsubsection{The CMS Dataset}
Manual evaluation was performed on results of 153 patients who either had postoperative cerebellar mutism syndrome (CMS) or was asymptomatic after surgery, for the purpose of studying CMS related surgical damage. Evaluation of the results is challenging, as it is extremely difficult for human experts to draw the damage VOI on the SUIT template by just looking at the 3D T1 in native space, blinded to the auto generated results (the normalized T1 and auto mask). To reduce bias, we asked two raters, an expert in cerebellar anatomy and a neuroradiologist, to evaluate the damage VOIs generated by the algorithm. During the evaluation, the raters first inspected the 3D T1 scan, and then examined the normalized image and the damage mask in the atlas space. Each damage mask was either considered accurate as is or edited manually and approved by both raters. Then, the performance of the algorithm was measured by the S$\phi$rensen-Dice Similarity Coefficient (Dice) between the auto generated mask (AM) and the approved ground truth (GT) mask. The Dice score is defined as
$$
Dice = \frac{2|GT\cap AM|}{|GT|+|AM|}.
$$
\subsubsection{Simulated Dataset}
The evaluation on the CMS dataset had some limitations. First, this dataset was relatively small, and only 13 out of 153 cases had a lateral damage. Thus, it's hard to conclude how well the algorithm does in detecting lateral damages. Second, the raters were potentially biased by the auto prediction due the natural of the evaluation process. Hence, we designed a simulation procedure to generate sufficient data that covers both ventricular and lateral damages of different sizes.
The main idea of the simulation was to generate a postoperative 3D T1 image in the native space, the normalized cerebellum in the SUIT atlas space, a damage VOI showing the missing tissue, and the transformation that maps the two spaces. To achieve this, we also used 3D T1 images of 69 healthy children (not necessarily age matched to the patients) recruited for a separate study. The healthy controls do not have a history of cancer or are a first degree relative or direct friend of the MB patients. The simulation contains the following steps: 1) Brain tissue segmentation and cerebellum isolation (described in Section~\ref{sec:method}) was performed on a 3D T1 of a healthy brain. 2) The SUIT template was registered to the isolated healthy cerebellum using ATNs SyN. The transformation of this registration was reliable because the anatomies of both the moving and fixed images were complete. The registration was also inspected to ensure quality. 3) A simulated damage VOI was created on the SUIT atlas by randomly sampling one VOI from all the 245 VOIs generated by the algorithm on the SJMB12 patients (including those that were not verified by human expects), followed by applying random deformation. 4) The simulated damage VOI was then overlaid on the native image by applying the transformation generated from step 2. 5) The simulated postoperative 3D T1 image was created by replacing the damage VOI in the native space with CSF intensities using Gaussian distribution $\mathcal{N}(\mu,10)$, followed by random deformation of the whole image. The mean of CSF $\mu$ can be set as the median intensity in the ``other'' mask from brain tissue segmentation, since the ``other'' class has no containment in healthy brains.
We generated 10 random damage VOIs for each of the 69 healthy children, which resulted in a total of 690 pairs of simulated 3D T1 and GT damage VOI. Our damage detection tool was applied on the simulated 3D T1 and the auto generated damage VOI was compared to GT using Dice score.
\subsection{Results}
\subsubsection{The CMS Dataset}
The average run time of the whole damage detection pipeline was about 40 minutes and the average memory usage was about 2 GB.
82 out of 153 damage masks were considered accurate as is, and for the remaining cases that required manual adjustment, the average time needed for editing was about 10 minutes. The mean Dice score of 153 patients was $0.892 (\pm 0.217)$. The GT mask size range from 0 (no damage) to 37210 $mm^3$, with a mean of 3455 $mm^3$. In this dataset, there were 4 patients that had no surigical damage observed by the raters. Because the algorithm always detect the missing voxels in the normalized image as damage, the Dice score for these 4 cases was 0. The masks that were approved as is had a Dice score of 1. We observed that the performance of the algorithm depends on the size of the damage. The smaller the damage, the harder to detect. This is because when the surgical cavity is small, the result can be skewed by MR artifact, abnormal anatomy not caused by resection, or misalignment during normalization, etc. The mean Dice scores of different damage size ranges are summarized in Table 1.
\begin{table}
\caption{Performance (Dice) of the cerebellar damage detection tool on patients with different damage sizes.} \label{table:dice_size}
\includegraphics[width=\textwidth]{figures/dice_size.png}
\end{table}
To better demonstrate our results, we selected one patient from each size range that had a Dice score close to the mean value (per Table~\ref{table:dice_size}). The results of the five example cases are shown in Fig.~\ref{fig:results}.
The first row shows the 3D T1 scans that the algorithm took as inputs. The second row shows the WM (red) and GM (green) segmentations from our brain tissue segmentation model. The third row shows the cerebellar mask (red) overlaid on the 3D T1 image. The fourth row shows the auto generated cerebellar damage mask (red) overlaid on the normalized cerebellum in the atlas space. The fifth row shows the GT mask (green) overlaid on the SUIT template. The results show the superior performance of our proposed approach under various scenarios (different damage sizes and locations).
\begin{figure}
\includegraphics[width=\textwidth]{figures/figure_CMS.png}
\caption{Representative examples of patients with different cerebellar damages in size and location. First row: cropped 3D T1. Second row: white matter (WM, red) and gray matter (GM, green) segmentations. Third row: cerebellum mask (red) overlaid on 3D T1. Fourth row: auto generated damage mask (AM, red) overlaid on the normalized cerebellum. Fifth row: approved ground truth (GT, green) mask overlaid on the SUIT template. Patient 1-4 had a damage near the fourth ventricle and results are shown in the sagittal plane. Patient 5 had a lateral damage and results are shown in the coronal plane.} \label{fig:results}
\end{figure}
Figure~\ref{fig:FN_FP} shows the heatmaps of false negative (FN) (Figure~\ref{fig:FN_FP}A) and false positive (FP) (Figure~\ref{fig:FN_FP}B) voxels detected by our approach. The count of FN is relatively small, with the maximum count being 7, and FN voxels appear to be randomly distributed. The FP voxels are centered near the superior cerebellar peduncles (SCPs). This is a result from misalignment in the SCP region during the normalization process. The SCPs have very thin structure and therefore are under emphasized by the registration algorithm.
\begin{figure}
\includegraphics[width=\textwidth]{figures/FN_FP.png}
\caption{The heatmap of false negative (A) and false positive (B) voxels detected by the proposed damage detection approach.} \label{fig:FN_FP}
\end{figure}
\subsubsection{Simulated Dataset}
The performance of the damage detection algorithm on the simulated data is summarized in Table~\ref{table:simulation}. The mean Dice score on ventricular damages was 0.723, higher than the score on lateral cases, which was 0.365. Similar to what was observed from the CMS dataset, the algorithm was more accurate on damages with larger size. Qualitative results are shown in Fig.~\ref{fig:simulation}. For each size range, we selected a representative example that had the closest Dice to the mean value (per Table~\ref{table:simulation}) for both ventricular (Fig.~\ref{fig:simulation}A) and lateral damages (Fig.~\ref{fig:simulation}B). The algorithm was able to accurately locate the damage, even when damage was small (first case in Fig.~\ref{fig:simulation}A and first two cases in Fig.~\ref{fig:simulation}B). Because the normalization was not perfect, a small degree of misalignment of the healthy brain tissues near the damage would dramatically lower the dice score if the missing VOI was small. These results suggest that inspection and manual editing by human expect is required to ensure the quality of damage detection. Overall, the simulation provided an unbiased evaluation of the proposed damage detection approach, which achieved reasonable normalization of the cerebellum and robust damage detection.
\begin{table}
\caption{Performance (Dice) of the cerebellar damage detection tool on simulated datasets categorized by size and location.} \label{table:simulation}
\includegraphics[width=\textwidth]{figures/table_simulation.png}
\end{table}
\begin{figure}
\includegraphics[width=\textwidth]{figures/figure_simulation.png}
\caption{Visualization of representative examples of the simulated data and damage detection results. First row: the healthy 3D T1 brain image used for simulation. Second row: the simulated postoperative 3D T1 image after applying the simulated damage VOI. Third row: the automatically detected damage mask (AM, red) on the normalized cerebellum. Fourth row: the ground truth (GT, green) simulated damage mask on the SUIT template. A) Ventricular damages. Results are shown in the saggittal plane. B) Lateral damages. Results are shown in the coronal plane.} \label{fig:simulation}
\end{figure}
\section{Discussions}
Existing cerebellum normalization tools are typically designed for healthy brain imaging data, thus tend to fail when lesions are present. As a by-product of this cerebellar damage detection approach, the normalization of cerebellum is also suitable for broader neuroimaging applications (e.g., fMRI preprocessing and analysis), and is more accurate than SUIT normalization~\cite{diedrichsen2011imaging} on postoperative images. Our normalization is more accurate because (1) it has a more accurate cerebellum mask and WM/GM segmentation to start with, (2) labeled inputs allow for easier and better optimization, and (3) the choice of state-of-the art nonlinear registration algorithm, Syn~\cite{klein2009evaluation}.
Our work has some limitations: (1) An important assumption of this approach is the absence of the tumor in the image. Hence, this method cannot handle cases with partial tumor resection or metastases. (2) Our brain tissue segmentation algorithm assumes everything with abnormal intensity (hyper or hypo) is not healthy brain tissue, which may not always be the case. For example, brain tissue might be stained by blood and therefore have an abnormal appearance in the image, but be otherwise healthy. However, this is not easy to tell from image at the time point immediately following surgery. MR scans at later time points may further inform healthy or damaged anatomy after the stain has been removed via natural processes. (3) Our damage detection algorithm is not learning-based. As a result, its performance can not be improved by collecting more accurate labels (manually edited masks). Our future work is to develop a deep learning based model to further improve the performance by taking advantage of the inputs from neuroradiologists.
\section{Conclusions}
In this paper, we present a fully automatic localization and measurement of postoperative damage in the cerebellum in the atlas space using postoperative MRI. The proposed framework has a novel brain tissue segmentation algorithms that considers surgery-related features (e.g., postsurgical cavity, absorbable hemostat, blood products) and other non-surgical features (e.g., enlarged ventricles), a more accurate cerebellum normalization procedure, and provides quantification of cerebellar damage in the atlas space. This automation greatly reduced human annotation time from hours to about 10 minutes. The normalization proposed in this work is also helpful to perform other neuroimaging analyses (e.g., fMRI analyses) where cerebellum normalization is needed, especially on postoperative images.
\bibliographystyle{splncs04}
| 7a93d768f330d9d289b11983b2839b3a2e32c173 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{section:intro}
\subsection{Overview}
\IEEEPARstart{W}{i-Fi} sensing has recently received extensive research interests~\cite{liu2019wireless,nirmal2021deep}. Wi-Fi is ubiquitous as it has been equipped in many consumable electronics including laptops, smartphones, tablets, wearable devices such as Fitbits, and smart home appliances, to name but a few.
Various interesting applications have been inspired, including large-scale movements such as human activity recognition~\cite{wang2014eyes,li2021two,wang2018spatial}, fall detection \cite{wang2016rt}, \cite{zhang2019commercial}, and gait recognition\cite{wang2016gait}, \cite{zhang2019widigr}, as well as small-scale movements such as gesture recognition~\cite{li2020wihf,gao2021towards}, sign language recognition\cite{signfi}, \cite{shang2017robust}, and vital sign detection \cite{zeng2020multisense}, \cite{wang2016human}.
These applications are very useful for our everyday life. For example, gesture recognition can be applied in human-computer interaction and smart home to enable smooth control of devices.
Wi-Fi transmissions experience line-of-sight propagation, reflection, refraction, and scattering, which are affected by the environment and the objects within it~\cite{goldsmith2005wireless}.
When a person performs an activity, e.g., walking or hand moving, the activity will cause unique changes to the radio propagation, which can be measured via the channel state information (CSI). Deep learning can be adopted to learn the unique mapping between the CSI patterns and activity types due to its effective feature extraction and classification capability~\cite{krizhevsky2012imagenet}. A deep learning-based sensing protocol involves two stages. In the training stage, a training dataset should be constructed and a deep learning model is trained offline. Specifically, when an activity is performed, the Wi-Fi signals will get perturbed.
The receiver will estimate the CSI whose variation is affected by the activity. For each activity, many CSI records are collected. The process is repeated for all the activities and the training dataset is established. A deep learning model can then be trained offline, which only needs to be done once.
During the testing stage, based on the captured CSI, the derived deep learning model will be utilized to infer the activity type.
\subsection{Limitations of Existing Systems}
Although the above deep learning-based sensing systems can recognise human activities with reasonable accuracy, there are still limitations that restrict practical applications.
\subsubsection{Limited Scalability}
A classic deep learning-based sensing system explores the uniqueness of the mapping between the CSI variations and activities. A common limitation of such systems is that they can only handle the same classes/activities as the ones used during the training process. Whenever a new class needs to be added to or an existing class needs to be removed from the well-trained model, the model should be retrained using data of the existing classes and the new classes. The retraining results in three types of overhead:
\begin{itemize}
\item The storage of the existing training data, which occupies hard disk space and requires extra maintenance.
\item The collection of the new data that is time consuming and labour extensive.
\item The computational overhead for retraining is quite high when there is a large amount of data, which cannot be completed in real time.
\end{itemize}
However, adjusting including adding and deleting classes is common in practical applications, e.g., a new sign gesture may need to be enrolled. The lack of scalability in the current approach and design makes it difficult to be deployed in real world scenarios.
\subsubsection{Domain Shift}
The radio signal experiences multipath propagation whose characteristics are determined by the environment and the sensing activity.
The same activity performed in different environments will cause deviated CSI variations.
In addition, the same activity performed by different users may also have various CSI variations because the users may perform the same activity in a slightly different way or their heights/body shapes are different.
The parameters uncorrelated to the activity can be denoted as a domain~\cite{kang2021context}, e.g., the environment and/or the user. Source and target domains refer to the domains of training and testing stages, respectively.
When the source and target domains are different, there is a domain shift. The obtained CSI will be varied, and the recognition will be impacted.
However, it is very common to apply Wi-Fi sensing in different environments and/or for different users.
Existing solutions can be classified into three categories, namely domain adversarial training-based~\cite{jiang2018towards,kang2021context}, transfer learning-based~\cite{CrossSense}, and domain-independent feature-based~\cite{widar} approaches.
However, these approaches are subject to the following limitations:
\begin{itemize}
\item The domain adversarial training-based approach requires the training dataset to cover data from many domains but there are numerous possible domains.
\item The transfer learning-based approach needs extensive data from the target domain, which sometimes may not be possible.
\item The domain-independent feature-based method relies on special knowledge to design unique features and may also need extra receivers, e.g., body-coordinate velocity profile (BVP) in~\cite{widar}. In addition, it may suffer from high computational costs, which make it impossible for practical use.
\end{itemize}
Hence, a lightweight cross-domain sensing approach is urgently required to reduce the overhead of collecting data from source or target domains as well as eliminate constrains of extra hardware and computational resources.
\subsection{Few-Shot Learning}
Few-shot learning (FSL) has been widely investigated in computer vision for image classification~\cite{vinyals2016matching}.
A classic deep learning model aims to learn the unique feature of each class and predict the label based on the feature mapping. In contrast, FSL is a meta-learning technique. Instead of learning unique features, it makes the prediction based on the similarity of the feature sets.
Specifically, a feature extractor will be pre-trained using the base set. Then, a support set will be constructed with a few pairs of data and labels.
Finally, a query set contains data whose label is to be inferred. The prediction is made by comparing the similarity between features of support and query sets.
FSL has recently been applied to Wi-Fi sensing as well~\cite{ding2020rf,xiao2021onefi} to address the scalability and cross-domain issues.
The support set can be flexibly adjusted because very few samples are required for each class, and no cumbersome training is needed.
The performance of FSL relies on the generalisation capability of the feature extractor, which however requires an extensive base set. Hence, the work in~\cite{xiao2021onefi} designed a virtual gesture generation algorithm to transform existing gestures into virtual gestures. The base set can thus be enriched and the data collection overhead can be mitigated.
However, designing the virtual gesture generation algorithm requires sophisticated knowledge which cannot be extended to other tasks straightforwardly.
\subsection{Contributions}
In the light of the initial success of using FSL in designing the virtual gesture generation algorithm~\cite{xiao2021onefi}, in this work, we further employ FSL to address the scalability issue and domain shift challenge faced by many practical Wi-Fi sensing applications.
We demonstrated that the feature extractor can be generalised to a totally different sensing task. Specifically, we exploited a public dataset as the base set, applied the trained feature extractor to a different dataset and achieved good classification performance, which is a different approach from~\cite{xiao2021onefi}.
In other word, our work aims to eliminate the need to construct an individual base set for every sensing task.
We adopt metric-based FSL for designing a scalable and cross-domain Wi-Fi sensing system. A CNN backbone is revised from the classic AlexNet architecture as the feature extractor. We design a collaborative sensing scheme to leverage the spatial diversity gain enabled by multiple receivers. The proposed model is evaluated using three public datasets, i.e, SignFi~\cite{signfi}, Widar 3.0~\cite{widar} and Wiar~\cite{guo2019wiar}. We study in-domain sensing, cross-domain sensing as well as cross-dataset sensing. Our results show that the proposed FSL-based sensing system has a better performance over the state-of-the-art works in the literature. Our contributions are summarised as follows:
\begin{itemize}
\item We propose an FSL-based Wi-Fi sensing system, FewSense, which is capable of novel class recognition and cross-domain sensing.
Specifically, the proposed system can achieve $76.3\%$ cross-domain accuracy for $76$ novel sign language gestures using only one labelled sample for each novel class from the target domain, when all the data is from the SignFi dataset.
\item FewSense can be fine-tuned to new sensing tasks using a few samples. When the feature extractor is trained on the SignFi dataset (sign language recognition), evaluation over the Widar (gesture recognition) and Wiar (human activity recognition) datasets can achieve an average accuracy of 96.5\% and 82.7\%, respectively, with five samples from each novel class.
\item A collaborative sensing approach is proposed to leverage the spatial diversity gain when there are multiple receivers. The classification accuracy can be boosted to $100\%$ with six receivers, when it is tested using the Widar dataset.
\end{itemize}
The rest of this paper is organised as follows.
Section~\ref{statement} shows the CSI model in the Wi-Fi sensing system and states the problem of the existing system.
Then we present the overview of our system design in Section~\ref{systemOverview}.
Section~\ref{sec:method} introduces our metric-based FSL method in details and
Section~\ref{sec:collab} presents our collaborative sensing model to enable multiple receivers to work collaboratively. The performance evaluation of FewSense is given in Section \ref{sec:exper}.
We review innovative works of FSL and cross-domain Wi-Fi sensing works in Section \ref{relate}.
Finally, Section \ref{conclusion} concludes the paper.
\section{Background and Motivation}\label{statement}
\subsection{Background}\label{sec:system_model}
As shown in Fig.~\ref{fig:system_model}, when a signal is sent from a transmitter and captured by a receiver, it experiences various propagation paths in a multipath environment. Specifically, there are static paths, which may include line-of-sight (LOS) and signals reflected by static objects such as walls and furniture, e.g., tables. There are also dynamic paths, which are reflected and/or scattered by moving objects in the environment~\cite{wang2015understanding}. Taking gesture recognition as an example, the movement of the palm will lead to dynamic paths of signal propagation.
\begin{figure}[!t]
\centering
\includegraphics[width =3.4in]{./HandDrawSystemModel.pdf}
\caption{Illustration of Wi-Fi sensing. Gesture recognition is shown as an example.}
\label{fig:system_model}
\end{figure}
The channel model can be mathematically given as
\begin{align}
h(\tau,t) = \sum_{s \in {S}} h(\tau_s,t) \delta(\tau - \tau_s) + \sum_{d \in {D}} h(\tau_d,t) \delta(\tau - \tau_d),
\end{align}
where $h(\tau_s,t)$ and $h(\tau_d,t)$ are the channel attenuation of the static (grouped as $\{\mathcal{S}\}$) and dynamic paths (grouped as $\{\mathcal{D}\}$), respectively, and $\delta(\cdot)$ is the Dirac delta function.
In the context of Wi-Fi sensing, orthogonal frequency-division multiplexing (OFDM) is the physical layer modulation used for IEEE 802.11a/g/n/ac/ax, which can obtain CSI in the frequency domain, given as
\begin{align}
\label{eqn:csimodel}
H(f, t) = \sum_{s \in {S}} h(\tau_s, t) e^{-j 2 \pi f \tau_{s}} + \sum_{d \in {D}} h(\tau_d, t) e^{-j 2 \pi f \tau_{d}}.
\end{align}
The CSI variations over time are directly affected by the objects and environments. Different gestures will have different moving patterns, which lead to unique CSI variations.
Deep learning can be adopted to reveal the relationship between the CSI variations and the corresponding gestures. By collecting a series of continuous packets when a gesture is performed, we can estimate and store records of CSI. During the training stage, the deep learning model can learn the specific patterns related to different gestures. Then in the testing stage, the receiver will obtain a collection of CSI and infer the gesture based on the pre-trained deep learning model.
\subsection{Motivation}
Deep learning has been widely adopted for Wi-Fi sensing and achieved excellent classification performance. Convolutional neural networks (CNN) \cite{signfi,ma2021location}, and long-short term memory (LSTM)~\cite{zhang2020data,Yang2019LearningGestures} can learn and reveal complex feature patterns in a supervised learning manner. However, they are still subject to novel class recognition and cross-domain sensing.
\subsubsection{Limited Scalability/Novel Class Recognition}
In real-life applications, it is common to add new gestures to and/or remove existing ones from a gesture recognition system, which will result in changes of the number of classes (gestures).
However, as already discussed, the classic deep learning techniques such as CNN and LSTM have limited scalability because they can only handle fixed-size classes once trained. Whenever new classes, e.g., unseen hand gestures, should be added, the neural network model needs to be retrained with a massive number of training samples.
Data collection and retraining overheads thus occur~\cite{ding2020rf,xiao2021onefi}, which are time-consuming and laborious.
\subsubsection{Cross-Domain Sensing}
In order to achieve a good classification performance, the training and test datasets should share the same data distribution. Most existing works let the same user perform the gestures in the same environment \cite{wang2016wifall,wang2015understanding}. In another word, the static CSI part in (\ref{eqn:csimodel}) remains the same but the dynamic part is uniquely caused by the gestures of the user. This is termed in-domain sensing; here the domains refer to the environment and the user.
However, cross-domain sensing is actually more common in real-life deployments.
The gestures will be probably performed by different users in different environments, i.e., cross-domain. The domain shift impacts the performance of Wi-Fi sensing systems due to two reasons.
\begin{itemize}
\item When the environment changes, the static parts of (\ref{eqn:csimodel}) for training and testing are different, which will result in different CSI patterns, i.e., varied data distributions.
\item Different users may perform the same gesture in slightly different ways, which will also lead to different CSI dynamic parts of (\ref{eqn:csimodel}).
\end{itemize}
In order to illustrate the domain shift, we use sign languages from a public dataset SignFi~\cite{signfi} that will be introduced in Section~\ref{sec:dataset}.
Specifically, we selected three sign gestures that are performed by two users at home and in a lab environment. Their distributions are visualised using t-SNE~\cite{van2008visualizing}, which maps the CSI samples to a two-dimensional space. As shown in Fig.~\ref{fig:t_sne}, the same sign performed in different environments or by different users has totally different distributions. Therefore, directly applying the neural network model trained on one domain to a new domain would not work.
\begin{figure}[!t]
\centering
\includegraphics[width =3.4in]{./t_SNE.pdf}
\caption{t-SNE visualisation of the selected data from different domains: the same sign language from the same user but different environments and different sign languages from the same environment but different users.}
\label{fig:t_sne}
\end{figure}
\section{System Design}\label{systemOverview}
\subsection{System Overview}
This paper employs metric-based FSL to address the above limitations. Our goal is to recognise novel classes in cross-domain scenarios, i.e., different environments and/or users, using only a few samples. Specifically, a three-stage approach is designed, namely feature extractor training, feature matrix generation, and classification, as shown in Fig.~\ref{fig:system_overview}. Note that the same signal preprocessing algorithm is used for all three stages.
\begin{figure}[!t]
\centering
\includegraphics[width =3.4in]{./Systemoverview.pdf}
\caption{The proposed metric-based FSL system for Wi-Fi sensing.}
\label{fig:system_overview}
\end{figure}
\subsection{Feature Extractor Training}
We will first sample and preprocess raw CSI records and obtain CSI tensors $\mathbf{x}$, whose construction will be elaborated in Section~\ref{sec:preprocess}.
A base set, $\mathbf{B}=\left\{\left(\mathcal{X}^{b}, \mathcal{Y}^{b}\right)\right\}$, will be constructed by obtaining CSI tensors for different classes (e.g., sign gestures),
where $\mathcal{X}^{b}=\left\{\mathbf{x}^{b}_{1}, \mathbf{x}^{b}_{2}, \cdots, \mathbf{x}^{b}_{\tilde{N}_b}\right\}$ are the CSI tensors and $\mathcal{Y}^{b}=\left\{y_{1}^b, y_{2}^b, \cdots, y_{\tilde{N}_b}^b\right\}$ are the corresponding labels.
The number of instances in the base set is denoted as $\tilde{N}_b$, and $\tilde{N}_b = N_b \times K_b$, where $N_b$ is the number of base classes and $K_b$ is the number of samples per base class.
The domain of the base set is referred as the source domain, denoted as $\mathcal{D}_{s}$.
We use a standard supervised learning manner to train a feature extractor $f_{\theta}$ and a classifier $C\left(\cdot \mid \mathbf{W_b}\right)$, parametrised by $\mathbf{W_{\theta}}$ and $\mathbf{W_{b}}$, respectively. A large number of classes in the base set enables the feature extractor to learn unique latent features from base classes so that the distances between instances of the same class are closer while instances of different classes are further apart.
Once the training is completed, the classifier will be removed to obtain the trained feature extractor. The training only needs to be done once.
\subsection{Feature Matrix Generation}
\begin{figure}[!t]
\centering
\includegraphics[width=3.4in]{./DirectFeaturematrix.pdf}
\centering
\caption{The structure of two feature matrix generation methods.(a) Direct feature matrix generation. (b) Fine-tuned feature matrix generation.}
\label{fig:twomethods}
\end{figure}
The data involved in this stage is called the support set, which is defined as $\mathbf{S}=\left\{\left(\mathcal{X}^{s}, \mathcal{Y}^{s}\right)\right\}$, where $\mathcal{X}^{s}=\left\{\mathbf{x}^{s}_{1}, \mathbf{x}^{s}_{2}, \cdots, \mathbf{x}^{s}_{\tilde{N}_s}\right\}$ and $\mathcal{Y}^{s}=\left\{y_{1}^s, y_{2}^s, \cdots, y_{\tilde{N}_s}^s\right\}$ are the CSI tensors and labels, respectively.
$\tilde{N}_s = N \times K$, where $N$ is the number of novel classes, and $K$ is the number of samples for each class. FSL is usually described as $N$-way $K$-shot classification.
The classes in the support set have no overlap with the classes in the base set, i.e., $\mathcal{Y}^{b} \cap \mathcal{Y}^{s}=\emptyset$. Therefore, they are called novel classes.
The domain of the support set is named as the target domain, denoted as $\mathcal{D}_t$. In practice the source domain and target domain are likely different, i.e., $\mathcal{D}_t \neq \mathcal{D}_s$, thus cross-domain sensing occurs.
We first collect $K$ labelled shots (samples) for each novel class to construct the support set, which will be denoised by the preprocessing scheme.
With samples from the support set, the pre-trained feature extractor $f_{\theta}$ will learn the latent feature representations of the novel classes, termed as the feature matrix in this paper. We propose two methods to generate it, i.e., direct feature matrix generation and fine-tuned feature matrix generation, as shown in Fig.~\ref{fig:twomethods}.
\begin{itemize}
\item Direct Feature Matrix Generation: The feature extractor will extract the feature embedding directly from the support set. The feature embedding is stored in the database and output as the feature matrix.
\item Fine-tuned Feature Matrix Generation: A new classifier will be introduced after the feature extractor. The weight matrix of the classifier and the feature extractor will be optimised using the samples in the support set. The adapted weight matrix will be the feature matrix.
\end{itemize}
\subsection{Classification }
The data involved in this stage is named as the query set
$\mathbf{Q}=\left\{\left(\mathcal{X}^{q}, \mathcal{Y}^{q}\right)\right\}$, where $\mathcal{X}^{q}=\left\{\mathbf{x}^{q}_{1}, \mathbf{x}^{q}_{2}, \cdots, \mathbf{x}^{q}_{\tilde{N}_q}\right\}$ are the CSI tensors and $\mathcal{Y}^{q} =\left\{\mathbf{y}^{q}_{1}, \mathbf{y}^{q}_{2}, \cdots, \mathbf{y}^{q}_{\tilde{N}_q}\right\}$ represents the labels that to be inferred.
Note that the support set and query set share the same label space and are performed in the same domain. Therefore, they have the same distribution.
The signal preprocessing scheme is used to remove the phase noise of the query set. The feature extractor, $f_{\theta}$, is used to extract the latent features from the denoised query set. The final results are determined by computing the cosine similarity score between the latent features of the query set and the feature matrix generated from the previous stage.
\section{Methodology}\label{sec:method}
In this section, we first introduce the signal preprocessing algorithm that is used in all three stages of the proposed FSL-based system. We then elaborate on each of these three stages.
\subsection{Signal Preprocessing} \label{sec:preprocess}
A Wi-Fi transmitter will continuously send packets that are captured by receivers within the communication range.
The CSI can be estimated but will be impacted by sampling frequency offset (SFO), packet detection delay (PDD) and carrier frequency offset (CFO)~\cite{xie2018precise}, which can be given as
\begin{align}
\widehat{H}(f,t) = e^{-j\phi(t)}H(f,t),
\end{align}
where $\phi(t)$ is the phase shift collectively caused by the above issues.
Most of the existing works only employ amplitude for sensing. The phase of CSI is more sensitive than the amplitude~\cite{wang2016rt} and will be beneficial to the system performance.
In this work, we leverage a linear transformation proposed in~\cite{wang2015phasefi} to sanitise the CSI. According to the specific implementation of the Intel 5300 Wi-Fi network interface card (NIC)~\cite{Halperin_csitool}, the process can be given as
\begin{equation}
\angle \widetilde{H}(f_i,t) = \angle \widehat{H}(f_i,t) - k m_i - b,
\end{equation}
where $i$ is the subcarrier index ranging from 1 to 30, $f_i$ is the corresponding subcarrier frequency, $m_i$ is the subcarrier index ranging from -28 to 28, and
\begin{align}
k&=\frac{\angle \widehat{H}(f_{30},t)-\angle \widehat{H}(f_{1},t)}{m_{30}-m_1},\\
b&=\frac{1}{30} \sum_{i=1}^{30} \angle \widehat{H}(f_i,t).
\end{align}
In this work, we concatenate the amplitude, $|\widehat{H}(f_i,t)|$, and the sanitised phase, $\angle \widetilde{H}(f_i,t)$. By combining the data from all the antennas, we finally construct a CSI tensor $\mathbf{x}\in \mathbb{R}^{U_s \times U_{ap} \times U_{ant}}$, where $U_s$ is the number of sampling points, $U_{ap}$ is the number of elements of amplitude plus phase, and the $U_{ant}$ is the number of antenna pairs.
\subsection{Feature Extractor Training Stage}
The main purpose of the feature extractor training stage is to train a feature extractor that can extract discriminative features from input samples.
The structure of the feature extractor is shown in Fig. \ref{fig:Alexnet}, which is revised from the classic AlexNet architecture~\cite{krizhevsky2012imagenet}. An $L_2$-norm layer~\cite{ranjan2017l2} is added before the classifier, which normalises the embedded latent feature vectors as follows:
\begin{equation}
\label{eqn:l2norm}
f_{\theta}(\mathbf{x})=\frac{f^{\prime}_{\theta}(\mathbf{x})}{\|f^{\prime}_{\theta}(\mathbf{x})\|_{2}},
\end{equation}
where $f^{\prime}_{\theta}(\mathbf{x})$ is the output of the previous fully connected (FC) layer, and $\|\cdot\|_{2}$ denotes the $L_2$-norm operation.
By adding an $L_2$-norm layer, the embedded features will be forced to lie on a hypersphere of a fixed radius~\cite{ranjan2017l2}. Adopting $L_2$-norm layer will improve the system's converging speed and accuracy, because it forces the embedded features from the same class closer and moves the features from different classes further apart.
\begin{figure}[!t]
\centering
\includegraphics[width = 3.4in]{./cnn_architecture.pdf}
\caption{The architecture of the feature extractor, modified AlexNet.
}
\label{fig:Alexnet}
\end{figure}
A classifier, $C\left(\cdot \mid \mathbf{W}\right)$, consists of a fully connected layer with softmax activation function. In this stage, the classifier is parametrised by the weight matrix $\mathbf{W_{b}}\in \mathbb{R}^{n \times N_b}$, where $n$ is the dimension of extracted features from the feature extractor. The classifier takes the normalised feature and then computes $(\mathbf{W_b})^{\top} f_{\theta}\left(\mathbf{x}^{b}_{i}\right)$, where $(\cdot)^{\top}$ denotes transpose operation. Finally, the prediction is made via the softmax function, mathematically given as
\begin{equation}\label{eqn:cos}
\sigma(\mathbf{v})_{i} = \frac{e^{v_i}}{\sum_{j=1}^{G} e^{v_{j}}},
\end{equation}
where $\mathbf{v}$ is the output of the fully connected layer, and $G$ is the number of classes. Here, $\mathbf{v} = (\mathbf{W_b})^{\top} f_{\theta}\left(\mathbf{x}^{b}_{i}\right)$ and $G = N_b$. The output of the softmax function is a probability vector that represents confidence levels over different classes. The class with the highest probability is selected.
We train the feature extractor using the samples of the base set, $\mathbf{B}$, in a supervised learning manner.
The feature extractor will learn to extract distinct features from the CSI samples of different classes.
It is crucial to carry out training with a sufficient number of base classes to enable the feature extractor to generalise to novel classes.
Once the training is completed, the classifier is removed and the feature extractor is attained.
\subsection{Feature Matrix Generation Stage}
This stage aims to generate discriminative features for novel classes in the support set.
Firstly, one or a few samples from each novel class are obtained for the feature matrix generation. The corresponding CSI instances are added into the support set, $\mathbf{S}$. The signal preprocessing scheme will be performed on the support set data to get the input for the feature extractor. Depending on different deployment scenarios, two methods can be used in the feature matrix generation stage.
\subsubsection{Direct Feature Matrix Generation}
The feature extractor extracts the latent feature vectors from the denoised data. The support set embedding $\mathbf{F}_{s}\in \mathbb{R}^{n \times N}$ is obtained by computing the mean of the feature vector for each class, given as
\begin{equation}
\label{supportembedding}
\mathbf{F}_{s}=\left[\begin{array}{c}
f_{\theta}{(\mathbf{x}^{s}_1)} ,
f_{\theta}{(\mathbf{x}^{s}_2}) ,
\dots ,
f_{\theta}{(\mathbf{x}^{s}_{N})}
\end{array}\right].
\end{equation}
The feature embeddings are saved in the database for future use. In this case, the output matrix $\mathbf{M}$ of this stage is the feature embedding, i.e., $\mathbf{M} = \mathbf{F}_{s}$, which will be fed into the next stage for classification. This method allows fast generating distinctive feature matrix directly from the support set with little overhead.
\subsubsection{Fine-Tuned Feature Matrix Generation}
In some challenging scenarios, the distributions of datasets or fundamental features may change dramatically, the feature extractor trained on the base set will fail in new sensing tasks. In order to adapt the feature extractor to new tasks,
we propose fine-tuned feature matrix generation scheme based on the fully labelled support set.
A classifier with the trainable weight matrix $\mathbf{W}_{\text{f}} \in \mathbb{R}^{n \times N}$, initialised by the support set embedding $\mathbf{F}_{s}$, is added after the pre-trained feature extractor.
The output of the softmax function is $\mathbf{\sigma}((\mathbf{W}_{\text{f}})^{\top} f_{\theta}{(\mathbf{x}^{s}_i)})$.
Fine-tuning is done by minimising the softmax loss based on the support set, which will update the parameters in the feature extractor $f_{\theta}$ and the weight matrix of the classifier $\mathbf{W}_\text{f}$.
The output matrix of this method is the weight matrix of the classifier, i.e., $\mathbf{M}$ = $ \mathbf{W }_\text{f}$. The fine-tuned feature extractor will be used in the next stage. The fine-tuned feature matrix generation method enhances the adaptivity of the model in a new sensing domain or scenario.
\subsection{Classification Stage}
In this stage, the sample in the query set will be classified based on the cosine similarity score between the feature matrix and the extracted latent features of the query set.
An instance from the query set $\mathbf{Q}$ will be firstly processed by the preprocessing scheme. The feature extractor will extract an embedded feature vector $f_{\theta}{(\mathbf{x}^{q})} \in \mathbb{R}^{n \times 1} $.
We compute the cosine similarity score of $f_{\theta}{(\mathbf{x}^{q})}$ pair-wisely with all the elements in $\mathbf{M}$ by
\begin{equation}\label{eqn:cos}
\mathbf{v} = \frac{\mathbf{M}^{\top} f_{\theta}{(\mathbf{x}^{q})}}{\|\mathbf{M}\|_2\|f_{\theta}{(\mathbf{x}^{q})}\|_2}.
\end{equation}
The softmax function $\sigma$ maps the score, $\mathbf{v}$, to a probability distribution, $\mathbf{P}$, and the query set will be classified to the class with the highest probability.
\subsection{Summary}
FSL-based sensing has a good scalability capacity. Whenever novel classes need to be added, a few samples of these classes should be collected and their feature representations can be generated with acceptable overhead.
Let us revisit Fig.~\ref{fig:t_sne} for analysing cross-domain sensing. Data of the same class from the same domain is well-clustered while data of different classes from different domains is well-separated.
As the query set and support set are obtained from the same target domain, these samples experience the same domain shift from the base set.
By using fine-tuning, we can adapt the feature extractor and classifier weights using a few samples from the support set, i.e., a transformation from the source domain to the target domain. This approach enables the model to adapt to the target domain with minimal efforts.
\section{Collaborative Sensing}\label{sec:collab}
In the previous sections, we have introduced a Wi-Fi sensing system using one transmitter-receiver setup, which can capture features from a particular direction. However, the position and orientation of the user cannot be predicted, hence some transmitter-receiver links may not be optimal to capture CSI variations.
The performance can be boosted by employing multiple receivers, which can enrich feature observations from different directions and work collaboratively to improve system robustness and classification accuracy.
As shown in Fig.~\ref{fig:collabrative}, there are $N_r$ receivers deployed for collaborative sensing. Each receiver will be equipped with the fine-tuning-FSL-based sensing approach introduced in Section~\ref{sec:method}. When a signal is sent by a transmitter, all the receivers will receive the signal simultaneously but from different directions.
Each receiver will be initially deployed with the same feature extractor.
For the $i$-th receiver, it can construct its own support set, $\mathbf{S}_{i} = {(\mathcal{X}^{s}_i, \mathcal{Y}^{s}_i)}$.
The $i$-th receiver can then carry out fine tuning to adapt its feature extractor $f_{\theta}^{i}$ with weights $\mathbf{W}_{\theta}^{i}$ and the classifier with the weight matrix $\mathbf{W}_{\text{f}}^{i}$.
\begin{figure}[t]
\centering
\includegraphics[width = 3.4in]{collaborative_sensing_2.pdf}
\caption{Collaborative sensing using multiple receivers.}
\label{fig:collabrative}
\end{figure}
In the classification stage, each receiver will also collect its own query set, $\mathbf{Q}_i$.
The $i$-th receiver will first carry out the independent inference and obtain a probability distribution, $\mathbf{P}_{i}$. They will then work together by calculating an averaged probability distribution of all their observations, given as
\begin{equation}\label{eqn:mojorityvote}
\mathbf{P}^{c} = \frac{1}{N_r}\sum_{i=1}^{N_{r}} {\mathbf{P}_{i}}.
\end{equation}
The class with the highest probability is predicted as the final decision.
\section{Experimental Evaluation}\label{sec:exper}
\subsection{Setup of Experiments}
\subsubsection{Dataset} \label{sec:dataset}
We used three public datasets for evaluation, i.e., SignFi~\cite{signfi}, Widar 3.0~\cite{widar} and Wiar~\cite{guo2019wiar}.
The CSI of all these three datasets is collected by Intel 5300 Wi-Fi NIC using the IEEE 802.11 CSI tool~\cite{Halperin_csitool}. This tool reports the CSI of 30 subcarrier groups for each packet at 20 MHz channel spacing.
The information of the three datasets is summarised in Table~\ref{tab:dataset}. The datasets are collected by different research groups hence their volunteers and experimental experiments are different.
For each activity, we selected $U_{s} = 200$ packets to obtain a uniform input.
As there are 30 subcarrier groups in each packet and each subcarrier has amplitude and phase information, thus $U_{ap} = 60$.
Finally, there are three transmitter-receiver antenna pairs, i.e., $U_{ant} = 3$.
Thus, we have the CSI tensors: $\mathbf{x}\in \mathbb{R}^{200 \times 60 \times 3}$.
\begin{table}[]
\centering
\caption{Summary of Datasets.}
\label{tab:dataset}
\begin{tabular}{|l|l|l|L{1.7cm}|L{1.3cm}|}
\hline
\multicolumn{1}{|l|}{} & \multicolumn{1}{l|}{\textbf{Environment}} & \textbf{User ID} & \textbf{\# classes $\times$ \# reps $\times$ \# Rx} & \textbf{\# samples} \\ \hline
\multirow{6}{*}{\textbf{SignFi}} & Lab & User s5 & 276$\times$20$\times$1 & 5520 \\ \cline{2-5}
& \multirow{4}{*}{Lab 2} & User s1 & 150$\times$10$\times$1 & 1500 \\ \cline{3-5}
& & User s2 & 150$\times$10$\times$1 & 1500 \\ \cline{3-5}
& & User s3 & 150$\times$10$\times$1 & 1500 \\ \cline{3-5}
& & User s4 & 150$\times$10$\times$1 & 1500 \\ \cline{2-5}
& Home & User s5 & 276$\times$10$\times$1 & 2760 \\ \hline
\multirow{3}{*}{\textbf{Widar}} & \multirow{3}{*}{Classroom} & User w1 & 6$\times$20$\times$6 & 720 \\ \cline{3-5}
& & User w2 & 6$\times$20$\times$6 & 720 \\ \cline{3-5}
& & User w3 & 6$\times$20$\times$6 & 720 \\ \hline
\multirow{6}{*}{\textbf{Wiar}} & \multirow{6}{*}{Meeting room} & User a1 & 16$\times$30$\times$1 & 480 \\ \cline{3-5}
& & User a2 & 16$\times$30$\times$1 & 480 \\ \cline{3-5}
& & User a3 & 16$\times$30$\times$1 & 480 \\ \cline{3-5}
& & User a4 & 16$\times$30$\times$1 & 480 \\ \cline{3-5}
& & User a5 & 16$\times$30$\times$1 & 480 \\ \cline{3-5}
& & User a6 & 16$\times$30$\times$1 & 480 \\ \hline
\end{tabular}
\end{table}
\textbf{SignFi dataset}. The SignFi dataset~\cite{signfi} contains CSI samples of 276 different sign language gestures, which involve head, arm, hand and finger gestures. The gestures are performed by five users and in three environments. Specifically, the user s5 performs all the 276 sign language gestures at home and in a lab environment. Users s1 to s4 also carry out a subset of the gestures in the same lab room, but the settings vary, e.g., with different laptop placements and desk and chair arrangements. Hence, the lab environment for users s1 to s4, denoted as Lab 2 in this paper, is deemed different from the lab environment for user s5. User s5 performs 276 sign languages each with 20 times in the lab and 10 times at home. Users s1 to s4 perform 150 sign languages 10 times in the lab 2 environment.
\textbf{Widar dataset}. The Widar dataset~\cite{widar} is a large dataset, and we only used part of it. Specifically, we selected six gestures, namely Push \& Pull, Sweep, Clap, Slide, Draw-Zigzag, and Draw-N. These gestures involve arm and hand movements. Six receivers are used to capture each gesture. We used data from three users, denoted as users w1, w2, and w3. Each user performs 6 gestures 20 times in a classroom environment.
\textbf{Wiar dataset}. The Wiar dataset~\cite{guo2019wiar} contains 16 different human motions, namely Horizontal Arm Wave, High Arm Wave, Two Hands Wave, High Throw, Draw X, Draw Tick, Toss Paper, Forward Kick, Side Kick, Bend, Hand Clap, Walk, Phone Call, Drink Water, Sit Down, and Squat. These activities involve torso, arm and hand movements.
We used the data of six users, denoted as a1 to a6, for our experiments. Each user performs each activity 30 times in a meeting room.
\subsubsection{Training Configuration}
The feature extractor was trained in a supervised learning manner. The data of user s5 lab environment of the SignFi dataset was used as the base set. We used the Adam optimiser with an initial learning rate of $1e^{-3}$ to minimise the cross-entropy loss function. The learning rate was multiplied by 0.1 whenever the validation loss stopped decreasing for 20 epochs. The training process of the feature extractor terminated when the validation loss stopped decreasing for 50 epochs. All experiments were run on a PC with i7-8700K 3.7 GHz CPU, 16 GB memory with NVIDIA GeForce GTX 2080Ti. Tensorflow and Keras were used.
\subsubsection{Evaluation Method}
This paper evaluated the in-domain sensing (Section~\ref{sec:indomain}), cross-domain sensing (Section~\ref{sec:crossdomain}), cross-dataset evaluation (Section~\ref{sec:crossdataset}), and collaborative sensing (Section~\ref{sec:collaborative}).
The base set was constructed by randomly selecting $N_b$ classes each with $K_b$ samples from the user s5 lab environment in the SignFi dataset. The feature extractor, denoted as $f_\theta^{N_b}$, was trained from scratch using the base set.
We randomly selected $N$ different classes (ways) each with $K$ samples (shots) from SignFi, Widar, or Wiar datasets to construct the support set $\mathbf{S}$. The remaining data of each selected class was used as the query set.
Classification accuracy and confusion matrix were used as the metrics.
The accuracy was defined as the number of correct predictions divided by the total predictions. The confusion matrix was used to show the number of correct and incorrect predictions.
\subsection{In-Domain Sensing}\label{sec:indomain}
In-domain sensing is evaluated in this section, i.e., the base set, support set and query set share the same domain.
Intuitively, a feature extractor will have better generalisation capability when more base classes are available.
We used the data from the user s5 collected from the lab environment in the SignFi dataset as the base set. We trained feature extractors, $f_\theta^{N_b}$, with $N_b$ base classes. Specifically, we studied $N_b = $ 50, 100, 150, 200, and 250. These $N_b$ classes were randomly selected from all the available classes, i.e., 276 classes in total.
The rest of the classes of the user s5 lab environment were used to evaluate the feature extractor's novel class recognition performance.
Fine-tuning was not applied.
We evaluated different number of ways, i.e., $N$ = 2 to 26, and one shot, i.e., $K$=1.
As shown in Fig. \ref{fig:comparebasedclasses}, a larger $N_b$, i.e., more base classes, leads to a higher accuracy, which is expected. When there are more novel classes in the support set, i.e., a higher $N$, the accuracy is decreasing, because it requires the feature extractor with better generalisation capability. Specifically, the average accuracies over all different numbers of novel classes are 97.9\% and 99.2\% when $N_b = 200$ and $N_b = 250$, respectively.
In the rest parts of this paper, we used $f_\theta^{200}$ as the feature extractor. To further evaluate the performance of the feature extractor $f_\theta^{200}$, we increased the number of novel classes to 76. As shown in Table~\ref{tab:user5InLab}, the classification accuracy is $91.7\%$ when $N=$ 76, which is still quite high.
\begin{figure}[!t]
\centering
\includegraphics[width = 3.4in]{./compareBaseClasses.pdf}
\caption{Performance of in-domain sensing. Feature extractors are trained with different numbers of base classes.}
\label{fig:comparebasedclasses}
\end{figure}
\begin{table}[]
\centering
\caption{In-Domain Sensing Performance of the Feature Extractor $f_\theta^{200}$.}
\label{tab:user5InLab}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
\textbf{\# ways} & \textbf{30} & \textbf{40} & \textbf{50} & \textbf{60} & \textbf{70} & \textbf{76} \\ \hline
\textbf{Accuracy} & 95.4\% & 95.8\% & 95\% & 93.3\% & 92.3\% & 91.7\% \\ \hline
\end{tabular}
\end{table}
\subsection{Cross-Domain Sensing}\label{sec:crossdomain}
In this section, we evaluated the cross-domain sensing performance. The feature extractor $f_\theta^{200}$ is trained with the user s5 lab environment but the support and query sets are from different domains.
We have considered two scenarios:
\begin{itemize}
\item \textbf{Scenario 1, Cross-Environment Evaluation}: The support and query sets are from the user s5 home environment, i.e., the same user but different environments. As $N_b$ = 200 classes are used for training the feature extractor, there are 76 remaining novel classes.
\item \textbf{Scenario 2, Cross-Environment and Cross-User Evaluation}: The support and query sets are from users s1 to s4 in the lab 2 environment, i.e., different users and different environments. In the SignFi dataset, these four users performed 150 classes of sign language and 25 of them are novel classes.
\end{itemize}
\subsubsection{Impact of Numbers of Novel Classes} \label{sec:novel_class}
We evaluated $N$-way 1-shot learning and fine-tuning was not used.
Only the data of novel classes was used, hence we evaluated $N$ = 2 to 76 and $N$ = 2 to 25 for scenarios 1 and 2, respectively.
As shown in Fig.~\ref{fig:NoClasses}, the accuracies of both scenarios decrease when the number of novel classes, $N$, is increasing, as a better generalisation capability of the feature extractor is required.
The accuracy of scenario 1 dropped to $57.9\%$ when $N$ = 76.
Regarding scenario 2, the results in Fig.~\ref{fig:crossDiffUserandEnvironment} demonstrate that the system can learn the distinctive representation of the novel classes, even though the feature extractor has not seen data from these environments and users before.
The accuracies vary for different users because each user may perform the same sign language in a different manner, which results in different data distribution.
\begin{figure}
\centering
\subfloat[]{\includegraphics[width=3.4in]{crossenvir_user5.pdf}
\label{fig:CrossDomainSameUser}}
\subfloat[]{\includegraphics[width=3.4in]{crossenvir_user1234.pdf}
\label{fig:crossDiffUserandEnvironment}}
\caption{Performance of cross-domain sensing. Impact of number of novel classes. Fine-tuning is not applied. The feature extractor $f_\theta^{200}$ is used. (a) Scenario 1, Cross-Environment Evaluation. (b) Scenario 2, Cross-Environment and Cross-User Evaluation.}
\label{fig:NoClasses}
\end{figure}
\subsubsection{Impact of Numbers of Shots} \label{sec:shots}
This section studied the impacts of different numbers of shots, $K$, in the support set. Fine-tuning was not used. We increased $K$ from 1 to 5 for both scenarios 1 and 2.
The results are shown in Fig.~\ref{fig:fsl}. For scenario 1, we used all the remaining 76 classes, i.e., 76-way $K$-shot.
The accuracy increased from $57.9\%$ to $71.5\%$ when $K$ increased from 1 to 5, respectively.
Regarding scenario 2, we tested 25-way $K$-shot.
Increasing the number of shots can enrich data diversity in the support set, thus increasing the generalisation capability of the feature extractor.
Again, the accuracies vary according to different users, about 20\% difference between the best and worse cases, probably due to various gesture patterns among users.
Note that the user s5 had the lowest accuracy compared to the other four users, because the number of novel classes of the user s5 was higher ($N_b$ = 76).
\begin{figure}[!t]
\centering
\includegraphics[width=3.4in]{./FSL.pdf}
\caption{Performance of cross-domain sensing. Impact of number of shots in scenario 1 (user s5) and scenario 2 (user s1-s4). Fine-tuning is not applied. The feature extractor $f_\theta^{200}$ is used.}
\label{fig:fsl}
\end{figure}
\subsubsection{Impact of Fine-Tuning} \label{sec:finetune}
We performed fine-tuning on the support set with only one and five samples for each novel class. We studied 76-way 1-shot and 25-way 1-shot for scenarios 1 and 2, respectively.
As shown in Fig.~\ref{fig:ft}, fine-tuning has improved the accuracies for all the tests. For scenario 1 (user s5), the accuracy is improved from $57.9\%$ to $72.8\%$ and 87.1\%, when one shot and five shots are applied, respectively. For scenario 2 (users s1 to s4), the performance is improved by $8\%$ on average when one shot is applied, and improved by 16.9\% on average when five shots are applied. The evaluation results indicated that our proposed fine-tuning method could significantly improve the cross-domain's accuracy and quickly adapt to the new domain without extensive data collection, because the feature extractor can be updated even using one shot.
\begin{figure}[!t]
\centering
\includegraphics[width=3.4in]{./compareFinetuning.pdf}
\caption{Performance of cross-domain sensing. Impact of fine-tuning in scenario 1 (user s5) and scenario 2 (user s1-s4). The feature extractor $f_\theta^{200}$ is used.}
\label{fig:ft}
\end{figure}
\subsection{Cross-Dataset Evaluation}\label{sec:crossdataset}
This section evaluated the classification performance when the base set and support \& query sets were from different datasets. The same feature extractor $f_\theta^{200}$ trained by the SignFi dataset was used.
The support and query sets were from the Widar or Wiar datasets.
This would be quite challenging. Firstly, there will be a significant domain shift between the datasets, as the users and environments are different.
Secondly, different datasets may have different sensing tasks, which result in totally different CSI variation patterns. For example, sign languages in SignFi are mainly a combination of finger, hand, and head movements, the gestures in the Widar dataset involve hand movements, and the human motions in the Wiar dataset are large-scale human activities that involve arm, hand, limb, and leg, etc.
The tasks of SignFi and Widar are similar, but the ones between SignFi and Wiar are quite different.
\subsubsection{Evaluation on Widar Dataset}\label{sec:widar}
This section evaluated the recognition performance when the tasks of the base sets (SignFi) and the support \& query sets (Widar) were different but similar. Specifically, the gesture data in the Widar dataset was used for evaluation.
As there are multiple receivers available in Widar, we selected data of one receiver from users w1, w2, and w3. We evaluated 6-way $K$-shot, where $K$ = 1 to 5. We studied both direct feature matrix generation and fine-tuned feature matrix generation. The results are shown in Fig.~\ref{fig:widarperform}.
\begin{figure}[!t]
\centering
\includegraphics[width=3.4in]{./widarperform_3user.pdf}
\caption{Cross-dataset evaluation on the Widar dataset. The feature extractor $f_\theta^{200}$ is trained using SignFi dataset.}
\label{fig:widarperform}
\end{figure}
When fine-tuning was used, the overall accuracy increased dramatically.
The minimum accuracy increase was around $16.5\%$, when the data was from the users w1 and w2 in the one-shot learning case.
The overall accuracy increased with an increase in the number of shots.
In the case of one-shot learning with fine-tuning, the accuracies were $51.8\%$, $86.3\%$, and $72.3\%$, for user w1, w2, and w3, respectively. The five shots learning gave the best performance, i.e., $90.6\%$, $99.0\%$ and $100\%$ for user w1, w2, and w3, respectively. The corresponding confusion matrix is shown in Fig.~\ref{fig:cfm}.
The experiment results showed that our proposed method can be scaled to similar sensing tasks, e.g., gesture recognition, with only a few samples.
\begin{figure*}[!t]
\centering
\subfloat[]{\includegraphics[width=2.2in]{./5shots_widar.pdf}\label{fig:widar_userW1}}
\hspace{0.5cm}
\subfloat[]{\includegraphics[width=2.2in]{./5shots_223_finetuned_0.98_user2.pdf}\label{fig:widar_userW2}}
\hspace{0.5cm}
\subfloat[]{\includegraphics[width=2.2in]{./5shots_223_finetuned_1.0_user3.pdf}\label{fig:widar_userW3}}
\centering
\caption{Cross-dataset evaluation on the Widar dataset. The confusion matrix for 5 shots learning. (a) User w1. (b) User w2. (c) User w3.}
\label{fig:cfm}
\end{figure*}
As the feature extractor was trained using the data of sign languages (SignFi), some latent features learned may not apply gesture recognition (Widar).
The performance was thus a little limited when fine-tuning was not used.
When fine-tuning was used, even with very few samples, the recognition accuracy was significantly improved.
This is because the feature extractor has already obtained informative knowledge about how to extract latent features for sign language, and sign languages and gestures share some common features.
\subsubsection{Evaluation on Wiar Dataset}\label{sec:wiar}
This section evaluated the recognition performance when the tasks of the base sets (SignFi) and the support \& query sets (Wiar) are quite different. Specifically, the human activity data in the Wiar dataset was used for evaluation.
We selected the data from six users in the Wiar dataset. We evaluated 16-way $K$-shot, where $K$ = 1 to 5. The experiments were conducted with fine-tuning applied.
The experiment results are shown in Fig.~\ref{fig:wiar}. The feature extractor $f^{200}_\theta$ achieved the highest accuracy for user a1, with an accuracy of $66.2\%$ and $94.2\%$ for one and five shots learning, respectively. The lowest accuracy was shown for user a3, with an accuracy of $31.0\%$ and $57.24\%$ for one and five shots learning, respectively. In the case of five-shot learning, the performance for most of the users (except for user a3) were over 80\% of accuracy.
The results demonstrate that our feature extractor could adapt to various sensing tasks without the burden of high data collection and model retraining, even when the tasks are very different to the original task.
\begin{figure}[t]
\centering
\includegraphics[width = 3.4in]{./wiar_dataset.pdf}
\caption{ Cross-dataset evaluation on the Wiar dataset. The feature extractor $f_\theta^{200}$ is trained using SignFi dataset.}
\label{fig:wiar}
\end{figure}
\subsubsection{Discussion}
The work~\cite{xiao2021onefi} also used FSL for Wi-Fi sensing and argued it is cumbersome to build a feature extractor with good generalisation capability. Hence they designed virtual gesture generation for reducing the effort of collecting data. However, as we showed in this Section, we can leverage the public dataset as the base set and train a versatile feature extractor. When the feature extractor is applied to a different sensing task, we can always fine-tune the feature extractor using very few samples from the new task. Therefore, the overhead for collecting the base set can be mitigated.
\subsection{Collaborative Sensing}\label{sec:collaborative}
In this section, we evaluated the performance of the collaborative sensing algorithm. The fine-tuning was applied in this section.
The same feature extraction, $f_\theta^{200}$, trained by the SignFi dataset was used.
The support and query sets were from the user w1 of the Widar dataset.
Six gestures were performed 20 times and captured by six receivers each time. Therefore, each receiver has 20 samples for each gesture. We constructed the support and query sets for each receiver. One to five samples were randomly selected and added to the support set and the remaining were used as the query set.
As shown in Fig.~\ref{fig:multiRx}, the accuracies increased as the number of receivers increased. For one-shot learning, the accuracy increased from $38.2\%$ to $75.5\%$ when the number of receivers increased from 1 to 6.
Exploiting more receivers enables the feature extractor to view gestures from different perspectives, hence rich features can be obtained for improving system performance.
Similarly, the classification accuracy in collaborative sensing can also be enhanced by more shots.
The highest accuracy achieved by the 5-shot learning is $100.0\%$ when the number of receivers was six.
The work in~\cite{xiao2021onefi} also used multiple receivers. However, the accuracy in their work is already relatively high only using one receiver, hence the accuracies were saturated with more receivers.
However, as demonstrated by our work, more receivers will be beneficial when the accuracy using a single receiver is not saturated.
\begin{figure}[t]
\centering
\includegraphics[width = 3.4in]{./multiRx_resultsuser1.pdf}
\caption{Performance of collaborative sensing using multiple receivers. The feature extractor $f_\theta^{200}$ is trained using the SignFi dataset. Support and query sets are from Widar dataset.}
\label{fig:multiRx}
\end{figure}
\subsection{Complexity}\label{sec:complexity}
The computational overhead involves three parts: training feature extractor, fine-tuning, and inference. The signal pre-processing algorithm used in all three stages took 0.3 seconds.
\textbf{Training feature extractor} probably is the most computationally expensive.
In order to achieve good generalisation capability, the feature extractor needs to be pre-trained with a large number of base classes.
Specifically, we trained the feature extractor using 200 classes each with 20 samples of user s5 lab environment in the SignFi dataset, which took around 10 minutes. Fortunately, the feature extractor only needs to be trained once and can be done offline. The pre-trained feature extractor takes up 45 MB memory, which can be deployed on embedded systems.
\textbf{Fine-tuning} is applied for adapting the pre-trained feature extractor and classifier to new sensing tasks or domains. The time cost depends on the amount of data. In all the experiments, the most time consuming one is 76-way 5-shots learning, which takes 82 seconds. Thus, the overhead in this stage is acceptable.
\textbf{Inference} is completed by the proposed collaborative sensing model when multiple receivers are available.
The time consumption of each inference is around 0.04 seconds, which can be done almost in real-time.
\subsection{Comparison with the-State-of-the-Art}\label{sec:adv}
\subsubsection{Comparison with Domain Adversarial Training}\label{sec:adv}
This section compared the performance of the proposed method to a state-of-the-art domain adversarial training-based method, also know as adversarial domain adaptation, which has been employed in the Wi-Fi sensing area, e.g., \cite{jiang2018towards,kang2021context}.
The basic idea of domain adversarial training is to train a network on data from different domains, which are termed source domains. A domain adversarial network includes a feature extractor, a classifier and a domain discriminator. The feature extractor is used to extract features. The classifier is used to identify the label of the input. The domain discriminator is used to distinguish the domain label of the input. The domain discriminator would force the feature extractor to extract domain-independent features. After training, the network is expected to perform well on target domain data.
To ensure a fair comparison, we used the same modified AlexNet architecture in this paper as the feature extractor of the adversarial network.
We trained the model with the data from two different environments, i.e., lab and lab 2, and five users of the SignFi dataset. Each environment and user pair forms a source domain. The data from the home environment, user s5, was used as the target domain data.
The comparison results are shown in Table~\ref{tab:compareADV}.
The domain adversarial training-based approach only achieved an overall accuracy of $4\%$, which failed to recognise the sign language gestures in the target domain. This is probably because the domain adversarial network requires a large number of domains to enable the feature extractor to learn domain-independent features. For example, the work in~\cite{jiang2018towards} revealed their classification accuracy increased from about 45\% to 75\% when the number of sources domains were increased from 2 to 22.
Instead of training a feature extractor to learn the domain-independent features, our proposed method adapted to the target domain by comparing the similarity of the extracted features between target domain instances. With one sample from each class, our method achieved $57.9\%$ accuracy without fine-tuning. The accuracy can be further increased to $72.8\%$ when fine-tuning is applied.
\begin{table}[]
\centering
\caption{Performance Comparison With Domain Adversarial Training }
\label{tab:compareADV}
\begin{tabular}{|l|l|}
\hline
\textbf{Methods} & \textbf{Accuracy} \\ \hline
\textbf{FewSense without fine-tuning (1-shot)} & 57.9\% \\ \hline
\textbf{FewSense with fine-tuning (1-shot)} & 72.8\% \\ \hline
\textbf{Domain adversarial training} & 4.0\% \\ \hline
\end{tabular}
\end{table}
\subsubsection{Comparison with Domain-Independent Feature-based Approach}
The work in~\cite{widar} proposed BVP, a domain-independent feature, to address the cross-domain sensing problem.
According to~\cite{widar}, extracting BVP requires at least three receivers. The Widar dataset provides BVP extracted from six receivers, which is used here for comparison.
We used the same deep learning model in~\cite{widar}. In order to perform cross-domain sensing, the BVP data of users w4 to w9 collected from the hall and office environments was used for training.
The BVP data of users w1 to w3 obtained from classroom was used for evaluation.
The results are shown in Table~\ref{tab:comparewidar}. The BVP-based approach achieved an average accuracy of 82.0\%, 90.5\% and 92.4\% for the user w1, w2 and w3, respectively.
\begin{table}[]
\centering
\caption{Cross-Domain Performance Comparison With Widar}
\label{tab:comparewidar}
\begin{tabular}{|l|l|l|l|}
\hline
\textbf{Methods} & \textbf{User w1} & \textbf{User w2} & \textbf{User w3} \\ \hline
\textbf{Widar} & 82.0\% & 90.5\% & 92.4\% \\ \hline
\textbf{FewSense (1-shot 6-Rx)} & 76.0\% & 94.3\% & 90.2\% \\ \hline
\textbf{FewSense (2-shot 6-Rx)} & 92.0\% & 98.3\% & 100\% \\ \hline
\textbf{FewSense (5-shot 2-Rx)} & 87.3\% & 91.0\% & 95.8\% \\ \hline
\end{tabular}
\end{table}
In comparison, we also evaluated the performance of the FewSense method. The CSI data from the same three users collected in the classroom environment via six receivers is used as the support and query set.
As can be observed from Table~\ref{tab:comparewidar}, one-shot learning can achieve comparable performance with the BVP-based method. When two shots are used, our method outperforms the BVP.
However, the applications of BVP feature may be limited in real life since it requires at least three receivers to resolve the ambiguity problem.
In comparison, our method provides an alternative solution when the number of receivers is limited. For example, by using five shots in the two receivers setting, our collaborative sensing model can still achieve comparable performance with the BVP features at the six receivers setting.
In terms of complexity, the BVP feature extraction is computationally expensive. For example, it took around 205 seconds to compute the BVP feature for one gesture from six receivers. The high computational complexity limits its practical use.
On the other hand, our model is lightweight, as analysed in Section~\ref{sec:complexity}.
\section{Related work} \label{relate}
\subsection{Few-Shot Learning}
FSL has been successfully implemented in many vision-based tasks~\cite{wang2020generalizing} and the similarity-based methods are widely investigated.
Matching networks \cite{vinyals2016matching} is proposed to use a cosine similarity function to solve the FSL problem.
The authors in \cite{chen2019closer} train a neural network with a large number of base classes and then replace the last fine-tuning layer with a cosine similarity function.
Researchers in \cite{snell2017prototypical} propose a prototypical network (PN) for FSL, which employs Euclidean distance as the distance metric.
The works above aim to learn transferable features and a fixed similarity function, the generalisation capability of them is limited. The work in~\cite{sung2018learning} uses a relation network with a trainable similarity metric to address the above limitation.
\subsection{Domain Robustness of Wi-Fi Sensing}
The authors in \cite{jiang2018towards,kang2021context} use an domain adversarial training-based domain adaptation method to extract domain-independent features from multiple domains.
For example, the work in~\cite{kang2021context} employs a gesture classifier to classify gestures, and a domain discriminator to recognise the domain label of the input samples. The feature extractor is trained to cheat the domain discriminator such that the domain discriminator cannot distinguish the domain labels of the input. The feature extractor is expected to map the input from different domains to the same feature space. However, this method requires a massive number of training samples from different domains to obtain satisfactory performance.
The transfer learning focuses on leveraging the knowledge of a pre-trained model and applying the model to a different but similar task. In order to reduce the model retraining effort in the target domain, the authors in \cite{CrossSense} propose a transfer learning-based method called CrossSense, which employs an ANN-based model to translate the CSI features of gestures in the source domain to the target domain.
Although transfer learning can reduce the data collection effort in a new domain, it still requires many data samples from the target domain to achieve a satisfying performance. Moreover, transfer learning only works when the source and target domains are similar enough. Otherwise, the problem of negative transfer may limit the applications of this method~\cite{weiss2016survey}.
The domain-independent feature-based method is another type of solution. Widar~\cite{widar} uses multiple receivers to extract BVP. However, at least three receivers are required to resolve the velocity direction ambiguity problem, and the feature extraction overhead cannot be ignored. Those factors may limit the application in real deployments.
The work in~\cite{li2020wihf} extracts the motion patterns as the input features which are experimentally demonstrated domain-independent. However, some similar gestures could have similar movement patterns. Therefore, the classes of the gestures need to be designed to avoid similarity.
\section{Conclusions}\label{conclusion}
This paper proposed FSL-based WiFi sensing, i.e., FewSense, to address the scalability and domain-dependent challenges. FewSense utilised a revised AlexNet architecture as the feature extractor to gain generalisation capability.
The features of query and support sets were compared and the inference was made based on their cosine similarity score.
Collaborative sensing was designed to fuse observations from multiple receivers for boosting classification accuracy.
Three public Wi-Fi sensing datasets, including SignFi (sign language), Widar (gesture recognition) and Wiar (human activity recognition), were leveraged for evaluation. We carried out extensive evaluation using a feature extractor trained by the SignFi dataset.
The experimental results indicated that FewSense could be adapted to new sensing tasks or domains with a low data collection and computational cost.
The FewSense with one-shot learning can recognise novel sign language gestures in SignFi with an average accuracy of 99.2\% and 84.2\% in in-domain and cross-domain scenarios, respectively. When applying FewSense to new sensing tasks, FewSense recognised novel gestures on the Widar dataset with an average accuracy of 69.9\% and 96.5\% for one-shot and five-shot learning, respectively.
It also achieved an average accuracy of 52.8\% and 82.7\% for one-shot and five-shot learning, respectively, for classifying novel human activities on the Wiar dataset.
Finally, our collaborative sensing approach can boost the classification accuracy by 30\% on average when there were six receivers. In summary, FewSense demonstrated that cross-dataset sensing is applicable. The generalisability of the feature extractor can be achieved using a publicly available dataset, alleviating the data collection overhead.
\bibliographystyle{IEEEtran}
| 27c01fde8032a640aed41e586dd056d93b86a446 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the AdS/CFT correspondence, where a precise non-perturbative definition of quantum gravity exists \cite{Maldacena:1997re,Witten:1998qj}, one of the most basic objects is the bulk-to-boundary map $V$ in the large $N$ and large t'Hooft coupling limit \cite{Balasubramanian:1998sn,Banks:1998dd,Balasubramanian:1999ri,Hamilton:2006az}. In this limit, the bulk description is well approximated by semi-classical supergravity, and the bulk-to-boundary map $V$ embeds bulk semi-classical states into the Hilbert space $\mathcal{H}_\text{CFT}$ of the dual, strongly coupled, quantum mechanical theory. The bulk semi-classical states in question are essentially comprised of effective field theory (EFT) excitations of bounded energy density around some chosen gravitational background $g$, which we can arrange in the form of a bulk EFT Hilbert space $\mathcal{H}_\text{bulk EFT}^g$. Two decades of work have explored aspects of this embedding of the bulk gravity EFT in the boundary CFT. One of the most mysterious aspects of the map is the emergence of locality in the bulk low-energy theory, especially in regions causally hidden behind horizons.
Some fundamental properties of the map $V$, such as the redundancy of bulk operator representations and radial locality, are most naturally understood by realizing that the embedding
\begin{equation}
V: \mathcal{H}_{\text{bulk EFT}}^g \to \mathcal{H}_\text{CFT} ,
\end{equation}
has the structure of a quantum error correcting code \cite{Almheiri:2014lwa,Pastawski:2015qua}. In the quantum error correction (QEC) language, states in the bulk effective field theory around the gravitational background $g$ are often referred to as the code subspace $\mathcal{H}_\text{code} \subset \mathcal{H}_\text{CFT}$ \cite{Harlow:2016vwg}.
The code subspace is intuitively the subspace of states in the full Hilbert space $\mathcal{H}_\text{CFT}$ upon which a bulk local EFT operator may be represented, and states outside the code subspace are interpreted as having lost access to the operator either due to absorption by a large black hole or a more fundamental breakdown of the emergent background metric $g$ \cite{Harlow:2018fse}. Quantum error correction in this holographic context means that the representation on a boundary subregion $A$ of bulk operators within a corresponding bulk subregion enclosed between $A$ and the quantum extremal surface (QES) for $A$ -- called the ``entanglement wedge" $\mathcal{W}(A)$ -- is robust against the erasure of the complementary subregion ${\overline{A}}$ \cite{Dong:2016eik}. Equivalently, the state of bulk quantum fields in the region $\mathcal{W}(A)$ is ``encoded'' in the boundary subregion $A$, and is protected from any and all ``errors'', i.e., operator actions, on ${\overline{A}}$. Conversely operators in the entanglement wedge $\mathcal{W}(A)$ can be reconstructed entirely from data in $A$. This modernization of the bulk-to-boundary map is referred to as entanglement wedge reconstruction (EWR).
\begin{figure}[t]
\centering
\includegraphics[height=5cm]{evapBH.png}
\caption{\small{
A cartoon of an evaporating black hole. \textbf{(Left)} The UV description consists of a holographic quantum mechanical system $B$ (green dot) in a highly excited state coupled to a bath $R$ (blue line). \textbf{(Right)} The dual semi-classical description is that of an evaporating black hole coupled to a bath. Past the Page time, there is a new Quantum Extremal Surface (black dot) relevant for the entropy of the bath, and the region in the black hole interior enclosed by it to its left is called the ``island''.
}
}
\label{fig:SC}
\end{figure}
The EWR paradigm has provided a powerful framework for bringing tools from quantum information theory to bear on problems in quantum gravity. For instance, EWR played an important role in recent progress on the black hole information problem \cite{Almheiri:2019psf,Penington:2019npb,Almheiri:2019qdq,Penington:2019kki,Chen:2019iro}. These papers studied an asymptotically AdS$_2$ evaporating black hole coupled to a non-gravitational bath, where the bath serves as a reservoir to collect the Hawking radiation (Fig.~\ref{fig:SC}). In this context, if one computes the von Neumann entropy of the bath past the Page time, the semi-classical description involves a new quantum extremal surface that develops in the black hole interior and cuts off the naive growth of the entropy, reproducing the expected Page curve. Crucially, the EWR paradigm in this context implies that the region in the black hole interior enclosed by this new QES -- the so called ``island'' -- is encoded in the bath, and thus bulk operators in the island can be represented as operators on the bath. Indeed, it was shown in \cite{Penington:2019kki} that a general tool from the theory of quantum error correction called the Petz map can be used to give an explicit reconstruction of operators in the black hole interior.
This is a surprising and radical claim -- the bath need not be a conventional, strongly coupled holographic quantum field theory, and could be a free field theory, for instance. Yet, it can encode semi-classical supergravity degrees of freedom in the same sense as holographic duality, i.e., a portion of the bath is dual to, or an equivalent description of, the gravity degrees of freedom in a distant part of space. In an evaporating black hole, this would appear to be at odds with the semi-classical causal structure of the black hole spacetime, as the encoded island lies behind the horizon. However, this encoding should be fundamentally different from the standard bulk-boundary map in holographic dualities: we expect it to be exponentially complex in the black hole entropy, i.e, implementing this encoding with a quantum computer should be exponentially complex in $N$, explaining why simple observables give the appearance of information loss in Hawking radiation. By contrast, the standard encodings of EFT excitations in regions of AdS space that are causally accessible from the boundary only have polynomial complexity since the reconstruction just involves solving equations of motion \cite{Balasubramanian:1998sn,Balasubramanian:1999ri,Hamilton:2006az}. Note that for entanglement wedges of a strongly coupled theory which include a ``Python's Lunch'', which we will discuss below, the encoding map is also expected to be highly complex since the main source of reconstructive power is entanglement between gravitational degrees of freedom.
Recently, progress was made toward a more detailed understanding of the quantum error correcting structure of the code in the Hawking radiation for the interior of the black hole \cite{Kim:2020cds}. In this paper, Kim, Preskill and Tang (KPT) suggested that the bulk degrees of freedom in the black hole interior are encoded in the radiation with robust QEC properties -- more so than the usual QEC familiar in encoding of ``causal wedges'', i.e., regions causally connected to the spacetime boundary in AdS/CFT. In particular, they found (in a model for black hole evaporation) that despite the fact that the ``island'' region in the black hole interior is encoded in the early Hawking radiation, this encoding is protected in the QEC sense from all ``low-complexity'' and ``low-rank'' errors on the early radiation as long as the remaining black hole is macroscopic. Here an ``error'' is any quantum operation that acts on the radiation. This is very different from the standard statement of QEC in AdS/CFT where the bulk entanglement wedge $\mathcal{W}(A)$ is protected from errors in ${\overline{A}}$, but not in general from errors on $A$ itself. KPT further argued that this robust error correction has an important bearing on the apparent tension between islands and the semi-classical causal structure of the black hole -- all low complexity quantum operations on the bath commute with the representations in the bath of bulk operators in the black hole interior. Thus, in order to affect the black hole interior explicitly, an observer in the bath will have to perform an exponentially complex operation. Crucial to the analysis of KPT was an assumption that the reduced state on the bath/radiation is \emph{pseudorandom}, i.e., cannot be reliably distinguished from the maximally mixed state (more generally, the thermal ensemble) with a polynomial complexity quantum operation; complexity theory enters their arguments through this assumption.
The main purpose of our work is to study these novel quantum error correction properties of the encoding of the black hole interior in Hawking radiation. Instead of using the KPT pseudorandomness assumption, we will directly study the QEC properties in a toy model for an evaporating black hole in Jackiw-Teitelboim (JT) gravity using the Euclidean gravity path integral. We will show that in this setting the bulk density matrix in the interior is approximately recoverable from the action of quantum operations on the bath which (i) do not have prior access to the details of the black hole microstates and their overlaps, and (ii) do not have a large, negative ``coherent information'' -- an information theoretic measure of the amount of noise a quantum operation generates -- with respect to the maximally mixed state on the bath (which we may interpret as the ``semi-classical state" of the radiation, i.e., the analog of Hawking's thermal state). The lower bound on the coherent information is controlled by the black hole entropy and code subspace dimension. This implies that the encoding of the black hole interior in the radiation is robust against generic, low-rank quantum operations on the bath which do not have access to details of the black hole microstate, thus providing a gravitational perspective on the results of KPT. Furthermore, for erasure errors on the bath, we show that gravity comes within an $O(1)$ distance of saturating the quantum Singleton bound on the error tolerance of quantum error correcting codes. We also show that typical errors on the bath must have a large rank compared to the product of the black hole Hilbert space and bath Hilbert space dimensions in order to affect the interior.
We comment on the interpretation of our coherent information criterion in light of an unusual non-isometry property of the interior encoding map that we describe.
While our analysis here is limited to the specific example of the black hole interior, we propose that even for a more general entanglement wedge $W(A)$ dual to a CFT subregion $A$, \emph{the encoding of semi-classical degrees of freedom in the python's lunch} \cite{Brown:2019rox, Engelhardt:2021mue, Engelhardt:2021qjs} \emph{(i.e., the part hidden behind a non-minimal QES) should be robust against generic, low-rank errors on the boundary subregion $A$.} By ``generic'', we mean quantum operations which do not have prior access to the details of the encoding map. Note that such robust error correction properties certainly do not hold in the causal wedge. Indeed, after the Page time the black hole interior is in the python's lunch because it is hidden behind a non-minimal QES, namely the empty surface. Finally, we discuss implications of our results for the consistency of the island program for black hole information recovery \cite{Almheiri:2019psf,Penington:2019npb,Almheiri:2019qdq,Penington:2019kki,Chen:2019iro} with the semi-classical causal structure of evaporating black hole spacetimes.
Four sections follow.
In Sec.~\ref{sec:prelim}, we review necessary ideas from quantum error correction as well as the toy model of black hole evaporation which serves as our primary example.
In Sec.~\ref{sec:qec-pssy}, we apply QEC tools to the toy model and, using the gravitational path integral, derive our criterion for correctable errors on the black hole interior code subspace. Euclidean wormholes play a key role in the action of errors on the encoded state which are not correctable. We also discuss the non-isometric nature of the interior encoding. In Sec.~\ref{sec:ErasuresTypical} we present a detailed analysis of erasure errors and typical/random, fixed-rank errors.
We conclude with a discussion of generalization to other entanglement wedges and implications for the semi-classical causal structure of the evaporating black hole spacetime in Sec.~\ref{sec:disc}.
\section{Preliminaries}\label{sec:prelim}
In this section, we review background material on quantum error correction, following \cite{NielsenChuang}, and the toy model of black hole evaporation \cite{Penington:2019kki} which we will use in later sections. The material in this section is not new, but is included here to keep the article somewhat self-contained.
\subsection{Conditions for quantum error correction}\label{sec:QECreview}
Let $\mathcal{H}_{\text{code}}$ denote the code subspace, and let $\{|i\rangle\}_{i=1}^d$ be a basis of states for $\mathcal{H}_{\text{code}}$. Furthermore, let $\mathcal{H}_{\text{phys}}$ be a larger Hilbert space in which we wish to encode $\mathcal{H}_{\text{code}}$, via the isometric encoding map:
\begin{equation}
V: \mathcal{H}_{\text{code}} \to \mathcal{H}_{\text{phys}} , \quad V\ket{i}_\text{code} = \ket{\Psi_i}_\text{phys},\quad V^{\dagger}V=\mathbb{1}_\text{code}.
\end{equation}
Thus $VV^\dagger$ projects the physical Hilbert onto encoding of the code subspace in the physical degrees of freedom. The general theory of quantum error correction deals with recovering the encoded state from the physical system after the action of quantum operations, which can potentially create ``errors'' in the encoding. Errors in the physical system can be modelled by the action of a quantum channel $\mathcal{E}$ which is a linear, completely positive\footnote{All operations of the form $\mathcal{E}(\rho) = \sum_m E_m \rho E_m^\dagger$ are completely positive.
Complete positivity of $\mathcal{E}$ means $(\mathbb{1}_E \otimes \mathcal{E})(A) \geq 0$ for any positive $A \geq 0$ which acts on the extended system $\mathcal{H}_E \otimes \mathcal{H}_\text{phys}$.
For any $\ket{\psi}_{E,\text{phys}}$, we define $\ket{\phi_m}_{E,\text{phys}} = (\mathbb{1}_E \otimes E_m^\dagger) \ket{\psi}_{E,\text{phys}}$.
By $A \geq 0$ we have $\bra{\phi_m} A \ket{\phi_m} \geq 0$.
Summing over $m$, we find the action of $\mathbb{1}_E \otimes \mathcal{E}$ on $A$:
\[ \sum_m \bra{\phi_m} A \ket{\phi_m} = \sum_m \bra{\psi} (\mathbb{1}_E \otimes E_m) A (\mathbb{1}_E \otimes E_m^\dagger) \ket{\psi} \geq 0 . \]
The middle expression is nothing but $\bra{\psi} (\mathbb{1}_E \otimes \mathcal{E})(A) \ket{\psi}$, and we have shown it is non-negative for any $\ket{\psi}$, so we have $(\mathbb{1}_E \otimes \mathcal{E})(A) \geq 0$. These conditions guarantee in our setting that density matrices on the system plus the environment are mapped to density matrices.
}
and trace preserving map which takes a state (i.e., a density matrix\footnote{In our notation, we will refer to both pure states $\ket{\psi}$ and density matrices $\rho$ as ``states". }) on $\mathcal{H}_{\text{phys}}$ to another state on $\mathcal{H}_{\text{phys}}$.
The action of a general quantum channel can always be expressed in terms of its Kraus operators $\{E_m\}$:
\begin{equation}
\mathcal{E}(\rho) = \sum_m E_m \rho E_m^{\dagger},
\end{equation}
where requiring $\mathcal{E}$ to be trace preserving implies
\begin{equation} \label{trC}
\sum_m E_m^{\dagger} E_m = \mathbb{1}.
\end{equation}
The minimum number of Kraus operators necessary to implement $\mathcal{E}$ is sometimes called the (Kraus) rank of the channel, and could be regarded as measure of how complex a channel is.\footnote{Obviously, the rank is not related to the circuit complexity of the channel since a channel of rank 1 has a single unitary Kraus operator which can have arbitrary circuit complexity.}
Another particularly useful fact is that the action of any quantum channel $\mathcal{E}$ on a state $\rho$ can be modeled by a unitary operation on an enlarged Hilbert space that includes an auxiliary environment. In detail, introduce an environment $\mathcal{H}_{\text{env}}$ spanned by $\{|e_m\rangle\}$ in some initial state $|e_0\rangle$. Then act on the enlarged state $\rho\otimes |e_0\rangle\langle e_0| \in \mathcal{H}_{\text{phys}}\otimes \mathcal{H}_{\text{env}}$ with a unitary operator $U_\mathcal{E}$, and trace out the environment to get
\begin{equation}\label{SD}
\mathcal{E}(\rho) = \mathrm{Tr}_{\text{env}}\left[ U_\mathcal{E}( \rho \otimes |e_0\rangle\langle e_0|_{\text{env}} ) U_\mathcal{E}^{\dagger} \right],
\end{equation}
where $E_m = \langle e_m | U_\mathcal{E} | e_0\rangle$, i.e., $U_\mathcal{E}|\psi\rangle\otimes |e_0\rangle = \sum_{m}E_m |\psi\rangle \otimes |e_m\rangle$.
This expression of the channel $\mathcal{E}$ as the trace over an environment after conjugation by a unitary $U_\mathcal{E}$ will be referred to as the \emph{isometric extension} (sometimes also called the Stinespring dilation) of $\mathcal{E}$. It is a way of representing a quantum channel as an open system.
We say that an error $\mathcal{E}$ can be corrected on $\mathcal{H}_{\text{code}}$ if there exists a recovery channel $\mathcal{R}$ such that
\begin{equation}
\mathcal{R}(\mathcal{E}(\rho)) = \rho,
\end{equation}
for any density matrix $\rho$ supported on $\mathcal{H}_{\text{code}}$, or more precisely, supported on the image of the code subspace inside the physical Hilbert space under $V$. One of the central results of the theory of quantum error correction is that the channel $\mathcal{E}$ is exactly correctable if and only if it satisfies the ``Knill-Laflamme'' conditions \cite{NielsenChuang}:
\begin{equation} \label{KLcond0}
P_{\text{code}} E_m^{\dagger}E_n P_{\text{code}} = \alpha_{mn}P_{\text{code}},
\end{equation}
where $P_{\text{code}} = VV^\dagger$ is the projector onto the code subspace.
The projection, trace preservation, and positivity properties of the operators on the left hand side of \eqref{KLcond0} imply that $\alpha$ is actually a density matrix on $\mathcal{H}_\text{env}$.
It is useful to restate this criterion in a different form. We introduce a reference system with a Hilbert space $\mathcal{H}_\text{ref}$ isomorphic to the code Hilbert space and construct the state
\begin{eqnarray}
|\Psi'\rangle &\equiv& \frac{1}{\sqrt{d}}\sum_{i}|i\rangle_{\text{ref}}\otimes U_\mathcal{E}(|\Psi_i\rangle_{\text{phys}}\otimes |e_0\rangle_{\text{env}}) \nonumber\\
&\equiv& \frac{1}{\sqrt{d}}\sum_{i,m}|i\rangle_{\text{ref}}\otimes (E_m|\Psi_i\rangle_{\text{phys}})\otimes |e_m\rangle_{\text{env}} \, ,
\end{eqnarray}
where $|\Psi_i\rangle_\text{phys}=V|i\rangle_{\text{code}}$ are the basis states of the code subspace of the physical Hilbert space.
The state $\ket{\Psi'}$ can be thought of as the action of the error channel on a state that is maximally entangled between the code subspace and an auxiliary reference system, and is engineered so that its entanglement properties are related to the validity of \eqref{KLcond0}.
Specifically, the error $\mathcal{E}$ is correctable if and only if the reduced state of $\ket{\Psi'}$ on $\mathcal{H}_\text{ref} \otimes \mathcal{H}_\text{env}$ factorizes as \cite{Schumacher:1996dy}
\begin{equation}
\rho'_{\text{ref},\text{env}} = \rho'_{\text{ref}}\otimes \rho'_{\text{env}} ,
\end{equation}
where the reduced density matrices above are defined as $\rho'_{\text{ref},\text{env}} = \Tr_\text{phys} \ket{\Psi'}\bra{\Psi'}$ and so on.
This is called the \emph{decoupling principle} \cite{Preskill:2016htv}.
We can diagnose decoupling by computing the mutual information:
\begin{equation}
I_{\Psi'}(\text{ref}:\text{env}) \equiv S_{\Psi'}(\text{ref})+S_{\Psi'}(\text{env})-S_{\Psi'}(\text{ref}\cup \text{env}),
\end{equation}
where each term on right hand side is the entropy of the reduced density matrix on the indicated factor. This mutual information vanishes if and only if the decoupling principle holds. The decoupling principle also generalizes to approximate quantum error correction \cite{2001quant.ph.12106S}. For instance, we will often be interested in cases where the above mutual information is not strictly zero, but small (i.e., exponentially small in the black hole entropy). In such a situation, the mutual information gives a bound on the decoupling:
\begin{equation}
||\rho'_{\text{ref},\text{env}}-\rho'_{\text{ref}}\otimes \rho'_{\text{env}}||_1 \leq \sqrt{2 I_{\Psi'}(\text{ref}:\text{env})},
\end{equation}
where $||A ||_1 = \mathrm{Tr}\,\sqrt{A^{\dagger}A}$ is the trace norm. This in turn implies that the error channel is approximately correctable \cite{Kim:2020cds}, i.e., there exists a recovery channel $\mathcal{R}$ such that
\begin{equation}
\max_{\rho}||\mathcal{R}(\mathcal{E}(\rho))-\rho||_1 \leq \epsilon,\;\;\;\epsilon = \left(\frac{1}{2}I_{\Psi'}(\text{ref}:\text{env})\right)^{1/4},
\label{eq:mutual-bound}
\end{equation}
where the maximization is over all states supported on the code subspace.
Using the decoupling principle, we can derive a simple bound on the number of maximal errors, i.e., erasures, a code can tolerate while remaining correctable. Let $t-1$ be the maximum number of qubits in the physical Hilbert space on which we can tolerate errors. Let $d$ be the dimension of the code space and let $2^{\mathcal{S}}$ be the dimension of the physical Hilbert space. Partition the physical Hilbert space into three subsets: $Q_1$ and $Q_2$ with $(t-1)$ qubits each and $Q_3$ with the remaining $(\mathcal{S}-2t+2)$ qubits. Now consider the state
\begin{equation}
|\chi\rangle = \frac{1}{\sqrt{d}}\sum_i |i\rangle_{\text{ref}}\otimes |\Psi_i\rangle_{\text{phys}}.
\end{equation}
Since both $Q_1$ and $Q_2$ have less than $t$ qubits each, we can, by definition, tolerate errors on these subsystems. For instance, we could bring in $(t-1)$ environment qubits and swap the state on $Q_1$ or $Q_2$ with the environment. For such operations to be correctable, by the decoupling principle there must be no correlations after the swap between the environment and the reference. This in turn means that the initial state cannot have correlations between the qubits in $Q_1$ and the reference, and similarly for $Q_2$. Using this fact, we get
\begin{equation}
S(\text{ref}) + S(Q_1) = S(\text{ref},Q_1) = S(Q_2Q_3) \leq S(Q_2) + S(Q_3),
\end{equation}
\begin{equation}
S(\text{ref}) + S(Q_2) = S(\text{ref},Q_2) = S(Q_1Q_3) \leq S(Q_1) + S(Q_3),
\end{equation}
where all the entropies refer to the state $|\chi\rangle$, and we have used the subadditivity of entropy in the last step of each inequality. The second equality follows because the entropy of a subsystem of a pure state equals the entropy of its complement. Adding these inequalities gives
\begin{equation}
S(\text{ref})\leq S(Q_3)\;\;\;\Rightarrow 2(t-1) \leq (\mathcal{S}-\log_2d).
\label{eq:singleton-bound}
\end{equation}
where we used the fact that the reference state is maximally entangled with the code subspace and so $S(\text{ref}) = \log_2{d}$ while the $S(Q_3)$ is bounded from above by the number of qubits in $Q_3$, namely $(\mathcal{S}-2t+2)$. This last inequality is known as the \emph{quantum Singleton bound.}
\noindent\textbf{Subsystem QEC condition}: We will be interested in a setting where the code subspace has a natural tensor factorization $\mathcal{H}_{\text{code}}=\mathcal{H}_a\otimes \mathcal{H}_{a'}$. In the black hole case this will correspond to factorization of some subset of the bulk quantum field theory degrees of freedom between the interior and the exterior of the black hole. With this additional structure, we can define a more general version of error correction \cite{2005quant.ph..4189K, Nielsen_2007}, which, following \cite{Harlow:2016vwg}, we will refer to as ``subsystem quantum error correction''. We say that the state on $\mathcal{H}_{a}$ is recoverable under the error $\mathcal{E}$ if, for any density matrix $\rho$ supported on $\mathcal{H}_{a}$ and any density matrix $\sigma$ supported on $\mathcal{H}_{a'}$, there exists a recovery channel $\mathcal{R}$ such that
\begin{equation}
\mathcal{R}(\mathcal{E}(\rho \otimes \sigma)) = \rho \otimes \sigma',
\end{equation}
for some density matrix $\sigma'$ on $\mathcal{H}_{a'}$. We again have in mind that the density matrix on the left hand side has been encoded in $\mathcal{H}_\text{phys}$ by the map $V$ before the action of $\mathcal{E}$. In analogy with equation \eqref{KLcond0} the channel $\mathcal{E}$ is correctable on $\mathcal{H}_a$ if and only if
\begin{equation} \label{KLcond1}
P_{\text{code}} E_m^{\dagger}E_n P_{\text{code}} = \mathbb{1}^{(a)}\otimes B^{(a')}_{mn},
\end{equation}
where $P_{\text{code}}$ is the projector onto the code subspace, and $B_{mn}$ are some operators with support only on $\mathcal{H}_{a'}$ \cite{2005quant.ph..4189K, Nielsen_2007} .
Following \cite{Nielsen_2007}, it is useful to restate this criterion in the form of a decoupling principle. In order to do this, we once again need to introduce a reference system. Let us in fact introduce two reference systems $\mathfrak{R}_a$ and $\mathfrak{R}_{a'}$ which have the same dimensions as $\mathcal{H}_{a}$ and $\mathcal{H}_{a'}$ respectively. For any basis state $|i, i'\rangle_{\text{code}} \equiv |i\rangle_a \otimes |i'\rangle_{a'}$, let $|\Psi_{i,i'}\rangle_{\text{phys}}$ be the image of $\ket{i,i'}_\text{code}$ in $\mathcal{H}_{\text{phys}}$ under the encoding map $V$:
\begin{equation}
V|i,i'\rangle_{\text{code}} = |\Psi_{i,i'}\rangle_{\text{phys}} .
\end{equation}
We now introduce a basis for the reference system $\ket{i,i'}_\text{ref} = \ket{i}_{\mathfrak{R}_a} \otimes \ket{i'}_{\mathfrak{R}_{a'}}$, and an environment $\mathcal{H}_{\text{env}}$ supporting isometric extension of the quantum channel $\mathcal{E}$. In terms of these quantities we define
\begin{align}
\label{psi'}
\ket{\Psi'} & \equiv \frac{1}{\sqrt{d}}\sum_{i,i'} \ket{i,i'}_\text{ref} \otimes U_{\mathcal{E}} (\ket{\Psi_{i,i'}}_\text{phys} \otimes \ket{e_0}_\text{env} ) \\
& = \frac{1}{\sqrt{d}} \sum_{i,i',m} \ket{i,i'}_\text{ref} \otimes E_m \ket{\Psi_{i,i'}}_\text{phys} \otimes \ket{e_m}_\text{env} .
\end{align}
We are now in a position to restate the condition for subsystem quantum error correction: an error $\mathcal{E}$ is correctable on $\mathcal{H}_a$ if and only if the reduced state $\rho'$ corresponding to $|\Psi'\rangle$ on the subsystem $\mathcal{H}_\text{ref} \otimes \mathcal{H}_\text{env}$ factorizes as
\begin{equation}
\rho'_{\mathfrak{R}_a,\mathfrak{R}_{a'},\text{env}} = \rho'_{\mathfrak{R}_a} \otimes \rho'_{\mathfrak{R}_{a'},\, \text{env}} \, .
\end{equation}
Equivalently,
\begin{equation} \label{SQEC}
I_{\Psi'}(\mathfrak{R}_a: \mathfrak{R}_{a'}\cup \text{env}) \equiv S_{\Psi'}(\mathfrak{R}_a) + S_{\Psi'}(\mathfrak{R}_{a'}\cup \text{env}) - S_{\Psi'}(\mathfrak{R}_a \cup \mathfrak{R}_{a'} \cup\text{env}) = 0.
\end{equation}
It is particularly useful to restate the QEC conditions in the form of a vanishing mutual information because it distills down the criterion to the computation of specific entanglement entropies. In AdS/CFT, the gravitational path integral can compute entanglement entropies via the replica trick. Thus, we can hope to get a handle on the error correction properties of subsystems in gravity by studying the above mutual information via the gravitational path integral. Indeed, for the black hole interior, the only way we can currently access the bulk-to-boundary encoding map $V$ is indirectly via the Euclidean path integral. In Sec.~\ref{sec:qec-pssy}, our primary goal is to compute this mutual information with $\mathfrak{R}_a$ and $\mathfrak{R}_{a'}$ being the reference systems for the interior and exterior code subspaces on an evaporating black hole background after the Page time.
\subsection{PSSY model}
In this section, we review a toy model for an evaporating black hole that was constructed by Penington, Shenker, Stanford and Yang (PSSY) \cite{Penington:2019kki}. We will use this model below. Consider an evaporating black hole in Jackiw-Teitelboim (JT) gravity with an end of the world (EOW) brane deep inside the black hole capping off the geometry, plus some propagating bulk degrees of freedom which constitute the code subspace. The Euclidean action is given by:
\begin{equation}
I=I_{\mathrm{JT}}+\mu \int_{\text {brane }} \mathrm{d} s + I_{\text{field theory}},
\end{equation}
where $I_{\text{field theory}}$ is a bulk effective field theory action and the JT gravity action is
\begin{equation}
I_{\mathrm{JT}}=-\frac{S_{0}}{2 \pi}\left[\frac{1}{2} \int_{\mathcal{M}} \sqrt{g} R+\int_{\partial \mathcal{M}} \sqrt{h} K\right]-\left[\frac{1}{2} \int_{\mathcal{M}} \sqrt{g} \phi(R+2)+\int_{\partial \mathcal{M}} \sqrt{h} \phi (K-1)\right],
\label{eq:jt-action}
\end{equation}
with the boundary conditions:
\begin{equation}
\left.\mathrm{d} s^{2}\right|_{\partial \mathcal{M}}=\frac{1}{\epsilon^{2}} \mathrm{~d} \tau^{2}, \quad \left.\phi\right|_{\partial\mathcal{M}}=\frac{1}{\epsilon}, \quad \epsilon \rightarrow 0.
\end{equation}
The boundary condition on the the EOW brane is given by
\begin{equation}
\partial_{n} \phi=\mu\geq 0, \quad K=0.
\end{equation}
Here $S_0$ can be thought of as the entropy of the higher-dimensional extremal black hole for which the JT theory is the two-dimensional reduction, or it can be thought of as the ground state entropy of the JT system.
The EOW brane hosts semi-classical, intrinsic degrees of freedom, which we label by Greek letters, $\alpha,\beta$ etc., running from 1 to $k$. These intrinsic EOW brane states are taken to be orthogonal in the bulk semi-classical description i.e., these indices along the brane are contracted with the propagator $\delta_{\alpha\beta}$.
We think of these states as arising from some sort of theory living on the brane, or alternatively as a flavor index for the brane which must be conserved along the worldline to ensure orthogonality.
The JT gravity black hole describes the semi-classical bulk dual to an excited state in a holographic quantum mechanical system, which we label $B$. More precisely, JT gravity describes an ensemble average over the Hamiltonian of the dual quantum mechanical system $B$ \cite{Saad:2019lba}; in this context, the EOW brane states correspond to effectively random superpositions of energy eigenstates,\footnote{These superpositions also depend in a controlled way on the brane tension $\mu$ and inverse temperature set by Euclidean boundary length, which renders them not strictly random \cite{Penington:2019kki}.} with the coefficients chosen from a Gaussian ensemble with zero mean and unit variance \cite{Penington:2019kki}.
In order to model an evaporating scenario, we entangle this system ($B$) with a quantum mechanical radiation system ($R$):
\begin{equation}
\label{eq: global states}
\left|\Psi_i\right\rangle= \frac{1}{\sqrt{k}} \sum_{\alpha=1}^{k}\left|\psi^\alpha_i\right\rangle_B \otimes |\alpha \rangle_R,
\end{equation}
where $\left|\psi^\alpha_i\right\rangle_B$ is a state in the quantum system $B$ whose semi-classical bulk dual consists of the black hole, including an EOW brane with its intrinsic degrees of freedom in the state $\alpha$, and the bulk code subspace fields in the code basis state $i$ ($i=1,...,d$). Note that the entanglement is mediated by the EOW brane degrees of freedom, and not by the code subspace states. We will choose the code subspace to consist of simple, qubit-like excitations in the effective field theory, as will be explained below. The strategy is to treat this one-parameter family (labelled by $k$) of entangled states as snapshots of an AdS black hole evaporating into a reservoir/bath, with $k$ playing the role of time. More accurately, rather than considering an evaporating black hole {\it per se}, we maximally entangle a black hole with a fixed number of degrees of freedom with an ever larger number of radiation states and study the consequences. It was shown in \cite{Penington:2019kki} that as the parameter $k$ becomes large, there is a phase transition in the entanglement entropy $S(R)$ at $k \sim e^{S_0}$, where $S_0$ is the extremal black hole entropy. The phase transition cuts off the naive growth of entropy and realizes the expected Page curve. The mechanism by which this happens in the replica trick computation of entropy involves the appearance of a new gravitational saddle, i.e., the replica wormhole, which makes the dominant contribution past the Page point \cite{Penington:2019kki}.
\section{Quantum error correction behind the horizon}\label{sec:qec-pssy}
We now come to the first main goal of this paper, which is to apply equation \eqref{SQEC} to the black hole interior in the PSSY toy model for an evaporating black hole. Following PSSY, we consider a black hole entangled with radiation \emph{past} the Page time:
\begin{equation}
|\Psi\rangle = \frac{1}{\sqrt{k}}\sum_{\alpha=1}^k |\psi^{\alpha}\rangle_B\otimes |\alpha\rangle_R,\;\;\;\cdots \;\;\;(k \gg e^{S_0})
\end{equation}
The physical Hilbert space is $\mathcal{H}_{\text{phys}} = \mathcal{H}_{B}\otimes \mathcal{H}_R$, where $B$ is the microscopic quantum mechanical system describing the black hole, and $R$ is the Hilbert space of the radiation bath which absorbs the Hawking radiation. Loosely speaking, our code subspace will be the Hilbert space of bulk quantum field theory excitations on the black hole background, but we will specify it more precisely below. We will take this Hilbert space to factorize as
\begin{equation}
\mathcal{H}_{\text{code}} = \mathcal{H}_{\text{int}}\otimes \mathcal{H}_{\text{ext}},
\label{eq:code-space-grav}
\end{equation}
i.e., it is a tensor product of the degrees of freedom in the black hole interior times degrees of freedom in the black hole exterior.
Let $\{|i\rangle_\text{int} \}$ be an orthonormal basis spanning the interior Hilbert space $\mathcal{H}_{\text{int}}$ and let $\{|i'\rangle_\text{ext} \}$ be an orthonormal basis spanning the exterior Hilbert space $\mathcal{H}_{\text{ext}}$.
As before, there is a basis for $\mathcal{H}_\text{code}$ given by $\ket{i,i'}_\text{code} = \ket{i}_\text{int} \otimes \ket{i'}_\text{ext}$.
We view this system as a quantum error correcting code:
\begin{equation}
V: \mathcal{H}_{\text{int}}\otimes \mathcal{H}_{\text{ext}} \to \mathcal{H}_{\text{phys}}.
\end{equation}
Corresponding to a basis state $|i,i'\rangle_\text{code} \in \mathcal{H}_{\text{code}}$, we have an image under the encoding in the physical Hilbert space:
\begin{equation}
|\Psi_{i,i'}\rangle_{\text{phys}} = V |i,i'\rangle_{\text{code}}
\end{equation}
given by
\begin{equation}
|\Psi_{i,i'}\rangle_\text{phys} =\frac{1}{\sqrt{k}}\sum_{\alpha=1}^k |\psi^{\alpha}_{i,i'}\rangle_B\otimes |\alpha\rangle_R,
\end{equation}
where $|\psi^{\alpha}_{i,i'}\rangle_B$ is a state in the Hilbert space of the CFT dual to the black hole whose corresponding bulk geometry has an EOW brane in the internal state $\alpha$ and bulk quantum fields in the state $|i,i'\rangle_{\text{code}}$. Thus, the encoding map leads to a representation of the code state as an excitation around a particular evaporating black hole (entangled with the radiation bath), which has a universal coarse description as a gravitating geometry, and a dual microscopic CFT representation. Past the Page time, the quantum extremal surface for the radiation bath lies at the bifurcation point in the black hole geometry, so that by the general entanglement wedge reconstruction paradigm, we expect the factor of the code subspace associated to the interior of the black hole to be encoded within the radiation Hilbert space.
Formally, the state $|\psi^{\alpha}_{i,i'}\rangle_B$ can be constructed in the boundary CFT via a Euclidean path integral with sources turned on, corresponding to asymptotic boundary conditions for the bulk quantum fields (Fig.~\ref{fig:state}). In Sec.~\ref{sec:non-iso}, we will show that this holographic encoding map is only an isometry up to non-perturbatively small $O(d e^{-S_0})$ corrections. Therefore, we can use the error-correcting code theorems discussed in Sec.~\ref{sec:QECreview} (which strictly speaking assume isometric encoding), so long as we are considering a sufficiently small code subspace with $d \ll e^{S_0}$.
Another subtlety involves our assumption of a factorized code subspace in (\ref{eq:code-space-grav}). In the path integral construction of the encoding map, smooth sources producing finite energy densities will not result in states that are products between $\mathcal{H}_{\text{int}}$ and $\mathcal{H}_{\text{ext}}$. Conversely, a state for bulk quantum fields that is a product between the interior and exterior will generically have large energy density near the horizon. However, we are interested in an effective field theory around the black hole background, which therefore comes equipped with a UV cutoff (much smaller than Planck scale). The Hilbert space thus does include product states between the interior and exterior, provided the excitations are separated by distances much larger than the UV cutoff scale. For instance, we could imagine constructing a qubit degree of freedom in the interior and another qubit degree of freedom in the exterior, from sufficiently localized, low-energy operators in the effective field theory, well-separated from the horizon. Our analysis is performed in this setting. By turning on various Euclidean sources on the asymptotic boundary \cite{Marolf:2017kvq}, we can at least formally construct such states $|\psi^{\alpha}_{i,i'}\rangle_B$ corresponding to low energy bulk modes of the effective field theory in product states $|i,i'\rangle_{\text{code}}$. We will ignore $O(G_N)$ back-reaction corrections to the geometry and the state of bulk matter fields in the code subspace.
\begin{figure}
\centering
\includegraphics[height=3.5cm]{state.pdf}
\caption{\small{\textbf{(Left)} The state $|\psi^{\alpha}_{i,i'}\rangle_B$ can be constructed in the CFT from a Euclidean path integral with appropriate superpositions of source configurations (indicated by a cross) and an EOW brane boundary condition $\alpha$ at $\tau = - \beta/2$ (red dot). \textbf{(Right)} The bulk description has a black hole with an EOW brane (shown in red) and bulk matter fields in the state $|i,i'\rangle_{\text{code}}$. The black dot is the black hole bifurcation point.}}
\label{fig:state}
\end{figure}
As mentioned before, past the Page time, the black hole interior is in the entanglement wedge of the radiation/bath. We will ask what type of quantum operations acting on the bath Hilbert space $\mathcal{H}_R$ are correctable on the code subspace in the black hole interior. By correctable, we mean that we should be able to recover the reduced density matrix of bulk modes in the interior of the black hole (i.e., with support on $\mathcal{H}_{\text{int}}$) after the action of such errors.
\subsection{Coherent information criterion}
The question we wish to address fits perfectly within the formalism of subsystem quantum error correction explained in the previous section. Following that logic, we model the error as a quantum channel $\mathcal{E}=\{E_m\}$ which acts on the physical Hilbert space $\mathcal{H}_{B}\otimes \mathcal{H}_{R}$. More precisely, we will only consider errors which act on the radiation bath Hilbert space $\mathcal{H}_R$, leaving $\mathcal{H}_B$ untouched. Next, we introduce two reference systems $\mathfrak{R}_i$ and $\mathfrak{R}_e$ which have the same dimensions as $\mathcal{H}_{\text{int}}$ and $\mathcal{H}_{\text{ext}}$ respectively.
Recall that $\mathcal{H}_\text{int}$ and $\mathcal{H}_\text{ext}$ are the two factors of $\mathcal{H}_\text{code}$.
Now define the state
\begin{equation}
|\Psi'\rangle = \sum_{i,i',m} \ket{i,i'}_\text{ref} \otimes E_m|\Psi_{i,i'}\rangle_{\text{phys}} \otimes \ket{e_m}_\text{env}
\end{equation}
where, following the discussion in Sec.~\ref{sec:QECreview}, we have introduced another factor $\mathcal{H}_\text{env}$ corresponding to an auxiliary environment, and the reference space is $\mathcal{H}_\text{ref} = \mathcal{H}_{\mathfrak{R}_i} \otimes \mathcal{H}_{\mathfrak{R}_e}$. In order to determine whether the reduced density matrix of bulk quantum fields in the black hole interior is correctable under the action of $\mathcal{E}$, we are instructed to compute the mutual information:
\begin{equation}
I_{\Psi'}(\mathfrak{R}_i: \mathfrak{R}_e\cup \text{env}) \equiv S_{\Psi'}(\mathfrak{R}_i) + S_{\Psi'}(\mathfrak{R}_e\cup \text{env}) - S_{\Psi'}(\mathfrak{R}_i \cup \mathfrak{R}_e \cup\text{env}).
\label{eq:mutualinfo}
\end{equation}
The error is correctable if and only if $I_{\Psi'}(\mathfrak{R}_i: \mathfrak{R}_e\cup \text{env}) = 0$, i.e., the corresponding density matrix factorizes as
\begin{equation}
\rho'_{\mathfrak{R}_i,\mathfrak{R}_{e},\text{env}} = \rho'_{\mathfrak{R}_i} \otimes \rho'_{\mathfrak{R}_{e},\text{env}}.
\end{equation}
For gravity computations, it is often more convenient to consider the R\'enyi version of the above mutual information:
\begin{equation}
I^{(n)}_{\Psi'}(\mathfrak{R}_i: \mathfrak{R}_e\cup \text{env}) \equiv S^{(n)}_{\Psi'}(\mathfrak{R}_i) + S^{(n)}_{\Psi'}(\mathfrak{R}_e\cup \text{env}) - S^{(n)}_{\Psi'}(\mathfrak{R}_i \cup \mathfrak{R}_e \cup\text{env}).
\label{eq:renyimutualinformation}
\end{equation}
More precisely, instead of computing the von Neumann entropies $S_{vN}=-\mathrm{Tr}\,\rho \log \rho$, we compute the R\'enyi entropy $S^{(n)} = \frac{1}{1-n}\log\,\mathrm{Tr}\,\rho^n$ for integer values of $n$. The von Neumann entropies can then be obtained by analytically continuing the R\'enyi entropies in $n$ and taking the limit $n\to 1^+$. So, our goal now is to compute the above R\'enyi mutual information.
In order to compute the R\'enyi mutual information, we first compute the reduced density matrix on $\mathfrak{R}_i \cup \mathfrak{R}_e \cup \text{env}$ by tracing out the physical Hilbert space:
\begin{equation}\label{redDM}
\rho'_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}} = \frac{1}{k\mathcal{N}} \sum_{i,i',j,j',m,n}|i,i',e_m\rangle\langle j,j',e_n|_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}\sum_{\alpha,\beta}\langle \psi^{\alpha}_{j,j'}|\psi^{\beta}_{i,i'}\rangle_B \langle \alpha| E_n^{\dagger}E_m | \beta\rangle_{R},
\end{equation}
where $\mathcal{N}$ is a normalization constant. To determine the normalization constant, we take the trace of the above expression. Using equation \eqref{trC}, we get
\begin{equation}
\mathrm{Tr}\,\rho'_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}} = \frac{1}{k\mathcal{N}} \sum_{i,i'}\sum_{\alpha}\langle \psi^{\alpha}_{i,i'}|\psi^{\alpha}_{i,i'}\rangle_B
\label{eq:normalization-defn}
\end{equation}
where $i$ and $i'$ are indices in the code space and $\alpha$ is an index of the EOW brane state.
Using the gravity description in the semi-classical limit, the overlap is computed by the geometry depicted in Fig.~\ref{fig:BH}.
This geometry is filling in a boundary condition which consists of two copies of Fig.~\ref{fig:state} that are glued together to produce the overlap
\begin{equation}
\langle \psi^{\alpha}_{i,i'}|\psi^{\alpha}_{i,i'}\rangle_B= e^{S_0}Z_1 \langle i,i'|i,i'\rangle_{\text{code}}= e^{S_0}Z_1 ,
\end{equation}
where $Z_1$ is the exponential of the on-shell gravitational action (\ref{eq:jt-action}) (or beyond the semi-classical approximation, the JT gravity path integral) on a disc capped off by an EOW brane, along with the non-code-subspace bulk field theory modes in their Hartle-Hawking vacuum state. We explicitly pulled out the factor of $e^{S_0}$ coming from the first term in the JT action. Since we are ignoring the back-reaction from the effective field theory, so on the disc it simply gives the path integral overlap $\langle i, i' | i,i'\rangle_\text{code}$.
Then, enforcing $\Tr \rho'_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}} = 1$ determines the normalization factor
\begin{equation}
\mathcal{N} = d_{i}d_ee^{S_0}Z_1,
\label{eq:norm-density-matrix}
\end{equation}
where $d_i$ and $d_e$ are the dimensions of the interior and exterior factors in the code subspace respectively.
We will continue to make these contributions of the code subspace portion of the effective field theory explicit, and absorb the remainder of the bulk field theory modes into partition functions like $Z_1$.
\begin{figure}[t]
\centering
\includegraphics[height=5cm]{BH.pdf}
\caption{\small{The Euclidean black hole geometry which enters in determining the normalization factor $\mathcal{N}$ defined in \eqref{eq:normalization-defn}, specifically contributing to the overlap $\inner{\psi_{i,i'}^\alpha}{\psi_{i,i'}^\alpha}_B$. The red line is the EOW brane with index $\alpha$, the blue line is an interior code subspace excitation with label $i$, and the green line is an exterior code subspace excitation with label $i'$. }}
\label{fig:BH}
\end{figure}
Let us now evaluate $S^{(n)}_{\Psi'}(\mathfrak{R}_i \cup \mathfrak{R}_e \cup \text{env})$. From equation \eqref{redDM}, we first obtain
\begin{eqnarray}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}} &=& \frac{1}{k^n\mathcal{N}^n} \sum_{i_1,\cdots,i_n}\sum_{i_1',\cdots,i'_n}\sum_{\alpha_1\cdots,\alpha_n}\sum_{\beta_1,\cdots,\beta_n}\sum_{m_1,\cdots, m_n}\langle \psi^{\alpha_1}_{i_2,i_2'}|\psi^{\beta_1}_{i_1,i_1'}\rangle_B \cdots \langle \psi^{\alpha_n}_{i_1,i_1'}|\psi^{\beta_n}_{i_{n},i_{n}'}\rangle_B\nonumber\\
&\times & \langle \alpha_1| E_{m_2}^{\dagger}E_{m_1}|\beta_1\rangle_R\cdots \langle\alpha_n| E_{m_1}^{\dagger}E_{m_n}|\beta_n\rangle_R.
\end{eqnarray}
Here $i$ and $i'$ are indices in the code subspace, $\alpha$ and $\beta$ are EOW brane indices, and the $m$ indices label Kraus operators.
Notice that the reference space and auxiliary environment have disappeared here because we are taking the trace on them. It is convenient to group together all the factors pertaining to the channel $\mathcal{E}$ and the corresponding sums over the environment indices to define a tensor which we will call $f$:
\begin{equation}
f_{\alpha_1,\beta_1,\cdots \alpha_n,\beta_n} \equiv \sum_{m_1,\cdots, m_n}\langle \alpha_1| E_{m_2}^{\dagger}E_{m_1}|\beta_1\rangle_R\cdots \langle\alpha_n| E_{m_1}^{\dagger}E_{m_n}|\beta_n\rangle_R.
\end{equation}
We can re-express the $f$-tensor in terms of the isometric extension of $\mathcal{E}$ as
\begin{equation}
\begin{split}
f_{\alpha_1,\beta_1,\cdots \alpha_n,\beta_n} & = \sum_{m_1,\cdots, m_n}\sum_{\gamma_1,\cdots, \gamma_n}\langle \alpha_1,e_0|U_\mathcal{E}^{\dagger}| \gamma_1,e_{m_2}\rangle_{R,\text{env}} \langle \gamma_1,e_{m_1}|U_\mathcal{E} |\beta_1,e_0\rangle_{R,\text{env}} \cdots \\
& \quad \times \langle \alpha_n,e_0|U_\mathcal{E}^{\dagger}| \gamma_n, e_{m_1}\rangle_{R,\text{env}} \langle \gamma_n, e_{m_n}|U_\mathcal{E} |\beta_n,e_0\rangle_{R,\text{env}} ,
\end{split}
\label{eq:f-tensor}
\end{equation}
where in the second sum we have also inserted complete sets of radiation states $\sum_\gamma \ket{\gamma}\bra{\gamma}_R = \mathbb{1}_R$.
In terms of the $f$-tensor defined above, we have
\begin{equation} \label{tr1}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}} = \frac{1}{k^n\mathcal{N}^n} \sum_{i_1,\cdots,i_n}\sum_{i_1',\cdots,i'_n}\sum_{\alpha_1\cdots,\alpha_n}\sum_{\beta_1,\cdots,\beta_n}\langle \psi^{\alpha_1}_{i_2,i_2'}|\psi^{\beta_1}_{i_1,i_1'}\rangle_B \cdots \langle \psi^{\alpha_n}_{i_1,i_1'}|\psi^{\beta_n}_{i_{n},i_{n}'}\rangle_B\,f_{\alpha_1,\beta_1,\cdots,\alpha_n,\beta_n}.
\end{equation}
Using the isometric extension, the $f$-tensor can be conveniently visualized in terms of a tensor network, as shown in Fig.~\ref{fig:f}. Similarly, we get
\begin{equation} \label{tr2}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}} = \frac{1}{k^n\mathcal{N}^n} \sum_{i_1,\cdots,i_n}\sum_{i_1',\cdots,i'_n}\sum_{\alpha_1\cdots,\alpha_n}\sum_{\beta_1,\cdots,\beta_n}\langle \psi^{\alpha_1}_{i_1,i_2'}|\psi^{\beta_1}_{i_1,i_1'}\rangle_B \cdots \langle \psi^{\alpha_n}_{i_n,i_1'}|\psi^{\beta_n}_{i_{n},i_{n}'}\rangle_B\,f_{\alpha_1,\beta_1,\cdots,\alpha_n,\beta_n}.
\end{equation}
Finally,
\begin{equation}\label{tr3}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i} = \frac{1}{k^n\mathcal{N}^n} \sum_{i_1,\cdots, i_n}\sum_{i_1',\cdots,i_n'}\sum_{\alpha_1,\cdots, \alpha_n}\langle \psi^{\alpha_1}_{i_2,i_1'}|\psi^{\alpha_1}_{i_1,i_1'}\rangle_B\cdots \langle\psi^{\alpha_n}_{i_1,i_n'}|\psi^{\alpha_n}_{i_n,i_n'}\rangle_B.
\end{equation}
The coefficients in these various expressions involving products of overlaps of the black hole states can be computed using the standard rules of JT gravity.
\begin{figure}
\centering
\begin{tabular}{ c c}
\includegraphics[height=6cm]{TN.pdf} & \includegraphics[height=6cm]{code7.pdf}
\end{tabular}
\caption{\small{\textbf{(Left)} A tensor network representation of the index structure $f_{\alpha_1,\beta_1,\cdots,\alpha_n,\beta_n}$ with $n=3$. The light grey boxes denote $U_\mathcal{E}$ while dark grey boxes denote $U_\mathcal{E}^{\dagger}$. Dashed blue and red internal legs denote environment and radiation indices respectively. External legs are outgoing (incoming) to denote uncontracted indices in a bra (ket). \textbf{(Right)} The contraction of the $f$-index structure with the gravitational overlaps (i.e., asymptotic boundaries), denoted with solid black lines, required to compute the sums in (\ref{tr1}, \ref{tr2}). Each term in these sums involves a product of overlaps between $n$ sets of asymptotic gravitational states. The states are defined in terms of operators acting at the asymptotic boundaries which are illustrated here as black lines. Below we will see that the gravitational path integral can fill in these boundaries in various ways that may connect them or not through a bulk geometry.
}
}
\label{fig:f}
\end{figure}
More precisely, JT gravity plus EOW branes computes an \emph{ensemble average} of microscopic theories \cite{Saad:2019lba,Penington:2019kki,Gao:2021uro}, but the error correction theorems we developed in Sec.~\ref{sec:QECreview}, refer to single theories, not an ensemble. We can nevertheless use the JT gravity approach because, in the limit of large $k$ and $S_0$, the ensemble average is a good approximation to the typical answer since the variance is suppressed in $k$ and $e^{-S_0}$. We must carefully to account for the various different bulk geometries which can fill in the fixed asymptotic boundary conditions. Below, we will first consider the two most important geometries, namely the fully disconnected geometry and the fully connected one; both of these preserve the $\mathbb{Z}_n$ symmetry of the asymptotic boundary conditions. We will also argue later that at least for erasure errors, the $\mathbb{Z}_n$-breaking bulk geometries do not dominate, at least far from the phase transition between the two $\mathbb{Z}_n$-symmetric solutions. However, we do not have a general proof for all errors.
\subsection*{An assumption about $\mathcal{E}$}
In the computations that follow, we will assume the error operation $\mathcal{E}$ does not have prior access to black hole microstates, or more precisely, the $f$-tensor structure, which contains all the details of the quantum channel $\mathcal{E}$, has no additional black hole microstate overlaps in it. Thus, we will only have to evaluate the products of gravitational overlaps $\langle \psi^{\alpha}_{i,i'}|\psi^{\beta}_{j,j'}\rangle_B$ that are explicitly written in equations \eqref{tr1}, \eqref{tr2} and \eqref{tr3} using JT gravity. If the assumption were not true, then the evaluation of the $f$-tensor would contain additional gravitational path integrals, and we would have to account for additional geometries in JT gravity which potentially connect between the overlaps appearing explicitly in (\ref{tr1},\ref{tr2},\ref{tr3}) and those contained inside the $f$-tensor structure.
A simple example of this situation would be if the Stinespring dilation of the channel $\mathcal{E}$ couples an external environment qubit to the Petz operator reconstruction on the bath of a bulk operator in the black hole interior \cite{Penington:2019kki}. Since the Petz operator is constructed with knowledge of black hole microstates (or more precisely, with knowledge of the encoding map $V$), this channel would violate our assumption.\footnote{Such channels may be related to the ``miracle operators" discussed in \cite{Qi:2021sxb}.}
Furthermore, we expect such channels to also be exponentially complex in the sense of circuit complexity. Indeed, \cite{Kim:2020cds} conjectured that the density matrix on the bath Hilbert space is \emph{pseudorandom}, i.e., it cannot be reliably distinguished from the maximally mixed state on the bath with polynomial complexity operations.
It would be interesting to better understand the interplay between this pseudorandom assumption and our assumption about the absence of microstate overlaps in $\mathcal{E}$.
\subsection*{Fully disconnected geometry}
We now turn to the evaluation of the various entropies in JT gravity. To begin we consider the fully disconnected contribution where each of the gravitational overlaps in (\ref{tr1}), (\ref{tr2}) and (\ref{tr3}) is evaluated by a separate bulk geometry. In this case, we can simply use the formula
\begin{equation}
\langle \psi^{\alpha}_{i,i'}|\psi^{\beta}_{j,j'}\rangle_B= e^{S_0}Z_1 \delta_{\alpha\beta} \delta_{ij}\delta_{i'j'} .
\label{eq:DiscOverlap}
\end{equation}
This is derived in an identical fashion to the calculation of the norm $\mathcal{N}$ in Fig.~\ref{fig:BH}, but with the more general code subspace overlap $\inner{i,i'}{j,j'}_\text{code} = \delta_{ij}\delta_{i'j'}$ and EOW brane index matching yielding $\delta_{\alpha\beta}$. The resulting computation is diagrammed in Figs.~\ref{fig:discode1} and \ref{fig:discode2}, and gives
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}\Big|_{\text{disconn.}} = d^{1-n}_id^{1-n}_e\;\mathrm{Tr}\,\sigma^n_{\text{env}},
\label{eq:disc-result1}
\end{equation}
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}\Big|_{\text{disconn.}} = d^{1-n}_e\;\mathrm{Tr}\,\sigma^n_{\text{env}},
\label{eq:disc-result2}
\end{equation}
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}\Big|_{\text{disconn.}} = d^{1-n}_i,
\label{eq:disc-result3}
\end{equation}
where we have defined
\begin{equation}
\sigma_{\text{env}} = \mathrm{Tr}_{R} \left\{ U_\mathcal{E}\left(\frac{1}{k}\sum_{\alpha=1}^k |\alpha\rangle\langle\alpha|_R \otimes |e_0\rangle\langle e_0|_{\text{env}}\right) U_\mathcal{E}^{\dagger}\right\}.
\label{eq:sigmaenv}
\end{equation}
For example, the code subspace dimension factors in the disconnected contribution to $\Tr {\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$ are computed by observing in Fig.~\ref{fig:discode1} that there is one green index loop yielding $d_e$, one blue index loop yielding $d_i$, and the normalization $\mathcal{N}$ contributes $(d_id_e)^{-n}$.
The $\mathrm{Tr}\,\sigma^n_{\text{env}}$ contribution comes from the $f$-tensor with appropriate index contractions dictated by the EOW brane propagators (Kronecker deltas matching indices) in the geometry.
\begin{figure}
\centering
\includegraphics[height=4.5cm]{code1.pdf}
\caption{\small{The disconnected gravitational contribution to the R\'enyi entropy of $\mathfrak{R}_i\cup \mathfrak{R}_e \cup \text{env}$. The grey region is the gravitational geometry filling in the boundary conditions, while the red lines are EOW branes. Green and blue lines denote bulk contractions of the exterior and interior factors in the code subspace respectively. Dotted lines indicate identifying and summing over the code-subspace indices.}}
\label{fig:discode1}
\end{figure}
\begin{figure}
\centering
\begin{tabular}{c c}
\includegraphics[height=5.5cm]{code2.pdf} & \includegraphics[height=5.5cm]{code3.pdf}
\end{tabular}
\caption{\small{The disconnected gravitational contributions to the R\'enyi entropies of $\mathfrak{R}_e \cup\text{env}$ \textbf{(left)} and $\mathfrak{R}_i$ \textbf{(right)}. The grey region is the gravitational geometry filling in the boundary conditions, while the red lines are EOW branes. Green and blue lines denote bulk contractions of the exterior and interior factors in the code subspace respectively. }}
\label{fig:discode2}
\end{figure}
The contribution to the R\'enyi mutual information coming from the fully disconnected geometry is given by taking the logarithm of \eqref{eq:disc-result1} and subtracting from it the logarithm of \eqref{eq:disc-result2} plus the logarithm of \eqref{eq:disc-result3}:
\begin{equation}
I^{(n)}_{\Psi'}(\mathfrak{R}_i: \mathfrak{R}_e, \text{env})\Big|_{\text{disconn.}}=0.
\label{eq:disconnMI}
\end{equation}
Thus, when the disconnected geometry dominates the gravity path integral, the encoding of the black hole interior is robust, i.e., error channels that act only on the bath $\mathcal{H}_R$ do not affect the encoding of the black hole interior. Of course, the mutual information is not strictly zero, because of sub-leading in $e^{-S_0}$ contributions.
Nevertheless, the state on the black hole interior is at least approximately correctable, i.e., there exists a recovery channel $\mathcal{R}$ such that
\begin{equation}
\text{max}_{\rho}|| \mathcal{R}( \mathcal{E}(\rho)) - \rho ||_1 \leq \epsilon,\;\;\;\epsilon \sim e^{-S_0/4},
\end{equation}
where the maximization is over all density matrices supported on the code subspace.
This result follows from the error correction theorems we reviewed in Sec.~\ref{sec:QECreview}, specifically \eqref{eq:mutual-bound}.
\subsection*{Fully connected geometry}
Let us now consider the fully connected contribution. This geometry is the analog of the replica wormhole saddle. By explicit computation, one can prove that this contribution never dominates in $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}$, and the dominant contribution is always the disconnected geometry. However, in the other two cases -- i.e., $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$ and $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}$ -- the replica wormhole (Figs.~\ref{fig:conncode1} and \ref{fig:conncode2}) can become dominant, as we will see below by explicit computation. In these cases, the gravitational path integral for general $n$ result is not easy to compute. But because we are interested in taking the $n\to 1^+$ limit we can use the procedure proposed by Lewkowycz and Maldacena \cite{Lewkowycz:2013nqa}: (1) assume that the relevant geometries that contribute are the replica symmetric ones, (2)
compute the gravitational action of the $\mathbb{Z}_n$ quotient of these geometries (Figs.~\ref{fig:conncode1} and~\ref{fig:conncode2}, as shown on the right in Fig.~\ref{fig:conncode1}), and (3) analytically continue to $n=1^+$.
Applying this procedure in the $n\to 1^+$ limit, the quotient geometry is (approximately) identical to a copy of a single Euclidean black hole, but with a conical singularity at the black hole bifurcation point, and in addition a twist operator insertion for the bulk fields which corresponds to a specific prescription for contracting the bulk code subspace indices, as shown in Figs.~\ref{fig:conncode1} and \ref{fig:conncode2}.
For example, the gravitational contribution for the connected phase of $\Tr {\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$
is computed by evaluating Fig.~\ref{fig:conncode1}, and this gives
\begin{equation}
\frac{1}{k^n\mathcal{N}^n} \sum_{i_1,\dots , i_n} \sum_{i'_1,\dots , i'_n} \inner{\psi^{\alpha_1}_{i_2,i_2'}}{\psi^{\beta_1}_{i_1,i_1'}} \dots \inner{\psi^{\alpha_n}_{i_1,i_1'}}{\psi^{\beta_n}_{i_n,i_n'}} = e^{S_0(1-n)} \frac{Z_n}{k^nZ_1^n} d_e^{1-n} \delta_{\alpha_1\beta_2} \dots \delta_{\alpha_n\beta_1} ,
\label{eq:conn-EOW-structure}
\end{equation}
where $Z_n$ is the exponential of the on-shell gravity action on the disk geometry with $n$ EOW branes, times the bulk field theory path integral on this geometry.
There is a similar formula for the gravitational contribution to the connected phase of $\Tr {\rho'}^n_{\mathfrak{R}_e,\text{env}}$, with the only difference being that we get a factor of $d_i^{1-n}d_e^{1-n}$
instead of $d_e^{1-n}$ because of the different structure of the bulk code subspace excitation loops (blue and green loops in Figs.~\ref{fig:conncode1} and \ref{fig:conncode2}).
Putting everything together, we get
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}\Big|_{\text{conn.}} \sim e^{(1-n)[S_0+\frac{2\pi}{\beta}\phi_0+s_{\text{bulk}}]}d^{1-n}_e\;\mathrm{Tr}\,\sigma^n_{R},
\label{eq:re-env-conn1}
\end{equation}
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}\Big|_{\text{conn}}\sim e^{(1-n)[S_0+\frac{2\pi}{\beta}\phi_0+s_{\text{bulk}}]}d^{1-n}_id^{1-n}_e\;\mathrm{Tr}\,\sigma^n_{R},
\label{eq:re-env-conn}
\end{equation}
where $\phi_0$ is the value of the dilaton at the bifurcation point, and $s_{\text{bulk}}$ is the bulk entropy for all the bulk modes in their Hartle-Hawking vacuum (i.e., those which are not excited in the code subspace). In the above formulas, we have used $Z_n/(Z_1)^n\sim e^{(1-n)[S_0 + (2\pi/\beta) \phi_0 + s_{\text{bulk}}]}$ because we are working in the approximation that $(n-1)$ is small and positive, where the left hand side can be evaluated semi-classically using the Lewkowycz-Maldecena method.
The factors of $S_0$ and $\phi_0$ come from JT gravity, while the $s_{\text{bulk}}$ comes from the bulk field theory.
The density matrix $\sigma_R$ is defined as
\begin{equation}
\sigma_{R} = \mathrm{Tr}_{\text{env}} \left\{ U_\mathcal{E}\left(\frac{1}{k}\sum_{\alpha=1}^k |\alpha\rangle\langle\alpha|_R \otimes |e_0\rangle\langle e_0|_{\text{env}}\right) U_\mathcal{E}^{\dagger}\right\} = \mathcal{E} \left( \frac{1}{k}\mathbb{1}_R \right).
\label{eq:sigmaR}
\end{equation}
The second equation follows because, upon tracing out the environment factor from the action of the Stinespring dilation, we obtain the action of the channel $\mathcal{E}$.
The particular state upon which the channel acts is the maximally mixed radiation state, i.e., the ``semi-classical state'' on the radiation.
As before, this $\mathrm{Tr}\,\sigma^n_{R}$ contribution comes from the $f$-tensor with appropriate index contractions dictated by the EOW brane propagators in the connected geometry.
Notice that here the reduced state $\sigma_R$ has appeared, whereas in the disconnected geometry it was $\sigma_\text{env}$.
This difference has to do with the manner in which the EOW brane indices are contracted, and is demonstrated by explicit calculation.
\begin{figure}[t]
\centering
\begin{tabular}{c c c}
\includegraphics[height=4.8cm]{code4.pdf} & \hspace{0.2cm} & \includegraphics[height=4.4cm]{BH2.pdf}
\end{tabular}
\caption{\small{\textbf{(Left)} The connected gravitational contribution to the R\'enyi entropy of $\mathfrak{R}_i\cup \mathfrak{R}_e \cup \text{env}$. Green and blue lines denote bulk contractions of the exterior and interior factors in the code subspace respectively. The dashed white lines separate fundamental domains of the $\mathbb{Z}_3$ replica symmetry. \textbf{(Right)} The $\mathbb{Z}_n$ quotient of the replica wormhole in the $n\to 1^+$ limit. The geometry of this manifold, with a cut on the interior, provides a specific prescription for the contractions of the bulk code subspace indices.
The bulk field theory state created on the $\mathbb{Z}_2$-symmetric slice is replicated and results in a bulk R\'enyi entropy for the black hole interior.
In order to compute this entropy, the interior code subspace excitation (blue line) must pass between replicas and the exterior code subspace excitation (green line) must remain within a given replica.
This pattern of contractions is drawn on the left figure.
The images of the cut on the $n\to 1^+$ geometry are the white dashed lines on the $n=3$ geometry.
}}
\label{fig:conncode1}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=5.6cm]{code5.pdf}
\caption{\small{The connected gravitational contribution to the R\'enyi entropy of $\mathfrak{R}_e \cup \text{env}$. The grey region is the manifold which fills in the boundary conditions defined by the gravitational overlaps in \eqref{tr2}. The red lines are the EOW branes. Green and blue lines denote bulk contractions of the exterior and interior factors in the code subspace respectively. }}
\label{fig:conncode2}
\end{figure}
We can take logarithms of the above formulae to show that when the fully connected geometry dominates in the R\'enyi entropies we get a non-trivial mutual information
\begin{equation}
I_{\Psi'}(\mathfrak{R}_i: \mathfrak{R}_e, \text{env})\Big|_{\text{conn.}}=2 \log d_i.
\label{eq:connMI}
\end{equation}
Therefore, in this case the bulk density matrix in the black hole interior is not correctable under the action of the error on the bath. The mechanism underlying this failure of error correction in gravity is the appearance of the replica wormhole.
Finally, we are interested in the conditions for the fully connected geometry to dominate over the disconnected one.
For $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}$, the criterion for the connected geometry to be significantly greater than the disconnected one when $(n-1)$ is small and positive is
\begin{equation}
\frac{\mathrm{Tr}\,\sigma^n_{\text{env}}}{\mathrm{Tr}\,\sigma^n_{R}} \ll e^{(1-n)[S_{BH} + \log d_i]} .
\label{eq:entropy-diff}
\end{equation}
The analogous criterion for $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$ is
\begin{equation}
\frac{\Tr \sigma_\text{env}^n}{\Tr \sigma_R^n} \ll e^{(1-n)[S_{BH} - \log d_i]} ,
\label{eq:correctability}
\end{equation}
where we have defined $S_{BH} = S_0+\frac{2\pi}{\beta}\phi_0 + s_{\text{bulk}}$. (As before, $S_0$ is the extremal entropy, $\frac{2\pi}{\beta}\phi_0$ is the classical horizon contribution, and $s_{\text{bulk}}$ is the entropy of bulk quantum fields.)
Notice that the first inequality actually implies the second, so if $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}$ is dominated by the connected geometry then so is $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$, but the converse is not true except when $d_i = 1$.
If we only have $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$ dominated by the connected geometry, the above equations show we still have $I_{\Psi'} \neq 0$.
Therefore, we ought to characterize correctability only by the connected phase of $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$.
Taking the logarithm on both sides of \eqref{eq:correctability}, multiplying by $\frac{1}{n-1}$, and taking $n\to 1^+$, we get
\begin{equation}\label{ci}
S(\sigma_{R})-S(\sigma_{\text{env}}) \ll -( S_{BH}- \log d_i),
\end{equation}
The left hand side above is entirely a property of the quantum channel $\mathcal{E}$, with no reference to gravity. In quantum information theory, this quantity is referred to as the \emph{coherent information}.
For a quantum channel $\mathcal{E}$ which can be isometrically extended to a unitary $U_\mathcal{E}$ acting jointly on the system plus an environment factor $\mathcal{H}_{\text{env}}$, the coherent information of the channel $\mathcal{E}$ with respect to an input state $\rho$ is defined as:
\begin{equation}
I_c(\mathcal{E};\rho) \equiv S(\mathcal{E}(\rho))-S(\rho'_{\text{env}}),
\label{eq:coherentinfo}
\end{equation}
where $\rho'_{\text{env}}$ is the reduced density matrix of the environment after the action of the joint unitary $U_\mathcal{E}$.
The coherent information is a measure for how much of the quantum information in the input state survives the action of the quantum channel; for instance, it decreases monotonically under composition of quantum channels. For our purposes, the input state is the maximally mixed state on $R$, which we can loosely think of as the ``semi-classical state'' of the radiation, i.e., a coarse-grained description of the state of the radiation obtained by keeping track of the coupled evolution of the bath with the black hole in the semi-classical gravity description.\footnote{In contrast, the microscopic state of the radiation obtained by keeping track of the coupled evolution with the quantum system dual to the black hole will not be maximally mixed/thermal and will have subtle correlations which restore unitarity.} Equation \eqref{ci} therefore states that if a quantum channel has a sufficiently negative coherent information with respect to the semi-classical state on the radiation Hilbert space, then the encoding of the black hole interior is no longer correctable under such a channel. Conversely, the encoding of the bulk state in the black hole interior is robust against the action of any quantum channel satisfying
\begin{equation} \label{CIcond}
I_c \left( \mathcal{E};\frac{1}{k}\mathbb{1} \right) \gg -( S_{BH}-\log d_i),
\end{equation}
together with our previous assumption that the channel should not alter the asymptotic Euclidean boundary conditions by sourcing additional boundaries. Therefore, as long as the remaining black hole is macroscopic, we find that the black hole interior is robustly encoded in the radiation.
If the black hole entropy is instead so small as to be comparable to the $O(1)$ code subspace entropy $\log d_i$, an error with very small negative coherent information could corrupt the interior encoding. In either case, once the bound \eqref{CIcond} is violated, the black hole interior moves over to the entanglement wedge of the environment factor $\mathcal{H}_{\text{env}}$ and QEC on the bath fails at that point.
It is worth emphasizing that a channel with large, negative coherent information need not be very complex. Consider the following example. Let the dimension of the environment factor be the same as the dimension of the bath Hilbert space, i.e., $\mathrm{dim}(\mathcal{H}_{R}) = \mathrm{dim}(\mathcal{H}_{\text{env}})=k= 2^M$, and let the unitary $U$ correspond to the \text{SWAP} operator. For such a channel, we have
\begin{equation}
S(\sigma_R) = 0,\;\;\;S(\sigma_\text{env}) = \log\,k = M \log 2,
\end{equation}
and so the coherent information is given by
\begin{equation}
I_c \left( \mathcal{E};\frac{1}{k}\mathbb{1} \right) = -M \log 2 \ll -(S_{BH}- \log\,d_i).
\end{equation}
Clearly, the SWAP channel violates the bound \eqref{CIcond}. This is not at all surprising because this channel essentially transfers all the entanglement from the radiation density matrix into the environment and replaces the original state of the radiation with $|e_0\rangle\langle e_0|_R$; the interior code subspace could not possibly be protected under such a drastic replacement. But importantly, the SWAP channel is \emph{not} very complex\footnote{By the complexity of a quantum channel, we mean the minimum circuit complexity of the unitary $U_\mathcal{E}$ in its isometric extension.} -- for every qubit in the radiation, we need one 2-qubit SWAP gate, and so the complexity of this channel should be $M\sim \log\,k$.
Note, however, that by unitarity of quantum mechanics, if we treat the radiation as a subsystem and implement the erasure channel via a SWAP operator, this procedure does not actually destroy any information.
All it does is move where the black hole interior has been encoded in the larger system.
This amounts to including the environment as a subsystem upon which we may try to reconstruct the interior.
\subsection*{Black hole exterior}
\begin{figure}
\centering
\includegraphics[height=5.6cm]{code6.pdf}
\caption{\small{The connected gravitational contribution to the R\'enyi entropy of $\mathfrak{R}_i \cup \text{env}$. Green and blue lines denote bulk contractions of the exterior and interior factors in the code subspace respectively. }}
\label{fig:ext}
\end{figure}
We can also ask whether the operation $\mathcal{E}$ can affect the black hole exterior. In this case, we need to compute the mutual information $I_{\Psi'}(\mathfrak{R}_e: \mathfrak{R}_i \cup \text{env})$. When the disconnected geometry dominates, the calculation is essentially the same as before and we get zero mutual information. When the connected geometry dominates, the calculation of $S(\mathfrak{R}_i\cup \mathfrak{R}_e \cup \text{env})$ is also the same as before. The main difference is in $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\text{env}}$ (Fig.~\ref{fig:ext}).
Instead of \eqref{eq:re-env-conn}, we get
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\text{env}}\Big|_{\text{conn.}}\sim e^{(1-n)S_{BH}} \;\mathrm{Tr}\,\sigma^n_{R},
\end{equation}
where $(n-1)$ is small and positive. Note also that $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e}$, which we need in place of \eqref{eq:disc-result3}, is always dominated by the disconnected geometry:
\begin{equation}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e}\Big|_{\text{disconn.}} = d^{1-n}_e.
\end{equation}
Using these expressions, we can check that $I_{\Psi'}(\mathfrak{R}_e: \mathfrak{R}_i \cup \text{env})=0$, and thus the black hole exterior is never affected by quantum operations on the radiation, which is of course consistent with the fact that the exterior is not in the entanglement wedge of the radiation.
\subsection{Approximate isometry of the interior encoding}\label{sec:non-iso}
The theorems concerning standard quantum error correction reviewed in Sec.~\ref{sec:QECreview} assumed an isometric embedding $V$ of the code Hilbert space in the physical Hilbert space.
However, the map $V\ket{i,i'}_\text{code} = \ket{\Psi_{i,i'}}_\text{phys}$ for the black hole interior is not actually an isometry.
This is because
the enormous number of effective field theory states in the growing interior will eventually outnumber the states of the physical Hilbert space as measured by the black hole entropy.
Thus, states that are orthogonal in the effective field theory cannot be orthogonal after embedding in the physical Hilbert space.
However, we might expect that
for code subspaces which are small compared to the black hole Hilbert space, the interior encoding map $V$ should be an approximate isometry \cite{Kim:2020cds}, and therefore the standard theorems should still be approximately valid and thus can be used to diagnose the existence of a recovery channel.
We can use the gravitational path integral to demonstrate that this expectation is indeed true. The map $V$ can be made arbitrarily close to an isometry by appropriately tuning the code subspace relative to the black hole Hilbert space.
To see this, we will evaluate the Frobenius norm of the operator $P^2 - P$, where $P = V V^\dagger$ is the code subspace projector.
If $V$ were an isometry, $P$ would be a projector and we would have $P^2 - P = 0$.
So, if the Frobenius norm of this quantity is small, $V$ is approximately an isometry.
The first step is to normalize $P$ by requiring that
\begin{equation}
\frac{1}{\mathcal{N}_P} \Tr P = d ,
\label{eq:projectornorm}
\end{equation}
where ${\mathcal{N}_P}$ is a normalization and $d$ is the code subspace dimension, as appropriate for normalizing a projector.
Using $P = VV^\dagger$ and the definition of $V$ as above, the left hand side reduces to a combination of gravitational overlaps
\begin{equation}
\frac{1}{\mathcal{N}_P} \sum_{i,i'} \inner{\Psi_{i,i'}}{\Psi_{i,i'}}_\text{phys} = d .
\end{equation}
These overlaps reduce to $\inner{\Psi_{i,i'}}{\Psi_{i,i'}}_\text{phys} = e^{S_0} Z_1$ where we separated out the $e^{S_0}$ factor from the JT gravity path integral on the disc as before in \eqref{eq:norm-density-matrix}.
Therefore, the normalization for $P$ is
\begin{equation}
\mathcal{N}_P = e^{S_0}Z_1 .
\end{equation}
As usual, we have ignored terms which are subleading in the genus expansion, which is justified as we have $S_0 \gg 1$.
This normalization essentially fixes the orthonormality of the code subspace states $\inner{i,i'}{j,j'}_\text{code} = \delta_{ij}\delta_{i'j'}$.
We are now interested in the Frobenius norm
\begin{equation}
\| \widehat{P}^2 - \widehat{P} \|_F^2 \equiv \Tr (\widehat{P}^2-\widehat{P})^2 ,
\end{equation}
where we normalized $\widehat{P} \equiv P/\mathcal{N}_P$ to have a proper projector (see Eq.~\ref{eq:projectornorm}).
The leading contribution to this norm in the genus expansion should arise from the fully disconnected contribution, where all gravitational boundaries are filled separately by disk geometries with an EOW brane. In fact, an explicit calculations shows that these contributions cancel in the
function of $\widehat{P}$ appearing in the Frobenius norm.
To see this, we make the replacement $\inner{\Psi_{i,i'}}{\Psi_{j,j'}} = e^{S_0}Z_1 \delta_{ij}\delta_{i'j'}$.
The completely disconnected contributions are then
\begin{equation}
\| \widehat{P}^2 - \widehat{P} \|_F^2 \Big|_{\text{total disconn.}} = \frac{d e^{4S_0} Z_1^4}{\mathcal{N}_P^4} - \frac{2d e^{3S_0} Z_1^3}{\mathcal{N}_P^3} + \frac{d e^{2S_0} Z_1^2}{\mathcal{N}_P^2} = 0 .
\end{equation}
where the numerators were computed by direct evaluation of the relevant path integrals in the same manner as in the calculation of (\ref{eq:norm-density-matrix}).
Since this vanishes, we look for the next order contribution in the genus expansion. This comes from the path integral geometries in which precisely one pair of asymptotic boundaries is connected by a wormhole bounded on either side by EOW branes; we will call the JT path integral on this connected pair geometry $Z_2$. In $\Tr \widehat{P}^n$ this term appears from the overall path integral geometry with $n-1$ connected components; so the result will be suppressed by $e^{-S_0}$ compared to each term in the disconnected contribution.
Thus,
\begin{equation}
\| \widehat{P}^2 - \widehat{P}\|_F^2 \Big|_{Z_2} = \frac{6d^2 e^{3S_0} Z_2Z_1^2}{\mathcal{N}_P^4} - \frac{6d^2 e^{2S_0}Z_2Z_1}{\mathcal{N}_P^3} + \frac{d^2e^{S_0}Z_2}{\mathcal{N}_P^2} = d^2 e^{-S_0} \frac{Z_2}{Z_1^2} .
\end{equation}
The total Frobenius norm is therefore
\begin{equation}
\| \widehat{P}^2 - \widehat{P}\|_F^2 = d^2 e^{-S_0} \frac{Z_2}{Z_1^2} + O(e^{-2S_0}) ,
\end{equation}
where the $O(e^{-2S_0})$ term includes contributions with fewer disconnected components, and the genus expansion of each component.
This result tells us that by taking $d \ll e^{S_0/2}$, we can ensure that $\widehat{P}$ is exponentially close to a true projector in the physical Hilbert space.
Thus the assumption in \cite{Kim:2020cds} that the encoding map for the interior is close to an isometry holds true in the PSSY model of two-dimensional gravity.
To use the matrix ensemble language for JT gravity plus EOW branes \cite{Saad:2019lba}, we have shown that a typical member of the boundary ensemble implements an approximately isometric bulk-to-boundary encoding map $V$. So we are justified in drawing conclusions about the error correction properties of the typical theory by using the decoupling principle and similar standard isometric error correction theorems.
A more refined analysis is possible.
We could imagine deriving completely general Knill-Laflamme conditions for the existence of a recovery channel when the encoding map is arbitrarily far from an isometry.
In this most general case, $V$ is simply a linear map, not necessarily isometric, which must be injective to provide a faithful encoding. Injectivity certainly requires $d \ll e^{S_0}$. However, whether this map is actually faithful depends on how it maps the code space into the actual physical Hilbert space. If we are dealing with a theory of gravity like JT gravity that is dual to an ensemble of microscopic theories, faithfulness of the encoding $V$ requires that the measure in the ensemble should not be concentrated on theories for which $V$ is degenerate. In JT gravity, this is likely the case, as the EOW brane states which we labeled by $\alpha$ are dual to states that include random coefficients in the energy basis \cite{Penington:2019kki}, and are thus highly unlikely to lead to degeneracies in $V$ for small $d$ compared to $e^{S_0}$. However, despite this, it does not follow that the encoding map is necessarily an isometry, or even close to one, because the encoding vectors may have large overlaps in individual members of the ensemble. If we could write generalized Knill-Laflamme conditions, along with a decoupling principle that applied to such non-isometric encodings, we could compute the relevant mutual information
to diagnose the existence of a recovery channel for some error $\mathcal{E}$ and given linear map $V$.
In this way, we could aim to derive entropic conditions for the existence of a recovery map in {\it every} member of the underlying ensemble for a given error, instead of showing its existence only for a typical member of the ensemble as we have done in this paper.
The non-isometric codes discussed in \cite{Akers:2021fut} may give some intuition for what the correct generalized Knill-Laflamme conditions should be.
In the language used in that work, in this paper we are studying a sort of state-independent sector of typical non-isometric codes where the dimension of the code subspace is small enough to ensure that more conventional error correction criteria are still approximately valid.
\section{Erasures and random errors}
\label{sec:ErasuresTypical}
The previous section developed general results about robustness against errors of the encoding of the black hole interior in the Hawking radiation. In this section we further investigate two kinds of error channels for which more detailed analyses are possible. In the first class of errors, in which some of the physical bits in the radiation are simply erased, we are able to explicitly sum over all planar gravitational geometries (following the resolvent method of \cite{Penington:2019kki}) contributing to mutual information that diagnoses error correction. In the second class consisting of typical errors, where a random unitary acts on the radiation and an auxiliary environment Hilbert space, we give a geometric formula for the mutual R\'enyi entropy in terms of gravitational path integrals, which when evaluated in JT gravity, reproduces the same coherent information bound obtained previously.
\subsection{Erasure errors}
Our analysis in the previous section was in the semi-classical approximation where we restricted attention to replica-symmetric saddles. In this section, we will work with the full gravitational path integral, but in order to do so, we will need to specialize to a particular quantum channel: partial erasure of the radiation.
Suppose the error channel $\mathcal{E}_\text{erase}$ simply erases $t$ of the $M$ qubits in $\mathcal{H}_R$.
The channel applied to the maximally mixed state on $\mathcal{H}_R$ then yields a maximally mixed state of dimension $2^{M-t}$, since the partial trace over $t$ qubits of a maximally mixed state on $M$ qubits gives a maximally mixed state on $M-t$ qubits.
A suitable isometric extension of $\mathcal{E}_\text{erase}$ is simply a partial analog of the SWAP operator discussed previously, where the unitary $U_\mathcal{E}$ exchanges $t$ ``clean'' environment qubits with $t$ radiation qubits. More precisely, by ``clean'' we mean that we take the environment to initially consist of $t$ qubits each in some fixed fiducial state $|0\rangle\langle 0|$, and the action of the channel swaps the state on $t$ of the radiation bath qubits with these $t$ environment qubits. Afterward, the remaining radiation entropy is $\log 2^{M-t}$ and the environment entropy is $\log 2^t$.
The coherent information is then
\begin{equation}
I_c \left( \mathcal{E}_\text{erase} , \frac{1}{k} \mathbb{1} \right) = (M-2t) \log 2 .
\end{equation}
The gravitational meaning of this result is surprising.
We have found that for, e.g., $t \approx M/2$, the coherent information vanishes or is at least very small in magnitude compared to a large black hole entropy.
This means that we can erase considerably more than half of the radiation and still reconstruct essentially all effective field theory operators in the black hole interior.
Hints of this robustness were seen in doubly holographic models of evaporation \cite{Balasubramanian:2020hfs}, where a multiboundary wormhole model was used to reach roughly the same conclusion.
But just how sharp is the transition between a correctable and non-correctable erasure error?
In other words, can we produce a more accurate criterion than our condition that the coherent information of the erasure channel must exceed minus the black hole entropy by ``a large amount'', as written in \eqref{CIcond}?
In order to study this transition more precisely, we must incorporate more gravitational geometries than just the disconnected and fully connected phases that we studied in Sec.~\ref{sec:qec-pssy}.
As discussed in Sec.~\ref{sec:qec-pssy}, in deriving our coherent information criterion we neglected the contribution from geometries breaking $\mathbb{Z}_n$ symmetry. For erasure errors, we can do better by resumming all the planar geometries; this is a good approximation when the parameters $k, e^{S_0}, d_i$ etc. running in loops are large, and it will give us a sharper picture of the correctability transition for erasure errors. We will perform the planar resummation by writing down a Schwinger-Dyson equation for a resolvent, following \cite{Penington:2019kki}.
The general idea of this method is to compute the entropy $S(\rho)$ of a density matrix $\rho$ indirectly by first computing a resolvent matrix $\mathcal{R}(w) = [w\mathbb{1} - \rho]^{-1}$.
As is standard practice in, e.g., random matrix theory, we may obtain the eigenvalue density $D(w)$ for $\rho$ by computing the discontinuity of the resolvent across the real axis, $2\pi \mathrm{i} D(w) = \Tr [ \mathcal{R}(w-\mathrm{i}\epsilon) - \mathcal{R}(w+\mathrm{i}\epsilon) ]$ with $\epsilon \to 0$.
With $D(w)$ in hand, the entropy $S(\rho)$ is easily obtained by writing $S(\rho) = -\int \mathrm{d}w\; D(w) \; w \log w$.
We may similarly evaluate \textit{any} function of the eigenvalues of $\rho$ using $D(w)$; if we wish to compute $\Tr \rho^n$, for instance, we write $\Tr \rho^n = \int \mathrm{d}w \; D(w) w^n$.
The resolvent is particularly useful in summing planar diagrams for a density matrix $\rho$ that involves gravitational path integrals because there is a recursive Schwinger-Dyson equation that determines $\mathcal{R}$ exactly in the planar approximation.
We are going to relax the Lewkowycz-Maldacena assumption of $\mathbb{Z}_n$ replica symmetry, and will replace it with three simplifying assumptions in order to keep the calculation tractable.
As we are most interested in subsystem correctability of $\mathcal{H}_\text{int}$, we will assume that (i) we restrict ourselves to a code subspace where $\mathcal{H}_{\text{ext}}$ is trivial (i.e., one-dimensional). Next, as we noticed in Sec.~\ref{sec:qec-pssy}, part of the utility of the Lewkowycz-Maldacena assumption was that we were able to choose a particular contraction pattern for the bulk code subspace indices based on the geometry of the replica symmetry quotient manifold (see the discussion around Fig.~\ref{fig:conncode1}). In relaxing the Lewkowycz-Maldacena assumption to incorporate replica symmetry breaking geometries, we want to continue choosing contraction patterns for the code subspace indices based only on the geometry of the replica manifold, so we assume that (ii) $\mathcal{H}_{\text{int}}$ corresponds to a degree of freedom moving on or close to the EOW brane. Finally, the canonical ensemble involves integral equations relating the partition functions $Z_n$ and the density of states, so we will assume that (iii) we work in the microcanonical ensemble for the black hole, since the density of states in the microcanonical ensemble is flat in the microcanonical window and is therefore related to the microcanonical partition function by a simple algebraic equation; this will greatly simplify our Schwinger-Dyson equation for the resolvent.
Let us be more explicit about the implication of assumption (ii).
In products of gravitational microstate overlaps, if $\mathcal{H}_\text{int}$ is a degree of freedom that is ``on or close to the EOW brane'', we can simply contract this index identically to the contraction pattern of the EOW branes.
For example, consider the overlap $\inner{\psi^{\alpha_1}_{i_1}}{\psi^{\beta_1}_{j_1}}\inner{\psi^{\alpha_2}_{i_2}}{\psi^{\beta_2}_{j_2}}_B$, where $i_1,j_1,i_2,j_2$ are the interior code subspace indices and $\alpha_1,\beta_1,\alpha_2,\beta_2$ are EOW brane indices.
Our assumption (ii) means that if we contract EOW brane indices to obtain $\delta_{\alpha_1\beta_1}\delta_{\alpha_2\beta_2}$, we require that the code subspace indices are contracted to yield $\delta_{i_1j_1}\delta_{i_2j_2}$.
Similarly, if we contract EOW brane indices as $\delta_{\alpha_1\beta_2}\delta_{\beta_1\alpha_2}$, we require that the code subspace contractions give $\delta_{i_1j_2}\delta_{j_1i_2}$.
This simplifies our analysis of replica symmetry breaking geometries considerably because it allows us to discard configurations where the EOW brane indices are contracted in a different pattern than the interior code subspace indices, a situation which was in principle possible in Sec.~\ref{sec:qec-pssy} but which did not actually arise due to the Lewkowycz-Maldacena assumption.
Here, we are relaxing the Lewkowycz-Maldacena assumption but we will keep one of its important implications by making assumption (ii) to keep the calculation tractable.
With these simplifications, we can work out the R\'enyi mutual information $I^{(n)}_{\Psi'}(\mathfrak{R}_i:\text{env})$ in the relevant state
\begin{equation}
\ket{\Psi'} = \frac{1}{\mathcal{N}^{1/2}}\sum_{i=1}^{d_i} \ket{i}_{\mathfrak{R}_i} \otimes U_\mathcal{E} (\ket{\Psi_i}_\text{phys} \otimes \ket{e_0}_\text{env} ) , \quad \mathcal{N} = d_i e^{S_0} Z_1 ,
\end{equation}
where the normalization is the $d_e=1$ specialization of the one we had previously \eqref{eq:norm-density-matrix}.
Note that the R\'enyi mutual information has the property that it vanishes if and only if the state factorizes. So, it is just as good as the mutual information $I_{\Psi'}(\mathfrak{R}_i : \text{env})$ as a diagnostic of decoupling between the reference system and the environment.
Let us first consider $\mathrm{Tr}\,\rho'^n_{\mathfrak{R}_i,\text{env}}$.
It is convenient to first write the matrix elements for the reduced density matrix, each of which involves a gravitational overlap
\begin{equation}
\bra{i,e_m} \rho'_{\mathfrak{R}_i,\text{env}} \ket{j,e_n} = \frac{1}{k \mathcal{N}} \sum_{\alpha,\beta,\gamma=1}^k \bra{\gamma,e_m} U_\mathcal{E} \ket{\beta,e_0} \inner{\psi^\alpha_i}{\psi^\beta_j} \bra{\alpha,e_0} U_\mathcal{E}^\dagger \ket{\gamma,e_n}.
\label{eq:Ri-env-element}
\end{equation}
We can sum over planar geometries by introducing the resolvent matrix $\mathcal{R}$:
\begin{equation}
\mathcal{R}(w) \equiv \frac{1}{(w \mathbb{1} - \rho'_{\mathfrak{R}_i,\text{env}})}= \frac{1}{w}\mathbb{1}+\frac{1}{w^2}\rho'_{\mathfrak{R}_i,\text{env}}+\frac{1}{w^3}\rho'^2_{\mathfrak{R}_i,\text{env}}+\cdots,
\label{eq:resolvent-expansion}
\end{equation}
where $w$ is a complex number and $\mathcal{R}$ has dimension $\dim (\mathcal{H}_\text{int} \otimes \mathcal{H}_\text{env})$.
We may represent this infinite sum pictorially using the same gravitational boundary conditions and tensor network constructions as in Fig.~\ref{fig:f}:
\begin{equation}
\includegraphics[height=0.8cm]{resolvent.pdf}
\label{eq:resolvent-bc-pic}
\end{equation}
Here the black lines denote asymptotic boundaries and come with a factor of $(k\mathcal{N})^{-1}$, dashed blue lines denote environment indices, and dashed red lines denote radiation indices; we have suppressed the $\mathfrak{R}_i$ index for simplicity. Further, $U_\mathcal{E}$ are denoted with light gray boxes and $U_\mathcal{E}^\dagger$ are denoted with dark gray boxes. The $n$th term additionally also comes with the appropriate power $w^{-n}$ which has been left implicit in the figure.
As an example of how \eqref{eq:resolvent-expansion} is translated to \eqref{eq:resolvent-bc-pic}, take the first non-trivial term in \eqref{eq:resolvent-bc-pic}, which is pictorially representing the matrix element \eqref{eq:Ri-env-element}.
The gravitational boundary condition is $\inner{\psi_i^\alpha}{\psi^\beta_j}_B$, represented by the solid black line oriented from ket to bra.
Implicit at the left and right endpoints of this condition are the $(\beta,j)$ and $(\alpha,i)$ indices, respectively.
The left endpoint with radiation index $\beta$ feeds into the light gray $U_\mathcal{E}$ operator, along with a hanging environment leg that is implicitly contracted with $e_0$.
The outgoing environment leg of $U_\mathcal{E}$ is $e_m$, while the outgoing radiation leg is $\gamma$, and this radiation index is matched to the incoming radiation index of the dark gray $U_\mathcal{E}^\dagger$ operator.
The incoming environment leg in $U_\mathcal{E}^\dagger$ is $e_n$, while the outgoing radiation leg is $\alpha$, matched to the EOW brane index at the right endpoint of the black line.
Finally, the outgoing radiation leg on $U_\mathcal{E}^\dagger$ is implicitly contracted with $e_0$.
We have suppressed the $\mathfrak{R}_i$ indices since these would just add an extra incoming and outgoing dashed line attaching to the endpoints of the black line, but not interacting with the $U_\mathcal{E}$ or $U_\mathcal{E}^\dagger$ operators.
The missing factors of $w$ and $k\mathcal{N}$ are implicit as mentioned under \eqref{eq:resolvent-bc-pic}.
Evaluating \eqref{eq:resolvent-bc-pic} exactly would be very difficult if we were to allow arbitrary gravitational geometries filling in the boundary conditions.
However, in the planar approximation, there is an interesting simplification, as exploited in \cite{Penington:2019kki}.
If we only consider planar gravitational geometries, namely those whose index loops and bulk gravitational components do not cross each other and do not include higher genus contributions, there is a recursive Schwinger-Dyson equation for this resolvent where the right hand side of \eqref{eq:resolvent-bc-pic} can be written pictorially as
\begin{equation}
\includegraphics[height=2.5cm]{resolvent2.pdf},
\label{eq:SD-pic}
\end{equation}
where the shaded regions denote bulk Euclidean spacetimes filling in the asymptotic boundary conditions with the solid red lines denoting EOW branes, and the orange blobs stand for insertions of the resolvent.
The logic behind why this sort of recursion captures all the planar contributions is as follows.
The first nontrivial term in \eqref{eq:SD-pic} captures all planar geometries in which the first asymptotic boundary is not connected through the bulk to any other asymptotic boundary.
Similarly, the second nontrivial term captures all planar geometries in which the first asymptotic boundary is connected through the bulk to just one other asymptotic boundary, but this second boundary does not necessarily have to be adjacent to the first, and it also does not necessarily have to be the last asymptotic boundary on the right.
To capture the situations in which they are not adjacent, the resolvent (orange blob) is inserted between the two boundaries, and to capture the situations in which the second boundary is not also the last boundary on the right, the resolvent also is inserted to the right of the second boundary.
These two resolvent insertions are independent because any connection through the bulk between them would render the geometry non-planar.
Iterating this sequence, and noting that every planar geometry can be put into exactly one such class, we arrive at the recursion relation \eqref{eq:SD-pic}.
We can translate \eqref{eq:SD-pic} back to an equation by recalling several facts.
First, $Z_m$ is the JT gravity path integral on a disk with $m$ EOW branes, and clearly the $m^{\text{th}}$ nontrivial term on the right hand side of \eqref{eq:SD-pic} carries a factor of $Z_m$.
Second, $m$ explicit gravitational boundary conditions come with a normalization of $(k\mathcal{N})^{-m}$ as described below \eqref{eq:resolvent-bc-pic}.
Also, dashed blue environment index lines come with a factor of $w^{-1}$, unless they have a resolvent insertion, in which case this factor is omitted since there is no factor of $w$ on the left hand side of \eqref{eq:resolvent-expansion}.
Because of this, from the form of \eqref{eq:SD-pic}, we see that every term on the right hand side is proportional to $w^{-1}$.
By our assumption (ii), the code subspace indices are contracted along the EOW branes, which means the $\mathfrak{R}_i$ indices of each resolvent insertion (except the last one on the right) in each term in \eqref{eq:SD-pic} are contracted together, yielding a trace $\Tr_{\mathfrak{R}_i} \mathcal{R}$.
Defining the $R \cup \text{env}$ density matrix $\sigma$ again as
\begin{equation}
\sigma = U_\mathcal{E} \left(\frac{1}{k}\sum_{\alpha=1}^k|\alpha\rangle\langle \alpha|_R \otimes |e_0\rangle\langle e_0|_{\text{env}}\right)U_\mathcal{E}^{\dagger} ,
\end{equation}
we see that because the resolvents are interspersed only on the environment factors, they multiply between the $\sigma$ factors to give $[\Tr_\text{env} (\sigma . \Tr_{\mathfrak{R}_i} \mathcal{R})]^{m-1}$. Here $m-1$ appears instead of $m$ since the last factor of $\sigma$ is generated by the leftmost $U_\mathcal{E}$ and the rightmost $U_\mathcal{E}^\dagger$, and is therefore not included in the environment traces generated by the adjacent EOW brane contractions, as can be seen from studying \eqref{eq:conn-EOW-structure} and \eqref{eq:f-tensor}.
This last factor of $\sigma$ instead has a radiation trace from the single large EOW brane index loop in $Z_m$.
Tacking on the last factor of $\mathcal{R}$ on the right, which multiplies the environment factor from the right, we finally find
\begin{equation}
\mathcal{R}
=\frac{1}{w}\mathbb{1}_{\mathfrak{R}_i,\text{env}}+ \frac{1}{w} \sum_{m=1}^{\infty}\frac{Z_m}{\mathcal{N}^m}\mathrm{Tr}_R\{\sigma. [\mathrm{Tr}_{\mathfrak{R}_i, \text{env}}(\sigma.\mathcal{R})]^{m-1}\}.\mathcal{R},
\label{eq:SD-eqn}
\end{equation}
The dot notation for products means that we multiply operators which may act on different Hilbert spaces by appending an identity operator to the extra factors.
Note the factor of $k^{-m}$ from the normalization was used to construct the $m$ factors of $\sigma$.
For a general quantum channel, solving \eqref{eq:SD-eqn} seems involved, although not impossible. But our restriction to the partial SWAP channel simplifies things quite a bit. For this channel, we have the factorization
\begin{equation}
\sigma = \sigma_R \otimes \sigma_\text{env}, \quad \sigma_\text{env} = \frac{1}{2^t}\mathbb{1}_\text{env}, \quad \sigma_R = \frac{1}{2^{M-t}}\mathbb{1}_{M-t}\otimes (|e_0\rangle\langle e_0|)^{\otimes t} ,
\end{equation}
so we can rewrite the Schwinger-Dyson equation \eqref{eq:SD-eqn} as
\begin{eqnarray}
w\mathcal{R} &=&\mathbb{1}_{\mathfrak{R}_i,\text{env}}+\sum_{m=1}^{\infty}\frac{Z_m}{\mathcal{N}^m}\mathrm{Tr}_R(\sigma_R^m) \left(\mathrm{Tr}_{\mathfrak{R}_i, \text{env}}(\sigma_\text{env}.\mathcal{R})\right)^{m-1} \sigma_\text{env}.\mathcal{R},\nonumber\\
&=&\mathbb{1}_{\mathfrak{R}_i,\text{env}}+\frac{1}{2^t}\sum_{m=1}^{\infty}\frac{Z_m}{\mathcal{N}^m2^{M(m-1)}} \left(\mathrm{Tr}_{\mathfrak{R}_i, \text{env}}(\mathcal{R})\right)^{m-1} \mathcal{R}
\end{eqnarray}
In fact, the first line also holds for any channel $U_\mathcal{E}$ which does not generate any mutual information when acting on the maximally mixed radiation state times the fiducial environment state $\ket{e_0}\bra{e_0}_\text{env}$; in other words when $I_{\sigma}(R : \text{env}) = 0$, as this implies $\sigma = \sigma_R \otimes \sigma_\text{env}$.
Note that this is a different mutual information from $I_{\Psi'}(\mathfrak{R}_i : \text{env})$, which we use to diagnose correctability of $U_\mathcal{E}$. To go further, we need expressions for the partition functions $Z_n$, which are found in \cite{Penington:2019kki}.
We have the integral expression
\begin{equation}
Z_n = \int_0^{\infty}ds\,\rho(s)[y(s)]^n,
\end{equation}
where
\begin{equation}
\rho(s) = \frac{s}{2\pi^2}\sinh(2\pi s),\;\;\;y(s)=e^{-\frac{\beta}{2}s^2}2^{1-2\mu}|\Gamma(\mu-\frac{1}{2}+is)|^2,
\end{equation}
after which the resolvent sum becomes geometric:
\begin{equation}
w\mathcal{R}=\mathbb{1}_{\mathfrak{R}_i,\text{env}}+\frac{1}{2^t}\int_0^{\infty}ds\frac{1}{\mathcal{N} }\rho(s)y(s)\frac{1}{1-\frac{y(s) \mathrm{Tr}_{\mathfrak{R}_i,\text{env}}(\mathcal{R})}{\mathcal{N} 2^{M}}}\mathcal{R} .
\end{equation}
The integral equation relating the density of states $\rho(s)$ to the JT path integral $Z_n$ is complicated due to the dependence of the integrand on the energy $s$.
By employing assumption (iii), we may work instead in the microcanonical ensemble where we define the gravitational path integral with boundary conditions that will fix the energy $s$ in a small window $\Delta s$.
This choice will change nothing in the above Schwinger-Dyson analysis except the explicit form of $Z_n$, which is now \textit{algebraically} related to a flat microcanonical density of states in the energy window $[s,s+\Delta s]$ with height $e^{S_0} \rho(s)$.
Specifically, we have the microcanonical entropy $\textbf{S}$ and microcanonical partition functions $\textbf{Z}_n$ defined by
\begin{equation}
e^{\textbf{S}} = e^{S_0}\rho(s) \Delta s , \quad \textbf{Z}_n = e^{S_0}\rho(s) y(s)^n \Delta s = e^{\textbf{S}} y(s)^n .
\end{equation}
These relations simplify the above integral equation as we replace $e^{S_0} Z_n$ by $\textbf{Z}_n$ and $\int \mathrm{d}s$ by $\Delta s$, and after taking a trace over $\mathfrak{R}_i \cup \text{env}$ the microcanonical Schwinger-Dyson equation for $\Tr \mathcal{R}$ is
\begin{figure}
\centering
\includegraphics[height=7.5cm]{MI1.pdf}
\caption{\small{The R\'enyi mutual information from $n=2$ to $n=5$ as a function of the number of qubits $t$ in the environment. We have taken $d_i=2^{10},\;e^{\mathcal{S}}=2^{100}$. We see that the phase transition happens precisely near $t\sim \frac{\mathcal{S}-\log\,d_i}{2 \log 2}= 45.$ The curves overlap substantially, except near the phase transition.}}
\label{fig:Renyi}
\end{figure}
\begin{equation}
w\,X= d_i2^t+\frac{1}{d_i2^t}\frac{X}{\left(1-\frac{ X}{d_ie^{\textbf{S}} 2^{M}}\right)}, \quad X=\mathrm{Tr}_{\mathfrak{R}_i,\text{env}}\;\mathcal{R}.
\end{equation}
This equation can be solved to obtain
\begin{equation}
\mathrm{Tr}_{\mathfrak{R}_i,\text{env}}\,\mathcal{R}^{(\mathfrak{R}_i,\text{env})} = \frac{d_ie^{\mathcal{S}}}{2}-\frac{\left(2^{-t}e^{\mathcal{S}}-d_i2^t\right)}{2w}-\frac{1}{2}\sqrt{\left(d_ie^{\mathcal{S}}-\frac{\left(2^{-t}e^{\mathcal{S}}-d_i2^t\right)}{w}\right)^2-4\frac{d_i^2e^{\mathcal{S}}2^t}{w}},
\end{equation}
where we have added a superscript $(\mathfrak{R}_i,\text{env})$ to $\mathcal{R}$ to denote the fact that this is the resolvent for the density matrix $\rho'_{\mathfrak{R}_i,\text{env}}$, and we have defined $e^{\mathcal{S}}=2^Me^{\textbf{S}}$. Similarly, we can also obtain the resolvent $\mathcal{R}^{(\text{env})}$ for $\rho'_{\text{env}}$:
\begin{equation}
\mathrm{Tr}_{\text{env}}\,\mathcal{R}^{(\text{env})} = \frac{d_ie^{\mathcal{S}}}{2}-\frac{\left(2^{-t}d_ie^{\mathcal{S}}-2^t\right)}{2w}-\frac{1}{2}\sqrt{\left(d_ie^{\mathcal{S}}-\frac{\left(2^{-t}d_ie^{\mathcal{S}}-2^t\right)}{w}\right)^2-4\frac{d_ie^{\mathcal{S}}2^t}{w}}.
\end{equation}
By expanding these resolvents in powers of $\frac{1}{w}$, we can compute all the R\'enyi entropies. We show a comparison of the R\'enyi mutual information as a function of $t$, the number of environment qubits, between several R\'enyi indices in Fig.~\ref{fig:Renyi}. From these plots we indeed see that $\mathbb{Z}_n$-symmetry breaking contributions only become important close to the phase transition and smooth out a relatively sharp transition between the fully disconnected and the fully connected geometries.
These two $\mathbb{Z}_n$-symmetric contributions were evaluated in general in Sec.~\ref{sec:qec-pssy} and together create a step function shape for the R\'enyi mutual information, so we interpret the near-step function shape of the full planar resummation as just a small correction to these contributions, at least for erasure errors.
Note also that the transition happens near, but slightly before
\begin{equation}
2t \sim \frac{\mathcal{S}}{\log 2} - \log_2 d_i ,
\label{eq:gravity-singleton}
\end{equation}
where $\mathcal{S}$ is the logarithm of the total physical Hilbert space dimension (microcanonical JT ensemble with entropy $\textbf{S}$ plus the radiation bath with entropy $M \log 2$) and $t \log 2$ is the logarithm of the environment Hilbert space dimension.
This is precisely the upper limit from the quantum Singleton bound \eqref{eq:singleton-bound} on the number of qubits that can be erased while leaving the code subspace protected.
More specifically, in \eqref{eq:singleton-bound} we called $\mathcal{S}$ the number of qubits in the physical Hilbert space, so we must replace it here with $\mathcal{S} / \log 2$, and we had a base-2 logarithm of the code subspace dimension, both of which appear in \eqref{eq:gravity-singleton}.
The R\'enyi mutual information starts to become $O(1)$ (as opposed to being exponentially small) when $t$ is only an $O(1)$ number less than the greatest number of qubit erasures that any code can protect against!
Thus, the JT gravity model comes within an $O(1)$ number of qubits of saturating the quantum Singleton bound.
It would be interesting to characterize the very slightly suboptimal performance of gravity; perhaps non-planar contributions are important to account for in this analysis.
\subsection{Typical errors}
\label{sec:randomerrors}
In this section we will consider how well the black hole interior is protected against the action of typical random errors on the radiation bath.
In more detail, suppose that the Stinespring dilation of $\mathcal{E}$ involves a unitary operator $U_\mathcal{E}$ acting on a Hilbert space of dimension $\dim (\mathcal{H}_R \otimes \mathcal{H}_\text{env}) = k\ell$, where $\ell$ is the dimension of the environment factor and $k$ controls the dimension of the radiation subspace that is entangled with the black hole microstates.
Then, by a typical/random error, we mean one where we can take $U_\mathcal{E}$ to be a random unitary transformation drawn from the Haar measure on the group $U(k\ell)$.
We will perform this random error calculation in two ways.
First, we will directly Haar average the $f$-tensor \eqref{eq:f-tensor} and re-compute the mutual information $I_{\Psi'}(\mathfrak{R}_i : \mathfrak{R}_e \cup \text{env})$.
Second, we will take the Haar average of the coherent information $I_c(\mathcal{E},\mathbb{1}_R/k)$ and compare its value with our correctability criterion \eqref{CIcond}.
A priori, these two approaches need not give the same result, because averaging the $f$-tensor first may affect which gravitational geometries dominate the path integral, whereas in averaging the coherent information we have assumed that the replica symmetric contributions are the relevant ones.
However, we will see that an interesting simplification occurs when we consider the two limits $1 \ll k \ll \ell$ and $1 \ll \ell \ll k$, and the two approaches in fact agree.
\subsubsection{$f$-Tensor}
\label{sec:f-tensor-haar}
We begin by directly taking the Haar average of \eqref{eq:f-tensor} and evaluating the resulting gravitational path integrals which appear in the mutual information (\ref{eq:renyimutualinformation}) between the interior and the union of the exterior and the environment.
This in turn involves the R\'enyi entropies of various combinations of the black hole interior, exterior and the environment (see Eqs.~\ref{tr1},\ref{tr2},\ref{tr3}).
We will first average the error over the unitary group and then carry out the gravitational path integral computing the microstate overlaps.
As we are most interested in the regimes where the environment is either large ($\ell \gg k \gg 1$) or small ($1 \ll \ell \ll k$) compared to the radiation Hilbert space with both large compared to 1, we will only retain the leading terms in $1/k$ or $1/\ell$.
The Haar average of the $f$-tensor is then:
\begin{align}
\begin{split}
\bar{f}_{\alpha_1,\beta_1,\cdots \alpha_n,\beta_n}
&=\int_{U(k\ell)} dU_\mathcal{E} f_{\alpha_1,\beta_1,\cdots \alpha_n,\beta_n}(U_\mathcal{E}) \\
& = \sum_{m_1,\cdots, m_n}\sum_{\gamma_1,\cdots, \gamma_n} \int_{U(k\ell)} dU_\mathcal{E} (U^{\dagger})^{\alpha_1,e_0}_{\gamma_1,e_{m_2}} U^{\gamma_1,e_{m_1}}_{\beta_1,e_0} \cdots (U^{\dagger})^{\alpha_n,e_0}_{\gamma_n, e_{m_1}} U^{\gamma_n,e_{m_n}}_{\beta_n,e_0}
\label{eq:HaarAverageU}
\\
& = (\ell^{1-n} + \cdots) (\delta_{\alpha_1\beta_1} \delta_{\alpha_2\beta_2} \ldots \delta_{\alpha_n\beta_n}) + (k^{1-n} + \cdots ) (\delta_{\alpha_1\beta_2} \delta_{\alpha_2\beta_3} \ldots \delta_{\alpha_n\beta_1} ) + \cdots ,
\end{split}
\end{align}
where we have defined the matrix elements
\begin{equation}
U^{\beta,e_m}_{\alpha,e_n} \equiv \bra{\beta,e_m}U_\mathcal{E} \ket{\alpha,e_n} , \quad (U^\dagger)^{\beta,e_m}_{\alpha,e_n} \equiv \bra{\beta,e_m}U_\mathcal{E}^\dagger \ket{\alpha,e_n} .
\end{equation}
We compute the integrals in the second line in Appendix~\ref{sec:haar} by applying standard results concerning Haar integration of products of unitary matrices.
The first term dominates when $1 \ll \ell \ll k$, while the second term dominates when $1 \ll k \ll \ell$.
Remarkably, the first term in (\ref{eq:HaarAverageU}) has the same microstate index structure as the product of disconnected path integrals (\ref{eq:DiscOverlap}) for the gravitational overlaps in the R\'enyi mutual informations (\ref{tr1}), (\ref{tr2}), and (\ref{tr3}). The second term has the same microstate index structure as the fully connected diagrams for computing the same product of overlaps, shown for instance in (Fig.~\ref{fig:conncode2} and Eq.~\ref{eq:conn-EOW-structure}). This means that averaging over $U_{\mathcal{E}}$, at least in the two limits $1 \ll \ell \ll k$ and $1 \ll k \ll \ell$, reinforces the two replica-symmetric, disconnected and connected geometries respectively in the gravitational path integral. Due to this simplification, we see that the relevant gravitational geometries that will enter in the calculation of the R\'enyi mutual information $I_{\Psi'}^{(n)}(\mathfrak{R}_i : \mathfrak{R}_e \cup \text{env})$ are precisely the ones we considered already in Sec.~\ref{sec:qec-pssy}. In the $n\to 1^+$ limit, the overlaps in (\ref{tr1}), (\ref{tr2}), and (\ref{tr3}) involve different propagators for the matter fields and thus lead to different powers of $d_i$ and $d_e$ (see discussion below Eq.~\ref{eq:conn-EOW-structure}), but in a manner identical to Sec..~\ref{sec:qec-pssy}.
Therefore, the resulting mutual information will be identical, as we will now confirm explicitly.
Using our result \eqref{eq:HaarAverageU} for the Haar-averaged $f$ tensor in (\ref{tr1}), (\ref{tr2}), and (\ref{tr3}), along with the gravitational overlaps from the disconnected saddle-points (\ref{eq:DiscOverlap}), we can find the leading disconnected contribution to the R\'enyi entropies.
The contraction of $\bar{f}$ with the tensor on the radiation Hilbert space coming from the gravitational overlaps gives
\begin{equation}
\frac{1}{k^n}
\sum_{\alpha_1...\alpha_n}
\sum_{\beta_1...\beta_n}
\left(
\delta_{\alpha_1\beta_1} \delta_{\alpha_2\beta_2} \cdots \delta_{\alpha_n\beta_n}
\right) \bar{f}_{\alpha_1\beta_1 \cdots \alpha_n\beta_n}
= \ell^{1-n} + k^{2(1-n)} + \cdots
\label{eq: g-f-contraction1}
\end{equation}
Carrying out the remaining sums over the internal and external code space indices gives
\begin{eqnarray}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}
&=&
d_i^{1-n} d_e^{1-n} \,
\left[
\ell^{1-n} + k^{2(1-n)}
+ \cdots
\right]
\label{eq:TrIEDisc}
\\
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}
&=&
d_e^{1-n} \,
\left[
\ell^{1-n} + k^{2(1-n)}
\right]
\label{eq:TrEDisc}
\\
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}
&=&
d_i^{1-n} \, .
\label{eq:TriDisc}
\end{eqnarray}
The terms in the square brackets give an explicit evaluation of the $\Tr{\sigma_{\rm env}^n}$ factors in (\ref{eq:disc-result1})-(\ref{eq:disc-result3}) for the disconnected gravitational contribution to a general error. We can now evaluate the $n$-th R\'enyi mutual information (\ref{eq:renyimutualinformation}) between the interior of the black hole and the exterior of the black hole plus the environment in the disconnected phase of the gravitational path integral. We find
\begin{equation}
I^{(n)}(\mathfrak{R}_i: \mathfrak{R}_e \cup \text{env}) =
\frac{1}{1-n} \left( \log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i} +
\log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}
-
\log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}
\right)
=
0 \, ,
\end{equation}
confirming (\ref{eq:disconnMI}) for random error channels.
Similarly, we can combine the Haar-averaged f-tensor with the gravitational overlaps from the connected saddlepoints in \eqref{eq:conn-EOW-structure}, along with its analogs for the other mutual informations, to computed the leading connected contribution to the R\'enyi entropies (\ref{tr1}), (\ref{tr2}), and (\ref{tr3}). The contraction of $\bar{f}$ with the tensor on the radiation Hilbert space coming from the connected gravitational overlaps gives
\begin{equation}
\frac{1}{k^n}
\sum_{\alpha_1...\alpha_n}
\sum_{\beta_1...\beta_n}
\left(\delta_{\alpha_1\beta_2} \delta_{\alpha_2\beta_3} \cdots \delta_{\alpha_n\beta_1} \right) \bar{f}_{\alpha_1\beta_1 \cdots \alpha_n\beta_n}
=
(k\ell)^{1-n} + k^{1-n} + \cdots
\label{eq: g-f-contraction2}
\end{equation}
Carrying out the remaining sums over the internal and external code space indices (in the $n\to 1^+$ limit) gives
\begin{eqnarray}
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}
&=& \frac{e^{S_0 (1-n)} Z_n}{Z_1^n}
\, d_e^{1-n} \,
\left[
(k \ell)^{1-n} + k^{1-n}
\cdots \right]
\label{eq:TrIEConn}
\\
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}
&=&\frac{e^{S_0 (1-n)} \, Z_n}{Z_1^n} \, d_i^{1-n} d_e^{1-n} \, \left[
(k \ell)^{1-n} + k^{1-n}
+
\cdots \right]
\label{eq:TrEConn}
\\
\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}
&=&
\frac{e^{S_0 (1-n)} Z_n}{Z_1^n} \, k^{1-2n} \ell^{-n}
\label{eq:TriConn}
\end{eqnarray}
The quantities $S_0$, $Z_n$ and $Z_1$ are discussed below (\ref{eq:conn-EOW-structure}). The terms in the square brackets give an explicit evaluation of the $\Tr{\sigma_{\rm R}^n}$ factors in (\ref{eq:re-env-conn1})-(\ref{eq:re-env-conn}) for the connected gravitational contribution to a general error.
We can now evaluate the R\'enyi mutual information (\ref{eq:renyimutualinformation}) between the interior and exterior plus the environment in the connected phase of the gravitational path integral in the $n\to 1$ limit. In evaluating this mutual information we will continue to use the disconnected expression for the $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}$ (\ref{eq:TriDisc}) because it is dominant for any choice of the error rank $\ell$ as the connected contribution (\ref{eq:TriConn}) is suppressed in powers of $k$. We get
\begin{equation}
\begin{split}
I(\mathfrak{R}_i: \mathfrak{R}_e \cup \text{env}) =
\lim_{n\to 1} \frac{1}{1-n} \left( \log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i} +
\log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}
-
\log \mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}
\right)
= 2 \log d_i
\end{split}
\end{equation}
confirming the general result (\ref{eq:connMI}) for random error channels.
Finally we can check the condition for the connected phase to dominate the R\'enyi entropies and hence the mutual information between the black hole interior and the exterior plus the environment. As discussed above, $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i}$ is always dominated by the disconnected component because the connected part (\ref{eq:TriConn}) is suppressed. A little algebra shows that the connected contribution to $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_e,\text{env}}$ in (\ref{eq:TrEConn}) dominates the disconnected contribution in (\ref{eq:TrEDisc}) when
\begin{equation}
\frac{\left(\frac{\ell}{k} \right)^{n-1} + k^{n-1}}{1 + \ell^{n-1}}
\approx
\left( \frac{k}{\ell} \right)^{n-1}
\ll
e^{(1-n)( S_{BH} + \log d_i )}
\end{equation}
where $e^{S_0(1-n)} Z_n / Z_1^n \sim e^{(1-n) S_{BH}}$ in the $n\to 1^+$ limit, as discussed below (\ref{eq:re-env-conn}). The approximate expression is valid when $\ell \sim O(k^2)$ or smaller. Similarly the connected contribution to $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$ in (\ref{eq:TrIEConn}) dominates the disconnected contribution in (\ref{eq:TrIEDisc}) when
\begin{equation}
\frac{\left(\frac{\ell}{k} \right)^{n-1} + k^{n-1}}{1 + \ell^{n-1}}
\approx
\left( \frac{k}{\ell} \right)^{n-1}
\ll
e^{(1-n)( S_{BH} - \log d_i )}
\label{eq:secondcondition}
\end{equation}
The approximate expression is again valid when $\ell \sim O(k^2)$ or smaller. This second condition is satisfied for smaller $\ell$ than the first condition.
It can be checked that the mutual information is nonzero if the connected phase dominates just $\mathrm{Tr}\,{\rho'}^n_{\mathfrak{R}_i,\mathfrak{R}_e,\text{env}}$. The condition (\ref{eq:secondcondition}) is satisfied when
\begin{equation}
\ell \gg \frac{k}{d_i} e^{S_{BH}} \gg k
\label{eq:thirdcondition}
\end{equation}
giving a precise condition on the rank of the error channel required to give an error that cannot be corrected.
\subsubsection{Coherent information criterion}
\label{sec:CIcond-haar}
Alternatively we can diagnose correctability of typical errors by directly computing the Haar average of the coherent information criterion in (\ref{eq:coherentinfo}).
A priori, this is a separate calculation from the one we discussed above, where we took the Haar average of the $f$-tensor before computing the gravitational path integral.
Here, we will take the general correctability criterion, which in particular is nonlinear in the error channel $U_\mathcal{E}$, and perform the Haar average.
However, in the case of the $f$-tensor average, because we found that the microstate index structures precisely matched the two relevant ones coming from the disconnected and fully connected gravitational path integral geometries, we ought to obtain the same answer by simply averaging the coherent information.
As we will now see, this is indeed what occurs, and we obtain the typical error condition \eqref{eq:secondcondition} in a more straightforward manner, without discussing gravitational path integrals.
The coherent information is given by the difference of two entropies, $S(\sigma_R)$ and $S(\sigma_\text{env})$, where we recall that $\sigma_R$ and $\sigma_\text{env}$ are partial traces of $\sigma$
\begin{equation}
\sigma = U_\mathcal{E} \left( \frac{1}{k} \mathbb{1}_R \otimes \ket{e_0}\bra{e_0}_\text{env} \right) U_\mathcal{E}^\dagger .
\end{equation}
over the complementary factor (see (\ref{eq:sigmaenv}) and (\ref{eq:sigmaR})).
Choosing explicit bases $\ket{\alpha}_R$ with $\alpha = 1, \dots , k$ and $\ket{e_m}_\text{env}$ with $m = 1, \dots , \ell$ for the radiation and environment respectively, the reduced density matrices are
\begin{equation}
\begin{split}
\sigma_R & = \frac{1}{k} \sum_{\alpha,m} \bra{e_m} U_\mathcal{E} \ket{\alpha, e_0} \bra{\alpha, e_0} U_\mathcal{E}^\dagger \ket{e_m} , \\
\sigma_\text{env} & = \frac{1}{k} \sum_{\alpha,\beta} \bra{\beta} U_\mathcal{E} \ket{\alpha, e_0} \bra{\alpha, e_0} U_\mathcal{E}^\dagger \ket{\beta} .
\end{split}
\end{equation}
We can compute the Haar averaged entropies by using the replica trick just as we did with the gravitational path integral.
Note that here we are not actually performing any gravitational path integrals; we are simply using the replica expressions for entropies and computing their Haar averages in order to find the typical value of the coherent information for a random channel.
The $n^{\text{th}}$ R\'enyi coherent information with respect to the maximally mixed state is given in terms of the $n^{\text{th}}$ R\'enyi entropies by
\begin{equation}
I^{(n)}_c \left( \mathcal{E}, \frac{1}{k} \mathbb{1}_R \right) = S^{(n)}(\sigma_R) - S^{(n)}(\sigma_\text{env}) .
\end{equation}
The necessary traces of powers of the above density matrices are given by
\begin{equation}
\begin{split}
\Tr \sigma_R^n & = \frac{1}{k^n} \sum_{\alpha_1, \cdots , \alpha_n} \sum_{\beta_1, \cdots , \beta_n} \sum_{m_1, \cdots , m_n} U^{\beta_1,e_{m_1}}_{\alpha_1,e_0} (U^\dagger)^{\alpha_1,e_0}_{\beta_2, e_{m_1}} \cdots U^{\beta_n,e_{m_n}}_{\alpha_n,e_0} (U^\dagger)^{\alpha_n,e_0}_{\beta_1, e_{m_n}} , \\
\Tr \sigma_\text{env}^n & = \frac{1}{k^n} \sum_{\alpha_1, \cdots , \alpha_n} \sum_{\beta_1,\cdots \beta_n} \sum_{m_1, \cdots , m_n} U^{\beta_1,e_{m_1}}_{\alpha_1,e_0} (U^\dagger)^{\alpha_1,e_0}_{\beta_1,e_{m_2}} \cdots U^{\beta_n,e_{m_n}}_{\alpha_n,e_0} (U^\dagger)^{\alpha_n,e_0}_{\beta_n,e_{m_1}} ,
\end{split}
\end{equation}
The Haar averages $\overline{\Tr \sigma_R^n}$ and $\overline{\Tr \sigma_\text{env}^n}$ can be evaluated in terms of Weingarten functions and combinations of Kronecker delta functions, as explained in Appendix~\ref{sec:haar}.
The results are
\begin{equation}
\begin{split}
\overline{\Tr \sigma_R^n} & = \frac{1}{k^{n-1}} , \\
\overline{\Tr \sigma_\text{env}^n} & = \frac{1}{\ell^{n-1}} .
\end{split}
\end{equation}
This is sensible, as it suggests that a random unitary operator will (at leading order in $k$ and $\ell$) create maximally mixed reduced states on the radiation and environment factors.
Using the standard formula for the entropy in terms of these traces as described above \eqref{eq:entropy-diff}, we find
\begin{equation}
\begin{split}
S(\sigma_R) & = \log k , \\
S(\sigma_\text{env}) & = \log \ell .
\end{split}
\end{equation}
Finally, the coherent information (\ref{eq:coherentinfo}) for a random error of fixed Kraus rank $\ell$ acting on a radiation space of dimension $k$ is
\begin{equation}
I_c \left( \mathcal{E}, \frac{1}{k}\mathbb{1}_R \right) = S(\sigma_R) - S(\sigma_\text{env}) = \log \frac{k}{\ell} .
\end{equation}
According to this result, a random error must be of sufficient Kraus rank in order to affect the interior.
In particular, utilizing \eqref{CIcond}, if
\begin{equation}
\log \frac{k}{\ell} \ll -(S_{BH} - \log d_i)
\end{equation}
then the error $\mathcal{E}$ on the interior is not quantum correctable based on our criterion on the coherent information.
This result is equivalent to \eqref{eq:thirdcondition}, as advertised.
If the Kraus rank of the randomizing error is scaling with a fractional power of the radiation dimension $k$, for instance $\ell \sim \sqrt{k}$, this shows that a typical random error will actually not affect the interior because the ratio $k/\ell$ will be large, leading to a positive coherent information.
This is surprising because after the Page time we have $k \gg e^{S_0}$, and a random error with $\ell = \sqrt{k}$ is modeled by a Stinespring dilation $U_\mathcal{E}$ that will almost certainly have exponential quantum complexity in $S_0$.
Our result shows that such exponential complexity errors cannot affect the interior.
It also shows that the erasure/SWAP example we discussed previously is somehow finely tuned, because the SWAP Stinespring dilation only requires an environment of dimension $\ell = k$, and for this relation $\log (k/\ell) = 0$.
Of course, for $\ell = k$ we are certainly missing contributions in our Haar integrals as we assumed either $\ell \gg k$ or $k \gg \ell$, so perhaps these terms provide large corrections to the coherent information for a typical error with $\ell = k$ that brings them closer to the SWAP error.
On the other hand, when the Kraus rank of the error is significantly greater than the radiation dimension, it is quite easy to affect the interior after the Page time $k \gg e^{S_0}$, as e.g. we may take $\ell = k^2$ maximal \cite{NielsenChuang} and then we have the condition $\log \frac{1}{k} \ll -S_0$ which is nothing but the statement that we have passed the Page time.
\section{Discussion}\label{sec:disc}
We used a toy model for black hole evaporation in JT gravity to analyze novel quantum error correction properties of the black hole interior. Similar properties were previously proposed by Kim, Preskill and Tang on the basis of a pseudorandom encoding model for Hawking radiation \cite{Kim:2020cds}, while we investigated these properties in a simple gravitational model.
Modeling errors in the encoded state as quantum channels $\mathcal{E}$, we found that any error is correctable provided (i) the channel does not have access to black hole microstate details (or more precisely, the encoding map for the code subspace), and (ii) its coherent information $I_c$ with respect to the maximally mixed radiation state is greater than a certain (negative) lower bound set by the black hole entropy. Thus, while the black hole is macroscopic,
its interior
is robustly protected against generic, low-rank errors (i.e. errors whose rank is smaller than the size of the radiation Hilbert space) on the radiation.
The breakdown of correctability upon violating the coherent information bound happens due to the dominance of the replica wormhole geometry in the gravitational path integral, which is the same effect that leads to restoration of the Page curve in entropy calculations.
\subsection{Robust QEC in the Python's Lunch}
In this work, we have focused on the error correction properties of the black hole interior, viewed past the Page time as being encoded in the radiation bath. It is interesting to ask how the robust error correction properties of the black hole interior that we have found may generalize to other entanglement wedges. Recent work \cite{Brown:2019rox, Engelhardt:2021mue, Engelhardt:2021qjs} has highlighted the rich structure of entanglement wedges in AdS/CFT. Let $A$ be a subregion, or, more precisely, the domain of dependence $D(A)$ of a subregion $A$, in the boundary CFT and let ${\overline{A}}$ be its complement. We can divide the entanglement wedge $\mathcal{W}(A)$ of $A$ in the dual geometry into two portions: (i) the \emph{simple wedge}, i.e., the portion of $\mathcal{W}(A)$ which lies between $A$ and the outermost quantum extremal surface homologous to $A$, and (ii) the \emph{Python's Lunch}, i.e., the complement of the simple wedge inside $\mathcal{W}(A)$. The simple wedge can further be divided into the \emph{causal wedge}, i.e., the portion of the simple wedge which is in causal contact with $D(A)$, and its complement inside the simple wedge. It is expected that the encoding map for semi-classical degrees of freedom in the Python's Lunch portion of an entanglement wedge will be exponentially complex, while the encoding map will not be complex in the simple wedge. Indeed, the black hole interior past the Page time is an example of a Python's Lunch, as it is hidden behind a non-minimal QES, namely the empty surface.
It is natural to conjecture that in more general entanglement wedges, \emph{the encoding of semi-classical degrees of freedom in the Python's Lunch should be robust against generic, low-rank errors on the boundary subregion $A$.}\footnote{This conjecture also makes precise the idea that the holographic encoding of black hole microstates in the AdS/CFT correspondence is exceedingly complex, being essentially random, and is therefore difficult to decode without access to operations that have direct knowledge of the microstates \cite{Balasubramanian:2005mg}.}
By ``generic'', we mean quantum operations which do not have prior access to the details of the microscopic encoding map.\footnote{One way to make this more concrete is by considering typical errors, namely errors whose Stinespring dilation consists of a typical unitary from the Haar ensemble.} In contrast, we do not expect robust quantum error correction in the simple wedge. Indeed, it is not difficult to construct examples of generic, low-rank quantum operations which corrupt the encoding of degrees of freedom in the simple wedge. For instance, in the causal wedge, even a one-qubit environment interacting with $A$ via coupling to some single-trace operator is sufficient. Further, the entire simple wedge can be brought in causal contact with $D(A)$ by a relatively simple unitary operation on $A$. Our conjecture can be analysed in simplified models of holography, such as random tensor networks or fixed area states, as we will report elsewhere \cite{coming_soon}. It would also be interesting to understand the implications of our results in cosmological settings \cite{Langhoff:2021uct,Bousso:2022gth}, de Sitter universes \cite{Shaghoulian:2022fop}, and for disjoint gravitating universes that share entanglement between their degrees of freedom \cite{Balasubramanian:2020coy,Anderson:2020vwi, Balasubramanian:2021wgd}.
\subsection{Pseudorandomness and complexity}
Our analysis of QEC in the black hole interior has relied entirely on the gravitational path integral, and provides a different perspective on the results of \cite{Kim:2020cds}, where the authors relied on quantum information theory arguments together with an assumption on the pseudorandomness of the radiation density matrix. More precisely, the authors of \cite{Kim:2020cds} assumed that any low-complexity (i.e., polynomial in the black hole entropy) quantum operation could not reliably distinguish the radiation density matrix from the maximally mixed state. This led them to conclude that the black hole interior was robust against low-complexity errors on the bath. On the other hand, our analysis relied upon a different key assumption: the action of the error channel $\mathcal{E}$ should be independent of gravitational overlaps between black hole microstates $\inner{\psi_{i,i'}^\alpha}{\psi_{j,j'}^\beta}_B$.
If this assumption were violated, our calculation in Sec.~\ref{sec:qec-pssy} would have to be modified because we would potentially need to include gravitational path integrals on additional manifolds in order compute the amplitudes associated to the action of $\mathcal{E}$.
We are thus led to conclude that the interior is robust against (low-rank) errors which do not have prior access to details of black hole microstates (or more generally, the encoding map for the code). It would be nice to further explore the connection between these two approaches.
It is worth noting that in our analysis, the maximally mixed state on the radiation arises naturally as the state relative to which the coherent information of $\mathcal{E}$ should be computed in order to determine whether or not the error is correctable. This is an output of the gravitational path integral in the specific toy model we are considering. We can interpret this maximally mixed state loosely as the semi-classical state of the radiation, analogous to Hawking's density matrix, i.e., the coarse-grained, random approximation to the true, pseudorandom density matrix on the radiation.
One curious aspect which deserves further exploration, given this observation, is the following.
In the toy model, the gravitational path integral is really computing an ensemble average $\mathbb{E}[\cdot]$ over many boundary theories \cite{Penington:2019kki}.
The maximally mixed radiation state which appears in our criterion happens to also be the ensemble average of the microscopic radiation states in these theories, because these states depend on gravitational overlaps that will vary in each individual theory in the ensemble.
So, we may write our criterion for interior state corruption as
\begin{equation}
I_c(\mathcal{E}, \mathbb{E}[\rho_R] ) \ll -(S_{BH} - \log d_i ) .
\end{equation}
Though a complete generalization of our criterion to include cases where $\mathcal{E}$ has access to microstate details was beyond the scope of our work, we may conjecture that a sufficient criterion for interior state corruption that includes cases where $\mathcal{E}$ has access to microstate details should simply move the ensemble average outside to reflect the fact that $\mathcal{E}$ also has nontrivial ensemble dependence.
\begin{equation}
\mathbb{E} [I_c(\mathcal{E} , \rho_R)] \ll -(S_{BH} - \log d_i) \quad \xRightarrow[]{?} \quad \rho_\text{int} \text{ is not recoverable from } \mathcal{E}(\rho_R) .
\label{eq:conjecture}
\end{equation}
Note that in the above conjecture it is crucial that the implication is only to the right.
We are not claiming that the left hand side is a necessary criterion for interior state corruption; we only mean that it may be a sufficient criterion in the presence of a channel $\mathcal{E}$ which depends on the microstate details.
This contrasts with the criterion \eqref{CIcond} derived in Sec.~\ref{sec:qec-pssy}, which was both necessary and sufficient for interior state corruption due to our no-prior-access assumption.
Indeed, the converse of \eqref{eq:conjecture} can be explicitly violated by a simple channel that couples a Petz operator on the radiation to a single environment qubit.
Understanding \eqref{eq:conjecture} in greater detail could shed light on properties of the microscopic radiation state, connecting with the work of \cite{Kim:2020cds} and potentially bringing complexity theory into the discussion.
Indeed, a simple counterexample to our no-prior-access-to-black-hole-microstate-overlaps assumption would be a channel which uses the Petz reconstruction of bulk operators on the radiation; however it seems plausible that such channels are exponentially complex \cite{Zhao:2020wgp,Gilyen:2020gmg}, and therefore also violate the assumptions of \cite{Kim:2020cds}.
Analyzing such channels which are allowed to access the microscopic state details in a controlled manner provides an avenue for connecting our assumptions with the pseudorandomness ideas of \cite{Kim:2020cds}.
In this language, the interesting question to ask is: what properties of the channel $\mathcal{E}$ with access to microstate details allow it to corrupt the interior while violating the left hand side of \eqref{eq:conjecture}?
It would be good to make a more precise connection between the pseudorandomness assumption of KPT and the assumptions which have gone into the present analysis, perhaps through an analysis of channels which are allowed to access the microscopic state details, as our setup allows us to relax the no-prior-access assumption in a controlled manner, while relaxing the pseudorandom assumption of KPT seems difficult.
\subsection{Implications for semi-classical causality}
We found that an error channel acting on the radiation can affect the black hole interior if it has a sufficiently negative coherent information. This situation can arise when the Euclidean gravity path integral associated to the error channel is dominated by wormhole contributions. Some of the expected properties of effective field theories can be challenged by the presence of such wormholes \cite{Geng:2021hlu}.
We focus here on semi-classical causality in the black hole spacetime, which states that an interior EFT operator $\phi$ should commute with EFT operators $\mathcal{Q}_R$ acting on the radiation:
\begin{equation}
[\phi, \mathcal{Q}_R] = 0 .
\label{eq:microcausal}
\end{equation}
This expression, much like the notion of bulk radial locality in AdS/CFT \cite{Harlow:2018fse}, ought to hold at least in a code subspace of the physical Hilbert space.
However, after the Page time, the interior is located within the entanglement wedge of $R$, and therefore we may reconstruct logical operators such as $\phi$ using operators which act only on $R$ in the physical Hilbert space.
So, it is not obvious whether microcausality in the sense of \eqref{eq:microcausal} will still be respected, even in the code subspace.
To check this, we must compute the commutator in the physical Hilbert space, using a global representation of $\phi$.
By global representation, we mean a representation of $\phi$ which acts on the physical Hilbert space $\mathcal{H}_B \otimes \mathcal{H}_R$.
In place of the radiation operator $\mathcal{Q}_R$, we will use the Kraus operators $E_n$, which model operations we can perform on the Hawking quanta in the radiation.
Correctability of the set of errors $\mathcal{E}=\{E_m\}$ has an important implication for this issue, namely consistency with the semi-classical causal structure of the black hole spacetime \cite{Kim:2020cds}. Let $\phi$ be a bulk operator in the code subspace localized in the black hole interior,
\begin{equation}
\phi = \sum_{k'=1}^{d_e} \sum_{i,j=1}^{d_i} \phi_{ij} \ket{i,k'}\bra{j,k'}_{\text{int},\text{ext}} \, .
\end{equation}
This is localized on the interior because we have traced over the exterior factor of the Hilbert space; in other words $\phi$ acts as the identity on $\mathcal{H}_\text{ext}$.
Recalling that the action of the encoding map is $V\ket{i,i'}_\text{bulk EFT} = \ket{\Psi_{i,i'}}_{B,R}$, let $\mathcal{O}$, the ``global'' representation of $\phi$ on the physical Hilbert space $\mathcal{H}_\text{phys} = \mathcal{H}_B \otimes \mathcal{H}_R$, be
\begin{equation}\label{GR}
\mathcal{O} = V\phi V^{\dagger} = \sum_{k'} \sum_{i,j}\phi_{ij}|\Psi_{i,k'}\rangle\langle \Psi_{j,k'}|_{B,R}.
\end{equation}
By global representation we simply mean the explicit lift of $\phi$ to an operator $\mathcal{O}$ on $\mathcal{H}_\text{phys}$ via conjugation by the encoding map $V$.
The exact correctability of $\mathcal{E}$ is equivalent to the Knill-Laflamme conditions \eqref{KLcond1}:
\begin{equation}
P E_m^{\dagger}E_n P = V( \mathbb{1}^{(\text{int})}\otimes B^{(\text{ext})}_{mn})V^\dagger,\;\;\;\forall m,n,
\label{eq:KLcond2}
\end{equation}
where the operators $B^{(\text{ext})}_{mn}$ are analogous to those appearing in \eqref{KLcond1} and $P = \sum_{i,i'}
\ket{\Psi_{i,i'}}\bra{\Psi_{i,i'}}_{B,R}$ is a projector onto the code subspace. Since $\mathcal{O}$ is localized on the interior factor on the code subspace, the isometry $V$ by definition obeys $V^\dagger V = \mathbb{1}_{\text{int},\text{ext}}$, and the combination $PE_m^\dagger E_n P$ acts as the identity on $\mathcal{H}_\text{int}$ after conjugation by $V$ \eqref{eq:KLcond2}. So we find
\begin{equation}
\left[\mathcal{O}, P E_m^{\dagger}E_n P\right] =0,\;\;\;\forall m,n \, .
\end{equation}
This is an operator algebra restatement of the error correction conditions \cite{2007PhRvA..76d2303B}. From equation \eqref{GR}, we also observe that $\mathcal{O}$ commutes with the code subspace projector $P= VV^\dagger$ and so this condition translates to
\begin{equation}
P\left[\mathcal{O}, E_m^{\dagger}E_n \right] P=0,\;\;\;\forall m,n .
\end{equation}
Suppose the class of errors $\mathcal{E}=\{E_m\}$ includes an operator proportional to the identity, and, if not, extend it with such an operator. Now require correctability under the extended set \cite{Kim:2020cds}). Then the Knill-Laflamme conditions immediately imply
\begin{equation}
P\left[\mathcal{O}, E_n \right] P=P\left[\mathcal{O}, E_n^{\dagger} \right] P= 0,\;\;\;\forall n,
\label{eq:code-subspace-causality}
\end{equation}
and so the Kraus operators respect the semi-classical causal structure of the black hole spacetime within the code subspace.
Thus, any correctable quantum operation $\mathcal{E}$ on the radiation bath will respect semi-classical causality with respect to operators in the black hole interior. For a correctable quantum operation to violate semi-classical causality, it therefore seems to need prior access to black hole microstate details through the overlaps $\inner{\psi_{i,i'}^\alpha}{\psi_{j,j'}^\beta}$, which would modify the results we obtained in Sec.~\ref{sec:qec-pssy}.
On the other hand, the failure of correctability due to a large Kraus rank for $\mathcal{E}$ \emph{does not} necessarily imply a violation of causality, for example if the information is simply moved to another physical environment.
In that case the original representation of $\mathcal{O}$ on $B\cup R$, the union of of the quantum system dual to the black hole and the radiation bath, fails to reconstruct the bulk operator in the interior. One must instead construct interior operators on the joint system $B\cup R\cup \text{env}$.
The global representation of $\phi$ need not commute with errors that are uncorrectable, as our derivation above concerning semi-classical causality relied on the Knill-Laflamme conditions for correctability.
A way out of this apparent violation of causality may be that the global representation of $\phi$ is simply inaccessible to any reasonable observer living in the radiation. In that case, violation of causality for observers in the radiation would not be diagnosed by non-commutation of the Kraus operators with the global representation. Instead causality would be violated for such observers if the Kraus operators did not commute with a representation of $\phi$ which lives purely on the radiation Hilbert space.
We will now demonstrate that, at least for one particular such representation, semi-classical causality is indeed never violated.
An explicit example of such a representation on $R$ alone is the Petz reconstruction
\begin{equation}
\mathcal{O}_\text{Petz} = \mathcal{E}(P)^{-1/2}\mathcal{E}(V\phi V^{\dagger})\,\mathcal{E}(P)^{1/2} .
\end{equation}
This operator makes explicit use of the quantum error channel acting exclusively on the radiation. If there is no recovery channel, i.e., if the error cannot be corrected, then the Petz operator will vanish \cite{Penington:2019kki}.
This is the case, for example, before the Page time.
Thus, this operator respects causality, because it vanishes precisely when the channel is not correctable. It is possible for this to happen because operator makes explicit use of the known error channel. When the error is correctable, the Petz operator has code subspace matrix elements which match the global reconstruction, and so causality is ensured in this case as well, because by (\ref{eq:code-subspace-causality}) the Kraus operators will also commute with $\mathcal{O}_{\text{Petz}}$ in the code subspace.
When the error channel is correctable, $\phi$ and $\mathcal{O}_{\text{Petz}}$ are identical as operators acting on the code subspace.
The global representation, on the other hand, is always nonzero on the physical Hilbert space if the bulk EFT operator $\phi$ is nonzero, and the Knill-Laflamme conditions are not enough to ensure causality for such operators.
The bottom line is that an operator represented purely on the radiation with access to the error channel, such as the Petz operator, may be constructed to vanish when the error may violate semi-classical causality with respect to the global representation, thus ensuring that this weaker version of semi-classical causality is preserved.
\subsection*{Acknowledgments}
We thank Abhijit Gadde, Lampros Lamprou, Raghu Mahajan, Gautam Mandal, Alexey Milekhin, Shiraz Minwalla, Daniel Ranard, Pratik Rath, Pushkal Shrivastava, Douglas Stanford, Sandip Trivedi and Zhenbin Yang for useful discussions. We thank Abhijit Gadde for comments on a previous version of the paper.
VB is supported in part by the Department of Energy through grant DE-SC0013528 and grant QuantISED DE-SC0020360, as well as the Simons Foundation through the It From Qubit Collaboration (Grant No. 38559).
AK is supported by the Simons Foundation through the It from Qubit Collaboration.
CL is supported by the Department of Energy through grants DE-SC0013528 and DE-SC0020360. OP is supported by the Department of Atomic Energy, Government of India, under project identification number RTI 4002.
| e6d8c3e8ff5e52b9c4242dfa176200f8f7c3cb9f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
Finding a time-optimal way to follow a reference path while respecting kinematic joint limits is one of the most prevalent tasks in industrial robotics.
However, just like a race driver needs to know the exact course of the race track to achieve an optimal lap time, time-optimal path parameterization (TOPP) requires the desired reference path to be fully known in advance.
For that reason, existing offline methods for time-optimal path parameterization are not applicable to reactive online scenarios in which the reference path needs to be adjustable during motion execution.
The same is true if sensory feedback is needed to successfully perform the path tracking task.
As shown in Fig.~\ref{fig:header}, this applies, for example, to the bipedal humanoid robot ARMAR-4, which can lose its balance when its arms are moved or to an industrial robot that balances a ball on a plate while following a reference path.
\begin{figure}[t]
\def-14.15cm{-14.15cm}
\def-6.35cm{-6.35cm}
\captionsetup[subfigure]{margin=80pt}
\begin{tikzpicture}
\def\xminPlot{0.151}
\def\xmaxPlot{0.949}
\SUBTRACT{\xmaxPlot}{\xminPlot}{\xdeltaPlot}
\DIVIDE{\xdeltaPlot}{3}{\xdeltaPlotNorm}
\def0.135{0.135}
\def0.985{0.985}
\def0.364{0.364}
\def0.535{0.535}
\SUBTRACT{0.985}{0.135}{\ydeltaPlot}
\SUBTRACT{0.535}{0.364}{\ydeltaAcc}
\MULTIPLY{\ydeltaAcc}{4}{\ydeltaGraph}
\SUBTRACT{\ydeltaPlot}{\ydeltaGraph}{\ydeltaTmp}
\DIVIDE{\ydeltaTmp}{3}{\ydeltaGap}
\node[text width=4cm] at (0.0cm, 0cm) (origin){};
\node[align=left, scale=0.9] at ($(origin.center)+(-0.15cm, -0.275cm)$) {Without sensory feedback};
\node[align=left, scale=0.9] at ($(origin.center)+(4.35cm, -0.275cm)$) {With sensory feedback};
%
\draw [draw=black!50, line width=1.0pt, scale=0.7] ($(origin.center)+(3.075cm, -0.1cm)$) -- + (0.0cm, -13.3cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{};
\end{tikzpicture}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-14.15cm}
%
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=800 340 800 220, clip, height=0.92\textwidth]{figures/header/one_industrial/header_2_1_small.jpg}};
\end{tikzpicture}
%
%
%
\vspace{-0.38cm}\hspace*{-2.5cm}\subcaptionbox{Kuka iiwa}[9cm]
%
%
\end{subfigure}
\hspace{0.0065\textwidth}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-14.15cm}
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) {\includegraphics[trim=800 340 800 220, clip, height=0.92\textwidth]{figures/header/ball_balancing/header_2_3_small.jpg}};
\end{tikzpicture}
%
\vspace{-0.38cm}\hspace*{-2.94cm}\subcaptionbox{KUKA with balance board}[10cm]
\end{subfigure}
\vspace{0.18cm}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-6.35cm}
\vspace{0.0cm}
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) {\includegraphics[trim=800 450 800 110, clip, height=0.92\textwidth]{figures/header/armar_6/header_2_1_small.jpg}};
\end{tikzpicture}
%
%
\vspace{-0.425cm}\hspace*{-2.05cm}\subcaptionbox{ARMAR-6}[8cm]
%
%
\end{subfigure}
\hspace{0.0065\textwidth}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-6.35cm}
\vspace{0.0cm}
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=790 435 760 80, clip, height=0.92\textwidth]{figures/header/armar_4/header_2_5_small.jpg}};
\end{tikzpicture}
%
%
\vspace{-0.425cm}\hspace*{-2.55cm}\subcaptionbox{ARMAR-4}[9cm]
\end{subfigure}
\vspace{-1.2cm}
\caption{%
Our evaluation environments for learning time-optimized path tracking.
Using sensory feedback, our method can incorporate additional goals such as ball balancing or maintaining balance with the bipedal \mbox{robot ARMAR-4}.
}%
\vspace{-0.75cm}
\label{fig:header}
\end{figure}
In this work, we address the problem of time-optimized path tracking for online scenarios like those shown in Fig. \ref{fig:header} by learning a well-performing trade-off between the execution speed, the deviation from the reference path and additional objectives like ball balancing via model-free reinforcement learning (RL).
For this purpose, a neural network is trained in a simulation environment using thousands of different reference paths.
At each time step, only the directly following part of the reference path is made available to the neural network as an input signal.
Once the training process is finished, the neural network can generate optimized actions even for paths not included in the training set. Our action space ensures that all generated trajectories comply with predefined kinematic joint limits.
Additional optimization objectives like ball balancing can be easily added to the reward function as they do not have to be differentiable with respect to the selected action. \\
Our main contributions can be summarized as follows:
\begin{itemize}
\item We introduce an online method to track reference paths based on reinforcement learning that ensures compliance with kinematic joint limits at all times.
\item We show that our approach can incorporate sensory feedback to account for additional objectives such as maintaining the balance of a bipedal humanoid robot.
\item We evaluate our method using robots with up to 30 degrees of freedom and demonstrate successful sim-2-real transfer for a time-optimized ball-on-plate task. %
\end{itemize}
Our source code will be made publicly available.\href{https://www.github.com/translearn}{$^2$}
\begin{figure*}[t]
\input{figures/overview_training}
\caption{%
The principle of online trajectory generation with our approach illustrated for a single time step $t$ of a \mbox{ball-on-plate task.}}
\label{fig:basic_principle}
\vspace{-0.6cm}
\end{figure*}
\section{Related work}
\subsection{Time-optimal path parameterization (TOPP)}
The problem of finding a time-optimal path parameterization subject to kinematic joint constraints has been studied for decades with methods proposed based on dynamic programming \cite{shin85Time}, numerical integration \cite{bobrow1985time}, convex optimization \cite{verscheure2009time} and reachability analysis \cite{pham2018Toppra}.
Widely used implementations that can consider velocity and acceleration constraints include \cite{kunz2012time} for paths consisting of line segments and circular blends and \cite{pham2018Toppra} for paths defined as cubic splines.
While our approach additionally supports jerk limits, TOPP subject to jerk constraints is still an open research problem~\cite{pham2017structure}.
Methods related to TOPP have also been used for ball balancing \cite{kiemel2020truerma} or to keep bipedal humanoid robots in balance e.g. by defining multi-contact friction constraints \cite{hauser2014fast, pham2018Toppra} or by imposing constraints on the zero moment point (ZMP) \cite{pham2012time}.
However, as time-optimal path parameterization requires the desired path to be known in advance, all calculations have to be performed offline. As a consequence, it is not possible to consider sensory feedback to compensate for external disturbances or model errors. The latter is particularly problematic as existing methods often require simplified contact models.
\subsection{Online trajectory generation (OTG)}
Time-optimal point-to-point motions subject to kinematic joint constraints can be computed online, for instance, with the Reflexxes motion library \cite{kroger2011opening}. This is done by specifying a desired kinematic target state at each control time step. In contrast to our work, this method does not allow to specify a desired reference path.
An online method for computing a feasible and jerk-limited path-accurate robot trajectory is presented in \cite{lange2015path}.
However, unlike our work, the method does not put focus on generating time-optimized trajectories. %
Another line of research aims at finding optimized robot trajectories based on model-predictive control (MPC) \cite{faulwasser2016implementation, carron2019data, nubert2020safe}.
Compared to MPC, model-free RL offers the advantage that no differentiable dynamics model is required and that the reward function does not have to be differentiable with respect to the selected action.
In addition, generating optimized motions with neural networks trained via RL is very fast which allows real-time execution even for systems with many degrees of freedom.
As a consequence, model-free RL can be used for a wide range of applications and has attracted increasing attention from the research community over the past few years. %
Exemplary learning tasks presented in the past include in-hand manipulation \cite{andrychowicz2020learning}, locomotion with bipedal robots \cite{siekmann2021sim} or collision-free target point reaching \cite{kiemel2021collision}, to name but a few.
Model-free RL has also been used to adjust reference trajectories in order to meet additional objectives \cite{peng2018deepmimic, kiemel2020trueadapt}. Compared to these approaches, we consider reference paths instead of reference trajectories, meaning that the velocity profile of the resulting trajectory is part of the optimization rather than being specified in advance.
Considering that additional objectives like balancing become more difficult the faster the reference path is traversed, optimizing the traversing speed during the training process is an important factor for learning well-performing trajectories.
\label{sec:citations}
\section{Approach}
\subsection{Overview}
Fig.~\ref{fig:basic_principle} illustrates the basic principle of our method based on a ball-on-plate task with sensory feedback.
The goal of the task is to quickly follow a reference path (shown in green) while avoiding a ball to fall from a balance board attached to the last link of the robot.
The position of the ball on the board is measured and provided as sensory feedback.
Using the mathematical framework of a Markov decision process, we formalize the online generation of optimized robot trajectories as a discrete-time control problem with a constant time~$\Delta t_N$ between decision steps. %
At each decision step~$t$, a state description $s_t$ containing information on the reference path, the kinematic state of the robot and the ball position is given as input to a neural network.
The neural network outputs an action~$\underline{a}_{t}$ that is used to compute a continuous trajectory from the current time step~$t$ to the subsequent time step $t+1$.
During training, the trajectory is executed in a physics simulator and a scalar reward is computed to evaluate the performance of the generated action~$\underline{a}_{t}$.
As part of the reward computation, the path resulting from the trajectory execution (shown in red) is assessed based on its length and its deviation to the reference path.
In addition, the balancing performance is rated based on the distance between the ball and the center of the board.
Using model-free RL, the neural network is trained to output actions that maximize the sum of rewards received over time.
This way, the network learns to generate robot trajectories that are optimized with respect to the execution time, the path deviation and the balancing performance.
The following subsection provides further details on the Markov decision process.
Subsequently, we explain how the reference paths used to train the neural network are generated.
\begin{figure}[b]
\centering
\vspace{0.0cm}
\begin{tikzpicture}[auto, node distance=5cm,>=latex', scale=1.0, every node/.style={scale=1.0}]
\def0.9{0.9}
\node[image_frame](observation_spline) {\includegraphics[trim=500 440 550 1300, clip, width=0.47\textwidth]{figures/observation_spline_sharpen_3_small.jpg}};
\node[minimum height=0cm, minimum width=0.0cm](annotation_center) at ($(observation_spline.center) + (0.28cm, -0.15cm)$){};
\draw [|-|, draw=black!80, thick] plot [smooth, tension=0.5] coordinates { ($(annotation_center.center) + (-1.7cm, 0.125cm)$) ($(annotation_center.center) + (-1.55cm, -0.02cm)$) ($(annotation_center.center) + (-1.4cm, -0.14cm)$)};
\draw [|-|, draw=black!80, thick] plot [smooth, tension=0.5] coordinates { ($(annotation_center.center) + (-1.85cm, -0.9cm)$) ($(annotation_center.center) + (-0.93cm, -1.35cm)$) ($(annotation_center.center) + (0.6cm, -1.15cm)$)
($(annotation_center.center) + (1.5cm, -0.4cm)$)
($(annotation_center.center) + (2.4cm, 0.7cm)$)};
\node[minimum height=0cm, minimum width=0.0cm, align=center, scale=0.9] at ($(annotation_center.center) + (2.8cm, -0.5cm)$){Path length \\$l_{\textrm{State}}$};
\node[minimum height=0cm, minimum width=0.0cm, align=center, scale=0.9] at ($(annotation_center.center) + (-3.25cm, -0.5cm)$){Reference \\ path}; %
\draw [-stealth, draw=black!80, thick] ($(annotation_center.center) + (-3.2cm, 0.0cm)$) -- ($(annotation_center.center) + (-2.7cm, 0.5cm)$);
\draw [-stealth, draw=black!80, thick] ($(annotation_center.center) + (-1.85cm, 0.975cm)$) -- ($(annotation_center.center) + (-1.85cm, 0.1cm)$) node[pos=0.0, above, minimum height=0cm, minimum width=0.0cm, xshift=0.05cm, scale=0.9]{Knot $1$};
\draw [-stealth, draw=black!80, thick] ($(annotation_center.center) + (1.4cm, 1.175cm)$) -- ($(annotation_center.center) + (1.85cm, 1.175cm)$) node[pos=0.0, left, minimum height=0cm, minimum width=0.0cm, scale=0.9]{Knot $N$};
\draw [-stealth, draw=black!80, thick, scale=0.9] ($(annotation_center.center) + (-1.15cm, 0.05cm)$) -- ($(annotation_center.center) + (-1.5cm, -0.33cm)$) node[pos=0.0, right=-0.0cm, minimum height=0cm, minimum width=0.0cm, align=center, yshift=0.3cm, scale=0.9]{Current \\path position};
\end{tikzpicture}
\caption{The figure illustrates how the following part of the reference path is included in the state using $N=5$ knots.}
\label{fig:observation_spline}
\end{figure}
\subsection{Details on the Markov decision process}
The problem of generating optimized robot trajectories is formalized as a Markov decision process $(\mathcal{S}, \mathcal{A}, P_{\underline{a}}, R_{\underline{a}})$, with $\mathcal{S}$ being the state space, $\mathcal{A}$ being the action space, $P_{\underline{a}}$ representing (unknown) transition probabilities and $R_{\underline{a}}$ being the reward resulting from action $\underline{a}$. In the following, details on the state space, the action space, the reward calculation and the termination of episodes are provided.
\subsubsection{Composition of states ${s} \in \mathcal{S}$}
A state ${s}$ contains information on the following part of the reference path and on the kinematic state of the robot.
For tasks that require sensory feedback, the sensor signals are included in the state as well.
In our work, the reference paths are described as cubic splines.
A spline is a mathematical representation that can be used to define a piecewise polynomial path between specified waypoints, called knots.
To provide the neural network with information about the reference path, $N$ of these knots are included in the state.
Note that the reference path is defined in joint space, meaning that each knot is a vector of joint positions.
For illustration purposes, however, we apply forward kinematics to visualize the reference path in Cartesian space.
Fig. \ref{fig:observation_spline} illustrates how the knots are selected.
The red cross indicates the current position on the reference path.
At the beginning of an episode, this position is set to the start of the reference path.
During an episode, each action causes the robot to move along a certain path. %
After each action, the current position on the reference path is shifted forward according to the length of the path generated by the action.
As shown in Fig. \ref{fig:observation_spline}, the state contains the knot preceding the current path position and $N\!-\!1$ following knots.
In addition, the path length labeled $l_{\textrm{State}}$ and the path length between the first knot and the current path position are included in the state.
If the remaining part of the reference path consists of less than $N\!-\!1$ knots, the last knot is inserted into the state multiple times.
In this case, the decreasing path length $l_{\textrm{State}}$ indicates to the network that the robot must be slowed down.
\begin{figure}[t]
\centering
\captionsetup[subfigure]{margin=0pt}
\vspace{0.2cm}
\begin{tikzpicture}
\def\xminPlot{0.151}
\def\xmaxPlot{0.949}
\SUBTRACT{\xmaxPlot}{\xminPlot}{\xdeltaPlot}
\DIVIDE{\xdeltaPlot}{3}{\xdeltaPlotNorm}
\def0.135{0.135}
\def0.985{0.985}
\def0.364{0.364}
\def0.535{0.535}
\SUBTRACT{0.985}{0.135}{\ydeltaPlot}
\SUBTRACT{0.535}{0.364}{\ydeltaAcc}
\MULTIPLY{\ydeltaAcc}{4}{\ydeltaGraph}
\SUBTRACT{\ydeltaPlot}{\ydeltaGraph}{\ydeltaTmp}
\DIVIDE{\ydeltaTmp}{3}{\ydeltaGap}
\hspace{0.23cm}
\node[text width=4cm] at (0.2cm, 4.35cm) (origin){};
\draw [draw=REFERENCE_PATH_GREEN, line width=1.5pt, scale=0.7] ($(origin.center)+(-9.3cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{Reference path};
\draw [draw=OBSERVATION_PATH_CYAN, line width=1.5pt, scale=0.7] ($(origin.center)+(-5.72cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{Waypoints in state};
\draw [draw=GENERATED_PATH_RED, line width=1.5pt, scale=0.7] ($(origin.center)+(-1.5cm, 0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.00cm, align=left, scale=0.85]{Generated path};
%
\end{tikzpicture}
\begin{subfigure}[c]{0.23\textwidth}
%
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) {\includegraphics[trim=450 300 1400 475, clip, height=0.85\textwidth]{figures/path_change/path_change_1_2_small.jpg}};
\end{tikzpicture}
%
\vspace{-0.40cm}\hspace*{-0.7cm}\subcaptionbox{Initial reference path}[5cm]
%
%
\end{subfigure}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=450 300 1400 475, clip, height=0.85\textwidth]{figures/path_change/path_change_3_2_small.jpg}};
\end{tikzpicture}
%
%
\vspace{-0.4cm}\hspace*{-0.645cm}\subcaptionbox{Adjusted reference path}[5cm]
\end{subfigure}
\caption{Adjusting the reference path during motion execution.}
\label{fig:spline_change}
\vspace{-0.38cm}
\end{figure}
As illustrated in Fig. \ref{fig:spline_change}, the reference path can be adjusted during motion execution.
We note that the reference path can be traversed faster if more knots are included in the state.
In return, however, a larger part of the reference path needs to be known in advance.
The kinematic state of the robot is described by the position~$p$, velocity~$v$ and acceleration~$a$ of each robot joint.
For the ball-on-plate task, the position of the ball on the plate is included in the state as sensory feedback.
To maintain balance with the humanoid robot \mbox{ARMAR-4}, the position and the rotation of the robot's pelvis relative to its initial upright pose is included in the state. While we extract the data from simulation, it could also be provided by an inertia measurement unit (IMU).
\vspace{0.15cm}
\subsubsection{Trajectory generation based on actions $\underline{a} \in \mathcal{A}$}
\hfill\\
Each action $\underline{a}$ defines a trajectory with a duration of $\Delta t_N$.
The generated trajectory must satisfy the following kinematic joint constraints for each robot joint and at all times:
\begin{alignat}{3}
p_{min} &{}\le{}& \theta &{}\le{}& p_{max} \label{eq:constraint_p} \\
v_{min} &{}\le{}& \dot{\theta} &{}\le{}& v_{max} \label{eq:constraint_v}\\
a_{min} &{}\le{}& \ddot{\theta}&{}\le{}& a_{max} \label{eq:constraint_a} \\
j_{min} &{}\le{}& \dddot{\theta} &{}\le{}& j_{max}, \label{eq:constraint_j}
\end{alignat}
where $\theta$ is the joint position and $p$, $v$, $a$ and $j$ stand for position, velocity, acceleration and jerk, respectively.
\begin{figure}[t]
\begin{tikzpicture}[auto, node distance=5cm,>=latex', scale=0.7, every node/.style={scale=0.85}]
\node[text width=4cm] at (0, 0) (origin){};
\draw [draw=LINE_COLOR_BLUE, line width=1.5pt] ($(origin.center)+(-3.8cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=-0.01cm, align=left]{\small{Generated trajectory}};
\draw [solid, draw=LINE_COLOR_RED, line width=1.5pt] ($(origin.center)+(0.4cm, 0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=-0.01cm, align=left]{\small{Kinematically feasible acceleration setpoints}};
%
\end{tikzpicture}
\hspace{-5.5cm}
\resizebox{0.8\textwidth}{!}{
\input{figures/action_space}
} %
\caption{%
The figure illustrates for a single joint how actions $\underline{a}$ are mapped to kinematically feasible accelerations.
}
\vspace{-0.5cm}
\label{fig:action_space}
\end{figure}
Each action $\underline{a} \in \mathcal{A}$ is a vector consisting of a scalar $\in [-1, 1]$ per robot joint controlled by the neural network.
At time step $t$, the action $\underline{a_t}$ specifies the joint acceleration $a_{t+1}$.
Compliance with the specified joint limits is ensured by mapping the action $\underline{a_t}$ along the range of kinematically feasible acceleration setpoints $[a_{t+1_{min}}, a_{t+1_{max}}]$:
\begin{align}
a_{t+1} = a_{t+1_{min}} + \frac{1 + \underline{a}_t}{2} \cdot \left(a_{t+1_{max}} - a_{t+1_{min}}\right)
\end{align}
The method to compute the range $[a_{t+1_{min}}, a_{t+1_{max}}]$ is explained in \cite{kiemel2020learning}.
Fig. \ref{fig:action_space} demonstrates how the acceleration of a single joint is controlled using three exemplary actions $\underline{a_t}$, $\underline{a_{t+1}}$ and $\underline{a_{t+2}}$.
To generate a continuous trajectory, the acceleration is linearly interpolated between the discrete time steps.
Once the course of the joint acceleration is known, velocity and position setpoints for a trajectory controller can be calculated by integration.
\subsubsection{Calculation of rewards $R_{\underline{a}}$}
We consider the tracking of reference paths as a multi-objective optimization problem.
The reward function formalizes the trade-off between the time to traverse the reference path, the deviation from the reference path and additional task-specific objectives:
\begin{align}
R_{\underline{a}} = \alpha \cdot R_l + \beta \cdot R_d + \gamma \cdot R_{s_1} + \ldots + \delta \cdot R_{s_n},
\label{eq:reward}
\end{align}
where $R_l$ is the path length reward, $R_d$ is the path deviation reward and $R_{s_1}$ to $ R_{s_n}$ represent task-specific objectives. %
The reward components $R_l$, $R_d$, $R_{s_1} \ldots R_{s_n}$ are defined to be in the range $[0.0, 1.0]$.
Non-negative weighting factors \mbox{$\alpha$, $\beta$, $\gamma \ldots \delta$} control the influence of each objective.
To determine the path length reward $R_l$ and the path deviation reward $R_d$, we calculate the path traversed as a result of action $\underline{a}$. In the reward calculation box of Fig.~\ref{fig:basic_principle}, this path is shown in red.
\begin{figure}[b]
\centering
\captionsetup[subfigure]{margin=0pt}
\vspace{-0.2cm}
\begin{subfigure}[c]{0.23\textwidth}
%
\vspace{-0.0cm}
%
\input{figures/reward/distance}
%
%
%
%
\end{subfigure}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-0.0cm}
%
\input{figures/reward/deviation}
%
%
\end{subfigure}
\vspace{-0.1cm}
\caption{Path length reward $R_l$ and path deviation reward $R_d$.}
\label{fig:rewards}
\vspace{-0.1cm}
\end{figure}
We then compute the length of the path $l$ and its average deviation from the \mbox{reference path $d$}.
To calculate $d$, the Euclidian distance between waypoints on the generated path and waypoints on the reference path is averaged.
The first waypoint of the generated path is compared with the waypoint of the reference path that is marked with a red cross in Fig.~\ref{fig:observation_spline}.
Further comparison points are shifted by the same arc length on each of the two paths.
Fig.~\ref{fig:rewards} illustrates how $l$ and $d$ are mapped to $R_l$ and $R_d$.
The path length reward $R_l$ controls the speed of traversing the reference path.
Under normal conditions, a fast traversal is preferred. For that reason, larger path lengths receive higher rewards.
However, towards the end of the reference path, the robot should slow down and finally come to a standstill.
To achieve this behavior, the reward is reduced if the path length exceeds $l_\text{State}$, the length of the path included in the state. %
Once the end of the reference spline is reached, $l_\text{State}$ is zero and the constant length $l_\text{End}$ controls how much the robot is penalized for further movements. The path deviation reward $R_d$ encourages the robot to stay close to the reference path. It is defined as a decreasing quadratic function yielding a value of zero if the deviation $d$ exceeds a predefined threshold $d_{\textrm{max}}$.
For the ball-on-plate task, an additional reward component $R_{s_1}$ is determined based on the distance between the ball and the center of the plate, using a decreasing quadratic function like that shown for $R_d$.
To maintain balance with the bipedal humanoid \mbox{ARMAR-4}, we reward small angles between the robot's pelvis and an upward pointing z-axis. This angle is \SI{0}{\degree} when the robot is upright and \SI{90}{\degree} when the robot is lying on the ground.
If the legs of the robot are not fixed, we additionally reward a small positional displacement of the pelvis to prevent the robot from moving around.
\subsubsection{Termination of episodes} In the case of a successful task execution, an episode is terminated after a specified number of time steps. However, during training we define additional conditions that lead to an early termination of an episode. For path tracking without sensory feedback, an episode is terminated if the deviation $d$ between the generated path and the reference path exceeds a predefined threshold.
The ball-on-plate task is additionally terminated if the ball falls from the plate.
In case of the bipedal humanoid \mbox{ARMAR-4}, an episode is terminated if the robot falls over.
Note that the reward assigned to an action is never negative.
For that reason, early termination is avoided during the learning process as it leads to a smaller sum of rewards.
\subsection{Generation of reference paths}
\subsubsection{Datasets used for the training process}
When using model-free RL, the aim of the training process is to find a policy that optimizes the sum of future rewards.
Transferred to our specific problem, this means that the learning algorithm tries to find an optimized tracking strategy for the entire reference path, although only a part of it is known.
The best optimization results can be achieved if the reference paths used within the training process are similar to those encountered during deployment.
For our evaluation, we use three datasets with different path characteristics.
Fig.~\ref{fig:datasets} visualizes some of the reference paths for each dataset.
The random dataset is generated by selecting random actions at each decision step and storing the paths resulting from the random robot motions. We use a method described in \cite{kiemel2021collision} to ensure that only collision-free paths are generated. As can be seen in Fig.~\ref{fig:datasets}a, the reference paths cover the entire working space of the robot.
In typical industrial applications, however, robots do not move randomly. Instead, they often move between Cartesian target points.
To recreate such a situation, we train a neural network to generate collision-free trajectories between randomly sampled target points using the method from \cite{kiemel2021collision}.
The target point dataset shown in Fig.~\ref{fig:datasets}b is composed of paths produced by this network.
For the ball-on-plate task, the reference paths are computed in such a way that the plate is always aligned horizontally. The corresponding dataset is visualized in Fig.~\ref{fig:datasets}c.
We note that the datasets used for training and testing are created in the same way, but do not contain the same paths.
\begin{figure}[t]
\captionsetup[subfigure]{margin=0pt}
\vspace{0.2cm}
\hspace{-0.0025\textwidth}
\begin{subfigure}[c]{0.155\textwidth}
%
\vspace{-0.0cm}
%
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=925 300 900 350, clip, height=1.025\textwidth]{figures/datasets/random_small.jpg}};
\end{tikzpicture}
\vspace{-0.40cm}\hspace*{-1.2cm}\subcaptionbox{Random}[5cm]
%
%
\end{subfigure}
\begin{subfigure}[c]{0.155\textwidth}
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=925 300 900 350, clip, height=1.025\textwidth]{figures/datasets/target_point_small.jpg}};
\end{tikzpicture}
%
%
\vspace{-0.4cm}\hspace*{-1.2cm}\subcaptionbox{Target point}[5cm]
\end{subfigure}
\begin{subfigure}[c]{0.155\textwidth}
\vspace{0.0cm}
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=925 300 900 350, clip, height=1.025\textwidth]{figures/datasets/balancing_small.jpg}};
\end{tikzpicture}
%
\vspace{-0.4cm}\hspace*{-1.2cm}\subcaptionbox{Ball balancing}[5cm]
%
%
\end{subfigure}
\caption{Visualization of the datasets used for our evaluation.}%
\label{fig:datasets}
\vspace{-0.28cm}
\end{figure}
\subsubsection{Spline knots included in the state}
\begin{figure}[h]
\captionsetup[subfigure]{margin=0pt}
\vspace{-0.5cm}
\centering
\begin{tikzpicture}
\definecolor{ORIGINAL_PATH_BLUE}{RGB}{0, 0, 255}
\definecolor{MATPLOTLIB_ORANGE}{RGB}{255, 127, 14}
\definecolor{MATPLOTLIB_DIMGREY}{RGB}{105, 105, 105}
\hspace{0.23cm}
\node[text width=4cm] at (0.2cm, 4.35cm) (origin){};
\draw [draw=ORIGINAL_PATH_BLUE, line width=1.5pt, scale=0.7] ($(origin.center)+(-9.3cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{Original path\strut};
\draw [draw=black, line width=0.9pt, scale=0.7] ($(origin.center)+(-6.0cm, -0.075cm)$) -- + (0.2cm, 0.2cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{};
\draw [draw=black, line width=0.9pt, scale=0.7] ($(origin.center)+(-6.0cm, +0.125cm)$) -- + (0.2cm, -0.2cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{};
\node[align=left, scale=0.85] at ($(origin.center)+(-3.12cm, -0.0cm)$) {Spline knots\strut};
\draw [draw=REFERENCE_PATH_GREEN, line width=1.5pt, scale=0.7] ($(origin.center)+(-3.0cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{Resulting reference path\strut};
%
\end{tikzpicture}
\begin{subfigure}[c]{0.23\textwidth}
%
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame](eight_equal_sampling) at (0, 0.0cm) {\includegraphics[trim=200 970 270 950, clip, height=0.6\textwidth]{figures/spline_sampling/eight_equal_sampling_original.png}};
\end{tikzpicture}
%
\vspace{-0.40cm}\hspace*{-0.65cm}\subcaptionbox{Distance-based sampling}[5cm]
%
%
\end{subfigure}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame](eight_curvature) at (0, 0.0cm) {\includegraphics[trim=200 970 270 950, clip, height=0.6\textwidth]{figures/spline_sampling/eight_curvature_original.png}};
\end{tikzpicture}
%
%
\vspace{-0.4cm}\hspace*{-0.645cm}\subcaptionbox{Curvature-based sampling}[5cm]
\end{subfigure}
\begin{subfigure}[c]{0.475\textwidth}
\vspace{0.1cm}
\begin{tikzpicture}
\definecolor{MATPLOTLIB_BLUE}{RGB}{31, 119, 180}
\definecolor{MATPLOTLIB_ORANGE}{RGB}{255, 127, 14}
\definecolor{MATPLOTLIB_DIMGREY}{RGB}{105, 105, 105}
\hspace{0.23cm}
\node[text width=4cm] at (0.2cm, 4.35cm) (origin){};
\draw [draw=MATPLOTLIB_BLUE, line width=1.5pt, scale=0.7] ($(origin.center)+(-9.3cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{$x(s)$};
\draw [draw=MATPLOTLIB_ORANGE, line width=1.5pt, scale=0.7] ($(origin.center)+(-7.4cm, -0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.01cm, align=left, scale=0.85]{$y(s)$};
\draw [draw=MATPLOTLIB_DIMGREY, line width=1.5pt, scale=0.7] ($(origin.center)+(-5.5cm, 0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.00cm, align=left, scale=0.85]{Curvature $\kappa(s)=\sqrt{x''(s)^2 + y''(s)^2}$};
%
\end{tikzpicture}
\vspace{0.1cm}
%
\includegraphics[trim=8 62 0 0, clip, width=\textwidth]{figures/spline_sampling/eight_equal_pos_curvature.pdf}
\includegraphics[trim=8 60 0 0, clip, width=\textwidth]{figures/spline_sampling/eight_curvature_pos_original.pdf}
\includegraphics[trim=8 0 0 54, clip, width=\textwidth]{figures/spline_sampling/eight_equal_pos_curvature.pdf}
\vspace{-0.4cm}\hspace*{-0.845cm}\subcaptionbox{Knot position over arc length for both sampling strategies.}[10cm]
\end{subfigure}
\caption{%
Strategies for the sampling of spline knots.}%
\label{fig:spline_sampling}
\vspace{-0.28cm}
\end{figure}
The reference path is described in the state based on a fixed number of following spline knots.
Different strategies can be used to place the knots along a given path.
Fig. \ref{fig:spline_sampling} illustrates two of them using an exemplary two-dimensional path shaped like a lemniscate.
With distance-based sampling, the knots are placed so that the arc length between them is equal.
When using curvature-based sampling, the curvature of the path is computed and integrated.
The knots are selected such that the integrated curvature between the knots is equal.
With distance-based sampling, the length of the path included in the state is always the same, except at the end of the reference path. When using curvature-based sampling, however, the included path length depends on the curvature of the reference path.
Sections of low curvature that can be traversed quickly lead to a larger path being included in the state.
\section{Evaluation}
We evaluate our method by learning to track reference paths with and without additional objectives.
Our evaluation environments are shown in Fig. \ref{fig:header}.
The KUKA iiwa is an industrial lightweight robot with 7 joints.
The humanoid robot \mbox{ARMAR-6}, shown in Figure \ref{fig:header}c, is controlled by 17~joints.
The bipedal humanoid ARMAR-4 has 30 joints, with 18 used for the upper body and 12 used for the legs.
In the case of ARMAR-4, the reference path specifies the joint positions of the upper body only.
For the other two robots, the positions of all joints are defined by the reference path.
Our neural networks are trained using proximal policy optimization (PPO) \cite{schulman2017proximal} based on data generated by the physics engine PyBullet\cite{coumans2016pybullet}.
We use networks with two hidden layers, the first one consisting of 256 neurons and the second one consisting of 128 neurons.
The time between decision steps is set to $\Delta t_{N}=$ \SI{0.1}{\s}.
The results shown in the following tables were obtained by averaging data from 1200 episodes, with reference paths taken from separate test datasets.
For the random dataset and the target point dataset, we use curvature-based sampling with $N=9$ knots included in the state.
In the case of the ball balancing dataset, distance-based sampling with $N=5$ knots is used.
\begin{figure}[b]
\captionsetup[subfigure]{margin=0pt}
\vspace{-0.3cm}
\centering
\begin{tikzpicture}
\def\xminPlot{0.151}
\def\xmaxPlot{0.949}
\SUBTRACT{\xmaxPlot}{\xminPlot}{\xdeltaPlot}
\DIVIDE{\xdeltaPlot}{3}{\xdeltaPlotNorm}
\def0.135{0.135}
\def0.985{0.985}
\def0.364{0.364}
\def0.535{0.535}
\SUBTRACT{0.985}{0.135}{\ydeltaPlot}
\SUBTRACT{0.535}{0.364}{\ydeltaAcc}
\MULTIPLY{\ydeltaAcc}{4}{\ydeltaGraph}
\SUBTRACT{\ydeltaPlot}{\ydeltaGraph}{\ydeltaTmp}
\DIVIDE{\ydeltaTmp}{3}{\ydeltaGap}
\node[text width=4cm] at (-1.2cm, 0cm) (origin){};
\draw [draw=REFERENCE_PATH_GREEN, line width=1.5pt, scale=0.7] ($(origin.center)+(-6.75cm, 0.075cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=-0.05cm, align=left, scale=0.85]{Reference path};
\draw [draw=REFERENCE_PATH_PURPLE, line width=1.5pt, scale=0.7] ($(origin.center)+(-6.75cm, -0.0cm)$) -- + (0.45cm, 0cm);
\draw [draw=GENERATED_PATH_RED, line width=1.5pt, scale=0.7] ($(origin.center)+(-0.75cm, 0.0cm)$) -- + (0.45cm, 0cm) node[pos=1, right, yshift=0.00cm, align=left, scale=0.85]{Generated path};
%
\end{tikzpicture}
\begin{subfigure}[c]{0.23\textwidth}
%
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=855 420 860 250, clip, height=0.85\textwidth]{figures/final_path/one_industrial/final_2.png}};
\end{tikzpicture}
%
\vspace{-0.40cm}\hspace*{-0.7cm}\subcaptionbox{KUKA iiwa}[5cm]
%
%
\end{subfigure}
\begin{subfigure}[c]{0.23\textwidth}
\vspace{-0.0cm}
%
\begin{tikzpicture}[scale=1.0, every node/.style={scale=1.0}, node distance=2cm]
\node[image_frame] at (0, 0.0cm) { \includegraphics[trim=800 390 800 170, clip, height=0.85\textwidth]{figures/final_path/armar_6/final_5_small.jpg}};
\end{tikzpicture}
%
%
\vspace{-0.4cm}\hspace*{-0.645cm}\subcaptionbox{ARMAR-6}[5cm]
\end{subfigure}
\caption{%
Path tracking without additional objectives.}%
\label{fig:example_path_no_feedback}
\vspace{-0.28cm}
\end{figure}
\begin{table*}[t]
\caption{Training results for time-optimized path tracking without additional objectives obtained based on 1200 episodes.}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{\textwidth}{p{34.5mm}p{19mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}p{8.7mm}}
\toprule
\multirow[t]{2}{*}{Configuration}
& \multicolumn{1}{c}{Duration [s]} & \multicolumn{3}{c}{Joint position deviation [rad]} & \multicolumn{3}{c}{Cart. position deviation [cm]} & \multicolumn{3}{c}{Cart. orientation deviation [°]}\\
& (robot stopped) & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} & \multicolumn{1}{c}{final} & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} & \multicolumn{1}{c}{final} & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} & \multicolumn{1}{c}{final}\\
\hline
Kuka iiwa & & & & \\
~~\llap{\textbullet}~~ Random dataset
& \hfil $4.28$ \hfil & \hfil $ 0.11$ \hfil & \hfil $0.19$ \hfil & \hfil $0.09$ \hfil & \hfil $3.3$ \hfil & \hfil $7.5$ \hfil & \hfil $2.7$ \hfil & \hfil $5.8$ \hfil & \hfil $11.8$ \hfil & \hfil $5.1$ \hfil \\
~~\llap{\textbullet}~~ Target point dataset %
& \hfil $4.99$ \hfil & \hfil $0.12$ \hfil & \hfil $0.21$ \hfil & \hfil $0.12$ \hfil & \hfil $3.7$ \hfil & \hfil $8.1$ \hfil & \hfil $3.8$ \hfil & \hfil $6.7$ \hfil & \hfil $13.5$ \hfil & \hfil $6.3$ \hfil \\
~~\llap{\textbullet}~~ Ball balancing dataset
& \hfil $2.44$ \hfil & \hfil $0.04$ \hfil & \hfil $0.08$ \hfil & \hfil $0.03$ \hfil & \hfil $1.4$ \hfil & \hfil $3.0$ \hfil & \hfil $1.4$ \hfil & \hfil $1.7$ \hfil & \hfil $3.9$ \hfil & \hfil $1.5$ \hfil \\
\hline
ARMAR-6 & & & & \\
~~\llap{\textbullet}~~ Random dataset
& \hfil $4.98$ \hfil & \hfil $0.14$ \hfil & \hfil $0.20$ \hfil & \hfil $0.16$ \hfil & \hfil $5.5$ \hfil & \hfil $13.6$ \hfil & \hfil $6.2$ \hfil & \hfil $5.3$ \hfil & \hfil $11.5$ \hfil & \hfil $6.0$ \hfil \\
\hline %
ARMAR-4 (fixed base) & & & & \\
~~\llap{\textbullet}~~ Random dataset
& \hfil $5.09$ \hfil & \hfil $0.14$ \hfil & \hfil $0.20$ \hfil & \hfil $0.14$ \hfil & \hfil $3.3$ \hfil & \hfil $7.8$ \hfil & \hfil $3.5$ \hfil & \hfil $5.6$ \hfil & \hfil $11.7$ \hfil & \hfil $5.7$ \hfil \\
~~\llap{\textbullet}~~ Target point dataset
& \hfil $5.48$ \hfil & \hfil $0.14$ \hfil & \hfil $0.21$ \hfil & \hfil $0.15$ \hfil & \hfil $3.6$ \hfil & \hfil $8.8$ \hfil & \hfil $3.8$ \hfil & \hfil $5.6$ \hfil & \hfil $12.4$ \hfil & \hfil $6.2$ \hfil \\
\bottomrule
\end{tabular*}
\label{table:tracking_no_feedback}
\vspace{-0.33cm}
\end{table*}
\subsection{Path tracking without additional objectives}
When tracking paths without additional targets, the networks are trained to minimize the duration of the trajectory and the average deviation from the reference path.
TABLE \ref{table:tracking_no_feedback} shows the training results obtained for the different robots and datasets.
Note that the learning algorithm tries to minimize the average joint position deviation, given in the second column.
However, to give a better idea of the results, we also specify deviations with respect to the position and orientation in Cartesian space.
For this purpose, the reference path and the generated path are converted to Cartesian space using forward kinematics.
The tool center point (TCP) is used as the reference point for the KUKA iiwa, whereas the fingertips are used for the humanoid robots.
Renderings of two exemplary episodes can be seen in Fig. \ref{fig:example_path_no_feedback}.
To compute the deviations, points on the reference path are compared with points on the generated path.
The comparison points are selected such that they are equally far away from the beginning of the respective path.
To specify a trajectory duration and a final position deviation, the robot joints are decelerated as soon as the end of the reference path is reached.
Using the ball balancing dataset, an average Cartesian deviation of \SI{1.4}{\cm} and \SI{1.7}{\degree} is obtained.
The paths in the random and the target point dataset show a larger variance, resulting in a Cartesian deviation of approximately \SI{4}{\cm} and~\SI{6}{\degree}.
In the following, we evaluate how the trajectory duration and the position deviation can be influenced.
\subsubsection{Trade-off traversing time vs. path deviaton}
\begin{table}[t]
\vspace{-0.0cm}
\caption{%
Trade-off traversing time vs. path deviation.
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{@{}p{27.0mm}p{17.5mm}p{8.7mm}p{8.7mm}p{8.7mm}}
\toprule
\hspace{0.05cm}Kuka iiwa with & \multicolumn{1}{c}{Duration [s]} & \multicolumn{3}{c}{Cart. position deviation [cm]} \\
\hspace{0.05cm}random dataset & (robot stopped) & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} & \multicolumn{1}{c}{final} \\
\hline
\hspace{0.05cm}~~\llap{\textbullet}~~ Length $<$ deviation
& \hfil $4.76$ \hfil & \hfil $2.7$ \hfil & \ \hfil $6.4$ \hfil & \hfil $2.3$ \hfil\\
\hspace{0.05cm}~~\llap{\textbullet}~~ Length $\approx$ deviation
& \hfil $4.28$ \hfil & \hfil $3.3$ \hfil & \ \hfil $7.5$ \hfil & \hfil $2.7$ \hfil\\
\hspace{0.05cm}~~\llap{\textbullet}~~ Length $>$ deviation
& \hfil $4.08$ \hfil & \hfil $3.6$ \hfil & \ \hfil $8.0$ \hfil & \hfil $3.1$ \hfil\\
\bottomrule
\end{tabular*}
\vspace{-0.54cm}
\label{table:trade_off}
\end{table}
Our reward function (\ref{eq:reward}) allows us to assign different weights to the path length reward $R_l$ and to the path deviation reward $R_d$.
TABLE \ref{table:trade_off} shows how the weighting affects the traversing time and the path deviation.
If the weighting of the deviation reward is increased, the path deviation decreases whereas the trajectory duration increases.
Likewise, faster trajectories with a higher path deviation are learned when the weighting of the path length reward is increased.
\subsubsection{Number of knots included in the state}
TABLE \ref{table:knots} shows how the training results are affected by the the number of knots $N$ used to describe the following part of the reference path. If more knots are included in the state, the reference path can be traversed faster. In return, however, a larger part of the reference path has to be known in advance.
\begin{table}[t]
\vspace{-0.0cm}
\caption{%
Impact of the knots included in the state.
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{p{22.5mm}p{19mm}p{8.7mm}p{8.7mm}p{8.7mm}}
\toprule
Kuka iiwa with & \multicolumn{1}{c}{Duration [s]} & \multicolumn{3}{c}{Cart. position deviation [cm]} \\
random dataset & (robot stopped) & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} & \multicolumn{1}{c}{final} \\
\hline
~~\llap{\textbullet}~~ 5 knots
& \hfil $5.20$ \hfil & \hfil $2.5$ \hfil & \ \hfil $5.6$ \hfil & \hfil $2.8$ \hfil\\
~~\llap{\textbullet}~~ 7 knots
& \hfil $4.52$ \hfil & \hfil $3.4$ \hfil & \ \hfil $7.5$ \hfil & \hfil $3.0$ \hfil\\
~~\llap{\textbullet}~~ 9 knots
& \hfil $4.28$ \hfil & \hfil $3.3$ \hfil & \ \hfil $7.5$ \hfil & \hfil $2.7$ \hfil\\
\bottomrule
\end{tabular*}
\vspace{-0.5cm}
\label{table:knots}
\end{table}
\begin{table}[b]
\vspace{-0.25cm}
\caption{%
Generalization ability between datasets.
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{@{}p{27.0mm}p{17.5mm}p{8.5mm}p{7.5mm}p{7.5mm}}
\toprule
\hspace{0.05cm} & \multicolumn{1}{c}{Duration [s]} & \multicolumn{3}{c}{Cart. position deviation [cm]} \\
\hspace{0.05cm} & (robot stopped) & \hfil mean \hfil & \hfil max \hfil & \hfil final \hfil \\
\hline
\hspace{0.05cm}Kuka iiwa & & & & \\
\hspace{0.05cm}~~\llap{\textbullet}~~ Random dataset to \newline \hspace*{0.05cm}\phantom{~~\llap{\textbullet}~~ }target point dataset
& \hfil $4.87$ \hfil & \hfil $3.9$ \hfil & \hfil $9.2$ \hfil & \hfil $2.7$ \hfil\\
\hspace{0.05cm}ARMAR-4 (fixed base) & & & & \\
\hspace{0.05cm}~~\llap{\textbullet}~~ Random dataset to \newline \hspace*{0.05cm}\phantom{~~\llap{\textbullet}~~ }target point dataset
& \hfil $5.40$ \hfil & \hfil $3.8$ \hfil & \hfil $9.2$ \hfil & \hfil $4.1$ \hfil\\
\bottomrule
\end{tabular*}
\vspace{-0.1cm}
\label{table:random_to_target}
\end{table}
\begin{table}[b]
\vspace{-0.1cm}
\caption{Feature comparison with TOPP-RA \cite{pham2018Toppra}}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{p{44mm}p{15.5mm}p{15.5mm}}
\toprule
\hspace{0.02cm} & \multicolumn{1}{c}{TOPP-RA} & \multicolumn{1}{c}{Ours} \\
\hline
Supports velocity limits & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil \\
Supports acceleration limits & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil \\
Supports jerk limits & \hfil \textcolor{PATH_COLLISION}{\ding{55}} \hfil & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil \\
Supports path adjustments & \hfil \textcolor{PATH_COLLISION}{\ding{55}} \hfil & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil \\
Supports sensory feedback & \hfil \textcolor{PATH_COLLISION}{\ding{55}} \hfil & \hfil \textcolor{TABLE_GOOD}{\ding{51}} \hfil \\
\bottomrule
\end{tabular*}
\vspace{0ex}
\label{table:feature_toppra}
\vspace{-0.0cm}
\end{table}
\vspace{-0.3cm}
\subsubsection{Generalization ability between datasets}
To analyze the generalization ability with respect to different path characteristics,
we evaluate networks trained using the random dataset with reference paths from the target point dataset.
TABLE \ref{table:random_to_target} shows the results for the Kuka iiwa and the humanoid \mbox{ARMAR-4}.
Compared to the networks trained directly with the target point dataset, the resulting trajectories are slightly faster but
also a little less accurate in tracking the reference path.
Overall, however, the differences are small, indicating that networks trained on random paths can also be used to track paths with different path characteristics.
In the accompanying \href{https://youtu.be/gCPN8mqPVHg}{\textcolor{blue}{video}}, we additionally show how a network trained on random paths performs on reference paths that resemble geometric shapes in Cartesian space.
\vspace{0.05cm}
\subsubsection{Comparison with time-optimal path parameterization}
We benchmark our approach with TOPP-RA \cite{pham2018Toppra}, a state-of-the-art offline method for time-optimal path parameterization.
TABLE \ref{table:feature_toppra} compares the features of both methods.
\begin{figure*}[t]
\input{figures/real_robot_balancing}
\caption{%
The ball-on-plate task performed by a real KUKA iiwa.
When the balancing performance is rewarded (bottom), the reference path is traversed less quickly, but the ball is kept close to the center of the plate.
}
\label{fig:real_robot_balancing}
\vspace{-0.45cm}
\end{figure*}
While our method additionally supports jerk limits and online adjustments of the reference path, the offline method TOPP-RA generates faster trajectories that track the reference path almost perfectly.
A quantitative analysis is shown in TABLE \ref{table:toppra_timing}.
Trajectories for the KUKA iiwa generated with TOPP-RA require around \SI{78}{\percent} of the time needed by our method.
For the humanoid \mbox{ARMAR-4}, around \SI{65}{\percent} of the time is needed.
In return, the reference paths need to be known in advance and it is not possible to consider additional objectives that require sensory feedback.
\subsection{Path tracking with objectives based on sensory feedback}
\subsubsection{Ball-on-plate task}
The goal of the ball-on-plate task is to traverse a reference path while balancing a ball.
During the motion execution, the position of the ball on the plate is provided as sensory feedback.
TABLE \ref{table:ball_balancing} shows the training results with and without an additional reward component based on the balancing performance.
Without the balancing reward, the ball falls off the plate in all evaluated episodes.
However, when the balancing performance is rewarded, the ball falls down in less than \SI{0.5}{\percent} of the episodes.
In return, the reference path is traversed less quickly and less precisely.
Fig. \ref{fig:real_robot_balancing} visualizes the results based on an exemplary reference path traversed by a real KUKA iiwa. The robot is controlled using the networks trained in simulation and the ball position is provided by a resistive touch panel.
\begin{table}[b]
\vspace{-0.1cm}
\caption{%
Comparison with the offline method \mbox{TOPP-RA.}
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{p{22.0mm}p{17.5mm}p{14.5mm}p{18.0mm}}
\toprule
& \hfil Duration [s] \hfil& \hfil Relative \hfil & \hfil Max. joint pos. \hfil \\
& \hfil (robot stopped) \hfil & \hfil duration [\%] \hfil & \hfil deviation [rad] \hfil \\
\hline
Kuka iiwa & & & \\
~~\llap{\textbullet}~~ Random
& \hfil $3.36$ \hfil & \hfil $78.5$ \hfil & \hfil $0.00$ \hfil \\
~~\llap{\textbullet}~~ Target point
& \hfil $3.79$ \hfil & \hfil $76.0$ \hfil & \hfil $0.00$ \hfil \\
~~\llap{\textbullet}~~ Ball balancing
& \hfil $1.91$ \hfil & \hfil $78.3$ \hfil & \hfil $0.00$ \hfil \\
\mbox{ARMAR-4 (fixed base)}& & & \\
~~\llap{\textbullet}~~ Random
& \hfil $3.31$ \hfil & \hfil $65.0$ \hfil & \hfil $0.00$ \hfil \\
~~\llap{\textbullet}~~ Target point
& \hfil $3.68$ \hfil & \hfil $67.2$ \hfil & \hfil $0.00$ \hfil \\
\bottomrule
\end{tabular*}
\vspace{-0.0cm}
\label{table:toppra_timing}
\end{table}
\begin{table}[t]
\vspace{0.05cm}
\caption{%
Training results for the ball-on-plate task.
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{@{}p{19.2mm}p{14.95mm}p{11.0mm}p{10.5mm}p{14.5mm}}
\toprule
\hspace{0.05cm}Ball balancing & \multicolumn{1}{c}{Duration [s]} & \multicolumn{1}{c}{Balancing} & \multicolumn{2}{c}{Cart. pos. deviation [cm]} \\
\hspace{0.05cm}dataset & \hfil (end of path) \hfil & \hfil error [\%] \hfil & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} \\
\hline
\hspace{0.05cm}Kuka iiwa & & & & \\
\hspace{0.00cm}~~\llap{\textbullet}~~ No balancing \newline \hspace*{0.00cm}\phantom{~~\llap{\textbullet}~~ }reward
& \hfil $2.24$ \hfil & \hfil $100.0$ \hfil & \hfil $1.4$ \hfil & \hfil $3.0$ \hfil\\
\hspace{0.00cm}~~\llap{\textbullet}~~ Balancing \newline \hspace*{0.00cm}\phantom{~~\llap{\textbullet}~~ }reward
& \hfil $2.99$ \hfil & \hfil $0.3$ \hfil & \hfil $2.3$ \hfil & \hfil $4.9$ \hfil\\
\bottomrule
\end{tabular*}
\vspace{-0.56cm}
\label{table:ball_balancing}
\end{table}
\begin{table}[b]
\vspace{-0.3cm}
\caption{%
Maintaining balance with ARMAR-4.
}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{@{}p{19.2mm}p{14.95mm}p{11.0mm}p{10.5mm}p{14.5mm}}
\toprule
\hspace{0.05cm}Target point & \multicolumn{1}{c}{Duration [s]} & \multicolumn{1}{c}{Balancing} & \multicolumn{2}{c}{Cart. pos. deviation [cm]} \\
\hspace{0.05cm}dataset & \hfil (end of path) \hfil & \hfil error [\%] \hfil & \multicolumn{1}{c}{mean} & \multicolumn{1}{c}{max} \\
\hline
\hspace{0.05cm}ARMAR-4 with \newline \hspace*{0.05cm}fixed legs& & & & \\
\hspace{0.00cm}~~\llap{\textbullet}~~ No balancing \newline \hspace*{0.00cm}\phantom{~~\llap{\textbullet}~~ }reward
& \hfil $5.28$ \hfil & \hfil $26.1$ \hfil & \hfil $3.6$ \hfil & \hfil $8.8$ \hfil\\
\hspace{0.00cm}~~\llap{\textbullet}~~ Balancing \newline \hspace*{0.00cm}\phantom{~~\llap{\textbullet}~~ }reward
& \hfil $5.63$ \hfil & \hfil $5.3$ \hfil & \hfil $4.1$ \hfil & \hfil $9.6$ \hfil\\
\hspace{0.05cm}ARMAR-4 with \newline \hspace*{0.05cm}controlled legs& & & & \\
\hspace{0.00cm}~~\llap{\textbullet}~~ Balancing \newline \hspace*{0.00cm}\phantom{~~\llap{\textbullet}~~ }reward
& \hfil $5.59$ \hfil & \hfil $0.8$ \hfil & \hfil $4.2$ \hfil & \hfil $9.8$ \hfil\\
\bottomrule
\end{tabular*}
\vspace{-0.0cm}
\label{table:armar_balancing}
\end{table}
\subsubsection{Maintaining balance with ARMAR-4}
The additional objective of this task is to prevent a bipedal robot from falling over.
For that purpose, the pose of the robot's pelvis is provided as sensory feedback.
TABLE \ref{table:armar_balancing} shows the training results for three different experimental configurations.
In the first two experiments, the legs of the robot are fixed in an outstretched position.
Without an additional balancing term in the reward function, the robot falls over in \SI{26}{\percent} of the episodes.
With an additional reward for standing upright, the robot loses balance in approximately \SI{5}{\percent} of the episodes.
In the third experiment, the 12 joints of the legs are also controlled by the neural network. %
Thus, the network can use the legs to stabilize the motions of the upper body. As a result, the robot falls over in less than \SI{1}{\percent} of the episodes.
\begin{figure}[t]
\hspace{-0.6cm}
\resizebox{0.53\textwidth}{!}{
\input{figures/armar_balancing}
} %
\caption{%
Top: Without a balancing reward, the robot falls over.
Bottom: With a reward for standing upright, the legs of the robot are used to keep the robot in balance.
}
\label{fig:armar_balancing}
\vspace{-0.45cm}
\end{figure}
An exemplary episode of the first and the third experiment is shown in Fig. \ref{fig:armar_balancing} and in the accompanying \href{https://youtu.be/gCPN8mqPVHg}{\textcolor{blue}{video}}.
\subsection{Real-time capability}
As shown in Fig. \ref{fig:real_robot_balancing} and in the accompanying \href{https://youtu.be/gCPN8mqPVHg}{\textcolor{blue}{video}},
we successfully applied our method to a real KUKA iiwa robot using networks trained in simulation.
The transfer is performed by sending the trajectory setpoints specified by each action to a real trajectory controller instead of a virtual one simulated by PyBullet. %
To analyze the computational requirements of our method, we calculate 1200 episodes for each of the three robots shown in this paper using an Intel i7-8700K CPU.
We then calculate the quotient of the computation time and the trajectory duration and provide the highest value of all episodes in TABLE \ref{table:computation_times}.
The results show that the computation time is significantly smaller than the trajectory duration, making our method well-suited for real-time trajectory generation.
\begin{table}[h]
\vspace{-0.15cm}
\caption{Evaluation of the computational effort.}
\vspace{-0.15cm}
\makegapedcells
\begin{tabular*}{0.49\textwidth}{@{}p{25.5mm}p{12mm}p{23.5mm}p{11mm}}
\toprule
\hspace{0.02cm} & \hfil Kuka iiwa \hfil & ARMAR-4 with legs & \multicolumn{1}{c}{ARMAR-6} \\
\hline
\hspace{0.05cm}\scalebox{1.4}{$\frac{\text{Computation time}}{\text{Trajectory duration}}$}
& \hfil \SI{7.50}{\percent} & \hfil \SI{34.97}{\percent} & \hfill \SI{10.59}{\percent} \\
\bottomrule
\end{tabular*}
\vspace{-0.5cm}
\label{table:computation_times}
\end{table}
\section{Conclusion and future work}
This paper presented a learning-based approach to follow reference paths that can be changed during motion execution.
Trajectories are generated by a neural network trained to maximize the traversing speed while minimizing the deviation from the reference path.
Additional task-specific objectives can be considered by including sensory feedback into the state.
The mapping of network actions to joint accelerations ensures that no kinematic joint limits are violated.
We evaluated our method with and without additional objectives on robotic systems with up to 30 degrees of freedom
showing that well-performing trajectories can be learned for reference paths with different path characteristics. %
We also demonstrated successful \mbox{sim-2-real} transfer for a ball-on-plate task performed by an industrial robot. %
In future work, we would like to investigate ways to additionally control the traversing speed during motion execution.
\section*{ACKNOWLEDGMENT}
This research was supported by the German Federal Ministry of Education and Research (BMBF) and the Indo-German Science \& Technology Centre (IGSTC) as part of the project TransLearn (01DQ19007A). We thank Tamim Asfour for his valuable feedback and advice.
\bibliographystyle{IEEEtran}
| 30ba33e393b54e687ffc8d035669e132969e2546 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Experimental details on CIFAR10}
\label{sec:app_cifar10}
In this section, we give the experimental details on the CIFAR10-based experiments shown in Figures \ref{fig:teaserplot} and \ref{fig:K_plot}. Moreover, we also conduct similar experiments using different neural network architectures. First, we give the full experimental details and then provide the results of the experiments using the different architectures.
\paragraph{Subsampling CIFAR10}
In all our experiments we subsample CIFAR10 to simulate the low sample size regime. We ensure that for all subsampled versions the number of samples of each class are equal. Hence, if we subsample to $500$ training images, then each class has exactly $50$ images, which are drawn uniformly from the $5k$ training images of the respective class.
\paragraph{Mask perturbation on CIFAR10}
We consider square black-mask perturbations; the attacker can set in the image a patch of size $2 \times 2$ to zero. The attack is a simplification of the patch-attack as considered in \cite{Wu20}. We show an example of a black-mask attack on each of the classes in CIFAR10 in Figure \ref{fig:cifar10_masks}. Clearly, the mask reduces the information about the class in the image as it occludes part of the object in the image.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.8\linewidth]{plotsAistats/cifar10_black_mask_attack.png}
\caption{We show an example of a mask perturbation for all $10$ classes of CIFAR10. Even though the attack occludes part of the images, a human can still easily classify all images correctly.}
\label{fig:cifar10_masks}
\end{figure}
During test time, we evaluate the attack exactly by means of a full grid search over all possible windows. Note that a full grid search requires $900$ forward passes to evaluate one image, which computationally too expensive during training time. Therefore, we use the same approximation as in \cite{Wu20} at training time. For each image in the training batch, we compute the gradient from the loss with respect to the input. Using that gradient, which is a tensor in $\mathbb{R}^{3 \times 32 \times 32}$, we compute the $l_1$-norm of each patch by a full grid search and save the upper left coordinates of the $K$ windows with largest $l_1$-norm. The intuition is that windows with high $l_1$-norm are more likely to change the prediction. Out of the $K$ identified candidate windows, we take the most loss worsening by means of a full list-search.
\begin{wrapfigure}{r}{0.4\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/K_plot_cifar.png}
\caption{We plot the standard error, robust error and susceptibility for varying attack strengths $K$. We see that the larger $K$, the lower the susceptibility, but the higher the standard error.}
\label{fig:K_plot}
\end{wrapfigure}
\paragraph{Experimental training details}
For all our experiments on CIFAR10, we adjusted the code provided by \cite{Phan21}. As typically done for CIFAR10, we augment the data with random cropping and horizontal flipping. For the experiments with results depicted in Figures \ref{fig:teaserplot} and \ref{fig:K_plot}, we use a ResNet18 network and train for $100$ epochs. We tune the parameters learning rate and weight decay for low robust error. For standard standard training, we use a learning rate of $0.01$ with equal weight decay. For adversarial training, we use a learning rate of $0.015$ and a weight decay of $10^{-4}$. We run each experiment three times for every dataset with different initialization seeds, and plot the average and standard deviation over the runs.
For the experiments in Figure \ref{fig:teaserplot} and \ref{fig:num_obs_CIFAR} we use an attack strength of $K = 4$. Recall that we perform a full grid search at test time and hence have a good approximation of the robust accuracy and susceptibility score.
\paragraph{Increasing training attack strength} We investigate the influence of the attack strength $K$ on the robust error for adversarial training. We take $\eps_{\text{tr}} = 2$ and $n = 500$ and vary $K$. The results are depicted in Figure \ref{fig:K_plot}. We see that for increasing $K$, the susceptibility decreases, but the standard error increases more severely, resulting in an increasing robust error.
\paragraph{Robust error decomposition}
In Figure \ref{fig:teaserplot}, we see that the robust error increases for adversarial training compared to standard training in the low sample size regime, but the opposite holds when enough samples are available. For completeness, we provide a full decomposition of the robust error in standard error and susceptibility for standard and adversarial training. We plot the decomposition in Figure \ref{fig:num_obs_CIFAR}.
\begin{figure*}[!b]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cifar10_robust_numobs.png}
\caption{Robust error}
\label{fig:RA_CIFAR_10_n}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cifar10_standard_numobs.png}
\caption{Standard error}
\label{fig:SA_CIFAR_10_n}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cifar10_sus_numobs.png}
\caption{Susceptibility}
\label{fig:Robustness_n}
\end{subfigure}
\caption{We plot the standard error, robust error and susceptibility of the subsampled datasets of
CIFAR10 after adversarial and standard training. For small sample size, adversarial
training has higher robust error then standard training. We see that the increase in standard error in comparison to the drop in susceptibility of standard versus robust training, switches between the low and high sample size regimes.}
\label{fig:num_obs_CIFAR}
\end{figure*}
\paragraph{Multiple networks on CIFAR10}
We run adversarial training for multiple network architectures on subsampled CIFAR10 ($n=500$) with mask perturbations of size $2 \times 2$ and an attack strength of $K=4$. We plot the results in Table \ref{CIFAR10_diffArchitectures}. For all the different architectures, we notice a similar increase in robust error when trained with adversarial training instead of standard training.
\begin{table}[!ht]
\centering
\caption{We subsample CIFAR10 to a dataset of sample size $500$ and perform both standard training (ST) and adversarial training (AT) using different networks. We evaluate the resulting susceptibility score and the robust and standard error. }
\begin{tabular}{ |p{2cm}||p{2cm}||p{1cm}||p{1cm}|p{2cm}|p{2cm}|p{2cm}|}
\hline
\multicolumn{7}{|c|}{Adversarial training on CIFAR10} \\
\hline
Architecture & learning rate & weight decay & Train type & standard error & robust error & Susceptibility\\
\hline
ResNet34 & $ 0.02$ & $0.025$ & ST & 44 & 64 & 50 \\
ResNet34 & $0.015$ & $10^{-4}$ & AT & 52 & 66 & 40\\
ResNet50 & $0.015$ & $0.03$ & ST & 45 & 62 & 47\\
ResNet50 & $0.015$ & $10^{-4}$ & AT & 53 & 68 & 45\\
VGG11bn & $0.03$ & $0.01$ & ST & 40 & 55 & 43\\
VGG11bn & $0.015$ & $10^{-4}$ & AT & 48 &63 & 34\\
VGG16bn & $0.02$ & $0.01$ & ST & 41 & 60 & 48\\
VGG16bn & $0.015$ & $10^{-4}$ & AT & 50 & 65 & 42\\
\hline
\end{tabular}
\label{CIFAR10_diffArchitectures}
\end{table}
\section{Static hand gesture recognition}
\label{sec:handgestures}
The goal of static hand gesture or posture recognition is to recognize hand gestures such as a pointing index finger or the okay-sign based on static data such as images \cite{Oudah20, Yang13}. The current use of hand gesture recognition is primarily in the interaction between computers and humans \cite{Oudah20}. More specifically, typical practical applications can be found in the environment of games, assisted living, and virtual reality \cite{Mujahid21}. In the following, we conduct experiments on a hand gesture recognition dataset constructed by \cite{Mantecon19}, which consists of near-infrared stereo images obtained using the Leap Motion device. First, we crop or segment the images after which we use logistic regression for classification. We see that adversarial logistic regression deteriorates robust generalization with increasing $\eps_{\text{tr}}$.
\paragraph{Static hand-gesture dataset}
We use the dataset made available by \cite{Mantecon19}. This dataset consists of near-infrared stereo images taken with the Leap Motion device and provides detailed skeleton data. We base our analysis on the images only. The size of the images is $640 \times 240$ pixels. The dataset consists of $16$ classes of hand poses taken by $25$ different people. We note that the variety between the different people is relatively wide; there are men and women with different posture and hand sizes. However, the different samples taken by the same person are alike.
We consider binary classification between the index-pose and L-pose, and take as a training set $30$ images of the users $16$ to $25$. This results in a training dataset of $300$ samples. We show two examples of the training dataset in Figure \ref{fig:original_examples}, each corresponding to a different class. Observe that the near-infrared images darken the background and successfully highlight the hand-pose. As a test dataset, we take $10$ images of each of the two classes from the users $1$ to $10$ resulting in a test dataset of size $200$.
\begin{figure}
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=.80\linewidth]{plotsAistats/Lpose.png}
\caption{L pose}
\label{fig:L_pose_or_example}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=.80\linewidth]{plotsAistats/Indexpose.png}
\caption{Index pose}
\label{fig:index_pose_or_example}
\end{subfigure}
\caption{We plot two images, where both correspond to the two different classes. We recognize the "L"-sign in Figure \ref{fig:L_pose_or_example} and the index sign in Figure \ref{fig:index_pose_or_example}. Observe that the near-infrared images highlight the hand pose well and blends out much of the non-useful or noisy background. }
\label{fig:original_examples}
\end{figure}
\paragraph{Cropping the dataset}
To speed up training and ease the classification problem, we crop the images from a size of $640 \times 240$ to a size of $200 \times 200$. We crop the images using a basic image segmentation technique to stay as close as possible to real-world applications. The aim is to crop the images such that the hand gesture is centered within the cropped image.
For every user in the training set, we crop an image of the L-pose and the index pose by hand. We call these images the training masks $\{\text{masks}_i \}_{i=1}^{20}$. We note that the more a particular window of an image resembles a mask, the more likely that the window captures the hand gesture correctly. Moreover, the near-infrared images are such that the hands of a person are brighter than the surroundings of the person itself. Based on these two observations, we define the best segment or window, defined by the upper left coordinates $(i,j)$, for an image $x$ as the solution to the following optimization problem:
\begin{equation}
\label{preprocessing}
\argmin_{i \in [440], \Hquad j \in [40]} \sum_{l=1}^{20}\|\text{masks}_l-x_{\{i:i+200,j:j+200\}}\|^2_2 - \frac{1}{2}\|x_{\{i+w,j+h\}}\|_1.
\end{equation}
Equation \ref{preprocessing} is solved using a full grid search. We use the result to crop both training and test images. Upon manual inspection of the cropped images, close to all images were perfectly cropped. We replace the handful poorly cropped training images with hand-cropped counterparts.
\begin{figure}[!ht]
\centering
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/L_147.png}
\caption{Cropped L pose}
\label{fig:cropped_L}
\end{subfigure}
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/index_28.png}
\caption{Cropped index pose}
\label{fig:cropped_index}
\end{subfigure}
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/L_pose_with_mask.png}
\caption{Black-mask perturbation}
\label{fig:cropped_L_mask}
\end{subfigure}
\caption{In Figure \ref{fig:cropped_L} and \ref{fig:cropped_index} we show an example of the images cropped using Equation \ref{preprocessing}. We see that the hands are centered and the images have a size of $200 \times 200$. In Figure \ref{fig:cropped_L_mask} we show an example of the square black-mask perturbation.}
\label{fig:preprocessing}
\end{figure}
\paragraph{Square-mask perturbations}
Since we use logistic regression, we perform a full grid search to find the best adversarial perturbation at training and test time. For completeness, the upper left coordinates of the optimal black-mask perturbation of size $\eps_{\text{tr}} \times \eps_{\text{tr}}$ can be found as the solution to
\begin{equation}
\label{square_perturbations_logistic_regression}
\text{arg}\max_{i \in [200-\eps_{\text{tr}}], \Hquad j \in [200-\eps_{\text{tr}}]} \sum_{l,m \in [\eps_{\text{tr}}]}\theta_{[i:i+l,j:j+m]}.
\end{equation}
The algorithm is rather slow as we iterate over all possible windows. We show a black-mask perturbation on an $L$-pose image in Figure \ref{fig:cropped_L_mask}.
\paragraph{Results} We run adversarial logistic regression with square-mask perturbations on the cropped dataset and vary the adversarial training budget and plot the result in Figure \ref{fig:eps_mask}. We observe attack that adversarial logistic regression deteriorates robust generalization.
Because we use adversarial logistic regression, we are able to visualize the classifier. Given the classifier induced by $\theta$, we can visualize how it classifies the images by plotting $\frac{\theta - \min_{i \in [d]}\theta_{[i]}}{\max_{i \in [d]}\theta_{[i]}} \in [0,1]^{d}$. Recall that the class-prediction of our predictor for a data point $(x,y)$ is given by $\text{sign}(\theta^{\top} x) \in \{\pm 1\}$. The lighter parts of the resulting image correspond to the class with label $1$ and the darker patches with the class corresponding to label $-1$.
\begin{wrapfigure}{r}{0.4\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/mask_plot_main.png}
\caption{We plot the standard error and robust error for varying adversarial training budget $\eps_{\text{tr}}$. We see that the larger $\eps_{\text{tr}}$ the higher the robust error.}
\label{fig:eps_mask}
\end{wrapfigure}
We plot the classifiers obtained by standard logistic regression and adversarial logistic regression with training adversarial budgets $\eps_{\text{tr}}$ of $10$ and $25$ in Figure \ref{fig:visulation_log}. The darker parts in the classifier correspond to patches that are typically bright for the $L$-pose. Complementary, the lighter patches in the classifier correspond to patches that are typically bright for the index pose. We see that in the case of adversarial logistic regression, the background noise is much higher than for standard logistic regression. In other words, adversarial logistic regression puts more weight on non-signal parts in the images to classify the training dataset and hence exhibits worse performance on the test dataset.
\newpage
\begin{figure}[!ht]
\centering
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/natural_log_regr_result.png}
\caption{$\eps_{\text{tr}} = 0 $}
\label{fig:log_natural}
\end{subfigure}
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/perTrain10logisticReg.png}
\caption{$\eps_{\text{tr}} = 10 $}
\label{fig:log_e10}
\end{subfigure}
\begin{subfigure}{0.31\textwidth}
\centering
\includegraphics[width=.80\linewidth]{plotsAistats/perTrain25logisticReg.png}
\caption{$\eps_{\text{tr}} = 25$}
\label{fig:log_e25}
\end{subfigure}
\caption{We visualize the logistic regression solutions. In Figure \ref{fig:log_natural} we plot the vector that induces the classifier obtained after standard training. In Figure \ref{fig:log_e10} and Figure \ref{fig:log_e25} we plot the vector obtained after training with square-mask perturbations of size $10$ and $25$, respectively. We note the non-signal enhanced background correlations at the parts highlighted with the red circles in the image projection of the adversarially trained classifiers. }
\label{fig:visulation_log}
\end{figure}
\section{Adversarial training hurts robust generalization for nonlinear feature learning}
\label{sec:app_theorycs}
\fy{i've no idea what your statement was, it was a whole blurb was redundant definitions so this butchering is the fastest guess after scanning it}
In this section, we give a mathematical explanation for the effect of
adversarial training with directed attacks\xspace increasing the robust error
for nonlinear feature learning models. In particular, we construct a
dataset, the concentric spheres dataset, that has exactly one discriminative
feature: the norm of the datapoints. Figure \ref{fig:app_cs_repeat}
and Figure \ref{fig:cs_numsamp_rob} show that the behaviour of the
feature learning model on our synthetic setting matches the behaviour
we observe on the linear synthetic \emph{and} on the real-world
datasets: in the low sample size regime, adversarial training
increasingly hurts robust generalization with increasing perturbation
set size.
More concretely, we discuss a two-layer neural network and conclude the same intuitive explanation as in the linear example. First, we introduce the dataset and model. Then, we discuss some theoretical results. Lastly, we plot and discuss experiments.
\subsection{Problem Setting}
In this subsection, we first introduce the concentric spheres distribution, 2-layer quadratic neural networks and the directed attack\xspace that we consider. Then, we show that the optimal robust classifier is included in our function space.
\paragraph{Distribution for concentric spheres}
We study the concentric spheres distribution as also used in
\cite{gilmer18, kolter19}. In particular, for
$0<\radius_{-1}<\radius_{1}$, we draw $(x,y) \sim \mathbb{P}_{\text{CS}}$ as follows: we draw a binary label $y\in \{+1, -1\}$
equiprobably and a covariate vector $x \in \mathbb{R}^{d}$ that is,
conditional on the label, distributed uniformly on the sphere of
radius $R_{y}$.
\paragraph{Perturbation sets}
In this example, the radius of the input corresponds to the signal,
hence, for a training perturbation size $0< \eps_{\text{tr}} <
\frac{\radius_{1}-\radius_{-1}}{2}$ and a covariate $x$ we may define
a directed attack\xspace as an attack out of the perturbation set
\begin{equation}
\label{eq:pertsetsphere}
\mathcal{S}_C(x, \eps_{\text{tr}}) = \left\{\delta \in \mathbb{R}^{d} \mid \delta = \frac{x}{\norm{x}_2}\eta, \Hquad |\eta|<\eps_{\text{tr}}\right\}.
\end{equation}
\paragraph{Neural network classifier}
Similar to prior work on concentric spheres such as \cite{gilmer18},
we consider two-layer neural networks with quadratic activations as
our parameterized function class with
\begin{equation*}
f_\theta(x) = \left(x^T W_1 \right)^2 W_2 + b,
\end{equation*}
where $\theta =(W_1 , W_2,b)$ and $W_1 \in \mathbb{R}^{d \times p}$, $ W_2 \in \mathbb{R}^p$, $b\in \mathbb{R}$. Every function induces a decision boundary defined by
\begin{equation}
\label{eq:decisionboundary}
db(f_\theta) = \{x \in \mathbb{R}^{d} \mid f_{\theta}(x) = 0\}.
\end{equation}
\fy{defined by} We note the function space of all neural networks as
$\mathcal{F}_{\text{QNN}} = \{f_\theta(x): W_1 \in \mathbb{R}^{d \times p}, W_2 \in
\mathbb{R}^p\}$.
In particular, the function space includes a \fy{ perfectly robust classifier: this an expression that is not defined}.
\fy{more like a fact than lemma}
\begin{lemma}
If $p>d$, the function space $\mathcal{F}_{\text{QNN}}$ contains a classifier that minimizes the robust error against perturbations~\eqref{eq:pertsetsphere} defined by the distribution $\mathbb{P}_{\text{CS}}$.
\end{lemma}
\fy{given $f_\theta$ what even is the $db(f_\theta)$? - i fixed it}
\begin{proof}
Clearly, for any consistent $\eps_{\text{te}}$, one perfectly robust classifier is a classifier with decision boundary ($db\left(f_{\theta}\right)$) the sphere with radius $R_{opt} = \frac{\radius_{-1}+\radius_{1}}{2}$. For a visualization see Figure \ref{fig:teaser_concentric_spheres}. Hence, it suffices to show that $\mathcal{F}_{\text{QNN}}$ includes a function that induces a decision boundary
that is the sphere with radius $R_{opt}$.
When $p>d$, choosing
\begin{equation*}
W_1 = \begin{pmatrix}
I_d & 0
\end{pmatrix},
\end{equation*}
$W_2 = 1_{\{p\}}$ and $b = -\radius_{-1}^2-\frac{\radius_{1}^2-\radius_{-1}^2}{2}$induces the decision boundary of $db\left(f_{\theta}\right)$ that is equivalent to a sphere of radius $R_{opt}$.
Note that this is only one particular parameter constellation. In fact, there exist infinitely many $\theta$ that induce the same decision boundary.
\end{proof}
\subsection{Geometric characterization of the two layer quadratic neural network}
\fy{this is again super poor language}
A decision boundary that is ellipsoid uses primarily the signal (norm), else hyperboloid, using angular information (useless features).
In experiments, we show that adversarial training
learns networks with hyperboloids as decision boundary. In
contrast, standard training leads to an ellipsoid.
This explains why the ``phenomenon'' also appears for CS
observed in experiments.
In this section we describe how we can quantify and
plot the ``hyperboloidity'' in learned
classifiers with respect to $\eps_{\text{tr}}$ \fy{why not numsamp}
\fy{this is A HUGE SECTION for just one plot of explanation ... }
\paragraph{Decision boundary of a two layer quadratic network.} To ease the flow of the text, we introduce a lemma close to the computation made in \cite{gilmer18}, which brings the quadratic neural network to a classical known form, here. We provide the proof in Subsection\ref{subsec:proof_lemma}.
\begin{lemma}
\label{lem:quadratic_symm_matrix}
For any 2-layer quadratic neural network with $p>d$, there exists a real symmetric matrix $A \in \mathbb{R}^{d \times d}$ such that
\begin{equation}
f_{\theta}(x) = x^{\top} A x + b,
\end{equation}
for any $x \in \mathbb{R}^{d}$.
\end{lemma}
Let $A, b$ be the characterization of a two-layer quadratic neural
network as per Lemma \ref{lem:quadratic_symm_matrix}. Then, recalling
the definition of a decision boundary ~\eqref{eq:decisionboundary}
induced by $f_\theta$, we can define $A_{db}= -\frac{A}{b}$ such that
\begin{equation}
\label{db_quadrnetwork}
db(f_\theta) = \{ x \in \mathbb{R}^{d} \mid x^{\top} A_{db} x = 1 \},
\end{equation}
where we note that $A_{db}$ is a real symmetric matrix.
\fy{another fact}
\begin{fact}
Let $\lambda^{\theta}$ be the vector with as entries all eigenvalues of
$A_{db}$ induced by $f_{\theta}$. If $\lambda^{\theta}_{i} > 0$ for all $i
\in [d]$, then $db(f_{\theta})$ is an ellipsoid, otherwise,
$db(f_{\theta})$ is an hyperboloid.
\end{fact}
\fy{this should go to experimental section, here its theory still} See Figure
\ref{fig:teaser_concentric_spheres} for a visualization of ellipsoids and hyperboloids \fy{this is actually already experimental}.
\fy{future work would be to show that this is indeed what happens}
\paragraph{The dissimilarity score.}
\fy{please use a macro for this score ...}
Since we cannot visualize the dec. boundaries in high dimensions, we can characterize how close the decision boundary is to the truth by calculating the ...
\fy{what is this notation $1_{d}$ -> please use macro and fix}
Observe that any robust optimal two-layer quadratic neural network has $\lambda^{\theta}_i = \lambda_{opt} = \frac{4}{(\radius_{-1}+\radius_{1})^2}$ for all $i \in [d]$. We define our dissimilarity score as follows
\begin{equation}
\text{dissim}(f_{\theta}) := \frac{1}{d}\norm{\lambda^{\theta}-1_{\{d\}}\lambda_{opt}}_2.
\end{equation}
We note the following properties of our dissimilarity score:
\fy{what the hell is happening here, the $R$ don't have values here yet do they?}
\begin{enumerate}
\item $\text{Dissim}(f_{\theta})=0$ if and only if $f_{\theta}$ is a perfect robust classifier.
\item If $f_{\theta}$ achieves perfect standard accuracy, then $\text{dissim}(f_{\theta})< \sqrt{\frac{1}{d}}(\lambda_{opt}-\frac{1}{\radius_{1}^2}) = 1.03 \cdot 10^{-3}$.
\item Given we classify a training dataset correctly, if $\text{dissim}(f_{\theta}) > \sqrt{\frac{1}{d}}(\lambda_{opt}-\frac{1}{\radius_{1}^2})$, then $db(f_{\theta})$ is necessarily a hyperboloid. Moreover, the larger $\text{dissim}(f_{\theta})$, the more skewed the hyperboloid.
\item If $\text{Dissim}(f_{\theta})$ is large, we have either a stretched out ellipsoid or a sharp hyperboloid. See Figure \ref{fig:teaser_concentric_spheres} for a visualisation of a 2D cut of a hyperboloid and an ellipsoid.
\end{enumerate}
Intuitively, the larger the dissimilarity score, the worse the robust accuracy of the classifier, because the classifier uses more angular information to interpolate the training points.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\linewidth]{plotsAistats/CS_teaser.png}
\caption{2D cut along the first two dimensions of the concentric spheres
example for $d=500$ to visualize the decision boundaries obtained via adversarial (left) and standard training (right) of a two-layer network with quadratic activations on training points not shown. The learned robust classifier has an hyperbolic decision boundary and uses angle information for classification, whereas the standard classifiers perfectly separates the classes.
}
\label{fig:teaser_concentric_spheres}
\end{figure}
\subsection{Experimental details on concentric spheres example}
\label{sec:app_expcs}
\fy{what did you do here before? how were these two sections? why figure 3b?}
In this section, we further study the concentric spheres example experimentally and give experimental details on Figure \ref{fig:eps_cs}. More precisely, we observe that attack-model overfitting on the concentric spheres dataset is possible for multiple adversarial test perturbation budgets $\eps_{\text{te}}$.
\paragraph{Experimental details to Figure \ref{fig:eps_cs}}
We sample $5$ datasets of size $n =10^5$ samples with varying dimensions $\dim{} = 350, \Hquad 500$ and $750$ of the concentric spheres distribution with radii $\radius_{-1} = 1$ and $\radius_{1} = 1.3$. The results we plot in Figure \ref{fig:eps_cs} are the average robust accuracies over the $5$ datasets and the shaded areas the respective standard deviations.
For optimization, we use Tensorflow \cite{tensorflow2015-whitepaper} and its Adam optimizer with a learning rate of $0.001$ and a batch size of $10$. We train for $100$ epochs or until all training points are correctly classified with a two-layer squared neural network of width $p = 1000$. We implement adversarial training by solving the inner maximization using
\begin{equation}
\label{eq:csATmaximization}
x' = x-\eps_{\text{tr}}\frac{x}{\norm{x}_2}\text{sign}\left(x^T \partial_x f_{\theta}(x)\right).
\end{equation}
\subsection{Experimental results and discussion}
In this subsection we show the experimental results and discuss their implications.
In all experiments, we adversarially train a two-layer neural network with quadratic
activations and width $1000$ on the concentric spheres dataset with
$\radius_{-1}=1$ and $\radius_{1}=11$. We minimize the cross-entropy loss until
convergence. More experimental details can be found in
Section~\ref{sec:app_expcs}.
\begin{figure}[!t]
\vskip 0.2in
\centering
\begin{subfigure}[b]{0.33\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cs_eps_standard_app.png}
\caption{Standard accuracy}
\label{fig:app_cs_dissim}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cs_eps_main.png}
\caption{Robust accuracy}
\label{fig:app_cs_repeat}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cs_dissim_app.png}
\caption{Dissimilarity score}
\label{fig:app_cs_dissim}
\end{subfigure}
\caption{Experiments with the 2-layer quadratic network on the concentric spheres dataset with different adversarial budgets (x-axis). We use an input dimension of $d = 500$. Note that the robust and standard accuracy monotonically decrease for increasing $\eps_{\text{tr}}$. Moreover, the dissimilarity score monotonically increases with increasing $\eps_{\text{tr}}$, meaning that the network converge to sharper hyperbolic solutions, and hence uses more angular information to classify the training points. See Subsection \ref{sec:app_expcs} for experimental details.}
\label{fig:app_eps_cs}
\end{figure}
\paragraph{AT hurts robust generalization}
\fy{show by Varying the perturbation set size $\eps_{\text{tr}}$}
To understand the effect of increasing adversarial training budget from standard training $(\eps_{\text{tr}} = 0)$ to training with a large adversarial budget, we perform several experiments. We fix the dimension to be $d=500$, choose varying dataset sizes $n = 50, 100$ and $200$ and log the standard accuracy, robust accuracy ($\eps_{\text{te}} = 3$) and dissimilarity score. We plot the results In Figure \ref{fig:app_eps_cs}. Observe that similar to the linear case and the real-world experiments the robust accuracy decreases with increasing adversarial training budget $\eps_{\text{tr}}$. Moreover, the aggravating trend is more severe when $\frac{d}{n}$ is large.
\paragraph{Explanation: AT makes DB more hyperboloid}
To understand the change in decision boundaries, we look at the dissimilarity score. In Figure \ref{fig:app_cs_dissim}, we note that the dissimilarity score monotonically increases with increasing $\eps_{\text{tr}}$. In particular, we see that the dissimilarity score is strictly larger than $1.03 \cdot 10^{-3}$ for all $\frac{d}{n}$ when $\eps_{\text{tr}} > 2.5$. By property 3 of the dissimilarity score, listed in the subsection above, this means that for large $\eps_{\text{tr}}$, the decision boundary is an hyperboloid; the classifier uses angular information the interpolate the training dataset. Moreover, the larger the dissimilarity score the more sharp the hyperboloid. For visualization of a hyperbolic and ellipsoidal decision boundary, we refer to Figure \ref{fig:teaser_concentric_spheres}.
Lastly, we also investigate the robustness score with varying $\eps_{\text{tr}}$. In Figure \ref{fig:cs_trade_off}, we plot the decomposition after adversarial training with increasing $\eps_{\text{tr}}$ and $n = 50$. Similar to the linear example, plotted in Figure \ref{fig:app_tradeoff_logreg}, we recognize an U-shape. Together with the increasing dissimilarity score, we can make the following arguments. For large $\eps_{\text{tr}}$, when the dissimilarity score is also large, we converge to networks with sharp hyperboloid decision boundaries. These are highly robust but have low standard accuracy. In contrast, when using standard training ($\eps_{\text{tr}} = 0$) we converge to decision boundaries close to the optimal robust one, which has high standard accuracy and robustness.
In summary, first steering away from the optimal decision boundary (increasing $\eps_{\text{tr}}$) decreases standard accuracy and robust accuracy. Thereafter, when the decision boundary is a hyperboloid, increasing $\eps_{\text{tr}}$ causes to further decrease standard accuracy, but increase the robustness. The increase in robustness is the result of converging to a sharper hyperboloid decision boundary, which uses less the norm of the samples to classify them.
\begin{figure}[!b]
\vskip 0.2in
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cs_rob_acc_numsamp_half.png}
\caption{Robust accuracy}
\label{fig:numobs}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/cs_trade_off_decomposition_app.png}
\caption{Standard-robust accuracy trade-off}
\label{fig:cs_trade_off}
\end{subfigure}
\caption{ We set $d = 500, \radius_{-1} = 1, \radius_{1} = 11$ and $\eps_{\text{te}} = 3$. (a) Adversarial training on the concentric spheres dataset with increasing sample size. We see that for low sample sizes, adversarial training hurts robust accuracy, but for high sample sizes, we recognize the known regime where it helps robust generalization. (b) Robust accuracy decomposition of adversarial training with increasing perturbation budget $\eps_{\text{tr}}$. For large $\eps_{\text{tr}}$, we note how the robust accuracy decreases, while the robustness increases. The decrease is hence a result of decreasing standard accuracy. See Subsection \ref{sec:app_expcs} for experimental details.}
\label{fig:numobs_trade_off}
\end{figure}
\subsection{Proof of Lemma \ref{lem:quadratic_symm_matrix}}
\label{subsec:proof_lemma}
Let us start by recalling the two-layer squared neural network. A two-layer neural network is a function $f:\mathbb{R}^d\xrightarrow{}\mathbb{R}$, of the form
\begin{equation*}
f_{\theta}(x) = \left(x^T W_1\right)^2 W_2+b.
\end{equation*}
We rewrite this equation in a quadratic form:
\begin{equation*}
\begin{split}
f(x) &= \left(\sum_{i=1}^d x_i W_{1,i}\right)^2 W_2 + b\\
&= \sum_{j=1}^p\left(\sum_{i=1}^d x_i W_{1,i}\right)^2_j W_{2,j} + b\\
&= \sum_{i,j = 1}^d a_{i,j}x_i x_j + b\\
&= x^T A x + b,
\end{split}
\end{equation*}
where
\begin{equation*}
a_{i,j} = \begin{cases}
\sum_{m=1}^p W_{1,i,m}^2 W_{2,m} & \text{if i = j},\\
\sum_{m=1}^p W_{1,i,m} W_{1,j,m} W_{2,m} & \text{if i $\neq$ j.}
\end{cases}
\end{equation*}
Hence, the decision boundary of $f$ is a quadratic equation and any two-layer quadratic neural network can be written in the form $f(x) = x^T A x + b$.
\subsection{Attack-model overfitting for multiple adversarial test budgets $\eps_{\text{te}}$}
The choice of $\eps_{\text{te}} = 0.075$ is reasonable but somewhat arbitrary. Hence, we also conduct the same experiment with experimental details as in Section \ref{app_csexpdetails_main} with a different adversarial test perturbation budget $\eps_{\text{te}}=0.1$ and include standard accuracy ($\eps_{\text{te}}=0$). We plot the results of the experiments in Figure \ref{fig:n_d_exp_robust}. Again, we observe attack-model overfitting.
\section{Bounds on the susceptibility score}
\label{app:susc}
In Theorem \ref{thm:linlinf}, we give non-asymptotic bounds on the robust and standard error of a linear classifier trained with adversarial logistic regression. Moreover, we use the robust error decomposition in susceptibility and standard error to gain intuition about how adversarial training may hurt robust generalization. In this section, we complete the result of Theorem \ref{thm:linlinf} by also deriving non-asymptotic bounds on the susceptibility score of the max $\ell_2$-margin classifier.
Using the results in Appendix \ref{sec:app_theorylinear}, we can prove following Corollary \ref{cor:robustness}, which gives non asymptotic bounds on the susceptibility score.
\begin{corollary}
\label{cor:robustness}
Assume $d-1>n$. For the $\eps_{\text{te}}$-susceptibility on test samples from $\mathbb{P}_{r}$ with $2 \eps_{\text{te}} < r$ and perturbation sets in Equation~\eqref{eq:linfmaxpert} and~\eqref{eq:l1maxpert} the following holds:
For $\eps_{\text{tr}} < \frac{r}{2} - \tilde{\gamma}_{\max}$, with probability at least $1-2\text{e}^{-\frac{\alpha^2 (d-1)}{2}}$ for any $0<\alpha<1$, over the draw of a dataset $D$ with $n$ samples from $\mathbb{P}_{r}$, the $\eps_{\text{te}}$-susceptibility is upper and lower bounded by
\begin{equation}
\begin{split}
&\suscept{\thetahat{\epstrain}} \leq \Phi \left(\frac{(r-2 \eps_{\text{tr}}) (\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}_{\max} \sigma}\right) - \Phi \left( \frac{(r-2 \eps_{\text{tr}})( -\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}_{\min}\sigma} \right)\\
&\suscept{\thetahat{\epstrain}} \geq \Phi \left(\frac{(r-2 \eps_{\text{tr}}) (\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}_{\min}\sigma}\right) - \Phi \left( \frac{(r-2 \eps_{\text{tr}})( -\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}_{\max} \sigma} \right)
\end{split}
\end{equation}
\end{corollary}
We give the proof in Subsection \ref{sec:proof_robust_cor}. Observe that the bounds on the susceptibility score in Corollary \ref{cor:robustness} consist of two terms each, where the second term decreases with $\eps_{\text{tr}}$, but the first term increases. We recognise following two regimes: the max $\ell_2$-margin classifier is close to the ground truth $e_1$ or not. Clearly, the ground truth classifier has zero susceptibility and hence classifiers close to the ground truth also have low susceptibility. On the other hand, if the max $l_2$-margin classifier is not close to the ground truth, then putting less weight on the first coordinate increases invariance to the perturbations along the first direction. Recall that by Lemma \ref{lem:maxmargin}, increasing $\eps_{\text{tr}}$, decreases the weight on the first coordinate of the max $\ell_2$-margin classifier. Furthermore, in the low sample size regime, we are likely not close to the ground truth. Therefore, the regime where the susceptibility decreases with increasing $\eps_{\text{tr}}$ dominates in the low sample size regime.
To confirm the result of Corollary \ref{cor:robustness}, we plot the mean and standard deviation of the susceptibility score of $5$ independent experiments. The results are depicted in Figure \ref{fig:logreg_robust}. We see that for low standard error, when the classifier is reasonably close to the optimal classifier, the susceptibility increases slightly with increasing adversarial budget. However, increasing the adversarial training budget, $\eps_{\text{tr}}$, further, causes the susceptibility score to drop greatly. Hence, we can recognize both regimes and validate that, indeed, the second regime dominates in the low sample size setting.
\begin{figure*}[!b]
\centering
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{plotsAistats/app_susceptibilty.png}
\caption{Susceptibility score decreases with $\eps_{\text{tr}}$}
\label{fig:app_robustness}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{plotsAistats/logreg_trade_off_plot.png}
\caption{Robust error decomposition}
\label{fig:app_tradeoff_logreg}
\end{subfigure}
\caption{We set $r = 6$, $d = 1000$, $n = 50$ and $\eps_{\text{te}} = 2.5$. (a) We plot the average susceptibility score and the standard deviation over 5 independent experiments. Note how the bounds closely predict the susceptibility score. (b) For comparison, we also plot the robust error decomposition in susceptibility and standard error. Even though the susceptibility decreases, the robust error increases with increasing adversarial budget $\eps_{\text{tr}}$.}
\vspace{-0.2in}
\label{fig:logreg_robust}
\end{figure*}
\subsection{Proof of Corollary \ref{cor:robustness}}
\label{sec:proof_robust_cor}
We proof the statement by bounding the robustness of a linear classifier. Recall that the robustness of a classifier is the probability that a classifier does not change its prediction under an adversarial attack. The susceptibility score is then given by
\begin{equation}
\label{eq:rob_sus}
\suscept{\thetahat{\epstrain}} = 1 - \robness{\thetahat{\epstrain}}.
\end{equation}
The proof idea is as follows: since the perturbations are along the first basis direction, $e_1$, we compute the distance from the robust $l_2$-max margin $\thetahat{\epstrain}$ to a point $(X,Y) \sim \mathbb{P}$. Then, we note that the robustness of $\thetahat{\epstrain}$ is given by the probability that the distance along $e_1$, from $X$ to the decision plane induced by $\thetahat{\epstrain}$ is greater then $\eps_{\text{te}}$. Lastly, we use the non-asymptotic bounds of Lemma \ref{lem:boundsmaxmargin}.
Recall, by Lemma \ref{lem:maxmargin}, the max $l_2$-margin classifier is of the form of
\begin{equation}
\label{eq:robustmaxmarg}
\thetahat{\epstrain} = \frac{1}{\sqrt{(r-2 \eps_{\text{tr}})^2 + 4 \tilde{\gamma}^{2}}}\left[r-2\eps_{\text{tr}}, 2 \tilde{\gamma} \tilde{\theta} \right].
\end{equation}
Let $(X, Y) \sim \mathbb{P}$. The distance along $e_1$ from $X$ to the decision plane induced by $\thetahat{\epstrain}$, $\decplanegen{\thetahat{\epstrain} }$, is given by
\begin{equation*}
d_{e_1}(X, \decplanegen{\thetahat{\epstrain}}) = \left| \indof{X}{1}+ \frac{1}{ \indof{\thetahat{\epstrain}}{0}} \sum_{i=2}^{ d } \indof{\thetahat{\epstrain}}{i} \indof{X}{i} \right|.
\end{equation*}
Substituting the expression of $\thetahat{\epstrain}$ in Equation \ref{eq:robustmaxmarg} yields
\begin{equation*}
d_{e_1}(X, \decplanegen{\thetahat{\epstrain}}) = \left| \indof{X}{1} + 2 \tilde{\gamma} \frac{1}{(r-\eps_{\text{tr}})} \sum_{i=2}^{d} \indof{\tilde{\theta}}{i} \indof{X}{i} \right|.
\end{equation*}
Let $N$ be a standard normal distributed random variable. By definition $\| \tilde{\theta}\|_2^2 = 1$ and using that a sum of Gaussian random variables is again a Gaussian random variable, we can write
\begin{equation*}
d_{e_1}(X,\decplanegen{\thetahat{\epstrain}}) = \left| \indof{X}{1} + 2 \tilde{\gamma} \frac{\sigma}{(r-\eps_{\text{tr}})} N \right|.
\end{equation*}
The robustness of $\thetahat{\epstrain}$ is given by the probability that $d_{e_1}(X,\decplanegen{\thetahat{\epstrain}}) > \eps_{\text{te}}$. Hence, using that $X_1 = \pm \frac{r}{2}$ with probability $\frac{1}{2}$, we get
\begin{equation}
\label{eq:robustness_form}
\robness{\thetahat{\epstrain}} = P\left[ \frac{r}{2} + 2 \tilde{\gamma} \frac{\sigma}{(r-2\eps_{\text{tr}})} N > \eps_{\text{te}} \right] + P \left[ \frac{r}{2} + 2 \tilde{\gamma} \frac{\sigma}{(r-\eps_{\text{tr}})} N < -\eps_{\text{te}} \right].
\end{equation}
We can rewrite Equation \ref{eq:robustness_form} in the form
\begin{equation*}
\robness{\thetahat{\epstrain}} = P \left[ N > \frac{(r-2\eps_{\text{tr}}) (\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}\sigma} \right] + P \left[ N < \frac{(r-2\eps_{\text{tr}})( -\eps_{\text{te}} -\frac{ r}{2})}{2 \tilde{\gamma}\sigma} \right].
\end{equation*}
Recall, that $N$ is a standard normal distributed random variable and denote by $\Phi$ the cumulative standard normal density. By definition of the cumulative denisity function, we find that
\begin{equation*}
\robness{\thetahat{\epstrain}} = 1 - \Phi \left(\frac{(r-2\eps_{\text{tr}}) (\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}\sigma}\right) + \Phi \left( \frac{(r-2 \eps_{\text{tr}})( -\eps_{\text{te}} - \frac{r}{2})}{2 \tilde{\gamma}\sigma} \right).
\end{equation*}
Substituting the bounds on $\tilde{\gamma}$ of Lemma \ref{lem:boundsmaxmargin} gives us the non-asymptotic bounds on the robustness score and by Equation \ref{eq:rob_sus} also on the susceptibility score.
\section{Experimental details on the linear model}
\label{sec:logregapp}
In this section, we provide detailed experimental details to Figures \ref{fig:main_theorem} and \ref{fig:lineartradeoff}.
We implement adversarial logistic regression using stochastic gradient descent with a learning rate of $0.01$. Note that logistic regression converges logarithmically to the robust max $l_2$-margin solution. As a consequence of the slow convergence, we train for up to $10^7$ epochs. Both during training and test time we solve $\max_{x_i' \in \pertset{x_i}{\eps_{\text{tr}}}} L(f_\theta(x_i') y_i)$ exactly. Hence, we exactly measure the robust error.
Unless specified otherwise, we set $\sigma= 1$, $r = 12$ and $\eps_{\text{te}} = 4$.
\paragraph{Experimental details on Figure \ref{fig:main_theorem}} (a) We draw $5$ datasets with $n= 50$ samples and input dimension $d=1000$ from the distribution $\mathbb{P}$. We then run adversarial logistic regression on all $5$ datasets with adversarial training budgets, $\eps_{\text{tr}} = 1$ to $5$. To compute the resulting robust error gap of all the obtained classifiers, we use a test set of size $10^{6}$. Lastly, we compute the lower bound given in part 2. of Theorem \ref{thm:linlinf}. (b) We draw $5$ datasets with different sizes $n$ between $50$ and $10^4$. We take an input dimension of $d = 10^4$ and plot the mean and standard deviation of the robust error after adversarial and standard logistic regression over the $5$ samples.(c) We again draw $5$ datasets for each $d/n$ constellation and compute the robust error gap for each dataset.
\paragraph{Experimental details on Figure \ref{fig:lineartradeoff}} For both (a) and (b) we set $d = 1000$, $\eps_{\text{te}} = 4$, and vary the adversarial training budget ($\eps_{\text{tr}}$) from $1$ to $5$. For every constellation of $n$ and $\eps_{\text{tr}}$, we draw $10$ datasets and show the average and standard deviation of the resulting robust errors. In (b), we set $n = 50$.
\section{Theoretical statements for the linear model}
\label{sec:app_theorylinear}
Before we present the proof of the theorem, we introduce two lemmas are of separate interest that are used throughout the proof of Theorem 1. Recall that the definition of the (standard normalized) maximum-$\ell_2$-margin solution (max-margin solution in short) of a dataset $D =\{(x_i, y_i)\}_{i=1}^n$ corresponds to
\begin{equation}
\label{eq:stdmaxmargin}
\thetahat{} := \argmax_{\|\theta\|_2\leq 1} \min_{i\in [n]} y_i \theta^\top x_i,
\end{equation}
by simply setting $\eps_{\text{tr}} = 0$ in Equation~\eqref{eq:maxmargin}. The $\ell_2$-margin of $\thetahat{}$ then reads $\min_{i\in[n]} y_i \thetahat{\top} x_i$. Furthermore for a dataset $D = \{(x_i, y_i)\}_{i=1}^n$ we refer to the induced dataset $\widetilde{D}$ as the dataset with covariate vectors stripped of the first element, i.e.
\begin{equation}
\widetilde{D} = \{(\tilde{x}_i, y_i)\}_{i=1}^n := \{ ((x_i)_{[2:d]}, y_i) \}_{i=1}^n,
\end{equation}
where $(x_i)_{[2:d]}$ refers to the last $d-1$ elements of the vector $x_i$. Furthermore, remember that for any vector $z$, $\indof{z}{j}$ refers to the $j$-th element of $z$ and $e_j$ denotes the $j$-th canonical basis vector.
Further, recall the distribution $\mathbb{P}_r$ as defined in
Section~\ref{logreg_linear_model}: the label $y \in \{+1, -1\}$ is
drawn with equal probability and the covariate vector is sampled as $x
= [y\frac{r}{2}, \tilde{x}]$ where $\tilde{x} \in \mathbb{R}^{d-1}$ is
a random vector drawn from a standard normal distribution,
i.e. $\tilde{x} \sim \mathcal{N}(0, \sigma^2 I_{d-1})$. We generally allow
$r$, used to sample the training data, to differ from $r_{\text{test}}$, which is
used during test time.
The following lemma derives a closed-form expression for the normalized max-margin solution for any dataset with fixed separation $r$ in the signal component, and that is linearly separable in the last $d-1$ coordinates with margin $\tilde{\gamma}$.
\begin{lemma}
\label{lem:maxmargin}
Let $D = \{(x_i,y_i)\}_{i=1}^{n}$ be a dataset that
consists of points $(x,y) \in \mathbb{R}^{d}\times\{\pm 1\}$ and
$\xind{1} = y\frac{r}{2}$, i.e. the covariates $x_i$ are
deterministic in their first coordinate given $y_i$ with
separation distance $r$. Furthermore, let the induced dataset
$\widetilde{D}$ also be linearly separable by the normalized
max-$\ell_2$-margin solution $\tilde{\theta}$ with an $\ell_2$-margin
$\tilde{\gamma}$. Then, the normalized max-margin solution of the
original dataset $D$ is given by
\begin{equation}
\label{eq:lemmaxmargin}
\thetahat{} = \frac{1}{\sqrt{r^2 + 4 \tilde{\gamma}^{2}}}\left[r, 2 \tilde{\gamma} \tilde{\theta} \right].
\end{equation}
Further, the standard accuracy of $\thetahat{}$ for data drawn from $\mathbb{P}_{r_{\text{test}}}$ reads
\begin{equation}
\label{eq:stdaccmaxmargin}
\mathbb{P}_{r_{\text{test}}}(Y \thetahat{\top} X > 0) = \Phi\left(
\frac{r \:r_{\text{test}} }{4\sigma\: \tilde{\gamma}} \right).
\end{equation}
\end{lemma}
The proof can be found in Section~\ref{sec:maxmarginproof}. The next lemma provides high probability upper and lower bounds
for the margin $\tilde{\gamma}$ of $\widetilde{D}$ when $\tilde{x}_i$ are drawn from the normal distribution.
\begin{lemma}
\label{lem:boundsmaxmargin}
Let $\widetilde{D}=\{(\Tilde{x}_i,y_i)\}_{i=1}^{n}$ be a random dataset where $y_i \in \{\pm 1\}$ are equally distributed and $\tilde{x}_i \sim \mathcal{N}(0,\sigma I_{d-1})$ for all $i$, and $\tilde{\gamma}$ is the maximum $\ell_2$ margin that can be written as
\begin{equation*}
\tilde{\gamma}= \max_{\|\tilde{\theta}\|_2 \leq 1} \min_{i \in [n]} y_i \tilde{\theta}^{\top} \Tilde{x}_i .
\end{equation*}
Then, for any $t \geq 0$, with probability greater than $1-2e^{-\frac{t^2}{2}}$, we have $\tilde{\gamma}_{\min}(t) \leq \tilde{\gamma} \leq \tilde{\gamma}_{\max}(t)$ where
\begin{align*}
\label{Crude_bounds_subsequent_maxmar}
&\tilde{\gamma}_{\max}(t) = \sigma \left( \sqrt{\frac{d-1}{n}} + 1 + \frac{t}{\sqrt{n}}\right), \:\: \tilde{\gamma}_{\min}(t)= \sigma \left( \sqrt{\frac{d-1}{n}} -1 - \frac{t}{\sqrt{n}}\right).
\end{align*}
\end{lemma}
\subsection{Proof of Theorem~\ref{thm:linlinf}}
\label{sec:thmproof}
Given a dataset $D = \{(x_i, y_i)\}$ drawn from $\mathbb{P}_r$, it is easy to see that the (normalized) $\eps_{\text{tr}}$-robust max-margin solution~\eqref{eq:maxmargin} of $D$ with respect to signal-attacking perturbations $\pertset{\eps_{\text{tr}}}{x_i}$ as defined in Equation~\eqref{eq:linfmaxpert}, can be written as
\begin{equation}
\begin{aligned}
\label{eq:robmaxmargin}
\thetahat{\eps_{\text{tr}}} &= \argmax_{\|\theta\|_2\leq 1} \min_{i\in [n], x_i' \in \pertset{x_i}{\eps_{\text{tr}}}} y_i \theta^\top x'_i \\
&= \argmax_{\|\theta\|_2\leq 1}\min_{i\in [n],|\beta|\leq \eps_{\text{tr}}}y_i \theta^\top (x_i + \beta e_1) \nonumber\\
&= \argmax_{\|\theta\|_2\leq 1} \min_{i\in [n]} y_i \theta^\top (x_i - y_i \eps_{\text{tr}} \sign(\thetaind{1}) e_1). \nonumber
\end{aligned}
\end{equation}
Note that by definition, it is equivalent to the (standard normalized)
max-margin solution $\thetahat{}$ of the shifted dataset ${D_{\epstrain} =
\{(x_i - y_i \eps_{\text{tr}} \sign(\thetaind{1}) e_1,
y_i)\}_{i=1}^n}$. Since $D_{\epstrain}$ satisfies the assumptions of
Lemma~\ref{lem:maxmargin}, it then follows directly that the
normalized $\eps_{\text{tr}}$-robust max-margin solution reads
\begin{equation}
\label{eq:appmaxmargin}
\thetahat{\eps_{\text{tr}}} = \frac{1}{\sqrt{(r -2\eps_{\text{tr}})^2 + 4 \tilde{\gamma}^{2}}}\left[r-2\eps_{\text{tr}}, 2 \tilde{\gamma} \tilde{\theta} \right],
\end{equation}
by replacing $r$ by $r - 2\eps_{\text{tr}}$ in
Equation~\eqref{eq:lemmaxmargin}. Similar to above, $\tilde{\theta} \in
R^{d-1}$ is the (standard normalized) max-margin solution of
$\{(\tilde{x}_i, y_i)\}_{i=1}^n$ and $\tilde{\gamma}$ the corresponding
margin.
\paragraph{Proof of 1.}
We can now compute the $\eps_{\text{te}}$-robust accuracy of the
$\eps_{\text{tr}}$-robust max-margin estimator $\thetahat{\eps_{\text{tr}}}$ for a
given dataset $D$ as a function of $\tilde{\gamma}$. Note that in
the expression of $\thetahat{\eps_{\text{tr}}}$, all values are fixed for a
fixed dataset, while $0\leq \eps_{\text{tr}}\leq r-2\tilde{\gamma}_{\max}$ can be chosen.
First note that for a test distribution $\mathbb{P}_r$, the
$\eps_{\text{te}}$-robust accuracy, defined as one minus the robust error (Equation~\eqref{eq:roberr}), for a classifier
associated with a vector $\theta$, can be written as
\begin{align}
\label{eq:robacc_closed}
\robacc{\theta} &= \mathbb{E}_{X,Y\sim \mathbb{P}_r} \left[\Indi{\min_{x'
\in \pertset{X}{\eps_{\text{te}}}} Y \theta^\top x'>0}\right] \\
&= \mathbb{E}_{X,Y\sim \mathbb{P}_{r}} \left[ \Indi{ Y \theta^\top X -
\eps_{\text{te}} \thetaind{1} >0}\right] = \mathbb{E}_{X,Y\sim \mathbb{P}_{r}}
\left[\Indi{ Y \theta^\top (X - Y\eps_{\text{te}} \sign(\thetaind{1}) e_1) >0}\right]
\nonumber
\end{align}
Now, recall that
by Equation~\eqref{eq:appmaxmargin} and the assumption in the
theorem, we have $r-2\eps_{\text{tr}}>0$, so that $\sign(\thetahat{\eps_{\text{tr}}})=1$.
Further, using the definition of the $\pertset{\eps_{\text{tr}}}{x}$ in
Equation~\eqref{eq:linfmaxpert} and by definition of the
distribution $\mathbb{P}_r$, we have $\indof{X}{1} = Y
\frac{r}{2}$.
Plugging into Equation~\eqref{eq:robacc_closed} then yields
\begin{align*}
\robacc{\thetahat{\eps_{\text{tr}}}}&= \mathbb{E}_{X,Y\sim \mathbb{P}_{r}} \left[\Indi{ Y \thetahat{\eps_{\text{tr}} \top} (X - Y\eps_{\text{te}} e_1) >0}\right] \\
&= \mathbb{E}_{X,Y\sim \mathbb{P}_{r}}\left[\Indi{ Y \thetahat{\eps_{\text{tr}} \top} (X_{-1} + Y\left(\frac{r}{2} - \eps_{\text{te}}\right) e_1) >0}\right] \\
&= \mathbb{P}_{r- 2 \eps_{\text{te}}} (Y\thetahat{\eps_{\text{tr}} \top} X >0 )
\end{align*}
where $X_{-1}$ is a shorthand for the random vector $X_{-1} = (0;
\indof{X}{2}, \dots, \indof{X}{d})$. The assumptions in
Lemma~\ref{lem:maxmargin} ($D_{\epstrain}$ is linearly separable) are
satisfied whenever the $n<d-1$ samples are distinct, i.e. with
probability one. Hence applying Lemma~\ref{lem:maxmargin} with
$r_{\text{test}} = r - 2\eps_{\text{te}}$ and $r = r -
2\eps_{\text{tr}}$ yields
\begin{equation}
\label{eq:arsenal}
\robacc{\thetahat{\eps_{\text{tr}}}} =
\Phi\left(\frac{r(r-2\eps_{\text{te}})}{4\sigma \tilde{\gamma}}
- \eps_{\text{tr}} \frac{r-2\eps_{\text{te}}}{2\sigma \tilde{\gamma}}\right).
\end{equation}
Theorem statement a) then follows by noting that
$\Phi$ is a monotically decreasing function in $\eps_{\text{tr}}$.
The expression for the robust error then follows by noting that $1-\Phi(-z) = \Phi(z)$ for any $z \in \mathbb{R}$
and defining
\begin{equation}
\label{eq:varphidef}
\tilde{\varphi} = \frac{\sigma \tilde{\gamma}}{r/2 - \eps_{\text{te}}}.
\end{equation}
\paragraph{Proof of 2.}
First define $\varphi_{\text{min}}, \varphi_{\text{max}}$ using $\tilde{\gamma}_{\min}, \tilde{\gamma}_{\max}$ as in Equation~\eqref{eq:varphidef}. Then we have by Equation~\eqref{eq:arsenal}
\begin{align*}
\roberr{\thetahat{\eps_{\text{tr}}}} - \roberr{\thetahat{0}} &= \robacc{\thetahat{0}} - \robacc{\thetahat{\eps_{\text{tr}}}}\\
&= \Phi\left(\frac{r/2}{\tilde{\varphi}}\right) - \Phi\left(\frac{r/2 - \eps_{\text{tr}}}{\tilde{\varphi}}\right)\\
&= \int_{r/2-\eps_{\text{tr}}}^{r/2} \frac{1}{\sqrt{2\pi}\tilde{\varphi}} \text{e}^{- \frac{x^2 }{\tilde{\varphi}^2}} d x
\end{align*}
By plugging in $t = \sqrt{\frac{2 \log 2/\delta}{n}}$ in
Lemma~\ref{lem:boundsmaxmargin}, we obtain that with probability at
least $1-\delta$ we have
\begin{equation*}
\tilde{\gamma}_{\min} := \sigma
\left[\sqrt{\frac{d-1}{n}} - \left(1+\sqrt{\frac{2 \log (2/\delta)}{n}}\right)\right] \leq \tilde{\gamma} \leq \sigma
\left[\sqrt{\frac{d-1}{n}} + \left(1+\sqrt{\frac{2 \log (2/\delta)}{n}}\right)\right] =: \tilde{\gamma}_{\max}
\end{equation*}
and equivalently $\varphi_{\text{min}} \leq \tilde{\varphi} \leq \varphi_{\text{max}}$.
Now note the general fact that for all
$\tilde{\varphi} \leq \sqrt{2} x$ the density function
$f(\tilde{\varphi}; x) = \frac{1}{\sqrt{2\pi}\tilde{\varphi}} \text{e}^{- \frac{x^2 }{\tilde{\varphi}^2}} $
is monotonically increasing in $\tilde{\varphi}$.
By assumption of the theorem, $\tilde{\varphi} \leq \sqrt{2} (r/2-\eps_{\text{tr}})(r/2-\eps_{\text{te}})$ so that $f(\tilde{\varphi}; x) \geq f(\varphi_{\text{min}};x)$ for all $x\in [r/2-\eps_{\text{tr}},r/2]$ and therefore
\begin{equation*}
\int_{r/2-\eps_{\text{tr}}}^{r/2} \frac{1}{\sqrt{2\pi}\tilde{\varphi}} \text{e}^{- \frac{x^2 }{\tilde{\varphi}^2}} d x \geq \int_{r/2-\eps_{\text{tr}}}^{r/2} \frac{1}{\sqrt{2\pi}\varphi_{\text{min}}} \text{e}^{- \frac{x^2 }{\tilde{\varphi}^2}} d x = \Phi\left(\frac{r/2}{\varphi_{\text{min}}}\right) - \Phi\left(\frac{r/2-\eps_{\text{tr}}}{\varphi_{\text{min}}}\right).
\end{equation*}
and the statement is proved.
\subsection{Proof of Corollary~\ref{cor:l1extension}}
We now show that Theorem~\ref{thm:linlinf} also holds for
$\ell_1$-ball perturbations with at most radius $\epsilon$. Following
similar steps as in Equation~\eqref{eq:appmaxmargin}, the
$\eps_{\text{tr}}$-robust max-margin solution for $\ell_1$-perturbations can
be written as
\begin{equation}
\label{eq:maxmarginl1}
\thetahat{\eps_{\text{tr}}} := \argmax_{\|\theta\|_2 \leq 1}\min_{i\in [n]} y_i \theta^\top (x_i - y_i \eps_{\text{tr}} \sign(\indof{\theta}{j^\star(\theta)}) e_{j^\star(\theta)})
\end{equation}
where $j^\star(\theta) := \argmax_j |\theta_j|$ is the index of the maximum absolute value of $\theta$.
We now prove by contradiction that the robust max-margin solution for
this perturbation set~\eqref{eq:l1maxpert} is equivalent to the solution~\eqref{eq:appmaxmargin} for the perturbation set~\eqref{eq:linfmaxpert}.
We start by assuming that $\thetahat{\epstrain}$ does not solve
Equation~\eqref{eq:appmaxmargin}, which is equivalent to assuming $1\not \in
j^\star(\thetahat{\epstrain})$ by definition. We now show how this assumption leads
to a contradiction.
Define the shorthand $\maxind := j^\star(\thetahat{\epstrain}) -1$. Since
$\thetahat{\epstrain}$ is the solution of~\eqref{eq:maxmarginl1}, by definition, we
have that $\thetahat{\epstrain}$ is also the max-margin solution of the shifted
dataset $D_{\epstrain} :=(x_i - y_i \eps_{\text{tr}} \sign(\thetaind{\maxind+1})
e_{\maxind+1}, y_i)$. Further, note that by the assumption that $1
\not \in j^\star(\thetahat{\epstrain})$, this dataset $D_{\epstrain}$ consists of input
vectors $x_i = (y_i \frac{r}{2}, \tilde{x}_i - y_i \eps_{\text{tr}}
\sign(\thetaind{\maxind+1}) e_{\maxind+1} )$. Hence via
Lemma~\ref{lem:maxmargin}, $\thetahat{\epstrain}$ can be written as
\begin{equation}
\label{eq:sml}
\thetahat{\epstrain} = \frac{1}{\sqrt{r^2 - 4 (\tilde{\gamma}^{\eps_{\text{tr}}})^2}} [r, 2 \tilde{\gamma}^{\eps_{\text{tr}}} \tilde{\theta}^{\eps_{\text{tr}}}],
\end{equation}
where $\tilde{\theta}^{\eps_{\text{tr}}}$ is the normalized max-margin solution
of $\widetilde{D} := (\tilde{x}_i
- y_i \eps_{\text{tr}} \sign(\indof{\tilde{\theta}}{\maxind}) e_{\maxind},
y_i)$.
We now characterize $\tilde{\theta}^{\eps_{\text{tr}}}$. Note that by
assumption, $\maxind = j^\star(\tilde{\theta}^{\eps_{\text{tr}}}) = \argmax_j
|\indof{\tilde{\theta}^{\eps_{\text{tr}}}}{j}|$. Hence, the normalized max-margin
solution $\tilde{\theta}^{\eps_{\text{tr}}}$ is the solution of
\begin{equation}
\label{eq:maxmarginsmall}
\tilde{\theta}^{\eps_{\text{tr}}} := \argmax_{\|\tilde{\theta}\|_2 \leq 1}
\min_{i\in [n]} y_i \tilde{\theta}^\top \tilde{x}_i - \eps_{\text{tr}}
|\indof{\tilde{\theta}}{\maxind}|
\end{equation}
Observe that the minimum margin of this estimator
$\tilde{\gamma}^{\eps_{\text{tr}}}=\min_{i\in [n]} y_i
(\tilde{\theta}^{\eps_{\text{tr}}})^\top \tilde{x}_i - \eps_{\text{tr}}
|\indof{\tilde{\theta}^{\eps_{\text{tr}}}}{\maxind}|$ decreases with
$\eps_{\text{tr}}$ as the problem becomes harder $\tilde{\gamma}^{\eps_{\text{tr}}}
\leq \tilde{\gamma}$, where the latter is equivalent to the margin of
$\tilde{\theta}^{\eps_{\text{tr}}}$ for $\eps_{\text{tr}} = 0$. Since $r >
2\tilde{\gamma}_{\max}$ by assumption in the Theorem, by Lemma~\ref{lem:boundsmaxmargin}
with probability at least $1-2\text{e}^{-\frac{\alpha^2 (d-1)}{n}}$, we then have that $r> 2\tilde{\gamma} \geq 2\tilde{\gamma}^{\eps_{\text{tr}}}$. Given
the closed form of $\thetahat{\epstrain}$ in Equation~\eqref{eq:sml}, it
directly follows that $\indof{\thetahat{\epstrain}}{1} = r >
2\tilde{\gamma}^{\eps_{\text{tr}}} \|\tilde{\theta}^{\eps_{\text{tr}}}\|_2 =
\|\indof{\thetahat{\epstrain}}{2:d}\|_2$ and hence $1\in j^\star(\thetahat{\epstrain})$. This
contradicts the original assumption $1\not \in j^\star(\thetahat{\epstrain})$ and
hence we established that $\thetahat{\eps_{\text{tr}}}$ for the
$\ell_1$-perturbation set~\eqref{eq:l1maxpert} has the same closed
form~\eqref{eq:robmaxmargin} as for the perturbation
set~\eqref{eq:linfmaxpert}.
The final statement is proved by using the analogous steps as in
the proof of 1. and 2. to obtain the closed form of the robust accuracy of
$\thetahat{\eps_{\text{tr}}}$.
\subsection{Proof of Lemma~\ref{lem:maxmargin}}
\label{sec:maxmarginproof}
We start by proving that $\thetahat{}$ is of the form
\begin{equation}
\label{Eq:max_margin_param_form_total_D}
\thetahat{} = \left[a_1, a_2 \tilde{\theta} \right],
\end{equation}
for $a_1, a_2 > 0$. Denote by $\decplanegen{\theta}$ the plane through the origin with normal $\theta$. We define $d\left((x,y), \decplanegen{\theta} \right)$ as the signed euclidean distance from the point $(x,y) \in D \sim \mathbb{P}_{r}$ to the plane $\decplanegen{\theta}$. The signed euclidean distance is the defined as the euclidean distance from x to the plane if the point $(x,y)$ is correctly predicted by $\theta$, and the negative euclidean distance from $x$ to the plane otherwise. We rewrite the definition of the max $l_2$-margin classifier. It is the classifier induced by the normalized vector $\thetahat{}$, such that
\begin{equation*}
\max_{\theta \in \mathbb{R}^{d}} \min_{(x,y) \in D}d\left( \left(x,y\right),\decplanegen{\theta}\right) = \min_{(x,y) \in D} d\left( \left(x,y \right),\decplanegen{\thetahat }\right).
\end{equation*}
We use that $D$ is deterministic in its first coordinate and get
\begin{equation*}
\begin{split}
\max_{\theta}\min_{(x,y) \in D}d\left(\left(x,y\right), \decplanegen{\theta} \right) &= \max_{\theta}\min_{(x,y) \in D} y (\thetaind{1} \xind{1} + \tilde{\theta}^{\top} \tilde{x})\\
&= \max_{\theta} \theta_1 \frac{r}{2} + \min_{(x,y) \in D} y \tilde{\theta}^{\top} \Tilde{x}.
\end{split}
\end{equation*}
Because $r >0$, the maximum over all $\theta$ has $\thetahatind{}{1} \geq 0$. Take any $a > 0$ such that $\|\tilde{\theta}\|_2 = a$. By definition the max $l_2$-margin classifier, $\tilde{\theta}$, maximizes $\min_{(x,y) \in D} d\left(\left(x,y\right), \decplanegen{\theta} \right)$. Therefore, $\thetahat{}$ is of the form of Equation \eqref{Eq:max_margin_param_form_total_D}.
Note that all classifiers induced by vectors of the form of Equation \eqref{Eq:max_margin_param_form_total_D} classify $D$ correctly. Next, we aim to find expressions for $a_1$ and $a_2$ such that Equation \eqref{Eq:max_margin_param_form_total_D} is the normalized max $l_2$-margin classifier. The distance from any $x \in D$ to $\decplanegen{\thetahat{}}$ is
\begin{equation*}
d\left(x,\decplanegen{\thetahat{}} \right) = \left| a_1 \xind{1} + a_2 \tilde{\theta}^{\top} \tilde{x} \right|.
\end{equation*}
Using that $\xind{1} = y \frac{r}{2}$ and that the second term equals $a_2 d\left(x, \decplanegen{\tilde{\theta}}\right)$, we get
\begin{equation}
\label{eq:distance_to_opt_intermidate}
d\left(x, \decplanegen{\thetahat{}}\right) = \left| a_1 \frac{r}{2} + a_2 d\left(x, \decplanegen{\tilde{\theta}}\right) \right| = a_1 \frac{r}{2} + \sqrt{1-a_1^2} d\left(x, \decplanegen{\tilde{\theta}}\right).
\end{equation}
Let $(\tilde{x},y) \in \widetilde{D}$ be the point closest in Euclidean
distance to $\tilde{\theta}$. This point is also the closest point in
Euclidean distance to $\decplanegen{\thetahat{}}$, because by Equation
\eqref{eq:distance_to_opt_intermidate} $d\left(x,
\decplanegen{\thetahat{}}\right)$ is strictly decreasing for
decreasing $d\left(x, \decplanegen{\tilde{\theta}}\right)$. We maximize
the minimum margin $d\left(x, \decplanegen{\thetahat{}} \right)$ with
respect to $a_1$. Define the vectors $a = \left[a_1,
a_2\right]$ and $v = \left[\frac{r}{2}, d\left(x,
\decplanegen{\tilde{\theta}}\right)\right]$. We find using the dual
norm that
\begin{equation*}
a = \frac{v}{\|v\|_2}.
\end{equation*}
Plugging the expression of $a$ into Equation
\eqref{Eq:max_margin_param_form_total_D} yields that $\thetahat{}$ is
given by
\begin{equation*}
\thetahat{} = \frac{1}{\sqrt{r^2 + 4 \tilde{\gamma}^2}}\left[r, 2 \tilde{\gamma}\tilde{\theta} \right].
\end{equation*}
For the second part of the lemma we first decompose
\begin{equation}
\label{eq:jacob}
\mathbb{P}_{r_{\text{test}}} (Y\thetahat{\top} X >0 ) = \frac{1}{2}\mathbb{P}_{r_{\text{test}}} \left[ \thetahat{\top} X >0 \mid Y=1 \right] +\frac{1}{2}\mathbb{P}_{r_{\text{test}}} \left[\thetahat{\top} X <0 \mid Y=-1\right]\nonumber
\end{equation}
We can further write
\begin{align}
\label{eq:cumul1}
\mathbb{P}_{r_{\text{test}}} \left[\thetahat{\top} X > 0 \mid
Y = 1\right] &=\mathbb{P}_{r_{\text{test}}} \left[\sum_{i=2}^{d}\indof{\thetahat{}}{i} \indof{X}{i} > -
\indof{\thetahat{}}{1} \: \indof{X}{1} \mid Y=1\right]\\
&= \mathbb{P}_{r_{\text{test}}} \left[2 \tilde{\gamma} \sum_{i=1}^{d-1}\indof{\tilde{\theta}}{i} \indof{X}{i} > -
r \: \frac{r_{\text{test}}}{2} \mid Y=1\right]\nonumber\\
&= 1-\Phi\left(-\frac{r\: r_{\text{test}}}{4\sigma \tilde{\gamma}} \right) =
\Phi\left(\frac{r \: r_{\text{test}}}{4\sigma \tilde{\gamma}} \right) \nonumber
\end{align}
where $\Phi$ is the cumulative distribution function. The second equality
follows by multiplying by the normalization constant on both sides and the
third equality is due to the fact that $\sum_{i=1}^{d-1}\indof{\tilde{\theta}}{i} \indof{X}{i}$ is
a zero-mean Gaussian with variance $\sigma^2\|\tilde{\theta}\|^2_2 = \sigma^2$ since $\tilde{\theta}$ is normalized.
Correspondingly we can write
\begin{align}
\label{eq:cumul2}
\mathbb{P}_{r_{\text{test}}} \left[\thetahat{\top} X < 0 \mid
Y = -1\right] &=\mathbb{P}_{r_{\text{test}}} \left[2\tilde{\gamma}
\sum_{i=1}^{d-1}\indof{\tilde{\theta}}{i} \indof{X}{i} < -
r \left(- \frac{r_{\text{test}}}{2}\right) \mid Y=-1\right] = \Phi\left(\frac{r \:r_{\text{test}}}{4\sigma \tilde{\gamma}}\right)
\end{align}
so that we can
combine~\eqref{eq:jacob} and~\eqref{eq:cumul1} and \eqref{eq:cumul2} to obtain
$\mathbb{P}_{r_{\text{test}}} (Y\thetahat{\top} X >0 ) = \Phi \left(\frac{r \:r_{\text{test}}}{4\sigma \tilde{\gamma}}\right)$. This concludes the proof of the lemma.
\subsection{Proof of Lemma \ref{lem:boundsmaxmargin}}
\label{sec:boundsmaxmargin}
The proof plan is as follows. We start from the definition of the max
$\ell_2$-margin of a dataset. Then, we rewrite the
max $\ell_2$-margin as an expression that includes a random matrix with independent
standard normal entries. This allows us to prove the upper and lower bounds for the
max-$\ell_2$-margin in Sections~\ref{sec:gammaupperbound} and ~\ref{sec:gammalowerbound}
respectively, using non-asymptotic estimates on the singular values of
Gaussian random matrices.
Given the dataset $\widetilde{D} = \{(\tilde{x}_i, y_i)\}_{i=1}^{n}$, we define the random matrix
\begin{equation}
\label{eq:randmatrixsamples}
X = \begin{pmatrix}
\tilde{x}_1^{\top}\\
\tilde{x}_2^{\top}\\
...\\
\tilde{x}_{n}^{\top}
\end{pmatrix}.
\end{equation}
where $\tilde{x}_i \sim \mathcal{N}(0,\sigma I_{d-1})$.
Let $\mathcal{V}$ be the class of all perfect predictors of $\widetilde{D}$. For a matrix $A$ and vector $b$ we also denote by $|Ab|$ the vector whose entries correspond to the absolute values of the entries of $Ab$.
Then, by definition
\begin{equation}
\label{maxmargindefgammaproof}
\tilde{\gamma} = \max_{v \in \mathcal{V}, \|v\|_2=1} \min_{j \in [n]} \indof{|X v|}{j} = \max_{v \in \mathcal{V}, \|v\|_2=1} \min_{j \in [n]} \sigma \indof{|Q v|}{j},
\end{equation}
where $Q = \frac{1}{\sigma} X$ is the scaled data matrix.
In the sequel we will use the operator norm of a matrix $A \in \mathbb{R}^{n \times d-1}$.
\begin{equation*}
\| A\|_2 = \sup_{v \in \mathbb{R}^{d-1} \mid \|v\|_2=1} \|A v \|_2
\end{equation*}
and denote the maximum singular value of a matrix $A$ as $s_{\text{max}} (A)$ and the minimum singular value as $s_{\text{min}}(A)$.
\subsubsection{Upper bound}
\label{sec:gammaupperbound}
Given the maximality of the
operator norm and since the minimum entry of the vector
$|Q v|$ must be smaller than $\frac{\|Q\|_2}{\sqrt{n}}$, we can upper bound
$\tilde{\gamma}$ by
\begin{equation*}
\tilde{\gamma} \leq \sigma \frac{1}{\sqrt{n}} \|Q{}\|_2.
\end{equation*}
Taking the expectation on both sides with respect to the draw of
$\widetilde{D}$ and noting $\|Q\|_2 \leq
s_{\text{max}}\left(Q\right)$,
it follows from
Corollary 5.35 of \cite{vershynin12}
that for all $t\geq 0$:
\begin{equation*}
\mathbb{P} \left[\sqrt{d-1}+\sqrt{n}+t \geq s_{\text{max}}\left(Q\right) \right] \geq 1-2e^{-\frac{t^2}{2}}.
\end{equation*}
Therefore, with a probability greater than $1-2e^{-\frac{t^2}{2}}$,
\begin{equation*}
\tilde{\gamma} \leq \sigma \left(1+ \frac{t+\sqrt{d-1}}{\sqrt{n}}\right).
\end{equation*}
\subsubsection{Lower bound}
\label{sec:gammalowerbound}
By the definition in Equation \eqref{maxmargindefgammaproof}, if we
find a vector $v \in \mathcal{V}$ with $\|v\|_2=1$ such that for an
$a>0$, it holds that $\Hquad \min_{j \in n} \sigma
\indof{|X v|}{j} > a$, then $\tilde{\gamma} > a$.
Recall the definition of the max-$\ell_2$-margin
as in Equation \ref{eq:randmatrixsamples}.
As $n < d-1$, the random matrix $Q$ is a wide
matrix, i.e. there are more columns than rows and therefore the
minimal singular value is $0$.
Furthermore, $Q$ has rank $n$ almost surely and hence
for all $c >0$, there exists a $v \in \mathbb{R}^{d-1}$ such that
\begin{equation}
\label{eq:existencerhs}
\sigma Q v= 1_{\{ n\}}c> 0,
\end{equation}
where $ 1_{\{ n \}}$ denotes the all ones vector of dimension $n$. The smallest non-zero singular value of $Q$, $s_{\text{min, nonzero}}(Q)$, equals the smallest non-zero singular value of its transpose $Q^{\top}$. Therefore, there also exists a $v \in \mathcal{V}$ with $\|v\|_2=1$ such that
\begin{equation}
\label{minimum_step_gamma}
\tilde{\gamma} \geq \min_{j \in [n]} \sigma \indof{|Q v|}{j} \geq \sigma s_{\text{min,nonzeros}}\left(Q^{\top}\right)\frac{1}{\sqrt{n}},
\end{equation}
where we used the fact that any vector $v$ in the span of non-zero eigenvectors satisfies $\|Q v \|_2 \geq s_{\text{min, nonzeros}}(Q)$ and the existence of a solution $v$ for any right-hand side as in Equation \ref{eq:existencerhs}.
Taking the expectation on both sides,
Corollary 5.35 of \cite{vershynin12} yields that with a probability greater than $1-2e^{-\frac{t^2}{2}}, t\geq 0$ we have
\begin{equation}
\tilde{\gamma} \geq \sigma\left( \frac{\sqrt{d-1}-t}{\sqrt{n}}-1\right).
\end{equation}
\section{Experimental details on the Waterbirds dataset}
\label{sec:waterbirds}
In this section, we discuss the experimental details and construction of the Waterbirds dataset in more detail. We also provide ablation studies of attack parameters such as the size of the motion blur kernel, plots of the robust error decomposition with increasing $n$, and some experiments using early stopping.
\paragraph{The waterbirds dataset}
To build the Waterbirds dataset, we use the CUB-200 dataset \cite{Welinder10}, which contains images and labels of $200$ bird species, and $4$ background classes (forest, jungle/bamboo, water ocean, water lake natural) of the Places dataset \cite{zhou17}.The aim is to recognize whether or not the bird, in a given image, is a waterbird (e.g. an albatros) or a landbird (e.g. a woodpecker). To create the dataset, we randomly sample equally many water- as landbirds from the CUB-200 dataset. Thereafter, we sample for each bird image a random background image. Then, we use the segmentation provided in the CUB-200 dataset to segment the birds from their original images and paste them onto the randomly sampled backgrounds. The resulting images have a size of $256 \times 256$. Moreover, we also resize the segmentations such that we have the correct segmentation profiles of the birds in the new dataset as well. For the concrete implementation, we use the code provided by \cite{Sagawa20}.
\paragraph{Experimetal training details}
Following the example of \cite{Sagawa20}, we use a ResNet50 pretrained on the ImageNet dataset for all experiments, a weight-decay of $10^{-4}$, and train for $300$ epochs using the Adam optimizer. Extensive fine-tuning of the learning rate resulted in an optimal learning rate of $0.006$ for all experiments in the low sample size regime. Adversarial training is implemented as suggested in \cite{madry18}: at each iteration we find the worst case perturbation with an exact or approximate method. In all our experiments, the resulting classifier interpolates the training set. We plot the mean over all runs and the standard deviation of the mean.
\paragraph{Specifics to the motion blur attack}
Fast moving objects or animals are hard to photograph due to motion blur. Hence, when trying to classify or detect moving objects from images, it is imperative that the classifier is robust against reasonable levels of motion blur. We implement the attack as follows. First, we segment the bird from the original image, then use a blur filter and lastly, we paste the blurred bird back onto the background. We are able to apply more severe blur, by enlarging the kernel of the filter. See Figure \ref{fig:motion_blur_panel} for an ablation study of the kernel size.
The motion blur filter is implemented as follows. We use a kernel of size $M \times M$ and build the filter as follows: we fill the row $(M-1)/2$ of the kernel with the value $1/M$. Thereafter, we use the 2D convolution implementation of OpenCV (filter2D) \cite{opencv_library} to convolute the kernel with the image. Note that applying a rotation before the convolution to the kernel, changes the direction of the resulting motion blur. Lastly, we find the most detrimental level of motion blur using a list-search over all levels up to $M_{max}$.
\begin{figure*}[!t]
\centering
\begin{subfigure}[b]{0.19\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbird_original_example.png}
\caption{Original}
\label{fig:motion_blur_or}
\end{subfigure}
\begin{subfigure}[b]{0.19\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/motion_blur_5.png}
\caption{$M = 5$}
\label{fig:motion_blur_5}
\end{subfigure}
\begin{subfigure}[b]{0.19\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/motion_blur_10.png}
\caption{$M = 10$}
\label{fig:motion_blur_10}
\end{subfigure}
\begin{subfigure}[b]{0.19\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/motion_blur_15.png}
\caption{$M = 15$}
\label{fig:motion_blur_15}
\end{subfigure}
\begin{subfigure}[b]{0.19\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/motion_blur_20.png}
\caption{$M = 20$}
\label{fig:motion_blur_20}
\end{subfigure}
\caption{We perform an ablation study of the motion blur kernel size, which corresponds to the severity level of the blur. We see that for increasing $M$, the severity of the motion blur increases. In particular, note that for $M = 15$ and even $M = 20$, the bird remains recognizable: we do not semantically change the class, i.e. the perturbations are consistent.}
\label{fig:motion_blur_panel}
\end{figure*}
\begin{figure*}[!b]
\centering
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_03_dark.png}
\caption{$\epsilon = -0.3$}
\label{fig:dark_03}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_02_dark.png}
\caption{$\epsilon = -0.2$}
\label{fig:dark_02}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_01_dark.png}
\caption{$\epsilon = -0.1$}
\label{fig:dark_01}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbird_original_example.png}
\caption{Original}
\label{fig:light_or}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_01_light.png}
\caption{$\epsilon = 0.1$}
\label{fig:light_01}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_02_light.png}
\caption{$\epsilon = 0.2$}
\label{fig:light_02}
\end{subfigure}
\begin{subfigure}[b]{0.136\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/bird_light_03_light.png}
\caption{$\epsilon = 0.3$}
\label{fig:light_03}
\end{subfigure}
\caption{We perform an ablation study of the different lighting changes of the adversarial illumination attack. Even though the directed attack\xspace attacks the signal component in the image, the bird remains recognizable in all cases.}
\label{fig:light_panel}
\end{figure*}
\paragraph{Specifics to the adversarial illumination attack}
An adversary can hide objects using poor lightning conditions, which can for example arise from shadows or bright spots. To model poor lighting conditions on the object only (or targeted to the object), we use the adversarial illumination attack.
The attack is constructed as follows: First, we segment the bird from their background. Then we apply an additive constant $\epsilon$ to the bird, where the absolute size of the constant satisfies $|\epsilon| < \eps_{\text{te}} = 0.3$. Thereafter, we clip the values of the bird images to $[0, 1]$, and lastly, we paste the bird back onto the background. See Figure \ref{fig:light_panel} for an ablation of the parameter $\epsilon$ of the attack. It is non-trivial how to (approximately) find the worst perturbation. We find an approximate solution by searching over all perturbations with increments of size $\eps_{\text{te}}/K_{\text{max}}$. Denote by seg\xspace, the segmentation profile of the image $x$. We consider all perturbed images in the form of
\begin{equation*}
x_{pert} = (1-seg) x + seg (x + \epsilon \frac{K}{K_{\text{max}}} 1_{255 \times 255}), \Hquad K \in [-K_{max}, K_{max}].
\end{equation*}
During training time we set $K_{max} = 16$ and therefore search over $33$ possible images. During test time we search over $65$ images ($K_{max} = 32$).
\paragraph{Early stopping} In all our experiments on the Waterbirds dataset, a parameter search lead to an optimal weight-decay and learning rate of $10^{-4}$ and $0.006$ respectively. Another common regularization technique is early stopping, where one stops training on the epoch where the classifier achieves minimal robust error on a hold-out dataset. To understand if early stopping can mitigate the effect of adversarial training aggregating robust generalization in comparison to standard training, we perform the following experiment. On the Waterbirds dataset of size $n = 20$ and considering the adversarial illumination attack, we compare standard training with early stopping and adversarial training $(\eps_{\text{tr}} = \eps_{\text{te}} = 0.3)$ with early stopping. Considering several independent experiments, early stopped adversarial training has an average robust error of $33.5$ a early stopped standard training $29.1$. Hence, early stopping does decrease the robust error gap, but does not close it.
\paragraph{Error decomposition with increasing $n$}
In Figure \ref{fig:waterbirds_light_numobs}, we see that adversarial training hurts robust generalization in the small sample size regime. For completeness, we plot the robust error composition for adversarial and standard training in Figure \ref{fig:light_numsamp_decomposition}. We see that in the low sample size regime, the drop in susceptibility that adversarial training achieves in comparison to standard training, is much lower than the increase in standard error. Conversely, in the high sample regime, the drop of susceptibility from adversarial training over standard training is much bigger than the increase in standard error.
\begin{figure*}[!t]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/numsamp_waterbirds_light.png}
\caption{Robust error}
\label{fig:app_waterbirds_robust_error}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbirds_standard_numsamp.png}
\caption{Standard error}
\label{fig:app_waterbirds_standard_error}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbirds_susceptibility_decomposition.png}
\caption{Susceptibility}
\label{fig:app_waterbirds_susceptibility}
\end{subfigure}
\caption{We plot the robust error decomposition of the experiments depicted in Figure \ref{fig:waterbirds_light_numobs}. The plots depict the mean and standard deviation of the mean over several independent experiments. We see that, in comparison to standard training, the reduction in susceptibility for adversarial training is minimal in the low sample size regime. Moreover, the increase in standard error of adversarial training is quite severe, leading to an overall increase in robust error in the low sample size regime.}
\label{fig:light_numsamp_decomposition}
\end{figure*}
\section{Future work}
This paper aims to caution the practitioner against blindly following
current widespread practices to increase the robust
performance of machine learning models.
Specifically, adversarial training is currently recognized to be one
of the most effective defense mechanisms for $\ell_p$-perturbations,
significantly outperforming robust performance of standard training. However, we prove that
this common wisdom is not applicable for directed attacks -- that are perceptible (albeit consistent) but efficiently focus their
attack budget to target ground truth class information -- in the low-sample size regime.
In particular, in such settings adversarial training can in fact yield worse accuracy than standard training.
In terms of follow-up work on directed attacks in the low-sample
regime, there are some concrete questions that would be interesting to
explore. For example, as discussed in Section~\ref{sec:relatedwork},
it would be useful to test whether some methods to mitigate the
standard accuracy vs. robustness trade-off would also relieve the
perils of adversarial training for directed attacks. Further, we
hypothesize, independent of the attack during test time, it is
important in the small sample-size regime to choose perturbation sets
during training that align with
the ground truth signal (such as rotations for data with inherent
rotation). If this hypothesis were to be confirmed, it would break
with yet another general rule that the best defense perturbation type
should always match the attack during evaluation. The insights from
this study might also be helpful in the context of searching for
good defense perturbations.
\section{Discussion and related work}
\section{Introduction}
\label{sec:intro}
\begin{wrapfigure}{r}{0.43\textwidth}
\centering
\vspace{-0.1in}
\includegraphics[width=0.99\linewidth]{plotsAistats/teaser_try_2.png}
\caption{
On subsampled \mbox{CIFAR10} attacked by $2\times 2$ masks, adversarial training yields higher robust error than standard training
when the sample size is small, even though it helps for large sample sizes.
(see Sec.~\ref{sec:app_cifar10} for details).}
\vspace{-0.2in}
\label{fig:teaserplot}
\end{wrapfigure}
Today's best-performing classifiers are vulnerable to adversarial attacks
\cite{goodfellow15, szegedy14} and exhibit high \emph{robust error}: for many inputs, their predictions change under adversarial perturbations,
even though the true class stays the same.
For example, in image classification tasks, we distinguish between two categories of
such attacks that are content-preserving \cite{gilmer18b} (or consistent \cite{raghunathan20}) if their strength is limited --- perceptible and imperceptible perturbations.
Most work to date studies imperceptible attacks such as
bounded $\ell_p$-norm perturbations \cite{goodfellow15, madry18, moosavi16}, small transformations using image processing
techniques \cite{ghiasi19, zhao20, laidlaw21, Luo18} or
nearby samples on the data manifold \cite{Lin20, Zhou20}.
They can often use their limited budget to successfully fool a learned classifier but, by definition, do not visibly reduce information about the actual class: the object in the perturbed image looks exactly the same as in the original version.
On the other hand, perceptible perturbations may occur more naturally in practice or are physically realizable.
For example, stickers can be placed on traffic signs \cite{Eykholt18},
masks of different sizes may cover important features of human faces
\cite{Wu20}, images might be rotated or translated \cite{Logan19},
animals in motion may appear blurred in photographs
depending on the shutter speed, or the lighting conditions could be poor (see Figure~\ref{fig:sig_att_examples}).
Some perceptible attacks can effectively
use the perturbation budget to reduce actual class information
in the input (the \emph{signal}) while still preserving the original class.
For example, a stop sign with a small sticker doesn't lose its semantic meaning
or a flying bird does not become a different species because it induces motion blur
in the image.
We refer to these attacks as \emph{directed attacks\xspace} (see
Section~\ref{sec:robustness} for a more formal
characterization).
In this paper, we
demonstrate that one of the most common beliefs
for adversarial attacks does not transfer to directed attacks\xspace, in
particular when the sample size is small. Specifically, it is widely acknowledged that adversarial training often achieves significantly lower adversarial error than standard
training. This holds in particular if
the perturbation type
\cite{madry18, zhang19, Bai21} and perturbation budget match the attack during test time.
Intuitively, the improvement is a result of decreased
\emph{attack-susceptibility}: independent of the true class,
adversarial training explicitly encourages the classifier to predict
the same class for all perturbed points.
\begin{figure}[t]
\vskip 0.2in
\begin{center}
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/CIFAR10_class_boat.png}
\caption{Masking}
\label{fig:CIFAR10_boat}
\end{subfigure}
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/light_darker.png}
\caption{Illumination}
\label{fig:WB_light_dark}
\end{subfigure}
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/water-bird_motion_blurred.png}
\caption{Motion blur}
\label{fig:WB_motion_blur}
\end{subfigure}
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbird_original_example.png}
\caption{Original}
\label{fig:fig:WB_original}
\end{subfigure}
\caption{Examples of directed attacks\xspace on CIFAR10 and the
Waterbirds dataset. In Figure \ref{fig:CIFAR10_boat}, we corrupt the image with a black mask of size $2 \times 2$ and in Figure \ref{fig:WB_light_dark} and \ref{fig:WB_motion_blur} we change the lighting conditions (darkening) and apply motion blur on the bird in the image respectively.
All perturbations effectively reduce the information about the class in the images: they are the result of directed attacks\xspace.}
\label{fig:sig_att_examples}
\end{center}
\vskip -0.2in
\end{figure}
In this paper, we question the efficacy of adversarial training to increase
robust accuracy for directed attacks\xspace.
In particular, we show that adversarial training not only increases standard test error as noted in \cite{zhang19, tsipras19, Stutz19, raghunathan20}), but surprisingly,
\begin{center}
\emph{adversarial training may even increase the robust test error compared to standard training!}
\end{center}
Figure \ref{fig:teaserplot} illustrates the main message of our paper for CIFAR10 subsets: Although adversarial training
outperforms standard training when enough training samples are available, it is inferior
in the low-sample regime.
More specifically, our contributions are as follows:
\begin{itemize}
\item We prove that, almost surely, adversarially training a linear classifier on separable data yields a monotonically increasing robust error as the perturbation budget grows.
We further establish high-probability non-asymptotic lower bounds on the robust error gap between adversarial and standard training.
\item Our proof provides intuition for why this phenomenon is particularly prominent for directed attacks\xspace in the small sample size regime.
\item We show that this phenomenon occurs on a variety of real-world datasets and perceptible directed attacks\xspace in the small sample size regime.
\end{itemize}
\section{Real-world experiments}
\label{sec:realworldexpapp}
In this section, we demonstrate that adversarial training may
hurt robust accuracy in a variety of image attack scenarios
on the Waterbirds and CIFAR10 dataset.
The corresponding experimental details and more experimental results (including
on an additional hand gestures dataset) can be found in Appendices
\ref{sec:waterbirds}, \ref{sec:app_cifar10} and \ref{sec:handgestures}.
\subsection{Datasets}
We now describe the datasets and models that we use for the
experiments. In all our experiments on CIFAR10, we vary the sample
size by subsampling the dataset and use a ResNet18 \cite{He16} as
model. We always train on the same (randomly subsampled) dataset,
meaning that the variances arise from the random seed of the model and
the randomness in the training algorithm. In Appendix
\ref{sec:app_cifar10}, we complement the results of this section by
reporting the results of similar experiments with different
architectures.
As a second dataset, we build a new version of the Waterbirds
dataset, consisting of images of water- and
landbirds of size $256 \times 256$ and labels that distinguish the
two types of birds. We construct the dataset as follows: First, we
sample equally many water- and landbirds from the CUB-200 dataset
\cite{Welinder10}. Then, we segment the birds and paste them onto a
background that is randomly sampled (without replacement) from the Places-256 dataset \cite{zhou17}.
For the implementation of the dataset we used the code provided by \citet{Sagawa20}. Also, following the choice of \citet{Sagawa20}, we use as model a ResNet50 that was pretrained on ImageNet and which achieves near perfect standard accuracy.
\subsection{Evaluation of directed attacks\xspace}
We consider three types of directed attacks\xspace on our real world datasets:
square masks, motion blur and adversarial illumination. The mask
attack is a model used to simulate sticker-attacks and general
occlusions of objects in images \cite{Eykholt18, Wu20}. On the other
hand, motion blur may arise naturally for example when photographing
fast moving objects with a slow shutter speed. Further, adversarial
illumination may result from adversarial lighting conditions or smart
image corruptions. Next, we describe the attacks in more detail.
\paragraph{Mask attacks}
On CIFAR10, we consider the square black mask attack: the adversary can set a mask
of size $\eps_{\text{te}} \times \eps_{\text{te}}$ to zero in the image. To ensure that the mask does not cover the whole signal in the image, we
restrict the size of the masks to be at most $2 \times 2$. Hence, the search space of the attack consists of all possible locations of the masks in the targeted image. For exact robust error evaluation, we perform a full grid search over all possible locations during test time. See Figure \ref{fig:CIFAR10_boat} for an example of a mask attack on CIFAR10.
\paragraph{Motion blur}
On the Waterbirds dataset we consider two directed attacks\xspace: motion blur and adversarial illumination. For the motion blur attack,
the bird may move at different speeds without changing the background.
The aim is to be robust against all motion blur severity levels up to $M_{max} = 15$.
To simulate motion blur, we first segment the birds and then use a filter with a kernel of size $M$ to apply motion blur on the bird only. Lastly, we paste the blurred bird back onto the background image. We can change the severity level of the motion blur by increasing the kernel size of the filter.
See Appendix \ref{sec:waterbirds} for an ablation study and concrete expressions of the motion blur kernel. At test time, we perform a full grid search over all kernel sizes to exactly evaluate the robust error. We refer to Figure \ref{fig:WB_motion_blur} and Section \ref{sec:waterbirds} for examples of our motion blur attack.
\paragraph{Adversarial illumination} As a second attack on the Waterbirds dataset, we consider adversarial illumination. The adversary can darken or brighten the bird without corrupting the background of the image. The attack aims to model images where the object at interest is hidden in shadows or placed against bright light.
To compute the adversarial illumination attack, we segment the bird, then darken or brighten the it, by adding a constant $a \in [-\eps_{\text{te}}, \eps_{\text{te}}]$, before pasting the bird back onto the background image. We find the most adversarial lighting level, i.e. the value of $a$, by equidistantly partitioning the interval $[-\eps_{\text{te}}, \eps_{\text{te}}]$ in $K$ steps and performing a full list-search over all steps.
See Figure \ref{fig:WB_light_dark} and Section \ref{sec:waterbirds} for examples of the adversarial illumination attack.
\subsection{Adversarial training procedure}
For all datasets, we run SGD until convergence on the \emph{robust} cross-entropy
loss~\eqref{eq:emploss}. In each iteration, we search for an adversarial example
and update the weights using a gradient with respect to the resulting
perturbed example \cite{goodfellow15, madry18}.
For every experiment, we choose the learning
rate and weight decay parameters that minimize the robust error on a
hold-out dataset. We now describe the implementation of the
adversarial search for the three types of
directed attacks\xspace.
\paragraph{Mask attacks}
Unless specified otherwise, we use an approximate attack similar to
\citet{Wu20} during training time:
First, we identify promising mask locations by analyzing the gradient, $\nabla_x L(f_\theta(x), y)$, of the cross-entropy loss with respect to the input. Masks that cover part of the image where the gradient is large, are more likely to increase the loss. Hence, we compute the $K$ mask locations $(i, j)$, where $\|\nabla_x L(f_\theta(x), y)_{[i:i+2, j:j+2]} \|_1$ is the largest and take using a full list-search the mask that incurs the highest loss.
Our intuition from the theory predicts that higher $K$,
and hence a more exact ``defense'', only increases the robust error of
adversarial training, since the mask could then more efficiently cover
important information about the class. We indeed confirm this effect
and provide more details in Section~\ref{sec:app_cifar10}.
\paragraph{Motion blur}
Intuitively the worst attack should be the most severe blur, rendering
a search over a range of severity superfluous. However, similar to
rotations, this is not necessarily true in practice since the training loss on
neural networks is generally nonconvex. Hence, during training time,
we perform a search over kernels with sizes $2i$ for $i = 1,\dots,
M_{max}/2$. Note that, at test time, we do an exact search
over all kernels of sizes in $[1, 2, \dots, M_{max}]$.
\paragraph{Adversarial illumination}
Similar to the motion blur attack, intuitively the worst perturbation
should be the most severe lighting changes; either darkening or
illuminating the object maximally. However, again this is not
necessarily the case, since finding the worst attack is a nonconvex
problem. Therefore, during training and testing we partition the
interval $[-\eps_{\text{tr}}, \eps_{\text{tr}}]$ in $33$ and $65$ steps
respectively, and perform a full grid-search to find the worst
perturbation.
\subsection{Adversarial training can hurt robust generalization}
Further, we perform the following experiments on the Waterbirds dataset using the motion blur and adversarial illumination attack. We vary the adversarial training budget $\eps_{\text{tr}}$, while keeping the number of samples fixed, and compute the resulting robust error.
We see in Figure \ref{fig:waterbirds_light_d_n} and \ref{fig:motion_lines} that, indeed, adversarial training can hurt robust generalization with increasing perturbation budget $\eps_{\text{tr}}$.
Furthermore, to gain intuition as described in Section~\ref{logreg_proof_sketch} and, we also plot the robust error decomposition (Equation~\ref{eq:decomposition}) consisting of the standard error and susceptibility in Figure \ref{fig:light_trade_off} and \ref{fig:motion_blur_trade_off}. Recall that we measure susceptibility as the fraction of data points in the test set for
which the classifier predicts a different class under an adversarial attack.
As in our linear example, we observe an increase in robust error despite a slight drop
in susceptibility, because of the more severe increase in standard error.
Similar experiments for the hand gesture dataset can be
found in~\ref{sec:handgestures}.
\begin{figure*}[!t]
\centering
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbirds_motion_d_n.png}
\caption{Robust error with increasing $\eps_{\text{tr}}$}
\label{fig:motion_lines}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/waterbirds_trade-off.png}
\caption{Robust error decomposition}
\label{fig:motion_blur_trade_off}
\end{subfigure}
\caption{ (a) We plot the robust error with increasing adversarial training budget $\eps_{\text{tr}}$ of $5$ experiments on the subsampled Waterbirds datasets of sample sizes $20$ and $30$. Even though adversarial training hurts robust generalization for low sample size ($n = 20$), it helps for $n = 50$. (b) We plot the decomposition of the robust error in standard error and susceptibility with increasing adversarial budget $\eps_{\text{tr}}$. We plot the mean and standard deviation of the mean of $5$ experiments on a subsampled Waterbirds dataset of size $n = 20$. The increase in standard error is more severe than the drop in susceptibility, leading to a slight increase in robust error. For more experimental details see Section \ref{sec:waterbirds}.}
\label{fig:motion_blur_real_world}
\end{figure*}
As predicted by our theorem, the phenomenon where adversarial training hurts robust generalization is most pronounced in the small sample size regime. Indeed, the experiments depicted in Figures \ref{fig:waterbirds_light_d_n} and \ref{fig:motion_lines} are conducted on small sample size datasets of $n = 20$ or $50$.
In Figure \ref{fig:teaserplot} and \ref{fig:waterbirds_light_numobs}, we
observe that the as sample size increases, adversarial training does improve robust generalization compared to standard training, even for directed attacks\xspace. Moreover, on the experiments of CIFAR10 using the mask perturbation, which can be found in Figure \ref{fig:teaserplot} and Appendix \ref{sec:app_cifar10}, we observe the same behaviour: Adversarial training hurts robust generalization in the low sample size regime, but helps when enough samples are available.
\subsection{Discussion}
In this section, we discuss how different algorithmic choices, motivated
by related work, affect when and how adversarial training hurts robust generalization.
\paragraph{Strength of attack and catastrophic overfitting}
In many cases, the worst case perturbation during adversarial training is found using an approximate algorithm such as projected gradient descent. It is common belief that using the strongest attack (in the mask-perturbation case, full grid search) during training should also result in better robust generalization.
In particular, the literature on catastrophic overfitting shows that weaker attacks during training lead to bad performance on stronger attacks during testing \cite{Wong20Fast, andriushchenko20, li21}.
Our result suggests the opposite is true in the low-sample size regime for
directed attacks\xspace : the weaker the attack, the better
adversarial training performs.
\paragraph{Robust overfitting}
Recent work observes empirically \cite{rice20} and theoretically
\cite{sanyal20, donhauser21}, that perfectly minimizing the
adversarial loss during training might in fact be suboptimal for
robust generalization; that is, classical regularization techniques
might lead to higher robust accuracy. The phenomenon is often referred
to as robust overfitting. May the phenomenon be mitigated using
standard regularization techniques? In Appendix \ref{sec:waterbirds} we shed light on
this question and show that adversarial training hurts robust generalization even with standard regularization methods such as early stopping are used.
\section{Related work}
\label{sec:relatedwork}
We now discuss how our results relate to phenomena that have been observed or proven in the literature before.
\paragraph{Robust and non-robust useful features}
In the words of \citet{ilyas19, springer21}, for
directed attacks, all robust features become less useful, but adversarial
training uses robust features more. In the small sample-size regime
$n<d-1$ in particular, robust learning assigns so much weight
on the robust (possibly non-useful) features, that the signal in the non-robust
features is drowned. This leads to an unavoidable and large increase
in standard error that dominates the decrease in susceptibility and
hence ultimately leads to an increase of the robust error.
\paragraph{Small sample size and robustness}
A direct consequence of Theorem~\ref{thm:linlinf} is that in order to
achieve the same robust error as standard training, adversarial
training requires more samples. This statement might remind the reader
of sample complexity results for robust generalization in
\citet{schmidt18, Yin19, Khim18}. While those results compare sample
complexity bounds for standard vs. robust error, our theorem
statement compares two algorithms, standard vs. adversarial training,
with respect to the robust error.
\paragraph{Trade-off between standard and robust error}
Many papers observed that even though adversarial training decreases robust error compared to standard training, it may lead
to an increase in standard test error \cite{madry18, zhang19}.
For example, \citet{tsipras19, zhang19, javanmard20, dobriban20, chen20} study settings where the Bayes optimal robust classifier is not equal to the Bayes optimal (standard)
classifier (i.e. the perturbations are inconsistent or the dataset is non-separable).
\cite{raghunathan20} study consistent perturbations, as in our paper,
and prove that for small sample size, fitting adversarial
examples can increase standard error even in the absence of
noise. In contrast to aforementioned works, which do not refute that
adversarial training decreases robust error, we prove that for
directed attacks\xspace perturbations, in the small sample regime adversarial training may also increase \emph{robust error}.
\paragraph{Mitigation of the trade-off}
A long line of work has proposed procedures to
mitigate the trade-off phenomenon. For example \citet{alayrac19,
Carmon19, zhai20, raghunathan20} study robust self training, which
leverages a large set of unlabelled data, while \citet{lee20, lamb19,
xu20} use data augmentation by interpolation. \citet{Ding20,
balaji19, Cheng20} on the other hand propose to use adaptive
perturbation budgets $\eps_{\text{tr}}$ that vary across inputs.
Our intuition from the theoretical analysis suggests that the standard
mitigation procedures for imperceptible perturbations may not work for
perceptible directed attacks\xspace, because all relevant features are non-robust.
We leave a thorough empirical study
as interesting future work.
\section{Robust classification}
\label{sec:robustness}
We first introduce our robust classification setting more formally by defining
the notions of adversarial robustness, directed attacks\xspace and adversarial training
used throughout the paper.
\paragraph{Adversarially robust classifiers}
For inputs $x \in \mathbb{R}^d$, we consider multi-class classifiers
associated with parameterized functions $f_\theta:\mathbb{R}^d \to
\mathbb{R}^K$, where $K$ is the number of labels. In the special case of binary classification ($K = 2$), we use the output predictions $y=\textrm{sign}(f_\theta(x))$. For example, $f_\theta(x)$ could be linear models (as in Section~\ref{sec:theoryresults}) or
neural networks (as in Section~\ref{sec:realworldexpapp}).
One key step to encourage deployment of machine learning based classification in real-world applications, is to increase the
robustness of classifiers against perturbations that do not
change the ground truth label.
Mathematically speaking, we would like to have a small
\emph{$\eps_{\text{te}}$-robust error}, defined as
\begin{equation}
\label{eq:roberr}
\roberr{\theta} := \mathbb{E}_{(x, y)\sim \mathbb{P}} \max_{x' \in \pertset{x}{\eps_{\text{te}}}} \ell(f_\theta (x'),y),
\end{equation}
where $\ell$ is the multi-class zero-one loss, which only equals $1$ if the predicted
output using $f_\theta(x)$ does not match the true label $y$.
Further, $\pertset{x}{\eps_{\text{te}}}$ is a perturbation set associated with a \emph{transformation type} and size $\eps_{\text{te}}$.
Note that the \emph{(standard) error} of a classifier corresponds to evaluating $\roberr{\theta}$ at $\eps_{\text{te}} = 0$, yielding the standard error $\stderr{\theta} =\mathbb{E}_{(x, y)\sim \mathbb{P}} \ell(f_\theta (x),y)$.
\paragraph{(Signal)-Directed attacks}
Most works in the existing literature consider consistent perturbations where
$\eps_{\text{te}}$ is small enough such that all samples in the perturbation set
have the same ground truth or expert label.
Note that the ground truth model $f_{\theta^{\star}}$ is therefore robust against perturbations and achieves the same error for standard and adversarial evaluation.
The inner maximization in Equation~\eqref{eq:roberr} is often called the adversarial \emph{attack} of the model $f_\theta$ and the corresponding solution is referred to as the adversarial example.
In this paper, we consider \emph{directed attacks\xspace}, as described in Section~\ref{sec:intro}, that effectively reduce the information about the ground truth classes.
Formally, we characterize \emph{directed attacks\xspace} by the following property:
for any model $f_\theta$ with low standard error, the corresponding adversarial example is well-aligned with the adversarial example found using the ground truth model
$f_{\theta^{\star}}$.
An example for such an attack are additive perturbations that are constrained to the direction of the ground truth decision boundary. We provide concrete examples for linear classification in Section~\ref{logreg_linear_model}.
\paragraph{Adversarial training}
In order to obtain classifiers with a good robust accuracy, it is
common practice to minimize a (robust) training objective $\mathcal{L}_{\eps_{\text{tr}}}$ with a surrogate
classification loss $L$ such as
\begin{equation}
\label{eq:emploss}
\robloss{\theta} := \frac{1}{n} \sum_{i=1}^n \max_{x_i' \in \pertset{x_i}{\eps_{\text{tr}}}} L(f_\theta(x_i') y_i),
\end{equation}
which is called adversarial training. In practice, we often use the
cross entropy loss $L(z) = \log (1+ \text{e}^{-z})$ and minimize the
robust objective by using first order optimization methods such as
(stochastic) gradient descent. SGD is also the algorithm that we
focus on in both the theoretical and experimental sections.
When the desired type of robustness is known in advance, it is
standard practice to use the same perturbation set for training as for
testing, i.e. $\pertset{x}{\eps_{\text{tr}}}=\pertset{x}{\eps_{\text{te}}}$. For example, \citet{madry18} shows that the robust error sharply increases for $\eps_{\text{tr}} < \eps_{\text{te}}$.
In this paper, we show that for directed attacks\xspace in the small sample size regime, in fact, the opposite is true.
\section{Theoretical results}
\label{sec:theoryresults}
In this section, we prove for linear functions $f_\theta(x) =
\theta^\top x$ that in the case of directed attacks, robust
generalization deteriorates with increasing $\eps_{\text{tr}}$.
The proof, albeit in a simple setting, provides
explanations for why adversarial training fails in the
high-dimensional regime for such attacks.
\subsection{Setting}
\label{logreg_linear_model}
We now introduce the precise linear setting used in our theoretical results.
\paragraph{Data model}
In this section, we assume that the ground truth and hypothesis class
are given by linear functions $f_\theta(x) = \theta^\top x$ and the
sample size $n$ is lower than the ambient dimension $d$. In
particular, the generative distribution $\mathbb{P}_r$ is similar to
\cite{tsipras19, kolter19}: The label $y \in \{+1, -1\}$ is drawn with
equal probability and the covariate vector is sampled as $x =
[y\frac{r}{2}, \tilde{x}]$ with the random vector $\tilde{x} \in
\mathbb{R}^{d-1}$ drawn from a standard normal distribution,
i.e. $\tilde{x} \sim \mathcal{N}(0, \sigma^2 I_{d-1})$. We would like to
learn a classifier that has low robust error by using a dataset
$D = {(x_i, y_i)}_{i=1}^n$ with $n$ i.i.d. samples from
$\mathbb{P}_{r}$.
Notice that the distribution $\mathbb{P}_{r}$ is noiseless: for a given input
$x$, the label $y = \sign(\xind{1})$ is deterministic. Further, the
optimal linear classifier (also referred to as the \emph{ground
truth}) is parameterized by $\theta^{\star} = e_1$.\footnote{Note that the result more generally holds for non-sparse models that are not axis aligned by way of a simple rotation $z = U x$. In that case the distribution is characterized by $\theta^\star = u_1$ and a rotated Gaussian in the $d-1$ dimensions orthogonal to $\theta^\star$.} By definition, the ground truth is
robust against all consistent perturbations and hence the optimal
robust classifier.
\paragraph{Directed attacks\xspace}
The focus in this paper lies on consistent directed attacks\xspace that by
definition efficiently concentrate their attack budget in the
direction of the signal. For our linear setting, we can model such
attacks by additive perturbations in the first dimension
\begin{equation}
\label{eq:linfmaxpert}
\pertset{x}{\epsilon} = \{x'=x+\delta \mid \delta = \beta e_1 \text{ and } -\epsilon \leq \beta\leq \epsilon\}.
\end{equation}
Note that this attack is always in the direction of the true signal dimension, i.e. the ground truth. Furthermore, when $\epsilon < \frac{r}{2}$, it is a consistent directed attack\xspace.
Observe how this is different from $\ell_p$ attacks - an $\ell_p$ attack, depending on the model, may add a perturbation that only has a very small component in the signal direction.
\paragraph{Robust max-$\ell_2$-margin classifier}
A long line of work studies the implicit bias of interpolators
that result from applying stochastic gradient descent on the logistic loss until convergence \cite{liu20, Ji19, Chizat20, nacson19}.
For linear models, we obtain the $\eps_{\text{tr}}$-robust maximum-$\ell_2$-margin solution (\emph{robust max-margin} in short)
\begin{equation}
\label{eq:maxmargin}
\thetahat{\eps_{\text{tr}}} := \argmax_{\|\theta\|_2\leq 1} \min_{i\in [n], x_i' \in \pertset{x_i}{\eps_{\text{tr}}}} y_i \theta^\top x_i'.
\end{equation}
This can for example be shown by a simple rescaling
argument using Theorem 3.4 in \cite{liu20}. Even though our result is proven for the max-$\ell_2$-margin classifier,
it can easily be extended to other interpolators.
\begin{figure*}[!t]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/gap_lower_final_main_theorem.png}
\caption{Robust error increase with $\eps_{\text{tr}}$}
\label{fig:main_lower_bound_eps}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/robust_error_ST_AT.png}
\caption{Standard-adversarial training}
\label{fig:main_numobs}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/gap_final_good_colours.png}
\caption{Effect of over-parameterization}
\label{fig:main_numobs_bound}
\end{subfigure}
\caption{Experimental verification of Theorem \ref{thm:linlinf}.
(a) We set $d = 1000$, $r = 12$, $n = 50$ and plot the robust error gap between standard and adversarial training with increasing adversarial budget $\eps_{\text{tr}}$ of $5$ independent experiments. For comparison, we also plot the lower bound given in Theorem \ref{thm:linlinf}. In (b) and (c), we set $d = 10000$ and vary the number of samples $n$. (b) We plot the robust error of standard and adversarial training ($\eps_{\text{tr}} = 4.5$). (c) We compute the error gap and the lower bound of Theorem \ref{thm:linlinf}. For more experimental details see Appendix~\ref{sec:logregapp}.}
\vspace{-0.2in}
\label{fig:main_theorem}
\end{figure*}
\subsection{Main results}
\label{logreg_main_theorem}
We are now ready to characterize the
$\eps_{\text{te}}$-robust error as a function of $\eps_{\text{tr}}$, the separation
$r$, the dimension $d$ and sample size $n$ of the
data. In the theorem statement we use the following quantities
\begin{align*}
\varphi_{\text{min}} &= \frac{\sigma}{r/2-\eps_{\text{te}}} \left( \sqrt{\frac{d-1}{n}} - \left(1 + \sqrt{\frac{2 \log (2/\delta)}{n}}\right)\right)\\
\varphi_{\text{max}} &= \frac{\sigma}{r/2-\eps_{\text{te}}} \left( \sqrt{\frac{d-1}{n}} + \left(1 + \sqrt{\frac{2 \log (2/\delta)}{n}}\right)\right)
\end{align*}
that arise from concentration bounds for the singular values of the random data matrix. Further, let $\tilde{\epsilon} := \frac{r}{2} - \frac{\varphi_{\text{max}}}{\sqrt{2}}$ and denote by
$\Phi$ the cumulative distribution function of a standard normal.
\begin{theorem}
\label{thm:linlinf}
Assume $d-1>n$.
For any $\eps_{\text{te}} \geq 0$, the $\eps_{\text{te}}$-robust error on test samples from $\mathbb{P}_{r}$ with $2 \eps_{\text{te}} < r$ and perturbation sets in Equation~\eqref{eq:linfmaxpert} and~\eqref{eq:l1maxpert}, the following holds:
\begin{enumerate}
\item
The $\eps_{\text{te}}$-robust error of the $\eps_{\text{tr}}$-robust max-margin estimator reads
\begin{equation}
\roberr{\thetahat{\eps_{\text{tr}}}} = \Phi \left( -\frac{\left( \frac{r}{2}-\eps_{\text{tr}} \right) }{\tilde{\varphi}} \right)
\end{equation}
for a random quantity $\tilde{\varphi}>0$ depending on $\sigma, r,\eps_{\text{te}}$, which is a strictly increasing function with respect to $\eps_{\text{tr}}$.
\item
With probability at least $1-\delta$, we further have $\varphi_{\text{min}} \leq \tilde{\varphi}\leq \varphi_{\text{max}}$ and the following lower bound on the robust error increase by adversarially training with size $\eps_{\text{tr}}$
\begin{equation}
\roberr{\thetahat{\eps_{\text{tr}}}} - \roberr{\thetahat{0}}
\geq
\Phi \left(\frac{r/2}{\varphi_{\text{min}}} \right) - \Phi \left( \frac{r/2 -\min\{\eps_{\text{tr}}, \tilde{\epsilon}\}}{ \varphi_{\text{min}}} \right).
\end{equation}
\end{enumerate}
\end{theorem}
The proof can be found in Appendix~\ref{sec:app_theorylinear} and
primarily relies on high-dimensional probability. Note that the
theorem holds for any $0\leq \eps_{\text{te}} <\frac{r}{2}$ and hence
also directly applies to the standard error by setting $\eps_{\text{te}} =
0$. In Figure~\ref{fig:main_theorem}, we empirically confirm the statements of Theorem \ref{thm:linlinf} by performing multiple experiments on synthetic datasets as described in Subsection \ref{logreg_linear_model} with different choices of $d/n$ and $\eps_{\text{tr}}$.
In the first statement, we prove that for small
sample-size ($n<d-1$) noiseless data,
almost surely, the robust error increases monotonically with
adversarial training budget $\eps_{\text{tr}} >0$.
In Figure~\ref{fig:main_lower_bound_eps}, we plot the robust error gap between standard and adversarial logistic regression in function of the adversarial training budget $\eps_{\text{tr}}$ for $5$ runs.
The second statement establishes a simplified lower bound on the
robust error increase for adversarial training (for a fixed
$\eps_{\text{tr}} = \eps_{\text{te}}$) compared to standard training.
In Figures~\ref{fig:main_lower_bound_eps} and \ref{fig:main_numobs_bound}, we show how the lower bound closely
predicts the robust error gap in our synthetic experiments.
Furthermore, by the dependence of $\varphi_{\text{min}}$ on the overparameterization ratio $d/n$, the lower bound on the robust error gap is amplified for large $d/n$.
Indeed, Figure~\ref{fig:main_numobs_bound} shows how the error gap increases with $d/n$
both theoretically and experimentally. However, when $d/n$ increases above a certain threshold, the gap decreases again, as standard training fails to learn the signal and yields a high error (see Figure~\ref{fig:main_numobs}).
\begin{figure*}[!t]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/d_n_logreg.png}
\caption{Robust error vs $\eps_{\text{tr}}$}
\label{fig:eps_logreg}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/logreg_trade_off_plot.png}
\caption{Robust error decomposition}
\label{fig:main_robust}
\end{subfigure}
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=0.99\linewidth]{plotsAistats/linear_intuition_try.png}
\caption{Intuition in 2D}
\label{fig:2D_dataset_intuition}
\end{subfigure}
\caption{(a) We set $d=1000$ and $r = 12$ and plot the robust error with increasing adversarial training budget ($\eps_{\text{tr}}$) and with increasing $d/n$. (b) We plot the robust error decomposition in susceptibility and standard error for increasing adversarial budget $\eps_{\text{tr}}$.
Full experimental details can be found in Section~\ref{sec:logregapp}. (c) 2D illustration providing intuition for the linear setting: Training on directed attacks\xspace (yellow) effectively corresponds to fiting the original datapoints (blue) after shifting them closer to the decision boundary. The robust max-$\ell_2$-margin (yellow dotted) is heavily tilted if the points are far apart in the non-signal dimension, while the standard max-$\ell_2$-margin solution (blue dashed) is much closer to the ground truth (gray solid). }
\label{fig:lineartradeoff}
\end{figure*}
\subsection{Proof idea: intuition and surprises}
\label{logreg_proof_sketch}
The reason that adversarial training hurts
robust generalization is based on an extreme robust vs. standard
error tradeoff. We provide intuition for the effect of
directed attacks\xspace and the small sample regime
on the solution of adversarial training by decomposing the
robust error $\roberr{\theta}$.
Notice that $\eps_{\text{te}}$-robust error $\roberr{\theta}$
can be written as the probability of the union of two events: the
event that the classifier based on $\theta$ is wrong and the event
that the classifier is susceptible to attacks:
\begin{equation}
\label{eq:decomposition}
\begin{aligned}
\roberr{\theta} &= \mathbb{E}_{x, y\sim \mathbb{P}} \left[\Indi{y f_\theta (x) <0} \vee \max_{x' \in \pertset{x}{\eps_{\text{te}}}} \Indi{f_\theta(x) f_\theta(x')<0} \right] \\
&\leq \stderr{\theta} + \suscept{\theta}
\end{aligned}
\end{equation}
where $\suscept{\theta}$ is the expectation of the maximization term in Equation \eqref{eq:decomposition}.
$\suscept{\theta}$ represents the $\eps_{\text{tr}}$-\emph{attack-susceptibility} of a classifier
induced by $\theta$ and $\stderr{\theta}$ its standard error.
Equation~\eqref{eq:decomposition} suggests that
the robust error can only be small if both the standard error and
susceptibility are small. In Figure~\ref{fig:main_robust}, we
plot the decomposition of the robust error in standard error and susceptibility for adversarial logistic regression with increasing $\eps_{\text{tr}}$. We observe that increasing $\eps_{\text{tr}}$
increases the standard error too drastically compared to the decrease
in susceptibility, leading to an effective drop in robust accuracy. For completeness, in Appendix \ref{app:susc}, we provide upper and lower bounds for the susceptibility score. We
now explain why, in the small-sample size regime, adversarial training
with directed attacks\xspace ~\eqref{eq:linfmaxpert} may increase standard
error to the extent that it dominates the decrease in susceptibility.
A key observation is that the robust max-$\ell_2$-margin solution of a
dataset $D= \{(x_i, y_i)\}_{i=1}^n$
maximizes the minimum margin that reads ${\min_{i\in [n]}
y_i \theta^\top (x_i - y_i \eps_{\text{tr}} |\thetaind{1}| e_1)}$, where
$\indof{\theta}{i}$ refers to the $i$-th entry of vector $\theta$. Therefore, it
simply corresponds to the max $\ell_2$-margin solution of the dataset
shifted towards the decision boundary ${D_{\epstrain} = \{(x_i - y_i \eps_{\text{tr}}
|\indof{\thetahat{\eps_{\text{tr}}}}{1}| e_1, y_i)\}_{i=1}^n}$.
Using this fact, we obtain
a closed-form expression of the (normalized) max-margin solution~\eqref{eq:maxmargin} as a function of
$\eps_{\text{tr}}$ that reads
\begin{equation}
\label{eq:maxmarginmaintext}
\thetahat{\eps_{\text{tr}}} = \frac{1}{(r-2\eps_{\text{tr}})^2 + 4 \tilde{\gamma}^2}
\left[r - 2\eps_{\text{tr}}, 2 \tilde{\gamma} \tilde{\theta} \right],
\end{equation}
where $\|\tilde{\theta}\|_2 = 1$ and $\tilde{\gamma} >0$ is a random quantity
associated with the max-$\ell_2$-margin solution of the
$d-1$ dimensional Gaussian inputs orthogonal
to the signal direction
(see Lemma~\ref{lem:maxmargin} in Section~\ref{sec:app_theorylinear}).
In high dimensions, with high probability any two
Gaussian random vectors are far apart -- in our
distributional setting, this corresponds to the vectors being far
apart in the non-signal directions. In
Figure~\ref{fig:2D_dataset_intuition}, we illustrate the phenomenon
using a simplified 2D cartoon, where the few samples
in the dataset are all far apart in the non-signal direction.
We see how shifting the dataset closer to the true decision boundary,
may result in a max-margin solution (yellow) that aligns much worse
with the ground truth (gray), compared to the estimator learned from
the original points (blue). Even though the new (robust max-margin)
classifier (yellow) is less susceptible to directed attacks in the
signal dimension, it also uses the signal dimension less.
Mathematically, this is directly
reflected in the expression of the max-margin solution in
Equation~\eqref{eq:maxmarginmaintext}: Even without the definition of
$\tilde{\gamma}, \tilde{\theta}$, we can directly see that the first
(signal) dimension is used less as $\eps_{\text{tr}}$ increases.
\subsection{Generality of the results}
In this section we discuss how the theorem might generalize to
other perturbation sets, models and training procedures.
\paragraph{Signal direction is known}
The type of additive perturbations used in Theorem~\ref{thm:linlinf},
defined in Equation~\eqref{eq:linfmaxpert}, is explicitly constrained
to the direction of the true signal. This choice is reminiscent of
corruptions where every possible perturbation in the set is directly
targeted at the object to be recognized, such as motion blur of moving
objects. Such corruptions are also studied in the context of domain
generalization and adaptation \cite{Schneider20}.
Directed attacks\xspace in general, however, may also consist of
perturbation sets that are only strongly biased towards the true
signal direction, such as mask attacks. They may find the true signal
direction only when the inner maximization is
exact. The following corollary extends Theorem~\ref{thm:linlinf} to
small $\ell_1$-perturbations
\begin{equation}
\label{eq:l1maxpert}
\pertset{x}{\epsilon} = \{x'=x+\delta \mid \|\delta\|_1 \leq \epsilon\},
\end{equation}
for $0<\epsilon<\frac{r}{2}$ that reflect such attacks. We state the corollary here and give the proof in Appendix \ref{sec:app_theorylinear}.
\begin{corollary}
\label{cor:l1extension}
Theorem~\ref{thm:linlinf} also holds for ~\eqref{eq:maxmargin} with perturbation sets defined in \eqref{eq:l1maxpert}.
\end{corollary}
The proof uses the fact that the inner maximization
effectively results in a sparse perturbation equivalent to the attack
resulting from the perturbation set~\eqref{eq:linfmaxpert}.
\paragraph{Other models}
Motivated by the implicit bias results of (stochastic)
gradient descent on the logistic loss, Theorem~\ref{thm:linlinf} is proven for the max-$\ell_2$-margin
solution. We would like to conjecture
that for the data distribution in Section \ref{sec:theoryresults},
adversarial training can hurt robust generalization also for other models with zero
training error (\emph{interpolators} in short).
For example, Adaboost is a widely used algorithm that converges to the max-$\ell_1$-margin classifier \cite{telgarsky13}. One might argue that for a sparse ground truth, the max-$\ell_1$-margin classifier should (at least in the noiseless case) have the right inductive bias to alleviate large bias in high dimensions. Hence, in many cases the (sparse) max-$\ell_1$-margin solution might align with the ground
truth for a given dataset. However, we conjecture that even in this
case, the \emph{robust} max-$\ell_1$-margin solution (of the dataset
shifted towards the decision boundary) would be misled to choose a
wrong sparse solution. This can be seen with the help of the cartoon
illustration in Figure \ref{fig:2D_dataset_intuition}.
| a329cc7a53475e6f2c1dd000859adabccc891d6d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Energy calibration is the first and most essential step in a newly installed accelerator facility. There are several experiments performed by different accelerator facilities across the globe for energy calibration such as $^{7}Li(p,n)$, $^{13}C(p,n)$, $^{27}Al(p,n)$, $^{27}Al(p,\gamma)$, $^{19}F(p,n)$ etc.. Among which neutron emitting reactions are very popular due to their relatively easier detections. Every neutron producing reaction has a particular energy, called neutron threshold energy which is the minimum required energy of the incident beam to produce neutron. The present study has been done keeping in mind the Facility for Research in Experimental Nuclear Astrophysics (FRENA)\cite{Ref1} located at Saha Institute of Nuclear Physics, INDIA. FRENA is a 3 MV Tandetron low energy high current accelerator dedicated for low energy Nuclear Astrophysics research. This is a new machine and energy calibration is needed before doing any experiment. Few neutron-producing reactions are generally chosen for calibration purposes. Neutrons generated from such experiments can interact with any elements present in the accelerator hall and produce different isotopes. Now, if produced isotopes are radioactive in nature and have a long half-life (for few months to several years) then those isotopes will contribute as background gamma peaks in the next performing experiments. These background peaks will overlap with the gammas from the actual experiments if the energy of those background gammas is very close to the gammas from actual astrophysical experiments. A systematic study has been done for all the possible radioactive isotopes that may produce during some experiments which are planned to perform in FRENA for energy calibration of the machine.
\section{Energy calibration for FRENA accelerator}
The terminal voltage of FRENA can be varied from 200 kV to 3 MV. The energy of the charged particle produced in the ion source is accelerated by this terminal voltage.
The energy of any charged particle in a potential difference is given by the relation:
\begin{equation}
E_{particle} = (q+1) \times V.
\end{equation}
Where, $E_{particle}$ is the charged particle energy, q is the charge state of the projectile, V is the terminal voltage.
In practical cases, the accelerated charge passes through several magnetic fields and finally bombard into a target material. Since every machine has some intrinsic deviation from the ideal value. Energy needs to be calibrated so that a more accurate relation between terminal voltage and beam energy can be established for this particular machine.
\\
\section{Results for different calibration reactions}
In FRENA, available proton energy varies from 400 keV to 6 MeV. For energy calibration purposes, those (p,n) experiments (Table. 1) are chosen for which neutron threshold energy lies within the available energy at FRENA.
\begin{table}
\begin{center}
\begin{tabular}{|p{4cm}|p{2cm}|p{2cm}|p{2cm}|p{2cm}|}
\hline
& $^{7}Li(p,n)$ &$^{13}C(p,n)$ &$^{19}F(p,n)$& $^{27}Al(p,n)$ \\ \hline
Neutron threshold energy (MeV) & 1.88036 & 3.2355 & 4.23513 & 5.80362 \\ \hline
\end{tabular}
\caption{Neutron threshold energy for different (p,n) reactions.}
\end{center}
\end{table}
Most of the components of the accelerators are made of 304 and 316 Stainless steel alloy along with copper (Cu) and Tantalum (Ta). Apart from that Zirconium (Zr), Aluminium (Al), Calcium (Ca), Silicon (Si) and Oxygen (O) are also present in the vicinity. Chemical composition for 304 and 316 stainless steel alloys along with maximum \% are listed in Table. 2 and Table. 3.
\begin{table}
\begin{center}
\begin{tabular}{|p{1cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{1.2cm}|p{0.8cm}|}
\hline
& C & Mn & Si & P & S & Cr & Ni & Fe & N \\ \hline
Max\% & 0.07 & 2.00 & 1.00 & 0.05 & 0.03 & 19.50 & 10.50 & Balance & 0.11\\ \hline
\end{tabular}
\caption{Maximum \% of chemical compositions present in 304 Stainless steel alloy \cite{Ref2}. }
\end{center}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|p{1cm}|p{0.5cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{0.8cm}|p{1.2cm}|p{0.8cm}|}
\hline
& C & Mn & Si & P & S & Cr & Mo &Ni & Fe & N \\ \hline
Max\% & 0.08 & 2.00 & 0.75 & 0.045 & 0.03 & 18.00 &3.00& 14.00& Balance & 0.10\\ \hline
\end{tabular}
\caption{Maximum \% of chemical compositions present in 316 Stainless steel alloy \cite{Ref3}. }
\end{center}
\end{table}
Reaction cross-section of isotopes present in the vicinity with produced neutrons from calibration experiments are calculated using a Hauser-Feshbach statistical model code TALYS 1.95 \cite{Ref4}. In this study, channels having cross-sections less than $10^{-7}$ mb are not considered. Gamma decay schemes of every produced isotope are taken from National Nuclear Data Center (NNDC)\cite{Ref5}. Radioactive isotopes produced by neutrons having a half-life of around 3 months or more are considered. Gammas having a relative intensity of more than 1\% from those radioactive isotopes are listed in this study.
\subsection{$^{7}Li(p,n)$ reaction:}
$^{7}Li(p,n)$ is one of the most common experiments done for energy calibration at accelerator facilities. Proton beam of energy 1.88036 MeV was considered for this study (Table. 1). From kinematics neutron energy for that particular proton beam energy has been calculated and it was found to be between 20 - 42 keV. In calculations, only maximum and minimum values of neutron energy are considered. For the interaction of produced neutron, all the stable isotopes of the elements present in the vicinity are considered. \\
Table. 4 also shows isotopic abundance in nature, reaction product, and its formation cross-section along with gamma energy and corresponding relative intensities are also listed.
\begin{figure}
\centering
\includegraphics[width=.8\linewidth]{fig1.pdf}
\caption{Distribution of neutron energy at threshold energy, $E_p$=1.88036 MeV}
\end{figure}
\subsection{$^{13}C(p,n)$ reaction:}
A proton beam of energy 3.2355 MeV was considered for the study of this reaction (Table. 1). From kinematics neutron energy for that particular proton beam energy has been calculated and it was found to be between 11 - 24 keV.
\begin{figure}
\centering
\includegraphics[width=.8\linewidth]{fig2.pdf}
\caption{Distribution of neutron energy at threshold energy, $E_p$=3.2355 MeV}
\end{figure}
\subsection{$^{19}F(p,n)$ reaction:}
Neutron threshold energy for this reaction is 4.23513 MeV. Proton of that energy is considered for this calculation. Neutrons emitted from this reaction ranges between 4 to 20 keV.
\begin{figure}
\centering
\includegraphics[width=.8\linewidth]{fig3.pdf}
\caption{Distribution of neutron energy at threshold energy, $E_p$=4.23513 MeV}
\end{figure}
\subsection{$^{27}Al(p,n)$ reaction:}
A proton of energy 5.80362 MeV was considered for this study (Table. 1). From kinematics available neutron energy for that particular proton beam energy is between 0 - 30 keV.
\begin{figure}[h]
\centering
\includegraphics[width=.8\linewidth]{fig4.pdf}
\caption{Distribution of neutron energy at threshold energy, $E_p$=5.80362 MeV}
\end{figure}
Kinematics calculations show that neutron energy for all reactions mentioned above is within the 10 to 40 keV range. Cross-sections of different stable isotopes present in the vicinity of the accelerator hall with the produced neutrons from the calibration reactions are listed in Table. 4. Half-lives of reaction products, daughter nucleus with corresponding gamma energies having relative intensities greater or equal to 1 are also listed in the table.
\begin{table}
\tiny
\begin{center}
\begin{tabular}{|p{.5cm}|p{.5cm}|p{.8cm}|p{.8cm}|p{1cm}|p{1cm}|p{.8cm}|p{1cm}|p{1.2cm}|p{1.5cm}|}
\hline
Iso-topes & Abun-dance (\%) & Neutron energy (keV)& Reaction product & Cross section (mb) & half life & Daughter nucleus & Gamma energy (keV) & Gamma intensity (\%) & Remarks\\ \hline \hline
\multirow{2}{*}{$^{13}C$} & \multirow{2}{*}{1.07} & 10 & \multirow{2}{*}{$^{14}C$} &\multirow{1}{*}{1.59$\times 10^{-1}$} & \multirow{2}{*}{5700 y} & \multirow{2}{*}{$^{14}N$} & & & \multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 8.61$\times 10^{-2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 5.44$\times 10^{-2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 4.38$\times 10^{-2}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{14}N$} & \multirow{2}{*}{99.6} & 10 & \multirow{2}{*}{$^{14}C$} &\multirow{1}{*}{1.28$\times 10^{3}$} & \multirow{2}{*}{5700 y} & \multirow{2}{*}{$^{14}N$} & & & \multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 8.68$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 5.83$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 5.34$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{17}O$} & \multirow{2}{*}{0.04} & 10 & \multirow{2}{*}{$^{14}C$} &\multirow{1}{*}{2.80$\times 10^{2}$} & \multirow{2}{*}{5700 y} & \multirow{2}{*}{$^{14}N$} & & &\multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 2.97$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 1.30$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 1.89$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{34}S$} & \multirow{2}{*}{4.25} & 10 & \multirow{2}{*}{$^{35}S$} &\multirow{1}{*}{2.00} & \multirow{2}{*}{87.37 d} & \multirow{2}{*}{$^{35}Cl$} & & & \multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 1.216 &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 9.41$\times 10^{-1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 7.82$\times 10^{-1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{40}Ca$} & \multirow{2}{*}{96.941} & 10 & \multirow{2}{*}{$^{41}Ca$} &\multirow{1}{*}{5.25$\times 10^{1}$} & \multirow{2}{*}{99400 y} & \multirow{2}{*}{$^{41}K$} & & & \multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 2.46$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 2.69$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 1.68$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{44}Ca$} & \multirow{2}{*}{2.086} & 10 & \multirow{2}{*}{$^{45}Ca$} &\multirow{1}{*}{6.50$\times 10^{2}$} & \multirow{2}{*}{162.61 d} & \multirow{2}{*}{$^{45}Sc$} & \multirow{2}{*}{12.47}& \multirow{2}{*}{3.0$\times 10^{-6}$}& \\ \cline{3-3} \cline{5-5}
&&20 && 4.15$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 3.25$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 2.74$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{54}Fe$} & \multirow{2}{*}{5.85} & 10 & \multirow{2}{*}{$^{55}Fe$} &\multirow{1}{*}{9.46$\times 10^{1}$} & \multirow{2}{*}{2.744 y} & \multirow{2}{*}{$^{55}Mn$} & \multirow{2}{*}{126}&\multirow{2}{*}{1.28 $\times 10^{-7}$} & \\ \cline{3-3} \cline{5-5}
&&20 && 6.35$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 5.17$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 4.56$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{58}Ni$} & \multirow{2}{*}{68.07} & 10 & \multirow{2}{*}{$^{59}Ni$} &\multirow{1}{*}{2.16$\times 10^{2}$} & \multirow{2}{*}{7.6$\times 10^{4}$ y} & \multirow{2}{*}{$^{59}Co$} & \multirow{2}{*}{511}&\multirow{2}{*}{7.4 $\times 10^{-5}$} & Annihilation gammas \\ \cline{3-3} \cline{5-5}
&&20 && 1.47$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 1.20$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 1.06$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{62}Ni$} & \multirow{2}{*}{3.635} & 10 & \multirow{2}{*}{$^{63}Ni$} &\multirow{1}{*}{5.45$\times 10^{1}$} & \multirow{2}{*}{101.2 y} & \multirow{2}{*}{$^{63}Cu$} & & & \multirow{2}{*}{100\% $\beta$ -decay} \\ \cline{3-3} \cline{5-5}
&&20 && 3.78$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 3.18$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 2.89$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{92}Mo$} & \multirow{2}{*}{14.65} & 10 & \multirow{2}{*}{$^{93}Mo$} &\multirow{1}{*}{1.17$\times 10^{2}$} & \multirow{2}{*}{4.0$\times 10^{3}$y} & \multirow{2}{*}{$^{93}Nb$} & \multirow{2}{*}{30.77}&\multirow{2}{*}{5.20 $\times 10^{-4}$} & \\ \cline{3-3} \cline{5-5}
&&20 && 6.80$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 5.50$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 4.05$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{2}{*}{$^{92}Zr$} & \multirow{2}{*}{17.15} & 10 & \multirow{2}{*}{$^{93}Zr$} &\multirow{1}{*}{8.36$\times 10^{1}$} & \multirow{2}{*}{1.61$\times 10^{6}$y} & \multirow{2}{*}{$^{93}Nb$} & \multirow{2}{*}{30.77}&\multirow{2}{*}{4.30 $\times 10^{-4}$} & \\ \cline{3-3} \cline{5-5}
&&20 && 4.77$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 3.49$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 2.83$\times 10^{1}$ &&&&& \\ \cline{3-3} \cline{5-5}
\hline
\hline
\multirow{20}{*}{$^{181}Ta$} & \multirow{20}{*}{99.988} & 10 & \multirow{20}{*}{$^{182}Ta$} &\multirow{1}{*}{2.09$\times 10^{3}$} & \multirow{20}{*}{114.74 d} & \multirow{20}{*}{$^{182}W$} & \multirow{1}{*}{65.722}&\multirow{1}{*}{3.01} & \\ \cline{3-3} \cline{5-5} \cline{8-9}
&&20 && 1.34$\times 10^{3}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&30 && 1.05$\times 10^{3}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&40 && 8.69$\times 10^{2}$ &&&&& \\ \cline{3-3} \cline{5-5}
&&&&&&& 67.749 & 42.9& \\ \cline{8-9}
&&&&&&& 84.680 & 2.654& \\ \cline{8-9}
&&&&&&& 100.106 & 14.2& \\ \cline{8-9}
&&&&&&& 113.672 & 1.871& \\ \cline{8-9}
&&&&&&& 152.429 & 7.02& \\ \cline{8-9}
&&&&&&& 156.386 & 2.671& \\ \cline{8-9}
&&&&&&& 179.393 & 3.119& \\ \cline{8-9}
&&&&&&& 198.352 & 1.465& \\ \cline{8-9}
&&&&&&& 222.109 & 7.57& \\ \cline{8-9}
&&&&&&& 229.321 & 3.644& \\ \cline{8-9}
&&&&&&& 264.074 & 3.612& \\ \cline{8-9}
&&&&&&& 1001.7 & 2.086& \\ \cline{8-9}
&&&&&&& 1121.29 & 35.24& \\ \cline{8-9}
&&&&&&& 1189.04 & 16.49& \\ \cline{8-9}
&&&&&&& 1221.395 & 27.23& \\ \cline{8-9}
&&&&&&& 1231.004 & 11.62& \\ \cline{8-9}
&&&&&&& 1257.407 & 1.51& \\ \cline{8-9}
&&&&&&& 1289.145 & 1.37& \\ \cline{8-9}
\hline
\hline
\end{tabular}
\caption{Cross-section of different isotopes present in the vicinity of the accelerator with neutrons of different energy, half-lives of produced isotopes and gamma energy with relative intensity
. }
\end{center}
\end{table}
\clearpage
\section{Discussion and conclusion}
In this work, a systematic detailed study has been done to identify any possible long-lived gamma emitting isotopes formed during the calibration study. Apart from the above listed isotopes two more isotopes $^{100}Mo$ and $^{180m}Ta$ (0.077 MeV, 9-) are produced in the vicinity during the energy calibration. These two isotopes have no decay record in NNDC. Decay of $^{180m}Ta$ having half-life $>4.5 \times 10^{16} y$ has never been observed till date\cite{Ref6} \cite {Ref7}. $^{100}Mo$ having half-life $6.8 \times 10^{18} y$ undergoes two-neutrino double $\beta$ -decay \citep{Ref8}. Reaction products such as $^{55}Fe$, $^{59}Ni$, $^{45}Ca$, $^{93}Mo$, $^{93}Zr$ have half-lives much greater than 3 months but intensities of gammas from those nuclei have very low intensity. Isotopes like $^{63}Ni$, $^{14}C$, $^{35}S$, $^{41}Ca $ have significantly long half-lives. These isotopes undergo
$\beta-$ decay, with no record of any gamma decay channel in NNDC. This study shows that experiments mentioned above for the energy calibration purpose are producing only $^{182}Ta$ which has significant long-lived activity (114.72 days) and decay gamma over a wide range of energies (60 - 1300 keV) with an intensity of more than 1\%. several gamma energy over a wide range. Gamma of energies 1121.29, 1189.04, 1221.395, 1231.004 keV having intensities 35.24, 16.49, 11.23, 11.62 respectively can have a significant effect on experimental data if performed just after the calibration experiments. Nuclear astrophysics experiments detecting gamma after the calibration study needs to be spaced enough to minimize the effect of such background gammas from long-lived isotopes especially if the expected gamma energy of the performing experiments overlap with the background energy.
\\
\section*{Acknowledgement}
| b6dde4709c6dbbb0a45e93a31d97d6891a24ee12 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Since the discovery of superconducting \cite{cao2018unconventional,yankowitz2019tuning} and insulating \cite{cao2018correlated} behavior in magic angle graphene, twistronics has quickly become one of the most active fields in condensed matter physics. One of the most recent advancements in the field is the discovery of ferroelectric behavior in twisted hetero-bilayers \cite{zheng2020unconventional,yasuda2021stacking}. The discovery was regarded as highly unconventional because it is uncommon (but not impossible \cite{ji2019freestanding}) for two-dimensional (2D) systems to exhibit ferroelectric behavior. In addition to the observations of ferroelectricity, polarization domains have been observed without an overall ferroelectric response \cite{woods2021charge}. A switching of polarization in a stacking domain of hexagonal boron nitride (hBN) via van der Waals sliding has also been observed \cite{stern2020interfacial}. Although claimed to be ferroelectricity, insufficient measurements were taken to show hysteric behavior of the sample. In Ref.~\cite{yasuda2021stacking}, a first-order switching of the polarization domains was observed in \textit{untwisted} hBN, but in twisted hBN the hysteresis loop was continuous and very narrow. Thus, how and when moir\'e superlattices exhibit ferroelectric behavior is still unclear.
It had been proposed a few years before the discoveries that two-dimensional layered systems could exhibit a local out-of-plane polarization domain structure \cite{li2017binary}, which could be inverted by a relative `van der Waals' sliding between the layers, changing the stacking arrangments. Applying this concept to a twisted bilayer, the stacking domains can be considered polar domains, where neighboring domains have opposite polarization. It has been suggested that the relative sizes of these domains can be tuned with the application of an electric field perpendicular to the bilayer, leading to a nonzero total polarization of the superlattice \cite{bennett2022electrically,enaldiev2021scalable}, which can be very nonlinear when the moir\'e period is large. Although experimental measurements clearly show a ferroelectric response in homo-bilayers, from a theoretical perspective an ideal homo-bilayer has zero total polarization at zero field \cite{bennett2022electrically}. Even in an imperfect sample, which would have a nonzero total polarization due to uneven stacking domains, the orientation at zero field cannot be inverted. Thus, an ideal homo-bilayer does not satisfy the criteria for ferroelectricity via this mechanism, and additional effects must be required to explain the experimentally observed behavior.
Separate to the question of whether or not the overall system exhibits a ferroelectric response is the question of whether or not the stacking domains themselves are ferroelectric. Several theoretical studies have claimed, without any justification, that the stacking domains in homo-bilayers are ferroelectric \cite{ferreira2021weak,enaldiev2021piezoelectric,enaldiev2021scalable}. This is simply incorrect: the domains are determined by the structure, and even if they change in area in response to an applied field, the orientation of the polarization in each domain is pinned cannot be inverted \cite{bennett2022electrically}.
It is possible that the domains in homo-bilayers, could become ferroelectric via van der Waals sliding. An electric field is applied locally with a biased tip which scans the sample, and the average polarization is calculated. It has been suggested that when the local polarization is anti-aligned with the field, a sliding by one third of a unit cell occurs in order to change the stacking configuration and invert the polarization. However, in order for the domains to be ferroelectric, a global sliding of the layers must occur so that the orientation at zero field is reversible. Because the field is applied locally by scanning a biased tip over a sample, it is not clear whether the sliding occurs locally beneath the tip, and reverses once the tip moves on, or globally, and persists when the field is removed. Thus, while possible, it is not immediately clear whether the stacking domains are actually ferroelectric via this mechanism. Furthermore, in Ref.~\onlinecite{stern2020interfacial}, a switching of the orientation of a domain due to an applied field was observed, but whether or not this switching persisted at zero field was not investigated. In Ref.~\onlinecite{yasuda2021stacking}, a first-order polarization switching was observed in untwisted samples, but not twisted samples. Thus, ferroelectricity in moir\'e heterostructures via van der Waals sliding has not yet been experimentally observed.
In order to address the question of if and when moir\'e heterostructures exhibit ferroelectricity, and when the stacking domains can be considered ferroelectric domains rather than just polar domains, a better theoretical understanding of polar phenomena in moir\'e heterostructures is needed. One possible reason for the misconceptions mentioned above is that the discovery of polar properties of moir\'e heterostructures is relatively recent. They are very different from conventional ferroelectric materials such as oxide perovskites (\chem{ABO_3}), where polarization arises from soft-mode lattice instabilities, e.g.~the off-centering of the B cations with respect to the \chem{O_6} octahedra \cite{megaw1952origin,migoni1976origin}. For example, in barium titanate (\chem{BaTiO_3}, BTO) and lead titanate (\chem{PbTiO_3}, PTO), the Ti off-centering and ferroelectricity have been attributed to the hybridization of the titanium $3d$ and oxygen $2p$ states \cite{cohen1992origin}. The physical mechanism for polarization in moir\'e superlattices, the local breaking of out-of-plane inversion symmetry, is completely different, and while two unique phenomena have already been attributed to the domains, they are not well-understood in general. For example, while the polar domains in homo-bilayers (two identical monolayers, regardless of orientation) have been studied, the properties of polar domains in hetero-bilayers (two different monolayers), have not been investigated.
The aim of this paper is to propose that polar domains are a fundamental property of \textit{all} moir\'e heterostrctures, and to establish a basis for the theoretical understanding of these domains. In order to avoid further confusion with conventional ferroelectric domains, I will refer to polar domains in moir\'e heterostructures as `moir\'e polar domains' (MPDs) throughout this paper. The out-of-plane MPDs in bilayer hBN and the transition-metal dichalcogenides (TMDs) \chem{MX_2/M'X'_2}, with \chem{M,M'=Mo,W} and \chem{X,X'=S,Se}, are measured from first-principles calculations. It is demonstrated that the MPDs lead to non-uniform dielectric properties in moir\'e superlattices, such as interlayer charge transfer, polarizability, and the evolution of the local band gap due to an applied field. Although the concept of a local band gap is in general ill-defined, it has been used to estimate the potential landscape for interlayer moir\'e excitons \cite{jung2014ab,wu2014tunable,zhang2017interlayer,yu2017moire,wu2018theory,seyler2019signatures}, giving rise to novel exciton dynamical, optical and many-body phenomena \cite{remez2021dark}. While the theory of lattice relaxation and its influence on structural properties of moir\'e superlattices is well-known \cite{nam2017lattice,carr2018relaxation}, the practical details required to perform efficient calculations are typically omitted. A complete study of lattice relaxation in moir\'e heterostructures is provided in the Appendix, with particular emphasis on the use of crystal symmetries to greatly reduce the computational effort of the calculations. The properties of MPDs in homo- and hetero-bilayers as a function of moir\'e period (twist angle) and electric field via lattice relaxation are then discussed. Finally, a discussion of MPDs as a fundamental property of all moir\'e heterostructures is provided. While electrically tunable domains via lattice relaxation and first-order polarization switching via van der Waals sliding are interesting and unique properties, they are not criteria for moir\'e heterostructures to be considered polar materials. In light of the understanding of MPDs established in this paper, the question of whether moir\'e heterostructures are ferroelectric, and how they differ from conventional ferroelectrics, is addressed.
\section{Methods}
\subsection{Theoretical model}
The model originally used to describe the polar response in homo-bilayers via electrically tunable lattice relaxation \cite{bennett2022electrically} is used to describe MPDs in a general moir\'e heterostructure. It is a continuum elastic model, described by a total free energy which is an integral of a local energy density over a moir\'e period:
\beq{eq:V_tot_main}
V_{\text{tot}} = \frac{1}{\Asc}\int_{\Asc} \V_{\text{tot}}(\bvec{r}+\bvec{U}(\bvec{r}))\dd\bvec{r}
\;,\end{equation}
where $\Asc$ is the area of the moir\'e period, and a displacement field $\bvec{U}(\bvec{r})$ is included in order to allow for lattice relaxation. This model is general, and can describe various phenomena in moir\'e superlattices at the continuum level, depending on the contributions included in $\mathcal{V}_{\text{tot}}$. Typically, the intralayer elastic energy and interlayer stacking energy are included, and Eq.~\eqref{eq:V_tot_main} is minimized with respect to $\bvec{U}(\bvec{r})$ in order to describe the reconstruction of the stacking domains \cite{jung2015origin,nam2017lattice,zhang2018structural,carr2018relaxation}. For large moir\'e periods (small twist angles), the effect of lattice relaxation becomes significant, and leads wide stacking domains separated by sharp domain walls. By including the electrostatic energy due to the coupling between the MPD and an applied field, the effect of an applied field on lattice relaxation can be described \cite{bennett2022electrically}. The total energy density is then given by the sum of elastic, stacking and electrostatic energy densities:
\beq{eq:V_tot_all
\begin{gathered}
\mathcal{V}_{\text{tot}}(\bvec{r}) = \V_{\text{elastic}}(\nabla\bvec{U}(\bvec{r})) + \V_{\text{stack}}(\bvec{r}) + \Velec(\bvec{r})\\[5pt]
\mathcal{V}_{\text{elastic}}(\nabla\bvec{U}(\bvec{r})) = C_{ijkl}\epsilon_{ij}\epsilon_{kl}\\
\Vstack(\bvec{r}) = \V_0(\bvec{r})\\
\V_{\text{elec}}(\bvec{r}) = - \Ef p_0(\bvec{r})
\end{gathered}
\end{equation}
where $\epsilon_{ij} =\frac{1}{2}\left( \partial_i U_j + \partial_j U_i\right)$ is the strain tensor, $C$ is the elastic tensor \cite{carr2018relaxation,zhu2020twisted,bennett2022electrically}, $\V_0$ is the cohesive energy, $\Ef$ is the electric field perpendicular to the bilayer and $p_0(\bvec{r})$ is the local out-of-plane dipole moment. It is assumed that the layer separation $d(\bvec{r})$ is at its equilibrium value $d_0(\bvec{r})$ everywhere in the supercell, neglecting the effects of higher order elastic contributions \cite{jung2015origin} and the dielectric response to the field \cite{bennett2022electrically}. For a given system, Eq.~\eqref{eq:V_tot_all} can be paramaterized using first-principles calculations without allowing for prohibitively expensive in-plane relaxations, and Eq.~\eqref{eq:V_tot_main} can then be used to estimate the strain fields, and the structure after lattice relaxation. Even without the in-plane relaxations, this parameterization is still prohibitively expensive. Expensive calculations can be avoided by using the mapping from real space to configuration space \cite{carr2018relaxation,bennett2022electrically}, where every configuration in real space is condensed into a single primitive cell:
\beq{eq:slide_map}
\bvec{s}(\bvec{r}) = \left( I - R_{\th}^{-1} \right) \bvec{r} \mod \{ \bvec{a}_1,\bvec{a}_{2}\}
\;,\end{equation}
where $\bvec{s}$ is the corresponding position in configuration space, $R_{\th}$ is a rotation matrix, and $\bvec{a}_i$ are the lattice vectors of a commensurate bilayer. Although this mapping is exact, for angles which are small deviations from the commensurate stackings 3R ($\th = 0 + \frac{n\pi}{3}$) and 2H ($\th=\frac{\pi}{6}+ \frac{n\pi}{3}$) the positions in configuration space are well-described by a relative translation between the layers. After mapping Eqs.~\eqref{eq:V_tot_main} and \eqref{eq:V_tot_all} to configuration space, Eq.~\eqref{eq:V_tot_all} can be parameterized with first-principles calculations using a commensurate bilayer and sliding one layer over the other, but only for twist angles close to the commensurate 3R and 2H stackings.
\subsection{First-principles calculations}
The structural properties and MPDs of bilayer hBN, four TMD homo-bilayers (\chem{MoS_2}, \chem{MoSe_2}, \chem{WS_2}, \chem{WSe_2}) and the six hetero-bilayers formed from different combintations of the aforementioned TMD monolayers were measured in configuration space using first-principles calculations, following the methodology in Ref.~\onlinecite{bennett2022electrically}. Density functional theory (DFT) calculations were performed using the {\sc siesta} method and code \cite{siesta}, using PSML \cite{psml} norm-conserving \cite{norm_conserving} pseudopotentials, obtained from pseudo-dojo \cite{pseudodojo}. {\sc siesta} employs a basis set of numerical atomic orbitals (NAOs) \cite{siesta,siesta_2}, and double-$\zeta$ polarized (DZP) orbitals were used for all calculations. The basis sets were optimized by hand, following the methodology in Ref.~\onlinecite{basis_water}.
A mesh cutoff of $1200 \ \text{Ry}$ was used for the real space grid in all calculations. A Monkhorst-Pack $k$-point grid \cite{mp} of $12 \times 12 \times 1$ was used for the initial geometry relaxations, and a mesh of $18 \times 18 \times 1$ was used to calculate the out-of-plane dipole moments. Calculations were converged until the relative changes in the Hamiltonian and density matrix were both less than $10^{-6}$. For the geometry relaxations, the atomic positions were fixed in the in-plane directions, and the vertical positions and in-plane stresses were allowed to relax until the force on each atom was less than $0.1 \ \text{meV/\AA}$. The layer separation $d$ was taken to be the distance between the two layers in bilayer hBN and the distance between the metals in the TMD bilayers, and the volume of the bilayers was taken to be $\Omega = Ad$, where $A$ is the in-plane area of the unit cell. The cohesive energy per unit cell is taken to be $\V_0 = \V_{\text{bilayer}} - 2\V_{\text{mono}}$, where $\V_{\text{bilayer}}$ and $\V_{\text{mono}}$ are the total energies of the bilayer and monolayers, respectively. The polarizabilities of the bilayers, defined in Ref.~\cite{bennett2022electrically}, were not included in Eq.~\eqref{eq:V_tot_all}, but were calculated for completeness. The polarizability to zeroth order in $d$, $\a_0$, was obtained by fixing the relaxed geometry and applying strong positive and negative fields, which is required for the cancellation of errors in the polarization at zero field. The polarizability to first order in $d$, $\a_1$, was obtained by changing the layer separation by $\pm 1 \%$ with respect to $d_0$ and measuring the relative change in the polarizability. When an out-of-plane electric field was applied, a dipole correction \cite{dipole_correction_1,dipole_correction_2,dipole_correction_3,dipole_correction_4} was used in the vacuum region to prevent dipole-dipole interactions between periodic images. \\
Calculations were repeated to parameterize $\V_0$, $d_0$, $\a_0$ and $\a_1$, as well as the dipole moment $p_0$ and out-of-plane polarization $P_0=\frac{p_0}{\Omega}$, in configuration space. The data were then fitted to a Fourier expansion, taking advantage of the $\mathcal{C}_3$ rotation symmetry of the bilayers. For a general scalar field $\phi$, the following Fourier expansion was used:
\beq{eq:fourier_fitting}\resizebox{0.91\columnwidth}{!}{$
\begin{gathered}
\phi(x,y) = \phi_0 + \sum_{i=1}^{3} \phi_i^{\text{even}}f_i^{\text{even}}(x,y) + \phi_i^{\text{odd}}f_i^{\text{odd}}(x,y) \\[5pt]
f_1^{\text{even}}(x,y) = \cos{(2\pi x)} + \cos{(2\pi y)} + \cos{(2\pi (x+y))} \\
f_2^{\text{even}}(x,y) = \cos{(2\pi (x-y))} + \cos{(2\pi (2x + y))} + \cos{(2\pi (x+2y))} \\
f_3^{\text{even}}(x,y) = \cos{(4\pi x)} + \cos{(4\pi y))} + \cos{(4\pi (x+y))} \\[5pt]
f_1^{\text{odd}}(x,y) = \sin{(2\pi x)} + \sin{(2\pi y)} - \sin{(2\pi (x+y))} \\
f_2^{\text{odd}}(x,y) = \sin{(2\pi (y-x)} + \sin{(2\pi (2x+y))} - \sin{(2\pi (x+2y))} \\
f_3^{\text{odd}}(x,y) = \sin{(4\pi x)} + \sin{(4\pi y)} - \sin{(4\pi (x+y))}\\
\end{gathered}
$}
\end{equation}
where $\bvec{s} = x \bvec{a}_1 + y \bvec{a}_2$. The first-principles calculations were performed along the line $s=x=y$, which includes all of the high symmetry stacking configurations, greatly reducing the number of calculations needed. However, it should be noted that $f_2^{\text{odd}}(s,s)=0$ for every value of $s$, so in order to be more precise, a full 2D parameterization of configuration space should be performed when considering properties which have large odd components. Plots of the aforementioned quantities along the configuration space diagonal as well as the corresponding parameterization using Eq.~\eqref{eq:fourier_fitting} for the 11 bilayers considered can be found in the supplementary information (SI). Tables of the fitting parameters are also given.
Several electronic properties were also measured as a function of electric field in configuration space for one homo-bilayer (\chem{MoS_2}), and one hetero-bilayer (\chem{MoS2/MoSe_2}). The `local band gap' was determined from the electronic band structure at each value of $s$. The average charge density in the out-of-plane direction was obtained with {\sc c2x} \cite{c2x} using macroscopic averaging techniques \cite{junquera2003first,nanosmooth,stengel2011band,c2x}:
\beq{}
\rho_{\text{av}}(z) = \frac{1}{A}\int_A \rho(x,y,z) \dd{A}
\;,\end{equation}
on a grid of $10^{6}$ points. The charge transferred between the layers is given by
\beq{}
\Delta q = A\int_{d_0/2}^{(d_0+c)/2} \rho_{\text{av}}(z) \dd{z} - q
\;,\end{equation}
where $c$ is the out-of-plane lattice vector, and $q$ is the formal charge of the layer.
\subsection{Lattice Relaxation}
Following the methodology described in the Appendix, lattice relaxation calculations were performed using the parameterizations obtained from first-principles and the elastic moduli calculated in Ref.~\onlinecite{carr2018relaxation}. 3R-stacked \chem{MoS_2} and both stackings of \chem{MoS_2/MoSe_2} were considered, for twist angles ranging from $\th=1.0^{\circ}$ to $\th=0.1^{\circ}$ in steps of $0.1^{\circ}$, and electric fields ranging from $\Ef=-2$ V/\AA \ to $\Ef = +2$ V/\AA \ in steps of 0.5 V/\AA. The following truncations were applied to the Brillouin zone (BZ): five shells (10 $\bvec{k}$-vectors) for $\th \geq 0.5^{\circ}$, six shells (21 $\bvec{k}$-vectors) for $0.5^{\circ} > \th \geq 0.3^{\circ}$ and and seven shells (28 $\bvec{k}$-vectors) for $0.3^{\circ} >\th \geq 0.1^{\circ}$. The total energy was optimized with respect to the displacement field using the {\sc Optim} package in {\sc Julia}, using the analytic expression for the gradient of the free energy derived in the Appendix. The total energy was optimized using the limited memory BFGS algorithm (L-BFGS), with a tolerance of $1\times 10^{-5}$ eV. The elastic part of the energy and gradient are given exactly by Eqs.~\eqref{eq:V_elastic_exact} and \eqref{eq:G_elastic}, respectively, and the rest of the energy and gradient were obtained by numerical integration using the {\sc Cubature} package. The integrations could be parallelized, and the individual components of of the gradient could also be calculated in parallel, which could be employed in calculations involving a very large number of shells. In this study, since a large number of calculations were performed for different materials and different values of $\th$ and $\Ef$, separate calculations were executed in parallel, each using a single core. Calculations with the same value of $\th$ but different values of $\Ef$ were chained together by using the output of calculations at $\Ef=0$ as the starting point for calculations at $\Ef= \pm 0.5$ \ V/\AA \, and so on, reducing the number of iterations in each calculation considerably. The same procedure was also used to chain together calculations at zero field but at different twist angles, leaving additional shells (if any) empty for the first iteration.
\section{Results}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\columnwidth]{./pol.pdf}
\caption{Polarization along the configuration space diagonal for \textbf{(a)}: homo-bilayers (3R stacking), \textbf{(b)}: hetero-bilayers (3R stacking), and \textbf{(c)}: hetero-bilayers (2H stacking). For the hetero-bilayers, one example of each of the following is shown: same metal, different chalcogen (\chem{MoS_2/MoSe_2}), same chalcogen, different metal (\chem{MoS_2/WS_2}), different metal, different chalcogen (\chem{MoS_2/WSe_2}), with the rest available in the SI.}
\label{fig:pol}
\end{figure}
\begin{figure*}[p!]
\centering
\includegraphics[width=\linewidth]{./contour-main.pdf}
\caption{MPDs for \textbf{(a)},\textbf{(d)},\textbf{(g)}: 3R-stacked \chem{MoS_2}, \textbf{(b)},\textbf{(e)},\textbf{(h)}: 3R-stacked \chem{MoS_2/MoSe_2} and\textbf{(c)},\textbf{(f)},\textbf{(i)}: 2H-stacked \chem{MoS_2/MoSe_2}, all at a twist angle of $\th = 0.1^{\circ}$ and electric field strengths of $\Ef = -2,0,2$ V/\AA. \textbf{(j)}-\textbf{(l)}: total polarization in configuration space as a function of $\Ef$ for several twist angles for each of the three bilayers.}
\label{fig:pol-contour}
\end{figure*}
The out-of-plane polarization of the TMD bilayers along the configuration space diagonal is shown in Fig.~\ref{fig:pol}. All of the homo-bilayers have an odd polarizaiton along the diagonal, with the magnitude being stronger in the sulfur TMDs. For hetero-bilayers with the same chalcogens but different metals, the MPD is nearly odd, but the polarization at the AB ($s=\frac{1}{2}$) and BA ($s=\frac{2}{3}$) stacking configurations are no longer equal in magnitude. When the chalcogens in each layer are different, the MPD has a very different form. It is no longer centered about zero because the out-of-plane inversion symmetry is broken everywhere by having two different monolayers. Unlike the 2H-stacked homo-bilayers, the 2H-stacked hetero-bilayers also exhibit MPDs.
Lattice relaxation calculations were performed for one homo-bilayer (\chem{MoS_2}) and one hetero-bilayer (\chem{MoS_2/MoSe_2}) for a range of twist angles and electric fields. The results are summarized in Fig.~\ref{fig:pol-contour}, with additional plots available in the SI. The behavior of the homo-bilayers is already known \cite{bennett2022electrically}: for a 2H stacking, there is no MPD and the structure is not influenced by an applied field. For a 3R stacking, any finite field breaks the $\mathcal{C}_6$ rotation symmetry of the model, causing the domains which are aligned with the field to grow, and the domains which are anti-aligned to shrink. For small angles and strong fields, the domain structure changes from a sharp triangular structure to a sharp hexagonal structure. For the hetero-bilayers, the shape of the MPDs was found not to be strongly influenced by an applied field. Changes in the 2H-stacked bilayers were negligible, and the 3R-stacked bilayers changed noticeably only for small twist angles and strong electric fields. Although Fig.~\ref{fig:pol-contour} (b) shows that the relative sizes of the domains change in 3R-stacked \chem{MoS_2/MoSe_2} for $\th=0.1^{\circ}$ and $\Ef = -2$ V/\AA, both domains have the same orientation, and thus the sign of the total polarization does not change.
The total polarization of \chem{MoS_2} (3R) and \chem{MoS_2/MoSe_2} (both stackings) as a function of $\th$ and $\Ef$ is summarized in Figs.~\ref{fig:pol-contour} (j)-(l). The 3R-stacked homo-bilayers show a nonlinear response to the field at small twist angles. The 3R-stacked hetero-bilayers only show a noticeable response when a strong field is applied in the direction opposite to the orientation of the MPDs, and the 2H-stacked hetero-bilayers do not show a noticeable response at any of the values of $\th$ and $\Ef$ used in this study.
\begin{figure}[ht!]
\centering
\includegraphics[width=\columnwidth]{./gap-charge.pdf}
\caption{Charge transfer $\Delta q$ for 3R-stacked \chem{MoS_2} as a function of \textbf{(a)}: $s$ along the diagonal through configuration space for several fixed values of $\Ef$ and \textbf{(b)}: $\Ef$, for several fixed values of $s$. \textbf{(c)} and \textbf{(d)}: the local band gap for the same values of $s$ and $\Ef$ as in \textbf{(a)} and \textbf{(b)}, respectively. \textbf{(e)}: Critical field $\Ef_{\text{crit}}$ required to close the local band gap in configuration space for both stackings of \chem{MoS_2} and \chem{MoS_2/MoSe_2}.}
\label{fig:gap-charge}
\end{figure}
The interlayer charge transfer in 3R-stacked \chem{MoS_2} along the configuration space diagonal is shown in Fig.~\ref{fig:gap-charge} (a) for several values of $\Ef$. At zero field, we can see that the interlayer charge transfer exactly matches the shape of the MPD. For stronger fields, the charge transfer resembles the MPDs of the 3R-stacked hetero-bilayers with different chalcogens. The charge transfer as a function of field strength for the AA ($s=0$), AB, SP ($s=\frac{1}{2}$) and BA stacking configurations is shown in Fig.~\ref{fig:gap-charge} (b). It is continuous and linear for all except the AA stacking, which experiences a change in slope at $\Ef \approx 1$ \ V/\AA. These results are contrary to similar measurements in Ref.~\onlinecite{bennett2022electrically}, where the charge transfer was estimated from the M{\"u}lliken charges instead of the macroscopic charge density. No transfer was observed below $\Ef \approx 0.27$ V/\AA \, above which the charge transfer grew linearly with the field strength. The origin of the discontinuity was unknown, and because it was not observed here, may be an artifact of the previous calculations, since the M{\"u}lliken charges are not a well-defined quantity in a periodic solid.
Fig.~\ref{fig:gap-charge} (c) shows the local band gap in configuration space for different field strengths. It has been suggested that the excitonic properties of moir\'e superlattices may be tuned with an electric field, which would couple to the interlayer exciton dipole \cite{yu2017moire}. However, additional effects from the applied field, such as electrically tunable lattice relaxation, and the decreasing, and possibly closing of the local gap, have not been considered. Fig.~\ref{fig:gap-charge} (c) shows that the local band gap changes significantly, and non-uniformly due to the MPD, which should be accounted for when considering the possibility of tuning the excitonic properties of moir\'e superlattices. Fig.~\ref{fig:gap-charge} (d) shows the local gap as a function of electric field strength for the high symmetry stacking configurations. The AA configuration decreases most quickly and goes to zero at $\Ef \approx 1.5$\ V/\AA, which may explain the change in slope in the interlayer charge transfer. For the other configurations, the gap closes for $\Ef > 2$ \ V/\AA. Fig.~\ref{fig:gap-charge} (e) shows the critical field required to close the gap in configuration space for both stackings of \chem{MoS_2} and \chem{MoS_2/MoSe_2}. Although the local band gap measured in configuration space is in general not a physically meaningful quantity, Figs.~\ref{fig:gap-charge} (c) - (e) may provide some insight into the possibility of tuning the excitonic properties of moir\'e heterostructures at small twist angles with an applied field. Because electric fields are typically applied to real samples locally via a biased tip, knowledge of the nonuniform dielectric properties of moir\'e superlattices as a result of MPDs may provide some helpful insight in understanding experimental measurements.
\section{Discussion and Conclusions}
MPDs were found in all bilayers considered in this study. They are determined by the local stacking arrangements, and thus should be considered a fundamental property of \textit{all} moir\'e heterostructures. Homo-bilayers exhibit MPDs which average to zero due to the local inversion symmetry of the AA and SP stacking regions. The hetero-bilayers exhibit MPDs which do not average to zero because the inversion symmetry is broken everywhere. In general, the magnitude and shape of the MPDs appears to be dominated by the chalcogens and is not strongly influenced by the transition metals. While the total polarization of the homo-bilayers (3R stacking) exhibits a nonlinear response to an applied field, the total polarization of the hetero-bilayers exhibits a negligible response in nearly all cases. The weak response to the field in the hetero-bilayers with different chalcogens is understandable: in the 3R-stacked homo-bilayers there is a $\mathcal{C}_6$ rotation symmetry at zero field. Any finite electric field breaks this symmetry, leading to the uneven relaxation of the AB and BA domains. In the hetero-bilayers however, this symmetry is already broken by having two different monolayers, and thus there is already a difference in energy between the AB and BA domains. Applying a field does not lead to much additional splitting, and lattice relaxation is not significantly altered. All MPDs are sensitive to the moir\'e period, however.
While the behavior of the MPDs at angles which are small deviations from 2H and 3R stackings has been illustrated, the general behavior at all angles is far from clear. Although the configuration space mapping is valid for arbitrary twist angles, in principle we can only measure MPDs by performing a Taylor expansion of Eq.~\eqref{eq:slide_map} about a commensurate reference structure and parameterize systems at twist angles which are small deviations from the reference structure by sliding one layer over the other in first-principles calculations. It is not clear if MPDs can be measured at arbitrary twist angles; this would require the measurement of the local polarization in a moir\'e superlattice without relying on the configuration space mapping, which, even if possible, would be extremely computationally expensive.
However, it is an interesting theoretical question: consider a bilayer at an arbitrary twist angle $\theta$, which ranges from a reference structure at zero, for which the unit cell is commensurate, up to the first angle which leaves the system invariant. For hBN and the TMD bilayers, if we take the 3R stacking to be the reference structure at $\th=0$, then the 2H stacking is realized at $\th = \frac{\pi}{6}$, and the 3R stacking is realized again at $\th = \frac{\pi}{3}$. This is sketched in Fig.~\ref{fig:pol-theta}. The vertical axis represents some measure of polarity, using the root mean squared polarization of a moir\'e period as an example. For $\th = 0,\frac{\pi}{6},\frac{\pi}{3}$ the bilayer is commensurate and hence non-polar. In homo-bilayers, for small rational twist angles, we know that there is a MPD structure about the 3R stacking, and zero polarization about the 2H stacking. If we consider the twist angle as a tunable parameter, there should be a transition from a polar state to non-polar state between the 3R and 2H stackings. However without the knowledge of the MPD structure at arbitrary twist angles, it is impossible to locate or determine the order of this transition.
\begin{figure}[ht!]
\centering
\includegraphics[width=\columnwidth]{./pol-theta.pdf}
\caption{Sketch of the polarity of a moir\'e homo-bilayer as a function of twist angle. The horizontal axis denotes the twist angle $\theta$ in radians, with the 3R commensurate stacking at $\theta=0$ and $\theta=\frac{\pi}{3}$, and the 2H stacking at $\theta=\frac{\pi}{6}$. The vertical axis denotes the root mean squared polarization, $P_{\text{RMS}} = \left(\frac{1}{A}\int \bvec{P}(\bvec{x})^2\dd\bvec{x} \right)^{\frac{1}{2}}$, which, without lattice relaxation, is constant about the high symmetry stackings. Exactly at the $\th = 0,\frac{\pi}{6},\frac{\pi}{3}$, the layers are perfectly aligned, and therefore non-polar, as indicated by the red dots. Using the configuration space mapping, we can measure $P_{\text{RMS}}$ for small rational twists with respect to these configurations. The behavior at larger deviations from these angles is unknown.}
\label{fig:pol-theta}
\end{figure}
\subsection{Are moir\'e heterostructures really ferroelectric?}
\subsubsection{Lattice relaxation}
To reiterate the conclusions of Ref.~\onlinecite{bennett2022electrically}, the total response of the spontaneous polarization arising from tunable lattice relaxation under an applied field does not qualify as ferroelectricity, since there is no remnant, switchable polarization at zero field. Despite this, ferroelectric behavior has been observed experimentally, and the reason for this discrepancy is unclear. The obvious difference between theory and experiment is that theoretical descriptions are of an ideal sample, whereas experimental samples are imperfect. Defects and strains induced by finite boundaries may lead to imperfect domains, and a remnant polarization at zero field. Even so, this remnant polarization would not be switchable via lattice relaxation alone, because the system would relax back to its original configuration at zero field. Perhaps electromechanical couplings such as piezoelectricity \cite{mcgilly2020visualization} and flexoelectricity \cite{mcgilly2020visualization,bennett2021flexoelectric,springolo2021direct} could lead to irreversible structural changes as a field is applied and removed. It is easier to see that this is also true of the hetero-bilayers, in which the MPDs exhibit a negligible change in shape in response to an applied field. Even for fields strong enough to change the relative sizes of the domains, both types of domains have the same orientation but different magnitude, so it would be impossible to change the orientation of the total spontaneous polarization by applying a strong field in the opposite direction.
\subsubsection{van der Waals sliding}
A first-order polarization switching has been observed in twisted hBN \cite{stern2020interfacial,yasuda2021stacking}. However, there are a number of conceptual problems which much be addressed before it can be determined if the observed behavior actually corresponds to a ferroelectric response. The images of the domains were made by scanning biased tips over the samples. It is thought that applying a strong electric field leads to a sliding by one third of a unit cell, switching the stacking configuration from e.g.~AB to BA, and aligning the local polarization with the field. Although claimed to be ferroelectric in this study and other theoretical studies \cite{wu2021sliding,zhong2021sliding}, whether or not this mechanism truly corresponds to ferroelectricity was not discussed. The field is applied locally, and it is not clear whether the sliding occurs locally, beneath the tip, and the atoms relax back to their initial configuration due to a large elastic energy penalty once the tip moves on, or whether a global sliding of one layer over the other occurs. The former corresponds to a local first-order switching of polarization, but not a ferroelectric response. The latter corresponds to a first-order switching of the entire domain structure, which in homo-bilayers would result in an inversion of the domains: $\bvec{P}(\bvec{r})\to -\bvec{P}(\bvec{r})$. For an ideal sample, this would still not correspond to ferroelectricity, since the polarization at zero field would always be zero, but for an imperfect sample, it is possible that the total polarization at zero field could be switched after sliding, and thus may qualify for ferroelectricity. In Ref.~\onlinecite{stern2020interfacial}, the orientation of the domains were measured at $\Ef = \pm 3$ V/\AA and a switching of the polarization was observed, although the orientation was not measured again after the fields were removed, so it is impossible to determine whether the orientation of the MPD at zero field was inverted or not. Thus, the claims of ferroelectricity were made without strong evidence. In Ref.~\onlinecite{yasuda2021stacking}, a first-order polarization switching was observed in untwisted hBN, but not in twisted hBN. Since neither study provides strong evidence of global van der Waals sliding in a twisted system, it is impossible to conclude whether moir\'e heterostructures can be ferroelectric, or whether MPDs can be considered ferroelectric domains via this mechanism.
Using Eq.~\eqref{eq:V_tot_all}, a rough estimate of when a first-order local sliding beneath a biased tip would occur can be obtained. Taking a 3R-stacked homo-bilayer for simplicity, we would expect a first order switching from AB $\to$ BA when $\Ef$ and $P_{\text{AB}}$ are anti-aligned, and when the energy gained from aligning the polarization with the field is greater than the cost of traversing the energy barrier, plus the elastic energy penalty associated with sliding:
\beq{eq:first-order-1}
\Ef \left( p_{\text{BA}} - p_{\text{AB}}\right) > \Delta\Vstack + \Delta\mathcal{V}_{\text{elastic}}
\;.\end{equation}
A sliding from AB to BA is achieved by a displacement $\bvec{u}(\bvec{x}_{\text{AB}}) = \begin{bmatrix} 1/3 \\ 1/3 \end{bmatrix}$, where $\bvec{x}$ is in units of $\bvec{a}_1$ and $\bvec{a}_2$ (see Appendix). From Eq.~\eqref{eq:V_elastic}, the elastic cost of such a displacement is $\Delta\mathcal{V}_{\text{elastic}} = \frac{4}{27}\mu\th^2$, and the energy barrier is taken to be the one at zero field for simplicity: $\Delta\Vstack = \V_{\text{SP}} - \V_{\text{AB}}$. Under these approximations, Eq.~\eqref{eq:first-order-1} predicts a phase boundary as a function of $\th$ for when a first-order sliding would occur in a given material:
\beq{eq:first-order-2}
\Ef_{\text{slide}} = \frac{1}{2\left| p_{\text{AB}}\right|} \left( \frac{4}{27}\mu\th^2 + \Delta\Vstack\right)
\;.\end{equation}
This is sketched in Fig.~\ref{fig:first-order-phase} (a). It should be noted that Eq.~\eqref{eq:first-order-1} is a very rough estimate, because it assumes that that the energy barrier is not reduced by the field, and only the point directly beneath the biased tip is displaced. In reality, a neighborhood around the tip would be displaced, although the precise behavior is difficult to estimate without a detailed study, which is beyond the scope of this paper.
We can also estimate the phase boundary for the materials considered in this paper. Fig.~\ref{fig:first-order-phase} (b) shows the parameterization of Eq.~\eqref{eq:first-order-2} for \chem{hBN} and \chem{MoS_2}. We can see that a first-order transition is certainly possible for hBN, although the field strengths required to achieve a sliding in \chem{MoS_2} are unrealistic. Even if the elastic energy penalty is neglected, the cost of traversing the energy barrier would require a field of $\Ef_{\text{slide}} > 7$ V/\AA.
\begin{figure}[h!]
\centering
\includegraphics[width=\columnwidth]{./first-order-phase.pdf}
\caption{\textbf{(a)}: Sketch of the phase boundary as a function of $\Ef$ and $\th$, separating a local first-order switching of the polarization via van der Waals sliding, and a second order change in the total polarization via lattice relaxation. \textbf{(b)}: parameterization of Eq.~\eqref{eq:first-order-2} for 3R-stacked hBN and \chem{MoS_2} using the values of $p_{\text{AB}}$ and $\Delta\Vstack$ calculated in this paper and the shear moduli from Refs.~\onlinecite{lin2019effective} and \onlinecite{carr2018relaxation}, respectively. }
\label{fig:first-order-phase}
\end{figure}
\subsection{Comparison with conventional ferroelectrics}
\begin{table*}[ht!]
\renewcommand*{\arraystretch}{2}
\setlength{\tabcolsep}{10pt}
\begin{center}
\begin{tabular}{| c | c | c |}
\hline\hline
\textbf{System} & \textbf{Conventional ferroelectrics} & \textbf{Moir\'e heterostructures} \\ \hline
Prototypical materials & \chem{ABO_3} oxide perovskites (\chem{BaTiO_3}, \chem{PbTiO_3}, \ldots) & hBN, TMDs (\chem{MoS_2}, \chem{WS_2}, \ldots) \\ \hline
Mechanism for polarization & Soft mode lattice instabilities & Local inversion symmetry breaking \\ \hline
Origin of polarization & Change in hybridization of B and O states & Interlayer charge transfer\\ \hline
Origin of domains & Screen depolarizing field & Local inversion symmetry breaking \\ \hline \hline
\end{tabular}
\end{center}
\caption{Comparison of the polar properties of conventional ferroelectrics (oxide perovskites) and moir\'e heterostructures, assuming ideal structures.}
\label{table:ferro}
\end{table*}
The formation of ferroelectric domains, in oxide perovskite thin films for example, is a well-known problem which has been understood for many years. A free-standing ferroelectric thin film can form a polydomain structure in order to screen the depolarizing field arising from the polar discontinuities at the surfaces, and the behavior of the domains as a function of film thickness and perpendicular applied field can be studied in the limit of infinitely thin domain walls \cite{springer,vacuum_1,bennett2020electrostatics}. When the width of the domains $w$ is less than the thickness $d$ of the film, the domains follow Kittel's law \cite{springer,vacuum_1}: $w = \sqrt{l_{\text{k}}d}$, where $l_{\text{k}}$ Kittel's length, the relevant length scale for the domains. In the limit of zero thickness, the width of the domains diverge, and a monodomain or paraelectric phase becomes more favourable \cite{bennett2020electrostatics}. Of course, in this limit, the approximation of infinitely thin walls is not appropriate. Ferroelectric domains with domain walls of finite width can be considered by generalizing to a Ginzburg-Landau theory \cite{kretschmer1979surface,chandra2007landau,luk2009universal}. At zero temperature, the free energy of a ferroelectric thin film can be given by
\beq{eq:GL_ferro}
V_{\text{tot}}^{\text{GL}} = \int \left( \lambda^2 \left( \nabla\bvec{P}(\bvec{r})\right)^2 + \frac{1}{4}\bvec{P}(\bvec{r})^4 - \frac{1}{2}\bvec{P}(\bvec{r})^2 - \Ef\cdot \bvec{P}(\bvec{r}) \right) \dd{\bvec{r}}
\;,\end{equation}
where $\lambda$ is the correlation length and $\bvec{P}$ is in units of the spontaneous polarization $\bvec{P}_0$ of the monodomain phase ($\lambda\to\infty$). Minimizing Eq.~\eqref{eq:GL_ferro} and solving the resulting equations, we find that the general solution for $\bvec{P}(\bvec{r})$ is in terms of elliptic functions \cite{luk2009universal}. At zero field, Eq.~\eqref{eq:GL_ferro} is clearly always invariant if we switch the orientation of the domains: $\bvec{P}(\bvec{r}) \to -\bvec{P}(\bvec{r})$, even in the monodomain and polydomain limits. This is not true of MPDs, although direct comparisons are difficult because in Eq.~\eqref{eq:V_tot_main} the polarization is not an order parameter, but rather it is determined by the structure.
It could be argued that the sign of the polarization in large regions of the moir\'e superlattice does change after relaxation when the domain walls move considerably. However, this argument relies on a poor definition of a polar domain. A domain can be defined as the area enclosed by a boundary on which an order parameter is zero. Although better definitions could be devised which account for the width of the domain walls, this definition can describe both MPDs and conventional ferroelectric domains. Using this definition, it is clear that, although the domain walls move and the area of the domains change, the sign of the polarization inside a domain never changes. Thus, it is clear that the MPDs in 3R-stacked homo-bilayers are \textit{not} ferroelectric in a static homo-bilayer or via lattice relaxation, as claimed in Refs.~\onlinecite{ferreira2021weak,enaldiev2021piezoelectric,enaldiev2021scalable}. An inversion of the signs of the polarization in the domains may be possible in homo-bilayers via van der Waals sliding. Currently, this is the only known mechanism by which the MPDs could be classed as ferroelectric domains. However as mentioned previously, this has not yet been experimentally observed.
It has also been suggested that MPDs are antiferroelectric \cite{shen2019emergent,li2021extremely}. Although in 3R-stacked homo-bilayers the domains have opposite orientation and the polarization averages to zero, the classical definition of antiferroelectricity refers to anti-aligned adjacent \textit{dipoles}, not adjacent \textit{domains}, analogous to an antiferromagnet. A more general definition of antiferroelectricity is a material which can be made ferroelectric with an applied field \cite{rabe2013antiferroelectricity}, resulting in a double hysteresis loop, which has not been experimentally observed for moir\'e superlattices. Using either definition, it is not appropriate to classify moir\'e heterostructures as antiferroelectric.
\subsection{Final remarks}
The recent experimental discoveries of ferroelectric behavior have prompted a change in thought about moir\'e heterostructures, and it is clear that they should be considered polar materials. Despite this, there have been misconceptions about what exactly corresponds to ferroelectric behavior. The domains have incorrectly been called ferroelectric, and it is not clear whether electrically tunable lattice relaxation can lead to ferroelectricity without the sample being imperfect, or if van der Waals sliding in a twisted bilayer is possible. This highlights the importance of establishing a general theoretical understanding of moir\'e heterostructures as polar materials. In particular, great care should be taken when comparing moir\'e heterostructures to conventional ferroelectric materials, because the origin of polarization and domain formation in each case, summarized in Table \ref{table:ferro}, is completely different.
The are still many unanswered questions about the polar properties of moir\'e heterostructures. Electrically tunable lattice relaxation is generally well understood for homo-bilayers \cite{bennett2022electrically}, and has been described in detail for general moir\'e heterostructures in this paper. However, to my knowledge, a theoretical study of polarization switching via van der Waals sliding has yet to be done. Additionally, the role of electromechanical coupling in both electrically tunable lattice relaxation and van der Waals sliding is not well-known. Finally, a complete picture of MPDs in moir\'e heterostructures at general twist angles is missing. Addressing these questions will lead to a better understanding of the polar properties of moir\'e heterostructures, and may lead to practical advancements in the field of twistronics.
\section*{Acknowledgements}
I would like to thank B.~Remez, E.~Artacho, M.~Stengel, E.~Bousquet and P.~Ghosez for helpful discussions. I would like to acknowledge funding from the EPSRC Centre for Doctoral Training in Computational Methods for Materials Science under grant number EP/L015552/1, St.~John's College (University of Cambridge), and the University of Li{\'e}ge under special funds for research (IPD-STEMA fellowship programme).\\
\section*{Appendix: Description of lattice relaxation calculations}
A complete overview of how lattice relaxation calculations are practically performed is presented here, i.e.~how to numerically minimize Eq.~\eqref{eq:V_tot_main} in an efficient manner. There are several viable approaches to perform lattice relaxation calculations: the first and perhaps most obvious option is to minimize the free energy of the model using variational methods and solve the resulting differential equations. However, solving nonlinear partial differential equations (PDEs) with periodic boundary conditions can be difficult, and in configuration space, increasing the moir\'e period (decreasing the twist angle) makes the PDEs more nonlinear and volatile, requiring the use of very fine grids. The second option is to make use of the periodic boundary conditions in configuration space by performing Fourier expansions of the displacement field and the stacking energy, then minimize the total energy with respect to the Fourier components of the strain field. However, a closed form solution cannot be obtained because the Fourier components of the stacking energy depend on the Fourier components of the displacement field: $\Vstack(\bvec{s} + \bvec{u}(\bvec{s}))$. One option is to update the Fourier components of the displacement field and total energy self-consistently until the total energy is minimized \cite{nam2017lattice}. Another option is to directly minimize the free energy with respect to Fourier components of the displacement field using optimization techniques \cite{carr2018relaxation,bennett2022electrically}, which is the method used in this paper.
\subsubsection{One-dimensional example}
To illustrate the method, consider a one-dimensional (1D) field theory, described by a free energy functional
\beq{eq:V_tot_1D}
V_{\text{tot}}[u] = \int_0^1 \mathcal{V}_{\text{tot}}(x,u(x),\partial_xu(x)) \dd x
\;,\end{equation}
where $\mathcal{V}_{\text{tot}}$ is nonlinear and periodic on the interval $[0,1]$. For example, if
\beq{}
\mathcal{V}_{\text{tot}}(x,u(x),\partial_xu(x)) = \frac{1}{2}B\left( \partial_x u\right)^2 + \V_0 \cos{(2\pi (x+u(x))}
\;,\end{equation}
then Eq.~\eqref{eq:V_tot_1D} would describe a Frenkel-Kontorova model, which has been used to model domain walls in moir\'e superlattices \cite{popov2011commensurate,lebedeva2016dislocations,lebedeva2019commensurate,bennett2022electrically}. We could minimize Eq.~\eqref{eq:V_tot_1D}: $\left( \partial_x\partial_{\partial_u} - \partial_u\right) V_{\text{tot}} = 0$, and solve the resulting ordinary differential equation (ODE) for $u(x)$, but this can become difficult if $\V$ is highly nonlinear, or for a higher number of dimensions. Instead, we can take advantage of the periodicity of the system by expanding the displacement field in a Fourier series:
\beq{eq:u_1D}
u(x) = \sum_k u_k e^{2\pi i k x}
\;,\end{equation}
after which, Eq.~\eqref{eq:V_tot_1D}, is no longer a functional of $u(x)$, but a function of $\{u_k\}$: $V_{\text{tot}}[u] \to V_{\text{tot}}(\{u_k\})$. Instead of minimizing Eq.~\eqref{eq:V_tot_1D} using variational methods and solving the resulting ODEs, we can simply minimize it by differentiating with respect to Fourier components $\{u_k\}$:
\beq{}
\bvec{G} \equiv \nabla_{u_k}V_{\text{tot}} = 0, \qquad \nabla_{u_k} \equiv \begin{bmatrix} \partial_{u_1} \\ \partial_{u_2} \\ \vdots \end{bmatrix}
\;.\end{equation}
The $k\textsuperscript{th}$ component of $\bvec{G}$ is
\beq{}
\partial_{u_k}V_{\text{tot}} = \int \partial_{u_k} \mathcal{V}_{\text{tot}} \dd x = \int \frac{\delta \mathcal{V}_{\text{tot}}}{\delta u} \left(\partial_{u_k} u\right) \dd x
\;,\end{equation}
where
\beq{}
\frac{\delta \mathcal{V}_{\text{tot}}}{\delta u} = \partial_x \frac{\partial \mathcal{V}_{\text{tot}}}{\partial(\partial u)} - \frac{\partial\mathcal{V}_{\text{tot}}}{\partial u} \equiv D
\end{equation}
is the functional derivative, and the partial derivative $\partial_{u_k}$ selects the $k\textsuperscript{th}$ Fourier basis function from $u$:
\beq{}
\partial_{u_k} u = e^{2\pi i k x}
\;.\end{equation}
Expanding the functional derivative in a Fourier series,
\beq{}
D = \sum_k D_k e^{2\pi i k x}
\;,\end{equation}
we get
\beq{}
\partial_{u_k}V_{\text{tot}} = \int D \ e^{2\pi i k x} \dd x = D_k
\;.\end{equation}
Thus the derivative of $V_{\text{tot}}$ with respect to the Fourier components $\{u_k\}$ is
\beq{}
\bvec{G} = \begin{bmatrix} D_1 \\ D_2 \\ \vdots \end{bmatrix}
\;,\end{equation}
where in principle there is an infinite number of components, but in practice they should be appropriately truncated to some number $N_{\text{max}}$ (all results should be well converged with respect to the number of components). We can now use root finding techniques to solve $\bvec{G} = 0$ for $\{u_k\}$. For example, using Newton's method, the $n\textsuperscript{th}$ approximation to the Fourier coefficients $\bvec{u}^n = \left[ u_1^n, u_2^n, \ldots , u_{N_{\text{max}}}^n\right]$ is
\beq{}
\bvec{u}^{n} = \bvec{u}^{n-1} - H^{-1}(\bvec{u}^{n-1})\bvec{G}(\bvec{u}^{n-1})
\end{equation}
where $H$ is the Hessian of $V_{\text{tot}}$. Although we have an analytic expression for $\bvec{G}$ for a given $\bvec{u}^n$, an analytic expression for $H$ cannot be obtained when $\mathcal{V}_{\text{tot}}$ is nonlinear. Thus in order to use Newton's method, $H$ would need to be calculated using finite differences, making the calculations more expensive. A better approach is to use a quasi-Newton method such as the Broyden–Fletcher–Goldfarb–Shanno algorithm (BFGS), where an approximation to the the inverse Hessian is used to choose the direction of the next iteration to update $\bvec{u}^n$.
\subsubsection{Lattice relaxation in moir\'e heterostructures}
The optimization techniques introduced in the previous section can be generalized to 2D and used to perform lattice relaxation calculations in moir\'e superlattices, or more generally any nonlinear field theory with a crystal symmetry. In configuration space, Eq.~\eqref{eq:V_tot_main} becomes
\beq{eq:V_tot_main_config}
\begin{split}
V_{\text{tot}} = \frac{1}{\As}\int_{\As}\bigg[\bigg. \mathcal{V}_{\text{elastic}}(\nabla\bvec{u}) +& \Vstack(\bvec{s}+\bvec{u}(\bvec{s})) \\
+& \Velec(\bvec{s}+\bvec{u}(\bvec{s}),\Ef)\bigg.\bigg] \dd\bvec{s}
\end{split}
\;.\end{equation}
The elastic energy is, explicitly \cite{carr2018relaxation,bennett2022electrically},
\beq{eq:V_elastic_cartesian}
\begin{split}
\mathcal{V}_{\text{elastic}} =& \frac{\th^2}{2} \bigg[\bigg. B\left( \partial_{s_x}u_{s_y} - \partial_{s_y}u_{s_x}\right)^2 + \\
&\mu \left( \left( \partial_{s_x}u_{s_y} + \partial_{s_y}u_{s_x} \right)^2
+ \left( \partial_{s_x}u_{s_x} - \partial_{s_x}u_{s_x}\right)^2 \right) \bigg.\bigg]
\end{split}
\end{equation}
where $B$ is the bulk modulus and $\mu$ is the shear modulus, both multiplied by the in-plane area and thus in units of energy per unit cell.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\columnwidth]{./BZ.pdf}
\caption{BZ containing the $\bvec{k}$-vectors over which the displacement field and total energy density are expanded. The BZ is divded into six sectors, with $S_{2n+1}$ (light blue) and $S_{2n}$ (light red) being related by $\mathcal{C}_3$ rotations. The expansion over the entire BZ is reduced to an expansion over the vectors in $S_1$, highlighted in red, using the $\mathcal{C}_3$ rotation symmetry. }
\label{fig:BZ}
\end{figure}
We write the displacement field as a Fourier series expansion,
\beq{}
\bvec{u}(\bvec{s}) = \sum_{\bvec{k}} \bvec{u}_{\bvec{k}} e^{i\bvec{k}\cdot\bvec{s}}
\;,\end{equation}
over the reciprocal lattice vectors $\bvec{k}$ in the Brillouin zone (BZ) of the configuration space unit cell, see Fig.~\ref{fig:BZ}. The number of independent $\bvec{k}$ vectors can be greatly reduced using crystal symmetries of the system, which for the bilayers considered in this paper is a $\mathcal{C}_3$ rotation symmetry. There is a mirror plane along the diagonal in configuration space which could be used to further reduce the number of independent $\bvec{k}$-vectors, but it is not used here. Scalar fields $\phi$ and vector fields $\bvec{u}$ transform as
\beq{}
\begin{split}
\phi_{\mathcal{C}_3\bvec{k}} & = \phi_{\bvec{k}} \\
\bvec{u}_{\mathcal{C}_3\bvec{k}} & = \mathcal{C}_3\bvec{u}_{\bvec{k}} \\
\end{split}
\;,\end{equation}
respectively. We split the sum over $\bvec{k}$ into 6 individual sectors $S_1,\ldots,S_6$, where the corresponding vectors in sectors $S_{2n}$ are equivalent via $\mathcal{C}_3$ rotations, as are the vectors in sectors $S_{2n+1}$, see Fig.~\ref{fig:BZ}. Thus, we can write
\beq{eq:u_full}
\begin{split}
\bvec{u}(\bvec{s}) &= \sum_{n=1}^{6}\sum_{\bvec{k}}^{S_n} \bvec{u}_{\bvec{k}} e^{i\bvec{k}\cdot\bvec{s}} \\
&= \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1} \mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} e^{i\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}} +
\sum_{n=1}^{3}\sum_{\bvec{k}}^{S_4} \mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} e^{i\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}}\\
&= \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1} \left( \mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} e^{i\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}} + \mathcal{C}_3^{n-1}\bvec{u}_{-\bvec{k}} e^{-i\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}} \right)\\
&= \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1} \mathcal{C}_3^{n-1}\Big(\Big. \left( \bvec{u}_{\bvec{k}} + \bvec{u}_{-\bvec{k}}\right) \cos{\left(\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}\right)} \\
& \qquad\qquad\quad + i\left( \bvec{u}_{\bvec{k}} - \bvec{u}_{-\bvec{k}}\right) \sin{\left(\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}\right)} \Big.\Big)
\end{split}
\end{equation}
Going from the first line to the second line, we use the $\mathcal{C}_3$ rotation symmetry to relate the Fourier components of equivalent $\bvec{k}$-vectors in sectors $S_{2n}$ and $S_{2n+1}$. The third line is arrived at by noting that $S_1$ and $S_4$ are related by $\bvec{k}\to -\bvec{k}$, although the Fourier components $u_{\bvec{k}}$ and $u_{-\bvec{k}}$ are independent unless there is an additional inversion symmetry. Finally, we rewrite the complex exponentials in terms of sine and cosine basis functions. In general, the displacement field has both even and odd terms, and it is convenient rewrite the Fourier components as
\beq{}
\begin{split}
\bvec{u}^{\text{even}}_{\bvec{k}} &= \bvec{u}_{\bvec{k}} + \bvec{u}_{-\bvec{k}} \\
\bvec{u}^{\text{odd}}_{\bvec{k}} &= i\left( \bvec{u}_{\bvec{k}} - \bvec{u}_{-\bvec{k}}\right) \\
\end{split}
\end{equation}
If there is also an inversion symmetry, giving the system an overall $\mathcal{C}_6$ rotation symmetry, then $\bvec{u}_{\bvec{k}} = - \bvec{u}_{-\bvec{k}}$, i.e.~the cosine terms vanish, and the displacement is purely odd:
\beq{eq:u_odd}
\bvec{u}(\bvec{s}) = \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1} \mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} \sin{\left(\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{s}\right)}
\;,\end{equation}
where we have dropped the `odd' superscript. This is the case for 3R-stacked homo-bilayers, although not for 2H-stacked homo-bilayers or any hetero-bilayers. In the interest of simplicity, we consider a system with $\mathcal{C}_6$ rotation symmetry and therefore a displacement field given by Eq.~\eqref{eq:u_odd}, for the rest of this section. The generalization to a system with $\mathcal{C}_3$ rotation symmetry, and hence a displacement field given by Eq.~\eqref{eq:u_full} is obtained following the same methodology since the even and odd terms are linearly independent, although the algebra is more involved.
It is also convenient to work in reduced coordinates, i.e.~in terms of the lattice vectors $\bvec{a}_1 = \begin{bmatrix}1 \\ 0\end{bmatrix}$ and $\bvec{a}_2 = \begin{bmatrix}1/2 \\ \sqrt{3}/2\end{bmatrix}$, achieved by the transformation:
\beq{}
\begin{gathered}
\bvec{s} = g \bvec{x} \leftrightarrow \bvec{x} = g^{-1}\bvec{x}\\
g = \begin{bmatrix}1 & 1/2\\ 0 & \sqrt{3}/2\end{bmatrix}
\end{gathered}
\;,\end{equation}
In reduced coordinates, the unit vectors are non-orthogonal, but the unit cell is a square, making numerical integrations more efficient. The displacement field transforms as ${\bvec{u}(\bvec{s}) \to g \bvec{u}(g\bvec{x})}$, and the Fourier components transform as $\bvec{u}_{\bvec{k}} \to g \bvec{u}_{\bvec{k}}$. The exponentials and trigonometric functions transform as
\beq{}
\begin{split}
\exp{\left( i \bvec{k}\cdot\bvec{s}\right)} &= \exp{\left( i \bvec{k}\cdot\left( g\cdot g^{-1}\right)\bvec{s}\right)} \\
& = \exp{\left( i \bvec{k}\cdot g\bvec{x}\right)} \\
& = \exp{\left( i g^T\bvec{k}\cdot \bvec{x}\right)}
\end{split}
\;.\end{equation}
Thus, Eq.~\eqref{eq:u_odd} becomes
\beq{}
\bvec{u}(\bvec{x}) = \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1} g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)}
\;.\end{equation}
Note that the factor of $g^{-1}$ cannot be absorbed into the Fourier components because the $\mathcal{C}_3$ rotation symmetry is not preserved after the transformation to reduced coordinates.
While $\Vstack$ and $\Velec$ are naturally obtained in reduced coordinates from first-principles calculations by using Eq.~\eqref{eq:fourier_fitting}, the elastic energy in Eq.~\eqref{eq:V_elastic} must be transformed from Cartesian to reduced coordinates:
\beq{eq:V_elastic}
\begin{split}
\mathcal{V}_{\text{elastic}} &= \frac{\th^2}{2} \bigg[\bigg. B\left( \partial_{x}u_{y} - \partial_{y}u_{x}\right)^2 \\
&+ \mu \left( \frac{4}{3}\left( \partial_{x}u_{y} + \partial_{y}u_{x} \right)^2 + \left( \partial_{x}u_{x} - \partial_{y}u_{y}\right)^2 \right) \bigg.\bigg]
\end{split}
\;.\end{equation}
Fortunately, we can write $\mathcal{V}_{\text{elastic}}$ exactly in terms of the Fourier components $\{\bvec{u}_{\bvec{k}}\}$. In order to do this, we need the derivatives of the displacement field:
\beq{eq:u_deriv}
\begin{split}
\partial_x \bvec{u} &= \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1}
\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_x\left(\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right) \cos{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)}\\
\partial_y \bvec{u} & = \sum_{n=1}^{3}\sum_{\bvec{k}}^{S_1}
\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_y\left(\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right) \cos{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)}\\
\end{split}
\;.\end{equation}
Inserting Eq.~\eqref{eq:u_deriv} into Eq.~\eqref{eq:V_elastic} we obtain, after some algebra,\vskip 0.1in
\begin{widetext}
\beq{eq:V_elastic_exact}
\begin{split}
\mathcal{V}_{\text{elastic}} = \frac{1}{4}\th^2 \sum_{\bvec{k}}\sum_{n=1}^3 & B\left( \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_x \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_y - \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_y \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_x \right)^2 \\
&+ \mu \left( \frac{4}{3}\left( \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_x \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_y + \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_y \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_x \right)^2 \right. \\
&\left.+ \left( \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_x \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_x - \left( g^T\mathcal{C}_3^{n-1}\bvec{k}\right)_y \left( g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}}\right)_y \right)^2 \right)
\end{split}
\;,\end{equation}
\end{widetext}
where we use the orthogonality relations,
\beq{}
\begin{split}
\int \sin{(g^T \bvec{k}\cdot\bvec{x})}\sin{(g^T\bvec{k}'\cdot\bvec{x})}\dd{\bvec{x}} & = \frac{1}{2}\d_{\bvec{k},\bvec{k}'} \\
\int \cos{(g^T \bvec{k}\cdot\bvec{x})}\cos{(g^T\bvec{k}'\cdot\bvec{x})}\dd{\bvec{x}} & = \frac{1}{2}\d_{\bvec{k},\bvec{k}'}
\end{split}
\;,\end{equation}
to perform the integration. When the $\mathcal{C}_6$ rotation symmetry is reduced to $\mathcal{C}_3$, and the displacement field is given by Eq.~\eqref{eq:u_full}, Eqs.~\eqref{eq:u_deriv} and \eqref{eq:V_elastic_exact} will contain equivalent contributions from $\{\bvec{u}_{\bvec{k}}^{\text{even}}\}$.
Eq.~\eqref{eq:V_elastic_exact} is an exact expression for the elastic energy, and doesn't need to be calculated numerically, which results in a considerable reduction in computational effort. Unfortunately, the stacking energy cannot be obtained analytically since, as mentioned previously, the Fourier components of $\Vstack$ are functions of $\{\bvec{u}_{\bvec{k}}\}$.
Now we must obtain an analytic expression for the gradient as we did in the 1D example. First, we truncate the expansion so that a suitable number of shells are included, containing $N_{\text{max}}$ vectors in total. For example, in Fig.~\ref{fig:BZ}, the first 5 shells are shown, containing $N_{\text{max}} = 15$ vectors in $S_1$. The gradient is then a $2\times N_{\text{max}}$ array,
\beq{eq:G_initial}
\begin{split}
G &\equiv \nabla_{\bvec{u}_{\bvec{k}}} V_{\text{tot}} = \nabla_{\bvec{u}_{\bvec{k}}}\int \mathcal{V}_{\text{tot}} \ \dd\bvec{x} = \int \frac{\delta\mathcal{V}_{\text{tot}}}{\delta\bvec{u}} \nabla_{\bvec{u}_{\bvec{k}}} \bvec{u}\ \dd\bvec{x} \\[10pt]
&= \int \begin{bmatrix} D_x\partial_{u_{x,1}} u_x & D_y\partial_{u_{y,1}} u_y\\ \vdots & \vdots \\ D_x\partial_{u_{x,Nmax}} u_x & D_y\partial_{u_{y,Nmax}} u_y \end{bmatrix} \dd{\bvec{x}}
\end{split}
\;,\end{equation}
where the components of $\bvec{D}$ are the functional derivatives of $\mathcal{V}_{\text{tot}}$ with respect to the components of $\bvec{u}$:
\beq{}
\begin{split}
\bvec{D} &= \begin{bmatrix} D_x \\ D_y\end{bmatrix}\\
D_x & \equiv \frac{\delta \mathcal{V}_{\text{tot}}}{\delta u_x}\\
D_y & \equiv \frac{\delta \mathcal{V}_{\text{tot}}}{\delta u_y}
\end{split}
\;.\end{equation}
Eq.~\eqref{eq:G_initial} is general, and we would like to write the gradient in terms of the reduced set of $\bvec{k}$-vectors in $S_1$. First, note that
\beq{}
\begin{split}
\mathcal{C}_3\partial_{u_{k_x}} \bvec{u}_{\bvec{k}} &= \mathcal{C}_3\begin{bmatrix} 1 \\ 0\end{bmatrix} = \begin{bmatrix} -1/2 \\ \sqrt{3}/2\end{bmatrix} \\
\partial_{u_{k_x}} \mathcal{C}_3\bvec{u}_{\bvec{k}} &= \partial_{u_{k_x}}\frac{1}{2}\begin{bmatrix} -u_{k_x} -\sqrt{3}u_{k_x} \\ \sqrt{3}u_{k_x} - u_{k_y}\end{bmatrix} = \begin{bmatrix} -1/2 \\ \sqrt{3}/2\end{bmatrix}
\end{split}
\end{equation}
i.e.~the derivative commutes with the $\mathcal{C}_3$ rotation operator (as well as the metric tensor). Now, the derivatives of $\bvec{u}$ with respect to the basis functions are
\beq{eq:f_k_x}
\begin{split}
\partial_{u_{k_x}}\bvec{u} & = \sum_{n=1}^3\sum_{\bvec{k}}^{S_1} \partial_{u_{k_x}}g^{-1}\mathcal{C}_3^{n-1}\bvec{u}_{\bvec{k}} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)}\\
& = \sum_{n=1}^3\sum_{\bvec{k}}^{S_1} g^{-1}\mathcal{C}_3^{n-1}\partial_{u_{k_x}}\bvec{u}_{\bvec{k}} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)}\\
& = \sum_{n=1}^3\sum_{\bvec{k}}^{S_1} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)} g^{-1}\mathcal{C}_3^{n-1}\hat{\bvec{x}} \equiv \bvec{f}_{\bvec{k}}^x(\bvec{x})
\end{split}
\;,\end{equation}
and similarly for $y$:
\beq{eq:f_k_y}
\partial_{u_{k_y}}\bvec{u} = \sum_{n=1}^3\sum_{\bvec{k}}^{S_1} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)} g^{-1}\mathcal{C}_3^{n-1}\hat{\bvec{y}} \equiv \bvec{f}_{\bvec{k}}^y(\bvec{x})
\;,\end{equation}
where $\hat{\bvec{x}} = \begin{bmatrix} 1 \\ 0 \end{bmatrix}$ and $\hat{\bvec{y}} = \begin{bmatrix} 0 \\ 1 \end{bmatrix}$. Re-writing the gradient in terms of Eqs.~\eqref{eq:f_k_x} and \eqref{eq:f_k_y}, we get
\beq{}
\begin{split}
\partial_{u_{k_x}}V_{\text{tot}} &= \partial_{u_{k_x}}\int \mathcal{V}_{\text{tot}}(x+u_x, y+u_y)\dd\bvec{x}\\
& = \int \left(\frac{\delta\mathcal{V}_{\text{tot}}}{\delta u_x}\partial_{u_{k_x}}u_x + \frac{\delta \mathcal{V}_{\text{tot}}}{\delta u_y}\partial_{u_{k_x}}u_y\right)\dd\bvec{x} \\
& = \int \bvec{D}\cdot \bvec{f}_{\bvec{k}}^x(\bvec{x})\dd \bvec{x}
\end{split}
\end{equation}
for the $x$ component, and
\beq{}
\partial_{u_{k_y}}V_{\text{tot}} = \int \bvec{D}\cdot \bvec{f}_{\bvec{k}}^y(\bvec{x})\dd \bvec{x}
\end{equation}
for the $y$ component. Thus, the gradient becomes
\beq{}
G = \int \begin{bmatrix} \bvec{D}\cdot \bvec{f}^x_{\bvec{k}_1} & \bvec{D}\cdot \bvec{f}^y_{\bvec{k}_1}\\ \vdots & \vdots \\ \bvec{D}\cdot \bvec{f}^x_{\bvec{k}_{N_{\text{max}}}} & \bvec{D}\cdot \bvec{f}^y_{\bvec{k}_{N_{\text{max}}}}\end{bmatrix}\dd\bvec{x}
\end{equation}
The functional derivatives are, explicitly,
\beq{}\resizebox{\columnwidth}{!}{$
\bvec{D} =
\begin{bmatrix}
\partial_x\mathcal{V}_{\text{stack}} - \frac{1}{3}\theta^2 \lb3\mu\partial_x^2u_x -(3B-\mu)\partial_x\partial_yu_y + (3B+4\mu)\partial_y^2u_x\right)\\[10pt]
\partial_y\mathcal{V}_{\text{stack}} - \frac{1}{3}\theta^2 \lb3\mu\partial_y^2u_y -(3B-\mu)\partial_x\partial_yu_x + (3B+4\mu)\partial_x^2u_y\right)\\
\end{bmatrix}
$}
\end{equation}
which can be verified using e.g.~the {\sc VariationalMethods} package in {\sc Mathematica}.
The first term in $\bvec{D}\cdot \bvec{f}^j_{\bvec{k}}$ is the projection of the derivtatives of $\mathcal{\Vstack}$ onto the basis function $\bvec{f}^j_{\bvec{k}}$. These must be obtained with numerical integration, which may be expensive. Fortunately, rest of the terms, i.e.~the elastic energy, is exact and can be calculated for free for any $\bvec{u}_{\bvec{k}}$. Writing out the elastic part of the functional derivative, and defining\vskip 0.1in
\begin{widetext}
\beq{}
\begin{split}
\bvec{\bar{D}}_{\text{el},\bvec{k}} &\equiv \sum_{n=1}^3\sum_{\bvec{k}}^{S_1} \int \bvec{D} \sin{\left( g^T\mathcal{C}_3^{n-1}\bvec{k}\cdot\bvec{x}\right)} \dd\bvec{x} \\
& =-\frac{1}{3}\theta^2\sum_{n=1}^3 \begin{bmatrix} 3\mu \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_x^2\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_x - (3B-\mu) \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_x\left( \mathcal{C}_3^{n-1}\bvec{k}\right)_y\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_y + (3B+4\mu) \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_y^2\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_x\\[10pt]
3\mu \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_y^2\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_y - (3B-\mu) \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_x\left( \mathcal{C}_3^{n-1}\bvec{k}\right)_y\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_x + (3B+4\mu) \left( \mathcal{C}_3^{n-1}\bvec{k}\right)_x^2\left( \mathcal{C}_3^{n-1}\bvec{u}\right)_y \end{bmatrix}
\end{split}
\end{equation}
\end{widetext}
the elastic part of the gradient is, after some algebra,
\beq{eq:G_elastic}\resizebox{\columnwidth}{!}{$
G_{\text{el}} = \frac{1}{2}\sum_{n=1}^3 \begin{bmatrix} \bvec{\bar{D}}_{\text{el},\bvec{k}_1} \cdot \left( g^{-1} \mathcal{C}_3^{n-1}\cdot\hat{\bvec{x}}\right) & \bvec{\bar{D}}_{\text{el},\bvec{k}_1} \cdot \left( g^{-1}\mathcal{C}_3^{n-1}\cdot\hat{\bvec{y}}\right) \\
\vdots & \vdots \\
\bvec{\bar{D}}_{\text{el},\bvec{k}_{N_{\text{max}}}} \cdot \left( g^{-1}\mathcal{C}_3^{n-1}\cdot\hat{\bvec{x}}\right) & \bvec{\bar{D}}_{\text{el},\bvec{k}_{N_{\text{max}}}} \cdot \left( g^{-1}\mathcal{C}_3^{n-1}\cdot\hat{\bvec{y}}\right)
\end{bmatrix}
$}
\end{equation}
which can be verified by calculating the gradient of Eq.~\eqref{eq:V_elastic} numerically using finite difference techniques.
\clearpage
| 2cc8d3b183a76d4400ae2770f957bc6369890faa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Many complex and significant optimization problems (such as all of Karp's 21 NP-complete problems~\cite{L2014}, the~travelling salesman problem~\cite{KT1985}, the~protein folding problem~\cite{BW1987}, financial portfolio management~\cite{quantumfinance}) can be mapped into the problem of finding the~ground state of the~Ising spin-glass. Sophisticated and promising new methods for solving Ising instances rely on quantum computation, particularly quantum annealing.
Quantum annealing is a form of quantum computing particularly well-tailored for optimization~\cite{KN1998,SMTC2002}. It is closely related to adiabatic quantum computation~\cite{M2014}, a paradigm of universal quantum computation which relies on the~adiabatic theorem~\cite{K1950} to perform calculations. It is equivalent (up to polynomial overhead) to the~better-known gate model of quantum computation~\cite{M2014}. Nevertheless, commercially available quantum annealers (i.e., D-Wave) are prone to various errors, and their ability to find low energetic states is limited.
Inspired by the~recently proposed deep reinforcement learning method for finding spin glass ground states~\cite{FSNLSL2021}, here, we propose a new post-processing error correction schema for quantum annealers called Simulated Annealing with Reinforcement (SAwR). In this procedure, we combine deep reinforcement learning with simulated annealing. We employ a graph neural network to encode the~Ising instance into an ensemble of low-dimensional vectors used for reinforcement learning. The~agent learns a strategy for improving (finding a lower energy state) solutions given by the~physical quantum annealer. The process of finding the lower energy state involves "flipping" spins one by one according to the learned strategy and recording the energy state after each step. The solution is defined as the lowest energy state found during this procedure. In Simulated Annealing with Reinforcement, we start with simulated annealing and, at low temperature, we replace the~Metropolis-Hasting criterion with a single pass of spin flipping procedure.
Unlike recent error-correcting schema~\cite{PAL2014,PAL2015,V2015} we do not utilize multiple physical qubits for representing single logical qubits. This approach allows for a far greater size of problems to which our method applies. However, the performance of SAwR is still inferior to a typical algorithm one could incorporate in this context, such as simulated annealing. Nevertheless, using reinforcement learning for post-error correction is still an open and promising avenue of research.
\begin{figure}[ht!]
\centering
\includegraphics[width = \textwidth]{schema.eps}
\caption{\small Overview of our method, arrows represent subsequent steps. First, we define the Ising instance by providing couplings strength and biases (external magnetic field strength). Then we obtain the proposed solution from a quantum annealer. Here, it is represented by a single Chimera unit cell rendered as a graph. Black nodes represent spin value $\sigma_i = -1$ and white nodes represent spin value $\sigma_i = 1$. Edges represent couplings between spins. In the next step, we encode such an instance using a graph neural network into the matrix of encodings, where each row corresponds to the embedding of a vertex. Then this matrix is passed through a decoder to obtain Q-values of actions associated with each vertex. The spin flipping procedure involves "flipping" spins one by one according to Q-Values, starting from the highest and recording the energy state after each step. The solution is defined as the lowest energy state found during this procedure.}
\label{schema}
\end{figure}
\section{Ising Spin Glass and Quantum Annealing}
\subsection{Quantum Annealing in D-Wave }
The Ising problem is defined on some arbitrary simple graph $G$. An Ising Hamiltonian is given by:
\begin{equation}
\label{ising}
H_{\text{Ising}}(\sigma) = \sum_{\langle i,j \rangle \in G} J_{i,j} \sigma_i \sigma_j + \sum_{i} h_i \sigma_i,
\end{equation}
\noindent where $\sigma_i \in \{+1,-1\}$ denotes the~$i$-th Ising spin, $\sigma = \{ \sigma_1, \ldots, \sigma_n \}$ is the~vector representation of a spin configuration. $\langle i,j \rangle$ denotes neighbours in graph $G$, $J_{i,j}$ strength of interaction (coupling coefficient) between $i$-th and $j$-th spin. $h_i$ is external magnetic field (bias) affecting $i$-th spin. The~goal is to find spin configuration, called ground state,
\begin{equation}
\label{goal}
\sigma^{*} = \argmin_{\sigma} \, H_{\text{Ising}}(\sigma).
\end{equation}
\noindent such that energy of $H_{\text{Ising}}$ is minimal. This is an NP-hard problem. Quantum annealing is a method for finding the~ground state of (\ref{ising}). This is done by adiabatic evolution from the initial Hamiltonian $H_X$ of the~Transverse-field Ising model to the final Hamiltonian $H_{\text{Ising}}$. The~Hamiltonian of this process is described by
\begin{equation}
\mathcal{H}(t) = A(t)H_{X} + B(t)H_{\text{Ising}},
\end{equation}
\noindent where $H_X = \sum_i \hat{\sigma_i}^x$ and $\hat{\sigma_i}^x$ is the~standard Pauli $X$ matrix acting on the~$i$-th qubit. The~function $A(t)$ decreases monotonically to zero, while $B(t)$ increases monotonically from zero, with $t \in [0, t_f]$, where $t_f$ denotes total time of anneal~\cite{JRS2007,SMTC2002}. For a closed system, the~adiabatic theorem guarantees that if the~initial state is the~ground state, then the~final state will be arbitrarily close to the~ground state, provided all technical requirements are met~\cite{K1950}. In summary, the~systems start with a set of qubits, each in a superposition state of $-1$ and $1$. By annealing, the~system collapses into the~classical state that represents the~minimum energy state of the~problem, or one very close to it.
The D-Wave 2000Q is a psychical realization of the~quantum annealing algorithm. Sadly, the~idealized conditions of the~adiabatic theorem are nearly impossible to be realized in a physical device. In such an open system, there is inevitable thermal noise which may cause decoherence~\cite{decoherence}. Furthermore, due to technical limitations, those device suffers from programming control errors on the~$h_i$ and $J_{i,j}$ terms, which can unintentionally cause the~annealer to evolve according to the~wrong Hamiltonian~\cite{control}.
\subsection{D-Wave 2000Q}
At the~heart of every D-wave quantum annealer lies the~Quantum Processing Unit (QPU), which is a lattice of interconnected qubits. While its physical details are beyond the scope of this paper\footnote{Interested reader may find details in~\cite{dwave_IEEE} and~\cite{dwave_nature1} }, it is necessary to mention some technical details. In QPU, qubits can be thought of as loops being "oriented" vertically or horizontally (see figure \ref{qubits}) and connected to each other via devices called couplers. How qubits and couplers are interconnected is described by QPU topology.
As of the~time of writing, there are two available architectures of D-Wave quantum annealers, namely 2000Q with Chimera topology deployed in 2017 and Advantage with Pegasus topology deployed in 2020. Third, called Advantage~2 with Zephyr topology~\cite{zephyr} is stated to release in 2023-2024~\cite{roadmap}. In this work, we will focus on the~2000Q device and Chimera topology.
\begin{figure}[t]
\centering
\includegraphics[scale = 0.34]{couplers.eps}
\caption{\small Qubits are represented as horizontal and vertical loops. This graphic shows three coupled unit cells with eight qubits each. Green dots represent internal couplers connecting qubits inside the~unit cell, while blue bars represent external couplers connecting different unit cells.}
\label{qubits}
\end{figure}
\subsubsection{Chimera Topology}
The basic building block of Chimera topology is a set of connected qubits called a unit cell. Each unit cell consists of four horizontal qubits connected to four vertical qubits via couplers which form bipartite connectivity as seen in figure \ref{unit_cell}. Unit cells are tiled vertically and horizontally with adjacent qubits connected, creating a lattice of sparsely connected qubits.
It is conceptually valuable to categorize couplers into \emph{internal couplers} which connect intersecting (orthogonal) qubits and \emph{external couplers} which connect colinear pairs of qubits (that is, pairs of qubits that lie in the~same row or column). The~notation $C_n$ describes the~Chimera grid composed of $n \times n$ coupled unit cells, consisting of $8n^2$ qubits. D-Wave 2000Q device is equipped with $C_{16}$ QPU, with more than 2000 qubits~\cite{manual}.
\begin{figure}
\centering
\begin{subfigure}[t]{0.3\textwidth}
\centering
\tikzstyle{node}=[circle,draw=blue!50,fill=blue!20]
\begin{tikzpicture}
\node at (0,0) [node] (1) {3};
\node at (0,1) [node] (2) {2};
\node at (0,2) [node] (3) {1};
\node at (0,3) [node] (4) {0};
\node at (2,0) [node] (5) {7}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (2,1) [node] (6) {6}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (2,2) [node] (7) {5}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (2,3) [node] (9) {4}
edge (1)
edge (2)
edge (3)
edge (4);
\end{tikzpicture}
\caption{\small Unit cell rendered as column}
\label{column}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.4\textwidth}
\centering
\tikzstyle{node}=[circle,draw=blue!50,fill=blue!20]
\begin{tikzpicture}[scale=0.8]
\node at (-2,0) [node] (1) {0};
\node at (-1,0) [node] (2) {1};
\node at (1,0) [node] (3) {2};
\node at (2,0) [node] (4) {3};
\node at (0,2) [node] (5) {4}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (0,1) [node] (6) {5}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (0,-1) [node] (7) {6}
edge (1)
edge (2)
edge (3)
edge (4);
\node at (0,-2) [node] (8) {7}
edge (1)
edge (2)
edge (3)
edge (4);
\end{tikzpicture}
\caption{\small Unit cell rendered as cross}
\label{cross}
\end{subfigure}
\hfill
\caption{\small Different renderings of Chimera unit cell as graph. Nodes represent qubits and edges represent internal couplers.}
\label{unit_cell}
\end{figure}
\section{Results}
\subsection{Reinforcement Learning Formulation}
We will consider standard reinforcement learning setting defined as Markov Decision Process~\cite{SB2018} where an agent interacts with an environment over a number of discrete time steps $t = 0,1, \ldots, T$. At each time step $t$, the~agent receives a state $s_t$ and selects an action $a_t$ from some set of possible actions $\mathcal{A}$ according to its policy $\pi$, where $\pi$ is a mapping from set of states $\mathcal{S}$ to set of actions $\mathcal{A}$. In return, the~agent receives a scalar reward $r_t$ and moves to next state $s_{t+1}$. The~process continues until the~agent reaches a terminal state $s_T$ after which the~process restarts. We call one pass of such process an episode. The~return at time step $t$, denoted $R_t = \sum_{k=0}^{T-t} \gamma^k r_{t+k}$ is defined as sum of rewards that agent will receive for rest of the~episode discounted by discount factor $\gamma \in (0, 1]$. The~goal of the~agent is to maximize the~expected return from each state $s_t$.
We will start by defining state, action, and reward in the~context of the~Ising spin-glass model.
\begin{itemize}
\item \textbf{State}: a state $s$ represents the~observed spin glass instance, including both the~spin configuration $\sigma$, the~coupling strengths $\{ J_{ij} \}$ and values of external magnetic field $\{ h_i \}$.
\item \textbf{Action}: an action $a^{(i)}$ means to flip spin $i$. By flipping spin we mean changing is value to opposite. For example, after agent performs action $a^{(i)}$, spin $\sigma_i = 1$ becomes $\sigma_i = -1$. Agent can flip each spin once.
\item \textbf{Reward}: the~reward $r(s_t; a^{(i)}_t; s_{t+1})$ is defined as the~energy change after flipping spin $i$ from state $s_t$ to a new state $s_{t+1}$.
\end{itemize}
Starting at $t = 0$, an agent flips one spin during each time step, which moves him to the~next state (different spin configuration). The~terminal state $s_T$ is met when the~agent has flipped each spin. The~solution is defined as spin configuration $\sigma$ corresponding lowest energy state found during this procedure.
An action-value function $Q^{\pi}(s,a) = \mathbb{E}(R_t ~\vert~ s_t = s,~ a_t = a)$ is the~expected return for selecting action $a$ in state $s$ and following policy $\pi$. The~value $Q^{\pi}(s,a)$ is often called $Q$-value of action $a$ in state $s$. The~optimal action-value function $Q^{*}(s,a) = \max_{\pi} Q^{\pi}(s,a)$ which gives the~maximum action value for state $s$ and action $a$ achievable by any policy. As learning optimal action-value function is in practice infeasible, We seek to learn function approximator $Q(s,a;\Theta) \approx Q^{*}(s,a)$ where $\Theta$ is set of learnable model parameters. We denote policy used in such aproximation as $\pi_{\Theta}$.
\begin{figure}[t]
\centering
\includegraphics[scale = 0.6]{procedure.eps}
\caption{\small Overview of single episode. To simplify, we show it on a single Chimera unit cell. White nodes represent $\sigma_i = 1$ and black nodes represent $\sigma_i = -1$ We start in some state $s_0$ and one by one we flip spins until all spins are flipped. The~solution is defined as spin configuration $\sigma^*$ corresponding lowest energy state found during this procedure.}
\label{fig:my_label}
\end{figure}
\subsection{Model Architecture}
Our model architecture is inspired by DIRAC (\textbf{D}eep reinforcement learning for sp\textbf{I}n-glass g\textbf{R}ound-st\textbf{A}te \textbf{C}alculation), an Encoder-Decoder architecture introduced in~\cite{FSNLSL2021}. It exploits the~fact that the~Ising spin-glass instance is wholly described by the~underlying graph. In this view, couplings $J_{i,j}$ become edge weights, external magnetic field $h_i$ and spin $\sigma_i$ become node weights. Employing DIRAC is a two-step process. At first, it encodes the~whole spin-glass instance such that every node is embedded into a low-dimensional vector, and then the~decoder leverages those embeddings to calculate the~$Q$-value of every possible action. Then, the agent chooses the~action with the~highest $Q$-value. In the~next sections, We will describe those steps in detail.
\subsubsection{Encoding}
As described above, the~Ising spin-glass instance can be described in the~language of graph theory. It allows Us to employ graph neural networks~\cite{GRL,GSRVD2017}, which are neural networks designed to take graphs as inputs. We use modified SGNN (\textbf{S}pin \textbf{G}lass \textbf{N}eural \textbf{N}etwork)~\cite{FSNLSL2021} to obtain node embedding. To capture the~coupling strengths and external field strengths (i.e., edge weights $J_{i,j}$ and node weights $h_i$), which are crucial to determining the~spin glass ground states, SGNN performs two updates at each layer, specifically, the~edge-centric update and the~node-centric update, respectively.
Lets $z_{(i,j)}$ denote embedding of edge $(i,j)$ and $z_{(i)}$ embedding of node $i$. The~edge-centric update aggregates embedding vectors from from its adjacent nodes (i.e. for edge $(i,j)$ this update aggregate emmbedings $z_{(i)}$ and $z_{(j)}$), and then concatenates it with self-embedding $z_{(i,j)}$. Vector obtained in this way is then subject to non-linear transformation (ex. ReLU$(x)= \max(0,x)$). Mathematically it can be described by following equation
\begin{equation}
\label{edge}
z_{(i,j)}^{k+1} = \text{ReLU}(\gamma_{\theta} (z_{(i,j)}^{k}) \oplus \phi_{\theta}(z_{(i)}^{k} + z_{(j)}^{k})),
\end{equation}
\noindent where $z_{(i,j)}^{k}$ denotes encoding of edge $(i,j)$ obtained after $k$ layers. Similary $z_{(i)}^{k}$ denotes encoding of node $i$ obtained after $k$ layers, $\gamma_{\theta}$ and $\phi_{\theta}$ are some differentiable functions (ex. feed-foward neural networks) whih demends on set of paramethers $\theta$. Symbol $\oplus$ is used to denote concatenation operation.
The~node-centric update is defined in similar fashion. It aggregates embedding of adjacent edges, and then concatenates it with self-embedding $z_{(i)}$. Later we transform this concatenated vector to obtain final embedding. Using notation from equation \ref{edge}, final result is following:
\begin{equation}
z_{(i)}^{k+1} = \text{ReLU}(\phi_\theta(z_{(i)}^k) \oplus \gamma_\theta(\text{E}_i^k)),
\end{equation}
\begin{equation}
\text{E}_i^k = \sum_{j} z_{(i, j)}^k .
\end{equation}
\noindent Edge features are initialized as edge weights $\{J_{i,j}\}$. It is not trivial to find adequate node features, as node weights $\{h_i \}$ and spins ${\sigma_i}$ are not enough.
It is worth noting that both those operations are \textit{message passing} schema~\cite{GRL}. Edge-centric update aggregate information about adjacent nodes of edge and edge itself and sends it as a "message" to this edge. Similarly, node-centric update aggregate information about edges adjacent to the~node and the~node itself. In those edges are also encoded information about neighbouring nodes.
We also included pooling layers not presented in the~original design. We reasoned that after concatenation, vectors start becoming quite big, so we employ pooling layers to not only reduce the~model size but also preserve the~most essential parts of every vector.
As every node is a potential candidate for action, we call the~final encoding of node $i$ its \textit{action embedding} and denote it as $Z_i$. To represent the~whole Chimera (state of our environment), we use \textit{state embedding}, denoted as $Z_s$, which is the~sum over all node embedding vectors, which is a straightforward but empirically effective way for graph-level encoding~\cite{KDZDS2017}.
\subsubsection{Decoding}
Once all action embeddings $Z_i$ and state embedding $Z_s$ are computed in the~encoding stage, the~decoder will leverage these representations to compute approximated state-action value function $Q(s, a;\Theta)$ which predicts the~expected future rewards of taking action $a$ in state $s$, and following the~policy $\pi_{\Theta}$ till the~end. Specifically, we concatenate the~embeddings of state and action and use it as decoder input. In principle, any decoder architecture may be used. Here, we use a standard feed-forward neural network. Formally, the~decoding process can be written as:
\begin{equation}
Q(s, a^{(i)}; \Theta) = \psi_{\Theta}(Z_s \oplus Z_i),
\end{equation}
\noindent where $\psi_{\Theta}$ is a dense feed-forward neural network.
\subsection{Training}
We train our model on randomly generated Chimera instances. We found that the~minimal viable size of the~training instance is $C_3$ (as a reminder, $C_3$ is Chimera architecture with nine unit cells arranged into a $3 \times 3$ grid, which gives us 72 spins). Smaller instances lack couplings between clusters, crucial in full Chimera, which leads to poor performance. We generate $\{ J_{i,j}\}$ and $\{ h_i\}$ from normal distribution $\mathcal{N}(0,1)$ and starting spin configuration $\sigma$ from uniform distribution. To introduce low-energy instances, we employed the~following pre-processing procedure. For each generated instance, with probability $p=10\%$, we perform standard simulated annealing before passing the~instance through SGNN.
We seek to learn approximation of optimal action-value function $Q(a,s;\Theta)$, so as reinforcement learning algorithm we used standard n-step deep $Q$ learning~\cite{PW1994,DQL} with memory replay buffer. During episode we collect sequence of states action and rewards $\tau = (s_0, a_0, r_0, \ldots, s_{T-1}, a_{T-1}, r_{T-1}, s_T)$ with terminal state as final element. From those we construct $n$-step transitions $\tau^{n}_{t} = (s_t, a_t, r_{t,t+n}, s_{t+n})$ which we collect in memory replay buffer $\mathcal{B}$. Here $r_{t,t+n} = \sum_{k =0}^{k=n} \gamma^{k} r_{t+k}$ is return after $n$-steps.
\subsection{Simulated Annealing with Reinforcement}
Simulated annealing with reinforcement (SAwR) combines machine learning and classical optimization algorithm. Simulated annealing (SA) takes its name from a process in metallurgy involving heating a material and then slowly lowering the~temperature to decrease defects, thus minimizing the~system energy. In SA, we start in some state $s$ and in each step, we move to a randomly chosen neighboring state $s'$. If a move lowers energy $E(s)$ of the~system, we accept it. If it doesn't, we use the so-called the~Metropolis-Hasting criterion.
\begin{equation}
\mathbb{P}(\text{accept} \, s' ~\vert~ s) = \min (1, e^{-\beta \Delta E}),
\end{equation}
\noindent where $\Delta E = E(s') - E(s)$ and $\beta$ denotes inverse temperature $1/T$. In our case, the~move is defined as a single-spin flip. Simulated annealing tends to accept all possible moves at high temperatures (i.e., lower $\beta$). However, it likely accepts only those moves that lower the~energy at low temperatures.
Our idea is to reinforce random sampling with a trained model. It means that at low temperatures, instead of using the~Metropolis-Hasting criterion, we perform a single pass of the~DIRAC episode.
\section{Experiments}
We collected data from the D-Wave 2000Q device using default parameters (number of samplings, annealing time, etc.). We have generated 500 random instances of sizes $C_4$, $C_8$, $C_{12}$ and $C_{16}$, corresponding to systems of sizes 128, 512, 1152 and 2048 spins respectively. We used identical distributions to training instances, so $\{ J_{i,j}\}$ and $\{ h_i\}$ was generated from normal distribution $\mathcal{N}(0,1)$ and starting spin configuration $\sigma$ from uniform distribution.
We then used quantum annealing to obtain the low energy states of generated instances.
We have used three methods - standard simulated annealing, a single pass of spin-flipping procedure and simulated annealing with reinforcement. Results are shown in figures \ref{a} and \ref{b}.
We tested for two metrics: the probability of finding lower energy states and the mean value of an improvement over starting energy state. To compute the probability for each Chimera size, we started with proposed solutions obtained from quantum annealer and tried to lower them using different tested methods. Then we counted those instances for which a lower energy state was found. We define the value of the improvement as the difference between starting energy state and the lowest energy state found by the tested method in abstract units of energy.
Simulated annealing with reinforcement achieved lower probabilities of finding a lower energy state. Although the difference between SAwR and traditional simulated annealing is slight, its consistency across all sizes suggests that it is systemic rather than random noise. The single pass of the spin-flipping procedure was the order of magnitude worse, reaching approximately $1\%$ success rate.
It is interesting that, on average, SAwR was able to find a better low energy state than simulated annealing, but still, the difference is not significant.
\begin{figure}
\centering
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{prob.eps}
\caption{\small The probability of finding a lower energy state was computed over 500 random instances for each Chimera size. Here we decided to omit results for a single pass of the spin flipping procedure because its results were the order of magnitude worse, making the figure hard to read.}
\label{a}
\end{subfigure}
\hfill
\begin{subfigure}{\textwidth}
\includegraphics[width=\textwidth]{imporv.eps}
\caption{\small Mean value of the improvement. We define the value of the improvement as the difference between starting energy state and the lowest energy state found by the tested method in abstract units of energy. The mean was computed over all instances where the method could find improvement}
\label{b}
\end{subfigure}
\caption{\small Results of our experiments for tested metrics. By model, we define a single pass of the spin flipping procedure. SAwR denotes simulated annealing with reinforcement.}
\label{fig:figures}
\end{figure}
\section{Discussion and Further Work}
The real-life Chimera graph is a much more complex problem than the~regular lattice employed in~\cite{FSNLSL2021}, which may be the reason for poor performance. However, we managed to replicate the excellent scaling of DIRAC. We trained our model on relatively small instances and employed it for large ones. We did not observe significant differences in performance relative to the size of the system. Much more work is needed because more complex architectures are deployed (Pegasus, Zephyr) by D-wave systems.
One possible avenue of research is changes in architecture. Right now goals of the~encoder and decoder are quite different. The~encoder tries to encode information while the~decoder leverages them for reinforcement learning. Training them both at the~same time might be difficult. One option is to divorce them from a single architecture. For example, we may use graph autoencoder~\cite{GAE,GAE2} to train the~encoder. Then for the~reinforcement learning part, we would use an already trained encoder and train only the~neural network responsible for approximating the~action-value function.
Another option is to use different reinforcement learning algorithms. Asynchronous methods have been shown to consistently beat their synchronous counterparts, especially asynchronous advantage actor-critic~\cite{A3C}, but asynchronous Sarsa or Q-learning also look promising. Methods based on Monte Carlo tree search inspired by AlphaZero~\cite{AZ} also seems promising. It has shown excellent performance on tasks involving large search space (ex. Chess, Go).
\section*{Acknowledgments}
This research was supported by the Foundation for Polish Science (FNP) under grant number TEAM NET POIR.04.04.00-00-17C1/18-00 (LP, ZP, and BG). TS acknowledges support from the National Science Centre (NCN), Poland, under SONATA BIS 10 project number 2020/38/E/ST3/00269. This research was partially funded by National Science Centre, Poland (grant no 2020/39/ B/ST6/01511 and 2018/31/N/ST6/02374) and Foundation for Polish Science (grant no POIR.04.04.00-00-14DE/ 18-00 carried out within the Team-Net program co-financed by the European Union under the European Regional Development Fund). For the purpose of Open Access, the author has applied a CC-BY public copyright license to any Author Accepted Manuscript (AAM) version arising from this submission.
\bibliographystyle{splncs04}
| 97bea8085b4fcf23b1eb7d1656cb95b0a37cd879 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\begin{figure}[t]
\begin{center}
\includegraphics[width=7cm]{intro4cycle.eps}
\caption{\label{fig:intro}
Schematic view of chemical enrichment in galaxies.
}
\end{center}
\end{figure}
Explaining the origin of the elements is one of the scientific triumphs linking nuclear physics with astrophysics. As Fred Hoyle predicted, carbon and heavier elements (`metals' in astrophysics) were not produced during the Big Bang but instead created inside stars. So-called $\alpha$ elements (O, Ne, Mg, Si, S, Ar, and Ca) are mainly produced by core-collapse supernovae, while iron-peak elements (Cr, Mn, Fe, and Ni) are more produced by thermonuclear explosions, observed as Type Ia supernovae (SNe Ia; \citealt{kob20sr}, hereafter K20). The production depends on the mass of white dwarf (WD) progenitors, and a large fraction of SNe Ia should come from near-Chandrasekhar (Ch) mass explosions (see \citealt{kob20ia} for constraining the relative contribution between near-Ch and sub-Ch mass SNe Ia).
Among core-collapse supernovae, hypernovae ($\gtsim 10^{52}$ erg) produce a significant amount of Fe as well as Co and Zn, and a significant fraction of massive stars ($\gtsim 20M_\odot$) should explode as hypernovae in order to explain the Galactic chemical evolution (GCE; \citealt{kob06}, hereafter K06).
Heavier elements are produced by neutron-capture processes. The slow neutron-capture process (s-process) occurs in asymptotic giant branch (AGB) stars \citep{kar16}, while the astronomical sites of rapid neutron-capture process (r-process) have been debated. The possible sites are neutron-star (NS) mergers \citep[NSM,][]{wan14,jus15}, magneto-rotational supernovae \citep[MRSNe,][]{nis15,rei21}, magneto-rotational hypernovae/collapsars \citep[MRHNe,][]{yon21a}, and common envelope jets supernovae \citep{gri22}. Light neutron-capture elements (e.g., Sr) are also produced by electron-capture supernovae (ECSNe, \citealt{wan13ecsn}), $\nu$-driven winds \citep{arc07,wan13nu}, and rotating-massive stars \citep{fri16,lim18}.
The cycles of chemical enrichment are schematically shown in Figure \ref{fig:intro}, where each cycle produces different elements and isotopes with different timescales.
In a galaxy, not only the total amount of metals, i.e. metallicity $Z$, but also elemental abundance ratios evolve as a function of time.
Therefore, we can use all of this information as fossils to study the formation and evolutionary histories of the galaxy. This approach is called Galactic archaeology, and several on-going and future surveys with multi-object spectrographs (e.g., APOGEE, HERMES-GALAH, Gaia-ESO, WEAVE, 4MOST, MOONS, Subaru Prime Focus Spectrograph (PFS), and Maunakea Spectroscopic Explorer (MSE)) are producing a vast amount of observational data of elemental abundances.
Moreover, integral field unit (IFU) spectrographs (e.g., SAURON, SINFONI, CALIFA, SAMI, MaNGA, KMOS, MUSE, and HECTOR) allow us to measure metallicity and elemental abundance ratios within galaxies. It is now possible to apply the same approach not only to our own Milky Way but also to other types of galaxies or distant galaxies. Let us call this extra-galactic archaeology.
One of the most important uncertainties in this quest is the input stellar physics, namely, the nucleosynthesis yields, which could directly affect conclusions on the formation and evolutionary histories of galaxies.
In this review, I will focus on the impact of stellar mass loss due to stellar rotation in Galactic and extra-galactic archaeology.
\vspace*{-2mm}
\section{Galactic chemical evolution}
\label{sec:gce}
Galactic chemical evolution (GCE) has been calculated analytically and numerically using the following equation:
\begin{equation}\label{eq:gce}
\frac{d(Z_if_{\rm g})}{dt}=E_{\rm SW}+E_{\rm SNcc}+E_{\rm SNIa}+E_{\rm NSM}-Z_i\phi+Z_{i,{\rm inflow}}R_{\rm inflow}-Z_iR_{\rm outflow}
\end{equation}
where the mass fraction of each element $i$ in gas-phase ($f_{\rm g}$ denotes the gas fraction) increases via element ejections from stellar winds ($E_{\rm SW}$), core-collapse supernovae ($E_{\rm SNcc}$), Type Ia supernovae ($E_{\rm SNIa}$), and neutron star mergers ($E_{\rm NSM}$). It also decreases by star formation (with a rate $\phi$), and modified by inflow (with a rate $R_{\rm inflow}$) and outflow (with a rate $R_{\rm outflow}$) of gas in/from the system considered.
It is assumed that the elemental abundance of gas is instantaneously well mixed in the system (called an one-zone model), but the instantaneous recycling approximation is not adopted nowadays.
The initial conditions are $f_{{\rm g},0}=1$ (a closed system) or $f_{{\rm g},0}=0$ (an open system) with the chemical composition ($Z_{i,0}$) from the Big Bang nucleosynthesis.
The first two terms depend only on nucleosynthesis yields, while the third and fourth terms also depend on modelling of the progenitor systems, which is uncertain. The last three terms are galactic terms, and should be determined from galactic dynamics, but are assumed with analytic formula in GCE models.
As in my previous works, star formation rate is assumed to be proportional to gas fraction as $\phi\!=\!\frac{1}{\tau_{\rm s}}f_{\rm g}$. The inflow rate is assumed to be $R_{\rm inflow}\!\!=\!\frac{t}{\tau_{\rm i}^2}\exp\frac{-t}{\tau_{\rm i}}$ (for the solar neighbourhood), or $R_{\rm inflow}\!\!=\!\frac{1}{\tau_{\rm i}}\exp\frac{-t}{\tau_{\rm i}}$ (for the rest). The outflow rate is assumed to be proportional to the star formation rate as $R_{\rm outflow}\!\!=\!\frac{1}{\tau_{\rm o}}f_{\rm g}\propto\phi$. In addition, star formation is quenched ($\phi=0$) in case with a galactic wind at an epoch $t_{\rm w}$.
The timescales are determined to match the observed metallicity distribution function (MDF) in each system (Fig.\,2 of K20). The parameter sets that have very similar MDFs give almost identical tracks of elemental abundance ratios (Fig.\,A1 of \citealt{kob20ia}).
Our nucleosynthesis yields of core-collapse supernovae were originally calculated in K06, 3 models of which are replaced in \citet{kob11agb} (which are also used in \citealt{nom13}), and a new set with failed supernovae is used in K20. This is based on the lack of observed progenitors at supernova locations in the HST data \citep{sma09}, and on the lack of successful explosion simulations for massive stars \citep{jan12,bur21}.
The yields for AGB stars were originally calculated in \citet{kar10} and \citet{kar16}, but a new set with the s-process is used in K20. The narrow mass range of super-AGB stars is also filled with the yields from \citet{doh14a}; stars at the massive end are likely to become ECSNe as Crab Nebula exploded in 1054. At the low-mass end, off-centre ignition of C flame moves inward but does not reach the centre, which remains a hybrid C+O+Ne WD. This might become a sub-class of SNe Ia \citep{kob15}.
For SNe Ia, the nucleosynthesis yields of near-Ch and sub-Ch mass models are newly calculated in \citet{kob20ia}, which used the same code as in \citet{leu18} and \citet{leu20} but with more realistic, solar-scaled initial composition. The initial composition gives significantly different (Ni, Mn)/Fe ratios.
r-process yields are taken from literature (see \S 2.1 of K20 for more details).
The Kroupa initial mass function (IMF) between $0.01M_\odot$ and $50M_\odot$ (or $120M_\odot$ in \S \ref{sec:f}, \ref{sec:iso}, and \S \ref{sec:hydro}) is used throughout this paper.
For comparison to observational data, the solar abundances are taken from \citet{asp09}, except for $A_\odot({\rm O}) = 8.76, A_\odot({\rm Th}) = 0.22$, and $A_\odot({\rm U}) = -0.02$ (\S 2.2 of K20 for the details).
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.95\textwidth]{origin.eps}
\caption{The time evolution (in Gyr) of the origin of elements in the periodic table: Big Bang nucleosynthesis (black), AGB stars (green), core-collapse supernovae including SNe II, HNe, ECSNe, and MRSNe (blue), SNe Ia (red), and NSMs (magenta). The amounts returned via stellar mass loss are also included for AGB stars and core-collapse supernovae depending on the progenitor mass. The dotted lines indicate the observed solar values.}
\label{fig:origin}
\end{center}
\end{figure*}
Using the K20 GCE model for the solar neighbourhood, we summarize the origin of elements in the form of a periodic table. In each box of Figure \ref{fig:origin}, the contribution from each chemical enrichment source is plotted as a function of time: Big Bang nucleosynthesis (black), AGB stars (green), core-collapse supernovae including SNe II, HNe, ECSNe, and MRSNe (blue), SNe Ia (red), and NSMs (magenta).
It is important to note that the amounts returned via stellar mass loss are also included for AGB stars and core-collapse supernovae depending on the progenitor star mass.
The x-axis of each box shows time from $t=0$ (Big Bang) to $13.8$ Gyrs, while the y-axis shows the linear abundance relative to the Sun, $X/X_\odot$.
The dotted lines indicate the observed solar values, i.e., $X/X_\odot=1$ and $4.6$ Gyr for the age of the Sun.
Since the Sun is slightly more metal-rich than the other stars in the solar neighborhood (see Fig.\,2 of K20), the fiducial model goes through [O/Fe]$=$[Fe/H]$=0$ slightly later compared with the Sun's age.
Thus, a slightly faster star formation timescale ($\tau_{\rm s}=4$ Gyr instead of 4.7 Gyr) is adopted in this model.
The evolutionary tracks of [X/Fe] are almost identical.
The adopted star formation history is similar to the observed cosmic star formation rate history, and thus this figure can also be interpreted as the origin of elements in the universe.
Note that Tc and Pm are radioactive.
The origin of stable elements can be summarized as follows:
\begin{itemize}
\item H and most of He are produced in Big Bang nucleosynthesis. As noted, the green and blue areas also include the amounts returned to the interstellar medium (ISM) via stellar mass loss in addition to He newly synthesized in stars.
Be and B are supposed to be produced by cosmic rays,
which are not included in the K20 model.
\item The Li model is very uncertain because the initial abundance and nucleosynthesis yields are uncertain. Li is supposed to be produced also by cosmic rays and novae, which are not included in the K20 model. The observed Li abundances show an increasing trend from very low metallicities to the solar metallicity, which could be explained by cosmic rays. Then the observation shows a decreasing trend from the solar metallicities to the super-solar metallicities, which might be caused by the reduction of the nova rate \citep{gri19}; this is also shown in theoretical calculation with binary population synthesis \citep{kem22}, where the nova rate becomes higher due to smaller stellar radii and higher remnant masses at low metallicities.
\item 49\% of C, 51\% of F, and 74\% of N are produced by AGB stars (at $t=9.2$ Gyr). Note that extra production from Wolf-Rayet (WR) stars is not included and may be important for F (see \S \ref{sec:f}).
For the elements from Ne to Ge, the newly synthesized amounts are very small for AGB stars, and the small green areas are mostly for mass loss.
\item $\alpha$ elements (O, Ne, Mg, Si, S, Ar, and Ca) are mainly produced by core-collapse supernovae, but 22\% of Si, 29\% of S, 34\% of Ar, and 39\% of Ca come from SNe Ia.
These fractions would become higher with sub-Ch-mass SNe Ia \citep{kob20ia} instead of 100\% Ch-mass SNe Ia adopted in the K20 model.
\item A large fraction of Cr, Mn, Fe, and Ni are produced by SNe Ia. In classical works, most of Fe was thought to be produced by SNe Ia, but the fraction is only 60\% in our model, and the rest is mainly produced by HNe. The inclusion of HNe is very important as it changes the cooling and star formation histories of the universe significantly \citep{kob07}.
Co, Cu, Zn, Ga, and Ge are largely produced by HNe.
In the K20 model, 50\% of stars at $\ge 20M_\odot$ are assumed to explode as hypernovae, and the rest of stars at $> 30M_\odot$ become failed supernovae.
\item Among neutron-capture elements, as predicted from nucleosynthesis yields, AGB stars are the main enrichment source for the s-process elements at the second (Ba) and third (Pb) peaks.
\item 32\% of Sr, 22\% of Y, and 44\% of Zr can be produced from ECSNe, which are included in the blue areas, even with our conservative mass ranges; we take the metallicity-dependent mass ranges from the theoretical calculation of super-AGB stars \citep{doh15}.
Combined with the contributions from AGB stars, it is possible to perfectly reproduce the observed trends, and no extra light element primary process (LEPP) is needed. The inclusion of $\nu$-driven winds in GCE simulation results in a strong overproduction of the abundances of the elements from Sr to Sn with respect to the observations.
\item For the heavier neutron-capture elements, contributions from both NS-NS/NS-black hole (BH) mergers and MRSNe are necessary, and the latter is included in the blue areas.
\end{itemize}
In this model, the O and Fe abundances go though the cross of the dotted lines, meaning [O/Fe] $=$ [Fe/H] $=0$ at 4.6 Gyr ago.
This is also the case for some important elements including N, Ca, Cr, Mn, Ni, Zn, Eu, and Th.
Mg is slightly under-produced in the model, although the model gives a $0.2-0.3$ dex higher [Mg/Fe] value than observed at low metallicities.
This [O/Mg] problem is probably due to uncertain nuclear reaction rates (namely, of $^{12}$C($\alpha$,$\gamma$)$^{16}$O) and/or the mass loss based by stellar rotation or binary interaction (see also Fig.\,9 of K06).
The under-production of the elements around Ti is a long-standing problem, which was shown to be enhanced by multi-dimensional effects (\citealt{sne16}; see also K15 model in K20).
The s-process elements are slightly overproduced even with the updated s-process yields.
Notably, Ag is over-produced by a factor of $6$, while Au is under-produced by a factor of $5$. U is also over-produced. These problems may require revising nuclear physics modelling (see \S \ref{sec:au}).
The contributions from SNe Ia depend on the mass of the progenitor WDs, and sub-Ch mass explosions produce less Mn and Ni, and more Si, S, and Ar than near-Ch mass explosions.
Figure \ref{fig:ofe} shows the [O/Fe]--[Fe/H] relations with varying the fraction of sub-Ch-mass SNe Ia. Including up to 25\% sub-Ch mass contribution to the GCE (dashed line) gives a similar relation as the K20 model (solid line), while the model with 100\% sub-Ch-mass SNe Ia (dotted line) gives too low an [O/Fe] ratio compared with the observational data. For Ch-mass SNe Ia, the progenitor model is based on the single-degenerate scenario with the metallicity effect due to optically thick winds \citep{kob98}. For sub-Ch-mass SNe Ia, the observed delay-time distribution is used since the progenitors are the combination of mergers of two WDs in double degenerate systems and low accretion in single degenerate systems; \citet{kob15}'s formula are for those in single degenerate systems only.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{sn-sch.ps}
\vspace*{-2mm}
\caption{The [O/Fe]--[Fe/H] relations in the solar neighborhood for the models with Ch-mass SNe Ia only (solid line), 75\% Ch plus 25\% sub-Ch-mass SNe Ia (dashed line), and sub-Ch-mass SNe Ia only (dotted line). The observational data (filled circles) are high-resolution non-local thermodynamic equilibrium (NLTE) abundances.
}
\label{fig:ofe}
\end{center}
\end{figure}
The [$\alpha$/Fe]--[Fe/H] relation is probably the most important diagram in GCE.
In the beginning of the universe, the first stars form and die, but the properties such as mass and rotation are uncertain and have been studied using the second generation, extremely metal-poor (EMP) stars.
Secondly, core-collapse supernovae occur, and their yields are imprinted in the Population II stars in the Galactic halo. The [$\alpha$/Fe] ratio is high and stays roughly constant with a small scatter. This plateau value does not depend on the star formation history but does on the IMF.
Finally, SNe Ia occur, which produce more Fe than O, and thus the [$\alpha$/Fe] ratio decreases toward higher metallicities; this decreasing trend is seen for the Population I stars in the Galactic disk.
Because of this [$\alpha$/Fe]--[Fe/H] relation, high-$\alpha$ and low-$\alpha$ are often used as a proxy of old and young age of stars in galaxies, respectively. Note that, however, that this relation is not linear but is a plateau and decreasing trend (called a `knee'). The location of the knee depends on the star formation timescale and is at a high metallicity in the Galactic bulge, followed by the Galactic thick disk, thin disk, and satellite galaxies show at lower metallicities.
\subsection{Mystery of gold}
\label{sec:au}
\begin{figure*}
\center
\includegraphics[width=0.95\textwidth]{xfe-sr-new.eps}
\caption{The [X/Fe]--[Fe/H] relations for neutron capture elements, comparing to the models in the solar neighborhood, with s-process from AGB stars only (blue long-dashed lines), with s-process and ECSNe (light-blue short-dashed lines), with s-process, ECSNe, and $\nu$-driven winds (green dotted-long-dashed lines), with s-process, ECSNe, and NS-NS mergers (olive dotted lines), with s-process, ECSNe, and NS-NS/NS-BH mergers (orange dotted-short-dashed lines), with s-process, ECSNe, NS-NS/NS-BH mergers, and MRSNe (red solid lines).
Observational data are updated from K20.}
\label{fig:xfe}
\end{figure*}
Figure \ref{fig:xfe} shows the evolutions of neutron-capture elements as [X/Fe]--[Fe/H] relations.
As known, AGB stars can produce the first (Sr, Y, Zr), second (Ba), and third (Pb) peak s-process elements, but no heaver elements (navy long-dashed lines). It is surprising that ECSNe from a narrow mass range ($\Delta M \sim 0.15-0.2M_\odot$) can produce enough of the first peak elements; with the combination of AGB stars, it is possible to reproduce the observational data very well (cyan short-dashed lines). This means that no other light element primary process (LEPP), such as rotating massive stars, is required. However, the elements from Mo to Ag seem to be overproduced, which could be tested with the UV spectrograph proposed for VLT, CUBES. Additional production from $\nu$-driven winds leads to further over-production of these elements in the model (green dot-long-dashed lines), but this should be studied with more self-consistent calculations of supernova explosions.
Neutron star mergers can produce lanthanides and actinides, but not enough (olive dotted lines); the rate is too low and the timescale is too long, according to binary population synthesis. This is not improved enough even if neutron-star black-hole mergers are included (orange dot-short-dashed lines). An r-process associated with core-collapse supernovae, such as MRSNe, is required. The same conclusion is obtained with other GCE models and more sophisticated chemodynamical simulations \citep[e.g.,][]{hay19}, as well as from the observational constraints of radioactive nuclei in the solar system \citep{wal21}.
In the GCE model with MRSNe (magenta solid lines), it is possible to reproduce a plateau at low metallicities for Eu, Pt, and Th, relative to Fe. However, even with including both MRSNe and neutron star mergers, the predicted Au abundance is more than ten times lower than observed. This underproduction is seen not only for the solar abundance but also for low metallicity stars although the observational data are very limited. UV spectroscopy is needed for investigating this problem further.
In the GCE models, we adopt the best available nucleosynthesis yields in order to explain the universal r-process pattern: $25M_\odot$ ``b11tw1.00'' 3D yields from \citet{nis15} and $1.3M_\odot$+$1.3M_\odot$ 3D/GR yields from \citet{wan14}.
These yields are sensitive to the electron fraction, which depends on hydrodynamics and $\nu$-processes during explosions. Post-process nucleosynthesis calculations of successful explosion simulations of massive stars are required.
However, it would not be easy to increase Au yields only since Pt is already in good agreement with the current model. There are uncertainties in some nuclear reaction rates and in the modelling of fission,
which might be able to increase Au yields only, without increasing Pt or Ag.
The predicted Th and U abundances are after the long-term decay, to be compared with observations of metal-poor stars, and the current model does not reproduce the Th/U ratio either.
It seems necessary to have the r-process associated with core-collapse supernovae, such as MRSNe.
Is there any direct evidence for such events?
There are a few magneto-hydrodynamical simulations and post-process nucleosynthesis that successfully showed enough neutron-rich ejecta to produce the 3rd peak r-process elements \citep{win12,mos18} as in the Sun. However, there is no observational evidence that support magneto-rotational supernovae. The predicted iron mass is rather small \citep{nis15,rei21}, and the astronomical object could be faint.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.65\textwidth]{Yong_Figure_2.eps}
\caption{The elemental abundance of a extremely metal-poor star SMSS J200322.54-114203.3, which has [Fe/H]$= -3.5$, comparing with mono-enrichment from a magneto-rotational hypernovae (blue dashed line and squares) and multi-enrichment including a neutron star merger (orange dot-dashed line and triangles).
The lower panel shows the differences, i.e., the observed values minus model values.
}
\label{fig:yong}
\end{center}
\end{figure}
Extremely metal-poor (EMP) stars have been an extremely useful tool in the Galactic archeology. At the beginning of galaxy formation, stars form from a gas cloud that was enriched only by one or very small number of supernovae, and hence the elemental abundances of EMP stars have offered observational evidences of supernovae in the past.
As a result, it is found that quite a large fraction of massive stars became faint supernovae that give a high C/Fe ratio leaving a relatively large black hole ($\sim 5M_\odot$, \citealt{nom13}). It might even possible to form a black hole from $10-20M_\odot$ stars \citep{kob14}.
It is challenging to find EMP stars due to the expected small number, and it is also challenging to measure elemental abundances in detail. Thus only a limited number of EMP stars have been analysed in previous works.
The Australian team has been using a strategic approach to increase EMP data; $\sim 26000$ candidates are found from photometric data on the SkyMapper telescope, 2618 of which have spectroscopic observations at ANU's 2.3m telescope \citep{dac19}, and for $\sim 500$ stars detailed elemental abundances are measured with higher-resolution spectra taken at larger telescopes such as Magellan, Keck, and VLT \citep{yon21b}. SMSS J200322.54-114203.3 was discovered in the SkyMapper EMP survey and reported in \citet{yon21a}.
Figure \ref{fig:yong} shows the elemental abundance of SMSS J200322.54-114203.3, which has [Fe/H]$= -3.5$, and is 2.3 kpc away with the Galactic halo orbit. This star has a very clear detection of uranium and thorium, and thus it is a so-called actinides boost star. The stars also clearly showed the solar r-process pattern from $Z \sim 60$ to 70. Surprisingly, the observed Zn abundance is very high ([Zn/Fe]$=0.72$), which indicates that the enrichment source was an energetic explosion ($\gtsim 10^{52}$ erg). It also showed normal $\alpha$ enhancement as for the majority of Galactic halo stars, although a normal supernova ($\sim 10^{51}$ erg) gives a range of $\alpha$ enhancement depending on the progenitor mass.
The star is not carbon enhanced ([C/Fe]$<0.07$) but shows nitrogen enhancement ([N/Fe]$=1.07$), which indicates that the progenitor stars were rotating. All of these features at $Z\le 30$ are consistent with the enrichment from a single hypernova.
The observed abundance pattern is compared with two theoretical models. The blue dashed line is for mono-enrichment from a single, magneto-rotational hypernova (MRHN), where the nucleosynthesis yields are obtained combining the iron-core of a magneto-rotational supernovae \citep{nis15} to the envelope of a $25M_\odot$ hypernova. The orange dot-dashed line shows a model of multi-enrichment where a neutron star merger occur in the interstellar matter chemically enriched by previous generations of core-collapse supernovae. The MRHN model gives better fit at and below Zn ($Z\le30$). Both models show overproduction around the 1st peak ($Z\sim40$) and around Ba ($Z=56$), which may be due to the uncertainties in nuclear astrophysics. The MRSN scenario is also consistent with the short timescale of the EMP star formation; due to the required long timescales of neutron star mergers, it is unlikely that this EMP star is enriched by a neutron star merger.
{\bf Future prospects}: This star provided the first observational evidence of the r-process associated with core-collapse supernovae. The r-process site could be in the jets as in the MRHN model, or could also be in the accretion disk around black hole as in the collapsar model \citep{sie19}. According to the chemical evolution models, this event is rare, only one in 1000 supernovae. MRHNe should be observable in future astronomical transient surveys such as with LSST/Rubin, since they eject a significant amount of iron; the ejected iron mass is $0.017M_\odot$ in the $25M_\odot$ model, while $0.33M_\odot$ in the $40M_\odot$ model \citep{yon21a}. It might be possible to detect the signature of Au and Pt production directly in the supernova spectra.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.7\textwidth]{Figure4b.eps}
\caption{The fluorine abundance in the galaxy NGP-190387 at redshift $z=4.4$, comparing with galactic chemical evolution models with only supernovae (blue lines), plus AGB stars (orange lines), plus WR stars (green lines).
}
\label{fig:franco}
\end{center}
\end{figure}
\subsection{Ramp-up of fluorine at high-redshifts}
\label{sec:f}
Fluorine is an intriguing element. The major production site has been debated for 30 years among low-mass AGB stars, stellar winds in massive stars, and the $\nu$-process in core-collapse supernovae \citep{kob11f}. The observational constraints have been obtained for stars, but accessible lines are only in infrared, and thus the sample number was limited. Moreover, there was confusion in the excitation energies and transition probabilities for the HF lines \citep{jon14}. The vibrational-rotational lines get too weak at low metallicities, and the available sample of EMP stars is only for carbon enhanced stars \citep{mur20}, which may not be representative of galactic chemical evolution.
\citet{fra21} opened a new window for constraining fluorine production in the early universe by directly measuring HF abundance in a galaxy at redshift $z=4.4$.
The galaxy NGP-190387 was discovered by the H-ATLAS survey with the Herschel satellite, with the redshift confirmed by the Northern Extended Millimetre Array (NOEMA). It is a gravitationally lensed galaxy with a magnification factor $\mu\simeq5$. The lowest rotational transitions from the HF molecule appeared as an absorption line in the Atacama Large Millimeter/submillimeter Array (ALMA) data. This molecule is very stable and the dominant gas-phase form of fluorine in the ISM.
Figure \ref{fig:franco} shows the fluorine abundance of NGP-190387 (yellow point), comparing with galactic chemical evolution models with various star formation timescales. In the models, it is assumed that star formation took place soon after the reionization, which was boosted probably due to galaxy merger 100 Myr before the observed redshift $z=4.4$. This assumption is based on the properties of other submillimetre galaxies (SMGs) at similar redshifts. The models show rapid chemical enrichment including fluorine, but in order to explain the observed fluorine abundance, the contribution from Wolf-Rayet stars (green lines) is required. The nucleosynthesis yields from \citet{lim18} are adopted in the models, assuming the rotational velocity of 300 km s$^{-1}$ for $13-120M_\odot$ stars.
{\bf Future prospects}: This galaxy provided the first observational evidence for rotating massive stars with initial masses $\gtsim 30M_\odot$ producing a significant amount of fluorine in the early universe.
Since it is not easy to explode these stars with $\nu$-driven mechanism, these stars are likely to remain black holes with masses $\gtsim 20 M_\odot$, which may be detected with gravitational wave missions.
It would be possible to constrain the formation history further if CNO abundances and/or $^{13}$C isotopic ratio are measured, because the yields depend on the mass of progenitor stars.
\subsection{GCE with rotating massive stars}
\label{sec:iso}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\textwidth]{iso.ps}
\vspace*{-1mm}
\caption{\label{fig:iso}
Evolution of isotopic ratios of GCE models for the Milky Way (red solid and blue long-dashed lines) and for a submillimetre galaxy (green short-dashed and magenta dotted lines), which are calculated with the K20 yields (red solid and green short-dashed lines) or the yields from \citet{lim18} with the rotational velocity of 300 km s$^{-1}$ (blue long-dashed and magenta dotted lines).
The solar symbol indicates the solar ratios \citep{asp09}.
See \citet{rom19} for the observational data sources of the ISM (filled squares and crosses).
}
\end{center}
\end{figure}
The evolution of isotopic ratios is shown in Figs.\,17-19 of \citet{kob11agb} and Fig.\,31 of K20 (see also \citealt{rom19}). $^{13}$C and $^{25,26}$Mg are produced from AGB stars ($^{17}$O might be overproduced in our AGB models), while other minor isotopes are more produced from metal-rich massive stars, and thus the ratios between major and minor isotopes (e.g., $^{12}$C/$^{13}$C, $^{16}$O/$^{17,18}$O) generally decrease as a function of time/metallicity. But this evolutionary trend is completely different if rotating massive stars are included.
Figure \ref{fig:iso} shows the evolution of isotopic ratios in the GCE models with and without rotating massive stars, comparing to observational data.
For the Milky Way (red solid and blue long-dashed lines), the adopted star formation history is the same as the solar neighbourhood model in \S \ref{sec:gce}.
For the SMG (green short-dashed and magenta dotted lines), rapid gas accretion and star formation with the timescales of 1 Gyr are assumed, and after $\sim 1.5$ Gyr the star formation is self-regulated with a constant gas fraction of 50\% (see M. Doherty et al., in prep. for more details).
The K20 model (red solid lines) is not bad, well reproducing C and O isotopic ratios in the Sun (the solar symbol) and in the present-day ISM (filled squares), but not the $^{13}$C/$^{18}$O ratio (cross).
The model with rotating massive stars (blue long-dashed lines) gives a better evolutionary track of the $^{13}$C/$^{18}$O ratio, showing a rapid increase in the past 5 Gyrs in the Milky Way. However, $^{12}$C is overproduced at all metallicities, and $^{16}$O is overproduced at high metallicities in this model.
Here the yields from \citet{lim18} with the rotational velocity of 300 km s$^{-1}$ are used. With 150 km s$^{-1}$, the $^{12}$C/$^{13}$C ratio becomes even larger, and the $^{13}$C/$^{18}$O ratio becomes slightly smaller. The model with 150 km s$^{-1}$ is worse than the model with 300 km s$^{-1}$ in order to reproduce the observations, and using an averaged rotational velocity as in \citet{pra18} does not solve this problem.
With these yields set, even excluding stellar rotation, the $^{12}$C/$^{13}$C ratio is systematically higher than in the K20 yields, and the difference may be caused by their choice of nuclear reaction rates, mass-loss, and/or convection.
It is possible that these effects could change the evolution of the $^{18}$O/$^{17}$O ratio.
The overproduction of $^{16}$O is probably due to their choice of remnant masses, and in fact these yields do not reproduce the observed [O/Fe]--[Fe/H] relation.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{rotBa.eps}
\vspace*{-2mm}
\caption{The [Ba/Fe]--[Fe/H] relation, comparing with K20 model (red line) and models with rotating massive stars with different rotational velocities (green and blue lines). See K20 for the other data points and model details.}
\label{fig:ba}
\end{center}
\end{figure}
The impact of stellar rotation should be studied from s-process elements, since the weak-s process forms s-process elements from the existing seeds at a much shorter timescale than AGB stars \citep{fri16,lim18}.
For s-process elements, there is some underproduction from [Fe/H] $\sim -3$ to $\sim -1$ in the K20 model, which could be improved in the inhomogeneous enrichment effect in chemodynamical simulations (\S \ref{sec:n}). Alternatively, the models with rotating massive stars can explain the observed average trend (Fig.\,\ref{fig:ba}); in the first model (green line, also plotted in \citealt{yon21b}) the same metallicity-dependent distribution of rotational velocity as in \citet{pra18} is used, while in the second and better model (blue line) 150 and 300 km s$^{-1}$ is assumed for HNe and MRHNe, respectively.
As already noted, the nucleosynthesis yields from \citet{lim18} cannot reproduce the observations of many elements, and thus we use the contributions from stellar envelopes and winds only in addition to the K20 yields (see Kobayashi et al., in prep. for more details).
\vspace*{-2mm}
\section{Chemodynamical evolution of galaxies}
\label{sec:hydro}
Thanks to the development of super computers and numerical techniques, it is now possible to simulate the formation and evolution of galaxies from cosmological initial conditions, i.e., initial perturbation based on $\Lambda$ cold dark matter (CDM) cosmology. Star formation, gas inflow, and outflow in Eq.(\ref{eq:gce}) are not assumed but are, in principle, given by dynamics. Due to the limited resolution, star forming regions in galaxy simulations, supernova ejecta, and active galactic nuclei (AGN) cannot be resolved, and thus it is necessary to model star formation and the subsequent effects -- feedback -- introducing a few parameters. Fortunately, there are many observational constraints, from which it is usually possible to choose a best set of parameters. To ensure this, it is also necessary to run the simulation until the present-day, $z=0$, and reproduce a number of observed relations at various redshifts.
Although hydrodynamics can be calculated with publicly available codes such as Gadget, RAMSES, and AREPO, modelling of baryon physics is the key and is different depending on the simulation teams/runs, such as EAGLE, Illustris, Horizon-AGN, Magneticum, and SIMBA simulations.
These are simulations of a cosmological box with periodic boundary conditions, and contain galaxies with a wide mass range (e.g., $10^{9-12} M_\odot$ in stellar mass) at $z=0$.
In order to study massive galaxies it is necessary to increase the size of the simulation box (e.g., $\gtsim 100$ Mpc), while in order to study internal structures it is necessary to improve the spacial resolution (e.g., $\ltsim 0.5$ kpc).
The box size and resolution are chosen depending on available computational resources.
On the other hand, zoom-in techniques allow us to increase the resolution focusing a particular galaxy,
but this also requires tuning the parameters with the same resolution, comparing to a number of observations in the Milky Way.
For the baryon physics, the first process to calculate is
{\bf radiative cooling}, and photo-heating by a uniform
and evolving UV background radiation.
We use a metallicity-dependent
cooling function computed with the MAPPINGS III
software
(see \citealt{kob07} for the details),
and find candidate gas particles that can form stars in a given timestep.
The effect of $\alpha$ enhancement relative to Fe is taken into account using the observed relation in the solar neighborhood.
Our {\bf star formation} criteria are:
(1)
converging flow, $(\nabla \cdot \mbox{\boldmath$v$})_i < 0$; (2) rapid
cooling, $t_{\rm cool} < t_{\rm dyn}$; and (3) Jeans unstable gas, $t_{\rm
dyn} < t_{\rm sound}$. The star formation timescale is taken to be
proportional to the dynamical timescale ($t_{\rm sf} \equiv
\frac{1}{c_*}t_{\rm dyn}$), where $c_*$ is a star formation timescale
parameter.
The parameter $c_*$ basically determines when to form stars following cosmological accretion, and has a great impact on the size--mass relation of galaxies (Fig.\,4 of \citealt{kob05}).
We found that it is better not to include a fixed density criterion.
Note that based on smaller-scale simulations,
more sophisticated analytic formula are proposed
including the effects of turbulence and magnetic fields.
A fractional part of the mass of the gas
particle turns into a new star particle \citep[see][for the details]{kob07}.
Note that an individual
star particle has a typical mass of $\sim 10^{5-7} M_\odot$, i.e.~it does not
represent a single star but an association of many stars. The masses of
the stars associated with each star particle are distributed according to an
IMF. We adopt Kroupa IMF, which is assumed to be
independent of time and metallicity, and
limit the IMF to the mass range
$0.01M_\odot \le m \le 120M_\odot$.
This assumption is probably valid for the resolution down to $\sim 10^4 M_\odot$.
We then follow the evolution of the star particle, as a simple stellar population, which is defined as a single generation of coeval and chemically homogeneous stars of various masses, i.e. it consists of a number of stars with various masses but the same age and chemical composition \citep{kob04}.
The production of each element $i$ from the star particle (with an initial mass of $m_*^0$) is calculated using a very similar equation as Eq.(\ref{eq:gce}):
\begin{equation}\label{eq:e-z}
E_Z(t)= m_*^0 \left[ E_{\rm SW}+E_{\rm SNcc}+E_{\rm SNIa}+E_{\rm NSM} \right] .
\end{equation}
Similarly, the energy production from the star particle is:
\begin{equation}\label{eq:e-e}
E_e(t) = m_*^0 \left[ e_{e,{\rm SW}}{\cal{R}}_{\rm SW}(t)+e_{e,{\rm SNcc}}{\cal{R}}_{\rm SNcc}(t)+e_{e,{\rm SNIa}}{\cal{R}}_{\rm SNIa}(t) \right]
\end{equation}
which includes the event rates of SWs (${\cal{R}}_{\rm SW}$), core-collapse supernovae (${\cal{R}}_{\rm SNcc}$), and SNe Ia (${\cal{R}}_{\rm SNIa}$), and the energy per event (in erg) for SWs: $e_{e,{\rm SW}}=0.2 \times 10^{51} ({Z}/{Z_\odot})^{0.8}$ for $>8M_\odot$ stars, Type II supernovae: $e_{e,{\rm SNII}}=1 \times 10^{51}$, hypernovae: $e_{e,{\rm HN}}=10,10,20,30 \times 10^{51}$ for $20,25,30,40 M_\odot$ stars, respectively, and for SNe Ia: $e_{e,{\rm SNII}}=1.3 \times 10^{51}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{chem2agn.eps}
\caption{\label{fig:hydro}
Schematic view of chemodynamical evolution of galaxies.
}
\end{center}
\end{figure}
We distribute this {\bf feedback} metal mass and energy from a star particle $j$
to a fixed number of neighbour gas particles $\ell$, $N_{\rm FB}$, weighted by the smoothing kernel as:
\begin{equation}\label{eq:fb-z}
\frac{d(Z_{i,k}\,m_{{\rm g},k})}{dt}= \sum_j \left[W_{kj} E_{Z,j}(t) / \sum_\ell^{N_{\rm FB}} W_{\ell j} \right]
\end{equation}
and
\begin{equation}\label{eq:fb-e}
\frac{du_k}{dt}= (1-f_{\rm kin}) \sum_j \left[W_{kj} E_{e,j}(t) / \sum_\ell^{N_{\rm FB}} W_{\ell j} \right]
\end{equation}
where $u_k$ denotes the internal energy of a gas particle $k$, and $f_{\rm kin}$ denotes the kinetic energy fraction, i.e., the fraction of energy that is distributed as a random velocity kick to the gas particle.
Note that in the first sum the number of neighbour star particles is not fixed to $N_{\rm FB}$.
To calculate these equations, the feedback neighbour search needs to be done twice in order to ensure proper
mass and energy conservation; first to compute the sum of weights for the
normalization, and a second time for the actual distribution.
The parameter $N_{\rm FB}$ determines the average energy distributed to gas particles; a large value of $N_{\rm FB}$ leads to inefficient feedback as the ejected energy radiatively cools away shortly, while a small value of $N_{\rm FB}$ results in a only small fraction of matter affected by feedback.
Alternatively, with good resolution in zoom-in simulations, feedback neighbors could be chosen within a fixed radius ($r_{\rm FB}$), which affects a number of observations including the MDF (Fig.\,12 of \citealt{kob11mw}).
The parameter $f_{\rm kin}$ has a great impact on radial metallicity gradients in galaxies, and $f_{\rm kin}=0$, i.e., purely thermal feedback, gives the best match to the observations (Fig.\,14 of \citealt{kob04}).
Note that various feedback methods are proposed such as
the stochastic feedback \citep[][used for EAGLE]{dal08} and the mechanical feedback \citep[][used for FIRE]{hop18}.
Eq.(\ref{eq:fb-z}) gives positive feedback that enhances radiative cooling, while Eq.(\ref{eq:fb-e}) gives negative feedback that suppresses star formation.
The mass ejection in Eq.(\ref{eq:fb-z}) never becomes zero, while the energy production in Eq.(\ref{eq:fb-e}) becomes small after 35 Myrs.
Therefore, it is not easy to control star formation histories depending on the mass/size of galaxies within this framework.
In particular, it is not possible to quench star formation in massive galaxies, since low-mass stars keep returning their envelope mass into the ISM for a long timescale, which will cool and keep forming stars.
Therefore, in order to reproduce observed properties of massive galaxies, additional feedback was required, and the discovery of the co-evolution of super-massive black holes and host galaxies provided a solution -- {\bf AGN feedback}.
Modelling of AGN feedback consists of (1) seed formation, (2) growth by mergers and gas accretion, and (3) thermal and/or kinetic feedback (see \citealt{tay14} for the details).
In summary, we introduced a seeding model where seed black holes originate from the formation of the first stars, which is different from the `standard' model by \citet{spr05agn} and from most large-scale hydrodynamical simulations.
In our AGN model,
seed black holes are generated if the metallicity of the gas cloud is zero ($Z=0$) and the density is higher than a threshold density ($\rho_{\rm g} > \rho_{\rm c}$).
The growth of black holes is the same as in other cosmological simulations, and is calculated with Bondi-Hoyle accretion, swallowing of ambient gas particles, and merging with other black holes.
Since we start from relatively small seeds, the black hole growth is driven by mergers at $z\gtsim 3$ (Fig.\,2 of \citealt{tay14}).
On a very small scale, it is not easy to merge two black holes, and an additional time-delay is applied in more recent simulations (P. Taylor et al. in prep.).
Proportional to the accretion rate, thermal energy is distributed to the surrounding gas, which is also the same as in many other simulations.
In more recent simulations, non-isotropic distribution of feedback area is used to mimic the small-scale jet \citep[e.g.,][]{dav19}.
There are a few parameters in our chemodynamical simulation code, but \citet{tay14} constrained the model parameters from observations, and determined the best parameter set: $\alpha=1$, $\epsilon_{\,\rm f}=0.25$, $M_{\rm seed}=h^{-1} 10^3M_\odot$, and $\rho_{\rm c}=0.1 h^2 m_{\rm H}\,\textrm{cm}^{-3}$.
Our black holes seeds are indeed the debris of the first stars although we explored a parameter space of $10^{1-5}M_\odot$.
This is not the only one channel for seeding, and the direct collapse of primordial gas and/or the collapse of dense stellar clusters ($\sim 10^5M_\odot$, \citealt{woo19}) are rarer but should be included in larger volume simulations.
Nonetheless, our model can successfully drive large-scale galactic winds from massive galaxies \citep{tay15b} and can reproduce many observations of galaxies with stellar masses of $\sim 10^{9-12}M_\odot$ \citep{tay15a,tay16,tay17}.
The movie of our fiducial run is available at \url{https://www.youtube.com/watch?v=jk5bLrVI8Tw}
\subsection{Extra-galactic archaeology}
\label{sec:n}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.35\textwidth]{average_vs_tau.eps}
\includegraphics[width=0.60\textwidth]{no_oh_all.eps}
\caption{(Left) Evolution of CNO abundance ratios in simulated disk galaxies with different formation timescales, taken in a cosmological simulation.
(Right) Evolution of the N/O--O/H relation of simulated galaxies in a cosmological simulation. The gray bar indicates the compilation of observational data.
}
\label{fig:cno}
\end{center}
\end{figure}
Elemental abundances and isotopic ratios provide additional constrains on the timescales of formation and evolution of galaxies. This extra-galactic archaeology will be possible in the near future with the James Webb Space Telescope (JWST), the wavelength coverage of which will allow us to measure CNO abundance simultaneously \citep{vin18a}.
At high redshifts, metallicities have been measured mainly with emission lines in star-forming galaxies, where Fe abundance is not accessible. Instead of $\alpha$/Fe ratios in Galactic archaeology, CNO abundances can be used for extra-galactic archaeology.
The left panel of Figure \ref{fig:cno} shows theoretical predictions of CNO abundance ratios for simulated disk galaxies that have different formation timescales. These galaxies are chosen from a cosmological simulation, which is run by a Gadget-3 based code that includes detailed chemical enrichment \citep{kob07}. In the simulation, C is mainly produced from low-mass stars ($\ltsim 4M_\odot$), N is from intermediate-mass stars ($\gtsim 4M_\odot$) as a primary process \citep{kob11agb}, and O is from massive stars ($\gtsim 13M_\odot$). C and N are also produced by massive stars; the N yield depends on the metallicity as a secondary process, and can be greatly enhanced by stellar rotation (as for F in \S \ref{sec:f}).
In the nearby universe, the N/O--O/H relation is known for stellar and ISM abundances, which shows a plateau ($\sim -1.6$) at low metallicities and a rapid increase toward higher metallicities. This was interpreted as the necessity of rotating massive stars by \citet{chi06}. However, this should be studied with hydrodynamical simulations including detailed chemical enrichment, and \citet{kob14mw} first showed the N/O--O/H relation in a chemodynamical simulation. \citet{vin18b} showed that both the global relation, which is obtained for average abundances of the entire galaxies, and the local relation, which is obtained for spatially resolved abundances from IFU data, can be reproduced by the inhomogeneous enrichment from AGB stars. Since N yield increases at higher metallicities, the global relation originates from the mass--metallicity relation of galaxies, while the local relation is caused by radial metallicity gradients within galaxies.
Moreover, the right panel of Figure \ref{fig:cno} shows a theoretical prediction on the time evolution of the N/O--O/H relation, where galaxies evolve along the relation. Recent observation with KMOS (KLEVER survey) confirmed a near redshift-invariant N/O-O/H relation \citep{hay22}.
{\bf Future prospects}:
It is possible to reproduce the observed N/O--O/H relation only with AGB stars and supernovae in our chemodynamical simulations. If we include N production from WR stars, the N/O plateau value at low metallicities becomes too high compared with observations.
The enrichment from the first stars is not included either, which can be significantly different because very massive stars ($\ge 140M_\odot$) explode as pair-instability supernovae without leaving any remnants, or massive stars ($\sim 25M_\odot$) explode as failed supernovae that leave relatively more massive black holes ($\sim 5M_\odot$). The former is based on theoretical prediction (but no observational evidence yet) and the latter is inferred from observational results \citep{nom13}.
It would be very useful if other elemental abundances (e.g., Ne, S) are measured in order to determine the chemical enrichment sources and constrain the formation histories of galaxies.
\subsection{Galactic archaeology}
\label{sec:mw}
Inhomogeneous enrichment in chemodynamical simulations leads to a paradigm shift on the chemical evolution of galaxies.
As in a real galaxy, i) the star formation history is not a simple function of radius, ii) the ISM is not homogeneous at any time, and iii) stars migrate \citep{vin20}, which are fundamentally different from one-zone or multi-zones GCE models.
As a consequence, (1) there is no age--metallicity relation, namely for stars formed in merging galaxies. It is possible to form extremely metal-poor stars from accretion of nearly primordial gas, or in isolated chemically-primitive regions, at a later time.
(2) Enrichment sources with long time-delays such as AGB stars, SNe Ia, and NSMs can appear at low metallicities. This is the reason why the N/O plateau is caused by AGB stars in \S \ref{sec:n}. However, this effect is not enough to reproduce the observed [Eu/(O,Fe)] ratios only with NSMs \citep{hay19}.
(3) There is a significant scatter in elemental abundance ratios at a given time/metallicity, as shown in Figs.\,16-18 of \citet{kob11mw}.
Figure \ref{fig:mwxfe} shows the frequency distribution of elemental abundance ratios in our chemodynamical zoom-in simulation of a Milky Way-type galaxy ($r\le100$ kpc). The colour indicates the number of stars in logarithmic scale. In order to compare the scatter, the range of y-axis is set to the same, from $-1.5$ to $+0.5$ for the elements except for He and Li.
In the simulation we know the ages and kinematics of star particles as well as the locations within the galaxy, which are used in the following summary:
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.95\textwidth]{snapshot_1000_xfefeh2-oball.eps}
\vspace*{-1mm}
\caption{Frequency distribution of elemental abundance ratios in a chemodynamical zoom-in simulation of a Milky Way-type galaxy.
The filled circles are observational data from high-resolution spectra, and the metal-rich and metal-poor contours show the HERMES-GALAH survey DR3 \citep{bud21} and the SkyMapper EMP survey \citep{yon21b}, respectively.
}
\label{fig:mwxfe}
\end{center}
\end{figure*}
\begin{itemize}
\item The Big Bang nucleosynthesis gives $Y=0.2449$.
He is also synthesized during stellar evolution and the He abundance increases up to $Y \sim 0.33$ showing a tight function of metallicity (see also \citealt{vin19}).
\item The initial Li abundance is set to be a theoretical value of the Big Bang nucleosynthesis, $A({\rm Li})=2.75$, although it is higher than the observed value of EMP stars.
The Li abundance increases by the production from AGB stars, but decreases at super-solar metallicities due to the destruction in AGB stars.
There is a large scatter from $A({\rm Li})\sim2.6$ to $0.3$ at [Fe/H] $\gtsim -1$.
\item The average [C/Fe] ratio is around $\sim 0$. There is only a small number of C-enhanced EMP stars with [C/Fe] $\ge0.7$ due to the inhomogeneous enrichment from AGB stars in our model. This number is much smaller than observed because faint supernovae are not included.
\item The [N/Fe] ratio shows a large scatter of $\pm0.5$ dex even at [Fe/H] $\sim -3$ due to the inhomogeneous enrichment, and it is possible to reproduce the observational data only with AGB stars, without rotating massive stars.
\item The $\alpha$ elements, O, Mg, Si, S, and Ca, show similar trends, a plateau with a small scatter at [Fe/H] $\ltsim -1$, and then bimodal decreasing trends both caused by the delayed Fe enrichment from SNe Ia. The high-$\alpha$ stars tend to have high $v/\sigma$ and old ages, forming the thick disk. On the other hand, the low $\alpha$ stars are rotationally supported and relatively younger, belonging to the thin disk.
The thick disk reaches only [Fe/H] $\sim 0$, while thin disk expends to [Fe/H] $\sim 0.2$. The stars with [Fe/H] $>0.2$ in this figure are located in the bulge region ($r\le 1$ kpc in this paper).
\item The odd-$Z$ elements, Na, Al, and Cu, show bimodal increasing trends from [Fe/H] $\sim -2$ to higher metallicities, which are predominantly caused by the metallicity dependent yields of core-collapse supernovae.
The higher [(Na,Al,Cu)/Fe] trend is composed of the high $\alpha$ stars in the thick disk, while the lower [(Na,Al,Cu)/Fe] trend is made of the low $\alpha$ stars in the thin disk; these differences are caused by SNe Ia.
Moreover, [Na/Fe] shows a greater increase with a steeper slope for the thick disk than that for the thin disk, which is due to an additional contribution from AGB stars. Such a difference in the slopes is not seen for Al and Cu.
\item The [Cr/Fe] ratio is almost constant at $\sim 0$ over all ranges of metallicities. In observations, EMP stars tend to show [Cr/Fe] $<0$ values affected by the NLTE effect, but Cr II observations show [Cr/Fe] $\sim 0$ (K06). No decrease of [Cr/Fe] is caused by SNe Ia either if we use the latest SN Ia yields \citep{kob20ia}.
\item In contrast to the [$\alpha$/Fe] ratios, [Mn/Fe] shows a plateau value of $\sim -0.5$ at [Fe/H] $\ltsim -1$, and then bimodal increasing trends because Ch-mass SNe Ia produce more Mn than Fe (K06). Unlike one-zone models, [Mn/Fe] does not reach a very high value ($\sim +0.5$) in chemodynamical simulations because not all SN Ia progenitors have high metallicities due to the inhomogeneity. The inclusion of sub-Ch mass SNe Ia would worsen the situation.
\item Co and Zn can be enhanced by higher energy explosions (K06). There are two plateaus with [Zn/Fe] $\sim 0.2$ for the thick disk and [Zn/Fe] $\sim 0$ for the thin disk. At [Fe/H] $>0.2$, [Zn/Fe] shows a rapid increase, which is seen only for the bulge stars in our model.
Very similar features are seen also for Co, although Co is under-produced overall, which is due to the lack of multi-dimensional effect in the nucleosynthesis yields \citep{sne16}.
\item The first-peak s-process elements (Sr, Y, and Zr) are produced by ECSNe and AGB stars. Sr and Y do not show a bimodality, while Zr shows a very similar trend as for the $\alpha$ elements.
\item The second-peak s-process elements (Ba, La, and Ce) show a very large scatter of $1.5$ dex especially at low metallicities. These elements are first produced by MRSNe/MRHNe in our model with a floor value of [(Ba,La,Ce)/Fe] $\sim -1$, and are greatly enhanced by AGB stars. The contribution from AGB stars appears from [Fe/H] $\sim -2$; [(Ba,La,Ce)/Fe] ratios reach $\sim 0$ at [Fe/H] $\sim -1$, peaks at [Fe/H] $\sim 0$, then decrease due to SNe Ia. The decrease is significant in the bulge stars at [Fe/H] $>0.2$.
\item The r-process elements (Nd, Eu, Dy, Er, Au, and Th in the figure) do show a bimodality in the model with MRSNe/MRHNe. The average [Nd/Fe] is about $\sim 0$, while Eu, Dy, Er, and Th behave very similar to the $\alpha$ elements.
As discussed in \S \ref{sec:au}, Au is underabundant overall, but the thick disk stars are systematically more gold-rich than for the thin disk stars!
\item Pb is one of the third-peak s-process elements, and the distribution at low metallicities are similar to the second-peak s-process elements. However, [Pb/Fe] shows a steeper decrease from [Fe/H] $\sim -1$ to higher metallicities, due to the metallicity dependence; Pb is produced more from metal-poor AGB stars.
\end{itemize}
Compared with observational data, the scatter of s-process elements seem too large around [Fe/H] $\sim -1$, which also supports the weak-s process rotating massive stars.
The elemental abundance distributions depend on the locations within galaxies, as shown in Fig.\,1 of \citet{kob16mw} for a different chemodynamical simulation.
These distributions should be compared with non-biased, homogeneous dataset of observational data and the comparison to APOGEE DR16 was shown in \citet{vin20}.
More analysis will be shown in Kobayashi (2022, in prep).
The movie of our fiducial run is available at \url{https://star.herts.ac.uk/~chiaki/works/Aq-C-5-kro2.mpg}
\vspace*{-2mm}
\section{Conclusions and Discussion}
Thanks to the long-term collaborations between nuclear and astrophysics, we have good understanding on the origin of elements (except for the elements around Ti and a few neutron-capture elements such as Au).
Inhomogeneous enrichment is extremely important for interpreting the elemental abundance trends (\S \ref{sec:mw}). It can reproduce the observed N/O--O/H relation only with AGB stars and supernovae (\S \ref{sec:n}), but not the observed r-process abundances only with NSMs; an r-process associated with core-collapse supernovae such as magneto-rotational hypernovae is required, although the explosion mechanism is unknown (\S \ref{sec:au}).
It is necessary to run chemodynamical simulations from cosmological initial conditions, including detailed chemical enrichment.
Theoretical predictions depend on input stellar physics, and the effects of stellar mass loss due to rotation and/or binary interaction should be investigated further.
The importance of WR stars is indicated by the high F abundance in $z\sim 4.4$ galaxy (\S \ref{sec:f}), as well as for some isotopic ratios and Ba and Pb abundances around [Fe/H] $\sim -1$ in the Milky Way (\S \ref{sec:iso}).
If WR stars are producing heavy elements on a very short timescale, it might be hard to find very metal-poor or Population III (and dust-free) galaxies at very high redshifts even with JWST.
Finally, mass loss from these stars will change the dust production as well, and thus it is also important to calculate element-by-element dust formation, growth, and destruction, as well as the detailed chemical enrichment.
Galactic archaeology is a powerful approach for reconstructing the formation history of the Milky Way and its satellite galaxies.
APOGEE and HERMES-GALAH surveys have provided homogeneous datasets of many elemental abundances that can be statistically compared with chemodynamical simulations. Future surveys with WEAVE and 4MOST will provide more.
Having said that, the number of EMP stars will not be increased so much in these surveys, and a target survey such as the SkyMapper EMP survey is also needed, in particular for constraining the early chemical enrichment from the first stars.
Reflecting the difference in the formation timescale, elemental abundances depend on the location within galaxies.
Although this dependence has been explored toward the Galactic bulge by APOGEE, the dependence at the outer disk is still unknown, which requires 8m-class telescopes such as the PFS on Subaru telescope. Despite the limited spectral resolution of the PFS, $\alpha$/Fe and a small number of elements will be available.
The PFS will also be able to explore the $\alpha$/Fe bimodality in M31; it is not yet known if M31 has a similar $\alpha$/Fe dichotomy or not.
The next step will be to apply the Galactic archaeology approach to external or distant galaxies. Although it became possible to map metallicity, some elemental abundances, and kinematics within galaxies with IFU, the sample and spacial resolution are still limited even with JWST.
Integrated physical quantities over galaxies, or stacked quantities at a given mass bin, will also be useful, which can be done with the same instruments developed for Galactic archaeology (although optimal spectral resolutions and wavelength coverages are different).
In addition, ALMA has opened a new window for elemental abundances and isotopic ratios in high-redshift galaxies.
It is very important to understand stellar physics when these observational data are translated into the formation timescales or physical processes of galaxies.
For example, the famous [$\alpha$/Fe] ratio is not a perfect clock.
Analysing low-$\alpha$ EMP stars, \citet{kob14} summarized various reasons that cause low $\alpha$/Fe ratios: (1) SNe Ia, (2) less-massive core-collapse supernovae ($\ltsim 20M_\odot$), which become more important with a low star formation rate (3) hypernovae, although the majority of hypernovae are expected to give normal [$\alpha$/Fe] ratios ($\sim 0.4$), and (4) pair-instability supernovae, which could be very important in the early universe.
Therefore, low-$\alpha$ stars are not necessarily enriched by SNe Ia in a system with a long formation timescale.
It is necessary to also use other elemental abundances (namely, Mn and neutron-capture elements) or isotopic ratios, with higher resolution ($>40,000$) multi-object spectroscopy on 8m class telescopes (e.g., cancelled WFMOS or planned MSE).
\vspace*{-1mm}
\acknowledgement
I thank my collaborators, D. Yong, M. Franco, F. Vincenzo, P. Taylor, A. Karakas, M. Lugaro, N. Tominaga, S.-C. Leung, M. Ishigaki, and K. Nomoto, and V. Springel for providing Gadget-3.
I acknowledge funding from the UK Science and Technology Facility Council through grant ST/M000958/1, ST/R000905/1, ST/V000632/1.
The work was also funded by a Leverhulme Trust Research Project Grant on ``Birth of Elements''.
\vspace*{-4mm}
| 87b643c49e792244c348ee9aa6df39bc0e02d140 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Method}
We conduct our work according to the methodology used by~\cite{Mentis2014}. The examples presented in this paper
come from the
Champalimaud Foundation in Lisbon, PT~\cite{Zorzal2020}.
At that site, we
observed laparoscopic surgeries on one of the Foundation's surgical rooms and
talked to surgeons before and after the procedures, who explained
what was about to happen or what took place. Also, during the surgery, nurses
provided insight into the several stages of the surgery, or what was happening at that time. We also had the opportunity to ask questions of the surgeons at appropriate moments during the surgery. Besides that, we video recorded for further analysis in addition to our field notes.
A total of five (5) laparoscopic surgeries were observed for a total of approximately 10 hours.
The observed cases included the surgical team is composed of mostly male members, with only one female surgeon in it, and their ages range from 34 to 42 years of age.
During surgery, there are at least six people involved in the procedure
A head surgeon, who coordinates the entire procedure, 1 or 2 assistant surgeons, who mostly observe but also participate in parts of the surgery, a nurse solely responsible for passing the surgeons tools they may require throughout the operation, an anesthetist keeping track of the patient's vital signs, a nurse supporting the anesthetist and a circulating nurse. Additionally, a senior surgeon may come in and serve as an advisor, providing insight and making remarks about what is being seen on camera.
\bgroup
We collected images and videos using smartphones and tablets, as well as the notes made during the observations. The inductive bottom-up approach to data analysis was used in which the authors analyzed their field notes and videos. The findings and the discussion are presented in the subsequent section.
\section{Results and Discussion}
Performing user and task analysis allowed us to better understand the existing problems in the procedure of laparoscopy, while identifying several constraints and design requirements, which a solution has to follow in order to address those problems. For the following, we will discuss the design requirements identified through the analysis. Problems statements, requirements, and the proposed design solutions also are summarized in Table~\ref{table:problem_requirements_solution}.
\begin{table*}[htb]
\caption{Problem statements, design requirements and design solution for a prototype that supports laparoscopy}\label{table:problem_requirements_solution}
\begin{tabular}{|p{0.31\textwidth}|p{0.31\textwidth}|p{0.31\textwidth}|}
\hline
{\bfseries \small Problem Statements} & {\bfseries \small Design Requirements} & {\bfseries \small Design Solution}\\
\hline
\small Visualizing the laparoscopic video during extended periods of time is exhausting for the neck. & \small The solution should allow the user to adopt more comfortable neck postures instead of forcing the user to look to the side to see what the other surgeons are seeing. & \small Following display: The laparoscopic video follows user head movement, so users can look around and assume a neck posture that is more comfortable for them. \\
\hline
\small Current interactions surgeons have, such as pointing or consulting patient data, require them to let go of their tools, which interrupts the procedure. & \small Surgeons should have hands-free interactions in order to operate in an uninterrupted fashion. & \small Hands-free interaction: Every interaction is either done with the head or using the feet. \\
\hline
\small Browsing patient data interoperatively takes too long because it requires to call in an assistant, who browses the images for the surgeon. & \small Users should be able to look at patient data by themselves, without interrupting and adding extra time to the surgery. & \small Patient data image browser: users can look to the side to see and browse magnetic resonance images from the patient. \\
\hline
\small Users may have to move around the patient in order to adopt better positions to hold their tools. & \small Interaction using the foot should not rely on pedals, as these would need to be moved around to cope with user movement. & \small Foot browsing: Users can use the foot to navigate the patient images, rotating it on its heel to change images faster or slower. \\
\hline
\small Pointing is unclear and ambiguous: different users have different interpretations of where a surgeon is pointing at. & \small Users should be able to point precisely and understand where other users are pointing at, regardless of position in the operating room. & \small Pointing reticle: users can place a reticle on both laparoscopic video and patient images, controlling it with head motion. This cursor is visible on other users’ headsets.\\
\hline
\small Surgeons operate in a crowded area, as they are usually very close together. & \small Augmented space should present information close to the surgeon to prevent it from appearing intersected with a colleague. & \small Close quarters: Positioning of interface elements is no further than at an elbow's reach.\\
\hline
\end{tabular}
\end{table*}
\subsection{Following display}
Laparoscopy is an intensive process, not just mentally but physically as well. The procedure is already very demanding in itself due to surgeons having to expend extra mental effort thanks to a lack of hand-eye coordination that is caused by indirect visualization (Fig.~\ref{fig:neck}). That effort extends to the physical plane when we consider that they have look at the screen all the time, which places a continuous strain on their necks. Laparoscopy currently faces the glaring problem of monitor positioning. During surgery, screens are usually placed far away and at an uncomfortable angle, causing neck and eye strain over the course of a surgery, especially if it drags for longer periods of time. Given this, it was important to allow the surgeons some freedom in how they want to see the video, and then the video, while visible, should follow user head movements so users do not have to reposition it in the augmented space, should they feel the need to assume another posture with the neck.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.65\textwidth,keepaspectratio]{Figure_3.png}
\caption{All doctors look to the same screen and, sometimes, this means having to assume an uncomfortable position. Visualizing the laparoscopic video during extended periods of time is exhausting for the neck.}
\label{fig:neck}
\end{figure*}
The surgeons have to look at screens placed outside the field of operation, which results in discomfort~\cite{Batmaz2017}, affecting the surgeon's efficiency due to a disconnect between the visual and motor axis, because the surgeon cannot look at the instruments or hands and the field of surgery simultaneously. To be successful, more training is required to adapt to this condition, as extra mental effort must be applied~\cite{Leite2016}.
In addition, almost all
display screens are limited in sense that they do not support techniques to improve communication and visual collaboration with the rest of the surgical team~\cite{HenryFuchs1MarkA.Livingston1RameshRaskar1DnardoColucci11963,HMentis2019}.
Muratore et al.~\cite{Muratore2007} suggest for the future,
the ideal display system would be a 3D high definition image HMD,
citing the comfort of looking at the endoscopic image in any preferred head position, improving ergonomics and reducing neck strain. The use of an HMD is also seen as beneficial in the sense that it alleviates equipment clutter in the operating room. It is further noted the usefulness of individualized image manipulation features like zooming, which allows each surgeon to see the endoscopic video in the way they find most comfortable.
Also, the works of~\cite{Walczak2015,Maithel2005,Batmaz2017,Kihara2012} seem to support the usage of a HMD for laparoscopic surgery, with the video following the user’s head movements. With this, users can assume their preferred head position instead of being forced to look sideways in order to see the video.
\subsection{Hands-free interaction}
From our field observations, we noticed that surgeons place down their tools to perform some secondary tasks, which interrupts the procedure (Fig.~\ref{fig:losing}). We suggest that projects avoid this type of situation, with a completely hands-free approach, using both head gaze and foot movement as sources of input.
\begin{figure*}[!bhtpb]
\centering
\includegraphics[width=1\textwidth,keepaspectratio]{Figure_4.png}
\caption{Surgeons can't look at their hands, thus losing hand-eye coordination.}
\label{fig:losing}
\end{figure*}
Other studies~\cite{Kim2017,Jayender}
looked at
using head movements, and gaze
to select targets, especially
when the content follows head movements~\cite{Grinshpoon}. A more elaborate approach takes the form of the eye gazing~\cite{Esteves2015,Velloso2016}, which was well-received by users, but may not transition well onto the surgical operating field: these controls would have to be displayed continuously and right in front of the user, unlike in the presented works, which could be distractive for users, but more importantly, it would take valuable space from the HMD's already limited field of view. In terms of feet, two different approaches emerge: using a foot pedal as a means to activate a selected control~\cite{Jayender} and using foot movement to select and activate controls~\cite{Muller2019}. After comparing the two, we conclude using foot movement would be a more flexible choice, as it does not rely on extra hardware that is situated in a given position in space.
\subsection{Patient data image browser and foot browsing}
Consulting patient data intra-operatively, such as MRIs and computed tomographies, may be unfeasible. On the one hand, the data are extensive and may require some time to identify the required set. On the other, surgeons must abandon the operating table to sit at a computer and browse the desired images. The surgeon then gives directions on where to look and when to stop while an assistant handles the computer. Surgeons usually do not browse the images themselves when sterilized. Indeed each interaction with non-sterile equipment such as the keyboard and mouse would entail a new sterilization procedure, wasting additional resources. This has been discussed in \cite{ohara-2014,Lopes-2019,Johnson2011}. This is a design opportunity for AR as discussed in \cite{Zorzal2019}.
In the work of
\cite{Muratore2007}, the authors emphasize the importance of the HMD, stating preoperative imaging could be individually manipulated through its use, as well as grant surgeons extra comfort by allowing them to see the laparoscopic image regardless of head positioning. Hands-free interaction is again considered, with the suggestion of using foot pedals instead. Also discussed is the issue of paradoxical imaging, which occurs when the camera faces the surgeon, causing movements with the tool to appear inverted compared to the hand movements. However, in an interview with the surgeons of the
Champalimaud Foundation, this did not appear to be an issue, since the surgeon can move around the
operating table, which ensures the camera always faces the opposite direction. This freedom to move around also impacts the practicality of using foot pedals to ensure hands-free interaction, as the authors suggest, as the surgeon would have to either have the same pedals on multiple sides, or move the pedals around. In this case, exploring foot movement, as proposed by
\cite{Muller2019}, could be more useful
\subsection{Pointing reticle and close quarters}
Although they are close quarters, surgeons also currently face problems in communication. In fact, according to the inquired surgeons, just as they complain about difficulty in maintaining proper posture, so do they complain about not being able to let other surgeons know what part of the video they are pointing at, or to understand what others are pointing at as well. Several works have approached this by looking at proxemics \cite{Mentis2012} and embodied vision~\cite{Mentis2013}.
\begin{figure*}[!htpb]
\centering
\includegraphics[width=0.65\textwidth,keepaspectratio]{Figure_6.png}
\caption{Doctor point at the screen to communicate. Communication is unclear and ambiguous: different users have different interpretations of where a surgeon is pointing at.}
\label{fig:pointing2}
\end{figure*}
The instructor can point at the screen
for the other surgeons to understand what anatomical structure he/she is referring to and
use gestures for others to understand the motion of the tools better and envision cutting lines. Sometimes, pointing can also be done with the tools themselves, but even though it may be effective, it is not always correct because if both hands are occupied, it implies letting go of a structure to point with the tool or asking someone else to hold it. Additionally, pointing from a distance with the hand is ambiguous at best, as there is no clear way to tell where precisely a surgeon is pointing at, as can be seen in Fig.~\ref{fig:pointing2}.
For Prescher et al.~\cite{Prescher}, the impact of the pointer in a real operating scenario may be lessened because target selection is not random but rather contextual, meaning that the following targets may be located through the description of what is being displayed on-screen. Also, the pointer is embedded in the laparoscope. Thus the camera must be displaced to move the pointer, causing the view plane to change and forcing the surgeon to readjust to the new perspective, losing perceived depth. Therefore, it would be more useful if the cursor moved independently of the camera, controlled via gestures or head tracking for a hands-free approach, as suggested by 3D interactions performed above a table~\cite{mendes2014,mendes2016}.
\section{Conclusion}
This paper presents the requirements and design solutions for laparoscopy through a user and task analysis. During a task analysis, we map out the sequence of activities surgeons go through and the actions required to achieve that goal. Drawing on observations and analysis of video recordings of laparoscopic surgeries, we identify several constraints and design requirements, which a solution will have to follow in order to address those problems. These requirements propose to inform the design solutions towards improved surgeons' comfort and make the surgical procedure less laborious.
\begin{comment}
\section*{Summary points}
\noindent
\setlength{\fboxsep}{10pt}
\fbox{\parbox{0.9\columnwidth}{\textbf{What is already known on the topic?}
\begin{itemize}
\setlength\itemsep{0pt}
\item Laparoscopy suffers some problems that can cause the surgical procedure more laborious.
\item A variety of different problems influence the performance of surgeons in laparoscopic procedures.
\end{itemize}
\textbf{What does this study add to our knowledge?}
\begin{itemize}
\setlength\itemsep{0pt}
\item Identification of problems related to design for laparoscopy.
\item Design requirements propose to inform the design solutions towards improved surgeons' comfort and make the surgical procedure less laborious.
\end{itemize}
}}
\end{comment}
\section*{Acknowledgments}
This work was supported by national funds through FCT, Fundação para a Ciência e a Tecnologia, under project UIDB/50021/2020. The authors would like to thank the Champalimaud Foundation for its collaboration and support in the development of the study.
\bibliographystyle{abbrv-doi}
\section{Introduction}
This template is for papers of VGTC-sponsored conferences which are \emph{\textbf{not}} published in a special issue of TVCG.
\section{Using the Style Template}
\begin{itemize}
\item If you receive compilation errors along the lines of ``\texttt{Package ifpdf Error: Name clash, \textbackslash ifpdf is already defined}'' then please add a new line ``\texttt{\textbackslash let\textbackslash ifpdf\textbackslash relax}'' right after the ``\texttt{\textbackslash documentclass[journal]\{vgtc\}}'' call. Note that your error is due to packages you use that define ``\texttt{\textbackslash ifpdf}'' which is obsolete (the result is that \texttt{\textbackslash ifpdf} is defined twice); these packages should be changed to use ifpdf package instead.
\item The style uses the hyperref package, thus turns references into internal links. We thus recommend to make use of the ``\texttt{\textbackslash autoref\{reference\}}'' call (instead of ``\texttt{Figure\~{}\textbackslash ref\{reference\}}'' or similar) since ``\texttt{\textbackslash autoref\{reference\}}'' turns the entire reference into an internal link, not just the number. Examples: \autoref{fig:sample} and \autoref{tab:vis_papers}.
\item The style automatically looks for image files with the correct extension (eps for regular \LaTeX; pdf, png, and jpg for pdf\LaTeX), in a set of given subfolders (figures/, pictures/, images/). It is thus sufficient to use ``\texttt{\textbackslash includegraphics\{CypressView\}}'' (instead of ``\texttt{\textbackslash includegraphics\{pictures/CypressView.jpg\}}'').
\item For adding hyperlinks and DOIs to the list of references, you can use ``\texttt{\textbackslash bibliographystyle\{abbrv-doi-hyperref-narrow\}}'' (instead of ``\texttt{\textbackslash bibliographystyle\{abbrv\}}''). It uses the doi and url fields in a bib\TeX\ entry and turns the entire reference into a link, giving priority to the doi. The doi can be entered with or without the ``\texttt{http://dx.doi.org/}'' url part. See the examples in the bib\TeX\ file and the bibliography at the end of this template.\\[1em]
\textbf{Note 1:} occasionally (for some \LaTeX\ distributions) this hyper-linked bib\TeX\ style may lead to \textbf{compilation errors} (``\texttt{pdfendlink ended up in different nesting level ...}'') if a reference entry is broken across two pages (due to a bug in hyperref). In this case make sure you have the latest version of the hyperref package (i.\,e., update your \LaTeX\ installation/packages) or, alternatively, revert back to ``\texttt{\textbackslash bibliographystyle\{abbrv-doi-narrow\}}'' (at the expense of removing hyperlinks from the bibliography) and try ``\texttt{\textbackslash bibliographystyle\{abbrv-doi-hyperref-narrow\}}'' again after some more editing.\\[1em]
\textbf{Note 2:} the ``\texttt{-narrow}'' versions of the bibliography style use the font ``PTSansNarrow-TLF'' for typesetting the DOIs in a compact way. This font needs to be available on your \LaTeX\ system. It is part of the \href{https://www.ctan.org/pkg/paratype}{``paratype'' package}, and many distributions (such as MikTeX) have it automatically installed. If you do not have this package yet and want to use a ``\texttt{-narrow}'' bibliography style then use your \LaTeX\ system's package installer to add it. If this is not possible you can also revert to the respective bibliography styles without the ``\texttt{-narrow}'' in the file name.\\[1em]
DVI-based processes to compile the template apparently cannot handle the different font so, by default, the template file uses the \texttt{abbrv-doi} bibliography style but the compiled PDF shows you the effect of the \texttt{abbrv-doi-hyperref-narrow} style.
\end{itemize}
\section{Bibliography Instructions}
\begin{itemize}
\item Sort all bibliographic entries alphabetically but the last name of the first author. This \LaTeX/bib\TeX\ template takes care of this sorting automatically.
\item Merge multiple references into one; e.\,g., use \cite{Max:1995:OMF,Kitware:2003} (not \cite{Kitware:2003}\cite{Max:1995:OMF}). Within each set of multiple references, the references should be sorted in ascending order. This \LaTeX/bib\TeX\ template takes care of both the merging and the sorting automatically.
\item Verify all data obtained from digital libraries, even ACM's DL and IEEE Xplore etc.\ are sometimes wrong or incomplete.
\item Do not trust bibliographic data from other services such as Mendeley.com, Google Scholar, or similar; these are even more likely to be incorrect or incomplete.
\item Articles in journal---items to include:
\begin{itemize}
\item author names
\item title
\item journal name
\item year
\item volume
\item number
\item month of publication as variable name (i.\,e., \{jan\} for January, etc.; month ranges using \{jan \#\{/\}\# feb\} or \{jan \#\{-{}-\}\# feb\})
\end{itemize}
\item use journal names in proper style: correct: ``IEEE Transactions on Visualization and Computer Graphics'', incorrect: ``Visualization and Computer Graphics, IEEE Transactions on''
\item Papers in proceedings---items to include:
\begin{itemize}
\item author names
\item title
\item abbreviated proceedings name: e.\,g., ``Proc.\textbackslash{} CONF\_ACRONYNM'' without the year; example: ``Proc.\textbackslash{} CHI'', ``Proc.\textbackslash{} 3DUI'', ``Proc.\textbackslash{} Eurographics'', ``Proc.\textbackslash{} EuroVis''
\item year
\item publisher
\item town with country of publisher (the town can be abbreviated for well-known towns such as New York or Berlin)
\end{itemize}
\item article/paper title convention: refrain from using curly brackets, except for acronyms/proper names/words following dashes/question marks etc.; example:
\begin{itemize}
\item paper ``Marching Cubes: A High Resolution 3D Surface Construction Algorithm''
\item should be entered as ``\{M\}arching \{C\}ubes: A High Resolution \{3D\} Surface Construction Algorithm'' or ``\{M\}arching \{C\}ubes: A high resolution \{3D\} surface construction algorithm''
\item will be typeset as ``Marching Cubes: A high resolution 3D surface construction algorithm''
\end{itemize}
\item for all entries
\begin{itemize}
\item DOI can be entered in the DOI field as plain DOI number or as DOI url; alternative: a url in the URL field
\item provide full page ranges AA-{}-BB
\end{itemize}
\item when citing references, do not use the reference as a sentence object; e.\,g., wrong: ``In \cite{Lorensen:1987:MCA} the authors describe \dots'', correct: ``Lorensen and Cline \cite{Lorensen:1987:MCA} describe \dots''
\end{itemize}
\section{Example Section}
Lorem\marginpar{\small You can use the margins for comments while editing the submission, but please remove the marginpar comments for submission.} ipsum dolor sit amet, consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam
voluptua~\cite{Kitware:2003,Max:1995:OMF}. At vero eos et accusam et
justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet. Lorem ipsum dolor sit
amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor
invidunt ut labore et dolore magna aliquyam erat, sed diam
voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est.
\section{Exposition}
Duis autem vel eum iriure dolor in hendrerit in vulputate velit esse
molestie consequat, vel illum dolore eu feugiat nulla facilisis at
vero eros et accumsan et iusto odio dignissim qui blandit praesent
luptatum zzril delenit augue duis dolore te feugait nulla
facilisi. Lorem ipsum dolor sit amet, consectetuer adipiscing elit,
sed diam nonummy nibh euismod tincidunt ut laoreet dolore magna
aliquam erat volutpat~\cite{Kindlmann:1999:SAG}.
\begin{equation}
\sum_{j=1}^{z} j = \frac{z(z+1)}{2}
\end{equation}
Lorem ipsum dolor sit amet, consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam
et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet.
\subsection{Lorem ipsum}
Lorem ipsum dolor sit amet (see \autoref{tab:vis_papers}), consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam
et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet. Lorem ipsum dolor sit
amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor
invidunt ut labore et dolore magna aliquyam erat, sed diam
voluptua. At vero eos et accusam et justo duo dolores et ea
rebum.
\begin{table}[tb]
\caption{VIS/VisWeek accepted/presented papers: 1990--2016.}
\label{tab:vis_papers}
\scriptsize%
\centering%
\begin{tabu}{%
r%
*{7}{c}%
*{2}{r}%
}
\toprule
year & \rotatebox{90}{Vis/SciVis} & \rotatebox{90}{SciVis conf} & \rotatebox{90}{InfoVis} & \rotatebox{90}{VAST} & \rotatebox{90}{VAST conf} & \rotatebox{90}{TVCG @ VIS} & \rotatebox{90}{CG\&A @ VIS} & \rotatebox{90}{VIS/VisWeek} \rotatebox{90}{incl. TVCG/CG\&A} & \rotatebox{90}{VIS/VisWeek} \rotatebox{90}{w/o TVCG/CG\&A} \\
\midrule
2016 & 30 & & 37 & 33 & 15 & 23 & 10 & 148 & 115 \\
2015 & 33 & 9 & 38 & 33 & 14 & 17 & 15 & 159 & 127 \\
2014 & 34 & & 45 & 33 & 21 & 20 & & 153 & 133 \\
2013 & 31 & & 38 & 32 & & 20 & & 121 & 101 \\
2012 & 42 & & 44 & 30 & & 23 & & 139 & 116 \\
2011 & 49 & & 44 & 26 & & 20 & & 139 & 119 \\
2010 & 48 & & 35 & 26 & & & & 109 & 109 \\
2009 & 54 & & 37 & 26 & & & & 117 & 117 \\
2008 & 50 & & 28 & 21 & & & & 99 & 99 \\
2007 & 56 & & 27 & 24 & & & & 107 & 107 \\
2006 & 63 & & 24 & 26 & & & & 113 & 113 \\
2005 & 88 & & 31 & & & & & 119 & 119 \\
2004 & 70 & & 27 & & & & & 97 & 97 \\
2003 & 74 & & 29 & & & & & 103 & 103 \\
2002 & 78 & & 23 & & & & & 101 & 101 \\
2001 & 74 & & 22 & & & & & 96 & 96 \\
2000 & 73 & & 20 & & & & & 93 & 93 \\
1999 & 69 & & 19 & & & & & 88 & 88 \\
1998 & 72 & & 18 & & & & & 90 & 90 \\
1997 & 72 & & 16 & & & & & 88 & 88 \\
1996 & 65 & & 12 & & & & & 77 & 77 \\
1995 & 56 & & 18 & & & & & 74 & 74 \\
1994 & 53 & & & & & & & 53 & 53 \\
1993 & 55 & & & & & & & 55 & 55 \\
1992 & 53 & & & & & & & 53 & 53 \\
1991 & 50 & & & & & & & 50 & 50 \\
1990 & 53 & & & & & & & 53 & 53 \\
\midrule
\textbf{sum} & \textbf{1545} & \textbf{9} & \textbf{632} & \textbf{310} & \textbf{50} & \textbf{123} & \textbf{25} & \textbf{2694} & \textbf{2546} \\
\bottomrule
\end{tabu}%
\end{table}
\subsection{Mezcal Head}
Lorem ipsum dolor sit amet (see \autoref{fig:sample}), consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam
et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet.
\subsubsection{Duis Autem}
Lorem ipsum dolor sit amet, consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam
et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet. Lorem ipsum dolor sit
amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor
invidunt ut labore et dolore magna aliquyam erat, sed diam
voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est. Lorem
ipsum dolor sit amet.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{paper-count-w-2015-new}
\caption{A visualization of the 1990--2015 data from \autoref{tab:vis_papers}. The image is from \cite{Isenberg:2017:VMC} and is in the public domain.}
\label{fig:sample}
\end{figure}
\subsubsection{Ejector Seat Reservation}
Duis autem~\cite{Lorensen:1987:MCA}\footnote{The algorithm behind
Marching Cubes \cite{Lorensen:1987:MCA} had already been
described by Wyvill et al. \cite{Wyvill:1986:DSS} a year
earlier.} vel eum iriure dolor in hendrerit
in vulputate velit esse molestie consequat,\footnote{Footnotes
appear at the bottom of the column.} vel illum dolore eu
feugiat nulla facilisis at vero eros et accumsan et iusto odio
dignissim qui blandit praesent luptatum zzril delenit augue duis
dolore te feugait nulla facilisi. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit, sed diam nonummy nibh euismod tincidunt
ut laoreet dolore magna aliquam erat volutpat.
\paragraph{Confirmed Ejector Seat Reservation}
Ut wisi enim ad minim veniam, quis nostrud exerci tation ullamcorper
suscipit lobortis nisl ut aliquip ex ea commodo
consequat~\cite{Nielson:1991:TAD}. Duis autem vel eum iriure dolor in
hendrerit in vulputate velit esse molestie consequat, vel illum dolore
eu feugiat nulla facilisis at vero eros et accumsan et iusto odio
dignissim qui blandit praesent luptatum zzril delenit augue duis
dolore te feugait nulla facilisi.
\paragraph{Rejected Ejector Seat Reservation}
Ut wisi enim ad minim veniam, quis nostrud exerci tation ullamcorper
suscipit lobortis nisl ut aliquip ex ea commodo consequat. Duis autem
vel eum iriure dolor in hendrerit in vulputate velit esse molestie
\section{Conclusion}
Lorem ipsum dolor sit amet, consetetur sadipscing elitr, sed diam
nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat,
sed diam voluptua. At vero eos et accusam et justo duo dolores et ea
rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem
ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur
sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et
dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam
et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea
takimata sanctus est Lorem ipsum dolor sit amet. Lorem ipsum dolor sit
amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor
invidunt ut labore et dolore magna aliquyam erat, sed diam
voluptua. At vero eos et accusam et justo duo dolores et ea
rebum.
\acknowledgments{
The authors wish to thank A, B, and C. This work was supported in part by
a grant from XYZ.}
\bibliographystyle{abbrv-doi}
| a2f4f799b3327d885654aba7e32b209950b41764 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{Heterogeneous graph}
\label{sec:preliminaries:hg}
Heterogeneous graphs are an important abstraction for modeling the relational data of multi-modal systems.
Formally, a heterogeneous graph is defined as $\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathcal{T}, \mathcal{R})$ where
the node set $\mathcal{V}: = \{1,\ldots,|\mathcal{V}|\}$;
the edge set $\mathcal{E}$ consisting of ordered tuples $e_{ij}:=(i, j)$ with $i, j\in\mathcal{V}$, where $e_{ij}\in\mathcal{E}$ iff an edge exists from $i$ to $j$;
the set of node types $\mathcal{T}$ with associated map $\tau:\mathcal{V}\mapsto\mathcal{T}$; the set of relation types $\mathcal{R}$ with associated map $\phi:\mathcal{E}\mapsto\mathcal{R}$.
This flexible formulation allows directed, multi-type edges.
We additionally assume the existence of a node feature vector $x_i\in\mathcal{X}_{\tau(i)}$ for each $i\in\mathcal{V}$, where $\mathcal{X}_{t}$ is a feature space specific to nodes of type $t$ .
This allows $\mathcal{G}$ to represent nodes with different feature modalities such as images, text, locations, or booleans.
Note that these modalities are not necessarily exclusive (e.g.\ two node types $s,t$ might share the same feature space, $\mathcal{X}_s = \mathcal{X}_t$).
\begin{figure*}[t!]
\centering
\vspace{-4mm}
\subfigure[Toy graph]
{
\label{fig:toy:hg}
\includegraphics[width=0.15\linewidth]{FIG/Toy_HG.png}
}
\hspace{10mm}
\subfigure[Gradient path for feature extractor $f_s(\cdot)$]
{
\label{fig:toy:source_cg}
\includegraphics[width=.26\linewidth]{FIG/Toy_CG_source.png}
}
\hspace{10mm}
\subfigure[Gradient path for feature extractor $f_t(\cdot)$]
{
\label{fig:toy:target_cg}
\includegraphics[width=.26\linewidth]{FIG/Toy_CG_target.png}
}
\caption
{
Illustration of a toy heterogeneous graph and the gradient paths for feature extractors $f_s$ and $f_t$. We see that the same HGNN nonetheless produces different feature extractors for each feature domain $\mathcal{X}_s$ and $\mathcal{X}_t$. Colored arrows in figures (b) and (c) show the gradient paths for feature domains $\mathcal{X}_s$ and $\mathcal{X}_t$, respectively.
Note the over-emphasis of the respective gradients in the (b) source and (c) target feature extractors, which can lead to poor generalization.
}
\label{fig:toy}
\end{figure*}
\begin{figure*}[t!]
\centering
\subfigure[Test accuracy across various feature extractors]
{
\label{fig:toy_exp:accuracy}
\includegraphics[width=0.32\linewidth]{FIG/Micro_accuracy.png}
}
\hspace{5mm}
\subfigure[L2 norms of gradients of $W_{\tau(\cdot)}^{(l)}$]
{
\label{fig:toy_exp:transform_w}
\includegraphics[width=.28\linewidth]{FIG/Transform_w.png}
}
\hspace{5mm}
\subfigure[L2 norms of gradients of $M_{\phi(\cdot)}^{(l)}$]
{
\label{fig:toy_exp:message_w}
\includegraphics[width=.28\linewidth]{FIG/Message_w.png}
}
\caption
{
HGNNs trained on a source domain underfit a target domain and perform poorly on a ``nice" heterogeneous graph.
Our theoretically-induced version of \textsc{HGNN-KTN}\xspace adapts the model to the target domain successfully.
In (a) we see performance on the simulated heterogeneous graph, for 4 kinds of feature extractors; (\textit{source}: source extractor $f_s$ on source domain,
\textit{target-src-path}: $f_s$ on target domain,
\textit{target-org-path}: target extractor $f_t$ on target domain,
and \textit{theoretical-KTN}: $f_t$ on target domain using \textsc{HGNN-KTN}\xspace.)
In (b-c), here L2 norms of gradients of parameters $W_{\tau(\cdot)}$ and $M_{\phi(\cdot)}$ in HGNN models.
}
\label{fig:toy_exp}
\end{figure*}
\subsection{Heterogeneous GNNs}
\label{sec:preliminaries:hgnn}
A graph neural network (GNN) can be regarded as a graph encoder which uses the input graph data as the basis for the neural network's computation graph \cite{chami2020machine}.
At a high-level, for any node $j$, the embedding of node $j$ at the $l$-\emph{th} GNN layer is obtained with the following generic formulation:
\begin{equation}\label{eq:gnn}
\small
h_j^{(l)} = \textbf{Transform}^{(l)}\left(\textbf{Aggregate}^{(l)}(\mathcal{E}(j))\right)
\end{equation}
where $\mathcal{E}(j) = \{(i, k)\in\mathcal{E}: i,k\in\mathcal{V}, k = j\}$ denotes all the edges which connect (directionally) to $j$.
HGNNs are a recently-introduced class of GNNs for modeling heterogeneous graphs.
For HGNNs, the above operations typically involve type-specific parameters to exploit the inherent multiplicity of modalities in heterogeneous graphs.
We now define the commonly-used versions of \textbf{Aggregate} and \textbf{Transform} for HGNNs, which we use throughout this paper.
First, we define a linear \textbf{Message} function
\begin{equation}
\label{eq:message}
\small
\textbf{Message}^{(l)}(i, j) = M_{\phi((i, j))}^{(l)}\cdot \left(h_i^{(l-1)}\mathbin\Vert h_j^{(l-1)}\right)
\end{equation}
where $M_r^{(l)}$ are the specific message passing parameters for each $r\in\mathcal{R}$ and each of $L$ GNN layers.
Then defining $\mathcal{E}_r(j)$ as the set of edges of type $r$ pointing to node $j$, our HGNN \textbf{Aggregate} function mean-pools messages by edge type, and concatenates:
\begin{equation}\label{eq:aggregate}
\small
\textbf{Aggregate}^{(l)}(\mathcal{E}(j)) = \underset{r \in\mathcal{R}}{\mathbin\Vert}\tfrac{1}{|\mathcal{E}_r(j)|}\sum_{e\in\mathcal{E}_r(j)}\textbf{Message}^{(l)}(e)
\end{equation}
Finally, \textbf{Transform} maps the message into a type-specific latent space:
\begin{equation}\label{eq:transform}
\small
\textbf{Transform}^{(l)}(j) = \alpha(W_{\tau(j)}\cdot\textbf{Aggregate}^{(l)}(\mathcal{E}(j)))
\end{equation}
The above formulation of HGNNs allows for full handling of the complexities of a real-world heterogeneous graph.
By stacking HGNN blocks for $L$ layers, each node aggregates a larger proportion of nodes --- with different types and relations --- in the full graph, which generates highly contextualized node representations.
The final node representations can be fed into another model to perform downstream heterogeneous network tasks, such as node classification or link prediction.
\subsection{Feature extractors for a toy heterogeneous graph}\label{sec:motivation:feature_extractor}
We first reason intuitively about the differences between $f_s(x_i)$ and $f_t(x_j)$ when $s\ne t$, using a toy heterogeneous graph shown in Figure~\ref{fig:toy:hg}. In that graph, consider nodes $v_1$ and $v_2$, noticing that $\tau(1)\ne \tau(2)$. Using Equations~\eqref{eq:message}-\eqref{eq:transform} from Section \ref{sec:preliminaries:hgnn}, for any $l\in\{0, \ldots, L-1\}$ we have
\begin{equation}\label{eq:tpynode1}
\small
h_1^{(l)} = W_s^{(l)}\left[M_{ss}^{(l)}\left(h_3^{(l-1)}\mathbin\Vert h_1^{(l-1)}\right)\mathbin\Vert M_{ts}^{(l)}\left(h_2^{(l-1)}\mathbin\Vert h_1^{(l-1)}\right)\right]
\end{equation}
and
\begin{equation}\label{eq:tpynode1}
\small
h_2^{(l)} = W_t^{(l)}\left[M_{st}^{(l)}\left(h_1^{(l-1)}\mathbin\Vert h_2^{(l-1)}\right)\mathbin\Vert M_{tt}^{(l)}\left(h_4^{(l-1)}\mathbin\Vert h_2^{(l-1)}\right)\right],
\end{equation}
where $h_j^{(0)} = x_j$. From these equations, we see that the features of nodes $v_1$ and $v_2$, which are of different types, are extracted using \emph{disjoint} sets of model parameters at each layer. In a 2-layer HGNN, this creates unique gradient backpropagation paths between the two node types, as illustrated in Figures~\ref{fig:toy:source_cg}-\ref{fig:toy:target_cg}. In other words, even though the same HGNN is applied to node types $s$ and $t$, the feature extractors $f_s$ and $f_t$ have different computational paths. Therefore they project node features into different latent spaces, and have different update equations during training. We study the consequences of this next.
\subsection{Empirical gap between $f_s$ and $f_t$}
\label{sec:motivation:experiments}
Here we study the experimental consequences of the above observation via simulation.
We first construct a synthetic graph extending the 2-type graph in Figure~\ref{fig:toy:hg} to have multiple nodes per-type, and multiple classes.
Next, we include well-separated classes in both the graph and feature space, where edges and Euclidean node features are well-clustered within-type and within-class
(more details available in Appendix~\ref{appendix:graph-generator:toy}).
On such a well-separated graph, without considering the observation in Section~\ref{sec:motivation:feature_extractor}, there may seem to be no need for domain adaptation from $f_t$ to $f_s$. However, when we train the HGNN model solely on $s$-type nodes, as shown in Figure~\ref{fig:toy_exp:accuracy} we find the test accuracy for $s$-type nodes to be high ($90\%$) and the test accuracy for $t$-type nodes to be quite low ($25\%$).
Now if instead we make the $t$-type nodes use the source feature extractor $f_s$, much more transfer learning is possible (${\sim}65\%$, orange line).
This shows the performance drop mainly comes from the different feature extractors present in the HGNN model, and so domain adaptation on it can not be solved by simply matching data distributions.
To analyze this phenomenon at the level of backpropagation, in Figures~\ref{fig:toy_exp:transform_w}-\ref{fig:toy_exp:message_w} we show the magnitude of gradients passed to parameters of source and target node types. As we intuited in Section~\ref{sec:motivation:feature_extractor}, and as illustrated in Figures~\ref{fig:toy:source_cg}-\ref{fig:toy:target_cg}, we find that the final-layer \textbf{Transform} parameter $W^{(2)}_t$ for type-$t$ nodes have zero gradients (Figure~\ref{fig:toy_exp:transform_w}), and similarly for the final-layer \textbf{Message} parameters (Figure~\ref{fig:toy_exp:message_w}). Additionally, those same parameters in the first-layer for $t$-type nodes have much smaller gradients than their $s$-type counterparts: $W^{(1)}_{t}$ (blue line in Figure~\ref{fig:toy_exp:transform_w}), $M^{(1)}_{st}$ and $M^{(1)}_{tt}$ (blue and orange lines in Figure~\ref{fig:toy_exp:message_w}) appear below than other lines. This is because they contribute to $f_s$ less than $f_t$
\subsection{Relationship between feature extractors in HGNNs}
\label{sec:motivation:theoretical_analysis}
The case study introduced in Section~\ref{sec:motivation:experiments} shows that even when an HGNN is trained on a relatively simple, balanced, and class-separated heterogeneous graph, a model trained only on the source domain node type cannot transfer to the target domain node type. Here, to rigorously describe this phenomenon and the intuition behind it, we derive a strict transformation between $f_s$ and $f_t$, which will motivate the core domain adaptation component of \textsc{HGNN-KTN}\xspace. The following theorem assumes an HGNN as in Equations \eqref{eq:message}-\eqref{eq:transform} without skip-connections, a simplification which we define and explain along with the proof in the Appendix:
\begin{theorem}\label{theorem}
Given a heterogeneous graph $\mathcal{G} = \{\mathcal{V}, \mathcal{E}, \mathcal{T}, \mathcal{R}\}$. For any layer $l>0$, define the set of $(l-1)$-\emph{th} layer HGNN parameters as
\begin{equation}
\label{eq:thm}
\small
\mathcal{Q}^{(l-1)} = \{M_r^{(l-1)}: r\in\mathcal{R}\}\cup\{W_t^{(l-1)}: t\in\mathcal{T}\}.
\end{equation}
Let $A$ be the total $n\times n$ adjacency matrix. Then for any $s,t\in\mathcal{T}$ there exist matrices $A_{ts}^\ast = a_{ts}(A)$ and $Q_{ts}^\ast = q_{ts}(\mathcal{Q}^{(l-1)})$ such that
\begin{equation}
\label{eq:relationship}
\small
H_s^{(l)} = A_{ts}^\ast H_t^{(l)} Q_{ts}^\ast
\end{equation}
where $a_{ts}(\cdot)$ and $q_{ts}(\cdot)$ are matrix functions that depend only on $s,t$.
\end{theorem}
The full proof of Theorem 1 can be found in Appendix \ref{appendix:theorem1}. Notice how in Equation~\ref{eq:relationship}, $Q_{ts}^{\ast}$ acts as a macro- $\textbf{Message}/\textbf{Transform}$ operator that maps $H^{(L)}_t$ into the source domain, then $A_{ts}^{\ast}$ aggregates the mapped embeddings into $s$-type nodes.
To examine the implications of Theorem~\ref{theorem}, we run the same experiment as described in Section~\ref{sec:motivation:experiments}, while this time mapping the target features $H^{(L)}_t$ into the source domain by multiplying with $Q_{ts}^{\ast}$ in Equation~\ref{eq:relationship} before passing over to a task classifier.
We see via the red line in Figure~\ref{fig:toy_exp:accuracy} that, with this mapping, the accuracy in the target domain becomes much closer to the accuracy in the source domain (${\sim}70\%$). Thus, we use this theoretical transformation as a foundation for our trainable HGNN domain adaptation module, introduced in the following section.
\subsection{Algorithm}
\label{sec:matching_loss:algorithm}
We minimize a classification loss $\mathcal{L}_{\text{CL}}$ and a transfer loss $\mathcal{L}_{\text{KTN}}$ jointly with regard to a HGNN model $\textbf{f}$, a classifier $\textbf{g}$, and a knowledge transfer network $\textbf{t}_{\text{KTN}}$ as follows:
\begin{align*}
\small
\underset{\textbf{f},~\textbf{g},~\textbf{t}_{\text{KTN}}}{min}\mathcal{L}_{\text{CL}}(\textbf{g}(\textbf{f}(X_{s})), Y_{s}) + \lambda \left\|\textbf{f}(X_{s}) - \textbf{t}_{\text{KTN}}(\textbf{f}(X_{t}))\right\|_{2}
\end{align*}
where $\lambda$ is a hyperparameter regulating the effect of $\mathcal{L}_{\text{KTN}}$.
Algorithm~\ref{alg:train} describes a training step with a minibatch.
After computing the node embeddings $H^{(L)}_s, H^{(L)}_t$ using a HGNN model $\textbf{f}$, we map $H^{(L)}_t$ to the source domain using $\textbf{t}_{\text{KTN}}$ and compute $\mathcal{L}_{\text{KTN}}$.
Finally, we update the models using gradients of $\mathcal{L}_{\text{CL}}$ and $\mathcal{L}_{\text{KTN}}$.
Algorithm~\ref{alg:test} describes the test phase on the target domain.
After we get node embeddings $H^{(L)}_t$ from the trained HGNN model $\textbf{f}$, we map $H^{(L)}_t$ into the source domain using the trained transformation matrix $T_{ts}$.
Finally we pass the transformed target embeddings $H^{*}_t$ into the classifier $\textbf{g}$ which was trained on the source domain.
\noindent \textbf{Indirect Connections}
We note that in practice, the source and target node types can be indirectly connected in heterogeneous graphs via other node types.
Appendix~\ref{appendix:indirect} describes how we can easily extend \textsc{HGNN-KTN}\xspace to cover domain adaption scenarios in this case.
\subsection{Datasets}
\label{sec:experiments:dataset}
\noindent \textbf{Open Academic Graph (OAG).}~ A dataset introduced in \cite{zhang2019oag} composed of five types of nodes: papers, authors, institutions, venues, fields and their corresponding relationships.
Paper and author nodes have text-based attributes, while institution, venue, and field nodes have text- and graph structure-based attributes.
Paper, author, and venue nodes are labeled with research fields in two hierarchical levels, L1 and L2.
To test the generalization of the proposed model, we construct three field-specific subgraphs from OAG: computer
science, computer networks, and machine learning academic graphs. \vspace{-5pt}
\noindent \textbf{PubMed.} A network of genes, diseases, chemicals, and species \citep{yang2020heterogeneous}, which has 11 types of edges.
Gene and chemical nodes have graph structure-based attributes, while disease and species nodes have text-based attributes.
Each gene or disease is labeled with a set of diseases they belong to or cause. \vspace{-5pt}
\noindent \textbf{Synthetic heterogeneous graphs.} We generate stochastic block models \citep{abbe2017community} with multiple classes and multiple node types. We control within-type edge signal-to-noise ratio by within/between-class edge probabilities, and multivariate Normal feature signal-to-noise ratio by within/between-class variance. We also control \emph{between}-type edge signal-to-noise ratio by allowing nodes of the different types to connect if they are in the same class. A complete definition of the generative model is given in Appendix~\ref{appendix:graph-generator}.
\subsection{Baselines}
\label{sec:experiments:baseline}
We compare \textsc{HGNN-KTN}\xspace with two MMD-based DA methods (DAN~\cite{long2015learning}, JAN~\cite{long2017deep}), three adversarial DA methods (DANN~\cite{ganin2016domain}, CDAN~\cite{long2017conditional}, CDAN-E~\cite{long2017conditional}), one optimal transport-based method (WDGRL~\cite{shen2018wasserstein}), and two traditional graph mining methods (LP and EP~\cite{zhu2005semi}).
For DA methods, we use a HGNN model as their feature extractors.
More information of each method is described in Appendix~\ref{appendix:baseline}.
\begin{table*}[]
\caption{
\textbf{Open Academic Graph on Computer Network field}
}
\label{tab:oag:cn}
\centering
\small
\begin{tabular}{l|l|c|cc|ccc|c|cc|r}
\toprule\hline
\textbf{Task} & \textbf{Metric} & \textbf{Source} & \textbf{DAN} & \textbf{JAN} & \textbf{DANN} & \textbf{CDAN} & \textbf{CDAN-E} & \textbf{WDGRL} & \textbf{LP} & \textbf{EP} & \textbf{KTN (gain)} \\ \hline\midrule
\multirow{2}{*}{\textbf{P-A (L2)}} & \textbf{NDCG} & 0.331 & 0.344 & o.o.m & 0.335 & o.o.m & o.o.m & 0.287 & 0.221 & 0.270 & \textbf{0.382 (16$\%$)} \\
& \textbf{MRR} & 0.250 & 0.277 & o.o.m & 0.280 & o.o.m & o.o.m & 0.199 & 0.130 & 0.270 & \textbf{0.360 (44$\%$)} \\ \hline
\multirow{2}{*}{\textbf{A-P (L2)}} & \textbf{NDCG} & 0.313 & 0.290 & o.o.m & 0.250 & 0.234 & 0.168 & 0.266 & 0.114 & 0.319 & \textbf{0.364 (17$\%$)} \\
& \textbf{MRR} & 0.250 & 0.233 & o.o.m & 0.130 & 0.116 & 0.051 & 0.212 & 0.038 & 0.296 & \textbf{0.368 (47$\%$)} \\ \hline
\multirow{2}{*}{\textbf{A-V (L2)}} & \textbf{NDCG} & 0.539 & 0.521 & 0.519 & 0.510 & 0.467 & 0.362 & 0.471 & 0.232 & 0.443 & \textbf{0.567 (5$\%$)} \\
& \textbf{MRR} & 0.584 & 0.528 & 0.461 & 0.510 & 0.293 & 0.294 & 0.365 & 0.000 & 0.406 & \textbf{0.628 (8$\%$)} \\ \hline
\multirow{2}{*}{\textbf{V-A (L2)}} & \textbf{NDCG} & 0.256 & 0.343 & 0.345 & 0.265 & 0.328 & 0.316 & 0.263 & 0.133 & 0.119 & \textbf{0.348 (33$\%$)} \\
& \textbf{MRR} & 0.117 & 0.296 & 0.286 & 0.151 & 0.285 & 0.275 & 0.147 & 0.000 & 0.000 & \textbf{0.296 (141$\%$)} \\ \hline \bottomrule
\end{tabular}
\normalsize
\end{table*}
\begin{table*}[]
\caption{
\textbf{Open Academic Graph on Machine Learning field}
}
\label{tab:oag:ml}
\centering
\small
\begin{tabular}{l|l|c|cc|ccc|c|cc|r}
\toprule\hline
\textbf{Task} & \textbf{Metric} & \textbf{Source} & \textbf{DAN} & \textbf{JAN} & \textbf{DANN} & \textbf{CDAN} & \textbf{CDAN-E} & \textbf{WDGRL} & \textbf{LP} & \textbf{EP} & \textbf{KTN (gain)} \\ \hline\midrule
\multirow{2}{*}{\textbf{P-A (L2)}} & \textbf{NDCG} & 0.268 & 0.290 & o.o.m & 0.291 & o.o.m & 0.249 & 0.232 & 0.272 & 0.215 & \textbf{0.318 (19$\%$)} \\
& \textbf{MRR} & 0.134 & 0.220 & o.o.m & 0.222 & o.o.m & 0.095 & 0.098 & 0.195 & 0.143 & \textbf{0.269 (102$\%$)} \\ \hline
\multirow{2}{*}{\textbf{A-P (L2)}} & \textbf{NDCG} & 0.261 & 0.225 & o.o.m & 0.234 & 0.228 & 0.241 & 0.241 & 0.119 & 0.267 & \textbf{0.319 (22$\%$)} \\
& \textbf{MRR} & 0.207 & 0.127 & o.o.m & 0.155 & 0.152 & 0.095 & 0.182 & 0.035 & 0.214 & \textbf{0.287 (39$\%$)} \\ \hline
\multirow{2}{*}{\textbf{A-V (L2)}} & \textbf{NDCG} & 0.465 & 0.493 & 0.463 & 0.477 & 0.408 & 0.422 & 0.393 & 0.224 & 0.424 & \textbf{0.538 (16$\%$)} \\
& \textbf{MRR} & 0.469 & 0.542 & 0.537 & 0.519 & 0.412 & 0.240 & 0.213 & 0.001 & 0.391 & \textbf{0.632 (35$\%$)} \\ \hline
\multirow{2}{*}{\textbf{V-A (L2)}} & \textbf{NDCG} & 0.252 & 0.293 & 0.292 & 0.237 & 0.242 & 0.255 & 0.250 & 0.137 & 0.119 & \textbf{0.302 (20$\%$)} \\
& \textbf{MRR} & 0.131 & 0.212 & 0.199 & 0.086 & 0.085 & 0.129 & 0.118 & 0.000 & 0.000 & \textbf{0.227 (73$\%$)} \\ \hline\bottomrule
\end{tabular}
\normalsize
\end{table*}
\subsection{Zero-shot domain adaptation}
\label{sec:experiments:zero-shot}
We run $18$ different zero-shot domain adaptation tasks across three OAG and PubMed graphs.
Each heterogeneous graph has node classification tasks for both source and target node types.
Only source node types have labels, while target node types have none during training.
The performance is evaluated by NDCG and MRR --- widely adopted ranking metrics~\cite{hu2020heterogeneous, hu2020gpt}.
In Tables~\ref{tab:oag:cs},~\ref{tab:pubmed},~\ref{tab:oag:cn}, and~\ref{tab:oag:ml}, our proposed method \textsc{HGNN-KTN}\xspace consistently outperforms all baselines on all tasks and graphs by up to $73.3\%$ higher in MRR (P-A(L1) task in OAG-CS, Table \ref{tab:oag:cs}).
When we compare with the original accuracy possible using the model pretrained on the source domain without any domain adaptation ($3$rd column, \textit{Source}), the results are even more impressive.
Here we see our method \textsc{HGNN-KTN}\xspace provides relative gains of up to $340\%$ higher MRR without using any labels from the target domain.
These results show the clear effectiveness of \textsc{HGNN-KTN}\xspace on zero-shot domain adaptation tasks on a heterogeneous graph.
We note that in OAG graphs, the paper and author node types have different modalities (text and graph embeddings), and in the PubMed graph, disease and gene node types have different modalities (text and graph embeddings).
In all cases, \textsc{HGNN-KTN}\xspace still transfers knowledge successfully while all baselines show poor performance even between domains of the same modalities (as they do not consider different feature extractors in HGNN models).
Finally, we mention that venue and author node types are not directly connected in the OAG graphs (Figure~\ref{fig:schema1:oag}), but \textsc{HGNN-KTN}\xspace successfully transfer knowledge by passing the intermediate nodes.
\noindent \textbf{Baseline Performance.}
\label{sec:experiments:zero-shot-analysis}
Among baselines, MMD-based models (DAN and JAN) outperform adversarial based methods (DANN, CDAN, and CDAN-E) and optimal transport-based method (WDGRL), unlike results reported in~\cite{long2017conditional, shen2018wasserstein}.
These reversed results are a consequence of HGNN's unique feature extractors for source and target domains.
DANN and CDAN define their adversarial losses as a cross entropy loss ($\mathbb{E}[log\textbf{f}_s(x_s)] - \mathbb{E}[log\textbf{f}_t(x_t)]$) where gradients of the subloss $\mathbb{E}[log\textbf{f}_s(x_s)]$ computed from the source feature extractor $f_s(x_s)$ are passed only back to $\textbf{f}_s(x_s)$, while gradients of the subloss $\mathbb{E}[log\textbf{f}_t(x_t)]$ computed from the target feature extractor $\textbf{f}_t(x_t)$ are passed only back to $\textbf{f}_t(x_t)$.
Importantly, source and target feature extractors do not share any gradient information, resulting in divergence.
This did not occur in their original test environments where source and target domains share a single feature extractor.
Similarly, WDGRL measures the first-order Wasserstein distance as an adversarial loss, which also brings the same effect as the cross-entropy loss we described above, leading to divergent gradients between source and target feature extractors.
On the other hand, DAN and JAN define a loss in terms of higher-order MMD between source and target features.
Then the gradients of the loss passed to each feature extractor contain both source and target feature information, resulting in a more stable gradient estimation.
This shows again the importance of considering different feature extractors in HGNNs.
More analysis can be found in Appendix~\ref{appendix:analysis}
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\linewidth]{FIG/synthetic-legend.png} \\
\subfigure[Edge probability (easy)]
{
\label{fig:2-type-simple:edge}
\includegraphics[width=0.47\linewidth]{FIG/2type-simple-edge.png}
}
\subfigure[Feature distribution (easy)]
{
\label{fig:2-type-simple:feat}
\includegraphics[width=.47\linewidth]{FIG/2type-simple-feat.png}
}
\subfigure[Edge probability (hard)]
{
\label{fig:2-type-hard:edge}
\includegraphics[width=0.47\linewidth]{FIG/2type-hard-edge.png}
}
\subfigure[Feature distribution (hard)]
{
\label{fig:2-type-hard:feat}
\includegraphics[width=.47\linewidth]{FIG/2type-hard-feat.png}
}
\caption
{
Effects of edge probabilities and feature distributions across classes and types in $2$-node type heterogeneous graphs.
}
\label{fig:2-type}
\end{figure}
\subsection{Sensitivity analysis}
\label{sec:experiments:sensitivity}
\newcommand{\sigma_e}{\sigma_e}
\newcommand{\sigma_f}{\sigma_f}
Using our synthetic heterogeneous graph generator described in Section \ref{sec:experiments:dataset}, we generate non-trivial 2-type heterogeneous graphs to examine how the feature and edge distributions of heterogeneous graphs affect the performance of \textsc{HGNN-KTN}\xspace and other baselines.
We generate a \emph{range} of test-case scenarios by manipulating (1) signal-to-noise ratio $\sigma_e$ of within-class edge probability and (2) signal-to-noise ratio $\sigma_f$ of within-class feature distributions (details in Appendix~\ref{appendix:graph-generator}) across all of the (a) source-source ($s\leftrightarrow s$), (b) target-target ($t\leftrightarrow t$), and (c) source-target ($s\leftrightarrow t$) relationships. A higher signal-to-noise ratio for a particular data dimension (edges vs features) across a particular relationship $r\in \{s\leftrightarrow s,\ t\leftrightarrow t,\ s\leftrightarrow s\}$ means that classes are more \emph{separable} in that data dimension, when comparing within $r$, and hence easier for HGNNs. Note that while tuning one $\sigma$ on the range $[1.0, 10.0]$ for one of the six $(\sigma, r)$ pairs, the $\sigma$ in all five other pairs are held at $10.0$. Additionally, we vary $\sigma$ across two scenarios: (I) ``easy": source and target node types have same number of classes and same feature dimensions, (II) ``hard" source and target node types have different number of classes and feature dimensions. At each unique value of $\sigma$ across the six ($\sigma, r$) pairs, we generate 5 heterogeneous graphs, train HGNN-KTN and other DA baselines using source class labels, and test using target class labels.
The findings from our synthetic data study are shown in Figure~\ref{fig:2-type}. Figures~\ref{fig:2-type-simple:edge} and \ref{fig:2-type-hard:edge} show results from changing $\sigma_e$ across the three relation types. We see that \textsc{HGNN-KTN}\xspace is affected only by $\sigma_e$ across the $s\leftrightarrow t$ relationship, which accords with our theory, since \textsc{HGNN-KTN}\xspace exploits the between-type computation (adjacency) matrix. Surprisingly, as seen in Figures~\ref{fig:2-type-simple:feat} and \ref{fig:2-type-hard:feat}, we do not find a similar dependence of \textsc{HGNN-KTN}\xspace on $\sigma_f$, which shows that \textsc{HGNN-KTN}\xspace is robust by learning purely from edge homophily in the absence of feature homophily. This robustness is a result of our theoretically-motivated formulation of KTN, allowing the full expressivity of HGNNs within the transfer-learning task.
Regarding the performance of other baselines, EP shows similar tendencies as \textsc{HGNN-KTN}\xspace --- only affected by cross-type $\sigma_e$ --- because EP also relies on cross-type propagation along edges. However, its accuracy is bounded above due to the fact that it does not model or propagate the (unlabelled) target features.
DAN and DANN, which do not exploit cross-type edges, are not affected by cross-type $\sigma_e$.
However, they show either low or unstable performance across different scenarios.
DAN shows especially poor performance in the ``hard" scenarios (Figure~\ref{fig:2-type-hard:edge} and~\ref{fig:2-type-hard:feat}), failing to deal with different feature spaces for source and target domains.
\begin{table}[]
\caption{
\textbf{Different types of HGNNs:}
sharing more parameters does not improve domain adaptation.
}
\label{tab:hgnn-type}
\centering
\small
\begin{tabular}{l|l|ll|ll}
\toprule \hline
\multirow{2}{*}{\textbf{Task}} & \multirow{2}{*}{\textbf{Model}} & \multicolumn{2}{c|}{\textbf{NDCG}} & \multicolumn{2}{c}{\textbf{MRR}} \\
& & \textbf{Source} & \textbf{Target} & \textbf{Source} & \textbf{Target} \\ \hline
\multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}P-A\\ (L1)\end{tabular}}} & HGNN-v1 & 0.634 & 0.564 & 0.604 & 0.519 \\
& HGNN-v2 & 0.794 & 0.613 & 0.788 & 0.617 \\
& HGNN & 0.792 & 0.623 & 0.785 & 0.629 \\ \hline
\multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}A-V\\ (L1)\end{tabular}}} & HGNN-v1 & 0.675 & 0.568 & 0.690 & 0.543 \\
& HGNN-v2 & 0.69 & 0.669 & 0.695 & 0.687 \\
& HGNN & 0.689 & 0.671 & 0.693 & 0.698 \\ \hline
\bottomrule
\end{tabular}
\normalsize
\end{table}
\subsection{Different types of HGNNs}
\label{sec:experiments:hgnn-types}
Using different parameters for each node and edge types in HGNNs result in different feature extractors for source and target node types.
By sharing more parameters among node/edge types, could we see domain adaptation effect?
Here, we design two variants of HGNNs.
HGNN-v1 provides node-wise input layer that maps different modalities into the shared dimension then shares all the remaining parameters across nodes and layers.
HGNN-v2 provides node-wise transformation matrices and edge-wise message matrices, but sharing them across layers.
In Table~\ref{tab:hgnn-type}, HGNN-v1 shows lower accuracy for both source and target node types.
More parameters specialized to each node/edge types, HGNN models show higher accuracy on source domain, thus higher performance could be transferred to target domain.
Regardless of HGNN model types, \textsc{HGNN-KTN}\xspace transfers knowledge between source and target node types consistently.
\subsection{Effect of trade-off coefficient $\lambda$}
\label{sec:experiments:lambda}
We examine the effect of $\lambda$ on the domain adaptation performance.
In Table~\ref{tab:lambda}, as $\lambda$ decreases, target accuracy decreases as expected.
Source accuracy also sees small drops since $\mathcal{L}_{\text{KTN}}$ functions as a regularizer; by removing the regularization effect, source accuracy decreases.
When $\lambda$ becomes large, both source and target accuracy drop significantly.
Source accuracy drops since the effect of $\mathcal{L}_{\text{KTN}}$ becomes bigger than the classification loss $\mathcal{L}_{\text{CL}}$.
Even the effect of transfer learning become bigger by having bigger $\lambda$, since the source accuracy which will be transferred to the target domain is low, the target accuracy is also low.
Thus we set $\lambda$ to $1$ throughout the experiments.
\begin{table}[]
\caption{
\textbf{Effect of $\lambda$}
}
\label{tab:lambda}
\centering
\small
\begin{tabular}{l|cccc}
\toprule \hline
\multicolumn{1}{c|}{\textbf{Task}} & \multicolumn{4}{c}{\textbf{P-A (L1)}} \\ \hline
\multicolumn{1}{c|}{\textbf{Metric}} & \multicolumn{2}{c|}{\textbf{NDCG}} & \multicolumn{2}{c}{\textbf{MRR}} \\ \hline
$\lambda$ & \textbf{source} & \multicolumn{1}{c|}{\textbf{target}} & \textbf{source} & \textbf{target} \\ \hline \midrule
\textbf{$10^{-4}$} & 0.780 & \multicolumn{1}{c|}{0.587} & 0.772 & 0.595 \\
\textbf{$10^{-2}$} & 0.788 & \multicolumn{1}{c|}{0.58} & 0.779 & 0.576 \\
\textbf{$1$} & 0.792 & \multicolumn{1}{c|}{0.621} & 0.788 & 0.633 \\
\textbf{$10^{2}$} & 0.75 & \multicolumn{1}{c|}{0.617} & 0.757 & 0.623 \\
\textbf{$10^{4}$} & 0.143 & \multicolumn{1}{c|}{0.177} & 0.007 & 0.031 \\ \hline \midrule
\multicolumn{1}{c|}{\textbf{Task}} & \multicolumn{4}{c}{\textbf{A-V (L1)}} \\ \hline
\multicolumn{1}{c|}{\textbf{Metric}} & \multicolumn{2}{c|}{\textbf{NDCG}} & \multicolumn{2}{c}{\textbf{MRR}} \\ \hline
\textbf{$\lambda$} & \textbf{source} & \multicolumn{1}{c|}{\textbf{target}} & \textbf{source} & \textbf{target} \\ \hline \midrule
\textbf{$10^{-4}$} & 0.689 & \multicolumn{1}{c|}{0.626} & 0.690 & 0.642 \\
\textbf{$10^{-2}$} & 0.687 & \multicolumn{1}{c|}{0.654} & 0.689 & 0.677 \\
\textbf{$1$} & 0.689 & \multicolumn{1}{c|}{0.67} & 0.692 & 0.696 \\
\textbf{$10^{2}$} & 0.654 & \multicolumn{1}{c|}{0.644} & 0.659 & 0.668 \\
\textbf{$10^{4}$} & 0.411 & \multicolumn{1}{c|}{0.432} & 0.373 & 0.421 \\ \hline \bottomrule
\end{tabular}
\normalsize
\end{table}
\subsection{Proof of Theorem 1}\label{appendix:theorem1}
The proof of Theorem \ref{theorem} is below. As stated in the assumptions of the theorem, we adopt a simplified version of our message-passing function that ignores the skip-connection:
\begin{equation}
\small
\textbf{Message}^{(l)}(i, j) = M_{\phi(i,j)}^{(l)}h_i^{(j)}.
\end{equation}
This lets the Theorem match the experimental results shown in Figure \ref{fig:toy_exp}, as the HGNN trained in that experiment does not use skip-connections and hence represents an ``idealized" HGNN without skip-connections, and with a theoretically-exact KTN component. In the real experiments, we use (1) skip-connections, exploiting their usual benefits~\cite{hamilton2017inductive}, and (2) the trainable version of KTN.
\begin{proof}
Without loss of generality, we prove the result for the case where $\mathcal{R} = \{(s, t): s,t\in\mathcal{T}\}$, meaning the type of an edge is identified with the (ordered) types of the neighbor nodes. In other words, there is only one edge modality possible, such as a social networks with multiple node types (e.g.\ ``users", ``groups") but only one edge modality (``friendship"). In the case of multiple edge modalities (e.g. ``friendship" and ``message"), the result is extended trivially (though with more algebraically-dense forms of $a_{ts}$ and $q_{ts}$).
Throughout this proof, we use the following notation for the set of all $j$-adjacent edges of relation type $r$:
\begin{equation}
\small
\mathcal{E}_r(j):=\{(i,j): i\in\mathcal{V}, (i,j) = r\}.
\end{equation}
We write $A_{x_1x_2}$ to denote the sub-matrix of the total $n\times n$ adjacency matrix $A$ corresponding to node types $x_1,x_2\in\mathcal{T}$, and $\bar{A}_{x_1x_2}$ to denote the same matrix divided by its sum. $H_x^{(l)}$ is the (row-wise) $n_x\times d_l$ embedding matrix of $x$-type nodes at layer $l$.
To begin, we first compute the $l$-\emph{th} output $g_j^{(l)}$ of the $\textbf{Aggregate}$ step defined for HGNNs in Equation \eqref{eq:aggregate}, for any node $j\in\mathcal{V}$ such that $\tau(j) = s$. The output of \textbf{Aggregate} is in fact a concatenation of edge-type-specific aggregations (see Equation~\ref{eq:aggregate}). Note that at most $T = |\mathcal{T}|$ elements of this concatenation are non-zero, since the node $j$ only participates in $T$ out of $T^2$ relation types in $\mathcal{R}$. Thus we can write $g_j^{(l)}$ as
\begin{align*}
\small
g_j^{(l)} &= \underset{r\in\mathcal{R}}{\mathbin\Vert}\tfrac{1}{|\mathcal{E}_r(j)|}\sum_{e\in\mathcal{E}_r(j)}\textbf{Message}^{(l)}(e)\\
&= \underset{x\in\mathcal{T}}{\mathbin\Vert}\tfrac{1}{|\mathcal{E}_{xs}(j)|}\sum_{e\in\mathcal{E}_{xs}(j)}\textbf{Message}^{(l)}(e)\\
&=\underset{x\in\mathcal{T}}{\mathbin\Vert}\tfrac{1}{|\mathcal{E}_{xs}(j)|}\sum_{(i,j)\in\mathcal{E}_{xs}(j)}M_{xs}^{(l)}h_i^{(l-1)}\\
&=\underset{x\in\mathcal{T}}{\mathbin\Vert}\tfrac{1}{|\mathcal{E}_{xs}(j)|}M_{xs}^{(l)}\sum_{(i,j)\in\mathcal{E}_{xs}(j)}h_i^{(l-1)}\\
&=\underset{x\in\mathcal{T}}{\mathbin\Vert}M_{xs}^{(l)}\left(H_x^{(l-1)}\right)'\bar{A}_{xs}^{(j)},
\end{align*}
where $\bar{A}_{xs}^{(j)}$ denotes the $j$-\emph{th} column of $\bar{A}_{xs}$. Notice that
\begin{equation}
\small
h_j^{(l)} = \textbf{Transform}^{(l)}(j) = W_s^{(l)}g_j^{(l)},
\end{equation}
and (again) at most $T$ elements of the concatenation $g_j^{(l)}$ are non-zero. Therefore let $W_{xs}^{(l)}$ be the columns of $W_s^{(l)}$ that select the concatenated element of $g_j^{(l)}$ corresponding to node type $x$. Then we can write
\begin{equation}
\small
h_j^{(l)} = \sum_{x\in\mathcal{T}}W_{xs}^{(l)}M_{xs}^{(l)}\left(H_x^{(l-1)}\right)'\bar{A}_{xs}^{(j)}.
\end{equation}
\begin{algorithm}[t!]
\caption{Test step for a target domain (indirect version)}
\label{alg:test-extend}
\begin{algorithmic}[1]
\small
\REQUIRE pretrained HGNN $\textbf{f}$, classifier $\textbf{g}$, \textsc{HGNN-KTN}\xspace $\textbf{t}_{\text{KTN}}$
\ENSURE target node label matrix $Y_t$
\STATE $H^{(L)}_t = \textbf{f}(H^{(0)} = X, \mathcal{G})$, $H^{*}_{t} = \textbf{0}$
\FOR{each meta-path $p = t \rightarrow s$}
\STATE $x = t$, $Z = H^{(L)}_t$
\FOR{each node type $y \in p$}
\STATE $X = ZT_{xy}$
\STATE $x = y$
\ENDFOR
\STATE $H^{*}_{t} = H^{*}_{t} + Z$
\ENDFOR
\RETURN $\textbf{g}(H^{*}_{t})$
\normalsize
\end{algorithmic}
\end{algorithm}
Defining the operator $Q_{xs}^{(l)} := \left(W_{xs}^{(l)}M_{xs}^{(l)}\right)'$, this implies that
\begin{align*}
\small
&H^{(l)}_s = \sum_{x\in\mathcal{T}}\bar{A}_{xs}H_x^{(l-1)}Q_{xs}^{(l)} \\
&= [\bar{A}_{x_{1}s},\ldots,\bar{A}_{x_{T}s}]
\begin{bmatrix}
H^{(l-1)}_{x_1} & 0 & 0\\
0 & \ldots & 0\\
0 & 0 & H^{(l-1)}_{x_T}
\end{bmatrix}
\begin{bmatrix}
Q_{x_1s}^{(l-1)}\\
\ldots\\
Q_{x_Ts}^{(l-1)}
\end{bmatrix}
\\
& = \bar{A}_{\cdot s}H_{\cdot}^{(l-1)}Q_{\cdot s}^{(l-1)}
\end{align*}
Similarly we have $H^{(l)}_t = \bar{A}_{\cdot t}H_{\cdot}^{(l-1)}Q_{\cdot t}^{(l-1)}$. Since $H^{(l)}_s$ and $H^{(l)}_t$ share the term $H_\cdot^{(l-1)}$, we can write
\begin{equation}
\label{eq:thoretical}
\small
H_s^{(l)} = \bar{A}_{\cdot s}\bar{A}^{-1}_{\cdot t} H^{(l)}_{t} (Q_{\cdot t}^{(l-1)})^{-1}Q_{\cdot s}^{(l-1)},
\end{equation}
where $X^{-1}$ denotes the pseudo-inverse. This proves the result.
\end{proof}
\subsection{Indirectly Connected Source and Target Node Types}
\label{appendix:indirect}
When source and target node types are indirectly connected by another node type $x$, we can simply extend $\textbf{t}_{\text{KTN}}(H^{(L)}_{t})$ to $(A_{xs}(A_{tx}H^{(L)}_{t}T_{tx})T_{xs})$ where $T_{tx}T_{xs}$ becomes a mapping function from target to source domains.
Algorithm~\ref{alg:train-extend} and~\ref{alg:test-extend} show how \textsc{HGNN-KTN}\xspace is extended.
For every step ($x \rightarrow y$) in a meta-path ($t \rightarrow \cdots \rightarrow s$) connecting from target node type $t$ to source node type $s$, we define a transformation matrix $T_{xy}$, run a convolution operation with an adjacency matrix $A_{xy}$, and map the transformed embedding to the source domain.
We run the same process for all meta-paths connecting from target node type $t$ to source node type $s$, and sum up them to match with the source embeddings.
In the test phase, we run the same process to get the transformed target embeddings, but this time, without adjacency matrices.
We run Algorithm~\ref{alg:train-extend} and~\ref{alg:test-extend} for domain adaptation tasks between author and venue nodes which are indirectly connected by paper nodes in OAG graphs (Figure~\ref{fig:schema1:oag}).
As shown in Tables~\ref{tab:oag:cs},~\ref{tab:oag:cn}, and~\ref{tab:oag:ml}, we successfully transfer HGNN models between author and venue nodes (A-V and V-A) for both L1 and L2 tasks.
Which meta-path between source and target node types should we choose?
Will lengths of meta-paths affect the performance?
We examine the performance of \textsc{HGNN-KTN}\xspace varying the length of meta-paths.
In Table~\ref{tab:meta-path}, accuracy decreases with longer meta-paths.
When we add additional meta-paths than the minimum path, it also brings noise in every edge types.
Note that author and venue nodes are indirectly connected by paper nodes; thus the minimum length of meta-paths in the A-V (L1) task is $2$.
The accuracy in the A-V (L1) task with a meta-path of length $1$ is low because \textsc{HGNN-KTN}\xspace fails to transfer anything with a meta-path shorter than the minimum.
Using the minimum length of meta-paths is enough for \textsc{HGNN-KTN}\xspace.
\subsection{Analysis for Baselines in Section~\ref{sec:experiments:zero-shot}}
\label{appendix:analysis}
JAN, CDAN, and CDAN-E often show out of memory issues in Tables~\ref{tab:oag:cs},~\ref{tab:oag:cn}, and~\ref{tab:oag:ml}.
These baselines consider the classifier prediction whose dimension is equal to the number of classes in a given task.
That is why JAN, CDAN, and CDAN-E fail at the L2 field prediction tasks in OAG graphs where the number of classes is $17,729$.
LP performs worst among the baselines, showing the limitation of relying only on graph structures.
LP maintains a label vector with the length equal to the number of classes for each node, thus shows out-of-memory issues on tasks with large number of classes on large-size graphs (L2 tasks with $17,729$ labels on the OAG-CS graph).
EP performs moderately well similar to other DA methods, but lower than \textsc{HGNN-KTN}\xspace up to $60\%$ absolute points of MRR, showing the limitation of not using target node attributes.
\begin{table}[]
\caption{
\textbf{Meta-path length in \textsc{HGNN-KTN}\xspace:}
increasing the meta-path longer than the minimum does not bring significant improvement to \textsc{HGNN-KTN}\xspace.
Note that the minimum length of meta-paths in the A-V (L1) task is $2$.
}
\label{tab:meta-path}
\centering
\small
\begin{tabular}{c|llll}
\toprule \hline
\textbf{Task} & \multicolumn{2}{c}{\textbf{P-A (L1)}} & \multicolumn{2}{c}{\textbf{A-V (L1)}} \\ \hline
\textbf{\begin{tabular}[c]{@{}c@{}}Meta-path\\ length\end{tabular}} & \textbf{NDCG} & \multicolumn{1}{l|}{\textbf{MRR}} & \textbf{NDCG} & \textbf{MRR} \\ \hline \midrule
\textbf{1} & 0.623 & \multicolumn{1}{l|}{0.621} & 0.208 & 0.010 \\
\textbf{2} & 0.627 & \multicolumn{1}{l|}{0.628} & 0.673 & 0.696 \\
\textbf{3} & 0.608 & \multicolumn{1}{l|}{0.611} & 0.627 & 0.648 \\
\textbf{4} & 0.61 & \multicolumn{1}{l|}{0.623} & 0.653 & 0.671 \\
\hline \bottomrule
\end{tabular}
\normalsize
\end{table}
\subsection{Synthetic Heterogeneous Graph Generator}
\label{appendix:graph-generator}
Our synthetic heterogeneous graph generator is based on attributed Stochastic Block Models (SBM)~\cite{tsitsulin2020graph,tsitsulin2021synthetic}, using clusters (blocks) as the node classes.
In the attributed SBM, graphs exhibit \emph{within-type} cluster homophily at the \emph{edge-level} (nodes most-frequently connect to other nodes in their cluster), and at the \emph{feature-level} (nodes are closest in feature space to other nodes in their cluster).
To produce heterogeneous graphs, we additionally introduce \emph{between-type} cluster homophily, which allows us to model real-world heterogeneous graphs in which knowledge can be shared across node types.
The first step in generating a heterogeneous SBM is to decide how many clusters will partition each node type. Assume within-type cluster counts $k_1, \ldots, k_T$. We allow for cross-type homophily with a $K_T:=\min_t\{k_t\}$-partition of clusters such that each cluster group has at least one cluster from each node type.
Secondly, edge-level homophily is controlled by signal-to-noise ratios $\sigma_e = p/q$ where nodes within-cluster are connected with probability $p$ and nodes between-cluster are connected with probability $q$. Additionally, nodes within the same cluster group across-types (see previous paragraph) can generate between-edges with some $\sigma_e>1.0$. In Section \ref{sec:experiments:sensitivity} we describe the manipulation of multiple $\sigma_e$ parameters within-and-across types.
Finally, node attributes are generated by a multivariate Normal mixture model, using the cluster partition as the mixture groups.
Thus feature-level homophily is controlled by increasing the variance of the cluster centers $\sigma_f$, while keeping the within-cluster variance fixed. Note that features of different types are allowed to have different dimensions, as we generate different mixture-model cluster centers for each cluster \emph{within each type}. Cross-type feature homophily is not necessary, since HGNN-KTN learns a transformation function between the type feature spaces.
\subsubsection{Toy Heterogeneous Graph in Section~\ref{sec:motivation:experiments}}
\label{appendix:graph-generator:toy}
Using the synthetic graph procedure described above, we used the following hyperparameters to simulate the toy heterogeneous graph shown in Figure~\ref{fig:toy_exp}.
We generate the graph with two node types and four edge types as described in Figure~\ref{fig:toy:hg}, then we divide each node type into $4$ classes of $400$ nodes.
To generate an easy-to-transfer scenario, signal-to-noise ratio $\sigma_f$ between means of feature distributions are all set to $10$.
The ratio $\sigma_e$ of the number of intra-class edges to the number of inter-class edges is set to $10$ among the same node types and across different node types.
The dimension of features is set to $24$ for both node types.
\subsubsection{Sensitivity test in Section~\ref{sec:experiments:sensitivity}}
\label{appendix:graph-generator:sensitivity}
Figure~\ref{fig:schema2} shows the structures of graphs we used in Section~\ref{sec:experiments:sensitivity}.
The dimension of features are set to $24$ for both node types for the ``easy" scenario and, and $32,48$ for types $s$ and $t$ (respectively) for the ``hard" scenario. Additionally, for the ``hard" scenario, we divide the $t$ nodes into 8 clusters instead of 4.
The other hyperparameters $\sigma_e$ and $\sigma_f$ are described in Section~\ref{sec:experiments:sensitivity}.
\begin{table*}[t!]
\caption{
\textbf{Statistics of Open Academic Graph}
}
\label{tab:oag:statistics}
\centering
\small
\begin{tabular}{l|l|l|l|l|l|l}
\toprule\hline
\textbf{Domain} & \textbf{\#papers} & \textbf{\#authors} & \textbf{\#fields} & \textbf{\#venues} & \textbf{\#institues} & \\ \hline\midrule
\textbf{Computer Science} & 544,244 & 510,189 & 45,717 & 6,934 & 9,097 & \\
\textbf{Computer Network} & 75,015 & 82,724 & 12,014 & 2,115 & 4,193 & \\
\textbf{Machine Learning} & 90,012 & 109,423 & 19,028 & 3,226 & 5,455 & \\ \hline
\textbf{Domain} & \textbf{\#P-A} & \textbf{\#P-F} & \textbf{\#P-V} & \textbf{\#A-I} & \textbf{\#P-P} & \textbf{\#F-F} \\ \hline\midrule
\textbf{Computer Science} & 1,091,560 & 3,709,711 & 544,245 & 612,873 & 11,592,709 & 525,053 \\
\textbf{Computer Network} & 155,147 & 562,144 & 75,016 & 111,180 & 1,154,347 & 110,869 \\
\textbf{Machine Learning} & 166,119 & 585,339 & 90,013 & 156,440 & 1,209,443 & 163,837 \\ \hline\bottomrule
\end{tabular}
\normalsize
\end{table*}
\begin{table*}[t!]
\caption{
\textbf{Statistics of PubMed Graph}
}
\label{tab:pubmed:statistics}
\centering
\small
\begin{tabular}{p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}}
\toprule\hline
\textbf{\#gene} & \textbf{\#disease} & \textbf{\#chemicals} & \textbf{\#species} & \\ \hline\midrule
13,561 & 20,163 & 26,522 & 2,863 & \\ \hline
\textbf{\#G-G} & \textbf{\#G-D} & \textbf{\#D-D} & \textbf{\#C-G} & \textbf{\#C-D} \\ \hline\midrule
32,211 & 25,963 & 68,219 & 31,278 & 51,324 \\ \hline
\textbf{\#C-C} & \textbf{\#C-S} & \textbf{\#S-G} & \textbf{\#S-D} & \textbf{\#S-S} \\ \hline\midrule
124,375 & 6,298 & 3,156 & 5,246 & 1,597 \\ \hline\bottomrule
\end{tabular}
\normalsize
\end{table*}
\subsection{Real-world Dataset}
\label{appendix:dataset}
\paragraph{Open Academic Graph (OAG)}~\cite{sinha2015overview, tang2008arnetminer, zhang2019oag} is the largest publicly available heterogeneous graph.
It is composed of five types of nodes: papers, authors, institutions, venues, fields and their corresponding relationships.
Papers and authors have text-based attributes, while institutions, venues, and fields have text- and graph structure-based attributes.
To test the generalization of the proposed model, we construct three field-specific subgraphs from OAG: the Computer
Science (OAG-CS), Computer Networks (OAG-CN), and Machine Learning (OAG-ML) academic graphs.
Papers, authors, and venues are labeled with research fields in two hierarchical levels, L1 and L2.
OAG-CS has both L1 and L2 labels, while OAG-CN and OAG-ML have only L2 labels (their L1 labels are all "computer science").
Domain adaptation is performed on the L1 and L2 field prediction tasks between papers, authors, and venues for each of the aforementioned subgraphs.
Note that paper-author (P-A) and paper-venue (P-V) are directly connected, while author-venue (A-V) are indirectly connected via papers.
The number of classes in the L1 task is $275$, while the number of classes in the L2 task is $17,729$.
The graph statistics are listed in Table~\ref{tab:oag:statistics}, in which P–A, P–F, P–V, A–I, P–P, and F-F denote the edges between paper and author, paper and field, paper and venue, author and institute, the citation links between two papers, the hierarchical links between two fields.
The graph structure is described in Figure~\ref{fig:schema1:oag}.
For paper nodes, features are generated from each paper's title using a pre-trained XLNet~\cite{wolf2020transformers}.
For author nodes, features are averaged over features of papers they published.
Feature dimension of paper and author nodes is $769$.
For venue, institution, and field node types, features of dimension $400$ are generated from their heterogeneous graph structures using metapath2vec~\cite{dong2017metapath2vec}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.36\linewidth]{FIG/Synthetic_2.png}
\caption
{
Schema of synthetic heterogeneous graphs used in the sensitivity test in Section~\ref{sec:experiments:sensitivity}.
}
\label{fig:schema2}
\end{figure}
\begin{figure}[t!]
\centering
\subfigure[OAG]
{
\label{fig:schema1:oag}
\includegraphics[width=0.55\linewidth]{FIG/oag.png}
}
\subfigure[PubMed]
{
\label{fig:schema1:pubmed}
\includegraphics[width=.39\linewidth]{FIG/Pubmed.png}
}
\caption
{
Schema of real-world heterogeneous graphs
}
\label{fig:schema1}
\end{figure}
\paragraph{PubMed}~\cite{yang2020heterogeneous} is a novel biomedical network constructed through text mining and manual processing on biomedical literature.
PubMed is composed of genes, diseases, chemicals, and species.
Each gene or disease is labeled with a set of diseases (e.g., cardiovascular disease) they belong to or cause.
Domain adaptation is performed on a disease prediction task between genes and disease node types.
The number of classes in the disease prediction task is $8$.
The graph statistics are listed in Table~\ref{tab:pubmed:statistics}, in which G, D, C, and S denote genes, diseases, chemicals, and species node types.
The graph structure is described in Figure~\ref{fig:schema1:pubmed}.
For gene and chemical nodes, features of dimension $200$ are generated from related PubMed papers using word2vec~\cite{mikolov2013distributed}.
For diseases and species nodes, features of dimension $50$ are generated based on their graph structures using TransE~\cite{bordes2013translating}.
\subsection{Baselines}
\label{appendix:baseline}
Zero-shot domain adaptation can be categorized into three groups --- MMD-based methods, adversarial methods, and optimal-transport-based methods.
MMD-based methods~\cite{long2015learning, sun2016return, long2017deep} minimize the maximum mean discrepancy (MMD)~\cite{gretton2012kernel} between the mean embeddings of two distributions in reproducing kernel Hilbert space.
DAN~\cite{long2015learning} enhances the feature transferability by minimizing multi-kernel MMD in several task-specific layers.
JAN~\cite{long2017deep} aligns the joint distributions of multiple domain-specific layers based on a joint maximum mean discrepancy (JMMD) criterion.
Adversarial methods~\cite{ganin2016domain, long2017conditional} are motivated by theory in~\cite{ben2007analysis, ben2010theory} suggesting that a good cross-domain representation contains no discriminative information about the origin of the input.
They learn domain invariant features by a min-max game between the domain classifier and the feature extractor.
DANN~\cite{ganin2016domain} learns domain invariant features by a min-max game between the domain classifier and the feature extractor.
CDAN~\cite{long2017conditional} exploits discriminative information conveyed in the classifier predictions to assist adversarial adaptation.
CDAN-E~\cite{long2017conditional} extends CDAN to condition the domain discriminator on the uncertainty of classifier predictions, prioritizing the discriminator on easy-to-transfer examples.
Optimal transport-based methods~\cite{shen2018wasserstein} estimate the empirical Wasserstein distance~\cite{redko2017theoretical} between two domains and minimizes the distance in an adversarial manner
Optimal transport-based method are based on a theoretical analysis~\cite{redko2017theoretical} that Wasserstein distance can guarantee generalization for domain adaptation.
WDGRL~\cite{shen2018wasserstein} estimates the empirical Wasserstein distance between two domains and minimizes the distance in an adversarial manner.
\subsection{Experimental Settings}
\label{appendix:experiment-setting}
All experiments were conducted on the same p2.xlarge Amazon EC2 instance.
Here, we describe the structure of HGNNs used in each heterogeneous graph.
\paragraph{Open Academic Graph:}
We use a $4$-layered HGNN with transformation and message parameters of dimension $128$ for \textsc{HGNN-KTN}\xspace and other baselines.
Learning rate is set to $10^{-4}$.
\paragraph{PubMed:}
We use a single-layered HGNN with transformation and message parameters of dimension $10$ for \textsc{HGNN-KTN}\xspace and other baselines.
Learning rate is set to $5 \times 10^{-5}$.
\paragraph{Synthetic Heterogeneous Graphs:}
We use a $2$-layered HGNN with transformation and message parameters of dimension $128$ for \textsc{HGNN-KTN}\xspace and other baselines.
Learning rate is set to $10^{-4}$.
We implement LP, EP and \textsc{HGNN-KTN}\xspace using Pytorch.
For the domain adaptation baselines (DAN, JAN, DANN, CDAN, CDAN-E, and WDGRL), we use a public domain adaptation library ADA~\footnote{\url{https://github.com/criteo-research/pytorch-ada}}.
\section{Introduction}
\label{sec:introduction}
\input{001introduction.tex}
\section{Related Work}
\label{sec:related_work}
\input{002related_work.tex}
\section{Preliminaries}
\label{sec:preliminaries}
\input{003preliminaries.tex}
\section{Cross-Type Transformations in HGNNs}
\label{sec:motivation}
\input{004motivation.tex}
\section{Method: \textsc{HGNN-KTN}\xspace}
\label{sec:matching_loss}
\input{005matching_loss.tex}
\section{Experiments}
\label{sec:experiments}
\input{006experiments.tex}
\section{Conclusion}
\label{sec:conclusion}
\input{007conclusion.tex}
\section*{Acknowledgement}
\label{sec:ackonwledgement}
\input{008acknowledgement.tex}
| 6ecbc1c6af963f296c5041cee6ab596a727f144d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Neuroimaging with MRI is of paramount importance in the understanding of the morphology and connectivity of the human brain. \textit{In vivo} MR imaging has been widely adopted in research, where studies rely most often on prospective datasets of high-quality brain scans. Meanwhile, clinical MRI datasets remain largely unexplored in neuroimaging studies, despite their much higher abundance (e.g., 10 million brain scans were acquired in the US in 2019~\cite{oren_curbing_2019}). Analysing such datasets is highly desirable, since it would enable sample sizes in the millions, which is much higher than the current largest research studies, which include tens of thousands subjects (e.g., ENIGMA~\cite{hibar_common_2015} or UK BioBank~\cite{alfaro-almagro_image_2018}).
The use of clinical data in neuroimaging studies has been mainly hindered by the high variability in acquisition protocols. As opposed to high resolution (HR) scans used in research, physicians usually prefer low resolution (LR) acquisitions with fewer slices (for faster inspection), which span a large range of orientations, slice spacings and slice thicknesses. Moreover, clinical scans employ numerous MRI contrasts to highlight different tissue properties.
Overall, no segmentation method can robustly adapt to such variability. Manual labelling is the gold standard in segmentation techniques, but it remains too tedious for large-scale clinical applications. An alternative would be to only consider subjects with high-quality acquisitions (e.g., \SI{1}{\milli\meter} T1 scans), as these can be easily analysed with neuroimaging softwares~\cite{ashburner_unified_2005,fischl_freesurfer_2012}. However, this would enormously decrease the effective sample size of clinical datasets, where such scans are seldom available. Hence, there is a clear need for a robust automated segmentation tool that can adapt to clinical scans of any MRI contrast and~resolution.
Contrast-invariance has traditionally been achieved via Bayesian segmentation strategies with unsupervised likelihood model~\cite{puonti_fast_2016}. Unfortunately, these methods are highly sensitive to partial volume effects (PV) caused by changes in resolution~\cite{choi_partial_1991}. This problem can partly be mitigated by directly modelling PV within the Bayesian framework~\cite{van_leemput_unifying_2003}. However, this strategy quickly becomes intractable for scans with decreasing resolutions and increasing number of labels, thus limiting its application to large clinical datasets.
Recent automated segmentation methods rely on supervised convolutional neural networks (CNNs)~\cite{kamnitsas_efficient_2017,milletari_v-net_2016,ronneberger_u-net_2015}. While CNNs obtain fast and accurate results on their training domain, they are fragile to changes in resolution~\cite{ghafoorian_transfer_2017,orbes-arteaga_multi-domain_2019} and MRI contrast~\cite{akkus_deep_2017,jog_psacnn_2019}, even within the same MRI modality~\cite{karani_test-time_2021,kushibar_supervised_2019}. Although data augmentation can improve robustness in intra-modality scenarios~\cite{zhang_generalizing_2020}, CNNs still need to be retrained for each new MRI contrast and resolution. This issue has sparked a vivid interest in domain adaptation schemes, where CNNs are trained to generalise to a specific target domain~\cite{chen_synergistic_2019,karani_test-time_2021}. However, these methods still need to be retrained for every new target resolution or MRI contrast, which makes them impractical to apply on highly heterogeneous clinical data.
Very recently, a publicly available method named \textit{SynthSeg}~\cite{billot_synthseg_2021} has been proposed for out-of-the-box segmentation of brain scans of any contrast and resolution. \textit{SynthSeg} relies on a 3D UNet trained on synthetic scans generated with a domain randomisation approach~\cite{tobin_domain_2017}. While \textit{SynthSeg} yields excellent generalisation compared with previous techniques, it still lacks robustness when applied to clinical scans with low signal-to-noise (SNR) ratio or poor tissue contrast.
Improvements in robustness have previously been tackled with hierarchical models, where a first CNN performs a simpler preliminary task (e.g., predicting an initial mask~\cite{roth_application_2018}, or pre-segmenting at low resolution~\cite{isensee_nnu-net_2021}), and the results are refined by a second network trained for the target task. However, these methods often remain insufficient to capture high-order topological relations, which is a well-known problem for CNNs~\cite{nosrati_incorporating_2016}. A possible solution is to use conditional random fields for postprocessing~\cite{kamnitsas_efficient_2017}, but these often struggle to model relations between multiple labels at different scales. Recent methods now seek to improve semantic correctness either by aligning predictions and ground truths in latent space during training~\cite{oktay_anatomically_2018}, or by using denoising CNNs~\cite{larrazabal_post-dae_2020}. Although these methods have shown promising results in relatively simple cases (i.e., 2D images with few labels), they are yet to be demonstrated in more complex setups.
In this work, we present \textit{SynthSeg$^{+}$}, a novel architecture for robust segmentation of clinical MRI scans of any contrast and resolution without retraining. Specifically, we build on the domain randomisation strategy introduced by \textit{SynthSeg}, and propose a hierarchy of conditional segmentation and denoising CNNs for improved robustness and semantic correctness. We evaluate this method on more than 10,000 highly heterogeneous clinical scans, directly taken from the picture archiving communication system (PACS) of Massachusetts General Hospital (MGH). \textit{SynthSeg$^{+}$} yields considerably enhanced robustness compared to \textit{SynthSeg}, while also outperforming cascaded networks and state-of-the-art denoising methods.
\section{Methods}
\subsection{Hierarchical conditional architecture}
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/architecture.pdf}
\caption{Overview of the proposed architecture. A first network $S_1$ outputs initial segmentations of four tissue classes. Robustness is then improved by refining these with a denoiser $D$ (e.g., red box). Final segmentations are obtained with a second segmenter $S_2$, which takes as inputs the image and the robust estimates of the four tissue classes.}
\label{fig:archi}
\end{figure}
We propose an architecture that relies on three hierarchical CNN modules (Fig.~\ref{fig:archi}). This design aims at efficiently subdividing the target segmentation task into intermediate steps that are easier to perform, and thus less prone to errors. For this purpose, a first network $S_1$ is trained to produce coarse initial segmentations of the input images. More precisely, these initial segmentations only contain four labels that group brain regions into classes of similar tissue types and intensities (cerebral white matter, cerebral grey matter, cerebrospinal fluid, and cerebellum). These classes are easier to discriminate than individual regions.
The output of $S_1$ is then fed to a denoising network $D$~\cite{larrazabal_post-dae_2020} in order to increase the robustness of the initial segmentations. By modelling high-level relations between tissue types, $D$ seeks to correct potential semantic inconsistencies introduced by $S_1$ (e.g, cerebral grey matter in the cerebellum). Moreover, it also enables recovery from large mistakes in the initial tissue classes, which sometimes occur for scans with low SNR, poor contrast, or very low resolution.
Final segmentations are obtained with a second segmenter $S_2$, which takes as input the image and the corrected tissue classes given by $D$, thus combining the robustness of $D$ with the accuracy of $S_2$. In practice, $S_2$ learns to subdivide the initial tissue predictions into the target labels, as well as to refine the boundaries given by $D$, which are often excessively smooth (e.g., the cortex in Fig.~\ref{fig:archi}).
\subsection{Training scheme for the segmentation modules}
\label{sec:training_s}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figures/generation.pdf}
\caption{
Generative model to train $S_1$ and $S_2$. A label map is deformed~(a) and used to generate an HR scan~(b), from which we simulate a LR scan of random resolution~(c). $S_1$ is trained to produce HR segmentations of the four coarse tissue classes from the LR image, which is upsampled to HR space for convenience~(d). $S_2$ is trained to estimate all target labels, using both the image and soft tissue segmentations~(e,f). During training, we corrupt these tissue maps to model the errors made by the denoiser $D$.
}
\label{fig:generation}
\end{figure}
$S_1$ and $S_2$ are trained separately with a domain randomisation strategy~\cite{tobin_domain_2017}.
Specifically, we use synthetic data sampled on the fly from a generative model that only needs a set of label maps as input, and whose parameters (contrast, resolution, artefacts, etc.) are drawn from uninformative priors (Supplementary Table~\ref{tab:priors}). As a result, $S_1$ and $S_2$ are exposed to vastly varying examples, which forces them to learn contrast- and resolution-agnostic features. Training image-target pairs for $S_1$ are generated with a procedure similar to \textit{SynthSeg}~\cite{billot_synthseg_2021}:
\textit{(a)}
We draw a label map from a set of 3D segmentations with $N$ labels. We assume that these are defined on a grid of $J$ voxels at high resolution $r_{HR}$ (here \SI{1}{\milli\meter} isotropic). We then spatially augment the segmentation with a nonlinear transform (a smooth stationary velocity field~\cite{arsigny_log-euclidean_2006}) composed with three rotations, scalings, shearings, and translations. We call this augmented map $L$ (Fig.~\ref{fig:generation}a).
\textit{(b)}
We obtain an image $G=\{G_j\}_{j=1}^J$ by sampling a Gaussian Mixture Model (GMM) conditioned on $L$. All means and variances $\{\mu_n, \sigma^2_n\}_{n=1}^N$ are sampled from uniform priors to obtain a different random contrast at each minibatch~\cite{billot_learning_2020}. We also corrupt $G$ by a random bias field $B$, sampled in logarithmic domain:
\begin{equation}
p(G| L, \{\mu_n, \sigma^2_n\}_{n=1}^N) =
\prod_{j=1}^J \frac{B_j}{\sqrt{2\pi \sigma^2_{L_j}}} \exp [ -\frac{1}{2 \sigma^2_{L_j}} ( G_j B_j - \mu_{L_j} )^2]
\end{equation}
An image $I_{HR}$ is then formed by normalising the intensities of $G$ in $[0, 1]$, and nonlinearly augmenting them with a random voxel-wise exponentiation (Fig.~\ref{fig:generation}b).
\textit{(c)}
We then simulate LR scans with PV. This is achieved by blurring $I_{HR}$ with a Gaussian kernel $K$ of random standard deviation (to model slice thickness), and subsampling it to a random low resolution $r_{sp}$ (to simulate slice spacing). Finally, we form an image $I_{LR}$ by modelling the scanner noise with an additive field $\mathcal{E}$, sampled from a zero-mean Gaussian of random variance~(Fig.~\ref{fig:generation}c):
\begin{equation}
I_{LR} = Resample(I_{HR} \ast K, r_{sp}) + \mathcal{E}.
\end{equation}
\textit{(d)}
The final training image $I$ is obtained by resampling $I_{LR}$ back to $r_{HR}$, while the target segmentation for $S_1$ is built by grouping the cerebral regions of $L$ into the four tissue classes (Fig.~\ref{fig:generation}d). We emphasise that test scans will also be resampled to $r_{HR}$, since this enables us to segment on the target HR grid.
Training data for $S_2$ are sampled from the same model, but with two differences. First, the ground truths now contain all target labels (Fig.~\ref{fig:generation}e). Second, we still build tissue segmentations as in \textit{(d)} (since these are needed as inputs to $S_2$), but we now represent them as soft probability maps, which we randomly dilate/erode and spatially deform to model $D$ imperfections at test-time (Fig.~\ref{fig:generation}f).
\subsection{Training scheme for the denoising module}
\label{sec:training_d}
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/degradation.pdf}
\caption{Degradation model to train $D$. (a) A real image and its labels are spatially~deformed (b, f). The image is degraded (c,d) and fed to $S_1$ to obtain the input for $D$~(e).}
\label{fig:degradation}
\end{figure}
Recent denoising methods are mostly based on supervised CNNs trained to recover ground truth segmentations from artificially corrupted versions of the same maps~\cite{karani_test-time_2021,khan_deep_2021,larrazabal_post-dae_2020}. However, the employed corruption strategies are often handcrafted (random erosion and dilation, swapping of labels, etc.), and thus do not accurately capture errors made by the segmentation method to correct.
Instead, we propose to employ examples representative of $S_1$ errors, obtained by degrading real images, and feeding them to the trained $S_1$. $D$ is then trained to map the outputs of $S_1$ back to their ground truths. Images are degraded on the fly with the same steps as Section~\ref{sec:training_s} (except for the GMM, since we now use real images): spatial deformation, bias field, voxel-wise exponentiation, simulation of low resolution, and noise injection (Fig.~\ref{fig:degradation}). In practice, these corruptions use considerably wider prior parameter distributions than in Section~\ref{sec:training_s} (Table~\ref{tab:priors}), in order to ensure a high probability of erroneous segmentations from $S_1$.
\subsection{Implementation details}
All modules are trained separately using the average soft Dice as loss function~\cite{milletari_v-net_2016}. The segmentation modules $S_1$ and $S_2$ use the same 3D UNet architecture as in~\cite{billot_synthseg_2021}. Briefly, it comprises 5 levels of 2 convolution layers. All layers use 3$\times$3$\times$3 kernels and ELU activation~\cite{clevert_fast_2016}, except for the last layer, which employs a softmax. The first layer has 24 feature maps; this number is doubled after each max-pooling, and halved after each upsampling. Meanwhile, the denoiser $D$ uses a lighter structure chosen with a validation set: one convolution per layer, with a constant number of 16 features. Importantly, we delete the skip connections between the top two levels, to reach a compromise between UNets, where high-level skip connections propagate potential errors in the input segmentations to correct; and auto-encoders, with excessive bottleneck-induced smoothness.
\section{Experiments and Results}
\subsection{Brain MRI Datasets}
The training dataset consists of 1020 \SI{1}{\milli\meter} T1 brain scans (500 from ADNI~\cite{jack_alzheimers_2008}, 500 from HCP~\cite{van_essen_human_2012}, 20 from a private dataset~\cite{fischl_whole_2002}), which are available with a combination of manual and automated labels for 44 regions (Table~\ref{tab:regions}). We note that $S_1$ and $S_2$ are only trained with the label maps, and that using subjects from varied populations enables us to increase robustness to morphological variability.
We evaluate \textit{\textit{SynthSeg}$^{+}$} on 10,520 uncurated scans from the PACS of MGH. These are obtained from 1,047 MRI sessions of distinct subjects (Fig. S1), using a huge range of resolutions and contrasts (T1, T2, FLAIR, B0 channels from diffusion, among others). Among all sessions, 62 include T1 scans at maximum \SI{1.3}{\milli\meter} resolution, for which we obtain label maps by running FreeSurfer~\cite{fischl_freesurfer_2012}. Next, we propagate the resulting labels with rigid registration to all the scans of the corresponding sessions, which provides us with silver standard segmentations for 520 scans. Finally, these are split between validation (20), and testing (500), while the other 10,000 scans are used for indirect evaluation.
\subsection{Competing Methods}
We evaluate \textit{\textit{SynthSeg}$^{+}$} (i.e., $S_1 + D + S_2$) against four approaches.
\noindent\textbf{\textit{SynthSeg}~\cite{billot_synthseg_2021}}: We use the publicly available model for testing.
\noindent\textbf{Cascaded networks~\cite{roth_application_2018} ($\boldsymbol{S_1 + S_2}$)}: we ablate the denoiser $D$ to obtain an architecture that is representative of classical cascaded networks.
\noindent\textbf{Denoiser~\cite{larrazabal_post-dae_2020} (\textit{SynthSeg} $\boldsymbol{+ D}$)}: A state-of-the-art method for denoising by postprocessing, where a denoiser ($D$) is appended to the method to correct (\textit{SynthSeg}). Here, $D$ is trained as in Section~\ref{sec:training_d} to correct \textit{all} target labels.
\noindent\textbf{Cascaded networks with appended denoiser ($\boldsymbol{S_1 + S_2 + D}$)}: A combination of the cascaded architecture with the denoising network $D$.
All networks are trained for 300,000 steps (7 days on a Nvidia RTX6000) with the Adam optimiser~\cite{kingma_adam_2017}. Based on~\cite{billot_synthseg_2021}, we use Keras~\cite{chollet_keras_2015} and Tensorflow~\cite{abadi_tensorflow_2016}. Inference takes between 8 and 12 seconds for all methods on the same GPU.
\subsection{Quantitative analysis}
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/segmentations.pdf}
\caption{Segmentations obtained by all methods for scans where \textit{SynthSeg} shows large (``Big Fail''), mild (``Small Fail''), or no errors (``Pass''). Arrows indicate major mistakes. \textit{SynthSeg$^{+}$} yields outstanding results given the low SNR, poor tissue contrast, or low resolution of the inputs. Note that appending $D$ considerably smooths segmentations.}
\vspace{0.1cm}
\label{fig:segmentations}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\textwidth]{figures/dice.pdf}
\caption{Dice scores for the 500 test scans with ground truth. Results are presented based on the outcome of a QC analysis performed on the segmentations of \textit{SynthSeg}.}
\vspace{-0.6cm}
\label{fig:dice}
\end{center}
\end{figure}
First, we evaluate all methods on the 500 scans with ground truth. For visualisation purposes, we subdivide these scans into three classes based on the outcome of a visual quality control (QC) performed on the segmentations of \textit{SynthSeg} (Fig.~\ref{fig:segmentations}, third column): large errors (``Big fails'', 82 cases), mild errors (``Small Fails'', 103 cases) and good segmentations (``Passes'', 315 cases).
Figure~\ref{fig:dice} reveals that decomposing the target segmentation task into easier steps considerably improves robustness~\cite{roth_application_2018}. Indeed, the cascaded networks $S_1 + S_2$ outperforms \textit{SynthSeg} by 16.5 Dice points on the worst scans (i.e., Big Fails), while presenting much less outliers. Moreover, this strategy slightly increases accuracy, and leads to improvements by up to 2 points for small fails.
In comparison, appending a denoiser (either to \textit{SynthSeg} or cascaded networks) also improves robustness, but leads to a non-negligible loss in accuracy for good cases (down to 2 Dice points below \textit{SynthSeg}). This is due to the fact that $D$ cannot accurately model the distribution of convoluted boundaries and returns very smoothed segmentations (e.g., the cortex in Fig.~\ref{fig:segmentations}). Moreover, since $D$ does not have access to the input scan, its segmentations may deviate from the original anatomy, which may introduce biases when analysing populations with morphologies different from the training set.
Remarkably, \textit{SynthSeg$^{+}$} yields the best scores in all three categories, with an outstanding improvement of 23.5 Dice points over \textit{SynthSeg} for Big Fails. In comparison with cascaded networks, inserting a denoiser $D$ between $S_1$ and $S_2$ enables us to obtain robust tissue segmentations, which consistently improves scores by 2 to 5 Dice points. Finally, integrating $D$ within our framework (rather than using it for postprocessing) enables \textit{SynthSeg$^{+}$} to exploit both images and prior information when predicting final segmentations. As a result, our approach is more accurate than the two methods using $D$ for postprocessing (Fig.~\ref{fig:segmentations}), and outperforms them by at least 3 points.
\subsection{Volumetric study}
\label{sec:vol_study}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figures/volumes.png}
\caption{
Trajectories of cortical and hippocampal volumes with age (10,000 scans).
}
\label{fig:volumes}
\end{figure}
We now conduct a proof-of-concept volumetric study on the held-out 10,000 scans. Specifically, we analyse age-related atrophy using the volumes estimated by all methods. Our ageing model includes: B-splines with 10 equally spaced knots and soft constraints for motonoticity, linear terms for slice spacing in each acquisition direction, and a bias for gender. We then fit this model numerically by minimising the sum of squares of the residuals with the L-BFGS-B method~\cite{byrd_limited_1995}.
Figure~\ref{fig:volumes} shows the scatter plot and B-spline fits for the cerebral cortex and hippocampus (see Fig. S2 for the other regions). It reveals that the two methods using $D$ for postprocessing detect no or small atrophy, and over-estimate the volumes relatively to the other methods (due to the smoothing effect). Remarkably, \textit{SynthSeg$^{+}$} yields atrophy curves very close to those obtained in recent studies conducted on scans of much higher quality (i.e., \SI{1}{\milli\meter} T1 scans)~\cite{coupe_towards_2017,dima_subcortical_2022}. While \textit{SynthSeg} and $S_1 + S_2$ yield similar average trajectories, \textit{SynthSeg$^{+}$} produces far fewer outliers (especially at LR, see Fig. S3), which suggests that it can be used to investigate other population effects with much higher statistical power.
\section{Conclusion}
We have presented \textit{SynthSeg$^{+}$}, a novel hierarchical architecture that enables large-scale robust segmentation of brain MRI scans in the wild, without retraining. Our method shows considerably improved robustness relatively to \textit{SynthSeg}, while outperforming cascaded CNNs and state-of-the-art denoising networks. We demonstrate \textit{SynthSeg$^{+}$} in a study of ageing using 10,000 highly heterogeneous clinical scans, where it accurately replicates atrophy patterns observed on research data of much higher quality. By releasing the trained model, we aim at greatly facilitating the adoption of neuroimaging studies in the clinic, which has the potential to highly improve our understanding of neurological disorders.
\subsubsection{Acknowledgement} This work is supported by the European Research Council (ERC Starting Grant 677697), the EPSRC-funded UCL Centre for Doctoral Training in Medical Imaging (EP/L016478/1), the Department of Health's NIHR-funded Biomedical Research Centre at UCLH, Alzheimer’s Research UK (ARUK-IRG2019A-003), and the NIH (1R01AG070988, 1RF1MH123195).
\bibliographystyle{splncs04}
\input{biblio}
\newpage
\renewcommand{\thetable}{S\arabic{table}}
\renewcommand{\thefigure}{S\arabic{figure}}
\makeatother
\makeatletter
\title{Robust Segmentation of Brain MRI in the Wild with Hierarchical CNNs and no Retraining \\
Supplementary Materials}
\titlerunning{Robust Segmentation of Brain MRI with a Single Hierarchical CNN}
\author{B. Billot, M. Colin, S. Arnold, S. Das and J.E. Iglesias}
\institute{}
\maketitle
\begin{table}
\setlength\tabcolsep{7.5pt}
\caption{Ranges of the uniform distributions employed for the generative model (for the segmenters) and the degradation model (for the denoiser). The GMM assumes intensities in [0, 255]. The nonlinear transform and bias field are modelled by sampling a small tensor from a zero-mean Gaussian (respectively of variance $\sigma_v$ and $\sigma_B$), and upsampling it to $r_{HR}$. $\gamma$ represents the value of the voxel-wise exponentiation. $\sigma_{th}$ designates the standard deviation of $K$, and $\sigma_\mathcal{E}$ is the standard deviation of $\mathcal{E}$. Pilot experiments showed that training the segmenters with the degradation instead of the generative priors led to a considerable loss in performance.}
\centering
\label{tab:priors}
\begin{tabular}{|l|c|c|}
\hline
Parameters & Generative model prior & Degradation model prior \\
\hline
Rotation $(^\circ)$ & [-15, 15] & [-25, 25] \\
Scaling & [0.85, 1.15] & [0.5, 1.5] \\
Shearing & [-0.012, 0.012] & [0.02, 0.02] \\
Translation & [-20,20] & [-50, 50] \\
Nonlinear transform $\sigma_v^2$ & [0, 1.5] & [0, 4] \\
GMM means $\mu_k$ & [0, 255] & - \\
GMM variances $\sigma_k^2$ & [0, 6] & - \\
Bias field $\sigma_B^2$ & [0, 0.25] & [0, 3] \\
Intensity corruption $\gamma$ & [0.9, 1.1] & [0.3, 3] \\
Slice thickness $\sigma_{th}$ & [0.5, 5] & [0.5, 8] \\
Slice spacing $r_{sp}$ & [1, 9] & [1, 12] \\
Noise injection $\sigma_\mathcal{E}$ & [0, 10] & [0, 50] \\
\hline
\end{tabular}
\end{table}
\begin{table}
\setlength\tabcolsep{4pt}
\caption{List of the structures present in the training maps. Labels with different contralateral values are marked with $^{\text{R/L}}$, whereas predicted regions are noted with *.}
\centering
\label{tab:regions}
\begin{tabular}{|c|c|c|c|}
\hline
Background* & Thalamus$^{\text{R/L}}$* & Hippocampus$^{\text{R/L}}$* & Artery\\
Cerebral white matter$^{\text{R/L}}$* & Caudate$^{\text{R/L}}$* & Amygdala$^{\text{R/L}}$* & Optic chiasm \\
Cerebral cortex$^{\text{R/L}}$* & Putamen$^{\text{R/L}}$* & Accumbens area$^{\text{R/L}}$* & Soft tissues \\
Lateral ventricle$^{\text{R/L}}$* & Pallidum$^{\text{R/L}}$* & Ventral DC$^{\text{R/L}}$* & Mucosa\\
Inferior Lateral Ventricle$^{\text{R/L}}$* & $3^{rd}$ ventricle* & Cerebral vessels$^{\text{R/L}}$ & Skin\\
Cerebellar white matter$^{\text{R/L}}$* & $4^{th}$ ventricle* & Choroid plexus$^{\text{R/L}}$ & Skull\\
Cerebellar grey matter$^{\text{R/L}}$* & Brainstem* & CSF & Eyes \\
\hline
\end{tabular}
\end{table}
\begin{center}
\justifying
\includegraphics[width=\textwidth]{figures/subjects.png} \\
\small \textbf{Figure S2.} (a) Age/gender and (b) resolution distributions (individual scans in pink) for 10,520 scans acquired at MGH during 1,047 subject sessions.
\end{center}
\vspace{0.2cm}
\begin{center}
\justifying
\includegraphics[width=\textwidth]{figures/volumes_all_regions.png} \\
\small \textbf{Figure S3.} Volume trajectories for representative regions obtained with \textit{SynthSeg$^{+}$} on the 10,000 scans of Section~\ref{sec:vol_study}. We emphasise that these curves are very similar to those obtained in recent research studies (Coup\'e et al., 2017; Dima et al., 2022), which employed much higher quality scans (\SI{1}{\milli\meter} isotropic MP-RAGE sequences).
\end{center}
\vspace{0.2cm}
\begin{center}
\justifying
\includegraphics[width=\textwidth]{figures/volumes_6mm.png} \\
\small \textbf{Figure S4.} Volume trajectories computed for scans with slice spacing above \SI{6.5}{\milli\meter} (N$=587$). The competing approaches yield very uniform volumes across the age span, and produce many outliers. In contrast, \textit{SynthSeg$^{+}$} is much more robust: it detects atrophy patterns very similar to those obtained on all available subjects, and presents much less outliers than the other methods.
\end{center}
\end{document} | 833892a8d149cb6f932323a90a1826c527a96965 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{General Bids} \label{sec:general-bids}
It is natural to consider negative values and bids in a matching market.
Negative values model \emph{costs} incurred for a match, such as in a job market.
When a worker is matched as an employee to a company, she experiences some cost for which she must be compensated.
The goal of the matching market is to find an efficient price for her labor.
However, negative costs complicate matters because they also introduce negative bids.
A participant can make it difficult for a match to occur by bidding an arbitrary negative amount.
The result is that equilibrium is no longer sufficient for good welfare, even in bipartite graphs.
In this section, we first formalize the failure of equilibrium.
We then turn to \emph{stability} as a possible saviour.
We observe that approximate \emph{ex post} stability indeed implies good welfare in \emph{any mechanism}, and that the MM satisfies approximate ex post stability when participants are truthful.
However, ex post stability is a very strong requirement.
We formulate an alternative, \emph{ex ante} stability, and show that a ``rebate'' variant of the MM has approximately optimal welfare in any ex-ante stable \emph{Nash} equilibrium.
We explain the benefit of the rebate payment rule in terms of aligning individual surplus with market clearing without requiring detailed knowledge or communication from the participants.
Finally, however, we observe that this result does not naturally extend to Bayes-Nash equilibrium and illustrate the apparent difficulty involving coordinated communication.
\subsection{Model and failure of equilibrium}
We make one restriction on the general model of Section \ref{sec:prelim}: we assume the graph is bipartite.
This is primarily for notational and narrative convenience.
To make our presentation more intuitive, we adopt terminology in which the two sides of the bipartite market are asymmetric: One side (e.g. employers) are \emph{bidders}, while the other side (e.g. workers) are \emph{askers}.
The bidders are indexed by $i$.
They have values $v_{ij}$ and make bids $b_{ij}$.
The askers are indexed by $j$.
We assume they have \emph{costs} $c_{ij}$ and make \emph{asks} $a_{ij}$.
The surplus of a match between $i$ and $j$ is $s_{ij} = v_{ij} - c_{ij}$, the utility of an asker $j$ for being matched to $i$ is $u_j = -c_{ij} - \pi_i$, and so on.
In other words, for the askers, all values and bids have been negated and renamed.
We generally picture values, bids, costs, and asks all as positive numbers, so a bidder has a positive value for matching to an asker, who incurs a positive cost from the match.
However, our results are all fully general and would allow for any value, bid, cost, or ask to be either positive or negative.
\bo{TODO decide if strategy profile should still just be $b$ for simplicity}
\paragraph{The $1/4$-rebate Marshallian Match.}
This section considers on a variant of the payment rule, which has been proposed in \cite{waggoner2019matching} without theoretical results.
Namely, in the $1/4$-rebate MM, after a pair $(i,j)$ are matched and make their respective payments $b_{ij}$ and $-a_{ij}$, the mechanism returns to each participant $1/4$ of the total payment, i.e. each participant receives a rebate of $(b_{ij} - a_{ij})/4$.
Thus, half the reported surplus is kept by the mechanism as payment and half is returned to the participants.
We will discuss the benefit of this payment rule in Section \ref{subsec:ex_ante} when we prove a price of anarchy result.
\paragraph{Failure of equilibrium.}
When asks are allowed, equilibrium becomes insufficient to provide welfare guarantees.
Participants in the mechanism can place bids and asks such as to effectively refuse matches with one another, by asking above value or bidding below cost.
If two players both ``refuse'' matches with one another, neither can unilaterally fix the situation.
We prove this result for the MM with 1/4 rebate, but the same profile is also a zero welfare equilibrium with no rebate.
\begin{proposition} \label{prop:bad-welfare-bipartite}
In the bipartite setting with general values and costs, there always exist Nash equilibria of the Marshallian Match with zero welfare.
\end{proposition}
\begin{proof}
Consider a complete bipartite graph in which each bidder $i$'s values all satisfy $v_{ij} = 2$ and each asker $j$'s costs all satisfy $c_{ji} = 1$.
The optimal welfare is $n$ if there are $n$ bidders and $n$ askers.
An equilibrium profile is for each bidder $i$ to bid $b_{ij} = 0$ on all askers while each asker $j$ asks $a_{ji} = 4$ on all bidders.
The net bid on each edge is $-4$, so no matches are made and welfare is zero.
It is an equilibrium because no agent can benefit from a unilateral deviation:
Bidders $i$ can only cause a match by bidding $b_{ij} \geq 4$, in which case $i$'s utility would be at most $v_{ij} - 3b_{ij}/4 \leq -1$.
Similarly, askers $j$ can only cause a match by asking $a_{ji} \leq 0 < c_{ji}$. They would be matched at price 0 and obtain no rebate, giving negative utility.
\end{proof}
In this example, any single player cannot deviate alone to improve her welfare.
However, any pair of bidders sharing an edge can coordinate a deviation together and guarantee themselves higher welfare.
This pairwise deviation for improved welfare mirrors \emph{stability}, a key concept in matching algorithm design.
\subsection{Ex Post Stability}
In classical ``stable matching'' problems~\cite{gale1962college}, the goal is that, once the mechanism produces a final matching, no two participants $i,j$ both prefer to leave their assigned partners and switch to matching each other instead.
We use \emph{ex post} to refer to the fact that this evaluation occurs after all randomness and the matching's outcome have been realized.
In our setting, if $i$ and $j$ chose to match each other, they would obtain a net utility of their surplus $s_{ij} = v_{ij} - c_{ij}$.
If this amount is larger than their total utility in the mechanism, then they could switch to each other and split the surplus so as to make them both better off.
On the other hand, making this switch presumably involves some amount of friction.
Therefore, we introduce an approximate version of stability, as a more lenient requirement of a mechanism.
\begin{definition}[Approximate ex post stability]
A strategy profile $(b,a)$ in a mechanism is $k$-ex post stable if, for all realizations of costs and values and for all feasible pairs of bidder $i$ and asker $j$,
\[u_i(b,a) + u_j(b,a)\geq \frac{1}{k}s_{ij} .\]
\end{definition}
We note that ex post stability is an extremely strong notion. For any matching mechanism, an ex post stable strategy profile produces an approximately optimal matching.
\begin{observation} \label{obs:ex-post-approx}
For any matching mechanism, any $k$-ex post stable strategy is a $\tfrac{1}{k}$-approximation of the first-best welfare.
\end{observation}
\begin{proof}
Recall that in the Bayes-Nash setting agents' types are drawn from from a joint distribution $\mathcal{D}$.
Let $(b,a)$ be a $k$-ex post stable strategy profile for some matching mechanism, and $M$ the associated matching.
The expected welfare over all realizations of types and strategies is at least the total utility of participants,
\begin{align*}
\mathrm{Welfare}(b,a) \geq \E\left[\sum_{\{i,j\}\in M}u_i(b,a) + u_j(b,a)\right] = \E\left[\sum_{\{i,j\}\in M^*}u_i(b,a) + u_j(b,a)\right],
\end{align*}
where $M^*$ is the maximal matching by surplus.
Ex post stability gives a lower-bound on utility for all pairs, regardless of realization, so
\begin{align*}
\mathrm{Welfare}(b,a) \geq \E\left[\sum_{\{i,j\}\in M^*}\frac{s_{ij}}{k}\right] = \frac{1}{k}\mathrm{Welfare}(\textsc{Opt})
\end{align*}
\end{proof}
Observation \ref{obs:ex-post-approx} also arises directly from a primal-dual analysis of the linear program for bipartite matching, in which the dual variables are the utilities of the agents and $k$-approximate satisfaction of the dual constraints is precisely $k$-ex post stability.
We note that while there exist deferred-acceptance style ``stable'' matching mechanisms that technically involve money, such as matching with contracts~\cite{hatfield2005matching}, we have not ascertained if they can be made to satisfy approximate ex post stability in our sense.
\paragraph{Ex post stability of the Marshallian Match under truthfulness.}
Ideally, a matching mechanism would admit equilibria in ex post stable strategies, so that welfare would be high and participants would adhere to the outcomes of the mechanism.
But as a weaker requirement, we would like an indication of whether approximately ex post stable strategies might be reasonable in a mechanism.
We present a simple $2$-ex post stable strategy profile for the $1/4$-rebate MM, for all realizations of costs and values.
\begin{proposition} \label{prop:truthful-ex-post}
The truthful strategy in which all participants bid or ask their true value or cost is $4$-ex post stable in the $1/4$-rebate Marshallian Match.
\end{proposition}
\begin{proof}
When all participants truthfully report their values and costs, Marshallian Match produces the greedy maximum weighted matching $M$, where edge weights are supluses $s_{ij} = v_{ij} - c_{ij}$ (and negative edges are discarded).
Let $(b^*,a^*)$ denote the strategy profile in which all agents are truthful.
Consider an edge $\{i,j\}$. If $\{i,j\}\in M$, then $i$ and $j$ each receive utility exactly $s_{ij}/4$ via the mechanism's split.
If $\{i,j\}\notin M$, then either $i$ or $j$ must have obtained a higher-surplus match before $s_{ij}$. Without loss of generality, say $i$ matched to $k$ before $s_{ij}$, with $s_{ik} > s_{ij}$. By truthfulness, $i$ obtains welfare $u_i(b^*,a^*) = s_{ik}/4 > s_{ij}/4$. The same argument applies to $j$ if she matches first, implying that $u_i(b^*,a^*) + u_j(b^*,a^*) \geq s_{ij}/4$ for all $i$, $j$.
\end{proof}
Again, once one connects the truthful MM to greedy matching, Proposition \ref{prop:truthful-ex-post} follows from a basic primal-dual analysis combined with the rebate payment rule.
While truthful reporting from all participants produces approximately optimal social welfare (and the participants capture half of it), truthfulness is not in general even an approximate equilibrium.
We provide an example in the $1/4$-rebate MM setting in which a player can increase her welfare arbitrarily by overstating her true value for a match.
If players can make significant gains by deviating from the truthful strategy, it is likely that they will no adhere to an ex post stable profile.
\begin{example}[Figure \ref{fig:non-truthful}]
Consider the bipartite graph with three participants: $A$ and $B$ place bids on matching to $C$.
$C$ has cost $c_{CA} = c_{CB} = 0$ for both matches, $A$ has value $v_{AC} = k+1$ and $B$ has value $v_{BC} = k$ for matching with $C$.
In the truthful profile, $A$ is matched to $C$, and $B$ goes unmatched with welfare $0$.
If $B$ deviates to the non-truthful strategy of bidding $b'_{BC} = k+2$, $B$ would be matched with $C$ and would receive utility $u_B(b_{-B}^*,a^*,b') = k - (k+2) + (k+2)/4 = k/4-3/2$.
Thus, picking $k$ appropriately, $B$ can benefit arbitrarily by deviating from the truthful strategy.
\begin{figure}
\begin{center}
\begin{tikzpicture}[
every node/.style={circle,draw},scale=.75]
\node (A) at (-2,1.5) {$A$};
\node (B) at (-2,-1.5) {$B$};
\node (C) at (2,0) {$C$};
\draw (A) node[right,rectangle,above,sloped,draw=none,fill=none,outer ysep=8pt]{$v_{AC} = k+1$} -- (C) node[left,rectangle,above ,sloped,draw=none,fill=none,outer ysep=8pt]{$c_{CA} = 0$};
\draw (B) node[right,rectangle,above,sloped,draw=none,fill=none,outer ysep=8pt]{$v_{BC} = k$} -- (C) node[left,rectangle,below ,sloped,draw=none,fill=none,outer ysep=8pt]{$c_{CB} = 0$};
\end{tikzpicture}
\caption{In the $1/4$-rebate MM, $B$ is incentivized to deviate from truthfulness to the non-truthful bid $k+2$.}
\label{fig:non-truthful}
\end{center}
\end{figure}
\end{example}
\paragraph{Drawbacks of ex post stability.}
If one can prove that a mechanism is ex post stable in equilibrium, that is an ideal result as a welfare guarantee immediately follows.
However, without such a result, the value of ex post stability is questionable.
Consider the following somewhat subtle point.
Intuitively, it may seem reasonable to suppose that participants prefer stable strategy profiles as a sort of equilibrium refinement.
In particular, if strategy profile $b$ is not approximately stable, then (one would think) there are two participants $i$ and $j$ who would prefer to jointly switch their strategies so as to match to each other.
However, \emph{ex post} stability does not give this kind of guarantee.
It can only tell $i$ and $j$ whether they are satisfied after the mechanism happens.
To capture the above intuition, we turn to \emph{ex ante} stability.
\subsection{Ex Ante Stability}\label{subsec:ex_ante}
We now consider a model of stability that generalizes equilibrium by supposing no \emph{pair} of participants has an incentive to \emph{bilaterally} deviate.
Importantly, the incentive is relative to expected utility, so it involves an \emph{ex ante} calculation by the participants rather than ex post.
\begin{definition}[ex ante stability] \label{def:ex-ante}
A strategy profile $(b,a)$ is $k$-ex ante stable if, for all feasible pairs of bidder $i$ and asker $j$, for all strategies $b_i'$ of $i$ and $a_j'$ of $j$,
\[\E[u_i(b,a) + u_j(b,a)]\geq \frac{1}{k}\E[u_i(b_{-i},a_{-j},b_i',a_j') + u_j(b_{-i},a_{-j},b_i',a_j')] , \]
with randomness taken over realizations of types and strategies.
\end{definition}
That is, a profile is $k$-ex ante stable if there exists no deviation for any pair of players by which they could expect to increase their collective welfare by a factor of more than $k$.
There are two primary differences between ex ante and ex post stability.
First, ex ante applies to preferences ``before the fact'', i.e. in expectation, while ex post applies to preferences ``after the fact''.
Second, ex post stability postulates the ability of two participants to completely bypass the rules of the mechanism and match to each other.
In ex ante stability, the participants are limited to deviating to strategies actually allowed by the mechanism.
We observe that an ex ante stable profile is by definition a Bayes-Nash equilibrium, as one can in particular consider profiles where only one of the two participants deviates.
We show that for deterministic values and costs, ex ante stability is in fact sufficient to guarantee an approximation of optimal welfare.
\paragraph{Benefit of the rebate mechanism.}
Our price of anarchy result has a similar structure but very different underpinnings compared to the result for positive bids.
It relies crucially on the rebate version of the Marshallian Match, which returns to each participant $1/4$ of the total bid on the edge.
This might not appear to be a significant difference from the original version with no rebate; after all, wouldn't participants simply rescale their bids to achieve the same result?
However, in our smoothness approach, we obtain a significant benefit.
In the positive-values setting, $i$ can deviate to bids $v_{ij}/2$, knowing that each corresponding $j$'s bid is nonnegative.
In deviation, $i$ is assured that her net utility is exactly the price at which she is matched; earlier (higher price) meaning higher utility; and at the latest, $i$ matches at time $\max_j v_{ij}/2$.
This fails to hold when values and bids may be negative.
For example, $i$ might be bidding $b_{ij} = v_{ij}/2$, but if $j$ is asking $a_{ji} = v_{ij}/2 - \epsilon$, then this match (with high utility for $i$) will not happen until time $\epsilon$.
By that time, $i$ might be matched along an edge generating much less utility.
So $i$ cannot bid defensively in order to align the timing of matches with her own utility.
In the $1/4$-rebate MM, however, we recover this property.
If $i$ deviates to truthful bids $v_{ij}$, then any match at price $p(t)$ will result in a utility of exactly $v_{ij} - \pi_i + (v_{ij} - a_{ij})/4 = (v_{ij} - a_{ij})/4 = p(t)/4$.
The analogous fact holds for askers $j$.
So the $1/4$ rebate allows participants to align their utility with the order of market clearing without needing to know details of other participants' bids.
The final ingredient needed is ex ante stability: when $i$ deviates to truthfulness, she needs some partner $j$ to also do so, ensuring that some reasonable match is available.
\paragraph{Smoothness and deviation strategies.}
We define a pairwise deviation of $\{i,j\}$ to truthfulness for bids and asks as follows: $b_{i\ell}' = v_{i\ell}$ for all feasible neighbors $\ell$, and similarly $a_{j\ell}' = c_{j\ell}$ for all neighbors $\ell$.
For any pairwise deviation to truthfulness, we show that that pair collectively achieves a constant fraction of their welfare in the $1/4$-rebate MM.
\begin{lemma}\label{mechanism-splits-smoothness}
For any pair of feasible neighbors $i$ and $j$, for any strategy profile $(b,a)$,
\begin{equation*}
u_i(b_{-i}a_{-j},b_i'a_j') + u_j(b_{-i}a_{-j},b_i'a_j') \geq \frac{s_{ij}}{4} .
\end{equation*}
\end{lemma}
\begin{proof}
If $i$ and $j$ are still unmatched in deviation when the price reaches $s_{ij}$, then they will be matched and each get utility $s_{ij}/4$, since their bid-ask spread is $s_{ij}$.
If $i$ matches to some $\ell\neq j$ before $s_{ij}$, we claim $i$ still achieves high utility. Since $i$ is truthful, she obtains utility $u_i(b_{-i}a_{-j},b_i'a_j') = v_{i\ell} - b_{i\ell}' + (b_{i\ell}' - a_{\ell i})/4 = (b_{i\ell}' - a_{\ell i})/4 > s_{ij}/4$, since in deviation the bid-ask spread on the edge $(i,\ell)$ is greater than that on edge $(i,j)$.
If $j$ matches to some $\ell\neq i$ before $s_{ij}$, then $j$ achieves high utility. Since $j$ is truthful, she obtains utility $u_j(b_{-i}a_{-j},b_i'a_j') = a_{\ell j}' - c_{\ell j} + (b_{\ell j} - a_{j \ell}')/4 = (b_{\ell j} - a_{j \ell}')/4 > s_{ij}/4$, again since $(\ell,j)$ has a higher bid-ask spread in deviation than $(i,j)$.
\end{proof}
\begin{theorem} \label{thm:mechanism-splits-PoA}
For deterministic values and costs, and strategy profile $(b,a)$ that is $k$-ex ante stable, the $1/4$-rebate Marshallian Match achieves a $\frac{1}{4k}$-approximation of the optimal expected welfare.
\end{theorem}
Surprisingly, in this result, the contribution of payments to overall welfare can be ignored.
We will actually obtain that total participant surplus is a constant fraction of the optimal welfare.
\begin{proof}[Proof of Theorem \ref{thm:mechanism-splits-PoA}]
Let $M$ denote the matching produced by the strategy profile $(b,a)$, and $M^*$ denote the first-best welfare matching. The welfare of $M$ is
\begin{align*}
\textsc{Welf}(b,a) &= \E\left[\sum_{\{i,j\}\in M}u_i(b,a) + u_j(b,a) + p_i(b,a)\right]\\
&\geq \E\left[\sum_{\{i,j\}\in M^*}u_i(b,a) + u_j(b,a)\right]\\
&\geq \E\left[\frac{1}{k}\sum_{\{i,j\}\in M^*}u_i(b_{-i},a_{-i},b_i',a_j') + u_j(b_{-i},a_{-i},b_i',a_j')\right]
\end{align*}
since the values and costs of agents are deterministic and common knowledge, each agent can compute her partner in the first-best welfare matching, and can coordinate to deviate as a pair, guaranteeing a lower bound on welfare for both. Applying Lemma \ref{mechanism-splits-smoothness},
\begin{align*}
\textsc{Welf}(b,a) \geq \frac{1}{k}\E\left[\sum_{\{i,j\}\in M^*} \frac{s_{ij}}{4}\right] = \frac{1}{4k}\textsc{Welf}(\textsc{Opt})
\end{align*}
\end{proof}
This result relies heavily on participants' ability to calculate the fixed matching with optimal social welfare, and coordinate deviations with their partner.
We exploit the deterministic costs and values considered to fix the optimal matching $M^*$ across realizations of potentially mixed strategies.
\subsection{Discussion: failure of the proof in the Bayes-Nash setting}
In this section, we discuss the main open problem: proving welfare guarantees in the Bayes-Nash setting under a reasonable stability definition.
\begin{openprob}
Give a ``reasonable'' stability assumption and a mechanism for the matching model of \ref{sec:prelim}, with general values, such that, in the Bayes Nash setting, every ``stable'' strategy profile generates a constant factor of the optimal expected welfare.
\end{openprob}
In the Bayes-Nash setting, the distributions over values and costs are common knowledge, but their realizations are not.
As a result, the optimal matching $M^*$ is dependent on the realizations of players' types.
This causes our proof of welfare approximation under ex ante stability to fail in an interesting way.
In fact, even the definition of ex ante stability (Definition \ref{def:ex-ante}) has nuanced implications.
Consider a complete bipartite $n \times n$ graph with a ``star-crossed lovers'' distribution on types\footnote{One can create a version where type distributions are independent that makes roughly the same point.}: there is a uniformly random perfect matching $M^*$ in which the edges have positive surplus, while all edges not in $M^*$ have high costs and low values.
Even if we take a very bad mechanism, such as one that always assigns the same matching $M$ regardless of types, it can be ex ante stable: any particular pair are getting low utility, but if they are able to jointly deviate to matching with each other deterministically, they also get low utility in expectation over the type distribution.
They are unlikely to be fated for each other once the types are realized.
Similarly, recall that the proof of Theorem \ref{thm:mechanism-splits-PoA} used ex ante stability in the following step:
\begin{align*}
\E\left[\sum_{(i,j)\in M^*}u_i(b,a) + u_j(b,a)\right]
&\geq \E\left[\frac{1}{k}\sum_{(i,j)\in M^*}u_i(b_{-i},a_{-i},b_i',a_j') + u_j(b_{-i},a_{-i},b_i',a_j')\right] .
\end{align*}
In the Nash setting, $M^*$ was fixed, and so were the deviation strategies $b_i',a_j'$ which depended on $M^*$.
So the expectation could move inside the sum, followed by an application of ex ante stability.
But in a Bayes-Nash setting, $M^*$ is a random variable depending on the realizations of the types.
We cannot move the expectation inside the sum.
A tempting fix is some sort of \emph{ex interim stability} assumption, where the deviation strategies of $i$ and $j$ can depend on the types of both players.
In the star-crossed lovers example, this is appealing: the pairs with high surplus know this fact from their types and can easily coordinate a deviation.
So it is reasonable to assume that strategy profiles played in the mechanism are not much worse than such coordinated deviations.
However, the amount of coordination required grows significantly if type distributions are more complicated.
For example, suppose each value and cost comes from an independent power-law distribution.
Finding blocking joint deviations seems to require significant knowledge by the agents, perhaps of global properties of the type space (such as who would be matched under the greedy matching or optimal matching).
To assume agents play strategies that eliminate such deviations appears to unjustly relieve the mechanism of responsibility to coordinate agents' information and decisionmaking.
\section{Conclusion and Future Work} \label{sec:conclusion}
Two-sided matching is difficult, even in such simplified abstract models as in this paper, for at least three reasons:
\begin{itemize}
\item We would like the process to accommodate information acquisition in a way that is compatible with optimal search theory.
\item The process is generally dynamic and sequential (for the previous reason), and strategic behavior in dynamic settings is complicated.
\item The constraints are complex and interlocking, i.e. $i$ can match to $j$ if and only $j$ matches to $i$, yet their preferences over this event can be conflicting and contextualized by their other options.
\end{itemize}
The variants of the Marshallian Match studied in this paper address each challenge to some extent.
At least in the nonnegative values setting, the MM is compatible with optimal search because matches are coordinated to occur approximately in order from highest surplus to lowest.
This allows inspections to occur approximately in order of their ``index'' (``strike price'') from the Pandora's box problem.
In particular, it enables the useful technical property of ``exercising in the money'', i.e. a bidder who inspects at a late stage and discovers a very valuable match is able to unilaterally lock in that match.
Dynamic strategizing is addressed by strictly limiting information leakage, i.e. each participant can only see their own bid and learn when they match.
It is unclear whether this feature actually makes the mechanism better, but it does make it easier to analyze.
We believe that all of our results extend if participants are able to observe when any match occurs, but this would require careful formalization as the game becomes truly dynamic in that case.
The complex constraints are addressed to an extent by the coordination of matches in order of value.
A bidder $i$ in the MM, whether pay-your-bid or $1/4$-rebate, has deviation strategies available in which the timing of their match corresponds to their utility.
If another participant bids high enough to lock in a match with $i$ early, this is out of $i$'s direct control, but $i$ can bid so that they are assured of enough utility to make the match worthwhile.
\paragraph{Future work.}
There are a number of appealing variants on the model and directions for future investigation.
In the job market application, an interview is a \emph{simultaneous} inspection event between a worker and employer.
Can the MM's welfare guarantees extend to a model where inspection is simultaneous, even in the positive-values setting?
Other variations can include multiple stages of inspection (an extension in \citet{kleinberg2016descending}) or matching where inspection is optional (studied algorithmically by \citet{beyhaghi2019pandora}).
Another direction involves reasonable stability assumptions and their impact on monetary mechanism design for matching.
For example, a strategy profile seems somewhat unlikely if it allows for the following sorts of ``Stackelberg deviations'': participant $i$ announces a deviation strategy to all of their neighbors on the other side of the market, and those neighbors respond with their own deviations (perhaps best responses).
Finally, the MM itself admits a number of possible variations.
One additional benefit of the rebate payment rule is that it disincentivizes overbidding and under-asking, because participants do not need to strategically bid to capture value: the payment rule returns it to them as a rebate.
A more sophisticated approach involves a two-part bid consisting of a reserve bid $b_{ij}$ and a ``surplus bid'' $\beta_{ij} \in (0,1)$ (respectively for askers, a reserve $a_{ji}$ and surplus $\alpha_{ji}$).
When a match occurs at $p(t) = b_{ij} - a_{ij}$, the bidder receives a rebate of $\beta_{ij} p(t)$ while the asker receives a rebate of $\alpha_{ji} p(t)$.
Perhaps a variant like this achieves a price of anarchy result, or a stability-based welfare guarantee, for the general setting.
\section{Introduction}
A primary goal of designing mechanisms is to coordinate groups to arrive at collectively good allocations or outcomes.
For example, in auctioning a set of items to unit-demand buyers, the problem is to coordinate among the varied preferences of the buyers to achieve an overall good matching of buyers to items.
In such auction settings, ``good'' is usually formalized via \emph{price of anarchy}~\citep{roughgarden2017price}: in any equilibrium of the auction game, the expected social welfare (total utility) should be approximately optimal.
Matching people to people, with preferences on both sides, appears to require even more coordination.
\citet{gale1962college} introduced the foundational deferred-acceptance algorithm -- a matching mechanism without money -- for participants with ordinal preferences.
The key ``good'' property it achieves (if all participants are truthful) is \emph{stability}: no pair prefers to switch away from the given matching and match to each other instead.
Variants of deferred acceptance have had significant impact in applications from kidney exchange to the National Residency Matching Program (NRMP) for doctors and hospitals~(e.g. \citet{iwama2008survey}).
We are motivated by two drawbacks of deferred-acceptance-style approaches.
First, the \emph{social welfare} generated by such mechanisms is unclear, even in settings where money is explicitly modeled such as matching with contracts~\citep{hatfield2005matching}.
While their criterion of stability is a nice property, its relationship to welfare is not obvious.
We would like to investigate this relationship and obtain explicit welfare guarantees.
Second, it is unclear the extent to which such mechanisms are compatible with \emph{inspection} stages in which participants must invest effort to discover their preferences.
For example, in practice, the design of the NRMP requires relatively expensive and time-constrained interviews, which must be completed before matching begins.
While such concerns have motivated significant work on information acquisition in matching markets, particularly variants of deferred acceptance~(e.g. \cite{immorlica2021designing}; see Section \ref{sec:related-work}), to our knowledge none of it incorporates monetary mechanisms with quantitative welfare guarantees.
On the other hand, prior work of \cite{kleinberg2016descending} has shown that even in the special case of matching people to items (which have no preferences), approximately optimal welfare \emph{requires} a market design with dynamically interspersed matching and information acquisition.
\cite{kleinberg2016descending} showed that descending-price mechanisms tend to be compatible with costly inspection stages and still yield high social welfare, due to a connection with the Pandora's box problem~\citep{weitzman1979optimal}.
\paragraph{The Marshallian Match.}
Inspired by \citet{kleinberg2016descending}, \citet{waggoner2019matching} propose the ``Marshallian Match'' (MM) for two-sided matching with money.
In \citet{marshall1920principles}, its namesake describes a theory of market clearing in which the buyer-seller matches that generate the largest surplus -- i.e. the most net utility between the pair -- occur first, and so on down.\footnote{This dynamic eventually leads to the market clearing price (e.g. \citet{plott2013marshall}), after which no more positive-surplus matches are possible. Indeed, in a simple commodity market it is possible to ignore Marshall's dynamics and focus on the calculation of the clearing price. But in a more complex two-sided matching problem, we appear to require dynamics in order to properly coordinate matches.}
A similar dynamic is observed in decentralized matching markets~\cite{chade2017sorting}, yet it does not directly underly standard centralized designs such as deferred acceptance.
Similarly, the MM begins with a high price that descends over time.
Participants maintain a bid on each of their potential matches, with the sum of the two bids ideally representing the total surplus generated by the match.
When the price reaches the sum of any pair's bids on each other, that pair is matched.
They pay their bids and drop out of the mechanism, which continues.
\citet{waggoner2019matching} speculate on the dynamics, strategy, and benefits of this mechanism, but do not obtain theoretical results.
\subsection{Our results}
We first consider a setting where all participants' values are nonnegative.
Under this restriction, we show a general \emph{price of anarchy (PoA)} guarantee for the Marshallian Match, i.e. in any Bayes-Nash equilibrium the expected social welfare is within a constant factor of the optimal possible.
We use a smoothness approach (e.g. \cite{roughgarden2017price}) along with several key properties of our variant of the MM, including limited information leakage.
The positive-bids setting can model, for example, matching of industrial plants to geographic areas (an application of \cite{koopmans1957assignment}) or matching local businesses to municipal-owned locations.
We also show robustness of this result via several extensions.
It not only holds for non-bipartite graphs, but also extends to a \emph{group formation} setting.
There, agents must be partitioned into subsets of size at most $k$, the edges of a hypergraph.
Each agent $i$ has a private valuation for joining each possible subset.
We modify the MM by keeping the global descending price and clearing a subset when the price reaches the sum of its bids.
We obtain a $\Omega(1/k^2)$ price of anarchy for this problem.
Next, we show that the welfare guarantee also extends to a model with information acquisition costs.
It is natural in many matching settings that participants do not initially know their values for a potential match.
It requires time, effort, and/or money to investigate and discover one's value.
A good mechanism should carefully coordinate these investigations to happen at appropriate times, or else significant welfare will be lost: participants will either waste too much utility on unnecessary inspections, or they will forego valuable matches due to the inspection cost and uncertainty about the match.
Adapting techniques of \cite{kleinberg2016descending} for analyzing welfare in models of inspection, we show that the price of anarchy of the MM is also constant in this setting.
\begin{theorem*}
With nonnegative values, the Marshallian Match has the following guarantees:
\begin{enumerate}
\item For matching on general graphs, a Bayes-Nash price of anarchy of at least $1/8$.
\item For matching on hypergraphs with group size at most $k$, a Bayes-Nash price of anarchy of at least $1/2k^2$.
\item For matching on general graphs with inspection costs, a Bayes-Nash price of anarchy of at least $1/8$.
\end{enumerate}
\end{theorem*}
\paragraph{Possibly-negative values.}
We then consider a general two-sided matching setting where values may be negative.
Understanding this setting is desirable because it more accurately captures job markets, where workers incur a cost (i.e. negative value) for being matched to a job and must be compensated more than that cost.
Here we see for natural reasons that a PoA result for the MM is impossible:
if all participants set their bids so as to refuse all matches, no single participant can deviate to cause any change and the equilibrium obtains zero welfare.
In general, this suggests that a kind of stability condition may be natural and necessary for high-welfare matching mechanisms of any kind.
We observe that \emph{approximate ex post} stability implies approximately optimal welfare, and show that the MM achieves approximate ex post stability if participants bid truthfully.
We use a simple ``rebate'' tweak to return half this surplus to the bidders, resulting in a generally favorable outcome.
\begin{proposition*}
In any $k$-approximate ex post stable strategy profile, the $(1/4)$-rebate Marshallian Match (in fact, \emph{any} mechanism) achieves at least $1/k$ of the optimal expected welfare.
Truthful bidding in the $(1/4)$-rebate Marshallian Match is $4$-approximately ex post stable.
\end{proposition*}
However, truthfulness is not generally an equilibrium.\footnote{We note that even for deferred acceptance, which is also stable \emph{if all participants are truthful}, in general only one side of the market optimizes their outcomes by being truthful.}
This raises the question of whether it is reasonable to assume participants will adopt approximately stable strategy profiles.
We argue that ex post stability is too strong an assumption, and instead propose \emph{ex ante} stability.
We show that in a Nash setting (not Bayes-Nash) with fixed valuations, if strategies are approximately \emph{ex ante stable}, then the welfare of MM is approximately optimal.
This result crucially relies on the ``rebate'' payment rule: it allows participants the option of a strategy deviation that aligns their personal utility with the order of market clearing.
This variant of the payment rule appears necessary to coordinate market clearing by amount of surplus while limiting the amount of communication and knowledge required.
\begin{theorem*}
In the general Nash setting with fixed valuations, in any strategy profile that is $k$-approximate ex ante stable, the $(1/4)$-rebate Marshallian Match achieves at least a $1/4k$ fraction of the optimal welfare.
\end{theorem*}
In fact, the stability property also ensures that participants keep a large fraction of the welfare; the proof uses that participant surplus alone (i.e. welfare minus payments) is at least $1/4k$ of the optimal welfare.
Unfortunately, this stable-price-of-anarchy result is fragile and the proof does not extend to the \emph{Bayes-Nash} setting where private valuations are drawn from a common-knowledge prior.
This is roughy due to the difficulty of coordinating and communicating deviations between ``blocking pairs''.
Therefore, the main problem of proving a welfare guarantee in a general negative-bids setting and under a reasonable stability assumption remains open.
\begin{openprob*}
Give a well-justified stability assumption and a mechanism such that, in the \emph{Bayes-Nash setting} with general values, stable strategy profiles guarantee a constant factor of the optimal expected welfare.
\end{openprob*}
\subsection{Related work} \label{sec:related-work}
We have not found any mechanisms in the literature involving a two-sided matching market with money and quantifiable welfare results.\footnote{One can always apply a general Vickrey-Clarke-Groves (VCG) mechanism, which has an equilibrium with optimal social welfare. But VCG is undesirable because it appears incompatible with both price of anarchy results and models with inspection costs, cf. \citet{kleinberg2016descending}.}
However, the literature on matching with strategic agents is very broad, including with inspection stages, and we highlight a number of papers that are related to our problem.
Probably closest to our work is \citet{immorlica2021designing}, which considers design of a platform to coordinate two-sided matchings with inspection costs and quantifiable welfare guarantees.
Motivated by e.g. matching platforms for romantic dating, the paper studies agents coming from specific populations with a known distribution of types.
The mechanism uses its knowledge of the type distributions to compute strategies for directing inspection stages (e.g. first dates).
The computational problem is challenging and intricate.
\citet{immorlica2021designing} is able to show the structure of equilibria and use this to give welfare guarantees, all in a setting without money.
In contrast, we are interested in monetary mechanisms, motivated (eventually) by e.g. labor markets.
We consider a very simple descending-price mechanism that has no access to knowledge about the agent types or distributions.
Beyond \citet{immorlica2021designing} there is an extensive literature on matching marketplaces and dynamics.
Work in that literature involving money (transferable utility) historically often takes a Walrasian equilibrium perspective, while ours is in the tradition of auction design.
We refer to the survey of \citet{chade2017sorting} for more on this literature.
As mentioned above, a number of recent works study matching markets with information acquisition, but generally consider variants of deferred acceptance without money.
Works of this kind include \citet{he2020application,che2019efficiency,ashlagi2020clearing,chen2021information,fernandez2021centralized,immorlica2020information,hakimov2021costly}.
Among these, \citet{immorlica2020information} uses a lens of optimal search theory similar to ours.
It studies matching of students to schools.
Its matching problem is almost one-sided, in the sense that schools have known and fixed preferences.
The focus is on coordinating efficient acquisition of information by students.
Unlike our social-welfare perspective, that work focuses on the more standard criterion of stability and introduces regret-free stable outcomes.
In particular, it does not involve money.
Similarly, \citet{hakimov2021costly} study a serial-dictator mechanism for coordinating student inspection without money.
We refer to \citet{immorlica2020information} for an extensive discussion of further work related to information acquisition in matching markets.
Our notions of stability are naturally closely related to others in the literature.
Ex post stability is only a quantitative version of stability in matching; a more sophisticated version of it is used by \citet{immorlica2020information}, for example.
\cite{fernandez2021centralized} also utilizes a similar notion of stability, in a setting of incomplete information.
\section{Results for Positive Valuations}
In this section, we consider a restriction of the general setting where all values $v_{ij}$ are nonnegative.
First, we show that the Marshallian Match achieves a constant approximation of optimal welfare for matching.
Second, we extend the result to the group formation (i.e. matchings on hypergraphs) setting.
Finally, we extend the result in a different direction to the case where participants do not initially know their valuations and can choose to expend effort to discover them.
\subsection{The vanilla positive valuations model}
Here, we take the general model of Section \ref{sec:prelim} and assume that each valuation satisfies $v_{ij} \geq 0$.
We modify the Marshallian Match to require all bids $b_{ij}$ to be nonnegative at all times.
\paragraph{Intuition.}
The nonnegative MM is similar to running multiple interlocking descending-price unit-demand auctions simultaneously.\footnote{Although general price of anarchy results are available for these kinds of auctions for goods, e.g. \citet{lucier2015greedy,feldman2016price}, we do not know of any that apply to two-sided matching.}
This parallel is most obvious in the bipartite case: any bidder $i$ competes against other bidders in the same set for her favorite matches.
Extending this perspective, each bidder $i$ could hypothetically bid as though her neighbors $j$ were simply items with no preferences.
By analyzing the failure of this hypothetical strategy, we find in the MM that it is connected with a different high-welfare event, namely $j$ matching early.
This intuition underlies our smoothness lemma, discussed next.
\paragraph{Smoothness for two-sided matching.}
We give a smoothness lemma that powers our price of anarchy result.
Recall (e.g. \citet{roughgarden2017price}) that smoothness proofs of PoA proceed by guaranteeing high-welfare events in a counterfactual world where $i$ deviates to a less-preferred strategy.
One challenge is that in a dynamic mechanism that proceeds over time, a deviation could cause chain reactions that make it impossible to reason about the outcomes.
Here, we rely on our variant of MM that does not leak any information about bids or strategies.
The only piece of information the agent receives from the mechanism comes at the moment they are matched, after which they cannot react.
A second key challenge is that in a matching market, $i$ may not match $j$ \emph{and} the prices that $i$ and $j$ each pay may still be low, obstructing an adaptation of a standard smoothness proof.
In the MM there is, however, a high \emph{total} price paid for an edge that obstructs the match.
Recall that while $\pi_i(b)$ is $i$'s payment, $p_i(b)$ is the \emph{total} payment on the edge containing $i$ that is matched (zero if $i$ is unmatched).
\paragraph{The deviation and smoothness lemma.}
Let $b$ be any strategy profile. For any bidder $i$, define the deviation strategy $b_i'$ as $b_{ij}'(t) = v_{ij}/2$ for all feasible neighbors $j$ and all times $t$.
Observe that if $i$ matches at any price $p(t)$ in deviation, their utility is $p(t)$.
In other words, the key property of the Marshallian Match in combination with this deviation strategy is that it coordinates surplus generation with the order of matching.
\begin{lemma} \label{lemma:nonneg-smooth}
In the nonnegative values setting, for any feasible pair $\{i,j\}$, any strategy profile $b$, and any realization of types,
\begin{align}
u_i(b_{-i},b_i') + \frac{p_i(b)}{4} + \frac{p_j(b)}{4} &\geq \frac{v_{ij}}{8} .
\end{align}
\end{lemma}
\begin{proof}
Observe that all three terms are nonnegative.
For $u_i(b_{-i},b_i')$, this follows by definition of $b_i'$.
Therefore, if $p_j(b) \geq v_{ij}/2$ or $p_i(b)\geq v_{ij}/2$, then the result is immediate.
Otherwise, under $b$, both $i$ and $j$ are unmatched by the time the price has dropped to $v_{ij}/2$.
Observe that $b$ and $(b_{-i},b_i')$ are indistinguishable to all participants until $i$ matches, which at the latest is to $j$ at time $v_{ij}/2$.
By definition of $b_i'$, it follows that $u_i(b_{-i},b_i') \geq v_{ij}/2$.
\end{proof}
\begin{theorem} \label{thm:nonneg-welfare}
In the nonnegative values setting, the Marshallian Match has a Bayes-Nash price of anarchy of at least $1/8$.
\end{theorem}
\begin{proof}
Letting $b$ be any Bayes-Nash equilibrium and $M^*$ be the optimal matching (a random variable), we have the following.
We will use that, in equilibrium, $i$ prefers $b_i$ to $b_i'$; that $M^*$ contains at most every participant, and utilities and payments are nonnegative; and Lemma \ref{lemma:nonneg-smooth}.
\begin{align*}
\mathrm{Welfare}(b) &= \E \sum_i \left(u_i(b) + \frac{p_i(b)}{2}\right) \\
&\geq \E \sum_i \left(u_i(b_{-i},b_i') + \frac{p_i(b)}{2}\right) \\
&\geq \E\left[\sum_{\{i,j\}\in M^*} \left(u_i(b_{-i},b_i') + u_j(b_{-j},b_j') + \frac{p_{i}(b)}{2} + \frac{p_{j}(b)}{2}\right)\right]\\
&= \E\left[\sum_{\{i,j\}\in M^*} \left(u_i(b_{-i},b_i') + \frac{p_{i}(b)}{4} + \frac{p_{j}(b)}{4} + u_j(b_{-j},b_j') + \frac{p_{i}(b)}{4} + \frac{p_{j}(b)}{4} \right) \right]\\
&\geq \E\left[\sum_{\{i,j\}\in M^*} \left( \frac{v_{ij}}{8} + \frac{v_{ji}}{8} \right) \right] = \frac{1}{8}\mathrm{Welfare}(\textsc{Opt}) .
\end{align*}
\end{proof}
\subsection{Matchings on hypergraphs}
We now extend to the problem of coordinating formation of groups of size up to $k$.
We will be brief because the proof is similar to the matching case above, i.e. the special case where $k=2$ and groups of size one are disallowed.
Instead of a graph, we are given a hypergraph where agents are vertices and a hyperedge $S$ represents a feasible group, i.e. subset of agents of size at most $k$.
The value of agent $i$ for being assigned to group $S$ is $v_{iS} \geq 0$.
A ``matching'' or assignment $M$ consists of a subset of the hyperedges such that no agent is in two different groups $S,S' \in M$.
The surplus of a group $S$ is $s_S = \sum_{i \in S} v_{iS}$.
The social welfare of an assignment $M$ is $\sum_{S \in M} s_S$.
The Marshallian Match for this setting is modified as follows.
Participants $i$ maintain bids $b_{iS}$ on all of their feasible groups $S$.
When the descending price matches the total sum of bids on any group, i.e. $p(t) = \sum_{i \in S} b_{iS}$, that group is matched and drops out of the mechanism; all members pay their bids.
Again considering the deviation $b_i'$ where $b_{iS}'(t) = v_{iS}/2$ for all feasible $S$ and all $t$, we have:
\begin{lemma} \label{lemma:group-smoothness}
For any strategy profile $b$, realizations of types, agent $i$, and hyperedge $S$ containing $i$,
\[ u_i(b_{-i},b_i') + \frac{1}{k^2}\sum_{j \in S} p_j(b) \geq \frac{v_{iS}}{2k^2} . \]
\end{lemma}
\begin{proof}
All terms are nonnegative.
If under $b$ there exists $j \in S$ with $p_j(b) \geq \frac{v_{iS}}{2}$, then we are done.
Otherwise, in $(b_{-i},b_i')$ all members of $S$ are unmatched at least until $i$ matches, which at the latest occurs at $p(t) = v_{iS}/2$, yielding $i$ utility at least $v_{iS}/2$.
\end{proof}
\begin{theorem} \label{thm:group-welfare}
In the hyperedge matching (group formation) setting with nonnegative values and group size up to $k$, the Marshallian Match has a Bayes-Nash price of anarchy of at least $\frac{1}{2k^2}$.
\end{theorem}
\begin{proof}
Letting $b$ be a Bayes-Nash equilibrium and $M^*$ the optimal assignment (a random variable), we have the following.
The first line follows because groups have size at most $k$, so the total payment of a group $S$ is at least $\sum_{i \in S} p_i(b)/k$.
\begin{align*}
\mathrm{Welfare}(b) &\geq \E \sum_i \left( u_i(b) + \frac{p_i(b)}{k} \right) \\
&\geq \E \sum_i \left( u_i(b_{-i},b_i') + \frac{p_i(b)}{k} \right) \\
&\geq \E \sum_{S^* \in M^*} \sum_{i \in S^*} \left( u_i(b_{-i},b_i') + \frac{p_i(b)}{k} \right) \\
&\geq \E \sum_{S^* \in M^*} \sum_{i \in S^*} \left( u_i(b_{-i},b_i') + \sum_{j \in S^*} \frac{p_j(b)}{k^2} \right) \\
&\geq \E \sum_{S^* \in M^*} \sum_{i \in S^*} \frac{v_{iS^*}}{2k^2}
\quad= \frac{\mathrm{Welfare}(\textsc{Opt})}{2k^2} .
\end{align*}
\end{proof}
It remains to be seen if the factor can be improved to $\Omega(1/k)$.
We appear to lose one factor of $k$ because the greedy algorithm (e.g. the MM where all participants are truthful) is only a $1/k$ approximation to optimal, and then another factor from strategic behavior.
\subsection{Inspection} \label{sec:inspection}
We now extend our welfare result for graphs to a model with information acquisition costs.
We again do not require the graph to be bipartite.
The model is augmented as follows, following e.g. \citet{kleinberg2016descending}.
The type of a participant consists of, for each feasible partner $j$, a cost of inspection $r_{ij} \geq 0$ and a distribution $D_{ij}$ over the nonnegative reals.
Our setting is Bayes-Nash, i.e. all types are drawn jointly from a common-knowledge prior.
When $i$ inspects an edge $\{i,j\}$, they incur a cost of $r_{ij}$ and observe a value $v_{ij} \sim D_{ij}$ independently of all other randomness in the game.
The cost $r_{ij}$ can model a financial investment, or the cost of time or effort required for $i$ to learn their value $v_{ij}$.
Let $I_{ij} \in \{0,1\}$ be the random variable indicator that $i$ inspects $j$ and let $A_{ij} \in \{0,1\}$ be the indicator that $i$ is matched to $j$.
We adopt the standard assumption (although recent algorithmic work of \citet{beyhaghi2019pandora} has weakened it) that $i$ must inspect $j$ prior to being matched to $j$; i.e. if the match occurs, $i$ must inspect and incur cost $r_{ij}$ if they haven't yet.
In other words, $A_{ij} \leq I_{ij}$ pointwise.
We also assume that, in the game, inspection is instantaneous with respect to the movement of the price $p(t)$.
An agent $i$'s utility is their value for their match (if any) minus the sum of all inspection costs and the price they pay.
Formally, we have $u_i(b) = \sum_j \left(A_{ij} v_{ij} - I_{ij} r_{ij}\right) - \pi_i(b)$, where the sum is over feasible neighbors.
Social welfare is the sum of all agent utilities and revenue, i.e. $\mathrm{Welfare} = \sum_{i,j} \left( A_{ij} v_{ij} - I_{ij} r_{ij} \right)$.
\paragraph{Covered call values and exercising in the money.}
\citet{kleinberg2016descending} give technical tools, based on a solution of the Pandora's box problem~\citep{weitzman1979optimal}, utilizing finance-inspired definitions.\footnote{We refer the reader to \citet{kleinberg2016descending} for explanation of the terminology, but in brief, the idea is to imagine that when $i$ inspects $j$, the cost $r_{ij}$ is subsidized by an investor in return for a ``call option'', i.e. the right to the excess surplus $v_{ij} - \sigma_{ij}$ beyond a threshold $\sigma_{ij}$, if any. The agent's surplus for that match becomes $\kappa_{ij}$, and the investor breaks even if $(i,j)$ exercises in the money, otherwise loses money.}
\begin{definition} \label{def:strike-covered-call}
Given a cost $r_{ij}$ and distribution $D_{ij}$, the \emph{strike price} is the unique value $\sigma_{ij}$ satisfying
\[ \E_{v_{ij}\sim D_{ij}} \left(v_{ij} - \sigma_{ij}\right)^+ = r_{ij} , \]
where $\left(\cdot\right)^+ = \max\{\cdot ~,~ 0\}$.
The \emph{covered call value} is the random variable $\kappa_{ij} = \min\{\sigma_{ij}, v_{ij}\}$.
\end{definition}
We assume $\E_{v_{ij} \sim D_{ij}} v_{ij} \geq r_{ij}$ if $\{i,j\}$ is a feasible match.
As we have nonnegative values, this is equivalent to the condition $\sigma_{ij} \geq 0$.
A \emph{matching process} is any procedure that involves sequentially inspecting some of the potential matches and making matches, according to any algorithm or mechanism.
The following property, along with Lemma \ref{lemma:covered-call}, allows us to relate surplus in a matching process to that of a world with zero inspection costs and values $\kappa_{ij}$.
\begin{definition} \label{def:exercise-money}
In any matching process, we say that the ordered pair $(i,j)$ \emph{exercises in the money} if, for all realizations of the process, if $I_{ij} = 1$ and $v_{ij} > \sigma_{ij}$ then $A_{ij} = 1$.
We say that agent $i$ exercises in the money if $(i,j)$ exercises in the money for all feasible partners $j$.
\end{definition}
\begin{lemma}[\cite{kleinberg2016descending}] \label{lemma:covered-call}
For any feasible partners $\{i,j\}$, any fixed types of all agents, and any matching process,
\[ \E\left[ A_{ij} v_{ij} - I_{ij} r_{ij} \right] \leq \E \left[ A_{ij} \kappa_{ij} \right], \]
with equality if and only if $(i,j)$ exercises in the money.
\end{lemma}
\begin{proof}
Using the definitions, independence of $v_{ij} \sim D_{ij}$ from the variable $I_{ij}$, and the assumption $A_{ij} \leq I_{ij}$,
\begin{align*}
\E \left[ A_{ij} v_{ij} - I_{ij} r_{ij} \right]
&= \E \left[ A_{ij} v_{ij} - I_{ij} \E_{v'_{ij} \sim D_{ij}} ( v_{ij}' - \sigma_{ij})^+ \right] \\
&= \E \left[ A_{ij} v_{ij} - I_{ij} ( v_{ij} - \sigma_{ij})^+ \right] \\
&\leq \E \left[ A_{ij} v_{ij} - A_{ij} ( v_{ij} - \sigma_{ij})^+ \right] \\
&= \E \left[ A_{ij} \min\{\sigma_{ij}, v_{ij}\} \right].
\end{align*}
We observe that the inequality is strict if and only if there is positive probability of the event that $I_{ij} = 1$, $A_{ij} = 0$, and $v_{ij} > \sigma_{ij}$ all occur, i.e. $(i,j)$ fails to exercise in the money.
\end{proof}
\paragraph{The deviation strategy.}
For a strategy profile $b$, define the random variable $\kappa_i(b)$ to be the covered call value of $i$ for its match in profile $b$, i.e. $\min\{\sigma_{ij}, v_{ij}\}$ when $i$ is matched to $j$.
Let $\bar{u}_i(b)$ denote $i$'s ``covered call utility'', i.e. $\bar{u}_i(b) = \kappa_i(b) - \pi_i(b)$.
Define the deviation strategy $b'_{i}$ for agent $i$ as follows: Initially bid $0$ on each neighbor $j$; when the clock reaches $\sigma_{ij}/2$, inspect neighbor $j$ and update bid to $v_{ij}/2$.
The key technical property of the deviation strategy, in combination with the Marshallian Match, is that it allows bidders to control their own destiny: If they inspect and discover a high value, they can bid high and immediately lock in the match.
That is, their strategy can ``exercise in the money''.
\begin{lemma} \label{lemma:inspect-exercise}
The deviation strategy $b_i'$ exercises in the money and ensures $\bar{u}_i(b_i',b_{-i}) \geq 0$, for any strategy profile $b$.
\end{lemma}
\begin{proof}
(Exercises in the money.)
We consider the two possible scenarios where $i$ inspects a neighbor $j$, i.e. when $I_{ij} = 1$.
If the inspection occurs because the mechanism has just matched $i$ to a previously-uninspected neighbor $j$, then $A_{ij} = 1$ and the requirement of exercising in the money is satisfied.
Otherwise, the clock $p(t)$ has reached $\sigma_{ij}/2$, and after inspecting, $i$ updates that bid to $v_{ij}/2$.
If $v_{ij} \geq \sigma_{ij}$, then $b_{ij} \geq p(t)$ and the match occurs immediately, so $A_{ij} = 1$.
(Nonnegative covered call utility.)
If $i$ is matched to some previously-uninspected $j$, then $\pi_i(b) = 0$, so $\bar{u}_i(b) \geq 0$.
If $i$ is matched to some $j$ that $i$ has already inspected, then $\pi_i(b) = \kappa_{ij}/2$, so $\bar{u}_i(b) = \kappa_{ij}/2 \geq 0$.
Finally, if $i$ is unmatched, then $\bar{u}_i(b) = 0$.
\end{proof}
\begin{lemma}[Covered call smoothness] \label{lemma:inspect-smooth}
For any feasible neighbors $\{i,j\}$ and any strategy profile $b$, for all realizations of types and values:
\[ \bar{u}_i(b_{-i},b_i') + \frac{p_i(b)}{4} + \frac{p_j(b)}{4} \geq \frac{\kappa_{ij}}{8} . \]
\end{lemma}
\begin{proof}
Fix a realization of all types and values.
The quantities $p_i(b)$, $p_j(b)$, and $\bar{u}_i(b_{-i},b_i')$ are all nonnegative.
If $p_i(b) \geq \kappa_{ij}/2$ or $p_j(b) \geq \kappa_{ij}/2$, then we are already done.
So suppose neither holds.
Then in profile $b$, $i$ is not yet matched at time $\kappa_{ij}/2$.
Observe that in $(b_{-i},b_i')$, all other agents' bids and behavior are unchanged until $i$ is matched or $\kappa_{ij}/2$, because they continue playing their strategies in $b$ and no information available to them indicates the change in $i$'s strategy.
The first possibility is that in $(b_{-i},b_i')$, agent $i$ is unmatched when the price reaches $\kappa_{ij}/2$.
Then by definition of $b_i'$ they match to $j$ at this time, obtaining $\kappa_i(b_{-i},b_i') = \kappa_{ij}$ and $\pi_i(b_{-i},b_i') = \kappa_{ij}/2$, and we are done.
The other possibility is that in $(b_{-i}, b_i')$, agent $i$ matches before the price reaches $\kappa_{ij}/2$.
In this case, we claim the match is on some neighbor that $i$ has inspected.
This follows because $i$ bids zero on all neighbors prior to inspection, which is weakly lower than whatever $i$ bids in profile $b_i$, where they are not matched until at least time $\kappa_{ij}/2$.
So $i$ can only match at price $p(t) \geq \kappa_{ij}/2$ on a neighbor $j'$ that $i$ has inspected, which means they are bidding $\kappa_{ij'}/2$ on $j'$, so we have $\bar{u}_i(b_{-i},b_i') = \kappa_i(b_{-i},b_i') - \pi_i(b_{-i},b_i) = \kappa_{ij'}/2 \geq \kappa_{ij}/2$, completing the proof.
\end{proof}
\begin{theorem} \label{theorem:inspect-welfare}
In the inspection setting with nonnegative values, the Marshallian Match guarantees a Bayes-Nash price of anarchy of at least $1/8$.
\end{theorem}
\begin{proof}
Let the random variable $M^*$ be the max-weight matching where the weight on edge $\{i,j\}$ is $\kappa_{ij} + \kappa_{ji}$.
Lemma \ref{lemma:covered-call} also implies that $\mathrm{Welfare}(\textsc{Opt})$ is at most the total weight of $M^*$, as follows.
Here $I_{ij},A_{ij}$ are the indicators under the $\textsc{Opt}$ procedure, and notice $A_{ij} = A_{ji}$ in any matching.
\begin{align*}
\mathrm{Welfare}(\textsc{Opt}) &= \E \sum_{\{i,j\} \in E} \left[ \left( A_{ij} v_{ij} - I_{ij} r_{ij} \right) + \left( A_{ji} v_{ji} - I_{ji} r_{ji} \right) \right] \\
&\leq \E \sum_{\{i,j\}} \left(A_{ij} \kappa_{ij} + A_{ji} \kappa_{ji}\right) & \text{Lemma \ref{lemma:covered-call}}\\
&\leq \E \sum_{\{i,j\} \in M^*} \left( \kappa_{ij} + \kappa_{ji} \right) &\text{the solution must form a matching.}
\end{align*}
(We note that we have no idea what $\textsc{Opt}$ actually is in this setting, or if it can even be computed in polynomial time; nevertheless, this upper bound cannot be too loose, since the MM approximates it.)
Lemma \ref{lemma:inspect-exercise} states that the deviation strategy $b_i'$ exercises in the money, so Lemma \ref{lemma:covered-call} implies $\E u_i(b_{-i},b_i') = \E \bar{u}_i(b_{-i}, b_i')$.
Therefore,
\begin{align*}
\mathrm{Welfare}(b) &= \E \sum_{i} \left[u_i(b) + \frac{p_i(b)}{2} \right] \\
&\geq \E\sum_{i} \left[ u_i(b_{-i},b_i') + \frac{p_{i}(b)}{2} \right] \\
&= \E\sum_{i} \left[ \bar{u}_i(b_{-i},b_i') + \frac{p_{i}(b)}{2} \right] \\
&\geq \E\sum_{\{i,j\}\in M^*} \left( \left[ \bar{u}_i(b_{-i},b_i') + \frac{p_i(b)}{4} + \frac{p_j(b)}{4} \right] + \left[ \bar{u}_j(b_{-j},b_j') + \frac{p_{i}(b)}{4} + \frac{p_{j}(b)}{4} \right] \right) \\
&\geq \E \sum_{\{i,j\}\in M^*} \left[ \frac{\kappa_{ij}}{8} + \frac{\kappa_{ji}}{8} \right]
\quad \geq \frac{1}{8} \mathrm{Welfare}(\textsc{Opt}) .
\end{align*}
\end{proof}
\section{Preliminaries} \label{sec:prelim}
We now define the model and the variant of the Marshallian Match mechanism studied in this paper, originally described by \citet{waggoner2019matching}.
We define a general model, in which the graph is possibly non-bipartite and bids and values are possibly negative.
Later, we will consider different special cases or variations.
There is a finite set of $n$ agents, forming vertices of an undirected graph $G = (\{1,\dots,n\}, E)$.
For now, we do not assume that $G$ is bipartite.
The presence of an edge $\{i,j\} \in E$ represents that it is feasible to match agents $i$ and $j$.
In this case, agent $i$ has a \emph{value} $v_{ij} \in \mathbb{R}$ for being matched to $j$, and symmetrically, $j$ has a value $v_{ji}$ for being matched to $i$.
If $i$ and $j$ are neighbors, we let $s_{ij} = v_{ij} + v_{ji}$ denote the \emph{surplus} of the edge $\{i,j\}$.
An agent $i$'s \emph{type} consists of their values $v_{ij}$ for each feasible partner $j$.
In the \emph{Nash} setting, each agent has a fixed type, and types are common knowledge.
In the better-motivated \emph{Bayes-Nash} setting, types are drawn from a common-knowledge joint distribution $\mathcal{D}$ and each agent observes their own type.
\subsection{The Marshallian Match}
In the MM, a global price $p(t)$ begins at $+\infty$, i.e. $p(0)=\infty$, and descends continuously in time until it reaches zero at time $1$, i.e. $p(1) = 0$.\footnote{This can be accomplished in theory by letting the price be e.g. $p(t) = \frac{1}{2t}$ for $t \in [0,1/2]$ and $p(t) = 2-2t$ for $t \in [1/2,1]$.}
Each agent $i$ maintains, for all neighbors $j$, a bid $b_{ij}(t)$ at each time $t$.
The mechanism can observe all bids at all times, but agents cannot observe any bids besides their own.
For convenience, we will generally drop the dependence on $t$ from the bid notation.
When the sum of bids on any edge exceeds the global price, i.e. $b_{ij} + b_{ji} \geq p(t)$, then $i$ and $j$ are immediately matched to each other.
Each player pays their respective bid to the mechanism.
(We will modify the payment rule in Section \ref{sec:general-bids}.)
\paragraph{Intuition for the mechanism.}
Why might this mechanism be good, and how might participants strategize?
We briefly describe some intuition, referring the reader to \citet{waggoner2019matching} for more detailed discussion.
Social welfare and price of anarchy will be formally defined below.
First, the edges of the graph are in competition with each other to match first.
When an edge $\{i,j\}$ is matched, it produces a surplus of $s_{ij} = v_{ij} + v_{ji}$.
The first-best solution, i.e. the optimal solution for a central planner who holds all information, is to select a maximum matching where $s_{ij}$ are the edge weights.
However, as discussed in \cite{kleinberg2016descending}, algorithms for maximum matching are complex and appear to interact poorly with inspection stages.
More robust is an approximate first-best solution: the \emph{greedy} matching, where the highest-weight edge is matched first, and so on down.
This procedure obtains at least half of the optimal welfare, and can be simulated by a Marshallian Match in which all participants bid truthfully.
In \cite{kleinberg2016descending}, this fact was used to obtain a constant price of anarchy for matching people to items, including in the presence of inspection costs.
However, two-sided matching introduces new strategic considerations.
``Within'' an edge $\{i,j\}$, there is a competition or bargaining for how to split the surplus generated.
An agent $i$ with many edges (many outside options) may be able to underbid significantly, while her counterpart $j$ with very few outside options must offer a very high bid along that edge.
The question is whether this strategizing and competition between $i$ and $j$ destroys the cooperation incentives for the overall bid on the edge $\{i,j\}$.
\subsection{Welfare and price of anarchy}
A full strategy profile is denoted $b$.
We let $v_i(b)$ be $i$'s value for their match when $b$ is played (zero if none), a random variable depending on the randomization in the strategies and, in the Bayes-Nash setting, on the random draw of the types.
Similarly, $\pi_i(b)$ denotes $i$'s payment.
Finally, $p_i(b)$ denotes the \emph{total} payment on the edge that $i$ is matched along, i.e. $p_i(b) = \pi_i(b) + \pi_j(b)$ when $i$ is matched to $j$.
We assume all agents are Bayesian, rational, and have preferences quasilinear in payment.
That is, given a mechanism and a particular strategy profile $b$, the \emph{utility} of a participant $i$ is the random variable
\[ u_i(b) = v_i(b) - \pi_i(b) . \]
A \emph{Nash equilibrium} of a mechanism with given types $\{v_{ij}\}$ is a strategy profile $b$, consisting of a plan for how to set bids at each moment in time\footnote{We observe that usual nuances around equilibrium in dynamic games, such as non-credible threats, refinements such as subgame perfect equilibrium, etc., do not arise here. In our variant of the MM, each agent $i$ observes nothing until they are matched, after which they can no longer affect the game. So without loss of generality, $i$ commits in advance to a plan $\{ b_{ij}(t) \}$ and follows it until matched.}, where each participant maximizes their expected utility, i.e.
\[ \E u_i(b) \geq \E u_i(b_{-i}, b_i') \]
for all $i$ and for any other strategy $b_i'$ of $i$, where the randomness is taken over the strategies.
A \emph{Bayes-Nash equilibrium} is defined in exactly the same way, but in a setting consisting of a joint distribution over types.
In that case, the randomness is taken over both types and strategies (which are maps from an agent's type to a plan for bidding over time).
We use $\mathrm{Welfare}(b)$ to denote the expected social welfare, i.e. sum of utilities and payments:
\[ \mathrm{Welfare}(b) = \E \left[ \sum_i v_i(b) + \sum_j v_j(b) \right] , \]
where the expectation is over all randomness.
In the setting with inspection costs, utility and social welfare also accounts for the loss of utility from the inspection processes; this will be formalized at the relevant point, Section \ref{sec:inspection}.
$\mathrm{Welfare}(\textsc{Opt})$ refers to the optimal social welfare.
In the Nash setting,
\[ \mathrm{Welfare}(\textsc{Opt}) = \max_M \sum_{\{i,j\} \in M} v_{ij} + v_{ji} , \]
where the maximum is over all matchings $M$.
In the Bayes-Nash setting,
\[ \mathrm{Welfare}(\textsc{Opt}) = \E \left[ \max_M \sum_{\{i,j\} \in M} v_{ij} + v_{ji} \right], \]
where the expectation is over the realizations of types.
The \emph{price of anarchy} measures the worst-case ratio of $\mathrm{Welfare}(b)$ to $\mathrm{Welfare}(\textsc{Opt})$ in any equilibrium $b$.
For Nash equilibrium, we have
\[ \text{PoA} = \min \frac{\mathrm{Welfare}(b)}{\mathrm{Welfare}(\textsc{Opt})} ,\]
where the minimum is taken over all settings (i.e. all types of the participants) and all Nash equilibria $b$.
The Bayes-Nash price of anarchy is defined in exactly the same way, but the minimum is now over all Bayes-Nash settings (i.e. joint distributions on types) and all Bayes-Nash equilibrium strategy profiles $b$.
Note that a Nash equilibrium is a special case of Bayes-Nash where the type distributions are degenerate.
Therefore, a Bayes-Nash price of anarchy result immediately implies a Nash price of anarchy.
| 66c309edb5502836fcf9a86ba4ddb372a23fed88 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 2pt}
\titlespacing\subsection{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 2pt}
\renewcommand*{\paragraph}[2]{\textbf{#2}}
\newcommand{}{}
\newcommand{\mycaption}[2][]{\renewcommand{}{#2}%
\caption[#1]{{\footnotesize #2}}}
\makeatletter
\renewcommand{\fnum@table}{\textbf{\tablename~\thetable}}
\renewcommand{\fnum@figure}{\textbf{\figurename~\thefigure}}
\makeatother
\makeatletter
\AtBeginDocument{\let\LS@rot\@undefined}
\makeatother
\newcommand{\Delta m^2}{\Delta m^2}
\newcommand{\ensuremath{\text{eV}^2}}{\ensuremath{\text{eV}^2}}
\newcommand{\mathop{\mathrm{diag}}}{\mathop{\mathrm{diag}}}
\newcommand{\overline{\nu}}{\overline{\nu}}
\newcommand{\delta_\mathrm{CP}}{\delta_\mathrm{CP}}
\newcommand{\sin^22\theta_{13}}{\sin^22\theta_{13}}
\newcommand{\mathrm{ND}}{\mathrm{ND}}
\newcommand{\mathrm{FD}}{\mathrm{FD}}
\newcommand\snowmass{\begin{center}\rule[-0.2in]{\hsize}{0.01in}\\\rule{\hsize}{0.01in}\\
\vskip 0.1in Submitted to the Proceedings of the US Community Study\\
on the Future of Particle Physics (Snowmass 2021)\\
\rule{\hsize}{0.01in}\\\rule[+0.2in]{\hsize}{0.01in} \end{center}}
\def\eq#1{{Eq.~(\ref{#1})}}
\def\eqs#1#2{{Eqs.~(\ref{#1})--(\ref{#2})}}
\def\fig#1{{Fig.~\ref{#1}}}
\def\figs#1#2{{Figs.~\ref{#1}--\ref{#2}}}
\def\Table#1{{Table~\ref{#1}}}
\def\Tables#1#2{{Tables~\ref{#1}--\ref{#2}}}
\def\sect#1{{Sect.~\ref{#1}}}
\def\sects#1#2{{Sects.~\ref{#1}--\ref{#2}}}
\def\app#1{{Appendix~\ref{#1}}}
\def\apps#1#2{{Apps.~\ref{#1}--\ref{#2}}}
\def\vev#1{\left\langle #1\right\rangle}
\def\abs#1{\left| #1\right|}
\def\mod#1{\abs{#1}}
\def\mbox{Im}\,{\mbox{Im}\,}
\def\mbox{Re}\,{\mbox{Re}\,}
\def\mbox{Tr}\,{\mbox{Tr}\,}
\def\mbox{det}\,{\mbox{det}\,}
\def\hbox{\it et al.}{\hbox{\it et al.}}
\def\hbox{\it i.e.}{}{\hbox{\it i.e.}{}}
\def\hbox{\it e.g.}{}{\hbox{\it e.g.}{}}
\def\hbox{\it etc}{}{\hbox{\it etc}{}}
\def\bra#1{\mathinner{\langle{#1}|}}
\def\ket#1{\mathinner{|{#1}\rangle}}
\def\Bra#1{\left\langle#1\right|}
\def\Ket#1{\left|#1\right\rangle}
\def\unskip,{\unskip,}
\newcommand\lsim{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$<$}}}
\newcommand\gsim{\mathrel{\rlap{\lower4pt\hbox{\hskip1pt$\sim$}}
\raise1pt\hbox{$>$}}}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{\begin{eqnarray}}{\begin{eqnarray}}
\newcommand{\end{eqnarray}}{\end{eqnarray}}
\newcommand{\begin{pmatrix}}{\begin{pmatrix}}
\newcommand{\end{pmatrix}}{\end{pmatrix}}
\newcommand{\noindent}{\noindent}
\newcommand{\nonumber}{\nonumber}
\newcommand{\color{red}}{\color{red}}
\newcommand{\color{blue}}{\color{blue}}
\newcommand{\ \mathrm{TeV}}{\ \mathrm{TeV}}
\newcommand{\ \mathrm{GeV}}{\ \mathrm{GeV}}
\newcommand{\ \mathrm{MeV}}{\ \mathrm{MeV}}
\newcommand{\ \mathrm{keV}}{\ \mathrm{keV}}
\newcommand{\ \mathrm{eV}}{\ \mathrm{eV}}
\setlength{\paperheight}{11in}
\setlength{\paperwidth}{8.5in}
\setlength{\topmargin}{0in}
\setlength{\headsep}{0in}
\setlength{\headheight}{11.74pt}
\setlength{\textheight}{9in}
\setlength{\footskip}{0.5in}
\setlength{\textwidth}{6.5in}
\setlength{\evensidemargin}{0in}
\setlength{\oddsidemargin}{0in}
\renewcommand{\baselinestretch}{1.17}
\pagestyle{fancy}\def{}
\def{}
\setlength{\belowcaptionskip}{-5pt}
\fancyfoot[L]{DRAFT \today}
\fancyfoot[C]{}
\fancyfoot[R]{\thepage}
\fancyhead{}
\renewcommand{\headrulewidth}{0pt}
\begin{document}
\title{\scshape\Large \bf Hyper-Kamiokande Experiment: A Snowmass White Paper\\
\vspace{5mm}
\normalsize Contributed Paper to Snowmass 2021
\vspace{5mm}
}
\author{J.~Bian}\affiliation{Department of Physics and Astronomy, University of California Irvine, Irvine CA 92697-4575, USA}
\author{F.~Di Lodovico}\affiliation{King's College London, Department of Physics, Strand Building, Strand, London, United Kingdom}
\author{S.~Horiuchi}\affiliation{Center for Neutrino Physics, Virginia Tech, Blacksburg, Virginia 24061, USA}
\author{J.~G.~Learned}\affiliation{Department of Physics and Astronomy, University of Hawaii, Honolulu, Hawaii 96822, USA}
\author{C.~Mariani}\email{Corresponding author: mariani@vt.edu}\affiliation{Center for Neutrino Physics, Virginia Tech, Blacksburg, Virginia 24061, USA}
\author{J.~Maricic}\affiliation{Department of Physics and Astronomy, University of Hawaii, Honolulu, Hawaii 96822, USA}
\author{J.~Pedro~Ochoa~Ricoux}\affiliation{Department of Physics and Astronomy, University of California Irvine, Irvine CA 92697-4575, USA}
\author{C.~Rott}\affiliation{Department of Physics and Astronomy, University of Utah, Salt Lake City, UT 84112, USA}\affiliation{Department of Physics, Sungkyunkwan University, Suwon 16419, Korea}
\author{M.~Shiozawa}\affiliation{University of Tokyo, Institute for Cosmic Ray Research, Kamioka Observatory, Kamioka, Japan}\affiliation{University of Tokyo, Kavli Institute for the Physics and Mathematics of the Universe (WPI), University of Tokyo Institutes for Advanced Study, Kashiwa, Japan}\affiliation{University of Tokyo, Next-generation Neutrino Science Organization, Kamioka, Japan}
\author{M.~B.~Smy}\affiliation{Department of Physics and Astronomy, University of California Irvine, Irvine CA 92697-4575, USA}
\author{H.~W.~Sobel}\affiliation{Department of Physics and Astronomy, University of California Irvine, Irvine CA 92697-4575, USA}
\author{R.~B.~Vogelaar}\affiliation{Center for Neutrino Physics, Virginia Tech, Blacksburg, Virginia 24061, USA}
\collaboration{on behalf of the Hyper-Kamiokande Collaboration}
\date{\today}
\begin{abstract}
\vspace{2cm}
Hyper-Kamiokande (HK) is the next generation underground water Cherenkov detector that builds on the highly successful Super-Kamiokande (SK) experiment. The 260,000-ton detector has an 8.4 times larger fiducial volume than its predecessor. HK's low energy threshold combined with the very large fiducial volume make the detector unique; HK is expected to acquire an unprecedented exposure of 3.8 Mton-year over a period of 20 years of operation. It has an extremely diverse science program including long-baseline neutrino oscillation measurements, nucleon decay searches, atmospheric neutrinos, neutrinos from the sun and supernova explosions, and neutrinos from other astrophysical origins. Like DUNE, the flagship project of the U.S. high-energy physics program, HK measures fundamental properties of neutrinos such as the search for leptonic CP violation and neutrino physics beyond the Standard Model.
\end{abstract}
\maketitle
\renewcommand{\familydefault}{\sfdefault}
\renewcommand{\thepage}{\roman{page}}
\setcounter{page}{0}
\pagestyle{plain}
\clearpage
\textsf{\tableofcontents}
\renewcommand{\thepage}{\arabic{page}}
\setcounter{page}{1}
\pagestyle{fancy}
\fancyhead{}
\fancyfoot{}
\fancyfoot[RO]{\textsf{\footnotesize \thepage}}
\fancyfoot[LO]{\textsf{\footnotesize Hyper-Kamiokande: A Snowmass White Paper}}
\fancypagestyle{plain}{}
\renewcommand{\footrulewidth}{0.4pt}
\newpage
\input{overview.tex}
\newpage
\section*{\bf \large References}
\renewcommand{\refname}{References}
\section{\bf \large Overview}
The HK collaboration is presently formed amongst groups from 19 countries including the United States, whose community has a long history of making significant contributions to the neutrino physics program in Japan. US physicists have played leading roles in the Kamiokande, KamLAND, SK, EGADS, K2K, and T2K programs. Currently, the HK design is being finalized while excavation has already been started. HK data taking is expected to start in 2027.
The US groups are interested in exploiting the complementarity of the HK and DUNE experiments. This complementarity is similar to the one of the T2K and the NO$\nu$A experiments: different baselines, beam energies (narrow-band vs. wide-band), detector technology, detection mechanism and detector size. While both DUNE and HK are also observatories for supernova neutrinos, they are complementary in their detection channels (electron neutrino detection vs. anti-electron neutrino detection). There is also complementarity in the nucleon decay channels.
The motivation for pursuing "natural" neutrino sources in addition to measuring neutrino beams remains strong: not only did the study of these neutrino sources lead to many groundbreaking discoveries in the past (including the discovery of neutrino flavor oscillation and mass, measurements of the mixing angles, discovery of the oscillation pattern, etc.), but future measurements of these sources will significantly contribute to determine the mass ordering and support and complement CP violation and other beam-related neutrino measurements.
{\vskip 0.25in}
In the following sections we give details on particular physics topics and objectives, highlighting the complementary between the HK and DUNE physics programs:
\begin{enumerate}
\item Nucleon Decay and Dark Matter
\item Solar Neutrinos
\item Supernova Neutrinos, Multi-messenger science, Relic Supernova Neutrinos
\item Atmospheric Neutrinos, Neutrino Mass Ordering, Non-standard interactions and CP violation
\item Accelerator Neutrinos - The J-PARC to Hyper-K long baseline experiment
\end{enumerate}
{\vskip 0.25in}
\noindent
\section{\bf \large Nucleon Decay and Dark Matter}
{\vskip 0.075in}
HK will be able to provide some of the most stringent tests of the Standard Model. The possibility that baryon number violation holds the key to a number of questions in fundamental physics and cosmology makes the continued search for proton decay and related processes a high priority. We expect the HK experiment to pick up where the SK program leaves off. For example, nucleon decay sensitivities will be extended by one order of magnitude [Fig.~\ref{fig:nucleon_decay}] beyond the current limits and could reveal grand unified theories (GUTs)~\cite{HyperKamiokande:2018ofw}.
\begin{figure}[tb!]
\centering
\includegraphics[width=1.0\columnwidth]{figures/HK_nucleon_decay.jpg}
\caption{Nucleon decay physics reach of Hyper-K~\cite{Itow:2021rnc}.}
\label{fig:nucleon_decay}
\end{figure}
With 20~years of data, Hyper-K will reach a proton decay sensitivity of $10^{35}$~years for $p \rightarrow \pi^0 e^+$ and $3\times 10^{34}$~years for $p \rightarrow \bar{\nu} K^+$. These decay modes are highly complementary to those accessible by liquid argon based detectors: HK has a huge total mass, and lower mass per nucleus (including free protons), but offers less detail in event reconstruction and some particles (e.g. Kaons) are near or below the Cherenkov threshold. HK proton decay searches are similarly complementary to JUNO which is also sensitive to Kaons {\it and} has a low mass per nucleus, but much lower total mass than HK and much better reconstruction capabilities than liquid Argon detectors.
\vspace{1cm}
The search for physics beyond the Standard Model of particle physics is one of the priorities of HK. Following SK's success, the indirect search for dark matter from the Sun, Earth, Galactic center, and halo is expected to continue to provide the most sensitive tests for dark matter nucleon scattering for masses of a few GeV in a model independent way~\cite{Choi:2013eda,SuperKamiokande:2015xms,Danninger:2014xza}. HK will search for boosted dark matter~\cite{SuperKamiokande:2017dch}, asymmetric dark matter~\cite{Murase:2016nwx}, or test models with predominantly hadronic annihilation channels that remain hidden to other neutrino detectors~\cite{Rott:2012qb,Bernal:2012qh,Rott:2015nma}. Searches for dark matter from the Galactic halo will provide some of the best sensitivities to dark matter for annihilation channels with large neutrino yields.
{\vskip 0.25in}
\noindent
\section{\bf \large Solar Neutrinos}
{\vskip 0.075in}
\par Hyper-K will detect solar neutrino interactions with unprecedented statistical power. Like Super-K, the interaction mode is elastic scattering of $^8$B neutrinos. The analysis of the recoil electron spectrum reveals the underlying neutrino spectrum. In particular, the ``upturn'', the vacuum--MSW transition, will be detectable at $5\sigma$ ($3\sigma$) based on 10 years of data with an energy threshold of $3.5$~MeV ($4.5$~MeV)~\cite{HyperKamiokande:2018ofw}. In addition to constraining solar neutrino oscillation parameters in the standard oscillation scenario, the recoil electron spectrum provides sensitivity to non-standard neutrino interactions. Hyper-K will also measure the diurnal variation of the solar neutrino interaction rate due to Earth matter effects. In addition to measuring oscillation parameters, this effect is sensitive to non-standard interactions as well. Also the comparison of solar neutrino oscillation parameters with reactor anti-neutrino oscillation parameters tests CPT invariance.
Hyper-K is particularly well suited to measure short-period flux variations in solar neutrinos, realizing a real-time monitoring of the Solar core temperature. Moreover, Hyper-K could achieve the first measurement of hep solar neutrinos, giving new insights in solar physics. Hyper-K solar neutrino measurements would be complementary to possible DUNE measurements using solar $^8$B neutrino absorption on argon: while the neutrino absorption process has better energy correlation between the detected electron and the incident neutrino, elastic scattering avoids the high nuclear energy threshold of $\approx5$ MeV. The cosmogenic background in DUNE will be lower (due to the greater depth), but elastic scattering measurements have better signal/noise ratio due to the good directional correlation of electrons and neutrinos.
Low energy interactions (less than $\sim$100~MeV where most electrons will start electromagnetic showers) are more challenging than higher energy events since the light yield in a Cherenkov detector (number of detected photo-electrons per MeV above Cherenkov threshold) is much smaller than in a liquid scintillator (or the ionization yield in a TPC): O(10) detected photo electrons per MeV~\cite{HyperKamiokande:2018ofw}. In fact, for Super-K the light yield is about six per MeV, while for a detector like JUNO it will be around 1300 per MeV~\cite{2022103927}. This makes event reconstruction and radioactive background discrimination more challenging, and results in a higher threshold compared to JUNO's expected $\sim$2~MeV. However, Hyper-K's strengths are its unmatched size and the ability to reconstruct the charged particle's direction, which makes it more resilient against backgrounds and is an important feature for astrophysics.
{\vskip 0.25in}
\noindent
\section{\bf \large Supernova Neutrinos, Multi-messenger Science}
{\vskip 0.075in}
Through the observation of $\sim$10~MeV neutrinos with time, energy and directional information, Hyper-K will take a unique role as multi-messenger observatory. Hyper-K will expand on the successful multi-messenger science program of Super-K~\cite{SuperKamiokande:2016jsv}. It has the potential to detect thermal neutrinos from nearby ($<10$~Mpc) neutron star merger events in coincidence with gravitational waves.
Given sufficient running time, the observation of a signal in HK due to a core collapse supernova is assured~\cite{Hirata:1987hu,Ikeda:2007sa,Nakamura:2016kkl}. It is estimated that the chance of a Type II supernova in the Milky Way galaxy is a few percent per year. What we might learn from such an event is relatively unknown, although there exist many models that make baseline predictions for comparison once we have the data. What is certain is that the statistical signal will dwarf the 19 events observed by IMB and Kamiokande from SN1987A. For a supernova at 10~kpc, somewhere between 16,000 to 64,000 neutrino events above 7~MeV would be detected by Hyper-K in a few seconds, depending on the supernova model. Hyper-Kamiokande will provide the largest sample of the most detailed events, including elastic scatter events that will reveal the direction of the supernova and enable the astronomical community to engage in multi-messenger discoveries. Hyper-K's reach extends to the Andromeda Galaxy M31 ($\sim$780~kpc) with about $\sim$10 events expected per supernova.
The direction of a supernova at 10~kpc can be reconstructed with an accuracy of about 1$^o$ assuming similar event reconstruction performance as Super-K~\cite{HyperKamiokande:2018ofw}, making Hyper-K essential for distributing early alerts and multi-messenger observations. Further, even a few neutrinos from nearby extra-galactic supernovae can reveal the nature of transients whose mechanism is uncertain~\cite{SuperKamiokande:2000kzn,Dragowsky:2001ax}.
With tens of thousands of events from a supernova in the Milky Way galaxy, it is anticipated that studying the time, energy, and flavor structure of the burst may reveal not only astrophysical secrets of the core collapse and subsequent explosion, but also information on neutrino properties. The observed burst signal may reveal the effects of neutrino-neutrino interactions in a regime inaccessible by any other means, as well as effects due to beyond-the-Standard-Model physics. Since the matter effect resonance is directly connected to the neutrino mass hierarchy, it is possible that structure in the time-energy spectrum may reveal whether the hierarchy is inverted or normal. If the supernova occurs when both Hyper-K and DUNE are running, the large sample of inverse-beta decay events from Hyper-K and the large sample of electron neutrino charge-current events from DUNE will yield highly complementary flavor information of the core-collapse supernova neutrinos, opening a novel flavor window to study a variety of effects from supernova shock effects to neutrino properties such as the neutrino mass ordering.
{\vskip 0.25in}
\noindent
\section{\bf \large Supernova Relic Neutrinos}
{\vskip 0.075in}
In addition to being prepared for a gravitational collapse supernova event, we have another way to study supernova neutrinos. Neutrinos emitted by supernovae at a distance beyond Hyper-Kamiokande's ability to uniquely localize in time, comprise the so-called ``diffuse supernova neutrino background'' (DSNB), also known as the ``supernova relic neutrinos'' (SRN). These are one of the few remaining neutrino signals that has not yet been detected by any experiment. When observed, the measured rate of SRN could provide information about stellar collapse, nucleosynthesis, and the rate of star formation in the universe. The addition of Gd to Hyper-Kamiokande or direct comparison between pure water HK and Gd loaded SK would address the backgrounds that have limited prior searches~\cite{Super-Kamiokande:2011lwo}, by allowing the signal selection to require the distinctive signature of inverse beta decay followed by neutron capture. The search window would lie between the background due to nuclear reactor antineutrinos and backgrounds due to atmospheric neutrinos [Fig.~\ref{fig:supernovae}].
\begin{figure}[htb!]
\centering
\includegraphics[width=0.7\columnwidth]{figures/supernovae.jpg}
\caption{The spectrum of anti-neutrinos due to relic supernova neutrinos compared with backgrounds from atmospheric neutrinos and reactors. The width of the SRN band represents the range of model predictions. The dashed atmospheric neutrino lines reflect the expected background without neutron capture on gadolinium~\cite{HyperKamiokande:2018ofw}.}
\label{fig:supernovae}
\end{figure}
{\vskip 0.25in}
\noindent
\section{\bf \large Atmospheric Neutrinos - Neutrino Mass Ordering and Non-standard Interactions}
{\vskip 0.075in}
Precise oscillation measurements with atmospheric neutrinos will be used to study the neutrino mass ordering as well as push limits for Lorentz symmetry violation~\cite{SuperKamiokande:2014exs,IceCube:2017qyp}, non-standard interaction~\cite{SuperKamiokande:2011dam,IceCube:2017zcu}, quantum decoherence~\cite{Stuttard:2020qfv}, sterile neutrino oscillation~\cite{T2K:2019efw,IceCube:2020tka}, and various dark sector particle searches such as heavy neutral lepton~\cite{T2K:2019jwa,Coloma:2019htx}, long lived particle~\cite{Arguelles:2019ziu}, and millicharged particles~\cite{Plestid:2020kdm}.
Atmospheric neutrinos span a wide range of energy and path lengths, and are born in both the neutrino and antineutrino form of electron and muon flavor. Neutrinos that travel through the earth pass through a substantial matter density that modifies the vacuum oscillation probability in a predictable way. This feature gives HK the potential to determine the neutrino mass ordering from atmospheric neutrinos alone. Hierarchy sensitivity predominantly comes from the ``Multi-GeV'' upward-going electron neutrinos as seen in Fig.~\ref{fig:atmospheric}. There are resonant oscillations between 2-10~GeV for $\nu_\mu$ or $\bar{\nu_\mu}$ depending on the mass hierarchy.
\begin{figure}[tb!]
\subfloat[]{\includegraphics[width=0.52\columnwidth]{figures/oscillation_right.jpg}}
\subfloat[]{\includegraphics[width=0.45\columnwidth]{figures/oscillation_left.jpg}}
\caption{(a) The oscillogram showing oscillation probability from $\nu_\mu$ to $\nu_e$ as a function of energy and cosine zenith angle. (b) The resultant sensitivity for wrong hierarchy rejection as a function of $\sin^2\theta_{23}$ and running time. Running time is for 1, 5, 10, 15, 30 years. Figures from Ref.~\cite{HyperKamiokande:2018ofw}.}
\label{fig:atmospheric}
\end{figure}
Although atmospheric neutrinos have limited sensitivity to CP-violation relative to the beam measurement, the sensitivity is largely complementary and the addition of atmospheric neutrino data to the beam measurement can improve the $\delta_{CP}$ measurement, particularly in regions of limited beam sensitivity due to degeneracies between oscillation parameters (i.e. mass hierarchy and CP effects). As an example, the sensitivity for T2K compared to T2K $+$ SK is shown in [Fig.~\ref{fig:sensitivity}]. The addition of SK atmospheric neutrino data improves rejection for $0 < \delta_{CP} < \pi$ since the T2K baseline~\cite{Ankowski:2015jya,Ankowski:2016bji}, like the HK baseline, is relatively short and feels no matter-effect. The final projected sensitivity for HK is shown in Fig.~\ref{fig:sensitivity}.
\begin{figure}[tb!]
\subfloat[]{\includegraphics[width=0.48\columnwidth]{figures/t2k_sensitivity_left.jpg}}
\subfloat[]{\includegraphics[width=0.52\columnwidth]{figures/CPV_scan_HK_NH_comb.jpg}}
\caption{(a) The sensitivity for T2K (solid blue) and T2K$+$SK (solid red) to reject the sin($\delta_{CP}$)=0 hypothesis assuming unknown hierarchy. (b) Projected significance for $\sin \delta_{CP}$=0 exclusion for Hyper-K assuming normal ordering.}
\label{fig:sensitivity}
\end{figure}
The unitarity of the PMNS matrix can be tested through $\nu_\tau$ appearance~\cite{Ankowski:2016bji,Benhar:2015wva,Ankowski:2016jdd}. The current determinations of PMNS mixing elements agree between atmospheric neutrinos and long-baseline or reactor experiments, but tensions could appear that reveal new research directions.
\begin{figure}[htb!]
\centering
\includegraphics[width=0.75\columnwidth]{figures/tau_neutrino.jpg}
\caption{Tau neutrino excess in upward going events in Super-K~\cite{HyperKamiokande:2018ofw}.}
\label{fig:tau_neutrinos}
\end{figure}
Super-K has demonstrated the ability to identify $\nu_\tau$ in the atmospheric neutrino data, excluding the hypothesis of no-tau-appearance with a significance of 4.6 $\sigma$ (Fig. ~\ref{fig:tau_neutrinos}). After 10 years of running HK, will have the order of 1000~$\nu_\tau$ events that can be used to study CC $\nu_\tau$ cross section, leptonic universality, etc.
With a significant increase of fiducial volume from Super-K, Hyper-K expects to observe 1,000 atmospheric neutrinos in the TeV range. These neutrinos can fill the gap between accelerator-based neutrino experiments and neutrino telescopes, and they open up new opportunities for physics beyond current scopes~\cite{Schneider:2021wzs}. Topics of TeV range atmospheric neutrinos at Hyper-K include neutrino cross-section measurements in the TeVs energy range competitive to the CERN FASER$\nu$ experiment~\cite{FASER:2021mtu}, and provide significant chance to discover new physics including but not limited to TeV boosted dark matter.
In a global context Hyper-K data combined with those from other long-baseline neutrino oscillation experiments will be essential to test non-standard oscillation scenarios, which can only be revealed by combining data from different experiments.
{\vskip 0.25in}
\noindent
\section{\bf \large Accelerator Neutrinos - The J-PARC to Hyper-K long baseline experiment}
{\vskip 0.075in}
The focus of the J-PARC to Hyper-K experiment is the measurements of $\Delta m^{2}_{32}$, $\sin^{2}\theta_{23}$, $\sin^{2}\theta_{13}$, and $\delta_{CP}$.
The neutrino energy spectrum of the J-PARC neutrino beam is tuned to the first oscillation maximum using the off-axis technique, which enhances the flux at the peak energy while reducing the higher energy component that produces background events. The peak energy, around 600~MeV, is well matched to the water Cherenkov detector technology, which has an excellent e/µ separation capability (miss-identification is less than 1\% for sub-GeV and $\sim$2\% for multi-GeV), high background rejection efficiency and high signal efficiency for sub-GeV neutrino events. Due to the relatively short baseline of 295~km and thus lower neutrino energy at the oscillation maximum, the contribution of the matter effect is smaller for the J-PARC to Hyper-Kamiokande experiment compared to DUNE. Thus, the CP asymmetry measurement with the J-PARC to Hyper-K long baseline experiment has less uncertainty related to the matter effect, while DUNE with its 1,300 km baseline has a much better sensitivity to the mass hierarchy. Nevertheless, as described earlier, Hyper-K can determine the mass hierarchy using atmospheric neutrinos and sensitivities for CP violation and mass hierarchy can be further enhanced by combining accelerator and atmospheric neutrino measurements.
Fig.~\ref{fig:Delta_CP_Error} shows the 68\% CL uncertainty of $\delta_{CP}$ as a function of running time (the integrated beam power).
\begin{figure}[htb!]
\centering
\includegraphics[width=0.75\columnwidth]{figures/Delta_CP_Error.jpg}
\caption{Expected 68\% CL uncertainty of $\delta_{CP}$ as a function of running time. Shown for the normal hierarchy case, with the mass hierarchy assumed to be known~\cite{HyperKamiokande:2018ofw}.}
\label{fig:Delta_CP_Error}
\end{figure}
| edcc77c5bad5a287bd784d1b68a7874139d65805 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Promoting a field theory Lagrangian from a Lorentz-invariant one to a generally-covariant one necessarily leads to an interaction between the fields of the theory and the gravitational field.
In the case of a scalar field, $S$, the natural generalization of this minimal interaction scenario is to introduce a non-minimal coupling term of the form
\begin{equation}
\label{eq:gen_nonmin_int}
\propto \xi_S S^2R \;.
\end{equation}
Here $R$ is the Ricci scalar and $\xi_S$ is a non-minimal coupling constant.
This non-minimal coupling to gravity proved to be useful in many applications to cosmology. Examples include Higgs inflation \cite{Bezrukov:2007ep,Lebedev:2021xey}, where $S$ is associated with the Higgs field degree of freedom $h$ --- the only scalar degree of freedom in the Standard Model, preheating \cite{Ema:2016dny}, where $S$ is associated with the inflaton field $\phi$, and non-perturbative production of dark matter \cite{nonminprod}, where $S$ represents the scalar dark matter particle $X$.
In the general case, when the fields $\phi$, $h$, and $X$ are all different, the question arises as to what extent they must interact with each other in order to successfully reheat the Universe and generate the right amount of dark matter.
Recent studies have shown that interactions via gravity alone, to which the fields are coupled minimally, is enough for these purposes. Indeed, the perturbative gravitational production of dark matter through graviton exchange can play a dominant role during reheating with processes involving the inflaton \cite{MO,CMOV,Barman:2021ugy} as well as thermal bath particles \cite{CMOV,Haque:2021mab}.
Further, the minimal gravitational coupling can lead to the completion of the reheating process for certain types of the inflationary potential, $V(\phi) \sim \phi^k$ with $k > 2$ \cite{CMOV,Haque:2022kez}. Thus, gravity is strong enough to mediate perturbative channels of reheating and dark matter production.
The purpose of this work is to study how the inclusion of the non-minimal coupling terms of the form (\ref{eq:gen_nonmin_int}) affect the gravitational production of dark matter and radiation during reheating.
Note that the presence of these terms is unavoidable: if there were no such couplings at tree level, they would still be generated by quantum corrections \cite{Callan:1970ze}.
We study particle production in the processes $hh\rightarrow XX$, $\phi\phi\rightarrow hh$, and $\phi\phi\rightarrow XX$ which are induced by the non-minimal couplings.
Here $\phi$ represents the inflaton background oscillating around its minimum after the end of inflation~\cite{gravprod}.
Since the scalar fields couple directly to the curvature scalar $R$, the oscillating background causes the effective masses of the fields to change non-adiabatically and leads to particle production.
This regime of particle creation has been considered in several different contexts, including gravitational production of scalar~\cite{gravscalar,ema}, fermion~\cite{gravferm}, and vector dark matter~\cite{gravvector}.
Our main interest is to compare the (dark) matter production channels induced by the non-minimal couplings with the production via the s-channel graviton exchange that sets minimal possible production rates. We will see for which values of the couplings the rates are enhanced, and what are the consequences on the dark matter density or the temperature attained during reheating.
Throughout the work we adopt the Starobinsky inflationary potential~\cite{staro}, although our results are largely independent of the particular form of the potential.
As for the potentials for the fields $h$ and $X$, we take them to be renormalizable polynomials. We also assume no direct interaction between $\phi$, $h$, and $X$.
Working in the perturbative regime implies that the non-minimal couplings must satisfy $|\xi_S| \ll M_P^2/\langle S \rangle^2$, where $\langle S \rangle$ is the vacuum expectation value of $S=\phi,h,X$. The value of $\xi_h$ is constrained from collider experiments as $|\xi_h|\lesssim 10^{15}$~\cite{higgscons1}.\footnote{Note that in the case of Higgs inflation, $\xi_h$ is fixed from CMB measurements~\cite{Bezrukov:2007ep}.} Furthermore, the lower bound on $\xi_h$ comes from the fact that the Standard Model electroweak vacuum may not be absolutely stable~\cite{HiggsStab}. To prevent the vacuum decay due to quantum fluctuations during inflation~\cite{fluc}, the effective mass of the Higgs field induced by the non-minimal coupling must be large enough; this gives $\xi_h\gtrsim 10^{-1}$~\cite{higgstree,higgsloop} (see also \cite{Markkanen:2018pdo}).\footnote{This estimate assumes no new physics interfering the RG running of the Higgs self-coupling constant until inflationary energy scales.}
The paper is organized as follows: The framework for our computation is presented
in Section~\ref{Sec:framework}. We discuss non-minimal gravitational couplings of the inflaton, the Higgs boson, and a dark matter scalar in detail. We calculate the dark matter production rates either from scattering in the thermal bath or from oscillations in the inflaton condensate. We compare similar processes obtained from the minimal gravitational particle production. We choose the Starobinsky model of inflation and discuss the reheating epoch when the inflaton begins oscillating. In Section~\ref{Sec:dm} we discuss the resulting abundance of dark matter produced from the thermal bath and directly from scattering of the inflaton condensate. We also compute the effects of the non-minimal couplings on the maximum temperature attained during reheating. We then compare different processes in Section~\ref{Sec:results}, before summarizing our results in Section~\ref{Sec:conclusion}.
\section{The framework}
\label{Sec:framework}
\subsection{Scalar-gravity Lagrangian}
\label{Ssec:lagr}
The theory we consider comprises 3 scalar fields non-minimally coupled to gravity: the inflaton $\phi$, the Higgs field\footnote{We consider the Higgs boson as a surrogate for any additional scalars with Standard Model couplings.} $H$, for which we adopt the Unitary gauge, $H=(0,h)^T/\sqrt{2}$, and the dark matter candidate $X$.
The relevant part of the action takes the form\footnote{The metric signature is chosen as $(+,-,-,-)$.}
\begin{equation}
\mathcal{S} \; = \; \int d^4 x \sqrt{-\tilde{g}} \left[-\frac{M_P^2}{2} \Omega^2 \tilde{R} +\mathcal{L}_{\phi} + \mathcal{L}_{h} + \mathcal{L}_{X} \right]
\label{eq:jordan}
\end{equation}
with the conformal factor $\Omega^2$ given by
\begin{equation}
\label{eq:conformalfact}
\Omega^2 \; \equiv \; 1 + \frac{\xi_{\phi} \phi^2}{M_P^2} + \frac{\xi_{h} h^2}{M_P^2} + \frac{\xi_{X} X^2}{M_P^2} \, .
\end{equation}
Here $M_P=2.4\times 10^{18}\,\rm{GeV}$ is the reduced Planck mass, and the tilde used in Eq.~(\ref{eq:jordan}) indicates that the theory is considered in the Jordan frame.
For the scalar field Lagrangians we have
\begin{equation}
\label{eq:L_S}
\mathcal{L}_{S} \; = \frac{1}{2} \tilde{g}^{\mu\nu} \partial_{\mu} S \partial_{\nu} S - V_{S} \, , ~~~ S=\phi,h,X \, .
\end{equation}
Next, we specify the scalar field potentials. For a model of inflation, we choose the well-motivated Starobinsky model for which~\cite{staro}
\begin{equation}
\label{eq:inflpot}
V_{\phi} \; = \; \frac{3}{4} m_{\phi}^2 M_P^2 \left(1 - e^{-\sqrt{\frac{2}{3}} \frac{\phi}{M_P}} \right)^2 \, .
\end{equation}
In what follows, we work in the perturbative regime with $\phi\ll M_P$, hence the potential is approximated as
\begin{equation}
\label{eq:inflpot_approx}
V_{\phi} \; \simeq \; \frac{1}{2} m_\phi ^2 \phi^2 \, .
\end{equation}
The inflaton mass, $m_\phi$, is fixed by the amplitude of scalar perturbations inferred from CMB measurements~\cite{Planck}; for the potential (\ref{eq:inflpot}) this gives $m_\phi = 3 \times 10^{13}$ GeV~\cite{building}.
The potential for the Higgs field is taken as follows
\begin{equation}
\label{eq:HiggsPot}
V_h \; = \; \frac{1}{2} m_h^2 h^2+\frac{1}{4} \lambda_h h^4 \, .
\end{equation}
Here $m_h$ and $\lambda_h$ are the Higgs mass and quartic coupling, correspondingly. Note that both parameters undergo the renormalization group (RG) running. In what follows we take a weak scale mass, which is a good approximation at the time of reheating and our results are insensitive to $\lambda_h$. Finally, the dark matter potential is simply given by
\begin{equation}
\label{eq:DMpot}
V_X \; = \; \frac{1}{2} m_X^2 X^2 \, .
\end{equation}
To study the reheating in the theory (\ref{eq:jordan}), it is convenient to remove the non-minimal couplings by performing the redefinition of the metric field.
Leaving the details to Appendix~\ref{sec:appendixA}, we write the action (\ref{eq:jordan}) in the Einstein frame,
\begin{equation}
\label{eq:einstein}
\begin{split}
\mathcal{S} \; = \; \int d^4 x \sqrt{-g} & \left[-\frac{M_P^2}{2}R + \frac{1}{2} K^{ij} g^{\mu\nu} \partial_{\mu}S_i \partial_{\nu}S_j \right. \\
& \qquad\qquad\qquad \left. - \frac{V_\phi+V_h+V_X}{\Omega^4} \right] \, .
\end{split}
\end{equation}
Here the indices $i,j$ enumerate the fields $\phi, h, X$, and the kinetic function is given by
\begin{equation}
\label{eq:kinfunc}
K^{i j} \; = \; 6 \frac{\partial \log \Omega}{\partial S_i} \frac{\partial \log \Omega}{\partial S_j} + \frac{\delta^{ij}}{\Omega^2} \, .
\end{equation}
Note that the scalar field kinetic term is not canonical.
In general, it is impossible to make a field redefinition that would bring it to the canonical form, unless all three non-minimal couplings vanish.\footnote{ Such a redefinition exists if the three-dimensional manifold spanned by the fields $\phi$, $h$ and $X$ is flat. One can show that it is not the case if at least one of the couplings is non-zero.}
For the theory~(\ref{eq:einstein}) to be well-defined, the kinetic function (\ref{eq:kinfunc}) must be positive-definite. Computing the eigenvalues, one arrives at the condition
\begin{equation}
\Omega^2 > 0 \, ,
\end{equation}
which is satisfied automatically for positive values of the couplings. Note that the negative couplings are also allowed for certain scalar field magnitudes.
In what follows, we will be interested in the small-field limit
\begin{equation}
\label{eq:smallfield}
\frac{|\xi_{\phi}| \phi^2}{M_P^2} \;, ~~ \frac{|\xi_{h}| h^2}{M_P^2} \;, ~~ \frac{|\xi_{X}| X^2}{M_P^2} \ll 1 \, .
\end{equation}
We can expand the kinetic and potential terms in the action (\ref{eq:einstein}) in powers of $M_P^{-2}$. We obtain a canonical kinetic term for the scalar fields and deduce the leading-order interactions induced by the non-minimal couplings.
The latter can be brought to the form
\begin{equation}
\label{lag4point}
\mathcal{L}_{\rm{non-min.}} \; = \; -\sigma_{hX}^{\xi} h^2 X^2 - \sigma_{\phi X}^{\xi} \phi^2 X^2 - \sigma_{\phi h}^{\xi} \phi^2 h^2 \, ,
\end{equation}
where the $\sigma_{ij}^{\xi}$ are functions of the couplings $\xi_i$, $\xi_j$, the masses $m_i$, $m_j$, and the Mandelstam variables; see Appendix~\ref{sec:appendixA} for details.
The small-field approximation~(\ref{eq:smallfield}) implies the bound $\sqrt{|\xi_{S}|} \lesssim M_P/\langle S\rangle$ with $S = \phi, h, X$.
Since the inflaton value at the end of inflation is $\phi_{\rm end}\sim M_P$ and afterwards $\langle \phi^2 \rangle \sim a^{-3} $, where $a$ is the cosmological scale factor, then $|\xi_\phi|\lesssim (a/a_{\rm end})^3$.
In particular, at the onset of inflaton oscillations
\begin{equation}
\label{eq:bound_on_xi_phi}
| \xi_\phi | \lesssim 1 \, .
\end{equation}
Note that since our calculations involve the effective couplings $\sigma_{\phi X}^{\xi}$ ($\sigma_{\phi h}^{\xi}$), which depend both on $\xi_\phi$ and $\xi_X$ ($\xi_h$), the relatively small value of $| \xi_{\phi} |$ can, in principle, be compensated by a large value of the other couplings.
In Fig.~\ref{fig:feyn1}, we show the scattering processes obtained from the Lagrangian~(\ref{lag4point}).
These contribute to reheating (when $h$ is in the final state) and dark matter production (when $X$ is in the final state).
\begin{figure}[ht!]
\includegraphics[width = 0.6\columnwidth]{feynman1.pdf}
\caption{\em \small Feynman diagram for the 4-point interactions between the inflaton $\phi$, the dark matter scalar candidate $X$, and the Higgs boson $h$, given by the Lagrangian~(\ref{lag4point}).}
\label{fig:feyn1}
\end{figure}
Finally, in evaluating the cosmological parameters, it is important to stay within the validity of the low-energy theory. The cutoff of the theory can be estimated as (see, e.g., \cite{higgscons2})
\begin{equation}
\Lambda\sim\frac{M_P}{\text{max}_i \:|\xi_i| } \, .
\end{equation}
In particular, the temperature of reheating must not exceed $\Lambda$.
\subsection{Graviton exchange}
\label{Ssec:gravExch}
Let us first consider the case of vanishing $\xi_{\phi, \, h, \, X}$, i.e., the case of the minimal coupling of the scalar fields to gravity~\cite{MO, CMOV,ema,Garny:2015sjg,Tang:2017hvq,Chianese:2020yjo,Redi:2020ffc}.
It was argued in~\cite{MO,CMOV} that the interaction between the dark and visible sectors induced by gravity leads to unavoidable contributions to reheating and dark matter production, in the thermal bath or via the scattering of the inflaton condensate, through the graviton exchange processes shown in Fig.~\ref{fig:feyn2}.
It is therefore important to compare the minimal gravitational particle
production to similar processes obtained from the
Lagrangian in Eq.~(\ref{lag4point}) with non-minimal couplings.
\begin{figure}[ht!]
\includegraphics[width = 1\columnwidth]{Feynman.pdf}
\caption{\em \small Feynman diagram for the (dark) matter production through the gravitational scattering of the inflaton or the Higgs boson from the thermal bath.
}
\label{fig:feyn2}
\end{figure}
To study the universal gravitational interactions in minimally coupled gravity, we expand the space-time metric around flat space using $g_{\mu \nu} \simeq \eta_{\mu \nu} + 2 h_{\mu \nu}/M_P$, where $h_{\mu \nu}$ is the canonically-normalized perturbation. The gravitational interactions are characterized by the following Lagrangian,
\begin{equation}
{\cal L}_{\rm min.}= -\frac{1}{M_P}h_{\mu \nu}
\left(T^{\mu \nu}_{h}+T^{\mu \nu}_\phi + T^{\mu \nu}_{X} \right) \, ,
\label{eq:lagrgrav}
\end{equation}
where the stress-energy tensor is given by
\begin{equation}
T^{\mu \nu}_{S} \; = \;
\partial^\mu S \partial^\nu S-
g^{\mu \nu}
\left[
\frac{1}{2}\partial^\alpha S \partial_\alpha S-V_S \right] \, .
\label{eq:tensors}
\end{equation}
Note that in this work, we consider only the Higgs field in the visible sector.
Generalization to the complete spectrum of the Standard Model is
straightforward, and we leave it for future work.
For models with minimally coupled gravity, the processes $\phi/h(p_1)+\phi/h(p_2) \rightarrow {h}/X(p_3)+{h}/X(p_4)$ can be parametrized by
\begin{equation}
\label{eq:partialamp}
\mathcal{M}^{00} \propto M_{\mu \nu}^0 \Pi^{\mu \nu \rho \sigma} M_{\rho \sigma}^0 \;,
\end{equation}
where the graviton propagator for the canonically-normalized field $h_{\mu\nu}$ with exchange momentum $k = p_1 + p_2$ is given by
\begin{equation}
\Pi^{\mu\nu\rho\sigma}(k) = \frac{\eta^{\mu \rho}\eta^{\nu\sigma} +
\eta^{\mu \sigma}\eta^{\nu \rho} - \eta^{\mu\nu}\eta^{\rho\sigma} }{2k^2} \, ,
\end{equation}
and the partial amplitude, $M_{\mu \nu}^0$, is given by
\begin{flalign}
\label{eq:partialamp2}
M_{\mu \nu}^{0} = \frac{1}{2}\left[p_{1\mu} p_{2\nu} + p_{1\nu} p_{2\mu} - \eta_{\mu \nu}p_1\cdot p_2 - \eta_{\mu \nu} V_S''\right] \, ,
\end{flalign}
with analogous expression for the final state in terms of outgoing momenta $p_{3, 4}$ and the final state potential.
In Fig.~\ref{fig:feyn2} we show the s-channel graviton exchange scattering obtained from the Lagrangian~(\ref{eq:lagrgrav}) for the production of dark matter from either the Higgs field or the inflaton condensate as well as the reheating process (the production of Higgs bosons from the inflaton condensate).
\subsection{Production rates}
In this work, we consider three processes:
\begin{enumerate}[A.]
\item{The production of dark matter from the
scattering of thermal Higgs bosons
(assuming reheating is produced by inflaton decay).
In this case, the dark matter
is populated via a freeze-in mechanism throughout the reheating period.}
\item{The production of dark matter from
direct excitations
of the inflaton condensate.
This process, which can be viewed as gravitational
inflaton scattering,
is independent of the presence of a thermal bath.}
\item{The creation of a radiative bath at the start of reheating arising from the Higgs boson production through gravitational inflaton scattering. Since such a process is unavoidable in
minimally coupled gravity,
it is interesting to know when such a process becomes
dominant in models with non-minimal couplings $\xi_i$.}
\end{enumerate}
The thermal dark matter production rate $R(T)$ for the process $h h \rightarrow X X$ can be calculated from\footnote{We include the symmetry factors associated with identical initial and final states in the definition of $|\overline{{\cal{M}}}|^2$, and a factor of 2 is explicitly included in the definition of the rate to account for the production of 2 identical particles.} \cite{gravitino}
\begin{equation}
\label{eq:rategen}
R(T) = \frac{2 \times N_h}{1024 \pi^6}\int f_1 f_2 E_1 \diff E_1 E_2 \diff E_2 \diff \cos \theta_{12}\int |\overline{{\cal M}}|^2 \diff \Omega_{13} \, ,
\end{equation}
where $E_i$ is the energy of particle $i=1,2$, $\theta_{13}$ and $\theta_{12}$ are the angles formed by momenta ${\bf{p}}_{1,3}$ and ${\bf{p}}_{1,2}$, respectively. $N_h = 4$ is the number of internal degrees of freedom for 1 complex Higgs doublet, $|\overline{{\cal M}}|^2$ is the matrix amplitude squared with all symmetry factors included. This accounts for the explicit factor of 2 in the numerator of Eq.~(\ref{eq:rategen}). The thermal distribution function of the incoming Higgs particles is given by the Bose-Einstein distribution
\begin{equation}
\label{eq:boseeinstein}
f_{i}= \frac{1}{e^{E_i/T} - 1} \, .
\end{equation}
The rate for minimal gravitational interactions from Eq.~(\ref{eq:lagrgrav})
was derived in \cite{CMOV,Bernal:2018qlk}. The rate we use here differs in two respects. As noted earlier, we only include Higgs scalars in the initial state whereas in \cite{CMOV,Bernal:2018qlk}, all Standard Model particle initial states were included. Secondly, we keep terms depending on the dark matter mass which had not previously been taken into account. This allows us to consider dark matter masses approaching the inflaton mass and/or the reheating temperature.
For minimal (non-minimal) gravitational interactions, we find that the thermal dark matter production rate can be expressed as
\begin{flalign}
\label{eq:ratefull1}
R_X^{T, \, (\xi)} (T) & = & \beta_1^{(\xi)} \frac{T^8}{M_P^4} + \beta_2^{(\xi)} \frac{m_X^2 T^6}{M_P^4} + \beta_3^{(\xi)} \frac{m_X^4 T^4}{M_P^4} \, ,
\end{flalign}
where the coefficients $\beta_{1, \, 2, \, 3}^{(\xi)}$ are given in Appendix~\ref{sec:appendixB} by Eqs.~(\ref{eq:b1}-\ref{eq:b3})
(Eqs.~(\ref{eq:b1xi}-\ref{eq:b3xi})).
The ratio of the non-minimal to minimal rate is shown in Fig.~\ref{plotsxi1}. However, we note that when $\xi_i \sim \mathcal{O}(1)$ both rates are comparable and interference effects become significant. The full coefficients $\beta_{1,\, 2,\, 3}$ including interference are given by Eqs.~(\ref{eq:b11}-\ref{eq:b33}) from Appendix~\ref{sec:appendixB}. We leave the comparison of the effects on dark matter production from the two rates for the next section.
\begin{figure}[ht!]
\includegraphics[width = 1.0\columnwidth]{plotsxi1.pdf}
\caption{\em \small Contours of the ratio of
the dark matter production rates from the thermal bath based on non-minimal gravitational interactions to those based on minimal interactions. The ratio is displayed in the $(\xi_h,\xi_X)$ plane. Note that as discussed in the Introduction, negative values of $\xi_h$ may require new physics (such as supersymmetry) to stabilize the Higgs vacuum. }
\label{plotsxi1}
\end{figure}
The rate for dark matter produced from inflaton oscillations of the inflaton condensate
for a potential of the form $V = \lambda \phi^k$ were considered in detail in \cite{GKMO2,CMOV}. The time-dependent inflaton can be written as $\phi(t) = \phi_0(t) \mathcal{P}(t)$, where $\phi_0(t)$ is the time-dependent amplitude that includes the effects of redshift and $\mathcal{P}(t)$ describes the periodicity of the oscillation.
The dark matter production rate is calculated by writing the potential in terms of the Fourier modes of the oscillations \cite{Ichikawa:2008ne,Kainulainen:2016vzv,GKMO2,CMOV}
\begin{equation}
V(\phi)=V(\phi_0)\sum_{n=-\infty}^{\infty} {\cal P}_n^ke^{-in \omega t}
=\rho_\phi\sum_{n = -\infty}^{\infty} {\cal P}_n^ke^{-in \omega t} \, .
\end{equation}
For $k=2$ (the only case considered here),
the frequency of oscillation is simply,
$\omega = m_\phi$.
The rate generated by non-minimal couplings
can be readily calculated using the Lagrangian~(\ref{lag4point}), which leads to
\begin{equation}
\label{eq:rateinfdm0}
R_X^{\phi, \, \xi} \; = \; \frac{2 \times \sigma_{\phi X}^{\xi~2}}{\pi} \frac{ \rho_{\phi}^2}{m_{\phi}^4} \Sigma_0^k \, ,
\end{equation}
where
\begin{equation}
\Sigma_0^k = \sum_{n = 1}^{\infty} |{\cal P}^k_n|^2
\sqrt{1-\frac{4m_X^2}{E_n^2}} \, ,
\label{Sigma0k}
\end{equation}
and $E_n = n \omega$ is the energy of the $n$-th inflaton oscillation mode.
For $k=2$, only the second Fourier mode in the sum contributes, with $\sum |\mathcal{P}^2_n|^2 = \frac{1}{16}$. Thus, the rate becomes
\begin{equation}
\label{eq:rateinfdm1}
R_X^{\phi, \, \xi} \; = \; \frac{2 \times \sigma_{\phi X}^{\xi~2}}{16\pi} \frac{ \rho_{\phi}^2}{m_{\phi}^4} \sqrt{1 - \frac{m_X^2}{m_{\phi}^2}}\, ,
\end{equation}
where $\rho_{\phi}$ is the energy density of the inflaton and the interaction term $\sigma_{\phi X}^{\xi}$ is given in Appendix~\ref{sec:appendixA} by Eq.~(\ref{appa:sigphiX}).
It was shown in~\cite{MO} that the dark matter production rate through
the exchange of a graviton, computed from the partial amplitude~(\ref{eq:partialamp}), is
\begin{equation}
\label{eq:ratescalar2}
R^{\phi}_X=\frac{ 2 \times \rho_\phi^2}{256 \pi M_P^4}
\left(1+\frac{m_X^2}{2m^2_\phi}\right)^2\sqrt{1-\frac{m_X^2}{m_\phi^2}} \, ,
\end{equation}
which can be written in the same form as (\ref{eq:rateinfdm1}) by defining an effective coupling $\sigma_{\phi X}$
\begin{equation}
\sigma_{\phi X} \; = \; -\frac{m_{\phi}^2}{4M_P^2} \left(1 + \frac{m_X^2}{2m_{\phi}^2} \right) \, .
\end{equation}
A comparison of the non-minimal to minimal rates for the production of dark matter from inflaton scattering is shown in Fig.~\ref{plotsxi2}.
\begin{figure}[t!]
\includegraphics[width = 1.0\columnwidth]{plotsxi2.pdf}
\caption{\em \small Contours of the ratio of
the dark matter production rates from oscillations in the inflaton condensate based on non-minimal gravitational interactions to those based on minimal interactions. The ratio is displayed in the $(\xi_\phi,\xi_X)$ plane.}
\label{plotsxi2}
\end{figure}
For the production of Higgs bosons through inflaton condensate scattering, we follow a similar procedure, and from the Lagrangian~(\ref{lag4point}) we find
\begin{equation}
\label{eq:ratehiggs}
R_h^{\phi, \, \xi} \; \simeq \; N_h \frac{2 \times \sigma_{\phi h}^{\xi~2}}{16 \pi} \frac{ \rho_{\phi}^2}{m_{\phi}^4} \, ,
\end{equation}
where we assumed that $m_h \ll m_{\phi}$, $N_h = 4$ is the number of internal degrees of freedom for 1 complex Higgs doublet, and $\sigma_{\phi h}^{\xi}$ is given in Appendix~\ref{sec:appendixA} by Eq.~(\ref{appa:sigphih}).
On the other hand, it was argued in~\cite{CMOV} that the scattering
$\phi \phi \rightarrow h h$ through the graviton exchange can
also be parameterized by an effective coupling
\begin{equation}
\mathcal{L}_h = - \sigma_{\phi h} \phi^2 h^2 \, ,
\end{equation}
with
\begin{equation}
\sigma_{\phi h} \; = \; -\frac{m_{\phi}^2}{4 M_P^2} \, ,
\label{sph}
\end{equation}
and the rate $R_h^{\phi}$ is given by the analogous expression to~(\ref{eq:ratehiggs}) with $\sigma_{\phi h}^{\xi}$ replaced by $\sigma_{\phi h}$.
The full four-point coupling of course is given by
the sum $\sigma_{\phi h/X}^{\xi} + \sigma_{\phi h/X}$.
However, except for values where the two are similar, which occurs when $12 \xi^2 + 5 \xi \simeq \frac12$ (assuming $m_X \ll m_\phi$ and taking the $\xi_i$ to be equal to $\xi$), either the minimal or the non-minimal contribution dominates. Thus, for the most part, we will consider separately the minimal and non-minimal contributions. Note that for two values of $\xi$ ($\xi \sim -1/2$ and 1/12) destructive interference could occur causing the entire rate to vanish (at the tree level).
\section{Particle Production with a Non-Minimal Coupling}
\label{Sec:dm}
Given the rates $R_i^j$ calculated in the previous section, we compute the
evolution for the gravitational (minimal and non-minimal)
contribution to the reheating processes
and the dark matter density for the three reactions outlined above.
\subsection{$h~h \rightarrow X~X$}
The gravitational scattering of thermal Higgs bosons leads to the production of massive scalar dark matter particles $X$. The dark matter number density $n_X$ can be calculated from the classical Boltzmann equation
\begin{equation}
\frac{dn_X}{dt} + 3Hn_X = R^T_X \,,
\label{eq:boltzmann1}
\end{equation}
where $H=\frac{\dot a}{a}$ is the Hubble parameter and the right-hand side of the equation represents the dark matter production rate.
It is more practical to rewrite the above equation in terms of the scale factor $a$ rather than the parameters $t$ or $T$.
We proceed by introducing the comoving number density $Y_X= n a^3$ and rewriting the Boltzmann equation as
\begin{equation}
\frac{dY_X}{da}=\frac{a^2R^T_X(a)}{H(a)} \, .
\label{eq:boltzmann3}
\end{equation}
Since the production rate (\ref{eq:ratefull1}) is a function of the temperature of the thermal bath, it is necessary to determine the relation between $T$ and $a$
in order to solve the Boltzmann equation as a function of the scale factor $a$.
For the Starobinsky potential in Eq.~(\ref{eq:inflpot}), at the end of inflation, the inflaton starts oscillating about a quadratic minimum, and we find the following energy conservation equations\footnote{For the inflaton scattering with $V(\phi) \sim \phi^k$, where $k >2$, see \cite{Bernal:2019mhf,GKMO1,GKMO2,GKMOV,Haque:2021mab,Haque:2022kez,Ahmed:2021fvt}.}
\begin{eqnarray}
&&
\frac{d \rho_\phi}{dt} + 3H\rho_\phi = -\Gamma_\phi \rho_\phi \, ,
\label{eq:diffrhophi}
\\
&&
\frac{d \rho_R}{dt} + 4H\rho_R = \Gamma_\phi \rho_\phi \, ,
\label{eq:diffrhor}
\end{eqnarray}
where $\rho_{\phi}$ and $\rho_R$ are the energy density of the
inflaton and radiation, respectively, $\Gamma_{\phi}$ is the inflaton decay rate, and for a quadratic minimum, we are able to set the equation of state parameter $w_{\phi} = \frac{P_{\phi}}{\rho_{\phi}}\simeq 0$.
We will assume that reheating occurs
due to an effective inflaton coupling to the Standard Model fermions, given by the interaction Lagrangian
\begin{center}
\begin{equation}
\mathcal{L}_{\phi-SM}^{y} \; = \; - y \phi \bar{f} f \, ,
\end{equation}
\end{center}
where $y$ is a Yukawa-like coupling, $f$ is a Standard Model fermion, and the inflaton decay rate is
\begin{equation}
\Gamma_{\phi} \; = \; \frac{y^2}{8 \pi} m_{\phi} \, .
\end{equation}
If we solve the Friedmann equations~(\ref{eq:diffrhophi},~\ref{eq:diffrhor}), we find~\cite{GKMO1,GKMO2,CMOV}
\begin{equation}
\rho_\phi(a) = \rho_{\rm end} \left(\frac{a_{\rm end}}{a} \right)^3
\label{rhophia}
\end{equation}
and
\begin{equation}
\rho_R(a)=\rho_{\rm RH}\left(\frac{a_{\rm RH}}{a}\right)^{\frac{3}{2}}
\frac{1-\left(\frac{a_{\rm end}}{a}\right)^{\frac{5}{2}}}
{1-\left(\frac{a_{\rm end}}{a_{\rm RH}}\right)^{\frac{5}{2}}} \, ,
\label{Eq:rhoR}
\end{equation}
where $a_{\rm{end}}$ is the scale factor at the end of inflation, $\rho_{\rm end} \equiv \rho_{\phi}(a_{\rm{end}})$ is the inflaton energy density at the end of inflation when there is no radiation present, $a_{\rm{RH}}$ is the scale factor at reheating, and $\rho_{\rm RH} \equiv \rho_R(a_{\rm{RH}}) = \rho_{\phi}(a_{\rm{RH}})$ is the energy density at reheating. We note that these equations are strictly valid for $a_{\rm end} \ll a \ll a_{\rm{RH}}$ and the end of inflation occurs when $\ddot{a} = 0$ which corresponds to $\rho_{\rm end} = \frac{3}{2} V(\phi_{\rm{end}})$. For the Starobinsky potential, $\rho_{\rm end} \simeq 0.175 m_\phi^2 M_P^2$~\cite{egno5}.
The radiation energy density can be parameterized as
\begin{equation}
\rho_R=\frac{g_T\pi^2}{30}T^4\equiv \alpha T^4 \, ,
\end{equation}
where $g_T$ is the number of relativistic degrees of freedom at the temperature $T$. The maximum temperature is attained when the radiation energy density reaches its peak at $\rho_R(a_{\rm{max}}) = \alpha T_{\rm{max}}^4$. It was shown in~\cite{GKMO1} that the ratio of $a_{\rm max}$ to $a_{\rm end}$
is given by
\begin{equation}
\label{Eq:amax}
\frac{a_{\rm max}}{a_{\rm end}}=\left(\frac{8}{3}\right)^{\frac{2}{5}} \simeq 1.48 \, .
\end{equation}
Using Eq.~(\ref{Eq:rhoR}) we can then express the production rate
from gravitational scattering of thermal particles~(\ref{eq:ratefull1})
as a function of the scale factor $a$
\begin{equation}
R_{X}^{T, \,( \xi)}(a) \simeq \beta_1^{(\xi)} \frac{\rho_{\rm RH}^2}{\alpha^2M_P^4} \left(\frac{a_{\rm RH}}{a}\right)^{3}
\left[\frac{1 - \left(\frac{a_{\rm end}}{a} \right)^\frac{5}{2}}
{1 - \left(\frac{a_{\rm end}}{a_{\rm RH}} \right)^\frac{5}{2}}
\right]^2 \, ,
\label{rateXT}
\end{equation}
where we assumed that $m_{X} \ll m_{\phi},\,T$, and thus neglected the terms $\beta_{2, \, 3}^{(\xi)}$. If we use $H \simeq \frac{\sqrt{\rho_\phi(a)}}{\sqrt{3}M_P}$,
which is valid for $a \ll a_{\rm RH}$, we can rewrite Eq.~(\ref{eq:boltzmann3}) as
\begin{equation}
\frac{dY_X^{\xi}}{da}=\frac{\sqrt{3}M_P}{\sqrt{\rho_{\rm RH}}}a^2\left(\frac{a}{a_{\rm RH}}\right)^{\frac{3}{2}}R_X^{T, \, (\xi)}(a) \, .
\label{Eq:boltzmann3}
\end{equation}
We find that the solution to this equation is
\begin{flalign}
\label{Eq:nxthermal}
&~~n^{T, \, \xi}_X(a_{\rm RH}) =
\frac{2 \beta_1^{\xi}}{\sqrt{3} \alpha^2M_P^3}
\frac{\rho_{\rm RH}^{3/2}}{(1-(a_{\rm end}/a_{\rm RH})^{\frac{5}{2}})^2} \times
\nonumber \\
& \left(1 + 3 \left(\frac{a_{\rm end}}{a_{\rm RH}} \right)^\frac{5}{2} - \frac{25}{7}\left(\frac{a_{\rm end}}{a_{\rm RH}} \right)^\frac{3}{2} - \frac{3}{7}\left(\frac{a_{\rm end}}{a_{\rm RH}} \right)^5\right) \, ,
\end{flalign}
where we integrated Eq.~(\ref{Eq:boltzmann3}) in the interval $a_{\rm end} < a < a_{\rm RH}$.
The relic abundance is given by~\cite{book}
\begin{equation}
\Omega_Xh^2 = 1.6\times 10^8\frac{g_0}{g_{\rm RH}}\frac{n(T_{\rm RH})}{T_{\rm RH}^3}\frac{m_X}{1~{\rm GeV}} \, ,
\end{equation}
and if we combine it with Eq.~(\ref{Eq:nxthermal}), we obtain
\begin{eqnarray}
\Omega_X^{T, \, (\xi)} h^2 \; = \; \frac{2}{3} \Omega_k^{(\xi)} &&\left[1 + 3 \left( \frac{\rho_{\rm{RH}}}{\rho_{\rm{end}}}\right)^{\frac{5}{6}}
-\frac{25}{7} \left( \frac{\rho_{\rm{RH}}}{\rho_{\rm{end}}}\right)^{\frac{1}{2}}
\right.
\nonumber
\\
&&- \left.
\frac{3}{7} \left( \frac{\rho_{\rm{RH}}}{\rho_{\rm{end}}}\right)^{\frac{5}{3}} \right] \, ,
\label{oh2Txi}
\end{eqnarray}
with
\begin{equation}
\Omega_k^{(\xi)} = 1.6 \times 10^8 \frac{g_0}{g_{\rm RH}}
\frac{\beta_1^{(\xi)}\sqrt{3}}{\sqrt{\alpha} }
\frac{m_X}{1~\rm{GeV}}
\frac{T_{\rm RH}^3}{M_P^3}\left[1-\left(\frac{\rho_{\rm RH}}{\rho_{\rm end}}\right)^{\frac{5}{6}}\right]^{-2} ,
\label{ok}
\end{equation}
where $g_0 = 43/11$ and we use the Standard Model value $g_{\rm RH} = 427/4$.
We observe that $\Omega_X^{T, \, \xi} \propto \beta_1^{\xi} \, T_{\rm{RH}}^3$. Therefore large values of the couplings $\xi_{h}$ and $\xi_{X}$ would require a decrease in the reheating temperature. In Section \ref{Sec:results} we compare the scattering rates and the dark matter abundances with the minimally coupled case.
\subsection{$\phi~\phi \rightarrow X~X$}
Another mode of dark matter production is through the scattering of the inflaton itself. Whereas the graviton exchange channel
was treated with care in \cite{MO, CMOV}, in the case of
non-minimal coupling
it suffices to replace $R_X^{T, \, \xi}$ in Eq.~(\ref{Eq:boltzmann3}) with the production rate~(\ref{eq:rateinfdm1}),
\begin{equation}
\frac{dY_X^{\xi}}{da}=\frac{\sqrt{3}M_P}{\sqrt{\rho_{\rm RH}}}a^2\left(\frac{a}{a_{\rm RH}}\right)^{\frac{3}{2}}R_X^{\phi, \, \xi}(a) \, ,
\label{Eq:boltzmann4}
\end{equation}
and to integrate
between $a_{\rm end}$ and $a_{\rm RH}$, which leads to
\begin{equation}
n_X^{\phi, \, \xi} (a_{\rm RH}) = \frac{\sigma_{\phi X}^{\xi~2} \rho_{\rm{RH}}^{3/2}M_P}{4 \sqrt{3} \pi m_{\phi}^4} \left[\left( \frac{a_{\rm{RH}}}{a_{\rm{end}}}\right)^{\frac{3}{2}} - 1 \right]\sqrt{1 - \frac{m_{X}^2}{m_{\phi}^2}} \, .
\label{Eq:nsphi}
\end{equation}
For $a_{\rm RH} \gg a_{\rm end}$, using Eq.~(\ref{rhophia})
we can express $n_X^{\phi,\,\xi}$ as a function of $\rho_{\rm end}$:
\begin{equation}
n_X^{\phi, \, \xi} (a_{\rm RH})\simeq \frac{\sigma_{\phi X}^{\xi~2} \rho_{\rm RH} \sqrt{\rho_{\rm end}} M_P}{4 \sqrt{3} \pi m_{\phi}^4} \sqrt{1 - \frac{m_{X}^2}{m_{\phi}^2}} \, ,
\label{n0phi}
\end{equation}
and we find
\begin{flalign}
\frac{\Omega_X^{\phi, \, \xi} h^2}{0.12}\simeq \frac{1.3 \times 10^{7} \sigma_{\phi X}^{\xi~2} \rho_{\rm{RH}}^{1/4} M_P^2}{m_{\phi}^3} \frac{m_X}{1 \, \rm{GeV}} \sqrt{1 - \frac{m_{X}^2}{m_{\phi}^2}} \, ,
\label{Eq:omegaphixi}
\end{flalign}
where we assumed the Starobinsky value for $\rho_{\rm end}$.
The analogous expression for models with minimally coupled gravity is found by replacing $\sigma_{\phi X}^{\xi} \rightarrow \sigma_{\phi X}$.
Up to this point we have assumed that the radiation is produced via the direct inflaton decay to a fermion pair. In the next subsection we discuss an unavoidable radiation production channel when the inflaton condensate scattering produces Higgs bosons in models with minimal and non-minimal coupling to gravity.
\subsection{$\phi~\phi \rightarrow h~h$}
Gravitational processes that produce dark matter can also
populate the
thermal bath in the same way. Even if this
Planck-suppressed production mechanism does not dominate throughout the entire reheating process,
it was shown in
\cite{CMOV} that for $T_{\rm RH} \lesssim 10^{9}$ GeV it is graviton
exchange that dominates the production of the thermal bath at the
very beginning of the reheating, when $\rho_\phi\sim \rho_{\rm end}$.
In fact, it was shown that the maximal temperature reached, $T_{\rm max}$,
(which can be considered as an absolute lower bound on $T_{\rm max}$)
is $T_{\rm max}\sim 10^{12}$ GeV.
It is therefore natural to determine the value of the couplings ($\xi_\phi$, $\xi_h$), for which
non-minimal gravitational processes generate the thermal bath at early times, and the
maximal temperature which can be attained by these processes.
Following the discussion in the previous subsection,
to compute the radiation energy density produced by gravitational couplings
we implement the rate
$R_h^{\phi,\,\xi}$ (\ref{eq:ratehiggs}) into the Friedmann equation~(\ref{eq:diffrhor})
\begin{equation}
\label{eq:rhorad1}
\frac{d \rho_R}{dt} + 4H\rho_R \; \simeq \; N_h \frac{\sigma_{\phi h}^{\xi~2}}{8\pi} \frac{ \rho_{\phi}^2}{m_{\phi}^3} \, ,
\end{equation}
where we took into account that each scattering corresponds
to an energy transfer of $2 m_\phi$.\footnote{Or equivalently that each Higgs quanta
carries an energy $m_\phi$.}
The solution to this equation is
\begin{equation}
\rho_R \; = \; N_h \frac{\sqrt{3} \sigma_{\phi h}^{\xi~2}}{4\pi} \frac{\rho_{\rm end}^{3/2}M_P}{m_{\phi}^3} \left[ \left( \frac{a_{\rm{end}}}{a}\right)^4 - \left(\frac{a_{\rm{end}}}{a}\right)^{\frac{9}{2}} \right] \, .
\label{rhoRphi}
\end{equation}
Note that the dependence on the scale factor $a$ is very different from
that found in Eq.~(\ref{Eq:rhoR}) due to inflaton decay.
Indeed, the Higgs bosons produced by gravitational
scattering
(minimal as well as non-minimal) are redshifted to a greater extent because of the high dependence
of the rate on their energy due to the form of the energy-momentum
tensor $T_{\mu \nu}^0$. Since $\rho_R \propto a^{-4}$ in Eq.~(\ref{rhoRphi}) (at large $a$) and $\rho_\phi \propto a^{-3}$ in Eq.~(\ref{rhophia}), reheating through this process does not occur (i.e., $\rho_R$ never comes to dominate the total energy at late times) and inflaton decay is necessary.\footnote{This conclusion is avoided if the inflaton potential about minimum is approximated by $\phi^k$ with a higher power of $k>4$ \cite{CMOV,Haque:2022kez}.}
However, as in the case of the reheating from the inflaton decay,
the energy density in Eq.~(\ref{rhoRphi})
exhibits a maximum when $a = a_{\rm max} = (81/64) a_{\rm end}$.
The maximum radiation density is then,
\begin{equation}
\rho_{\rm{max}}^{\xi} \; \simeq \; N_h \frac{\sigma_{\phi h}^{\xi~2}}{12 \sqrt{3} \pi} \frac{\rho_{\rm end}^{3/2}M_P}{m_{\phi}^3} \left(\frac{8}{9} \right)^8 \, ,
\end{equation}
and from this expression we find that the maximum temperature produced by gravitational interactions is given by
\begin{eqnarray}
T_{\rm{max}}^{\xi} &\simeq& 6.5 \times 10^{11} \left(\frac{|\sigma_{\phi h}^{\xi}|}{10^{-11}} \right)^{\frac{1}{2}}
{\rm GeV}
\label{tximax}
\\
&\simeq&1.8 \times 10^{12}\sqrt{|\xi|}\left(|5+12 \xi|\right)^{\frac{1}{2}} \left(\frac{m_{\phi}}{3\times10^{13}\,{\rm GeV}}\right)
{\rm GeV}\, ,
\nonumber
\end{eqnarray}
where we took $\xi_\phi=\xi_h=\xi$ in the last equality.
The analogous expression for models with minimally coupled gravity is found by replacing $\sigma_{\phi h}^{\xi} \rightarrow \sigma_{\phi h}$.
To compare the maximum temperature obtained by non-minimal interactions with respect to minimal gravitational interactions, we can rewrite Eq.~(\ref{tximax}) (now including minimal interactions in $T_{\rm max}^\xi$) as
\begin{equation}
T_{\rm{max}}^{\xi} \simeq 1.3 \times 10^{12} \left(\frac{|\sigma_{\phi h}^{\xi}+\sigma_{\phi h}|}{\sigma_{\phi h}} \right)^{\frac{1}{2}} {\rm GeV} \, .
\label{Eq:tximaxbis}
\end{equation}
The value of $\xi$ for which the maximum temperature
generated by the non-minimal coupling surpasses the one from graviton
exchange is shown in Fig.~\ref{plotTmaxxi} and is determined using
\begin{equation}
\sqrt{\frac{|\sigma_{\phi h}^\xi|}{|\sigma_{\phi h}|}}
=\sqrt{2 |\xi|}\left(|5 +12 \xi|\right)^{\frac{1}{2}} >1
\label{Eq:ratiotmax}
\end{equation}
which is satisfied when $\xi> 1/12$ or $\xi < -1/2$, as discussed earlier.
\begin{figure}[ht!]
\includegraphics[width = 1.0\columnwidth]{Tmax_xi.pdf}
\caption{\em \small The maximum temperature during reheating generated separately by minimal and non-minimal gravitational scattering of Higgs bosons in the thermal bath. }
\label{plotTmaxxi}
\end{figure}
As noted above and discussed in \cite{CMOV}, minimal (and non-minimal) gravitational interactions for a quadratic inflaton potential do not lead to the completion of the reheating process, thus requiring additional inflaton interactions for decay. Although radiation density produced in scattering falls off faster than that from decay,
at early time, the radiation density may in fact dominate and determine $T_{\rm max}$.
To determine when the $\phi~\phi \rightarrow h~h$ process leads to the maximum temperature, we rewrite Eq.~(\ref{Eq:rhoR}) as:
\begin{equation}
\rho^y_R \;=\; \frac{\sqrt{3} y^2 m_{\phi} M_P^3}{20 \pi} \left(\frac{\rho_{\rm end}}{M_P^4} \right)^{\frac{1}{2}} \left[
\left(\frac{a_{\rm end}}{a}\right)^{\frac{3}{2}} - \left(\frac{a_{\rm end}}{a}\right)^4 \right] \, .
\label{Eq:rhorsigma_sol2}
\end{equation}
Using Eq.~(\ref{Eq:amax}), we find that the maximum radiation density
produced by the inflaton decay is given by
\begin{equation}
\rho_{\rm max}^y \; = \; \frac{\sqrt{3} y^2 m_{\phi} M_P^3}{32 \pi} \left(\frac{\rho_{\rm end}}{M_P^4} \right)^{\frac{1}{2}} \left(\frac{3}{8} \right)^{\frac{3}{5}} \, .
\end{equation}
The maximum temperature is therefore determined by (non-minimal) gravitational
interactions when
\begin{equation}
y^2 \lesssim N_h \frac{8 \rho_{\rm end} \sigma_{\phi h}^{\xi~2}}{9 m_{\phi}^4} \left(\frac{8}{9} \right)^8 \left(\frac{8}{3} \right)^{\frac{3}{5}}
\end{equation}
or
\begin{equation}
y\lesssim 1.6~\sigma_{\phi h}^{\xi}
\sqrt{\frac{\rho_{\rm end}}{m_\phi^4}}
\simeq
5.4 \times 10^{4}~\sigma_{\phi h}^{\xi} \left(\frac{3 \times 10^{13} \, \rm{GeV}}{m_{\phi}} \right) \, .
\label{Eq:ymin}
\end{equation}
This leads to the following reheating temperature:
\begin{eqnarray}
T_{\rm{RH}} \; &\lesssim& \; 3.1 \times 10^{19} \sigma_{\phi h}^{\xi} \left(\frac{m_{\phi}}{3 \times 10^{13} \, \rm{GeV}} \right)^{-1/2}
{\rm GeV} \nonumber
\\
&\lesssim& 2.4\times 10^9 \left(\frac{m_\phi}{3 \times 10^{13}} \right)^{\frac{3}{2}}\xi(5 + 12 \xi)~{\rm GeV}
\end{eqnarray}
where $T_{\rm RH}$ is given by \cite{GKMO2}
\begin{equation}
\rho_\phi(a_{\rm RH})=\alpha T_{\rm RH}^4=
\frac{12}{25}\Gamma_\phi^2M_P^2=
\frac{3 y^4 m_{\phi}^2 M_P^2}{400 \pi^2} \, ,
\label{rhoRH}
\end{equation}
when the reheating temperature is determined by inflaton decay.
The primary effect of the gravitational scattering processes on reheating is the augmentation of $T_{\rm max}$ for sufficiently small
inflaton decay coupling, $y$. This can be seen
in Fig.~\ref{enden1} where we show the evolution of the energy density of radiation from scattering and decay as well as the energy density of the inflaton as a function of $a/a_{\rm end}$ for $\sigma_{\phi h}^{\xi} = 0$ and $\sigma_{\phi h}^{\xi}/\sigma_{\phi h} = 100$, respectively.
\begin{figure}[ht!]
\includegraphics[width = 1.0\columnwidth]{enden3.pdf}
\caption{\em \small Evolution of the inflaton density (blue) and the total radiation density (red), with radiation density produced from inflaton decays (dashed orange) and $\phi~\phi \rightarrow h~h$ scattering processes $\rho_R^{\sigma, \, \xi}$ (dotted green) and $\rho_R^{\sigma}$ (dash-dotted purple) with $\sigma_{\phi h}^{\xi}/\sigma_{\phi h} = 100$ (or $\xi_{\phi} = \xi_h = \xi \simeq -2.3~{or}~1.8$), as a function of $a/a_{\rm end}$ for a Yukawa-like coupling $y = 10^{-8}$ and $\rho_{\rm end} \simeq 0.175 \, m_{\phi}^2 M_P^2$ $\simeq 9 \times 10^{62} \, \rm{GeV}^4$. The black dashed lines corresponds to the ratios $a_{\rm{int}}/a_{\rm{end}} \simeq 150$ and $6500$, which agrees with Eq.~(\ref{aint}). The numerical solutions are obtained from Eqs.~(\ref{eq:diffrhophi}), (\ref{eq:diffrhor}), and (\ref{eq:rhorad1}).}
\label{enden1}
\end{figure}
As we saw in Eq.~(\ref{Eq:ratiotmax}), minimal gravitational interactions dominate over non-minimal interactions when $\sigma_{\phi h}^{\xi} < \sigma_{\phi h}$ or when
\begin{equation}
12 \xi_\phi \xi_h + 3 \xi_h + 2 \xi_\phi < \frac12 \, ,
\label{xicond1}
\end{equation}
when we neglect contributions proportional to the Higgs mass. In this case, the maximum temperature
is determined by gravitational interactions when $y \lesssim 2.1 \times 10^{-6}$ from Eq.~(\ref{Eq:ymin}) using $\sigma_{\phi h}$ from Eq.~(\ref{sph}). The evolution of the energy densities in this case is shown in Fig.~\ref{enden1} with $y = 10^{-8}$.
However as the energy density of radiation
after the maximum falls faster than $\rho_\phi$,
reheating in the Universe is determined by the inflaton decay.
For a sufficiently small coupling $y$, the energy density from
the decay dominates the radiation density at $a > a_{\rm int}$, where
\begin{equation}
\frac{a_{\rm int}}{a_{\rm end}} \simeq \left(\frac{5 \sigma_{\phi h}^2 N_h \rho_{\rm end}}{y^2 m_\phi^4} \right)^{2/5} \simeq 1.6 \left( \frac{\sigma_{\phi h} M_P}{y m_\phi} \right)^{4/5}\, .
\label{aint}
\end{equation}
For $\sigma_{\phi h} = 3.8 \times 10^{-11}$,
$m_\phi = 3 \times 10^{13}$ GeV, and $y = 10^{-8}$ we have $a_{\rm int} \approx 160 a_{\rm end}$,
as seen in the figure.
When Eq.~(\ref{xicond1}) is not satisfied,
non-minimal interactions may dominate as shown in the bottom panel of Fig.~\ref{enden1}, for $\sigma_{\phi h}^{\xi}=100 \sigma_{\phi h}$ and $y = 10^{-8}$. The cross-over can be determined from Eq.~(\ref{aint}) with the replacement $\sigma_{\phi h} \rightarrow \sigma_{\phi h}^{\xi}$.
In this example, $a_{\rm int} \approx 6500 a_{\rm end}$.
\section{Results}
\label{Sec:results}
We now turn to some general results that may be obtained from the framework described above.
Concerning the gravitational production of dark matter from
the thermal bath, the difficulty of populating the Universe
via the exchange of a graviton was already known \cite{Bernal:2018qlk,CMOV}.
Summing the minimal and non-minimal contributions in
Eq.~(\ref{oh2Txi}), we find for $\rho_{\rm RH}\ll\rho_{\rm end}$
\begin{eqnarray}
\frac{\Omega_X^T}{0.12}&\simeq&
\left[1+30f(\xi_h,\xi_X)\right]
\left(\frac{T_{\rm RH}}{10^{14}~{\rm GeV}}\right)^3\left(\frac{m_X}{4.0 \times 10^9~ {\rm GeV}}\right)
\label{Eq:omegaTtot}
\nonumber
\\
&=&\left[1+120\xi^2(1+6 \xi+12\xi^2) \right]
\nonumber \\
&& \times
\left(\frac{T_{\rm RH}}{10^{14}~{\rm GeV}}\right)^3\left(\frac{m_X}{4.0 \times 10^9~{\rm GeV}}\right)
\end{eqnarray}
with
\begin{equation}
f(\xi_h,\xi_X)=\xi_h^2 + 2\xi_h \xi_X + \xi_X^2 + 12\xi_h \xi_X \left(\xi_h + \xi_X + 4\xi_h \xi_X \right)
\nonumber
\end{equation}
where we assumed $\xi_h=\xi_X=\xi$ in the last equality,
for simplicity.
It is clear that, if we set $\xi=0$, i.e. if we consider
only graviton exchange, the reheating temperature necessary
to obtain a reasonable density respecting the data \cite{Planck}
is dangerously close to
the mass of the inflaton, even
for extremely large dark matter masses.
This problem had already been raised in \cite{Bernal:2018qlk}
and resolved in \cite{MO,CMOV} by considering the dark matter produced from
the (minimal) gravitational inflaton scattering.
On the other hand, from Eq.~(\ref{Eq:omegaTtot})
we see that there is another solution to this tension if one allows
for non-minimal gravitational couplings. Indeed, it is easy to see that for values of
$\xi_i \gtrsim 0.1$ ($f(\xi_h,\xi_X)\gtrsim{\frac{1}{30}}$), non-minimal gravitational production
dominates over graviton exchange. In this case, it becomes easier to obtain the
correct dark matter density for more reasonable values of
$T_{\rm RH}$ and/or $m_X$. For example, for a common value $\xi=\xi_h=\xi_X=1$,
a temperature of $T_{\rm RH}\sim 1.2\times 10^{13}$ GeV, thus
slightly below the inflaton mass, is sufficient
to produce an EeV dark matter candidate, whereas for
$\xi=1000$, $T_{\rm RH} \sim 10^{11}$ GeV will saturate the relic density for a 2.6 TeV dark matter mass. We show this result in Fig.~\ref{Fig:omegaT} where we plot the reheating temperature
needed to satisfy the relic density constraint as
function of $m_X$ for different value of $\xi$. For each value of $\xi$, the relic density exceeds $\Omega_X h^2 = 0.12$ above the corresponding curve.
As one can see, the line for $\xi=0$ is in the upper corner of the figure at high values of $T_{\rm RH}$ and $m_X$ and these drop significantly at higher values of $\xi$.
\begin{figure}[ht!]
\includegraphics[width = 1.0\columnwidth]{omegaT.pdf}
\caption{\em \small Region of parameter
space respecting the relic density constraint
$\Omega_X h^2=0.12$ in the plane ($m_X$,$T_{\rm RH}$) for different values of $\xi=\xi_h=\xi_X$ and
$\rho_{\rm end} \simeq 0.175 \, m_{\phi}^2 M_P^2$ in the
case of gravitational production from the thermal bath
$h~h \rightarrow X~X$. Both minimal and non-minimal
contributions are taken into account.
}
\label{Fig:omegaT}
\end{figure}
As was shown in \cite{MO, CMOV}, another possibility to avoid the necessity of
high reheating temperatures and/or dark matter masses is the production of matter from the oscillations within the inflaton condensate when the energy stored
in the condensate is much larger than the reheating temperature. A simple comparison between Eqs.~(\ref{oh2Txi}) and (\ref{Eq:omegaphixi}) shows that the
production of dark matter via inflaton scattering when $\xi_i \ne 0$ generally dominates over the production of dark matter from the thermal bath:
\begin{eqnarray}
\frac{\Omega_X^{\phi, \, \xi}}{\Omega_X^{T, \, \xi}}
&\simeq& 34\frac{(\sigma_{\phi X}^\xi)^2}{\beta_1^\xi}
\frac{M_P^5}{T_{\rm RH}^2 m_\phi^3} \nonumber
\\
&\simeq&
185 \frac{M_P m_\phi}{T_{\rm RH}^2}\frac{(5+12 \xi)^2}{1+6 \xi +12 \xi^2}
\gg 1 \, ,
\
\end{eqnarray}
where we took $\xi=\xi_\phi=\xi_h=\xi_X$ and
$m_X \ll m_\phi$ in the last equality.
We are therefore able to state that the relic density of dark matter
generated by the non-minimal gravitational scattering of the inflaton is
always much more abundant than that produced by the thermal bath.
Dark matter production from inflaton scattering via minimal graviton exchange also dominates over minimal gravitational thermal production \cite{CMOV}. This state of affairs is anything but surprising.
Indeed, the energy available in the inflaton condensate at the onset of
oscillations is much greater than that available in
the thermal bath during the reheating process.
As the scattering cross-sections are themselves highly dependent
on the energies through the energy-momentum tensor,
it is quite normal that inflaton scattering is the
dominant process for both minimal and non-minimal gravitational couplings.
Since inflaton scattering dominates in both the minimal and non-minimal gravitational interactions we can compare the two. We obtain
\begin{equation}
\frac{\Omega_X^{\phi, \, \xi}}{\Omega_X^\phi} = \frac{\sigma_{\phi X}^{\xi~2}}{\sigma_{\phi X}^{2}}
\simeq 4 \xi^2(5 +12 \xi)^2 \, ,
\end{equation}
and we see again that non-minimal interactions dominate when $\xi > 1/12$ or $< -1/2$.
We show in Fig.~\ref{Fig:omegatot} the region of the parameter space in the
($m_X$, $T_{\rm RH}$) plane allowed by the relic density constraint, adding all of the minimal and non minimal gravitational contributions, from inflaton scattering
and as well as Higgs scattering from the thermal bath taking
$\xi_\phi=\xi_h=\xi_X=\xi$. As expected, for
$\xi=0$ we recover the result found in \cite{CMOV}. As one can see, the difficulty in the
gravitational production from the thermal bath
is indeed alleviated as a reheating temperature $T_{\rm RH}\simeq 10^{11}$ GeV
allows for the production of a PeV scale dark matter candidate. If
in addition we introduce the
non-minimal couplings $\xi$, the necessary reheating temperature
to fit the Planck data may be as low as the electroweak scale
for a GeV candidate if $\xi\gtrsim 1000$.
\begin{figure}[ht!]
\includegraphics[width = 1.0\columnwidth]{omegatot.pdf}
\caption{\em \small Region of parameter
space respecting the relic density constraint
$\Omega_X h^2=0.12$ in the plane ($m_X$,$T_{\rm RH}$) for different values of $\xi_\phi=\xi_h=\xi_X=\xi$ and
$\rho_{\rm end} \simeq 0.175 \, m_{\phi}^2 M_P^2$ in the
case of production from gravitational inflaton scattering
$\phi~\phi \rightarrow X~X$. Both minimal and non-minimal
contributions are taken into account.}
\label{Fig:omegatot}
\end{figure}
Finally, we note that given the dark matter mass and
reheating temperature (if that sector of beyond the Standard Model physics were known), the contours in Fig.~\ref{Fig:omegatot} allow us to place an upper bound on the non-minimal couplings, $\xi$.
We can rewrite Eq.~(\ref{Eq:omegaphixi}) as
\begin{eqnarray}
\frac{\Omega_X h^2}{0.12} &=& 4.1\times 10^{-7}
(12 \xi^2 + 5\xi +\frac12)^2
\left(\frac{T_{\rm RH}}{10^{10} {\rm GeV}} \right)
\nonumber \\
&& \times \left(\frac{m_X}{1 {\rm GeV}} \right) \left(\frac{m_\phi}{3\times 10^{13} {\rm GeV}} \right) \, ,
\end{eqnarray}
when $m_X \ll m_\phi$ and $\xi = \xi_\phi = \xi_X$.
Then, for example, if $m_X = 1$ TeV, and $T_{\rm RH} = 10^9$ GeV, we obtain an upper limit of $|\xi| \lesssim 4$.
\section{Conclusions}
\label{Sec:conclusion}
In this paper, we have generalized the
{\it minimal} gravitational interactions
in the early Universe, i.e., the s-channel exchange of a graviton, to include {\it non-minimal}
couplings of all scalars to the Ricci curvature $R$. We consider
a scalar sector $S_i$ consisting of the inflaton condensate $\phi$, the Higgs field $H$ and a dark matter candidate $X$, and
we have analyzed the impact of couplings of the type
$\xi_i S_i^2 R$ on the reheating
process and dark matter production.
The latter can be generated
by the thermal Higgs
scattering or excitations of the inflaton,
both through minimal and non-minimal gravitational couplings.
Whereas the Higgs scattering through the exchange
of a graviton necessitates
a very large reheating temperature and/or dark matter mass in order to fulfill Planck CMB
constraints ($T_{\rm RH} \simeq 10^{14}$ GeV with $m_X \simeq 10^9$ GeV), for $\xi \gtrsim 0.1$,
the non-minimal coupling dominates the process and
alleviates the tension.
For $\xi \simeq 1000$, a dark matter mass of $\sim 1$ PeV with
$T_{\rm RH}\simeq 10^{10}$ GeV will satisfy the constraint, see Fig.~\ref{Fig:omegaT}.
However, thermal production is not the sole source of dark matter production through gravity.
When we include the contribution (necessarily present) of the
inflaton scattering,
we showed that the energy stored in the condensate
at the end of inflation compensates largely the reduced
gravitational Planck coupling. These processes yield
the correct relic abundance through minimal graviton exchange
for a dark matter mass of $\sim 10^8$ GeV
with $T_{\rm RH}\simeq10^{10}$ GeV, and the constraint is satisfied for a dark matter mass of $\sim 100$ GeV
and $T_{\rm RH} \gtrsim 10^4$ GeV if one adds non-minimal
couplings of the order
$\xi \simeq 100$ as we show in Fig.~\ref{Fig:omegatot}.
Gravitational inflaton scattering also affects the reheating process, producing a maximum temperature
$ \simeq 10^{12}$ GeV with minimal couplings, reaching as large as
$T_{\rm max}^\xi\simeq 5 |\xi| T_{\rm max} \simeq 10^{14}$ GeV for $\xi=100$ as one can see in Fig.~\ref{plotTmaxxi}.
This result can be re-expressed as an upper limit to
$|\xi|$ given values of $m_X$ and $T_{\rm RH}$.
We can not over-emphasize
that all of our results are unavoidable, in the sense that
they are purely gravitational, and do not rely on physics beyond the Strandard Mode.
The relic density of dark matter, and maximum temperature of the thermal bath computed here should be considered as lower bounds, that should be implemented in any extension of the Standard Model,
whatever is its nature.
Note added : During the completion of the manuscript,
some overlapping results were presented in \cite{Aoki:2022dzd}.
\noindent {\bf Acknowledgements. } The authors want
to thank Emilian Dudas for useful discussions. This work was made possible by with the support of the Institut Pascal at
Université Paris-Saclay during the
Paris-Saclay Astroparticle Symposium 2021, with the support of the P2IO Laboratory of Excellence (program “Investisse ments d’avenir” ANR-11-IDEX-0003-01 Paris-Saclay and ANR-10-LABX-0038), the P2I axis of the Graduate School Physics of Université Paris-Saclay, as well as IJCLab, CEA, IPhT, APPEC, the IN2P3 master projet UCMN and EuCAPT
ANR-11-IDEX-0003-01 Paris-Saclay and ANR-10-LABX-0038). This project has received support from the European Union’s Horizon 2020 research and innovation programme under the Marie Sk$\lslash$odowska-Curie grant agreement No 860881-HIDDeN.
The work of K.A.O. and A.S.~was supported in part by DOE grant DE-SC0011842 at the University of
Minnesota.
\section*{Appendix}
| 4cc0f5dbc67321bce1a0f88ee2646363363657bf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Let~$g\geq 1$ be an integer, and let~$\mathcal{H}_g$ be the Siegel upper
half space of degree~$g$, which consists of all symmetric $g\times g$
complex matrices with positive definite imaginary
part. Let~$a,b\in\{0,1\}^g$. Then the \emph{theta function} of
genus~$g$ and characteristic~$(a,b)$ is defined
on~$\mathbb{C}^g\times \mathcal{H}_g$ by the following exponential series:
\begin{equation}
\label{eq:theta-function}
\theta_{a,b}(z, \tau) = \sum_{m\in \mathbb{Z}^g}
\exp\left( i \pi (m + \tfrac a2)^t\tau (m+\tfrac a2)
+ 2 i \pi (m+\tfrac a2) ^t (z + \tfrac b2)\right).
\end{equation}
Theta functions appear in many areas of mathematics, from partial
differential equations to arithmetic geometry; an overview is given
in~\cite{igusa_ThetaFunctions1972, mumford_TataLecturesTheta1983,
mumford_TataLecturesTheta1984}. They have symmetries with respect to
the action of the modular
group~$\Sp_{2g}(\mathbb{Z})$~\cite[§II.5]{mumford_TataLecturesTheta1983}, and
they also satisfy the Riemann relations, a broad generalization of the
well-known duplication
formula~\cite[§II.6]{mumford_TataLecturesTheta1983}. \emph{Theta
constants} are the values of these functions taken at~$z=0$, and are
of interest in number theory. Each theta constant is a Siegel modular
form, and every Siegel modular form for~$\Sp_{2g}(\mathbb{Z})$ has an
expression as a rational fraction in terms of theta
constants~\cite{igusa_GradedRingThetaconstants1964}; even a
polynomial, if~$g\leq 3$~\cite{igusa_GradedRingThetaconstants1964,
freitag_VarietyAssociatedRing2019}.
In this paper, we are interested in algorithms to evaluate theta
constants at a given point~$\tau\in \mathcal{H}_g$, or more generally theta
functions at a given point~$(z, \tau)$, to precision~$N$ for some
integer~$N\geq 0$. In the whole paper, we consider absolute precision:
the output will be a finitely encodable (for instance, dyadic) complex
number~$x$ such that~$\abs{\theta(z,\tau) - x} \leq 2^{-N}$.
Two main approaches to computing theta functions exist. The first one,
sometimes called the \emph{naive algorithm}, consists in computing
partial sums of the series~\eqref{eq:theta-function} and obtaining an
upper bound on the modulus of its
tail~\cite{deconinck_ComputingRiemannTheta2004,
enge_ShortAdditionSequences2018,
frauendiener_EfficientComputationMultidimensional2019,
agostini_ComputingThetaFunctions2021}. The resulting algorithm can
be applied in any genus; its complexity is~$O(\M(N) N^{g/2})$
if~$(z,\tau)$ is
fixed~\cite[Prop.~4.2]{labrande_ComputingThetaFunctions2016}, and can
be made uniform in~$(z,\tau)$ if this input is suitably
reduced~\cite[Thm.~3 and Thm.~8]{deconinck_ComputingRiemannTheta2004}.
The second approach was first described by
Dupont~\cite{dupont_FastEvaluationModular2011,
dupont_MoyenneArithmeticogeometriqueSuites2006} in the case of theta
constants of genus~$g\leq 2$. It combines the arithmetic-geometric
mean (AGM), and higher-dimensional analogues of the AGM called
Borchardt means, with Newton iterations, and claims a complexity
of~$O\paren[\big]{\M(N)\log N}$ binary operations. Extensions to theta
functions in genus~$g\leq 2$, as well as higher genera, were then
described in~\cite{labrande_ComputingJacobiTheta2018,
labrande_ComputingThetaFunctions2016}. In practice, these algorithms
beat the naive method for precisions greater than a few hundred
thousand bits for~$g=1$, and a few thousand bits
for~\mbox{$g=2$}. This improvement is especially welcome in
number-theoretic applications, where huge precisions are often
necessary to recognize rational numbers from their complex
approximations~\cite{enge_ComputingModularPolynomials2009,
enge_ComplexityClassPolynomial2009,
enge_ComputingClassPolynomials2014,
kieffer_EvaluatingModularEquations2021}, although the naive method
remains superior for~$g=1$ in the current range of practical
applications.
In order to prove the correctness of an algorithm based on Newton's
method, and establish an upper bound on its complexity, the first step
is usually to show that the linearized system that Newton's method
uses is actually invertible. A proof of this fact is currently missing
for~$g\geq
2$~\cite[§10.2]{dupont_MoyenneArithmeticogeometriqueSuites2006},
\cite[Conj.~3.6]{labrande_ComputingThetaFunctions2016}. For~$g=1$,
the invertibility of this linear system was
proved~\cite[Prop.~11]{dupont_FastEvaluationModular2011},
\cite[Prop.~4.4]{labrande_ComputingJacobiTheta2018}, but the rate of
convergence of the resulting Newton scheme was not made explicit. This
makes these algorithms difficult to implement in a provably correct
way.
The purpose of the present paper is to turn the quasi-linear time
algorithms for theta constants in genus~$1$ and~$2$, as well as theta
functions in genus~$1$, into provably correct algorithms. This is done
by giving explicit upper bounds on derivatives of certain analytic
functions derived from Borchardt sequences on explicit polydisk
neighborhoods of the points where Newton's method is applied.
In the case of genus~$2$ theta constants, we also show how to combine
Newton's method with the naive algorithm to obtain a uniform
quasi-linear complexity on the Siegel fundamental domain, thus
generalizing earlier constructions in
genus~$1$~\cite[Thm.~5]{dupont_FastEvaluationModular2011},
\cite[§4.2]{labrande_ComputingJacobiTheta2018}. In the case of theta
functions in genus~$2$, and higher genera, we are no longer able to
prove that Newton's method will succeed for all inputs. However, if it
succeeds, then the same methods can be applied to certify the
correctness of the result.
The paper is organized as follows. In \cref{sec:newton}, we give a
general result of explicit convergence for Newton schemes involving
multivariate analytic functions. We study Borchardt means as analytic
functions in detail in \cref{sec:borchardt}. In \cref{sec:dupont}, we
review the existing Newton schemes for the computation of theta
functions; then, we obtain explicit values for the magnitudes and
radii of convergence of the analytic functions defining them, and thus
explicit convergence results. Finally, we present the uniform
algorithm to compute genus~$2$ theta constants in \cref{sec:unif}.
\section{Certified multivariate Newton iterations}
\label{sec:newton}
In this section, we are interested in designing provably correct
Newton schemes for multivariate analytic functions, assuming that the
system is linearized using finite differences at each step. More
precisely, let~$\mathcal{U}$ be an open set in~$\mathbb{C}^r$, let~$f\colon \mathcal{U}\to \mathbb{C}^r$
be an analytic function, and let~$x_0\in \mathcal{U}$; assuming that~$f(x_0)$
is known and that~$f$ can be evaluated at any point, we are interested
in building a Newton scheme to compute~$x_0$ itself.
First, we give an explicit convergence result provided that the first
and second derivatives of~$f$ are locally bounded around~$x_0$, and
that~$df(x_0)$ is invertible. Using Cauchy's formula, we also obtain
explicit convergence estimates if we simply assume that~$f$ is bounded
on a certain polydisk around~$x_0$. Finally, we translate these
theoretical results into the concrete world of finite-precision
arithmetic. All these results are certainly well-known in spirit, but
we were unfortunately unable to find sufficiently explicit results in
the literature.
Let us introduce some notation. We always consider~$\mathbb{C}^r$ as a normed
vector space for the~$L^\infty$ norm, denoted simply
by~$\norm{\cdot}$: in terms of coordinates, we have
\begin{displaymath}
\norm{(x_1,\ldots,x_r)} = \max_{1\leq j\leq r} \abs{x_j}.
\end{displaymath}
If~$\rho>0$ and~$x\in \mathbb{C}^r$, we denote by~$\mathcal{D}_\rho(z)$ the open ball
(i.e.~the polydisk) centered in~$z$ of radius~$\rho$. We also denote the induced
norm of (multi-)linear operators by~$\nind{\cdot}$.
Let~$(e_i)$ be the canonical basis of~$\mathbb{C}^r$, and denote the
coordinates by~$x_1,\ldots, x_r$. If~$x\in \mathcal{U}$, then we have
\begin{displaymath}
df(x) = \sum_{i=1}^r \frac{\partial f}{\partial x_i}(x) \,dx_i,
\end{displaymath}
where~$dx_i$ is seen as the linear form~$x\mapsto x_i$.
For~$\eta>0$ such that~$\mathcal{D}_\eta(x) \subset \mathcal{U}$, we also define
\begin{displaymath}
\FD_\eta f(x) = \sum_{i=1}^r \frac{f(x+\eta e_i) - f(x)}{\eta} \,dx_i.
\end{displaymath}
This linear operator is an approximation of~$df(x)$ using finite
differences.
Assume we already know~$x\in \mathcal{U}$ such that~$\norm{x-x_0}\leq \varepsilon$ for
some~$\varepsilon>0$. Then we can formulate a Newton iteration step to refine
the approximation~$x$ of~$x_0$ as follows: simply replace~$x$
by~$x + h$, where
\begin{displaymath}
h = df(x)^{-1} (f(x_0)-f(x)).
\end{displaymath}
In the finite differences version, we take instead:
\begin{displaymath}
h = \FD_\eta f (x)^{-1} (f(x_0)-f(x)),
\end{displaymath}
where~$\eta>0$ is a suitably chosen small parameter. Then, provided
that~$\varepsilon$ is small enough,~$\norm{x+h-x_0}$ will be of the order
of~$\varepsilon^2$, ensuring quadratic convergence of the Newton
iteration.
\begin{prop}
\label{prop:theoretical-newton}
Let~$\mathcal{U}\subset\mathbb{C}^r$ be an open set, let~$f\colon \mathcal{U}\to\mathbb{C}^r$ be an
analytic function, and let~$x_0\in\mathcal{U}$. Let~$\rho>0$
and~$B_1,B_2,B_3 \geq 1$ be real numbers such
that~$\mathcal{D}_\rho(x_0)\subset\mathcal{U}$ and the following inequalities are
satisfied:
\begin{enumerate}
\item $\nind{df(x)}\leq B_1$ and~$\nind{d^2f(x)}\leq B_2$ for all~$x\in \mathcal{D}_\rho(x_0)$;
\item $df(x_0)$ is invertible and $\nind{df(x_0)^{-1}}\leq B_3$.
\end{enumerate}
Let~$\varepsilon,\eta>0$ be such that
\begin{displaymath}
\varepsilon \leq \min\set[\Big]{\frac{\rho}{2}, \frac{1}{2B_2B_3}}
\quad \text{and}\quad
\eta \leq \frac{\varepsilon}{4r B_1 B_3}.
\end{displaymath}
Then, for each~$x\in \mathbb{C}^r$ such that~$\norm{x - x_0}\leq \varepsilon$, if
we set
\begin{displaymath}
h = \FD_\eta f (x)^{-1} (f(x_0)-f(x)),
\end{displaymath}
we will have
\begin{displaymath}
\norm{x + h - x_0} \leq 2 B_2 B_3 \varepsilon^2.
\end{displaymath}
\end{prop}
\begin{proof}
First, note that
\begin{displaymath}
\nind{df(x) - df(x_0)} \leq B_2 \norm{x-x_0} \leq B_2\varepsilon \leq \frac{1}{2\nind{df(x_0)^{-1}}},
\end{displaymath}
so~$df(x)$ is also invertible, with~$\nind{df(x)^{-1}}\leq 2B_3$.
We can now study the ``usual'' Newton scheme. Let us write
\begin{displaymath}
f(x_0) = f(x) + df(x)(x_0-x) + v,
\end{displaymath}
for some vector~$v$ such that~$\norm{v}\leq \frac12 B_2\varepsilon^2$.
Let~$h_0 = df(x)^{-1}(f(x_0)-f(x))$. Then
\begin{equation}
\label{eq:x+h0-x0}
\norm{x+h_0 - x_0} = \norm{df(x)^{-1} v} \leq B_2 B_3 \varepsilon^2.
\end{equation}
Finally, we show that~$h$ is close
to~$h_0$. Since~$\mathcal{D}_\eta(x)\subset \mathcal{D}_\rho(x_0)$
(because~$\eta\leq \varepsilon\leq \rho/2$), we have for
each~$1\leq j\leq r$:
\begin{displaymath}
\abs[\Big]{\frac{f(x+\eta e_i) - f(x)}{\eta}
- \frac{\partial f}{\partial x_j}(x)} \leq \frac12 B_2\eta.
\end{displaymath}
Therefore,
\begin{displaymath}
\nind{\FD_\eta f(x) - df(x)} \leq \frac{r}{2} B_2 \eta \leq \frac{1}{4 B_3}
\leq \frac{1}{2\nind{df(x)^{-1}}},
\end{displaymath}
so that
\begin{displaymath}
\nind{\FD_\eta f(x)^{-1} - df(x)^{-1}} \leq 2 \nind{df(x)^{-1}}^2 \cdot \nind{\FD_\eta f(x) - df(x)}
\leq 4r B_2 B_3^2 \eta,
\end{displaymath}
and
\begin{equation}
\label{eq:h-h0}
\norm{h - h_0} \leq 4r B_2 B_3^2 \eta \norm{f(x) - f(x_0)}
\leq 4r B_1B_2B_3^2 \eta \varepsilon \leq B_2 B_3 \varepsilon^2.
\end{equation}
We obtain the result from~\eqref{eq:x+h0-x0},~\eqref{eq:h-h0}, and
the triangle inequality.
\end{proof}
Cauchy's integration
formula~\cite[Thm.~2.2.1]{hormander_IntroductionComplexAnalysis1966}
provides uniform upper bounds on~$\nind{df(x)}$ and~$\nind{d^2f(x)}$
for~$x\in \mathcal{D}_\rho(x_0)$ whenever a uniform upper bound on~$\norm{f}$
on a slightly larger polydisk is known; this makes the necessary data
in \cref{prop:theoretical-newton} easier to collect.
\begin{prop}
\label{prop:cauchy}
Let~$r,s\geq 1$, let~$x_0\in \mathbb{C}^r$, let~$\rho>0$, and
let~$f\colon \mathcal{D}_\rho(x_0)\to \mathbb{C}^s$ be an analytic
function. Let~$M\geq 0$ such that~$\norm{f(x)}\leq M$ for
all~$x\in D_\rho(x_0)$. Then for every~$n\geq 0$ and
every~$x\in \mathcal{D}_{\rho/2}(x_0)$, we have
\begin{displaymath}
\nind{d^nf(x)} \leq \frac{2^n n!}{\rho^n} \binom{n+r}{r} M.
\end{displaymath}
\end{prop}
\begin{proof}
It is enough to prove that
\begin{displaymath}
\nind{d^n f(x_0)} \leq \frac{n!}{\rho^n} \binom{n+r}{r} M
\end{displaymath}
for all~$n$; afterwards, we simply note
that~$D_{\rho/2}(x)\subset \mathcal{D}_\rho(x_0)$ for
each~$x\in \mathcal{D}_{\rho/2}(x_0)$. Write~$x_0 = (z_1,\ldots, z_r)$. We
compute the Taylor expansion of~$f$ at~$x_0$ using Cauchy's formula.
For each~$\zeta = (\zeta_1,\ldots,\zeta_r)\in \mathcal{D}_{\rho/2}(x_0)$, we
have
\begin{displaymath}
f(\zeta) = \sum_{n = (n_1,\ldots, n_r)\in \mathbb{N}^r} a_n(f) \prod_{j=1}^r (\zeta_j - z_j)^{n_j},
\end{displaymath}
where the Taylor coefficients~$a_n(f)\in \mathbb{C}^s$ are computed as
follows:
\begin{displaymath}
a_n(f) = \frac{1}{(2\pi i)^r} \int_{\partial \mathcal{D}_\rho(z_1)}\cdots \int_{\partial \mathcal{D}_\rho(z_r)}
\frac{f(x_1,\ldots,x_r)}{\prod_{j=1}^r (x_j - z_j)^{n_j+1}}
\ dx_1\cdots dx_r.
\end{displaymath}
In particular,
\begin{displaymath}
\norm{a_n(f)} \leq \frac{M}{\rho^{\sum_j n_j}}.
\end{displaymath}
Now, for each~$v\in \mathbb{C}^r$, the value of~$d^nf(x_0)(v,\ldots,v)\in \mathbb{C}^s$ is
given by all terms of total degree~$n$ in the Taylor expansion, up
to a factor of~$n!$:
\begin{displaymath}
d^nf(x_0)(v,\ldots,v) = n! \sum_{m\in \mathbb{N}^r,\ \sum m_j = n} a_m(f) \prod_{j=1}^r v_j^{m_j}.
\end{displaymath}
There are exactly~$\binom{n+r}{r}$ terms in the sum.
Since~$d^n f(x_0)$ is a symmetric operator, the result follows
easily.
\end{proof}
In order to run certified Newton iterations on a computer, showing a
theoretical convergence result is not enough: we also have to consider
precision losses, which will for instance prevent us from
choosing~$\eta$ too close to zero. Thankfully, Newton iterations are
self-correcting, and precision losses can be controlled by taking an
additional, explicit safety margin.
We adopt the following computational model for complex numbers. Dyadic
elements of~$\mathbb{C}^r$ (i.e.~elements of~$2^{-N}\mathbb{Z}[i]^r$ for
some~$N\in \mathbb{Z}$) are represented exactly; and for a
general~$z\in \mathbb{C}^r$, we call an \emph{approximation of~$z$ to
precision~$N$} a dyadic~$z'$ such that~$\norm{z-z'}\leq
2^{-N}$. Elementary operations on approximations of complex numbers
can be carried out using ball
arithmetic~\cite{tucker_ValidatedNumericsShort2011}. Recall that a
function~$C\colon \mathbb{Z}_{\geq 1}\to \mathbb{R}_{\geq 0}$ is called
\emph{superlinear} if~$C(m+n)\geq C(m)+C(n)$ for
all~$m,n\in \mathbb{Z}_{\geq 1}$.
\begin{thm}
\label{thm:practical-newton}
Let~$\mathcal{U}\subset\mathbb{C}^r$ be an open set, let~$f\colon \mathcal{U}\to\mathbb{C}^r$ be an
analytic function, and let~$x_0\in\mathcal{U}$. Let~$\rho\leq 1,M \geq 1,$
and~$B_3\geq 1$ be real numbers such that~$D_\rho(x_0)\subset \mathcal{U}$,
$\nind{f(x)}\leq M$ for each~$x\in \mathcal{D}_\rho(x_0)$,
and~$\nind{df(x_0)^{-1}}\leq B_3$. Let~$C\colon \mathbb{Z}_{\geq 1}\to \mathbb{R}$ be
a superlinear function such that the following holds:
\begin{itemize}
\item there exists an algorithm~$\mathcal{A}$ which, given an
exact~$x\in \mathcal{D}_\rho(x_0)$ and~$N\geq 0$, computes an approximation
of~$f(x)$ to precision~$N$ in~$C(N)$ binary operations;
\item two~$N$-bit integers can be multiplied in~$C(N)$ binary
operations;
\item we have~$C(2N)\leq KC(N)$ for some~$K\geq 1$ and for all~$N$
sufficiently large.
\end{itemize}
Then, given~$N\geq 0$,
an approximation of~$f(x_0)$ to precision~$N$, and an approximation
of~$x_0$ to precision
\begin{displaymath}
n_0 = 2\ceil[\big]{\log_2(2(r+1)M/\rho)} + 2\ceil[\big]{\log_2(B_3)} + 4,
\end{displaymath}
\cref{algo:newton} below computes an approximation of~$x_0$ to
precision~$N - \ceil[\big]{\log_2(B_3)} - 1$
in~$\bigO\paren[\big]{C(N)}$ binary operations; the hidden constant
in this complexity bound depends only on~$r,\rho,M,B_3,$ and~$K$.
\end{thm}
We now describe the algorithm. Let
\begin{displaymath}
B_1 = \frac{2(r+1)M}{\rho}
\quad\text{and}\quad
B_2 = \frac{2(r+1)(r+2)M}{\rho^2}.
\end{displaymath}
By \cref{prop:cauchy}, the real numbers $\rho/2,B_1,B_2,B_3$ meet the
conditions of \cref{prop:theoretical-newton}. Up to decreasing~$\rho$
and increasing~$B_1,B_2,B_3$, we may assume that they are all powers
of~$2$. Denote the given dyadic approximation of~$f(x_0)$ by~$z_0$.
\begin{algo}[Certified Newton iterations for analytic functions]~
\label{algo:newton}
\begin{enumerate}
\item Let~$n = n_0$, and let~$x$ be the given dyadic approximation
of~$x_0$ to precision~$n$.
\item While~$n < N$, do:
\label{step:loop}
\begin{enumerate}
\item
Let~$m = n + \log_2(B_1) + \log_2(B_3) + \ceil{\log_2(r)} + 2$,
and~$\eta = 2^{-m}$;
\item Using algorithm~$\mathcal{A}$, compute approximations of~$f(x)$
and~$f(x+\eta e_j)$ for all~$1\leq j\leq r$ to precision
$p = 2n + 2\ceil{\log_2(r)} + 2\log_2(B_1) + 2\log_2(B_3) + 9$;
\item Compute an approximation of the~$r\times r$ matrix~$M_1$
whose~$j$th column contains the finite
difference~$\frac 1\eta \paren[\big]{f(x+\eta e_j)-f(x)}$, for
all~$j$, to precision $p - \log_2(1/\eta)-1$ (entrywise);
\item
\label{step:inverse}
Compute an approximation of the~$r\times r$ matrix~$M_2 = M_1^{-1}$ to
precision $p' = p - \log_2(1/\eta) - 2\log_2(B_3) - 7$;
\item
\label{step:h}
Compute an approximation of the
vector~$h = M_2\paren[\big]{z_0 - f(x)}$ to precision
$p' + n - 1 - \log_2(B_1) - \ceil{\log_2(r)}$;
\item Let~$n' = 2n - \log(B_2) - \log(B_3) - 2$; replace~$x$ by a
dyadic approximation of~$x+h$ to precision~$n'+1$, and
replace~$n$ by~$n'$.
\end{enumerate}
\item Return~$x$.
\end{enumerate}
\end{algo}
\begin{proof}[Proof of \cref{thm:practical-newton}] We will show that
the different quantities appearing in \cref{algo:newton} can be
computed to the claimed precisions, and that~$x$ remains an
approximation of~$f^{-1}(z_0)$ to
precision~$n$. Since~$\nind{df(x)^{-1}}\leq 2B_3$ for
all~$x\in \mathcal{D}_\rho(x_0)$, the result will be an approximation
of~$x_0$ to precision~$N - \log_2(B_3) - 1$, as claimed.
At the beginning of each loop,~$x$ is dyadic, and so are
the~$x+\eta e_j$ for each $1\leq j\leq r$. Therefore, each entry
of~$M_1$ can be computed to precision~$p - \log_2(1/\eta) - 1$. Note
that~$\nind{\FD_\eta f(x)^{-1}}\leq 4B_3$ as a linear
operator. Let~$M_1'$ be a dyadic approximation of~$M_1$ to
precision~$p - \log_2(1/\eta)$; then we have
\begin{displaymath}
\nind{M_1 - M_1'} \leq \frac{1}{2\nind{M_1^{-1}}},
\end{displaymath}
so
that~$\nind{M_1^{-1}- M_1'^{-1}} \leq 2 \nind{M_1^{-1}}^2 \nind{M_1
- M_1'} \leq 32 B_3^2\, 2^{-p}/\eta$. This shows that~$M_1^{-1}$
can be computed to the required precision~$p'$ in
step~\eqref{step:inverse}. In step~\eqref{step:h}, we perform the
matrix-vector product using the schoolbook formula. The entries
of~$M_2$ have modulus at most~$4 B_3$, and are known up to
precision~$p'$; the entries of~$z_0-f(x)$ have modulus at
most~$2^{-n}B_1$, and are known up to precision~$p-1$. The total
error on the product can be bounded above by
\begin{displaymath}
r(4B_3 2^{-p+1} + 2^{-n} B_1 2^{-p'} + 2^{-p'-p-1}) \leq 2^{-n+1}r B_12^{-p'}.
\end{displaymath}
The precision~$p$ was chosen in such a way that we obtain, at the
end of the loop, an approximation of~$x+h$ to
precision~$2n - \log(B_3) - 1 \geq n'+1$. By
\cref{prop:theoretical-newton}, the result is also an approximation
of~$f^{-1}(z_0)$ to precision~$n'$.
The initial value of~$n_0$ ensures that~$n' > 3n/2$, so that number
of steps in the loop is~$O(\log N)$. Each loop involves a finite
number of elementary operations with complex numbers of
modulus~$O(1)$ at precision~$2n + O(1)$, where the hidden constants
depend only on~$r,\rho, M$, and~$B_3$; the cost of these
computations is~$O\paren[\big]{C(n)}$ binary operations. Since~$C$
is superlinear, the cost of the last loop dominates the cost of the
whole algorithm, a well-known feature of Newton's method.
\end{proof}
\section{Borchardt means as analytic functions}
\label{sec:borchardt}
The existing Newton schemes for the computation of theta
functions~\cite{dupont_FastEvaluationModular2011,
dupont_MoyenneArithmeticogeometriqueSuites2006,
labrande_ComputingJacobiTheta2018,
labrande_ComputingThetaFunctions2016} are based on \emph{Borchardt
means}, a higher-dimensional analogue of the classical
arithmetic-geometric mean
(AGM)~\cite{cox_ArithmeticgeometricMeanGauss1984}. Additional
references for the study of Borchardt means, especially in genus~$2$,
are~\cite{bost_MoyenneArithmeticogeometriquePeriodes1988,
jarvis_HigherGenusArithmeticgeometric2008}. Our goal in this
section is to study Borchardt means as analytic functions in detail,
obtaining explicit bounds on their magnitudes and radii of
convergence.
\subsection{Borchardt sequences}
\label{subsec:borchardt-def}
Fix~$g\geq 1$, and let~$\mathcal{I}_g = (\mathbb{Z}/2\mathbb{Z})^g$. A \emph{Borchardt
sequence} of genus~$g$ is by definition a sequence of complex
numbers
\begin{displaymath}
s = \paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g, n\geq 0}
\end{displaymath}
that satisfy the following recurrence relation: for every~$n\geq 0$,
there exists a choice of square roots~$\paren[\big]{t_b^{(n)}}_{b\in \mathcal{I}_g}$
of~$\paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g}$ such that for all~$b\in \mathcal{I}_g$, we have
\begin{equation}
\label{eq:borchardt-step}
s_b^{(n+1)} = \frac{1}{2^g} \sum_{b_1 + b_2 = b} t_{b_1}^{(n)} t_{b_2}^{(n)}.
\end{equation}
We say that~$\paren[\big]{s_b^{(n+1)}}_{\smash{b\in \mathcal{I}_g}}$ is the
result of a Borchardt step given by the choice of square
roots~$\paren[\big]{t_b^{(n)}}_{b\in\mathcal{I}_g}$ at the~$n$\th term. This
recurrence relation emulates the duplication formula satisfied by
theta constants~\cite[p.\,214]{mumford_TataLecturesTheta1983}, after
identifying~$\smash{\set{0,1}^g}$ with~$\mathcal{I}_g$ in the natural way: for
every~$\tau\in \mathcal{H}_g$, the sequence of squared theta constants
\begin{equation}
\label{eq:theta-as-borchardt}
\paren[\big]{\theta^2_{0,b}(0,2^n\tau)}_{b\in \mathcal{I}_g, n\geq 0}
\end{equation}
is a Borchardt sequence.
The convergence behavior of Borchardt sequences is similar to that of
the classical
AGM~\cite[§7.2]{dupont_MoyenneArithmeticogeometriqueSuites2006}. Let
us define a set of complex numbers to be \emph{in good position} if it
is included in an open quarter plane seen from the origin, i.e.~a set
of the form
\begin{displaymath}
\set[\big]{r \exp(i\theta) \colon r>0, \alpha < \theta < \alpha + \tfrac{\pi}{2}}
\end{displaymath}
for some angle~$\alpha\in \mathbb{R}$. We say that the~$n$\th step of a
Borchardt sequence is given by \emph{good sign choices} (or for short,
is \emph{good}) if the square
roots~$\paren[\big]{t_b^{(n)}}_{\smash{b\in\mathcal{I}_g}}$ are in good position;
otherwise, we say that this step is \emph{bad}. Then a Borchardt
sequence~$s$ will converge to~$(0,\ldots,0)$ if and only if~$s$
contains infinitely many bad steps. On the other hand, a Borchardt
sequence~$s$ in which all steps are good after a while converges to a
limit of the form~$(\mu,\ldots,\mu)$ for some~$\mu\neq 0$, and the
speed of convergence is quadratic; we call~$\mu = \mu(s)$ the
\emph{Borchardt mean} of the sequence. Borchardt sequences given by
theta functions as in~\eqref{eq:theta-as-borchardt} are of this second
type: see for
instance~\cite[Prop.~6.1]{dupont_MoyenneArithmeticogeometriqueSuites2006}.
A related kind of recurrent sequence is used in the context of
computing theta functions. Let us call an \emph{extended Borchardt
sequence} of genus~$g$ a pair~$(u,s)$ of sequence of complex numbers
\begin{displaymath}
u = \paren[\big]{u_b^{(n)}}_{b\in \mathcal{I}_g, n\geq 0},\qquad
s = \paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g, n\geq 0}
\end{displaymath}
satisfying the following recurrence relation: for every~$n\geq 0$,
there exists a choice of square
roots~$\paren[\big]{v_b^{(n)}}_{b\in \mathcal{I}_g}$
of~$\paren[\big]{u_b^{(n)}}_{b\in \mathcal{I}_g}$
and~$\paren[\big]{t_b^{(n)}}_{b\in \mathcal{I}_g}$
of~$\paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g}$ such that for all~$b$,
\begin{equation}
\label{eq:ext-borchardt-step}
u_b^{(n+1)} = \frac{1}{2^g} \sum_{b_1 + b_2 = b} v_{b_1}^{(n)} t_{b_2}^{(n)}
\quad \text{and}\quad
s_b^{(n+1)} = \frac{1}{2^g} \sum_{b_1 + b_2 = b} t_{b_1}^{(n)} t_{b_2}^{(n)}.
\end{equation}
In particular,~$s$ is a regular Borchardt sequence. We say that
the~$n$\th step in~$(u,s)$ is \emph{good} if both of the
sets~$\paren[\big]{v_b^{(n)}}_{b\in \mathcal{I}_g}$
and~$\paren[\big]{t_b^{(n)}}_{b\in \mathcal{I}_g}$ are independently in good
position, and \emph{bad} otherwise. For each~$\tau\in \mathcal{H}_g$
and~$z\in \mathbb{C}^g$, the duplication formula for theta functions implies
that the sequence
\begin{displaymath}
\paren[\big]{\theta^2_{0,b}(z, 2^n\tau), \theta^2_{0,b}(0, 2^n\tau)}_{b\in \mathcal{I}_g, n\geq 0}
\end{displaymath}
is an extended Borchardt sequence; it contains only finitely many bad
steps as well.
It is not true in general that an extended Borchardt sequence
containing finitely many bad steps converges quadratically. Instead,
following~\cite{labrande_ComputingThetaFunctions2016}, we define the
\emph{extended Borchardt mean} of such a sequence~$(u,s)$ to be
\begin{equation}
\label{eq:def-ext-borchardt}
\lambda(u, s) = \mu(s) \cdot
\lim_{n\to +\infty} \left(\frac{u_0^{(n)}}{s_0^{(n)}}\right)^{2^n}
= \mu(s) \cdot
\lim_{n\to +\infty} \left(\frac{u_0^{(n)}}{\mu(s)}\right)^{2^n}.
\end{equation}
These associated sequences do converge
quadratically~\cite[Prop.~3.7]{labrande_ComputingThetaFunctions2016}.
Assume that we are given a Borchardt sequence~$s$ containing finitely
many bad steps. Then we may try to construct a function~$\mu_s$,
defined at any point~$\smash{x = (x_b)_{b\in \mathcal{I}_g}}$ in some
neighborhood of~$\paren[\big]{s_b^{(0)}}_{\smash{b\in \mathcal{I}_g}}$, by the following procedure:
``construct a modified Borchardt sequence whose first term is~$(x_b)$
that follows same choices of square roots as in~$s$, and take its
Borchardt mean''. The Newton schemes we want to study are precisely
built around this kind of functions~$\mu_s$, and their analogues for
extended Borchardt means. In the rest of this section, we show that
these functions indeed exist as analytic functions defined on explicit
polydisks, provided that all terms in the relevant Borchardt sequences
are bounded away from zero.
\subsection{The case of good sign choices}
\label{subsec:borchardt-good}
Let~$s$ be a Borchardt sequence containing good steps only. Then we
can find real numbers~$0< m_0 < M_0$ and~$\alpha$ such that such that
the first term~$\paren[\big]{s_b^{(0)}}_{\smash{b\in \mathcal{I}_g}}$ of~$s$ lies in the open
set~$\mathcal{U}_g(m_0, M_0)$ of~$\mathbb{C}^{2^g}$ defined as follows:
\begin{displaymath}
\mathcal{U}_g(m_0, M_0) = \bigcup_{\smash{\alpha\in [0,2\pi]}} \mathcal{U}_g(m_0,M_0,\alpha),
\end{displaymath}
where
\begin{displaymath}
\mathcal{U}_g(m_0, M_0, \alpha) = \set[\big]{(x_b)_{b\in \mathcal{I}_g} \colon
\forall b\in \mathcal{I}_g, m_0 < \re(e^{-i\alpha}x_b) < M_0}.
\end{displaymath}
\begin{prop}
\label{prop:good-borchardt-analytic}
Let~$0< m_0 < M_0$ be real numbers. Then there exists a unique
analytic function~$\mu \colon \mathcal{U}_g(m_0, M_0)\to \mathbb{C}$ with the
following property: for every point
$x=(x_b)_{b\in \mathcal{I}_g} \in \mathcal{U}_g(m_0, M_0)$, the value of~$\mu$ at~$x$
is the Borchardt mean of the unique Borchardt sequence with first
term~$x$ given by good steps only. Moreover, the inequalities
$m_0 \leq \abs{\mu(x)} \leq M_0$ hold for all~$x\in \mathcal{U}_g(m_0,M_0)$.
\end{prop}
\begin{proof}
For each~$x\in \mathcal{U}_g(m_0, M_0,\alpha)$, there is a unique way of
making a good Borchardt step starting from~$x$; moreover the result
of this Borchardt step still lands in~$\mathcal{U}_g(m_0, M_0,\alpha)$
by~\cite[Lem.~7.3]{dupont_MoyenneArithmeticogeometriqueSuites2006}. Therefore
we may define~$\mu(x)$ as the limit of the resulting Borchardt
sequence; we have~$m_0\leq \abs{\mu(x)}\leq M_0$. Since there exists
an analytic square root function on~$\mathcal{U}_1(m_0,M_0,\alpha)$, the
function~$\mu$ on~$\mathcal{U}_g(m_0,M_0,\alpha)$ is the pointwise limit of a
sequence of analytic functions. The convergence is uniform on
compact sets
by~\cite[Prop.~7.2]{dupont_MoyenneArithmeticogeometriqueSuites2006},
so~$\mu$ is analytic on the whole of~$\mathcal{U}_g(m_0,M_0)$.
\end{proof}
We now consider the case of extended Borchardt means given by good
choices of square roots only. This case is easier to analyse if we
assume that the truly Borchardt part of the sequence already starts in
the quadratic convergence
area. By~\cite[Prop.~7.1]{dupont_MoyenneArithmeticogeometriqueSuites2006},
if we have
\begin{equation}
\label{eq:quadratic-area}
\abs[\big]{s_b^{(n)} - s_0^{(n)}} < \tfrac{\varepsilon}{4} \abs[\big]{s_0^{(n)}}
\end{equation}
for some~$\varepsilon \leq 1/2$, then we have
\begin{displaymath}
\abs[\big]{s_b^{(n+k)} - s_0^{(n+k)}} \leq
\frac{2}{7}\left(\frac{7\varepsilon}{8}\right)^{2^k}\cdot \max_{b\in \mathcal{I}_g}\ \abs[\big]{s_b^{(n)}}
\end{displaymath}
for all~$k\geq 0$ and~$b\in \mathcal{I}_g$. If we assume that the first term
of~$s$ lies in a ball of the form $\mathcal{D}_{\rho}(z_0)$ for
some~$z_0\in \mathbb{C}^\times$ and $0< \rho < \tfrac{1}{17} \abs{z_0}$, then
inequality~\eqref{eq:quadratic-area} will be satisfied
with~$\varepsilon = \frac12$ at~$n=0$.
\begin{prop}
\label{prop:good-ext-borchardt-analytic}
Let~$0< m_0<M_0$ be real numbers, fix a nonzero~$z_0\in \mathbb{C}$,
and let $0 < \rho < \tfrac{1}{17} \abs{z_0}$. Then there exists a
unique analytic function
\begin{displaymath}
\lambda\colon \mathcal{U}_g(m_0,M_0)\times \mathcal{D}_\rho(z_0)^{2^g}
\to \mathbb{C}
\end{displaymath}
with the following property: for every~$(x,y)$ in this open
set,~$\lambda(x,y)$ is equal to the extended Borchardt mean of any
extended Borchardt sequence with first term~$(x,y)$ given by good
steps only. Moreover, we have
\begin{displaymath}
\exp\left(-28\log^2(4M/m)\right) \leq \abs{\lambda(x,y)}
\leq \exp\left(20\log^2(4M/m)\right)
\end{displaymath}
where~$M = \max\{\abs{z_0}+\rho, M_0, 1\}$
and~$m = \min\{\abs{z_0}-\rho, m_0, 1\}$.
\end{prop}
\begin{proof}
We follow the proof
of~\cite[Thm.~3.10]{labrande_ComputingJacobiTheta2018}, and hints on
how to generalize it to higher genera given
in~\cite[Prop.~3.7]{labrande_ComputingThetaFunctions2016}. We may
fix~$\alpha\in\mathbb{R}$ and restrict our attention
to~$\mathcal{U}_g(m_0,M_0,\alpha)$.
First of all, by the proof
of~\cite[Lem.~3.8]{labrande_ComputingJacobiTheta2018},
each~$(x,y)\in \mathcal{U}_g(m_0,M_0,\alpha) \times \mathcal{D}_\rho(z_0)^{2^g}$ is
the starting point of at least one extended Borchardt
sequence~$(u,s)$ with good sign choices at all steps. Any two such
sequences differ at the~$n$\th term by global
multiplication~$\paren[\big]{u_b^{(n)}}_{\smash{b\in\mathcal{I}_g}}$ by a~$2^n$-th
root of unity; therefore, their extended Borchardt means are
equal. Note that~$M$ (resp.~$m$) is an upper (resp.~lower) bound on
the modulus of all complex numbers appearing in these extended
Borchardt sequences. In the rest of this proof, we
fix~$\theta_0\in\mathbb{R}$ such that~$\theta_0 = \arg(z_0)$ mod~$2\pi$, and
consider the unique such sequence~$(u,s)$ whose $n$\th term lies in
\begin{displaymath}
\mathcal{U}_g\left(m,M,\frac{\alpha + (2^n-1)\theta_0}{2^n}\right) \times
\mathcal{U}_g(m,M,0).
\end{displaymath}
Each term of~$(u,s)$ is an analytic function of its starting
point~$(x,y)$.
By construction, we have for all~$n\geq 0$:
\begin{displaymath}
\abs[\big]{s_b^{(n)} - s_0^{(n)}} < 2^{-2^n} \abs{z_0}.
\end{displaymath}
Let~$\mu$ be the Borchardt mean of~$s$. For~$n\geq 1$, write
\begin{displaymath}
q_n = \frac{(u_0^{(n)})^2}{u_0^{(n-1)} \mu},
\end{displaymath}
so that for all~$k\geq 0$, we have
\begin{displaymath}
\lambda(x,y) = \left(\frac{u_0^{(k)}}{\mu}\right)^{2^k}\prod_{n\geq k} q_{n+1}^{2^n}.
\end{displaymath}
These complex numbers~$q_n$ converge quadratically fast to~$1$. To
be more explicit, we have for all~$n\geq 1$:
\begin{align*}
\abs[\big]{u_0^{(n+1)} - v_0^{(n)} t_0^{(n)}}
&\leq \frac{\sqrt{M}}{2^g} \sum_{b\in\mathcal{I}_g} \left(\abs[\big]{t_b^{(n)} - t_0^{(n)}}
+ \abs[\big]{v_b^{(n)} - v_0^{(n)}}\right) \\
&\leq \frac{\sqrt{M}}{2^g \cdot 2\sqrt{m}}
\sum_{b\in\mathcal{I}_g} \left(\abs[\big]{s_b^{(n)} - s_0^{(n)}}
+ \abs[\big]{u_b^{(n)} - u_0^{(n)}}\right) \\
&\leq \frac{\sqrt{M}}{2\sqrt{m}}
\Bigl(2^{-2^n}\abs{z_0} + \frac{1}{2^g}\sum_{b\in \mathcal{I}_g} \abs[\big]{u_b^{(n)} - u_0^{(n)}} \Bigr).
\end{align*}
To bound the remaining sum, we write
\begin{align*}
\abs[\big]{u_b^{(n)} - u_0^{(n)}}
&\leq \frac{\sqrt{M}}{2^g}
\sum_{b'\in\mathcal{I}_g} \abs[\big]{t_{b+b'}^{(n-1)} - t_b^{(n-1)}} \\
&\leq \frac{\sqrt{M}}{2^g\cdot 2\sqrt{m}} \sum_{b'\in \mathcal{I}_g}\abs[\big]{s_{b+b'}^{(n-1)} - s_b^{(n-1)}}
\leq \frac{\sqrt{M}}{\sqrt{m}}\, 2^{-2^{n-1}} \abs{z_0}.
\end{align*}
Therefore, we have for all~$n\geq 1$
\begin{displaymath}
\abs[\big]{u_0^{(n+1)} - v_0^{(n)} t_0^{(n)}}
\leq \frac{5}{4} \sqrt{\frac{M}{m}}\ 2^{-2^{n-1}}\abs{z_0} =: B\cdot 2^{-2^{n-1}}.
\end{displaymath}
We deduce as in~\cite[Thm.~3.10]{labrande_ComputingJacobiTheta2018}
that
\begin{displaymath}
\abs{q_{n+1}-1} \leq B'\cdot 2^{-2^{n-1}}
\end{displaymath}
where
\begin{displaymath}
B' = 2\abs{z_0} + \frac{1}{m^2}(2MB + B^2) \leq \frac{5M^3}{m^3}.
\end{displaymath}
Let~$k\geq 1$ be minimal such that~$B' \cdot 2^{-2^{k-1}}\leq \frac12$. Then we have
\begin{displaymath}
\sum_{n\geq k} 2^n\log\abs{q_{n+1}} \leq \sum_{n\geq k}
2^n \cdot \frac{1}{2}\cdot 2^{2^{k-1}} \cdot 2^{-2^{n-1}} \leq 2^{k}.
\end{displaymath}
This proves that the sequence~\eqref{eq:def-ext-borchardt}
converges; since our estimates are uniform,~$\lambda$ must be
analytic. Moreover,
\begin{displaymath}
\abs{\lambda(x,y)} = \abs[\Big]{\frac{u_0^{(k)}}{\mu}}^{2^{k}}
\prod_{n\geq k} \abs{q_{n+1}}^{2^n}\leq \exp\left(2^{k}\paren[\big]{1+ \log(M/m)}\right).
\end{displaymath}
We obtain the final upper bound on~$\abs{\lambda(x,y)}$ from the
inequality~$2^{k}\leq 4(1+\log_2(B'))$, after some further
simplifications. The lower bound comes from the inequality
\begin{displaymath}
\sum_{n\geq k} 2^n\log\abs{q_{n+1}} \geq -2^k \cdot 2\log(2)
\end{displaymath}
in a similar way.
\end{proof}
\subsection{The general case}
\label{subsec:borchardt-general}
Let~$s$ be a Borchardt sequence containing finitely many bad steps. We
now construct the ``Borchardt mean following~$s$'' in a neighborhood
of the first term of~$s$ as an analytic function, provided that~$s$
contains no zero value. To make things explicit, we introduce the
following quantities:
\begin{itemize}
\item a real number~$M_0>0$ such that $\abs[\big]{s_b^{(0)}} < M_0$ for
all~$b\in \mathcal{I}_g$;
\item an integer~$n_0$ such that all steps in~$s$ of index~$n\geq n_0$
are good;
\item a real number~$m_\infty>0$ such
that~$\paren[\big]{s_b^{(n_0)}}_{b\in \mathcal{I}_g}\in \mathcal{U}_g(m_\infty,M_0)$ in the
notation of~§\ref{subsec:borchardt-good};
\item for each~$0\leq n\leq n_0-1$, a real number~$m_n>0$ such
that~$\abs[\big]{s_b^{(n)}} > m_n$ for all~$b\in \mathcal{I}_g$.
\end{itemize}
For each~$n \leq n_0-1$, we also
let~$\smash{\paren[\big]{t_b^{(n)}}_{b\in \mathcal{I}_g}}$ be a collection of
square roots of~$\smash{\paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g}}$ such that
the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$s$ is given by the recurrence
relation~\eqref{eq:borchardt-step}.
It will be useful to introduce Borchardt steps as analytic maps,
besides the case of good sign choices.
Let~$z = (z_b)_{b\in\mathcal{I}_g} \in \mathbb{C}^{2^g}$; assume that~$0 < m < M$ are
real numbers such that~$m < \abs{z_b}^2 < M$ for all~$b$. Then for
each~$b\in \mathcal{I}_g$, there exists a unique analytic square root
map~$\Sqrt_{z_b}$ on the disk~$\mathcal{D}_{m/2}(z_b^2)$ which maps~$z_b^2$
to~$z_b$. Thus, we have a well-defined analytic map
\begin{displaymath}
\Bstep_z \colon \prod_{b\in \mathcal{I}_g} \mathcal{D}_{m/2}(z_b^2) \to \mathbb{C}^{2^g}
\end{displaymath}
A quick calculation shows that~$\nind{d\Bstep_z}\leq \sqrt{(2M+m)/m}$
uniformly on its open set of definition.
\begin{lem}
\label{lem:borchardt-following-s}
Given~$s$ and the quantities listed above, let
\begin{equation}
\label{eq:borchardt-rho}
\rho = \min\left\{\frac{m_0}{2}, \frac{m_1}{2}\sqrt{\frac{m_0}{2M_0 + m_0}},\cdots,
\frac{m_{\infty}}{2} \prod_{j=0}^{n_0-1}\sqrt{\frac{m_j}{2M_0 +m_j}}
\right\}.
\end{equation}
Let~$s^{(0)} = \paren[\big]{s_b^{(0)}}_{b\in \mathcal{I}_g}$ be the first term of~$s$, and
let~$x\in \mathcal{D}_\rho(s^{(0)})$. Then there exists a unique Borchardt
sequence~$s'$ with the following properties:
\begin{enumerate}
\item the first term of~$s'$ is~$x$;
\item for all~$0\leq n\leq n_0-1$ and all~$b\in \mathcal{I}_g$, we have
$\abs[\big]{{s'}_b^{(n)} - s_b^{(n)}} < \frac{1}{2}m_{n}$; moreover
the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$s'$ is the result of a Borchardt step with
choice of square roots~$\Sqrt_{t_b^{(n)}}({s'}_b^{(n)})$ for
all~$b\in \mathcal{I}_g$;
\item for all~$n\geq n_0$, the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$s'$ is the result
of a Borchardt step from the previous term with good sign choices.
\end{enumerate}
\end{lem}
\begin{proof}
We proceed by induction, using the above estimate on derivatives of
Borchardt steps for~$n\leq n_0-1$.
\end{proof}
\begin{prop}
\label{prop:bad-borchardt-analytic}
Given~$s$ and the quantities listed above,
let~$s^{(0)} = \paren[\big]{s_b^{(0)}}_{b\in \mathcal{I}_g}$ be the first
term of~$s$, and define~$\rho>0$ as
in~\eqref{eq:borchardt-rho}. Then there exists a unique analytic
function~$\mu_s\colon \mathcal{D}_\rho(s^{(0)})\to\mathbb{C}$ with the following
property: for each~$x\in \mathcal{D}_\rho(s^{(0)})$, the value of~$\mu_s$
at~$x$ is the Borchardt mean of the sequence defined in
\cref{lem:borchardt-following-s}. We
have~$\frac12 m_\infty \leq \abs{\mu_s(x)}\leq M_0+\rho$ for
all~$x\in \mathcal{D}_\rho(s^{(0)})$.
\end{prop}
\begin{proof}
By \cref{lem:borchardt-following-s}, the function~$\mu_s$ is
obtained as the composition of a finite number of analytic Borchardt
steps, followed by an analytic Borchardt mean as defined in
\cref{prop:good-borchardt-analytic}. The upper bound
on~$\abs{\mu_s(x)}$ comes from the fact
that~$\norm{x}\leq M_0+\rho$. For the lower bound, we remark that
the~$n_0$\th term of the Borchardt sequence of
\cref{lem:borchardt-following-s} lands
in~$\mathcal{U}_g(\frac12 m_\infty, M_0+\rho)$.
\end{proof}
We extend this result to the case of extended Borchardt
means. Let~$(u,s)$ be an extended Borchardt sequence containing
finitely many bad steps. Assume that we are given:
\begin{itemize}
\item a disk~$\mathcal{D}_\rho(z_0)\subset\mathbb{C}$ such
that~$\rho < \tfrac {1}{17}\abs{z_0}$ (for instance,~$z_0$
and~$\rho$ may be dyadic)
\item An integer~$n_0$ such that all values in~$s_b^{(n_0)}$ lie
in~$\mathcal{D}_\rho(z_0)$, and after which all sign choices in~$(u,s)$ are
good;
\item A real number~$M_0>1$ such that~$\abs[\big]{s_b^{(0)}} < M_0$
and~$\abs[\big]{u_b^{(0)}} < M_0$ for all~$b\in\mathcal{I}_g$,
and~$M_0 > \abs{z_0} +\rho$;
\item a real number $0 < m_\infty < 1$ such that the~$n_0$\th term
of~$u$ lies in~$\mathcal{U}_g(m_\infty,M_0)$,
and~$m_\infty < \abs{z_0}-\rho$;
\item For each~$0\leq n\leq n_0-1$, a real number $m_n >0$ such
that~$\abs[\big]{s_b^{(n)}} > m_n$ and~$\abs[\big]{u_b^{(n)}} > m_n$
for all~$b\in\mathcal{I}_g$.
\end{itemize}
For each~$n \leq n_0-1$, we also
let~$\paren[\big]{t_b^{(n)}}_{\smash{b\in \mathcal{I}_g}}$
and~$\paren[\big]{v_b^{(n)}}_{\smash{b\in \mathcal{I}_g}}$ be collections of
square roots of~$\paren[\big]{s_b^{(n)}}_{b\in \mathcal{I}_g}$
and~$\paren[\big]{u_b^{(n)}}_{b\in \mathcal{I}_g}$ respectively such that
the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$(u,s)$ is given by the recurrence
relation~\eqref{eq:ext-borchardt-step}.
The following lemma and proposition are proved by the same methods we
used for regular Borchardt means, and we omit their proofs.
\begin{lem}
\label{lem:ext-borchardt-following-s}
Given~$(u,s)$ and the quantities listed above, let
\begin{equation}
\label{eq:ext-borchardt-rho}
\rho = \min_{0\leq n\leq n_0}\left(\frac{m_n}{2}
\prod_{j=0}^{n-1} \sqrt{\frac{m_n}{2M_0+m_n}}\right),
\end{equation}
with the convention that~$m_{n_0}= m_\infty$. Let~$(u^{(0)},s^{(0)})$ be
the first term of~$(u,s)$, and
let~$(x,y)\in \mathcal{D}_\rho\paren[\big]{(u^{(0)},s^{(0)})}$. Then there exist
extended Borchardt sequences~$(u',s')$ with the following
properties:
\begin{enumerate}
\item the first term of~$(u',s')$ is~$(x,y)$;
\item for each~$0\leq n\leq n_0-1$ and each~$b\in \mathcal{I}_g$, we
have
\begin{displaymath}
\abs[\big]{{s'}_b^{(n)} - s_b^{(n)}} < \frac12 m_n
\quad\text{and}\quad
\abs[big]{{u'}_b^{(n)} - u_b^{(n)}} < \frac12 m_n;
\end{displaymath}
moreover the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$(u',s')$ is the result of an
extended Borchardt step with choices of square
roots~$\Sqrt_{t_b^{(n)}}({s'}_b^{(n)})$
and~$\Sqrt_{v_b^{(n)}}({u'}_b^{(n)})$ for all~$b\in \mathcal{I}_g$;
\item for all~$n \geq n_0$, the~$n+1$\ensuremath{{}^{\mathrm{st}}}\ term of~$(u,s)$ is obtained
from the previous one by an extended Borchadt step with good sign
choices.
\end{enumerate}
These extended Borchardt sequences coincide up to their~$n_0$\th
terms, and their extended Borchardt means are equal.
\end{lem}
\begin{prop}
\label{prop:bad-ext-borchardt-analytic}
Given~$(u,s)$ and the quantities listed above,
let~$(u^{(0)},s^{(0)})$ be the first term of~$(u,s)$, and
define~$\rho>0$ as in~\eqref{eq:ext-borchardt-rho}. Then there
exists a unique analytic
function~$\lambda_{(u,s)}\colon \mathcal{D}_\rho(z_0)\to\mathbb{C}$ with the following
property: for each~$(x,y)\in \mathcal{D}_\rho(z)$, the value
of~$\lambda_{(u,s)}$ at~$x$ is the extended Borchardt mean of any of
the extended Borchardt sequences defined in
\cref{lem:ext-borchardt-following-s}. Moreover, we have
\begin{displaymath}
\exp\paren[\big]{-28\log^2(4M/m)} \leq
\abs{\lambda_{u,s}(x,y)} \leq \exp\paren[\big]{20\log^2(4M/m)}
\end{displaymath}
where~$m = \frac12 m_\infty$ and~$M = M_0+\rho$.
\end{prop}
\begin{rem}
In~\cite[§6.1]{dupont_FastEvaluationModular2011},
\cite[§7.4.2]{dupont_MoyenneArithmeticogeometriqueSuites2006},
\cite[§3.4]{labrande_ComputingJacobiTheta2018},
and~\cite[Prop.~3.7]{labrande_ComputingThetaFunctions2016} it is
shown that the analytic functions~$\mu,\lambda,\mu_s$
and~$\lambda_{(u,s)}$ that we just defined can be evaluated at any
given complex point in quasi-linear
time~$O\paren[\big]{\M(N)\log N}$ in the required precision,
where~$\M(N)$ denotes the cost of multiplying $N$-bit
integers. In fact, these proofs show that these analytic functions
can be evaluated in \emph{uniform} quasi-linear time. In the case
of~$\mu_s$ and~$\lambda_{(u,s)}$, the implied constant only depends
on the auxiliary data listed in this section, not on the Borchardt
sequences themselves.
\end{rem}
\section{Newton schemes for theta functions}
\label{sec:dupont}
In this section, we present the different Newton schemes used for the
computation of theta constants and theta functions in genus~$1$
and~$2$ as well as possible extensions to higher genera,
following~\cite{dupont_FastEvaluationModular2011,
dupont_MoyenneArithmeticogeometriqueSuites2006,
labrande_ComputingJacobiTheta2018,
labrande_ComputingThetaFunctions2016}. We formulate them in terms of
the analytic Borchardt functions introduced
in~§\ref{sec:borchardt}. In the three cases of theta functions in
genus~$1$ and theta constants in genus~$1$ and~$2$, we are able to
write down the inverse of the analytic function~$\mathbb{C}^r\to \mathbb{C}^r$ used in
the Newton scheme in an explicit way. This provides us with all the
necessary data to apply the results of~§\ref{sec:newton} and obtain
explicit convergence results for these Newton schemes.
\subsection{General picture}
\label{subsec:general-picture}
The Newton schemes we consider to compute theta constants at a given
point~$\tau\in \mathcal{H}_g$ use increasingly better approximations of the
point
\begin{equation}
\label{eq:theta-quotients}
\Theta(\tau) =
\left(\frac{\theta_{0,b}(0,\tau/2)}{\theta_{0,0}(0,\tau/2)}\right)_{b\in \mathcal{I}_g\setminus\set{0}}
\in \mathbb{C}^{2^g-1}.
\end{equation}
From this input, computing certain Borchardt means will provide
approximations of the quantities~$\theta^2_{0,b}(0,N\tau)$, for any
symplectic matrix~$N\in \Sp_{2g}(\mathbb{Z})$ that we might choose. Recall
that a matrix~$N\in \Sp_{2g}(\mathbb{Z})$ with~$g\times g$
blocks~$ \left(\begin{smallmatrix} a&b\\c&d
\end{smallmatrix}\right)$ acts on~$\mathcal{H}_g$
as $N\tau = (a\tau+b)(c\tau+d)^{-1}$, and on~$\mathbb{C}^g\times \mathcal{H}_g$
as~$N\cdot(z,\tau) = \paren[\big]{(c\tau+d)^{-t}z, N\tau}$,
where~${}^{-t}$ denotes inverse transposition. The next proposition,
derived from the works mentioned above, is key.
\begin{prop}
\label{prop:theta-as-borchardt}
Let~$\tau\in \mathcal{H}_g,$ let~$z\in \mathbb{C}^g$, and let
$\lambda,\mu\in \mathbb{C}^\times$. Then
\begin{enumerate}
\item The sequence
\label{it:lambda}
\begin{equation}
\label{eq:theta-as-borchardt-4}
\left(\frac{\theta_{0,b}^2(0,2^n\tau)}{\mu}\right)_{b\in \mathcal{I}_g,n\geq 0}
\end{equation}
is a Borchardt sequence with Borchardt mean~$1/\mu$, obtained
from the choice of square roots
\begin{displaymath}
\left(\frac{\theta_{0,b}(0,2^n\tau)}{\sqrt{\mu}}\right)_{b\in \mathcal{I}_g}
\end{displaymath}
for some choice of~$\sqrt{\mu}$, at each step.
\item All sequences of the form
\label{it:mu}
\begin{equation}
\label{eq:theta-as-ext-borchardt}
\left(\frac{\theta_{0,b}^2(z, 2^n\tau)}{\lambda^{2^{-n}}\mu^{1-2^{-n}}},
\frac{\theta_{0,b}^2(0,2^n\tau)}{\mu}\right)_{b\in \mathcal{I}_g,n\geq 0}
\end{equation}
with compatible choices of~$2^{-n}$-th roots
(i.e.~such that $(\lambda^{2^{-n-1}})^2 = \lambda^{2^{-n}}$
and $(\mu^{1-2^{-n-1}})^2 = \mu\cdot \mu^{1-2^{-n}}$
for all~$n$) are extended Borchardt sequences with extended
Borchardt mean~$1/\lambda$; they precisely are the sequences obtained
from choices of square roots of the form
\begin{displaymath}
\left(\frac{\theta_{0,b}(z,
2^n\tau)}{\lambda^{2^{-n-1}}\mu^{(1-2^{-n})/2}},
\frac{\theta_{0,b}(0,2^n\tau)}{\sqrt{\mu}}\right)_{b\in
\mathcal{I}_g}
\end{displaymath}
for some choice of square roots of~$\mu, \lambda^{2^{-n}}$
and~$\mu^{1-2^{-n}}$, at each step.
\end{enumerate}
\end{prop}
Consider first the case of theta constants. From the theta
quotients~\eqref{eq:theta-quotients}, one can compute all squared
theta quotients of the
form~$\theta_{a,b}^2(0,\tau)/\theta^2_{0,0}(0,\tau/2)$ using the
duplication formula. Then, applying the transformation formulas
under~$\Sp_{2g}(\mathbb{Z})$~\cite[§II.5]{mumford_TataLecturesTheta1983}
allows us to compute all theta quotients of the
form~$\theta_{0,b}^2(0, N\tau)/\theta_{0,0}^2(0, N\tau)$
for~$b\in \mathcal{I}_g$. Finally, applying
\cref{prop:theta-as-borchardt},~\eqref{it:lambda} gives us access
to~$\mu = \theta_{0,0}^2(0,N\tau)$, so that we can recover
all~$\theta_{0,b}^2(0,N\tau)$, as promised. At the end of the
algorithm, we apply the transformation formulas once more: the
relations between squared theta values~$\theta_{a,b}^2(0,\tau)$
and~$\theta_{a,b'}^2(0,N\tau)$ involve a factor~$\det(C\tau+D)$
where~$C,D$ are the lower~$g\times g$ blocks of~$N$. These
determinants are simple functions of the entries of~$\tau$, and we use
this feedback in a Newton scheme to compute a better approximation of
the initial theta quotients~\eqref{eq:theta-quotients}. When an
appropriate precision is reached, we repeat the above process one last
time to return approximations of the squared theta
values~$\theta_{a,b}^2(0,\tau)$.
In the case of theta functions, we consider the following larger set
of theta quotients:
\begin{equation}
\label{eq:theta-quotients-2}
\Theta'(\tau) =
\left(\frac{\theta_{0,b}(0,\tau/2)}{\theta_{0,0}(0,\tau/2)},
\frac{\theta_{0,b}(z,\tau/2)}{\theta_{0,b}(z,\tau/2)}
\right)_{b\in \mathcal{I}_g\setminus\set{0}}
\in \mathbb{C}^{2^{g+1}-2}.
\end{equation}
We obtain the theta
quotients~$\theta^2_{0,b}\paren[\big]{N\cdot(z,\tau)}/\theta_{0,0}^2\paren[\big]{N\cdot(z,\tau)}$
from the transformation formulas, and
\cref{prop:theta-as-borchardt},~\eqref{it:mu} allows us to compute
$\lambda = \theta_{0,0}^2\paren[\big]{N\cdot(z,\tau)}$. The feedback
is again provided by transformation formulas, and involves simple
functions (determinants and exponentials) in the entries of~$z$
and~$\tau$.
In order to run this algorithm, one has to make the correct choices of
square roots each time \cref{prop:theta-as-borchardt} is applied. At
the end of the loop, when using feedback on~$z$ and~$\tau$ to obtain
theta values at a higher precision, one \emph{assumes} that the
Jacobian matrix of the system is well-defined and invertible; in
particular, it must be a square matrix. In practice, one computes an
approximation of this Jacobian matrix using finite differences; the
resulting Newton scheme is of the type studied in~§\ref{sec:newton}.
We close this presentation with a discussion on argument
reduction. Before attempting to run these Newton schemes, one should
\emph{reduce} the input~$(z,\tau)$ using symmetries of theta
functions. Performing this reduction is necessary to even hope for
algorithms with uniform complexities in~$(z,\tau)$. If~$g\leq 2$, it
is possible to use the action of~$\Sp_{2g}(\mathbb{Z})$ on~$\tau$ to reduce it
to the \emph{Siegel fundamental domain}~$\mathcal{F}_g\subset\mathcal{H}_g$
defined by the following
conditions~\cite[§I.3]{klingen_IntroductoryLecturesSiegel1990}:
\begin{itemize}
\item $\im(\tau)$ is Minkowski-reduced;
\item $\abs{\re(\tau_{i,j})}\leq 1/2$ for all~$1\leq i,j\leq g$;
\item $\abs{\det(C\tau+D)}\geq 1$ for all~$g\times g$ matrices~$C,D$
forming the lower blocks of a symplectic matrix~$N\in \Sp_{2g}(\mathbb{Z})$;
in particular we have~$\abs{\tau_{i,i}}\geq 1$ for
all~$1\leq i\leq g$, so that~$\im(\tau_{i,i})\geq \sqrt{3}/2$.
\end{itemize}
The reduction algorithm is described
in~\cite[§6]{streng_ComputingIgusaClass2014}.
In fact, it is possible to obtain useful information on values of
theta functions, and to study the Newton schemes described above,
without assuming that all the conditions defining~$\mathcal{F}_g$ hold: see
for instance~\cite[Prop.~7.6]{streng_ComputingIgusaClass2014}. On the
other hand, we will additionally assume that the imaginary part
of~$\tau$ is bounded; this assumption is necessary to show that the
Newton schemes converge uniformly. Other inputs can be handled using
duplication formulas and the naive algorithm:
see~\cite[§6.3]{dupont_FastEvaluationModular2011}
and~\cite[§4.2]{labrande_ComputingJacobiTheta2018} in the genus~1
case. We will adapt this strategy to obtain a uniform algorithm for
genus~2 theta constants in~§\ref{sec:unif}.
The argument~$z$ can be reduced as well. By periodicity of the theta
function~$\theta(\cdot,\tau)$ with respect to the
lattice~$\mathbb{Z}^g + \tau\mathbb{Z}^g$~\cite[§II.1]{mumford_TataLecturesTheta1983},
it is always possible to assume that~$\abs{\re(z_i)}\leq \frac12$ for
each~$i$, and that
\begin{displaymath}
\left(
\begin{matrix}
\im(z_1)\\ \vdots\\ \im(z_g)
\end{matrix}
\right)
= \im(\tau)
\left(
\begin{matrix}
v_1\\ \vdots\\ v_g
\end{matrix}
\right)
\end{displaymath}
for some vector~$v\in \mathbb{R}^g$ such that~$\abs{v_i}\leq \frac12$ for
all~$i$. Since duplication formulas relate the values of theta
functions at~$z$ and~$2z$, we can in fact assume that~$z$ is very
close to zero, for instance~$\abs{z_i} < 2^{-n}$ for some fixed~$n$.
In the rest of this section, we analyze the Newton systems more
closely in the case of theta constants of genus~$1$ and~$2$, as well
as theta functions in genus~$1$, for suitably reduced inputs; our goal
is to apply \cref{thm:practical-newton}. We also discuss the situation
in higher genera.
\subsection{Genus~1 theta constants}
\label{subsec:g1-cst}
In the case of genus~$1$ theta constants, the Newton system is
univariate, and~$\tau$ is simply a complex number with positive
imaginary part. Let~$\mathcal{R}_1\subset \mathcal{H}_1$ be the compact set defined
by the following conditions:
\begin{itemize}
\item $\abs{\re(\tau)}\leq \frac12$;
\item $\abs{\tau}\geq 1$;
\item $\im(\tau)\leq 2$.
\end{itemize}
Thus,~$\mathcal{R}_1$ is a truncated, closed version of the usual fundamental
domain~$\mathcal{F}_1$. The only matrix in~$\Sp_{2}(\mathbb{Z}) = \SL_2(\mathbb{Z})$ that we
consider is
\begin{displaymath}
N = \mat{0}{-1}{1}{0},
\end{displaymath}
so that~$N\tau =
-1/\tau$. By~\cite[§I.7]{mumford_TataLecturesTheta1983}, we have
\begin{equation}
\label{eq:feedback-g1-cst}
\theta_{0,0}^2(0,N\tau) = -i\tau \theta_{0,0}^2(0,\tau).
\end{equation}
It turns out that the two Borchardt sequences used in the algorithm,
namely
\begin{equation}
\label{eq:borchardt-g1-cst}
s_1 = \paren*{\frac{\theta_{0,b}^2(0,2^n\tau)}{\theta_{0,0}^2(0,\tau)}}_{b\in \mathbb{Z}/2\mathbb{Z}, n\geq 0}
\quad \text{and}
\quad
s_2 = \paren*{\frac{\theta_{0,b}^2(0,2^n N\tau)}{\theta_{0,0}^2(0,N\tau)}}_{b\in \mathbb{Z}/2\mathbb{Z}, n\geq 0}
\end{equation}
are given by good sign choices only:
see~\cite[Thm.~2]{dupont_FastEvaluationModular2011}
and~\cite[Lem.~2.9]{cox_ArithmeticgeometricMeanGauss1984}. Our first
aim is to collect the data listed in~§\ref{subsec:borchardt-good} for
these sequences. This can be done by looking at the theta
series~\eqref{eq:theta-function} directly; see for
instance~\cite[Lem.~3.3]{labrande_ComputingJacobiTheta2018}. We
formulate the following result in the more general context of theta
functions, since it will also be useful in~§\ref{subsec:g1-fun}.
\begin{lem}
\label{lem:theta-inequality-g1}
Let~$(z,\tau)\in \mathbb{C}\times\mathcal{H}_1$ be such that
that~$\abs{\im(z)} < 2\im(\tau)$, and write $q =
\exp(-\pi\im(\tau))$. Then we have
\begin{displaymath}
\abs[\big]{\theta_{0,b}(z,\tau)-1} < 2q\cosh(2\pi\im(z))
+ \frac{2q^4\exp(4\pi\abs{\im(z)})}
{1 - q^5\exp(2\pi\abs{\im(z)})}
\end{displaymath}
for all~$b\in \mathbb{Z}/2\mathbb{Z}$, and
\begin{displaymath}
\abs[\bigg]{\frac{\theta_{1,0}(z,\tau)}{\exp(\pi i\tau/4)}
- \paren[\big]{\exp(\pi i z) + \exp(-\pi i z)}}
< \frac{2q^2 \exp(3\pi \abs{\im z})}{1 - q^4 \exp(2\pi \abs{\im z})}.
\end{displaymath}
\end{lem}
\begin{proof}
For the first inequality, write
\begin{displaymath}
\theta_{0,b}(z,\tau) = 1 + \exp(\pi i\tau + 2\pi iz) + \exp(\pi i\tau - 2\pi iz)
+ \sum_{n\in \mathbb{Z},\, \abs{n}\geq 2} \exp(\pi i n^2\tau + 2\pi i n z).
\end{displaymath}
The modulus of this last sum can be bounded above by
\begin{equation}
\label{eq:theta-tail-g1}
2\sum_{n\geq 2} \exp\paren[\big]{-\pi n^2 \im(\tau) + 2\pi n \abs{\im z}},
\end{equation}
and we conclude by comparing~\eqref{eq:theta-tail-g1} with the sum
of a geometric series matching its first two terms. The proof of the
second inequality is similar and omitted.
\end{proof}
In particular, for each~$\tau\in \mathcal{R}_1$, we
have~$\abs[\big]{\theta_{0,0}(0,\tau/2) - 1} < 0.53$, so
that~$\theta_{0,0}(0,\tau/2)$, which appears as the denominator
of~$\Theta(\tau)$ in~\eqref{eq:theta-quotients}, is indeed
nonzero. More numerical computations will appear in subsequent proofs;
we will only write down the first few digits of all real numbers
involved.
\begin{prop}
\label{lem:theta-g1-cst}
Let~$\tau\in \mathcal{R}_{1}$. Then, in the notation
of~\emph{§\ref{subsec:borchardt-good}}, the following bounds apply
to the Borchardt sequence~\eqref{eq:borchardt-g1-cst}
with~$\mu = \theta_{0,0}^2(0,\tau)$:
\begin{displaymath}
m_0 = 0.56 \quad\text{and}\quad M_0 = 1.7.
\end{displaymath}
The following bounds apply to the
sequence~\eqref{eq:borchardt-g1-cst} taken at~$N\tau$
with~$\mu = \theta_{0,0}^2(0,N\tau)$:
\begin{displaymath}
m_0 = 0.13 \quad\text{and}\quad M_0 = 1.38.
\end{displaymath}
\end{prop}
\begin{proof}
For the first sequence, we note that $\exp(-\pi \im(\tau)) < 0.066$,
and conclude using \cref{lem:theta-inequality-g1}. For the second
sequence, we invoke the transformation formula: we have
\begin{displaymath}
\frac{\theta_{0,1}(0,-1/\tau)}{\theta_{0,0}(0,-1/\tau)}
= \frac{\theta_{1,0}(0,\tau)}{\theta_{0,0}(0,\tau)}.
\end{displaymath}
By \cref{lem:theta-inequality-g1}, the angle
between~$\theta_{0,0}(0,\tau)$ and~$\theta_{1,0}(0,\tau)$ seen from
the origin is at most~$0.95 < \pi/2$; moreover we have
$0.41 < \abs{2\exp(i\pi\tau/4)} < 1.02$, from which
the claimed bounds follow.
\end{proof}
\begin{thm}
\label{thm:analytic-g1-cst}
Let~$\rho = 1.4\cdot 10^{-4}$, define~$\Theta$ as
in~\eqref{eq:theta-quotients} for~$g=1$, and let
\begin{displaymath}
\mathcal{V} = \bigcup_{\tau\in \mathcal{R}_1} \mathcal{D}_\rho\paren[\big]{\Theta(\tau)}.
\end{displaymath}
Then the operations described in~\emph{§\ref{subsec:general-picture}},
taking good choices of square roots always, combined
with~eq.~\eqref{eq:feedback-g1-cst} define an analytic
function~$F\colon \mathcal{V}\to \mathbb{C}$ such that
\begin{displaymath}
F\paren[\big]{\Theta(\tau)}
= \tau
\end{displaymath}
for each~$\tau\in \mathcal{R}_1$. We have~$\abs{F(x)}\leq 27$ for all~$x\in\mathcal{V}$.
\end{thm}
\begin{proof}
We backtrack from the result of the previous
proposition. Let~$\tau\in \mathcal{R}_1$. Then the Borchardt means we take
are well-defined as analytic functions on any open set where the
theta quotients
\begin{equation}
\label{eq:first-terms-g1-cst}
\frac{\theta_{0,1}^2(\tau)}{\theta_{0,0}^2(\tau)}
\quad\text{and}\quad
\frac{\theta_{1,0}^2(\tau)}{\theta_{0,0}^2(\tau)}
\end{equation}
are perturbed by a complex number of modulus at most~$m = 0.13$. We
construct~$\mathcal{V}$ in such a way that the maximal perturbation will not
exceed~$m/2$. The quantities~\eqref{eq:first-terms-g1-cst} are
obtained as quotients of the form:
\begin{equation}
\label{eq:intermediate-g1-cst}
\frac{\theta_{a,b}^2(\tau)/\theta_{0,0}^2(\tau/2)}{\theta_{0,0}^2(\tau)/\theta_{0,0}^2(\tau/2)}.
\end{equation}
By \cref{lem:theta-inequality-g1}, the modulus of the denominator is
at least~$0.32$, the modulus of the numerator is at
most~$5.7$. Hence, each of the individual theta
quotients~\eqref{eq:intermediate-g1-cst} may be perturbed by any
complex number of modulus at most~$6.2\cdot 10^{-4}$. In turn, these
quotients are obtained from the duplication formula applied to~$1$
and~$\Theta(\tau)$; the modulus of these two complex numbers are at
most~$2.13$, hence they may be perturbed
by~$\rho = 1.4\cdot 10^{-4}$. By construction, the value taken by
the resulting Borchardt means at any~$x\in \mathcal{V}$ has modulus at
least~$0.066$ and at most~$1.8$, hence the final bound
on~$\abs{F(x)}$.
\end{proof}
Since the inverse of~$F$ is given by theta constants, we easily see
that the Jacobian of~$F$ is invertible at all the relevant points, in
a uniform way.
\begin{prop}
\label{prop:inverse-g1-cst}
For each~$\tau\in \mathcal{R}_1$, we have
\begin{displaymath}
\nind{d\Theta(\tau)} \leq 125.
\end{displaymath}
\end{prop}
\begin{proof}
By \cref{lem:theta-inequality-g1}, the denominator of this function
has modulus at least~$0.47$, and its numerator has modulus at
most~$1.53$. The result will then follow from an upper bound on the
quantities~$\norm{d\theta_{0,b}(\tau/2)}$. We can derive such bounds
from \cref{prop:cauchy}, noting that~$\theta_{0,b}$ is an analytic
function defined on~$\mathcal{D}_{1/4}(\tau/2)$, and has modulus at
most~$2.34$ on this disk by \cref{lem:theta-inequality-g1}.
\end{proof}
Combining \cref{thm:analytic-g1-cst,prop:inverse-g1-cst} with the
results of~§\ref{sec:newton}, we obtain:
\begin{cor}
For all~$\tau\in \mathcal{R}_1$, the Newton scheme described
in~\emph{§\ref{subsec:general-picture}} to compute theta constants
at~$\tau$ will converge starting from approximations
of~$\Theta(\tau)$ to~$60$ bits of precision.
\end{cor}
\subsection{Genus~1 theta functions}
\label{subsec:g1-fun}
In the case of genus~$1$ theta functions, Newton iterations are
performed using two complex variables. As in~§\ref{subsec:g1-cst}, we
only use the symplectic matrix
\begin{displaymath}
N = \mat{0}{-1}{1}{0}.
\end{displaymath}
The Newton scheme involves the extended Borchardt mean of the
sequence~\eqref{eq:theta-as-ext-borchardt} with
$\lambda = \theta_{0,0}^2(z,\tau)$ and~$\mu = \theta_{0,0}^2(0,\tau)$,
as well as the analogous sequence taken at
$N\cdot(z,\tau) = (z/\tau, -1/\tau)$ instead of~$(z,\tau)$. Feedback
is then provided by the two following
equalities~\cite[§I.7]{mumford_TataLecturesTheta1983}:
\begin{equation}
\begin{aligned}
\label{eq:feedback-g1-fun}
\theta_{0,0}^2(0, N\tau) &= -i\tau \theta_{0,0}^2(0,\tau), \\
\theta_{0,0}^2\paren[\big]{N\cdot(z,\tau)} &= -i\tau \exp(2\pi i
z^2/\tau) \theta_{0,0}^2(z,\tau).
\end{aligned}
\end{equation}
Both extended Borchardt sequences are given by good sign choices
only~\cite[Prop.~4.1]{labrande_ComputingJacobiTheta2018}, provided
that the following reductions are met: $\tau\in \mathcal{R}_{1}$, and the
inequalities
\begin{equation}
\label{eq:g1-z-reduction}
\abs{\im (z)} \leq \frac18 \im(\tau), \quad \abs{\re(z)} \leq \frac18
\end{equation}
are satisfied. Let~$\S_1\subset \mathbb{C}\times\mathcal{H}_1$ be the compact set of
such~$(z,\tau)$. Our first goal is to collect the necessary data to apply
\cref{prop:bad-ext-borchardt-analytic} in this context.
\begin{lem}
\label{lem:theta-inequalities-g1-fun}
Let~$(z,\tau)\in \S_{1}$, and let~$a,b\in \mathbb{Z}/2\mathbb{Z}$. Then the
following inequalities hold:
\begin{align*}
\abs{\theta_{0,b}(z,\tau/2)-1} &< 0.68, \\
\abs{\theta_{a,b}(z,\tau)} &< 1.17,\quad\text{and}\\
\abs{\theta_{1,0}(z, \tau)} &> 0.37.
\end{align*}
\end{lem}
\begin{proof}
These inequalities are direct consequences of
\cref{lem:theta-inequality-g1}. Let us only detail the lower bound
on~$\abs{\theta_{1,0}(z,\tau)}$. Since~$\abs{\re z} \leq \frac18$,
we have
\begin{displaymath}
\re(\exp(\pi i z) + \exp(-\pi i z)) \geq 2\cos(\pi/8) \cosh(\pi\im(z)) > 1.84.
\end{displaymath}
Therefore,
\begin{displaymath}
\abs{\theta_{1,0}(z,\tau)} > \exp(-\pi \im(\tau)/4) (1.84 - 0.025) > 0.37.\qedhere
\end{displaymath}
\end{proof}
\begin{prop}
\label{prop:data-g1-fun}
Let~$(z,\tau)\in \S_{1}$. Then, in the notation
of~\emph{§\ref{subsec:borchardt-general}}, the following bounds
apply to the extended Borchardt
sequence~\eqref{eq:theta-as-ext-borchardt}, where
$\lambda = \theta_{0,0}^2(z,\tau)$ and
$\mu = \theta_{0,0}^2(0,\tau)$:
\begin{displaymath}
n_0=1,\quad M_0 = 1.94, \quad m_0 =0.51 \quad\text{and} \quad m_\infty = 0.72.
\end{displaymath}
The following bounds apply to the extended Borchardt
sequence~\eqref{eq:theta-as-ext-borchardt} taken
at~$N\cdot(z,\tau)$,
with~$\lambda = \theta_{0,0}^2\paren[\big]{N\cdot(z,\tau)}$
and~$\mu = \theta_{0,0}^2(0,N\tau)$:
\begin{displaymath}
n_0=1,\quad M_0 = 1.69,\quad m_0 = 0.1, \quad \text{and} \quad
m_\infty = 0.51.
\end{displaymath}
\end{prop}
\begin{proof}
These explicit values are also derived from
\cref{lem:theta-inequality-g1}. In the case of the second Borchardt
sequence, we analyze the first term using the transformation formula
for theta functions under~$\SL_2(\mathbb{Z})$. For the next terms, we use
the following inequalities:
\begin{displaymath}
\im(-1/\tau) = \frac{\im(\tau)}{\abs{\tau}^2}
\geq \frac{\sqrt{\abs{\tau}^2 - \frac12}}{\abs{\tau}^2} \geq 0.46,
\end{displaymath}
\begin{displaymath}
\im(z/\tau) = \frac{1}{\abs{\tau}^2}
\abs[\big]{\im(z)\re(\tau) - \re(z)\im(\tau)}
\leq \frac{3}{16}\im(-1/\tau),
\end{displaymath}
so that for instance
\begin{displaymath}
\abs[\big]{\theta_{0,0}^2(z/\tau, -1/\tau)} < 1.78.\qedhere
\end{displaymath}
\end{proof}
\begin{thm}
\label{thm:analytic-g1-fun}
Let~$\rho = 2.9\cdot 10^{-5}$, define~$\Theta'$ as
in~\eqref{eq:theta-quotients-2} for~$g=1$, and let
\begin{displaymath}
\mathcal{V} = \bigcup_{(z,\tau)\in \S_1} \mathcal{D}_\rho \paren[\big]{\Theta'(\tau)}.
\end{displaymath}
Then the operations described
in~\emph{§\ref{subsec:general-picture}}, taking good choices of
square roots always, combined with the
formulas~\eqref{eq:feedback-g1-fun} define an analytic
function~$F\colon \mathcal{V}\to \mathbb{C}^2$ such that
\begin{displaymath}
F(\Theta'(\tau)) = \paren*{\tau, \exp(2\pi i z^2/\tau)}
\end{displaymath}
for each~$(z,\tau)\in \S_1$. We
have~$\norm{F(x)}\leq 4.3\cdot 10^{221}$ uniformly on~$\mathcal{V}$.
\end{thm}
\begin{proof}
We apply \cref{prop:bad-ext-borchardt-analytic} with the explicit
values provided above; to find an acceptable~$\rho$, we follow a
backtracking strategy as in the proof of \cref{thm:analytic-g1-cst},
using the first two inequalities of
\cref{lem:theta-inequalities-g1-fun}. The upper bound on~$\norm{F}$
comes from \cref{prop:bad-ext-borchardt-analytic}.
\end{proof}
The upper bound on~$\norm{F}$ could certainly be improved in this
situation, but the above value will be sufficient for our purposes.
This function~$F$ admits an analytic reciprocal. Here it is essential
that the theta constants~$\theta_{0,b}(z,\tau)$ for~$b\in \mathbb{Z}/2\mathbb{Z}$ are
invariant under~$z\mapsto -z$; this implies that they can be rewritten
as analytic functions of~$z^2$.
\begin{prop}
\label{prop:theta-from-square}
Let~$b\in \mathbb{Z}/2\mathbb{Z}$. Then there exists a unique analytic function
\begin{displaymath}
\zeta_{0,b}\colon \mathbb{C}\times\mathcal{H}_1\to \mathbb{C}
\end{displaymath}
such that for all~$(z,\tau) \in\mathbb{C}\times\mathcal{H}_1$, we
have~$\theta_{0,b}(z,\tau) = \zeta_{0,b}(z^2,\tau)$.
\end{prop}
\begin{proof}
Consider the following reorganization of the theta series:
\begin{displaymath}
\theta_{0,b}(z,\tau) =
1 + \sum_{n\geq 1} (-1)^{nb}\exp(\pi i n^2\tau) (\exp(2\pi i n z) + \exp(-2\pi i n z)).
\end{displaymath}
Each factor~$\exp(2\pi i nz) + \exp(-2\pi i nz)$, as an even entire
function, has only powers of~$z^2$ in its Taylor series. We obtain a
candidate~$\zeta_{0,b}$ as a formal power series, easily seen to
converge uniformly on compact sets of~$\mathbb{C}\times\mathcal{H}_1$.
\end{proof}
Note that for every~$(z,\tau)\in \S_1$, we
have~$\re(z^2/\tau) < 1/2$; therefore~$\exp(2\pi i z^2/\tau)$ lands in
the domain of definition of the principal branch of the complex
logarithm, denoted by~$\mathcal{U}$. Consider the two following maps:
\begin{displaymath}
\begin{matrix}
\mathcal{H}_1 \times \mathcal{U} &\to &\mathbb{C}\times \mathcal{H}_1&\to &\mathbb{C}\times\mathbb{C}\\[6pt]
(\tau, x) &\mapsto &\paren[\big]{\frac{1}{2\pi i}\log(x), \tau)} \\
& & (y, \tau) &\mapsto & \displaystyle\paren*{\frac{\theta_{0,1}(0,\tau/2)}
{\theta_{0,0}(0,\tau/2)},
\frac{\zeta_{0,1}(y,\tau/2)}{\zeta_{0,0}(y,\tau/2)}}.
\end{matrix}
\end{displaymath}
Call~$G$ their composition; it is well-defined on an open neighborhood
of the image of~$\S_1$
by~$(z,\tau)\mapsto \paren[\big]{\tau,\exp(2\pi i z^2/\tau)}$, and is
the reciprocal of~$F$.
\begin{prop}
\label{prop:inverse-g1-fun}
For each~$(z,\tau) \in \S_1$, we have
\begin{displaymath}
\nind[\big]{dG\paren[\big]{\tau, \exp(2\pi i z^2/\tau)}} \leq 8.6\cdot 10^4.
\end{displaymath}
\end{prop}
\begin{proof}
Let~$x = \exp(2\pi i z^2/\tau)$, and~$y = z^2$. We have
\begin{displaymath}
\abs[\big]{\im(z^2/\tau)} = \frac{1}{\abs{\tau}^2}
\cdot\frac{1}{64}\paren[\big]{\im(\tau) + \im(\tau)^3 + \im(\tau)}
\leq \frac{1}{16},
\end{displaymath}
showing that~$\abs{x}$ is close to~$1$. It only remains to obtain
explicit upper bounds on the derivative
of~$\zeta_{0,1}/\zeta_{0,0}$. To obtain such bounds, we consider the
polydisk of radius~$1/16$ centered in~$(y,\tau/2)$;
by~\cref{lem:theta-inequality-g1}, we
have~$\abs{\zeta_{0,b}} < 5.8$ on this disk for each~$b\in \mathbb{Z}/2\mathbb{Z}$,
so that~$\nind{d\zeta_{0,b}(y,\tau/2)} < 277$ by \cref{prop:cauchy}. We
can conclude using the lower bound on~$\abs{\zeta_{0,0}(y,\tau/2)}$
provided by \cref{lem:theta-inequalities-g1-fun}.
\end{proof}
\begin{cor}
For all~$(z,\tau)\in \S_1$, the Newton scheme described
in~\emph{§\ref{subsec:general-picture}} to compute theta functions
at~$(z,\tau)$ will converge starting from approximations
of~$G(\tau)$ to~$1600$ bits of precision.
\end{cor}
\subsection{Genus~2 theta constants}
\label{subsec:g2}
In the case of genus~$2$ theta constants, Newton iterations are
performed on three variables, and feedback is provided by the action
of three symplectic matrices.
For general~$g$, certain interesting symplectic matrices can be
written down explicitly. Denote the elementary~$g\times g$ matrices
by~$E_{i,j}$ for~$1\leq i,j\leq g$, and let~$I$ be the identity
matrix. Let~$M_i$ and~$N_{i,j}$ ($i\neq j$) be the following
symplectic matrices, written in~$g\times g$ blocks:
\begin{displaymath}
M_i = \mat{-I}{-E_i}{E_i}{-I + E_i}, \quad
N_{i,j} = \mat{-I}{-E_{i,j}-E_{j,i}}{E_{i,j}+E_{j,i}}{-I + E_{i,i} + E_{j,j}}
\end{displaymath}
The matrices~$M_i$ and~$N_{i,j}$ are precisely engineered so that
determinants of the form~$\det(C\tau+D)$ give us direct access to the
entries of~$\tau$. In the case of genus~$2$ theta constants,
considering these three matrices~$M_1, M_2$ and~$N_{1,2}$ is enough to
run the Newton scheme, using the following
formulas~\cite[Prop.~8]{enge_ComputingClassPolynomials2014}:
\begin{equation}
\label{eq:feedback-g2}
\begin{aligned}
\theta_{00,00}^2(0,M_1\tau) &= -\tau_{1,1}\theta_{01,00}^2(0,\tau), \\
\theta_{00,00}^2(0,M_2\tau) &= -\tau_{2,2} \theta_{10,00}^2(0,\tau), \quad\text{and}\\
\theta_{00,00}^2(0,N_{1,2}\tau) &= (\tau_{1,2}^2-\tau_{1,1}\tau_{2,2}) \theta_{00,00}^2(0,\tau).
\end{aligned}
\end{equation}
In~\cite{kieffer_SignChoicesAGMtoappear}, it is shown that all four
Borchardt sequences of the form~\eqref{eq:theta-as-borchardt-4} taken
at~$\tau$,~$M_1\tau$,~$M_2\tau$ and~$N_{1,2}\tau$ are given by good
choices of square roots only, provided that~$\tau$ satisfies the
following conditions:
\begin{itemize}
\item $\abs{\re(\tau_{i,j})} \leq \frac12$ for all~$1\leq i,j\leq 2$,
\item $2\abs{\im(\tau_{1,2})} \leq \im(\tau_{1,1}) \leq \im(\tau_{2,2})$,
\item $\abs{\tau_{j,j}}\geq 1$ for~$j = 1,2$.
\end{itemize}
These inequalities hold in particular whenever~$\tau$ lies in the
Siegel fundamental domain~$\mathcal{F}_2$. Let~$\mathcal{R}_2$ be the compact set
of such matrices~$\tau$, with the additional assumption
that~$\im(\tau_{1,1})\leq 2$ and~$\im(\tau_{2,2})\leq 8$. This choice
of upper bounds will be explained by the construction of a uniform
algorithm in~§\ref{sec:unif}.
As in the previous sections, we will collect the explicit data we need
to apply \cref{prop:good-borchardt-analytic} using inequalities
satisfied by genus~$2$ theta constants. Many such inequalities already
appear in~\cite[§9]{klingen_IntroductoryLecturesSiegel1990},
\cite[§6.2.1]{dupont_MoyenneArithmeticogeometriqueSuites2006},
\cite[§7.2]{streng_ComputingIgusaClass2014},
and~\cite{kieffer_SignChoicesAGMtoappear}; we will use one more.
\begin{lem}
\label{lem:theta-inequalities-g2}
For each~$\tau\in \mathcal{R}_2$, we have
$0.44 < \abs{\theta_{0,0}(0,\tau/2)} < 2.66$.
\end{lem}
\begin{proof} Let
\begin{displaymath}
\xi_0(\tau/2) = 1 + 2\exp\paren[\big]{i\pi \im(\tau_{1,1})/2}
+ 2\exp\paren[\big]{i\pi \im(\tau_{2,2})/2}.
\end{displaymath}
Since~$\tau\in\mathcal{R}_2$, the complex number~$\xi_0(\tau/2)$ has
modulus at least~$1$ and at
most~$2.1$. By~\cite[Lem.~4.4]{kieffer_SignChoicesAGMtoappear}, we
have $\abs{\theta_{0,0}(0,\tau/2) - \xi_0(\tau/2)} < 0.56$.
\end{proof}
\begin{prop}
\label{prop:data-g2}
Let~$\tau\in \mathcal{R}_2$. Then, in the notation
of~\emph{§\ref{subsec:borchardt-good}}, the following bounds apply.
\begin{enumerate}
\item In the case of the Borchardt
sequence~\eqref{eq:theta-as-borchardt-4}
with~$\lambda = \theta_{0,0}^2(0,\tau)$, we can take $m_0 = 0.069$
and~$M_0 = 13$.
\item In the case of the Borchardt
sequence~\eqref{eq:theta-as-borchardt-4} taken at~$M_j\tau$
with~$\lambda = \theta_{0,0}^2(0,M_j\tau)$, for
each~$j\in\{1,2\}$, we can take~$m_0 = 9.7\cdot 10^{-7}$
and~$M_0= 13$.
\item In the case of the Borchardt
sequence~\eqref{eq:theta-as-borchardt-4} taken at~$N_{1,2}\tau$
with~$\lambda = \theta_{0,0}^2(0,N_{1,2}\tau)$, we can take
$m_0 = 2.2\cdot 10^{-9}$ and~$M_0 = 13$.
\end{enumerate}
\end{prop}
\begin{proof}
We only have to analyze the first term of each of these Borchardt
sequences. These explicit constants are then derived from the proof
in~\cite{kieffer_SignChoicesAGMtoappear} that these complex numbers
are in good position.
\end{proof}
\begin{thm}
\label{thm:analytic-g2}
Let $\rho = 1.9\cdot 10^{-23}$, define~$\Theta$ as
in~\eqref{eq:theta-quotients} for~$g=2$, and let
\begin{displaymath}
\mathcal{V} = \bigcup_{\tau\in \mathcal{R}_2} D_\rho\paren[\big]{\Theta(\tau)}
\subset \mathbb{C}^3.
\end{displaymath}
Then the operations described in~\emph{§\ref{subsec:general-picture}},
taking good choices of square roots always, define an analytic
function~$F\colon \mathcal{V}\to \mathbb{C}^2$ such that
\begin{displaymath}
F\paren[\big]{\Theta(\tau)}
= (\tau_{1,1},\tau_{2,2}, \tau_{1,2}^2- \tau_{1,1}\tau_{2,2})
\end{displaymath}
for each~$\tau\in \mathcal{R}_{2}$. We have~$\norm{F}\leq 4.5\cdot 10^4$
uniformly on~$\mathcal{V}$.
\end{thm}
\begin{proof}
The first terms of each of the Borchardt sequences analyzed in
\cref{prop:data-g2} is obtained as quotients of the quantities
\begin{displaymath}
\frac{\theta_{a,b}^2(0,\tau)}{\theta_{0,0}^2(0,\tau/2)},
\end{displaymath}
for all even theta characteristics~$(a,b)$ (i.e.~such that
$a^tb = 0\mod 2$), except~$(11,11)$. The numerator and denominator
of these quantities is bounded, both above and away from zero, by
\cref{lem:theta-inequalities-g2}
and~\cite[Cor.~7.7]{streng_ComputingIgusaClass2014}. Using these
inequalities combined
with~\cref{prop:data-g2,prop:good-borchardt-analytic} is sufficient
to obtain an explicit value of~$\rho$.
\end{proof}
To conclude, we show that the Jacobian of~$F$ is uniformly invertible
by writing its inverse in terms of theta functions. Since~$F$ only
recovers the square of~$\tau_{1,2}$, we use the fact that each of the
fundamental theta constants~$\theta_{0,b}(0,\cdot)$ for~$b\in\mathcal{I}_2$ is
invariant under change of sign of~$\tau_{1,2}$.
\begin{lem}
\label{lem:theta-from-square-g2}
Let~$\mathcal{V}\subset\mathbb{C}^3$ be the image of~$\mathcal{H}_2$
under~$\tau\mapsto (\tau_{1,1},\tau_{2,2},-\det \tau)$. Then, for
each~$b\in \mathcal{I}_2$, there exists a unique analytic
function~$\xi_{0,b}\colon \mathcal{V}\to \mathbb{C}$ such
that
\begin{displaymath}
\theta_{0,b}(0,\tau/2) = \xi_{0,b}(
\tau_{1,1},\tau_{2,2},-\det \tau)
\end{displaymath}
for all~$\tau\in \mathcal{H}_2$.
\end{lem}
\begin{proof}
In the theta series~\eqref{eq:theta-function} for~$z=0$, the only
terms involving~$\tau_{1,2}$ are those associated
with~$(n_1,n_2)\in\mathbb{Z}^2$ both nonzero. Write~$b = (b_1,b_2)$. Then,
the terms associated with~$(n_1,n_2)$ and~$(n_1,-n_2)$ are
\begin{displaymath}
\exp\paren[\big]{i\pi(\tau_{1,1}n_1^2 + \tau_{2,2}n_2^2 \pm 2\tau_{1,2}n_1n_2)}
(-1)^{n_1b_1 + n_2 b_2},
\end{displaymath}
so their sum can be written as a power series in~$\tau_{1,2}^2$
only.
\end{proof}
Let~$G$ be the following analytic function:
\begin{displaymath}
G(x,y,z) = \paren*{\frac{\xi_{0,b}(x,y,z)}{\xi_{0,0}(x,y,z)}}_{b\in \mathcal{I}_2\setminus\set{0}}.
\end{displaymath}
It is well-defined on a neighborhood of the image of~$\mathcal{R}_2$
by~$\tau\mapsto (\tau_{1,1},\tau_{2,2}, -\det\tau)$, and is the
reciprocal of~$F$.
\begin{prop}
We have~$\nind{dG
(\tau_{1,1},\tau_{2,2},-\det\tau)}\leq 1.3\cdot 10^4$ for
all~$\tau\in \mathcal{R}_2$.
\end{prop}
\begin{proof}
Fix~$\rho = 1/4$, and let us compute an upper bound
on~$\abs{\xi_{0,b}(x,y,z)}$ for each point
$(x,y,z) \in
\mathcal{D}_\rho\paren[\big]{(\tau_{1,1},\tau_{2,2},-\det\tau)}$.
Then~$x,y,z$ are of the form~$(\tau'_{1,1},\tau'_{2,2}, -\det\tau')$
for some~$\tau'\in \mathcal{H}_2$; the smallest eigenvalue of~$\im(\tau')$
is bounded from below by
\begin{displaymath}
\frac{\det(\tau')}{\tr(\tau')} \geq 0.12.
\end{displaymath}
By the proof of~\cite[Lem.~4.7]{kieffer_SignChoicesAGMtoappear}, the
function~$\abs{\xi_{0,b}}$ is uniformly bounded above by~$9.28$ on
the disk we consider. By \cref{prop:cauchy}, we have
\begin{displaymath}
\nind{d\xi_{0,b}
(\tau_{1,1},\tau_{2,2},-\det\tau)} \leq 149
\end{displaymath}
for each~$b\in \mathcal{I}_2$. The upper bound on~$\nind{dG}$ then
follows from \cref{lem:theta-inequalities-g2}.
\end{proof}
\begin{cor}
For all~$\tau\in \mathcal{R}_2$, the Newton scheme described
in~\emph{§\ref{subsec:general-picture}} to compute theta constants
at~$\tau$ will converge starting from approximations
of~$\Theta(\tau)$ to~$300$ bits of precision.
\end{cor}
\subsection{Higher genera}
In higher genera, including the case of genus~$2$ theta functions, we
are no longer able to show that the linear systems appearing in the
Newton schemes are invertible, nor a fortiori are we able to give an
explicit upper bound on the norm of their inverse Jacobians. Let us
shortly explain what the obstacle is.
In order to build a Newton scheme, the linearized system must be
square; however, as~$g$ grows, the number~$r$ of theta quotients
(either~$2^g-1$ in the case of theta constants, or~$2^{g+1}-2$ in the
case of theta functions) becomes greater than the dimension
of~$\mathcal{H}_g$ or~$\mathbb{C}^g\times\mathcal{H}_g$ respectively. Two ways around
this issue are suggested
in~\cite[§3.5]{labrande_ComputingThetaFunctions2016}:
\begin{enumerate}
\item One could keep all theta quotients as variables, and simply
consider more symplectic matrices~$N$ to provide suitable feedback; or
\item One could perform Newton iterations not on the whole of~$\mathbb{C}^r$,
but rather on the algebraic subvariety of~$\mathbb{C}^r$ obtained as the
image of~$\mathcal{H}_g$ or~$\mathbb{C}^g\times\mathcal{H}_g$ by the fundamental theta
quotients~\eqref{eq:theta-quotients}
or~\eqref{eq:theta-quotients-2}.
\end{enumerate}
A fundamental obstacle to the second idea seems to be that the
algebraic subvariety of~$\mathbb{C}^r$ on which the theta quotients lie is not
smooth everywhere in general: consider for instance the Kummer
equation~\cite[§3.1]{gaudry_FastGenusArithmetic2007} in the case of
genus~$2$ theta constants. On the other hand, it seems very likely
that the first possibility can give rise to suitably invertible
systems, since much freedom is allowed in the choice of symplectic
matrices. However, the inverse of~$F$ will no longer be described
completely by theta functions, so the method we employed above to
prove the invertibility of the linearized systems no longer applies.
Despite the current lack of a uniform algorithm, the following
approach is available to certify the result of Newton's method to
evaluate theta constants (resp.~functions) at a given
$\tau\in \mathcal{H}_g$ (resp.~$(z,\tau)\in \mathbb{C}^g\times\mathcal{H}_g$). A finite
amount of precomputation, along with the results
of~\cref{sec:borchardt}, will allow us to compute real
numbers~$\rho>0$ and~$M>0$ such that the function~$F$ appearing in
Newton's method is analytic with~$\abs{F}\leq M$ on a polydisk of
radius~$\rho$ around the desired theta values. This gives upper bounds
on the norms of derivatives of~$F$ on a slightly smaller polydisk; in
particular, we can compute a certified approximation of~$dF^{-1}$
using finite differences, and check that it is indeed invertible. This
provides all the necessary data to run certified Newton
iterations.
\section{A uniform algorithm for genus~2 theta constants}
\label{sec:unif}
We have shown in~§\ref{subsec:g2} that genus~$2$ theta constants can
be evaluated on the compact subset~$\mathcal{R}_2$ of~$\mathcal{H}_2$ in uniform
quasi-linear time in the required precision, in a certified way. Using
this algorithm as a black box, we now design an algorithm to evaluate
genus~$2$ theta constants on the whole Siegel fundamental
domain~$\mathcal{F}_2$ in uniform quasi-linear time, generalizing the
strategy presented in~\cite[Thm.~5]{dupont_FastEvaluationModular2011},
\cite[§4.2]{labrande_ComputingJacobiTheta2018} in the genus~$1$ case:
we use duplication formulas to replace the input by another period
matrix which either lies in~$\mathcal{R}_2$, or is sufficiently close to the
cusp, in which case the naive algorithm can be applied. We will use
the following transformations: for every $\tau \in\mathcal{H}_2$, write
\begin{displaymath}
D_1(\tau) = \frac\tau2
\quad\text{and}\quad
D_2(\tau) = \mat{2 \tau_{1,1}}{\tau_{1,2}}{\tau_{1,2}}{\frac12 \tau_{2,2}}.
\end{displaymath}
Recall that every~$\tau\in\mathcal{F}_2$ satisfies the following inequalities:
\begin{equation}
\label{eq:F2}
\begin{cases}
\abs{\re(\tau_{i,j})} \leq \frac12 \quad \text{for each } 1\leq i, j\leq 2,\\[2pt]
2\abs{\im(\tau_{1,2})} \leq \im(\tau_{1,1})\leq \im(\tau_{2,2}),\\[2pt]
\abs{\tau_{i,i}}\geq 1 \quad \text{for each } i\in\{1,2\}. \phantom{\frac12}
\end{cases}
\end{equation}
We also define
\begin{displaymath}
\mathcal{J} = \paren[\big]{(00,00),(00,01),(10,00),(10,01)} \in (\mathcal{I}_2\times\mathcal{I}_2)^4,
\end{displaymath}
which is the tuple of theta characteristics corresponding to the
indices~$0,2,4,6$ in Dupont's
indexation~\cite[§6.2]{dupont_MoyenneArithmeticogeometriqueSuites2006}. For
each~$\tau\in \mathcal{H}_2$, the duplication formula allows us to compute
all squares of theta constants at~$\tau$ given the theta
values~$\theta_{0,b}\paren[\big]{D_1(\tau)}$ for all~$b\in \mathcal{I}_2$. By
applying the theta transformation formula to the symplectic matrix
\begin{displaymath}
\left(
\begin{matrix}
0&0&1&0\\0&1&0&0\\-1&0&0&0\\0&0&0&1
\end{matrix}
\right),
\end{displaymath}
we also see that all squares of theta constants at~$\tau$ can be
computed from the theta values~$\theta_{a,b}\paren[\big]{D_2(\tau)}$
for~$(a,b)\in\mathcal{J}$. It turns out that these complex numbers are
in good position; hence, they are easily determined from their squares
up to a harmless global change of sign.
\begin{lem}
\label{lem:dupl}
Let $\tau\in\mathcal{H}_2$ be a matrix satisfying~\eqref{eq:F2}.
\begin{enumerate}
\item If~$D_1(\tau)$ satisfies~\eqref{eq:F2}, then the complex
numbers~$\paren[\big]{\theta_{0,b}(D_1(\tau))}_{b\in \mathcal{I}_2}$ are in good position.
\item If $D_2(\tau)$ satisfies~\eqref{eq:F2}, except that the real
part of $D_2(\tau)_{1,1}$ is allowed to be smaller than~$1$
instead of~$\frac12$, then the complex numbers
$\paren[\big]{\theta_{a,b}(D_2(\tau))}_{(a,b)\in\mathcal{J}}$ are in good position.
\end{enumerate}
\end{lem}
\begin{proof}
See~\cite[Prop.~7.7]{streng_ComplexMultiplicationAbelian2010}
and~\cite[Lem.~5.2]{kieffer_SignChoicesAGMtoappear}.
\end{proof}
\begin{thm}
\label{thm:theta-F2}
There exists an algorithm which, given $\tau\in\mathcal{H}_2$
satisfying~\eqref{eq:F2} and given $N\geq 1$, computes the squares
of theta constants at~$\tau$ to precision~$N$ within
$O(\M(N)\log N)$ binary operations, uniformly in~$\tau$.
\end{thm}
\begin{proof}
Fix an arbitrary absolute constant~$C>0$ (for instance~$10$); in
practice, this constant should be adjusted to minimize the
algorithm's running time. First, let~$k_2$ be the smallest integer
such that
\begin{displaymath}
2^{k_2} \im(\tau_{1,1}) \geq \min\{CN,\, 2^{-k_2-2} \im(\tau_{2,2})\},
\end{displaymath}
and let $\tau'$ be the matrix obtained after applying~$k_2$
times~$D_2$ to~$\tau$ and reducing the real part at each step. In
order to compute theta constants at~$\tau$ to precision~$N$, we can
compute theta constants at~$\tau'$ to some precision $N'\geq N$,
then apply~$k_2$ times the duplication formula; all sign choices are
good by \cref{lem:dupl}. We have $k_2 = O(\log N)$, and the total
precision loss taken in extracting square roots is~$O(N)$
by~\cite[Prop.~7.7]{streng_ComputingIgusaClass2014}. Therefore, the
total precision loss is~$O(N)$ bits, and we can choose~$N' = C' N$
where $C'$ is an absolute constant.
Two cases arise now. If $\im(\tau'_{1,1}) \geq CN$, then we also
have $\im(\tau'_{2,2})\geq CN$; therefore we can compute theta
constants at~$\tau'$ to precision~$N'$ using~$O\paren[\big]{\M(N)}$
operations with the naive algorithm. Otherwise, we have
\begin{displaymath}
\im(\tau'_{1,1}) \leq \im(\tau'_{2,2}) \leq 4 \im(\tau'_{1,1}) \leq 4CN.
\end{displaymath}
Therefore we can find an integer $k_1 = O(\log N)$ such that
$\tau'' = D_1^{k_1}(\tau')$ belongs to~$\mathcal{R}_2$, by definition of
this compact set. We will compute theta constants at~$\tau''$ to
some precision~$N''\geq N'$ using the Newton scheme described
in~§\ref{sec:dupont}, then use the duplication formula~$k_1$
times. Since~$O(1)$ bits of precision are lost each time we apply
the duplication formula
by~\cite[Prop.~7.7]{streng_ComputingIgusaClass2014}, we can also
take~$N'' = C''N$ where~$C''$ is an absolute constant. Therefore,
the whole algorithm can be executed in~$O(\M(N)\log N)$ binary
operations.
\end{proof}
\begin{rem}
In order to implement this algorithm in a certified way, one could
use~\cite[Prop.~7.7]{streng_ComputingIgusaClass2014} more explicitly
to track down an acceptable value of~$C''$. Another possibility is to
start with~$C'' = 1.1$, say, and attempt to run this algorithm using
interval arithmetic to obtain real-time upper bounds on the
precision losses incurred. If the final precision we obtain is not
satisfactory, we may simply double~$C''$ and restart. The resulting
algorithm still has a uniform quasi-linear cost.
\end{rem}
\bibliographystyle{abbrv}
| d4bfc583e75cbb59bb35ca27c0620388f1142d7f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $G=(V,E)$ be a graph with maximum degree $\Delta$ and minimum degree $\delta>0$. Consider an edge $C$-colouring $c$ of $G$, i.e. an assignment of elements from a set $C$ of potential colours to the edges of $G$, which does not need to be proper. Let $E_G(v)$ denote the set of edges incident to a vertex $v$ in $G$ and set $E_G[uv]:=E_G(u)\cup E_G(v)$ for every edge $uv\in E$. Let further $P_c(uv)$, $P_c(v)$ be the \emph{palettes} of colours of $uv$ and $v$, i.e. the multisets of colours in $E(uv)$ and $E[v]$, respectively. We say that $uv$, resp. $v$, is \emph{satisfied} by $c$ if there is a unique colour in its palette,
i.e. a colour associated to exactly one edge in $E[uv]$, rsp. $E(v)$. For convenience we shall be writing up each multiset in the following form: $c_1^{m_1}\ldots c_p^{m_p}$ where $c_i\in C$ and $m_i$ indicates the number of repetitions of $c_i$ in the multiset, and we shall be abbreviating $c_i^1$ as $c_i$. An edge $e$ is thus in particular satisfied by $c$ if $P_c(e)$ is of the form $c_1c_2^{m_2}\ldots c_p^{m_p}$. The colouring $c$ is said to be \emph{conflict-free} if all edges of $G$ are satisfied, i.e. contain unique colours in their closed neighbourhoods. The least number of colours admitting such a colouring is called the \emph{conflict-free chromatic index} of $G$ and denoted $\chi'_{CF}(G)$.
By definition, $\chi'_{CF}(G)\leq \chi'(G)$, but $\chi'_{CF}(G)$ is usually much smaller than $\chi'(G)$.
Recently D\k{e}bski and Przyby{\l}o~\cite{DebskiPrzybylo} exhibited that $\chi'_{CF}\geq (1-o(1))\log_2\Delta$ for the family of complete graphs and used the probabilistic method to prove that there is a constant $C_2$ such that
$\chi'_{CF}(G)\leq 9\log_{\frac{1}{1-e^{-4}}}\Delta+C_2 = (9/\log_2\frac{1}{1-e^{-4}})\log_2\Delta+C_2$,
where $9/\log_2\frac{1}{1-e^{-4}}\approx 337.5$. The main goal of the present note is to provide a simple direct argument implying that $\chi'_{CF}(G)\leq 3\log_2\Delta+1$. In fact we shall prove something even stronger,
relating the graph invariant under investigation with the chromatic number $\chi(G)$ rather than $\Delta(G)$,
namely that $\chi'_{CF}(G)\leq 3\log_2\chi(G)+1$, cf. Corollaries~\ref{CorollaryGeneralChi} and~\ref{CorollaryGeneralDelta}. For this purpose we study bipartite graphs first and a few related issues.
The problem above grew out of its earlier variant concerning vertex colourings. The naturally defined graph parameter in vertex setting is called the \emph{conflict-free chromatic number} and denoted $\chi_{CF}(G)$. In fact investigating $\chi'_{CF}(G)$ goes down to study of $\chi_{CF}(G)$ for the family of line graphs, cf.~\cite{DebskiPrzybylo}. Originally the concept of $\chi_{CF}(G)$ was motivated by channel assignment in wireless networks, where colours represent available frequencies, which potentially interfere, and thus one should always have in range a transmitter with a unique frequency associated, see e.g.~\cite{EvenEtAl,SmorodinskyPhd,SmorodinskyApplications}.
In 2009 Pach and Tardos~\cite{PachTardos} showed that $\chi_{CF}(G) = O(\ln^{2+\epsilon}\Delta)$.
This was later improved to $\chi_{CF}(G) = O(\ln^2\Delta)$ by Bhyravarapu, Kalyanasundaram and Mathew~\cite{Hindusi},
while Glebov, Szab\'o and Tardos~\cite{GlebovEtAl} constructed a family of graphs with $\chi_{CF}(G) = \Omega(\ln^2\Delta)$, legitimating the optimal order of the bound from~\cite{Hindusi}. See also~\cite{Hindusi2,KellerEtAl,KostochkaEtAl} for other related results.
\section{Results}
A palette including just one colour shall be called \emph{monochromatic}.
Moreover, ginen a graph $G=(V,E)$ and an edge $uv$ (which does not have to belong to $E$), we set $G+uv:=(V\cup\{u,v\},E\cup\{uv\})$.
In~\cite{DebskiPrzybylo} it was in particular proved that $\chi'_{CF}(T)\leq 3$ for every tree $T$.
Here we rephrase this result in order to emphasize some detailed features of the colouring implying this upper bound.
By \emph{levels} of a tree $T$ rooted at $r$ we mean $L_0,L_1,L_2,\ldots$ such that $L_i=\{v\in V(T):{\rm dist}(v,r)=i\}$, where ${\rm dist}(v,r)$ denotes the distance of $v$ and $r$ in $T$. If $d_T(r)=1$, i.e. $r$ is a leaf, we refer to its only incident edge as the \emph{rooted edge}.
A tree of order $1$ is called \emph{trivial}.
\begin{lemma}\label{TreeColouring}
Let $L_0,\ldots,L_q$ be levels of a nontrivial tree $T$ rooted at its leaf $r$.
Then there exists an edge $\mathbb{Z}_3$-colouring $c$ of $T$ such that for every $i\geq 1$ and $u\in L_i$:
\begin{enumerate}
\item[($1^\circ$)] $P_c(u)=(i-1)i^{k_u}$, where $k_u\geq 0$ (with $i,i-1$ taken modulo $3$);
\item[($2^\circ$)] if moreover $uv\in E(T)$ where $v\in L_{i+1}$, then $P_c(uv)=(i-1)i^{k_e}(i+1)^{k'_e}$ where $k_e,k'_e\geq 0$.
\end{enumerate}
Consequently, $\chi'_{CF}(T)\leq 3$.
\end{lemma}
\begin{pf}
Set $c(e)=i$ (mod $3$) for every edge $e$ between levels $L_i$ and $L_{i+1}$.
It is straightforward to verify ($1^\circ$) for such $c$, while ($2^\circ$) is directly implied by ($1^\circ$).
As $P_c(e)=01^{k_e}$ for the rooted edge $e$, $\chi'_{CF}(T)\leq 3$ follows by ($2^\circ$).
\qed
\end{pf}
We note that it is easy to construct trees for which $\chi'_{CF}(T)= 3$, consider e.g. the complete binary tree of height $3$.
In order to prove our main result we shall in a way need to extend the upper bound from Lemma~\ref{TreeColouring} towards bipartite graphs in Theorem~\ref{ChiSCF-bipartite}.
Let us first briefly discuss the base case of complete bipartite graphs, which settle a ``lower bound'' for such a problem.
\begin{observation}\label{ObservComplBip}
If $G=K_{n,m}$ with $n,m\geq 1$, then
$\chi'_{CF}(G)\leq 2$ if $\min\{n,m\}\leq 2$ and $\chi'_{CF}(G)= 3$ if $\min\{n,m\}\geq 3$.
We also have $\chi'_{CF}(C_k)= 2$ for every $k\geq 3$.
\end{observation}
\begin{pf}
Let us assume that $G=K_{n,m}$ with $n,m\geq 3$, as the other case as well as the observation for cycles are straightforward.
A colouring $c'$ implying $\chi'_{CF}(G)\leq 3$ may be constructed by selecting any vertex $v$ in $G$ and painting its incident edges with $0$ and $1$ so that $P_{c'}(v)=01^{d(v)-1}$, and assigning $2$ to all remaining edges.
In order to finally show that $\chi'_{CF}(G)> 2$, suppose on the contrary that $c$ is an edge $\mathbb{Z}_2$-colouring of $K_{n,m}$ satisfying all its edges.
It is easy to see that each colour must be used on at least two edges.
Consider any edge. Since it is satisfied, at least one of its end-vertices, say $v$ must have a unique colour, say $0$ in its palette.
Now consider any other edge coloured $0$ in $G$ and denote by $u$ its end belonging to a different set of the bipartition of $K_{n,m}$ than $v$. Then however $uv$ is incident with at least two edges coloured $0$ and at least two edges coloured $1$, a contradiction.
\qed
\end{pf}
From now on it shall be more convenient for us to focus on a slightly modified graph invariant.
Suppose $c$ is an edge colouring of some subgraph $H$ of a given graph $G=(V,E)$.
Similarly as before we say that an edge $e\in E$ or a vertex $v\in V$ is satisfied by $c$ if $P_c(e)$ or $P_c(v)$, resp., contains a unique colour, where
$P_c(e)=\{c(e'):e'\in E_H[e]\}$ and $P_c(v)=\{c(e'):e'\in E_H(v)\}$ for $v\in V$.
Then by $\chi'_{sCF}(G)$ we denote the least number of colours admitting an edge colouring $c$ of some subgraph of $G$ (i.e. minimized over all subgraphs) so that all edges of $G$ are satisfied by $c$.
As given such a colouring one may use yet another additional colour to obtain a desired colouring of entire graph, for every graph $G$ without isolated edges we have:
\begin{equation}\label{ChiCFandChiSCF_relation}
\chi'_{sCF}(G)\leq \chi'_{CF}(G)\leq \chi'_{sCF}(G)+1.
\end{equation}
Note that the colouring used in the proof of Observation~\ref{ObservComplBip} actually yields something more.
\begin{observation}\label{ObservComplBipSubgraph}
For every $n,m\geq 1$, $\chi'_{sCF}(K_{n,m})\leq 2$.
\end{observation}
\begin{theorem}\label{ChiSCF-bipartite}
If $G$ is a bipartite graph without isolated vertices, then $\chi'_{sCF}(G)\leq 3$.
\end{theorem}
\begin{pf}
We may assume $G=(V,E)$ is connected.
Let $T$ be a BFS spanning tree of $G$ rooted at $r\in V$ with $d_G(r)=\delta(G)$
and let $L_0,L_1,\ldots,L_q$ be the levels of $T$, i.e. $L_0=\{r\}$. Recall that every edge of $G$ must join vertices from two consecutive levels of $T$.
Suppose first that $d_G(r)=1$. Consider a $\mathbb{Z}_3$-colouring $c'$ of $T$ complying with Lemma~\ref{TreeColouring}.
To show that $c'$ satisfies all edges, consider any $uv\in E$ with $u\in L_i,v\in L_{i+1}$ where $i\geq 1$. Then however, analogously as in Lemma~\ref{TreeColouring}, by ($1^\circ$), $P_c(uv)=(i-1)i^{k_e}(i+1)^{k'_e}$ where $k_e,k'_e\geq 0$ (i.e. ($2^\circ$) holds for the edges of $G$). Thus we are done, as the rooted edge is satisfied by $c'$ as well.
Assume therefore that $d_G(r)=d_T(r)\geq 2$. Let $s$ be the last neighbour of $r$ within the BFS-vertex ordering corresponding to $T$. Let $T_r,T_s$ be the subtrees obtained from $T$ by deleting the edge $rs$ such that $r\in V(T_r)$ and $s\in V(T_s)$. The choice of $s$ shall be crucial in our argument -- note it implies that:
\begin{eqnarray}
&&{\rm ~if} ~uv\in E\smallsetminus\{st\} {\rm ~is ~an ~edge ~joining ~levels} ~L_i {\rm ~and} ~L_{i+1} \nonumber\\
&&{\rm ~such ~that } ~u\in V(T_r) {\rm ~and} ~v\in (T_s), {\rm ~then} ~u\in L_{i+1} {\rm ~and} ~v\in L_i. \label{CrucialFeature}
\end{eqnarray}
Let $G_r:=G[V(T_r)]$, $G_s:=G[V(T_s)]$, and set $T'_r:=T_r+rs$, $T'_s:=T_s+rs$, $G'_r:=G_r+rs$, $G'_s:=G_s+rs$. Note that all edges of $G$ outside $E(G'_r)\cup E(G'_s)$ must satisfy~(\ref{CrucialFeature}). Note also that $T'_r$ can be regarded as a BFS spanning tree of $G'_r$ rooted at $s$, while $T'_s$ -- as a BFS spanning tree of $G'_s$ rooted at $r$, where $d_{T'_r}(s)=1$, $d_{T'_s}(r)=1$.
Let $c_r,c_s$ be $\mathbb{Z}_3$-colourings of $T'_r,T'_s$, resp., complying with Lemma~\ref{TreeColouring}, and let $c$ be a concatenation of $c_r$ and $c_s$ (note these agree on $rs$), which is an edge $\mathbb{Z}_3$-colouring of $T$.
Since $T_r$ and $T_s$ are vertex disjoint, then analogously as in the first case, $c$ satisfies all edges in $G_r$ and all edges in $G_s$, while $P_c(rs)=01^{k_{rs}}$ for some $k_{rs}>1$. It thus remains to show that the edges joining $V(T_r)$ and $V(T_s)$, other than $rs$ are satisfied by $c$ as well.
Let $L^r_0,L^r_1,\ldots,L^r_{q_r}$ and $L^s_0,L^s_1,\ldots,L^s_{q_s}$ be the levels in $T'_r$ and $T'_s$, resp., hence in particular $L^r_0=\{s\}$ and $L^s_0=\{r\}$. Note that for every $j\geq 0$:
\begin{eqnarray}\label{LevelsInclusions}
L^s_j\subseteq L_j~~~~~~ {\rm and}~~~~~~L^r_{j+1}\subseteq L_j.
\end{eqnarray}
Suppose $uv\neq rs$ is an edge of $G$ joining levels $L_i$ and $L_{i+1}$, $i\geq 1$, such that $u\in V(T_r)$ and $v\in (T_s)$. By~(\ref{CrucialFeature}), $u\in L_{i+1}$ and $v\in L_i$. Thus by~(\ref{LevelsInclusions}), $u\in L^r_{i+2}$ and $v\in L^s_i$. Therefore, as $c_r,c_s$ comply with Lemma~\ref{TreeColouring}, by ($1^\circ$), $P_c(u)=P_{c_r}(u)=(i+1)(i+2)^{k_u}=(i+1)(i-1)^{k_u}$ (modulo 3) and $P_c(v)=P_{c_s}(v)=(i-1)i^{k_v}$ for some $k_u,k_v\geq 0$. Hence, $P_c(uv)=(i+1)i^{k_{uv}}(i-1)^{k'_{uv}}$ for some $k_{uv},k'_{uv}\geq 0$, and thus $c$ satisfies $uv$.
\qed
\end{pf}
By~(\ref{ChiCFandChiSCF_relation}) we thus obtain the following corollary.
\begin{corollary}\label{ChiCF-bipartite}
If $G$ is a bipartite graph without isolated vertices, then $\chi'_{CF}(G)\leq 4$.
\end{corollary}
We may now prove our main result.
\begin{theorem}\label{ChiSCF-general}
For every graph $G$ without isolated vertices, $\chi'_{sCF}(G)\leq 3\lceil\log_2\chi(G)\rceil$.
\end{theorem}
\begin{pf}
Let $G=(V,E)$. We prove the theorem by induction. If $\chi(G)=2$, then the assertion follows by Theorem~\ref{ChiSCF-bipartite}.
Let us thus assume that $\chi(G)\geq 3$ and that the assertion holds for graphs with smaller chromatic numbers.
Colour the vertices of $G$ properly with $\chi(G)$ colours and let $V_i$ denote the vertices coloured with $i$th colour, $1\leq i\leq \chi(G)$. Set $V_i=\emptyset$ for $\chi(G)< i\leq 2^{\lceil\log_2\chi(G)\rceil}$, if any; set $k:=2^{\lceil\log_2\chi(G)\rceil-1}$.
Consider a bipartite subgraph $H$ induced by all edges between $A:=V_1\cup\ldots\cup V_k$ and $B:=V_{k+1}\cup\ldots\cup V_{2k}$ in $G$
and let $F$ be the graph induced by all the remaining edges of $G$ (we do not include potential isolated vertices in the vertex sets of $H$ or $F$). Note that $\chi(F)\leq k$, as we may in a way identify colours of the vertices in $V_i$ and $V_{k+i}$ for $1\leq i\leq k$ because there are no edges between these vertices in $F$ (these were included in $F$).
Since $k<\chi(G)$, by induction hypothesis, $\chi'_{sCF}(G)\leq 3\lceil\log_2\chi(F)\rceil \leq 3\lceil\log_2k\rceil = 3\lceil\log_2\chi(G)\rceil - 3$. There thus exists a subgraph of $F$ and its edge colouring with at most $3\lceil\log_2\chi(G)\rceil - 3$ colours which satisfies all edges of $F$. On the other hand, as $H$ is bipartite, by Theorem~\ref{ChiSCF-bipartite} there exists a colouring of a subgraph of $H$ with $3$ additional colours which satisfies all edges in $H$. The concatenation of the both colourings implies the theorem.
\qed
\end{pf}
By~(\ref{ChiCFandChiSCF_relation}), Theorem~\ref{ChiSCF-general} immediately implies the following corollary.
\begin{corollary}\label{CorollaryGeneralChi}
For every graph $G$ without isolated vertices, $\chi'_{CF}(G)\leq 3\lceil\log_2\chi(G)\rceil+1$.
\end{corollary}
In~\cite{DebskiPrzybylo} it is proved that
$\chi'_{CF}(K_n) > \lfloor \log_2n-\log_2\log_2n-1\rfloor = \lfloor \log_2\chi(K_n)-\log_2\log_2\chi(K_n)-1\rfloor$.
Thus in general one cannot expect the multiplicative constant of $3$ in the general upper bound from Corollary~\ref{CorollaryGeneralChi} to be pushed down below $1$. Note on the other hand that the same reasoning as applied within the proof of Theorem~\ref{ChiSCF-general} yields a slightly stronger upper bound in the case of the complete graphs.
\begin{observation}\label{UpperBoundCompleteGr}
For every $n\geq 2$, $\chi'_{sCF}(K_n)\leq 2\lceil\log_2 n\rceil$.
\end{observation}
\begin{pf}
The result follows by the same proof as the one of Theorem~\ref{ChiSCF-general}, as for $G=K_n$, $H$ is the complete bipartite graph, and thus by Observation~\ref{ObservComplBipSubgraph}, $\chi'_{sCF}(H)\leq 2$, while $F$ is a union of two disjoint complete graphs.
\qed
\end{pf}
\begin{corollary}\label{CorollaryGeneralDelta}
For every graph $G$ without isolated vertices, $\chi'_{CF}(G)\leq 3\lceil\log_2\Delta(G)\rceil+1$.
\end{corollary}
\begin{pf}
We may assume $G$ is connected.
Due to Brook's Theorem the result follows directly by Corollary~\ref{CorollaryGeneralChi} for all graphs except cycles and complete graphs. The remaining cases follow by Observations~\ref{ObservComplBip} and~\ref{UpperBoundCompleteGr}.
\qed
\end{pf}
\section{Concluding remarks}
By Corollary~\ref{ChiCF-bipartite} and Observation~\ref{ObservComplBip} we know that the optimal general upper bound for the conflict-free chromatic index within the family of bipartite graphs is either $3$ or $4$. We propose the following conjecture.
\begin{conjecture}
If $G$ is a bipartite graph without isolated vertices, then $\chi'_{CF}(G)\leq 3$.
\end{conjecture}
Note that even settling this conjecture in the affirmative might not provide
a significant improvement of the general bound in Corollary~\ref{CorollaryGeneralChi}. This would decrease merely by $1$ on the basis of our approach utilized in the proof of Theorem~\ref{ChiSCF-general}, unless we knew something specific on the corresponding colourings of bipartite graphs, e.g. implying that.$\chi'_{sCF}(G)\leq 2$ for these.
We believe that it would also be interesting to solve the following problem
\begin{problem}
Characterize the family of all trees $T$ with $\chi'_{CF}(T)=3$.
\end{problem}
| 8df82f7921565982d3517707c783cbf53b79e1bd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section[Introduction]{Introduction}
We show that linear chaos in the space $c(\N)$ ($\N:=\left\{1,2,3,\dots\right\}$ is the set of \textit{natural numbers}) of convergent sequences cannot be arrived at by merely extending the weighted backward shifts in the space $c_0(\N)$ of vanishing sequences, \textit{bounded}
\[
c_0\ni x:=\left(x_k\right)_{k\in \N}\mapsto A_wx:=w\left(x_{k+1}\right)_{k\in \N}\in c_0
\quad (|w|>1),
\]
introduced in \cite{Rolewicz} (see also \cite{Godefroy-Shapiro1991}), as well as \textit{unbounded}
\[
A_wx:=\left(w^kx_{k+1}\right)_{k\in \N} \quad (|w|>1)
\]
with maximal domain
\[
D(A_w):=\left\{ x:=\left(x_k\right)_{k\in \N} \in c_0 \,\middle|\, \left(w^kx_{k+1}\right)_{k\in \N}\in c_0 \right\},
\]
introduced in \cite{arXiv:1811.06640}.
Applying the newly found \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}) \cite[Theorem $3.2$]{arXiv:2106.14872}, we furnish concise proofs of the chaoticity of the aforementioned weighted backward shifts along with their powers and also itemize their spectral structure.
We further construct bounded and unbounded linear chaotic operators in $c(\N)$ as conjugate to the chaotic backward shifts in $c_0(\Z_+)$ ($\Z_+:=\left\{0,1,2,\dots\right\}$ is the set of \textit{nonnegative integers}) via a homeomorphic isomorphism between the two spaces.
As follows from the inclusions
\[
c_0(\N)\subset c(\N)\subset l_\infty(\N),
\]
the space $c(\N)$ lives between the space $c_0(\N)$, where linear chaos is known, and the space $l_\infty(\N)$ of bounded sequences, where linear chaos does not exist.
\begin{rem}
Henceforth, we use the notations $c_0(\N)$, $c(\N)$ for the spaces of vanishing and convergent sequences over $\N$, respectively, and the notations $c_0(\Z_+)$, $c(\Z_+)$ for their counterparts over $\Z_+$. We also use the shorter notations $c_0$ and $c$ whenever the indexing set is implied contextually.
\end{rem}
\section[Preliminaries]{Preliminaries}
The subsequent preliminaries are essential for our discourse.
\subsection{Spaces $c_0$ and $c$}\
The spaces
\[
c_0:= \left\{ x:= (x_k)_{k \in \N}\in \F^\N \,\middle|\, \lim_{k \to \infty}x_k = 0 \right\}
\]
of \textit{vanishing sequences} and
\[
c:= \left\{ x:= (x_k)_{k \in \N}\in \F^\N \,\middle|\, \exists \lim_{k \to \infty}x_k \in \F \right\}
\]
of \textit{convergent sequences} are infinite-dimensional separable Banach spaces relative to $\infty$-norm
\[
x:= (x_k)_{k\in \N} \mapsto \|x\|_\infty := \sup_{k \in \N} |x_k|,
\]
the former being a closed \textit{hyperplane}, which is a \textit{nowhere dense} subspace, of the latter (see, e.g., \cite{Markin2018EFA,Markin2020EOT}).
The \textit{limit functional}
\begin{equation}\label{lf}
c \ni x:= (x_n)_{n\in \N} \mapsto l(x):= \lim_{n \to \infty} x_n \in \F,
\end{equation}
is a bounded linear functional on $c$ with $\ker l = c_0$ (see, e.g., \cite{Markin2018EFA,Markin2020EOT}).
Relative to the standard Schauder basis $\left\{e_n:=\left(\delta_{nk}\right)_{k\in \N}\right\}_{n\in \N}$ for $c_0$, where $\delta_{nk}$ is the \textit{Kronecker delta}, each $x := \left(x_k\right)_{k \in \N}\in c_0$ allows the Schauder expansion
\[
x = \sum_{k=1}^\infty c_k(x) e_k
\]
with the \textit{coordinates} $c_k(x)=x_k$, $k\in \N$.
Relative to the standard Schauder basis $\left\{e_n\right\}_{n\in \Z_+}$ for $c$, where
\[
e_0 := (1,1,1,\dots)\quad \text{and}\quad e_n:=\left(\delta_{nk}\right)_{k\in \N},\ n\in\N,
\]
each $x := \left(x_k\right)_{k \in \N}\in c$ has the Schauder expansion
\begin{equation*
x = \sum_{k=0}^\infty c_k(x) e_k
\end{equation*}
with the \textit{coordinates}
\begin{equation*
c_0(x)=l(x)\quad \text{and}\quad c_k(x)=x_k-l(x),\ k\in \N.
\end{equation*}
See, e.g., \cite{Markin2018EFA,Markin2020EOT,MarkSogh2021}.
\subsection{Spectrum}\
The spectrum $\sigma(A)$ of a closed linear operator $A$ in a complex Banach space $X$ is the union of the following pairwise disjoint sets:
\begin{equation*}
\begin{split}
& \sigma_p(A):=\left\{\lambda\in \C \,\middle|\,A-\lambda I\ \text{is \textit{not injective}, i.e., $\lambda$ is an \textit{eigenvalue} of $A$} \right\},\\
& \sigma_c(A):=\left\{\lambda\in \C \,\middle|\,A-\lambda I\ \text{is \textit{injective},
\textit{not surjective}, and $\overline{R(A-\lambda I)}=X$} \right\},\\
& \sigma_r(A):=\left\{\lambda\in \C \,\middle|\,A-\lambda I\ \text{is \textit{injective} and $\overline{R(A-\lambda I)}\neq X$} \right\}
\end{split}
\end{equation*}
($R(\cdot)$ is the \textit{range} of an operator and $\overline{\cdot}$ is the \textit{closure} of a set), called the \textit{point}, \textit{continuous} and \textit{residual spectrum} of $A$, respectively (see, e.g., \cite{Dun-SchI,Markin2020EOT}).
\subsection{Hypercyclicity and Linear Chaos}\
\begin{defn}[Hypercyclic and Chaotic Linear Operators]\ \\ \noindent
For a (bounded or unbounded) linear operator $A$ in a (real or complex) Banach space $X$, a nonzero vector
\begin{equation*}
x\in C^\infty(A):=\bigcap_{n=0}^{\infty}D(A^n)
\end{equation*}
($D(\cdot)$ is the \textit{domain} of an operator, $A^0:=I$, $I$ is the \textit{identity operator} on $X$) is called \textit{hypercyclic} if its \textit{orbit} under $A$
\[
\orb(x,A):=\left\{A^nx\right\}_{n\in\Z_+}
\]
is dense in $X$, i.e.,
\[
\bar{\orb(x,A)}=X.
\]
Linear operators possessing hypercyclic vectors are said to be \textit{hypercyclic}.
If there exist an $N\in \N$ and a vector
\[
x\in D(A^N)\quad \text{with}\quad A^Nx = x,
\]
such a vector is called a \textit{periodic point} for the operator $A$ of period $N$. If $f\ne 0$, we say that $N$ is a \textit{period} for $A$.
Hypercyclic linear operators with a dense in $X$ set $\Per(A)$ of periodic points, i.e.,
\[
\bar{\Per(A)}=X,
\]
are said to be \textit{chaotic}.
\end{defn}
See \cite{Devaney,Godefroy-Shapiro1991,B-Ch-S2001}.
\begin{samepage}
\begin{exmps}\
\begin{enumerate}[label=\arabic*.]
\item On the infinite-dimensional separable Banach space $X:=c_0$ or $X:=\ell_p$ ($1\le p<\infty$), the classical Rolewicz weighted backward shifts
$$ X \ni x:=\left(x_k\right)_{k\in \N} \mapsto A_wx:= w\left(x_{k+1}\right)_{k\in \N} \in X, $$
where $w\in \F$ ($\F:=\R$ or $\F:=\C$) with $|w|>1$ are \textit{chaotic} \cite{Rolewicz,Godefroy-Shapiro1991}.
\item On the sequence space
\[
X:=\left\{ \left(x_n\right)_{n\in \N}\in \F^\N\,\middle|\,
\sum_{k=1}^\infty \left|\frac{x_{k}}{k}- \frac{x_{k+1}}{k+1}\right|<\infty\ \text{and}\ \lim_{k\to\infty}\frac{x_k}{k} = 0\right\},
\]
which is an infinite-dimensional separable Banach space relative to the norm
$$X \ni x:=\left(x_k\right)_{k\in \N}\mapsto \|x\| :=\sum_{k=1}^\infty \left| \frac{x_{k}}{k}- \frac{x_{k+1}}{k+1} \right|, $$
the backward shift
$$ X \ni x:=\left(x_k\right)_{k\in \N} \mapsto
Ax:= \left(x_{k+1}\right)_{k\in \N} \in X $$
is \textit{hypercyclic} but \textit{not} densely periodic,
and hence, \textit{not} chaotic \cite[Exercise $4.1.3$]{Grosse-Erdmann-Manguillot}.
\item On an infinite-dimensional separable Banach space $(X,\|\cdot\|)$,
the identity operator $I$ is densely periodic but \textit{not}
hypercyclic, and hence, \textit{not} chaotic.
\end{enumerate}
\end{exmps}
\end{samepage}
\begin{samepage}
\begin{rems}\label{HCrems}\
\begin{itemize}
\item In the prior definition of hypercyclicity, the underlying space is necessarily
\textit{infinite-dimensional} and \textit{separable} (see, e.g., \cite{Grosse-Erdmann-Manguillot}).
\item For a hypercyclic linear operator $A$, the set $HC(A)$ of its hypercyclic vectors is necessarily dense in $X$, and hence, the more so, is the subspace $C^\infty(A)\supseteq HC(A)$.
\item Observe that
\[
\Per(A)=\bigcup_{N=1}^\infty \Per_N(A),
\]
where
\[
\Per_N(A)=\ker(A^N-I),\ N\in \N
\]
is the \textit{subspace} of $N$-periodic points of $A$.
\item As immediately follows from the inclusions
\begin{equation*}
HC(A^n)\subseteq HC(A),\ \Per(A^n)\subseteq \Per(A), n\in \N,
\end{equation*}
if, for a linear operator $A$ in an infinite-dimensional separable Banach space $X$ and some $n\ge 2$, the operator $A^n$ is hypercyclic or chaotic, then $A$ is also hypercyclic or chaotic, respectively.
\end{itemize}
\end{rems}
\end{samepage}
Prior to \cite{B-Ch-S2001,deL-E-G-E2003}, the notions of linear hypercyclicity and chaos had been studied exclusively for \textit{continuous} linear operators on Fr\'echet spaces, in particular for \textit{bounded} linear operators on Banach spaces (for a comprehensive survey, see \cite{Bayart-Matheron,Grosse-Erdmann-Manguillot}).
The following extension of \textit{Kitai's ctriterion} for bounded linear operators (see \cite{Kitai1982,Gethner-Shapiro1987}) is a useful shortcut for establishing hypercyclicity for (bounded or unbounded) linear operators without explicitly furnishing a hypercyclic vector as in \cite{Rolewicz}.
\begin{thm}[Sufficient Condition for Hypercyclicity {\cite[Theorem $2.1$]{B-Ch-S2001}}]\ \\
Let $X$ be a (real or complex) infinite-dimensional separable Banach space and A be a densely defined linear operator in X such that each power $A^n \ (n \in \N)$ is a closed operator. If there exists a set
\[
Y\subseteq C^\infty(A):=\bigcap_{n=1}^\infty D(A^n)
\]
dense in $X$ and a mapping $B:Y\to Y$ such that
\begin{enumerate}
\item $\forall\, x\in Y:\ ABx=x$ and
\item $\forall\, x\in Y:\ A^n x, B^n x \to 0,\ n \to \infty,$
\end{enumerate}
then the operator $A$ is hypercyclic.
\end{thm}
The subsequent newly established sufficient condition for linear chaos \cite{arXiv:2106.14872}, obtained via strengthening one of the hypotheses of the prior sufficient condition for hypercyclicity, serves as a shortcut for establishing chaoticity for (bounded or unbounded) linear operators without explicitly furnishing both a hypercyclic vector and a dense set of periodic points and is fundamental for our discourse.
\begin{thm}[Sufficient Condition for Linear Chaos {\cite[Theorem $3.2$]{arXiv:2106.14872}}]\label{SCC}\ \\
Let $(X,\|\cdot\|)$ be a (real or complex) infinite-dimensional separable Banach space and $A$ be a densely defined linear operator in $X$ such that each power $A^{n}$ ($n\in\N$) is a closed operator. If there exists a set
\[
Y\subseteq C^\infty(A):=\bigcap_{n=1}^\infty D(A^n)
\]
dense in $X$ and a mapping $B:Y\to Y$ such that
\begin{enumerate}
\item $\forall\, x\in Y:\ ABx=x$ and
\item $\forall\, x\in Y\ \exists\, \alpha=\alpha(x)\in (0,1),\ \exists\, c=c(x,\alpha)>0\ \forall\, n\in \N:$
\begin{equation*}
\max\left(\|A^nx\|,\|B^nx\|\right)\le c\alpha^n,
\end{equation*}
\end{enumerate}
then the operator $A$ is chaotic.
\end{thm}
For applications, see \cite{arXiv:2106.09682}.
We also need the following statements.
\begin{cor}[Chaoticity of Powers {\cite[Corollary $4.3$]{arXiv:2106.14872}}]\label{CP}\ \\
For a chaotic linear operator $A$ in a (real or complex) infinite-dimensional separable Banach space subject to the \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}), each power $A^n$ ($n\in \N$) is chaotic.
\end{cor}
\begin{thm}[Bourdon {\cite[Theorem $2.54$]{Grosse-Erdmann-Manguillot}}]\label{BT}\ \\
For a bounded linear hypercyclic operator $A$ on an infinite-dimensional separable Banach space $X$ and a nonzero polynomial $p(\lambda):=\sum_{k=0}^n c_k\lambda^k$
($n\in \Z_+$, $c_k\in \F$, $k=0,\dots,n$), the range $R(p(A))$ of the operator
$p(A):=\sum_{k=0}^n c_kA^k$ is dense in $X$, i.e.,
\[
\bar{R(p(A))}=X.
\]
\end{thm}
\begin{rem}
Consistently with more general necessary conditions for hypercyclicity, found in \cite{arXiv:2106.14872}, the latter implies that, for a bounded linear hypercyclic operator $A$ on an infinite-dimensional separable Banach space $X$ and an arbitrary $\lambda\in \F$, the range $R(A-\lambda I)$ of the operator $A-\lambda I$ is \textit{dense} in $X$, i.e.,
\[
\bar{R(A-\lambda I)}=X.
\]
\end{rem}
\section{Linear Chaos in $c_0$}
\begin{thm}[Bounded Linear Chaos on $c_0$]\label{BLCc0}\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the bounded linear weighted backward shift operator
\[
c_0\ni x:=\left(x_k\right)_{k\in \N}\mapsto A_wx:=w\left(x_{k+1}\right)_{k\in \N}\in c_0
\]
on the space $c_0$ is chaotic as well as its every power $A_w^n$ $(n\in \N)$ and, provided the underlying space is complex (i.e., $\F=\C$),
\[
\sigma\left(A_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|\le |w| \right\}
\]
with
\[
\sigma_p\left(A_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|<|w| \right\}\quad \text{and}\quad
\sigma_c\left(A_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|=|w| \right\}.
\]
\end{thm}
\begin{proof}
Here, we provide a concise proof based on the \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}) (cf. the original proofs for hypercyclicity and chaoticity \cite{Rolewicz,Godefroy-Shapiro1991}).
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
Consider the subspace
\[
Y:=c_{00}:=\left\{ x:= (x_k)_{k \in \N}\in \F^\N \,\middle|\, \exists\, N\in \N\ \forall\,k\ge N:\ x_k=0 \right\}
\]
\textit{dense} in $c_0$ (see, e.g., \cite{Markin2020EOT,Markin2018EFA}) and the mapping $B:Y\to Y$, which is the restriction to $Y$ of the \textit{bounded linear operator}
on $c_0$
\[
c_0\ni x:=(x_k)_{k\in \N}\mapsto Bx:=w^{-1}\left(x_{k-1}\right)_{k\in \N}\in c_0 \quad (x_0:=0),
\]
(the right inverse of $A$) for which
\begin{equation}\label{BBS}
\|B\|={|w|}^{-1}<1
\end{equation}
(here and henceforth, whenever appropriate, $\|\cdot\|$ stands for the \textit{operator norm}) (see, e.g., \cite{Markin2020EOT}) and
\begin{equation}\label{RI1}
ABx=x,\ x\in Y.
\end{equation}
Let us show that
$\forall\, x\in Y\ \exists\, \alpha=\alpha(f)\in (0,1),\ \exists\, c=c(f,\alpha)>0\ \forall\, n\in \N:$
\begin{equation*}
\max\left(\|A^nx\|,\|B^nx\|\right)\le c\alpha^n,
\end{equation*}
Let $x:= (x_k)_{k\in \N} \in Y$ be arbitrary. Then
\[
\exists\, N\in \N\ \forall\,k\ge N:\ x_k=0,
\]
and hence,
\[
\forall\,n\ge N:\ A^nx = 0,
\]
which, implies that
\[
\forall\, \alpha\in (0,1),\ \exists\, c=c(x,\alpha)>0,\ \forall\, n\in \N:\
\|A^nx\|_\infty\le c\alpha^n.
\]
By the submultiplicativity of the operator norm, in view of \eqref{BBS}, we also have:
\begin{equation*}
\| B^n x\| \leq \| B^n \| \| x \| \leq \| B \|^n \| x \| = |w|^{-n} \| x \|.
\end{equation*}
By the \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}) and the \textit{Chaoticity of Powers Corollary} (Corollary \ref{CP}), we conclude that the operator $A$ is \textit{chaotic} as well as every power $A^n$ $(n\in \N)$.
Provided the underlying space is complex, the spectral part of the statement
immediately follows from the fact that
\[
A=wL,
\]
where
\[
c_0\ni x:=\left(x_k\right)_{k\in \N}\mapsto Lx:=\left(x_{k+1}\right)_{k\in \N}\in c_0
\]
is the backward shift operator on $c_0$, for which
\[
\sigma\left(L\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|\le 1 \right\}
\]
with
\[
\sigma_p\left(L\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|<1 \right\}\quad \text{and}\quad
\sigma_c\left(L\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|=1 \right\}
\]
(see, e.g., \cite{Dun-SchI,Markin2020EOT}).
\end{proof}
\begin{lem}\label{lem}\ \\
Let $w\in \F$ and $|w|>1$. Then, for the weighted backward shift operator
\[
A_wx:=\left(w^kx_{k+1}\right)_{k\in \N}
\]
in the space $c_0$ with maximal domain
\[
D(A_w):=\left\{ x:=\left(x_k\right)_{k\in \N} \in c_0 \,\middle|\, \left(w^kx_{k+1}\right)_{k\in \N}\in c_0 \right\},
\]
each power
\[
A_w^{n}x= \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \N}
\]
($n\in\N$) with domain
\[
D(A_w^{n})= \left\{ x := \left(x_k \right)_{k \in \N} \in c_0 \,\middle|\, \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \N} \in c_0 \right\}
\]
is a densely defined unbounded closed linear operator and the subspace
\[
C^\infty(A):=\bigcap_{n=1}^\infty D(A^n)
\]
of infinite differentiable relative to $A$ vectors is dense in $c_0$.
\end{lem}
\begin{proof}
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
Since
\[
A^2x= \left( w^{k} w^{k+1}x_{k+2} \right)_{k \in \N}
\]
with domain
\begin{align*}
D(A^{2})
&= \left\{ x := \left(x_k \right)_{k \in \N} \in D(A) \,\middle|\, Ax \in D(A) \right\}
\\
&= \left\{ x := \left(x_k \right)_{k \in \N} \in c_0 \,\middle|\, \left( w^{k} w^{k+1}x_{k+2} \right)_{k \in \N} \in c_0 \right\}
\end{align*}
and
\[
A^3x= \left( w^{k} w^{k+1}w^{k+2}x_{k+3} \right)_{k \in \N}
\]
with domain
\begin{align*}
D(A^{3})
&= \left\{ x := \left(x_k \right)_{k \in \N} \in D(A^2) \,\middle|\, A^2x \in D(A) \right\}
\\
&= \left\{ x := \left(x_k \right)_{k \in \N} \in c_0 \,\middle|\, \left( w^{k} w^{k+1}w^{k+2}x_{k+3} \right)_{k \in \N} \in c_0 \right\}
\end{align*}
we infer inductively that, for each $n \in \N$
\[
A^nx = \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \N},
\]
with domain
\[
D(A^{n})= \left\{ x := \left(x_k \right)_{k \in \N} \in c_0 \,\middle|\, \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \N} \in c_0 \right\}
\]
We have:
\[
D(A^{n+1}) \subseteq D(A^n),\ n \in \N.
\]
Since the subspace $c_{00}$ is \textit{dense} in $c_0$ and
\[
c_{00} \subseteq D(A^n),\ n \in \N,
\]
then each power $A^n$ ($n\in \N$) is densely defined and furthermore
\[
C^\infty(A) := \bigcap_{n=1}^\infty D(A^n)
\]
is also \textit{dense} in $c_0$.
Let $n\in \N$ and $e_m:=\left(\delta_{mk} \right)_{k\in \N}$, $m\in\N$, with
$\|e_m\|_\infty=1$, $m\in \N$. Then, in view of $|w|>1$,
\begin{align*}
\forall\, m\in \N:\ \|A^ne_{n+m}\|& = \left\| \left( \left[ \prod_{j=k}^{k+n-1} w^j \right] \delta_{(n+m)(k+n)} \right)_{k\in\N} \right\|_\infty = \prod_{j=1}^{m+n-1} |w|^j\\
&= |w|^{ \sum_{j=1}^{m+n-1}j }= |w|^{\frac{(m+n)(m+n-1)}{2}} \to \infty,\ m \to \infty,
\end{align*}
which implies that the linear operator $A^n$ is \textit{unbounded}.
Let $n\in \N$ and a sequence $\left(x^{(m)}:=\left(x_k^{(m)} \right)_{k\in \N} \right)_{n\in \N}$ in $D(A^n)$ be such that
\[
x^{(m)} \to x:=\left(x_k \right)_{m\in \N}\in c_0,\ m \to \infty,
\]
and
\[
A^nx^{(m)} = \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n}^{(m)} \to y:=\left(y_k \right)_{k\in \N}\in c_0,\ m \to \infty.
\]
Then, for each $k \in \N$ (see, e.g., \cite{Markin2018EFA,Markin2020EOT,MarkSogh2021}),
\[
x_k^{(m)} \to x_k,\ m \to \infty,
\]
and
\[
\left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n}^{(m)} \to y_k, \ m \to \infty.
\]
Whence we infer that, for each $k\in \N$,
\[
\left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} = y_k,
\]
which means that
\[
\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \N}=y\in c_0.
\]
Therefore, $x \in D(A^n)$ and $y = A^nx$, which, by the \textit{Sequential Characterization of Closed Linear Operators} (see, e.g., \cite{Markin2018EFA,Markin2020EOT}), implies the operator $A^n$ is closed.
\end{proof}
\begin{thm}[Unbounded Linear Chaos in $c_0$]\label{ULCc0}\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the unbounded linear weighted backward shift operator
\[
A_wx:=\left(w^kx_{k+1}\right)_{k\in \N}
\]
in the space $c_0$ with maximal domain
\[
D(A_w):=\left\{ x:=\left(x_k\right)_{k\in \N} \in c_0 \,\middle|\, \left(w^kx_{k+1}\right)_{k\in \N}\in c_0 \right\}
\]
is chaotic as well as every power $A_w^n$ $(n\in \N)$.
Furthermore, each $\lambda \in \F$ is an eigenvalue for $A_w$ of geometric multiplicity $1$, i.e.,
\[
\dim\ker(A_w-\lambda I)=1.
\]
In particular, provided the underlying space is complex,
\[
\sigma_p\left(A_w\right)=\C.
\]
\end{thm}
\begin{proof}
Here, we also provide a concise proof based on the \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}) (cf. the original proof \cite{arXiv:1811.06640}).
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
Consider the subspace
\[
Y:=c_{00}
\]
\textit{dense} in $c_0$ and the mapping $B:Y\to Y$, which is the restriction to $Y$ of the
\textit{bounded linear operator} on $c_0$
\[
c_0\ni x:=(x_k)_{k\in \N}\mapsto Bx:=\left(w^{-(k-1)}x_{k-1}\right)_{k\in \N}\in c_0\quad (x_0:=0),
\]
(the right inverse of $A$) for which
\begin{equation}\label{RI2}
ABx=x,\ x\in Y.
\end{equation}
With
\[
c_0\ni x:=(x_k)_{k\in \N}\mapsto B^2x= \left( w^{-(k-1)} w^{-(k-2)} x_{k-2} \right)_{k \in \N}\quad (x_{k-2}:=0,\ k=1,2)
\]
and
\[
c_0\ni x:=(x_k)_{k\in \N}\mapsto B^3x= \left( w^{-(k-1)} w^{-(k-2)} w^{-(k-3)} x_{k-3} \right)_{k \in \N}
\]
($x_{k-3}:=0,\ k=1,2,3$), we infer inductively that, for any $n\in \N$,
\[
c_0\ni x:=(x_k)_{k\in \N}\mapsto B^nx = \left(\left[ \prod_{j=1}^{n} w^{-(k-j)} \right] x_{k-n} \right)_{k \in \N}\quad (x_{k-n}:=0,\ k=1,\dots,n),
\]
or equivalently, in view of
\[
\prod_{j=1}^{n} w^{-(k-j)}=w^{-\sum_{j=1}^n (k-j)}=w^{-nk + \frac{n(n+1)}{2}},
\]
we have:
\[
c_0\ni (x_k)_{k\in \N}\mapsto B^nx = \left( w^{-nk + \frac{n(n+1)}{2}}x_{k-n} \right)_{k \in \N}\quad (x_k:=0,\ k=1,\dots,n)
\]
with
\begin{equation*}
\|B^n\| ={|w|}^{-\frac{n(n+1)}{2}}.
\end{equation*}
Indeed, for any $x:=\left(x_k\right)_{k\in \N}\in c_0$, in view of $|w|>1$,
\begin{align*}
\|B^nx\|_\infty&=\sup_{k\in \N}\left|w^{-nk + \frac{n(n+1)}{2}}x_{k-n}\right|
=\sup_{k\ge n+1}\left|w^{-nk + \frac{n(n+1)}{2}}x_{k-n}\right|
\\
&\le \sup_{k\ge n+1}{\left|w\right|}^{-nk + \frac{n(n+1)}{2}}
\sup_{k\ge n+1}\left|x_{k-n}\right|
={\left|w\right|}^{-n(n+1) + \frac{n(n+1)}{2}}\|x\|_\infty
\\
&={\left|w\right|}^{-\frac{n(n+1)}{2}}\|x\|_\infty,
\end{align*}
and hence,
\[
\|B^n\| \le {\left|w\right|}^{-\frac{n(n+1)}{2}}.
\]
Further, since, for $e_{1}:=\left(\delta_{1k}\right)_{k\in \N}$ with $\|e_{1}\|_\infty=1$,
\[
\|B^ne_1\|_\infty=\sup_{k\ge n+1}\left|w^{-nk + \frac{n(n+1)}{2}}\delta_{1(k-n)}\right|
={\left|w\right|}^{-n(n+1) + \frac{n(n+1)}{2}}={\left|w\right|}^{-\frac{n(n+1)}{2}},
\]
we infer that
\[
\|B^n\| ={\left|w\right|}^{-\frac{n(n+1)}{2}}.
\]
Thus,
\begin{equation}\label{UBS}
\lim_{n\to \infty}{\|B^n\|}^{1/n}=\lim_{n\to \infty}{\left|w\right|}^{-\frac{n+1}{2}}=0,
\end{equation}
i.e., the operator $B$ is \textit{quasinilpotent} (cf. \cite{arXiv:1811.06640}), which implies that
\[
\forall\, x\in Y,\ \forall\, \alpha\in (0,1)\ \exists\, c=c(x,\alpha)>0\ \forall\, n\in \N:\
\|B^nx\|_\infty\le c\alpha^n.
\]
Let $x:= (x_k)_{k\in \N} \in Y$ be arbitrary. Then
\[
\exists\, N\in \N\ \forall\,k\ge N:\ x_k=0,
\]
and hence,
\[
\forall\,n\ge N:\ A^nx = 0,
\]
which, implies that
\[
\forall\, \alpha\in (0,1)\ \exists\, c=c(x,\alpha)>0\ \forall\, n\in \N:\ \|A^nx\|_\infty\le c\alpha^n.
\]
By Lemma \ref{lem}, the \textit{Sufficient Condition for Linear Chaos} (Theorem \ref{SCC}), and the \textit{Chaoticity of Powers Corollary} (Corollary \ref{CP}), we conclude that the operator $A$ is \textit{chaotic} as well as every power $A^n$ $(n\in \N)$.
Here, we reproduce the proof of the spectral part of the statement given in \cite{arXiv:1811.06640}.
For arbitrary $\lambda \in \F$ ($\F:=\R$ or $\F:=\C$) and $x:=(x_k)_{\N} \in D(A)$, the equation
\begin{equation}\label{ev}
Ax=\lambda x
\end{equation}
is equivalent to
\[
(w^k x_{k+1})_{k\in \N}=\lambda(x_k)_{k\in \N},
\]
i.e.,
\[
w^k x_{k+1}=\lambda x_k,\ k\in \N
\]
Whence, we recursively infer that
\[
x_k=\left[\prod_{j=1}^{k-1}\frac{\lambda}{w^{k-j}}\right]x_1
=\dfrac{\lambda^{k-1}}{w^{\sum_{j=1}^{k-1}(k-j)}}x_1
=\dfrac{\lambda^{k-1}}{w^{\frac{k(k-1)}{2}}}x_1
=\left(\dfrac{\lambda}{w^{\frac{k}{2}}}\right)^{k-1}x_1,\
k\in \N,
\]
where for $\lambda=0$, $0^0:=1$.
Considering that $|w|>1$, for all sufficiently large $k\in \N$, we have:
\[
\left|\dfrac{\lambda}{w^{\frac{k}{2}}}\right|^{k-1}
=\left(\dfrac{|\lambda|}{|w|^{\frac{k}{2}}}\right)^{k-1}
\le \left(\dfrac{1}{2}\right)^{k-1},
\]
which implies that
\[
y:=(y_k)_{k\in \N}:=\left(\left(\dfrac{\lambda}{w^{\frac{k}{2}}}\right)^{k-1}\right)_{k\in \N}\in c_0.
\]
Further, since
\[
w^ky_{k+1}=w^k\dfrac{\lambda^{k}}{w^{\sum_{j=1}^{k}(k+1-j)}}
=\dfrac{\lambda^{k}}{w^{\sum_{j=2}^{k}(k+1-j)}}=\dfrac{\lambda^{k}}{w^{\frac{k(k-1)}{2}}}x_1
=\left(\dfrac{\lambda}{w^{\frac{k-1}{2}}}\right)^{k},\ k\in \N,
\]
we similarly conclude that
\[
(w^ky_{k+1})_{k\in \N}\in c_0,
\]
and hence,
\[
y\in D(A)\setminus \{0\}.
\]
Thus, we have shown that, for any $\lambda \in \F$, all solutions of equation \eqref{ev} are of the form
\[
x:=(x_k)_{\N}=cy\in D(A),
\]
where $c\in \F$ is arbitrary. They form the \textit{one-dimensional} subspace of $c_0$ spanned by the sequence $y$, which completes the proof.
\end{proof}
\begin{rem}\label{NZ}
Theorem \ref{BLCc0}, Lemma \ref{lem}, and Theorem \ref{ULCc0} naturally extend from $c_0(\N)$ to $c_0(\Z_+)$ for the bounded weighted backward shifts:
\[
c_0(\Z_+)\ni x:=\left(x_k\right)_{k\in \Z_+}\mapsto A_wx:=w\left(x_{k+1}\right)_{k\in \Z_+}\in c_0(\Z_+) \quad (|w|>1)
\]
and the unbounded weighted backward shifts:
\[
A_wx:=\left(w^kx_{k+1}\right)_{k\in \Z_+}\quad (|w|>1)
\]
with maximal domain
\[
D(A_w):=\left\{ x:=\left(x_k\right)_{k\in \Z_+} \in c_0(\Z_+) \,\middle|\, \left(w^kx_{k+1}\right)_{k\in \Z_+}\in c_0(\Z_+) \right\}
\]
and the powers
\[
A_w^{n}x= \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \Z_+},\ n\in \N,
\]
defined on
\[
D(A_w^{n})= \left\{ x := \left(x_k \right)_{k \in \Z_+} \in c_0(\Z_+) \,\middle|\, \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]x_{k+n} \right)_{k \in \Z_+} \in c_0(\Z_+) \right\}
\]
(see the proof of Lemma \ref{lem}).
In the former case, the bounded right inverse of $A_w$ is
\[
c_0(\Z_+)\ni x:=(x_k)_{k\in \Z_+}\mapsto B_w x:=w^{-1}\left(x_{k-1}\right)_{k\in \Z_=}\in c_0(\Z_+)\quad (x_{-1}:=0),
\]
for which $\|B\|={|w|}^{-1}<1$, and, in the latter case, the bonded right inverse of $A_w$ is
\[
c_0(\Z_+)\ni x:=(x_k)_{k\in \Z_+}\mapsto B_wx:=\left(w^{-(k-1)}x_{k-1}\right)_{k\in \Z_+}\in c_0(\Z_+)\quad (x_0:=0),
\]
with
\begin{align*}
c_0(\Z_+)\ni x:=(x_k)_{k\in \Z_+}\mapsto B_w^nx &= \left(\left[ \prod_{j=1}^{n} w^{-(k-j)} \right] x_{k-n} \right)_{k \in \Z_+}\\
&
=\left( w^{-nk + \frac{n(n+1)}{2}}x_{k-n} \right)_{k \in \Z_+},\ n\in \N,
\end{align*}
($x_{k-n}:=0$, $k=0,1,\dots,n-1$), for which
\[
\|B_w^n\|={\left|w\right|}^{-n^2 + \frac{n(n+1)}{2}}={\left|w\right|}^{-\frac{n(n-1)}{2}},
\]
and hence,
\begin{equation*}
\lim_{n\to \infty}{\|B_w^n\|}^{1/n}=\lim_{n\to \infty}{\left|w\right|}^{-\frac{n-1}{2}}=0,
\end{equation*}
i.e., $B_w$ is \textit{quasinilpotent} (cf. the proof of Theorem \ref{ULCc0}).
\end{rem}
\section{Weighted Backward Shifts in $c$}\label{WBSc}
The answer to the natural question of whether one can obtain linear chaos in the space $c(\N)$ of convergent sequences by merely extending the foregoing chaotic weighted backward shifts from the space $c_0(\N)$ is given in the negative by the subsequent statements.
\begin{prop}[Bounded Weighted Backward Shifts on $c$]\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the bounded linear weighted backward shift operator
\begin{equation*
c\ni x:=\left(x_k\right)_{k\in \N}\mapsto A_wx:=w\left(x_{k+1}\right)_{k\in \N}\in c
\end{equation*}
on the space $c$ is not hypercyclic.
\end{prop}
\begin{proof}
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
It is obvious that the operator $A$ is well defined on $c$ and also is linear and bounded with
\[
\|A\|=|w|.
\]
Since, for any $x:=\left(x_{k}\right)_{k \in \N}\in c$,
\begin{equation*}
(A - wI)x = w\left(x_{k+1}\right)_{k \in \N} - w\left(x_k\right)_{k \in \N}
= w\left(x_{k+1} - x_k\right)_{k \in \N}.
\end{equation*}
and
\[
\lim_{k\to \infty}w(x_{k+1} - x_k)=w\left(\lim_{k\to \infty}x_{k+1} - \lim_{k\to \infty}x_k\right)
=w\left(l(x) - l(x)\right)=0,
\]
we infer that
\[
R(A-wI)\subseteq c_0.
\]
Since $c_0$ is a closed proper subspace of $c$, it is nowhere dense in $c$
(see, e.g., \cite{Markin2018EFA,Markin2020EOT}) and, as follows from the prior inclusion,
so is $R(A-wI)$.
Hence,
\[
\overline{R(A_w-wI)}\neq c,
\]
which, by \textit{Bourdon's Theorem} (Theorem \ref{BT}) with $p(\lambda):=\lambda -w$, $\lambda\in \F$ (see also \cite[Proposition $4.1$]{arXiv:2106.14872}), the latter implies that
the operator $A$ is not hypercyclic.
\end{proof}
\begin{prop}[Unbounded Weighted Backward Shifts in $c$]\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the unbounded linear weighted backward shift operator
\[
A_wx:=\left(w^kx_{k+1}\right)_{k\in \N}
\]
in the space $c$ with maximal domain
\[
D(A_w):=\left\{ x:=\left(x_k\right)_{k\in \N} \in c \,\middle|\, \left(w^kx_{k+1}\right)_{k\in \N}\in c\right\}
\]
is not hypercyclic.
\end{prop}
\begin{proof}
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
As follows from the definition, for any $x:=\left(x_k\right)_{k\in \N}\in D(A)$,
\[
y:=\left(y_k:=w^kx_{k+1}\right)_{k\in \N}\in c
\]
and hence, in view of $|w|>1$,
\[
x_{k+1}=w^{-k}y_k\to 0,\ k\to \infty.
\]
Therefore,
\[
C^\infty (A) := \bigcap_{n=1}^\infty D(A^n)\subseteq D(A)\subseteq c_0.
\]
Since $c_0$ is a closed proper subspace of $c$, it is nowhere dense in $c$
(see, e.g., \cite{Markin2018EFA,Markin2020EOT}) and, as follows from the prior inclusion,
so is $C^\infty (A)$.
Hence,
\[
\overline{C^\infty (A)}\neq c,
\]
which immediately implies that the operator $A$ is not hypercyclic (see Remarks \ref{HCrems}).
\end{proof}
\section{Linear Chaos in $c$}
With the hypercyclicity by extension compromised, here, we construct bounded and unbounded chaotic linear operators in $c(\N)$ based on the chaotic backward shifts in $c_0(\Z_+)$ via establishing a homeomorphic isomorphism between the two spaces (i.e., an isomorphism which is also a homeomorphism).
\begin{lem}[Homeomorphic Isomorphism]\label{lem1}\ \\ \noindent
The mapping
\[
c(\N) \ni x:= (x_k)_{k \in \N} \mapsto Jx:= (y_k)_{k \in \Z_+} \in c_0(\Z_+),
\]
assigning to each $x:= (x_k)_{k \in \N}\in c(\N)$ the sequence $\left(y_k\right)_{k\in \Z_+}\in c_0(\Z_+)$ of the coordinates of $x$ relative to the standard Schauder basis
$\left\{e_n\right\}_{n\in \Z_+}$ for $c(\N)$, where
\[
e_0 := (1,1,1,\dots)\quad \text{and}\quad e_n:=\left(\delta_{nk}\right)_{k\in \N},\ n\in\N,
\]
i.e.,
\[
y_0:=l(x)\quad \text{and}\quad y_k:=x_k-l(x),\ k\in \N,
\]
where $l$ is the limit functional, is a homeomorphic isomorphism from $c(\N)$ to $c_0(\Z_+)$.
\end{lem}
\begin{proof}
In view of the uniqueness of the Schauder expansion, we infer that the mapping $J$ is \textit{linear} and further, since, for $x\in c(\N)$,
\[
Jx = 0 \iff y_k = 0,\ k \in \Z_+ \iff x = \sum^\infty_{k=0}y_k e_k = 0 \in c(\N),
\]
$J$ is also \textit{injective} (see, e.g., \cite{Markin2018EFA,Markin2020EOT,MarkSogh2021}).
Further, for any $(y_k)_{k \in \Z_+} \in c_0(\Z_+)$, let
\[
x:=\left( y_k+y_0\right)_{k\in \N}.
\]
Since
\[
\lim_{k\to \infty}y_k=0,
\]
we infer that
\[
\lim_{k\to \infty}x_k=\lim_{k\to \infty}(y_k+y_0)=y_0,
\]
Thus,
\[
x\in c(\N)\quad \text{and}\quad Jx=y,
\]
which implies that the mapping $J$ is also \textit{surjective}, and hence $J:c(\N)\to c_0(\Z_+)$ is an \textit{isomorphism} between the spaces $c(\N)$ and $c_0(\Z_+)$.
Since, for an arbitrary $x:=\left(x_k\right)_{k\in \N}\in c(\N)$,
\[
|l(x)|=\left|\lim_{k\to \infty}x_k\right|=\lim_{k\to \infty}\left|x_k\right|\le \sup_{k\in \N}|x_k|=:\|x\|_\infty,
\]
we also have:
\begin{equation*}
\| Jx \|_\infty := \sup_{k \in \Z+} \left|y_k\right|
=\max\left[|l(x)|,\sup_{k \in \N} \left|x_k - l(x)\right|\right]
\leq 2 \| x \|_\infty.
\end{equation*}
Thus, the linear mapping $J$ is \textit{bounded}, and hence \textit{continuous}, which, by the \textit{Inverse Mapping Theorem} (see, e.g., \cite{Markin2018EFA,Markin2020EOT}), implies that so is its inverse $J^{-1}:c_0(\Z_+)\to c(\N)$:
\begin{equation}\label{JI}
c_0(\Z_+) \ni x:= (y_k)_{k \in \Z_+} \mapsto J^{-1}x:= (y_k+y_0)_{k \in \N} \in c(\N).
\end{equation}
We conclude that the mapping $J:c(\N)\to c_0(\Z_+)$ is both \textit{isomorphic} and \textit{homeomorphic}.
\end{proof}
\begin{thm}[Bounded Linear Chaos in $c$]\label{BLC}\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the bounded linear operator
\[
c\ni x:=\left(x_k\right)_{k\in \N}\mapsto \hat{A}_wx:=w\left(x_{k+1}+x_1-2l(x)\right)_{k\in \N}\in c
\]
on the space $c$ is chaotic as well every power
\[
c\ni x:=\left(x_k\right)_{k\in \N}\mapsto {\hat{A}}^nx = w^n \left( x_{k+n} + x_n -2l(x) \right)_{k \in \N},\ n \in \N,
\]
and, provided the underlying space is complex
\[
\sigma\left(\hat{A}_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|\le |w| \right\}
\]
with
\[
\sigma_p\left(\hat{A}_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|<|w| \right\}\quad \text{and}\quad
\sigma_c\left(\hat{A}_w\right)=\left\{ \lambda\in \C \,\middle|\, |\lambda|=|w| \right\}.
\]
\end{thm}
\begin{proof}
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
On $c(\N)$, consider the linear operator $\hat{A}$ defined as follows:
\begin{equation}\label{sim}
\hat{A}:=J^{-1}AJ,
\end{equation}
i.e., via the commutative diagram
\begin{equation*}
\begin{tikzcd}[sep=large]
c_0(\Z_+) \arrow[r, "A"] & c_0(\Z_+) \\
c(\N)\arrow[u, "J"] \arrow[r, "\hat{A}"]& c(\N)\arrow[u, "J"]
\end{tikzcd}.
\end{equation*}
Since, by \eqref{sim},
\[
{\hat{A}}^n:=J^{-1}A^nJ,\ n\ \in \N,
\]
where
\[
c_0(\Z_+)\ni y:=\left(y_k\right)_{n\in \Z_+} \mapsto A^ny:=w^n\left(y_{k+n}\right)_{k\in \Z_+}\in c_0(\Z_+)
\]
(see Remark \ref{NZ}), for any $x:=\left(x_k\right)_{k\in \N} \in c(\N)$,
\[
A^nJx=w^n\left(x_{k+n}-l(x)\right)_{k\in \Z_+}=:\left(y_k\right)_{k\in \Z_+},
\]
and hence, in view of \eqref{JI}, we have:
\[
{\hat{A}}^nx:=J^{-1}A^nJ=\left(y_k+y_0\right)_{k\in \N}
=w^n\left(x_{k+n}+x_n-2l(x)\right)_{k\in \N}\in c(\N).
\]
Observe that
\[
\lim_{k\to \infty}w^n\left(x_{k+n}+x_n-2l(x)\right)=w^n\left(l(x)+x_n-2l(x)\right)
=w^n\left(x_n-l(x)\right),\ n\in \N.
\]
Since, by Lemma \ref{lem1}, $J:c(\N)\to c_0(\Z_+)$ is a \textit{homeomorphic isomorphism}, the operator ${\hat{A}}^n$ ($n\in \N$) inherits its \textit{linearity}, \textit{boundedness}, \textit{chaoticity}, and \textit{spectral structure} directly from its conjugate $A^n$ via $J$. Therefore, the statement follows immediately from Theorem \ref{BLCc0}.
\end{proof}
\begin{thm}[Unbounded Linear Chaos in $c$]\label{ULC}\ \\
For an arbitrary $w\in \F$ with $|w|>1$, the linear operator
\[
\hat{A}_wx:=\left(w^k(x_{k+1}-l(x))+x_1-l(x)\right)_{k\in \N}
\]
in the space $c$ with domain
\[
D(\hat{A}_w):=\left\{ x:=\left(x_k\right)_{k\in \N} \in c \,\middle|\, \left(w^k\left(x_{k+1}-l(x)\right)\right)_{k\in \N}\in c_0 \right\}
\]
is unbounded, closed, and chaotic as well every power
\[
\hat{A}_w^nx=\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right)
+\left[ \prod_{j=0}^{n-1} w^j \right]\left(x_{n}-l(x)\right) \right)_{k \in \N},\ n\in \N,
\]
with domain
\[
D({\hat{A}}^n)=\left\{ x:=\left(x_k\right)_{k\in \N} \in c \,\middle|\,
\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right) \right)_{k \in \N}\in c_0 \right\}
\]
Furthermore, each $\lambda \in \F$ is an eigenvalue for $\hat{A}_w$ of geometric multiplicity $n$., i.e.,
\[
\dim\ker(\hat{A}_w-\lambda I)=1.
\]
\end{thm}
\begin{proof}
Let $w\in \F$ with $|w|>1$ be arbitrary and, for the simplicity of notation, let $A:=A_w$.
In $c(\N)$, consider the linear operator $\hat{A}$ defined as follows:
\begin{equation}\label{sim1}
\hat{A}:=J^{-1}AJ,
\end{equation}
i.e., via the commutative diagram
\begin{equation*}
\begin{tikzcd}[sep=large]
c_0(\Z_+) &\hspace{-1.1cm} \supseteq D(A) \arrow[r, "A"] & c_0(\Z_+) \\
c(\N)& \hspace{-1.4cm} \supseteq D(\hat{A})\arrow[u, "J"] \arrow[r, "\hat{A}"]& c(\N)\arrow[u, "J"]
\end{tikzcd},
\end{equation*}
for which
\[
J\left(D(\hat{A})\right):=D(A).
\]
Since, by \eqref{sim1},
\[
{\hat{A}}^n:=J^{-1}A^nJ,\ n\ \in \N,
\]
where
\[
A^{n}y= \left( \left[ \prod_{j=k}^{k+n-1} w^j \right]y_{k+n} \right)_{k \in \Z_+}
\]
with domain
\[
D(A_w^{n})= \left\{ y := \left(y_k \right)_{k \in \Z_+} \in c_0(\Z_+) \,\middle|\,
\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]y_{k+n} \right)_{k \in \Z_+} \in c_0(\Z_+) \right\}
\]
(see Remark \ref{NZ}), we have:
\[
D({\hat{A}}^n)=\left\{ x:=\left(x_k\right)_{k\in \N} \in c(\N) \,\middle|\,
\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right) \right)_{k \in \N}\in c_0(\N) \right\}
\]
and, considering that, for any $x:=\left(x_k\right)_{k\in \N} \in D({\hat{A}}^n)$,
\[
A^nJx=\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right) \right)_{k \in \Z_+}=:\left(y_k\right)_{k\in \Z_+},
\]
in view of \eqref{JI},
\begin{align*}
{\hat{A}}^nx&=J^{-1}A^nJ=\left(y_k+y_0\right)_{k\in \N}
\\
&=\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right)
+\left[ \prod_{j=0}^{n-1} w^j \right]\left(x_{n}-l(x)\right) \right)_{k \in \N}\in c(\N).
\end{align*}
Observe that
\begin{align*}
&\lim_{k\to \infty}\left( \left[ \prod_{j=k}^{k+n-1} w^j \right]\left(x_{k+n}-l(x)\right)
+\left[ \prod_{j=0}^{n-1} w^j \right]\left(x_{n}-l(x)\right) \right)\\
&=\left[ \prod_{j=0}^{n-1} w^j \right]\left(x_{n}-l(x)\right).
\end{align*}
Since, by Lemma \ref{lem1}, $J:c(\N)\to c_0(\Z_+)$ is a \textit{homeomorphic isomorphism}, the operator ${\hat{A}}^n$ ($n\in \N$) inherits its \textit{linearity}, \textit{unboundedness}, \textit{closedness}, \textit{chaoticity}, and \textit{eigenvalues} along with their \textit{geometric multiplicities} directly from its conjugate $A^n$ via $J$. Therefore, the statement follows immediately from Theorem \ref{ULCc0}.
\end{proof}
| f06d9995727c1ed20bf8984dee81c8d547bee10b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Goldman \cite{MR846929} and Turaev \cite{MR1142906} defined a Lie bialgebra structure on the $\mathbb Z$-module generated by
the free homotopy classes of loops of an oriented surface $M$ (i.e. the conjugacy classes of $\pi_1(M)$). The
bracket is defined by
\begin{equation}\label{goldman}
[X ,Y ]_{\pi_1(M)}=\sum_{p\in \alpha\cap \beta}\epsilon(p; \alpha, \beta) \{\alpha_p \beta_p\}
\end{equation}
where $\alpha$ and $\beta$ are representatives of $X$ and $Y$
respectively, intersecting at most at transversal double points, the
number $\epsilon(p; \alpha, \beta)=\pm 1$ denotes the oriented
intersection number of $\alpha$ and $\beta$ at $p$, and $ \{\alpha_p \beta_p\} $ is the conjugacy class of the element $\alpha_p\cdot \beta_p\in
\pi_1(M, p)$ where $\alpha_p$ and $\beta_p$ are the elements
of $\pi_1(M, p)$ that correspond to reparametrize $\alpha$ and $\beta$ to start (and end) at $p$.
Turaev showed in \cite{MR1142906} that there is a colagebra
structure that gives rise to a Lie bialgebra. Chas in
\cite{MR2041630} proposed a combinatorial model for this bialgebra
structure. The aim of this paper is to develop a
generalization of the Goldman-Turaev construction replacing
homotopies by thin homotopies.
Let $M$ be an oriented surface endowed with any Riemannian metric. We denote by $\mathcal E(M)$ the set of classes of piecewise geodesic curves in $M$ modulo endpoint-preserving {\em thin homotopies}, which we shall define with precision in section \ref{s.prelim}. For each $x\in M$ we define ${\mathcal L}_x(M)$ as the elements of $\mathcal E(M)$ that start and end at $x$. It will be easy to notice that this is a group under concatenation, and $\pi_1(M,x)$ is a quotient of it.
Let $S(M)$, the {\em space of strings} of $M$, be the set of conjugacy classes of
${\mathcal L}_x(M)$ in $\mathcal E(M)$, i.e. $g, h\in \mathcal L_x(M)$ are conjugate if there is $p\in \mathcal E(M)$ such that $p^{-1}gp=h$. Let $\mathcal S(M)$ be the free abelian group generated by $S(M)$. In this paper we shall define a bracket $[ \; , \; ]$ on
$\mathcal S(M)$ following the lines of Chas in \cite{MR2041630}. Then we shall show
\begin{theorem}\label{jacobi}
$(\mathcal S(M), [\;,\;])$ is a Lie algebra.
\end{theorem}
It is also possible to give a coalgebra structure and show that $\mathcal S(M)$ is a Lie bi-algebra using the same techniques, though we will not present this construction explicitly. We will see that the Goldman-Turaev structure on $M$ is the quotient of $\mathcal S(M)$ obtained by taking regular (non-thin) homotpies. This fact will emerge naturally from our construction of the bracket, and the case for the co-bracket is analogous. The Goldman algebra is relevant in the study of spaces of representations of the fundamental groups of $2$-manifolds, which in turn, can be regarded as moduli spaces of flat connections on orientable $2$-manifolds (\cite{jeffrey2005flat}). Our construction may also play a role in the study of the space of all connections but this aspect is not going to be treated here.
Chas and Krongold obtained an algebraic characterization of homotopy classes of simple curves in terms of the bracket (\cite{chas2010algebraic}) and the cobracket (\cite{chas2016algebraic}) .
(See also \cite{chas2016extended} and \cite{chas2021lie}). In the present paper we give a
geometric proof of a different characterization of simple curves in terms of the bracket, which was conjectured by Chas in \cite{MR2041630}. For other related results see \cite{chas2010self}, \cite{chas2012self} \cite{chas2015goldman} \cite{chas2010minimal}, \cite{chas2019almost},
\cite{chas2015relations}.
A string $X$ is {\em primitive} if every representative $g\in \mathcal L_x(M)$ (for any $x\in M$) of $X$ is primitive in the sense of group theory: if there is no $h\in \mathcal L_x(M)$ such that $g=h^n$ with $n>1$. A piecewise geodesic closed curve will be called {\em simple} if it has no stable self-intersections, i.e. if there is a small perturbation that has no self-intersections (namely, that is simple in the usual sense).
We shall show the following theorem:
\begin{theorem}\label{simple}
Let $M$ be an oriented surface. A primitive $X\in S(M)$ has a simple
representative if and only if $[X , X^{-1}]=0$.
\end{theorem}
The geometric group theory approach of our proof of Theorem \ref{simple} may have some independent interest beyond the result itself. The aforementioned proof includes some tools, namely, the notion of $\alpha$-oriented subgroups (see section \ref{alpha_oriented}), which could be useful in other contexts.
Theorem \ref{simple} will allow us to prove a conjecture posed by Chas (\cite{MR2041630}):
\begin{corollary}
Let $M$ be a compact oriented surface with boundary. A primitive free
homotopy class $X$ of $M$ has a simple representative if and only if
$[X , X^{-1}]_{\pi_1(M)}=0$.
\end{corollary}
\begin{proof}
Let $p:\tilde M\to M$ be the universal covering of $M$. Let
$D\subset \tilde M$ be a fundamental polygon as in
\cite{MR2041630}. Since $M$ has boundary we can choose on $\tilde M$ a metric of constant non positive curvature.
The fundamental group is freely generated by the
set $T$ of those $g\in \pi_1(M)$ such that $gD\cap D\neq \emptyset$. Let
$\mathcal L_{x,0}(M)\subset \mathcal L_x(M)$ be the subgroup of $\mathcal L_x(M)$
generated by those elements of the form $h_g=[p\circ\alpha_g]$ where
$\alpha_g$ is a geodesic (corresponding to the chosen metric) joining $x$ to $gx$ for each $g\in T$. Denote by $S_0(M)$ the set of strings corresponding to the loops in $\mathcal L_{x,0}(M)$. Note that
$\pi_1(M)$ is isomorphic to $\mathcal L_{x, 0}(M)$ via the standard quotient, i.e. taking non-thin homotopies. As we mentioned below Theorem \ref{jacobi}, this quotient is a Lie algebra homomorphism, thus it gives an isomorphism between the subalgebra of $\mathcal S(M)$ generated by $S_0(M)$ and the Goldman-Turaev algebra on the free homotopy classes of $M$. Then we apply Theorem \ref{simple} to strings in $S_0(M)$ to conclude this proof.
\end{proof}
\section{The loop bundle}
\subsection{Definitions of thin homotopies and the spaces of loops} \label{s.prelim}
Let $I$ be the unit interval and $M$ a Riemannian manifold. We begin by recalling some standard notations. A {\em path} in $M$ is a continuous function from $I$ to $M$, and we say that two paths $a,b:I \to M$ are equivalent modulo
reparametrization if there is an orientation preserving homeomorphism $\sigma:I\to I$ such that $a\circ \sigma=b$.
Denote by $\Omega_0$ the quotient set under this
equivalence relation. If $a(1)=b(0)$ we define $ab$ and $a^{-1}$ as
follows: $ab(t)=a(2t)$ if $t\in [0,1/2]$ and $ab(t)=b(2t-1)$ if
$t\in [1/2,1]$; $a^{-1}(t) = a(1-t)$ for every $t\in [0,1]$. Let
$e_x \in \Omega_0$ be the contant path at $x$, i.e. $e_x(t) =x$ for every $t\in [0,1]$.
In order to define what we call {\em thin homotpy} between piecewise geodesic paths we need to consider another preliminary equivalence, which amounts to collapse constant sub-paths. Let $a$ be a non-constant path in $M$. We shall define a {\em minimal form} $a_r$ for $a$ as follows: let $I_i \subset I$ be the maximal subintervals in which
$a$ is constant, and let $\sigma: I \to I$ be a surjective non-decreasing continuous function, constant in each $I_i$ and strictly
increasing in $I - \bigcup_i I_i$. Then there is $a_r: I \to M$ such that $a =a_r \circ\sigma$, which is non-constant on any subinterval of $I$ (this map is obtained by a universal property of quotients). Different choices of the function $\sigma$ give rise to minimal forms that are equivalent modulo reparametrization, and moreover, if two paths $a$ and $b$ are equivalent, so are any of their minimal forms $a_r$ and $b_r$. This allows us to define the {\em minimal class} of an element of $\Omega_0$ (as the class of any minimal form of any representative), and take a quotient $\Omega_1$ where we identify two elements of $\Omega_0$ if they have the same minimal class (extending the definition to constant paths in the trivial way). The product and inverse are well defined on $\Omega_1$, and the classes of constant paths are units for the product.
Let $\Omega\subset \Omega_1$ be the set of classes of either constant paths or paths that are {\em piecewise geodesic}, i.e. a finite concatenation of geodesic segments.
Notice that for $\alpha\in \Omega$ there are well defined notions of endpoints $\alpha(0)$ and $\alpha(1)$, of image $\alpha(I)$, and of length $l(\alpha)$. Throughout the paper we will refer to the elements $\alpha \in \Omega$ as {\em curves}, and say that $\alpha$ is a {\em closed curve} if $\alpha(0)=\alpha(1)$.
In the set $\Omega$ we consider the equivalence
relation generated by the identifications $\alpha aa^{-1}\beta\sim
\alpha\beta$. This is what we call equivalence under {\em thin homotopies}. With the formal definition in hand, we recall the concepts from the introduction: Let $\mathcal E(M)$ denote the quotient set of $\Omega$ under thin homotopies, and
let ${\mathcal L}_x(M)$ be the projection onto $\mathcal E(M)$ of the set of closed curves starting and ending at $x$. Note that
${\mathcal L}_x(M)$ is a group under concatenation
whose identity element, $id_x$, is the equivalence class of $e_x$,
the constant path at $x$.
\subsection{Reductions and basic properties}
A {\em reduction} for $\alpha\in\Omega$ is a factorization of the form $\alpha=acc^{-1}d$ with nontrivial $c$. We say that $\alpha$ is {\em reduced} if it admits no such reduction. Since the curves in $\Omega$ are classes of piecewise geodesic (or constant) paths, it is easy to show that every element of $\mathcal E(M)$ has a unique reduced representative in $\Omega$ (though the proof of uniqueness may be a bit cumbersome). The {\em reduced form} of $\alpha\in\Omega$ is the unique reduced curve that is equivalent to $\alpha$ under thin homotopy.
Next we point out some basic facts that will be used without explicit reference throughout the article. Firstly notice that a curve $\gamma\in\Omega$ satisfies $\gamma=\gamma^{-1}$ only when it is of the form $\gamma=cc^{-1}$, thus for a reduced curve $\gamma$ it happens only when $\gamma$ is constant. Next we see that the concept of length in $\Omega$ satisfies the expected properties, namely that:
\begin{itemize}
\item $l(ab)=l(a)+l(b)$, and
\item if $ab=cd$ with $l(a)=l(c)$, then $a=c$ and $b=d$.
\end{itemize}
For $\gamma,\,\delta\in \Omega$ we shall write $\gamma\subset\delta$ if we have $\delta=a\gamma b$ for $a,b\in\Omega$. In case $a$ is trivial we say that $\gamma$ is an {\em initial segment} of $\delta$, and if $b$ is trivial that $\gamma$ is a {\em final segment} of $\delta$. Note that if $\delta$ is reduced, so must be $\gamma$.
We say that two curves $\gamma$ and $\delta$ {\em overlap} if an initial segment of one of them agrees with a final segment of the other, i.e. if we can write either $\gamma=ab$, $\delta=bc$ with $b$ non-constant, or $\gamma=ab$, $\delta=ca$ with $a$ non-constant. Note that if $\gamma$ is reduced and non constant, then $\gamma$ and $\gamma^{-1}$ cannot overlap: for instance, if $\gamma=ab$ and $\gamma^{-1}=bc$, we get $b=b^{-1}$ where $b$ is reduced, so $b$ must be constant.
\subsection{Definitions of loop bundle and horizontal lift}
Consider the space $\mathcal E(M)$ defined in the previous section, and let $[\alpha]\in \mathcal E(M)$ stand for the equivalence class of $\alpha\in \Omega$. Let
$\mathcal E_x(M)$ be the set of the $[\alpha]\in \mathcal E(M)$ such that $\alpha(0)=x$; define $\pi:\mathcal E_x(M)\to M$ by
$\pi([\alpha])=\alpha(1)$, and observe that $\mathcal L_x(M)=\pi^{-1}(x)$. The group $\mathcal L_x(M)$ acts on $\mathcal E_x(M)$ by left multiplication and for all $[\alpha]\in \mathcal L_x(M)$ and $[\gamma]\in
\mathcal E_x(M)$ we have $\pi\big([\alpha][\gamma]\big)=
\pi([\gamma])$; hence the quadruple $(\mathcal E_x(M) , \mathcal
L_x(M), M,\pi) $ is a principal fiber bundle over $M$, with
structure group $\mathcal L_x(M)$.
Let $\gamma$ be a path in $M$, and take $p\in\mathcal E_x(M)$ with
$\pi(p)=\gamma(0)$. We define the {\em horizontal lift} of $\gamma$
at $p$ to be the path $\tilde{\gamma}$ in $\mathcal E_x(M)$ which
is obtained in the following way. Take $\beta$ any representative of
$p$ (i.e. $p=[\beta]$), and for each $s\in I$ set $\gamma_s$ to be
the path in $M$ defined by $\gamma_s(t)=\gamma(st)$. Then
$\tilde{\gamma}(s)=[\beta\gamma_s]$. This horizontal lift can be seen as a topological
connection in the bundle $(\mathcal E_x (M), \mathcal L_x(M),
M,\pi)$.
We will say that a path in $\mathcal
E_x(M)$ is {\em horizontal} if it can be obtained by horizontal lift (see
\cite{MR0077122}, \cite{MR0124068}). Note that the concept of horizontal lift is well defined at the level of curves (i.e. in $\Omega_1$), thus we may speak of {\em horizontal curves}.
We define the {\em length of an horizontal curve} as the length of the projection. Observe that the action of $\mathcal L_x(M)$ preserves the set of horizontal curves, as well as their lengths (by definition). We should clarify that we are not giving a metric on $\mathcal E_x(M)$.
\subsection{Conjugacy classes in ${\mathcal L}_x$ and the space of strings.}
Recall that the {\em space of strings} $S(M)$ is the set of conjugacy classes of
${\mathcal L}_x(M)$ in $\mathcal E(M)$, i.e. $g, h\in \mathcal L_x(M)$ are conjugate if there is $p\in \mathcal E(M)$ such that $p^{-1}gp=h$. This does not depend on $x$ because of the following remark.
\begin{remark}\label{bpoint} (Change of basepoint)
If $x,y\in M$ and $\gamma_0:I\to M$ has
$\gamma_0(0)=x$ and $\gamma_0(1)=y$, let $p_0=[\gamma_0]$ and define
the maps $\psi: \mathcal E_x(M) \to \mathcal E_y(M)$ by
$\psi(p)=p_0^{-1}p$ and $\phi: \mathcal L_x(M) \to \mathcal L_y(M)$ by $\phi(g)=p_0^{-1}gp_0$. Then $\phi$ is an isomorphism of groups, and $(\psi,\phi)$ is an isomorphism of fiber bundles over $M$, commuting with the horizontal lift.
\end{remark}
We say that a closed curve $\alpha$ is {\em cyclically reduced}
if $\alpha$ is reduced and it cannot be factorized as $cac^{-1}$
with non trivial $c$. For a cyclically reduced curve $\gamma$, we say that $\beta\in\Omega$ is a {\em permutation} (or {\em cyclical permutation}) of $\gamma$ if there are $r,s\in\Omega$ such that $\gamma=rs$ and $\beta=sr$. If $s$ and $r$ are non constant we say that $\beta$ is a {\em non trivial permutation} of $\gamma$. Note that permutation is an equivalence relation among the cyclically reduced curves in $\Omega$.
For a string $X\in S(M)$ we can find $x\in M$ and a cyclically
reduced curve $\alpha$ based at $x$, such that $X$ is the conjugacy class of $[\alpha]\in\mathcal L_x(M)$. On the other hand, permutation agrees with conjugacy in $\mathcal E_x(M)$ among cyclically reduced curves, therefore we have
\begin{remark}
There is a bijection between $S(M)$
and the permutation classes of cyclically reduced curves.
\end{remark}
Throughout the paper, when we take representatives of strings we will always assume them to be cyclically reduced. If $X\in S(M)$ and $\alpha$ is a representative of it, we define $X^{-1}$ as the permutation class of $\alpha^{-1}$. It is straightforward to check that if $\alpha$ is non-constant then $\alpha^{-1}$ is not a permutation of $\alpha$. Thus $X\neq X^{-1}$ unless $X$ is trivial.
A cyclically reduced curve $\alpha$ is {\em primitive} if there is no $\gamma\in\Omega$ such that $\alpha=\gamma^n$ with $n>1$. The following easy result is well known.
\begin{lemma}\label{primitive}
Let $\alpha$ be a cyclically reduced curve. Then $\alpha$ is not primitive if and only if $\alpha$ has a non trivial permutation $\hat\alpha$ such that
$\alpha=\hat \alpha$.
\end{lemma}
Note that a string $X\in S(M)$ is primitive, as defined in the introduction, if it has a cyclically reduced representative that is primitive.
\section{Lie bialgebra structure}
\subsection{Linked pairs} \label{LPs}
In order to define the bracket in $\mathcal S(M)$ we need a way of encoding the intersections of curves in $\Omega$ that are stable under local homotopy. We do this by adapting the notion of {\em linked pairs} from Chas \cite{MR2041630} to our context.
Let $\alpha_1, \alpha_2$ and $\gamma$ in $\Omega$ be classes of geodesic segments contained in a normal ball such that $\alpha_1(1)=\alpha_2(0)=:y$ and either $\gamma(0)=y$ or $\gamma(1)=y$. Assume that $\alpha=\alpha_1\alpha_2$ is reduced and $\gamma$ only meets $\alpha$ at $y$. Take $\rho>0$ small enough so that $B(y, \rho)$ is a normal ball and $\alpha_1$, $\alpha_2$ and $\gamma$ are not contained in it, and let $z_1, z_2, z$ be the intersection points of
$\alpha_1$, $\alpha_2$ and $\gamma$ with $\partial B(y, \rho)$ respectively. Since $M$ is oriented, the orientation of $B(y, \rho)$ induces an orientation of $\partial B(y, \rho)\cong S^1$, which is equivalent to giving a circular order on $\partial B(y, \rho)$.
We write $sign(\alpha, \gamma)=1$ if either $\gamma(0)=y$ and the order of the sequence $z_2, z, z_1$ coincides with the circular order of $\partial B(y, \rho)$, or $\gamma(1)=y$
and the order of the sequence $z_2, z_1, z$ coincides with the circular order of $\partial B(y, \rho)$. Otherwise we write $sign(\alpha, \gamma)=-1$.
Notice that this sign does not depend on the choice of $\rho$.
Informally one could say that $sign(\alpha, \gamma)=1$ if $\gamma$ is either outgoing at the ``left'' side of $\alpha$ or incoming at the ``right'' side of $\alpha$, while $sign(\alpha, \gamma)=-1$ if one of the reverse situations happens.
Since elements of $\Omega$ are piecewise geodesic curves, the intersections between two elements are either transversal or along an interval. Taking this into account, we discuss the general forms of these intersections and indicate which ones will constitute linked pairs. A factorization of a curve $\alpha\in \Omega$ is a sequence $(\alpha_1,\ldots,
\alpha_n)$ such that $\alpha=\alpha_1\cdots\alpha_n$ where
$\alpha_i\in \Omega$.
\begin{definition} \label{def:lp}
Consider the following factorizations (of some curves)
\[ A=(a, \eta, b) \]
\[B= (c, \xi, d) \]
where $a,b,c,d$ are geodesics contained in normal balls. We say that
$(A, B)$ is a {\em linked pair} if any of the following conditions hold
\begin{enumerate}
\item
$\eta=\xi=\mbox{point}$, $d$ meets $ab$ only at $d(0)$ and $c$ meets $ab$ only at $c(1)$, and
\[sign(ab, d)=sign(ab, c)\]
\item $\eta=\xi$ (non constant), if we factorize $\eta=\gamma_1 \eta_1\gamma_2$ such that $\gamma_1$ and $\gamma_2$ are contained in normal balls, we have that $d$ meets $\gamma_2 b$ only at $d(0)$ and $c$ meets $a\gamma_1$ only at $c(1)$, and
\[sign(\gamma_2b, d)= sign(a\gamma_1, c) \]
\item
$\eta=\xi^{-1}$ (non constant), if we factorize $\eta=\gamma_1 \eta_1\gamma_2$ such that $\gamma_1$ and $\gamma_2$ are contained in normal balls, we have that $c$ meets $\gamma_2 b$ only at $c(1)$ and $d$ meets
$a\gamma_1$ only at $d(0)$, and
\[sign(a\gamma_1, d)= sign(\gamma_2b, c) \]
\end{enumerate}
\end{definition}
We define the sign of the linked pair as follows: In case (1) we set $sign(A, B)=sign(ab, d)$, in case (2) we set $sign(A, B)=sign(\gamma_2b, d)$ and in case (3) set $sign(A, B)=sign(a\gamma_1, d)$.
If all but the orientation (sign) conditions hold we say that $(A,B)$ is an {\em intersection pair}. Notice that the intersections between two cyclically reduced curves in $\Omega$ can locally be written in the form of intersection pairs. The orientation conditions say that an intersection pair $(A,B)$ is a linked pair exactly when the intersection between the underlying curves is stable under small perturbations. We shall refer to linked pairs of {\em type} (1), (2) or (3) according to which one of the conditions they satisfy in Definition \ref{def:lp}, and we do the same for intersection pairs.
Next we turn to the intersections of cyclically reduced curves in $\Omega$ in a global sense, i.e. in a way that takes account of multiplicities.
If $\alpha$ is a closed curve, we say that $P=(\xi,\eta)$ is a {\em cyclic factorization} of $\alpha$ if either $\alpha=\xi_1 \eta \xi_2$ with $\xi = \xi_2 \xi_1$ or $\alpha = \eta_1 \xi \eta_2$ with $\eta = \eta_2 \eta_1$. There is a slight abuse of notation here, as the decompositions $\xi = \xi_2 \xi_1$ or $\eta = \eta_2 \eta_1$ are needed for recovering $\alpha$, and are indeed intended as part of the definition, though we drop them from the notation to make it less cumbersome. If $\alpha$ is cyclically reduced and $\beta$ is a permutation of it, then there is a bijection between the cyclic factorizations of $\alpha$ and those of $\beta$. Notice though, that in order to talk about cyclic factorizations of a string, we need to choose a curve representative first. The reason for this choice of definition is that we want to keep track of the position of the sub-curves ($\xi$ and $\eta$) with respect to a chosen parameter-basepoint (which is well defined in $\Omega$). This detail will make a difference for strings that are not primitive.
\begin{definition} \label{def:lp cyclic}
Let $\alpha,\beta\in \Omega$ be cyclically reduced closed curves. A {\em linked pair between
$\alpha$ and $\beta$} is a pair $(P, Q)$ of cyclic factorizations
$P=(\alpha_1,\eta)$ and $Q=(\beta_1, \xi)$ of $\alpha$ and $\beta$ respectively, such that if we write
\begin{itemize}
\item $\alpha_1=b \hat
\alpha_1 a$ and $\beta_1= d \hat \beta_1 c$, where $a,b,c,d$ are geodesics contained in normal balls, and
\item $A=(a,\eta, b)$ and $B=(c,\xi, d)$,
\end{itemize}
then $(A,B)$ is a linked pair.
\end{definition}
Notice that the concatenations $a\eta b$ and $c\xi d$ are well defined, so the above definition makes sense. Moreover, they are sub-curves of some permutations of $\alpha$ and $\beta$ respectively, thus saying that $(A,B)$ is a linked pair means that there is a stable intersection between $\alpha$ and $\beta$, or the strings they represent. Defining $P$ and $Q$ as cyclic factorizations keeps track of the position of the intersection segments relative to the parameter-basepoints of $\alpha$ and $\beta$, so intersections that repeat count as different linked pairs. This amounts to counting multiplicity, just as is usual in differential topology for the intersection between transversal smooth paths. Notice also that taking a permutation of $\alpha$ or $\beta$ induces a natural bijection between the sets of linked pairs.
We define the {\em length} of a linked pair as $l(P,Q)=l(\eta)=l(\xi)$, i.e. as the length of the intersection segment. The {\em type} of $(P,Q)$ shall be the type of $(A,B)$ in Definition \ref{def:lp cyclic}.
\subsection{Definition of the string bracket}
In this section we define the bracket following closely the
presentation in \cite{MR2041630}. Since the definition of the
co-bracket involves no new ideas we omit it. Recall that $S(M)$,
the space of strings, is the set of conjugacy classes of ${\mathcal
L}_x(M)$. Also recall that $\mathcal S(M)$ is the free abelian group generated by
$S(M)$, in which we shall define the bracket.
For $X\in S(M)$ and an integer $n>0$, let $X^n$ be the conjugacy class of $[\alpha]^n$, where $[\alpha]$ represents $X$. Define also $l(X)=l(\alpha)$ where $\alpha$ is a cyclically reduced representative of $X$, noting that different choices of such representative have the same length. Since $\alpha$ is cyclically reduced, we have that $l(X^n)=nl(X)$.
Although we will not present the definition of the co-bracket, we give the main definition in which it is based, for the sake of completeness.
\begin{definition} \label{lp1} Let $X$ be a string. We define $LP_1(X)$, the set of linked pairs
of $X$, as the set of linked pairs of any two representatives of $X$.
\end{definition}
By the discussion at the end of the previous section, the choice of representatives of $X$ in Definition \ref{lp1} does not affect $LP_1(X)$. It is possible to show, similarly as in \cite{MR2041630}, that this definition reflects the stable self-intersections of $X$, at least when $X$ is primitive, in a 2 to 1 correspondence: each stable self-intersection corresponds to two linked pairs of the form $(P,Q)$ and $(Q,P)$. Non-primitive closed curves have stable self-intersections, in the sense of the self-intersections of a transversal perturbation, that do not arise from linked pairs. Since we will not focus on the co-bracket, we shall not prove these assertions. Next we turn to the case of linked pairs between two strings, that will be the key for the construction of the bracket.
\begin{definition} \label{lp2} Let $X$ and $Y$ be strings. Define $LP_2(X,Y)$, the set of linked pairs of $X$ and $Y$, as the set of linked pairs $(P,Q)$ between representatives of $X^n$ and $Y^m$ for $n,m\geq 1$, where
$l(X^{n-1})\leq l(P, Q)< l(X^{n})$ and $l(Y^{m-1})\leq l(P, Q)< l(Y^{m})$. (With the convention that $l(X^0)=l(Y^0)=0$).
\end{definition}
Again, different choices of representatives for the strings in Definition \ref{lp2} yield sets $LP_2(X,Y)$ that are in natural bijection.
\begin{remark} \label{r:lp2} The powers are necessary:
Consider $\alpha$ and $\beta$, closed geodesics starting and ending at the same point $x$ and meeting transversally at $x$. Let $X$ be the conjugacy class of $[\alpha]$ and $Y$ the conjugacy class of $[\beta] [\alpha^2]$.
There is no linked pair between $X$ and $Y$ but there is a linked pair between $X^3$ and $Y$. Note that the core segment of the linked pair is $\alpha^2$.
\end{remark}
On the other hand, it can be shown that $LP_2(X,X)=LP_1(X)$, i.e. the powers are not needed in the case $X=Y$. We shall see later, in Lemma \ref{chas20}, that $LP_2(X,Y)$ captures the notion of stable intersections between $X$ and $Y$. It is not inmediate from Definition \ref{lp2} that $LP_2(X,Y)$ is finite, the proof of this fact will be based in the following result.
\begin{proposition} \label{alphabet} Let $\mathcal{U}=\{\alpha_1,\ldots,\alpha_n \}$ be a finite set of piecewise geodesic curves. There are
factorizations $\alpha_i=a_{i,1}\cdots a_{i,n_i}$ such that whenever ${a_{i,j}\cap a_{k,l} \neq \emptyset}$, either
\begin{enumerate}
\item $a_{i,j}$ and $a_{k,l}$ meet only at one endpoint.
\item $a_{i,j}=a_{k,l}$
\item $a_{i,j}=a_{k,l}^{-1}$
\end{enumerate}
\end{proposition}
\begin{proof} Subdivide any factorization of the curves $\alpha_i$ until the desired properties are obtained. This will happen because
of the transversality properties of the geodesics.
\end{proof}
Given strings $X$ and $Y$, Proposition \ref{alphabet} allows us to find a finite set of curves that works as an alphabet for writing some representatives of $X$ and $Y$, as well as all the core curves of the intersection pairs between (powers of) these representatives. Thus we can write the cyclic factorizations that make up the elements of $LP_2(X,Y)$ as words in this alphabet.
\begin{lemma} For any strings $X$ and $Y$, $LP_2(X,Y)$ is finite.
\end{lemma}
\begin{proof} Using Proposition \ref{alphabet} as indicated above, this becomes a straightforward adaptation of Lemma 2.9 of \cite{MR2041630}.
\end{proof}
\begin{definition} \label{dotproduct} Let $X$ and $Y$ be strings and
$(P,Q)\in LP_2(X,Y)$. Write $P=(\alpha_1, \eta)$ and $Q=(\beta_1, \xi)$ as in the definition of linked pairs, and let $\alpha$ and $\beta$ be the representatives of $X$ and $Y$ that satisfy the following:
\begin{itemize}
\item If $(P, Q)$ is of type (1) or (2), then \[ \alpha^n=\alpha_1\eta \mbox{ and } \beta^m=\beta_1\xi \]
(where $n,m\geq 1$ are the powers of $X$ and
$Y$ that correspond to $(P,Q)$ in Definition \ref{lp2}).
\item If $(P, Q)$ is of type (3), then
\[ \alpha^n=\alpha_1\eta \mbox{ and } \beta^m=\xi\beta_1 \]
(for the same $n,m\geq 1$).
\end{itemize}
In any of the above cases, define $(X\cdot_{(P,Q)} Y)$ to be the conjugacy class of $[\alpha][\beta]$.
\end{definition}
We say that $(X\cdot_{(P,Q)} Y)$ is the {\em dot product} of $X$ and
$Y$ at $(P,Q)$. Notice that $\alpha$ and $\beta$ are the representatives of $X$ and $Y$ that we get by choosing parameter-basepoints at the ``ending'' of the linked pair's core curve. They are indeed loops based at the same point, so the concatenation $[\alpha][\beta]$ is well defined.
\begin{definition} \label{d:bracket} Let $X$ and $Y$ be strings, we define their bracket as
\[ [X,Y]=\sum_{(P,Q)\in LP_2(X,Y)}\mbox{sign}(P,Q)(X\cdot_{(P,Q)} Y) \]
Then we extend the definition to $\mathcal{S}(M)$ so that the bracket is bilinear.
\end{definition}
\subsection{Linked pairs and differentiable curves}\label{dif}
Our next goal is to show the correspondence between linked pairs and stable intersections, which will lead to the relationship between the bracket in Definition \ref{d:bracket} and the Goldman-Turaev bracket given by equation \eqref{goldman}. This will in turn allow us to prove Theorem \ref{jacobi}, i.e. that Definition \ref{d:bracket} gives a Lie algebra.
Let $C$ be a compact one dimensional complex on an oriented Riemannian surface $M$ whose edges are geodesic arcs, and consider a basepoint $x\in C$. Let $\tilde C$ be the universal covering of $C$. Then the following lemma is straightforward.
\begin{lemma}\label{lem:ident}
${\mathcal L}_x(C)\cong \pi_1(C, x)$ and ${\mathcal E}_x(C)\cong \tilde C$.
Moreover, these correspondences give an isomorphism of fiber bundles $$({\mathcal E}_x(C),{\mathcal L}_x(C),C,\pi) \cong (\tilde C,\pi_1(C, x),C,\pi)$$ where $\pi_1(C, x)$ acts on $\tilde C$ by deck transformations.
\end{lemma}
Let $S(C)$ be the set of strings contained in $C$, and note that the string
bracket of Definition \ref{d:bracket} can be restricted to
$\mathcal S(C)$, the free abelian group on $S(C)$. We will denote this bracket by $[\;,\;]_C$.
For any set $V\subset M$ let $V_{\varepsilon}$ denote an
$\varepsilon$-neighborhood of $V$. The following lemma is well
known, see \cite{MR621981} for a very general construction.
\begin{lemma}\label{tubular}
If $\varepsilon$ is small enough, there is a retraction
$\chi:C_{\varepsilon}\to C$. It induces an isomorphism $\chi_*:\pi_1(C_{\varepsilon},x)\to\pi_1(C,x)$ for any $x\in C$.
\end{lemma}
Let ${\mathcal S}_1(M)$ be the free abelian group generated by the
set of conjugacy classes of $\pi_1(M)$, where the Goldman-Turaev bracket is defined. By Lemma \ref{lem:ident} we see that $\chi$ induces an isomorphism of abelian groups $\chi_*:{\mathcal S}_1(C_{\varepsilon})\to {\mathcal S}(C) $. We shall prove that $\chi_*$ sends the Goldman-Turaev bracket of the surface $C_{\varepsilon}$ to $[\;,\;]_C$.
Proposition 3.11 in \cite{MR2041630} can be rephrased as
\begin{lemma}\label{chas0} Let $[\alpha]$ and $[\beta]$ be representatives of strings in $S(C)$. Then there are differentiable curves $\gamma$ and $\delta$ in $C_{\varepsilon}$ such that
\begin{itemize}
\item $\gamma$ and $\delta$ are $\varepsilon$-perturbations of $\alpha$ and $\beta$ respectively,
\item $\alpha$ and $\beta$ are the reduced forms of $\chi\circ \gamma$ and $\chi\circ \delta$ respectively, and
\item $\gamma$ and $\delta$ intersect transversally, in at most double points, and determine no bigons.
\end{itemize}
Moreover, $\gamma$ and $\delta$ have minimal intersection among the curves in their homotopy classes.
\end{lemma}
We remark that in Lemma \ref{chas0} the curves $\chi\circ \gamma$ and $\chi\circ \delta$ need not be reduced, but the curves removed in their reduction have length less than $\varepsilon$, and each one is contained in some geodesic edge of the complex $C$, assuming $\varepsilon$ is small enough. Now we can relate the linked pairs of two strings in $S(C)$ with the intersections that are stable under perturbation in $C_{\varepsilon}$ .
\begin{lemma}\label{chas20} Let $[\alpha]$ and $[\beta]$ be representatives of strings $X$ and $Y$ in $S(C)$, and let $\gamma$ and $\delta$ be curves given by Lemma \ref{chas0}.
Then for each intersection point $p$ of $\gamma$ and $\delta$ there are $n,m\geq 1$, curves $a\subset\gamma^n$ and $b\subset\delta^m$ meeting at $p$, and a linked pair $(P,Q)$ between $\alpha^n$ and $\beta^m$ that satisfy the following:
If $P=(\alpha_1,\xi)$ and $Q=(\beta_1,\eta)$, then the reduced forms of $\chi\circ a$ and $\chi\circ b$ can be written as $a_1\xi a_2$ and $b_1\eta b_2$ respectively, where $a_1,a_2,b_1,b_2$ are geodesic segments.
Moreover, this correspondence is a bijection between the intersection points of $\gamma$ and $\delta$, and $LP_2(X,Y)$.
\end{lemma}
\begin{proof}
This is done with the techniques of \cite{MR2041630}. We
consider $\pi:\tilde{C}_{\varepsilon}\to C_{\varepsilon}$, the universal cover of
$C_{\varepsilon}$. Then $\tilde{C}$ is embedded in
$\tilde{C}_{\varepsilon}$ as a tree, made of piecewise geodesic curves,
since $C$ is piecewise geodesic. Consider also
$\tilde{\chi}:\tilde{C}_{\varepsilon}\to\tilde{C}$ the lift of $\chi$, and
for each geodesic arc $c$ in the decomposition of $\tilde{C}$ let
$V(c)=\tilde{\chi}^{-1}(c)$. Then the sets $V(c)$ are homeomorphic to closed disks, they cover $\tilde{C}_{\varepsilon}$, and their interiors are disjoint.
Now, for each $p$ in the intersection between $\gamma$ and $\delta$, we pick $\hat{p}\in\tilde{C}_{\varepsilon}$ projecting to $p$, and consider $\hat{\gamma}$ and $\hat{\delta}$ the infinite lifts of $\gamma$ and $\delta$ that meet only at $\hat{p}$ (recalling that $\gamma$ and $\delta$ have no bigons). Let $\hat \alpha$ and $\hat \beta$ be the respective reductions of $\tilde \chi \circ \hat \gamma$ and $\tilde \chi \circ \hat \delta$, which are infinite lifts of $\alpha$ and $\beta$. The set $\hat \alpha(I) \cap \hat \beta(I)$ is compact, since $\hat{\gamma}$ and $\hat{\delta}$ meet only once and are lifts of closed curves, and it is an arc, since $\hat \alpha$ and $\hat \beta$ are reduced. We shall write $\hat \alpha(I) \cap \hat \beta(I)$ as a curve in two ways, with possibly different orientations: we call $\hat\xi$ and $\hat\eta$ to the curves spanning $\hat \alpha(I) \cap \hat \beta(I)$ with the orientations given by $\hat \alpha$ and $\hat \beta$ respectively. It may be the case that $\hat \alpha(I) \cap \hat \beta(I)$ is just a point, then $\hat \xi$ and $\hat \eta$ will be constant (this will result in a linked pair of type (1) ).
Let
\[ V = \cup \{V(c): c\cap \hat \alpha(I) \cap \hat \beta(I) \neq \emptyset \} \]
Then $V$ is, topologically, a closed disk, and we have $\hat \alpha \cap V = c_1 \hat\xi c_2$ and $\hat \beta \cap V = d_1 \hat\eta d_2$, where $c_1,c_2,d_1,d_2$ are pairwise different geodesic segments.
For each geodesic segment $c$ that meets $\hat \alpha(I) \cap \hat \beta(I)$ but is not contained in it, we define the set $B(c)=\partial V(c)\cap\partial V$. Note that $B(c)$ is a segment in $\partial V \cong S^1$, and that $c$ has an endpoint in $B(c)$ and the other in $\hat \alpha(I) \cap \hat \beta(I)$. The segments $B(c)$ just defined are pairwise disjoint, and their union is the relative boundary of $V$ in $\tilde{C}_{\varepsilon}$. Taking $\varepsilon$ small enough, we may assume that the $\varepsilon$-neighborhood of $\hat \alpha$ only meets the relative boundary of $V$ at the arcs $B(c_1)$ and $B(c_2)$. Thus $\hat \gamma \cap V$ is an arc that enters $V$ through $B(c_1)$ and exits through $B(c_2)$, meeting no other segment of the relative boundary of $V$. Similarly we get that $\hat \delta\cap V$ is an arc that traverses $V$ from $B(d_1)$ to $B(d_2)$.
Let $\hat a = \hat \gamma \cap V$ and $\hat b = \hat \delta \cap V$. They must intersect at $\hat p$, in particular $\hat p \in V$, since the complementary arcs of $\hat \gamma$ and $\hat \delta$ are in different components of $\tilde{C}_{\varepsilon}-V$. Note that $\tilde\chi\circ\hat a$ and $\tilde\chi\circ\hat b$ can be reduced, respectively, to $\hat \alpha \cap V = c_1 \hat\xi c_2$ and $\hat \beta \cap V = d_1 \hat\eta d_2$. We define the curves $a,b,\xi,\eta$ in the statement as the respective projections of $\hat a, \hat b, \hat \xi, \hat \eta$.
By compactness, there are $n,m\geq 1$ such that $\hat a$ and $\hat b$ are contained in lifts of $\gamma^n$ and $\delta^m$ inside $\hat \gamma$ and $\hat \delta$ respectively. We choose $n, m$ minimal for these inclusions to be strict. Thus $\xi$ and $\eta$ induce cyclical factorizations of $\alpha^n$ and $\beta^m$ respectively, namely $P$ and $Q$. It only remains to show that \[ (a_1,\xi,a_2) \mbox{ and } (b_1,\eta,b_2) \]
is a linked pair, where $a_1,a_2,b_1,b_2$ are the respective projections of $c_1,c_2,d_1,d_2$. This is because $\hat\gamma$ and $\hat \delta$ meet transversally, and only once in $V$, thus $B(c_1)\cup B(c_2)$ separates $B(d_1)$ from $B(d_2)$ in $\partial V \cong S^1$. Using the orientation of $V$ induced by lifting the one of $C_{\varepsilon}\subset M$, the last fact allows us to verify the sign conditions in the definition of linked pair. It also yields that $sign(P,Q)=\epsilon(p,\gamma,\delta)$, which will be useful later.
Note that, by the minimality of $n$ and $m$, we have $(P,Q)\in LP_2(X,Y)$. The reciprocal construction is now straightforward, and so is checking bijectivity.
\end{proof}
\begin{figure}[htbp]
\input{fig4.pic}
\caption{Proof of lemma \ref{chas20}. We depict the simple case when $\hat \alpha(I)\cap\hat\beta(I)$ does not contain branching vertices of $\tilde C$ in its interior. For this figure, $a_1,\ldots,a_k$ denote the geodesic segments in $\hat \alpha(I)\cap\hat\beta(I)$.}
\end{figure}
\begin{lemma}\label{cor:inc}
$({\mathcal S}(C), [\; ,\; ]_C)$ is a Lie algebra.
\end{lemma}
\begin{proof}
Let $X,Y\in S(C)$ and let $\alpha$, $\beta$, and $\gamma$, $\delta$ be as in Lemma \ref{chas0}. We need
to show that $\chi_*([\gamma ,
\delta]_{\pi_1(C_{\varepsilon})})=[X,Y]_C$, where $[\gamma , \delta]_{\pi_1(C_{\varepsilon})}$ stands for the Goldman-Turaev bracket between the free homotopy classes of $\gamma$ and $\delta$. This is a
consequence of Lemma \ref{chas20}: If $p$ corresponds to
the linked pair $(P,Q)$, then we have seen that $sign(P,Q)=\epsilon(p,\gamma,\delta)$ in the proof of Lemma \ref{chas20}. On the other hand, $(X\cdot_{(P, Q)}Y)$ is the conjugacy class of the image under $\chi$
of $\gamma\cdot_p\delta$, which follows from Definition \ref{dotproduct} and the properties of the correspondence between $p$ and $(P,Q)$ given by Lemma \ref{chas20}.
\end{proof}
We remark that we have obtained the isomorphism of Lie algebras \[\chi_*:({\mathcal S}_1(C_{\varepsilon}),[\;,\;]_{\pi_1(C_{\varepsilon})}) \to ({\mathcal S}(C), [\; ,\; ]_C) \] that we desired.
\subsection{Proof of Theorem\ref{jacobi}}
We shall check that the bracket on $\mathcal S(M)$ given by Definition \ref{d:bracket} satisfies the axioms of a Lie algebra. It is bilinear by definition, and we would like to remark that anti-symmetry can be checked directly, showing that the bijection between $LP_2(X,Y)$ and $LP_2(Y,X)$ that sends $(P,Q)$ to $(Q,P)$ verifies that \[ (X\cdot_{(P, Q)}Y)=(Y\cdot_{(Q, P)}X) \mbox{ and } sign(P,Q)=-sign(Q,P)\] It can also be proved by the same method we use for the Jacobi identity, which we shall check next.
Consider strings $X,Y,Z\in S(M)$, and cyclically reduced representatives $\alpha,\beta,\gamma$ of them. Applying Proposition \ref{alphabet} to $\mathcal U=\{\alpha,\beta,\gamma\}$ we see that the set $C=\alpha(I)\cup\beta(I)\cup\gamma(I)$ is a one dimensional complex with geodesic edges. Recall that the string bracket $[\;,\;]$ restricts to $\mathcal S(C)\subset \mathcal S(M)$, where it defines a Lie algebra by Lemma \ref{cor:inc}. By construction we have that $X,Y,Z$ are in $\mathcal S(C)$, thus the Jacobi identity between $X,Y,Z$ is obtained.
That shows Theorem \ref{jacobi}. We would also like to point out that there is a natural quotient $\mathcal S(M)\to\mathcal S_1(M)$, since free homotopy of closed curves is a coarser equivalence than the one defining $S(X)$, and we can show that this map is a homomorphism \[ (\mathcal S(M),[\;,\;])\to (S_1(M),[\;,\;]_{\pi_1(M)}) \]
To check this we can consider $X,Y\in S(M)$, take representatives $\alpha,\beta$ and let $\gamma,\delta$ be the curves given by Lemma \ref{chas0} for $C=\alpha(I)\cup\beta(I)$. Then $\gamma$ and $\delta$ are freely homotopic to $\alpha$ and $\beta$, and the same argument for Lemma \ref{cor:inc} shows that $[X,Y]$ maps to $[\gamma,\delta]_{\pi_1(M)}$ under the natural quotient.
As we commented in the introduction, it is also possible to define a co-bracket in a similar fashion as we did for the bracket in Definition \ref{d:bracket}, this time involving $LP_1$. That gives a Lie bi-algebra structure on $\mathcal S(M)$, and the axioms can also be verified using one dimensional complexes and results of \cite{MR2041630}.
\section{Infinite lifts and intersections} \label{s.lift int}
With the goal of proving Theorem \ref{simple} in mind, we will study the intersections of a cyclically reduced curve with its inverse by looking at the horizontal lifts in the loop bundle.
Throughout this section we fix a cyclically reduced, non-trivial closed curve $\alpha$, and write $x=\alpha(0)$. Let $\tilde \alpha$ be the horizontal lift of
$\alpha$ to ${\mathcal E}_x(M)$ such that $\tilde \alpha(0)=id_x$. We consider the set
\[ \Lambda_{\alpha}=\bigcup_{n\in {\mathbb Z}} [\alpha]^n \tilde \alpha(I) \]
which is nothing but the infinite lift of $\alpha$ through $id_x$. Since $\alpha$ is cyclically reduced, $\Lambda_{\alpha}$ is a {\em line} in ${\mathcal E}_x(M)$, i.e. is an embedding of $\field{R}$ (it has no ``spikes''). We give it a standard orientation induced by the orientation of $\tilde{\alpha}$. Note that $\Lambda_{\alpha^{-1}}$ agrees with $\Lambda_{\alpha}$ as a set, but has the opposite orientation.
\begin{figure}[htbp]
\input{fig1.pic}
\caption{The set $\Lambda_{\alpha}$}
\end{figure}
\subsection{Intersections as elements of $\mathcal L_x$.} \label{ss.lift int1}
For every $A\subset\mathcal E_x(M)$ define
\[ T(A)=\left\{ g \in {\mathcal L}_x(M) \;\; \mbox{such that} \;\;
g A\cap A \neq \emptyset\right\}\]
Consider $g\in T(\Lambda_{\alpha})$ so that $g\Lambda_{\alpha}\neq \Lambda_{\alpha}$. We show that $g\Lambda_{\alpha}\cap \Lambda_{\alpha}$ must be a compact arc (or a point): Note that $\mathcal E_x(M)$ contains no non-trivial horizontal loops, and if $g\Lambda_{\alpha}\cap \Lambda_{\alpha}$ contains a ray, then we note that $g[\alpha]^n\tilde{\alpha}(I)\subset \Lambda_{\alpha}$ for some $n$. Thus the projection of $g[\alpha]^n\tilde{\alpha}$ is some permutation of $\alpha$, which leads to an equation of the form $gw^{n+1}=w^m$ where $[\alpha]=w^k$ (for some such powers), and this implies $g\Lambda_{\alpha}= \Lambda_{\alpha}$.
Then to each $g\in T(\Lambda_{\alpha})$ with $g\Lambda_{\alpha}\neq \Lambda_{\alpha}$ we can associate a horizontal curve $b_g\subset \Lambda_{\alpha}$, with the same orientation as $\Lambda_{\alpha}$, such that $b_g(I)=g\Lambda_{\alpha}\cap \Lambda_{\alpha}$. Let $a_g \subset \Lambda_{\alpha}$ be the horizontal curve such that $ga_g(I)=b_g(I)$ and $a_g$ has the orientation carried from $b_g$ by the action of $g$ (thus we may write $ga_g=b_g$). Note that the pair $(a_g,b_g)$ determines $g$, since $\mathcal L_x(M)$ acts freely on $\mathcal E_x(M)$.
\begin{definition} \label{def:orient} Let $g\in T(\Lambda_{\alpha})$.
\begin{itemize}
\item We say that $g$ {\em preserves orientation} if either $g\Lambda_{\alpha}=\Lambda_{\alpha}$ or the orientation of $a_g$ agrees with that of $\Lambda_{\alpha}$. Let $T^+(\Lambda_{\alpha})$ be the set of orientation preserving elements of $T(\Lambda_{\alpha})$.
\item We say that $g$ {\em reverses orientation} if the orientation of $a_g$ is opposite to that of $\Lambda_{\alpha}$. We denote by $T^-(\Lambda_{\alpha})$ the set of orientation reversing elements of $T(\Lambda_{\alpha})$.
\end{itemize}
The case when $a_g$ is constant shall be regarded as both orientation preserving and reversing. When we want to exclude this case we say that $g$ {\em strictly} preserves or reverses orientation.
\end{definition}
\begin{remark} \label{inv T} The sets $T(\Lambda_{\alpha})$, $T^+(\Lambda_{\alpha})$ and $T^-(\Lambda_{\alpha})$ are closed under taking inverses.
\end{remark}
\begin{remark} \label{rm: no potencias} If $g\in T(\Lambda_{\alpha})$ and $g\Lambda_{\alpha}\neq \Lambda_{\alpha}$, then we have $l(a_g)=l(b_g)<l(\alpha)$.
\end{remark}
\begin{proof}
Assume the contrary, i.e. that $l(b_g)\geq l(\alpha)$. Then if $g\in T^+(\Lambda_{\alpha})$ we get a contradiction by showing that $g\Lambda_{\alpha}= \Lambda_{\alpha}$, with a similar argument as in the case when $g\Lambda_{\alpha}\cap \Lambda_{\alpha}$ contained a ray. On the other hand, if $g\in T^-(\Lambda_{\alpha})$ we can deduce that $\alpha^{-1}$ is a permutation of $\alpha$, which would mean that $\alpha$ is trivial.
\end{proof}
Next we shall see that each $g\in T(\Lambda_{\alpha})$ with $g\Lambda_{\alpha}\neq \Lambda_{\alpha}$ defines naturally an intersection pair between $\alpha$ and $\alpha^{-1}$. Let $t_g=\pi\circ a_g=\pi\circ b_g$ and $\epsilon = \pm 1$ according to whether $g$ preserves or reverses orientation (in case $a_g$ is constant the pick makes no difference). Recalling that $b_g(I)=g\Lambda_{\alpha}\cap \Lambda_{\alpha}$ and projecting what we see at a neighborhood of $b_g$ in $g\Lambda_{\alpha}\cup \Lambda_{\alpha}$, we can find geodesic curves $r,s,u,v$ such that $(r, t_g, s), (u, t^{-\epsilon}_g, v)$ is an intersection pair with $rt_gs\subset\alpha$ and $ut^{-\epsilon}_gv\subset \alpha^{-1}$, where the inclusions are modulo permutation (we do not have to consider powers of $\alpha$ or $\alpha^{-1}$ because of Remark \ref{rm: no potencias}). We depict this situation in Figure \ref{fig:2}.
\begin{figure}[htbp]
\input{fig2.pic}
\caption{A schematic example of $g\Lambda_{\alpha}\cup \Lambda_{\alpha}$ for $g\in T(\Lambda_{\alpha)}$. All shown curves are
horizontal.
We assume $\alpha=abc=ecaf$ and $g=[c^{-1}e^{-1}]$, note that $t_g= ca$.} \label{fig:2}
\end{figure}
Then we can give cyclic factorizations $P_g=(\alpha_1,t_g)$ of $\alpha$ and $Q_g=(\beta_1,t_g^{-\epsilon})$ of $\alpha^{-1}$ so that:
\begin{itemize}
\item the horizontal lift of $t_g\alpha_1$ at $b_g(0)$ is contained in $\Lambda_{\alpha}$, and
\item the horizontal lift of $t_g^{-\epsilon}\beta_1$ at $a_g^{-\epsilon}(0)$ is contained in $\Lambda_{\alpha}$.
\end{itemize}
In other words, $\alpha_1$ can be obtained by projecting a curve spanning a component of $$\Lambda_{\alpha}-\bigcup_{n\in\field{Z}}[\alpha]^nb_g(I)$$ with the orientation given by $\Lambda_{\alpha}$, while $\beta_1$ is the analog for the translates of $a_g(I)$ and the reverse orientation to that of $\Lambda_{\alpha}$.
Notice that if $g$ is orientation preserving, i.e. when $\epsilon=1$, we get an intersection pair of type either (3) or (1) between $\alpha$ and $\alpha^{-1}$. In the orientation reversing case, when $\epsilon=-1$, we get an intersection pair of type either (2) or (1). In Figure \ref{fig:2} we depict a situation where $g$ is orientation preserving.
Observe that an element of the form $h=[\alpha]^n g [\alpha]^m$ induces the same intersection pair as $g$, since $b_h=[\alpha]^nb_g$ and $a_h=[\alpha]^{-m}a_g$ induce the same cyclic factorizations of $\alpha$ and $\alpha^{-1}$. With this in mind we define $$T_1(\alpha) =\{g\in T(\tilde{\alpha}(I)) \,:\, g\Lambda_{\alpha}\neq \Lambda_{\alpha} \mbox{ and }\tilde{\alpha}(1)\notin a_g(I)\cup b_g(I) \} $$
noting that the conditions amount to ask that $a_g$ and $b_g$ meet $\tilde{\alpha}(I)$ but not $\tilde{\alpha}(1)$. Since $\tilde{\alpha}([0,1))$ is a fundamental domain for $\Lambda_{\alpha}$ under translations by powers of $[\alpha]$, we get the following:
\begin{remark}\label{mn}
Let $g\in T(\Lambda_{\alpha})$ such that $g\Lambda_{\alpha}\neq \Lambda_{\alpha} $. Then there is a unique $g_1\in T_1(\alpha)$ and integers $m,n$ such that
\[g=[\alpha]^n g_1[\alpha]^m. \]
\end{remark}
The next result describes the intersections of a string with its inverse in terms of elements of the loop group. Let $\mathcal I(\alpha, \alpha^{-1})$ denote the set of intersection pairs between $\alpha$ and $\alpha^{-1}$.
\begin{lemma}\label{bijection}
The map \[ T_1(\alpha)\to \mathcal I(\alpha,\alpha^{-1})\] that takes $g\to (P_g, Q_g)$ is a bijection.
\end{lemma}
\begin{proof}
For injectivity, consider the way in which an horizontal curve $\nu\subset\Lambda_{\alpha}$ with $l(\nu)<l(\alpha)$ defines a cyclic factorization of $\alpha$ or $\alpha^{-1}$, as was used in the construction of the map $g\to (P_g, Q_g)$. Then observe that two such curves $\nu_1$ and $\nu_2$ yield the same cyclic factorization iff $\nu_1=[\alpha]^n\nu_2$ for some $n\in \field{Z}$. Combining this fact with Remark \ref{mn} gives injectivity.
To show surjectivity consider $(P,Q)\in \mathcal I(\alpha,\alpha^{-1})$ and write $P=(\alpha_1,\xi)$, $Q=(\beta_1,\eta)$ as in Definition \ref{def:lp cyclic}. Let $\epsilon=\pm 1$ so that $\eta=\xi^{-\epsilon}$, i.e. $\epsilon=1$ for $(P,Q)$ of type (3), while $\epsilon=-1$ for type (2), and either one for type (1). Since $\xi\alpha_1$ is a permutation of $\alpha$, there are horizontal curves $\tilde\xi$ and $\tilde\alpha_1$, projecting to $\xi$ and $\alpha_1$ respectively, such that the concatenation $\tilde\xi\tilde\alpha_1$ is well defined and contained in $\Lambda_{\alpha}$. Note that $\tilde\xi$ and $\tilde\alpha_1$ have the orientation of $\Lambda_{\alpha}$, and that $\Lambda_{\alpha}$ is the infinite lift of $\xi\alpha_1$ that continues $\tilde\xi\tilde\alpha_1$. Now we consider the infinite horizontal lift of $\eta\beta_1$ starting at $\tilde\xi^{-\epsilon}(0)$. Since $\eta\beta_1$ is a permutation of $\alpha^{-1}$, this infinite lift must be of the form $g\Lambda_{\alpha}$ for some $g\in T(\Lambda_{\alpha})$. Notice that the definition of intersection pair implies, by taking the appropriate horizontal lifts, that $\Lambda_{\alpha}\cap g\Lambda_{\alpha} = \tilde\xi(I)$. Thus it becomes direct to verify that $(P_g,Q_g)=(P,Q)$, and we can use Remark \ref{mn} to finish the proof.
\end{proof}
Through the proof of Lemma \ref{bijection} we see that $T_1(\alpha)$ is a choice of a restriction of domain, in order to obtain a bijection from the construction that associates $g\to (P_g, Q_g)$. This choice satisfies the following nice property:
\begin{remark}\label{inverse}
$g\in T_1(\alpha)$ iff $g^{-1}\in T_1(\alpha)$. Moreover, $$b_{g^{-1}}=a_g^{\epsilon} \mbox{ and } a_{g^{-1}} = b_g^{\epsilon}$$ where $\epsilon=\pm 1$ according to whether $g$ is orientation preserving or reversing.
\end{remark}
In later sections we shall focus on the linked pairs, i.e. the intersection pairs that are relevant for the bracket.
\begin{definition} \label{def:T0} We define $T_0(\alpha)\subseteq T_1(\alpha)$ as the set of elements that correspond to linked pairs under the bijection of Lemma \ref{bijection}.
\end{definition}
Note that by Remark \ref{rm: no potencias} and Lemma \ref{bijection}, the set $T_0(\alpha)$ is in bijection with $LP_2(X,X^{-1})$, where $X$ is the conjugacy class of $[\alpha]$. From these same results we also get that $LP_1(X)=LP_2(X,X)$, which is in natural bijection with $LP_2(X,X^{-1})$.
Observe also that by Lemma \ref{chas20} a string $X$ is simple, as defined in the introduction, iff $LP_1(X)=\emptyset$, or equivalently, iff $LP_2(X,X^{-1})=\emptyset$.
\subsection{Orientation reversing elements and unique intersections.}
The orientation properties of the elements of $T_0(\alpha)$, which correspond to the type of their associated linked pairs, will play a major role in proving Theorem \ref{simple}. Next we study the key properties of the orientation reversing case.
\begin{lemma} \label{reversing0} Let $\xi$ and $\eta$ be non-constant segments of $\Lambda_{\alpha}$ going in the positive orientation. Then if $\pi\circ\xi=(\pi\circ\eta)^{-1}$ we have $\xi(I)\cap\eta(I)=\emptyset$ and $l(\alpha)>2l(\xi)$.
\end{lemma}
\begin{proof}
Write $\gamma=\pi\circ\xi$. Then an overlap between $\xi$ and $\eta$ would project to an overlap between $\gamma$ and $\gamma^{-1}$, and if $\xi$ and $\eta$ meet at an endpoint, that would project to a reduction of $\alpha$, of the form $\gamma\gamma^{-1}$ or $\gamma^{-1}\gamma$. Thus we get the first claim. The second one comes from considering a permutation $\alpha_0$ of $\alpha$ so that its horizontal lift starting at $\xi(0)$ is contained in $\Lambda_{\alpha}$. Note that such lift ends at $[\alpha]\xi(0)$, which is not in $\eta(I)$ by the first claim applied to $\eta$ and $[\alpha]\xi$. Thus we obtain $\alpha_0=\gamma a \gamma^{-1} b$ with $a$ and $b$ non-constant, so $$l(\alpha)=l(\alpha_0)>2l(\gamma)=2l(\xi).$$
\end{proof}
\begin{lemma}\label{reversing}
Let $g\in T^-(\Lambda_{\alpha})$. Then $[\alpha^n]a_g(I)\cap [\alpha^m]b_g(I)=\emptyset$ for all $n,m\in\field{Z}$.
\end{lemma}
\begin{proof}
If $g$ reverses orientation strictly, we apply lemma \ref{reversing0} to the curves $[\alpha^n]a_g^{-1}$ and $[\alpha^m]b_g$.
In case $a_g$ is constant, say $a_g=e_p$ for $p\in\Lambda_{\alpha}$, we get that $[\alpha^n]g[\alpha^{-m}]p=p$ which is absurd because the action of $\mathcal L_x(M)$ is free and $g$, being orientation reversing, is not a power of $\alpha$.
\end{proof}
\begin{remark} Note that for $g\in T^-(\Lambda_{\alpha})$ we have $l(t_g)<l(\alpha)/2$.
\end{remark}
We say that $\alpha$ has {\em unique intersection} if $T_1(\alpha)$ consists of only two elements, $g$ and $g^{-1}$ (by Remark \ref{inverse}). By Lemma \ref{bijection} this is equivalent to say that $\mathcal I(\alpha,\alpha^{-1})$ has two elements. Observe that the definition of intersection pair makes sense for a curve in a general one dimensional complex, i.e. not necessarily embedded in a surface. Thus we may speak of unique intersection for curves in this more general setting.
Given a subgroup $G\subseteq \mathcal L_x(M)$ we can consider $\Lambda_{\alpha}/G$, the image of $\Lambda_{\alpha}$ in the quotient $\mathcal E_x(M)/G$, which is a one dimensional complex since $\Lambda_{\alpha}/\mathcal L_x(M)=\alpha(I)$ and $\alpha$ is piecewise geodesic. For $g\in \mathcal L_x(M)$, let $G_g$ be the subgroup generated by $g$ and $[\alpha]$.
\begin{lemma}\label{quotient}
Let $g\in T^-(\Lambda_{\alpha})$ and $\bar \alpha$ be the projection of $\tilde \alpha$ onto $\Lambda_{\alpha}/G_g$.
Then $\bar \alpha$ has a unique intersection.
\end{lemma}
\begin{proof}
By Lemma \ref{reversing} we have that all the translates $[\alpha]^na_g(I)$ and $[\alpha]^mb_g(I)$, for $n,m\in\field{Z}$, are pairwise disjoint. Consider first the quotient $\Lambda_{\alpha}/\langle [\alpha]\rangle$, which is a circle that can be parametrized by the projection of $\tilde \alpha$. Then $a_g(I)$ and $b_g(I)$ project to $\Lambda_{\alpha}/\langle [\alpha]\rangle$ as two disjoint intervals, and $\Lambda_{\alpha}/G_g$ is the further quotient obtained by identifying these two intervals (with the appropriate orientation). The interval resulting from this identification is the core of the only self-intersection pairs of $\bar\alpha$, which are only two (of the form $(P,Q)$ and $(Q,P)$), and this shows the lemma.
\end{proof}
\subsection{$\alpha$-oriented subgroups}\label{alpha_oriented}
Now we turn our attention to the orientation preserving elements of $T(\Lambda_{\alpha})$. We will be showing that they generate subgroups with the following property.
\begin{definition}
Let $G\subset \mathcal L_x(M)$ be a subgroup. We say that $G$ is {\em $\alpha$-oriented} if for every $g, h\in G$ with $g^{-1}h\in T(\Lambda_{\alpha})$ we have
$g^{-1}h\in T^+(\Lambda_{\alpha})$.
\end{definition}
Let us explain this definition in geometric terms. First note that each $g\in \mathcal L_x(M)$ induces an orientation on $g\Lambda_{\alpha}$ by carrying the standard orientation of $\Lambda_{\alpha}$ through the action of $g$. For two elements $g,h\in \mathcal L_x(M)$, the condition that $g^{-1}h\in T(\Lambda_{\alpha})$ is equivalent to saying that $g\Lambda_{\alpha}\cap h\Lambda_{\alpha}\neq\emptyset$, and we have $g^{-1}h\in T^+(\Lambda_{\alpha})$ exactly when the orientations of $g\Lambda_{\alpha}$ and $h\Lambda_{\alpha}$ agree on their (non-empty) intersection. If $G\subset \mathcal L_x(M)$ is a subgroup we have that $$ G\Lambda_{\alpha} = \bigcup_{g\in G}g\Lambda_{\alpha}$$ is a one dimensional complex whose connected components are simplicial trees, recalling the form of the intersections between translates of $\Lambda_{\alpha}$ by elements of $\mathcal L_x(M)$. Then we have:
\begin{remark}
If $G$ is $\alpha$-oriented we can give $G\Lambda_{\alpha}$ a $G$-invariant orientation that extends the standard orientation of $\Lambda_{\alpha}$.
\end{remark}
If a subgroup $G\subset \mathcal L_x(M)$ is generated by some elements of $T(\Lambda_{\alpha})$ then $G\Lambda_{\alpha}$ is connected, thus is a simplicial tree. The following technical result will be useful in this context.
\begin{lemma} \label{convexity} Suppose $g,h_1,\ldots,h_n\in \mathcal L_x(M) $ satisfy that:
\begin{itemize}
\item $g^{-1}h_1, g^{-1}h_n \in T(\Lambda_{\alpha})$,
\item $g^{-1}h_i\notin T(\Lambda_{\alpha})$ for $i=2,\ldots,n-1$,
\item $h_{i-1}^{-1}h_i \in T(\Lambda_{\alpha})$ for $i=2,\ldots,n$.
\end{itemize}
Then $h_1^{-1}h_n \in T(\Lambda_{\alpha})$.
\end{lemma}
\begin{proof}
Interpreting the hypotheses in terms of intersections of translates of $\Lambda_{\alpha}$, we can find a curve $\beta\subset h_1\Lambda_{\alpha}\cup\cdots\cup h_n\Lambda_{\alpha}$ that only meets $g\Lambda_{\alpha}$ at its endpoints, with $\beta(0)\in h_1\Lambda_{\alpha}\cap g\Lambda_{\alpha}$ and $\beta(1)\in h_n\Lambda_{\alpha}\cap g\Lambda_{\alpha}$. Since $g\Lambda_{\alpha}\cup h_1\Lambda_{\alpha}\cup\cdots\cup h_n\Lambda_{\alpha}$ has no non-trivial loops, we must have $\beta(0)=\beta(1)$ (and $\beta$ must be a trivial loop), which provides a point in $h_1\Lambda_{\alpha}\cap h_n\Lambda_{\alpha}$ as desired.
\end{proof}
The following is a straightforward observation.
\begin{lemma}\label{fundamental}
If $g, h\in T^+(\Lambda_{\alpha})$ and $g^{-1}h\in T(\Lambda_{\alpha})$, then $g^{-1}h\in T^+(\Lambda_{\alpha})$.
\end{lemma}
Next we present the main result of this subsection, concerning the subgroups generated by orientation preserving elements.
\begin{lemma}\label{oriented}
Let $G\subset\mathcal L_x(M)$ be a finitely generated subgroup whose generators belong to $T^+(\Lambda_{\alpha})$. Then $G$ is $\alpha$-oriented.
\end{lemma}
\begin{proof}
By hypothesis we can write
\[G=\bigcup_i G_i, \mbox{ where } G_0=\{id_x\}, \mbox{ and } G_{i+1}=G_i\cup\{ h_{i+1}\}, \]
such that there is $h'_i\in G_i$ with $h_{i+1}^{-1}h'_i\in T^+(\Lambda_{\alpha})$ for every $i\geq 0$. We are going to show the lemma by induction on $i$: assuming that $g^{-1}h\in T(\Lambda_{\alpha})$ implies $g^{-1}h\in T^+(\Lambda_{\alpha})$ for $g, h\in G_i$, we shall show that this same property holds for $g,h\in G_{i+1}$. The base case of this induction is trivial.
By Remark \ref{inv T}, we only need to consider the case when $g=h_{i+1}$ and $h\in G_i$. So let $h\in G_i$ be such that $h^{-1}_{i+1}h\in T(\Lambda_{\alpha})$, and we are going to show that $h^{-1}_{i+1}h\in T^+(\Lambda_{\alpha})$. By construction of the set $G_i$ there is a sequence $h'_i=k_1,\ldots,k_n=h$ in $G_i$ such that $k_j^{-1}k_{j+1}\in T^+(\Lambda_{\alpha})$ for $j=1,\ldots,n-1$. Since $k_j\in G_i$ for every $j$, the induction hypothesis gives us that $k_j^{-1}k_l\in T^+(\Lambda_{\alpha})$ whenever
$k_j^{-1}k_l\in T(\Lambda_{\alpha})$.
By lemma \ref{convexity} we can assume, maybe after taking a subsequence, that
$h^{-1}_{i+1}k_j\in T(\Lambda_{\alpha})$ for every $j$. Now write
\[ h^{-1}_{i+1}k_2= (h^{-1}_{i+1}k_1) (k^{-1}_{1}k_2) \]
and note that $h^{-1}_{i+1}k_1$ and $k^{-1}_{1}k_2$ are orientation preserving by construction. Therefore $h^{-1}_{i+1}k_2$ is orientation preserving by
Lemma \ref{fundamental}. Proceeding inductively we conclude that $h^{-1}_{i+1}k_n$ is also orientation preserving, as desired.
\end{proof}
\section{Formulas for the terms of the bracket}
In this section we study the dot products between a string and its inverse applying what we developed in section \ref{s.lift int}. Again we fix a cyclically reduced, non-trivial closed curve $\alpha$, and let $x=\alpha(0)$ and $X\in S(M)$ be the conjugacy class of $\alpha$.
\subsection{Expressions for the dot product}
Let $g\in T_1(\alpha)$ and recall the horizontal curves $(a_g,b_g)$ defined in subsection \ref{ss.lift int1}. We introduce the following curves:
\begin{itemize}
\item $\tilde\alpha_g$ is the segment of $\Lambda_{\alpha}$ starting at $b_g(1)$ and ending at $[\alpha]b_g(1)$. Let $\alpha_g=\pi\circ\tilde\alpha_g$.
\item $\tilde\beta_g$ is the segment of $g\Lambda_{\alpha}$ starting at $b_g(1)$ and ending at $g[\alpha^{-1}]a_g(1)$. Let $\beta_g=\pi\circ\tilde\beta_g$.
\item $\tilde\gamma_g$ is the segment of $\Lambda_{\alpha}$ starting at $id_x$ and ending at $b_g(1)$. Let $\gamma_g=\pi\circ\tilde\gamma_g$.
\end{itemize}
We observe $\alpha_g$ and $\beta_g$ are permutations of $\alpha$ and $\alpha^{-1}$ respectively, and that $\tilde\alpha_g$ and $\tilde\beta_g$ are their respective horizontal lifts starting at $b_g(1)$. This is easy to see for $\alpha_g$, and in the case of $\beta_g$ note that $\tilde\beta_g=g\tilde\beta'$ where $\tilde\beta'$ is the segment of $\Lambda_{\alpha}$ starting at $a_g(1)$ and ending at $[\alpha^{-1}]a_g(1)$. According to Definition \ref{dotproduct} (which also makes sense for intersection pairs), the conjugacy class of $\alpha_g\beta_g$ is the dot product $(X\cdot_{(P,Q)}X^{-1})$ where $(P,Q)$ is the intersection pair corresponding to $g$.
On the other hand, $\gamma_g$ is the curve that gives the change of basepoint conjugation so that $[\gamma_g\alpha_g\gamma_g^{-1}]$ and $[\gamma_g\beta_g\gamma_g^{-1}]$ belong to $\mathcal L_x(M)$. We also have that $\tilde\gamma_g$ is the horizontal lift of $\gamma$ at $id_x$. For an example of these curves, in figure \ref{fig:2} we have $\alpha_g=bca$, $\beta_g=a^{-1}c^{-1}e^{-1}f^{-1}$ and $\gamma_g=a$.
\begin{remark} \label{rm:inv a b} By Remark \ref{inverse} we have
\begin{itemize}
\item If $g$ preserves orientation, then $\alpha_{g^{-1}} = \beta_g^{-1} \mbox{ and } \beta_{g^{-1}} = \alpha_g^{-1}$.
\item If $g$ reverses orientation, $\alpha_{g^{-1}}$ and $\beta_{g^{-1}}$ are the respective reductions of $t_g\beta_g^{-1}t_g^{-1}$ and $t_g\alpha_g^{-1}t_g^{-1}$.
\end{itemize}
(Note that in the first case $t_{g^{-1}}=t_g$, while in the second case $t_{g^{-1}}=t_g^{-1}$).
\end{remark}
This gives a relationship between the dot products between $X$ and $X^{-1}$ associated to $g$ and $g^{-1}$.
\begin{remark} \label{dot product of inverses} Let $(P,Q)\in \mathcal I(\alpha,\alpha^{-1})$ correspond to $g\in T_1(\alpha)$, and denote by $(Q',P')$ the intersection pair corresponding to $g^{-1}$. Then we have $$(X\cdot_{(Q',P')}X^{-1})=(X\cdot_{(P,Q)}X^{-1})^{-1} $$
\end{remark}
In particular, these dot products cannot be equal, as a non-trivial loop is not conjugate to its inverse. Next we define the curves that will help us write reduced forms for the dot products.
\begin{definition} \label{d:c1c2} Let $g\in T_1(\alpha)$. We define the curves $c_1(\alpha,g)$ and $c_2(\alpha,g)$ according to whether $g$ preserves of reverses orientation:
\begin{itemize}
\item If $g$ preserves orientation, $$c_1(\alpha,g) \mbox{ is the reduced form of } \alpha_gt_g^{-1} \mbox{ and }$$ $$ c_2(\alpha,g) \mbox{ is the reduced form of } t_g\beta_g$$
\item If $g$ reverses orientation, let $$c_1(\alpha,g)= \alpha_g \mbox{ and } c_2(\alpha,g)= \beta_g $$
\end{itemize}
\end{definition}
As an example, in Figure \ref{fig:2} we have $c_1(\alpha, g)=b$ and
$c_2(\alpha, g)=e^{-1}f^{-1}$.
Let us interpret this definition in terms of the construction of the intersection pair $(P_g,Q_g)$ given in subsection \ref{ss.lift int1}. In the orientation preserving case we had $P_g=(\alpha_1,t_g)$ and $Q_g=(\beta_1,t_g^{-1})$, and recalling the construction we get $$\alpha_g=\alpha_1 t_g \mbox{ and } \beta_g = t_g^{-1}\beta_1, \mbox{ thus } c_1(\alpha,g)=\alpha_1 \mbox{ and } c_2(\alpha,g)=\beta_1$$
In the orientation reversing case we had $P_g=(\alpha_1,t_g)$ and $Q_g=(\beta_1,t_g)$ and we get $$\alpha_g=\alpha_1 t_g \mbox{ and } \beta_g = \beta_1 t_g, \mbox{ and so } c_1(\alpha,g)=\alpha_1 t_g \mbox{ and } c_2(\alpha,g)=\beta_1 t_g$$
In both cases we have that the concatenation $c_1(\alpha,g)c_2(\alpha,g)$ is the cyclically reduced form of $\alpha_g\beta_g$. Also note that:
\begin{remark}\label{pair2}
If $g\in T_1(\alpha)$ is orientation reversing, then $c_1(\alpha,g)$ is a permutation of $\alpha$ and $c_2(\alpha,g)$ is a permutation of $\alpha^{-1}$. In particular $l(c_1(\alpha,g))=l(c_2(\alpha,g))=l(\alpha)$.
\end{remark}
\begin{remark}\label{pair3}
On the other hand, if $g\in T_1(\alpha)$ is orientation preserving we have $l(c_1(\alpha,g))=l(c_2(\alpha,g))=l(\alpha)-l(t_g)$.
\end{remark}
So in any case the lengths of $c_1(\alpha,g)$ and $c_2(\alpha,g)$ agree.
We can also deduce the length of the dot product (i.e. the length of a cyclically reduced form), as follows:
\begin{remark} \label{r:int length}
For $g\in T_1(\alpha)$ we have:
\begin{itemize}
\item ${l(c_1(\alpha,g)c_2(\alpha,g))= 2l(\alpha)-2l(t_g)}$ if $g$ preserves orientation, and
\item $l(c_1(\alpha,g)c_2(\alpha,g))= 2l(\alpha)$ if $g$ reverses
orientation.
\end{itemize}
\end{remark}
Recalling the relationship between orientation and the type of intersection pairs from subsection \ref{ss.lift int1}, Remark \ref{r:int length} implies that if $c_1(\alpha,g)c_2(\alpha,g)$ is a permutation of $c_1(\alpha,h)c_2(\alpha,h)$ for $g,h\in T_1(\alpha)$ then either:
\begin{itemize}
\item $(P_g,Q_g)$ and $(P_h,Q_h)$ are both of types (2) or (1) (i.e. $g$ and $h$ are orientation reversing), or
\item $(P_g,Q_g)$ and $(P_h,Q_h)$ are both of type (3) (i.e. $g$ and $h$ are strictly orientation preserving), and $l(t_g)=l(t_h)$.
\end{itemize}
This observation is an example of recovering information about the intersection pair from the corresponding dot product. In the following sections we will be proving stronger results within this same idea, which will ultimately lead us to Theorem \ref{simple} by showing there can be no cancellations in the formula for $[X,X^{-1}]$.
Next we record the behaviour of the curves from Definition \ref{d:c1c2} under taking inverses in $T_1(\alpha)$, which we can compute from Remark \ref{rm:inv a b}.
\begin{remark} \label{rm:inv c1c2} Let $g\in T_1(\alpha)$, then
\begin{itemize}
\item if $g$ preserves orientation, $$c_1(\alpha,g^{-1})=c_2(\alpha,g)^{-1} \mbox{ and } c_2(\alpha,g^{-1})= c_1(\alpha,g)^{-1}$$
\item if $g$ reverses orientation,
$$c_1(\alpha,g^{-1}) \mbox{ is the reduced form of } t_g c_2(\alpha,g)^{-1} t_g^{-1} \mbox{ and }$$ $$ c_2(\alpha,g^{-1}) \mbox{ is the reduced form of } t_g c_1(\alpha,g)^{-1} t_g^{-1}$$
\end{itemize}
\end{remark}
The next result lets us write the dot products as commutators in $\mathcal L_x(M)$.
\begin{lemma}\label{formula2}
Let $g\in T_1(\alpha)$, and put $c_i=c_i(\alpha,g)$ for $i=1,2$. Then we have
\begin{enumerate}
\item
$[\alpha]g[\alpha]^{-1}g^{-1}= [\gamma_g c_1 c_2 \gamma_g^{-1}] $, and
\item
if $(P,Q)\in \mathcal I(\alpha,\alpha^{-1})$ corresponds to $g$ by the bijection of Lemma \ref{bijection}, then
$(X \cdot_{(P,Q)}X^{-1})$ is the conjugacy class of $[\alpha]g[\alpha]^{-1}g^{-1}$.
\end{enumerate}
\end{lemma}
\begin{proof} The second point follows from the first and the fact that the cyclically reduced form of $[\gamma_g c_1 c_2 \gamma_g^{-1}]$, which is $c_1c_2$, represents the dot product $(X \cdot_{(P,Q)}X^{-1})$ as discussed previously in this subsection.
We show the first point in the statement for $g$ corresponding to an intersection pair of type (3), as the other cases result from a straightforward adaptation of the same computations. In fact, type (1) can be considered within types either (2) or (3) by allowing constant core curves.
Whithin the case of $(P,Q)$ being of type (3), we distinguish 3 subcases according to whether $id_x$ belongs to both $a_g$ and $b_g$, to only one of them, or to neither of them. First we assume that $id_x$ is in $b_g$ but not in $a_g$. This is the situation shown in Figure \ref{fig:2}. Then we can write $\alpha=abc=ecaf$ with $g=[c^{-1}e^{-1}]$, noting that it corresponds to an intersection pair of type (3). We have $\gamma_g=a$, $c_1=b$ and $c_2=e^{-1}f^{-1}$. Thus we get $[\gamma_g c_1 c_2 \gamma_g^{-1}]=abe^{-1}f^{-1}a^{-1}$. On the other hand we compute
\begin{equation}\label{eq1}
g[\alpha]^{-1}g^{-1}=[(c^{-1}e^{-1}) (f^{-1}a^{-1}c^{-1}e^{-1})(ec)]
\end{equation}
Thus
\[[\alpha]g[\alpha]^{-1}g^{-1}=[(abc)(c^{-1}e^{-1}) (f^{-1}a^{-1}c^{-1}e^{-1})(ec) ]= [\gamma_gc_1c_2\gamma_g^{-1}] \]
as desired. The situation is symmetrical for $id_x$ in $a_g$ but not in $b_g$.
In the second subcase, when $id_x$ is in both $a_g$ and $b_g$, we have that $g\in b_g(I)$. If $g$ lies before $id_x$ in the orientation of $\Lambda_{\alpha}$ we can write $\alpha= abcc' = c'afc$, where $t_g=cc'a$ and $g=[c'^{-1}]$. Now $\gamma_g=a$, $c_1=b$ and $c_2=f^{-1}$, and we compute \begin{equation}\label{eq2}
g[\alpha]^{-1}g^{-1}=[c'^{-1}(c^{-1}f^{-1}a^{-1}c'^{-1})c']
\end{equation}
and
\[ [\alpha]g[\alpha]^{-1}g^{-1} = [(abcc')c'^{-1}(c^{-1}f^{-1}a^{-1}c'^{-1})c'] = [abf^{-1}a^{-1}], \]
that is $[\gamma_g c_1c_2\gamma_g^{-1}]$. When $g$ lies after $id_x$ in $\Lambda_{\alpha}$ we write $\alpha=aa'bc = a'fca$, where $t_g=caa'$ and $g=[a']$, and the computation is similar.
Finally, if $id_x$ is neither in $a_g$ nor $b_g$, we have $\alpha=abc=ebf$ with $t_g=b$ and $g=[ae^{-1}]$. We see that $\gamma_g=ab$, $c_1=ca$ and $c_2=e^{-1}f^{-1}$. On the other hand
\begin{equation}\label{eq3}
g[\alpha]^{-1}g^{-1}=[(ae^{-1}) (f^{-1}b^{-1}e^{-1})(ea^{-1})]
\end{equation}
thus we get
\[ [\alpha]g[\alpha]^{-1}g^{-1}= [ab(ca e^{-1} f^{-1})b^{-1} a^{-1} ]=[\gamma_g c_1c_2\gamma_g^{-1}] \]
\end{proof}
\subsection{Conjugate dot products}
Our strategy for Theorem \ref{simple} is to show that for a primitive string $X$ there can be no cancellations in the formula for $[X,X^{-1}]$ given in Definition \ref{d:bracket}. This would imply that $[X,X^{-1}]=0$ only when $LP_2(X,X^{-1})=\emptyset$, provided that $X$ is primitive, thus proving Theorem \ref{simple}. So we need to study what happens if two linked pairs between $X$ and $X^{-1}$ yield the same dot product. Here we shall focus on what we can achieve for general intersection pairs, leaving the discussion of linked pairs and their signs for the next section.
Assume that $g, h\in T_1(\alpha)$ are such that $c_1(\alpha,g)c_2(\alpha, g)$ is a permutation (maybe trivial) of the curve $c_1(\alpha,
h)c_2(\alpha,h)$, which is to say that their corresponding intersection pairs yield the same dot product. So we have
\[[c_1(\alpha, h)c_2(\alpha, h)]=[rc_1(\alpha,g)c_2(\alpha, g) r^{-1}] \]
where $r\in\Omega$ is an initial segment of $c_1(\alpha, h)c_2(\alpha, h)$.
By Remark \ref{r:int length} and the discussion preceeding it, the curves $c_1(\alpha, h)$, $c_2(\alpha, h)$, $c_1(\alpha, g)$ and $c_2(\alpha, g)$ have all the same length. Thus we may assume that $r$ is an initial segment of $c_1(\alpha, h)$ (otherwise we exchange the roles of $g$ and $h$), and find $s,\,t,\,u\in\Omega$ such that
\begin{equation}\label{c1h}
c_1(\alpha, h)=rs \;\;\;\;\; c_2(\alpha, h)=tu
\end{equation}
\begin{equation}\label{c1g}
c_1(\alpha, g)=st \;\;\;\;\; c_2(\alpha, g)=ur
\end{equation}
where $l(r)=l(t)$ and $l(s)=l(u)$. In particular, $r$ is constant iff $t$ is constant (trivial permutation case), and $s$ is constant iff $u$ is constant. (Note: $t$ is not to be confused with $t_g$ nor $t_h$). Let
\begin{equation}\label{8}
\phi=[\gamma_h r \gamma_g^{-1}]\in\mathcal L_x(M)
\end{equation}
Then by Lemma \ref{formula2} we have
\begin{equation} \label{phi conj}
\phi[\alpha]g [\alpha]^{-1}g^{-1}\phi^{-1}= [\alpha]h[\alpha]^{-1}h^{-1}
\end{equation}
Lemma \ref{formula2} also gives the converse: if $g, h\in T_1(\alpha)$ are so that $[\alpha]g [\alpha]^{-1}g^{-1}$ and $[\alpha]h [\alpha]^{-1}h^{-1}$ are conjugate in $\mathcal L_x(M)$, then $c_1(\alpha,g)c_2(\alpha, g)$ is a permutation of $c_1(\alpha,h)c_2(\alpha,h)$.
\begin{lemma}\label{reversing2} Assume that $\alpha$ is primitive, and that $g, h\in T_1(\alpha)$ are orientation reversing and so that
$[\alpha]h[\alpha]^{-1}h^{-1}$ and $[\alpha]g[\alpha]^{-1}g^{-1}$
are conjugate. Then
\[g=h \] in particular
\[ c_i(\alpha,g)= c_i(\alpha,h) \qquad \mbox{ for }i=1,2.\]
\end{lemma}
\begin{proof}
Recall that since $g$ and $h$ are orientation reversing we have $c_1(\alpha,g)=\alpha_g$, $c_2(\alpha,g)=\beta_g$, $c_1(\alpha,h)=\alpha_h$ and $c_2(\alpha,h)=\beta_h$, which will be useful through the proof.
Let $\bar \alpha$ be the projection of $\tilde \alpha$ onto $\Lambda_{\alpha}/G_h$ as in Lemma \ref{quotient}. Note that $\mathcal E_x(M)/G_h$ is an intermediate bundle over $M$, and has a notion of horizontal lift, by projecting the one in $\mathcal E_x(M)$ (which is equivariant). Throughout this proof we shall consider horizontal lifts to $\mathcal E_x(M)/G_h$ repeatedly, and refer to them simply as ``lifts''.
Let $\bar \gamma_h$ be the lift of $\gamma_h$ starting at $\bar\gamma_h(0)=\bar \alpha(0)$, which is the same as the projection of $\tilde\gamma_h$, thus it is contained in $\Lambda_{\alpha}/G_h$ and $\bar\gamma_h(1)$ is the projection of $b_h(1)$. Let $\bar c_1(\alpha,h)\bar c_2(\alpha,h)$ be the lift of $c_1(\alpha,h) c_2(\alpha,h)$ beginning at $\bar \gamma_h(1)$. We see that this curve is closed and contained in $\Lambda_{\alpha}/G_h$, by observing that $\bar c_1(\alpha,h)$ and $\bar c_2(\alpha,h)$ are the respective projections of $\tilde\alpha_h\subset \Lambda_{\alpha}$ and $h^{-1}\tilde\beta_h\subset \Lambda_{\alpha}$, thus each one is a closed curve at $\bar\gamma_h(1)$.
Next we take $\bar r$ the lift of $r$ starting at $\bar\gamma_h(1)$, noting that it is an initial segment of $\bar c_1(\alpha,h)$. Consider $\bar c_1(\alpha,g)\bar c_2(\alpha,g)$ the lift of $c_1(\alpha,g)c_2(\alpha,g)$ that starts at $\bar r(1)$ (recalling that $c_1(\alpha,g)$ begins at $r(1)$).
{\sl Claim 1}: $\bar c_1(\alpha,g)\bar c_2(\alpha,g)$ is closed and contained in $\Lambda_{\alpha}/G_h$.
To show this claim, write $c_1(\alpha,g)c_2(\alpha,g) = stur$, and start by taking the lift $\bar s$ of $s$ starting at $\bar r(1)$. Since $rs = c_1(\alpha,h)$ we have that $\bar r\bar s = \bar c_1(\alpha,h)$, and thus $\bar s\subset \Lambda_{\alpha}/G_h$ and $\bar s(1) = \bar\gamma_h(1)$. We continue lifting $c_1(\alpha,g)c_2(\alpha,g) = stur$ by taking the lift of $tu$ beginning at $\bar s(1) =\bar\gamma_h(1)$, and we notice that this lift agrees with $\bar c_2(\alpha,h)$ since $tu=c_2(\alpha,h)$. In particular it ends at $\bar\gamma_h(1)$, so $\bar r$ is the lift of $r$ that we need to complete the lifting of $c_1(\alpha,g)c_2(\alpha,g) = stur$. Thus we get that $\bar c_1(\alpha,g)\bar c_2(\alpha,g)=\bar s \bar c_2(\alpha,h) \bar r $, which is closed at $\bar r(1)$ and clearly contained in $\Lambda_{\alpha}/G_h$.
{\sl Claim 2:} For $i=1,2$ we have
\begin{equation}\label{firsteq}
\bar c_i(\alpha,g)=\bar c_i(\alpha,h).
\end{equation}
We can consider $\mathcal L_{\bar\alpha(0)}(\Lambda_{\alpha}/G_h)$ as a subgroup of $\mathcal L_x(M)$, since projection induces an injective homomorphism, and then Lemma \ref{quotient} implies that $T_1(\bar\alpha)=\{h,h^{-1}\}$. By construction we have $$ \bar c_i(\alpha,h)= c_i(\bar\alpha,h) \mbox{ for } i=1,2.$$
On the other hand, by Claim 1 we have that $\bar c_1(\alpha,g)$ and $\bar c_2(\alpha,g)$ lie inside $\Lambda_{\alpha}/G_h$, and we recall that they are lifts of $\alpha_g$ and $\beta_g$ respectively, which are permutations of $\alpha$ and $\alpha^{-1}$. Therefore $\bar c_1(\alpha,g)$ and $\bar c_2(\alpha,g)$ are permutations of $\bar\alpha$ and $\bar\alpha^{-1}$ respectively (in particular they are closed).
Then we get that $\bar c_1(\alpha,g)\bar c_2(\alpha,g)$ represents the dot product for an intersection pair of $\bar\alpha$, and since it is cyclically reduced we must have $$\bar c_i(\alpha,g)= c_i(\bar\alpha,h^{\epsilon}) \mbox{ for } i=1,2 \mbox{ and for some } \epsilon=\pm 1 $$
Since we have $\bar c_1(\alpha,h)\bar c_2(\alpha,h)=\bar r \bar c_1(\alpha,g)\bar c_2(\alpha,g) \bar r^{-1}$, we can apply Remark \ref{dot product of inverses} to get that $\epsilon=1$, proving this claim.
To finish the proof of the lemma recall that, since we are in the orientation reversing case, we have $c_1(\alpha,g)=\alpha_g$ and $c_1(\alpha,h)=\alpha_h$, which are permutations of $\alpha$. Equation \eqref{firsteq} implies, by projecting, that $\alpha_g=\alpha_h$, and then Lemma \ref{primitive} gives $g=h$, since $\alpha$ is primitive.
\end{proof}
Putting Lemma \ref{reversing2} together with Remark \ref{r:int length}, we see that a dot product of the form $(X\cdot_{(P,Q)}X^{-1})$ with length $2l(\alpha)$ comes from a unique intersection pair $(P,Q)$, which is of type either (1) or (3).
One would like to remove the orientation reversal condition from the hypothesis of Lemma \ref{reversing2}, for that would give a stronger result than Theorem \ref{simple}, namely that the strings in the terms of the formula for $[X,X^{-1}]$ cannot repeat. Unfortunately this remains open. Next is the result we can get when dropping said orientation condition, which will suffice for our purpose.
\begin{lemma}\label{general1} Assume that $\alpha$ is primitive, and that $g, h\in T_1(\alpha)$ are such that $[\alpha]h[\alpha]^{-1}h^{-1}$ and $[\alpha]g[\alpha]^{-1}g^{-1}$
are conjugate. Then
\[ c_i(\alpha,g)= c_i(\alpha,h) \qquad \mbox{ for }i=1,2.\]
\end{lemma}
\begin{proof}
We shall assume $g$ and $h$ preserve orientation strictly, since Remark \ref{r:int length} implies that the only other possible case is the one covered by Lemma \ref{reversing2}.
Recall the notation from Equations \eqref{c1h} and \eqref{c1g}, i.e. the curves $r,s,t, u$ and their properties. Notice that we can prove the lemma by showing that $t$, and hence $r$, are constant.
Firstly we shall define some horizontal lifts in $\mathcal E_x(M)$ that will be useful through the proof: By Equation \eqref{c1g} we can take $\bar s$ the lift of $s$ starting at $\tilde \gamma_g(1)=b_g(1)$, and $\bar t$ the lift of $t$ starting at $\bar s(1)$. Note that $\bar s\bar t$ is the lift of $c_1(\alpha,g)$ that is an initial segment of $\tilde{\alpha}_g$, thus it is contained in $\Lambda_{\alpha}$ and ends at $\bar t(1)=[\alpha]b_g(0)$, recalling Definition \ref{d:c1c2} in the orientation preserving case. We also see that $\bar s$ and $\bar t$ are positively oriented in $\Lambda_{\alpha}$.
We also define $\tilde r$ as the lift of $r$ starting at $\tilde\gamma_h(1)=b_h(1)$, and $\tilde s$ as the lift of $s$ starting at $\tilde r(1)$, which are well defined by Equation \eqref{c1h}. Again we see that $\tilde r\tilde s$ is the lift of $c_1(\alpha,g)$ which is an initial segment of $\tilde \alpha_h$, and so it is contained in $\Lambda_{\alpha}$ and ends at $\tilde s(1)=[\alpha]b_h(0)$. Also, $\tilde r$ and $\tilde s$ are positively oriented in $\Lambda_{\alpha}$.
Finally, let $\tilde t$ be the lift of $t$ starting at $\tilde s(1)$, which is well defined since $t(0)=s(1)$ from Equation \eqref{c1g}. We see from Equation \eqref{c1h} that $\tilde t$ is an initial segment of a lift of $c_2(\alpha,h)$, namely the one contained in $[\alpha]\tilde\beta_h$, since it starts at $[\alpha]b_h(0)$. Thus $\tilde t$ is contained in $[\alpha]h\Lambda_{\alpha}$, meeting $\Lambda_{\alpha}$ only at $\tilde t(0)$, and it goes in the negative direction with respect to the orientation of $[\alpha]h\Lambda_{\alpha}$ induced by $[\alpha]h$.
Next we recall Equation \eqref{8}, defining $\phi=[\gamma_h r \gamma_g^{-1}]$. Notice that, by the above definitions, we have $\phi \tilde \gamma_g(1) = \tilde r(1)$, in particular $\phi\in T(\Lambda_{\alpha})$. Since the action of $\mathcal L_x(M)$ preserves horizontal lifting, we also get that $\phi \bar s = \tilde s$ and $\phi \bar t= \tilde t$.
{\sl Claim 1}: If $\phi\in T^+(\Lambda_{\alpha})$ then $t$ is constant.
Let $G$ be the subgroup generated by $g,h,\phi$ and $[\alpha]$. Then $G$ is $\alpha$-oriented by Lemma \ref{oriented}, so there is a $G$-invariant orientation on $G\Lambda_{\alpha}$. On the other hand, $\phi \bar t=\tilde t$ where $\bar t$ has positive orientation in $\Lambda_{\alpha}$ but $\tilde t$ has negative orientation in $[\alpha]h\Lambda_{\alpha}$. This is a contradiction unless $t$ is constant.
{\sl Claim 2}: If $s$ is non-constant, then $\phi\in T^+(\Lambda_{\alpha})$.
We have $\phi \bar s=\tilde s$ where both $\bar s$ and $\tilde s$ are contained in $\Lambda_{\alpha}$, thus $\bar s\subset b_{\phi}(I)$ and $\tilde s\subset a_{\phi}(I)$. Recall that both $\bar s$ and $\tilde s$ are positively oriented in $\Lambda_{\alpha}$, thus showing that $\phi$ preserves orientation, if $s$ is non-constant.
Recall that we prove the lemma by showing that $t$ is constant. Thus, in light of these claims, the only case that remains to be considered is when $s$ is constant and $\phi$ reverses orientation strictly. We shall show that this case is void, which makes sense as $s$ and $t$ cannot be both constant in Equation \eqref{c1g}. Thus we assume that $s$ is constant and $\phi$ strictly reverses orientation, aiming to reach a contradiction.
This case is the most complex part of this proof, the key will be to consider the dot product defined by $\phi$. Note that this can be defined even if $\phi$ is not in $T_1(\alpha)$, through Lemma \ref{mn}, and the same is true for the curves $c_1(\alpha,\phi)$ and $c_2(\alpha,\phi)$ from Definition \ref{d:c1c2}. In order to simplify notation we put $c_i(\phi)=c_i(\alpha,\phi)$ for $i=1,2$, and define $$ \omega = [\alpha] \phi [\alpha]^{-1}\phi^{-1} = [\gamma_{\phi}c_1(\phi)c_2(\phi)\gamma_{\phi}^{-1} ] $$
{\sl Claim 3:} $\omega$ is primitive.
By change of basepoint, i.e. Lemma \ref{bpoint}, this is equivalent to show that $c_1(\phi)c_2(\phi)$ is primitive. Assume the contrary, i.e. that there is a closed curve $\tau$ and $k>1$ so that $\tau^k=c_1(\phi)c_2(\phi)$. Note that $\tau$ must be cyclically reduced, since $c_1(\phi)c_2(\phi)$ is, and that $l(c_1(\phi))=l(c_2(\phi))$ by Remark \ref{pair2}.
If $k$ is even, we write $k=2j$ and get that $c_1(\phi)=\tau^j=c_2(\phi)$, which is absurd by Remark \ref{pair2} and the assumption that $\phi$ reverses orientation, since a permutation of $\alpha$ cannot agree with a permutation of $\alpha^{-1}$. If $k$ is odd, we write $k=2j+1$ and we get that $$\tau=vw \,\,\mbox{ with }\,\, c_1(\phi)=(vw)^jv \,\,\mbox{ and }\,\, c_2(\phi)=w(vw)^j $$ where we also have $l(v)=l(w)$. Again by Remark \ref{pair2}, we get that $(w^{-1}v^{-1})^jw^{-1}$ is a permutation of $(vw)^jv$. Since $l(v)=l(w)$ and $v$ cannot overlap $v^{-1}$ (nor $w$ overlap $w^{-1}$), we must have $v=w^{-1}$, which is absurd since $\tau$ cannot be trivial. Thus we have shown Claim 3.
{\sl Claim 4:} There is a non-trivial $\theta\in T(\Lambda_{\alpha})$ such that $\theta\omega\theta^{-1}=\omega$.
To show this claim it will be useful to write Equations \eqref{c1h} and \eqref{c1g} for the case when $s$ is constant:
\begin{equation}\label{c1c2 s cte}
c_1(\alpha,h)=c_2(\alpha,g)=r \,\, \mbox{ and }\,\, c_1(\alpha,g)=c_2(\alpha,h)=t,
\end{equation}
and we also get that $\bar t(0) = \tilde \gamma_g(1)=b_g(1)$ and $\tilde r(1)=\tilde t(0)=[\alpha]b_h(0)$. By Remark \ref{rm:inv c1c2} we deduce that
\begin{equation}\label{c1c2 -1 s cte}
c_1(\alpha,h^{-1})=c_2(\alpha,g^{-1})=t^{-1} \,\, \mbox{ and }\,\, c_1(\alpha,g^{-1})=c_2(\alpha,h^{-1})=r^{-1}
\end{equation}
Define $$\psi = h^{-1}\phi g. $$
Let us check that $\psi\in T^-(\Lambda_{\alpha})$.
Note first that, since $g$ preserves orientation, we have $\tilde\gamma_{g^{-1}}(1)=b_{g^{-1}}(1)=a_g(1)$ by Remark \ref{inverse}. By Equation \eqref{c1c2 -1 s cte} we may take $\hat r$ the lift of $r$ that ends at $\hat r(1)=a_g(1)$, i.e. so that $\hat r^{-1}$ lifts $c_1(\alpha,g^{-1})$ starting at $a_g(1)=\tilde\gamma_{g^{-1}}(1)$. Thus we see that $\hat r^{-1}$ is contained in $\Lambda_{\alpha}$ and is positively oriented, since it is a segment of $\tilde \alpha_{g^{-1}}$. We also see that $\hat r(0)=\hat r^{-1}(1) = [\alpha]a_g(0)$.
We shall describe the action of $\psi = h^{-1}\phi g$ on $\hat r$. First we notice that $g\hat r$ ends at $ga_g(1)=b_g(1)=\tilde\gamma_g(1)$. Next we recall that $\phi \tilde\gamma_g(1)=\tilde r(1)$, so by equivariance of the horizontal lifting we get that $\phi g\hat r=\tilde r$. Finally we get that $\psi\hat r =h^{-1}\tilde r$. Recalling Equation \eqref{c1c2 -1 s cte} we see that this curve lifts $c_2(\alpha,h^{-1})^{-1}$ starting at $h^{-1}\tilde r(0)=h^{-1}b_h(1)=a_h(1)$. Thus $\psi\hat r$ is contained in $h^{-1}\Lambda_{\alpha}$ and meets $\Lambda_{\alpha}$ exactly at $\psi\hat r(0)=a_h(1)$. This shows that $\psi\in T(\Lambda_{\alpha})$, and it must reverse orientation strictly by Lemma \ref{oriented} since $\phi=h\psi g^{-1}$.
Now we set $h^{\ast}=[\alpha]h[\alpha]^{-1}$, and a simple computation from Equation \ref{phi conj} and the definition of $\psi=h^{-1}\phi g$ gives us
\begin{equation} \label{dot pr psi} h^{\ast} [\alpha] \psi [\alpha]^{-1}\psi^{-1}{h^{\ast}}^{-1}=[\alpha] \phi [\alpha]^{-1}\phi^{-1}
\end{equation}
By Remark \ref{mn} there are integers $i, j, k, l$
such that $[\alpha]^i\phi[\alpha]^j$ and $[\alpha]^k\psi[\alpha]^l$ belong to $T_1(\alpha)$. Now we apply Lemma \ref{reversing2} to these elements and Equation \eqref{dot pr psi}, hence for $n=i-k$, $m=k-l$ we obtain
$$
\psi=[\alpha]^n\phi[\alpha]^m
$$
and Equation \eqref{dot pr psi} becomes
$$
(h^{\ast} [\alpha]^n)[\alpha]\phi [\alpha]^{-1}\phi^{-1}({h^{\ast}}[\alpha]^n)^{-1}=[\alpha] \phi [\alpha]^{-1}\phi^{-1}
$$
So we set $\theta=h^{\ast}[\alpha]^n$ and get that $\theta\omega\theta^{-1}=\omega$. Note that $\theta=[\alpha]h[\alpha]^{n-1}$ is not trivial, since $h$ is not a power of $[\alpha]$, and belongs to $T(\Lambda_{\alpha})$, since $\theta\Lambda_{\alpha}=[\alpha]h\Lambda_{\alpha}$ intersects $\Lambda_{\alpha}$ in the segment $[\alpha]b_h(I)$. Thus we have shown Claim 4.
Putting $C=\alpha(I)$ and recalling Lemma \ref{lem:ident}, we see that $\omega,\theta\in \mathcal L_x(C)$ which is a free group, so Claims 3 and 4 imply that
\begin{equation} \label{c-phi}
\theta=\omega^i \,\,\mbox{ for some } i\in\field{Z}, i\neq 0
\end{equation} We shall reach a contradiction by showing that $\theta\Lambda_\alpha$ is disjoint from $\Lambda_{\alpha}$, i.e.that $\theta\notin T(\Lambda_{\alpha})$, against Claim 4. Since $T(\Lambda_{\alpha})$ is closed under taking inverses (Remark \ref{inv T}), we may assume that $i>0$.
Changing the basepoint if necessary, we may assume that $b_{\phi}(1)=id_x$. Note that such a change of basepoint ammounts to repalce $\alpha$ by a permutation of it, and does not change the curves $c_1(\phi)$ and $c_2(\phi)$. By Lemma \ref{bpoint} this change of basepoint also preserves Claims 3 and 4, and Equation \eqref{c-phi}. From $b_{\phi}(1)=id_x$ we get that $\gamma_{\phi}$ is constant, $c_1(\phi)=\alpha$ and $c_2(\phi)$ is a non-trivial permutation of $\alpha^{-1}$.
Let $\tilde c_2(\phi)$ be the horizontal lift of $c_2(\phi)$ that ends at $b_{\phi}(1)=id_x$. Since $\phi$ reverses orientation, the discussion after Definition \ref{d:c1c2} implies that $\tilde c_2(\phi)$ intersects $\Lambda_{\alpha}$ exactly in the segment $b_{\phi}(I)$, and in particular $\tilde c_2(\phi)(0)\notin \Lambda_{\alpha}$. On the other hand we take $\eta$ the horizontal lift of $(c_1(\phi)c_2(\phi))^i$ that begins at $id_x$, so we have $\eta(1)=\omega^i=\theta$. Note that since $c_1(\phi)=\alpha$, we have a reduced factorization $$\eta=\tilde\alpha\nu$$ where $\nu$ meets $\Lambda_{\alpha}$ only at $\nu(0)$, since the lift of $c_2(\phi)$ starting at $[\alpha]=[\alpha]b_{\phi}(1)$ is an initial segment of $\nu$. Observe also that $\theta\tilde c_2(\phi)$, being the lift of $c_2(\phi)$ that ends at $\theta=\nu(1)$, is a final segment of $\nu$ (if $i=1$ it agrees with the initial segment just discussed). Since $\mathcal E_x(C)$ is a tree, the line $\theta\Lambda_{\alpha}$ cannot intersect $\Lambda_{\alpha}$ without containing $\nu(I)$, but acting by $\theta$ we see that $\theta\tilde c_2(\phi)$ intersects $\theta\Lambda_{\alpha}$ only at $\theta b_h(I)$, which does not contain $\theta \tilde c_2(\phi)(0)\in\nu(I)$. Therefore $\theta\notin T(\Lambda_{\alpha})$ and we have a contradiction, concluding the proof.
\end{proof}
\section{Signs of the terms of the bracket}
Here we shall study the signs of the linked pairs in $LP_2(X,X^{-1})$ for $X\in S(M)$, showing that linked pairs yielding the same dot product have also the same sign, and finally arriving at the proof of Theorem \ref{simple}. In order to do so, we may need to consider a small deformation of a curve $\alpha$ representing $X$.
\subsection{Deformations of 1-complexes}
Let $C$ be a one dimensional complex, and consider $p\in C$. Then a small enough neighborhood $B$ of $p$ in $C$ is homeomorphic to a wedge of intervals, each one with an endpoint at $p$ and the other in $\partial B$. We call the {\em valence} of $p$ in $C$ to the number of such segments. A point of valence $2$ will be called a {\em regular point} of $C$. Observe that the set of non-regular points is discrete, while the components of the set of regular points are arcs.
We will be interested in complexes of the form $C=\alpha(I)$ where $\alpha\in \Omega$ is a cyclically reduced, non-constant closed curve. Then $C$ has no points of valence $1$, and by compactness, the set of non-regular points (i.e. {\em branching points}) is finite. The closure of a component of the set of regular points is a segment with endpoints at non-regular points. Note that, replacing $\alpha$ by a permutation if necessary, we can assume that $x=\alpha(0)$ is a regular point.
Before introducing the perturbation of $\alpha$ that we need, it is worth recalling that $C_{\varepsilon}$ is the $\varepsilon$-neighborhood of $C$ in $M$, and for $\varepsilon$ small enough, Lemma \ref{tubular} states that $C_{\varepsilon}$ retracts by deformation onto $C$. Then the map induced by the inclusion $i_\ast: \pi_1(C, x)\rightarrow \pi_1( C_\varepsilon, x)$ is an isomorphism, and its inverse is $\chi_\ast: \pi_1(C_\varepsilon, x)\rightarrow \pi_1( C, x) $, which is induced by a retraction $\chi:C_{\varepsilon}\to C$.
\begin{lemma}\label{approximation}
Let $\alpha$ be a cyclically reduced closed curve and $C=\alpha(I)$. Assume that $x=\alpha(0)$ is a regular point of $C$.
Then there are $\varepsilon>0$ and a closed, cyclically reduced curve $\gamma\subset C_{\varepsilon}$ such that
\begin{enumerate}
\item $\gamma$ is an $\varepsilon$-perturbation of $\alpha$, with $\gamma(0)=x$,
\item $\Gamma=\gamma(I)$ has no points of valence greater than $3$, and
\item The map $i_\ast: \pi_1(\Gamma, x)\rightarrow \pi_1( C_\varepsilon, x)$ induced by the inclusion is an isomorphism.
\end{enumerate}
Moreover, $\gamma$ only differs from $\alpha$ in an $\varepsilon$-neighborhood of the non-regular points of $C$.
\end{lemma}
\begin{proof}
Set $\varepsilon$ as in Lemma \ref{tubular}, though we may need to reduce it further. For each non-regular point $p$ of $C$ we consider $B_p$ a closed ball in $M$ centered at $p$ with radius $\varepsilon/2$. Note that if $\varepsilon$ is small enough, then $B_p$ is a normal ball and $B_p\cap C$ is a union of geodesic segments joining $p$ to $\partial B_p$. Reducing $\varepsilon$ if necessary, we can also make the sets $B_p$ pairwise disjoint, and disjoint from $x$ by our assumption.
First we will construct the complex $\Gamma$. For each non-regular point $p$ of $C$, we take a segment $\eta\subset\alpha$ with endpoints in $\partial B_p$ and so that $\eta(I)\subset B_p$, i.e. $\eta$ is a segment of $\alpha$ that traverses $B_p$. Then we consider a piecewise geodesic complex $Y_p$ contained in $B_p$, with $Y_p\cap\partial B_p = C\cap\partial B_p$, and so that $Y_p$ is a finite tree containing $\eta(I)$, whose branching points lie in $\eta(I)$ and have valence $3$. Figure \ref{fig:6} shows an example of this construction. It can be interpreted as a deformation of $C\cap B_p$ that spreads out the segments that are not in $\eta$, so that their endpoints are spread along $\eta$ instead of converging at $p$. Then we define $\Gamma$ so that it agrees with $C$ in the complement of $B=\bigcup_p B_p$, and that $\Gamma\cap B_p=Y_p$ for every non-regular point $p$ in $C$.
\begin{figure}[htbp]
\input{fig9.pic}
\caption{proof of lemma \ref{approximation}}\label{fig:6}
\end{figure}
By construction, $\Gamma$ is a connected piecewise geodesic complex and all its non-regular points are of valence $3$. Note also that $B_p$ retracts by deformation to $Y_p$ for each $p$. These maps can be extended to a retraction by deformation $\chi_1:C_{\varepsilon}\to\Gamma$, thus obtaining point (3) in the statement.
Now we describe the curve $\gamma$. Write
\begin{equation}\label{alpha}
\alpha=\alpha_0\beta_1\alpha_1\cdots \beta_{n}\alpha_n
\end{equation}
where $\alpha_i$ is contained in the closure of the complement of $B$ for all $i=0,\ldots,n$, and $\beta_j$ is contained in $B$, for $j=1,\ldots,n$. None of these curves is constant since $x$ is outside $B$.
Then for each $j$ there is some $p$ with $\beta_j\subset B_p$, and we have $\beta_j(0),\beta_j(1)\in\partial B_p$. Note that, by construction, there is a unique reduced curve $\bar \beta_j$ joining $\beta_j(0)$ to $\beta_j(1)$ in $Y_p$. Thus $\bar\beta_j\subset \Gamma\cap B_p$ with $\bar\beta_j(0)=\beta_j(0)$ and $\bar\beta_j(1)=\beta_j(1)$. We define
\begin{equation}\label{gamma}
\gamma=\alpha_0\bar\beta_1\alpha_1\cdots \bar\beta_{n}\alpha_n
\end{equation}
i.e. we replace each $\beta_j$ with $\bar\beta_j$. This is an $\varepsilon$-perturbation of $\alpha$, since for each $j$, $\beta_j$ and $\bar\beta_j$ are in the same ball of radius $\varepsilon/2$. It is also clear that $\gamma$ admits no reductions, and that $\gamma(0)=\gamma(1)=x$. It only remains to show that $\gamma(I)=\Gamma$. By construction we have $\gamma(I)\subseteq\Gamma$, and it is also clear that the closure of $\Gamma-B$ is contained in $\gamma(I)$, writing this set as $\bigcup_i\alpha_i(I)$. So we must show that $Y_p\subset\gamma(I)$ for each non-regular point $p$ of $C$. For one such $p$ we note that the curve $\eta$ used in the construction of $Y_p$ is in $\gamma(I)$: by its definition $\eta=\beta_j$ for some $j$ (maybe more than one), and in that case we also have $\bar\beta_j=\eta$. On the other hand, since we have $\partial B_p\cap\Gamma\subset\gamma(I)$, the rest of the segments making up $Y_p$ must also be contained in $\gamma(I)$.
\end{proof}
\subsection{Signs and lifts}\label{liftsign}
Now we shall interpret the signs of linked pairs in terms of horizontal lifting, in the same fashion as we did for intersection pairs and dot products in the previous sections. Let $\alpha\in\Omega$ be a non-trivial cyclically reduced closed curve and $x=\alpha(0)$.
\begin{definition} \label{def:liftsign} For $g\in T_0(\alpha)$ we write $$\epsilon_g(\alpha)=sign(P_g,Q_g), $$ i.e. the sign of the linked pair of $\alpha$ corresponding to $g$. When the curve $\alpha$ is clear from the context, we just write $\epsilon_g$.
\end{definition}
Let $C=\alpha(I)$ and recall from Lemma \ref{lem:ident} that the universal cover $\tilde C$ is isomorphic, as a principal fiber bundle, to $\mathcal E_x(C)$. We choose such an isomorphism by picking $\tilde x\in \tilde C$ a lift of $x$, and setting that $id_x$ corresponds to $\tilde x$. Then we can identify $\Lambda_{\alpha}$ with the infinite lift of $\alpha$ to $\tilde C$ that starts at $\tilde x$.
Let $C_{\varepsilon}$ be a neighborhood of $C$ satisfying Lemma \ref{tubular}. Then its universal cover $\tilde C_{\varepsilon}$ contains $\tilde C$, and retracts onto it by lifting the retraction $\chi:C_{\varepsilon}\to C$. Note that the complement of $\Lambda_{\alpha}$ in $\tilde C_{\varepsilon}$ has two connected components. We give $\tilde C_{\varepsilon}$ the orientation lifted from that of $C_{\varepsilon}\subset M$, and let $C^+(\alpha)$ be the {\em left side} of $\Lambda_{\alpha}$. More precisely, $C^+(\alpha)$ is the component of $\tilde C_{\varepsilon}-\Lambda_{\alpha}$ so that the standard orientation of $\Lambda_{\alpha}$, given by $\tilde\alpha$, agrees with the orientation induced on $\Lambda_{\alpha}$ as part of the boundary $\partial C^+(\alpha)$. The other component, i.e. the {\em right side} of $\Lambda_{\alpha}$, will be denoted by $C^-(\alpha)$.
Let $a_1$, $a_2$, $\nu$ be small geodesic segments in $\tilde C_{\varepsilon}$, so that $a_1(1)=a_2(0)$ and $\nu$ meets $a_1a_2$ only at this point, which is an endpoint of $\nu$. Note that we have $sign(a_1a_2,\nu)=sign(\pi(a_1a_2),\pi(\nu))$, by definition of the orientation of $\tilde C_{\varepsilon}$.
In case that $a_1a_2$ is contained in $\Lambda_{\alpha}$ with positive orientation, then we have $sign(a_1a_2,\nu)=1$ if either $\nu(0)=a_1(1)$ and $\nu(1)\in C^+(\alpha)$, or $\nu(1)=a_1(1)$ and $\nu(0)\in C^-(\alpha)$. In the reverse cases we have $sign(a_1a_2,\nu)=-1$.
Now consider $g\in T_1(\alpha)$. We identify $\mathcal L_x(C)$ with $\pi_1(C_{\varepsilon},x)$ acting on $\tilde C_{\varepsilon}$ by deck transformations, using Lemmas \ref{lem:ident} and \ref{tubular}. Then, by definition of $T_1(\alpha)$, we have that $g[\alpha]\tilde x$ and $g[\alpha^{-1}]\tilde x$ are not contained in $\Lambda_{\alpha}$. Due to the previous observations, we see that $g$ corresponds to a linked pair, i.e. $g\in T_0(\alpha)$, iff $g[\alpha]\tilde x$ and $g[\alpha^{-1}]\tilde x$ are in different components of $\tilde C_{\varepsilon}-\Lambda_{\alpha}$. In that case, we have $\epsilon_g=1$ if $g[\alpha^{-1}]\tilde x\in C^-(\alpha)$, and thus $g[\alpha]\tilde x\in C^+(\alpha)$, which is to say that $g\Lambda_{\alpha}$ crosses $\Lambda_{\alpha}$ going from its right side and towards its left side. We have $\epsilon_g=-1$ if the reverse holds. We compile these results for future reference
\begin{remark} \label{rm:signs} Let $g\in T_1(\alpha)$, then
\begin{itemize}
\item $g\in T_0(\alpha)$ iff $g[\alpha]\tilde x$ and $g[\alpha^{-1}]\tilde x$ are on different sides of $\Lambda_{\alpha}$.
\item In that case, $\epsilon_g=1$ iff $g[\alpha^{-1}]\tilde x\in C^-(\alpha)$, i.e. is at the right side of $\Lambda_{\alpha}$.
\end{itemize}
\end{remark}
In a similar manner, we see that for $g\in T_0(\alpha)$ we have $\epsilon_g=1$ iff $\tilde\beta_g(1)\in C^-(\alpha)$. Equivalently, iff $g[\alpha]g^{-1}\tilde\beta_g(0)\in C^+(\alpha)$. Noting that $c_2(\alpha,g)$ ends at $t_g(1)$ by Definition \ref{d:c1c2}, we have the following.
\begin{remark} \label{rm:signs2} For $g\in T_0(\alpha)$ let $\tilde c_2(\alpha,g)$ be the lift of $c_2(\alpha,g)$ that ends at $b_g(1)$ (namely, the one contained in $g[\alpha]g^{-1}\tilde\beta_g$). Then $\epsilon_g=1$ iff $\tilde c_2(\alpha,g)$ is at the left of $\Lambda_{\alpha}$ (more precisely, is contained in the closure of $C^+(\alpha)$).
\end{remark}
By the comments after Definition \ref{d:c1c2} we see that the intersection of $\tilde c_2(\alpha,g)$ with $\Lambda_{\alpha}$ is either the endpoint $b_g(1)$, in case $g$ preserves orientation, or the segment $b_g(I)$, if $g$ reverses orientation. We also point out that the terminology of orientation preserving or reversing elements of Definition \ref{def:orient} does not relate to the orientation of $\tilde C_{\epsilon}$, which is preserved by every deck transformation.
\subsection{Signs and deformations}
Let $\alpha\in\Omega$ be a cyclically reduced closed curve, and assume that $x=\alpha(0)$ is a regular point of $C=\alpha(I)$. Take $\varepsilon$ and $\gamma$ as given by Lemma \ref{approximation}, and let $\Gamma=\gamma(I)$. Then we can identify $\mathcal E_x(\Gamma)$ with $\tilde \Gamma$, the universal cover of $\Gamma$, which is embedded in $\tilde C_{\varepsilon}$. This identifies $\mathcal L_x(\Gamma)$ with $\pi_1(C_{\varepsilon},x)$, and thus with $\mathcal L_x(C)$. We will be assuming these identifications in the sequel, though we should make clear that $\mathcal L_x(C)$ and $\mathcal L_x(\Gamma)$ are different as subgroups of $\mathcal L_x(M)$, and that $\mathcal E_x(C)\cup \mathcal E_x(\Gamma)$, as a subspace of $\mathcal E_x(M)$, is not homeomorphic to $\tilde C\cup\tilde\Gamma$. In fact, this correspondence identifies $[\gamma]\in \mathcal L_x(\Gamma)$ with $[\alpha]\in\mathcal L_x(C)$.
We also identify $\Lambda_{\gamma}$ with the infinite lift of $\gamma$ at $\tilde x$, the lift of $x$ used to define both correspondences $\mathcal E_x(C)\cong \tilde C$ and $\mathcal E_x(\Gamma)\cong\tilde\Gamma$. Thus the sets $T(\Lambda_{\alpha})$ and $T(\Lambda_{\gamma})$ can be considered as subsets of the same group $\pi_1(\tilde C_{\varepsilon},x)$, which we shall see as the group of deck transformations of $\tilde C_{\varepsilon}$.
\begin{lemma}\label{sign}
Let $\alpha$ and $\gamma$ be as in Lemma \ref{approximation}, and take $g\in \pi_1(\tilde C_{\varepsilon},x)$. Then $g\in T_0(\alpha)$ iff $g\in T_0(\gamma)$, in which case $\epsilon_g(\alpha)=\epsilon_g(\gamma)$.
\end{lemma}
\begin{proof}
Let $B$ be the $\varepsilon$-neighborhood of $\Lambda_{\alpha}$ in $\tilde C_{\varepsilon}$, which also contains $\Lambda_{\gamma}$. Moreover, we have $C^{\pm}(\alpha)\cap(\tilde C_{\varepsilon}-B)=C^{\pm}(\gamma)\cap(\tilde C_{\varepsilon}-B)$, i.e. points outside $B$ are in the same side with respect to both $\Lambda_{\alpha}$ and $\Lambda_{\gamma}$. Note that the construction of Lemma \ref{approximation} allows for reducing $\varepsilon$ as necessary, so we can assume that the translates of $\tilde x$ that do not belong to $\Lambda_{\alpha}$ are outside $B$.
If $g\in T_0(\alpha)$ we observed in Remark \ref{rm:signs} that $g[\alpha^{-1}]\tilde x$ and $g[\alpha]\tilde x$ are on different sides of $\Lambda_{\alpha}$, and since they are not in $B$, they are also on different sides of $\Lambda_{\gamma}$. Since $[\alpha]$ and $[\gamma]$ agree when seen as elements of $\pi_1(\tilde C_{\varepsilon},x)$, we see that $g\Lambda_{\gamma}$ contains both $g[\alpha^{-1}]\tilde x$ and $g[\alpha]\tilde x$, implying that $g\Lambda_{\gamma}$ and $\Lambda_{\gamma}$ intersect. This intersection corresponds to a linked pair by Remark \ref{rm:signs}. We get that $g\in T_1(\gamma)$ by recalling the definition of this set, together with the fact that $\alpha$ and $\gamma$ agree on a neighborhood of their basepoint, so every lift of $\gamma$ is an $\varepsilon$-perturbation of the corresponding lift of $\alpha$ that agrees with it in a neighborhood of its endpoints. With this we conclude that $g\in T_0(\gamma)$.
The reciprocal argument is analogous. We have that $\epsilon_g(\alpha)=\epsilon_g(\gamma)$ by Remark \ref{rm:signs}, together with the facts that $C^{\pm}(\alpha)\cap(\tilde C_{\varepsilon}-B)=C^{\pm}(\gamma)\cap(\tilde C_{\varepsilon}-B)$ and that $g[\alpha^{-1}]\tilde x$ and $g[\alpha]\tilde x$ lie outside $B$.
\end{proof}
Let $X$ and $Y$ be the strings represented by $\alpha$ and $\gamma$ as in Lemma \ref{approximation}. Then Lemma \ref{sign} gives a bijection between $LP_2(X,X^{-1})$ and $LP_2(Y,Y^{-1})$ that preserves the sign. It also preserves the dot product, by the isomorphism between $\mathcal L_x(C)$ and $\mathcal L_x(\Gamma)$ and Lemma \ref{formula2}. Thus proving Theorem \ref{simple} for $Y$ also implies it for $X$, i.e. we may replace $\alpha$ with $\gamma$ whenever necessary in our proof. We will only be doing this replacement at the points of the argument that require it, namely in the next Lemma \ref{general2}.
The proof of Lemma \ref{sign} clearly does not generalize for intersection pairs that are not linked. It is also possible to show Lemma \ref{sign} by following what happens to an intersection pair during the construction of $\gamma$ in Lemma \ref{approximation}, though some cases may get cumbersome, as well as the assertion about the signs. That approach would give that intersection pairs of types (2) and (3) of $\alpha$ are also present in $\gamma$, maintaining their types. Since $\Gamma$ has valence $3$ at every branching point, $\gamma$ has no intersection pairs of type (1), so the type (1) linked pairs of $\alpha$ will become linked pairs of type either (2) or (3) in $\gamma$. Some of the unlinked intersection pairs of type (1) of $\alpha$ may indeed be removed when passing to $\gamma$ (e.g. a suitable parametrization of a circle with three points identified). We do not need these assertions to show Theorem \ref{simple}, so we will not prove them.
\subsection{Signs and conjugation}
Next we show the main lemma that implies there are no cancellations in the formula for $[X,X^{-1}]$ when $X$ is primitive. After that, we shall finish the details of the proof of Theorem \ref{simple}. Let $\alpha\in\Omega$ be a cyclically reduced non-trivial closed curve, and $x=\alpha(0)$. We consider $C=\alpha(I)$ and $C_{\varepsilon}$ as in the rest of this section, identifying $\mathcal E_x(C)$ with $\tilde C$ embedded in $\tilde C_{\varepsilon}$.
\begin{lemma}\label{general2} Assume that $\alpha$ is primitive, and that $g, h\in T_0(\alpha)$ are such that $[\alpha]h[\alpha]^{-1}h^{-1}$ and $[\alpha]g[\alpha]^{-1}g^{-1}$ are conjugate. Then
\[\epsilon_g=\epsilon_h. \]
\end{lemma}
\begin{proof}
If $g$ and $h$ reverse orientation this is a consequence of Lemma \ref{reversing2}. So we assume that $g$ and $h$ preserve orientation, since by Remark \ref{r:int length} this is the other possible case. Recall the notation of Equations \eqref{c1h} and \eqref{c1g}, and let $\phi$ be the element defined in equation \eqref{8}. By Lemma \ref{general1} we have \begin{equation} \label{c1c2 - final}
c_i(\alpha,g)=c_i(\alpha,h) \qquad \mbox{for } i=1,2, \end{equation}
so $r$ and $t$ are constant in Equations \eqref{c1h} and \eqref{c1g}, and we have $\phi=[\gamma_h\gamma_g^{-1}]$. Therefore we get that $\phi b_g(1)=b_h(1)$.
Let $\tilde c_1(\alpha, g)$ be the lift of $c_1(\alpha, g)$ starting at $b_g(1)$, and $\tilde c_2(\alpha, g)$ the lift of $\tilde c_2(\alpha, g)$ ending at $b_g(1)$. Then $\tilde c_1(\alpha, g)$ is contained in $\Lambda_{\alpha}$ with positive orientation, and $\tilde c_2(\alpha, g)$ only meets $\Lambda_{\alpha}$ at its endpoint $b_g(1)$, since $g$ preserves orientation. By Remark \ref{rm:signs2}, the sign $\epsilon_g$ is decided by the side of $\Lambda_{\alpha}$ that $\tilde c_2(\alpha, g)$ is on. We may write
\begin{equation} \label{e-g}
\epsilon_g = sign(b_g\tilde c_1(\alpha, g), \tilde c_2(\alpha, g)) \end{equation} where we understand we are taking the intersections of these curves with a small enough ball centered at $b_g(1)$, as to follow the definition of sign in Subsection \ref{LPs}.
Similarly, we take $\tilde c_1(\alpha, h)$ as the lift of $c_1(\alpha, h)$ starting at $b_h(1)$, and $\tilde c_2(\alpha, h)$ as the lift of $\tilde c_2(\alpha, h)$ ending at $b_h(1)$. The same observations hold, in particular \begin{equation} \label{e-h}
\epsilon_h = sign(b_h\tilde c_1(\alpha, h), \tilde c_2(\alpha, h)) \end{equation}
Since $\phi b_g(1)=b_h(1)$ and deck transformations preserve horizontal lifting, Equation \eqref{c1c2 - final} implies that
\begin{equation} \label{tripod}
\phi \tilde c_1(\alpha, g)= \tilde c_1(\alpha, h)\;\;\;\;\; \mbox{and} \;\;\;\;\; \phi \tilde c_2(\alpha, g)= \tilde c_2(\alpha, h). \end{equation}
The situation is depicted in Figure \ref{fig:5}.
\begin{figure}[htbp]
\input{fig5.pic}
\caption{Proof of lemma \ref{general2}}\label{fig:5}
\end{figure}
First we shall complete the proof assuming the following extra condition:
{\sl Assumption 1:} There exists $\eta$ a final segment of $b_g$ so that $\phi\eta$ is a final segment of $b_h$.
Under this assumption, we note that Equation \eqref{tripod} gives that $$ sign(\eta\tilde c_1(\alpha, g), \tilde c_2(\alpha, g))= sign(\phi\eta\cdot\tilde c_1(\alpha, h), \tilde c_2(\alpha, h)) $$ since $\phi$, as a deck transformation, preserves the orientation of $\tilde C_{\varepsilon}$. When we combine this with Equations \eqref{e-g} and \eqref{e-h}, Assumption 1 gives us $$\epsilon_g=\epsilon_h $$ as desired.
Now observe that Assumption 1 holds if $t_g(1)=t_h(1)$ has valence $3$ in $C$: for then $b_g(1)$ and $b_h(1)$ would have valence $3$ in $\tilde C$, and since $\phi$ is bijective and satisfies Equation \ref{tripod}, we get that a small enough final segment of $b_g$ must be mapped by $\phi$ to a final segment of $b_h$. Figure \ref{fig:5} depicts this case.
Changing the basepoint of $\alpha$ if necessary, we consider the curve $\gamma$ in Lemma \ref{approximation}. By Lemma \ref{sign} we may replace $\alpha$ with $\gamma$ if needed to ensure Assumption 1.
\end{proof}
We finally complete the proof of Theorem \ref{simple}. Let $X\in S(M)$ be non-trivial and primitive, and $\alpha$ be a cyclically reduced representative of $X$.
{\sl Proof of Theorem \ref{simple}:} As we pointed out after Definition \ref{def:T0}, we have that $LP_1(X)=LP_2(X,X)\cong LP_2(X,X^{-1})$ which is in correspondence with $T_0(\alpha)$, recalling Lemma \ref{bijection} and Definition \ref{def:T0}. We rewrite the formula for $[X,X^{-1}]$ given in Definition \ref{d:bracket} using this correspondence, Lemma \ref{formula2} and Definition \ref{def:liftsign}, to get
\begin{equation} \label{corchete final}
[X,X^{-1}]=\sum_{g\in T_0(\alpha)} \epsilon_g\{[\alpha]g[\alpha]^{-1}g^{-1} \}
\end{equation} where $\{[\alpha]g[\alpha]^{-1}g^{-1} \}$ stands for the conjugacy class of $[\alpha]g[\alpha]^{-1}g^{-1}$.
We must show that if $[X,X^{-1}]=0$ then $LP_1(X)=\emptyset$, or equivalently, $T_0(\alpha)=\emptyset$. We show it by contradiction: we assume that $[X,X^{-1}]=0$ and $T_0(\alpha)$ is nonempty. Since $\mathcal S(M)$ is a free abelian group of basis $S(M)$, there must be cancellations in the second term of Equation \eqref{corchete final}, so there must be $g,h\in T_0(\alpha)$ with
$$\{[\alpha]g[\alpha]^{-1}g^{-1} \}=\{[\alpha]h[\alpha]^{-1}h^{-1}\} \quad \mbox{ and } \quad \epsilon_g=-\epsilon_h $$ contradicting Lemma \ref{general2}.
\begin{flushright}
$\Box$
\end{flushright}
\vskip 1cm
| 39619ee64383fffde5cc64fd64242cdb5902f696 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Time-dependent manipulation of few and many-particle quantum systems is important across all implementations of quantum computing and simulation. In such processes, decoherence and undesired transitions reducing the state fidelity are relatively ubiquitous. One important example is given by the undesired transitions that can occur between instantaneous eigenstates of the dynamical Hamiltonian upon the application of an external drive. This is why many driving protocols rely on adiabatic dynamics, where the system follows the instantaneous eigenstates and transitions are naturally suppressed. Ideal adiabatic processes are reversible making them - in principle - robust. However, to approach ideal adiabatic processes the dynamics must always be very slow, requiring compromises on the time-scales of competing heating and decoherence processes.
Speeding up adiabatic protocols to enable their completion within the system's coherence time is important for the development of any quantum technologies relying on such protocols \cite{acin2018quantum}. One approach to do this is the implementation of optimal driving protocols, which aim to end up with the system in a desired final state. For example, numerically optimised paths can be employed to avoid points where gaps in the spectrum of the system become small, or additional control fields can be tuned to increase the size of these gaps \cite{kirk2004optimal,glaser2015training,AlessandroBook2007}. In broad terms, this is the goal of protocols collectively referred to as quantum optimal control. Another option is to design techniques which speed up the adiabatic dynamics, often termed shortcuts to adiabaticity (STA). The primary aim of STA is to entirely remove or suppress diabatic transitions between instantaneous eigenstates of the dynamical Hamiltonian \cite{Torrontegui2013,GueryOdelin2019}. One particularly successful technique is counterdiabatic driving (CD), which was first utilised in physical chemistry by Demirplak and Rice \cite{demirplak2003,demirplak2005}, and was independently introduced by Berry \cite{berry_transitionless_2009} under the name `transitionless driving'. CD aims to suppress losses that arise due to fast deformations of the system far from the adiabatic limit by analytically compensating for them in the Hamiltonian. In general, to suppress diabatic losses exactly, the full analytical or numerical solutions of the Schr\"odinger equation are required. This makes the implementation of CD in complex systems - \@e.g.~for many-body dynamics - difficult and requires the need for new techniques to be introduced.
Links between optimal control and STA have existed throughout the development of both approaches \cite{stefanatos2021,Zhang2021connection}, but there are few examples of their explicit combination in a way that exploits their complementary nature. Some attempts to achieve this have included an emulation of CD through fast oscillations of the original Hamiltonian \cite{Petiziol2018fast, Petiziol2019accelerating} as well as through recent advances in reinforcement learning methods aimed at optimizing quantum protocols~\cite{bukov_reinforcement_2018}. Such methods have been shown to achieve a significant improvement in performance when implemented using concepts borrowed from CD~\cite{yao_reinforcement_2020}. In this work, we offer a significantly different new approach in combining elements from STA and quantum optimal control which we will call \textit{counterdiabatic optimised local driving} (COLD).
A key ingredient in the development of COLD is a recent approach designed for implementing CD in the setting of larger, more complex systems: local counterdiabatic driving (LCD) \cite{sels_minimizing_2017,Kolodrubetz2017geometry, gjonbalaj2021counterdiabatic}. LCD offers a method to derive \emph{approximate} CD protocols, with the aim of suppressing undesired transitions instead of fully eliminating them. This allows it to account for some physical constraints of the system, \@e.g.~locality conditions. However, the approximate nature of the LCD protocol can lead to poor performance, necessitating the introduction of additional non-local, long-range corrections \cite{sels_minimizing_2017}. If all possible corrections are added, then LCD is equivalent to the normal analytical approaches of CD, but the additional terms are generally difficult to control experimentally. COLD offers an alternative approach, with additional control fields which allow for an optimisation of the dynamical Hamiltonian for a given local form of LCD. The impact of more complex corrections can then be radically reduced, giving a corresponding improvement in the desired protocol.
The structure of this paper is as follows: first, we give a detailed description of the new method, COLD, with a focus on the elements of quantum optimal control and CD required for its implementation. In Sec.~\ref{sec:TwoSpin} we explore a 2-spin annealing protocol, that showcases the strengths of COLD. Sec.~\ref{sec:1dIsing} describes and analyses the improvements gained with COLD and its combination with other optimal control techniques in the case of state preparation in the Ising model. Then in Sec.~\ref{sec:lattice} we show the improvement that COLD can achieve on the recently realised example of LCD for state transfer on a synthetic lattice in ultracold atoms. A list of abbreviations used in this work can be found in Table.~\ref{table} for reference.
\begin{table}[h]
\begin{tabular}{p{2cm} | p{6cm}} \toprule
\multicolumn{1}{m{2cm}}{\centering Abbreviation}
& \multicolumn{1}{m{6cm}}{\centering Meaning} \\
\midrule
STA & shortcuts to adiabaticity \\ \hline
CD & counterdiabatic driving \\ \hline
LCD & local counterdiabatic driving \\ \hline
COLD & counterdiabatic optimised local driving \\ \hline
BPO & bare Powell optimisation \\ \hline
CRAB & chopped randomised basis \\ \hline
ARP & adiabatic rapid passage \\
\bottomrule
\end{tabular}
\caption{List of abbreviations used throughout the manuscript.}\label{table}
\end{table}
\section{An Introduction to Counterdiabatic Optimised Local Driving}\label{sec:intro_olcd}
\subsection{Quantum Optimal Control}\label{sec:OptCont}
In the context we consider, we employ quantum optimal control to optimise the function $f(\psi,\boldsymbol{\beta})$ in the Schr\"odinger equation
\begin{equation}
\dot{\psi} = f(\psi,\boldsymbol{\beta}),
\label{eq:Control}
\end{equation}
where $\psi$ is the quantum wave function and $\boldsymbol{\beta}$ is the set of optimisable control parameters. Optimisation of Eq.~\eqref{eq:Control} in most cases means taking the system from an initial state $\ket{\psi_0}$ to a final target state $\ket{\psi_T}$ by finding the optimal values of $\boldsymbol{\beta}$ with respect to some target metric (e.g.~the time taken to evolve the system from $\ket{\psi_0}$ to $\ket{\psi_T}$). There is a large variety of techniques available to achieve this goal \cite{glaser2015training,koch2016controlling}.
The success/target metric needs to be defined prior to the optimisation of $\boldsymbol{\beta}$. Often this is done by constructing a \emph{cost function}, which in turn defines the optimisation landscape. In general, we can optimise for any desired property of the final state of the system, with some examples being the entropy, energy, energy fluctuations or some other observable. A commonly used cost function in state preparation is related to the fidelity of the final, post-evolution state $\ket{\psi_f}$ with respect to the target state:
\begin{align} \label{eq:lossfunc}
\mathcal{C}(\ket{\psi_f}) = 1 - \left|\braket{\psi_T}{\psi_f}\right|^2.
\end{align}
In performing such a numerical optimisation, it is common to take the target state to be parameterised via a Hamiltonian split into two parts. The first is the so-called \emph{bare} Hamiltonian $H_0(t)$, which can be time-dependent and describes the dynamics of the quantum system in question. The second part is then an additional driving term that includes the control parameters $\boldsymbol{\beta}(t)$ and operators $\mathcal{O}_{\rm opt}$ which provide additional degrees of freedom in the dynamics of the system. The full Hamiltonian of the control system is then:
\begin{align} \label{eq:h_optimal_control}
H_{\beta}(t,\boldsymbol{\beta}) = H_0(t) + \boldsymbol{\beta}(t)\mathcal{O}_{\rm opt}.
\end{align}
The parameters $\boldsymbol{\beta}(t)$ can then be optimised for the optimal dynamics with respect to the metric defined by the cost function.
In this work, we use the Powell minimization approach \cite{powell1964efficient} for the numerical optimisation as implemented in Python's \textit{SciPy} \cite{Pauli2020SciPy}. When performing this optimisation without any CD terms in the Hamiltonian, we refer to this approach as bare Powell optimisation (BPO), with bare referring to the lack of CD.
Furthermore, we implement the chopped rendomised basis (CRAB) approach \cite{Caneva_chopped_2011,muller2021} and combine its methodology with that of COLD, to obtain the method of COLD-CRAB. CRAB expands the size of the parameter landscape by employing randomisation, usually in the optimised pulse driving the system. The approach was first developed for quantum many-body problems whose simulation requires the time-dependent density matrix renormalization group, despite the fact that these were thought to not be tractable in the quantum control setting \cite{brif2010,muller2021}. CRAB has benefits in that it can avoid traps in the control landscape \cite{Rach2015}, and has built-in flexibility for open-loop or closed-loop optimisation \cite{heck2018,muller2021} although these advantages come at a higher computational cost due to requiring a far larger search-space for the optimisation.
\subsection{Counterdiabatic Driving}\label{sec:LCD}
An important class of optimisation problems deals with the case where the initial and final states are ground states of a Hamiltonian $H_0(t)$ at some initial $t=t_i$ and final $t=t_f$ time. In these cases, the adiabatic theorem guarantees that for an infinitesimally slow transformation of the system $t_f-t_i\to\infty$, it should follow the instantaneous (non-degenerate) ground state of $H_0(t)$ and hence reach the target state with unit fidelity. This process is generally known as quantum annealing.
In large, complex systems with many degrees of freedom, quantum annealing tends to require prohibitively long protocol times due to vanishingly small gaps typically present in such systems. This often makes annealing protocols impractical \cite{farhi2008how}. It has been found that this problem can be formally overcome by using CD, where velocity-dependent terms are added to the Hamiltonian analytically enforcing the adiabatic wave function to be the solution of the time-dependent Schr\"odinger equation \cite{demirplak2003,demirplak2005,berry_transitionless_2009}. In this case, the dynamical state will follow the instantaneous eigenstate with no transitions regardless of the driving time. The form of the dynamical Hamiltonian enforcing this is \cite{berry_transitionless_2009}:
\begin{equation}\label{eq:counterdiabatic}
\begin{aligned}
& H_{\mathrm{CD}}(t) = H_0 (t) \\ & + i\hbar \sum_n (\ket{\partial_t n}\bra{n} - \bra{n}\ket{\partial_t n}\ket{n}\bra{n}),
\end{aligned}
\end{equation}
with $\ket{n}\equiv \ket{n(t)}$ the $n$-th eigenstate of the instantaneous Hamiltonian $H_0(t)$. The last term enforces the phases ($\bra{n}\ket{\partial_t n}$) on the instantaneous eigenstates, which are arbitrary and thus will be omitted. In general, knowledge of the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} requires knowledge of the full spectrum of $H_0(t)$ at each instantaneous moment in time.
\subsection{Counterdiabatic Optimised Local Driving} \label{sec:OLCD}
We will now introduce the main idea of COLD. The principle is to take the same approach as Sec.~\ref{sec:LCD} but with the original Hamiltonian given by $H_\beta(t,\boldsymbol{\beta})$, see Eq.~\eqref{eq:h_optimal_control}. Quantum Annealing then applies to the whole family of control Hamiltonians $ H_{\beta}(t,\boldsymbol{\beta})$ as long as the additional control terms $\boldsymbol{\beta}(t)$ vanish at the protocol boundaries: $\boldsymbol{\beta}(t_i)=\boldsymbol{\beta}(t_f)=0$. This flexibility was explored in finding the optimal adiabatic path characterized by e.g. the shortest distance between the initial and the final states, i.e. a geodesic \cite{tomka2016geodesic}. A similar geodesic approach was developed in the context of dissipative systems to minimize energy losses~\cite{Sivak2012Geodesic}. During the protocol, a dynamical Hamiltonian $H_\beta(t,\boldsymbol{\beta})$ generally induces transitions between the quantum states that it drives and the question about what is the optimal path remains open.
The Hamiltonian $H_\beta(t,\boldsymbol{\beta})$ and its eigenstates depend on time only through the driving parameters, which include $\boldsymbol{\beta}$ and any additional control terms in the particular protocol. This makes it convenient to introduce a path in the coupling space parametrized by a dimensionless parameter $\lambda\in [0,1]$ such that both $H_0$ and $\boldsymbol{\beta}$ are functions of $\lambda$ satisfying $H_\beta(\lambda=0)=H_{0}(t_i)$ and $H_\beta(\lambda=1)=H_{0}(t_f)$, i.e. being equal to the initial and the final Hamiltonian at the protocol boundaries. By construction this implies that any additional fields introduced to the bare Hamiltonian $H_0$ must go to zero at the boundaries. In this way, any protocol can be uniquely characterized by first specifying the path $\boldsymbol{\beta}(\lambda)$ in the coupling space manifold and then the time dependence $\lambda(t)$ along it. The path determines the sequence of couplings of the Hamiltonian during time evolution and hence the sequence of ground state wave functions followed by the driven state. Furthermore, the time dependence encodes the speed of traversing this path. We can then introduce a hermitian operator called the (path-dependent) adiabatic gauge potential~\cite{sels_minimizing_2017}: $\mathcal A_\lambda=i \hbar \sum_n \ket{\partial_\lambda n}\bra{n}$, which satisfies a closed form equation,
\begin{equation}
\label{eq:AGP_eq}
\left[\partial_\lambda H_\beta+{i\over \hbar} [ \mathcal A_\lambda,H_\beta],H_\beta\right]=0,
\end{equation}
where $H_\beta \equiv H_\beta(\lambda,\boldsymbol{\beta}(\lambda))$ and $\mathcal{A}_\lambda \equiv \mathcal{A}_\lambda(\lambda,\boldsymbol{\beta}(\lambda))$. Then the CD Hamiltonian reads
\begin{align}\label{eq:counterdiabaticLCD}
H_{\mathrm{CD}}(\lambda,\boldsymbol{\beta}(\lambda)) = H_\beta(\lambda,\boldsymbol{\beta}(\lambda)) +\dot \lambda \mathcal A_\lambda(\lambda,\boldsymbol{\beta}(\lambda)),
\end{align}
and is equivalent to the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} given knowledge of the exact adiabatic gauge potential. However, generally the adiabatic gauge potential is a very non-local object and solutions of Eq.~\eqref{eq:AGP_eq} are unstable to small perturbations containing exponentially many terms in the number of degrees of freedom.
LCD aims to find approximate gauge potentials that satisfy particular requirements like robustness and locality, thus circumventing many of the difficulties in determining the second component in Eq.~\eqref{eq:counterdiabatic} and~\eqref{eq:counterdiabaticLCD} exactly. The goal, in essence, is to suppress the most relevant diabatic effects rather than completely eliminate them. This method has recently been experimentally implemented to speed up state transfer for synthetic lattices in ultracold atoms \cite{meier_counterdiabatic_2020}, for preparing states in nuclear-magnetic-resonance systems \cite{zhou_experimental_2020}, and annealing protocols on an IBM quantum computer \cite{Hegade2021Shortcuts}.
Following the methods of Ref.~\cite{sels_minimizing_2017}, the problem of finding the optimal adiabatic gauge potential can be cast as the minimisation of the Hilbert-Schmidt norm of the operators
\begin{equation}\label{eq:Goperator}
G_{\lambda}= \partial_{\lambda}H_\beta + i\comm{\mathcal{A}_\lambda}{H_\beta},
\end{equation}
which is equivalent to minimisation of the action
\begin{equation}\label{eq:actionCD}
\mathcal{S}(\mathcal{A}_{\lambda}) = \Trace{\left[G_{\lambda}(\mathcal{A}_{\lambda})^2\right]},
\end{equation}
with respect to $\mathcal{A}_{\lambda}$. In most cases, this is achieved by first choosing an operator ansatz - \@i.e. a set of linearly independent operators $\{\mathcal{O}_{\rm LCD}\}$ - and then using this set as an operator basis for the adiabatic gauge potential $\mathcal A_\lambda=\sum_j \alpha_j \mathcal O_{\rm LCD}^{(j)}$. The action can then be minimized with respect to the the set of coefficients, ${\bm \alpha}$. In the example of an
Ising spin chain we may take $\mathcal{A}_\lambda = \sum_j^N \alpha_j \sigma^y_j$, where $j$ labels the $N$ chain sites, and $\{\mathcal{O}_{\rm LCD}\}$ is a set the $y$-pauli matrices.
Without any additional control fields $\boldsymbol{\beta}$, LCD is essentially an informed choice of the operator set $\{\mathcal{O}_{\rm LCD}\}$ in a way that the resulting control protocol from the minimisation of Eq.~\eqref{eq:actionCD} is optimal for a given $H_0(\lambda)$. In this case we explore the family of Hamiltonians
\begin{equation}
H_{\rm LCD}(\lambda) = H_0(\lambda) +\sum_j \alpha_j(\lambda) \mathcal{O}_{\rm LCD}^{(j)}.
\end{equation}
The performance of such LCD protocols is determined by how accurately the variational manifold spanned by the set $\{\mathcal{O}_{\rm LCD}\}$ can approximate an exact $\mathcal{A}_{\lambda}$ such that Eq.~\eqref{eq:AGP_eq} holds.
In the case of the new protocol COLD, we allow for extra exploration of the family of Hamiltonians due to the additional control fields as in Eq.~\eqref{eq:h_optimal_control}. This expands the family of Hamiltonians to
\begin{equation}\label{eq:expandedH}
H_\beta(\lambda,\boldsymbol{\beta}) = H_0(\lambda) + {\bm \alpha}(\lambda,\boldsymbol{\beta}) \mathcal{O}_{\rm LCD} + \boldsymbol{\beta}(\lambda) \mathcal{O}_{\rm opt}.
\end{equation}
Note, that the coefficients of the optimal control field change the form of the LCD driving coefficients, i.e. $\bm \alpha = f(\lambda,\boldsymbol{\beta})$. The aim of COLD is then to choose $\boldsymbol{\beta}(\lambda)$ and $\mathcal{O}_{\rm opt}$ such that the LCD term is optimal for the dynamical Hamiltonian $H_0(\lambda)+\boldsymbol{\beta}(\lambda)\mathcal{O}_{\rm opt}$. We will focus on the optimisation of the control parameters $\boldsymbol{\beta}(\lambda)$ for a given choice of $\mathcal{O}_{\rm opt}$, although the choice of operators $\mathcal{O}_{\rm opt}$ can also be optimised over as an extension.
With COLD, we have two methods to improve on the existing LCD protocol. As previously shown in Refs.~\cite{claeys_floquet-engineering_2019,prielinger_diabatic_2020}, there is a possibility to add more terms to the LCD making it less local, \@e.g.~through long-range interactions. In the spin chain case, we could take the aforementioned sum over $\sigma^y$ terms to be the \textit{first-order} anzatz for the LCD, where higher-order ans\"{a}tze might contain sets of operators $\{\mathcal{O}_{\rm LCD}\}$ with terms odd in $\sigma^y$ such as $\sigma^y_j\sigma^{(z,x)}_{j+1}$. This procedure generally improves the performance of CD protocols at the expense of adding more complex operators which may be experimentally impractical depending on the scenario. Alternately, with COLD and the introduction of additional local control fields to the Hamiltonian, we can improve the performance of LCD at a fixed complexity of the CD term. If these extra control fields vanish at the beginning and at the end of the protocol, they do not affect the adiabatically connected states, but they can significantly modify the adiabatic landscape at intermediate couplings to enhance the performance of the given order of LCD.
In order to optimize for the additional fields, in this work, we use methods of quantum optimal control. However, we note that such optimization can be done locally by requiring that the next order variational correction to the CD terms is small along the optimal landscape. This local optimization may be advantageous in that it does not require knowledge of the wave function and in this sense is not limited by small system sizes.
Furthermore, we compare COLD to the use of CRAB, as discussed in Sec.~\ref{sec:OptCont}. An advantage of COLD is that it can be combined with many advanced optimal control procedures, owing to the standard way additional control fields are introduced to the Hamiltonian. In this work we find the combination of COLD and CRAB particularly useful and we will refer to this as COLD-CRAB.
\section{Two Spin Quantum Annealing}\label{sec:TwoSpin}
To showcase and explore the use of COLD in a relatively simple setting we will consider a two spin quantum annealing problem with bare Hamiltonian
\begin{equation}\label{eq:Hanneal}
H_0(t) = -2J \sigma_1^z \sigma_2^z - h ( \sigma_{1}^z + \sigma_{2}^z) + 2h \lambda(t) (\sigma_{1}^x + \sigma_{2}^x),
\end{equation}
where $\sigma^a_j$, $a \in \{x,y,z\}$ are the Pauli matrices applied to spins indexed by $j$. For the scaling function $\lambda(t)$ we pick the form
\begin{equation}\label{eq:Scalingfunc}
\lambda(t) = \sin^2\left(\frac{\pi}{2} \sin^2 \left( \frac{\pi t}{2 \tau} \right) \right)
\end{equation}
such that $\lambda(0) = 0$ and $\lambda(\tau) = 1$. We consider the case of $J/h=0.5$, which means the spins start in the initial state of $\ket{\uparrow \uparrow}$ and finish in a superposition of all of the symmetric states.
As discussed in Ref.~\citep{sels_minimizing_2017}, since $H_0$ has a standard Ising spin chain form, the first-order LCD terms are given by the following ansatz for the adiabatic gauge potential:
\begin{equation}\label{eq:LCD1st}
\mathcal{A}(\lambda) = \alpha \sum_{i=1}^2 \sigma_i^y,
\end{equation}
with the sum being over the full length of the $N$ spin chain. Minimising Eq.~\eqref{eq:actionCD} for this $\mathcal{A}_{\lambda}$ with respect to the coefficient $\alpha$ gives
\begin{equation}
\alpha = - \frac{h^2}{4(h\lambda)^2 + h^2 + 4J^2}.
\end{equation}
To further improve on the first-order LCD we can implement COLD, as we will discuss shortly, or we can introduce higher-order terms to the ansatz for $\mathcal{A}_{\lambda}$. This second method serves as a good benchmark against COLD, since it offers an improvement to first-order LCD in the same way as COLD does, but requires more complicated interactions between the two spins increasing the implementation overhead. The second-order LCD can be found by taking an ansatz for the adiabatic gauge potential:
\begin{equation}
\begin{aligned}
\mathcal{A}^{(2)}(\lambda) =& \alpha \sum_{j} \sigma_j^y + \gamma (\sigma_1^x \sigma_{2}^y + \sigma_1^y \sigma_{2}^x) \\ & + \zeta (\sigma_1^z \sigma_{2}^y + \sigma_1^y \sigma_{2}^z),
\end{aligned}
\end{equation}
where to solve for $\alpha$, $\gamma$, and $\zeta$ we once again minimize the action given by Eq.~\eqref{eq:actionCD} and obtain three coupled equations which can be solved numerically (see Appendix \ref{app:derivation} for a detailed derivation).
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{TwoSpin.png}
\caption{Optimisation of the annealing protocol for two spin Hamiltonian given by Eq.~\eqref{eq:Hanneal} for $h/J=2$. (a) Final fidelities of the annealing protocol with triangles (black) representing the case where no CD is applied and circles showing the case of first-order LCD (pink) as well as the combination of first- and second-order LCD (orange). (b) Final fidelities achieved when using the optimal control method BPO (red diamonds) and the new approach of COLD (blue circles), both with $N_k=1$.}\label{fig:TwoSpin}
\end{figure}
We now consider three distinct cases in this two spin quantum annealing example: no LCD, first-order LCD, and second-order LCD. The fidelity of the final state for each case over a wide range of driving times is shown in Fig.~\ref{fig:TwoSpin}(a), with an easily distinguishable advantage in the case of LCD. The final fidelity where no LCD is implemented decreases rapidly as the ramp times are made short, with the system getting stuck in its initial state. On the contrary, first-order LCD retains good final state fidelities into short times, as the driving Hamiltonian becomes that of only the LCD term. The second-order LCD then gives unit fidelity, in agreement with previous observations \cite{claeys_floquet-engineering_2019}, as for a two spin Hamiltonian the highest order corrections are that including two spin terms.
We now add an optimisable term, as described in Sec.~\ref{sec:OptCont}, so that the new Hamiltonian reads:
\begin{equation}\label{eq:HannealOpt}
H_\beta(t) = H_0(t) + \sum_{k=1}^{N_k} \beta^k \sin (\pi k t / \tau) \sum_i \sigma_i^z,
\end{equation}
\noindent with $N_k$ the number of optimisation coefficients, and $\beta^k$ the coefficient of the $k$th frequency of the control function. Note that we consider
\begin{equation}
\boldsymbol{\beta}(t) = \sum_{k=1}^{N_k} \beta^k \sin (\pi k t / \tau) = \sum_{k=1}^{N_k} \beta^k \sin (\pi k f(\lambda)),
\end{equation}
with
\begin{equation}\label{eq:lambda}
f(\lambda) = \frac{2}{\pi}\arcsin\left(\sqrt{\frac{2}{\pi} \arcsin\left(\sqrt{\lambda}\right)}\right).
\end{equation}
The form of the additional control field fulfils the requirement that the boundary conditions are $H(0) = H_0(0)$ and $H(\tau) = H_0(\tau)$. Numerically optimising the $\beta^k$ for the best final state fidelity \emph{without} adding LCD terms results in the BPO method introduced in Sec.~\ref{sec:OptCont}. We show the results of BPO in Fig.~\ref{fig:TwoSpin}(b), where it is observed that BPO gives better results than the case of no LCD in Fig.~\ref{fig:TwoSpin}(a). However, for short times the BPO approach still results in the system getting stuck in the initial state.
\begin{figure*}[t]
\includegraphics[width=0.9\linewidth]{IsingUnconstrained.png}
\caption{Optimisation of the annealing protocol for the Ising model given by Eq.~\eqref{eq:h0_ising} for $N=5$ spins. (a) A comparison of final state fidelities for different driving times using the optimal control technique BPO (blue diamonds), first-order LCD (black dash-dot line) and COLD (red circles). The same is shown in (b) for CRAB (green diamonds) and COLD-CRAB (purple circles). CD enhanced techniques (COLD and COLD-CRAB) introduced in this work show a clear convergence to good fidelities at short driving times. All results are for the best (lowest) fidelity achieved over $500$ optimisations.}\label{fig:IsingUnconstrained}
\end{figure*}
Finally we present and compare the results of the new method, COLD. In this case the Hamiltonian before adding LCD terms is given by Eq.~\eqref{eq:HannealOpt} and the coefficient of the first-order LCD is
\begin{equation}
\alpha = -\frac{h(h+\boldsymbol{\beta}) + h\frac{\lambda}{\dot{\lambda}} \dot{\boldsymbol{\beta}}}{4(h\lambda)^2 + (h+\boldsymbol{\beta})^2 + 4J^2}.
\end{equation}
Note that the optimisation of the additional control field also feeds into the coefficient of the adiabatic gauge potential during the dynamics as discussed in Sec.~\ref{sec:OLCD}. The results of the COLD approach for this two spin annealing protocol are shown in Fig.~\ref{fig:TwoSpin}(b), where we observe an improvement of the final state fidelity beyond what is possible with first-order LCD alone in Fig.~\ref{fig:TwoSpin}(a). In this example, LCD alone reaches a final state fidelity of $1-F=3\%$ at short times, however COLD improves this error in the final state to $1-F=0.005\%$. This is due to the extended family of dynamical Hamiltonians in Eq.~\eqref{eq:expandedH} owing to the addition of an optimisable control field. This result shows that COLD can provide an advantageous alternative to the addition of higher-order LCD which may be experimentally impractical.
We have found that COLD performs better than LCD of the same order or BPO when the system dynamics are calculated numerically. This does not, however, imply anything about the performance of COLD in more complex scenarios, like in the case of an unknown target ground state. In that case the fidelity is a poor optimisation metric. There is, however, a way to come to the same conclusions as those presented in Fig.~\ref{fig:TwoSpin} without the need to compute the dynamics exactly. We can do this by first using a guess for the COLD protocol to find the approximate adiabatic gauge potential and then minimising the integral of the norm of the second-order correction to the adiabatic gauge potential along the path. Note, that the ground state can be in turn obtained through first order COLD, so there is no need to diagonalize the Hamiltonian. This integral should be small if COLD has implemented a dynamical Hamiltonian that makes the first-order adiabatic gauge potential the leading term. It is effectively a measure of the error of COLD and can be given by
\begin{equation}
\begin{aligned}
\mathcal{I}_1 = \int_0^\tau dt^\prime \Big[& \bra{\psi_g(t^\prime)} \Gamma^2(t^\prime) \ket{\psi_g(t^\prime)} \\ &- (\bra{\psi_g(t^\prime)} \Gamma(t^\prime) \ket{\psi_g(t^\prime)})^2\Big]^{1/2},
\end{aligned}
\end{equation}
\noindent with $\ket{\psi_g(t)}$ the instantaneous ground state along the path and
\begin{equation}
\Gamma(t) = \gamma(t) \left( \sigma_1^y \sigma_2^x + \sigma_1^x \sigma_2^y \right),
\end{equation}
\noindent one of the second-order correction terms. In order to confirm this is the case, we compare the different paths -- COLD and LCD only -- in the two-spin example in order to determine if $\mathcal{I}_1$ is small for COLD. If $\mathcal{I}_1$ is small when compared to the same measure for lower-order LCD as $t\rightarrow 0$, then we know that COLD is enforcing a better dynamical Hamiltonian. In the case of the two spin annealing protocol we find that as $t\rightarrow0$, $\mathcal{I}_1\rightarrow 0.04$ for COLD and $\mathcal{I}_1\rightarrow 0.2$ for LCD, showing that COLD is minimising the second-order correction along the path. A simpler integral
\begin{equation}
\mathcal{I}_2 = \int_0^\tau dt^\prime |\gamma(t^\prime)|,
\end{equation}
also reflects this correction in this two spin example, with $\mathcal{I}_2 \rightarrow 0.03$ for COLD and $\mathcal{I}_2 \rightarrow 0.1$ for LCD as $t\rightarrow0$. This could be useful in more complex scenarios as $\mathcal{I}_2$ will be relatively simple to calculate. We also observe the reduction of the corresponding integrals of the $(\sigma^y_1\sigma^z_2 + \sigma^z_1\sigma^y_2)$ term of the second-order LCD. By minimising these integrals, it could be possible to extend the COLD approach to more complex scenarios, including where the exact calculation of the dynamics is not possible.
\section{1D Ising Model} \label{sec:1dIsing}
In this section we apply COLD for state preparation on a 1D Ising spin chain in the presence of a transverse and longitudinal field. We consider an annealing protocol where the aim is to prepare the ground state across the Ising phase transition. The annealing Hamiltonian is given by
\begin{align}\label{eq:h0_ising}
\begin{split}
H_{0}(t) &= - J \sum_{j}^{N-1} \sigma^z_j\sigma_{j+1}^z + Z_0\sum_j^N \sigma_j^z \\ &+ \lambda(t) X_f \sum_j^N \sigma_j^x,
\end{split}
\end{align}
where $Z_0$ is a small offset parameter to break ground state degeneracies and $X_f$ is the final x-field strength. Note, the breaking of the ground state degeneracies is not a requirement but allows for easier consideration of the adiabatic path. As before, $\lambda(t)$ is a scaling function that has the boundary conditions $\lambda(0) = 0$ and $\lambda(\tau) = 1$, with $\tau$ the driving time. This means we start from the ground state of all spins up and drive across the quantum phase transition to the ground state which is a superposition of all basis states. We again take the scaling function to be given by Eq.~\eqref{eq:Scalingfunc}. In this example, we use $X_f = 10J$ and $Z_0 = 0.02J$.
For the Hamiltonian of Eq.~\eqref{eq:h0_ising}, the LCD to first and second order is well known, as the wave functions are entirely real. We take the first-order adiabatic gauge potential to be given by
\begin{equation}\label{eq:LCD1}
\mathcal{A}(\lambda) = \alpha \sum_{j}^N\sigma_j^y,
\end{equation}
\noindent where the coefficients for the general periodic spin chain of Eq.~\eqref{eq:h0_ising} are
\begin{align} \label{eq:alphas}
\alpha(\lambda) = \frac{1}{2} \frac{Z_0 X_f}{Z_0^2 + \lambda^2 X_f^2 + 2J^2}.
\end{align}
Note, that the quoted $\alpha$ above is technically for a periodic or infinite size system, with $J^2 \rightarrow J^2(1-1/N)$ for a finite system. However, we find that the inclusion of the factor for the finite system sizes we consider only changes the final achieved converged fidelities at short times by $\sim 10^{-6}\%$. The second-order adiabatic gauge potential is of the form
\begin{equation}
\begin{aligned}
\mathcal{A}^{(2)}(\lambda) =& \alpha \sum_{j} \sigma_j^y + \gamma \sum_{j} (\sigma_j^x \sigma_{j+1}^y + \sigma_j^y \sigma_{j+1}^x) \\ & + \zeta \sum_{j} (\sigma_j^z \sigma_{j+1}^y + \sigma_j^y \sigma_{j+1}^z) ,
\end{aligned} \label{eq:SecondLCD}
\end{equation}
with the coefficients $\alpha$, $\gamma$ and $\zeta$ again obtained by minimising the action given by Eq.~\eqref{eq:actionCD} and solving the coupled set of equations numerically (see Appendix \ref{app:derivation} for a detailed derivation).
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{MaxAmp.png}
\caption{Maximum amplitudes of CD terms in the Ising model annealing protocol for (a) first- and second-order LCD only with no additional optimal control fields and (b) the COLD approach optimised for the best final state fidelity implementing first-order LCD as shown in Fig.~\ref{fig:IsingUnconstrained} (a). The plot shows the maximum amplitude at each driving time for the first-order $\alpha$ (red circles) and the two second-order terms $\gamma$ (blue diamonds) and $\zeta$ (green triangles) as given in Eq.~\eqref{eq:SecondLCD} (although the second-order LCD is not actually implemented in COLD). An inversion in the strength of the second-order and first-order LCD terms for (a) no additional optimal control fields and (b) the addition of optimal control fields shows that COLD implements a dynamical Hamiltonain which is favourable for the applied order of LCD (first-order in this case).}\label{fig:MaxAmp}
\end{figure}
In this example, optimal control is implemented by introducing an additional driving field so that the dynamical Hamiltonian is given by
\begin{align}\label{eq:h_beta}
H_{\beta}(t, \boldsymbol{\beta}) = H_0(t) + \sum_j \boldsymbol{\beta}(t)\sigma^z_j,
\end{align}
with $\boldsymbol{\beta}$ being the terms to optimise over. We take our additional terms to again respect the boundary conditions $\boldsymbol{\beta}(0) = 0$ and $\boldsymbol{\beta}(\tau) = 0$, meaning a natural choice is
\begin{align}\label{eq:optimsable_1}
\boldsymbol{\beta}(t) = \sum_k^{N_k} \beta^k \sin(\omega_k t / \tau) = \sum_k^{N_k} \beta^k \sin(\omega_k f(\lambda)),
\end{align}
with $\omega_k = k 2\pi$ the $k$th principal frequency and $f(\lambda)$ given by Eq.~\eqref{eq:lambda}. To implement the CRAB algorithm discussed in Sec.~\ref{sec:OptCont}, we will use $k \rightarrow k(1+r)$ instead with $r$ drawn from a uniform random distribution $r \in [-0.5,0.5]$.
As before, we choose the first order adiabatic gauge potential given by Eq.~\eqref{eq:LCD1} and find that the coefficients are
\begin{align}
\alpha(\lambda,\boldsymbol{\beta}) = \frac{X_f}{2} \frac{(Z_0 + \boldsymbol{\beta}) - \lambda\dot{\boldsymbol{\beta}}/\dot{\lambda}}{(Z_0 + \boldsymbol{\beta})^2 + \lambda^2 X_f^2 + 2 J^2}.
\end{align}
Note, with the introduction of the additional control fields $\boldsymbol{\beta}$ it is possible for $\alpha$ to be non-zero at the start or end of the protocol, as $\dot{\boldsymbol{\beta}}$ is not fixed to be zero. However, this can be enforced by a suitable choice of the additional control field, we will consider replacing $\alpha \rightarrow S(\lambda)\alpha$ where $S(\lambda)$ is a scaling function that tends to zero as $\lambda \rightarrow 0$ and $\lambda \rightarrow 1$. We find that the scaling function only has a minimal effect on the final fidelities observed. This issue could also be resolved by a suitable choice of $\boldsymbol{\beta}$, with our example drive being an extreme case as $\dot{\boldsymbol{\beta}}$ is maximal at the boundaries of the protocol. The suitable choice of the form of $\boldsymbol{\beta}$ in a given example is a problem we will leave for future work, with our focus being on the introduction of the COLD protocol.
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{ScalingN.png}
\caption{Scaling of fidelities in the annealing protocol for the Ising model with (a) system size $N$ and (b) optimisation parameters $N_k$ at driving time $\tau=10^{-2}J^{-1}$. Plots show a comparison between BPO (blue diamonds) and COLD (red circles). In (a) we see that the COLD fidelity decreases as a function of $N$ but remains quite high when compared to BPO while (b) shows the non-existent improvement for both BPO and COLD with an increasing number of parameters in the $N=5$ spin case. Once again, plotted best fidelities are obtained across 500 optimisations.}\label{fig:ScalingN}
\end{figure}
We first compare the final state fidelity when using COLD versus BPO as shown in Fig.~\ref{fig:IsingUnconstrained}(a) for different driving times in a system of $N=5$ spins and a single $N_k=1$ optimisation coefficient. As expected, at long timescales the two methods agree as we approach the adiabatic limit of the dynamics. However, at shorter time scales the difference in behaviour is dramatic. We observe that the BPO approach fails in the case of very fast driving as the state gets stuck in the initial state but the COLD approach converges to $1-F \sim 10^{-3}$. Note, this is not achieved by the introduction of first-order LCD terms alone, as this will result in $F=0.0440$ for $\tau=10^{-3}J^{-1}$. COLD is instead achieving this by making the LCD term dominant for the dynamical Hamiltonian through the additional control fields.
\begin{figure*}[t]
\includegraphics[width=0.9\textwidth]{IsingcConstrained.png}
\caption{Optimisation of the \emph{constrained} annealing protocol for the Ising model for $N=5$ spins with a maximum amplitude limit on each term in the Hamiltonian of Eq.~\eqref{eq:h0_ising} of $10J$. (a) Shows a comparison between BPO (blue diamonds) and COLD (red circles) which both give much lower fidelities than in the unconstrained case in Fig.~\ref{fig:IsingUnconstrained}, although COLD persists in giving better results. In (b) the comparison is between CRAB (green diamonds) and COLD-CRAB (purple circles) which show orders of magnitude better fidelities than those in (a), with COLD-CRAB eking out higher fidelities at short driving times. The plotted best results are obtained from 200 optimisations for each method.}\label{fig:IsingcConstrained}
\end{figure*}
To confirm this, we plot the maximum amplitudes of both the first- and second-order adiabatic gauge potentials in Fig.~\ref{fig:MaxAmp}, where Fig.~\ref{fig:MaxAmp}(a) shows the case of no optimisation and Fig.~\ref{fig:MaxAmp}(b) the case of applying COLD. We can see that without COLD the second-order $(\sigma^x_j\sigma^y_{j+1} + \sigma^y_j\sigma^x_{j+1})$ corrections to the LCD are far larger than the first-order, resulting in the small final state fidelities when only first-order LCD is implemented. In the case of COLD, this relationship reverses and the first-order LCD terms dominate the dynamics. These results are further evidence that the implementation of COLD through the minimisation of the second-order corrections discussed in Sec.~\ref{sec:TwoSpin} could be fruitful in more complex and/or larger systems, where the dynamics can not be calculated.
We find that the results of BPO and COLD at short driving times are stable against increasing system size, as shown for $\tau = 10^{-2}J^{-1}$ in Fig.~\ref{fig:ScalingN}(a), with only a small decrease in final state fidelity for larger systems with COLD. Similarly, increasing the number of optimisation coefficients $N_k$ results in little improvement in the values obtained at short times for this example, as shown in Fig.~\ref{fig:ScalingN}(b). It is possible that in more complex systems, more optimisation coefficients will be needed to gain a larger advantage. We also note that by increasing the number of coefficients, we are increasing the complexity of the cost function landscape to be explored by the minimisation procedure, hence leading to slightly worse final fidelities This can mean that alternative approaches than the Powell minimisation used so far, \@e.g.~that of CRAB, could be better suited to probing the cost function for high $N_k$. We also note that this lack of improvement in the results is likely the consequence of the form of the control field given by Eq.~\eqref{eq:optimsable_1} rather than due to a failure of the optimiser in the face of a complex parameter space. We find that the parameter space is relatively smooth in the case of $N_k = 1,2,3$ and a better solution for this form of control field simply does not exist.
We now consider the combined method of COLD-CRAB for this annealing example as shown in Fig.~\ref{fig:IsingUnconstrained}(b). We point out that with our application of CRAB in this scenario we are not enforcing $\boldsymbol{\beta}$ to be zero at the start and end of the dynamics, allowing for their to be a tuning of the $z$-field offset. This is consistent between CRAB and COLD-CRAB and therefore does not influence our comparison of the two. First, it is important to note that CRAB alone results in a overall speedup of the dynamics for a high final state fidelity $1-F\sim 10^{-3}$ at long time-scales. However, CRAB still suffers from getting stuck in the initial state at fast driving times and the final state fidelity again tends to zero. This is not the case for COLD-CRAB, which converges to large final state fidelities $1-F \sim 10^{-3}$ at short driving times $\tau \leq 10^{-1}J^{-1}$. Note that the difference between the convergence to final state fidelities are only marginally different between COLD and COLD-CRAB at longer times, but at short times COLD-CRAB performs a lot better. Further improvement could be gained by combining COLD with more advanced versions of CRAB or other optimal control methods.
As shown in Fig.~\ref{fig:MaxAmp}, the amplitude of the driving required to achieve the fidelities discussed so far scales with the driving time. Practical scenarios will necessarily place limits both on achievable driving times and the maximum amplitude of any term that is being driven. However, the scaling of the drivings shown do not mean that everything diverges in the limit of $\tau \rightarrow 0$. To see this we can first write the Scr\"odinger equation for COLD as
\begin{equation}
i \hbar d_t \ket{\psi}=\left(H_\beta+\dot\lambda \mathcal{A}_\lambda\right) \ket{\psi},
\end{equation}
we then divide through by $\dot{\lambda}$ to get
\begin{equation}
i \hbar d_\lambda \ket{\psi}=\left(\frac{H_\beta}{\dot \lambda}+\mathcal{A}_\lambda\right) \ket{\psi},
\end{equation}
in the limit of $\tau \rightarrow 0$ then $\dot\lambda\to \infty$ to result in the Hamiltonian term disappearing, or in other words, we turn off the Hamiltonian. We then only drive the system in the $\tau \to 0$ limit with the COLD or LCD driving term:
\begin{equation} \label{eq:OnlyA}
i \hbar d_\lambda \ket{\psi}=\mathcal{A}_\lambda \ket{\psi}.
\end{equation}
In this limit then $\lambda$ plays the role of time, and this could then be implemented in a practical scenario in finite time as it corresponds to some manipulation of the couplings in the system. This renormalised time cannot then be infinitesimally short if the couplings are bounded but we have shown that the protocol does not diverge as $\tau \rightarrow 0$. In the case of a spin chain, evolving under Eq.~\eqref{eq:OnlyA} is effectively to first order in LCD implementing independent single spin rotations along the chain, and COLD can be easily applied.
If it is not possible to switch off the Hamiltonian as discussed above then as an alternative we can implement COLD with experimental constraints accounted for directly in the optimal control minimisation. We consider an extreme example of constraints to show that even in this scenario COLD can provide an advantage and corresponding speed-up. In the constrained case the annealing protocol remains that of Hamiltonian~\eqref{eq:h0_ising} but we choose to introduce a bound of $X_f$ on the maximum amplitude of all drivings. This makes it so that no optimal control or LCD term can go beyond the original amplitude of the $x$-field drive. We show the final state fidelities achieved for the constrained example in Fig.~\ref{fig:IsingcConstrained}. As can be seen in Fig.~\ref{fig:IsingcConstrained}(a), COLD provides a substantial improvement beyond what is achievable with BPO. BPO manages $F < 0.5$ for $\tau < 1J^{-1}$, but COLD can reach final state fidelities $F\sim 0.9$ for $\tau < 1J^{-1}$. The real improvement, however, comes with the application of CRAB and COLD-CRAB. CRAB already improves the fidelities substantially, and would allow for a speed up in the annealing protocol but with COLD-CRAB the final state fidelities are even better, with $F\sim 0.99$ achievable when approaching $\tau\sim 0.1J^{-1}$. Signs are seen of the onset of the convergence to small values for COLD-CRAB in Fig.~\ref{fig:IsingcConstrained}(b) before the maximum amplitude required becomes too large and the short time results tend towards zero fidelity and the state being stuck again. With this example and the discussion on implementation via turning off the Hamiltonian, we have shown that COLD is capable of delivering improvements beyond other schemes even for practical problems with strict and rather extreme constraints.
\section{Transport in a Synthetic Lattice}\label{sec:lattice}
\begin{figure*}[t]
\includegraphics[width=0.98\textwidth]{synthetic.png}
\caption{Optimisation of state transfer in a synthetic lattice. In (a) we compare the fidelities obtained via the bare ARP protocol (pink dashed line) and first-order LCD previously implemented in Ref.~\cite{meier_counterdiabatic_2020} (purple dash-dot line) to BPO (blue diamonds) and the COLD method (red circles). (c) Maximum amplitude of the tunneling term at each driving time for LCD (green diamonds) as given by Eq.~\eqref{eq:tunneling} as well as COLD (red triangles) which includes additional control parameters as shown in Eq.~\eqref{eq:tunneling_opt} and BPO (blue triangles) which omits the modifications due to CD but retains the control terms $\boldsymbol{\beta}$. In both (a) and (c) we simulate $N = 7$ lattice sites and use $N_k = 1$ parameter for optimisation of BPO and COLD. (b) Scaling of fidelities with increasing number of lattice sites (where $N_k = 1$) for both COLD (red circles) and BPO (blue diamonds) noting that the latter performs very poorly for $N > 9$. (d) does the same for the number of parameters while keeping $N=7$, with the trend indicating that increasing $N_k$ does not lead to better fidelities in either the BPO or COLD case. Note that both (b) and (d) are simulated for driving time $\tau = 0.5 J^{-1}$ and the best fidelities are obtained across 500 optimisations.}\label{fig:Synthetic}
\end{figure*}
The efficient transfer of states between opposite ends of a lattice could have future applications in the settings of quantum computation and simulation due to its promise of efficient transport of information \cite{lang2017}. This objective is often tackled in the setting of ultracold atoms in optical lattices. While the problem can be tuned to be a single-particle system and the analytical solutions of the corresponding instantaneous Schr\"odinger equation are known \cite{Hatsugai1993,Hugel2014} even for a finite system \cite{Duncan2018exact}, efficient evolution for state transfer is not straight-forward. This is due to the fact that the majority of the states are delocalised across the lattice, meaning that the $\ket{\psi}\bra{\psi}$ terms of the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} are global in reach. It is normal to consider this system in the tight-binding limit where the implementation of global terms is not straightforward. Such terms can be generated via the interactions of the atoms with cavity modes \cite{landig2016,keller2017} or from dipolar interactions \cite{baranov2002,menotti2008,trefzger2011}. However, it would be ambitious to expect this control to be general enough to implement the CD Hamiltonian of the exact solutions. This is one of the reasons that LCD has been pursued in this setting.
Recently, LCD has been successfully applied to improve an adiabatic rapid passage (ARP) protocol for population transfer across a synthetic lattice \cite{meier_counterdiabatic_2020}. In this realisation, population transfer was achieved in a synthetic tight-binding lattice of laser coupled atomic momentum states. We will consider the same problem as in Ref.~\cite{meier_counterdiabatic_2020} but with the improvement that can be gained by COLD. This system is described by the Hamiltonian
\begin{align}\label{eq:Hlattice}
\begin{split}
H_0(t) &= - \sum_n J_n(t)(c_n^{\dag}c_{n+1} + H.c.) \\ &+ \sum_n V_n(t) c_n^{\dag}c_n,
\end{split}
\end{align}
where $J_n(t)$ is the time-dependent tunnelling that describes the nearest-neighbour coupling, $V_n(t)$ is the on-site energy offset with respect to neighbouring sites and $c_n^{\dag}$($c_{n}$) is the creation(annihilation) operator on a given synthetic lattice site. In the ARP protocol, the population gets moved from one end of the lattice to the other by linearly ramping the lattice from a positive tilt to a negative tilt via
\begin{align}\label{eq:t_and_v}
J_n(t) = J_0(1.1 - \lambda) = J_0\Big(0.1 + \frac{t}{\tau}\Big), \\
V_n(t) = n V_0 2 (\lambda - 1/2) = nV_0\Big(1 - \frac{2t}{\tau}\Big),
\end{align}
where $V_0 = 4J_0$ is the initial site energy slope and $J_0$ is the characteristic tunnelling scale of the lattice. The scaling function in this case is given by
\begin{equation}
\lambda(t) = 1-\frac{t}{\tau}.
\end{equation}
In order to implement LCD as shown in Ref.~\cite{meier_counterdiabatic_2020}, the first order LCD can be accounted for by taking
\begin{align}\label{eq:tunneling}
J_n(t) \rightarrow J_{n, \mathrm{CD}}(t) e^{-i\phi_{n, \mathrm{CD}}(t)},
\end{align}
\noindent where
\begin{align}\label{eq:t_phi_cd}
J_{n, \mathrm{CD}}(t) = \sqrt{J_n(t)^2 + (\alpha_n(t)/\tau)^2}, \\
\phi_{n, \mathrm{CD}}(t) = \arctan\left(-\frac{J_n(t)\tau}{\alpha_n(t)}\right),
\end{align}
\noindent and $\alpha_n(t)$ is the CD terms which can be found by solving a set of linear equations
\begin{align}
\begin{split}
&-3(J_n J_{n+1})\alpha_{n+1} + (J_{n-1}^2 + 4J^2_n + J_{n+1}^2)\alpha_j \\ &- 3(J_n J_{n-1})\alpha_{n-1} + (V_{n+1} - V_n)^2 \alpha_n \\ &= -\partial_{\lambda}J_n (V_{n+1} - V_{n}).
\end{split}
\end{align}
In order to implement COLD we include additional terms to the tunnelling of the lattice
\begin{align}\label{eq:tunneling_opt}
J_n(t) \rightarrow J_n(t, \boldsymbol{\beta}) = J_n(t) + \boldsymbol{\beta}(t),
\end{align}
which can then be incorporated into the forms of both $J_{n, CD}(t)$ and $\phi_{n, CD}(t)$. We again want the additional control terms to go to zero around the problem boundaries and a natural choice is the same as in the Ising spin chain example in Eq.~\eqref{eq:optimsable_1}. The parameters $\boldsymbol{\beta}$ are optimised as before by minimizing with respect to the fidelity of the final state, where the population has been fully transferred to the opposite lattice site.
We first consider a system size of $N=7$ sites which was successfully experimentally probed in Ref.~\cite{meier_counterdiabatic_2020}, where final state fidelities of $0.75$ were achieved for $\tau = 1$ms with a final tunnelling strength of $J/\hbar = 1/2\pi kHz$ (equivalent to $\tau \sim 1 J^{-1}$ in our units). We initially confirm the breakdown of ARP in this setting for fast times, and the success of the LCD protocol at short times, as shown in Fig.~\ref{fig:Synthetic} (a) and found in Ref.~\cite{meier_counterdiabatic_2020}. Implementing BPO on its own manages to enhance the achievable fidelities at intermediate times of $\tau > 0.03 J^{-1}$. However, eventually, as observed in all scenarios in this work, BPO becomes stuck in the initial state at fast times, and the fidelity goes to zero. Implementing the newly introduced COLD protocol achieves an order of magnitude improvement in the fidelity over LCD. This is also plotted in Fig.~\ref{fig:Synthetic}(a) alongside previous results of ARP and first-order LCD.
One concern could be that COLD is achieving this improvement by simply pumping power into the tunnelling term, but as we can see in Fig.~\ref{fig:Synthetic}(c) the maximum amplitude of the tunnelling term tracks that of LCD. A key issue for experiments is the maximum amplitude achievable by a driving term and with this result we can stipulate that COLD is likely to be feasible in the same regimes as LCD in this synthetic lattice system. There is single outlier at intermediate times as indicated by the single point peaking in maximum amplitude in Fig.~\ref{fig:Synthetic}(c), this is the exception to the rule, where the optimisation has found a marginally higher fidelity (see the offset point in Fig.~\ref{fig:Synthetic}(a)) by pumping in power.
A large concern for state transfer techniques is the robustness of a protocol with respect to an increasing system size. We show the best achievable fidelities with increasing system size for both BPO and COLD in Fig.~\ref{fig:Synthetic}(c). While both protocols show a decreasing fidelity with system size as is to be expected, once again COLD does not suffer from getting stuck in the initial state. This is shown by the BPO fidelities going to unity for large systems in Fig.~\ref{fig:Synthetic}(c), and is the same mechanism for this as for the short driving times in Fig.~\ref{fig:Synthetic}(a).
Another concern could be that BPO will beat COLD if enough parameters are allowed for the optimisation, i.e. if we increase $N_k$ enough. We observed no evidence of this for the Ising model example and we again do not observe this in this synthetic lattice example, as is shown in Fig.~\ref{fig:Synthetic}(d). Small improvements are made in the fidelities achieved with BPO and COLD for larger $N_k$ but this is not substantial.
\section{Discussion and outlook} \label{sec:conclusions}
We have introduced a new hybrid approach combining quantum optimal control and shortcuts to adiabaticity: COLD. Inspired by the successes of LCD, where diabatic transitions are suppressed and locality conditions can be met, COLD improves on its methodology by combining it with quantum optimal control. The natural way to enhance the performance of LCD is by introducing higher order CD terms, but these are often non-local and difficult to engineer in experiments. COLD circumvents this by allowing for additional control fields that extend the family of dynamical Hamiltonians which can be explored. In this way, our method may find the best possible path where the effect of lower-order LCD is most relevant and higher order corrections are suppressed.
COLD has a clear potential in efficiently speeding up adiabatic evolution in various settings. We demonstrate this numerically via several example protocols which indicate improvements beyond a classical optimisation approach BPO as well as LCD of different orders. Our work shows that COLD reduces the strength of higher order LCD corrections, and that it performs well for increasing system sizes. We have shown that COLD can be implemented in the limit of fast driving by a `switching off' of the original dynamical Hamiltonian. For scenarios where removing the Hamiltonian is not possible, we have shown that an alternative way to implement COLD is to use a bounded optimisation where amplitudes are restricted. We find that both the COLD and COLD-CRAB protocols perform extremely well in this setting.
COLD will be most beneficial when the LCD is only realisable to a certain order but the higher order corrections are large. This means the diabatic transitions are not being sufficiently suppressed by the choice of LCD and COLD can be used to find the dynamical Hamiltonian for which the required order of LCD term dominates. Note, that this goes the other way too, with COLD not providing substantial improvements when the chosen lower order LCD is small across the path. This can be thought of as being the case in two limits. First is the adiabatic limit, for which any CD correction is small and COLD will tend towards the adiabatic result. Second, the low-order LCD terms can be small compared to the driving as the exact CD would be correcting transitions due to interactions at longer ranges. In this scenario, the order of LCD being implemented with COLD needs to be increased, so that the CD term is accounting for the longer range terms. While we have only utilised COLD with first-order LCD in this work, it can be applied with higher-order LCD terms. We found this to be the case when we tried to apply COLD to the protocols for \@e.g.~Greenberger-Horne-Zeilinger state generation in arrays of Rydberg atoms in Ref.~\cite{omran_generation_2019}. COLD with first- or second-order LCD could not improve upon the results of BPO or CRAB. This was due to both the first- and second-order corrections being smaller than the dynamical terms of the original Hamiltonian by two orders of magnitude. It is perhaps not surprising when we consider that the Hamiltonian of this protocol includes long-range couplings. In this case, higher-order corrections would need to be implemented with COLD, and finding methods for executing these now highly non-local terms may be a future path of research.
COLD may also be applied to more complex systems where exact dynamics are not possible, \@e.g. due to an excessively large Hilbert space. This may be achieved by variationally minimising the integrals of the coefficients for the higher order corrections to the LCD. This was briefly sketched out in the example of two spin annealing, and we have shown that COLD functions by minimising the strength of the higher order LCD corrections along a given path. A further option is to combine COLD with one of a large variety of numerical optimal control methods, as we have done for the example of CRAB. We have shown a substantial improvement for state preparation in the Ising model that can be obtained from the COLD-CRAB combination - particularly in the constrained case. Both fusions of COLD with advanced optimal control methods as well as the minimisation of higher order LCD term amplitudes for complex systems are two potential extensions which could prove fruitful with further study.
\begin{acknowledgements}
Work at the University of Strathclyde was supported by the EPSRC Quantum Technologies Hub for Quantum Computing and Simulation (EP/T001062/1), and the European Union’s Horizon 2020 research and innovation program under grant agreement No. 817482 PASQuanS. A.P. acknowledges support from NSF under Grant DMR-2103658 and by the AFOSR under Grants No. FA9550-16-1-0334 and FA9550-21-1-0342.
\end{acknowledgements}
\section{Introduction}\label{sec:intro}
Time-dependent manipulation of few and many-particle quantum systems is important across all implementations of quantum computing and simulation. In such processes, decoherence and undesired transitions reducing the state fidelity are relatively ubiquitous. One important example is given by the undesired transitions that can occur between instantaneous eigenstates of the dynamical Hamiltonian upon the application of an external drive. This is why many driving protocols rely on adiabatic dynamics, where the system follows the instantaneous eigenstates and transitions are naturally suppressed. Ideal adiabatic processes are reversible making them - in principle - robust. However, to approach ideal adiabatic processes the dynamics must always be very slow, requiring compromises on the time-scales of competing heating and decoherence processes.
Speeding up adiabatic protocols to enable their completion within the system's coherence time is important for the development of any quantum technologies relying on such protocols \cite{acin2018quantum}. One approach to do this is the implementation of optimal driving protocols, which aim to end up with the system in a desired final state. For example, numerically optimised paths can be employed to avoid points where gaps in the spectrum of the system become small, or additional control fields can be tuned to increase the size of these gaps \cite{kirk2004optimal,glaser2015training,AlessandroBook2007}. In broad terms, this is the goal of protocols collectively referred to as quantum optimal control. Another option is to design techniques which speed up the adiabatic dynamics, often termed shortcuts to adiabaticity (STA). The primary aim of STA is to entirely remove or suppress diabatic transitions between instantaneous eigenstates of the dynamical Hamiltonian \cite{Torrontegui2013,GueryOdelin2019}. One particularly successful technique is counterdiabatic driving (CD), which was first utilised in physical chemistry by Demirplak and Rice \cite{demirplak2003,demirplak2005}, and was independently introduced by Berry \cite{berry_transitionless_2009} under the name `transitionless driving'. CD aims to suppress losses that arise due to fast deformations of the system far from the adiabatic limit by analytically compensating for them in the Hamiltonian. In general, to suppress diabatic losses exactly, the full analytical or numerical solutions of the Schr\"odinger equation are required. This makes the implementation of CD in complex systems - \@e.g.~for many-body dynamics - difficult and requires the need for new techniques to be introduced.
Links between optimal control and STA have existed throughout the development of both approaches \cite{stefanatos2021,Zhang2021connection}, but there are few examples of their explicit combination in a way that exploits their complementary nature. Some attempts to achieve this have included an emulation of CD through fast oscillations of the original Hamiltonian \cite{Petiziol2018fast, Petiziol2019accelerating} as well as through recent advances in reinforcement learning methods aimed at optimizing quantum protocols~\cite{bukov_reinforcement_2018}. Such methods have been shown to achieve a significant improvement in performance when implemented using concepts borrowed from CD~\cite{yao_reinforcement_2020}. In this work, we offer a significantly different new approach in combining elements from STA and quantum optimal control which we will call \textit{counterdiabatic optimised local driving} (COLD).
A key ingredient in the development of COLD is a recent approach designed for implementing CD in the setting of larger, more complex systems: local counterdiabatic driving (LCD) \cite{sels_minimizing_2017,Kolodrubetz2017geometry, gjonbalaj2021counterdiabatic}. LCD offers a method to derive \emph{approximate} CD protocols, with the aim of suppressing undesired transitions instead of fully eliminating them. This allows it to account for some physical constraints of the system, \@e.g.~locality conditions. However, the approximate nature of the LCD protocol can lead to poor performance, necessitating the introduction of additional non-local, long-range corrections \cite{sels_minimizing_2017}. If all possible corrections are added, then LCD is equivalent to the normal analytical approaches of CD, but the additional terms are generally difficult to control experimentally. COLD offers an alternative approach, with additional control fields which allow for an optimisation of the dynamical Hamiltonian for a given local form of LCD. The impact of more complex corrections can then be radically reduced, giving a corresponding improvement in the desired protocol.
The structure of this paper is as follows: first, we give a detailed description of the new method, COLD, with a focus on the elements of quantum optimal control and CD required for its implementation. In Sec.~\ref{sec:TwoSpin} we explore a 2-spin annealing protocol, that showcases the strengths of COLD. Sec.~\ref{sec:1dIsing} describes and analyses the improvements gained with COLD and its combination with other optimal control techniques in the case of state preparation in the Ising model. Then in Sec.~\ref{sec:lattice} we show the improvement that COLD can achieve on the recently realised example of LCD for state transfer on a synthetic lattice in ultracold atoms. A list of abbreviations used in this work can be found in Table.~\ref{table} for reference.
\begin{table}[h]
\begin{tabular}{p{2cm} | p{6cm}} \toprule
\multicolumn{1}{m{2cm}}{\centering Abbreviation}
& \multicolumn{1}{m{6cm}}{\centering Meaning} \\
\midrule
STA & shortcuts to adiabaticity \\ \hline
CD & counterdiabatic driving \\ \hline
LCD & local counterdiabatic driving \\ \hline
COLD & counterdiabatic optimised local driving \\ \hline
BPO & bare Powell optimisation \\ \hline
CRAB & chopped randomised basis \\ \hline
ARP & adiabatic rapid passage \\
\bottomrule
\end{tabular}
\caption{List of abbreviations used throughout the manuscript.}\label{table}
\end{table}
\section{An Introduction to Counterdiabatic Optimised Local Driving}\label{sec:intro_olcd}
\subsection{Quantum Optimal Control}\label{sec:OptCont}
In the context we consider, we employ quantum optimal control to optimise the function $f(\psi,\boldsymbol{\beta})$ in the Schr\"odinger equation
\begin{equation}
\dot{\psi} = f(\psi,\boldsymbol{\beta}),
\label{eq:Control}
\end{equation}
where $\psi$ is the quantum wave function and $\boldsymbol{\beta}$ is the set of optimisable control parameters. Optimisation of Eq.~\eqref{eq:Control} in most cases means taking the system from an initial state $\ket{\psi_0}$ to a final target state $\ket{\psi_T}$ by finding the optimal values of $\boldsymbol{\beta}$ with respect to some target metric (e.g.~the time taken to evolve the system from $\ket{\psi_0}$ to $\ket{\psi_T}$). There is a large variety of techniques available to achieve this goal \cite{glaser2015training,koch2016controlling}.
The success/target metric needs to be defined prior to the optimisation of $\boldsymbol{\beta}$. Often this is done by constructing a \emph{cost function}, which in turn defines the optimisation landscape. In general, we can optimise for any desired property of the final state of the system, with some examples being the entropy, energy, energy fluctuations or some other observable. A commonly used cost function in state preparation is related to the fidelity of the final, post-evolution state $\ket{\psi_f}$ with respect to the target state:
\begin{align} \label{eq:lossfunc}
\mathcal{C}(\ket{\psi_f}) = 1 - \left|\braket{\psi_T}{\psi_f}\right|^2.
\end{align}
In performing such a numerical optimisation, it is common to take the target state to be parameterised via a Hamiltonian split into two parts. The first is the so-called \emph{bare} Hamiltonian $H_0(t)$, which can be time-dependent and describes the dynamics of the quantum system in question. The second part is then an additional driving term that includes the control parameters $\boldsymbol{\beta}(t)$ and operators $\mathcal{O}_{\rm opt}$ which provide additional degrees of freedom in the dynamics of the system. The full Hamiltonian of the control system is then:
\begin{align} \label{eq:h_optimal_control}
H_{\beta}(t,\boldsymbol{\beta}) = H_0(t) + \boldsymbol{\beta}(t)\mathcal{O}_{\rm opt}.
\end{align}
The parameters $\boldsymbol{\beta}(t)$ can then be optimised for the optimal dynamics with respect to the metric defined by the cost function.
In this work, we use the Powell minimization approach \cite{powell1964efficient} for the numerical optimisation as implemented in Python's \textit{SciPy} \cite{Pauli2020SciPy}. When performing this optimisation without any CD terms in the Hamiltonian, we refer to this approach as bare Powell optimisation (BPO), with bare referring to the lack of CD.
Furthermore, we implement the chopped rendomised basis (CRAB) approach \cite{Caneva_chopped_2011,muller2021} and combine its methodology with that of COLD, to obtain the method of COLD-CRAB. CRAB expands the size of the parameter landscape by employing randomisation, usually in the optimised pulse driving the system. The approach was first developed for quantum many-body problems whose simulation requires the time-dependent density matrix renormalization group, despite the fact that these were thought to not be tractable in the quantum control setting \cite{brif2010,muller2021}. CRAB has benefits in that it can avoid traps in the control landscape \cite{Rach2015}, and has built-in flexibility for open-loop or closed-loop optimisation \cite{heck2018,muller2021} although these advantages come at a higher computational cost due to requiring a far larger search-space for the optimisation.
\subsection{Counterdiabatic Driving}\label{sec:LCD}
An important class of optimisation problems deals with the case where the initial and final states are ground states of a Hamiltonian $H_0(t)$ at some initial $t=t_i$ and final $t=t_f$ time. In these cases, the adiabatic theorem guarantees that for an infinitesimally slow transformation of the system $t_f-t_i\to\infty$, it should follow the instantaneous (non-degenerate) ground state of $H_0(t)$ and hence reach the target state with unit fidelity. This process is generally known as quantum annealing.
In large, complex systems with many degrees of freedom, quantum annealing tends to require prohibitively long protocol times due to vanishingly small gaps typically present in such systems. This often makes annealing protocols impractical \cite{farhi2008how}. It has been found that this problem can be formally overcome by using CD, where velocity-dependent terms are added to the Hamiltonian analytically enforcing the adiabatic wave function to be the solution of the time-dependent Schr\"odinger equation \cite{demirplak2003,demirplak2005,berry_transitionless_2009}. In this case, the dynamical state will follow the instantaneous eigenstate with no transitions regardless of the driving time. The form of the dynamical Hamiltonian enforcing this is \cite{berry_transitionless_2009}:
\begin{equation}\label{eq:counterdiabatic}
\begin{aligned}
& H_{\mathrm{CD}}(t) = H_0 (t) \\ & + i\hbar \sum_n (\ket{\partial_t n}\bra{n} - \bra{n}\ket{\partial_t n}\ket{n}\bra{n}),
\end{aligned}
\end{equation}
with $\ket{n}\equiv \ket{n(t)}$ the $n$-th eigenstate of the instantaneous Hamiltonian $H_0(t)$. The last term enforces the phases ($\bra{n}\ket{\partial_t n}$) on the instantaneous eigenstates, which are arbitrary and thus will be omitted. In general, knowledge of the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} requires knowledge of the full spectrum of $H_0(t)$ at each instantaneous moment in time.
\subsection{Counterdiabatic Optimised Local Driving} \label{sec:OLCD}
We will now introduce the main idea of COLD. The principle is to take the same approach as Sec.~\ref{sec:LCD} but with the original Hamiltonian given by $H_\beta(t,\boldsymbol{\beta})$, see Eq.~\eqref{eq:h_optimal_control}. Quantum Annealing then applies to the whole family of control Hamiltonians $ H_{\beta}(t,\boldsymbol{\beta})$ as long as the additional control terms $\boldsymbol{\beta}(t)$ vanish at the protocol boundaries: $\boldsymbol{\beta}(t_i)=\boldsymbol{\beta}(t_f)=0$. This flexibility was explored in finding the optimal adiabatic path characterized by e.g. the shortest distance between the initial and the final states, i.e. a geodesic \cite{tomka2016geodesic}. A similar geodesic approach was developed in the context of dissipative systems to minimize energy losses~\cite{Sivak2012Geodesic}. During the protocol, a dynamical Hamiltonian $H_\beta(t,\boldsymbol{\beta})$ generally induces transitions between the quantum states that it drives and the question about what is the optimal path remains open.
The Hamiltonian $H_\beta(t,\boldsymbol{\beta})$ and its eigenstates depend on time only through the driving parameters, which include $\boldsymbol{\beta}$ and any additional control terms in the particular protocol. This makes it convenient to introduce a path in the coupling space parametrized by a dimensionless parameter $\lambda\in [0,1]$ such that both $H_0$ and $\boldsymbol{\beta}$ are functions of $\lambda$ satisfying $H_\beta(\lambda=0)=H_{0}(t_i)$ and $H_\beta(\lambda=1)=H_{0}(t_f)$, i.e. being equal to the initial and the final Hamiltonian at the protocol boundaries. By construction this implies that any additional fields introduced to the bare Hamiltonian $H_0$ must go to zero at the boundaries. In this way, any protocol can be uniquely characterized by first specifying the path $\boldsymbol{\beta}(\lambda)$ in the coupling space manifold and then the time dependence $\lambda(t)$ along it. The path determines the sequence of couplings of the Hamiltonian during time evolution and hence the sequence of ground state wave functions followed by the driven state. Furthermore, the time dependence encodes the speed of traversing this path. We can then introduce a hermitian operator called the (path-dependent) adiabatic gauge potential~\cite{sels_minimizing_2017}: $\mathcal A_\lambda=i \hbar \sum_n \ket{\partial_\lambda n}\bra{n}$, which satisfies a closed form equation,
\begin{equation}
\label{eq:AGP_eq}
\left[\partial_\lambda H_\beta+{i\over \hbar} [ \mathcal A_\lambda,H_\beta],H_\beta\right]=0,
\end{equation}
where $H_\beta \equiv H_\beta(\lambda,\boldsymbol{\beta}(\lambda))$ and $\mathcal{A}_\lambda \equiv \mathcal{A}_\lambda(\lambda,\boldsymbol{\beta}(\lambda))$. Then the CD Hamiltonian reads
\begin{align}\label{eq:counterdiabaticLCD}
H_{\mathrm{CD}}(\lambda,\boldsymbol{\beta}(\lambda)) = H_\beta(\lambda,\boldsymbol{\beta}(\lambda)) +\dot \lambda \mathcal A_\lambda(\lambda,\boldsymbol{\beta}(\lambda)),
\end{align}
and is equivalent to the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} given knowledge of the exact adiabatic gauge potential. However, generally the adiabatic gauge potential is a very non-local object and solutions of Eq.~\eqref{eq:AGP_eq} are unstable to small perturbations containing exponentially many terms in the number of degrees of freedom.
LCD aims to find approximate gauge potentials that satisfy particular requirements like robustness and locality, thus circumventing many of the difficulties in determining the second component in Eq.~\eqref{eq:counterdiabatic} and~\eqref{eq:counterdiabaticLCD} exactly. The goal, in essence, is to suppress the most relevant diabatic effects rather than completely eliminate them. This method has recently been experimentally implemented to speed up state transfer for synthetic lattices in ultracold atoms \cite{meier_counterdiabatic_2020}, for preparing states in nuclear-magnetic-resonance systems \cite{zhou_experimental_2020}, and annealing protocols on an IBM quantum computer \cite{Hegade2021Shortcuts}.
Following the methods of Ref.~\cite{sels_minimizing_2017}, the problem of finding the optimal adiabatic gauge potential can be cast as the minimisation of the Hilbert-Schmidt norm of the operators
\begin{equation}\label{eq:Goperator}
G_{\lambda}= \partial_{\lambda}H_\beta + i\comm{\mathcal{A}_\lambda}{H_\beta},
\end{equation}
which is equivalent to minimisation of the action
\begin{equation}\label{eq:actionCD}
\mathcal{S}(\mathcal{A}_{\lambda}) = \Trace{\left[G_{\lambda}(\mathcal{A}_{\lambda})^2\right]},
\end{equation}
with respect to $\mathcal{A}_{\lambda}$. In most cases, this is achieved by first choosing an operator ansatz - \@i.e. a set of linearly independent operators $\{\mathcal{O}_{\rm LCD}\}$ - and then using this set as an operator basis for the adiabatic gauge potential $\mathcal A_\lambda=\sum_j \alpha_j \mathcal O_{\rm LCD}^{(j)}$. The action can then be minimized with respect to the the set of coefficients, ${\bm \alpha}$. In the example of an
Ising spin chain we may take $\mathcal{A}_\lambda = \sum_j^N \alpha_j \sigma^y_j$, where $j$ labels the $N$ chain sites, and $\{\mathcal{O}_{\rm LCD}\}$ is a set the $y$-pauli matrices.
Without any additional control fields $\boldsymbol{\beta}$, LCD is essentially an informed choice of the operator set $\{\mathcal{O}_{\rm LCD}\}$ in a way that the resulting control protocol from the minimisation of Eq.~\eqref{eq:actionCD} is optimal for a given $H_0(\lambda)$. In this case we explore the family of Hamiltonians
\begin{equation}
H_{\rm LCD}(\lambda) = H_0(\lambda) +\sum_j \alpha_j(\lambda) \mathcal{O}_{\rm LCD}^{(j)}.
\end{equation}
The performance of such LCD protocols is determined by how accurately the variational manifold spanned by the set $\{\mathcal{O}_{\rm LCD}\}$ can approximate an exact $\mathcal{A}_{\lambda}$ such that Eq.~\eqref{eq:AGP_eq} holds.
In the case of the new protocol COLD, we allow for extra exploration of the family of Hamiltonians due to the additional control fields as in Eq.~\eqref{eq:h_optimal_control}. This expands the family of Hamiltonians to
\begin{equation}\label{eq:expandedH}
H_\beta(\lambda,\boldsymbol{\beta}) = H_0(\lambda) + {\bm \alpha}(\lambda,\boldsymbol{\beta}) \mathcal{O}_{\rm LCD} + \boldsymbol{\beta}(\lambda) \mathcal{O}_{\rm opt}.
\end{equation}
Note, that the coefficients of the optimal control field change the form of the LCD driving coefficients, i.e. $\bm \alpha = f(\lambda,\boldsymbol{\beta})$. The aim of COLD is then to choose $\boldsymbol{\beta}(\lambda)$ and $\mathcal{O}_{\rm opt}$ such that the LCD term is optimal for the dynamical Hamiltonian $H_0(\lambda)+\boldsymbol{\beta}(\lambda)\mathcal{O}_{\rm opt}$. We will focus on the optimisation of the control parameters $\boldsymbol{\beta}(\lambda)$ for a given choice of $\mathcal{O}_{\rm opt}$, although the choice of operators $\mathcal{O}_{\rm opt}$ can also be optimised over as an extension.
With COLD, we have two methods to improve on the existing LCD protocol. As previously shown in Refs.~\cite{claeys_floquet-engineering_2019,prielinger_diabatic_2020}, there is a possibility to add more terms to the LCD making it less local, \@e.g.~through long-range interactions. In the spin chain case, we could take the aforementioned sum over $\sigma^y$ terms to be the \textit{first-order} anzatz for the LCD, where higher-order ans\"{a}tze might contain sets of operators $\{\mathcal{O}_{\rm LCD}\}$ with terms odd in $\sigma^y$ such as $\sigma^y_j\sigma^{(z,x)}_{j+1}$. This procedure generally improves the performance of CD protocols at the expense of adding more complex operators which may be experimentally impractical depending on the scenario. Alternately, with COLD and the introduction of additional local control fields to the Hamiltonian, we can improve the performance of LCD at a fixed complexity of the CD term. If these extra control fields vanish at the beginning and at the end of the protocol, they do not affect the adiabatically connected states, but they can significantly modify the adiabatic landscape at intermediate couplings to enhance the performance of the given order of LCD.
In order to optimize for the additional fields, in this work, we use methods of quantum optimal control. However, we note that such optimization can be done locally by requiring that the next order variational correction to the CD terms is small along the optimal landscape. This local optimization may be advantageous in that it does not require knowledge of the wave function and in this sense is not limited by small system sizes.
Furthermore, we compare COLD to the use of CRAB, as discussed in Sec.~\ref{sec:OptCont}. An advantage of COLD is that it can be combined with many advanced optimal control procedures, owing to the standard way additional control fields are introduced to the Hamiltonian. In this work we find the combination of COLD and CRAB particularly useful and we will refer to this as COLD-CRAB.
\section{Two Spin Quantum Annealing}\label{sec:TwoSpin}
To showcase and explore the use of COLD in a relatively simple setting we will consider a two spin quantum annealing problem with bare Hamiltonian
\begin{equation}\label{eq:Hanneal}
H_0(t) = -2J \sigma_1^z \sigma_2^z - h ( \sigma_{1}^z + \sigma_{2}^z) + 2h \lambda(t) (\sigma_{1}^x + \sigma_{2}^x),
\end{equation}
where $\sigma^a_j$, $a \in \{x,y,z\}$ are the Pauli matrices applied to spins indexed by $j$. For the scaling function $\lambda(t)$ we pick the form
\begin{equation}\label{eq:Scalingfunc}
\lambda(t) = \sin^2\left(\frac{\pi}{2} \sin^2 \left( \frac{\pi t}{2 \tau} \right) \right)
\end{equation}
such that $\lambda(0) = 0$ and $\lambda(\tau) = 1$. We consider the case of $J/h=0.5$, which means the spins start in the initial state of $\ket{\uparrow \uparrow}$ and finish in a superposition of all of the symmetric states.
As discussed in Ref.~\citep{sels_minimizing_2017}, since $H_0$ has a standard Ising spin chain form, the first-order LCD terms are given by the following ansatz for the adiabatic gauge potential:
\begin{equation}\label{eq:LCD1st}
\mathcal{A}(\lambda) = \alpha \sum_{i=1}^2 \sigma_i^y,
\end{equation}
with the sum being over the full length of the $N$ spin chain. Minimising Eq.~\eqref{eq:actionCD} for this $\mathcal{A}_{\lambda}$ with respect to the coefficient $\alpha$ gives
\begin{equation}
\alpha = - \frac{h^2}{4(h\lambda)^2 + h^2 + 4J^2}.
\end{equation}
To further improve on the first-order LCD we can implement COLD, as we will discuss shortly, or we can introduce higher-order terms to the ansatz for $\mathcal{A}_{\lambda}$. This second method serves as a good benchmark against COLD, since it offers an improvement to first-order LCD in the same way as COLD does, but requires more complicated interactions between the two spins increasing the implementation overhead. The second-order LCD can be found by taking an ansatz for the adiabatic gauge potential:
\begin{equation}
\begin{aligned}
\mathcal{A}^{(2)}(\lambda) =& \alpha \sum_{j} \sigma_j^y + \gamma (\sigma_1^x \sigma_{2}^y + \sigma_1^y \sigma_{2}^x) \\ & + \zeta (\sigma_1^z \sigma_{2}^y + \sigma_1^y \sigma_{2}^z),
\end{aligned}
\end{equation}
where to solve for $\alpha$, $\gamma$, and $\zeta$ we once again minimize the action given by Eq.~\eqref{eq:actionCD} and obtain three coupled equations which can be solved numerically (see Appendix \ref{app:derivation} for a detailed derivation).
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{TwoSpin.png}
\caption{Optimisation of the annealing protocol for two spin Hamiltonian given by Eq.~\eqref{eq:Hanneal} for $h/J=2$. (a) Final fidelities of the annealing protocol with triangles (black) representing the case where no CD is applied and circles showing the case of first-order LCD (pink) as well as the combination of first- and second-order LCD (orange). (b) Final fidelities achieved when using the optimal control method BPO (red diamonds) and the new approach of COLD (blue circles), both with $N_k=1$.}\label{fig:TwoSpin}
\end{figure}
We now consider three distinct cases in this two spin quantum annealing example: no LCD, first-order LCD, and second-order LCD. The fidelity of the final state for each case over a wide range of driving times is shown in Fig.~\ref{fig:TwoSpin}(a), with an easily distinguishable advantage in the case of LCD. The final fidelity where no LCD is implemented decreases rapidly as the ramp times are made short, with the system getting stuck in its initial state. On the contrary, first-order LCD retains good final state fidelities into short times, as the driving Hamiltonian becomes that of only the LCD term. The second-order LCD then gives unit fidelity, in agreement with previous observations \cite{claeys_floquet-engineering_2019}, as for a two spin Hamiltonian the highest order corrections are that including two spin terms.
We now add an optimisable term, as described in Sec.~\ref{sec:OptCont}, so that the new Hamiltonian reads:
\begin{equation}\label{eq:HannealOpt}
H_\beta(t) = H_0(t) + \sum_{k=1}^{N_k} \beta^k \sin (\pi k t / \tau) \sum_i \sigma_i^z,
\end{equation}
\noindent with $N_k$ the number of optimisation coefficients, and $\beta^k$ the coefficient of the $k$th frequency of the control function. Note that we consider
\begin{equation}
\boldsymbol{\beta}(t) = \sum_{k=1}^{N_k} \beta^k \sin (\pi k t / \tau) = \sum_{k=1}^{N_k} \beta^k \sin (\pi k f(\lambda)),
\end{equation}
with
\begin{equation}\label{eq:lambda}
f(\lambda) = \frac{2}{\pi}\arcsin\left(\sqrt{\frac{2}{\pi} \arcsin\left(\sqrt{\lambda}\right)}\right).
\end{equation}
The form of the additional control field fulfils the requirement that the boundary conditions are $H(0) = H_0(0)$ and $H(\tau) = H_0(\tau)$. Numerically optimising the $\beta^k$ for the best final state fidelity \emph{without} adding LCD terms results in the BPO method introduced in Sec.~\ref{sec:OptCont}. We show the results of BPO in Fig.~\ref{fig:TwoSpin}(b), where it is observed that BPO gives better results than the case of no LCD in Fig.~\ref{fig:TwoSpin}(a). However, for short times the BPO approach still results in the system getting stuck in the initial state.
\begin{figure*}[t]
\includegraphics[width=0.9\linewidth]{IsingUnconstrained.png}
\caption{Optimisation of the annealing protocol for the Ising model given by Eq.~\eqref{eq:h0_ising} for $N=5$ spins. (a) A comparison of final state fidelities for different driving times using the optimal control technique BPO (blue diamonds), first-order LCD (black dash-dot line) and COLD (red circles). The same is shown in (b) for CRAB (green diamonds) and COLD-CRAB (purple circles). CD enhanced techniques (COLD and COLD-CRAB) introduced in this work show a clear convergence to good fidelities at short driving times. All results are for the best (lowest) fidelity achieved over $500$ optimisations.}\label{fig:IsingUnconstrained}
\end{figure*}
Finally we present and compare the results of the new method, COLD. In this case the Hamiltonian before adding LCD terms is given by Eq.~\eqref{eq:HannealOpt} and the coefficient of the first-order LCD is
\begin{equation}
\alpha = -\frac{h(h+\boldsymbol{\beta}) + h\frac{\lambda}{\dot{\lambda}} \dot{\boldsymbol{\beta}}}{4(h\lambda)^2 + (h+\boldsymbol{\beta})^2 + 4J^2}.
\end{equation}
Note that the optimisation of the additional control field also feeds into the coefficient of the adiabatic gauge potential during the dynamics as discussed in Sec.~\ref{sec:OLCD}. The results of the COLD approach for this two spin annealing protocol are shown in Fig.~\ref{fig:TwoSpin}(b), where we observe an improvement of the final state fidelity beyond what is possible with first-order LCD alone in Fig.~\ref{fig:TwoSpin}(a). In this example, LCD alone reaches a final state fidelity of $1-F=3\%$ at short times, however COLD improves this error in the final state to $1-F=0.005\%$. This is due to the extended family of dynamical Hamiltonians in Eq.~\eqref{eq:expandedH} owing to the addition of an optimisable control field. This result shows that COLD can provide an advantageous alternative to the addition of higher-order LCD which may be experimentally impractical.
We have found that COLD performs better than LCD of the same order or BPO when the system dynamics are calculated numerically. This does not, however, imply anything about the performance of COLD in more complex scenarios, like in the case of an unknown target ground state. In that case the fidelity is a poor optimisation metric. There is, however, a way to come to the same conclusions as those presented in Fig.~\ref{fig:TwoSpin} without the need to compute the dynamics exactly. We can do this by first using a guess for the COLD protocol to find the approximate adiabatic gauge potential and then minimising the integral of the norm of the second-order correction to the adiabatic gauge potential along the path. Note, that the ground state can be in turn obtained through first order COLD, so there is no need to diagonalize the Hamiltonian. This integral should be small if COLD has implemented a dynamical Hamiltonian that makes the first-order adiabatic gauge potential the leading term. It is effectively a measure of the error of COLD and can be given by
\begin{equation}
\begin{aligned}
\mathcal{I}_1 = \int_0^\tau dt^\prime \Big[& \bra{\psi_g(t^\prime)} \Gamma^2(t^\prime) \ket{\psi_g(t^\prime)} \\ &- (\bra{\psi_g(t^\prime)} \Gamma(t^\prime) \ket{\psi_g(t^\prime)})^2\Big]^{1/2},
\end{aligned}
\end{equation}
\noindent with $\ket{\psi_g(t)}$ the instantaneous ground state along the path and
\begin{equation}
\Gamma(t) = \gamma(t) \left( \sigma_1^y \sigma_2^x + \sigma_1^x \sigma_2^y \right),
\end{equation}
\noindent one of the second-order correction terms. In order to confirm this is the case, we compare the different paths -- COLD and LCD only -- in the two-spin example in order to determine if $\mathcal{I}_1$ is small for COLD. If $\mathcal{I}_1$ is small when compared to the same measure for lower-order LCD as $t\rightarrow 0$, then we know that COLD is enforcing a better dynamical Hamiltonian. In the case of the two spin annealing protocol we find that as $t\rightarrow0$, $\mathcal{I}_1\rightarrow 0.04$ for COLD and $\mathcal{I}_1\rightarrow 0.2$ for LCD, showing that COLD is minimising the second-order correction along the path. A simpler integral
\begin{equation}
\mathcal{I}_2 = \int_0^\tau dt^\prime |\gamma(t^\prime)|,
\end{equation}
also reflects this correction in this two spin example, with $\mathcal{I}_2 \rightarrow 0.03$ for COLD and $\mathcal{I}_2 \rightarrow 0.1$ for LCD as $t\rightarrow0$. This could be useful in more complex scenarios as $\mathcal{I}_2$ will be relatively simple to calculate. We also observe the reduction of the corresponding integrals of the $(\sigma^y_1\sigma^z_2 + \sigma^z_1\sigma^y_2)$ term of the second-order LCD. By minimising these integrals, it could be possible to extend the COLD approach to more complex scenarios, including where the exact calculation of the dynamics is not possible.
\section{1D Ising Model} \label{sec:1dIsing}
In this section we apply COLD for state preparation on a 1D Ising spin chain in the presence of a transverse and longitudinal field. We consider an annealing protocol where the aim is to prepare the ground state across the Ising phase transition. The annealing Hamiltonian is given by
\begin{align}\label{eq:h0_ising}
\begin{split}
H_{0}(t) &= - J \sum_{j}^{N-1} \sigma^z_j\sigma_{j+1}^z + Z_0\sum_j^N \sigma_j^z \\ &+ \lambda(t) X_f \sum_j^N \sigma_j^x,
\end{split}
\end{align}
where $Z_0$ is a small offset parameter to break ground state degeneracies and $X_f$ is the final x-field strength. Note, the breaking of the ground state degeneracies is not a requirement but allows for easier consideration of the adiabatic path. As before, $\lambda(t)$ is a scaling function that has the boundary conditions $\lambda(0) = 0$ and $\lambda(\tau) = 1$, with $\tau$ the driving time. This means we start from the ground state of all spins up and drive across the quantum phase transition to the ground state which is a superposition of all basis states. We again take the scaling function to be given by Eq.~\eqref{eq:Scalingfunc}. In this example, we use $X_f = 10J$ and $Z_0 = 0.02J$.
For the Hamiltonian of Eq.~\eqref{eq:h0_ising}, the LCD to first and second order is well known, as the wave functions are entirely real. We take the first-order adiabatic gauge potential to be given by
\begin{equation}\label{eq:LCD1}
\mathcal{A}(\lambda) = \alpha \sum_{j}^N\sigma_j^y,
\end{equation}
\noindent where the coefficients for the general periodic spin chain of Eq.~\eqref{eq:h0_ising} are
\begin{align} \label{eq:alphas}
\alpha(\lambda) = \frac{1}{2} \frac{Z_0 X_f}{Z_0^2 + \lambda^2 X_f^2 + 2J^2}.
\end{align}
Note, that the quoted $\alpha$ above is technically for a periodic or infinite size system, with $J^2 \rightarrow J^2(1-1/N)$ for a finite system. However, we find that the inclusion of the factor for the finite system sizes we consider only changes the final achieved converged fidelities at short times by $\sim 10^{-6}\%$. The second-order adiabatic gauge potential is of the form
\begin{equation}
\begin{aligned}
\mathcal{A}^{(2)}(\lambda) =& \alpha \sum_{j} \sigma_j^y + \gamma \sum_{j} (\sigma_j^x \sigma_{j+1}^y + \sigma_j^y \sigma_{j+1}^x) \\ & + \zeta \sum_{j} (\sigma_j^z \sigma_{j+1}^y + \sigma_j^y \sigma_{j+1}^z) ,
\end{aligned} \label{eq:SecondLCD}
\end{equation}
with the coefficients $\alpha$, $\gamma$ and $\zeta$ again obtained by minimising the action given by Eq.~\eqref{eq:actionCD} and solving the coupled set of equations numerically (see Appendix \ref{app:derivation} for a detailed derivation).
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{MaxAmp.png}
\caption{Maximum amplitudes of CD terms in the Ising model annealing protocol for (a) first- and second-order LCD only with no additional optimal control fields and (b) the COLD approach optimised for the best final state fidelity implementing first-order LCD as shown in Fig.~\ref{fig:IsingUnconstrained} (a). The plot shows the maximum amplitude at each driving time for the first-order $\alpha$ (red circles) and the two second-order terms $\gamma$ (blue diamonds) and $\zeta$ (green triangles) as given in Eq.~\eqref{eq:SecondLCD} (although the second-order LCD is not actually implemented in COLD). An inversion in the strength of the second-order and first-order LCD terms for (a) no additional optimal control fields and (b) the addition of optimal control fields shows that COLD implements a dynamical Hamiltonain which is favourable for the applied order of LCD (first-order in this case).}\label{fig:MaxAmp}
\end{figure}
In this example, optimal control is implemented by introducing an additional driving field so that the dynamical Hamiltonian is given by
\begin{align}\label{eq:h_beta}
H_{\beta}(t, \boldsymbol{\beta}) = H_0(t) + \sum_j \boldsymbol{\beta}(t)\sigma^z_j,
\end{align}
with $\boldsymbol{\beta}$ being the terms to optimise over. We take our additional terms to again respect the boundary conditions $\boldsymbol{\beta}(0) = 0$ and $\boldsymbol{\beta}(\tau) = 0$, meaning a natural choice is
\begin{align}\label{eq:optimsable_1}
\boldsymbol{\beta}(t) = \sum_k^{N_k} \beta^k \sin(\omega_k t / \tau) = \sum_k^{N_k} \beta^k \sin(\omega_k f(\lambda)),
\end{align}
with $\omega_k = k 2\pi$ the $k$th principal frequency and $f(\lambda)$ given by Eq.~\eqref{eq:lambda}. To implement the CRAB algorithm discussed in Sec.~\ref{sec:OptCont}, we will use $k \rightarrow k(1+r)$ instead with $r$ drawn from a uniform random distribution $r \in [-0.5,0.5]$.
As before, we choose the first order adiabatic gauge potential given by Eq.~\eqref{eq:LCD1} and find that the coefficients are
\begin{align}
\alpha(\lambda,\boldsymbol{\beta}) = \frac{X_f}{2} \frac{(Z_0 + \boldsymbol{\beta}) - \lambda\dot{\boldsymbol{\beta}}/\dot{\lambda}}{(Z_0 + \boldsymbol{\beta})^2 + \lambda^2 X_f^2 + 2 J^2}.
\end{align}
Note, with the introduction of the additional control fields $\boldsymbol{\beta}$ it is possible for $\alpha$ to be non-zero at the start or end of the protocol, as $\dot{\boldsymbol{\beta}}$ is not fixed to be zero. However, this can be enforced by a suitable choice of the additional control field, we will consider replacing $\alpha \rightarrow S(\lambda)\alpha$ where $S(\lambda)$ is a scaling function that tends to zero as $\lambda \rightarrow 0$ and $\lambda \rightarrow 1$. We find that the scaling function only has a minimal effect on the final fidelities observed. This issue could also be resolved by a suitable choice of $\boldsymbol{\beta}$, with our example drive being an extreme case as $\dot{\boldsymbol{\beta}}$ is maximal at the boundaries of the protocol. The suitable choice of the form of $\boldsymbol{\beta}$ in a given example is a problem we will leave for future work, with our focus being on the introduction of the COLD protocol.
\begin{figure}[t]
\includegraphics[width=0.98\linewidth]{ScalingN.png}
\caption{Scaling of fidelities in the annealing protocol for the Ising model with (a) system size $N$ and (b) optimisation parameters $N_k$ at driving time $\tau=10^{-2}J^{-1}$. Plots show a comparison between BPO (blue diamonds) and COLD (red circles). In (a) we see that the COLD fidelity decreases as a function of $N$ but remains quite high when compared to BPO while (b) shows the non-existent improvement for both BPO and COLD with an increasing number of parameters in the $N=5$ spin case. Once again, plotted best fidelities are obtained across 500 optimisations.}\label{fig:ScalingN}
\end{figure}
We first compare the final state fidelity when using COLD versus BPO as shown in Fig.~\ref{fig:IsingUnconstrained}(a) for different driving times in a system of $N=5$ spins and a single $N_k=1$ optimisation coefficient. As expected, at long timescales the two methods agree as we approach the adiabatic limit of the dynamics. However, at shorter time scales the difference in behaviour is dramatic. We observe that the BPO approach fails in the case of very fast driving as the state gets stuck in the initial state but the COLD approach converges to $1-F \sim 10^{-3}$. Note, this is not achieved by the introduction of first-order LCD terms alone, as this will result in $F=0.0440$ for $\tau=10^{-3}J^{-1}$. COLD is instead achieving this by making the LCD term dominant for the dynamical Hamiltonian through the additional control fields.
\begin{figure*}[t]
\includegraphics[width=0.9\textwidth]{IsingcConstrained.png}
\caption{Optimisation of the \emph{constrained} annealing protocol for the Ising model for $N=5$ spins with a maximum amplitude limit on each term in the Hamiltonian of Eq.~\eqref{eq:h0_ising} of $10J$. (a) Shows a comparison between BPO (blue diamonds) and COLD (red circles) which both give much lower fidelities than in the unconstrained case in Fig.~\ref{fig:IsingUnconstrained}, although COLD persists in giving better results. In (b) the comparison is between CRAB (green diamonds) and COLD-CRAB (purple circles) which show orders of magnitude better fidelities than those in (a), with COLD-CRAB eking out higher fidelities at short driving times. The plotted best results are obtained from 200 optimisations for each method.}\label{fig:IsingcConstrained}
\end{figure*}
To confirm this, we plot the maximum amplitudes of both the first- and second-order adiabatic gauge potentials in Fig.~\ref{fig:MaxAmp}, where Fig.~\ref{fig:MaxAmp}(a) shows the case of no optimisation and Fig.~\ref{fig:MaxAmp}(b) the case of applying COLD. We can see that without COLD the second-order $(\sigma^x_j\sigma^y_{j+1} + \sigma^y_j\sigma^x_{j+1})$ corrections to the LCD are far larger than the first-order, resulting in the small final state fidelities when only first-order LCD is implemented. In the case of COLD, this relationship reverses and the first-order LCD terms dominate the dynamics. These results are further evidence that the implementation of COLD through the minimisation of the second-order corrections discussed in Sec.~\ref{sec:TwoSpin} could be fruitful in more complex and/or larger systems, where the dynamics can not be calculated.
We find that the results of BPO and COLD at short driving times are stable against increasing system size, as shown for $\tau = 10^{-2}J^{-1}$ in Fig.~\ref{fig:ScalingN}(a), with only a small decrease in final state fidelity for larger systems with COLD. Similarly, increasing the number of optimisation coefficients $N_k$ results in little improvement in the values obtained at short times for this example, as shown in Fig.~\ref{fig:ScalingN}(b). It is possible that in more complex systems, more optimisation coefficients will be needed to gain a larger advantage. We also note that by increasing the number of coefficients, we are increasing the complexity of the cost function landscape to be explored by the minimisation procedure, hence leading to slightly worse final fidelities This can mean that alternative approaches than the Powell minimisation used so far, \@e.g.~that of CRAB, could be better suited to probing the cost function for high $N_k$. We also note that this lack of improvement in the results is likely the consequence of the form of the control field given by Eq.~\eqref{eq:optimsable_1} rather than due to a failure of the optimiser in the face of a complex parameter space. We find that the parameter space is relatively smooth in the case of $N_k = 1,2,3$ and a better solution for this form of control field simply does not exist.
We now consider the combined method of COLD-CRAB for this annealing example as shown in Fig.~\ref{fig:IsingUnconstrained}(b). We point out that with our application of CRAB in this scenario we are not enforcing $\boldsymbol{\beta}$ to be zero at the start and end of the dynamics, allowing for their to be a tuning of the $z$-field offset. This is consistent between CRAB and COLD-CRAB and therefore does not influence our comparison of the two. First, it is important to note that CRAB alone results in a overall speedup of the dynamics for a high final state fidelity $1-F\sim 10^{-3}$ at long time-scales. However, CRAB still suffers from getting stuck in the initial state at fast driving times and the final state fidelity again tends to zero. This is not the case for COLD-CRAB, which converges to large final state fidelities $1-F \sim 10^{-3}$ at short driving times $\tau \leq 10^{-1}J^{-1}$. Note that the difference between the convergence to final state fidelities are only marginally different between COLD and COLD-CRAB at longer times, but at short times COLD-CRAB performs a lot better. Further improvement could be gained by combining COLD with more advanced versions of CRAB or other optimal control methods.
As shown in Fig.~\ref{fig:MaxAmp}, the amplitude of the driving required to achieve the fidelities discussed so far scales with the driving time. Practical scenarios will necessarily place limits both on achievable driving times and the maximum amplitude of any term that is being driven. However, the scaling of the drivings shown do not mean that everything diverges in the limit of $\tau \rightarrow 0$. To see this we can first write the Scr\"odinger equation for COLD as
\begin{equation}
i \hbar d_t \ket{\psi}=\left(H_\beta+\dot\lambda \mathcal{A}_\lambda\right) \ket{\psi},
\end{equation}
we then divide through by $\dot{\lambda}$ to get
\begin{equation}
i \hbar d_\lambda \ket{\psi}=\left(\frac{H_\beta}{\dot \lambda}+\mathcal{A}_\lambda\right) \ket{\psi},
\end{equation}
in the limit of $\tau \rightarrow 0$ then $\dot\lambda\to \infty$ to result in the Hamiltonian term disappearing, or in other words, we turn off the Hamiltonian. We then only drive the system in the $\tau \to 0$ limit with the COLD or LCD driving term:
\begin{equation} \label{eq:OnlyA}
i \hbar d_\lambda \ket{\psi}=\mathcal{A}_\lambda \ket{\psi}.
\end{equation}
In this limit then $\lambda$ plays the role of time, and this could then be implemented in a practical scenario in finite time as it corresponds to some manipulation of the couplings in the system. This renormalised time cannot then be infinitesimally short if the couplings are bounded but we have shown that the protocol does not diverge as $\tau \rightarrow 0$. In the case of a spin chain, evolving under Eq.~\eqref{eq:OnlyA} is effectively to first order in LCD implementing independent single spin rotations along the chain, and COLD can be easily applied.
If it is not possible to switch off the Hamiltonian as discussed above then as an alternative we can implement COLD with experimental constraints accounted for directly in the optimal control minimisation. We consider an extreme example of constraints to show that even in this scenario COLD can provide an advantage and corresponding speed-up. In the constrained case the annealing protocol remains that of Hamiltonian~\eqref{eq:h0_ising} but we choose to introduce a bound of $X_f$ on the maximum amplitude of all drivings. This makes it so that no optimal control or LCD term can go beyond the original amplitude of the $x$-field drive. We show the final state fidelities achieved for the constrained example in Fig.~\ref{fig:IsingcConstrained}. As can be seen in Fig.~\ref{fig:IsingcConstrained}(a), COLD provides a substantial improvement beyond what is achievable with BPO. BPO manages $F < 0.5$ for $\tau < 1J^{-1}$, but COLD can reach final state fidelities $F\sim 0.9$ for $\tau < 1J^{-1}$. The real improvement, however, comes with the application of CRAB and COLD-CRAB. CRAB already improves the fidelities substantially, and would allow for a speed up in the annealing protocol but with COLD-CRAB the final state fidelities are even better, with $F\sim 0.99$ achievable when approaching $\tau\sim 0.1J^{-1}$. Signs are seen of the onset of the convergence to small values for COLD-CRAB in Fig.~\ref{fig:IsingcConstrained}(b) before the maximum amplitude required becomes too large and the short time results tend towards zero fidelity and the state being stuck again. With this example and the discussion on implementation via turning off the Hamiltonian, we have shown that COLD is capable of delivering improvements beyond other schemes even for practical problems with strict and rather extreme constraints.
\section{Transport in a Synthetic Lattice}\label{sec:lattice}
\begin{figure*}[t]
\includegraphics[width=0.98\textwidth]{synthetic.png}
\caption{Optimisation of state transfer in a synthetic lattice. In (a) we compare the fidelities obtained via the bare ARP protocol (pink dashed line) and first-order LCD previously implemented in Ref.~\cite{meier_counterdiabatic_2020} (purple dash-dot line) to BPO (blue diamonds) and the COLD method (red circles). (c) Maximum amplitude of the tunneling term at each driving time for LCD (green diamonds) as given by Eq.~\eqref{eq:tunneling} as well as COLD (red triangles) which includes additional control parameters as shown in Eq.~\eqref{eq:tunneling_opt} and BPO (blue triangles) which omits the modifications due to CD but retains the control terms $\boldsymbol{\beta}$. In both (a) and (c) we simulate $N = 7$ lattice sites and use $N_k = 1$ parameter for optimisation of BPO and COLD. (b) Scaling of fidelities with increasing number of lattice sites (where $N_k = 1$) for both COLD (red circles) and BPO (blue diamonds) noting that the latter performs very poorly for $N > 9$. (d) does the same for the number of parameters while keeping $N=7$, with the trend indicating that increasing $N_k$ does not lead to better fidelities in either the BPO or COLD case. Note that both (b) and (d) are simulated for driving time $\tau = 0.5 J^{-1}$ and the best fidelities are obtained across 500 optimisations.}\label{fig:Synthetic}
\end{figure*}
The efficient transfer of states between opposite ends of a lattice could have future applications in the settings of quantum computation and simulation due to its promise of efficient transport of information \cite{lang2017}. This objective is often tackled in the setting of ultracold atoms in optical lattices. While the problem can be tuned to be a single-particle system and the analytical solutions of the corresponding instantaneous Schr\"odinger equation are known \cite{Hatsugai1993,Hugel2014} even for a finite system \cite{Duncan2018exact}, efficient evolution for state transfer is not straight-forward. This is due to the fact that the majority of the states are delocalised across the lattice, meaning that the $\ket{\psi}\bra{\psi}$ terms of the CD Hamiltonian of Eq.~\eqref{eq:counterdiabatic} are global in reach. It is normal to consider this system in the tight-binding limit where the implementation of global terms is not straightforward. Such terms can be generated via the interactions of the atoms with cavity modes \cite{landig2016,keller2017} or from dipolar interactions \cite{baranov2002,menotti2008,trefzger2011}. However, it would be ambitious to expect this control to be general enough to implement the CD Hamiltonian of the exact solutions. This is one of the reasons that LCD has been pursued in this setting.
Recently, LCD has been successfully applied to improve an adiabatic rapid passage (ARP) protocol for population transfer across a synthetic lattice \cite{meier_counterdiabatic_2020}. In this realisation, population transfer was achieved in a synthetic tight-binding lattice of laser coupled atomic momentum states. We will consider the same problem as in Ref.~\cite{meier_counterdiabatic_2020} but with the improvement that can be gained by COLD. This system is described by the Hamiltonian
\begin{align}\label{eq:Hlattice}
\begin{split}
H_0(t) &= - \sum_n J_n(t)(c_n^{\dag}c_{n+1} + H.c.) \\ &+ \sum_n V_n(t) c_n^{\dag}c_n,
\end{split}
\end{align}
where $J_n(t)$ is the time-dependent tunnelling that describes the nearest-neighbour coupling, $V_n(t)$ is the on-site energy offset with respect to neighbouring sites and $c_n^{\dag}$($c_{n}$) is the creation(annihilation) operator on a given synthetic lattice site. In the ARP protocol, the population gets moved from one end of the lattice to the other by linearly ramping the lattice from a positive tilt to a negative tilt via
\begin{align}\label{eq:t_and_v}
J_n(t) = J_0(1.1 - \lambda) = J_0\Big(0.1 + \frac{t}{\tau}\Big), \\
V_n(t) = n V_0 2 (\lambda - 1/2) = nV_0\Big(1 - \frac{2t}{\tau}\Big),
\end{align}
where $V_0 = 4J_0$ is the initial site energy slope and $J_0$ is the characteristic tunnelling scale of the lattice. The scaling function in this case is given by
\begin{equation}
\lambda(t) = 1-\frac{t}{\tau}.
\end{equation}
In order to implement LCD as shown in Ref.~\cite{meier_counterdiabatic_2020}, the first order LCD can be accounted for by taking
\begin{align}\label{eq:tunneling}
J_n(t) \rightarrow J_{n, \mathrm{CD}}(t) e^{-i\phi_{n, \mathrm{CD}}(t)},
\end{align}
\noindent where
\begin{align}\label{eq:t_phi_cd}
J_{n, \mathrm{CD}}(t) = \sqrt{J_n(t)^2 + (\alpha_n(t)/\tau)^2}, \\
\phi_{n, \mathrm{CD}}(t) = \arctan\left(-\frac{J_n(t)\tau}{\alpha_n(t)}\right),
\end{align}
\noindent and $\alpha_n(t)$ is the CD terms which can be found by solving a set of linear equations
\begin{align}
\begin{split}
&-3(J_n J_{n+1})\alpha_{n+1} + (J_{n-1}^2 + 4J^2_n + J_{n+1}^2)\alpha_j \\ &- 3(J_n J_{n-1})\alpha_{n-1} + (V_{n+1} - V_n)^2 \alpha_n \\ &= -\partial_{\lambda}J_n (V_{n+1} - V_{n}).
\end{split}
\end{align}
In order to implement COLD we include additional terms to the tunnelling of the lattice
\begin{align}\label{eq:tunneling_opt}
J_n(t) \rightarrow J_n(t, \boldsymbol{\beta}) = J_n(t) + \boldsymbol{\beta}(t),
\end{align}
which can then be incorporated into the forms of both $J_{n, CD}(t)$ and $\phi_{n, CD}(t)$. We again want the additional control terms to go to zero around the problem boundaries and a natural choice is the same as in the Ising spin chain example in Eq.~\eqref{eq:optimsable_1}. The parameters $\boldsymbol{\beta}$ are optimised as before by minimizing with respect to the fidelity of the final state, where the population has been fully transferred to the opposite lattice site.
We first consider a system size of $N=7$ sites which was successfully experimentally probed in Ref.~\cite{meier_counterdiabatic_2020}, where final state fidelities of $0.75$ were achieved for $\tau = 1$ms with a final tunnelling strength of $J/\hbar = 1/2\pi kHz$ (equivalent to $\tau \sim 1 J^{-1}$ in our units). We initially confirm the breakdown of ARP in this setting for fast times, and the success of the LCD protocol at short times, as shown in Fig.~\ref{fig:Synthetic} (a) and found in Ref.~\cite{meier_counterdiabatic_2020}. Implementing BPO on its own manages to enhance the achievable fidelities at intermediate times of $\tau > 0.03 J^{-1}$. However, eventually, as observed in all scenarios in this work, BPO becomes stuck in the initial state at fast times, and the fidelity goes to zero. Implementing the newly introduced COLD protocol achieves an order of magnitude improvement in the fidelity over LCD. This is also plotted in Fig.~\ref{fig:Synthetic}(a) alongside previous results of ARP and first-order LCD.
One concern could be that COLD is achieving this improvement by simply pumping power into the tunnelling term, but as we can see in Fig.~\ref{fig:Synthetic}(c) the maximum amplitude of the tunnelling term tracks that of LCD. A key issue for experiments is the maximum amplitude achievable by a driving term and with this result we can stipulate that COLD is likely to be feasible in the same regimes as LCD in this synthetic lattice system. There is single outlier at intermediate times as indicated by the single point peaking in maximum amplitude in Fig.~\ref{fig:Synthetic}(c), this is the exception to the rule, where the optimisation has found a marginally higher fidelity (see the offset point in Fig.~\ref{fig:Synthetic}(a)) by pumping in power.
A large concern for state transfer techniques is the robustness of a protocol with respect to an increasing system size. We show the best achievable fidelities with increasing system size for both BPO and COLD in Fig.~\ref{fig:Synthetic}(c). While both protocols show a decreasing fidelity with system size as is to be expected, once again COLD does not suffer from getting stuck in the initial state. This is shown by the BPO fidelities going to unity for large systems in Fig.~\ref{fig:Synthetic}(c), and is the same mechanism for this as for the short driving times in Fig.~\ref{fig:Synthetic}(a).
Another concern could be that BPO will beat COLD if enough parameters are allowed for the optimisation, i.e. if we increase $N_k$ enough. We observed no evidence of this for the Ising model example and we again do not observe this in this synthetic lattice example, as is shown in Fig.~\ref{fig:Synthetic}(d). Small improvements are made in the fidelities achieved with BPO and COLD for larger $N_k$ but this is not substantial.
\section{Discussion and outlook} \label{sec:conclusions}
We have introduced a new hybrid approach combining quantum optimal control and shortcuts to adiabaticity: COLD. Inspired by the successes of LCD, where diabatic transitions are suppressed and locality conditions can be met, COLD improves on its methodology by combining it with quantum optimal control. The natural way to enhance the performance of LCD is by introducing higher order CD terms, but these are often non-local and difficult to engineer in experiments. COLD circumvents this by allowing for additional control fields that extend the family of dynamical Hamiltonians which can be explored. In this way, our method may find the best possible path where the effect of lower-order LCD is most relevant and higher order corrections are suppressed.
COLD has a clear potential in efficiently speeding up adiabatic evolution in various settings. We demonstrate this numerically via several example protocols which indicate improvements beyond a classical optimisation approach BPO as well as LCD of different orders. Our work shows that COLD reduces the strength of higher order LCD corrections, and that it performs well for increasing system sizes. We have shown that COLD can be implemented in the limit of fast driving by a `switching off' of the original dynamical Hamiltonian. For scenarios where removing the Hamiltonian is not possible, we have shown that an alternative way to implement COLD is to use a bounded optimisation where amplitudes are restricted. We find that both the COLD and COLD-CRAB protocols perform extremely well in this setting.
COLD will be most beneficial when the LCD is only realisable to a certain order but the higher order corrections are large. This means the diabatic transitions are not being sufficiently suppressed by the choice of LCD and COLD can be used to find the dynamical Hamiltonian for which the required order of LCD term dominates. Note, that this goes the other way too, with COLD not providing substantial improvements when the chosen lower order LCD is small across the path. This can be thought of as being the case in two limits. First is the adiabatic limit, for which any CD correction is small and COLD will tend towards the adiabatic result. Second, the low-order LCD terms can be small compared to the driving as the exact CD would be correcting transitions due to interactions at longer ranges. In this scenario, the order of LCD being implemented with COLD needs to be increased, so that the CD term is accounting for the longer range terms. While we have only utilised COLD with first-order LCD in this work, it can be applied with higher-order LCD terms. We found this to be the case when we tried to apply COLD to the protocols for \@e.g.~Greenberger-Horne-Zeilinger state generation in arrays of Rydberg atoms in Ref.~\cite{omran_generation_2019}. COLD with first- or second-order LCD could not improve upon the results of BPO or CRAB. This was due to both the first- and second-order corrections being smaller than the dynamical terms of the original Hamiltonian by two orders of magnitude. It is perhaps not surprising when we consider that the Hamiltonian of this protocol includes long-range couplings. In this case, higher-order corrections would need to be implemented with COLD, and finding methods for executing these now highly non-local terms may be a future path of research.
COLD may also be applied to more complex systems where exact dynamics are not possible, \@e.g. due to an excessively large Hilbert space. This may be achieved by variationally minimising the integrals of the coefficients for the higher order corrections to the LCD. This was briefly sketched out in the example of two spin annealing, and we have shown that COLD functions by minimising the strength of the higher order LCD corrections along a given path. A further option is to combine COLD with one of a large variety of numerical optimal control methods, as we have done for the example of CRAB. We have shown a substantial improvement for state preparation in the Ising model that can be obtained from the COLD-CRAB combination - particularly in the constrained case. Both fusions of COLD with advanced optimal control methods as well as the minimisation of higher order LCD term amplitudes for complex systems are two potential extensions which could prove fruitful with further study.
\begin{acknowledgements}
Work at the University of Strathclyde was supported by the EPSRC Quantum Technologies Hub for Quantum Computing and Simulation (EP/T001062/1), and the European Union’s Horizon 2020 research and innovation program under grant agreement No. 817482 PASQuanS. A.P. acknowledges support from NSF under Grant DMR-2103658 and by the AFOSR under Grants No. FA9550-16-1-0334 and FA9550-21-1-0342.
\end{acknowledgements}
| 30a0701b2d7f2d3f82acdac04bc19081a3cb2f11 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Deep Generative Networks (DGNs) have emerged as the go-to framework for generative modeling of high-dimensional dataset, such as natural images. Within the realm of DGNs, different frameworks can be used to produce an approximation of the data distribution e.g. Generative Adversarial Networks (GANs) \cite{goodfellow2020generative}, Variational AutoEncoders (VAEs) \cite{kingma2013auto} or flow-based models \cite{prenger2019waveglow}. But despite the different training settings and losses that each of these frameworks aim to minimize, the evaluation metric of choice that is used to characterize the overall quality of generation is the {\em Fr\'echet Inception Distance} (FID) \cite{heusel2017gans}. The FID is obtained by taking the Fr\'echet Distance in the InceptionV3 \cite{szegedy2016rethinking} embedding space between two distributions; the distributions are usually taken to be the training dataset and samples from a DGN trained on the dataset. It has been established in prior work \cite{sajjadi2018assessing} that FID non-linearly combines measures of quality and diversity of the samples, which has inspired further research into disentanglement of these quantities as {\em precision} and {\em recall} \cite{sajjadi2018assessing, kynkaanniemi2019improved} metrics respectively.
Recent state-of-the-art DGNs such as BigGAN \cite{brock2018large}, StyleGAN2/3 \cite{karras2020analyzing,karras2021alias}, and NVAE \cite{vahdat2020nvae} have reached FIDs nearly as low as one could get when comparing subsets of real data with themselves. This has led to the deployment of DGNs in a variety of applications, such as real-world high-quality content generation or data-augmentation. However, it is clear that {\bf depending on the domain of application, generating samples from the best FID model could be suboptimal}. For example, realistic content generation might benefit more from high-quality (precision) samples, while data-augmentation might benefit more from samples of high-diversity (recall), even if in each case, the overall FID slightly diminishes \cite{giacomello2018doom,islam2021crash}. Therefore, a number of state-of-the-art DGNs have introduced a controllable parameter to trade-off between the precision and recall of the generated samples, e.g., truncated latent space sampling \cite{brock2018large}, interpolating truncation \cite{karras2019style, karras2020analyzing}. However, those methods do not always work ``out-of-the-box'' \cite{brock2018large} as they require orthogonal regularization of the DGN's parameters during training; those methods also lack a clear theoretical understanding which can limit their deployment for sensitive applications.
{\bf In this paper, we propose a principled solution to control the quality (precision) and diversity (recall) of DGN samples that does not require retraining nor specific conditioning of model training.}
Our proposed method, termed {\em Polarity Sampling}, is based on the analytical form of the learned DGN sample distribution and introduces a new hyperparameter, that we dub the {\em polarity} $\rho \in \mathbb{R}$, that adapts the latent space distribution for post-training control.
{\bf The polarity parameter provably forces the latent distribution to concentrate on the modes of the DGN distribution, i.e., regions of high probability ($\rho<0$), or on the anti-modes, i.e., regions of low-probability ($\rho>0$); with $\rho=0$ recovering the original DGN distribution.}
The Polarity Sampling process depends only on
the top singular values of the DGN's output Jacobian matrices evaluated at each input sample and can be implemented to perform online sampling. A crucial benefit of Polarity Sampling lies in its theoretical derivation from the analytical DGN data distribution, where the product of the DGN Jacobian matrices singular values -- raised to the power $\rho$ -- provably controls the DGN samples distribution as desired.
See Fig.~\ref{fig:teaser} for an initial example of Polarity Sampling in action.
Our main contributions are as follows.
\noindent[C1]
We first provide the theoretical derivation of Polarity Sampling based on the singular values of the generator Jacobian matrix. We provide pseudocode for Polarity Sampling and an approximation scheme to control its computational complexity as desired (\cref{sec:method}).
\noindent[C2]
We demonstrate on a range of DGNs and and datasets that polarity sampling not only allows one to move on the precision-recall Pareto frontier (\cref{sec:pareto}), i.e., it controls the quality and diversity efficiently, but it also reaches improved FID scores for each model (\cref{sec:sota}).
\noindent[C3]
We leverage the fact that negative polarity sampling provides access to the modes of the learned DGN distribution, enabling us to explore several timely and important questions regarding DGNs. We provide visualization of the modes of trained GANs and VAEs (\cref{sec:mode}) and assess whether latent space truncation brings samples closer to the DGN modes, using the relation between DGN modes and perceptual-path-length (\cref{sec:PPL}).
\\
\section{Related Work}
\label{sec:related}
\noindent{\bf Deep Generative Networks as Piecewise-Linear Mappings.}~In most DGN settings, once training has been completed, sampling new data points is done by first sampling latent space samples ${\bm{z}}_i \in \mathbb{R}^{K}$ from a latent space distribution ${\bm{z}}_i\sim p_{{\bm{z}}}$ and then processing those samples through a DGN $G:\mathbb{R}^{K} \mapsto \mathbb{R}^{D}$ to obtain the sample ${\bm{x}}_i\triangleq G({\bm{z}}_i),\forall i$. One recent line of research that we will rely on through our study consists in formulating DGNs as Continuous Piecewise Affine (CPA) mappings \cite{montufar2014number,balestriero2018spline}, that be expressed as
\begin{align}
G({\bm{z}}) = \sum_{\omega \in \Omega}({\bm{A}}_{\omega}{\bm{z}}+{\bm{b}}_{\omega})1_{\{{\bm{z}} \in \omega\}},\label{eq:CPA}
\end{align}
where $\Omega$ is the input space partition induced by the DGN architecture, $\omega$ is a partition-region where $z$ resides, and ${\bm{A}}_{\omega}, {\bm{b}}_{\omega}$ are the corresponding slope and offset parameters. The CPA formulation of \cref{eq:CPA} either represents the exact DGN mapping, when the nonlinearities are CPA e.g. (leaky-)ReLU, max-pooling, or represents a first-order approximation of the DGN mapping. For more background on CPA networks, see \cite{balestriero2020mad}. {\em The key result from \cite{daubi} that we will leverage is that \cref{eq:CPA} is either exact, or can be made close enough to the true mapping $G$, to be considered exact for practical purposes}.
\noindent{\bf Post-Training Improvement of a DGN's Latent Distribution.}
The idea that the training-time latent distribution $p_{{\bm{z}}}$ might be sub-optimal for test-time evaluation has led to multiple research directions to improve quality of samples post-training. \cite{tanaka2019discriminator,che2020your} proposed to optimize the samples ${\bm{z}} \sim p_{{\bm{z}}}$ based on a Wasserstein discriminator, leading to the {\em Discriminator Optimal Transport} (DOT) method. That is, after sampling a latent vector ${\bm{z}}$, the latter is repeatedly updated such that the produced datum has greater quality. \cite{tanielian2020learning} proposes to simply remove the samples that produce data out of the true manifold. This can be viewed as a binary rejection decision of any new sample ${\bm{z}} \sim p_{{\bm{z}}}$. \cite{azadi2018discriminator} were the first to formally introduce rejection sampling based on a discriminator providing a quality estimate used for the rejection sampling of candidate vectors ${\bm{z}} \sim p_{{\bm{z}}}$. Replacing rejection sampling with the Metropolis-Hasting algorithm \cite{hastings1970monte} led to the method of
\cite{turner2019metropolis}, coined MH-GAN. An improvement made by \cite{grover2019bias} was to use the {\em Sampling-Importance-Resampling} (SIR) algorithm \cite{rubin1988using}. \cite{issenhuth2021latent} proposes {\em latentRS} which consists in training a WGAN-GP \cite{gulrajani2017improved} on top of any given DGN to learn an improved latent space distribution producing higher-quality samples. \cite{issenhuth2021latent} also proposes {\em latentRS+GA}, where the generated samples from that learned distribution are further improved through gradient ascent.
\noindent{\bf Truncation of the Latent Distribution.}
Latent space truncation was introduced for high-resolution face image generation by \cite{marchesi2017megapixel} as a method of removing generated artifacts. The authors employed a latent prior of ${\bm{z}} \sim \mathcal{U}[-1,1]$ during training and ${\bm{z}} \sim \mathcal{U}[-.5,.5]$ for qualitative improvement during evaluation. The ``truncation trick'' was formally introduced by \cite{brock2018large} where the authors propose resampling latents $z$ if they exceed a specified threshold as truncation. The authors also use weight orthogonalization during training to make truncation amenable. It is also speculated in \cite{brock2018large} that reducing the truncation parameter $\psi$, makes the samples converge to the DGN distribution modes, we will argue against that in \cref{sec:truncation}. Style-based architectures \cite{karras2019style,karras2020analyzing} introduce a linear interpolation based truncation in the style-space, which is also designed to converge to the average of the dataset \cite{karras2019style}. Ablations for truncation for style-based generators are provided in \cite{kynkaanniemi2019improved}.
\section{Introducing The Polarity Parameter From First Principles}
\label{sec:method}
In this section, we introduce \textit{Polarity Sampling} method that enables us to control generation quality and diversity of DGNs. This is done by first expressing the analytical form of DGNs' output distribution, and parametrizing the latent space distribution by the singular values of its Jacobian matrix and our {\em polarity} parameter (\cref{sec:theory}). We provide pseudo-code and an approximation strategy to allow for fast sampling (\cref{sec:pseudocode}).
\subsection{Analytical Output-Space Density Distribution}
\label{sec:theory}
Given a DGN $G$, samples are obtained by sampling $G({\bm{z}})$ with a given latent space distribution, as in ${\bm{z}} \sim p_{{\bm{z}}}$. This produces samples that will lie on the {\em image} of $G$, the distribution of which is subject to $p_{{\bm{z}}}$, the DGN latent space partition $\Omega$ and per-region affine parameters ${\bm{A}}_{\omega},{\bm{b}}_{\omega}$. We denote the DGN output space distribution as $p_{G}$. Under an injective DGN mapping assumption ($g(z)=g(z') \implies z=z'$) (which holds for various architectures, see, e.g., \cite{puthawala2020globally}) it is possible to obtain the analytical form of the DGN output distribution $p_{G}$ \cite{humayun2021magnet}. For a reason that will become clear in the next section, we only focus here on the case ${\bm{z}} \sim U(\mathcal{D})$ i.e., using a Uniform latent space distribution over the domain $\mathcal{D}$. Leveraging the Moore-Penrose pseudo inverse \cite{trefethen1997numerical} ${\bm{A}}^{\dag}\triangleq ({\bm{A}}^T{\bm{A}})^{-1}{\bm{A}}^T$, we obtain the following.
\begin{thm}
\label{thm:general_density}
For ${\bm{z}} \sim U(\mathcal{D})$, the probability density $p_{G}({\bm{x}})$ is given by
\begin{equation}
p_{G}({\bm{x}}) \propto \sum_{\omega \in \Omega} \det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{-\frac{1}{2}}\mathds{1}_{\{{\bm{A}}^{\dag}_{\omega}({\bm{x}}-{\bm{b}}_{\omega}) \in \omega \cap \mathcal{D}\}},
\end{equation}
where $\det$ is the pseudo-determinant, i.e., the product of the nonzero eigenvalues of ${\bm{A}}_{\omega}^T{\bm{A}}_{\omega}$. (Proof in Appendix~\ref{proof:general_density}.)
\end{thm}
Note that one can also view $\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{1/2}$ as the product of the nonzero singular values of ${\bm{A}}_{\omega}$. Theorem \ref{thm:general_density} is crucial to our development since it demonstrates that the probability of a sample ${\bm{x}} = g({\bm{z}})$ is proportional to the change in volume $(\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{1/2})$ produced by the coordinate system ${\bm{A}}_{\omega}$ of the region $\omega$ in which ${\bm{z}}$ lies in (recall \cref{eq:CPA}). If
a region $\omega \in \Omega$ has a slope matrix ${\bm{A}}_{\omega}$ that contracts the space ($\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})<1$) then the output density on that region --- mapped to the output space region $\{{\bm{A}}_{\omega}{\bm{u}}+{\bm{b}}_{\omega}:{\bm{u}} \in \omega\}$ --- is increased, as opposed to other regions that either do not contract the space as much, or even expand it ($\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})>1$). Hence, the concentration of samples in each output space region depends on how that region's slope matrix contracts or expands the space, relative to all other regions.
\subsection{Controlling the Density Concentration with a Single Parameter}
Given \cref{thm:general_density}, we directly obtain an explicit parametrization of $p_{{\bm{z}}}$ that enables us to control the distribution of samples in the output space, i.e., to control $p_{G}$. In fact, note that one can sample from the mode of the DGN distribution by employing ${\bm{z}} \sim U(\omega^*), \omega^* = \argmin_{\omega \in \Omega}\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})$. Alternatively, one can sample from the region of lowest probability, i.e., the anti-mode, by employing ${\bm{z}} \sim U(\omega^*), \omega^* = \argmax_{\omega \in \Omega}\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})$.
This directly leads to our Polarity Sampling formulation that adapts the latent space distribution based on the per-region pseudo-determinants.
\begin{cor}
\label{cor:pol_density}
The latent space distribution
\begin{equation}
p_{\rho}({\bm{z}}) \propto \sum_{\omega \in \Omega}\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{\frac{\rho}{2}} \mathds{1}_{\{{\bm{z}} \in \omega\}},\label{eq:latent_distribution}
\end{equation}
where ${\rho}\in\mathbb{R}$ is the polarity parameter, produces the DGN output distribution
\begin{equation}
p_{G}({\bm{x}})\propto \sum_{\omega \in \Omega} \det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{\frac{{\rho}-1}{2}}\mathds{1}_{\{{\bm{A}}^{\dag}_{\omega}({\bm{x}}-{\bm{b}}_{\omega}) \in \omega \cap \mathcal{D}\}},\label{eq:density_rho}
\end{equation}
which falls back to the standard DGN distribution for ${\rho}=0$, to sampling of the mode(s)
for ${\rho} \to -\infty$ and to sampling of the the anti-mode(s) for ${\rho} \to \infty$. (Proof in Appendix~\ref{proof:pol_density}.)
\end{cor}
Polarity Sampling consists of using the latent space distribution \cref{eq:latent_distribution} with a polarity parameter $\rho$, that is either negative, concentrating the samples toward the mode(s) of the DGN distribution $p_{G}$, positive, concentrating the samples towards the anti-modes(s) of the DGN distribution $p_{G}$ or zero, which removes the effect of polarity. Intuitively, $\rho<0$ will produce samples from ``higher-confidence'' regions of the data manifold, and $\rho>0$ will produce samples from ``lower-confidence'' regions. Note that Polarity Sampling provides a method to change the output density in a continuous fashion. Its practical effect, as we will see in Sec.~\ref{sec:pareto}, is to control the quality and diversity of the obtained samples.
\subsection{Approximation and Implementation}
\label{sec:pseudocode}
We now provide the details and pseudo-code for the Polarity Sampling procedure that implements \cref{cor:pol_density}.
\noindent{\bf Computing the ${\bm{A}}_{\omega}$ Matrix.}~The per-region slope matrix as in \cref{eq:CPA}, can be obtained given any DGN by first sampling a latent vector ${\bm{z}} \in \omega$, and then obtaining the Jacobian matrix of the DGN ${\bm{A}}_{\omega} = {\bm{J}}_{{\bm{z}}}G({\bm{z}}),\forall {\bm{z}} \in \omega$. This has the benefit of directly employing automatic differentiation libraries and thus does not require any exhaustive implementation nor derivation. Computing ${\bm{J}}_{{\bm{z}}}G({\bm{z}})$ of a generator is not uncommon in practice, e.g., it is employed during path length regularization of StyleGAN2 \cite{karras2020analyzing}.
\noindent{\bf Discovering the Regions $\omega \in \Omega$.}~As per \cref{eq:latent_distribution}, we need to obtain the singular values of ${\bm{A}}_{\omega}$ (see next paragraph) for each region $\omega \in \Omega$. This is often a complicated task, especially for state-of-the-art DGNs that can have a partition $\Omega$ made of a number of regions growing with the architecture depth and width \cite{montufar2021sharp}.
Furthermore, checking if ${\bm{z}} \in \omega$ requires one to solve a linear program \cite{fischetti2017deep}, which is expensive. As a result, we develop an approximation which consists of sampling many ${\bm{z}} \sim U(\mathcal{D})$ vectors from the latent space (hence our uniform prior assumption in \cref{cor:pol_density}), and computing their corresponding matrices ${\bm{A}}_{\omega({\bm{z}})}$. This way, we are guaranteed that ${\bm{A}}_{\omega({\bm{z}})}$ corresponds to the slope of the region $\omega$ in which ${\bm{z}}$ falls in, removing the need to check whether ${\bm{z}} \in \omega$. We do so over $N$ samples obtained uniformly from the DGN latent space (based on the original latent space domain). Selection of $N$ can impact performance as this exploration needs to discover as many regions from $\Omega$ as possible.
\noindent{\bf Singular Value Computation.}~Computing the singular values of ${\bm{A}}_{\omega}$ is an $\mathcal{O}(\min(K,D)^3)$ operation \cite{golub1971singular}. However, not all singular values might be relevant, e.g., the smallest singular values that are nearly constant across regions $\omega$ can be omitted without altering \cref{cor:pol_density}. Hence, we employ only the top-$k$ singular values of ${\bm{A}}_{\omega}$ to speed up singular value computation to $\mathcal{O}(k^3)$. (Further approximation could be employed if needed, e.g., power iteration \cite{packard1988power}).
While the required number of latent space samples $N$ and the number of top singular values $k$ might seem to be a limitation of Polarity Sampling, we have found in practice that $N$ and $k$ for state-of-the-art DGNs can be set at $N \approx 200K$, $k \in [30,100]$. We conduct a careful ablation study and demonstrate the impact of different choices for $N$ and $k$ in \cref{fig:topk_dets,tab:ablation_N,tab:ablation_topk} in \cref{sec:effect_N_k}. Computation times and software/hardware details are provided in \cref{sec:times}. To reduce round-off errors that can occur for extreme values of $\rho$, we compute the product of singular values in log-space, as shown in \cref{alg:sampling}.
\begin{algorithm}[t!]
\caption{
Polarity Sampling procedure with polarity $\rho$; the online version is given in \cref{alg:online} in \cref{sec:online_algorithm}. For implementation details, see \cref{sec:pseudocode}.}\label{alg:sampling}
\begin{algorithmic}
\Require $K>0,S>0,N \gg S,G,\mathcal{D},\rho \in \mathbb{R}$
\State ${\mathcal{Z}},{\mathcal{S}},{\mathcal{R}} \gets [],[],[]$
\For{$n=1,\dots,N$}
\State ${\bm{z}} \sim U(\mathcal{D})$
\State $\sigma = {\rm SingularValues}({\bm{J}}_{{\bm{z}}}G({\bm{z}}),{\rm decreasing=True})$
\State ${\mathcal{Z}}.{\rm append}({\bm{z}})$
\State ${\mathcal{S}}.{\rm append}(\rho \sum_{k=1}^{K}\log(\sigma[k]+\epsilon))$
\EndFor
\For{$n=1,\dots,S$}
\State $i \sim {\rm Categorical}({\rm prob}={\rm softmax}({\mathcal{S}}))$
\State ${\mathcal{R}}.{\rm append}({\mathcal{Z}}[i])$
\EndFor
\Ensure ${\mathcal{R}}$
\end{algorithmic}
\end{algorithm}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.02\linewidth}
\centering
\rotatebox{90}{\footnotesize \hspace{0.7cm}FID}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_incpol_topksweep.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_incpol_ndetz.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{ Effect of Polarity Sampling on FID of a StyleGAN2-F model pretrained on FFHQ for varying number of top-$k$ singular values ({\bf left}) and varying number of latent space samples $N$ used to obtain per-region slope matrix ${\bm{A}}_{\omega}$ singular values ({\bf right}) (recall \cref{sec:pseudocode,alg:sampling}). The trend in FIDs to evaluate the impact of $\rho$ stabilizes when using around $k=40$ singular values and $N\approx$200,000 latent space samples. For the effect of $k$ and $N$ on precision and recall, see \cref{fig:topk_prec_recall}.
}
\label{fig:topk_dets}
\end{figure}
We summarize how to obtain $S$ samples using the above steps in the pseudo-code given in \cref{alg:sampling} and provide an efficient solution to reduce the memory requirement incurred when computing the large matrix ${\bm{A}}_{\omega}$ in \cref{sec:memory}. We also provide an implementation that enables online sampling in \cref{alg:online} (\cref{sec:online_algorithm}). It is also possible to control the DGN prior $p_{{\bm{z}}}$ with respect to a different space than the data-space e.g. inception-space, or with a different input space than the latent-space e.g. style-space in StyleGAN2/3. This incurs no changes in \cref{alg:sampling} except that the DGN is now considered to either be a subset of the original one, or to be composed with a VGG/InceptionV3 network. We provide implementation details for style-space, VGG-space, and Inception-space in \cref{sec:other_space}. In fact, in those cases, the partition $\Omega$ and the per-region mapping parameters ${\bm{A}}_{\omega},{\bm{b}}_{\omega}$ are the ones of the corresponding sub-network or composition of networks (recall \cref{eq:CPA}). Polarity Sampling adapts the DGN prior distribution to get the modes or anti-modes with respect to the considered output spaces.
\section{Controlling Precision, Recall, and FID via Polarity}
\label{sec:experiments}
We now provide empirical validation of Polarity Sampling with an extensive array of experiments.
Since calculation of distribution metrics such as FID, precision, and recall are sensitive to image processing nuances, we use each model's original code repository except for BigGAN-deep on ImageNet \cite{deng2009imagenet}, for which we use the evaluation pipeline specified for
ADM \cite{dhariwal2021diffusion}.
For NVAE (trained on colored-MNIST \cite{arjovsky2019invariant}), we use a modified version of the StyleGAN3 evaluation pipeline. Precision and recall metrics are all based on the implementation of \cite{kynkaanniemi2019improved}. Metrics in \cref{tab:main_result} are calculated for 50K training samples to be able to compare with existing latent reweighing methods. For all other results, the metrics are calculated using $\min\{N_{D},100K\}$ training samples, where $N_{D}$ is the number of samples in the dataset.
\subsection{Polarity Efficiently Parametrizes the Precision-Recall Pareto Frontier}
\label{sec:pareto}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.02\linewidth}
\centering
\rotatebox{90}{\small
\hspace{1cm}Precision\hspace{2.7cm}Precision\hspace{2.7cm}Precision}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{pareto/pareto_lsun_car_stg2f_vggpol.pdf}\\
\includegraphics[width=\linewidth]{sec/pareto_afhq_stg3_pixelpol.pdf}\\
\includegraphics[width=\linewidth]{pareto/pareto_lsun_church_stg2f_vggpol.pdf}\\[-0.4em]
{\small Recall}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{pareto/pareto_ffhq_stg3_pixelpol.pdf}\\
\includegraphics[width=\linewidth]{pareto/pareto_lsun_cats_stg2f_stypol.pdf}\\
\includegraphics[width=\linewidth]{pareto/pareto_ffhq_stg2f_vggpol.pdf}\\[-0.4em]
{\small Recall}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{Pareto frontier of the precision-recall metrics can be obtained solely by varying the polarity parameter, for any given truncation level. We depict here six different models and datasets. Results for additional models and datasets are provided in \cref{fig:teaser} and \cref{fig:more_pareto}.
}
\label{fig:pareto}
\end{figure}
\begin{figure*}[t!]
\centering
\begin{minipage}{0.02\linewidth}
\hfill
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering
{\scriptsize (modes\reflectbox{$\to$})}\\
$\rho=-2$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=-1$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=-0.5$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=-0.2$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering
{\scriptsize (baseline)}\\
$\rho=0$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=0.2$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=0.5$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering\vspace{0.28cm}
$\rho=1$
\end{minipage}
\begin{minipage}{0.10\linewidth}
\centering
{\scriptsize ($\to$ anti-modes)}\\
$\rho=2$
\end{minipage}\\
\begin{minipage}{0.02\linewidth}
\rotatebox[origin=c]{90}{LSUN Cars}
\end{minipage}
\begin{minipage}{0.97\textwidth}
\includegraphics[width=\linewidth]{alpha_sweep/cars_alpha_stylespace_sweep.png}
\end{minipage}\\
\begin{minipage}{0.02\linewidth}
\rotatebox[origin=c]{90}{LSUN Cats}
\end{minipage}
\begin{minipage}{0.97\textwidth}
\includegraphics[width=\linewidth]{alpha_sweep/cats_alpha_stylespace_sweep_blurred.png}
\end{minipage}\\
\begin{minipage}{0.02\linewidth}
\rotatebox[origin=c]{90}{LSUN Church}
\end{minipage}
\begin{minipage}{0.97\textwidth}
\includegraphics[width=\linewidth]{alpha_sweep/church_alpha_sweep_pixel_blurred.png}
\end{minipage}
\vspace{-0.2cm}
\caption{Curated samples of cars and cats for Polarity Sampling in style-space, and church for Polarity Sampling in pixel-space. Full batch submitted with the supplementary materials.
(Qualitative comparison with truncation sweep in \cref{fig:truncation_sweep_images} and nearest training samples in \cref{fig:nearest} in the Appendix.) None of the images correspond to training samples, as we discuss in \cref{sec:mode}.
}
\label{fig:qualitative_alpha_sweep}
\end{figure*}
As we have discussed above, Polarity Sampling can explicitly sample from the modes or anti-modes of any learned DGN distribution. Since the DGN is trained to fit the training distribution, sampling from the modes and anti-modes correspond to sampling from regions of the data manifold that are approximated better/worse by the DGN
. Therefore Polarity Sampling is an efficient parameterization of the trade-off between precision and recall of generation \cite{kynkaanniemi2019improved} since regions with higher precision are regions where the manifold approximation is more accurate.
As experimental proof, we provide in \cref{fig:pareto} the precision-recall trade-off when sweeping polarity, and compare it with truncation \cite{karras2019style} for pretrained StyleGAN\{2,3\} architectures. We see that Polarity Sampling offers a competitive alternative to truncation for controlling the precision-recall trade-off of DGNs across datasets and models.
For any given precision, the $\rho$ parameter allows us to reach greater recall than what is possible via latent space truncation \cite{karras2019style}. And conversely, for any given recall, it is possible to reach a higher precision than what can be attained using latent space truncation. We see that diversity collapses rapidly for latent truncation compared to Polarity Sampling, across all architectures, which is a major limitation. In addition to that, controlling both truncation and polarity allows us to further extend the Pareto frontier for all of our experiments.
Apart from the results presented here, we also see that polarity can be used to effectively control the precision-recall trade-off for BigGAN-deep \cite{brock2018large} and ProGAN \cite{karras2017progressive}. ProGAN unlike BigGAN and StyleGAN, is not compatible with truncation based methods, i.e., latent space truncation has negligible effect on precision-recall. Hence polarity offers a great benefit over those existing solutions: Polarity Sampling can be applied regardless of training or controllability factors that are preset in the DGN design. We provide additional results in \cref{appendix:extra}.
\subsection{Polarity Improves Any DGN's FID}
\label{sec:sota}
We saw in \cref{sec:pareto} that polarity can be used to control quality versus diversity in a meaningful and controllable manner. In this section, we connect the effect of polarity with FID. Recall that the FID metric nonlinearly combines quality and diversity \cite{sajjadi2018assessing} into a distribution distance measure. Since polarity allows us to control the output distribution of the DGN, an indirect result of polarity is the reduction of FID by matching the inception embedding distribution of the DGN with that of the training set distribution. Recall that $\rho=0$ recovers the baseline DGN sampling; for all the state-of-the-art methods in question, we reach lower (better) FID by using a nonzero polarity. In \cref{tab:prior_art}, we compare Polarity Sampling with state-of-the-art solutions that propose to improve FID by learning novel DGN latent space distributions, as were discussed in \cref{sec:related}. We see that for a StyleGAN2 pre-trained on the LSUN church \cite{yu2015lsun} dataset, by increasing the diversity ($\rho=0.2$) of the Vgg embedding distribution, Polarity Sampling surpasses the FID of methods reported in literature that post-hoc improves quality of generation.
\begin{table}[t!]
\centering
\def\arraystretch{0.6}%
\begin{tabular}{@{}lrrr@{}}
\toprule
\multicolumn{4}{r}{\textbf{LSUN Church 256$\times$256}} \\ \midrule \hline
StyleGAN2 variant & \multicolumn{1}{l}{FID $\downarrow$} & \multicolumn{1}{l}{Prec $\uparrow$} & \multicolumn{1}{l}{Recall $\uparrow$} \\ \midrule
Standard & 6.29 & .60 & .51 \\
SIR$^\dagger$ \cite{rubin1988using} & 7.36 & .61 & \textbf{.58} \\
DOT$^\dagger$ \cite{tanaka2019discriminator} & 6.85 & .67 & .48 \\
latentRS$^\dagger$ \cite{issenhuth2021latent} & 6.31 & .63 & .58 \\
latentRS+GA$^\dagger$ \cite{issenhuth2021latent} & 6.27 & \textbf{.73} & .43 \\
$\rho$-sampling 0.2 & \textbf{6.02} & .57 & .53 \\ \bottomrule
\end{tabular}
\vspace{-0.2cm}
\caption{
Comparison of
Polarity Sampling with latent reweighting techniques from literature. FID, Precision and Recall is calculated using 50,000 samples. $^\dagger$Metrics reported from papers due to unavailability of code. $^\dagger$Precision-recall is calculated with $1024$ samples only.
}
\label{tab:prior_art}
\end{table}
\begin{table*}[t!]
\centering
\setlength\tabcolsep{0.5em}
\begin{tabular}{llrrrllrr}
\hline
Model & FID $\downarrow$ & \multicolumn{1}{l}{Precision $\uparrow$} & \multicolumn{1}{l}{Recall $\uparrow$} & & Model & FID $\downarrow$ & \multicolumn{1}{l}{Precision $\uparrow$} & \multicolumn{1}{l}{Recall $\uparrow$} \\ \hline
\multicolumn{4}{r}{\textbf{LSUN Church 256$\times$256}} & & \multicolumn{4}{r}{\textbf{LSUN Cat 256$\times$256}} \\ \cline{1-4} \cline{6-9}
DDPM$^\dagger$ \cite{ho2020denoising} & 7.86 & - & - & & ADM (dropout)$^\dagger$ & {\bf 5.57} & 0.63 & \textbf{0.52} \\
StyleGAN2 & 3.97 & 0.59 & 0.39 & & StyleGAN2 & 6.49 & 0.62 & 0.32 \\
+ $\rho$-sampling Vgg 0.001 & 3.94 & 0.59 & 0.39 & & + $\rho$-sampling Pix 0.01 & 6.44 & 0.62 & 0.32 \\
+ $\rho$-sampling Pix -0.001 & \textbf{3.92} & \textbf{0.61} & \textbf{0.39} & & + $\rho$-sampling Sty -0.1 & 6.39 & \textbf{0.64} & 0.32 \\ \cline{1-4} \cline{6-9}
\multicolumn{4}{r}{\textbf{LSUN Car 512$\times$384}} & & \multicolumn{4}{r}{\textbf{FFHQ 1024$\times$1024}} \\ \cline{1-4} \cline{6-9}
StyleGAN$^\dagger$ & 3.27 & \textbf{0.70} & 0.44 & & StyleGAN2-E & 3.31 & \textbf{0.71} & 0.45 \\
StyleGAN2 & 2.34 & 0.67 & 0.51 & & Projected GAN$^\dagger$ \cite{sauer2021projected} & 3.08 & 0.65 & 0.46 \\
+ $\rho$-sampling Vgg -0.001 & 2.33 & 0.68 & 0.51 & & StyleGAN3-T & 2.88 & 0.65 & 0.53 \\
+ $\rho$-sampling Sty 0.01 & \textbf{2.27} & 0.68 & \textbf{0.51} & & + $\rho$-sampling Vgg -0.01 & 2.71 & 0.66 & \textbf{0.54} \\
+ $\rho$-sampling Pix 0.01 & 2.31 & 0.68 & 0.50 & & & & & \\
\cline{1-4}
\multicolumn{4}{r}{\textbf{ImageNet 256$\times$256}} & & StyleGAN2-F & 2.74 & 0.68 & 0.49 \\ \cline{1-4}
DCTransformer$^\dagger$ \cite{nash2021generating} & 36.51 & 0.36 & 0.67 & & + $\rho$-sampling Ic3 0.01 & \textbf{2.57} & 0.67 & 0.5 \\
VQ-VAE-2$^\dagger$ \cite{razavi2019generating} & 31.11 & 0.36 & 0.57 & & + $\rho$-sampling Pix 0.01 & 2.66 & 0.67 & 0.5 \\
SR3 $^\dagger$\cite{saharia2021image} & 11.30 & - & - & & & & & \\ \cline{6-9}
IDDPM$^\dagger$\cite{nichol2021improved} & 12.26 & 0.70 & 0.62 & & \multicolumn{4}{r}{\textbf{AFHQv2 512$\times$512}} \\ \cline{6-9}
ADM$^\dagger$\cite{dhariwal2021diffusion} & 10.94 & 0.69 & \textbf{0.63} & & StyleGAN2$^\dagger$ & 4.62 & - & - \\
ICGAN+DA$^\dagger$\cite{casanova2021instanceconditioned} & 7.50 & - & - & & StyleGAN3-R$^\dagger$ & 4.40 & - & - \\
BigGAN-deep & 6.86 & 0.85 & 0.29 & & StyleGAN3-T & 4.05 & 0.70 & 0.55 \\
+ $\rho$-sampling Pix 0.0065 & 6.82 & \textbf{0.86} & 0.29 & & + $\rho$-sampling Vgg -0.001 & \textbf{3.95} & \textbf{0.71} & \textbf{0.55} \\
ADM+classifier guidance & \textbf{4.59} & 0.82 & 0.52 & & & & & \\ \hline
\end{tabular}
\vspace{-0.2cm}
\caption{$^\dagger$Paper reported metrics. We observe that moving away from $\rho=0$, Polarity Sampling improves FID across models and datasets, empirically validating that the top singular values of a DGN's Jacobian matrices contain meaningful information to improve the overall quality of generation}.
\label{tab:main_result}
\end{table*}
In \cref{tab:main_result}, we present for LSUN \{Church, Car, Cat\} \cite{yu2015lsun}, ImageNet \cite{deng2009imagenet}, FFHQ \cite{karras2019style}, and AFHQv2 \cite{choi2020stargan,karras2021alias} improved FID obtained \textit{solely by changing the polarity $\rho$} of a state-of-the-art DGN. This implies that Polarity Sampling
provides an efficient solution to adapt the DGN latent space.
We observe that, given any specific setting, $\rho \not = 0$ always improves a model's FID. We see that in a case specific manner, both positive and negative $\rho$ improves the FID.For StyleGAN2-F trained on FFHQ, \textit{increasing the diversity} of the inception space embedding distribution helps reach a new state-of-the-art FID. By \textit{increasing the precision} of StyleGAN3-T via Polarity Sampling in the Vgg space, we are able to surpass the FID of baseline StyleGAN2-F \cite{karras2021alias}. We observe that controlling the polarity of the InceptionV3 embedding distribution of StyleGAN2-F gives the most significant gains in terms of FID. This is due to the fact that the Frechet distance between real and generated distributions is directly affected while performing Polarity Sampling in the Inception space. We provide generated samples in \cref{fig:qualitative_alpha_sweep} varying the style-space $\rho$ for LSUN cars and LSUN cats, whereas varying the pixel-space $\rho$ for LSUN Church. It is clear that $\rho<0$ i.e. sampling closer to the DGN distribution modes produce samples of high visual quality, while $\rho>0$ i.e. sampling closer to the regions of low-probability produce samples of high-diversity, with some samples which are off the data manifold due to the approximation quality of the DGN in that region. Using Polarity Sampling, we are able to advance the state-of-the-art performance on three different settings: for StyleGAN2 on the FFHQ \cite{karras2019style} Dataset to FID 2.57, StyleGAN2 on the LSUN \cite{yu2015lsun} Car Dataset to FID 2.27, and StyleGAN3 on the AFHQv2 \cite{karras2021alias} Dataset to FID 3.95. For additional experiments with NVAE and color-MNIST under controlled training and reference dataset distribution shift, see \cref{sec:shift}.
\vspace{-0.2cm}
\section{New Insights into DGN Distributions}
\vspace{-0.2cm}
In \cref{sec:experiments} we demonstrated that Polarity Sampling
is a practical method to manipulate DGN output distributions to control their quality and diversity.
We now demonstrate that Polarity Sampling has more foundational theoretical applications as well.
In particular, we dive into several timely questions regarding DGNs that can be probed using our framework.
\vspace{-0.3cm}
\subsection{Are GAN/VAE Modes Training Samples?}
\label{sec:mode}
\vspace{-0.3cm}
Mode collapse \cite{metz2016unrolled,srivastava2017veegan,bang2021mggan} has complicated GAN training for many years. It consists of the entire DGN collapsing to generating a few different samples or modes.
For VAEs, modes can be expected to be related to the modes of the empirical dataset distribution, as reconstruction is part of the objective. But this might not be the case with GANs e.g., the modes can correspond to parts of the space where the discriminator is the least good at differentiating between true and fake samples. There has been no reported methods in literature that allows us to observe the modes of a trained GAN. Existing visualization techniques focus on finding the role of each DGN unit \cite{bau2018gan} or finding images that GANs cannot generate \cite{bau2019seeing}.
Using Polarity Sampling, we can visualize the DGN modes of DGNs for the first time. In \cref{fig:modes}, we present samples from the modes of BigGAN-deep trained on ImageNet, StyleGAN3 trained on AFHQv2, and NVAE trained on colored-MNIST. We observe that BigGAN modes tend to reproduce the unique features of the class, removing the background and focusing more on the object the class is assigned to. AFHQv2 modes on the other hand, focus on younger animal faces and smoother textures. NVAE modes predominately samples the digit `1' which corresponds to the dataset mode (digit with the least intra-class variation).
We also provide in \cref{fig:nn_hist} the distribution of the $l_2$ distances between generated samples and their $3$ nearest training samples for modal ($\rho=$ $-5$) and anti-modal ($\rho=1$) polarity. We see that changing the polarity has negligible effect on the StyleGAN2 nearest neighbor distribution, but for NVAE the modes come closer to the training samples.
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.32\linewidth}
\begin{minipage}{\columnwidth} \centering {\tiny BigGAN Samoyed} \end{minipage}
\includegraphics[width=.99\columnwidth]{sec/samoyed_modes.png}\\
\includegraphics[width=.99\columnwidth]{sec/tinca_modes_blurred.png}
\begin{minipage}{\columnwidth} \centering \vspace{-1em} {\tiny BigGAN Tench} \end{minipage}
\end{minipage}
\begin{minipage}{0.32\linewidth}
\begin{minipage}{\columnwidth} \centering {\tiny BigGAN Flamingo} \end{minipage}
\includegraphics[width=.99\columnwidth]{sec/flamingo_modes.png}
\includegraphics[width=.99\columnwidth]{sec/afhqmode.PNG}
\begin{minipage}{\columnwidth} \centering \vspace{-1em} {\tiny StyleGAN3 AFHQv2} \end{minipage}
\end{minipage}
\begin{minipage}{0.32\linewidth}
\begin{minipage}{\columnwidth} \centering {\tiny BigGAN Egyptian cat} \end{minipage}
\includegraphics[width=.99\columnwidth]{sec/egyptianmodes.png}\\
\includegraphics[width=.99\columnwidth]{sec/nvae_modes.png}
\begin{minipage}{\columnwidth} \centering \vspace{-1em} {\tiny NVAE colored-MNIST} \end{minipage}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{Modes for BigGAN-deep, StyleGAN3-T and NVAE obtained via $\rho\ll0$ Polarity Sampling. {\bf This is, to the best of our knowledge, the first visualization of the modes of DGNs in pixel space.}}
\label{fig:modes}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.49\linewidth}
\includegraphics[width=.99\columnwidth]{nn_hist_lsun_church_stg2.pdf}
\end{minipage}
\begin{minipage}{0.49\linewidth}
\includegraphics[width=.99\columnwidth]{nn_hist_nvae.pdf}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{Distribution of $l_2$ distance to 3 training set nearest neighbors, for 1000 generated samples from LSUN Church StyleGAN2 {\bf (left)} and colored-MNIST NVAE {\bf (right)}. Samples closer to the modes ($\rho<0$) for NVAE get significant closer to the training sampling while nearly no changes occurs for StyleGAN2. This behavior is expected as {\bf VAE models are encouraged to position their modes on the training samples, as opposed to GANs whose modes depend on the discriminator}.
}
\label{fig:nn_hist}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.49\linewidth}
\includegraphics[width=.99\columnwidth]{ffhq/ppl_zendpoints_ffhqstg2_vggpol.pdf}
\end{minipage}
\begin{minipage}{0.49\linewidth}
\includegraphics[width=.99\columnwidth]{ffhq/ppl_Wendpoints_ffhqstg2_vggpol.pdf}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{Distribution of PPL for StyleGAN2-F trained on FFHQ with varying Polarity Sampling (in VGG space) setting ($\rho$ given in the legend) for endpoints in the input latent space ({\bf left}) and endpoints in style-space ({\bf right}). The means of the distributions (PPL score) are provided as markers on the horizontal axis.
}
\label{fig:PPL}
\end{figure}
\vspace{-0.3cm}
\subsection{Perceptual Path Length Around Modes}
\label{sec:truncation}
\label{sec:PPL}
\vspace{-0.1cm}
Perceptual Path Length (PPL) is the VGG space distance between two latent space points, that is used as a measure of perceptual distance \cite{karras2020analyzing}. In \cref{fig:PPL}, we report the PPL of a StyleGAN2-F trained on FFHQ, for an interpolation step of length $10^{-4}$ between endpoints from the latent/style space. We sample points using Polarity Sampling varying $\rho \in [1,-1]$, essentially measuring the PPL for regions of the data manifold with increasing density. We see that for negative values of polarity (close to the modes), we have significantly lower PPL compared to positive polarity or even baseline sampling ($\rho=0$). This result shows that for StyleGAN2, there are smoother perceptual transitions closer to modes. We tabulate PPL values in \cref{appendix:extra}.
While truncation also reduces the PPL, it essentially does so by sampling points closer to the style space mean \cite{karras2019style}. Polarity Sampling on the other hand can be used to sample from Vgg modes, making it the first method that can be used to explicitly sample regions that are perceptually smoother. Polarity Sampling may therefore lead to the development of new sophisticated interpolation methods where, instead of linear or spherical interpolation between latent space points, the interpolation is done along a low PPL region of the manifold.
\section{Conclusions}
\vspace{-0.3cm}
We have demonstrated in this paper that the analytical form of a DGN's output space distribution provides a new mechanism to parameterize the DGN prior $p_{{\bm{z}}}$ in terms of a single additional hyperparameter --- the polarity $\rho$--- and force the DGN samples to concentrate near the distribution modes or anti-modes (\cref{sec:method}).
Our experiments with Polarity Sampling have demonstrated convincingly that the polarity $\rho$ provides a novel solution to control the quality and diversity of samples, improving upon the usual solution of latent space truncation (\cref{sec:pareto}). As a significant practical bonus, $\rho$ also improved a range of DGNs' FIDs to reach state-of-the-art performance.
On the theoretical side, Polarity Sampling's guarantee that it samples from the modes of a DGN enabled us to explore some timely open questions,
including the relation between distribution modes and training samples (\cref{sec:mode}), the impact of truncation on producing samples (not) near the distribution modes, and the effect of going from mode to anti-mode generation on the perceptual-path-length (\cref{sec:PPL}).
We believe that Polarity Sampling
will open the door to even more research directions and applications in the future.
\section{Introduction}
\label{sec:intro}
Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press.
This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version.
\subsection{Language}
All manuscripts must be in English.
\subsection{Dual submission}
Please refer to the author guidelines on the CVPR\ 2022\ web page for a
discussion of the policy on dual submissions.
\subsection{Paper length}
Papers, excluding the references section, must be no longer than eight pages in length.
The references section will not be included in the page count, and there is no limit on the length of the references section.
For example, a paper of eight pages with two pages of references would have a total length of 10 pages.
{\bf There will be no extra page charges for CVPR\ 2022.}
Overlength papers will simply not be reviewed.
This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide.
Note that this \LaTeX\ guide already sets figure captions and references in a smaller font.
The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts.
The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven.
\subsection{The ruler}
The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review.
The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution.
If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages.
The presence or absence of the ruler should not change the appearance of any other content on the page.
The camera-ready copy should not contain a ruler.
(\LaTeX\ users may use options of cvpr.sty to switch between different versions.)
Reviewers:
note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly.
Just use fractional references (\eg, this line is $087.5$), although in most cases one would expect that the approximate location will be adequate.
\subsection{Paper ID}
Make sure that the Paper ID from the submission system is visible in the version submitted for review (replacing the ``*****'' you see in this document).
If you are using the \LaTeX\ template, \textbf{make sure to update paper ID in the appropriate place in the tex file}.
\subsection{Mathematics}
Please number all of your sections and displayed equations as in these examples:
\begin{equation}
E = m\cdot c^2
\label{eq:important}
\end{equation}
and
\begin{equation}
v = a\cdot t.
\label{eq:also-important}
\end{equation}
It is important for readers to be able to refer to any particular equation.
Just because you did not refer to it in the text does not mean some future reader might not need to refer to it.
It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''.
(Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers).
All authors will benefit from reading Mermin's description of how to write mathematics:
\url{http://www.pamitc.org/documents/mermin.pdf}.
\subsection{Blind review}
Many authors misunderstand the concept of anonymizing for blind review.
Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available.
Blind review means that you do not use the words ``my'' or ``our'' when citing previous work.
That is all.
(But see below for tech reports.)
Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith;
it says that you are building on her work.
If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work.
An example of a bad paper just asking to be rejected:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods.
Why the previous paper was accepted without this analysis is beyond me.
[1] Removed for blind review
\end{quote}
An example of an acceptable paper:
\begin{quote}
\begin{center}
An analysis of the frobnicatable foo filter.
\end{center}
In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods.
Why the previous paper was accepted without this analysis is beyond me.
[1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213.
\end{quote}
If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work.
In such cases, include the anonymized parallel submission~\cite{Authors14} as supplemental material and cite it as
\begin{quote}
[1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as supplemental material {\tt fg324.pdf}.
\end{quote}
Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report.
For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a tech report for further details.
Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''.
Then submit the tech report as supplemental material.
Again, you may not assume the reviewers will read this material.
Sometimes your paper is about a problem which you tested using a tool that is widely known to be restricted to a single institution.
For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the CVPR70 audience would like to hear about your
solution.
The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal.
You can handle this paper like any other.
Do not write ``We show how to improve our previous work [Anonymous, 1968].
This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''.
That would be silly, and would immediately identify the authors.
Instead write the following:
\begin{quotation}
\noindent
We describe a system for zero-g frobnication.
This system is new because it handles the following cases:
A, B. Previous systems [Zeus et al. 1968] did not handle case B properly.
Ours handles it by including a foo term in the bar integral.
...
The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know.
It displayed the following behaviours, which show how well we solved cases A and B: ...
\end{quotation}
As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors.
A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess.
He or she would have to be sure that no other authors could have been contracted to solve problem B.
\medskip
\noindent
FAQ\medskip\\
{\bf Q:} Are acknowledgements OK?\\
{\bf A:} No. Leave them for the final copy.\medskip\\
{\bf Q:} How do I cite my results reported in open challenges?
{\bf A:} To conform with the double-blind review policy, you can report results of other challenge participants together with your results in your paper.
For your results, however, you should not identify yourself and should not mention your participation in the challenge.
Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\
\begin{figure}[t]
\centering
\fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
\caption{Example of caption.
It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.}
\label{fig:onecol}
\end{figure}
\subsection{Miscellaneous}
\noindent
Compare the following:\\
\begin{tabular}{ll}
\verb'$conf_a$' & $conf_a$ \\
\verb'$\mathit{conf}_a$' & $\mathit{conf}_a$
\end{tabular}\\
See The \TeX book, p165.
The space after \eg, meaning ``for example'', should not be a sentence-ending space.
So \eg is correct, {\em e.g.} is not.
The provided \verb'\eg' macro takes care of this.
When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word).
If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal.
However, use it only when there are three or more authors.
Thus, the following is correct:
``Frobnication has been trendy lately.
It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by
Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.''
This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors.
\begin{figure*}
\centering
\begin{subfigure}{0.68\linewidth}
\fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}}
\caption{An example of a subfigure.}
\label{fig:short-a}
\end{subfigure}
\hfill
\begin{subfigure}{0.28\linewidth}
\fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}}
\caption{Another example of a subfigure.}
\label{fig:short-b}
\end{subfigure}
\caption{Example of a short caption, which should be centered.}
\label{fig:short}
\end{figure*}
\section{Formatting your paper}
\label{sec:formatting}
All text must be in a two-column format.
The total allowable size of the text area is $6\frac78$ inches (17.46 cm) wide by $8\frac78$ inches (22.54 cm) high.
Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them.
The main title (on the first page) should begin 1 inch (2.54 cm) from the top edge of the page.
The second and following pages should begin 1 inch (2.54 cm) from the top edge.
On all pages, the bottom margin should be $1\frac{1}{8}$ inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper;
for A4 paper, approximately $1\frac{5}{8}$ inches (4.13 cm) from the bottom edge of the
page.
\subsection{Margins and page numbering}
All printed material, including text, illustrations, and charts, must be kept
within a print area $6\frac{7}{8}$ inches (17.46 cm) wide by $8\frac{7}{8}$ inches (22.54 cm)
high.
Page numbers should be in the footer, centered and $\frac{3}{4}$ inches from the bottom of the page.
The review version should have page numbers, yet the final version submitted as camera ready should not show any page numbers.
The \LaTeX\ template takes care of this when used properly.
\subsection{Type style and fonts}
Wherever Times is specified, Times Roman may also be used.
If neither is available on your word processor, please use the font closest in
appearance to Times to which you have access.
MAIN TITLE.
Center the title $1\frac{3}{8}$ inches (3.49 cm) from the top edge of the first page.
The title should be in Times 14-point, boldface type.
Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs;
do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word).
Leave two blank lines after the title.
AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title
and printed in Times 12-point, non-boldface type.
This information is to be followed by two blank lines.
The ABSTRACT and MAIN TEXT are to be in a two-column format.
MAIN TEXT.
Type main text in 10-point Times, single-spaced.
Do NOT use double-spacing.
All paragraphs should be indented 1 pica (approx.~$\frac{1}{6}$ inch or 0.422 cm).
Make sure your text is fully justified---that is, flush left and flush right.
Please do not place any additional blank lines between paragraphs.
Figure and table captions should be 9-point Roman type as in \cref{fig:onecol,fig:short}.
Short captions should be centred.
\noindent Callouts should be 9-point Helvetica, non-boldface type.
Initially capitalize only the first word of section titles and first-, second-, and third-order headings.
FIRST-ORDER HEADINGS.
(For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after.
SECOND-ORDER HEADINGS.
(For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after.
If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line.
\subsection{Footnotes}
Please use footnotes\footnote{This is what a footnote looks like.
It often distracts the reader from the main flow of the argument.} sparingly.
Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence).
If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced.
Use Times 8-point type, single-spaced.
\subsection{Cross-references}
For the benefit of author(s) and readers, please use the
{\small\begin{verbatim}
\cref{...}
\end{verbatim}} command for cross-referencing to figures, tables, equations, or sections.
This will automatically insert the appropriate label alongside the cross-reference as in this example:
\begin{quotation}
To see how our method outperforms previous work, please see \cref{fig:onecol} and \cref{tab:example}.
It is also possible to refer to multiple targets as once, \eg~to \cref{fig:onecol,fig:short-a}.
You may also return to \cref{sec:formatting} or look at \cref{eq:also-important}.
\end{quotation}
If you do not wish to abbreviate the label, for example at the beginning of the sentence, you can use the
{\small\begin{verbatim}
\Cref{...}
\end{verbatim}}
command. Here is an example:
\begin{quotation}
\Cref{fig:onecol} is also quite important.
\end{quotation}
\subsection{References}
List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper.
When referenced in the text, enclose the citation number in square brackets, for
example~\cite{Authors14}.
Where appropriate, include page numbers and the name(s) of editors of referenced books.
When you cite multiple papers at once, please make sure that you cite them in numerical order like this \cite{Alpher02,Alpher03,Alpher05,Authors14b,Authors14}.
If you use the template as advised, this will be taken care of automatically.
\begin{table}
\centering
\begin{tabular}{@{}lc@{}}
\toprule
Method & Frobnability \\
\midrule
Theirs & Frumpy \\
Yours & Frobbly \\
Ours & Makes one's heart Frob\\
\bottomrule
\end{tabular}
\caption{Results. Ours is better.}
\label{tab:example}
\end{table}
\subsection{Illustrations, graphs, and photographs}
All graphics should be centered.
In \LaTeX, avoid using the \texttt{center} environment for this purpose, as this adds potentially unwanted whitespace.
Instead use
{\small\begin{verbatim}
\centering
\end{verbatim}}
at the beginning of your figure.
Please ensure that any point you wish to make is resolvable in a printed copy of the paper.
Resize fonts in figures to match the font in the body text, and choose line widths that render effectively in print.
Readers (and reviewers), even of an electronic copy, may choose to print your paper in order to read it.
You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic.
When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below
{\small\begin{verbatim}
\usepackage{graphicx} ...
\includegraphics[width=0.8\linewidth]
{myfile.pdf}
\end{verbatim}
}
\subsection{Color}
Please refer to the author guidelines on the CVPR\ 2022\ web page for a discussion of the use of color in your document.
If you use color in your plots, please keep in mind that a significant subset of reviewers and readers may have a color vision deficiency; red-green blindness is the most frequent kind.
Hence avoid relying only on color as the discriminative feature in plots (such as red {\bm{s}} green lines), but add a second discriminative feature to ease disambiguation.
\section{Final copy}
You must include your signed IEEE copyright release form when you submit your finished paper.
We MUST have this form before your paper can be published in the proceedings.
Please direct any questions to the production editor in charge of these proceedings at the IEEE Computer Society Press:
\url{https://www.computer.org/about/contact}.
\section{Implementation Details and Online Sampling Solution}
\label{sec:implementation}
\subsection{Online Algorithm}
\label{sec:online_algorithm}
One important aspect of polarity-sampling, as summarized in \cref{alg:sampling} is the need to first sample the DGN latent space to obtain the top singular values of as much per-region slope matrices ${\bm{A}}_{\omega}$ as possible. This might seem as a bottleneck if one wants to repeatedly apply polarity-sampling on a same DGN. However, this is to provide an estimate of the DGN per-region change of variables, and as the DGN is not retrained nor fine-tuned, it only needs to be done once. Furthermore, this allows for an online sampling algorithm that we provide in \cref{alg:online}. In short, one first perform this task of estimating as much per-region top singular values as possible, once this is completed, only sampling of latent vectors ${\bm{z}}$ and rejection sampling based on the corresponding ${\bm{A}}_{\omega}$ matrix is done online, ${\bm{A}}_{\omega}$ of the sampled ${\bm{z}}$ being obtained easily via ${\bm{A}}_{\omega}={\bm{J}} G ({\bm{z}})$.
\begin{algorithm}[]
\caption{Online Rejection Sampling Algorithm}
\begin{algorithmic}
\Require Latent space domain, $\mathcal{D}$; Generator ${\bm{G}}$; $N$ change of volume scalars $\{\sigma_1, \sigma_2,...,\sigma_N \}$; Number of singular values $K$;
\While{True}{}
\State $z \sim U(\mathcal{D})$
\State $\alpha \sim U[0,1]$
\State ${\bm{A}} = {\bm{J}}_{{\bm{G}}}(z)$
\State $\sigma_z = \prod^K_{k=1} K$-$SingularValues({\bm{A}},K)$
\vspace{.5em}
\State \algorithmicif $\frac{\sigma_z^\rho}{\sigma_z^\rho + \sum_{i=1}^N \sigma_i^\rho} \geq \alpha$ \algorithmicthen\
\vspace{.5em}
$x \gets {\bm{G}}(z)$
\Return $x$
\EndWhile
\end{algorithmic}
\label{alg:online}
\end{algorithm}
\subsection{Effect of $N$ and $k$}
\label{sec:effect_N_k}
One important aspect of our algorithm comes from the two hyper-parameters $N$ and $k$. They represent respectively the number of latent space samples to use to estimate as much ${\bm{A}}_{\omega}$ as possible (recall \cref{alg:sampling}), and the number of top singular values to compute. Both represent a trade-off between exact polarity-sampling, and computation complexity. We argue that in practice, $N\approx 150K$ and $k\approx 100$ is enough to obtain a good estimate of the polarity-sampling distribution (\cref{eq:density_rho}). To demonstrate that, we first provide an ablation study of the number of $N$ and $k$ used for polarity sampling in \cref{tab:ablation_N} and \cref{tab:ablation_topk}. We also present a visual inspection of the impact of $N$ and $k$ on the precision and recall in \cref{fig:topk_prec_recall}.
\begin{table}[]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}crrrccc@{}}
\toprule
& \multicolumn{3}{c}{FFHQ 1024$\times$1024} & \multicolumn{3}{c}{LSUN Cat 256$\times$256} \\ \midrule
$N$ & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}FID\\ (lowest)\end{tabular}} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Precision\\ (max)\end{tabular}} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Recall\\ (max)\end{tabular}} & \begin{tabular}[c]{@{}c@{}}FID\\ (lowest)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Precision\\ (max)\end{tabular} & \begin{tabular}[c]{@{}c@{}}Recall\\ (max)\end{tabular} \\ \midrule
100K & 2.63 & 0.80 & 0.59 & 6.38 & 0.69 & 0.31 \\
200K & 2.62 & 0.82 & 0.63 & 6.38 & 0.71 & 0.32 \\
250K & 2.59 & 0.84 & 0.64 & 6.39 & 0.74 & 0.31 \\
300K & 2.61 & 0.87 & 0.65 & 6.37 & 0.75 & 0.33 \\
500K & 2.58 & 0.90 & 0.67 & 6.34 & 0.77 & 0.33
\end{tabular}
}
\caption{Ablation of $N$ and its effect on best FID, Precision and Recall values that can be obtained by a StyleGAN2 ($\psi=1$) on the FFHQ and LSUN Cat dataset. We vary the polarity in the VGG space for FFHQ dataset, and style space for LSUN Cats, for number of singular values $k=30$.}
\label{tab:ablation_N}
\end{table}
\begin{table}[]
\centering
\begin{tabular}{@{}crrr@{}}
\toprule
& \multicolumn{3}{c}{FFHQ 1024$\times$1024} \\ \midrule
$K$ & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}FID\\ (lowest)\end{tabular}} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Precision\\ (max)\end{tabular}} & \multicolumn{1}{c}{\begin{tabular}[c]{@{}c@{}}Recall\\ (max)\end{tabular}} \\ \midrule
10 & 2.71 & 0.89 & 0.65 \\
20 & 2.67 & 0.90 & 0.66 \\
40 & 2.57 & 0.90 & 0.67 \\
60 & 2.62 & 0.90 & 0.66 \\
80 & 2.67 & 0.90 & 0.66 \\
100 & 2.70 & 0.90 & 0.67 \\ \bottomrule
\end{tabular}
\caption{Ablation of $K$ and its effect on the best FID, Precision and Recall values that can be obtained by a StyleGAN2 ($\psi=1$) on the FFHQ dataset. We vary the polarity in the inception space for FFHQ dataset}
\label{tab:ablation_topk}
\end{table}
\subsection{Computation Times and Employed Software/Hardware}
\label{sec:times}
All the experiments were run on a Quadro RTX 8000 GPU, which has 48 GB of high-speed GDDR6 memory and 576 Tensor cores. For the software details we refer the reader to the provided codebase. In short, we employed TF2 (2.4 at the time of writing), all the usual Python scientific libraries such as NumPy and PyTorch. We employed the official repositories of the various models we employed with official pre-trained weights.
As a note, most of the architectures can not be run on GPUs with less or equal to 12 GB of memory.
We report here the Jacobian computation times for Tensorflow 2.5 with CUDA 11 and Cudnn 8 on an NVIDIA Titan RTX GPU. For StyleGAN2 pixel space, 5.03s/it; StyleGAN2 style-space, 1.12s/it; BigGAN 5.95s/it; ProgGAN 3.02s/it. For NVAE on Torch 1.6 it takes 20.3s/it. Singular value calculation for StyleGAN2 pixel space takes .005s/it, StyleGAN2 style space .008s/it, BigGAN .001s/it, ProgGAN .004s/it and NVAE .02s/it on NumPy. According to this, for StyleGAN2-e, N=250,000 requires ~14 days to obtain. This only needs to be done once, and it is also possible to perform online sampling once it is calculated. The time required for this is relatively small compared to the training time required for only one set of hyperparameters, which is ~35 days and 11 hours\footnote{https://github.com/NVlabs/stylegan2}. We have added pseudocode for MaGNET sampling and online sampling in Appendix G.
\subsection{Reducing Memory Requirements}
\label{sec:memory}
The core of polarity-sampling relies on computing the top-singular values of the possibly large matrix ${\bm{A}}_{\omega}$ for a variety of regions $\omega \in \Omega$, discovered through latent space sampling (recall \cref{alg:sampling}). One challenge for state-of-the-art DGNs lies in the size of the matrices ${\bm{A}}_{\omega}$. Multiple solutions exist, such as computing the top singular values through block power iterations. Doing so, the matrices ${\bm{A}}_{\omega}$ do not need to be computed entirely, only the matrix-matrix product ${\bm{A}}_{\omega}{\bm{W}}$ and ${\bm{A}}^T_{\omega}{\bm{V}}$ needs to be performed repeatedly (interleaved with QR decompositions). After many iterations, ${\bm{W}}$ estimate the top right-singular vectors of ${\bm{A}}_{\omega}$, and ${\bm{V}}$ the corresponding top left-singular vectors from which the singular values can be obtained. However, we found this solution to remain computationally extensive, and found that in practice, a simpler approximation that we now describe provided sufficiently accurate estimates.
Instead of the above iterative estimation, one can instead compute the top-singular values of ${\bm{W}}{\bm{A}}_{\omega}$ with ${\bm{W}}$ a semi-orthogonal matrix of shape $D'\times D$ with $D'<D$ (recall that ${\bm{A}}_{\omega}$ is of shape $D \times K$). Doing so, we are now focusing on the singular values of ${\bm{A}}_{\omega}$ whose left-singular vectors are not orthogonal with the right singular vectors of ${\bm{W}}$. While this possible incur an approximation error, we found that the above was sufficient to provide polarity-sampling an adequate precision-recall control.
\subsection{Applying Polarity Sampling in Style, VGG and Inception Space}
\label{sec:other_space}
We call the ambient space of the images the pixel-space, because each dimension in this space corresponds to individual pixels of the images. Apart from controlling the density of the pixel-space manifold, polarity can also be used to control the density of the style-space manifold for style based architectures such as StyleGAN\{1,2,3\} \cite{karras2019style,karras2020analyzing,karras2021alias}. We also extend the idea of intermediate manifolds to feature space manifolds such as VGG or InceptionV3 space, which can be assumed continuous mappings of the pixel space to the corresponding models' bottleneck embedding space. In \cref{fig:more_pareto}-left we present comparisons between Style, Pixel, VGG and Inception space precision-recall curves for StyleGAN2-F FFHQ with $\psi=1$, top-$k=30$ and $\rho=[-2,2]$. We see that the VGG and InceptionV3 curves trace almost identically. This is expected behavior since both these feature spaces correspond to perceptual features, therefore the transform they induce on the pixel space distribution is almost identical. On the other hand, the pixel space distribution saturates at high polarity at almost equal values. The point of equal precision and recall for both the Inception and VGG spaces, occur at a polarity of 0.1. Its clear from the figures that feature space polarity changes have a larger effect on precision and recall compared to pixel-space and style-space has the least effect on precision and recall. This could be due to the number of density transforms the style-space distribution undergoes until the VGG space, where precision and recall is calculated. In \cref{fig:more_pareto}-right we present the polarity characteristics for StyleGAN2-E, StyleGAN2-F and StyleGAN3. For each model, we choose the best space w.r.t the pareto frontier, VGG and Inception space for StyleGAN2-E and StyleGAN2-F, and pixel-space for StyleGAN3. Notice that StyleGAN3 exceeds the recall of the other two models for negative polarity, while matching the precision for StyleGAN2-E.
\section{Proofs}
\label{sec:proofs}
The proofs of the two main claims of the paper heavily rely on the spline form of the DGN input-output mapping from \cref{eq:CPA}. For more background on the form of the latent space partition $\Omega$, the per-region affine mappings ${\bm{A}}_{\omega},{\bm{b}}_{\omega},\forall \omega \in \Omega$ and further discussion on how to deal with DGN including smooth activation functions, we refer the reader to \cite{balestriero2020mad}.
\subsection{Proof of \cref{thm:general_density}}
\label{proof:general_density}
\begin{proof}
We will be doing the change of variables ${\bm{z}}=({\bm{A}}^T_{\omega}{\bm{A}}_{\omega})^{-1}{\bm{A}}_{\omega}^T({\bm{x}} - {\bm{b}}_{\omega})\triangleq {\bm{A}}_{\omega}^\dagger({\bm{x}} - {\bm{b}}_{\omega})$, also notice that $J_{G^{-1}}({\bm{x}})=A^\dagger$.
First, we know that
$
P_{G({\bm{z}})}({\bm{x}} \in w)= P_{{\bm{z}}}({\bm{z}} \in G^{-1}(w))= \int_{G^{-1}(w)}p_{{\bm{z}}}({\bm{z}}) d{\bm{z}}
$ which is well defined based on our full rank assumptions. We then proceed by
\begin{align*}
P_{G}({\bm{x}} \in w)=& \sum_{\omega \in \Omega}\int_{\omega \cap w} p_{{\bm{z}}}(G^{-1}({\bm{x}})) \\
&\times \sqrt{ \det(J_{G^{-1}}({\bm{x}})^TJ_{G^{-1}}({\bm{x}}))} d{\bm{x}}\\
=& \sum_{\omega \in \Omega}\int_{\omega \cap w}p_{{\bm{z}}}(G^{-1}({\bm{x}})) \\
&\times \sqrt{ \det(({\bm{A}}_{\omega}^{+})^T{\bm{A}}_{\omega}^{+})} d{\bm{x}}\\
=& \sum_{\omega \in \Omega}\int_{\omega \cap w} p_{{\bm{z}}}(G^{-1}({\bm{x}}))\frac{1}{\sqrt{\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})}} d{\bm{x}},\\
\end{align*}
where the second to third equality follows by noticing that $\sigma_i(A^\dagger)=(\sigma_i(A))^{-1}$ which can be showed easily be replacing ${\bm{A}}_{\omega}$ with its SVD and unrolling the product of matrices.
Now considering a uniform latent distribution case on a bounded domain $U$ in the DGN latent space we obtain by substitution in the above result
\begin{align}
p_{G}({\bm{x}}) = \frac{\sum_{\omega \in \Omega}\mathrm{1}_{{\bm{x}} \in \omega}\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{-\frac{1}{2}}}{Vol(U)},
\end{align}
leading to the desired result.
\end{proof}
\subsection{Proof of \cref{cor:pol_density}}
\label{proof:pol_density}
\begin{proof}
The proof of this result largely relies on \cref{thm:general_density}. Taking back our previous result, we know that
\begin{equation}
p_{G}({\bm{x}})=\sum_{\omega \in \Omega} p_{{\bm{z}}}(G^{-1}({\bm{x}}))\mathrm{1}_{\{G^{-1}({\bm{x}}) \in \omega\}}\frac{1}{\sqrt{\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})}} d{\bm{x}}.
\end{equation}
However, recall that polarity sampling leverages the prior probability given by
\begin{equation}
p_{\rho}({\bm{z}}) =\frac{1}{\kappa} \sum_{\omega \in \Omega}\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{\frac{\rho}{2}} \mathds{1}_{\{{\bm{z}} \in \omega\}},
\end{equation}
which, after replacing $G^{-1}({\bm{x}})$ with its corresponding ${\bm{z}}$ becomes
\begin{equation}
p_{G}({\bm{x}})=\sum_{\omega \in \Omega} \frac{1}{\kappa} \frac{\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{\frac{\rho}{2}}}{\sqrt{\det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})}}\mathds{1}_{\{{\bm{z}} \in \omega'\}} d{\bm{x}},
\end{equation}
and simplifies to
\begin{equation}
p_{G}({\bm{x}})=\sum_{\omega \in \Omega} \frac{1}{\kappa} \det({\bm{A}}_{\omega}^T{\bm{A}}_{\omega})^{\frac{\rho-1}{2}}\mathds{1}_{\{{\bm{z}} \in \omega'\}} d{\bm{x}},
\end{equation}
leading to the desired result. Note that when $\rho=1$ then the density is uniform onto the DGN manifold, when $\rho=0$, one recovers the original DGN density onto the manifold, and in the extrema cases, only the region with highest or lowest probability would be sample i.e. the modes or anti-modes.
\end{proof}
\section{Extra experiments}
\label{appendix:extra}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.02\linewidth}
\centering
\rotatebox{90}{\footnotesize \hspace{0.3cm}Precision/Recall}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_vgg-pix-inc_precrecall.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_stg3-2f-2e_precrecall.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\includegraphics[width=0.49\linewidth]{pareto/image.png}
\includegraphics[width=0.49\linewidth]{pareto/image-2.png}
\caption{ Top: Precision Recall tradeoff for polarity sweep on VGG, Inception and Pixel space distributions.
Bottom: BigGAN-deep Imagenet pareto curves obtained for a few classes, in red is the baseline while each other scatter point can be reach by varying truncation and $\rho$ Calculated with $1300$ real and generated samples.
}
\label{fig:more_pareto}
\end{figure}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.02\linewidth}
\centering
\rotatebox{90}{\footnotesize \hspace{0.7cm}Precision}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_vggpol_num_detz_prec.pdf}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_vggpol_topksweep_prec.pdf}
\end{minipage}
\begin{minipage}{0.02\linewidth}
\centering
\rotatebox{90}{\footnotesize \hspace{0.7cm}Recall}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_vggpol_num_detz_recall.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\centering
\includegraphics[width=\linewidth]{ffhq/ffhq_vggpol_topksweep_recall.pdf}\\[-0.5em]
{\footnotesize polarity $(\rho)$}
\end{minipage}
\end{center}
\vspace{-0.5cm}
\caption{ Effect of Polarity Sampling on Precision (top) and Recall (bottom) of a StyleGAN2-F model pretrained on FFHQ for varying number of top-$k$ singular values (right) and varying number of latent space samples $N$ (left) used to obtain per-region slope matrix ${\bm{A}}_{\omega}$ singular values (recall \cref{sec:pseudocode,alg:sampling}). The trend in metrics stabilizes when using around $N\approx$300,000 latent space samples. Increasing the number of top-$k$ singular values to use, amplifies the effect of polarity, saturating at around $k=50$.
}
\label{fig:topk_prec_recall}
\end{figure}
\begin{figure*}[t!]
\begin{center}
\begin{minipage}{\linewidth}
\includegraphics[width=.49\columnwidth]{modes/pug_modes.png}
\includegraphics[width=.49\columnwidth]{modes/lionmodes.png}\\
\includegraphics[width=.49\columnwidth]{modes/cheeseburger_modes.png}
\includegraphics[width=.49\columnwidth]{modes/pomerian_modes.png}\\
\end{minipage}
\end{center}
\caption{Modes for BigGAN-deep trained on Imagenet, and conditioned on the a specific class: ``pug'' (top left), ``lion'' (top right), ``cheeseburger'' (bottom left) and ``pomerian'' (bottom right). We observe that the modes correspond to nearly aligned faces with little to no background. Variation of colors and sizes can be seen across the modes. The same observation can be made for the cheeseburger, nearly no background is present, and the shape is consistent to a typical cheeseburger ''template''. See \cref{fig:modes_per_class} for additional classes.
}
\label{fig:truncation_sweep_images}
\end{figure*}
\begin{figure}[t!]
\begin{center}
\begin{minipage}{0.48\linewidth}
\includegraphics[width=.99\columnwidth]{nn_vae_biased_both.png}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\includegraphics[width=.99\columnwidth]{nn_vae_biased_both2.png}
\end{minipage}
\end{center}
\caption{Modes for VAE with 8 nearest neighbors. Notice the higher prevalence of digit 1. Due to low pixel variations, digit 1 samples have high density on the manifold.}
\label{fig:nearest}
\end{figure}
\subsection{Polarity Helps Under Distribution Shift}
\label{sec:shift}
\begin{figure}[t!]
\centering
\begin{minipage}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{NVAE/biasedNVAE-biasedTest.pdf}\\[-0.5em]
{\small polarity ($\rho$)}
\end{minipage}
\begin{minipage}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{NVAE/biasedNVAE-unifTest.pdf}\\[-0.5em]
{\small polarity ($\rho$)}
\end{minipage}
\vspace{-0.2cm}
\caption{{\bf Left:} training of NVAE on color-MNIST with hue bias and computing FID, precision and recall with the test set of color-MNIST that has the same bias. {\bf Right:} reproduction of the left experiment but now using a uniform hue for the test set. We see that {\bf the polarity allow to adapt the DGN distribution to balance possible distribution shifts}.}
\label{fig:nvae}
\end{figure}
A last benefit of Polarity-sampling is to adapt a sampling distribution to a reference distribution that suffered a distribution shift. For example, this could occur when training a DGN on a training set, and using it for content generation with a slightly different type target samples. In fact, as long as the distribution shift remains reachable by the model i.e. in the support of $p_{G}$, altering the value of $\rho$ will help to shift the sampling distribution, possible to better match the target one. In all generality, there is no guarantee that any benefit would happen for $\rho \not = 0$, however, for the particular case where the distribution shift only changes the way the samples are distributed (on the same domain), we observe in \cref{fig:nvae} that $\rho \not = 0$ can provide benefit. To control the experimental setting, we took the color-MNIST dataset and the NVAE DGN model \cite{vahdat2020nvae} and produce a training set with a Gaussian hue distribution favoring blue
and two test set, one with same hue distribution and one with uniform hue distribution. We observe that $\rho$ can provide a beneficial distribution shift to go from the biased-hue samples to the uniform-hue one.
\subsection{ProGAN Polarity Sweep}
\label{sec:progan}
Previously in \cref{sec:pareto}, we have drawn note to the fact that ProGAN \cite{karras2017progressive}, an architecture which is widely used, but does not incorporate truncation, can also be controlled via polarity sampling. In \cref{tab:progan} we present precision-recall characteristics for polarity sweep on ProGAN. As control, we also perform latent space truncation as in \cite{brock2018large} by sampling a truncated gaussian distribution, parameterized by its support $[-\beta,\beta]$. We change $\beta$ between $[10^{-10},10]$ and notice that for $\beta$ smaller than $10^{-4}$, the generator collapses to 0 precision and recall. Other than that, it maintains a precision of $0.72$ and recall of $0.34$. Using polarity sweep, we also exceed the baseline FID on CelebAHQ 1024x1024 attained by ProGAN; polarity of $-.01$ in pixel-space reduces the FID from $7.37$ to $7.28$.
\begin{table}[]
\centering
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}ccccccc@{}}
\toprule
& \multicolumn{6}{r}{CelebAHQ 1024$\times$1024} \\ \midrule
\multicolumn{1}{l}{} & \multicolumn{3}{c}{$\rho \leq 0$} & \multicolumn{3}{c}{$\rho > 0$} \\
$|\rho|$ & FID & Precision & \multicolumn{1}{c|}{Recall} & FID & Precision & Recall \\ \midrule
0 & 7.37 & .73 & .34 & - & - & - \\
0.01 & \textbf{7.28} & .73 & .34 & 7.45 & .73 & .35 \\
0.1 & 7.41 & .76 & .31 & 8.95 & .68 & .38 \\
1 & 12.65 & .85 & .19 & 17.96 & .58 & .48 \\
2 & 13.09 & \textbf{.86} & .19 & 18.54 & .58 & \textbf{.48} \\
\hline
\end{tabular}
}
\caption{FID, Precision and Recall metrics of ProGAN \cite{karras2017progressive} with polarity sweep in the pixel space.}
\label{tab:progan}
\end{table}
\section{Dataset description}
\label{sec:datasets}
\subsection{colored-MNIST}
\label{sec:hue_dataset}
We perform controlled experiments on NVAE \cite{vahdat2020nvae} by training on datasets with and without controllable distribution shifts. To control the shift, we colorize MNIST with hue ranging $[0,\pi]$ by 1) uniformly sampling the hue and 2) sampling the hue for each image from a truncated normal distribution, with a truncation scale of $2$.
\subsection{LSUN Dataset}
We use the LSUN dataset \cite{yu2015lsun} available at the official website\footnote{https://www.yf.io/p/lsun}. We preprocess the dataset using the StyleGAN2 repository\footnote{https://github.com/NVlabs/stylegan2}.
\subsection{AFHQv2 and FFHQ}
We use the version 2 of AFHQ that was released prepackaged with StyleGAN3 for our experiments. For FFHQ we use also use TFRecords provided with StyleGAN2.
\subsection{License}
NVAE, StyleGAN2 and StyleGAN3 are released under Nvidia Source Code License-NC. The ADM repository\footnote{https://github.com/openai/guided-diffusion/blob/main/LICENSE} is licensed under MIT license.
\section{Qualitative results}
\label{sec:quality}
We provide in this section \cref{fig:qualitative_alpha_sweep,fig:qualitative_alpha_sweep_cars,fig:qualitative_alpha_sweep_afhq} that correspond to LSUN Cats, LSUN Cars and AFHQv2 samples with varying polarity $\rho$ values.
\begin{figure*}[t!]
\centering
\begin{center}
\begin{minipage}{0.97\textwidth}
\centering
\includegraphics[width=.86\linewidth]{sweeps/cats_trunc_pol_matched_blurred.png}
\end{minipage}\\
\begin{minipage}{0.97\textwidth}
\centering
\includegraphics[width=.86\linewidth]{sweeps/cat_trunc_sweep_blurred.png}
\end{minipage}
\end{center}
\vspace{-0.2cm}
\caption{Uncurated samples of LSUN Cats using (top) $\rho=\{-1,-.5,-.2,-.1,0\}, \psi=.8$ and (bottom) $\psi=\{.7,.73,.75,.77,.8\}$; both representing regions with roughly an equal span of recall score on \cref{fig:pareto}. Notice the significant precision of the left-most columns of top compared to the left-most of bottom, where at equal diversity, top has significantly higher precision score.}
\label{fig:qualitative_alpha_sweep}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{center}
\begin{minipage}{\textwidth}
\includegraphics[width=\linewidth]{sweeps/lsun_car_equiprec_polsweep.png}
\end{minipage}\\
\begin{minipage}{\textwidth}
\includegraphics[width=\linewidth]{sweeps/lsun_car_equiprec_truncsweep.png}
\end{minipage}
\end{center}
\vspace{-0.2cm}
\caption{Uncurated samples of LSUN Cars using (top) $\rho=\{-.1,-.075,-.05,-.025,0\}, \psi=.7$ and (bottom) $\psi=\{.5,.55,.6,.65,.7\}$; both representing regions with roughly an equal span of precision score on \cref{fig:pareto} Notice the significant diversity in (top) especially in the leftmost columns, where the recall score is significantly higher than that of the leftmost column of bottom left, with equal precision.
}
\label{fig:qualitative_alpha_sweep_cars}
\end{figure*}
\begin{figure*}[t!]
\centering
\begin{center}
\begin{minipage}{0.97\textwidth}
\centering
\includegraphics[width=0.93\linewidth]{sweeps/afhqalphasweep.png}
\end{minipage}
\end{center}
\vspace{-0.2cm}
\caption{Uncurated samples of AFHQv2 using $\rho=\{-2,-1,-.5,-.2,0,.01,.1,.2,.5\}$ and $\psi=.9$ in pixel-space. As we move right from baseline (middle column) we see an increase in texture diversity of images, whereas, moving left, we see images with smoother textures.
}
\label{fig:qualitative_alpha_sweep_afhq}
\end{figure*}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.8\linewidth]{modes/imagenet_blurred.png}
\caption{Depiction of a single mode (large negative $\rho$) for each class of the first $800$ Imagenet classes.}
\label{fig:modes_per_class}
\end{figure*} | a10cf1152f51f11a5f9d2444a5cc1ed875dbcc20 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The quasi-geostrophic shallow-water (QGSW) equation \cite{CaCr13} is a transport equation in two space dimensions for an active scalar $q(t, \mathbf{x})$, $\mathbf{x} = (x, y)$,
\begin{equation}
\left\{
\begin{aligned}
\label{QGSW}
\begin{split}
& q_t + \u \cdot \nabla q = 0,\\
&\u=\nabla^\perp \Psi,\\
&-\Delta\Psi+F\Psi= q,
\end{split}
\end{aligned}\right.
\end{equation}
where $q$ is the potential vorticity, $\u$ is the velocity, $\Psi$ is the stream function, and $F$ is a positive constant.
This equation describes the dynamics of a single fluid layer of ocean or atmosphere in the mid-latitude region.
On the earth, the horizontal scale of the ocean or the atmosphere is about the order of thousands kilometers, which is much larger than the vertical scale, which is around $10$km. The time scale for fluid motion is same or longer than the scale of the earth rotation. In this situation, shallow water approximation can provide a reasonable reduction. This equation is also known as the Charney equation \cite{Charney48} in some geophysical context \cite{Ped87}, and its mathematical justification can be found in Majda’s book \cite{Maj03}. Another interesting point is that the same equation is also derived from the plasma physics \cite{HM78} and it is called Hasegawa-Mima equation in the plasma context, for example \cite{CNQ15}.
The mathematical structure of QGSW equation is close to incompressible Euler equation, which is the case when $F=0$. Its local and global wellposedness of weak and strong solutions in $\mathbb{R}^2$ has been proved in \cite{Pau04, GH04} in the suitable Sobolev spaces. In these results, the vanishing far-field limit of potential velocity $q$ is a requirement. This contrasts the result we will present below in this paper, where the far-field limit of $q$ is nonzero.
The main content of the current paper is to study a family of piecewise constant solutions for QGSW equation. This kind of solutions has been studied for several models related to 2D incompressible fluid. These models share the similar form of governing equations:
\begin{equation}\label{vorEq}
\left\{~\begin{aligned}
&\theta_t+\mathbf{u}\cdot\nabla\theta=0,\\
& \mathbf{u}=\nabla^\perp \Psi,\\
& \mathbf{A}\Psi=\theta,
\end{aligned}\right.
\end{equation}
where $\theta$ is the (potential) vorticity, $\mathbf{u}$ is the velocity, and $\Psi$ is the stream function. In the third equation, $\mathbf{A}$ is an elliptic operator connecting the (potential) vorticity and the stream function.
\begin{itemize}
\item When $\mathbf{A}=-\Delta$, it is the incompressible Euler equation. \item When $\mathbf{A}=(-\Delta)^{1/2}$, it is the surface quasi-geostrophic (SQG) equation.
\item When $\mathbf{A}=(-\Delta)^{\alpha/2}$ with $\alpha\in(0,2]$, it is the generalized surface quasi-geostrophic (GSQG) equation.
\item When $\mathbf{A}=F-\Delta$, where $F$ is a positive constant, it is the quasi-geostrophic shallow water (QGSW) equation.
\end{itemize}
In all these models, the (potential) vorticity satisfies a transport equation. Therefore when the initial data is a piecewise constant function, it will remain so for all time. This is the main reason that we can study the piecewise constant solutions for these equations. When we study these solutions, we usually treat them as a free boundary problem for the curves of the vorticity discontinuities transporting along the fluid flows.
According to the different geometries of these curves, we classify these problems as the vortex patch problem and the vortex front problem separately. We use the definitions in \cite{HSZ20b} to define the patch solution and the front solution.
Consider the piecewise constant solutions of \eqref{vorEq}
\begin{align}
\label{piecewisetheta}
\theta(t, \mathbf{x}) = \sum_{k = 1}^N \theta_k \mathds{1}_{\Omega_k(t)}(\mathbf{x}),
\end{align}
where $N \geq 2$ is a positive integer, $\theta_1, \dotsc, \theta_N\in \mathbb{R}$ are constants, and $\Omega_1(t),\dotsc, \Omega_N(t) \subset \mathbb{R}^2$ are disjoint domains such that
\[
\bigcup_{k = 1}^N \overline{\Omega_k(t)} = \mathbb{R}^2,
\]
and their boundaries $\partial \Omega_1(t), \dots, \partial \Omega_N(t)$ are smooth curves, whose components either coincide or are a positive distance apart. In \eqref{piecewisetheta}, $\mathds{1}_{\Omega_k(t)}$ denotes the indicator function of $\Omega_k(t)$.\\
\eqref{piecewisetheta} is a {\it patch solution} if it satisfies the following assumptions:
\begin{enumerate}
\item $N \geq 2$;
\item $\theta_N = 0$, but $\theta_k \in \mathbb{R} \setminus \{0\}$ for each $1 \leq k \leq N - 1$;
\item for each $1 \leq k \leq N - 1$, the region $\Omega_k(t)$ is bounded, and its boundary $\partial \Omega_k(t)$
is a smooth, simple, closed curve that is diffeomorphic to the circle $\mathbb{T}$;
\item the region $\Omega_N(t)$ is unbounded.
\end{enumerate}
\eqref{piecewisetheta} is an {\it (N-1)-front solution} if it satisfies the following assumptions:
\begin{enumerate}
\item $N \geq 2$;
\item $\theta_1,\cdots, \theta_N \in \mathbb{R}$ are constants such that $\theta_i\neq \theta_{i+1}$ for $i=1,\cdots, N-1$;
\item Each $\Omega_i(t)$, $i=1,\cdots, N,$ is unbounded and each of their boundaries is a simple, smooth curve diffeomorphic to $\mathbb{R}$.
\end{enumerate}
{\noindent\bf Vortex patch problems.}
For incompressible Euler equation, the contour dynamics equation is derived by Zabusky et.~al.~\cite{Zabusky}. The boundary of a vortex patch remains smooth globally in time \cite{BC93, Che93, Che98}. Some special types of nontrivial global-in-time smooth vortex patch solutions are constructed in \cite{Bur82, CCGS16b, dlHHMV16, HM16, HM17, HMV13}.
For SQG and generalized SQG equation, local well-posedness of the contour dynamics equations is proved in \cite{CCG18, Gan08, GP21, CCCGW12}. The question of whether finite-time singularities can form in smooth boundaries of SQG or GSQG patches remains open.
Numerical simulation suggests that the boundary of the SQG patch may develop self-similar singularities and provides evidence that two separated SQG patches can touch in finite time \cite{SD14, SD19}.
It is proved in \cite{GS14} that splash singularities cannot form. Some nontrivial global solutions for SQG and GSQG patches have been constructed in \cite{CCGS16a, dlHHH16, GS19, HH15, HM17, GPSY19p}. Local existence of the vortex patch problem for singular SQG equation is studied in \cite{KR20a, KR20b}.
When a rigid boundary is present, the local existence of smooth GSQG patches is shown in \cite{GP21, KRYZ16, KYZ17, JZ21p}, and the formation of finite-time singularities is proved for a range of $\alpha$ close to $2$. By contrast, vortex patches in this setting (with $\alpha=2$) have global regularity \cite{KRYZ16}.
For QGSW equation, the vortex patch problem has been studied in \cite{HR21, JD20, PD12, PD13, DHC19}.
\vspace{0.2cm}
{\noindent\bf Vortex front problems.}
For incompressible Euler equation, the vortex front problem is studied in \cite{BH10, Ray1895}, and the approximate models of the Euler front is justified in \cite{HSZ22}.
For GSQG equation with $\alpha\in[1,2]$, the vortex front equations are derived by a regularization procedure in \cite{HS18}. The SQG front equation is derived by a perturbation method in \cite{HSZ20a}.
For SQG equation, local existence and uniqueness for spatially periodic fronts are proved for $C^\infty$ initial data in \cite{Rod05} and analytic initial data in \cite{FR11}. Local well-posedness in Sobolev spaces for spatially periodic solutions of a cubically nonlinear approximation of the SQG front equation is proved in \cite{HSZ18}. Almost sharp SQG fronts are studied in \cite{CFR04, FLR12, FR12, FR15}, and smooth $C^\infty$ solutions for spatially periodic GSQG fronts with $1 < \alpha < 2$ also exist locally in time \cite{CFMR05}.
In the non-periodic setting, smooth solutions to the GSQG front equations with $0 < \alpha < 1$ on $\mathbb{R}$ are shown to exist globally in time for small initial data in \cite{CGI19}.
The analogous result is proved for the SQG front equation with $\alpha = 1$ in \cite{HSZ21}, and for GSQG fronts with $1<\alpha<2$ in \cite{HSZ20p}. The two-front problem of GSQG equation for different values of $\alpha\in(0,2]$ is studied in \cite{HSZ20b} and local well-posedness is proved.
\vspace{0.2cm}
{\noindent\bf QGSW front problem.}
In the current paper, we will derive the equation of the QGSW front. We consider a family of piecewise constant solutions of the form
\begin{equation}
q(t, \mathbf{x}) = \begin{cases} q_+ & \text{if $y >\varphi(t,x)$},\\q_- &\text{if $y<\varphi(t,x)$},\end{cases}
\label{front}
\end{equation}
where $q_+$ and $q_-$ are two constants. The set of discontinuities of $q$ is $\{(x,y)\mid y=\varphi(t,x)\}$, which is called the QGSW front. Without loss of generality, we set $F=1$. We will show that the QGSW front satisfies the equation
\begin{equation}\label{QGSWfront}
\varphi_t(t,x)=\frac1\pi\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)\,\mathrm{d} \zeta,
\end{equation}
where $K_0(\cdot)$ is the modified Bessel function of the second kind. After simplifying this equation, we will see that this is a nonlinear, nonlocal dispersive equation. We will mainly study the local and global existence of solutions.
Since the equation \eqref{QGSWfront} is a quasilinear dispersive equation, the local existence and uniqueness of solutions in Sobolev space follow from Kato's theory \cite{Kat75a, Kat75b}. The idea is to use the linearized equation to construct a sequence of approximate solutions. Then using the uniform bound of the sequence in the higher regularity space and the contraction in the lower regularity space to obtain a convergent limit. This procedure is similar as the SQG front in \cite{HSZ21}.
When proving global existence, we need the dispersive estimates.
This can be achieved by the method of space-time resonances introduced by Germain, Masmoudi and Shatah \cite{Germain,GMS09, GMS12}, together with estimates for weighted $L^\infty_\xi$-norms --- the so-called $Z$-norms --- developed by Ionescu and his collaborators \cite{CGI19, DIP17, DIPP17, IP12, IP13, IP14, IP15, IPu16}.
One may have two additional difficulties compared with the previous results of the GSQG front problem \cite{HSZ21, CGI19, HSZ20p}.
First, the dispersive relation $p(\xi)=-\xi(1+\xi^2)^{-1/2}$ satisfies $p''(0)=0$ and $p'''(0)\neq 0$, which implies that the frequency $\xi=0$ is a stationary phase and it will leads to slower dispersive decay rate, approximately $O(t^{-1/3})$, when doing oscillatory integral estimate \cite{Ste93}. To obtain $O(t^{-1/2})$ decay, we require some extra derivative in the lower frequency. Fortunately, the structure of the nonlinear term can provide the extra derivative in the low frequency. This helps to gain $O(t^{-1/2})$ decay.
Another difficulty is that the dispersive relation $p(\xi)=-\xi(1+\xi^2)^{-1/2}$ does not satisfy scaling property. So there is no scaling vector field to use. To solve this problem, we need to introduce the profile function $\hat h(\xi)=e^{-ip(\xi)t}\hat \varphi(\xi)$, and estimate $\partial_\xi\hat h$ by using the profile equation. In this procedure, there may be loss of derivatives when estimating the nonlinear terms. It can be solved by symmetrizing the nonlinearities and using the cancellation property of the symbol.
The paper is organized as follows. In Section \ref{sec:prelim}, we list some facts and notations to be used in the later sections. In Section \ref{sec:Der}, we derive the front equation. In Section \ref{sec:Loc}, we prove the local well-posedness of the front equation. After that, we consider the global solutions in Sections \ref{sec:Glo} to \ref{sec:Znorm}. The global existence theorem is stated in Theorem \ref{Thm:glo}. Sections \ref{sec:Dis}, \ref{sec:WE}, \ref{sec:Znorm} are mainly for dispersive estimates, weighted energy estimates and the space-time resonance analysis, separately.
\section{Preliminaries}
\label{sec:prelim}
We denote the Fourier transform of $f \colon \mathbb{R}\to \mathcal{C}$ by $\hat f \colon \mathbb{R}\to \mathcal{C}$, where $\hat f= \mathcal{F} f$ is given by
\[
f(x)=\int_{\mathbb{R}} \hat f(\xi) e^{i\xi x} \,\mathrm{d}\xi, \qquad \hat f(\xi)=\frac1{2\pi} \int_{\mathbb{R}}f(x) e^{-i\xi x}\,\mathrm{d}{x}.
{\color{red}}
\]
For $s\in \mathbb{R}$, we denote by $H^s(\mathbb{R})$ the space of Schwartz distributions $f$ with $\|f\|_{H^s} < \infty$, where
\[
\|f\|_{H^s} = \left[\int_\mathbb{R} \left(1+|\xi|^2\right)^s |\hat{f}(\xi)|^2\, \,\mathrm{d}{\xi}\right]^{1/2}.
\]
Throughout this paper, we use $A\lesssim B$ to mean there is a constant $C$ such that $A\leq C B$, and $A\gtrsim B$ to mean there is a constant $C$ such that $A\geq C B$. We use $A\approx B$ to mean that $A\lesssim B$ and $B\lesssim A$.
Let $\psi \colon\mathbb{R}\to [0,1]$ be a smooth function supported in $[-8/5, 8/5]$ and equal to $1$ in $[-5/4, 5/4]$.
For any $k\in \mathbb Z$, we define
\begin{align}
\label{defpsik}
\begin{split}
\psi_k(\xi)&=\psi(\xi/2^k)-\psi(\xi/2^{k-1}), \qquad \psi_{\leq k}(\xi)=\psi(\xi/2^k),\qquad \psi_{\geq k}(\xi)=1-\psi(\xi/2^{k-1}),\\
\tilde\psi_k(\xi)&=\psi_{k-1}(\xi)+\psi_k(\xi)+\psi_{k+1}(\xi).
\end{split}
\end{align}
So we have the homogeneous dyadic decomposition
\[
1=\sum\limits_{k\in\mathbb Z}\psi_k(\xi),
\]
and the non-homogeneous dyadic decomposition
\[
1=\sum\limits_{k\in\mathbb N}\psi_k(\xi)+\psi_{\leq 0}(\xi).
\]
We denote the sub-index of the non-homogeneous dyadic decomposition as $\{\leq 0\}\cup \mathbb N$.
We denote by $P_k$, $P_{\leq k}$, $P_{\geq k}$, and $\tilde{P}_k$ the Fourier multiplier operators with symbols $\psi_k, \psi_{\leq k}, \psi_{\geq k}$, and $\tilde{\psi}_k$, respectively. Notice that $\psi_k(\xi)=\psi_0(\xi/2^k)$, $\tilde\psi_k(\xi)=\tilde\psi_0(\xi/2^k)$.
It is easy to check that
\begin{equation}
\label{psi-L2}
\|\psi_k\|_{L^2}\approx 2^{k/2}, \qquad \|\psi_k'\|_{L^2}\approx 2^{-k/2}.
\end{equation}
We will need the following interpolation lemma, whose proof can be found in \cite{IPu16}.
\begin{lemma}\label{interpolation}
For any $k\in\mathbb Z$ and $f\in L^2(\mathbb{R})$, we have
\[
\|\widehat{P_kf}\|_{L^\infty}^2\lesssim \|P_k f\|_{L^1}^2\lesssim 2^{-k}\|\hat f\|_{L^2_\xi}\left[2^k\|\partial_\xi\hat f\|_{L^2_\xi}+\|\hat f\|_{L^2_\xi}\right].
\]
\end{lemma}
Define the symbol class
\[
S^\infty:=\{m: \mathbb{R}^d\to\mathcal{C}: m \text{ is continuous and } \|m\|_{S^\infty}:=\|\mathcal{F}^{-1}(m)\|_{L^1}<\infty\}.
\]
\begin{lemma}[Algebraic property of $S^\infty$]
If $m$, $m'\in S^\infty$, then $m\cdot m'\in S^\infty$.
\end{lemma}
Given $m \in S^\infty(\mathbb{R}^{n})$, we define a multilinear operator $M_m$ acting on Schwartz functions $f_1,\dots, f_m \in \mathcal{S}(\mathbb{R})$ by
\[
M_{m}(f_1,\dotsc,f_n)(x)=\int_{\mathbb{R}^{n}} e^{ix(\xi_1+\dotsb+\xi_n)}m(\xi_1, \dotsc, \xi_n)\hat f_1(\xi_1) \dotsm \hat f_n(\xi_n)\,\mathrm{d}{\xi_1} \dotsm \,\mathrm{d}{\xi_n}.
\]
\begin{lemma}[Estimates of multilinear Fourier integral operators]\label{multilinear}
(i)~ Suppose that $1<p_1, \dotsc, p_n\leq \infty$, $0<p<\infty$, satisfy
\[
\frac1{p_1}+\frac1{p_2}+ \dotsb +\frac1{p_n}=\frac1p,
\]
for every nonempty subset $J\subset \{1,2, \dotsc, n\}$. If $m \in S^\infty(\mathbb{R}^{n})$, then
\[
\|M_{m}\|_{L^{p_1}\times \dotsb \times L^{p_n}\to L^p}\lesssim \|m\|_{S^\infty}.
\]
(ii)~ Assume $p,q,r\in[1,\infty]$ satisfy $\frac1p+\frac1q=\frac1r$, and $m\in S^\infty$. Then for any $f,g\in L^2(\mathbb{R})$,
\[
\left|\int_{\mathbb{R}^2}m(\xi,\eta)\hat f(\xi)\hat g(\eta)\hat h(-\xi-\eta)\,\mathrm{d} \xi\,\mathrm{d}\eta\right|\lesssim \|m\|_{S^\infty}\|f\|_{L^p}\|g\|_{L^q}\|h\|_{L^r}.
\]
\end{lemma}
We define the $B^{a,b}$ semi-norm as
\begin{equation}\label{bnorm}
\|f\|_{B^{a,b}}=\sum\limits_{j\in\mathbb{Z}}(2^{aj}+2^{bj})\|P_jf\|_{L^\infty}.
\end{equation}
Here $a<b$ and $a,b$ are the indices of higher and lower frequencies. Compared with the $L^\infty$ norm, $B^{a,b}$ satisfies one property not shared with the $L^\infty$ norm, that is
\[
\|f\|_{B^{a,b}}\approx \sum\limits_{j\in\mathbb{Z}}\|P_jf\|_{B^{a,b}}.
\]
\subsection{Some facts about the function $K_0$}
We list some properties of the function $K_0$ here for convenience. They are from the handbook \cite{OLBC}.
The modified Bessel function of the second kind $K_0(x)$ is one solution of the equation
\begin{equation}\label{BesselEq}
xy''+ y'-xy=0,
\end{equation}
which satisfies
\[
K_0(x)\sim \sqrt{\frac{\pi}{2x}}e^{-x}, \quad \text{ when } x\to \infty.
\]
It satisfies the identities
\[
K_0(x)
=
\frac12
\int_{\mathbb{R}}
\frac{e^{i\xi x}}{\sqrt{1+\xi^2}}
\,\mathrm{d}\xi
=
\frac12
\mathcal{F}^{-1}\left(\frac{1}{\sqrt{1+\xi^2}}\right),\quad \text{and} \quad
K_0(x)=\int_0^\infty e^{-x\cosh t} \,\mathrm{d} t.
\]
The power series of $K_0$ around $0$ is
\begin{align
\label{K0-near0}
K_0(x)
&=-\left[\ln\left(\frac{|x|}2\right)+\gamma\right]I_0(x)+\sum\limits_{k=1}^\infty b_kx^{2k},
\end{align}
where $\gamma=0.577\cdots$ is Euler's constant and
\[
b_k=(\sum\limits_{i=1}^k \frac1i)(k!)^{-2}4^{-k}, \quad I_0(x)=\sum\limits_{k=0}^\infty \frac{1}{(k!)^2 4^k} x^{2k}.
\]
Its derivative satisfies
\[
K_0'(x)=-\int_0^\infty (\cosh t) e^{-x\cosh t} \,\mathrm{d} t=-e^{-x}\int_0^\infty (\cosh t) e^{x(1-\cosh t)} \,\mathrm{d} t.
\]
When $x>1$,
\[
0<\int_0^\infty (\cosh t) e^{x(1-\cosh t)} \,\mathrm{d} t\leq \int_0^\infty (\cosh t) e^{(1-\cosh t)} \,\mathrm{d} t,
\]
where the integral $\int_0^\infty (\cosh t) e^{(1-\cosh t)} \,\mathrm{d} t$ is convergent and it is a constant independent with $x$. Therefore
\[
|K_0'(x)|\lesssim e^{-x}, \quad \text{when } \quad x>1.
\]
Since $K_0$ and $K_0'$ both satisfy the exponential bound, by the equation \eqref{BesselEq}, $K_0''$ also satisfies $|K_0''(x)|\lesssim e^{-x}$ when $x>1$. Taking n-th order derivative to \eqref{BesselEq}, we have
\[
\frac{\,\mathrm{d}^n}{\,\mathrm{d} x^n}(xy''(x))=\frac{\,\mathrm{d}^n}{\,\mathrm{d} x^n}[xy(x)]-y^{(n+1)}(x),
\]
which leads to
\[
xy^{(n+2)}(x)=-(n+1)y^{(n+1)}(x)+xy^{(n)}(x)+ny^{(n-1)}(x).
\]
By induction, we conclude
\begin{equation}\label{decayK0}
|K_0^{(n)}(x)|\lesssim O(n!)e^{-x}, \quad \text{when } \quad x>1.
\end{equation}
\section{The QGSW front equation}\label{sec:Der}
In this section we derive the QGSW front equation. Assume the domain
\begin{equation}
\label{defomega}
\Omega_t^\pm = \{(x,y)\in \mathbb{R}^2 : y \gtrless \varphi(t,x)\}
\end{equation}
is an upper/lower half-space whose boundary is the graph of the function $y=\varphi(t,x)$, and
\begin{equation}
q(t, \mathbf{x}) = \begin{cases} q_+ & \text{if $y >\varphi(t,x)$},\\q_- &\text{if $y<\varphi(t,x)$}.\end{cases}
\label{sqg_front}
\end{equation}
Assume the initial data is denoted as
\begin{equation}
q_0(\mathbf{x}) = \begin{cases} q_+ & \text{if $y >\varphi_0(x)$},\\q_- &\text{if $y<\varphi_0(x)$}.\end{cases}
\end{equation}
We take $F=1$. We mainly care about the solution of QGSW equation\eqref{QGSW} when the front $\varphi(t,x)$ is close to the $x$-axis.
\subsection{Steady state solutions}
As a start, we consider a family of solutions which are translation-invariant in $x$-direction. The corresponding front $\varphi(t, x)\equiv 0$. By removing the $x$-derivative in the system, we can rewrite the system as
\begin{equation}
\begin{aligned}
-\partial_y^2\Psi+\Psi=q_\pm, \quad& \text{ in } y\gtrless0,\\
\u_s=(u_s, v_s)^T=(-\partial_y\Psi,0)^T, \quad& \text{ in } \mathbb{R}^2.
\end{aligned}
\end{equation}
By solving the above equation, we have the general solution
\[
\Psi(y)=c_1^\pm e^y+c_2^\pm e^{-y}+q_\pm, \quad \text{ in } y\gtrless0,
\]
where $c_1^\pm$ and $c_2^\pm$ are constants. These four constants are not independent. Across the front, the velocity is continuous.
So we propose the boundary conditions $\Psi_+=\Psi_-$ and $\partial_y\Psi_+=\partial_y\Psi_-$ on $y=0$, which lead to
\begin{equation}
\left\{\begin{array}{ll}
c_1^++c_2^++q_+=c_1^-+c_2^-+q_-,\\
c_1^+-c_2^+=c_1^--c_2^-.
\end{array}\right.
\end{equation}
If we assume the velocity is bounded at $\pm\infty$, we can furthermore set $c_1^+=c_2^-=0$. Then $c_1^-=-c_2^+=\frac12 \mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}$, where $\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}:=q_+-q_-$. Thus
\[
\Psi(y)=\left\{
\begin{array}{ll}
\displaystyle-\frac12\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]} e^{-y}+q_+,\quad y>0,\cr\cr
\displaystyle\frac12\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]} e^{y}+q_-,\quad y<0.
\end{array}
\right.
\]
The horizontal velocity is $u_s(y)=-\partial_y\Psi(y)=-\displaystyle\frac12\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]} e^{-|y|}$.
\subsection{Equation of the front}
We derive the equation of the front when the front is the graph of a function which is close to the steady state $y=0$.
Denote the Green's function for the elliptic operator $-\Delta +1$ as $G(x,y)$, defined as
\[
G(x,y)=\iint_{\mathbb{R}^2} \frac1{1+\xi^2+\eta^2}e^{i(x\xi+y\eta)}\,\mathrm{d} \xi\,\mathrm{d}\eta.
\]
To write the function and the integral into the polar coordinate, we denote $\xi=\rho\cos \alpha, \eta=\rho\sin\alpha$, $x=r\cos\theta, y=r\sin\theta$.
Then
\begin{align*}
G(r\cos\theta, r\sin\theta)&=\int_{0}^\infty\frac\rho{1+\rho^2}\int_0^{2\pi}e^{i\rho r\cos(\theta-\alpha)}\,\mathrm{d} \alpha \,\mathrm{d} \rho\\
&=\int_{0}^\infty\frac\rho{1+\rho^2}\int_0^{2\pi}e^{i\rho r\cos \alpha}\,\mathrm{d} \alpha \,\mathrm{d} \rho\\
&=\int_{0}^\infty\frac\rho{1+\rho^2}\int_0^{2\pi}\cos(\rho r\cos \alpha)\,\mathrm{d} \alpha \,\mathrm{d} \rho\\
&=2\pi\int_{0}^\infty\frac\rho{1+\rho^2}\left(\frac1{\pi}\int_0^{\pi}\cos(\rho r\cos \alpha)\,\mathrm{d} \alpha\right) \,\mathrm{d} \rho\\
&=2\pi\int_{0}^\infty\frac\rho{1+\rho^2}J_0(\rho r) \,\mathrm{d} \rho=2\pi K_0(r),
\end{align*}
where $J_0$ is the Bessel function of the first kind.
Notice that $J_0(z)\to 1$ as $z\to 0$, and $J_0(z)=\sqrt{2/(\pi z)}\big(\cos(z-\frac\pi4)+o(1)\big)$ as $z\to\infty$. The above integral converges. In fact, this is the Hankel transform of $1/(1+\rho^2)$, which equals to $K_0(r)$, the modified Bessel function of the second kind.
Using the facts that $K_0(r)\sim \sqrt{\pi/(2r)}e^{-r}$ as $r\to \infty$, and $K_0(r)\sim -\ln r$ as $r\to 0+$, we can obtain
\[
\Psi(x,y,t)=q_-\iint_{y'<\varphi(t, x')}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y'+q_+\iint_{y'>\varphi(t, x')}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y',
\]
in which the integral converges.
\noindent Denote the normal vector $\mathbf{n}=\frac{(-\varphi_x, 1)}{\sqrt{1+\varphi_x^2}}$. The front $\varphi(t,x)$ satisfies
\[
\varphi_t=\sqrt{1+\varphi_x^2}\mathbf{n}\cdot\u=(-\varphi_x,1)\cdot\u.
\]
Here the velocity
\[
\u=\nabla^\perp\Psi=\left(\begin{aligned}
q_-\iint_{y'<\varphi(t, x')}\partial_{y'}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y'+q_+\iint_{y'>\varphi(t, x')}\partial_{y'}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y'\\
q_-\iint_{y'<\varphi(t, x')}-\partial_{x'}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y'-q_+\iint_{y'>\varphi(t, x')}\partial_{x'}G(x-x',y-y')\,\mathrm{d} x'\,\mathrm{d} y'
\end{aligned}\right).
\]
By Green's theorem,
\begin{align*}
\u&=\left(\begin{aligned}
q_-\int_{\mathbb{R}}G(x-x',y-\varphi(t, x'))\,\mathrm{d} x'-q_+\int_{\mathbb{R}}G(x-x',y-\varphi(t, x'))\,\mathrm{d} x'\\
q_-\int_{\mathbb{R}}\varphi_{x'}(t, x')G(x-x',y-\varphi(t, x'))\,\mathrm{d} x'-q_+\int_{\mathbb{R}}\varphi_{x'}(t, x')G(x-x',y-\varphi(t, x'))\,\mathrm{d} x'
\end{aligned}\right)\\
&=-\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]} \int_{\mathbb{R}}G(x-x',y-\varphi(t, x'))(1, \varphi_{x'}(t, x'))^T\,\mathrm{d} x'.
\end{align*}
Therefore, the equation of $\varphi$ is
\begin{align*}
\varphi_t(t,x)&=\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x'}(t, x'))G(x-x', \varphi(t, x)-\varphi(t, x'))\,\mathrm{d} x'\\
&=2\pi\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x'}(t, x'))K_0(\sqrt{(x-x')^2+(\varphi(t, x)-\varphi(t, x'))^2})\,\mathrm{d} x'\\
&=2\pi\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))K_0(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})\,\mathrm{d} \zeta.
\end{align*}
The equation is a nonlinear nonlocal equation. Without loss of generality, we assume $\mathrlap{[}{\hspace{0.15em}[\hspace{0.15em}} q\hspace{0.15em}\mathrlap{]}{\hspace{0.15em}]}=1/\pi$
in the following, and the equation is written as
\[
\varphi_t(t,x)=2\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))K_0(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})\,\mathrm{d} \zeta.
\]
\subsection{Structure of the equation}
To investigate the structure of the right-hand-side of the above equation, we split it into the linear and nonlinear terms:
\[
\varphi_t=\mathcal{L}(\varphi)+\partial_x\mathcal{N}(\varphi),
\]
with
\begin{align*}
\mathcal{L}(\varphi)&:=2\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))K_0(\zeta)\,\mathrm{d} \zeta=c\varphi_x(t,x)-\frac1\pi\int_{\mathbb{R}}\varphi_{x'}(t, x')K_0({x'-x})\,\mathrm{d} x',\\
c&:=2\int_{\mathbb{R}}K_0(r)\,\mathrm{d} r=2\pi,
\\
\partial_x\mathcal{N}(\varphi)&:=
2\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))\left[K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\right]\,\mathrm{d} \zeta.
\end{align*}
Here we have used the fact that $K_0$ is an even function, so the absolute value in $K_0$ can be removed. Since
\[
K_0(x)
=
\frac{1}{2}\int_{\mathbb{R}}\frac{1}{\sqrt{1+\xi^2}}e^{ix\xi}\,\mathrm{d}\xi
=
\frac12
\mathcal{F}^{-1}\left(\frac{1}{\sqrt{1+\xi^2}}\right),
\]
we can rewrite the linear term as
\[
\mathcal{L}(\varphi)=2\pi\varphi_x- \frac{\partial_x}{\sqrt{1-\partial_x^2}}\varphi.
\]
$\mathcal{L}$ is a linear differential operator with symbol $2\pi i\xi-\frac{i\xi}{\sqrt{1+\xi^2}}$.
By a coordinate transform from $(t,x)$ to $(t, x-2\pi t)$, one can remove the linear transport term $2\pi \varphi_x$.
In the new coordinate, the equation is written as
\begin{equation}\label{FrEq}
\varphi_t=-\frac{\partial_x}{\sqrt{1-\partial_x^2}}\varphi+\partial_x\mathcal{N}(\varphi).
\end{equation}
\subsection{Expansion of the nonlinear term}
For the nonlinear term $\mathcal{N}(\varphi)$, split it into two terms $\mathcal{N}_L(\varphi)$ and $\mathcal{N}_S(\varphi)$ with
\[
\partial_x\mathcal{N}_L(\varphi)=2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))\left[K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\right]\,\mathrm{d} \zeta,
\]
\[
\partial_x\mathcal{N}_S(\varphi)=2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))\left[K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\right]\,\mathrm{d} \zeta,
\]
which correspond to the long-range interaction and the short-range interaction separately. Here $\chi_{\{|\zeta|>1\}}$ and $\chi_{\{|\zeta|<1\}}$ are the smooth cut-off functions of the sets $\{|\zeta|>1\}$ and $\{|\zeta|<1\}$ separately, and $\chi_{|\{\zeta|>1\}}+\chi_{\{|\zeta|<1\}}=1$.
{\bf 1.} By Taylor expansion,
\begin{align*}
&K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\\
=~&\sum\limits_{k=1}^\infty \frac1{k!}K_0^{(k)}(\zeta)\left[\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}-|\zeta|\right]^k.
\end{align*}
Since
\begin{align*}
\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}-|\zeta|
&=-|\zeta|\left(1-\sqrt{1+\frac{(\varphi(t,x)-\varphi(t, x+\zeta))^2}{\zeta^2}}\right)\\
&=|\zeta|\sum\limits_{l=1}^\infty \left(\begin{array}{c}\frac12\cr\cr l
\end{array}\right)\left|\frac{\varphi(t,x)-\varphi(t, x+\zeta)}{\zeta}\right|^{2l},
\end{align*}
we can write
\begin{align*}
&\sum\limits_{k=1}^\infty\frac1{k!}K_0^{(k)}(\zeta)\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}-|\zeta|\right)^k\\
=&\sum\limits_{k=1}^\infty\frac1{k!}K_0^{(k)}(\zeta)|\zeta|^k\left(\sum\limits_{l=1}^\infty \left(\begin{array}{c}\frac12\cr\cr l
\end{array}\right)\left|\frac{\varphi(t,x)-\varphi(t, x+\zeta)}{\zeta}\right|^{2l}\right)^k\\
=&\sum\limits_{k=1}^\infty\frac1{k!}K_0^{(k)}(\zeta)|\zeta|^k \left(\sum\limits_{|\vec{i}|=k}^\infty \left(\begin{array}{c} k\\ \vec{i}
\end{array}\right)\prod\limits_{l=1}^\infty \left(\begin{array}{c}\frac12\\ l
\end{array}\right)^{i_l}\left|\frac{\varphi(t,x)-\varphi(t, x+\zeta)}{\zeta}\right|^{2i_ll}\right)\\
=&\sum\limits_{k=1}^\infty\frac1{k!}K_0^{(k)}(\zeta)|\zeta|^k \left(\sum\limits_{|\vec{i}|=k}^\infty \left(\begin{array}{c} k\\ \vec{i}
\end{array}\right)\left[\prod\limits_{l=1}^\infty \left(\begin{array}{c}\frac12\\ l
\end{array}\right)^{i_l}\right]\left|\frac{\varphi(t,x)-\varphi(t, x+\zeta)}{\zeta}\right|^{2\sum\limits_{l=1}^{\infty}i_ll}\right)\\
=&\sum\limits_{\mu=1}^\infty A_{\mu}(\zeta)\left|\varphi(t,x)-\varphi(t, x+\zeta)\right|^{2\mu},
\end{align*}
where
\[
A_{\mu}(\zeta)=\sum\limits_{k=1}^\infty\frac1{k!}K_0^{(k)}(\zeta)|\zeta|^{k-2\mu} \left(\sum\limits_{|\vec{i}|=k, \sum i_ll=\mu}^\infty \left(\begin{array}{c} k\\ \vec{i}
\end{array}\right)\left[\prod\limits_{l=1}^\infty \left(\begin{array}{c}\frac12\\ l
\end{array}\right)^{i_l}\right]\right).
\]
We remark here that $k-2\mu\leq0$, and
\begin{align*}
\left(\begin{array}{c} k\\ \vec{i}
\end{array}\right)=\frac{k!}{i_1!i_2!\cdots i_l!\cdots},
\end{align*}
\begin{align*}
\left(\begin{array}{c} \frac12\\ l
\end{array}\right)=\frac{(-1)^{l-1}(2l-3)!!}{2^l l!}\in \left[-\frac18, \frac12\right].
\end{align*}
We observe that when $k>\mu$, there is no vector $\vec{i}$ such that $|\vec{i}|=k$, $\sum i_l l=\mu$. The summation in $A_\mu(\zeta)$ is a finite sum, thus it must be convergent.
Therefore,
\begin{align*}
K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)
=\sum\limits_{\mu=1}^\infty A_{\mu}(\zeta)\left|\varphi(t,x)-\varphi(t, x+\zeta)\right|^{2\mu}.
\end{align*}
We notice that by \eqref{decayK0}, $|A_{\mu}(\zeta)|\lesssim e^{-|\zeta|}$ when $|\zeta|>1$.
Therefore, the long-range interaction term can be written as
\begin{align*}
\partial_x\mathcal{N}_L(\varphi)=&2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))\left[K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\right]\,\mathrm{d} \zeta\\
=&2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta)) \sum\limits_{\mu=1}^\infty A_\mu(\zeta) \left|\varphi(t,x)-\varphi(t, x+\zeta)\right|^{2\mu}\,\mathrm{d} \zeta.
\end{align*}
{\bf 2.} By \eqref{K0-near0}, we can calculate the Taylor expansion
\begin{align*}
&K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\\
=~&\left[\ln\left(\frac{|\zeta|}2\right)+\gamma\right]I_0(\zeta)-\left[\ln\left(\frac{\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}}2\right)+\gamma\right]I_0(
\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})\\
&+\sum\limits_{k=1}^\infty b_k \left\{\left[\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2\right]^k-\zeta^{2k}\right\}\\
=~&\gamma \left(I_0(\zeta)-I_0(
\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})\right)+\left[\ln\left(\frac{|\zeta|}2\right)-\ln\left(\frac{\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}}2\right)\right]I_0(\zeta)\\
&-\ln\left(\frac{\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}}2\right)\left[I_0(
\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})-I_0(\zeta)\right]\\
&+\sum\limits_{k=1}^\infty b_k \left\{\left[\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2\right]^k-\zeta^{2k}\right\}.
\end{align*}
Taking into the Taylor expansion of $I_0$ and $\log$, we have
\begin{align*}
&K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\\
=~&-\left[\gamma+\ln\left(\frac{\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}}2\right)\right] \sum\limits_{k=1}^\infty \frac{1}{(k!)^2 4^k} \left[(\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2)^k-\zeta^{2k}\right]\\
&-\frac12I_0(\zeta)\sum\limits_{k=1}^\infty \frac{(-1)^{k-1}}{k}\left(\frac{\varphi(t, x+\zeta)-\varphi(t,x)}{\zeta}\right)^{2k}+\sum\limits_{k=1}^\infty b_k \left\{\left[\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2\right]^k-\zeta^{2k}\right\}\\
=~&-\left[\gamma+\ln\left(\frac{|\zeta|}2\right)+\frac12\sum\limits_{k=1}^\infty \frac{(-1)^{k-1}}{k}\left(\frac{\varphi(t, x+\zeta)-\varphi(t,x)}{\zeta}\right)^{2k}\right] \sum\limits_{k=1}^\infty \frac{1}{(k!)^2 4^k} \left[(\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2)^k-\zeta^{2k}\right]\\
&-\frac12I_0(\zeta)\sum\limits_{k=1}^\infty \frac{(-1)^{k-1}}{k}\left(\frac{\varphi(t, x+\zeta)-\varphi(t,x)}{\zeta}\right)^{2k}+\sum\limits_{k=1}^\infty b_k \left\{\left[\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2\right]^k-\zeta^{2k}\right\}\\
=~&
\sum\limits_{\mu=1}^\infty
B_\mu(\zeta)
(\varphi(t,x)-\varphi(t, x+\zeta))^{2\mu},
\end{align*}
where
\begin{align*}
B_\mu(\zeta)=&-\left[\gamma+\ln\left(\frac{|\zeta|}{2}\right)\right]
\sum\limits_{k=\mu}^\infty
\frac{1}{(k!)^24^k}
{k\choose \mu}|\zeta|^{2k-2\mu}
-
\frac12\sum\limits_{k=1}^\mu\frac{(-1)^{k-1}}{k}
\cdot
\sum\limits_{j=\mu-k}^\infty \frac{1}{(j!)^2 4^j}
{j\choose \mu-k}|\zeta|^{2j-2\mu}
\\
&~
-\frac12I_0(\zeta)
\frac{(-1)^{\mu-1}}{\mu}|\zeta|^{-2\mu}
+
\sum\limits_{k=\mu}^\infty b_k
{k\choose \mu}|\zeta|^{2k-2\mu}.
\end{align*}
We remark that the most singular term in $B_{\mu}(\zeta)$ is $|\zeta|^{-2\mu}$.
Therefore,
\begin{align*}
\partial_x\mathcal{N}_S(\varphi)=&2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))\left[K_0\left(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2}\right)-K_0(\zeta)\right]\,\mathrm{d} \zeta
\\
=&
2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta)) \sum\limits_{\mu=1}^\infty B_\mu(\zeta) \left(\varphi(t,x)-\varphi(t, x+\zeta)\right)^{2\mu}\,\mathrm{d} \zeta.
\end{align*}
{\bf 3.} Then we write the nonlinear terms into the multi-linear Fourier integral operators. By combining the cubic terms from $\mathcal{N}_L$ and $\mathcal{N}_S$, we define the cubic nonlinear term as
\begin{align}\nonumber
\partial_x\mathcal{N}_3(\varphi)=&2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))(\varphi(t,x)-\varphi(t, x+\zeta))^2A_{\mu}(\zeta)\,\mathrm{d} \zeta\\\nonumber
&+2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))(\varphi(t, x+\zeta)-\varphi(t,x))^2B_\mu(\zeta)\,\mathrm{d}\zeta\\\label{N3}
=&\frac13\partial_x\iiint_{\mathbb{R}^3} \mathbf{T}_1(\eta_1,\eta_2,\eta_3)e^{i(\eta_1+\eta_2+\eta_3)x}\hat\varphi(\eta_1)\hat\varphi(\eta_2)\hat\varphi(\eta_3)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d} \eta_3,
\end{align}
where
\begin{equation}\label{T1}
\begin{aligned}
&\mathbf{T}_1(\eta_1,\eta_2,\eta_3)=2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}\prod_{j=1}^3(1-e^{i\eta_j\zeta})A_{\mu}(\zeta)\,\mathrm{d} \zeta
+2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}} \prod_{j=1}^3(1-e^{i\eta_j\zeta})B_{\mu}(\zeta)\,\mathrm{d} \zeta.
\end{aligned}
\end{equation}
We remark that here and in the following, when there is no ambiguity, we may suppress the variable $t$ in the expression of the multilinear Fourier integral to make the expression shorter.
For the $m$-th ($m=2\mu+1$) degree term, we have
\begin{align*}
\partial_x\mathcal{N}_m(\varphi)=~&2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta)) \left(\varphi(t,x)-\varphi(t, x+\zeta)\right)^{2\mu} e^{-|\zeta|}A_\mu(\zeta)\,\mathrm{d} \zeta
\\
&+2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))(\varphi(t,x)-\varphi(t, x+\zeta))^{2\mu} B_{\mu}(\zeta)
d\zeta
\\
=~&\frac1m\partial_x\iiint_{\mathbb{R}^{m}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_m)e^{i\left(\sum\limits_{j=1}^m\eta_j\right)x}\hat\varphi(\eta_1)\cdots\hat\varphi(\eta_m)\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{m},
\end{align*}
where
\begin{align}\label{Tmu}
\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_m)
=2\int_{\mathbb{R}}\chi_{\{|\zeta|>1\}}\prod_{j=1}^m(1-e^{i\eta_j\zeta}) A_\mu(\zeta)\,\mathrm{d} \zeta
+2\int_{\mathbb{R}}\chi_{\{|\zeta|<1\}} \prod_{j=1}^m(1-e^{i\eta_j\zeta})B_\mu(\zeta)
\,\mathrm{d}\zeta.
\end{align}
Therefore
\begin{align}\label{nonlinearity}
\mathcal{N}(\varphi)=\sum\limits_{\mu=1}^\infty \mathcal{N}_{2\mu+1}(\varphi)=\sum\limits_{\mu=1}^\infty
\frac1{2\mu+1}\iiint_{\mathbb{R}^{{2\mu+1}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu+1})e^{i\left(\sum\limits_{j=1}^{2\mu+1}\eta_j\right)x}\hat\varphi(\eta_1)\cdots\hat\varphi(\eta_{2\mu+1})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu+1}},
\end{align}
or equivalently,
\[
\widehat{\mathcal{N}(\varphi)}(\xi)=\sum\limits_{\mu=1}^\infty\frac{1}{2\pi(2\mu+1)}\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})\hat\varphi(\eta_1)\cdots\hat\varphi(\eta_{2\mu})\hat\varphi(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\]
In the following, we also denote $\mathcal{N}_{\geq5}(\varphi):=\sum\limits_{\mu=2}^\infty \mathcal{N}_{2\mu+1}(\varphi)$. So we can split the nonlinear term as $\mathcal{N}(\varphi)=\mathcal{N}_3(\varphi)+\mathcal{N}_{\geq5}(\varphi)$.
\section{Local solutions}\label{sec:Loc}
In this section, we prove the local existence and uniqueness of the Cauchy problem
\begin{equation}
\label{front}
\left\{
\begin{aligned}
&\varphi_t(t,x)=2\int_{\mathbb{R}}(\varphi_x(t,x)-\varphi_{x}(t, x+\zeta))K_0(\sqrt{\zeta^2+(\varphi(t, x)-\varphi(t, x+\zeta))^2})\,\mathrm{d} \zeta
{
-2\pi\varphi_x(t,x)
}
\\
&\hspace{1.1cm}=-\frac{\partial_x}{\sqrt{1-\partial_x^2}}\varphi+\partial_x\mathcal{N}(\varphi),\\
&\varphi(x,0)=\varphi_0(x),
\end{aligned}
\right.
\end{equation}
where $\mathcal{N}(\varphi)=\sum\limits_{\mu=1}^\infty\mathcal{N}_{2\mu+1}(\varphi)$ is the nonlinear term defined in \eqref{nonlinearity}.
The strategy of the proof mainly follows Kato's theory of quasilinear equations \cite{Kat75b}. Firstly, we linearize the equation and construct a solution map. Then we prove the map is bounded in a higher regularity space and contraction in a lower regularity space. Therefore, the approximate sequence constructed by this solution map is a Cauchy sequence, thus convergent. Finally, we prove the continuity in time.
By taking the dyadic decomposition,
\begin{align*}
\widehat{\mathcal{N}_{2\mu+1}(\varphi)}(\xi)=&\frac1{2\pi(2\mu+1)}\sum\limits_{j_1,\cdots,j_{2\mu+1}\in\mathbb{Z}}
\iint_{\mathbb{R}^{{2\mu}}}
\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})
\\
&\cdot\hat\varphi_{j_1}(\eta_1)\cdots\hat\varphi_{j_{2\mu}}(\eta_{2\mu})\hat\varphi_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\end{align*}
Since the symbol $\mathbf{T}_\mu$ is symmetric with its variables, by a potential change of variables, we can assume $j_1\ge j_2\ge\cdots\ge j_{2\mu+1}$. The summation of the ordered indices is denoted by $\sum\limits_Q$. So
\begin{align*}
\widehat{\mathcal{N}_{2\mu+1}(\varphi)}(\xi)=&\frac1{2\pi(2\mu+1)}\sum_{Q}
\iint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})
\\
&\cdot\hat\varphi_{j_1}(\eta_1)\cdots\hat\varphi_{j_{2\mu}}(\eta_{2\mu})\hat\varphi_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\end{align*}
Then we linearize the equation as
\begin{align}
\label{eqn4_2}
&\hat\varphi_t
+
\frac{i\xi}{\sqrt{1+\xi^2}}\hat\varphi
=
\frac{1}{2\pi}
\sum_{\mu=1}^\infty \frac1{2\mu+1}i\xi
\\
&\cdot \sum_Q\iint_{\mathbb{R}^{2\mu}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})\hat \varphi_{j_1}(\eta_1)\hat u_{j_2}(\eta_2)\cdots\hat u_{j_{2\mu}}(\eta_{2\mu})\hat{u}_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\nonumber
\end{align}
This equation can be written in an abstract form as
\[
\varphi_t=\mathcal{A}(u)\varphi,\quad \varphi(0,x)=\varphi_0(x)\in H^s(\mathbb{R}),~~s\geq 8,
\]
where $\mathcal{A}(u)$ is a first order pseudo-differential operator.
The right-hand-side has one-order derivative loss, which can be controlled by Kato-Ponce type commutator estimate \cite{kato-ponce}.
Denote the map $\mathcal{B}: L^\infty(0,T; H^s(\mathbb{R})) \to L^\infty(0,T; H^s(\mathbb{R})), u\mapsto \varphi$.
In the following, we use a symmetrization argument to show that the map $\mathcal{B}$ is bounded in Sobolev space. Assume $\|\varphi_0\|_{H^s}<\bar C$, $\|\varphi_0\|_{H^3}<1$. Denote
\[
X_T=\{u\in L^\infty(0,T; H^s(\mathbb{R}))\mid \|u\|_{L^\infty_TH^s}\leq 2\bar C, \|u\|_{L^\infty_TH^3}<1\}.
\]
We will prove in the following that there is a positive $T$ such that if $u\in X_T$, the solution of the linearized equation $\varphi=\mathcal{B} u$ is also in $X_T$.
\subsection{Sobolev energy estimate}
For the higher order energy estimate, multiply $(1+|\xi|^2)^s\hat \varphi(t, -\xi)$ to \eqref{front}, and take integral with $\xi$,
\begin{align*}
\frac{\,\mathrm{d}}{\,\mathrm{d} t}
&\int_{\mathbb{R}}\frac12(1+|\xi|^2)^s|\hat\varphi(t, \xi)|^2 \,\mathrm{d} \xi
\\
=&
\frac{1}{2\pi}
\sum\limits_{\mu=1}^\infty \frac1{2\mu+1}i\sum_{Q}\iiint_{\mathbb{R}^{{2\mu+1}}}\xi(1+|\xi|^2)^s\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})\\
&\qquad\qquad\cdot\hat \varphi_{j_1}(\eta_1)\hat u_{j_2}(\eta_2)\cdots\hat u_{j_{2\mu}}(\eta_{2\mu})\hat{u}_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\hat \varphi(-\xi)\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\,\mathrm{d}\xi.
\end{align*}
Furthermore, making the change of variables $\eta_{2\mu+1}
=\eta_{2\mu+1}(\eta_1)=\xi-\eta_1-\cdots-\eta_{2\mu}$, we get
\begin{align*}
&\frac{\,\mathrm{d}}{\,\mathrm{d} t}
\int_{\mathbb{R}}\frac12(1+|\xi|^2)^s|\hat\varphi(t, \xi)|^2 \,\mathrm{d} \xi
\\
=&
\frac{1}{2\pi}
\sum\limits_{\mu=1}^\infty \frac1{2\mu+1}i\sum_{Q}\iiint_{\mathbb{R}^{{2\mu+1}}}\xi(1+|\xi|^2)^s
\,
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1},\eta_{2\mu},\cdots,\eta_{2\mu+1})
\\
&\hat \varphi_{j_1}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\hat u_{j_{2}}(\eta_2)
\cdots \hat u_{j_{2\mu+1}}(\eta_{2\mu+1})\hat{\varphi}(-\xi)\,\mathrm{d} \eta_2\cdots\,\mathrm{d} \eta_{{2\mu+1}}\,\mathrm{d}\xi.
\end{align*}
Interchanging the variables $-\xi$ and $\xi-\eta_2-\cdots-\eta_{2\mu+1}$, and then taking the average, we have
\begin{align*}
&
\iiint_{\mathbb{R}^{{2\mu+1}}}\xi(1+|\xi|^2)^s
\,
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1}, \eta_{2}, \cdots,\eta_{2\mu+1})
\\
&\hat \varphi_{j_1}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\hat u_{j_{2}} (\eta_2)
\cdots \hat u_{j_{2\mu+1}}(\eta_{2\mu+1})\hat{\varphi}(-\xi)\,\mathrm{d} \eta_2\cdots\,\mathrm{d} \eta_{{2\mu+1}}\,\mathrm{d}\xi
\\
=&
\iiint_{\mathbb{R}^{{2\mu+1}}}-(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s
\mathbf{T}_{\mu}(-\xi,\eta_2,\cdots, \eta_{2\mu+1})
\\
&\qquad\hat \varphi_{j_1}(-\xi) \hat u_{j_2}(\eta_2)\cdots\hat u_{j_{2\mu+1}}(\eta_{2\mu+1})\hat \varphi(\xi-\eta_2-\cdots-\eta_{2\mu+1})\,\mathrm{d} \eta_2\cdots\,\mathrm{d} \eta_{{2\mu+1}}\,\mathrm{d}\xi
\\
=&\frac12\iiint_{\mathbb{R}^{{2\mu+1}}}
\Big[\xi(1+|\xi|^2)^s
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1}, \eta_{2}, \cdots,\eta_{2\mu+1})
\psi_{j_{1}}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\\
&\qquad\qquad -(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s \mathbf{T}_{\mu}(-\xi, \eta_2,\cdots,\eta_{2\mu+1})\psi_{j_{1}}(-\xi)\Big]
\\
&\qquad\cdot\hat u_{j_2}(\eta_2)\cdots\hat u_{j_{2\mu+1}}(\eta_{2\mu+1})\hat\varphi(-\xi)\hat \varphi(\xi-\eta_2-\cdots-\eta_{2\mu+1})\,\mathrm{d} \eta_2\cdots\,\mathrm{d} \eta_{{2\mu+1}}\,\mathrm{d}\xi.
\end{align*}
We can split the symbol into three parts
\begin{align*}
&\Big[\xi(1+|\xi|^2)^s
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1}, \eta_{2}, \cdots,\eta_{2\mu+1})
\psi_{j_{1}}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\\
&\qquad\qquad -(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s \mathbf{T}_{\mu}(-\xi, \eta_2,\cdots,\eta_{2\mu+1})\psi_{j_{1}}(-\xi)\Big]
\\
=&\Big(\xi(1+|\xi|^2)^s-(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s\Big)
\\
&\cdot
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1}, \eta_{2}, \cdots,\eta_{2\mu+1})
\psi_{j_{1}}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\\
&+(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s
\\
&\cdot\Big(
\mathbf{T}_{\mu}(\xi-\eta_2-\cdots-\eta_{2\mu+1}, \eta_{2}, \cdots,\eta_{2\mu+1})
-
\mathbf{T}_{\mu}(-\xi, \eta_{2}, \cdots,\eta_{2\mu+1})
\Big)\psi_{j_{1}}(\xi-\eta_2-\cdots-\eta_{2\mu+1})
\\
&+(\xi-\eta_2-\cdots-\eta_{2\mu+1})(1+|\xi-\eta_2-\cdots-\eta_{2\mu+1}|^2)^s
\mathbf{T}_{\mu}(-\xi, \eta_{2}, \cdots,\eta_{2\mu+1})
\Big(\psi_{j_{1}}(\xi-\eta_2-\cdots-\eta_{2\mu+1})-\psi_{j_{1}}(-\xi)\Big).
\end{align*}
When writing it as the symbol of a multilinear Fourier integral operator, we use $\eta_{1}$ to replace $\xi-\eta_2-\cdots-\eta_{2\mu+1}$. By Proposition \ref{Prop_symb_Tmu} and the algebraic property of $S^\infty$ norm,
\begin{align*}
&
\left\|\left(\left(\sum\limits_{i=1}^{2\mu+1}\eta_i\right)\left(1+\left|\sum\limits_{i=1}^{2\mu+1}\eta_i\right|^2\right)^s
-\eta_{1}\left(1+|\eta_{1}|^2\right)^s\right)\cdot\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \eta_{2\mu+1})\psi_{j_1}(\eta_1)\cdots\psi_{j_{2\mu}}(\eta_{2\mu})\psi_{j_{2\mu+1}}(\eta_{2\mu+1})\right\|_{S^\infty}
\\
\lesssim~&
2^{2sj_{1}+(j_2+\cdots+j_{2\mu+1})}
(1+2^{2j_{2}})
(1+2^{j_2+\cdots+j_{2\mu+1}})(2^{j_2}+\cdots+2^{j_{2\mu+1}}).
\end{align*}
Since the symbol $\mathbf{T}_\mu$ is real, it is an even function with its variables, i.e. $$\mathbf{T}_\mu(\eta_1, \cdots, \eta_i,\cdots,\eta_{2\mu+1}) = \mathbf{T}_\mu(\eta_1, \cdots, -\eta_i,\cdots, \eta_{2\mu+1}),$$ for any $i=1,\cdots, 2\mu+1$. So by \eqref{Tmu_est5} we have
\begin{align*}
&\left\|\eta_{1}\left(1+|\eta_{1}|^2\right)^s
\cdot
\Big(\mathbf{T}_{\mu}(\eta_1, \eta_{2}, \cdots, \eta_{2\mu+1})
-
\mathbf{T}_{\mu}(-\sum\limits_{i=1}^{2\mu+1}\eta_i, \eta_{2}, \cdots, \eta_{2\mu+1}) \Big)\psi_{j_1}(\eta_1)\cdots\psi_{j_{2\mu}}(\eta_{2\mu})\psi_{j_{2\mu+1}}(\eta_{2\mu+1})\right\|_{S^\infty}
\\
=&\Big\|\eta_{1}(1+|\eta_{1}|^2)^s\cdot
\Big(
\mathbf{T}_{\mu}(\eta_1,\eta_{2}, \cdots, \eta_{2\mu+1})-\mathbf{T}_{\mu}(\sum\limits_{i=1}^{2\mu+1}\eta_i, \eta_{2}, \cdots, \eta_{2\mu+1}) \Big)\psi_{j_1}(\eta_1)\cdots\psi_{j_{2\mu}}(\eta_{2\mu})\psi_{j_{2\mu+1}}(\eta_{2\mu+1})\Big\|_{S^\infty}
\\
\lesssim&
(1+2^{2sj_{1}})(1+2^{3j_{2}})
\prod\limits_{k=2}^{2\mu+1}[(1+2^{j_k})2^{j_k}].
\end{align*}
Also, using $\psi_{j_1}'(\xi)=\frac1{2^{j_1}}\psi'(\xi/2^{j_1})$ and Proposition \ref{Prop_symb_Tmu} we get
\begin{align*}
&\Big\|\eta_{1}(1+|\eta_{1}|^2)^s
\mathbf{T}_{\mu}(-\sum\limits_{i=1}^{2\mu+1}\eta_i,\eta_{2}\cdots,\eta_{2\mu+1}) \psi_{j_2}(\eta_2)\cdots\psi_{j_{2\mu+1}}(\eta_{2\mu+1})\cdot\big(\psi_{j_{1}}(\eta_{1})-\psi_{j_{1}}(-\sum\limits_{i=1}^{2\mu+1}\eta_i)\big)\Big\|_{S^\infty}
\\
\lesssim & 2^{2sj_{1}+(j_2+\cdots+j_{2\mu+1})}(1+2^{2j_{2}})(1+2^{j_2+\cdots+j_{2\mu+1}})(2^{j_2}+\cdots+2^{j_{2\mu+1}}).
\end{align*}
Combining the above three parts, we have
\begin{align*}
&\left\|
\left(
\sum\limits_{i=1}^{2\mu+1}\eta_i
\right)
\left(1+\left|\sum\limits_{i=1}^{2\mu+1}\eta_i\right|^2\right)^s
\mathbf{T}_{\mu}(\eta_1, \eta_{2}, \cdots,\eta_{2\mu+1})
\psi_{j_{1}}(\eta_1)
\right.
\\
&\qquad\qquad
\left.-(\eta_1)(1+|\eta_1|^2)^s \mathbf{T}_{\mu}\left(-\sum\limits_{i=1}^{2\mu+1}\eta_i, \eta_2,\cdots,\eta_{2\mu+1}\right)\psi_{j_{1}}\left(-\sum\limits_{i=1}^{2\mu+1}\eta_i\right)
\right\|_{S^\infty}
\\
\lesssim~~& 2^{2sj_{1}+(j_2+\cdots+j_{2\mu+1})}(1+2^{4j_{2}})(1+2^{j_2+\cdots+j_{2\mu+1}}).
\end{align*}
Therefore,
\begin{align*}
\bigg|&\iiint_{\mathbb{R}^{{2\mu+1}}}\xi(1+|\xi|^2)^s
\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})
\\
&\qquad \hat \varphi_{j_1}(\eta_1)\hat u_{j_{2}}(\eta_2)\cdots\hat u_{j_{2\mu}}(\eta_{2\mu})\hat u_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\hat \varphi(-\xi)\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\,\mathrm{d}\xi\bigg|
\\
\lesssim~&
\|\varphi_{j_{1}}\|_{H^s}^2\|\partial_x^2u_{j_{2}}\|_{W^{4,\infty}}\prod_{i=3}^{2\mu+1}\|\partial_x u_{j_i}\|_{W^{1,\infty}}.
\end{align*}
After taking the summation for integers $j_1\geq \cdots\geq j_{2\mu+1}$, we obtain
\begin{align*}
\frac{\,\mathrm{d}}{\,\mathrm{d} t} \|\varphi(t)\|_{H^s}^2\lesssim \sum_{\mu=1}^\infty \|\varphi(t)\|_{H^s}^2\sum
\limits_{\tiny
\begin{array}{c}
j_i\in\mathbb{Z}
\\
i=2,\cdots,2\mu+1
\end{array}}
\left(\|\partial_x^2u_{j_{2}}\|_{W^{4,\infty}}\prod_{i=3}^{2\mu+1}\|\partial_x u_{j_i}\|_{W^{1,\infty}}\right).
\end{align*}
Replacing $u$ by $\varphi$, it leads to the energy inequality \ref{EnergyIneq}.
By Sobolev embedding theorem, we can obtain
\begin{align*}
\frac{\,\mathrm{d}}{\,\mathrm{d} t} \|\varphi(t)\|_{H^s}^2
\lesssim &
\sum_{\mu=1}^\infty \|\varphi(t)\|_{H^s}^2\sum\limits_{\tiny
\begin{array}{c}
j_i\in\mathbb{Z}
\\
i=2,\cdots,2\mu+1
\end{array}}
\left(\|\partial_x^2u_{j_{2}}\|_{H^5}\prod_{i=3}^{2\mu+1}\|\partial_x u_{j_i}\|_{H^2}
\right)
\\
\lesssim~&
\sum_{\mu=1}^\infty \|\varphi(t)\|_{H^s}^2 \|\partial_x^2 u(t)\|_{H^5}\|\partial_xu(t)\|_{H^2}^{2\mu-1}.
\end{align*}
After taking integral with $t$, we obtain
\[
\|\varphi(t)\|_{H^s}\leq \|\varphi_0\|_{H^s}+C\int_0^t \sum_{\mu=1}^\infty \|\varphi(\tau)\|_{H^s}^2 \|\partial_x^2 u(\tau)\|_{H^5}\|\partial_xu(\tau)\|_{H^2}^{2\mu-1}\,\mathrm{d} \tau,
\]
for some constant $C$.
Since the initial data satisfies $\|\varphi_0\|_{H^s}<\bar C$, and $\|\varphi_0\|_{H^3}<1$, there exist $T_1>0$, such that $\varphi\in X_{T_1}$.
\subsection{Contraction in lower space}
Assume for $i=1,2$, $u_i \in X_T$, and $\varphi_i$ is the solution of
\[
\partial_t\varphi_i=\mathcal{A}(u_i)\varphi_i, \quad \varphi_i(0,x)=\varphi_0(x)\in H^s(\mathbb{R}), s\geq 8.
\]
We will prove the contraction of $\mathcal{B}$ in $L^\infty(0,T; H^7(\mathbb{R}))$.
By taking difference of the two equations,
\[
\partial_t(\varphi_1-\varphi_2)=\mathcal{A}(u_1)(\varphi_1-\varphi_2)+[\mathcal{A}(u_1)-\mathcal{A}(u_2)]\varphi_2.
\]
Multiply $(1+|\xi|^2)^7(\hat \varphi_1(t, -\xi)-\hat \varphi_2(t, -\xi))$ to the Fourier transform of the above equation, and take integral with $\xi$,
\begin{align*}
&\frac{\,\mathrm{d}}{\,\mathrm{d} t}\|(\varphi_1-\varphi_2)(t)\|_{H^7}^2\\
\lesssim~ &\sum_{\mu=1}^\infty \|(\varphi_1-\varphi_2)(t)\|_{H^7}^2 \|\partial_x^2 u_1(t)\|_{H^5}\|\partial_xu_1(t)\|_{H^2}^{2\mu-1}+\Big[\|\partial_x^2(u_1-u_2)(t)\|_{H^5}(\|\partial_xu_1(t)\|_{H^2}^{2\mu-1}+\|\partial_xu_2(t)\|_{H^2}^{2\mu-1})\\
&+(\|\partial_x^2 u_1(t)\|_{H^5}+\|\partial_x^2 u_2(t)\|_{H^5})\|\partial_x(u_1-u_2)(t)\|_{H^2}(\|\partial_xu_1(t)\|_{H^2}^{2\mu-2}+\|\partial_xu_2(t)\|_{H^2}^{2\mu-2})\Big]\|\varphi_2(t)\|_{H^8}\|(\varphi_1-\varphi_2)(t)\|_{H^7}.
\end{align*}
By Gronwall inequality,
\begin{align*}
\|(\varphi_1-\varphi_2)(t)\|_{H^7}\leq \int_0^t e^{\int_0^s \|\partial_x^2 u_1(\tau)\|_{H^5}\sum_{\mu}\|\partial_xu_1(\tau)\|_{H^2}^{2\mu-1}\,\mathrm{d} \tau}\Big[\|\partial_x^2(u_1-u_2)(s)\|_{H^5}(\|\partial_xu_1(s)\|_{H^2}^{2\mu-1}+\|\partial_xu_2(s)\|_{H^2}^{2\mu-1})\\
+(\|\partial_x^2 u_1(s)\|_{H^5}+\|\partial_x^2 u_2(s)\|_{H^5})\|\partial_x(u_1-u_2)(s)\|_{H^2}(\|\partial_xu_1(s)\|_{H^2}^{2\mu-2}+\|\partial_xu_2(s)\|_{H^2}^{2\mu-2})\Big]\|\varphi_2(s)\|_{H^8}\,\mathrm{d} s.
\end{align*}
Since $u_1, u_2$ are both bounded in $L^\infty(0,T; H^s(\mathbb{R})), s\geq 8$, there exists $T_2>0$, such that there exist $L<1$,
\[
\|(\varphi_1-\varphi_2)\|_{L^\infty_{T_2}H^7}\leq L\|u_1-u_2\|_{L^\infty_{T_2}H^7}.
\]
This contraction argument also leads to the uniqueness of solutions to Cauchy problem \eqref{front}.
\subsection{Iteration scheme}
We then construct a sequence $\{\varphi^{(i)}\}$ of approximate solutions of \eqref{front} by
\begin{equation}
\varphi^{(0)}(x,t)=\varphi_0(x),\qquad \varphi^{(i)}=\mathcal{B}(\varphi^{(i-1)})\quad \text{for $i\in \mathbb N$}.
\label{defseq}
\end{equation}
For sufficiently small $T > 0$, we have proved that this sequence is bounded in $L^\infty([0,T]; H^s(\mathbb{R}))$ and Cauchy in
$L^\infty([0,T]; H^7(\mathbb{R}))$, which implies that its limit exists in $L^\infty([0,T]; H^7(\mathbb{R}))$ and is a local solution of the initial value problem \eqref{front}.
\subsection{Continuity in time}
Next, we prove that the solution constructed above is a continuous function of time with values in $H^s$.
First, we notice that $\varphi_t\in L^{\infty}([0,T];H^{s'}(\mathbb{R}))$ for any
$s' < s-1$, which implies that $\varphi\in C([0,T]; H^{s'}(\mathbb{R}))$.
The equation is time reversible and translation invariant in time, so it suffices to prove that
\[
\lim\limits_{t\to0+}\|\varphi(t)-\varphi(0)\|_{H^s}=0.
\]
Since $\|\varphi(t)\|_{H^s}$ is bounded on $[0,T]$ and $\varphi(t) \to \varphi(0)$ strongly in $H^{s'}$, the weak $H^s$-limit of any convergent subsequence is unique, and we see that
$\varphi(t)$ converges to $\varphi(0)$ in the weak $H^s$-topology.
To show convergence in the strong $H^s$-topology, we only need to prove the norm-convergence
\begin{equation}
\lim\limits_{t\to0+}\|\varphi(t)\|_{H^s}=\|\varphi(0)\|_{H^s}.
\label{normcon}
\end{equation}
This is directly from the energy inequality and the Gronwall's inequality.
Finally, we obtained the following existence and uniqueness theorem of Cauchy problem \eqref{front}.
\begin{theorem}[Local solutions]
If the initial data $\varphi_0\in H^s(\mathbb{R})$ for $s\geq 8$ satisfying $\| \varphi_0\|_{H^3}<1$, then there exist $T>0$ such that there is a unique solution to the Cauchy problem \eqref{front} in $C([0,T]; H^s(\mathbb{R}))$. The solution satisfies the following energy inequality
\begin{equation}\label{EnergyIneq}
\|\varphi(t)\|_{H^s}^2\lesssim \|\varphi_0\|_{H^s}^2+\int_0^t \|\varphi(\tau)\|_{H^s}^2\sum_{\mu=1}^\infty \|\varphi(\tau)\|_{B^{2,6}}\|\varphi(\tau)\|_{B^{1,2}}^{2\mu-1}\,\mathrm{d} \tau.
\end{equation}
\end{theorem}
\section{Global solutions for small localized initial data}\label{sec:Glo}
In the remaining sections, we will prove the global boundedness of the energy. We formulate the main theorem first.
Define the profile function $h(t,x)=e^{t\partial_x(1-\partial_x^2)^{-1/2}}\varphi(t,x)$ and the dispersion relation $p(\xi)=-\xi(1+\xi^2)^{-1/2}$. Thus $\hat h(t, \xi)=e^{-itp(\xi)}\hat \varphi(t, \xi)$. Define some constants:
\begin{align}\label{param_vals}
r=0.4, ~~w=11,~~ s=130,~~ p_0=10^{-4}.
\end{align}
$Z$-norm is defined as
\begin{equation}\label{Znorm}
\|f\|_Z:=\|(|\xi|^{r}+|\xi|^{w})\hat f(\xi)\|_{L^\infty_\xi}.
\end{equation}
Our main theorem is
\begin{theorem}\label{Thm:glo}
Let $s$, $p_0$ be defined as in \eqref{param_vals}. There exists a constant $0<\varepsilon \ll 1$, such that if $\varphi_0\in H^s(\mathbb{R})$ and satisfies
\[
\|\varphi_0\|_{H^s} + \|\partial_\xi\hat\varphi_0(\xi)\|_{L^2}+\|\varphi_0\|_{Z}\leq \varepsilon_0,
\]
for some $0 < \varepsilon_0 \leq \varepsilon$, then there exists a unique global solution $\varphi\in C([0,\infty); H^s(\mathbb{R}))$ of \eqref{front}.
Moreover, this solution satisfies
\[
(t+1)^{-p_0} \|\varphi(t)\|_{H^s}+(t+1)^{-0.01}\|\partial_\xi\hat h(t,\xi)\|_{L^2_\xi} + \|\varphi\|_{Z} \lesssim\varepsilon_0.
\]
\end{theorem}
The proof of this theorem follows directly from the following bootstrap proposition.
\begin{proposition}[Bootstrap]\label{bootstrap}
Let $T>1$ and suppose that $\varphi\in C([0,T]; H^s)$ is a solution of \eqref{front}, where the initial data satisfies
\[
\|\varphi_0\|_{H^s} + \|\partial_\xi\hat\varphi_0(\xi)\|_{L^2}+\|\varphi_0\|_{Z}\leq \varepsilon_0
\]
for some $0 < \varepsilon_0 \ll 1$. If there exists $\varepsilon_1$ with $\varepsilon_0 \leq \varepsilon_1 \lesssim \varepsilon_0^{2/3}$ such that the solution satisfies
\[
(t+1)^{-p_0} \|\varphi(t)\|_{H^s}+(t+1)^{-0.01}\|\partial_\xi\hat h(t,\xi)\|_{L^2_\xi} +\|\varphi(t)\|_{Z}\leq \varepsilon_1,
\]
for every $t\in [0,T]$, then the solution satisfies an improved bound
\[
(t+1)^{-p_0} \|\varphi(t)\|_{H^s}+(t+1)^{-0.01}\|\partial_\xi\hat h(t,\xi)\|_{L^2_\xi} + \|\varphi(t)\|_{Z} \lesssim\varepsilon_0.
\]
\end{proposition}
\begin{lemma}[Nonlinear pointwise decay]\label{NonDis}
Under the bootstrap assumptions,
\[
\|\varphi\|_{B^{1,6}} \lesssim (t+1)^{-1/2}\varepsilon_1.
\]
\end{lemma}
\begin{lemma}[Weighted energy estimate]\label{weightedE}
Under the bootstrap assumptions,
\[
\|\partial_\xi\hat h(t,\xi)\|_{L^2_\xi}\lesssim \varepsilon_0(t + 1)^{0.01}.
\]
\end{lemma}
\begin{lemma}[Z-norm estimate]\label{Z-norm_Est}
Under the bootstrap assumptions,
\[
\|\varphi(t)\|_{Z}\lesssim \varepsilon_0.
\]
\end{lemma}
\begin{proof}[Proof of Proposition \ref{bootstrap}]
By Lemma \ref{NonDis} and the energy inequality \eqref{EnergyIneq}, we obtain
\[
\|\varphi(t)\|_{H^s}\lesssim \varepsilon_0(t+1)^{p_0}.
\]
Then using Lemma \ref{weightedE} and Lemma \ref{Z-norm_Est}, we complete the proof.
\end{proof}
In the following sections, we will prove Lemma \ref{NonDis}, Lemma \ref{weightedE} and Lemma \ref{Z-norm_Est} separately.
\section{Dispersive estimate}\label{sec:Dis}
\begin{lemma}[Linear dispersive estimate]\label{disp}
For $t> 0$ and $h \in L^2$, we have the linear dispersive estimates
\begin{align}\nonumber
&\|e^{itp(D)} P_kh \|_{L^\infty}\\\label{LocDis}
\lesssim~& \frac1{\sqrt t} 2^{-\frac12k}(1+2^{2k})^{5/2} \|\psi_k(\xi)\hat h(t, \xi)\|_{L^\infty_\xi}+(t+1)^{-3/4} 2^{-\frac34k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (t, \xi)\psi_k(\xi)\|_{L^2_\xi},
\end{align}
where $D=-i\partial_x$.
\end{lemma}
\begin{proof} Using the inverse Fourier transform, we can write the solution as
\begin{align*}
e^{itp(D)} P_kh
= \int_\mathbb{R} e^{ix\xi+ip(\xi)t} \psi_k(\xi)\hat h (\xi)\,\mathrm{d}{\xi}.
\end{align*}
Notice that
\begin{equation}\label{exp}
\partial_\xi e^{ix\xi+ip(\xi)t}=i[x+tp'(\xi)] e^{ix\xi+ip(\xi)t}.
\end{equation}
Denote $\xi_0^\pm$ to be the solution of $p'(\xi_0^\pm)=-x/t$. Then we discuss different situations for $\xi_0^\pm$.
\noindent{\bf Case I.} When $|\xi_0^\pm|>\frac85 2^{k+1}$ or $|\xi_0^\pm|<\frac58 2^{k-1}$, integral by part with respect to $\xi$,
\begin{align}\nonumber
&\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\\\label{eqn6.3}
=&\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{-ip''(\xi)}{-t[\frac{x}t+p'(\xi)]^2}\hat h(\xi)\psi_k(\xi)\,\mathrm{d}\xi+\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{ 1}{it(\frac{x}t+p'(\xi))} \partial_\xi \left(\hat h (\xi)\psi_k(\xi)\right)\,\mathrm{d}\xi.
\end{align}
Since $p'(\xi)=-\frac1{(1+\xi^2)^{3/2}}$ is an even function and is increasing on $[0,\infty)$,
\begin{align*}
\Big|\frac{x}t+p'(\xi)\Big|=|p'(\xi)-p'(\xi_0)|\gtrsim |p''(2^k)|2^k\gtrsim 2^{2k}(1+2^{2k})^{-5/2}.
\end{align*}
Therefore, the two integrals in \eqref{eqn6.3} satisfy
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{-ip''(\xi)}{-t[\frac{x}t+p'(\xi)]^2}\hat h(\xi)\psi_k(\xi)\,\mathrm{d}\xi\right|\\
\lesssim~& \frac1t 2^{-4k}(1+2^{2k})^{5} 2^k(1+2^{2k})^{-5/2}2^{k} \|\psi_k(\xi)\hat h(\xi)\|_{L^\infty}\\
\lesssim~& \frac1t 2^{-2k}(1+2^{k})^{5} \|\psi_k\hat h\|_{L^\infty},
\end{align*}
and
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{ 1}{it(\frac{x}t+p'(\xi))} \partial_\xi \left(\hat h (\xi)\psi_k(\xi)\right)\,\mathrm{d}\xi\right|\\
\lesssim~&\frac1t 2^{-k}(1+2^{2k})^{5/2}2^{-\frac{k}2}\|\partial_\xi(\hat h (\xi)\psi_k(\xi))\|_{L^2_\xi}\\
\lesssim~&\frac1t 2^{-\frac32k}(1+2^{2k})^{5/2}(\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}+\|\hat h (\xi)\psi_k(\xi)\|_{L^\infty_\xi}2^{-k/2})\\
\lesssim~&\frac1t 2^{-\frac32k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}+ \frac1t 2^{-2k}(1+2^{2k})^{5/2} \|\psi_k\hat h\|_{L^\infty}.
\end{align*}
So we have
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\right|\\
\lesssim~&\frac1t 2^{-\frac32k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}+ \frac1t 2^{-2k}(1+2^{2k})^{5/2} \|\psi_k\hat h\|_{L^\infty}.
\end{align*}
When $t>2^{-3k}$, we have $t^{1/4}>2^{-\frac34k}$, $t^{1/2}>2^{-\frac32k}$, thus
\begin{align}\nonumber
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\right|\\\label{case1}
\lesssim~&\frac1{t^{3/4}} 2^{-\frac34k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}+ \frac1{t^{1/2}} 2^{-\frac12k}(1+2^{2k})^{5/2} \|\psi_k\hat h\|_{L^\infty}.
\end{align}
When $t<2^{-3k}$, we have $2^{\frac32k}<t^{-1/2}$, thus
\begin{align}\label{lowfre}
\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\right|
\lesssim~ 2^{k} \|\psi_k\hat h\|_{L^\infty}\lesssim \frac1{t^{1/2}} 2^{-\frac12k}(1+2^{2k})^{5/2} \|\psi_k\hat h\|_{L^\infty}.
\end{align}
\noindent{\bf Case II.} When $\frac58 2^{k-1}\leq |\xi_0^\pm|\leq \frac85 2^{k+1}$, we make further dyadic decomposition
\begin{align*}
\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}
=\sum\limits_{l}\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\,\mathrm{d}{\xi}.
\end{align*}
Considering the support of the cut-off functions, we can assume the above summation is for $l\leq k+1$.
After integration by part, we have
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\,\mathrm{d}{\xi}\right|\\
\lesssim~& \left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{-ip''(\xi)}{-t[\frac{x}t+p'(\xi)]^2}\hat h(\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\,\mathrm{d}\xi\right|\\
&+\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{1}{it(\frac{x}t+p'(\xi))} \partial_\xi \left(\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\right)\,\mathrm{d}\xi\right|.
\end{align*}
On the support of the cut-off function $\psi_{l}(|\xi|-|\xi_0^\pm|)$,
\begin{align*}
\left|\frac{x}t+p'(\xi)\right|=|p'(\xi)-p'(\xi_0)|\gtrsim |p''(2^k)|2^l\gtrsim 2^{k+l}(1+2^{2k})^{-5/2}.
\end{align*}
Therefore,
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{-ip''(\xi)}{-t[\frac{x}t+p'(\xi)]^2}\hat h(\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\,\mathrm{d}\xi\right|\\
\lesssim~& \frac1t 2^{-2k-2l}(1+2^{2k})^{5} 2^k(1+2^{2k})^{-5/2}2^l \|\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\hat h(\xi)\|_{L^\infty_\xi}\\
\lesssim~& \frac1t 2^{-k-l}(1+2^{2k})^{5/2} \|\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\hat h(\xi)\|_{L^\infty_\xi},
\end{align*}
and
\begin{align*}
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\frac{1}{it(\frac{x}t+p'(\xi))} \partial_\xi \left(\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\right)\,\mathrm{d}\xi\right|\\
\lesssim~&\frac1t 2^{-k-l}(1+2^{2k})^{5/2}2^{\frac{l}2}\|\partial_\xi (\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|))\|_{L^2_\xi}\\
\lesssim~&\frac1t 2^{-k-\frac l2}(1+2^{2k})^{5/2}\left[\|\partial_\xi\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\|_{L^2_\xi}+\|\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\|_{L^\infty_\xi}2^{-l/2})\right].
\end{align*}
Take the summation for $l$ from $-\frac12\log_2(t+1)-\frac k2$ to $k+1$,
\begin{align*}
&\left|\sum\limits_{l=-\frac12\log_2(t+1)}^{k+1}\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\right|\\
\lesssim~& t^{-1/2} 2^{-\frac12 k}(1+2^{2k})^{5/2} \|\psi_k(\xi)\hat h(\xi)\|_{L^\infty_\xi}+(t+1)^{-3/4} 2^{-\frac34 k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}.
\end{align*}
For $l<-\frac12\log_2(t+1)-\frac k2$,
\[
\left|\sum\limits_{l<-\frac12\log_2(t+1)-\frac k2}\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\psi_{l}(|\xi|-|\xi_0^\pm|)\,\mathrm{d}{\xi}\right|\lesssim (t+1)^{-1/2} 2^{-\frac12 k}\|\psi_k(\xi)\hat h(\xi)\|_{L^\infty_\xi}.
\]
So we have
\begin{align}\nonumber
&\left|\int_\mathbb{R} e^{ix\xi+ip(\xi)t}\hat h (\xi)\psi_k(\xi)\,\mathrm{d}{\xi}\right|\\
\lesssim~& \frac1{\sqrt t} 2^{-\frac12k}(1+2^{2k})^{5/2} \|\psi_k(\xi)\hat h(\xi)\|_{L^\infty_\xi}+(t+1)^{-3/4} 2^{-\frac34k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}.\label{case2}
\end{align}
Hence we obtain the conclusion by combining \eqref{case1}, \eqref{lowfre} and \eqref{case2}.
\end{proof}
\begin{proof}[Proof of Lemma \ref{NonDis}]
By Lemma \ref{disp}, for $t>1$, we have
\begin{align*}
\|\varphi\|_{B^{1,6}}\lesssim &\sum\limits_{2^k\leq t^{1/100}}(2^k+2^{6k})\|e^{t\partial_x(1-\partial_x^2)^{-1/2}} P_kh \|_{L^\infty}+\sum\limits_{2^k> t^{1/100}}\|\partial_x^{6}P_k\varphi\|_{L^\infty}\\
\lesssim &\sum\limits_{2^k\leq t^{1/100}}(1+2^{5k})\left[ \frac1{\sqrt t} 2^{\frac12k}(1+2^{2k})^{5/2} \|\psi_k(\xi)\hat h(\xi)\|_{L^\infty_\xi}+(t+1)^{-3/4} 2^{\frac14k}(1+2^{2k})^{5/2}\|\partial_\xi\hat h (\xi)\psi_k(\xi)\|_{L^2_\xi}\right]\\
&+\sum\limits_{2^k> t^{1/100}} 2^{6k}2^{k/2-sk}\||\xi|^s\psi_k\hat\varphi\|_{L^2_\xi}\\
\lesssim &t^{-1/2} \|(|\xi|^{0.4}+|\xi|^{11}) \hat h(\xi)\|_{L^\infty_\xi}+(t+1)^{-\frac34+\frac{11}{100}}\|\partial_\xi \hat h\|_{L^2_\xi}+t^{-1}\|\varphi\|_{H^s}\\
\lesssim &t^{-1/2} \varepsilon_1,
\end{align*}
where the last step is from the bootstrap assumption.
\end{proof}
\section{Weighted energy estimate}\label{sec:WE}
In this section, we prove Lemma \ref{weightedE}.
Recall that $\hat h(t, \xi)=e^{-itp(\xi)}\hat \varphi(t, \xi)$. Then the equation \eqref{front} can be expressed as
\begin{equation}\label{eqhhat}
\hat h_t= e^{-itp(\xi)}i\xi\widehat{\mathcal{N}(\varphi)} = e^{-itp(\xi)}i\xi\left(\widehat{\mathcal{N}_3(\varphi)}+\widehat{\mathcal{N}_{\geq5}(\varphi)}\right).
\end{equation}
Denote
\begin{equation}\label{defPhi}
\Phi(\eta_1,\eta_2,\xi)=p(\eta_1)+p(\eta_2)+p(\xi-\eta_1-\eta_2)-p(\xi)=
-\frac{\eta_1}{\sqrt{1+\eta_1^2}}-\frac{\eta_2}{\sqrt{1+\eta_2^2}}-\frac{\xi-\eta_1-\eta_2}{\sqrt{1+(\xi-\eta_1-\eta_2)^2}}+\frac{\xi}{\sqrt{1+\xi^2}}.
\end{equation}
By \eqref{N3}, we have
\begin{align*}
e^{-itp(\xi)}\widehat{\mathcal{N}_3(\varphi)}(\xi)&=\frac13 e^{-itp(\xi)}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\hat\varphi(\eta_1)\hat\varphi(\eta_2)\hat\varphi(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&=\frac13 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)e^{it\Phi}\hat h(\eta_1)\hat h(\eta_2)\hat h(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&=\frac13 \sum\limits_{j_1,j_2,j_3\in\mathbb{Z}}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)e^{it\Phi}\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Since this integral is symmetric with the three variables $\eta_1, \eta_2, \xi-\eta_1-\eta_2$, after the possible change of variables, we assume $j_1\geq j_2\geq j_3$.
Denote the set $\P$ as all the indices $j_1,j_2,j_3\in\mathbb{Z}$, such that $j_1\geq j_2\geq j_3$ with possible repetition.
Take $\partial_\xi$ to \eqref{eqhhat},
\begin{equation}\label{eqn7.2}
\begin{aligned}
\partial_{\xi}\hat h_t
=~&\frac i3\xi \iint_{\mathbb{R}^2}\partial_{\xi}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h(\eta_1)\hat h(\eta_2)\hat h(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac i3\xi \sum\limits_{\P}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\xi}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&-\frac13\xi t\sum\limits_{\P}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac i3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h(\eta_1)\hat h(\eta_2)\hat h(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+i\partial_\xi[ \xi e^{-itp(\xi)}\widehat{\mathcal{N}_{\geq5}(\varphi)}]\\
=:~&\rm{I+II+III+IV+V}.
\end{aligned}\end{equation}
We split the index set $\P$ into $\P_1\bigcup\P_2$, where $\P_1$ includes the indices satisfying $j_1\geq j_2\geq j_3$ and $ j_1-j_3> 1$, which correspond to nonresonant frequencies. $\P_2$ includes indices satisfying $j_1\geq j_2\geq j_3$ and $ j_1-j_3\leq 1$, which contains the resonant frequencies.
Thus the term $\rm{III}$ can be written as
\begin{align*}
\rm{III}=&-\frac 13\xi t \sum\limits_{\P_1}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&-\frac 13\xi t \sum\limits_{\P_2}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=:&\rm{III}_1+\rm{III}_2.
\end{align*}
For $\rm{III}_2$, we need the further decomposition. Define
\[
\upsilon_\pm(\eta)=\left\{
\begin{aligned}
&1,\quad {\text {if } } \pm\eta\geq 0,\\
&0, \quad {\text {if } } \pm\eta<0,
\end{aligned}
\right.
\]
and $\hat h_{j}^\pm(\eta)=\hat h_j(\eta)\upsilon_\pm(\eta)$.
The integrals in $\rm{III}_2$ can be written as
\begin{align*}
&-\frac 13\xi t \sum\limits_{\P_2}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&-\sum\limits_{\iota_1, \iota_2, \iota_3\in\{\pm\}}\frac 13\xi t \sum\limits_{\P_2}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Define the nonresonant part
\begin{align*}
{\rm{III}}_{21}=-\sum\limits_{\mbox{$\scriptsize\begin{array}{cc}(\iota_1, \iota_2, \iota_3)=\\ (+,+,-)\\
\text{ or } (-, -, +)\end{array}$}}\frac 13\xi t \sum\limits_{\P_2}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align*}
and the space-time resonance terms around $(\xi, -\xi)$ or $(-\xi, \xi)$,
\begin{align*}
{\rm{III}}_{22}=-\sum\limits_{\scriptsize\begin{array}{cc}\mbox{$(\iota_1, \iota_2)=(\pm, \mp)$}\\ \mbox{$\iota_3\in \{+, -\}$}\end{array}}\frac 13\xi t \sum\limits_{\P_2}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
When $(\iota_1, \iota_2, \iota_3)=(+,+,+)$ or $(-,-,-)$, we separate the space resonance and the space-time resonance:
\begin{align*}
&\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j_1-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\\
&+\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Define the space-time resonance around $(\xi, \xi)$ as
\begin{align*}
{\rm{III}}_{23}=&-\sum\limits_{\scriptsize\begin{array}{cc}\mbox{$(\iota_1, \iota_2,\iota_3)=$}\\ \mbox{$(+, +, +)$ or $(-, -, -)$}\end{array}}\frac 13\xi t \sum\limits_{\P_2}\\
&\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j_1-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
After taking out the time derivative, we can write the space resonance term as
\begin{align*}
&
\frac13\xi t
\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\\
=&\frac{\,\mathrm{d}}{\,\mathrm{d} t}
\left(
\frac13\xi t
\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\right)
\\
&-\frac13\xi
\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\\
&-\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \frac{\,\mathrm{d}}{\,\mathrm{d} t}[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)]\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Define
\begin{align*}
&{\rm{VI}}:=-\frac {\xi}3 t \sum\limits_{\scriptsize\begin{array}{cc}\mbox{$(\iota_1, \iota_2,\iota_3)=$}\\ \mbox{$(+, +, +)$ or $(-, -, -)$}\end{array}}\sum\limits_{\P_2}\\
&\quad\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align*}
\begin{equation}\label{III241}
\begin{aligned}
&{\rm{III}}_{241}:=
-\frac 13\xi \sum\limits_{\scriptsize\begin{array}{cc}\mbox{$(\iota_1, \iota_2,\iota_3)=$}\\ \mbox{$(+, +, +)$ or $(-, -, -)$}\end{array}}\sum\limits_{\P_2}\\
&\quad\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{aligned}
\end{equation}
\begin{equation}\label{III242}
\begin{aligned}
&{\rm{III}}_{242}=-\frac 13\xi t\sum\limits_{\scriptsize\begin{array}{cc}\mbox{$(\iota_1, \iota_2,\iota_3)=$}\\ \mbox{$(+, +, +)$ or $(-, -, -)$}\end{array}}\sum\limits_{\P_2}\\
&\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\Phi} e^{it\Phi} \frac{\,\mathrm{d}}{\,\mathrm{d} t}[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)]\upsilon_{\iota_3}(\xi)[1-\psi_{j_1-3}(\eta_1+\eta_2-2\xi)]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{aligned}
\end{equation}
So, we write the equation as
\begin{align*}
\partial_t[\partial_\xi\hat h+\rm{VI}]=\rm{I}+\rm{II}+\rm{III}_1+\rm{III}_{21}+\rm{III}_{22}+\rm{III}_{23}+\rm{III}_{241}+\rm{III}_{242}+\rm{IV}+\rm{V}.
\end{align*}
We multiply the equation by $\partial_\xi\hat h+\rm{VI}$ and then take integral.
\begin{equation}\label{7.5}
\begin{aligned}
&\left|\frac{\,\mathrm{d}}{\,\mathrm{d} t}\int \frac12[\partial_\xi \hat h(\xi)+{\rm{VI}}]\overline{[\partial_\xi \hat h(\xi)+{\rm{VI}}]}\,\mathrm{d}\xi\right|\\
=~&\left|\Re\int\partial_t[\partial_\xi \hat h(\xi)+{\rm{VI}}]\overline{[\partial_\xi \hat h(\xi)+{\rm{VI}}]}\,\mathrm{d}\xi\right|
\\
=~&\left|\Re\int[{\rm{I}+\rm{II}+\rm{III}_1+\rm{III}_{21}+\rm{III}_{22}+\rm{III}_{23}+\rm{III}_{241}+\rm{III}_{242}+\rm{IV+V}}]\overline{[\partial_\xi \hat h(\xi)+{\rm{VI}}]}\,\mathrm{d}\xi\right|
\\
\lesssim~&\|[{\rm{I}+\rm{II}+\rm{III}_{21}+\rm{III}_{22}+\rm{III}_{23}+\rm{III}_{241}+\rm{III}_{242}+\rm{IV+V}}]\|_{L^2}\|\partial_\xi \hat h(\xi)+{\rm{VI}}\|_{L^2}\\
&+\left|\Re\int {\rm{III}}_1 \overline{[\partial_\xi \hat h(\xi)+{\rm{VI}}]}
d\xi
\right|.
\end{aligned}
\end{equation}
We claim that
\begin{equation}\label{claim}
\begin{aligned}
\|\partial_\xi\hat h(t, \xi)\|_{L^2}^2\lesssim ~&\|\partial_\xi\hat h(0, \xi)\|_{L^2}^2+t\|\varphi\|_{B^{1,6}}^2\|\varphi\|_{L^2}\\
&\quad+\int_0^t \|\varphi(\tau)\|_{B^{1,6}}^2 F(\|\varphi(\tau)\|_{B^{1,6}})(\|\partial_\xi\hat h(\tau, \xi)\|_{L^2}^2+\|\varphi(\tau)\|_{L^2}^2)\,\mathrm{d} \tau,
\end{aligned}\end{equation}
where $F$ is a positive polynomial.
In the following, we prove this claim by estimating each term on the right of \eqref{7.5}, and then estimating
the term $\rm{VI}$ to show that $\|\partial_\xi\hat h+\rm{VI}]\|_{L^2_\xi}$ is equivalent to $\|\partial_\xi\hat h\|_{L^2_\xi}$.
\subsection{Term $\rm{I}$ estimate}
By the symbol estimate of $\partial_{\eta_3}\mathbf{T}_1$ in Proposition \ref{Prop_symb_Tmu}, we have
\begin{align*}
&\|{\rm{I}}\|_{L^2}\\
=~&\left\|\frac13\xi \iint_{\mathbb{R}^2}\partial_{\xi}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h(\eta_1)\hat h(\eta_2)\hat h(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~& \sum\limits_{j_1,j_2,j_3} (1+2^{\min\{j_1,j_2,j_3\}+\mathrm{med}\{j_1,j_2,j_3\}})2^{4\max\{0, \mathrm{med}\{j_1,j_2,j_3\}\}}\\
&\qquad \cdot\|\partial_x\varphi_{\min\{j_1,j_2,j_3\}}\|_{L^\infty}\|\partial_x\varphi_{\mathrm{med}\{j_1,j_2,j_3\}}\|_{L^\infty}\|\varphi_{\max\{j_1,j_2,j_3\}}\|_{H^1}\\
\lesssim~& \left(\sum\limits_{j\in\mathbb{Z}}\|\partial_x\varphi_j\|_{L^{\infty}}\right)^2\|\varphi\|_{H^{7}}\\
\lesssim~&\|\varphi\|_{B^{1,6}}^2\|\varphi\|_{H^7}.
\end{align*}
\subsection{Term $\rm{II}$ estimate}
By the definition of $\P$, $j_1=\max\{j_1, j_2, j_3\}$.
\begin{align*}
&\left\|\frac13\xi \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)e^{it\Phi(\eta_1,\eta_2,\xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\xi}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~&2^{j_1}\|\partial_\xi\hat h_{j_3}\|_{L^2}\|\mathbf{T}_1(\eta_1,\eta_2, \eta_3)\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\|_{S^{\infty}}\| h_{j_1}\|_{L^\infty}\| h_{j_2}\|_{L^\infty}\\
\lesssim~&2^{j_1+j_2+j_3}(1+2^{j_2})(1+2^{j_3})2^{2\max\{j_2, j_3, 0\}}\|\partial_\xi\hat h_{j_3}\|_{L^2}\| \varphi_{j_1}\|_{L^\infty}\| \varphi_{j_2}\|_{L^\infty}\\
\lesssim~&2^{(j_1+j_2)/2}\|\partial_\xi\hat h_{j_3}\|_{L^2}\|\partial_x\varphi_{j_1}\|_{L^{\infty}}\| \partial_x\varphi_{j_2}\|_{W^{4,\infty}}.
\end{align*}
By taking the summation for $(j_1, j_2, j_3)\in \P$,
\begin{align*}
&\sum\limits_{\P}\left\|\frac13 \xi\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi(\eta_1,\eta_2,\xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\xi}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~&\|\varphi\|_{B^{1,6}}^2\|\partial_\xi\hat h(\xi)\|_{L^2_\xi}.
\end{align*}
\subsection{Term $\rm{III}$'s estimates}
We estimate $\rm{III}_1, \rm{III}_{21}, \rm{III}_{22}, \rm{III}_{23}, \rm{III}_{241}, \rm{III}_{242}$ separately. Among these terms, $\rm{III}_1, \rm{III}_{21}$ are the terms away from resonances. $\rm{III}_{22}, \rm{III}_{23}$ are the terms near the space-time resonances. $\rm{III}_{241}, \rm{III}_{242}$ are related to the terms close to the space resonance.
\subsubsection{$\rm{III}_1$ estimate}
When $(j_1, j_2, j_3)\in \P_1$, $j_1-j_3>1$, this is the nonresonant case.
\begin{align}
\label{eqn7_1}
&\frac 13\xi t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\nonumber
\\
\nonumber
=~&\frac 13\xi \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \partial_{\eta_1}e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\\
=~&-\frac 13\xi \iint_{\mathbb{R}^2}\partial_{\eta_1}\left[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \right]e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\\
&-\frac 13\xi \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2
\nonumber
\\
\nonumber
&-\frac 13\xi \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align}
where
$\Phi$ is defined in \eqref{defPhi} and
\begin{align*}
\partial_\xi\Phi(\eta_1,\eta_2,\xi)=-\frac{1}{[1+(\xi-\eta_1-\eta_2)^2]^{3/2}}+\frac{1}{(1+\xi^2)^{3/2}},\\
\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)=-\frac{1}{(1+\eta_1^2)^{3/2}}+\frac{1}{[1+(\xi-\eta_1-\eta_2)^2]^{3/2}}.
\end{align*}
By direct calculation
\begin{equation}\label{eqn7_2}
\begin{aligned}
&\partial_{\eta_1}\left[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \right]\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\xi-\eta_1-
\eta_2)\\
=~&(\partial_{1}\mathbf{T}_1-\partial_3\mathbf{T}_1)(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\xi-\eta_1-
\eta_2)\\
&+\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\eta_1}\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\xi-\eta_1-
\eta_2)\\
&-\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\eta_1}^2\Phi\partial_{\xi}\Phi}{i(\partial_{\eta_1}\Phi)^2} \psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\xi-\eta_1-
\eta_2).
\end{aligned}
\end{equation}
The first term on the right-hand-side of equation \eqref{eqn7_1} is a Fourier multilinear integral with the symbol
\begin{align*}
&-i\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\\
&\cdot\Big[(\partial_{1}\mathbf{T}_1-\partial_3\mathbf{T}_1)(\eta_1,\eta_2, \eta_3)\frac{\frac{1}{[1+\eta_3^2]^{3/2}}-\frac{1}{(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}}}{\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}}}
+\mathbf{T}_1(\eta_1,\eta_2, \eta_3)\frac{\frac{3\eta_3}{2[1+\eta_3^2]^{5/2}}}{\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}}}\\
&\qquad -\mathbf{T}_1(\eta_1,\eta_2, \eta_3)\frac{\left[-\frac{3\eta_1}{(1+\eta_1^2)^{5/2}}-\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\left[\frac{1}{[1+\eta_3^2]^{3/2}}-\frac{1}{(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}}\right]}{\left[\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}}\right]^2}\Big].
\end{align*}
By direct calculation, we have the following estimates for the symbols
\begin{align}
\label{symbolphixi}
&\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\left[\frac{1}{[1+\eta_3^2]^{3/2}}-\frac{1}{(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}}\right]\right\|_{S^\infty}
\hskip-0.2in
\lesssim
2^{2j_1}(1+2^{2j_1})^{-1}(1+2^{3j_3})^{-1},
\end{align
\begin{align}\label{symbolpartialphi}
\left\|
\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)
\left[\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}} \right]^{-1}\right\|_{S^\infty}
\lesssim&
(1+2^{3j_3})(1+2^{-2j_1}),
\end{align}
\begin{align}\label{symbol2d}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\left[-\frac{3\eta_1}{(1+\eta_1^2)^{5/2}}-\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\right\|_{S^\infty}
\lesssim&
2^{j_1}(1+2^{j_1})^{-1}(1+2^{j_3})^{-4}
,
\end{align}
\begin{align*}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\left[\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\right\|_{S^\infty}\lesssim 2^{j_3}(1+2^{j_3})^{-5}.
\end{align*}
By the algebraic property of $S^\infty$ norm, combining the above estimates and Proposition \ref{Prop_symb_Tmu},
\begin{align*}
&\left\|\Big\{\partial_{\eta_1}[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi}]\Big\}_{\xi=\eta_1+\eta_2+\eta_3}\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\right\|_{S^\infty}\\
\lesssim~& [(1+2^{j_3})(1+2^{5j_2})2^{j_2+j_3}][1+2^{3j_3}](1+2^{3j_3})^{-1}\\
& +[(1+2^{j_3})(1+2^{3j_2})2^{j_2+j_3}][2^{j_3}(1+2^{j_3})^{-1}][(1+2^{3j_3})(1+2^{-2j_1})]\\
&+[(1+2^{j_3})(1+2^{3j_2})2^{j_2+j_3}][2^{j_1}(1+2^{j_1})^{-1}][2^{2j_1}(1+2^{2j_1})^{-1}(1+2^{3j_3})^{-1}][(1+2^{3j_3})(1+2^{-2j_1})]^2\\
\lesssim~&2^{j_2+j_3}(1+2^{5j_3})(1+2^{5j_2})(1+2^{-j_1}).
\end{align*}
The first term of equation \eqref{eqn7_1} satisfies
\begin{align*}
\left\|\frac13\xi\iint_{\mathbb{R}^2}\partial_{\eta_1}\left[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \right]e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim \|\varphi_{j_1}\|_{H^{1}}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
The second term of equation \eqref{eqn7_1} is
\begin{align*}
&\xi\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
For this term, we consider the following two cases:
Case 1: $j_2\le j_1\le j_2+1$,
and
Case 2: $j_2< j_1-1$.
\vskip.1in
\noindent
{\bf Estimate in Case 1.} Since $|\xi|\approx 2^{j_2}\approx 2^{j_1}>2^{j_3+1}$, using Proposition \ref{Prop_symb_Tmu}
we have
\begin{align*}
&\left\|
\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)
(\eta_1+\eta_2+\eta_3)\mathbf{T}_1(\eta_1,\eta_2,\eta_3)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi}
\right\|_{S^{\infty}}
\\
\lesssim~&
2^{j_2}
\|
\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)
\mathbf{T}_1(\eta_1,\eta_2,\eta_3)
\|_{S^{\infty}}
\lesssim
2^{(2j_2+j_3)}(1+2^{3 j_2}) (1+2^{j_3}),
\end{align*}
which implies that
\begin{align*}
&
\Big|\Re\int_{\mathbb{R}} \overline{[\partial_\xi \hat h(\xi)]}\iint_{\mathbb{R}^2}\xi\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \psi_{j_1}(\eta_1)\partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi
\Big|
\\
\lesssim~&
\|\psi_{j_1}(\xi)\partial_\xi\hat h(\xi)\|_{L^2}\|\tilde\psi_{j_1}(\xi)\partial_\xi\hat h(\xi)\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
{\bf Estimate in Case 2.} In this case, the second term of equation \eqref{eqn7_1} has the trouble of loss of derivatives in the high frequency part. One can transfer one-order of the high frequency derivative to the low frequency by doing the symmetrization. Thus we need to estimate
\begin{align*}
&\Re\int_{\mathbb{R}} \overline{[\partial_\xi \hat h(\xi)]}\iint_{\mathbb{R}^2}\xi\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \psi_{j_1}(\eta_1)\partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi
\\
=~&-\Im \iiint_{\mathbb{R}^3}\xi\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)} \psi_{j_1}(\eta_1)\\
& \cdot e^{it\Phi(\eta_1,\eta_2,\xi)}\overline{[\partial_\xi \hat h(\xi)]}\partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi\\
=~-&\Im\iiint_{\mathbb{R}^3}-\eta_1 \mathbf{T}_1(-\xi,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)} {p'(\xi-\eta_1-\eta_2)-p'(\xi)}\psi_{j_1}(\xi) \\
&\cdot e^{it\Phi(\eta_1,\eta_2,\xi)}\overline{[\partial_\xi \hat h(\xi)]} \partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi\\
=~-&\Im\iiint_{\mathbb{R}^3} m(\eta_1, \eta_2, \xi-\eta_1-\eta_2))e^{it\Phi(\eta_1,\eta_2,\xi)}\overline{[\partial_\xi \hat h(\xi)]} \partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi,
\end{align*}
where
\begin{align*}
&m(\eta_1, \eta_2, \xi-\eta_1-\eta_2)\\
=~&\frac12\Bigg[ -\eta_1\mathbf{T}_1(-\xi,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}{p'(\xi-\eta_1-\eta_2)-p'(\xi)}\psi_{j_1}(\xi)\\
&\qquad+ \xi\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}\psi_{j_1}(\eta_1)\Bigg]\\
=~& \frac12 \Bigg[-\eta_1\big(\mathbf{T}_1(-\xi,\eta_2,\xi-\eta_1-\eta_2)-\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\big)\frac{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}{p'(\xi-\eta_1-\eta_2)-p'(\xi)}\psi_{j_1}(\xi)\\
&~~\quad +\eta_1 \mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\left(\frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}-\frac{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}{p'(\xi-\eta_1-\eta_2)-p'(\xi)}\right)\psi_{j_1}(\xi)\\
&~~\quad +\eta_1 \mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}(\psi_{j_1}(\eta_1)-\psi_{j_1}(\xi))\\
&~~\quad+(\xi-\eta_1)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}\psi_{j_1}(\eta_1)\Bigg]\\
=:~&m_1+m_2+m_3+m_4.
\end{align*}
We will estimate the $S^\infty$ norm of each symbol in the dyadic support.
For the symbol $m_1$, using \eqref{Tmu_est5},
\begin{align*}
\|m_1\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}&\lesssim 2^{j_1}[2^{j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})]\\
&\lesssim 2^{j_2+j_3}(1+2^{j_2})^3(1+2^{j_3}).
\end{align*}
For the symbol $m_2$, we calculate the difference of the fractions
\[
\frac{p'(\eta_3)-p'(\eta_1+\eta_2+\eta_3)}{p'(\eta_1)-p'(\eta_3)}-\frac{p'(\eta_1)-p'(\eta_3)}{p'(\eta_3)-p'(\eta_1+\eta_2+\eta_3)}
=:\frac{\mathcal{A}}{\mathcal{B}},
\]
where
\begin{align*}
\mathcal{A}=~&[(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}-(1+\eta_3^2)^{3/2}]^2(1+\eta_1^2)^{3}-[(1+\eta_3^2)^{3/2}-(1+\eta_1^2)^{3/2}]^2(1+(\eta_1+\eta_2+\eta_3)^2)^{3}\\
=~&(1+\eta_3^2)^{3}(1+\eta_1^2)^{3}-2(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}(1+\eta_3^2)^{3/2}(1+\eta_1^2)^{3}\\
&-(1+\eta_3^2)^{3}(1+(\eta_1+\eta_2+\eta_3)^2)^{3}+2(1+\eta_3^2)^{3/2}(1+\eta_1^2)^{3/2}(1+(\eta_1+\eta_2+\eta_3)^2)^{3}\\
=~&(1+\eta_3^2)^{3}[(1+\eta_1^2)^{3}-(1+(\eta_1+\eta_2+\eta_3)^2)^{3}]\\
&+2(1+\eta_3^2)^{3/2}(1+\eta_1^2)^{3/2}(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}[(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}-(1+\eta_1^2)^{3/2}],
\end{align*}
\begin{align*}
\mathcal{B}=~[(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}-(1+\eta_3^2)^{3/2}](1+\eta_1^2)^{3/2}[(1+\eta_3^2)^{3/2}-(1+\eta_1^2)^{3/2}](1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}.
\end{align*}
It's easy to estimate the $S^\infty$ norm of $\mathcal{A}$ and $\mathcal{B}$ as
\[
\|\mathcal{A}\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}\lesssim 2^{2j_1+2j_2}(1+2^{j_1})^7(1+2^{j_2}),
\
\[
\|\mathcal{B}\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}\approx (1+2^{j_1})^{8}2^{4j_1}.
\
Therefore, by algebraic property of $S^\infty$ norm,
\[
\Big\|\frac{\mathcal{A}}{\mathcal{B}}\psi_{j_1}\psi_{j_2}\psi_{j_3}\Big\|_{S^\infty}\lesssim (1+2^{j_1})^{-1}(1+2^{j_2}).
\]
By above estimate and proposition \ref{Prop_symb_Tmu},
\begin{align*}
\|m_2\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}&\lesssim 2^{j_1}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})][(1+2^{j_1})^{-1}(1+2^{j_2})]\\
&\lesssim 2^{j_2+j_3}(1+2^{j_2})^4(1+2^{j_3}).
\end{align*}
For the symbol $m_3, m_4$, they satisfy the estimates
\begin{align*}
\|m_3\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}&\lesssim 2^{j_1}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})]2^{-j_1}\\
&\lesssim 2^{j_2+j_3}(1+2^{j_2})^3(1+2^{j_3}),
\end{align*}
\begin{align*}
\|m_4\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}&\lesssim 2^{j_2}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})]\\
&\lesssim 2^{2j_2+j_3}(1+2^{j_2})^3(1+2^{j_3}).
\end{align*}
Therefore the $S^\infty$-norm of $m(\eta_1, \eta_2, \eta_3)\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)$ satisfies
\begin{align*}
&\|m\psi_{j_1}\psi_{j_2}\psi_{j_3}\|_{S^\infty}\lesssim 2^{j_2+j_3}(1+2^{j_2})^4(1+2^{j_3}).
\end{align*}
The integral is bounded by
\begin{align*}
&\left|\Re\int_{\mathbb{R}} \overline{[\partial_\xi \hat h(\xi)]}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \psi_{j_1}(\eta_1)\partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi\right|\\
\lesssim~& \|\psi_{j_1}(\xi)\partial_\xi\hat h(\xi)\|_{L^2}\|\tilde\psi_{j_1}(\xi)\partial_\xi\hat h(\xi)\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
The other term for $\rm{III}_{1}$ estimate in \eqref{7.5} is
\begin{align*}
&\left|\Re\int_{\mathbb{R}} \overline{\rm{VI}(\xi)}\iint_{\mathbb{R}^2}\xi \mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \psi_{j_1}(\eta_1)\partial_{\eta_1}\hat h(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d}\xi\right|\\
\lesssim~& (1+2^{j_1})\|\tilde \psi_{j_1}(\xi){\rm{VI}}\|_{L^2}\|\tilde\psi_{j_1}\partial_\xi\hat h\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
The fourth term in \eqref{eqn7_1} satisfies
\begin{align*}
&\left|\iiint_{\mathbb{R}^3} \frac13\overline{[\partial_\xi \hat h(\xi)]}\xi\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\,\mathrm{d} \xi \right|\\
\lesssim~& \|\partial_\xi\hat h_{j_3}\|_{L^2_\xi}\|\partial_\xi\hat h \tilde \psi_{j_1}\|_{L^2} \|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_1}\|_{B^{1,6}}.
\end{align*}
\subsubsection{Term $\rm{III}_{21}$ estimate}
We will estimate the $L^2$-norm of
\begin{align*}
\frac \xi3 t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align*}
when $(\iota_1, \iota_2, \iota_3)=(+,+,-)$ or $(-, -, +)$ and $j_1, j_2, j_3\in \P_2$.
In the support of the integrand, $|\eta_1^2-(\xi-\eta_1-\eta_2)^2|\gtrsim \min\{1, 2^{2j_1}\}$. This is away from resonances.
By the same operation as \eqref{eqn7_1} and \eqref{eqn7_2}, we need to estimate
\begin{equation}\label{eqn714}
\begin{aligned}
&\frac\xi3 \iint_{\mathbb{R}^2}\partial_{\eta_1}[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} ]e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{aligned}\end{equation}
The symbols satisfy
\begin{align*}
&\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\left[\frac{1}{[1+\eta_3^2]^{3/2}}-\frac{1}{(1+(\eta_1+\eta_2+\eta_3)^2)^{3/2}}\right]\right\|_{S^\infty}\\
\lesssim~& 2^{2j_1}(1+2^{5j_1})^{-1},
\end{align*
\begin{align}\nonumber
&\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\left[\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}} \right]^{-1}\right\|_{S^\infty}\\\label{eqn715}
\lesssim~ &(1+2^{2j_1})^{5/2}\max\{1, 2^{-2j_1}\}\lesssim (1+2^{j_1})^{7}2^{-2j_1},
\end{align}
\begin{align}\label{eqn716}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\left[-\frac{3\eta_1}{(1+\eta_1^2)^{5/2}}-\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\right\|_{S^\infty}\lesssim 2^{j_1}(1+2^{j_1})^{-5},
\end{align}
\begin{align}\label{eqn717}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\left[\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\right\|_{S^\infty}\lesssim 2^{j_1}(1+2^{j_1})^{-5}.
\end{align}
Thus
\begin{align*}
&\left\|\frac\xi3 \iint_{\mathbb{R}^2}\partial_{\eta_1}\left[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} \right]e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~& \left\|\frac\xi3 \iint_{\mathbb{R}^2}\partial_{\eta_1}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
& +\left\|\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi \partial_{\eta_1}^2\Phi}{i(\partial_{\eta_1}\Phi)^2} e^{it\Phi}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
& +\left\|\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\partial_{\eta_1}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~&2^{j_1}\Big\{[2^{2j_1} (1+2^{j_1})^6][(1+2^{2j_1})]\\
&+[2^{3j_1}(1+2^{j_1})^3][2^{2j_1}(1+2^{5j_1})^{-1}][(1+2^{7j_1})2^{-2j_1}]^2[2^{j_1}(1+2^{5j_1})^{-1}]\\
&+[2^{3j_1}(1+2^{j_1})^3] [(1+2^{7j_1})2^{-2j_1}][2^{j_1}(1+2^{5j_1})^{-1}]\Big\}\cdot\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
Similarly, the other two integrals satisfy
\begin{align*}
&\left\|\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
&+\left\|\frac\xi3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{\partial_{\xi}\Phi}{i\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~&2^{j_1}[2^{3j_1}(1+2^{j_1})^3][(1+2^{2j_1})]
\big(\|\partial_\xi\hat h_{j_1}(\xi)\|_{L^2_\xi}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^\infty}+\|\partial_\xi\hat h_{j_3}(\xi)\|_{L^2_\xi}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_1}\|_{L^\infty}\big)\\
\lesssim~&\|\partial_{\xi}\hat h_{j_1}(\xi)\|_{L^2_\xi}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}+\|\partial_{\xi}\hat h_{j_3}\|_{L^2_\xi}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_1}\|_{B^{1,6}}.
\end{align*}
Then after taking the summation for the related sub-indices $(j_1, j_2, j_3)\in\P_2$, we obtain
\begin{align*}
\|{\rm{III}}_{21}\|_{L^2}\lesssim \|\varphi\|_{B^{1,6}}^2(\|\varphi\|_{L^2}+\|\partial_\xi\hat h(\xi)\|_{L^2_\xi}).
\end{align*}
\subsubsection{Term $\rm{III}_{22}$ estimate}
In this case, $(\eta_1, \eta_2)$ is around $(\xi, -\xi)$ or $(-\xi, \xi)$. Since $\eta_1$ and $\eta_2$ are symmetric in the integral, we assume $(\eta_1, \eta_2)$ is around $(\xi, -\xi)$ in the following. Under this assumption, $0\leq |\eta_1+\eta_2|\lesssim 2^{j_1}$, $|\eta_1-\eta_2|\approx 2^{j_1}$.
Since
\begin{equation}\label{eqn718}
(\partial_{\eta_1}-\partial_{\eta_2})e^{it\Phi(\eta_1, \eta_2, \xi)}=(p'(\eta_1)-p'(\eta_2))ite^{it\Phi(\eta_1, \eta_2, \xi)},
\end{equation}
we have
\begin{align*}
&-\frac\xi3 t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&\frac {i\xi}3 \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi \frac{1}{p'(\eta_1)-p'(\eta_2)}(\partial_{\eta_1}-\partial_{\eta_2})e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
=~&\frac {i\xi}3 \iint_{\mathbb{R}^2}e^{it\Phi} (\partial_{\eta_1}-\partial_{\eta_2})\left[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\eta_2)}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\right]\nonumber\\
&\qquad\cdot\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
=~&\frac {i\xi}3 \iint_{\mathbb{R}^2}e^{it\Phi} \Big[(\partial_1-\partial_2)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\eta_2)}\\
&+\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{(p''(\eta_1)+p''(\eta_2))\left(p'(\xi-\eta_1-\eta_2)-p'(\xi)\right)}{(p'(\eta_1)-p'(\eta_2))^2}\Big]\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\nonumber\\
&\qquad\cdot\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
&+\frac {i\xi}3 \iint_{\mathbb{R}^2}e^{it\Phi}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p'(\xi-\eta_1-\eta_2)-p'(\xi)}{p'(\eta_1)-p'(\eta_2)} (\partial_1-\partial_2)\left[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\right]\\
&\qquad\cdot\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
=:~& I_1+I_2.
\end{align*}
By estimates \eqref{eqnB7} and \eqref{eqnB8}, we obtain the bounds
\begin{align*}
\|I_1\|_{L^2}&\lesssim 2^{2j_1}(1+2^{j_1})^6 \|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
&\lesssim \|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2},
\end{align*}
and
\begin{align*}
\|I_2\|_{L^2}&\lesssim (2^{4j_1}(1+2^{j_1})^6 )\cdot(\|\partial_\xi\hat h_{j_1}(\xi)\|_{L^2_\xi}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^\infty}+\|\partial_\xi\hat h_{j_2}(\xi)\|_{L^2_\xi}\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_3}\|_{L^\infty})\\
&\lesssim \|\partial_{\xi}\hat h_{j_1}(\xi)\|_{L^2_\xi}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}+\|\partial_{\xi}\hat h_{j_2}\|_{L^2_\xi}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
Combining the estimate of each term, we obtain
\begin{equation}
\|{\rm{III}}_{22}\|_{L^2}\lesssim \|\varphi\|_{B^{1,6}}^2(\|\partial_\xi\hat h(t,\xi)\|_{L^2}+\|\varphi\|_{L^2}).
\end{equation}
\subsubsection{Term $\rm{III}_{23}$ estimate}
$\rm{III}_{23}$ is the space-time resonance around $(\xi, \xi)$. We will estimate the $L^2$-norm of
\begin{align*}
\frac \xi3 t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align*}
when $(\iota_1, \iota_2,\iota_3)=(+, +, +)$ or $(-, -, -)$ and $(j_1, j_2, j_3)\in\P_2$.
After integration by part, we split the integral into several terms.
\begin{align*}
&t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\xi}\Phi e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~& t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)[p'(\xi-\eta_1-\eta_2)-p'(\eta_1)] e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\\
&\qquad\cdot \psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
& +t \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)[p'(\eta_1)-p'(\xi)] e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\\
&\qquad\cdot \psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&i \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\partial_{\eta_1} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+i \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} \partial_{\eta_2}e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\\
&\qquad\qquad\psi_{j-3}(\eta_1+\eta_2-2\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&i \iint_{\mathbb{R}^2}\partial_{\eta_1}[\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)] e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+i \iint_{\mathbb{R}^2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+i \iint_{\mathbb{R}^2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&-i \iint_{\mathbb{R}^2}\partial_{\eta_2}[\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)]\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} \\
&\qquad \cdot e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+i \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{(p'(\eta_1)-p'(\xi))[p''(\eta_2)+p''(\xi-\eta_1-\eta_2)]}{(p'(\eta_2)-p'(\xi-\eta_1-\eta_2))^2} e^{it\Phi}\\
&\qquad\cdot \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\psi_{j-3}(\eta_1+\eta_2-2\xi)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&-i \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\psi_{j-3}(\eta_1+\eta_2-2\xi)\upsilon_{\iota_3}(\xi)\\
&\qquad\cdot \left[\partial_{\eta_2}\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)+h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_2}\hat h_{j_3}(\xi-\eta_1-\eta_2)\right]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Then we estimate each term. By the symbol estimate Proposition \ref{Prop_symb_Tmu},
\begin{align*}
&\left\|\xi \iint_{\mathbb{R}^2}\psi_{j_3}(\eta_1+\eta_2-2\xi)\partial_{\eta_1}[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)] e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim ~& 2^{3j_1}(1+2^{j_1})^6 \|\varphi_{j_1}\|_{L^\infty} \|\varphi_{j_2}\|_{L^\infty} \|\varphi_{j_3}\|_{L^2}\\
\lesssim ~& \|\varphi_{j_1}\|_{B^{1,6}} \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\| \xi \iint_{\mathbb{R}^2}\partial_{\eta_1}\psi_{j_3}(\eta_1+\eta_2-2\xi)[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)] e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim ~& 2^{4j_1}(1+2^{j_1})^4 2^{-j_1} \|\varphi_{j_1}\|_{L^\infty} \|\varphi_{j_2}\|_{L^\infty} \|\varphi_{j_3}\|_{L^2}\\
\lesssim ~& \|\varphi_{j_1}\|_{B^{1,6}} \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\|\xi\iint_{\mathbb{R}^2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~&2^{4j_1}(1+2^{j_1})^4 \|\varphi_{j_2}\|_{L^\infty} \|\varphi_{j_3}\|_{L^\infty} \|\partial_{\xi}\hat h_{j_1}(\xi)\|_{L^2_\xi}\\
\lesssim~& \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{B^{1,6}} \|\partial_{\xi}\hat h_{j_1}(\xi)\|_{L^2_\xi},
\end{align*}
\begin{align*}
&\left\|\xi\iint_{\mathbb{R}^2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^2}\\
\lesssim~& \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_1}\|_{B^{1,6}} \|\partial_{\xi}\hat h_{j_3}(\xi)\|_{L^2_\xi}.
\end{align*}
By Proposition \ref{Prop_symb_Tmu} and estimates \eqref{eqnB9},
\begin{align*}
&\Bigg\|\xi\iint_{\mathbb{R}^2}\partial_{\eta_2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} \\
&\qquad \cdot e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\Bigg\|_{L^2}\\
\lesssim ~& \|\varphi_{j_1}\|_{B^{1,6}} \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\Bigg\|\xi\iint_{\mathbb{R}^2}\psi_{j-3}(\eta_1+\eta_2-2\xi)\partial_{\eta_2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} \\
&\qquad \cdot e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\Bigg\|_{L^2}\\
\lesssim ~& \|\varphi_{j_1}\|_{B^{1,6}} \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\Bigg\|\xi\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{p'(\eta_1)-p'(\xi)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\psi_{j-3}(\eta_1+\eta_2-2\xi)\upsilon_{\iota_3}(\xi)\\
&\qquad\cdot \left[\partial_{\eta_2}\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)+\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_2}\hat h_{j_3}(\xi-\eta_1-\eta_2)\right]\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\Bigg\|_{L^2}\\
\lesssim~&\|\varphi_{j_1}\|_{B^{1,6}}( \|\varphi_{j_3}\|_{B^{1,6}} \|\partial_{\xi}\hat h_{j_2}(\xi)]\|_{L^2_\xi}+\|\varphi_{j_2}\|_{B^{1,6}} \|\partial_{\xi}\hat h_{j_3}(\xi)\|_{L^2_\xi}).
\end{align*}
By Proposition \ref{Prop_symb_Tmu}, estimates \eqref{eqnB9} and \eqref{eqnB10},
\begin{align*}
&\Bigg\| \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{(p'(\eta_1)-p'(\xi))[p''(\eta_2)+p''(\xi-\eta_1-\eta_2)]}{(p'(\eta_2)-p'(\xi-\eta_1-\eta_2))^2} e^{it\Phi}\\
&\qquad\cdot \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\psi_{j-3}(\eta_1+\eta_2-2\xi)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\Bigg\|_{L^2}\\
\lesssim ~& \|\varphi_{j_1}\|_{B^{1,6}} \|\varphi_{j_2}\|_{B^{1,6}} \|\varphi_{j_3}\|_{L^2}.
\end{align*}
Combining the estimate of each term, we obtain
\begin{equation}
\|{\rm{III}}_{23}\|_{L^2}\lesssim \|\varphi\|_{B^{1,6}}^2(\|\partial_\xi\hat h(t,\xi)\|_{L^2}+\|\varphi\|_{L^2}).
\end{equation}
\subsubsection{Terms $\rm{III}_{241}$ and $\rm{III}_{242}$ estimates}
By the definition \eqref{III241} and the symbol estimates Proposition \ref{Prop_symb_Tmu} and Proposition \ref{PropB3},
\begin{align*}
\|(1+|\xi|)^r{\rm{III}_{241}}\|_{L^2}&\lesssim \sum\limits_{(j_1,j_2,j_3)\in\P_{2}}2^{2j_1}(1+2^{j_1})^4 \|\varphi_{j_1}\|_{L^\infty} \|\varphi_{j_2}\|_{L^\infty}\|\hat h_{j_3}(\xi)\|_{L^2_\xi}\\
&\lesssim \|\varphi\|_{B^{1,6}}^{2} \|\varphi\|_{L^{2}}.
\end{align*}
Since
\begin{align*}
\|\partial_t \hat h_j\|_{L^2}\lesssim& \sum\limits_{\mu=1}^\infty\sum\limits_{j_1, \cdots, j_{2\mu+1}}2^{(j_2+\cdots+j_{2\mu+1})}\prod\limits_{k=2}^{2\mu+1} (1+2^{j_k})2^{2\max\{0, j_2, \cdots, j_{2\mu+1}\}} \|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\cdots\|\varphi_{j_{2\mu+1}}\|_{L^\infty}\\
\lesssim& \sum\limits_{\mu=1}^\infty \|\varphi\|_{B^{1,6}}^{2\mu}\|\varphi\|_{H^1},
\end{align*}
by the definition \eqref{III242} and the symbol estimates Proposition \ref{Prop_symb_Tmu} and Proposition \ref{PropB3},
\begin{align*}
\|{\rm{III}_{242}}\|_{L^2}&\lesssim t \sum\limits_{(j_1,j_2,j_3)\in\P_{2}}2^{2j_1}(1+2^{j_1})^4 \|\varphi_{j_1}\|_{L^\infty} \|\varphi_{j_2}\|_{L^\infty}\|\partial_t \hat h_{j_3}\|_{L^2}\\
&\lesssim t\|\varphi\|_{B^{1,6}}^2\sum\limits_{\mu=1}^\infty \|\varphi\|_{B^{1,6}}^{2\mu} \|\varphi\|_{H^{r+1}}.
\end{align*}
\subsection{Term $\rm{V}$ estimate}
Now we estimate the term
\begin{align*}
{\rm{V}}=~&\partial_\xi[\xi e^{-it\frac{\xi}{\sqrt{1+\xi^2}}}\widehat{\mathcal{N}_{\geq5}(\varphi)}],
\end{align*}
where $e^{-it\frac{\xi}{\sqrt{1+\xi^2}}}\widehat{\mathcal{N}_{\geq5}(\varphi)}$ is the sum of the following terms with $\mu=2,3, \cdots$,
\begin{align*}
e^{-it\frac{\xi}{\sqrt{1+\xi^2}}}&\widehat{\mathcal{N}_{2\mu+1}}(\xi)=\frac1{2\mu+1}i\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})\\
&\cdot e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\hat h(\eta_1)\cdots\hat h(\eta_{2\mu})\hat h(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\end{align*}
By dyadic decomposition, this integral can be written as the sum of
\begin{align*}
&\mathfrak N_{2\mu+1}^{j_1\cdots j_{2\mu+1}}(\xi)\\
=~&\frac1{2\mu+1}i\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\\
&\cdot \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}},
\end{align*}
for $j_1\ge j_2\ge \cdots\ge j_{2\mu+1}$, with possible repetetion.
Then
\begin{align*}
&\partial_\xi\mathfrak N_{2\mu+1}^{j_1\cdots j_{2\mu+1}}(\xi)\\
=~&
i\iiint_{\mathbb{R}^{{2\mu}}} \partial_{2\mu+1}\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\\
&\qquad\cdot \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\\
&+i\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\\
&\qquad\cdot(\xi-\eta_1-\cdots-\eta_{2\mu})\hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\partial_\xi\frac{\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})}{\xi-\eta_1-\cdots-\eta_{2\mu}}\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\\
&+i\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\\
&\qquad\cdot \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\frac{\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})}{\xi-\eta_1-\cdots-\eta_{2\mu}}\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\\
&-t\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\\
&\quad\cdot \left[(1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2)^{-3/2}-(1+\xi^2)^{-3/2}\right] \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}.
\end{align*}
By the symbol estimates Proposition \ref{Prop_symb_Tmu} and the symbol $(1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2)^{-3/2}-(1+\xi^2)^{-3/2}$ is bounded and smooth, we have the following estimate for each term.
\begin{align*}
&\left\|\xi\iiint_{\mathbb{R}^{{2\mu}}} \partial_{2\mu+1}\mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\right.\\
&\qquad\left.\cdot \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\right\|_{L^2}\\
\lesssim~ & \|\varphi_{j_1}\|_{H^1}\|\varphi_{j_2}\|_{B^{1,6}}\cdots\|\varphi_{j_{2\mu+1}}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|\xi\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\right.\\
&\qquad\left.\hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\partial_\xi \hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\right\|_{L^2}\\
\lesssim~ &\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{B^{1,6}}\cdots\|\varphi_{j_{2\mu}}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_{2\mu+1}}(\xi)\|_{L^2_\xi},
\end{align*}
\begin{align*}
&\left\| it\xi\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\right.\\
&~~\left.\left[(1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2)^{-3/2}-(1+\xi^2)^{-3/2}\right] \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\right\|_{L^2}\\
\lesssim~ &t\|\varphi_{j_1}\|_{H^1}\|\varphi_{j_2}\|_{B^{1,6}}\cdots\|\varphi_{j_{2\mu+1}}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|\xi\iiint_{\mathbb{R}^{{2\mu}}} \mathbf{T}_{\mu}(\eta_1,\cdots,\eta_{2\mu}, \xi-\eta_1-\cdots-\eta_{2\mu})e^{it\left[\frac{\eta_1}{\sqrt{1+\eta_1^2}}+\cdots+\frac{\eta_{2\mu}}{\sqrt{1+\eta_{2\mu}^2}}+\frac{\xi-\eta_1-\cdots-\eta_{2\mu}}{\sqrt{1+(\xi-\eta_1-\cdots-\eta_{2\mu})^2}}-\frac{\xi}{\sqrt{1+\xi^2}}\right]}\right.\\
&\qquad\left.\cdot \hat h_{j_1}(\eta_1)\cdots\hat h_{j_{2\mu}}(\eta_{2\mu})\hat h_{j_{2\mu+1}}(\xi-\eta_1-\cdots-\eta_{2\mu})\,\mathrm{d} \eta_1\cdots\,\mathrm{d} \eta_{{2\mu}}\right\|_{L^2}\\
\lesssim~ &\|\varphi_{j_1}\|_{H^1}\|\varphi_{j_2}\|_{B^{1,6}}\cdots\|\varphi_{j_{2\mu}}\|_{B^{1,6}}\|\varphi_{j_{2\mu+1}}\|_{B^{1,6}}.
\end{align*}
Combining the above estimates, by taking summation for suitable sub-indices,
\[
\|{\rm{V}}\|_{L^2}\lesssim t\sum\limits_{\mu=2}^\infty \|\varphi\|_{B^{1,6}}^{2\mu}(\|\varphi\|_{H^1}+\|\partial_\xi\hat h(\xi)\|_{L^2}).
\]
\subsection{Term $\rm{VI}$ estimate}
The term $\rm{VI}$ is bounded by
\begin{align*}
\|{\rm{VI}}\|_{L^2}&\lesssim \sum\limits_{(j_1,j_2,j_3)\in\P_{2}}2^{2j_1}(1+2^{j_1})^2 2^{2\max\{0, j_1\}} \|\varphi_{j_1}\|_{L^\infty} \|\varphi_{j_2}\|_{L^\infty}\| \varphi_{j_3}\|_{L^2}\\
&\lesssim \|\varphi\|_{B^{1,6}}^{2} \|\varphi\|_{H^{6}}.
\end{align*}
Combining all the estimates above, we have proved the claim \eqref{claim}.
\section{$Z$-norm estimates}\label{sec:Znorm}
In this section, we prove Lemma \ref{Z-norm_Est}. We will estimate $|(|\xi|^{r}+|\xi|^{w})\hat\varphi(t,\xi)|$ for each fixed $\xi$. According to different values of $\xi$, the estimate is done in different ways.
\subsection{Large and small frequencies}
\label{sec:large}
When $|\xi|$ is very big or small, we can estimate $|(|\xi|^{r}+|\xi|^{w})\hat\varphi(t,\xi)|$ from the bootstrap assumption without using the equation. Let $p_1=10^{-5}$. When $|\xi|\geq (t + 1)^{p_1}$, we get similarly from Lemma~\ref{interpolation} and the bootstrap assumptions that \begin{align*}
|(|\xi|^{r}+|\xi|^{w})\hat\varphi(t,\xi)|^2&\lesssim (|\xi|^{r}+|\xi|^{w})^2[\|\varphi\|_{L^2}\|\partial_\xi\hat h\|_{L^2}+|\xi|^{-1}\|\varphi\|_{L^2}^2]\\
&\lesssim\frac{(|\xi|^{r}+|\xi|^{w})^2}{|\xi|^{s}}\|\varphi\|_{H^s}\|\partial_\xi\hat h\|_{L^2_\xi}+\frac{(|\xi|^{r}+|\xi|^{w})^2}{|\xi|^{s+1}}\|\varphi\|_{H^s}\|\varphi\|_{L^2}\\
&\lesssim |\xi|^{2w-s}(t+1)^{2p_0} \varepsilon_0^2\\
&\lesssim (t+1)^{-(s-2w)p_1+2p_0}\varepsilon_0^2 \\
&\lesssim \varepsilon_0^2,
\end{align*}
since $-(s-2w)p_1+2p_0<0$ for the parameter values in \eqref{param_vals}.
When $|\xi|<(t + 1)^{- 0.01/r}=(t+1)^{-0.025}$, by Sobolev embedding, the bootstrap assumptions, Lemma~\ref{weightedE}, and the conservation of the $L^2$-norm of $\varphi$ gives
\begin{align*}
|(|\xi|^{r} + |\xi|^{w})\hat\varphi(t, \xi)|^2&\lesssim (t+1)^{-0.01}(\|\hat\varphi\|_{L^2_\xi}^2+\|\partial_\xi\hat h\|_{L^2_\xi}^2)\\
&\lesssim \varepsilon_0^2.
\end{align*}
Thus, we only need to consider the frequency range $(t+1)^{-0.025}\leq |\xi|\leq (t + 1)^{p_1}$ in the following. From now on, we fix $\xi$ in this range, and use $\mathfrak{d}$ to denote a smooth cutoff function such that
\begin{align}
\begin{split}
&\text{$\mathfrak{d}(t, \xi) = 1$ on $\left\{(t, \xi) \mid (t+1)^{-0.025}\leq|\xi|\le(t+1)^{p_1}\right\}$,}
\\
&\text{$\mathfrak{d}(t, \xi)$ is supported on a small neighborhood of $\left\{(t, \xi) \mid (t+1)^{-0.025}\leq|\xi|\le(t+1)^{p_1}\right\}$.}
\end{split}
\label{defdelta}
\end{align}
\subsection{Modified scattering}
Define $\mathfrak{A}:=\frac{3\xi}{(1+\xi^2)^{5/2}}$ and the phase correction
\begin{equation}
\Theta(t, \xi) =- \frac{\pi \xi }{3\mathfrak{A}} [\mathbf{T}_1(\xi, \xi, -\xi) + \mathbf{T}_1(\xi,-\xi,\xi) + \mathbf{T}_1(-\xi, \xi, \xi)] \int_0^t \frac{|\hat \varphi(\tau, \xi)|^2}{\tau + 1} \,\mathrm{d}{\tau}.
\label{defTheta}
\end{equation}
We then let
\[
\hat v(t, \xi)=e^{i\Theta(t, \xi)}\hat{h}(t, \xi).
\]
Using \eqref{eqhhat} and \eqref{defTheta}, we find that
\begin{align}
\hat v_t(t, \xi) = e^{i \Theta(t, \xi)} [\hat{h}_t(t, \xi) + i \Theta_t(t, \xi) \hat{h}(t, \xi)] = U(t, \xi) - e^{- i t \xi (1+\xi^2)^{-1/2}} e^{i \Theta(t, \xi)}i\xi \widehat{\mathcal{N}_{\geq 5}(\varphi)}(t, \xi),
\label{vteq}
\end{align}
where
\begin{align}
\begin{split}
U(t, \xi) &=e^{i\Theta(t, \xi)}\bigg\{ \frac13 i\xi \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h(\xi-\eta_1-\eta_2) \hat h(\eta_1)\hat h(\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
&\qquad - \frac{\pi i \xi }{3\mathfrak{A}} \left[\mathbf{T}_1(\xi, \xi, -\xi) + \mathbf{T}_1(\xi, -\xi, \xi) + \mathbf{T}_1(-\xi,\xi,\xi)\right] \frac{|\hat{h}(t, \xi)|^2\hat{h}(t, \xi)}{t + 1}\bigg\}.
\end{split}
\label{defU}
\end{align}
Then we obtain from \eqref{vteq} that
\begin{align*}
\|\varphi\|_{Z}&=\|(|\xi|^{r}+|\xi|^{w}) \hat\varphi(t, \xi)\|_{L^\infty_\xi}=\|(|\xi|^{r}+|\xi|^{w}) \hat v(t, \xi)\|_{L^\infty_\xi}
\\
&\lesssim \int_0^t \left\{\|(|\xi|^{r}+|\xi|^{w})U(\xi, \tau)\|_{L^\infty_\xi}+\|(|\xi|^{r+1}+|\xi|^{w+1})\widehat{\mathcal{N}_{\geq5}(\varphi)}(\xi, \tau)\|_{L^\infty_\xi}\right\} \,\mathrm{d} \tau.
\end{align*}
We first estimate the term
\begin{align*}
&\|(|\xi|^{r}+|\xi|^{w})U(\tau, \xi)\|_{L^\infty_\xi}\\
=~&\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\bigg\{ \frac13 \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h(\xi-\eta_1-\eta_2) \hat h(\eta_1)\hat h(\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
& - \frac{\pi }{3\mathfrak{A}} \left[\mathbf{T}_1(\xi, \xi, -\xi) + \mathbf{T}_1(\xi, -\xi, \xi) + \mathbf{T}_1(-\xi,\xi,\xi)\right] \frac{|\hat{h}(t, \xi)|^2\hat{h}(t, \xi)}{t + 1}\bigg\}\Bigg\|.
\end{align*}
Taking the dyadic decomposition and assuming $j_1\geq j_2\geq j_3$, we can rewrite the following integral as
\begin{align*}
\frac13 \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h(\xi-\eta_1-\eta_2) \hat h(\eta_1)\hat h(\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=\frac13 \sum\limits_{\P}\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2,
\end{align*}
where the summation with $\P$ same as previous section in \eqref{eqn7.2}.
\subsection{Nonresonant frequencies}\label{sec:nonres}
Since $|\xi|\approx 2^{j_1}$, by the assumption of $\mathfrak{b}$, we only need to consider $j_1<p_1\log_2(t+1)=10^{-5}\log_2(t+1)$.
The indices satisfying $j_1-j_3>1$ correspond to nonresonant frequencies. Denote the set of these indices as $\P'_1$. We will estimate
\[
\left\|(|\xi|^{r}+|\xi|^{w})\xi\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}.
\]
After integrating by parts, we have
\begin{align*}
& \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=~& \iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{it\partial_{\eta_1}\Phi(\eta_1,\eta_2, \xi)}\partial_{\eta_1} e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=~& -W_1-W_2-W_3,
\end{align*}
where
\begin{align*}
W_1(t, \xi) & = \iint_{\mathbb{R}^2} \partial_{\eta_1}\left[ \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{it\partial_{\eta_1}\Phi(\eta_1,\eta_2, \xi)}\right] e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},\\
W_2(t, \xi) & = \iint_{\mathbb{R}^2} \left[ \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{it\partial_{\eta_1}\Phi(\eta_1,\eta_2, \xi)}\right] e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},\\
W_3(t, \xi) & = \iint_{\mathbb{R}^2} \left[ \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{it\partial_{\eta_1}\Phi(\eta_1,\eta_2, \xi)}\right] e^{it\Phi(\eta_1,\eta_2, \xi)} \partial_{\eta_1} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
In the first integral $W_1$,
\begin{align*}
&\partial_{\eta_1}\left[ \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2, \xi)}\right] \\
=~&\frac{\partial_1\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)-\partial_3\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{p'(\eta_1)-p'(\xi-\eta_1-\eta_2)}-\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)[p''(\eta_1)+p''(\xi-\eta_1-\eta_2)]}{(p'(\eta_1)-p'(\xi-\eta_1-\eta_2))^2}.
\end{align*}
By the symbol estimate \eqref{symbolpartialphi}, \eqref{symbol2d} and \eqref{Tmu_est1} the $S^\infty$ norm of the symbol is bounded by
\begin{align*}
&[(1+2^{2j_1})(1+2^{3j_3})2^{-2j_1}][(1+2^{j_2})^5(1+2^{j_3})2^{j_2+j_3}]\\
&+[(1+2^{2j_1})(1+2^{3j_3})2^{-2j_1}]^2[(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})2^{j_1+j_2+j_3}][2^{j_1}(1+2^{j_1})^{-1}]\\
\lesssim~&2^{j_2+j_3-2j_1}(1+2^{j_1})^2(1+2^{j_2})^5(1+2^{j_3})^4+2^{j_2+j_3-2j_1}(1+2^{j_2})^3(1+2^{j_3})^{4}\\
\lesssim~&2^{j_2+j_3-2j_1}(1+2^{j_1})^2(1+2^{j_2})^5(1+2^{j_3})^{4}.
\end{align*}
By the symbol estimate \eqref{symbolpartialphi} and \eqref{Tmu_est1}, we obtain the estimate for $W_1$:
\begin{align*}
\|(|\xi|^{r+1}+|\xi|^{w + 1})W_1\|_{L^\infty_\xi}\lesssim~& (1+t)^{wp_1-1} 2^{j_2+j_3-j_1}(1+2^{j_1})^2(1+2^{j_2})^5(1+2^{j_3})^{4}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~& (1+t)^{wp_1-1}\|(1+2^{j_1})^7\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|2^{j_3}(1+2^{j_3})^{4}\varphi_{j_3}\|_{L^\infty}\\
\lesssim~& (1+t)^{wp_1-1}\|\varphi_{j_1}\|_{H^s}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
By the symbol estimate \eqref{symbolpartialphi} and \eqref{Tmu_est1}, the symbol of $W_2$ and $W_3$ is bounded by
\[
[(1+2^{2j_1})(1+2^{3j_3})2^{-2j_1}][(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})2^{j_1+j_2+j_3}]\lesssim (1+2^{j_1})(1+2^{j_2})^3(1+2^{j_3})^42^{j_2+j_3-j_1}.
\]
Therefore,
\begin{align*}
&\|(|\xi|^{r+1}+|\xi|^{w + 1})W_2\|_{L^\infty_\xi}\\
\lesssim &(1+t)^{wp_1}2^{j_1}[(1+2^{j_1})(1+2^{j_2})^3(1+2^{j_3})^42^{j_2+j_3-j_1}]\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi \hat h_{j_3}\|_{L^2_\xi}\\
\lesssim &(1+t)^{(w+4)p_1}2^{j_2+j_3}(1+2^{j_1})^4\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi \hat h_{j_3}\|_{L^2}\\
\lesssim &(1+t)^{(w+4)p_1}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi \hat h_{j_3}\|_{L^2},
\end{align*}
and
\begin{align*}
&\|(|\xi|^{r+1}+|\xi|^{w + 1})W_3\|_{L^\infty_\xi}\\
\lesssim &(1+t)^{wp_1}2^{j_1}[(1+2^{j_1})(1+2^{j_2})^3(1+2^{j_3})^42^{j_2+j_3-j_1}]\|\varphi_{j_3}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi \hat h_{j_1}\|_{L^2_\xi}\\
\lesssim &(1+t)^{(w+5)p_1}\|\varphi_{j_3}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi \hat h_{j_1}\|_{L^2}.
\end{align*}
By taking the summation for $\P_1'$,
\begin{align*}
&\sum\limits_{\P_1'}\left\|(|\xi|^{r+1}+|\xi|^{w + 1}) \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim &(1+t)^{(w+5)p_1-1}\|\varphi\|_{B^{1, 6}}\|\varphi\|_{L^2}(\|\varphi\|_{H^s}+\|\partial_\xi \hat h(\xi)\|_{L^2})\\
\lesssim &(1+t)^{-1.5+(w+5)p_1}\varepsilon_1^3(t+1)^{p_0+0.01},
\end{align*}
which is integrable for $t\in(0,\infty)$.
\subsection{Close to resonance}\label{sec:near}
When $j_1\geq j_2\geq j_3$, $j_1-j_3\leq 1$ and $ j_1<10^{-5}\log_2(t+1)$, we denote the set of these indices as $\P'_2$.
For $\iota_1, \iota_2, \iota_3\in \{+,-\}$, we split the integral into several parts
\begin{align*}
\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
The we estimate each case in the following.
{\noindent\bf 1. $(\iota_1, \iota_2, \iota_3)=(+, +, -)$ or $(-, -, +)$.} This is the nonresonant case. We integrate by part with respect to $\eta_1$ and obtain
\begin{equation}
\begin{aligned}
& \iint_{\mathbb{R}^2}\partial_{\eta_1}[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} ]e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{aligned}\end{equation}
By using the symbol estimates \eqref{eqn715}, \eqref{eqn716}, and \eqref{eqn717}, we obtain the estimate for each integrals.
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2}\partial_{\eta_1}[\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{\partial_{\eta_1}\Phi} ]e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\left\{[2^{j_2+j_3}(1+2^{j_2})^5(1+2^{j_3})][(1+2^{j_1})^72^{-2j_1}]\right.\\
&\left.+[(1+2^{j_1})^72^{-2j_1}]^2[2^{j_1}(1+2^{j_1})^{-5}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})]\right\}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}(1+2^{j_1})^{13}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+8)p_1}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})][(1+2^{j_1})^72^{-2j_1}]\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}2^{2j_1}(1+2^{j_1})^{10}\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+6)p_1}\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+6)p_1}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\partial_\xi\hat h_{j_3}\|_{L^2}.
\end{align*}
Therefore taking the summation for corresponding indices $(j_1, j_2, j_3)\in\P'_2$, we have
\begin{align*}
&\sum\limits_{\P'_2}\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|\\
\lesssim ~&(t+1)^{(w+8)p_1-1}\|\varphi\|_{B^{1,6}}\|\varphi\|_{L^2}(\|\varphi\|_{L^2}+\|\partial_\xi\hat h(\xi)\|_{L^2})\\
\lesssim &(1+t)^{-1.5+(w+8)p_1+0.01}\varepsilon_1^3,
\end{align*}
which is integrable in time $t\in (0, \infty)$.
{\noindent\bf 2. $(\iota_1, \iota_2, \iota_3)=(+, +, +)$ or $(-, -, -)$.}
We split the integral into two parts:
\begin{align}\label{eq8.6}
\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align}
and
\begin{align}\label{eq8.7}
\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align}
The first integral \eqref{eq8.6} includes the space-time resonance point $(\eta_1, \eta_2)=(\xi, \xi)$ and the second one include the space resonance point $(\eta_1, \eta_2)=(\xi/3, \xi/3)$.
For each resonance, we define the cut-off functions
\begin{align*}
\mathfrak{b}_{\mathbf{l}}^{1}(\eta_1, \eta_2, \xi):=\psi_{l_{1,1}}(\eta_1-\xi)\psi_{l_{1,2}}(\eta_2-\xi),
\end{align*}
\begin{align*}
\mathfrak{b}_{\mathbf{l}}^{2}(\eta_1, \eta_2, \xi):=\psi_{l_{2,1}}(\eta_1-\frac\xi3)\psi_{l_{2,2}}(\eta_2-\frac\xi3),
\end{align*}
where $(l_{1,1}, l_{1,2}, l_{2,1}, l_{2,2})\in \mathbb{Z}^4$. Define
\begin{equation}\label{varrho}
\varrho(t):=(t+1)^{-0.45}.
\end{equation}
We first consider the case when $\max\{l_{1,1}, l_{1,2}\}\geq \log_2(\varrho(t))$.
Considering the support of the integrand, we write the first integral as
\begin{align*}
&\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=~&\sum\limits_{\max\{l_{1,1}, l_{1,2}\}\leq j+5}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\\
&\qquad\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
Since $\eta_1$ and $\eta_2$ are symmetric, we assume $l_{1,1}\geq l_{1,2}$. We integrate by part with respect to $\eta_1$ and obtain
\begin{equation}\label{eqn8.6}
\begin{aligned}
& \iint_{\mathbb{R}^2}\partial_{\eta_1}[\mathfrak{b}_{\mathbf{l}}^{1}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} ]e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathfrak{b}_{\mathbf{l}}^{1}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathfrak{b}_{\mathbf{l}}^{1}\psi_{j-3}(\eta_1+\eta_2-2\xi)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{it\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{aligned}\end{equation}
By the symbol estimates Proposition \ref{Prop_symb_Tmu}, \eqref{eqnB11}, and
\begin{align*}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)\psi_{j-3}(\eta_1+\eta_2-2\xi)\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\mathfrak{b}_{\mathbf{l}}^1\left[\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}} \right]^{-1}\right\|_{S^\infty}\\\lesssim (1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,1}},
\end{align*}
we have the Z-norm estimate for each integral terms in \eqref{eqn8.6},
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\partial_{\eta_1} \left[\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\right]\mathfrak{b}_{\mathbf{l}}^1\psi_{j-3}(\eta_1+\eta_2-2\xi)\right.\\
&\qquad \left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\Big\{ [(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{j_2})^3(1+2^{j_3})]\\
&+[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,1}}][2^{j_2+j_3}(1+2^{5j_2})(1+2^{j_3})]\Big\}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1}2^{2j_1}(1+2^{j_1})^{11}2^{-l_{1,1}}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+6)p_1}2^{j_1-l_{1,1}}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\left[\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\right]\partial_{\eta_1} \mathfrak{b}_{\mathbf{l}}^1\psi_{j-3}(\eta_1+\eta_2-2\xi)\right. \\
&\qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\Big\{ [(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]2^{-l_{1,1}}\Big\}\|\psi_{l_{1,1}}(\eta_1-\xi)\hat\varphi_{j_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\\
&\qquad\cdot\|\psi_{l_{1,2}}(\eta_2-\xi)\hat \varphi_{j_2}(\eta_2)\|_{L^2_{\eta_2}L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^{8}2^{-l_{1,1}}
\|\hat\varphi_{j_1}\|_{L^\infty_\xi}\|\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+4)p_1}2^{j_1-l_{1,1}}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\left[\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\right] \mathfrak{b}_{\mathbf{l}}^1\partial_{\eta_1}\psi_{j_1-3}(\eta_1+\eta_2-2\xi) \right.\\
&\qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\Big\{ [(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,2}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]2^{-j_{1}}\Big\}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1}2^{2j_1}(1+2^{j_1})^82^{-l_{1,2}}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+4)p_1}2^{j_1-l_{1,2}}\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\| (|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)} \mathfrak{b}_{\mathbf{l}}^1\psi_{j-3}(\eta_1+\eta_2-2\xi)\right.\\
&\qquad\left. \partial_{\eta_1}\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{1,2}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1} 2^{3j_1}(1+2^{j_1})^82^{-l_{1,2}} \|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+4)p_1}2^{j_1-l_{1,2}} \|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2}.
\end{align*}
Similarly,
\begin{align*}
&\left\| (|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)} \mathfrak{b}_{\mathbf{l}}^1\psi_{j-3}(\eta_1+\eta_2-2\xi)\right.\\
&\qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+4)p_1}2^{j_1-l_{1,2}} \|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}.
\end{align*}
Then we take the summation for $l_{1,1}$ and $l_{1,2}$ from $\log_2(\varrho(t))$ to $j_1+2$, considering the support of the cut-off functions,
\begin{align*}
&\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\sum\limits_{\min\{l_{1,1}, l_{1,2}\}\geq \log_2[\varrho(t)]}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}\psi_{j-3}(\eta_1+\eta_2-2\xi)\\
&\qquad\qquad\qquad\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1}\,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(|j_1|+|\log_2[\varrho(t)]|)2^{j_1}\max\{[\varrho(t)]^{-1}, 2^{-j_1}\}(t+1)^{(w+6)p_1-1} \|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}(\|\partial_\xi\hat h_{j_3}\|_{L^2}+\|\varphi_{j_3}\|_{L^2})\\
&~~+(|j_1|+|\log_2[\varrho(t)]|)2^{j_1}\max\{[\varrho(t)]^{-1}, 2^{-j_1}\}(t+1)^{(w+6)p_1-1}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
Notice that $|\log_2[\varrho(t)]|=O(\log_2(t+1))$ and $|2^{j_1}|\lesssim (t+1)^{p_1}$. So we have
\begin{align*}
&\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\sum\limits_{\min\{l_{1,1}, l_{1,2}\}\geq \log_2[\varrho(t)]}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}\psi_{j-3}(\eta_1+\eta_2-2\xi)\\
&\qquad\qquad\qquad\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1}\,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+8)p_1-0.51}\Big\{ \|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}(\|\partial_\xi\hat h_{j_3}\|_{L^2}+\|\varphi_{j_3}\|_{L^2})+\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}\Big\}.
\end{align*}
Taking summation for corresponding $j_1, j_2, j_3$ and using the bootstrap assumption, we have
\begin{align*}
&\sum\limits_{j_1,j_2,j_3}\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\sum\limits_{\min\{l_{1,1}, l_{1,2}\}\geq \log_2[\varrho(t)]}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}\psi_{j-3}(\eta_1+\eta_2-2\xi)\\
&\qquad\qquad\qquad\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1}\,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+8)p_1-0.55-0.5}\varepsilon_1^3,
\end{align*}
which is integrable in time $t\in(0,\infty)$.
The other term to be estimated is \eqref{eq8.7}. We consider the case when $\max\{l_{2,1}, l_{2,2}\}\geq \log_2(\varrho(t))$.
Since $\eta_1$ and $\eta_2$ are symmetric, we assume $l_{2,1}\geq l_{2,2}$. We integrate by part with respect to $\eta_1$ and obtain
\begin{equation}
\begin{aligned}
& \iint_{\mathbb{R}^2}\partial_{\eta_1}[\mathfrak{b}_{\mathbf{l}}^{2}(1-\psi_{j-3}(\eta_1+\eta_2-2\xi))\frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{it\partial_{\eta_1}\Phi} ]e^{it\Phi}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathfrak{b}_{\mathbf{l}}^{2}(1-\psi_{j-3}(\eta_1+\eta_2-2\xi))\frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{it\partial_{\eta_1}\Phi} e^{it\Phi} \partial_{\eta_1}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+ \iint_{\mathbb{R}^2}\mathfrak{b}_{\mathbf{l}}^{2}(1-\psi_{j-3}(\eta_1+\eta_2-2\xi))\frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{it\partial_{\eta_1}\Phi} e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_1}\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{aligned}\end{equation}
By the symbol estimates Proposition \ref{Prop_symb_Tmu} and
\begin{multline*}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)]\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\mathfrak{b}_{\mathbf{l}}^2\right.\\
\left.\cdot\left[\frac{1}{(1+\eta_1^2)^{3/2}}-\frac{1}{[1+\eta_3^2]^{3/2}} \right]^{-1}\right\|_{S^\infty}\lesssim (1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{2,1}},
\end{multline*
\begin{multline*}
\left\|\psi_{j_1}(\eta_1)\psi_{j_2}(\eta_2)\psi_{j_3}(\eta_3)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)]\upsilon_{\iota_1}(\eta_1)\upsilon_{\iota_2}(\eta_2)\upsilon_{\iota_3}(\eta_1+\eta_2+\eta_3)\mathfrak{b}_{\mathbf{l}}^2\right.\\
\left.\cdot\left[-\frac{3\eta_1}{(1+\eta_1^2)^{5/2}}-\frac{3\eta_3}{(1+\eta_3^2)^{5/2}}\right]\right\|_{S^\infty}
\lesssim 2^{j_{1}}(1+2^{2j_1})^{-5/2},
\end{multline*}
we obtain the estimates of each integrals as following
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\partial_{\eta_1} \left[\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\right]\mathfrak{b}_{\mathbf{l}}^2(1-\psi_{j-3}(\eta_1+\eta_2-2\xi))\right.\\
& \qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\Big\{ [(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{2,1}}]^2[2^{j_{1}}(1+2^{2j_1})^{-5/2}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\\
&+[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{2,1}}][2^{j_2+j_3}(1+2^{5j_2})(1+2^{j_3})]\Big\}\|\psi_{l_{2,1}}(\eta_1-\frac\xi3)\hat\varphi_{j_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\\
&\qquad\cdot\|\psi_{l_{2,2}}(\eta_2-\frac\xi3)\hat \varphi_{j_2}(\eta_2)\|_{L^2_{\eta_2}L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}[(1+2^{j_1})^{11}2^{-l_{2,1}}2^{2j_1}]\|\hat\varphi_{j_1}\|_{L^\infty_\xi}\|\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+7)p_1}2^{j_1-l_{2,1}}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\left[\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\right]\partial_{\eta_1} \mathfrak{b}_{\mathbf{l}}^2[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)]\right.\\
&\qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}\Big\{ [(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{2,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]2^{-l_{2,1}}\Big\}\\
&\qquad\cdot\|\psi_{l_{2,1}}(\eta_1-\xi)\hat\varphi_{j_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\|\psi_{l_{2,2}}(\eta_2-\xi)\hat \varphi_{j_2}(\eta_2)\|_{L^2_{\eta_2}L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^82^{-l_{2,1}}\|\hat\varphi_{j_1}\|_{L^\infty_\xi}\|\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1-l_{2,1}}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat \varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
and
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\mathfrak{b}_{\mathbf{l}}^2[1-\psi_{j-3}(\eta_1+\eta_2-2\xi) ]\right.\\
&\qquad\left. \partial_{\eta_1}\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{2,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^82^{-l_{2,1}} \|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1-l_{2,1}} \|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2}.
\end{align*}
Similarly,
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi(\eta_1,\eta_2, \xi)}\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\partial_{\eta_1}\Phi(\eta_1,\eta_2,\xi)}\mathfrak{b}_{\mathbf{l}}^2[1-\psi_{j-3}(\eta_1+\eta_2-2\xi) ]\right.\\
&\qquad\left. \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2)\partial_{\eta_1} \hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1-l_{2,1}} \|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}.
\end{align*}
Then we take the summation for $l_{2,1}$ and $l_{2,2}$ from $\log_2\varrho(t)$ to $j_1+2$, considering the support of the cut-off functions,
\begin{align*}
&\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\sum\limits_{\max\{l_{2,1}, l_{2,2}\}\geq \log_2[\rho(t)]}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \\
&\qquad\qquad\qquad\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+7)p_1-1}2^{j_1}(|j_1|+|\log_2[\varrho(t)]|)\max\{[\varrho(t)]^{-1}, 2^{-j_1}\}\Big(\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}\\
&\qquad+\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}\Big)\\
\lesssim~&(t+1)^{(w+9)p_1-0.55}\Big(\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}\\
&\qquad+\|\partial_\xi\hat h_{j_1}\|_{L^2}\|2^{j_2}\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}\Big).
\end{align*}
Taking summation for corresponding $j_1, j_2, j_3$ and using the bootstrap assumption, we have
\begin{align*}
&\sum\limits_{j_1,j_2,j_3}\Bigg\|(|\xi|^{r}+|\xi|^{w})\xi\sum\limits_{\min\{l_{2,1}, l_{2,2}\}\geq \log_2[\varrho(t)]}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)]\\
&\qquad\qquad\qquad\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1}\,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+9)p_1-0.55-0.5+0.01}\varepsilon_1^3,
\end{align*}
which is integrable in time $t\in(0,\infty)$.
{\noindent\bf 3. $(\iota_1, \iota_2, \iota_3)=(+, -, +)$ or $(-, +, -)$. }
We define the cut-off function
\begin{align*}
\mathfrak{b}_{\mathbf{l}}^{3}(\eta_1, \eta_2, \xi):=&\psi_{l_{3,1}}(\eta_1-\xi)\psi_{l_{3,2}}(\eta_1+\eta_2),
\end{align*}
where $\mathbf{l}=(l_{3,1}, l_{3,2})\in \mathbb{Z}^2$.
Considering the support of the integrand, we write the integral as
\begin{align*}
&\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=~&\sum\limits_{\max\{l_{3,1}, l_{3,2}\}\leq j+2}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
When $l_{3,1} \geq l_{3,2}$, using \eqref{eqn718} and doing integration by part, we obtain
\begin{align*}
& \iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&-\frac it \iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{1}{p'(\eta_1)-p'(\eta_2)}(\partial_{\eta_1}-\partial_{\eta_2})e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
=~&\frac it \iint_{\mathbb{R}^2}e^{it\Phi} (\partial_{\eta_1}-\partial_{\eta_2})\left[ \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_1)-p'(\eta_2)}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\right]\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
=~&\frac it \iint_{\mathbb{R}^2}e^{it\Phi}\Big[ \mathfrak{b}_{\mathbf{l}}^{3} (\partial_1-\partial_2)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) +(\partial_1-\partial_2) \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \\
&- \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p''(\eta_1)+p''(\eta_2)}{p'(\eta_1)-p'(\eta_2)}\Big]\frac{1}{p'(\eta_1)-p'(\eta_2)}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\\nonumber
&+\frac it \iint_{\mathbb{R}^2}e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_1)-p'(\eta_2)}(\partial_1-\partial_2)\left[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\right]\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
By the symbol estimates Proposition \ref{Prop_symb_Tmu} and Proposition \ref{PropA3}, we obtain the estimates for each integral above.
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2}e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} (\partial_1-\partial_2)\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{p'(\eta_1)-p'(\eta_2)}\right.\\
&\qquad\left.\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{- l_{3,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{5j_2})(1+2^{j_3})]\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^{10}2^{- l_{3,1}}\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+7)p_1}2^{j_1- l_{3,1}}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\| (|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2}e^{it\Phi} (\partial_1-\partial_2)\mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{1}{p'(\eta_1)-p'(\eta_2)}\right.\\
&\qquad\left.\cdot\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}2^{-l_{3,1}}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]
[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{3,1}}]\\
&\quad\cdot\|\psi_{l_{3,1}}(\eta_1-\xi)\hat\varphi_{j_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\|\psi_{l_{3,2}}(\xi-\eta_3)\hat\varphi_{j_3}(\eta_3)\|_{L^2_{\eta_3}L^\infty_\xi}\|\varphi_{j_2}\|_{L^\infty}\\
\lesssim~& (t+1)^{wp_1}2^{3j_1-l_{3,1}}(1+2^{j_1})^{8}\|\hat\varphi_{j_1}\|_{L^\infty_\xi}\|\hat\varphi_{j_3}\|_{L^\infty_\xi}\|\varphi_{j_2}\|_{B^{1,6}}\\
\lesssim~& (t+1)^{(w+5)p_1}2^{j_1-l_{3,1}}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_3/2}\hat\varphi_{j_3}\|_{L^\infty_\xi}\|\varphi_{j_2}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2}e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)\frac{(p''(\eta_1)+p''(\eta_2))}{(p'(\eta_1)-p'(\eta_2))^2}\right.\\
&\qquad\left.\cdot\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{3,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}[(1+2^{j_1})^{8}2^{-l_{3,1}}\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1- l_{3,1}}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2}e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_1)-p'(\eta_2)} (\partial_1-\partial_2)\left[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\right]\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\Bigg\|_{L^\infty_\xi}\\
\lesssim~& (t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{- l_{3,1}}][2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\\
&\quad \cdot\Big(\|\partial_{\eta_1}\hat h_{j_1}(\eta_1)\|_{L^2_{\eta_1}}\|\varphi_{j_2}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\partial_{\eta_2}\hat h_{j_2}(\eta_2)\|_{L^2_{\eta_2}}\Big)\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~& (t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^{8}2^{- l_{3,1}}\Big(\|\partial_{\eta_1}\hat h_{j_1}(\eta_1)\|_{L^2_{\eta_1}}\|\varphi_{j_2}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\partial_{\eta_2}\hat h_{j_2}(\eta_2)\|_{L^2_{\eta_2}}\Big)\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1- l_{3,1}}\Big(\|\partial_{\xi}\hat h_{j_1}\|_{L^2_{\xi}}\|\varphi_{j_2}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\partial_{\xi}\hat h_{j_2}\|_{L^2_{\xi}}\Big)\|\varphi_{j_3}\|_{B^{1,6}}.
\end{align*}
When $l_{3,1} < l_{3,2}$, we integrate by part with $\eta_2$,
\begin{align*}
& \iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&-\frac it \iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{1}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\partial_{\eta_2}e^{it\Phi} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&\frac it \iint_{\mathbb{R}^2} e^{it\Phi}\partial_{\eta_2}\left[ \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\right]\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
=~&\frac it \iint_{\mathbb{R}^2} e^{it\Phi}\partial_{\eta_2} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac it \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\partial_{\eta_2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&-\frac it \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p''(\eta_2)+p''(\xi-\eta_1-\eta_2)}{[p'(\eta_2)-p'(\xi-\eta_1-\eta_2)]^2}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac it \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\partial_{\eta_2}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\\
&+\frac it \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_2}\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2.
\end{align*}
Then we estimate each integrals in the following.
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2} e^{it\Phi}\partial_{\eta_2} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~& (t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{3,2}}]2^{-l_{3,2}}[2^{j_1+j_2+j_3}(1+2^{j_1})^{-1}(1+2^{3j_2})(1+2^{j_3})]\\
&\quad\cdot\|\psi_{l_{3,1}}(\eta_1-\xi)\hat\varphi_{j_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\|\psi_{l_{3,2}}(\xi-\eta_3)\hat\varphi_{j_3}(\eta_3)\|_{L^2_{\eta_3}L^\infty_\xi}\|\varphi_{j_2}\|_{L^\infty}\\
\lesssim~&(t+1)^{wp_1}2^{3j_1}(1+2^{j_1})^{8}2^{- l_{3,2}}\|\hat\varphi_{j_1}\|_{L^\infty_\xi}\|\hat\varphi_{j_3}\|_{L^\infty_\xi}\|\varphi_{j_2}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1- l_{3,2}}\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_3/2}\hat\varphi_{j_3}\|_{L^\infty_\xi}\|\varphi_{j_2}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\partial_{\eta_2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2 \right\|_{L^\infty_\xi}\\
\lesssim~& (t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{- l_{3,2}}][2^{j_2+j_3}(1+2^{5j_2})(1+2^{j_3})]\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~&(t+1)^{(w+7)p_1}2^{j_1- l_{3,2}}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}.
\end{align*}
By the symbol estimate \eqref{eqnB10} and Proposition \ref{Prop_symb_Tmu},
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1}) \iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) \frac{p''(\eta_2)+p''(\xi-\eta_1-\eta_2)}{[p'(\eta_2)-p'(\xi-\eta_1-\eta_2)]^2}\right.\\
&\qquad\left.\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2\right\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{-l_{3,2}}][2^{j_2+j_3}(1+2^{3j_2})(1+2^{j_3})]\|\varphi_{j_1}\|_{L^\infty}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
\lesssim~& (t+1)^{(w+5)p_1}2^{j_1- l_{3,2}}\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\partial_{\eta_2}\hat h_{j_2}^{\iota_2}(\eta_2)\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2 \right\|_{L^\infty_\xi}\\
\lesssim~& (t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{- l_{3,2}}][2^{j_2+j_3}(1+2^{3j_2})(1+2^{j_3})]\|\varphi_{j_1}\|_{L^2}\|\partial_{\xi}\hat h_{j_2}\|_{L^2_{\xi}}\|\varphi_{j_3}\|_{L^\infty}\\
\lesssim~& (t+1)^{(w+5)p_1}2^{j_1- l_{3,2}}\|\varphi_{j_1}\|_{L^2}\|\partial_{\xi}\hat h_{j_2}\|_{L^2_{\xi}}\|\varphi_{j_3}\|_{B^{1,6}},
\end{align*}
\begin{align*}
&\left\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} e^{it\Phi} \mathfrak{b}_{\mathbf{l}}^{3} \frac{\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2)}{p'(\eta_2)-p'(\xi-\eta_1-\eta_2)}\hat h_{j_2}^{\iota_2}(\eta_2)\partial_{\eta_2}\hat h_{j_3}(\xi-\eta_1-\eta_2)\hat h_{j_1}^{\iota_1}(\eta_1)\upsilon_{\iota_3}(\xi)\,\mathrm{d} \eta_1\,\mathrm{d} \eta_2 \right\|_{L^\infty_\xi}\\
\lesssim~& (t+1)^{wp_1}2^{j_1}[(1+2^{2j_1})^{5/2}2^{-j_1}2^{- l_{3,2}}][2^{j_2+j_3}(1+2^{3j_2})(1+2^{j_3})]\|\partial_{\xi}\hat h_{j_3}\|_{L^2_{\xi}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_1}\|_{L^\infty}\\
\lesssim~&(t+1)^{(w+5)p_1}2^{j_1- l_{3,2}}\|\partial_{\xi}\hat h_{j_3}\|_{L^2_{\xi}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_1}\|_{B^{1,6}}.
\end{align*}
Then we take the summation for $\log_2\varrho(t)<\max\{l_{3,1}, l_{3,2}\}$,
\begin{align*}
&\sum\limits_{l_{3,1}, l_{3,2}\geq\log_2[\varrho(t)] }\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\\
&\hspace{3cm}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+7)p_1-1}2^{j_1}(|j_1|+|\log_2[\varrho(t)]|)\max\{[\varrho(t)]^{-1}, 2^{-j_1}\}\Big(\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}\\
&\qquad+\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}+\|\partial_\xi\hat h_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\varphi_{j_3}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}\Big)\\
\lesssim~&(t+1)^{(w+9)p_1-0.55}\Big(\|2^{j_1/2}\hat\varphi_{j_1}\|_{L^\infty_\xi}\|2^{j_2/2}\hat\varphi_{j_2}\|_{L^\infty_\xi}\|\varphi_{j_3}\|_{B^{1,6}}+\|\varphi_{j_1}\|_{B^{1,6}}\|\varphi_{j_2}\|_{L^2}\|\varphi_{j_3}\|_{L^2}\\
&\qquad+\|\partial_\xi\hat h_{j_1}\|_{L^2}\|2^{j_2}\varphi_{j_2}\|_{L^\infty}\|\varphi_{j_3}\|_{L^2}+\|\varphi_{j_1}\|_{L^2}\|\varphi_{j_2}\|_{B^{1,6}}\|\partial_\xi\hat h_{j_3}\|_{L^2}\Big).
\end{align*}
Taking the summation for corresponding $j_1, j_2, j_3$ and using the bootstrap assumption, we have
\begin{equation}\label{eqn8.9}
\begin{aligned}
&\sum\limits_{j_1,j_2,j_3}\sum\limits_{ l_{3,1}, l_{3,2}\geq\log_2[\varrho(t)] }\Bigg\|(|\xi|^{r+1}+|\xi|^{w+1})\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \\
&\qquad \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\Bigg\|_{L^\infty_\xi}\\
\lesssim~&(t+1)^{(w+9)p_1-0.55-0.5+0.01}\varepsilon_1^3,
\end{aligned}\end{equation}
which is integrable in time $t\in(0,\infty)$.
{\noindent\bf 4. $(\iota_1, \iota_2, \iota_3)=(-, +, +)$ or $(+, -, -)$.} We define the cut-off function
\begin{align*}
\mathfrak{b}_{\mathbf{l}}^{4}(\eta_1, \eta_2, \xi):=&\psi_{l_{4,1}}(\eta_2-\xi)\psi_{l_{4,2}}(\eta_1+\eta_2),
\end{align*}
where $\mathbf{l}=(l_{4,1}, l_{4,2})\in \mathbb{Z}^2$.
Considering the support of the integrand, we write the integral as
\begin{align*}
&\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
=~&\sum\limits_{\max\{l_{4,1}, l_{4,2}\}\leq j+2}\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{4} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
Since $\eta_1$ and $\eta_2$ are symmetric in the expression of the integral, by interchanging $\eta_1$ and $\eta_2$, we can transform this case into the previous case. So we can obtain the similar estimate as \eqref{eqn8.9}.
\subsection{Space resonance estimate}
Then we estimate
\begin{align}\label{eqn89}
\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align}
when $(\iota_1, \iota_2, \iota_3)=(+, +, +)$ or $(-, -, -)$, and $\max\{l_{2,1}, l_{2,2}\}\leq \log_2\rho(t)=-0.3\log_2(t+1)$.
In this case, we can expand $\mathbf{T}_1 / \Phi$ around $(\eta_1,\eta_2)=(\xi/3, \xi/3)$ as
\begin{equation}\label{T1Phi}
\frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\eta_1,\eta_2, \xi) }=
\frac{9\mathbf{T}_1(\frac\xi3,\frac\xi3,\frac\xi3)\sqrt{1+\xi^2}\sqrt{1+\xi^2/9}(\sqrt{1+\xi^2}+\sqrt{1+\xi^2/9})}{8\xi^3}.
\end{equation}
We will prove
\begin{equation}\label{eqn814}
\left\|(|\xi|^{r+1}+|\xi|^{w+1})\int_0^\infty\sum\limits_{\max\{l_{2,1}, l_{2,2}\}\leq \log_2\rho(t)}\eqref{eqn89}\,\mathrm{d} t\right\|_{L^\infty_\xi}<\infty.
\end{equation}
By interchanging the order of the summation and the time integral, it is equivalent to
\[
\sum\limits_{\max\{l_{2,1}, l_{2,2}\}\leq 0}\left\|(|\xi|^{r+1}+|\xi|^{w+1})\int_0^{t_\mathbf{l}}\eqref{eqn89}\,\mathrm{d} t\right\|_{L^\infty_\xi}<\infty.
\]
where $t_\mathbf{l}:=2^{-\frac{10}3\max\{l_{2,1}, l_{2,2}\}}$.
After writing
\begin{align*}
e^{i\tau \Phi(\xi, \eta_1,\eta_2)}=\frac{1}{i\Phi(\xi, \eta_1,\eta_2)} \left[\partial_\tau e^{i\tau\Phi(\xi, \eta_1,\eta_2)}\right],
\end{align*}
and integrating by parts with respect to $\tau$ in each time interval between the time discontinuities, we get that
\[
\begin{aligned}
&\int_0^{t_\mathbf{l}} \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)e^{i\tau\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2} \,\mathrm{d}{\tau}
\\[1ex]
&\qquad\qquad= J_1-\int_0^{t_\mathbf{l}} J_2(\tau)\,\mathrm{d}{\tau},
\end{aligned}
\]
where
\begin{align*}
J_1:=\iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{i\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)}&\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\\
&\cdot\upsilon_{\iota_3}(\xi)[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2}\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\bigg|_{t=0}^{t_\mathbf{l}},
\end{align*}
\begin{align*}
&J_2(\tau) := \iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)} \partial_\tau \left[\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\right][1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
For $J_1$, we have from \eqref{T1Phi} that
\[
\begin{aligned}
&\bigg|(|\xi|^{r+1}+|\xi|^{w + 1})\iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)}\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\\
&\qquad\qquad[1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2}\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\bigg|\\
\lesssim~~&(2^{rj_1}+2^{wj_1}) 2^{j_1}[2^{3j_1}(1+2^{j_1})^3][2^{-3j_1}(1+2^{j_1})^{3}] 2^{l_{2,1}+l_{2,2}} \|\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\|_{L^\infty_{\eta_1\eta_2\xi}}\\
\lesssim~&2^{1.4j_1}(1+2^{j_1})^{17}2^{l_{2,1}+l_{2,2}} \|\hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\|_{L^\infty_{\eta_1\eta_2\xi}}\\
\end{aligned}
\]
which are summable for $\max\{l_{2,1}, l_{2,2}\}\leq 0$.
After taking the time derivative, the term $J_2$ can be written as a sum of three terms.
\begin{align*}
& \iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)} \left[\partial_\tau\hat h_{j_1}(\eta_1, \tau)\hat h_{j_2}(\eta_2, \tau) \hat h_{j_3}(\xi-\eta_1-\eta_2, \tau)\right][1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},\\[1ex]
& \iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)} \left[\hat h_{j_1}(\eta_1, \tau) \partial_\tau \hat h_{j_2}(\eta_2, \tau) \hat h_{j_3}(\tau, \xi-\eta_1-\eta_2)\right][1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align*}
and
\begin{align*}
& \iint_{\mathbb{R}^2} \frac{\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2)}{\Phi(\xi, \eta_1,\eta_2)} e^{i\tau\Phi(\xi, \eta_1,\eta_2)} \left[\hat h_{j_1}(\eta_1, \tau)\hat h_{j_2}(\eta_2, \tau) \partial_\tau\hat h_{j_3}(\xi-\eta_1-\eta_2, \tau)\right][1-\psi_{j-3}(\eta_1+\eta_2-2\xi)] \mathfrak{b}_{\mathbf{l}}^{2} \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
Notice that by \eqref{eqhhat},and the bootstrap assumptions and Lemma \ref{NonDis}, we have
\begin{align*}
\|\partial_t h_j\|_{L^\infty}\lesssim& \bigg\|\psi_j(\xi)\xi\iint_{\mathbb{R}^2}\mathbf{T}_1(\eta_1,\eta_2,\xi-\eta_1-\eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}\hat h(\xi-\eta_1-\eta_2)\hat h(\eta_1)\hat h(\eta_2)\,\mathrm{d}\eta_1\,\mathrm{d}\eta_2\bigg\|_{L^1_\xi}+\left\|\psi_j(\xi)\xi\widehat{\mathcal N_{\geq 5}(\varphi)}\right\|_{L^1_\xi}\\
\lesssim~~ &\|\varphi_j\|_{B^{1,1}} \sum\limits_{j=1}^{\infty} \left(\|\varphi\|_{B^{1,6}}^{2 j } \right)\\
\lesssim~~ &\|\varphi_j\|_{B^{1,1}}(t + 1)^{-1}\varepsilon_1^2.
\end{align*}
Therefore, we obtain
\[
\begin{aligned}
&\int_0^{t_{\mathbf{l}}}\left| (|\xi|^{r+1}+|\xi|^{w+ 1}) J_2(\tau)\right|\,\mathrm{d}\tau\\
\lesssim &\sum t_{\mathbf{l}}^{(w-r)p_1}2^{(r+1)j_1}\int_0^{t_{\mathbf{l}}}(1+2^{6j_1}) \| \psi_{l_{2,1}}(\eta_1-\frac\xi3)\hat h_{\ell_1}(\eta_1)\|_{L^2_{\eta_1}L^\infty_\xi}\|\psi_{l_{2,2}}(\eta_2-\frac\xi3)\hat h_{\ell_2}(\eta_2)\|_{L^2_{\eta_2}L^\infty_\xi}\|\partial_\tau h_{\ell_3}\|_{L^\infty}\,\mathrm{d} \tau \\
\lesssim &\sum t_{\mathbf{l}}^{wp_1}\int_0^{t_{\mathbf{l}}}(1+2^{6j_1}) 2^{(l_{2,1}+l_{2,2})/2}\|2^{j_1/2} \hat h_{\ell_1}\|_{L^\infty_\xi}\|2^{j_1/2}\hat h_{\ell_2}\|_{L^\infty_\xi}\|\partial_\tau h_{\ell_3}\|_{L^\infty}\,\mathrm{d} \tau \\
\lesssim &t_{\mathbf{l}}^{(w+1)p_1-0.5} 2^{(l_{2,1}+l_{2,2})/2}\|2^{j_1/2} \hat h_{\ell_1}\|_{L_t^\infty L^\infty_\xi }\|2^{j_1/2}\hat h_{\ell_2}\|_{L_t^\infty L^\infty_\xi }\|\varphi_{\ell_3}\|_{L_t^\infty B^{1,6}}\cdot\varepsilon_1^2
\end{aligned}
\]
where the summation is for all permutations $(\ell_1, \ell_2, \ell_3)$ of $(j_1, j_2,j_3)$. Notice that the coefficients for $\mathbf{l}$ is positive, thus it is summable for $\max\{l_{2,1}, l_{2,2}\}\leq 0$.
Combining the estimates for $J_1$ and $J_2$, we obtain \eqref{eqn814}.
\subsection{Space-time resonances}
The cubic nonlinear terms related to the space-time resonances can be formulated as
\begin{itemize}
\item[CASE 1.] When $(\iota_1, \iota_2, \iota_3)=(+, +, +)$ or $(-, -, -)$, $\max\{l_{1,1}, l_{1,2}\}\leq \log_2\varrho(t)$,
\begin{align*}
&\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{1} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi)\psi_{j-3}(\eta_1+\eta_2-2\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align*}
with $\mathfrak{b}_{\mathbf{l}}^{1}=\psi_{l_{1,1}}(\eta_1-\xi)\psi_{l_{1,2}}(\eta_2-\xi)$. Here the cut-off function $\psi_{j-3}(\eta_1+\eta_2-2\xi)$ can be removed when $t$ is large.
\item[CASE 2.] When $(\iota_1, \iota_2, \iota_3)=(+, -, +)$ or $(-, +, -)$, $\max\{l_{3,1}, l_{3,2}\}\leq \log_2\varrho(t)$,
\begin{align*}
\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{3} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align*}
with $\mathfrak{b}_{\mathbf{l}}^{3}=\psi_{l_{3,1}}(\eta_1-\xi)\psi_{l_{3,2}}(\eta_1+\eta_2)$.
\item[CASE 3.] When $(\iota_1, \iota_2, \iota_3)=(-, +, +)$ or $(+, -, -)$, $\max\{l_{4,1}, l_{4,2}\}\leq \log_2\varrho(t)$,
\begin{align*}
\iint_{\mathbb{R}^2} \mathfrak{b}_{\mathbf{l}}^{4} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \hat h_{j_1}^{\iota_1}(\eta_1)\hat h_{j_2}^{\iota_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\upsilon_{\iota_3}(\xi) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2},
\end{align*}
with $\mathfrak{b}_{\mathbf{l}}^{4}=\psi_{l_{4,1}}(\eta_2-\xi)\psi_{l_{4,2}}(\eta_1+\eta_2)$.
\end{itemize}
We notice that after interchanging $\eta_1$ and $\eta_2$, CASE 3 will be transformed to CASE 2. After interchanging $\eta_2$ and $\xi-\eta_1-\eta_2$, CASE 2 will be transformed to CASE 1. So we only estimate CASE 1 in the following.
By taking the summation for $\max\{l_{1,1}, l_{1,2}\}\leq \log_2\varrho(t)$, we write the integral as
\begin{align*}
\iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)} \psi\left(\frac{\eta_1-\xi}{2^{\lfloor\log_2\varrho(t)\rfloor}}\right)\psi\left(\frac{\eta_2-\xi}{2^{\lfloor\log_2\varrho(t)\rfloor}}\right)\hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2)\,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}.
\end{align*}
Denote
\[
\varrho_1(t)=2^{\lfloor\log_2\varrho(t)\rfloor}.
\]
Then it satisfies $\varrho(t)/2\leq\varrho_1(t)<\varrho(t)$.
Considering the modified scattering and \eqref{defU}, in the following, we will estimate
\begin{align*}
& \frac13 i\xi \iint_{\mathbb{R}^2} \mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) e^{it\Phi(\eta_1,\eta_2, \xi)}\psi\left(\frac{\eta_1-\xi}{\varrho_1(t)}\right)\psi\left(\frac{\eta_2-\xi}{\varrho_1(t)}\right) \hat h_{j_1}(\eta_1)\hat h_{j_2}(\eta_2) \hat h_{j_3}(\xi-\eta_1-\eta_2) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
&- \frac{\pi i \xi }{3A}\mathbf{T}_1(\xi, \xi, -\xi) \frac{|\hat{h}(t, \xi)|^2\hat{h}(t, \xi)}{t + 1}.
\end{align*}
We can split it into two parts:
\begin{align}
\label{modscat1}
\begin{split}
& \frac{i \xi}{6} \iint_{\mathbb{R}^2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg)\\*
&\qquad \cdot \bigg[\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) \hat{h}_{j_1}(\eta_1)\hat{h}_{j_2}(\eta_2) \hat{h}_{j_3}(\xi-\eta_1-\eta_2) - \mathbf{T}_1(\xi, \xi, -\xi)|\hat{h}(\xi)|^2\hat{h}(\xi)\bigg] \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
\end{split}
\end{align}
and
\begin{align}
\label{modscat2}
\frac{i \xi}{6} \mathbf{T}_1(\xi, \xi, - \xi) |\hat{h}(t,\xi)|^2\hat{h}(t,\xi) \bigg[\iint_{A_2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2} - \frac{2 \pi }{A(t + 1)}\bigg].
\end{align}
The estimates for \eqref{modscat1} are achieved by a Taylor expansion, \eqref{Tmu_est2} and \eqref{Tmu_est3}.
\begin{align*}
&\bigg|(|\xi|^{r}+|\xi|^{w})\frac16 i\xi \iint_{\mathbb{R}^2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg)\\*
&\qquad \cdot \Big[\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) \hat{h}_{j_1}(\eta_1)\hat{h}_{j_2}(\eta_2) \hat{h}_{j_3}(\xi-\eta_1-\eta_2) - \mathbf{T}_1(\xi, \xi, -\xi)|\hat{h}(\xi)|^2\hat{h}(\xi) \Big] \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\bigg|\\
\lesssim~& (|\xi|^{r}+|\xi|^{w}) |\xi| \iint_{\mathbb{R}^2} \left|\partial_{\eta_1}\left[\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) \hat{h}_{j_1}(\eta_1)\hat{h}_{j_2}(\eta_2) \hat{h}_{j_3}(\xi-\eta_1-\eta_2)\right] \bigg|_{\eta_1 = \eta_1'} (\xi-\eta_1)\right|\\
& \hspace {.75in}+\left|\partial_{\eta_2}\left[\mathbf{T}_1(\eta_1, \eta_2, \xi - \eta_1 - \eta_2) \hat{h}_{j_1}(\eta_1)\hat{h}_{j_2}(\eta_2) \hat{h}_{j_3}(\xi-\eta_1-\eta_2)\right] \bigg|_{\eta_2 = \eta_2'} (\xi-\eta_2)\right| \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2}\\
\lesssim~&(t+1)^{(w+8)p_1} \Big( \||\xi|^{\frac{1+r}3}\hat\varphi_{j_1}\|_{L^\infty_\xi}\||\xi|^{\frac{1+r}3}\hat\varphi_{j_2}\|_{L^\infty_\xi}\||\xi|^{\frac{1+r}3}\hat\varphi_{j_3}\|_{L^\infty_\xi} [ \varrho_1(t)]^{3}\\
&\qquad+ \sum\||\xi|^{\frac{1+r}3}\hat\varphi_{\ell_1}\|_{L^\infty_\xi}\||\xi|^{\frac{1+r}3}\hat\varphi_{\ell_2}\|_{L^\infty_\xi}\|\partial_\xi\hat h_{\ell_3}\|_{L^2_\xi}[ \varrho_1(t)]^{5/2}\Big).
\end{align*}
As for \eqref{modscat2}, it suffices to estimate
\begin{align*}
& \bigg|(|\xi|^{r}+|\xi|^{w}) \frac{i \xi}{6} \mathbf{T}_1(\xi, \xi, - \xi) |\hat{h}(t,\xi)|^2\hat{h}(t,\xi) \bigg[\iint_{\mathbb{R}^2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2} - \frac{2 \pi }{\mathfrak{A}t}\bigg]\bigg|\\
\lesssim ~& \|(|\xi|^{r}+|\xi|^{w}) \hat{\varphi}(\xi)\|_{L^\infty_\xi} \||\xi|^{1/2} \hat{\varphi}(\xi)\|_{L^\infty_\xi} \|(|\xi|^{1/2}+|\xi|^w)\hat{\varphi}(\xi)\|_{L^\infty_\xi}\\
& \qquad \cdot \bigg\|\iint_{\mathbb{R}^2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2} - \frac{2 \pi }{\mathfrak{A}t}\bigg\|_{L^\infty_\xi}\\
\lesssim ~& \|\varphi\|_{Z}^3 \bigg\|\iint_{\mathbb{R}^2} e^{i t \Phi(\xi, \eta_1, \eta_2)} \psi\bigg(\frac{\eta_1 - \xi}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\eta_2 - \xi}{ \varrho_1(t)}\bigg) \,\mathrm{d}{\eta_1} \,\mathrm{d}{\eta_2} - \frac{2 \pi }{\mathfrak{A}t}\bigg\|_{L^\infty_\xi}.
\end{align*}
Writing $(\eta_1, \eta_2) = (\xi + \zeta_1, \xi + \zeta_2)$, we find from \eqref{defPhi} that
\begin{align*}
\Phi(\xi, \eta_1, \eta_2)
& = \frac{3\xi}{(1+\xi^2)^{5/2}}\zeta_1 \zeta_2 + O\bigg(\zeta_1^3 + \zeta_2^3\bigg)
=\frac{3\xi}{(1+\xi^2)^{5/2}}\zeta_1 \zeta_2 + O\left([ \varrho_1(t)]^3 \right).
\end{align*}
Since $p_0 = 10^{-4}$ and $ \varrho_1(t)$ satisfies \eqref{varrho}, the error term is integrable in time, so
we now only need to estimate
\begin{align}
\label{modscat}
\begin{split}
J_3 &:= \bigg\|\iint_{\mathbb{R}^2} e^{- i t \mathfrak{A} \zeta_1 \zeta_2} \psi\bigg(\frac{\zeta_1}{ \varrho_1(t)}\bigg) \cdot \psi\bigg(\frac{\zeta_2}{ \varrho_1(t)}\bigg) \,\mathrm{d}{\zeta_1} \,\mathrm{d}{\zeta_2} - \frac{2 \pi }{\mathfrak{A}t}\bigg\|_{L^\infty_\xi}.
\end{split}
\end{align}
Making the change of variables
\[
\zeta_1 = \sqrt{\frac{1}{\mathfrak{A}t}} x_1,\qquad \zeta_2 = \sqrt{\frac{1}{\mathfrak{A}t}} x_2,
\]
in \eqref{modscat} and using the fact that $|\xi| \le (t + 1)^{p_1}$, we find that
\begin{align}
\label{modscatest}
\begin{split}
J_3 &\le \frac{(t + 1)^{p_1}}{t} \bigg\|\iint_{\mathbb{R}^2} e^{- i x_1 x_2} \psi\bigg(\frac{1}{\sqrt{\mathfrak{A}t} \varrho_1(t)} x_1\bigg) \cdot \psi\bigg(\frac{1}{\sqrt{\mathfrak{A}t} \varrho_1(t)} x_2\bigg) \,\mathrm{d}{x_1} \,\mathrm{d}{x_2} - 2 \pi\bigg\|_{L^\infty_\xi}.
\end{split}
\end{align}
The integral identity
\[
\int_\mathbb{R} e^{- a x^2 - b x} \,\mathrm{d}{x} =\sqrt{ \frac{\pi}{a}} e^{\frac{b^2}{4 a}}, \qquad \text{for all $a, b \in \mathbb{C}$ with $\Re{a} > 0$}
\]
gives that
\[
\iint_{\mathbb{R}^2} e^{- i x_1 x_2} e^{- \frac{x_1^2}{\mathfrak{B}^2}} e^{- \frac{x_2^2}{\mathfrak{B}^2}} \,\mathrm{d}{x_1} \,\mathrm{d}{x_2} = \sqrt{\pi} \mathfrak{B} \int_\mathbb{R} e^{- \frac{x_2^2}{\mathfrak{B}^2}} e^{- \frac{\mathfrak{B}^2 x_2^2}{4}} \,\mathrm{d}{x_2} = 2 \pi + O(\mathfrak{B}^{-1}),
\qquad \text{as $\mathfrak{B}\to \infty$},
\]
and therefore
\begin{align}
\label{2piint}
\left|\iint_{\mathbb{R}^2} e^{- i x_1 x_2} \psi\bigg(\frac{x_1}{\mathfrak{B}}\bigg) \psi\bigg(\frac{x_2}{\mathfrak{B}}\bigg) \,\mathrm{d}{x_1} \,\mathrm{d}{x_2} -2 \pi\right|\lesssim \mathfrak{B}^{- 1 / 2},
\qquad \text{as $\mathfrak{B}\to \infty$}.
\end{align}
Thus
Using \eqref{2piint} with $\mathfrak{B} = \sqrt{\mathfrak{A}t} \varrho_1(t) = O(t^{0.05})$ in \eqref{modscatest} then yields
\begin{align*}
J_3 \lesssim (t + 1)^{ p_1-1.025}.
\end{align*}
Since $p_1 = 10^{-5}$, the right-hand side decays faster in time than $1 / t$, which implies that \eqref{modscat2} is integrable in time and bounded by a constant multiple of $\varepsilon_0^3$.
Putting all the above estimates together, we conclude that
\[
\int_0^{\infty} \|(|\xi|^{r}+|\xi|^{w}) U(t, \xi)\|_{L^\infty_\xi} \,\mathrm{d} t\lesssim \varepsilon_0.
\]
\subsection{Higher-degree terms}\label{higherorder}
In this subsection, we prove that
\[
\left\|\mathfrak{d}(t, \xi)(|\xi|^{r+1}+|\xi|^{w+1})\widehat{\mathcal{N}_{\geq5}(\varphi)}\right\|_{L^\infty_\xi}
\]
is integrable in time.
By Lemma \ref{multilinear} and the symbol estimate Proposition \ref{Prop_symb_Tmu}, we have
\begin{align*}
\left\|\mathfrak{d}(t, \xi)(|\xi|^{r+1}+|\xi|^{w+1})\widehat{\mathcal{N}_{\geq5}(\varphi)}\right\|_{L^\infty_\xi}\lesssim& ~(t+1)^{(w+1)p_1}\|\mathcal{N}_{\geq 5}(\varphi)\|_{L^1} \lesssim (t+1)^{(w+1)p_1} \|\varphi\|_{L^2}^2\sum\limits_{n=2}^\infty \left(\|\varphi\|_{B^{1,6}}^{2n-1}\right).
\end{align*}
Using the dispersive estimate Lemma~\ref{NonDis}, we see that the right-hand-side is integrable in $t$, which leads to
\[
\int_0^\infty \left\| \mathfrak{d}(t, \xi)(|\xi|^{r+1}+|\xi|^{w+1})\widehat{\mathcal N_{\geq5}(\varphi)} \right\|_{L_\xi^\infty}\,\mathrm{d} t\lesssim \varepsilon_0.
\]
This completes the proof of Theorem~\ref{bootstrap}.
| b77b36f8ef8bea217564c00fbc8258988e3ab49f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction \& Related Work}
There is a substantial gap between how many users dream of interacting with intelligent robots and how those robots are programmed in reality. The dream is for the human user to instruct their robot in something not too far from natural language, e.g. ``please visit both the gas station and the grocery store, and make sure you get back here within 30 minutes'', or ``land at one of three landing pads, but stay clear of other aircraft''. Unfortunately, robots today usually expect much more concrete guidance, such as a specific trajectory or feedback controller. As the tasks we wish to assign our robots grow increasingly complex, there is a correspondingly increased need for flexible specification of robot programs and tools to automatically derive concrete plans from those specifications. Moreover, since the real world is unavoidably messy, any plan thus derived must also be robust to unforeseen variation in the environment; the robot must be able to accomplish its plan even when the environment changes.
Luckily, when it comes to flexibly specifying complex tasks, we have a convenient tool in the form of \textit{temporal logic}. There are many flavors of temporal logic, but most relevant to many robotics problems is \textit{signal temporal logic} (or STL), which provides a flexible language for specifying requirements for continuous real-valued signals~\cite{donze10,Sun2022,Pant2017}. STL allows a user to specify a wide range of planning problems by combining logical and temporal operators to express requirements about ordering and dependencies between subtasks. In addition, although the formal syntax of STL can seem opaque at first, it is often quite easy to translate STL formulae into readily-understood natural language. Due to its flexibility, STL is a common choice for specifying robotics problems such as trajectory planning~\cite{Pant2018,Pantazides2022} and combined task and motion planning~\cite{Plaku2016,Sun2022,Takano2021}.
A number of classical methods exist for planning from STL specifications, the most common being abstraction-based methods~\cite{Plaku2016}, mixed-integer optimization-based methods~\cite{Sun2022,yang20}, and nonlinear optimization methods~\cite{Pant2018, Pantazides2022, Leung2020} (other approaches include sampling-based methods such as~\cite{kantaros20} and~\cite{vasile17}). Abstraction-based methods have the longest history; these methods first construct a discrete abstraction (in the form of a graph or automaton) of the continuous state space, then plan over this discrete abstraction~\cite{Plaku2016}. The drawback of abstraction-based methods is that the size of the discrete abstraction grows exponentially with the dimension of the state space, limiting the scalability of these methods.
Other methods, based on mixed-integer optimization, exploit the fact that STL specifications can be expressed as linear constraints with integer variables, and the resulting optimization formulation provide soundness and completeness guarantees. Unfortunately, although mixed-integer optimization is sound and complete, these mixed-integer programs quickly become intractable as the planning horizon increases~\cite{raman15,sadraddini15,yang20}. Some works reduce the size of the program by using timed waypoints instead of a receding horizon~\cite{Sun2022}, but this requires assumptions (such as access to a bounded-error tracking controller) that can be restrictive.
A more recent line of work has focused on solving STL planning problems using nonlinear optimization~\cite{Pant2017,Pant2018,Pantazides2022,Leung2020,Takano2021}. In these approaches, the STL specification is replaced with a continuously differentiable approximation and optimized using local gradient-based methods. These approaches achieve increased generality and scalability by sacrificing completeness and optimality guarantees.
A significant gap in the state of the art is that existing optimization-based STL planners~\cite{Pant2017,Pant2018,Takano2021,Leung2020,Pantazides2022,Sun2022} do not explicitly consider the effects of environmental disturbances while planning. These approaches include some amount of robustness implicitly, typically by maximizing the margin by which a plan satisfies the STL specification, but this is often not sufficient in practice to prevent the plan from failing in response to small changes in the environment. Some methods do explicitly consider robustness to disturbances~\cite{raman15}, but our experiments show that they yield mixed-integer problems that are intractable in practice.
In this paper, we fill this gap by developing a robust planner that uses counterexamples (examples of environmental changes that cause the plan to fail) to refine its plan using nonlinear optimization. This planner relies on an iterative optimization process, inspired by solution methods for multi-player games, that alternates between finding a plan that performs well for all counterexamples seen so far and finding new counterexamples to guide the optimization process. Our framework relies on differentiable simulation and differentiable temporal logic to derive gradients of the plan's performance with respect to both the planning parameters and the environmental disturbance, enabling an efficient search for new plans and counterexamples.
We compare our approach against state-of-the-art methods, including both mixed-integer methods~\cite{raman15} and nonlinear optimization methods~\cite{Pant2018,Pantazides2022,Leung2020}. We find that our method not only finds plans that succeed despite worst-case disturbances from the environment, but it also requires less than half the time of the next-most-successful method. Our approach easily scales to handle long-horizon tasks with complex STL specifications that are not tractable for mixed-integer programming, and the plans found using our method are consistently more robust than those found using existing methods.
\section{Preliminaries}\label{preliminaries}
We begin by introducing the syntax and semantics of signal temporal logic, or STL. STL defines properties about real-valued functions of time $s: \R^+ \mapsto \R^n$ called \textit{signals}. For our purposes, a signal is defined by a finite number of sampled points $(t_i, x_i)$, and we assume that the signal is piecewise-affine in between sampled points and constant after the last sample. Syntactically, an STL formula is constructed from predicates based on functions $\mu: \R^n \mapsto \R$, logical connectives, and temporal operators~\cite{donze13}. The syntax of an STL formula $\psi$ is defined inductively as:
\begin{align*}
\psi = \text{true}\ |\ \mu(x) \geq 0\ |\ \neg \psi\ |\ \psi_1 \wedge \psi_2\ |\ \psi_1 \ \until_I\ \psi_2
\end{align*}
where $I$ is a closed (but potentially unbounded) time interval and $\until_I$ is the ``until'' operator (read as: within interval $I$, $\psi_1$ must be true until $\psi_2$ becomes true). For convenience, when $I$ is omitted it is assumed to be $[0, \infty)$. Additional temporal operators such as eventually $\eventually_I\ \psi = \text{true } \until_I\ \psi$ and always $\always_I = \neg \eventually_I\ \neg \psi$ follow from this basic syntax, as do logical operators such as $\vee$ and $\implies$.
For any signal $s$, an STL formula is satisfied at a given time $t$ according to the following Boolean semantics~\cite{donze13}:
\begin{alignat*}{3}
&s, t &&\models \text{true} && \\
&s, t &&\models \mu(x) \geq 0\quad &&\text{iff}\ \mu(s(t)) \geq 0 \\
&s, t &&\models \neg \psi &&\text{iff}\ s, t \not\models \psi \\
&s, t &&\models \psi_1 \wedge \psi_2 &&\text{iff}\ s, t \models \psi_1 \text{ and } s, t \models \psi_2 \\
&s, t &&\models \psi_1\ \until_I\ \psi_2 &&\text{iff } \exists\ t' \in t + I \text{ s.t. } w, t' \models \psi_2 \\
& && && \phantom{iff} \text{ and } w, t'' \models \psi_1\ \forall\ t'' \in [t, t']
\end{alignat*}
A useful feature of STL is that, in addition to the Boolean semantics defined above, it also admits a \textit{quantitative semantics} giving the margin of satisfaction (or robustness margin) of an STL formula, denoted $\rho$. The formula is satisfied when $\rho > 0$ and not satisfied when $\rho < 0$. The robustness margin can also be defined inductively:
\begin{align*}
\rho(\text{true}, s, t) &= \top \\
\rho(\mu(x) \geq 0, s, t) &= \mu(s(t)) \\
\rho(\neg\psi, s, t) &= -\rho(\psi, s, t) \\
\rho(\psi_1 \wedge \psi_2, s, t) &= \min\{\rho(\psi_1, s, t), \rho(\psi_2, s, t)\} \\
\rho(\psi_1 \until_I\ \psi_2, s, t) &= \sup_{t' \in t + I} \min\{\rho(\psi_2, s, t'), \inf_{t'' \in [t, t']} \rho(\psi_1, s, t'')
\end{align*}
where $\top$ is a constant taken to be greater than all other real values. In practice, linear-time algorithms exist for evaluating $\rho$ given a piecewise-affine signal $s$~\cite{donze13}.
It is important to make a distinction between the robustness margin of the specification, $\rho$, and the robustness of a plan designed to satisfy that specification. $\rho$ measures the margin by which the specification is met for a particular execution of a plan, but it does not provide much information about whether the specification will hold across multiple executions, particularly when external disturbances can affect those executions. In the next section, we formalize the robust planning problem, which aims at finding a plan that will satisfy the STL specification even when affected by external disturbances.
STL syntax may appear opaque at first glance, but its myriad symbols belie the fact that it is often straightforward to translate an STL formula into easily-understood natural language. For example, $\eventually_{[10, 20]} ((\always_{[0, 5]}\ x \geq 0)\ \until\ y \leq 0)$ can be read as ``between 10--\SI{20}{s} from now, $x$ must be positive for \SI{5}{s} before $y$ becomes negative.'' We provide more examples of STL formulae for robotics problems in Section~\ref{experiments}.
\section{Problem Statement}
In this paper, we focus on the problem of robust planning from an STL specification, which we view as a sequential two-player zero-sum game between the planner and its environment. In the first step of this game (planning time), the planner has the opportunity to tune a set of \textit{design parameters} $\theta$, but in the second step (run-time) the environment can change a distinct set of \textit{exogenous parameters} $\chi$ to degrade the performance of the plan. Together, $\theta \in \Theta$ and $\chi \in \mathcal{X}$ define the behavior of an autonomous system $\xi: \Theta \times \mathcal{X} \mapsto X^T$, which we assume is a known simulator function mapping design and exogenous parameters to a length-$T$ trace of states $x_t \in X$. We assume that $\xi$ is deterministic, so all uncertainty must be imported via $\chi$, but we assume that $\chi$ may be chosen \textit{adversarially} to degrade the performance of our chosen $\theta$ as much as possible. We also assume that the designer must commit to a choice of $\theta$ before the adversary chooses $\chi$
The performance of a plan is given by a cost function $J: X^T \mapsto \R$ assigning a scalar cost to a behavior trace. To accommodate STL specifications, we deal mainly with cost functions of the form
\begin{align*}
J_\psi(\theta, \chi) = -\rho\pn{\psi, \xi(\theta, \chi)} + \lambda J_{other}(\theta, \chi)
\end{align*}
where $\rho\pn{\psi, \xi(\theta, \chi)}$ is the robustness margin of the behavior trace with respect to a given STL specification $\psi$. We negate $\rho$ so that minimizing $J$ maximizes the robustness margin, and the $\lambda J_{other}$ term permits us to consider other factors in the plan's performance (e.g. fuel use). The scaling factor $\lambda$ is typically small to prioritize satisfying the STL specification.
Since we assume that $\chi$ can vary adversarially to impose worst-case performance for any plan $\theta$, our goal is to find $\theta$ that are robust to this variation. Concretely, our goal is to solve an optimization problem representing a two-step sequential zero-sum game with two players:
\begin{align}
\max_{\chi \in \mathcal{X}}\ \min_{\theta \in \Theta}\ J(\theta, \chi) \label{planning_problem}
\end{align}
To make this discussion concrete, consider a simple example of path planning for an aerial robot. In this case, $\psi$ might specify that we eventually ($\eventually$) reach a goal and always ($\always$) avoid some obstacles, $\theta$ might represent the locations of waypoints along the path and the parameters of a trajectory-tracking controller to follow those waypoints, and $\chi$ might represent the force from wind that attempts to drive the robot off course. The behavior $\xi$ might be a function that simulates the dynamics of the robot flying through wind, and the additional cost $J_{other}$ might impose a small penalty on large control inputs to conserve battery life. We provide more in-depth examples in Section~\ref{experiments}.
Our formulation differs from that presented in~\cite{Pant2018} and~\cite{Pantazides2022}; although both of these works seek to maximize the robustness margin $\rho$, neither consider the effect of disturbances $\chi$. Our formulation is also distinct from the mixed-integer formulation in~\cite{Sun2022}, since we consider $\rho$ as part of an objective rather than as a constraint. Our unconstrained approach does not provide the same completeness guarantees as a mixed-integer constrained optimization (used in~\cite{raman15,sadraddini15,Sun2022}), but empirical results in Section~\ref{experiments} demonstrate that our approach scales much better.
Of course, solving~\eqref{planning_problem} to global optimality in the general nonlinear case is intractable. Instead, we take advantage of this game structure to design an iterative algorithm to find the \textit{generalized Nash equilibrium}: the design parameters $\theta$ and corresponding $\chi$ such that neither the planner nor the adversary have an incentive to change their choice~\cite{Facchinei2007}. The next section describes this iterative algorithm, which we implement using nonlinear programming with differentiable simulation and differentiable temporal logic.
\section{Approach}
To solve the robust STL planning problem~\eqref{planning_problem}, we need to address two key points. First, we must develop a meta-heuristic to find a generalized Nash equilibrium of the sequential game~\eqref{planning_problem}, taking care that we do not overfit to any particular value of $\chi$.
We solve this challenge by developing an iterative counterexample-guided nonlinear optimization framework. Second, in order to solve this problem using nonlinear optimization, we need an efficient way to compute gradients of $J$ with respect to both $\theta$ and $\chi$, which requires us to differentiate not only the behavior function $\xi$ but also the robustness margin computation $\rho$. We address this challenge using differentiable programming, which we discuss next before introducing our high-level counterexample-guided optimization strategy.
\subsection{Differentiable Simulation and Temporal Logic}\label{autodiff}
Although it is possible to solve nonlinear optimization problems without access to the gradients of the objective or constraint functions, either by estimating gradients~\cite{suh2021_bundled_gradients} or using zero-order methods~\cite{nevergrad}, it is often much faster to use exact gradient information when it is available. However, exact gradients can be difficult to derive symbolically for complex optimization problems. Instead, recent works have turned to \textit{automatic differentiation} using differentiable programming to automatically compute gradients in problems such as 3D shape optimization~\cite{cascaval2021differentiable}, aircraft design optimization~\cite{sharpe_thesis}, robot design optimization~\cite{dawson2022architect1,du2021underwater}, and machine learning~\cite{jax2018github}.
Inspired by this trend, we implement $\xi$ using the JAX framework for automatic differentiation~\cite{jax2018github}, yielding a differentiable simulation of the underlying autonomous system. For a system where the behavior is defined by continuous-time dynamics $\dot{x} = f(x, \theta, \chi, t)$, implementing numerical integration in a differentiable language such as JAX allows us to automatically back-propagate through the simulator to find the gradients $\nabla_\theta \xi$ and $\nabla_\chi \xi$. These gradients can typically be computed much more quickly using automatic differentiation than by finite-difference methods~\cite{dawson2022architect1}.
We can use a similar differentiable programming approach to obtain gradients through the quantitative semantics of an STL specification. Before doing so, we must replace the discontinuous $\max$ and $\min$ operators used to compute $\rho$ with smooth approximations:
\begin{align*}
\widetilde{\max}(x_1, x_2, \ldots) &= \frac{1}{k}\log\pn{e^{kx_1} + e^{kx_2} + \ldots} \\
\widetilde{\min}(x_1, x_2, \ldots) &= -\widetilde{\max}(-x_1, -x_2, \ldots)
\end{align*}
where $k$ is a smoothing parameter and $\lim_{k\to\infty} \widetilde{\max} = \max$. This differentiable relaxation was introduced in~\cite{Pant2017} and later used in~\cite{Pant2018,Pantazides2022}; \cite{Leung2020} uses a slightly different approximation.
Using these smooth approximations, we implement the fast, linear-time algorithms for computing the robustness margin proposed by~\cite{donze13}, using the JAX framework to enable efficient automatic differentiation. In contrast to~\cite{Leung2020}, our method achieves computational complexity that is linear in the length of the state trace $T$ (the complexity of the \texttt{stlcg} framework in \cite{Leung2020} is quadratic in $T$ for the $\mathcal{U}$ operator).
By combining smooth approximations of STL quantitative semantics with differentiable programming, we can efficiently compute the gradients $\nabla_\theta \rho$ and $\nabla_\xi \rho$. By combining these gradients with those found using differentiable simulation, we can efficiently compute the gradient of the objective $J$ with respect to both the design parameters $\theta$ and the adversary's response $\chi$. Usefully, our use of differentiable programming means that we are not restricted to considering trajectory planning separately from the design of a tracking controller, as in~\cite{Pant2018} and~\cite{Sun2022}. Instead, we can consider an end-to-end gradient that combines the planned trajectory and controller parameters in $\theta$ and optimizes them jointly (see Section~\ref{experiments} for an example of this end-to-end optimization). In the next section, we discuss how end-to-end gradients enable an iterative algorithm for counterexample-guided robust optimization.
\subsection{Counterexample-guided Optimization}
To solve the planning problem in~\eqref{planning_problem}, we need to find a generalized Nash equilibrium between the planner and the adversary; i.e. values of $\theta$ and $\chi$ where neither we nor the adversary has any local incentive to change. A common solution strategy for such problems is the family of nonlinear Gauss-Seidel-type methods~\cite{Facchinei2007}. These methods solve max-min problems like~\eqref{planning_problem} by alternating between $\theta$ and $\chi$, tuning one set of parameters while keeping the others constant; i.e. alternating between the two optimization problems:
\begin{subequations}
\begin{align}\label{eq:gauss_seidel}
\theta^* &= \argmin_\theta J(\theta, \chi^*) \\
\chi^* &= \argmax_\chi J(\theta^*, \chi)
\end{align}
\end{subequations}
Although these methods are not guaranteed to converge, it is known that if they do, then the convergence point $(\theta^*, \chi^*)$ is a Nash equilibrium~\cite{Facchinei2007}.
A risk of applying such a simple alternating scheme is that the nonlinear optimization for both $\theta$ and $\chi$ can easily get caught in local minima. Such local minima not only reduce the performance of the optimized plan, but also increase the risk of ``overfitting'' to a particular value of $\chi^*$. This risk is particularly salient because the planner must commit to a choice of $\theta$ before the adversary has a final opportunity to choose $\chi$. To mitigate this risk and improve the robustness of our optimized plan, we extend a standard Gauss-Seidel method with two ideas from the machine learning and optimization literature. First, we take inspiration from the success of domain randomization in robust machine learning~\cite{tobin2017}: instead of optimizing $\theta$ with respect to a single fixed $\chi^*$, we can maintain a dataset $\mathcal{X}_N = \set{\chi_i}_{i=1,\ldots,N}$ and optimize the performance of $\theta$ across all of these samples:
\begin{subequations}
\begin{align}\label{eq:gauss_seidel_domain_randomization}
\theta^* &= \argmin_\theta \mathbb{E}_{\mathcal{X}_N} \left[ J(\theta, \chi_i) \right]\\
\chi^* &= \argmax_\chi J(\theta^*, \chi)
\end{align}
\end{subequations}
Incorporating domain randomization into the Gauss-Seidel method has the potential to improve the robustness of the resulting equilibria, but it is relatively sample inefficient; it may require a large number of random samples $\chi_i$. To address this sample inefficiency, we take inspiration from a second idea in the optimization and learning literature: learning from counterexamples~\cite{Chang2019}. The key insight here is that we can do better than simply randomly sampling $\chi_i$; we can use the values of $\chi^*$ found during successive iterations of the Gauss-Seidel process as high-quality counterexamples to guide the optimization of $\theta$. This insight results in our counterexample-guided Gauss-Seidel optimization method, which is outlined in pseudocode in Algorithm~\ref{alg:cg_gs}.
Our algorithm proceeds as follows. We begin by initializing the dataset with $N_0$ i.i.d. examples $\chi_i$, then we alternate between solving the two optimization problems in~\eqref{eq:gauss_seidel_domain_randomization}. At each iteration, we add our current estimate of the adversary's best response $\chi^*$ to the dataset, and we stop either when the algorithm reaches a fixed point (the adversary's best response after solving~\eqref{eq:gauss_seidel_domain_randomization} is the same as the best response from the previous round) or when a maximum number of iterations is reached. As we show experimentally in Section~\ref{experiments}, this counterexample-guided optimization achieves a higher sample efficiency than simple domain randomization, in that it finds plans that are more robust to adversarial disturbance while considering a much smaller dataset. Although our use of nonlinear optimization means that our algorithm is not complete, we find empirically that it succeeds in finding a satisfactory plan in the large majority of cases.
It is important to note that this algorithm is fundamentally enabled by the automatic differentiation approach detailed in Section~\ref{autodiff}; without access to the gradients of $J$ it would be much more difficult to solve the subproblems in lines~\ref{alg:opt_theta} and~\ref{alg:opt_chi} of Algorithm~\ref{alg:cg_gs}. Although some previous approaches obtain gradients of STL satisfaction with respect to $\theta$ using standard trajectory optimization formulations, as in~\cite{Pant2018}, we are not aware of any approaches that make use of gradients with respect to disturbance parameters. There has been some work on using counterexamples to guide mixed-integer planning~\cite{raman15}, but our experiments in the next section demonstrate that these mixed-integer programs are intractable for long horizon problems. Specifically, we find that solving even a single mixed-integer program can take more than an hour, so solving multiple programs to derive counterexamples is not a practical solution. In the next section, we demonstrate that our gradient-based counterexample-guided approach outperforms these existing approaches, not only finding more robust plans but requiring substantially less computation time to do so.
\begin{algorithm}
\caption{Counterexample-guided Gauss-Seidel method for solving robust planning problems}\label{alg:cg_gs}
\DontPrintSemicolon
\KwInput{Starting dataset size $N_0$\\\phantom{Input: } Maximum number of iterations $M$}
\KwOutput{Optimized design parameters $\theta^*$\\\phantom{Output: } Dataset of counterexamples $\mathcal{X}_N$}
$\mathcal{X}_N \gets $ $N_0$ examples $\chi_i \in \mathcal{X}$ sampled uniformly i.i.d.\;
$\chi^*_{prev} \gets \varnothing$\;
\For{$i \in \set{1, \ldots, M}$}
{
$\theta^* = \argmin_\theta \mathbb{E}_{\mathcal{X}_N} \left[ J(\theta, \chi_i) \right] \label{alg:opt_theta}$ \;
$\chi^* = \argmax_\chi J(\theta^*, \chi)$ \label{alg:opt_chi} \;
\If{$\chi^* = \chi^*_{prev}$}{\Break}
$\chi^*_{prev} \gets \chi^*$\;
Append $\chi^*$ to $\mathcal{X}_N$\;
}
\KwRet{$\theta^*$, $\mathcal{X}_N$}
\end{algorithm}
\section{Experiments}\label{experiments}
We validate our approach by means of two case studies involving the satellite rendezvous problem posed in~\cite{Jewison2016}. We benchmark against state-of-the-art planning algorithms to show the robustness and scalability benefits of our approach.
In this satellite rendezvous problem, the goal is to maneuver a chaser satellite to catch a target satellite. We can express this problem in the Clohessy-Wiltshire-Hill coordinate frame~\cite{Jewison2016}, which assumes that the target's orbit is circular and constructs a coordinate frame with the origin at the target, the $x$-axis pointing away from the Earth, the $y$-axis pointing along the target's orbit, and the $z$-axis pointing out of the orbital plane. In this frame, the chaser's dynamics are approximately linear, with positions $p_x$, $p_y$, $p_z$ and velocities $v_x$, $v_y$, $v_z$ varying according to controlled thrust in each direction $u_x$, $u_y$, $u_z$:
\begin{align*}
\mat{\dot{p}_x \\ \dot{p}_y \\ \dot{p}_z \\ \dot{v}_x \\ \dot{v}_y \\ \dot{v}_z} = \mat{
v_x \\
v_y \\
v_z \\
3n^2 p_x + 2n v_y + u_x / m\\
-2n v_x + u_y / m\\
-n^2 p_z + u_z / m
}
\end{align*}
$n = \sqrt{\mu / a^3}$ is the mean-motion of the target, determined by the Earth's gravitational constant $\mu = \SI{3.986e14}{m^3/s^2}$ and the target's altitude $a$ (i.e. the length of the semi-major orbital axis, \SI{353}{km} in low Earth orbit). $m = \SI{500}{kg}$ is the mass of the chaser satellite~\cite{Jewison2016}.
In this setting, we construct STL specifications for two rendezvous missions: a simple low-speed rendezvous and a more complex loiter-then-rendezvous mission, illustrated in Fig.~\ref{fig:mission_specs}. The STL specifications for each mission, $\psi_1$ and $\psi_2$, are given formally as:
\begin{align*}
\psi_1 &= \psi_\text{reach target} \wedge \psi_\text{speed limit} \\
\psi_2 &= \psi_\text{reach target} \wedge \psi_\text{speed limit} \wedge \psi_\text{loiter} \\
\psi_\text{reach target} &= \eventually \pn{r \leq 0.1} \\
\psi_\text{speed limit} &= \pn{r \geq 2.0} \until\ \always \pn{v \leq 0.1} \\
\psi_\text{loiter} &= \eventually \always_{[0, T_{obs}]} \pn{2.0 \leq r \wedge r \leq 3.0}
\end{align*}
where $r = \sqrt{p_x^2 + p_y^2 + p_z^2}$ and $v = \sqrt{v_x^2 + v_y^2 + v_z^2}$. Informally, $\psi_1$ requires that ``the chaser eventually comes within \SI{0.1}{m} of the target and does not come within \SI{2}{m} of the target until its speed is less than \SI{0.1}{m/s}'', and $\psi_2$ additionally requires that ``the chaser eventually spends at least $T_{obs}$ seconds in the region between 2--\SI{3}{m} from the target.'' For each mission, the design parameters $\theta$ include both state/input waypoints along a planned trajectory and the feedback gains used to track that trajectory, and the exogenous parameters $\chi$ represent bounded uncertainty in the initial state of the chaser ($p_x(0), p_y(0) \in [10, 13]$, $p_z(0) \in [-3, 3]$, $v_x(0), v_y(0), v_z(0) \in [-1, 1]$). We use a \SI{200}{s}-long simulation with a \SI{2}{s} timestep for both missions, and $T_{obs} = \SI{10}{s}$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/satellite_missions.png}
\caption{Two satellite rendezvous missions used to test our framework. In the first mission, the chaser satellite must eventually reach the target while respecting a maximum speed constraint in the region immediately around the target. In the second mission, the chaser must still reach the target and obey the speed limit, but it must also loiter in an observation region for some minimum time before approaching. The first mission requires an STL formula with three predicates and three temporal operators, while the second mission requires five predicates and five temporal operators.}
\label{fig:mission_specs}
\end{figure}
For each mission $i=1, 2$, we define a cost function as $J_i = \rho_i + \lambda I$, where $\rho_i = \rho(\psi_i, \xi(\theta, \chi), 0)$ is the STL robustness margin at the start of the trajectory, $I$ is the total impulse required to execute the maneuver (in Newton-seconds), and $\lambda = 5\times10^{-5}$. By applying our iterative counterexample-guided optimization strategy to this problem, we find the optimized trajectories for mission 1 and 2 shown in Figs.~\ref{fig:mission_1_traj} and~\ref{fig:mission_2_traj} along with the worst-case $\chi$. In these examples, we use $N_0=8$ initial examples and $M=10$ maximum rounds, but the algorithm converges in less than 10 rounds in all trials. In both missions, our approach reliably finds a solution that remains feasible despite worst-case variation in the exogenous parameters, achieving a positive STL robustness margin in $>90\%$ of trials in each case. Our counterexample-guided approach requires an average of \SI{53.7}{s} to solve mission 1 and \SI{194.2}{s} to solve mission 2 (averaged across 50 trials).
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{figs/satellite_mission1_traj.png}
\caption{The optimized trajectory found using our counterexample-guided optimization strategy for mission 1 (rendezvous with speed constraint). The chaser satellite only enters the speed-limit zone once it has slowed down sufficiently.}
\label{fig:mission_1_traj}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/mission_1_comparison.png}
\caption{Comparison of different STL planning methods on the first example mission, averaged over 50 random seeds. Left: the robustness margin $\rho(\psi_1)$ computed for the optimized design parameters and worst-case exogenous parameters. Right: the planning time required by each method. Our method (CG) achieves much higher robustness than all other methods (satisfying the STL specification despite adversarial perturbations in all but 3 instances) and runs twice as fast as the next-most-robust method.}
\label{fig:mission_1_comparison}
\end{figure}
We can quantitatively compare our approach against two state of the art approaches: a mixed-integer STL planner based on that in~\cite{raman15} and~\cite{sadraddini15} and the nonlinear optimization approach in~\cite{Pant2018,Pantazides2022}. The mixed-integer planner (MIP) in~\cite{raman15} uses a model-predictive control formulation and proposes to add counterexamples after solving each instance of the mixed-integer program; however, we found that even a single instance could not be solved to optimality within 1 hour for either mission, and so we compare with the best solution found within a given period of time. Even though the size of the mixed-integer program in~\cite{raman15} grows linearly with the horizon of the problem, the complexity of solving the resulting MIP grows exponentially in the number of integer variables (these problems require between 2800--4500 integer variables when solved using Gurobi). Since it was not tractable to solve even once instance of the MIP, we were unable to use any MIP-generated counterexamples as proposed in~\cite{raman15}; instead, we take the best feasible solution found after \SI{500}{s} for the first mission and after \SI{1000}{s} for the second mission.
We also compare with an extension of the nonlinear optimization from~\cite{Pant2018,Pantazides2022}, where we add domain randomization to the authors' existing trajectory optimization formulation, averaging the objective over either 32 or 64 different values of $\chi$. We note that these methods rely on a similar optimization approach as those proposed in~\cite{Leung2020}, but we re-implement the authors' method to ensure a fair comparison (our implementation makes use of just-in-time compilation to speed the optimization process, and so comparing with off-the-shelf implementations like \texttt{stlcg}~\cite{Leung2020} would not be fair).
A comparison of our counterexample-guided approach (CG), nonlinear optimization with domain randomization (NLopt), and the mixed-integer planner (MIP) is shown in Fig.~\ref{fig:mission_1_comparison} for the first mission and Fig.~\ref{fig:mission_2_comparison} for the second mission. In all cases, we compare across 50 random trials, computing the time required to solve each instance and the robustness of the optimized plan when subject to adversarial disturbances. All experiments were run on a laptop computer with \SI{8}{GB} RAM and a \SI{1.8}{GHz} 8-core processor.
We find that our method is consistently more robust than prior methods; in the first mission, it satisfies the STL specification in all but 3 trials, despite adversarial disturbances. For comparison, the next-best method (NLopt with domain randomization across 64 examples) failed to solve the first mission in 14 out of 50 trials and took more than twice as long on average to find a plan (\SI{114.3}{s} as opposed to \SI{53.7}{s} for our method). This advantage is due to the quality of the examples used during optimization; instead of 64 random samples, our method uses 8 initial random samples and between 1 and 4 counterexamples (median 2) representing worst-case variation in $\chi$, making our method much more sample-efficient.
We also find that our method finds more robust solutions than the MIP method, since MIP cannot tractably consider variation in $\chi$ (the MIP method is also unable to find a feasible solution within \SI{500}{s} in 16 out of 50 trials). MIP's performance also suffers due to discretization error, since we were forced to discretize the continuous-time dynamics with relatively few knot points (one every \SI{2}{s}) to yield a tractable MIP optimization problem.
Our method also performs well on the second mission planning problem: only our method consistently finds solutions that are robust to variation in $\chi$ (see Fig.~\ref{fig:mission_2_comparison}). Due to the increased complexity of this example, the MIP method finds a feasible solution within \SI{1000}{s} in only 16 out of 50 trials (the MIP encoding of this mission requires 7769 continuous variables, 2806 binary variables, and 1479 constraints), and the feasible solutions found within \SI{1000}{s} tend to be of poor quality. The second-most-robust method, NLopt with 64 random examples, takes more than twice as long as our method and fails to satisfy the STL specification in 17 out of 50 trials (compared to only 4 failures for our method). Our method required a median of 2 counterexamples in addition to the 8 initial examples to solve this planning problem (the slowest trial required 7 additional examples).
These data demonstrate that our counterexample-guided approach to planning from STL specifications is faster, more sample efficient, and more robust to adversarial disturbances than state-of-the-art approaches. A software implementation of our method is available online at \url{https://github.com/MIT-REALM/architect}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/satellite_mission2_traj.png}
\caption{The optimized trajectory found using our counterexample-guided optimization strategy for mission 2 (loiter then rendezvous with speed constraint). The optimized plan satisfies the additional mission requirement of spending time in the observation region before approaching the target.}
\label{fig:mission_2_traj}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/mission_2_comparison.png}
\caption{Comparison of different STL planning methods on the second example mission, averaged over 50 random seeds. Left: the robustness margin $\rho(\psi_2)$ computed for the optimized design parameters and worst-case exogenous parameters. Right: time required by each method to find a plan. Our method (CG) finds much more robust plans, satisfying the specification in all but 4 instances compared to 17 failures for the next-best method (NLopt with 64 examples). Our method also runs more than twice as fast as the next-most-robust method.}
\label{fig:mission_2_comparison}
\end{figure}
\subsection{Hardware Demonstration}
We also validate our approach in hardware by solving the loiter-then-rendezvous mission for a Turtlebot 3 ground robot. We replace the satellite dynamics with nonlinear Dubins dynamics and plan a trajectory that satisfies $\psi = \psi_\text{reach target} \wedge \psi_\text{loiter}$ (we do not include the speed limit because the Turtlebot already has a relatively small maximum speed). Our counterexample-guided planner solves this problem in \SI{26.22}{s}, yielding the trajectory shown in Fig.~\ref{fig:hw}. A video of this demonstration is included in the supplementary materials.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/hw_mission_2.png}
\caption{The trajectory found using our counterexample-guided planner successfully moves the robot through the observation zone (where it must spend at least \SI{10}{s}) and into the docking zone. Odometry data indicate that the planned trajectory achieves an STL robustness margin of $0.0255$.}
\label{fig:hw}
\end{figure}
\section{Conclusion}
In this paper, we introduce a novel framework for robust optimization-based planning from temporal logic specifications. We frame the robust planning problem as a sequential two-player game between the planner, which chooses a set of design parameters at planning time, and the environment, which picks disturbances adversarially in response to the planner's choice. We develop an iterative counterexample-guided algorithm to find plans that robustly satisfy temporal logic specifications despite worst-case disturbances from the environment. Our method, which relies on differentiable programming for simulation and evaluating the temporal logic specification, not only finds more robust plans than state-of-the-art methods but also runs substantially faster. We apply our method to two planning problems involving time horizons $>$\SI{100}{s} and STL specifications with multiple nested temporal operators, and we provide source code for our approach at \url{https://github.com/MIT-REALM/architect}.
In future work, we hope to extend our framework to offer completeness guarantees by combining our counterexample-guided optimization with complete model checking and STL falsification methods, such as the Breach framework~\cite{donze10}. We also look forward to exploring applications of this framework to problems involving multiple agents, particularly for human-robot interaction, and to combining local gradient-based optimization with sampling based methods to solve more complex task-and-motion planning problems.
\bibliographystyle{IEEEtran}
\section{Introduction \& Related Work}
There is a substantial gap between how many users dream of interacting with intelligent robots and how those robots are programmed in reality. The dream is for the human user to instruct their robot in something not too far from natural language, e.g. ``please visit both the gas station and the grocery store, and make sure you get back here within 30 minutes'', or ``land at one of three landing pads, but stay clear of other aircraft''. Unfortunately, robots today usually expect much more concrete guidance, such as a specific trajectory or feedback controller. As the tasks we wish to assign our robots grow increasingly complex, there is a correspondingly increased need for flexible specification of robot programs and tools to automatically derive concrete plans from those specifications. Moreover, since the real world is unavoidably messy, any plan thus derived must also be robust to unforeseen variation in the environment; the robot must be able to accomplish its plan even when the environment changes.
Luckily, when it comes to flexibly specifying complex tasks, we have a convenient tool in the form of \textit{temporal logic}. There are many flavors of temporal logic, but most relevant to many robotics problems is \textit{signal temporal logic} (or STL), which provides a flexible language for specifying requirements for continuous real-valued signals~\cite{donze10,Sun2022,Pant2017}. STL allows a user to specify a wide range of planning problems by combining logical and temporal operators to express requirements about ordering and dependencies between subtasks. In addition, although the formal syntax of STL can seem opaque at first, it is often quite easy to translate STL formulae into readily-understood natural language. Due to its flexibility, STL is a common choice for specifying robotics problems such as trajectory planning~\cite{Pant2018,Pantazides2022} and combined task and motion planning~\cite{Plaku2016,Sun2022,Takano2021}.
A number of classical methods exist for planning from STL specifications, the most common being abstraction-based methods~\cite{Plaku2016}, mixed-integer optimization-based methods~\cite{Sun2022,yang20}, and nonlinear optimization methods~\cite{Pant2018, Pantazides2022, Leung2020} (other approaches include sampling-based methods such as~\cite{kantaros20} and~\cite{vasile17}). Abstraction-based methods have the longest history; these methods first construct a discrete abstraction (in the form of a graph or automaton) of the continuous state space, then plan over this discrete abstraction~\cite{Plaku2016}. The drawback of abstraction-based methods is that the size of the discrete abstraction grows exponentially with the dimension of the state space, limiting the scalability of these methods.
Other methods, based on mixed-integer optimization, exploit the fact that STL specifications can be expressed as linear constraints with integer variables, and the resulting optimization formulation provide soundness and completeness guarantees. Unfortunately, although mixed-integer optimization is sound and complete, these mixed-integer programs quickly become intractable as the planning horizon increases~\cite{raman15,sadraddini15,yang20}. Some works reduce the size of the program by using timed waypoints instead of a receding horizon~\cite{Sun2022}, but this requires assumptions (such as access to a bounded-error tracking controller) that can be restrictive.
A more recent line of work has focused on solving STL planning problems using nonlinear optimization~\cite{Pant2017,Pant2018,Pantazides2022,Leung2020,Takano2021}. In these approaches, the STL specification is replaced with a continuously differentiable approximation and optimized using local gradient-based methods. These approaches achieve increased generality and scalability by sacrificing completeness and optimality guarantees.
A significant gap in the state of the art is that existing optimization-based STL planners~\cite{Pant2017,Pant2018,Takano2021,Leung2020,Pantazides2022,Sun2022} do not explicitly consider the effects of environmental disturbances while planning. These approaches include some amount of robustness implicitly, typically by maximizing the margin by which a plan satisfies the STL specification, but this is often not sufficient in practice to prevent the plan from failing in response to small changes in the environment. Some methods do explicitly consider robustness to disturbances~\cite{raman15}, but our experiments show that they yield mixed-integer problems that are intractable in practice.
In this paper, we fill this gap by developing a robust planner that uses counterexamples (examples of environmental changes that cause the plan to fail) to refine its plan using nonlinear optimization. This planner relies on an iterative optimization process, inspired by solution methods for multi-player games, that alternates between finding a plan that performs well for all counterexamples seen so far and finding new counterexamples to guide the optimization process. Our framework relies on differentiable simulation and differentiable temporal logic to derive gradients of the plan's performance with respect to both the planning parameters and the environmental disturbance, enabling an efficient search for new plans and counterexamples.
We compare our approach against state-of-the-art methods, including both mixed-integer methods~\cite{raman15} and nonlinear optimization methods~\cite{Pant2018,Pantazides2022,Leung2020}. We find that our method not only finds plans that succeed despite worst-case disturbances from the environment, but it also requires less than half the time of the next-most-successful method. Our approach easily scales to handle long-horizon tasks with complex STL specifications that are not tractable for mixed-integer programming, and the plans found using our method are consistently more robust than those found using existing methods.
\section{Preliminaries}\label{preliminaries}
We begin by introducing the syntax and semantics of signal temporal logic, or STL. STL defines properties about real-valued functions of time $s: \R^+ \mapsto \R^n$ called \textit{signals}. For our purposes, a signal is defined by a finite number of sampled points $(t_i, x_i)$, and we assume that the signal is piecewise-affine in between sampled points and constant after the last sample. Syntactically, an STL formula is constructed from predicates based on functions $\mu: \R^n \mapsto \R$, logical connectives, and temporal operators~\cite{donze13}. The syntax of an STL formula $\psi$ is defined inductively as:
\begin{align*}
\psi = \text{true}\ |\ \mu(x) \geq 0\ |\ \neg \psi\ |\ \psi_1 \wedge \psi_2\ |\ \psi_1 \ \until_I\ \psi_2
\end{align*}
where $I$ is a closed (but potentially unbounded) time interval and $\until_I$ is the ``until'' operator (read as: within interval $I$, $\psi_1$ must be true until $\psi_2$ becomes true). For convenience, when $I$ is omitted it is assumed to be $[0, \infty)$. Additional temporal operators such as eventually $\eventually_I\ \psi = \text{true } \until_I\ \psi$ and always $\always_I = \neg \eventually_I\ \neg \psi$ follow from this basic syntax, as do logical operators such as $\vee$ and $\implies$.
For any signal $s$, an STL formula is satisfied at a given time $t$ according to the following Boolean semantics~\cite{donze13}:
\begin{alignat*}{3}
&s, t &&\models \text{true} && \\
&s, t &&\models \mu(x) \geq 0\quad &&\text{iff}\ \mu(s(t)) \geq 0 \\
&s, t &&\models \neg \psi &&\text{iff}\ s, t \not\models \psi \\
&s, t &&\models \psi_1 \wedge \psi_2 &&\text{iff}\ s, t \models \psi_1 \text{ and } s, t \models \psi_2 \\
&s, t &&\models \psi_1\ \until_I\ \psi_2 &&\text{iff } \exists\ t' \in t + I \text{ s.t. } w, t' \models \psi_2 \\
& && && \phantom{iff} \text{ and } w, t'' \models \psi_1\ \forall\ t'' \in [t, t']
\end{alignat*}
A useful feature of STL is that, in addition to the Boolean semantics defined above, it also admits a \textit{quantitative semantics} giving the margin of satisfaction (or robustness margin) of an STL formula, denoted $\rho$. The formula is satisfied when $\rho > 0$ and not satisfied when $\rho < 0$. The robustness margin can also be defined inductively:
\begin{align*}
\rho(\text{true}, s, t) &= \top \\
\rho(\mu(x) \geq 0, s, t) &= \mu(s(t)) \\
\rho(\neg\psi, s, t) &= -\rho(\psi, s, t) \\
\rho(\psi_1 \wedge \psi_2, s, t) &= \min\{\rho(\psi_1, s, t), \rho(\psi_2, s, t)\} \\
\rho(\psi_1 \until_I\ \psi_2, s, t) &= \sup_{t' \in t + I} \min\{\rho(\psi_2, s, t'), \inf_{t'' \in [t, t']} \rho(\psi_1, s, t'')
\end{align*}
where $\top$ is a constant taken to be greater than all other real values. In practice, linear-time algorithms exist for evaluating $\rho$ given a piecewise-affine signal $s$~\cite{donze13}.
It is important to make a distinction between the robustness margin of the specification, $\rho$, and the robustness of a plan designed to satisfy that specification. $\rho$ measures the margin by which the specification is met for a particular execution of a plan, but it does not provide much information about whether the specification will hold across multiple executions, particularly when external disturbances can affect those executions. In the next section, we formalize the robust planning problem, which aims at finding a plan that will satisfy the STL specification even when affected by external disturbances.
STL syntax may appear opaque at first glance, but its myriad symbols belie the fact that it is often straightforward to translate an STL formula into easily-understood natural language. For example, $\eventually_{[10, 20]} ((\always_{[0, 5]}\ x \geq 0)\ \until\ y \leq 0)$ can be read as ``between 10--\SI{20}{s} from now, $x$ must be positive for \SI{5}{s} before $y$ becomes negative.'' We provide more examples of STL formulae for robotics problems in Section~\ref{experiments}.
\section{Problem Statement}
In this paper, we focus on the problem of robust planning from an STL specification, which we view as a sequential two-player zero-sum game between the planner and its environment. In the first step of this game (planning time), the planner has the opportunity to tune a set of \textit{design parameters} $\theta$, but in the second step (run-time) the environment can change a distinct set of \textit{exogenous parameters} $\chi$ to degrade the performance of the plan. Together, $\theta \in \Theta$ and $\chi \in \mathcal{X}$ define the behavior of an autonomous system $\xi: \Theta \times \mathcal{X} \mapsto X^T$, which we assume is a known simulator function mapping design and exogenous parameters to a length-$T$ trace of states $x_t \in X$. We assume that $\xi$ is deterministic, so all uncertainty must be imported via $\chi$, but we assume that $\chi$ may be chosen \textit{adversarially} to degrade the performance of our chosen $\theta$ as much as possible. We also assume that the designer must commit to a choice of $\theta$ before the adversary chooses $\chi$
The performance of a plan is given by a cost function $J: X^T \mapsto \R$ assigning a scalar cost to a behavior trace. To accommodate STL specifications, we deal mainly with cost functions of the form
\begin{align*}
J_\psi(\theta, \chi) = -\rho\pn{\psi, \xi(\theta, \chi)} + \lambda J_{other}(\theta, \chi)
\end{align*}
where $\rho\pn{\psi, \xi(\theta, \chi)}$ is the robustness margin of the behavior trace with respect to a given STL specification $\psi$. We negate $\rho$ so that minimizing $J$ maximizes the robustness margin, and the $\lambda J_{other}$ term permits us to consider other factors in the plan's performance (e.g. fuel use). The scaling factor $\lambda$ is typically small to prioritize satisfying the STL specification.
Since we assume that $\chi$ can vary adversarially to impose worst-case performance for any plan $\theta$, our goal is to find $\theta$ that are robust to this variation. Concretely, our goal is to solve an optimization problem representing a two-step sequential zero-sum game with two players:
\begin{align}
\max_{\chi \in \mathcal{X}}\ \min_{\theta \in \Theta}\ J(\theta, \chi) \label{planning_problem}
\end{align}
To make this discussion concrete, consider a simple example of path planning for an aerial robot. In this case, $\psi$ might specify that we eventually ($\eventually$) reach a goal and always ($\always$) avoid some obstacles, $\theta$ might represent the locations of waypoints along the path and the parameters of a trajectory-tracking controller to follow those waypoints, and $\chi$ might represent the force from wind that attempts to drive the robot off course. The behavior $\xi$ might be a function that simulates the dynamics of the robot flying through wind, and the additional cost $J_{other}$ might impose a small penalty on large control inputs to conserve battery life. We provide more in-depth examples in Section~\ref{experiments}.
Our formulation differs from that presented in~\cite{Pant2018} and~\cite{Pantazides2022}; although both of these works seek to maximize the robustness margin $\rho$, neither consider the effect of disturbances $\chi$. Our formulation is also distinct from the mixed-integer formulation in~\cite{Sun2022}, since we consider $\rho$ as part of an objective rather than as a constraint. Our unconstrained approach does not provide the same completeness guarantees as a mixed-integer constrained optimization (used in~\cite{raman15,sadraddini15,Sun2022}), but empirical results in Section~\ref{experiments} demonstrate that our approach scales much better.
Of course, solving~\eqref{planning_problem} to global optimality in the general nonlinear case is intractable. Instead, we take advantage of this game structure to design an iterative algorithm to find the \textit{generalized Nash equilibrium}: the design parameters $\theta$ and corresponding $\chi$ such that neither the planner nor the adversary have an incentive to change their choice~\cite{Facchinei2007}. The next section describes this iterative algorithm, which we implement using nonlinear programming with differentiable simulation and differentiable temporal logic.
\section{Approach}
To solve the robust STL planning problem~\eqref{planning_problem}, we need to address two key points. First, we must develop a meta-heuristic to find a generalized Nash equilibrium of the sequential game~\eqref{planning_problem}, taking care that we do not overfit to any particular value of $\chi$.
We solve this challenge by developing an iterative counterexample-guided nonlinear optimization framework. Second, in order to solve this problem using nonlinear optimization, we need an efficient way to compute gradients of $J$ with respect to both $\theta$ and $\chi$, which requires us to differentiate not only the behavior function $\xi$ but also the robustness margin computation $\rho$. We address this challenge using differentiable programming, which we discuss next before introducing our high-level counterexample-guided optimization strategy.
\subsection{Differentiable Simulation and Temporal Logic}\label{autodiff}
Although it is possible to solve nonlinear optimization problems without access to the gradients of the objective or constraint functions, either by estimating gradients~\cite{suh2021_bundled_gradients} or using zero-order methods~\cite{nevergrad}, it is often much faster to use exact gradient information when it is available. However, exact gradients can be difficult to derive symbolically for complex optimization problems. Instead, recent works have turned to \textit{automatic differentiation} using differentiable programming to automatically compute gradients in problems such as 3D shape optimization~\cite{cascaval2021differentiable}, aircraft design optimization~\cite{sharpe_thesis}, robot design optimization~\cite{dawson2022architect1,du2021underwater}, and machine learning~\cite{jax2018github}.
Inspired by this trend, we implement $\xi$ using the JAX framework for automatic differentiation~\cite{jax2018github}, yielding a differentiable simulation of the underlying autonomous system. For a system where the behavior is defined by continuous-time dynamics $\dot{x} = f(x, \theta, \chi, t)$, implementing numerical integration in a differentiable language such as JAX allows us to automatically back-propagate through the simulator to find the gradients $\nabla_\theta \xi$ and $\nabla_\chi \xi$. These gradients can typically be computed much more quickly using automatic differentiation than by finite-difference methods~\cite{dawson2022architect1}.
We can use a similar differentiable programming approach to obtain gradients through the quantitative semantics of an STL specification. Before doing so, we must replace the discontinuous $\max$ and $\min$ operators used to compute $\rho$ with smooth approximations:
\begin{align*}
\widetilde{\max}(x_1, x_2, \ldots) &= \frac{1}{k}\log\pn{e^{kx_1} + e^{kx_2} + \ldots} \\
\widetilde{\min}(x_1, x_2, \ldots) &= -\widetilde{\max}(-x_1, -x_2, \ldots)
\end{align*}
where $k$ is a smoothing parameter and $\lim_{k\to\infty} \widetilde{\max} = \max$. This differentiable relaxation was introduced in~\cite{Pant2017} and later used in~\cite{Pant2018,Pantazides2022}; \cite{Leung2020} uses a slightly different approximation.
Using these smooth approximations, we implement the fast, linear-time algorithms for computing the robustness margin proposed by~\cite{donze13}, using the JAX framework to enable efficient automatic differentiation. In contrast to~\cite{Leung2020}, our method achieves computational complexity that is linear in the length of the state trace $T$ (the complexity of the \texttt{stlcg} framework in \cite{Leung2020} is quadratic in $T$ for the $\mathcal{U}$ operator).
By combining smooth approximations of STL quantitative semantics with differentiable programming, we can efficiently compute the gradients $\nabla_\theta \rho$ and $\nabla_\xi \rho$. By combining these gradients with those found using differentiable simulation, we can efficiently compute the gradient of the objective $J$ with respect to both the design parameters $\theta$ and the adversary's response $\chi$. Usefully, our use of differentiable programming means that we are not restricted to considering trajectory planning separately from the design of a tracking controller, as in~\cite{Pant2018} and~\cite{Sun2022}. Instead, we can consider an end-to-end gradient that combines the planned trajectory and controller parameters in $\theta$ and optimizes them jointly (see Section~\ref{experiments} for an example of this end-to-end optimization). In the next section, we discuss how end-to-end gradients enable an iterative algorithm for counterexample-guided robust optimization.
\subsection{Counterexample-guided Optimization}
To solve the planning problem in~\eqref{planning_problem}, we need to find a generalized Nash equilibrium between the planner and the adversary; i.e. values of $\theta$ and $\chi$ where neither we nor the adversary has any local incentive to change. A common solution strategy for such problems is the family of nonlinear Gauss-Seidel-type methods~\cite{Facchinei2007}. These methods solve max-min problems like~\eqref{planning_problem} by alternating between $\theta$ and $\chi$, tuning one set of parameters while keeping the others constant; i.e. alternating between the two optimization problems:
\begin{subequations}
\begin{align}\label{eq:gauss_seidel}
\theta^* &= \argmin_\theta J(\theta, \chi^*) \\
\chi^* &= \argmax_\chi J(\theta^*, \chi)
\end{align}
\end{subequations}
Although these methods are not guaranteed to converge, it is known that if they do, then the convergence point $(\theta^*, \chi^*)$ is a Nash equilibrium~\cite{Facchinei2007}.
A risk of applying such a simple alternating scheme is that the nonlinear optimization for both $\theta$ and $\chi$ can easily get caught in local minima. Such local minima not only reduce the performance of the optimized plan, but also increase the risk of ``overfitting'' to a particular value of $\chi^*$. This risk is particularly salient because the planner must commit to a choice of $\theta$ before the adversary has a final opportunity to choose $\chi$. To mitigate this risk and improve the robustness of our optimized plan, we extend a standard Gauss-Seidel method with two ideas from the machine learning and optimization literature. First, we take inspiration from the success of domain randomization in robust machine learning~\cite{tobin2017}: instead of optimizing $\theta$ with respect to a single fixed $\chi^*$, we can maintain a dataset $\mathcal{X}_N = \set{\chi_i}_{i=1,\ldots,N}$ and optimize the performance of $\theta$ across all of these samples:
\begin{subequations}
\begin{align}\label{eq:gauss_seidel_domain_randomization}
\theta^* &= \argmin_\theta \mathbb{E}_{\mathcal{X}_N} \left[ J(\theta, \chi_i) \right]\\
\chi^* &= \argmax_\chi J(\theta^*, \chi)
\end{align}
\end{subequations}
Incorporating domain randomization into the Gauss-Seidel method has the potential to improve the robustness of the resulting equilibria, but it is relatively sample inefficient; it may require a large number of random samples $\chi_i$. To address this sample inefficiency, we take inspiration from a second idea in the optimization and learning literature: learning from counterexamples~\cite{Chang2019}. The key insight here is that we can do better than simply randomly sampling $\chi_i$; we can use the values of $\chi^*$ found during successive iterations of the Gauss-Seidel process as high-quality counterexamples to guide the optimization of $\theta$. This insight results in our counterexample-guided Gauss-Seidel optimization method, which is outlined in pseudocode in Algorithm~\ref{alg:cg_gs}.
Our algorithm proceeds as follows. We begin by initializing the dataset with $N_0$ i.i.d. examples $\chi_i$, then we alternate between solving the two optimization problems in~\eqref{eq:gauss_seidel_domain_randomization}. At each iteration, we add our current estimate of the adversary's best response $\chi^*$ to the dataset, and we stop either when the algorithm reaches a fixed point (the adversary's best response after solving~\eqref{eq:gauss_seidel_domain_randomization} is the same as the best response from the previous round) or when a maximum number of iterations is reached. As we show experimentally in Section~\ref{experiments}, this counterexample-guided optimization achieves a higher sample efficiency than simple domain randomization, in that it finds plans that are more robust to adversarial disturbance while considering a much smaller dataset. Although our use of nonlinear optimization means that our algorithm is not complete, we find empirically that it succeeds in finding a satisfactory plan in the large majority of cases.
It is important to note that this algorithm is fundamentally enabled by the automatic differentiation approach detailed in Section~\ref{autodiff}; without access to the gradients of $J$ it would be much more difficult to solve the subproblems in lines~\ref{alg:opt_theta} and~\ref{alg:opt_chi} of Algorithm~\ref{alg:cg_gs}. Although some previous approaches obtain gradients of STL satisfaction with respect to $\theta$ using standard trajectory optimization formulations, as in~\cite{Pant2018}, we are not aware of any approaches that make use of gradients with respect to disturbance parameters. There has been some work on using counterexamples to guide mixed-integer planning~\cite{raman15}, but our experiments in the next section demonstrate that these mixed-integer programs are intractable for long horizon problems. Specifically, we find that solving even a single mixed-integer program can take more than an hour, so solving multiple programs to derive counterexamples is not a practical solution. In the next section, we demonstrate that our gradient-based counterexample-guided approach outperforms these existing approaches, not only finding more robust plans but requiring substantially less computation time to do so.
\begin{algorithm}
\caption{Counterexample-guided Gauss-Seidel method for solving robust planning problems}\label{alg:cg_gs}
\DontPrintSemicolon
\KwInput{Starting dataset size $N_0$\\\phantom{Input: } Maximum number of iterations $M$}
\KwOutput{Optimized design parameters $\theta^*$\\\phantom{Output: } Dataset of counterexamples $\mathcal{X}_N$}
$\mathcal{X}_N \gets $ $N_0$ examples $\chi_i \in \mathcal{X}$ sampled uniformly i.i.d.\;
$\chi^*_{prev} \gets \varnothing$\;
\For{$i \in \set{1, \ldots, M}$}
{
$\theta^* = \argmin_\theta \mathbb{E}_{\mathcal{X}_N} \left[ J(\theta, \chi_i) \right] \label{alg:opt_theta}$ \;
$\chi^* = \argmax_\chi J(\theta^*, \chi)$ \label{alg:opt_chi} \;
\If{$\chi^* = \chi^*_{prev}$}{\Break}
$\chi^*_{prev} \gets \chi^*$\;
Append $\chi^*$ to $\mathcal{X}_N$\;
}
\KwRet{$\theta^*$, $\mathcal{X}_N$}
\end{algorithm}
\section{Experiments}\label{experiments}
We validate our approach by means of two case studies involving the satellite rendezvous problem posed in~\cite{Jewison2016}. We benchmark against state-of-the-art planning algorithms to show the robustness and scalability benefits of our approach.
In this satellite rendezvous problem, the goal is to maneuver a chaser satellite to catch a target satellite. We can express this problem in the Clohessy-Wiltshire-Hill coordinate frame~\cite{Jewison2016}, which assumes that the target's orbit is circular and constructs a coordinate frame with the origin at the target, the $x$-axis pointing away from the Earth, the $y$-axis pointing along the target's orbit, and the $z$-axis pointing out of the orbital plane. In this frame, the chaser's dynamics are approximately linear, with positions $p_x$, $p_y$, $p_z$ and velocities $v_x$, $v_y$, $v_z$ varying according to controlled thrust in each direction $u_x$, $u_y$, $u_z$:
\begin{align*}
\mat{\dot{p}_x \\ \dot{p}_y \\ \dot{p}_z \\ \dot{v}_x \\ \dot{v}_y \\ \dot{v}_z} = \mat{
v_x \\
v_y \\
v_z \\
3n^2 p_x + 2n v_y + u_x / m\\
-2n v_x + u_y / m\\
-n^2 p_z + u_z / m
}
\end{align*}
$n = \sqrt{\mu / a^3}$ is the mean-motion of the target, determined by the Earth's gravitational constant $\mu = \SI{3.986e14}{m^3/s^2}$ and the target's altitude $a$ (i.e. the length of the semi-major orbital axis, \SI{353}{km} in low Earth orbit). $m = \SI{500}{kg}$ is the mass of the chaser satellite~\cite{Jewison2016}.
In this setting, we construct STL specifications for two rendezvous missions: a simple low-speed rendezvous and a more complex loiter-then-rendezvous mission, illustrated in Fig.~\ref{fig:mission_specs}. The STL specifications for each mission, $\psi_1$ and $\psi_2$, are given formally as:
\begin{align*}
\psi_1 &= \psi_\text{reach target} \wedge \psi_\text{speed limit} \\
\psi_2 &= \psi_\text{reach target} \wedge \psi_\text{speed limit} \wedge \psi_\text{loiter} \\
\psi_\text{reach target} &= \eventually \pn{r \leq 0.1} \\
\psi_\text{speed limit} &= \pn{r \geq 2.0} \until\ \always \pn{v \leq 0.1} \\
\psi_\text{loiter} &= \eventually \always_{[0, T_{obs}]} \pn{2.0 \leq r \wedge r \leq 3.0}
\end{align*}
where $r = \sqrt{p_x^2 + p_y^2 + p_z^2}$ and $v = \sqrt{v_x^2 + v_y^2 + v_z^2}$. Informally, $\psi_1$ requires that ``the chaser eventually comes within \SI{0.1}{m} of the target and does not come within \SI{2}{m} of the target until its speed is less than \SI{0.1}{m/s}'', and $\psi_2$ additionally requires that ``the chaser eventually spends at least $T_{obs}$ seconds in the region between 2--\SI{3}{m} from the target.'' For each mission, the design parameters $\theta$ include both state/input waypoints along a planned trajectory and the feedback gains used to track that trajectory, and the exogenous parameters $\chi$ represent bounded uncertainty in the initial state of the chaser ($p_x(0), p_y(0) \in [10, 13]$, $p_z(0) \in [-3, 3]$, $v_x(0), v_y(0), v_z(0) \in [-1, 1]$). We use a \SI{200}{s}-long simulation with a \SI{2}{s} timestep for both missions, and $T_{obs} = \SI{10}{s}$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/satellite_missions.png}
\caption{Two satellite rendezvous missions used to test our framework. In the first mission, the chaser satellite must eventually reach the target while respecting a maximum speed constraint in the region immediately around the target. In the second mission, the chaser must still reach the target and obey the speed limit, but it must also loiter in an observation region for some minimum time before approaching. The first mission requires an STL formula with three predicates and three temporal operators, while the second mission requires five predicates and five temporal operators.}
\label{fig:mission_specs}
\end{figure}
For each mission $i=1, 2$, we define a cost function as $J_i = \rho_i + \lambda I$, where $\rho_i = \rho(\psi_i, \xi(\theta, \chi), 0)$ is the STL robustness margin at the start of the trajectory, $I$ is the total impulse required to execute the maneuver (in Newton-seconds), and $\lambda = 5\times10^{-5}$. By applying our iterative counterexample-guided optimization strategy to this problem, we find the optimized trajectories for mission 1 and 2 shown in Figs.~\ref{fig:mission_1_traj} and~\ref{fig:mission_2_traj} along with the worst-case $\chi$. In these examples, we use $N_0=8$ initial examples and $M=10$ maximum rounds, but the algorithm converges in less than 10 rounds in all trials. In both missions, our approach reliably finds a solution that remains feasible despite worst-case variation in the exogenous parameters, achieving a positive STL robustness margin in $>90\%$ of trials in each case. Our counterexample-guided approach requires an average of \SI{53.7}{s} to solve mission 1 and \SI{194.2}{s} to solve mission 2 (averaged across 50 trials).
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{figs/satellite_mission1_traj.png}
\caption{The optimized trajectory found using our counterexample-guided optimization strategy for mission 1 (rendezvous with speed constraint). The chaser satellite only enters the speed-limit zone once it has slowed down sufficiently.}
\label{fig:mission_1_traj}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/mission_1_comparison.png}
\caption{Comparison of different STL planning methods on the first example mission, averaged over 50 random seeds. Left: the robustness margin $\rho(\psi_1)$ computed for the optimized design parameters and worst-case exogenous parameters. Right: the planning time required by each method. Our method (CG) achieves much higher robustness than all other methods (satisfying the STL specification despite adversarial perturbations in all but 3 instances) and runs twice as fast as the next-most-robust method.}
\label{fig:mission_1_comparison}
\end{figure}
We can quantitatively compare our approach against two state of the art approaches: a mixed-integer STL planner based on that in~\cite{raman15} and~\cite{sadraddini15} and the nonlinear optimization approach in~\cite{Pant2018,Pantazides2022}. The mixed-integer planner (MIP) in~\cite{raman15} uses a model-predictive control formulation and proposes to add counterexamples after solving each instance of the mixed-integer program; however, we found that even a single instance could not be solved to optimality within 1 hour for either mission, and so we compare with the best solution found within a given period of time. Even though the size of the mixed-integer program in~\cite{raman15} grows linearly with the horizon of the problem, the complexity of solving the resulting MIP grows exponentially in the number of integer variables (these problems require between 2800--4500 integer variables when solved using Gurobi). Since it was not tractable to solve even once instance of the MIP, we were unable to use any MIP-generated counterexamples as proposed in~\cite{raman15}; instead, we take the best feasible solution found after \SI{500}{s} for the first mission and after \SI{1000}{s} for the second mission.
We also compare with an extension of the nonlinear optimization from~\cite{Pant2018,Pantazides2022}, where we add domain randomization to the authors' existing trajectory optimization formulation, averaging the objective over either 32 or 64 different values of $\chi$. We note that these methods rely on a similar optimization approach as those proposed in~\cite{Leung2020}, but we re-implement the authors' method to ensure a fair comparison (our implementation makes use of just-in-time compilation to speed the optimization process, and so comparing with off-the-shelf implementations like \texttt{stlcg}~\cite{Leung2020} would not be fair).
A comparison of our counterexample-guided approach (CG), nonlinear optimization with domain randomization (NLopt), and the mixed-integer planner (MIP) is shown in Fig.~\ref{fig:mission_1_comparison} for the first mission and Fig.~\ref{fig:mission_2_comparison} for the second mission. In all cases, we compare across 50 random trials, computing the time required to solve each instance and the robustness of the optimized plan when subject to adversarial disturbances. All experiments were run on a laptop computer with \SI{8}{GB} RAM and a \SI{1.8}{GHz} 8-core processor.
We find that our method is consistently more robust than prior methods; in the first mission, it satisfies the STL specification in all but 3 trials, despite adversarial disturbances. For comparison, the next-best method (NLopt with domain randomization across 64 examples) failed to solve the first mission in 14 out of 50 trials and took more than twice as long on average to find a plan (\SI{114.3}{s} as opposed to \SI{53.7}{s} for our method). This advantage is due to the quality of the examples used during optimization; instead of 64 random samples, our method uses 8 initial random samples and between 1 and 4 counterexamples (median 2) representing worst-case variation in $\chi$, making our method much more sample-efficient.
We also find that our method finds more robust solutions than the MIP method, since MIP cannot tractably consider variation in $\chi$ (the MIP method is also unable to find a feasible solution within \SI{500}{s} in 16 out of 50 trials). MIP's performance also suffers due to discretization error, since we were forced to discretize the continuous-time dynamics with relatively few knot points (one every \SI{2}{s}) to yield a tractable MIP optimization problem.
Our method also performs well on the second mission planning problem: only our method consistently finds solutions that are robust to variation in $\chi$ (see Fig.~\ref{fig:mission_2_comparison}). Due to the increased complexity of this example, the MIP method finds a feasible solution within \SI{1000}{s} in only 16 out of 50 trials (the MIP encoding of this mission requires 7769 continuous variables, 2806 binary variables, and 1479 constraints), and the feasible solutions found within \SI{1000}{s} tend to be of poor quality. The second-most-robust method, NLopt with 64 random examples, takes more than twice as long as our method and fails to satisfy the STL specification in 17 out of 50 trials (compared to only 4 failures for our method). Our method required a median of 2 counterexamples in addition to the 8 initial examples to solve this planning problem (the slowest trial required 7 additional examples).
These data demonstrate that our counterexample-guided approach to planning from STL specifications is faster, more sample efficient, and more robust to adversarial disturbances than state-of-the-art approaches. A software implementation of our method is available online at \url{https://github.com/MIT-REALM/architect}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/satellite_mission2_traj.png}
\caption{The optimized trajectory found using our counterexample-guided optimization strategy for mission 2 (loiter then rendezvous with speed constraint). The optimized plan satisfies the additional mission requirement of spending time in the observation region before approaching the target.}
\label{fig:mission_2_traj}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/mission_2_comparison.png}
\caption{Comparison of different STL planning methods on the second example mission, averaged over 50 random seeds. Left: the robustness margin $\rho(\psi_2)$ computed for the optimized design parameters and worst-case exogenous parameters. Right: time required by each method to find a plan. Our method (CG) finds much more robust plans, satisfying the specification in all but 4 instances compared to 17 failures for the next-best method (NLopt with 64 examples). Our method also runs more than twice as fast as the next-most-robust method.}
\label{fig:mission_2_comparison}
\end{figure}
\subsection{Hardware Demonstration}
We also validate our approach in hardware by solving the loiter-then-rendezvous mission for a Turtlebot 3 ground robot. We replace the satellite dynamics with nonlinear Dubins dynamics and plan a trajectory that satisfies $\psi = \psi_\text{reach target} \wedge \psi_\text{loiter}$ (we do not include the speed limit because the Turtlebot already has a relatively small maximum speed). Our counterexample-guided planner solves this problem in \SI{26.22}{s}, yielding the trajectory shown in Fig.~\ref{fig:hw}. A video of this demonstration is included in the supplementary materials.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/hw_mission_2.png}
\caption{The trajectory found using our counterexample-guided planner successfully moves the robot through the observation zone (where it must spend at least \SI{10}{s}) and into the docking zone. Odometry data indicate that the planned trajectory achieves an STL robustness margin of $0.0255$.}
\label{fig:hw}
\end{figure}
\section{Conclusion}
In this paper, we introduce a novel framework for robust optimization-based planning from temporal logic specifications. We frame the robust planning problem as a sequential two-player game between the planner, which chooses a set of design parameters at planning time, and the environment, which picks disturbances adversarially in response to the planner's choice. We develop an iterative counterexample-guided algorithm to find plans that robustly satisfy temporal logic specifications despite worst-case disturbances from the environment. Our method, which relies on differentiable programming for simulation and evaluating the temporal logic specification, not only finds more robust plans than state-of-the-art methods but also runs substantially faster. We apply our method to two planning problems involving time horizons $>$\SI{100}{s} and STL specifications with multiple nested temporal operators, and we provide source code for our approach at \url{https://github.com/MIT-REALM/architect}.
In future work, we hope to extend our framework to offer completeness guarantees by combining our counterexample-guided optimization with complete model checking and STL falsification methods, such as the Breach framework~\cite{donze10}. We also look forward to exploring applications of this framework to problems involving multiple agents, particularly for human-robot interaction, and to combining local gradient-based optimization with sampling based methods to solve more complex task-and-motion planning problems.
\bibliographystyle{IEEEtran}
| 2687f3354e169a41239f3cbb08a2b2681b896056 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\@startsection{section}{2}%
{\z@}{.5\linespacing\@plus.7\linespacing}{.5em}%
{\normalfont\bfseries\centering}}
\def\@secnumfont{\bfseries}
\makeatother
\theoremstyle{definition}
\newtheorem{example}[theorem]{\indent Example}
\newtheorem{definition}[theorem]{\indent Definition}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{\indent Remark}
\newtheorem{notation}[theorem]{\indent Notations}
\title[Kleinian sphere packings, reflection groups, and arithmeticity]{Kleinian sphere packings, reflection groups, \\ and arithmeticity}
\author{Nikolay Bogachev}
\thanks{Bogachev was partially supported by the Russian Science Foundation, grant no. 22-41-02028.}
\address{The Kharkevich Institute for Information Transmission Problems, Moscow, Russia}
\address{Moscow Institute of Physics and Technology, Dolgoprudny, Russia}
\email{nvbogach@mail.ru}
\author{Alexander Kolpakov}
\thanks{Kolpakov was partially supported by the Swiss National Science Foundation, project no. PP00P2-202667 }
\address{Institut de Math\'ematiques, Universit\'e de Neuch\^atel, CH--2000 Neuch\^atel, Switzerland}
\email{kolpakov.alexander@gmail.com}
\author{Alex Kontorovich}
\thanks{Kontorovich was partially supported by NSF grant DMS-1802119, BSF grant 2020119, and the
2020-2021 Distinguished Visiting Professorship at the National Museum of Mathematics.}
\address{Department of Mathematics, Rutgers University, 110 Frelinghuysen Rd, Piscataway, NJ 08854, USA}
\email{alex.kontorovich@rutgers.edu}
\begin{document}
\begin{abstract}
In this paper we study crystallographic sphere packings and Kleinian sphere packings, introduced first by Kontorovich and Nakamura in 2017 and then studied further by Kapovich and Kontorovich in 2021. In particular, we solve the problem of existence of crystallographic sphere packings in certain
higher dimensions posed by Kontorovich and Nakamura. In addition, we present a geometric doubling procedure allowing to obtain sphere packings from some Coxeter polyhedra without isolated roots, and study ``properly integral'' packings (that is, ones which are integral but not superintegral).
\end{abstract}
\large
\maketitle
\section{Introduction}
\subsection{Sphere packings}
For $n\ge2$, let $\mathbb{H}^{n+1}$ be the $(n+1)$--di\-men\-sio\-nal hyperbolic space. Then the set of all ideal points of hyperbolic space, or its ideal boundary, is $\partial \mathbb{H}^{n+1} \approx \mathbb{S}^{n}$ and can be identified topologically with $\overline{\mathbb{R}^n}:= \mathbb{R}^n \cup \{\infty\}$. One can consider different coordinate systems in $\mathbb{S}^n$, and thus some of the definitions below depend on this choice of coordinates.
\begin{definition}
By a \textit{sphere packing} (or just \textit{packing}) $\mathscr{P}$ of $\mathbb{S}^n$
we mean an infinite collection $\{B_\alpha\}$ of
round balls in $\mathbb{S}^n$ so that:
\begin{itemize}
\item[\bf (1)] The interiors of the balls are disjoint, and
\item[\bf (2)] The union of the balls is dense in $\mathbb{S}^n$.
\end{itemize}
\end{definition}
Recall that, given a packing $\mathcal{P} = \{B_\alpha\}$ of $\mathbb{S}^n \approx \partial \mathbb{H}^{n+1}$, every sphere $S_\alpha = \partial B_\alpha$ also is the boundary of a hyperplane $H_\alpha \cong \mathbb{H}^n$ of $\mathbb{H}^{n+1}$, i.e.,
$S_\alpha = \partial_{\infty} H_\alpha$. Any such hyperplane $H_\alpha$ is associated with a reflection $R_\alpha \in \mathrm{Isom}(\mathbb{H}^{n+1})$, i.e., $H_\alpha$ is the mirror of the reflection $R_\alpha$. Let $\Gamma_R$ be the group generated by the reflections in these hyperplanes, $\Gamma_R:=\langle R_\alpha \mid B_\alpha \in P \rangle.$
\begin{definition}
The \textit{superpacking} of $\mathcal{P}$ is its orbit under the $\Gamma_R$--action:
$$
\widetilde{\mathcal{P}} = \Gamma_R \cdot \mathcal{P}.
$$
\end{definition}
\begin{definition}
The {\it bend} of a ball $B_\alpha\in \mathcal{P}$ is the inverse of its (signed) radius, which is defined after choosing an identification $\mathbb{S}^n=\mathbb{R}^n\cup\{\infty\}$.
A packing is {\it integral} if there is a choice of identification such that all balls $B_\alpha$ have integer bends. A packing is \textit{superintegral} if its superpacking has all integer bends.
We say a packing is \textit{properly integral} if it is integral but not superintegral.
\end{definition}
\begin{definition}
A packing $\mathscr{P}$ is \textit{Kleinian} if its set of accumulation points is the limit set of some geometrically finite discrete group
$\Gamma_S < \mathrm{Isom}(\mathbb{H}^{n+1})$. We call such a $\Gamma_S$ a {\it symmetry group} of $P$.
\end{definition}
\begin{definition}
Let $\mathcal{P} = \{B_\alpha\}$ be a Kleinian sphere packing with a symmetry group $\Gamma_S < \mathrm{Isom}(\mathbb{H}^{n+1})$. The \textit{supergroup} of $\mathcal{P}$ is the group $\widetilde{\Gamma} = \Gamma_R \rtimes \Gamma_S$ generated by the symmetry group $\Gamma_S$ and all reflections $R_\alpha$.
\end{definition}
It is a theorem that $\widetilde{\Gamma}$ is a lattice, that is, acts on $\mathbb{H}^{n+1}$ with finite covolume.
\begin{definition}
A packing is \textit{crystallographic} if it has a {\it reflective} symmetry group. That is, the symmetry group, and hence the supergroup, is generated by reflections in codimension $1$ facets.
\end{definition}
\subsection{Outline and main theorems} The main focus of this paper is on determining the dimensions in which superintegral crystallographic sphere packings exist. It is worth mentioning that Kleinian sphere packings exist in all dimensions, even those that are superintegral \cite{K^2}.
The existence of a superintegral crystallographic packing in $\mathbb{S}^n$, $n\geq 2$, implies that there exists a reflective nonuniform arithmetic lattice of simplest type defined over $\mathbb{Q}$ that acts on $\mathbb{H}^{n+1}$. (That it is reflective and arithmetic over $\mathbb Q$ follows directly from the Structure Theorem and Arithmeticity Theorem in \cite{KN}. That it is nonuniform is proved in \cite{K^2}.) A result of Esselmann \cite{E} implies that no such lattice is possible for $n\geq 21$. Previously, superintegral crystallographic sphere packings were known in dimensions $n\leq 13$ and $n=17$ by applying the techniques of \cite{KN}. We also remark that the existence of such a packing in dimension $n=18$ is claimed in \cite{KN}, although a further argument is needed (given below in Lemma~\ref{lemma:doubling}) to substantiate this claim.
The difficulty in producing a crystallographic sphere packing stems from the Structure Theorem of \cite{KN}. There it is shown that a finite covolume reflection group $\widetilde \Gamma$ gives rise to a sphere packing of $\mathbb{S}^n$ if and only if its Coxeter--Vinberg diagram (see \S\ref{sec:Coxeter}) has at least one ``isolated root,'' that is, a root which is orthogonal to all its neighbors. In terms of the Coxeter--Vinberg diagram, this means that the corresponding (isolated) vertex is connected to its neighbors only by dotted (common perpendiculars) or bold edges (parallel, i.e. tangent at the ideal boundary $\partial \mathbb{H}^{n+1}$).
Our main result about dimensions of crystallographic sphere packings is as follows.
\begin{theoremA}\label{theorem:dimension-bound}
There are no crystallographic sphere packings in $\mathbb{S}^{n}$ for $n\geq 19$.
\end{theoremA}
The theorem above is implied by the results of Esselman \cite{E}, together with the following fact.
\begin{proposition}\label{prop:lower-dimension}
If a superintegral crystallographic sphere packing exists in $\mathbb{S}^{n}$, then there exists a non-uniform arithmetic reflective lattice acting on $\mathbb{H}^n$.
\end{proposition}
As for the existence of crystallographic sphere packings, we can summarise our findings together with the previous results of \cite{KN}, in order to formulate the following statement.
\begin{theoremB}\label{theorem:crystallographic}
Superintegral crystallographic sphere packings in $\mathbb{S}^n$ exist only for $n\leq 18$. Examples are known for all $2 \leq n \leq 14$, and $n = 17, 18$.
\end{theoremB}
What is new here is the addition of dimensions $n=14$ and $n=18$. We construct these additional examples in higher dimensions by studying the lower--dimensional faces of the $21$--dimensional hyperbolic polyhedron discovered by Borcherds \cite{Bor}.
The proof of Theorem~\ref{theorem:crystallographic} relies on the following fact that allows to combine Coxeter polyhedra with many ``even'' angles into one that has a facet orthogonal to all its neighbours.
\begin{lemma}\label{lemma:doubling}
Let $P$ be a finite volume Coxeter polyhedron in $\mathbb{H}^n$ and $P_0$ be a facet that meets other facets of $P$ at angles $\pi/2$ or $\pi/4$. Let $P_1, \ldots P_k$ be all the facets of $P$ that meet $P_0$ at $\pi/4$.
Assume that each $P_j$ meets its neighbors at angles $\pi/2^m$, with $m \geq 1$, an integer.
Then the reflection group of $P$ is commensurable to the supergroup of a crystallographic packing.
\end{lemma}
Dimensions $n = 15$ and $16$ are still missing from our consideration: the gap seems non--trivial to fill, as at present, there are rather few candidate arithmetic lattices in high dimensions to explore for packings. For example, we studied another polyhedron discovered by Borcherds \cite{Borcherds2000} in $\mathbb{H}^{17}$ (with $960$ facets), only to discover that it has no isolated roots, and no Coxeter facets either. The present methods of obtaining sphere packings from polyhedra (\cite{KN} and Lemma~\ref{lemma:doubling}~loc.~cit.) cannot be applied to the polyhedron itself. We would like to thank Daniel Allcock for his invaluable help in our efforts to transfer this example into a computer--readable format.
\begin{question}
Do there exist crystallographic sphere packings in $\mathbb{S}^n$ for $n = 15$ and $16$?
\end{question}
We should stress the fact that answering the above question in the negative even in the more narrow superintegral setting will require a complete classification of reflective quadratic Lorentzian forms over $\mathbb{Q}$, which seems rather difficult at the moment.
Another interesting direction to pursue is the possible connection between arithmetic reflection groups and sphere packings. Namely, Kontorovich and Nakamura \cite{KN} asked the following question:
\begin{question}
Is it true that every arithmetic non--uniform lattice of simplest type defined over $\mathbb{Q}$ is commensurable to the supergroup of a superintegral crystallographic packing?
\end{question}
The answer turns out to be affirmative in low dimensions: here we confirm it for $n = 2, 3, 4$ by using the classification results of Scharlau, Walhorn, and Turkalj.
\begin{theoremC}\label{theorem:arithmetic-super}
Any reflective arithmetic non--uniform lattice $\Gamma < \mathrm{Isom}\,\mathbb{H}^{n+1}$, $n = 2,3,4$, of simplest type defined over $\mathbb{Q}$ is commensurable to the supergroup of a crystallographic packing in $\mathbb{S}^{n}$.
\end{theoremC}
For $n=2$, Kontrovich and Nakamura \cite{KN} confirmed the same for the finite list of reflective extended Bianchi groups. However, the Borcherds polyhedron in dimension $21$ provides a clear counterexample.
\subsection*{Acknowledgements}
The authors would like to thank Daniel Allcock for stimulating discussions.
\section{Preliminaries}\label{sec:Coxeter}
\subsection{Hyperbolic Lobachevsky space and convex polyhedra}
Let $\mathbb{R}^{n,1}$ be the $(n+1)$-dimensional real
\emph{Minkowski space}
equipped with the inner product
$$
(x, y) = -x_0 y_0 + x_1 y_1 + \ldots + x_n y_n
$$
of signature $(n, 1)$.
A \emph{vector model} of the \emph{$n$-dimensional hyperbolic Lobachevsky space} $\mathbb{H}^n$ is the above component of the standard hyperboloid lying in the \emph{future light
cone}:
$$
\mathbb{H}^n = \{x \in \mathbb{R}^{n,1} \mid (x, x) = -1, x_0 > 0\}.
$$
The points of $\mathbb{H}^n$ are called
\emph{proper points}.
The \emph{points at infinity} (or on \emph{the boundary} $\partial
\mathbb{H}^n$) in this model correspond to \emph{isotropic
one-dimensional subspaces} of $\mathbb{R}^{n,1}$, that is, rays $\{ \lambda x \mid \lambda \geq 0\}$, for $x \in \mathbb{R}^{n,1}$ such that $(x,x) = 0$.
The hyperbolic metric $\rho$ is given by
$$
\cosh \rho (x, y) = -(x, y).
$$
Let $\mathbf{O}_{n,1} (\mathbb{R})$ be the group of
orthogonal transformations of the space $\mathbb{R}^{n,1}$, and let $\mathbf{PO}_{n,1}(\mathbb{R})$ be
its subgroup of index $2$ preserving $\mathbb{H}^n$.
The \emph{isometry group} of the hyperbolic $n$-space $\mathbb{H}^n$ is
$\mathrm{Isom}(\mathbb{H}^n) \cong \mathbf{PO}_{n,1}(\mathbb{R})$.
Suppose that $e \in \mathbb{R}^{n,1}$ is a non--zero vector. Then the set
$$
H_{e} = \{x \in \mathbb{H}^n \mid (x,e) = 0\}
$$
is a \emph{hyperplane} in $\mathbb{H}^n$ and it divides the entire space into the \emph{half-spaces}
$$
H^-_e = \{x \in \mathbb{H}^n \mid (x,e) \leq 0\}, \qquad H^+_e = \{x \in \mathbb{H}^n \mid (x,e) \geq 0\}.
$$
The orthogonal transformation given by the formula
$$
R_e (x) = x - 2 \frac{(e, x)}{(e, e)} e,
$$
is called the \emph{reflection in the hyperplane}
$H_e$, which
is called the \emph{mirror} of $R_e$.
All the facts stated here for the standard Lorentzian form hold also for any inner product form $f(x, y)$ of signature $(n, 1)$, once we put $(x, y) = f(x, y)$. This makes some of our practical computations easier rather than each time passing to the standard form.
\begin{definition}
A convex polyhedron in $\mathbb{H}^n$ is the intersection of finitely many half-spaces that has non-empty interior.
A generalized convex polyhedron is the intersection (with non-empty interior) of a family (possibly, infinite)
of half-spaces such that every ball intersects only finitely many of their boundary hyperplanes.
\end{definition}
\begin{definition}
A generalized convex polyhedron is said to be acute-angled if all its dihedral angles do not exceed $\pi/2$. A generalized convex polyhedron is called a Coxeter polyhedron if all its dihedral angles are of the form $\pi/k$, where $k \in \{2,3,4,\ldots,+\infty\}$.
\end{definition}
\begin{definition}
The Coxeter--Vinberg diagram of a Coxeter polyhedron is a graph having a vertex corresponding to each facet, and vertices connected by edges of multiplicity $k$, if the corresponding facets meet at dihedral angle $\pi/(k+2)$, with $k\geq 0$, or by dotted edges. This includes the following interpretations:
\begin{itemize}
\item An edge of multiplicity $k=\infty$, represented by a bold edge. This happens when the respective facets meet at the ideal boundary $\partial \mathbb{H}^n$;
\item An edge of multiplicity $k=0$, i.e. no edge being present between a pair of vertices. This corresponds to the facets being orthogonal;
\item A dotted edge, corresponding to facets which diverge at infinity and admit a common perpendicular of length $\ell > 0$. Sometimes, such an edge may also be labelled with $\cosh \ell$ for convenience.
\end{itemize}
\end{definition}
It is known that the fundamental domains of discrete reflection groups are generalized Coxeter polyhedra (see \cite{Vin67, Vin85}).
A convex polyhedron has finite volume if and only if
it is equal to the
convex hull of finitely many points of the closure $\overline{\mathbb{H}^n} = \mathbb{H}^n \cup \partial \mathbb{H}^n$. If a polyhedron is compact, then it is a convex hull of finitely many proper points in $\mathbb{H}^n$.
It is also known that compact acute-angled hyperbolic polyhedra, in particular, compact Coxeter polyhedra in
$\mathbb{H}^n$, are \emph{simple}, i.e. each of their vertices
belongs to exactly $n$ facets (and $n$ edges).
\subsection{Kleinian sphere packing and the Structure Theorem}
In this subsection we discuss some details regarding Kleinian sphere packing and computational aspects of working with them.
In practice this works as follows. We have a quadratic form $f$ of signature $(n+1, 1)$ which defines a quadratic space $V=\mathbb{R}^{n+1,1}$. This gives us the two--sheeted convex cone $C:=\{x\in V: f(x,x) < 0\}$, and its projectivization: $\mathbf{P} C := C/\mathbb{R}_+$. Then we identify one sheet $\mathbf{P} C_+$ of this projectivized two--sheeted cone with $\mathbb{H}^{n+1}$, and use the inner product $(x,y) = f(x,y)$ to induce the hyperbolic metric on it. Here we use the same notation for the quadratic form and the induced bilinear form as it does not create much ambiguity.
Given a vector $e_\alpha\in \{v \in V: (v, v) > 0\}$, we have the hyperplane $H_\alpha = e_\alpha^\perp\cap \mathbf{P} C_+$, and a reflection $R_\alpha: x \mapsto x - 2\frac{(x, e_\alpha)}{(e_\alpha, e_\alpha)} e_\alpha$ with mirror $H_\alpha$. This reflection induces an inversion in the sphere $S_\alpha = \partial H_\alpha$.
Dual to $V$ is the vector space $V^*$ which is a quadratic space under $f^*$. Then to construct an inversive coordinate system, we simply proceed as follows. To choose coordinates, we first choose any two $f^*$--isotropic (over $\mathbb{R}$) vectors $b, \hat{b} \in V^*$, such that $f^*(b,\hat{b}) = -2$.
Then take any system of orthonormal vectors $b_j$, $j\in\{1,\dots,n\}$ in the space orthogonal to that generated by $b$ and $\hat{b}$. The inversive coordinate system of a sphere $S_\alpha$ corresponding to the vector $e_\alpha\in\{v\in V: (v,v)>0\}$
is then given by
$$
v
_\alpha = ( b(e_\alpha),\hat b(e_\alpha), b_1(e_\alpha),\dots, b_n(e_\alpha) ).
$$
This vector $v_\alpha$ contains all of the data of the root $e_\alpha$. Indeed, consider the sphere $S_\alpha$ in $\mathbb{S}^n$ corresponding to the ideal boundary of the hyperplane $H_\alpha$ which is $f-$orthogonal to $e_\alpha$; the ``bend'' (inverse radius) of the sphere is $b(e_\alpha)$ and its co-bend is $\hat b(e_\alpha)$. Moreover, its center is $(b_1(e_\alpha),\dots,b_n(e_\alpha))/b(e_\alpha).$
If the bend $b(e_\alpha)=0$, that is, $S_\alpha$ is a hyperplane, then this ``center'' is re-interpreted to mean the limit as spheres converge to this hyperplane; in that case, this ratio becomes the (Euclidean) unit normal to this hyperplane in $\partial \mathbb{H}^{n+1}$.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.2]{Fig2.jpg}
\caption{A facet $F_\alpha$ of a polyhedron $P$ has supporting hyperplane $H_\alpha$.}
\label{fig:struct-th}
\end{figure}
\begin{theorem}[Structure Theorem \cite{KN, K^2}] \label{theorem:structure-kleinian} \quad
\begin{itemize}
\item[\bf (1)] Let $\mathcal{P}$ be a Kleinian packing of $\mathbb{S}^n \simeq \partial \mathbb{H}^{n+1}$. Then its supergroup $\widetilde{\Gamma}$ is a lattice in $\mathrm{Isom}(\mathbb{H}^{n+1})$.
\item[\bf (2)] Suppose that $\widetilde{\Gamma} < \mathrm{Isom}(\mathbb{H}^{n+1})$ is a lattice (containing at least one reflection) with a convex fundamental polyhedron $D$ and a minimal generating set $\widetilde{S}$ for $\widetilde{\Gamma}$, where elements of $\widetilde{S}$ correspond to face-pairing transformations of $D$. Let $R \subset \widetilde{S}$ be a non-empty set of generators corresponding to reflections $\gamma_\alpha$ in certain facets $F_\alpha$, $\alpha \in A$, of $D$. For each $\alpha \in A$ let $S_\alpha = \partial_\infty H_\alpha \subset \partial \mathbb{H}^{n+1} = \mathbb{S}^n$ be the invariant sphere for the reflection $\gamma_\alpha$ with respect to the supporting hyperplane $H_\alpha$ of the facet $F_\alpha$ (see Fig.~\ref{fig:struct-th}). Assume that all the hyperplanes $H_\alpha$ are pairwise disjoint or parallel to each other, and that, if they meet other supporting hyperplanes of facets of $D$, then they do so orthogonally. Let $S = \widetilde{S}\setminus R$ and let $\Gamma = \langle S \rangle$.
Then the orbit of $\{S_\alpha \mid \alpha \in A\}$ under $\Gamma$ is a Kleinian packing of $\mathbb{S}^n$ with symmetry group $\Gamma$ and supergroup $\widetilde{\Gamma}$.
\end{itemize}
\end{theorem}
We remark that there is an even simpler description of the above construction in the case of crystallographic sphere packings. Suppose as above that $\widetilde\Gamma$ is generated by a set $\widetilde S$ of reflections, with Coxeter--Vinberg diagram with vertices $\widetilde S$. Suppose $\widetilde S $ can be partitioned into a ``cluster'' $R$ and ``cocluster'' $S=\widetilde S\setminus R$, with the following properties: (i) Any pair of vertices of $R$ are either connected by a dashed edge or a bold edge, and (ii) any vertex of $R$ and a vertex of $S$ are either connected as above, or disconnected (that is, orthogonal).
Then, as in the Structure Theorem, the orbit of the spheres in $R$ under the action of the group $\Gamma $ generated by inversions in $S$ is a crystallographic packing. (And moreover, every such arises in this way.)
\subsection{Vinberg's arithmeticity criterion}
In order to understand if a given lattice generated by reflections is arithmetic or not, we will use a handy criterion developed by Vinberg \cite{Vin67}. Let $H_e^- = \{x \in \mathbb{R}^{n,1} \mid (x,e) \le 0\}$ be the half-space associated to a hyperplane $H_e$ with normal $e \neq 0$. If $P$ is a Coxeter polyhedron in $\mathbb{H}^n$, such that
$$P = \bigcap_{j=1}^N H_{e_j}^-, $$
then let $G(P) = \{g_{ij}\}^N_{i,j=1} = \{(e_i, e_j)\}^N_{i,j=1}$ denote its Gram matrix, and let the field $K = \mathbb{Q}(\{g_{ij}\}^N_{i,j=0})$ be generated by its entries.
The set of all cyclic products of entries of a matrix $A$ (i.e. the set consisting of all possible products of the form $a_{i_1 i_2} a_{i_2 i_3} \ldots a_{i_k i_1}$) will be denoted by $\mathrm{Cyc}(A)$. Let $k = \mathbb{Q}(\mathrm{Cyc}(G(P))) \subset K$, and let $\mathcal{O}_k$ be the ring of integers of $k$.
Given the lattice $\Gamma$ generated by reflections in the facets of a Coxeter polyhedron $P$, we can determine if $\Gamma$ is arithmetic, quasi--arithmetic, or neither.
\begin{theorem}[Vinberg's arithmeticity criterion \cite{Vin67}]\label{V}
Let $\Gamma$ be a reflection group acting on $\mathbb{H}^n$ with finite volume fundamental Coxeter polyhedron $P$. Then $\Gamma$ is arithmetic if and only if
the following conditions hold:
\begin{itemize}
\item[{\bf(1)}] $K$ is a totally real algebraic number field;
\item[{\bf(2)}] for any embedding $\sigma \colon K \to \mathbb{R}$, such that $\sigma \mid_{k} \ne \mathrm{id}$, $G^\sigma(P)$ is positive semi-definite;
\item[{\bf(3)}] $\mathrm{Cyc}(2 \cdot G(P)) \subset \mathcal{O}_{k}$.
\end{itemize}
The group $\Gamma$ is quasi--arithmetic if and only if it satisfies conditions \textbf{(1)}--\textbf{(2)}, but not necessarily \textbf{(3)}.
\end{theorem}
Whenever $\Gamma$ is quasi--arithmetic or arithmetic, $k$ is called its \textit{field of definition} (or \textit{ground field}). The above criterion simplifies greatly if $k = \mathbb{Q}$. Then the lattice $\Gamma$ will always be quasi--arithmetic, and it will be arithmetic if and only if the cyclic products in Theorem~\ref{V}~(3) are integers.
\subsection{Vinberg's algorithm} In addition to his study of arithmeticity of reflective lattices, Vinberg also produced an algorithm that produces the set of generating reflections for a reflective quadratic from. Namely, if $f$ is a quadratic form of signature $(n, 1)$ over a totally real number field $k$ with the ring of integers $\mathcal{O}_k$.
Moreover, assume that $f$ is admissible, that is for each non--trivial Galois embedding $\sigma: k \rightarrow \mathbb{R}$ the form $f^\sigma$ (obtained by applying $\sigma$ to the coefficients of $f$) is positive definite. Then the group $\mathbf{O}_f(\mathcal{O}_k)$ of integer points of the orthogonal group $\mathbf{O}_f(\mathbb{R})$ is a lattice \cite[II, Chapter 6]{AVS93}.
The form $f$ as above is called reflective if $\mathbf{O}_f(\mathcal{O}_k)$ is generated by a finite number of reflections, up to a finite index. The algorithm devised by Vinberg in \cite{Vin72} produces a set of generating reflections for any reflective quadratic form. We shall refer to it as \textit{the Vinberg algorithm}. Several computer realizations of the algorithm are available \cite{Gug, BP, Bot}.
In contrast, if the form $f$ is not reflective, the algorithm never halts. The later work by Bugaenko \cite{Bug} also provided a way to certify that a given Lorentzian form $f$ is non--reflective. This ``method of infinite symmetry'' has been also implemented in \cite{Gug, BP, Bot} to various extent.
\section{Proofs}
\subsection{Proof of Lemma~\ref{lemma:doubling}}
Let $P$ be a finite volume Coxeter polyhedron in $\mathbb{H}^n$ and $P_0$ be a facet that meets other facets of $P$ at angles either $\pi/4$ or $\pi/2$. Let $P_1, \ldots P_k$ be all the facets of $P$ that meet $P_0$ at $\pi/4$. Assume that each $P_j$ meets its neighbors at ``even'' angles of the form $\pi/2^m$, with $m \geq 1$, an integer.
If $P$ is a polyhedron as above, and $F$ is a facet of $P$, let $R_F$ denote the reflection in the hyperplane of $F$.
Then the following inductive steps will produce a polyhedron with an orthogonal facet.
1. The double $P^{(1)}$ of $P$ along the facet $F_1 = P_1$ has the facet $R_{F_1} (P_0)$ that forms $(k-1)$ angles $\pi/4$ with its neighbours. Observe that $P^{(1)}$ is again a Coxeter polyhedron, since doubling of even Coxeter angles produces even Coxeter angles again or, if we double a right angle, it produces a new facet.
2. The facet $F_2 = R_{F_1} (P_2)$ of $P^{(1)}$ forms even Coxeter angles with its neighbours. Thus the double $P^{(2)}$ of $P^{(1)}$ along $F_2$ has the facet $R_{F_2} R_{F_1}(P_0)$ that forms $(k-2)$ angles $\pi/4$ with its neighbours. All the conclusions of the previous step apply respectively to $P^{(2)}$.
3. Continuing in this way, we obtain a polyhedron $P^{(k)}$ from $P^{(k-1)}$ by doubling along $F_k = R_{F_{k-1}} (P_k)$, and the facet
$R_{F_k}\cdots R_{F_1}(P_0)$ of $P^{(k)}$ will be orthogonal to all of its neighbours. Such a polyhedron produces a crystallographic sphere packing by the Structure Theorem (Theorem~\ref{theorem:structure-kleinian}).
\subsection{Proof of Theorem~\ref{theorem:dimension-bound}}
The proof of Theorem~\ref{theorem:dimension-bound} follows from the results of Esselmann \cite{E} and the following simple lemma which is well known to experts in hyperbolic reflection groups (and also follows from more general results of \cite{BK21}).
\begin{lemma}\label{lemma:stabilizer}
Let $\Gamma<\mathrm{Isom}(\mathbb{H}^{n+1})$ be the symmetry group of a crystallographic packing $\mathscr{P}$ of $\mathbb{S}^n$, and let $\Gamma_0$ be the stabilizer of a sphere in $\mathscr{P}$. Then $\Gamma_0$ is a reflective lattice acting on $\mathbb{H}^n$. If $\Gamma$ is arithmetic, then $\Gamma_0$ is also arithmetic
\end{lemma}
\begin{proof}
Indeed, the sphere corresponds to a facet $P_0$ of the Coxeter polyhedron $P$ associated to the packing $\mathscr{P}$, and $P_0$ is orthogonal to all the adjacent facets. Then $\Gamma_0$ coincides with the finite covolume hyperbolic group generated by reflections in the facets of $P_0$. The orthogonality of $P_0$ to its neighbours together with Vinberg's criterion (Theorem~\ref{V}) implies that $\Gamma$ being arithmetic makes $\Gamma_0$ arithmetic too.
\end{proof}
\medskip
It is easy to see that Lemma~\ref{lemma:stabilizer} implies Proposition~\ref{prop:lower-dimension} by applying it to the setting of superintegral crystallographic sphere packings.
In \cite{E}, Esselmann showed that Borcherds' $21$--dimensional polyhedron from \cite{Bor} is a unique, up to commensurablity, arithmetic non--compact Coxeter polyhedron in $\mathbb{H}^{21}$ and that there are no arithmetic non--compact Coxeter polyhedra in $\mathbb{H}^{20}$.
If we would have a superintegral sphere packing of $\mathbb{S}^{20}$, then it would produce a superintegral crystallographic packing of $\mathbb{S}^{19}$, which contradicts the above mentioned results by Esselmann.
\subsection{Proof of Theorem~\ref{theorem:crystallographic}} Our higher--dimensional examples of superintegral sphere packings come from the Borcherds' $21$--di\-men\-sio\-nal polyhedron $\mathcal{P}_{21}$ \cite{Bor}. By describing its lower--dimensional faces we can fill in some of the dimensions. For example, in $\mathbb{H}^{19}$ we have a face $\mathcal{P}_{19}$ (first found by Kaplinskaya and Vinberg in \cite{KV}) that satisfies the conditions of Theorem~\ref{lemma:doubling}.
There are two faces $\mathcal{P}^{(i)}_{18}$, $i=1,2$, each with an orthogonal facet, in $\mathbb{H}^{18}$. Here, $\mathcal{P}^{(1)}_{18}$ denotes the polyhedron described by Kaplinskaya and Vinberg in \cite{KV}. The other polyhedron $\mathcal{P}^{(2)}_{18}$ was found by Allcock in \cite{A}.
Be descending to lower--dimensional Coxeter faces (sometimes we have to pass via non-Coxeter faces of lower codimension in order to find their faces of higher codimension that turn out to be Coxeter), we obtain another face $\mathcal{P}_{15}$ satisfying Lemma~\ref{lemma:doubling} in $\mathbb{H}^{15}$.
The polyhedron $\mathcal{P}_{15}$ has $29$ facets and generates an arithmetic reflection group, as follows from Vinberg's arithmeticity criterion (Theorem~\ref{V}). Indeed, conditions (1) and (2) are vacuously satisfied. The only condition that needs to be checked is (3). As could be easily observed, if $G = (g_{ij})^{29}_{i,j=1}$ is the Gram matrix of $\mathcal{P}_{15}$, then each $g_{ij}$ satisfies $4\,g^2_{ij} \in \mathbb{Z}$, and thus either $\mathrm{Cyc}(G)$ are all integer, or contain a quadratic irrationality. In order to exclude the latter, it is enough to check that $g_{i_1 i_2} \ldots g_{i_{k-1} i_{k}} g_{i_k i_1} \in \mathbb{Z}$, for $i_1 i_2 \ldots i_k i_1$ being edge cycles in the Coxeter diagram of $\mathcal{P}_{15}$ belonging to a $\mathbb{Z}_2$--cycle basis of the diagram (viewed as a simple undirected graph). Indeed, if a cycle product $p$ corresponds to a cycle $c$ that is $\mathbb{Z}_2$--sum of $c_1$ and $c_2$ (with the corresponding cycle products $p_1$ and $p_2$) then $p_1 \cdot p_2 = p \cdot q^2$, where $q$ corresponds to the overlap of $c_1$ and $c_2$ (i.e. $q$ is a trivial back-and-forth cycle unless $c_1 = c_2$ as sets, which is a trivial situation). As mentioned above $q^2 \in \mathbb{Z}$, and thus $p$ cannot be a quadratic irrationality if $p_1, p_2 \in \mathbb{Z}$. This argument easily generalizes to an arbitrary $\mathbb{Z}_2$--sum of cycles.
However, there are no suitable Coxeter faces of $\mathcal{P}_{21}$ in $\mathbb{H}^n$ for $n = 16$ and $17$.
The remaining packings in $\mathbb{S}^n$, $2 \leq n \leq 13$, can be easily obtained from Vinberg's examples \cite{Vin67} for the series of quadratic forms~$f_n = -2 x^2_0 + x^2_1 + \ldots + x^2_n$, for $3 \leq n \leq 14$, respectively. Each of the polyhedra in \cite[Table 7]{Vin67} has a facet that is orthogonal to its neighbours\footnote{For packings in $\mathbb{S}^n$, $2 \leq n \leq 7$, one can also use the totally right--angled polyhedra from \cite{PV}.}.
All the computations mentioned above (faces of Borcherds' $\mathcal{P}_{21}$ and their properties of being Coxeter and/or having ``even'' angles) are accessible on Github \cite{github1} as a SageMath \cite{sagemath} worksheet.
\subsection{Proof of Theorem~\ref{theorem:arithmetic-super}} In \cite{SW} Sharlau and Walhorn provide the list of maximal arithmetic non--uniform reflective lattices (up to conjugation) acting on $\mathbb{H}^{n+1}$ for $n = 2, 3$. In \cite{Turkalj} Turkalj performs an analogous classification in $\mathbb{H}^{n+1}$ for $n = 4$.
It is worth mentioning that in the list of lattices acting on $\mathbb{H}^4$ has a few entries where the $\mathbb{H}$ direct summand apparently has to be replaced by ${}^2 \mathbb{H}$. Such entries are easily identifiable by running the Vinberg algorithm (the number of roots indicated in the respective tables seems to be correct). The lattices described by Turkalj require more work for the information to be extracted from the list provided in \cite{Turkalj}. We used instead an equivalent list provided by Scharlau and Kirschmer \cite{SK}.
In each case we reduce each list up to rational isometry between the respective lattices. Then the proof amounts to a lengthy computation carried our by using the Julia \cite{julia} implementation of the Vinberg algorithm due to Bottinelli \cite{Bot}.
For each lattice, we consider its reflection subgroup: if the corresponding polyhedron has an orthogonal facet (or, in term of Vinberg's work \cite{Vin15}, the lattice has an \textit{isolated root}), then we are done. Otherwise, we choose another lattice, up to an appropriate rational isometry, and use it instead: we get a commensurable lattice having an isolated root. The necessary rational isometry is usually guessed by trial and error: reflectivity is checked by using the Vinberg algorithm.
All the data necessary to verify our computation \cite{github1} and Bottinelli's implementation of the Vinberg algorithm \cite{github2} are available on Github.
\subsection{A remark on reflectivity}
Here we consider another question in the context of properly integral sphere packings.
In the proof of the Arithmeticity Theorem (see \cite[Theorem 19]{KN}), given an integral orbit of spheres $\mathcal O = S_0 \cdot \Gamma$, and a choice of $n+2$ linearly independent normals to spheres in the orbit, one constructs an arithmetic lattice $\Omega$ which has (a conjugate of) $\Gamma$ as a subgroup; therefore $\Gamma$ is sub--arithmetic, in the terminology of \cite[p. 3]{K^2}.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1.5]
\coordinate (one) at (0, 0);
\coordinate (two) at (1,1);
\coordinate (three) at (1,-1);
\coordinate (four) at (2,0);
\coordinate (five) at (3,0);
\coordinate (midpt) at (4,0);
\coordinate (six) at (5,0);
\draw[dashed] (five) -- (six) ;
\draw[thick] (five) -- (four);
\draw[thick, double distance = 2pt] (four) -- (three);
\draw[thick, double distance = 2pt] (one) -- (two);
\draw[thick] (one) -- (three);
\draw[thick] (two) -- (four);
\filldraw (midpt) circle (0pt) node[above] { $\sqrt{\nicefrac{8\, }7}$};
\filldraw (one) circle (2pt) node[left] {\scriptsize 1};
\filldraw (two) circle (2pt) node[above] {\scriptsize 2};
\filldraw (three) circle (2pt) node[below] {\scriptsize 3} ;
\filldraw (four) circle (2pt) node[above] {\scriptsize 4};
\filldraw (five) circle (2pt) node[above] {\scriptsize 5};
\filldraw (six) circle (2pt) node[above] {\scriptsize 6};
\end{tikzpicture}
\caption{The Coxeter--Vinberg diagram of the prism $P$.}
\label{fig:P}
\end{figure}
One may wonder, since $\Gamma$ is generated by reflections, whether $\Omega$ should itself be reflective. Here we record an example which answers this question in the negative.
Let $P$ be the $4$--dimensional hyperbolic prism described by the Coxeter-Vinberg diagram in Figure \ref{fig:P}.
The facet normals of $P$ are
\begin{align*}
&e_1 = \left(-\frac{\sqrt{2}}{8}, \sqrt{2}, 0, \frac{1}{2}, \frac{\sqrt{2}}{2}\right),\;
e_2 = (0, 0, 0, 0, -1),\\
&e_3 = (0, 0, 0, -1, 0),\;
e_4 = \left(1, 0, \frac{1}{2}, \frac{\sqrt{2}}{2}, \frac{1}{2}\right),\\
&e_5 = (0, 0, -1, 0, 0),\;
e_6=\left(\frac{\sqrt{14}}{28}, \frac{2 \sqrt{14}}{7}, \frac{2 \sqrt{14}}{7}, 0, 0\right),
\end{align*}
where the isolated vector is $e_6$. By applying Vinberg's criterion (Theorem~\ref{V}) to the Gram matrix of $P$
\begin{equation*}
G(P) =
\left(
\begin{array}{cccccc}
-1 & \frac{1}{\sqrt{2}} & \frac{1}{2} & 0 & 0 & 0 \\
\frac{1}{\sqrt{2}} & -1 & 0 & \frac{1}{2} & 0 & 0 \\
\frac{1}{2} & 0 & -1 & \frac{1}{\sqrt{2}} & 0 & 0 \\
0 & \frac{1}{2} & \frac{1}{\sqrt{2}} & -1 & \frac{1}{2} & 0 \\
0 & 0 & 0 & \frac{1}{2} & -1 & \sqrt{\frac{8}{7}} \\
0 & 0 & 0 & 0 & \sqrt{\frac{8}{7}} & -1 \\
\end{array}
\right)
\end{equation*}
we can conclude that $P$ generates a quasi--arithmetic but not arithmetic reflection group. Hence the sphere packing corresponding to $e_6$ cannot be superintegral.
The orbit of $e_6$ under the group $\Gamma$ generated by reflections through the normals $e_1,\dots,e_5$ is a crystallographic sphere packing (see Figure \ref{fig:pack}), which we claim is properly integral.
\begin{figure}
\centering
\includegraphics[width=.6\textwidth]{sphPack.pdf}
\caption{Some of the spheres of the packing $\mathcal{P}$ shown in blue, and the walls of the generators $e_1,\dots,e_5$ of $\Gamma$ shown in red.}
\label{fig:pack}
\end{figure}
To see this, we take for a list of $n+2=5$ linearly independent vectors the choice $V = \left( e_6, e_6 R_5, e_6 R_5 R_4, e_6 R_5 R_4 R_2, e_6 R_5 R_4 R_3 \right)$ as described in \cite[Equation (2)]{KN}.
These are all in the packing $\mathcal{P} = e_6\cdot\Gamma$.
After rescaling by $\sqrt {8/7}$, these spheres all have bend $1$.
Then the generators $R_1,\dots,R_5$, conjugated by $V$, become the matrices:
$$
VR_1V^{-1}=
\left(
\begin{array}{ccccc}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
9/4 & 9/4 & 3/2 & -1/4 & \
-3/4 \\
9/4 & 9/4 & 3/2 & -5/4 & 1/4 \
\\
\end{array}
\right),
\
VR_2V^{-1}=
\left(
\begin{array}{ccccc}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & -1 & 1 & 1 \\
\end{array}
\right),
$$
$$
VR_3V^{-1}=
\left(
\begin{array}{ccccc}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 \\
0 & 0 & -1 & 1 & 1 \\
0 & 0 & 1 & 0 & 0 \\
\end{array}
\right),\
VR_4V^{-1}=\left(
\begin{array}{ccccc}
1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & -1 & 1 & 0 & 1 \\
\end{array}
\right),$$
and
$$
VR_5V^{-1}= \left(
\begin{array}{ccccc}
0 & 1 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 1 \\
\end{array}
\right).
$$
These act on the {\it left} on the set of bends of the packing; the bends which arise in the packing are those in the action of the above matrices on the initial bends vector $(1,1,1,1,1)$ of all bends equal to $1$.
All but the first matrix are themselves integral, and hence evidently preserve integrality. Moreover, inspection shows that they preserve the congruence $b_1+b_2+2b_3+3b_4+b_5\equiv0\pmod 4$, which is satisfied by the initial bends vector $(1,1,1,1,1)$.
This congruence ensures that the action of the first matrix $VR_1 V^{-1}$ always gives integer bends in the orbit, thus proving the proper integrality of the packing.
Finally, following the proof of the Arithmeticity Theorem, we exhibit an arithmetic lattice $\Omega$ containing a conjugate of $\Gamma$. We find in this case that $\Omega$ is {\it not} reflective.
To begin, we record that the Gram matrix of $V$:
$$
G = G(V) = VQV^T=-\frac{1}{7}
\left(
\begin{array}{ccccc}
-7 & 9 & 9 & 9 & 9 \\
9 & -7 & 9 & 9 & 9 \\
9 & 9 & -7 & 9 & 25 \\
9 & 9 & 9 & -7 & 41 \\
9 & 9 & 25 & 41 & -7 \\
\end{array}
\right),
$$
where $Q = \mathbb H \perp \langle 1,1,1\rangle$.
The inverse Gram matrix equals
$$
G^{-1}=
-\frac{7}{256}
\left(
\begin{array}{ccccc}
25 & 9 & 0 & -9 & -9 \\
9 & 25 & 0 & -9 & -9 \\
0 & 0 & 12 & -8 & -4 \\
-9 & -9 & -8 & 9 & 1 \\
-9 & -9 & -4 & 1 & 5 \\
\end{array}
\right),
$$
that gives rise to the Lorentzian quadratic form $f(x) = x^T G^{-1} x$.
The proof of the Arithmeticity Theorem \cite{KN} shows that $V\Gamma V^{-1} < \Omega := \mathbf{O}_f(\mathbb{Z}).$
Finally, Vinberg's algorithm applied to $\Omega$ produces the following first $31$ roots:
\begin{align*}
&e_1 = (1, -1, 0, 0, 0),\ e_2 = (0, 1, 2, 1, 0),\ e_3 = (0, 1, 0, 1, 0), \\
&e_4 = (-1, 0, -1, -2, 1),\ e_5 = (0, 1, 2, 3, 0),\ e_6 = (2, 4, 9, 15, -1), \\
&e_7 = (1, 2, 8, 11, 0),\ e_8 = (1, 3, 6, 12, 0),\ e_9 = (5, 6, 20, 29, -2), \\
&e_{10} = (8, 8, 21, 35, -3),\ e_{11} = (7, 10, 23, 36, -3),\ e_{12} = (4, 4, 15, 25, -1), \\
&e_{13} = (6, 7, 16, 31, -2),\ e_{14} = (3, 3, 7, 13, -1),\ e_{15} = (5, 6, 13, 28, -1), \\
&e_{16} = (5, 6, 18, 29, -2),\ e_{17} = (8, 8, 23, 35, -3),\ e_{18} = (2, 3, 11, 20, 1), \\
&e_{19} = (4, 4, 17, 25, -1),\ e_{20} = (1, 1, 2, 5, 0),\ e_{21} = (3, 6, 19, 26, -1), \\
\end{align*}
\begin{align*}
&e_{22} = (3, 6, 15, 26, -1),\ e_{23} = (4, 4, 9, 23, 1),\ e_{24} = (3, 3, 14, 22, 0), \\
&e_{25} = (3, 3, 11, 17, -1),\ e_{26} = (3, 4, 12, 23, 0),\ e_{27} = (2, 5, 8, 21, 2), \\
&e_{28} = (3, 3, 7, 12, -1),\ e_{29} = (11, 12, 34, 59, -4),\ e_{30} = (10, 11, 32, 49, -4), \\
&e_{31} = (8, 11, 35, 54, -3
\end{align*}
There is an infinite order symmetry $\sigma$ mapping the elliptic subdiagram spanned by the roots $\{ e_{13}, e_3, e_{14}, e_1, e_6 \}$ to another elliptic subdiagram spanned by the roots $\{ e_{12}, e_{19}, e_{25}, e_{31}, e_7 \}$, exactly in this order
\begin{equation*}
e_{13} \mapsto e_{12}, e_3 \mapsto e_{19}, e_{14} \mapsto e_{25}, e_1 \mapsto e_{31}, e_6 \mapsto e_7.
\end{equation*}
The matrix from $\mathbf{O}_f(\mathbb{Z})$ representing $\sigma$ in the respective bases is
\begin{equation*}
S = \left(
\begin{array}{ccccc}
54 & 41 & -3 & -10 & 79 \\
64 & 48 & -3 & -12 & 96 \\
159 & 122 & -8 & -30 & 238 \\
276 & 210 & -15 & -51 & 413 \\
-21 & -16 & 1 & 4 & -31
\end{array}
\right).
\end{equation*}
Thus, the arithmetic group $\Omega$ containing $\Gamma$ is \textit{not} reflective according to Vinberg's criterion (Theorem \ref{V}).
However, we would like to remark that the group $\Omega$ is defined up to commensurability (because the choice of normals described above), and reflectivity is not commensurability invariant. Indeed, the lattice
\begin{equation*}
\left(
\begin{array}{ccccc}
0 & 0 & 49 \\
0 & 49 & 7 \\
49 & 7 & 3
\end{array}
\right)
\end{equation*}
has \textit{no} roots \cite{BK22, Col}, while it is easily seen to be commensurable to the diagonal lattice
\begin{equation*}
\left(
\begin{array}{ccccc}
-1 & 0 & 0 \\
0 & 2 & 0 \\
0 & 0 & 1
\end{array}
\right),
\end{equation*}
that is reflective and corresponds to the $(2,4,\infty)$ hyperbolic triangle reflection group.
| be040daeafe90b907636e8a1174f419a81584cc5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Quantum antiferromagnetism has been a topic of intense study for many decades, and has led to many new insights into quantum many-body phenomenon. In particular, a new class of quantum phases, known as \textit{quantum spin liquids} (QSLs) \cite{Anderson1973, Zhou2017, Savary2016}, are predicted to emerge in certain parameter regimes of antiferromagnets due to the combination of geometric frustration and quantum fluctuations. A generic feature of these QSLs is the existence of fractionalized excitations, which cannot be created individually by local operators.
A particularly well-studied antiferromagnetic model is the $J_1$-$J_2$ model on the square lattice~\cite{IoffeLarkin88,GSH89,Dagotto89,CCL90,RS91,SR91}, which has nearest-neighbor and next-nearest-neighbor antiferromagnetic exchange interactions with coefficients $J_1$ and $J_2$, respectively. It is known that the ground state of the nearest-neighbor square lattice antiferromagnet ($J_2 = 0$) has long-range N\'eel order {\it i.e.\/} global $\operatorname{SU}(2)$ spin rotation symmetry is broken with the spin expectation value $\langle {\bm S}_{{\boldsymbol{i}}} \rangle = \eta_{\boldsymbol{i}} {\bm N}_0$ , where ${\bm S}_{\boldsymbol{i}}$ is the spin operator on site ${\boldsymbol{i}}$, $\eta_{\boldsymbol{i}} = \pm 1$ on the two checkerboard sublattices, and ${\bm N}_0$ is the antiferromagnetic moment. The next-nearest-neighbor antiferromagnetic interactions compete against this N{\'e}el order, and the nature of this model as a function of $J_2 / J_1$, in particular in the regime of high frustration, $J_2 / J_1 \approx 0.5$, remain a key open question.
An early proposal \cite{NRSS89,NRSS90,RS91,SR91} was that there was a direct transition from the N\'eel state to a valence bond solid (VBS) (see Fig.~\ref{fig:becca}) which restores spin rotation symmetry but breaks lattice symmetries. This is followed by a first order transition at larger $J_2/J_1$ to a `columnar' state which breaks spin rotation symmetry, and which we do not address in the present work. This led to the development of a theory of `deconfined criticality'~\cite{OMAV04,senthil1,senthil2} which allowed for a direct transition between two symmetry-broken phases, a phenomenon disallowed by conventional Landau-Ginzberg theory. Numerical evidence has since accumulated for the presence of a VBS phase in the $J_1$-$J_2$ model - in particular, it was recently shown that a non-zero antiferromagnetic third-nearest-neighbor interaction $J_3$ stabilizes a clear VBS phase in a large parameter range of $J_2/J_1$~\cite{Liu2021}, and this phase is argued to be stable down to $J_3=0$. The nature of the N\'eel-VBS transition in this model has remained a question of significant debate. However, in the past two years, a consensus appears to have emerged \cite{BeccaKITP} among groups investigating this question by different numerical methods \cite{Sandvik18,Becca20,Imada20,Gu20}, and is summarized in Fig.~\ref{fig:becca}: there is a narrow window with a gapless spin liquid phase between the N\'eel and VBS states. This gapless phase has been identified \cite{Becca01,SenthilIvanov,Becca13,Becca18,Becca20} as a $\mathbb{Z}_2$ spin liquid \cite{RS91,SR91,wen1991,Kitaev1997} with gapless, fermionic, $S=1/2$ spinon excitations with a Dirac-like dispersion \cite{TSMPAF99,WenPSG,SenthilIvanov,SenthilLee05,Kitaev2006}. Although there is less of a consensus over the precise nature of this spin liquid, variational wavefunction studies using a Gutzwiller projection~\cite{Becca18, Becca20} identify the ground state as corresponding to the spin liquid Z2A$zz13$ in Wen's classification \cite{WenPSG}.
In a recent work with A. Thomson~\cite{Shackleton21}, we proposed a unified theory that contains multiple instabilities of the gapless $\mathbb{Z}_2$ spin liquid Z2A$zz13$, which we conjecture correspond to both the neighboring VBS and N\'eel orders shown in Fig.~\ref{fig:becca}. The mechanism for these instabilities consists of considering the $\mathbb{Z}_2$ spin liquid as a condensed phase of a parent $\operatorname{SU}(2)$ gauge theory coupled to Higgs bosons. This $\operatorname{SU}(2)$ theory, identified as the $\pi$-flux spin liquid, also has a proximate phase with $\operatorname{U}(1)$ gauge symmetry, known as the staggered flux or Dirac spin liquid. Both these phases are conjectured to be unstable on the square lattice and ultimately lead to ordered phases, which gives us a route for explaining the N\'eel and VBS ordered phases predicted to exist alongside the $\mathbb{Z}_2$ spin liquid in the $J_1$-$J_2$ model. In our previous work \cite{Shackleton21}, we studied the transition between the $\operatorname{SU}(2)$ and $\mathbb{Z}_2$ gauge theory using field-theoretic techniques; in this work, we complete our study by an analysis of the $\operatorname{U}(1)$ to $\mathbb{Z}_2$ transition.
The contents of this work are summarized as follows. In Section~\ref{sec:gaplessz2}, we review the motivation behind our continuum model and explain its connection to the microscopic lattice theory. In Section~\ref{sec:largeN}, we derive a large-$N_f$ effective theory for the $\operatorname{U}(1)$ to $\mathbb{Z}_2$ transition, with $N_f$ the number of fermion flavors. This allows us to study the critical theory in a systematic ${1}/{N_f}$ expansion. A renormalization-group analysis is performed in Section~\ref{sec:rg}, where we extract critical exponents of the theory to leading order in ${1}/{N_f}$.
\begin{figure}
\centering
\includegraphics[width=5in]{becca.pdf}
\caption{Phases of the $S=1/2$ $J_1$-$J_2$ antiferromagnet on the square lattice, from the numerical results of Refs.~\cite{Sandvik18,Becca20,Imada20,Gu20}, all of which agree that the spin liquid is gapless. Each ellipse in the valence bond solid (VBS) represents a singlet pair of electrons. Lower part of figure adapted from Ref.~\cite{BeccaKITP}.}
\label{fig:becca}
\end{figure}
The structure of our critical theory bears resemblence to prior studies of transitions between Dirac spin liquids and a gapped $\mathbb{Z}_2$ gauge theory~\cite{Rufus2018, Dupuis2021}. The difference between this theory and ours is reflected in different forms of Yukawa couplings between the fermions and the Higgs fields. This difference turns out to drastically change the qualitative features of the critical theory. The most notable difference is the lack of Lorentz invariance in our theory. It is known~\cite{Hermele05} that the symmetries of the square lattice permit a velocity anisotropy term in the fermion action. In the absence of additional gapless degrees of freedom, it has been shown~\cite{Franz2002, Vafek2002, Hermele05} that this anisotropy is irrelevant in a $1 / N_f$ expansion. The Yukawa couplings of previously-studied transitions preserve Lorentz invariance. However, we will show that the choice of Yukawa coupling necessary to realize our specific $\mathbb{Z}_2$ spin liquid of interest will lead to Lorentz symmetry breaking and a dynamical critical exponent $z = 1 + 0.225 / N_f + \order{1 / N_f^2}$, where $N_f = 1$ is the physical case.
\section{Summary of prior work and derivation of continuum theory}
\label{sec:gaplessz2}
This paper is a continuation of prior work~\cite{Shackleton21}, which describes multiple possible instabilities of gapless $\mathbb{Z}_2$ spin liquid through a parent $\pi$-flux phase coupled to various Higgs fields. We present a brief summary of this derivation - further details may be found in Ref.~\cite{Shackleton21}.
Our starting point is the fermionic spinon theory of spin liquids, which is derived by re-epressing the spin operators in terms of spinons $f_{{\boldsymbol{i}} \alpha}$, $\alpha = \uparrow\,, \downarrow$ at site ${\boldsymbol{i}} = (i_x\,, i_y)$ of the square lattice using
\begin{equation}
\begin{aligned}
\vb{S}_{{\boldsymbol{i}}} &= \frac{1}{2} \sum_{\alpha\,, \beta} f_{{\boldsymbol{i}} \alpha}^\dagger \vb{\sigma}_{\alpha \beta} f_{{\boldsymbol{i}} \beta} \\
\end{aligned}
\end{equation}
along with the constraints
\begin{equation}
\begin{aligned}
f_{{\boldsymbol{i}} \alpha}^\dagger f_{{\boldsymbol{i}} \alpha } = 1 \,, \quad f_{{\boldsymbol{i}} \alpha} f_{{\boldsymbol{i}} \beta} \epsilon_{\alpha \beta } = 0\,.
\label{eq:halfFillingConstraint}
\end{aligned}
\end{equation}
Introducing the Nambu spinor,
\begin{equation}
\begin{aligned}
\psi_{{\boldsymbol{i}}} &= \begin{pmatrix} f_{{\boldsymbol{i}} \uparrow} \\ f_{{\boldsymbol{i}} \downarrow}^\dagger \end{pmatrix} \,, \\
\end{aligned}
\end{equation}
and Pauli matrices $\tau^\ell$ which act on spinor indices, we can write a mean-field ansatz for our Hamiltonian
\begin{equation}
\begin{aligned}
H &= - \sum_{{\boldsymbol{i}} {\boldsymbol{j}}} \psi_{{\boldsymbol{i}}}^\dagger u_{{\boldsymbol{i}} {\boldsymbol{j}}} \psi_{{\boldsymbol{j}}}\,, \\
\end{aligned}
\end{equation}
where the hoppings $u_{{\boldsymbol{i}} {\boldsymbol{j}}}$ must be determined self-consistently. The additional degrees of freedom in this representation is reflected by an $\operatorname{SU}(2)_g$ gauge symmetry, under which
\begin{equation}
\begin{aligned}
\operatorname{SU}(2)_g : \psi_{{\boldsymbol{i}}} \rightarrow U_{g, {\boldsymbol{i}}} \psi_{{\boldsymbol{i}}} \quad \,, \quad U_{g, {\boldsymbol{i}}} \in \operatorname{SU}(2)
\end{aligned}
\end{equation}
and a corresponding transformation for $u_{{\boldsymbol{i}} {\boldsymbol{j}}}$. Including gauge fluctuations are necessary to enforce the constraint in Eq.~(\ref{eq:halfFillingConstraint}), which in the Nambu spinor variables becomes $\psi_{{\boldsymbol{i}}}^\dagger \tau^\ell \psi_{{\boldsymbol{i}}} = 0$. Different spin liquids may be described by different mean-field ansatzes $u_{{\boldsymbol{i}} {\boldsymbol{j}}}$, which may also spontaneously break the relevant gauge fluctuations from $\operatorname{SU}(2)$ down to $\operatorname{U}(1)$ or $\mathbb{Z}_2$.
The particular spin liquid of relevance to the $J_1$-$J_2$ model can be labeled as Z2A$zz$13 following Wen's classification~\cite{WenPSG}. This spin liquid, along with two relevant proximate spin liquids, can be described by the mean-field ansatz,
\begin{equation}
\begin{aligned}
u_{{\boldsymbol{i}},{\boldsymbol{i}}+\hat{x}} &= \chi \, \tau^x - \eta \, \tau^y \\
u_{{\boldsymbol{i}},{\boldsymbol{i}}+\hat{y}} &= \chi \, \tau^x + \eta \, \tau^y \\
u_{{\boldsymbol{i}},{\boldsymbol{i}}+\hat{x}+ \hat{y}} &= - \gamma \, \tau^x \\
u_{{\boldsymbol{i}},{\boldsymbol{i}}-\hat{x}+ \hat{y}} &= \gamma \, \tau^x
\label{eq:mfAnsatz}
\end{aligned}
\end{equation}
The three relevant spin liquids are:
\begin{itemize}
\item The $\pi$-flux phase with $\operatorname{SU}(2)$ gauge symmetry corresponds to $\chi = \eta \neq 0$ and $\gamma = 0$.
\item The staggered flux phase with $\operatorname{U}(1)$ gauge symmetry corresponds to $\chi\,,\eta \neq 0$, $\gamma = 0$, and $\chi \neq \eta$.
\item The $\mathbb{Z}_2$ spin liquid Z2A$zz$13 is obtained from the staggered flux phase by turning on a non-zero $\gamma$.
\end{itemize}
The dispersion relation of all three phases hosts 4 Dirac cones at low energy, and hence, all three phases are described by $N = 4$ massless Dirac fermions minimally coupled to the corresponding gauge field. The Dirac cones of the $\pi$-flux phase have an emergent Lorentz invariance, whereas the staggered flux and $\mathbb{Z}_2$ phase have anisotropic dispersion relations on a mean-field level (note that prior studies~\cite{Franz2002, Vafek2002, Hermele05} show an emergent Lorentz invariance of the staggered flux phase upon including gauge fluctuations).
The primary mechanism for our theory of the $J_1$-$J_2$ model is the assumption that both the $\pi$-flux and staggered flux phases on the square lattice are ultimately unstable to ordered phases, either N\'eel or VBS. For the staggered flux phase, the instability arises due to the presence of monopoles, which are allowed by the compactness of the $\operatorname{U}(1)$ gauge theory. The scaling dimension of these monopoles have been calculated to second-order in a $1/N$ expansion~\cite{Borokhov2002, Pufu2014}, with $N$ the number of fermions, and are relevant for $N=4$. Moreover, there exists a ``trivial'' monopole operator that respects the microscopic symmetries of the square lattice, and hence is allowed by symmetry in an effective Lagrangian~\cite{Alicea2008}. Proliferation of these monopoles is conjectured to lead to ordered phases, including N\'eel and VBS order~\cite{Song1, Song2}. The $\pi$-flux phase, on the other hand, has been conjectured to be a dual description of the DQCP separating N\'eel and VBS order~\cite{Wang17}, and hence is generically unstable to these phases. A unified framework for describing the gapless spin liquid Z2A$zz$13 as well as these two instabilities can hence be obtained from $\text{QCD}_3$ with $N=2$ fermion doublets and an $\operatorname{SU}(2)$ gauge group, and coupling it to Higgs fields whose condensation breaks the gauge group to either $\operatorname{U}(1)$ or $\mathbb{Z}_2$. The precise manner in which these Higgs fields must couple to the Dirac fermions in order to realize the specific spin liquids of interest can be determined by taking the continuum limit of Eq.~(\ref{eq:mfAnsatz}) and demanding that the condensation of the Higgs fields modifies the Dirac dispersion relation in a manner consistent with the lattice theory. This is the procedure carried out in Ref.~\cite{Shackleton21}. These couplings may also be determined by matching the fractionalization of the square lattice symmetries on the microscopic level with the symmetries of the continuum theory, which is done in Appendix A of Ref.~\cite{Shackleton21}. We refer the reader to Ref.~\cite{Shackleton21} for further details of this calculation; we will simply present the resulting Lagrangian in this work.
This procedure ultimately yields a Lagrangian consisting of four Dirac fermions $\psi$, an $\operatorname{SU}(2)$ gauge field $A_\mu^a$, and three three-component adjoint Higgs fields $\Phi_{1, 2, 3}^a$, $a=x, y, z$. The $\operatorname{SU}(2)$ gauge ($\sigma$) and valley ($\mu$) Pauli matrices both rotate between the four fermion flavors.
\begin{equation}
\begin{aligned}
\mathcal{L} &= \mathcal{L}_{\psi} + \mathcal{L}_{\Phi} + \mathcal{L}_{\Phi \psi}
\\
\mathcal{L}_{\psi} &= i \bar{\psi} \gamma^\mu \left( \partial_\mu - i A^a_\mu \sigma^a \right) \psi\,.
\\
\mathcal{L}_{\Phi} &= \sum_{i=1}^3 D_\mu \Phi_i^a D^\mu \Phi_i^a + V(\Phi)
\\
\mathcal{L}_{\Phi \psi} &= \Phi_1^a \, \bar{\psi} \mu^z \gamma^x \sigma^a \psi + \Phi_2^a \, \bar{\psi} \mu^x \gamma^y \sigma^a \psi + \Phi_3^a \bar{\psi} \mu^y \sigma^a (\gamma^x D_y + \gamma^y D_x) \psi
\\
V(\Phi) &= s \left( \Phi_1^a \Phi_1^a + \Phi_2^a \Phi_2^a \right) + \widetilde{s} \, \Phi_3^a \Phi_3^a + w \, \epsilon_{abc} \, \Phi_{1}^a \Phi_{2}^b \Phi_{3}^c \nonumber \\
&+ u \left( \Phi_1^a \Phi_1^a + \Phi_2^a \Phi_2^a \right)^2 + \widetilde{u} \left( \Phi_3^a \Phi_3^a \right)^2 + v_1 \left( \Phi_1^a \Phi_2^a \right)^2 + v_2 \left( \Phi_1^a \Phi_1^a \right)\left( \Phi_2^b \Phi_2^b \right) \nonumber \\
&+ v_3 \left[ \left( \Phi_1^a \Phi_3^a \right)^2 + \left( \Phi_2^a \Phi_3^a \right)^2 \right] + v_4 \left( \Phi_1^a \Phi_1^a + \Phi_2^a \Phi_2^a \right)\left( \Phi_3^b \Phi_3^b \right) + \ldots
\label{eq:fullLocalLagrangian}
\end{aligned}
\end{equation}
We have absorbed the coefficients of the Yukawa couplings into the Higgs fields.
The general form of the Higgs potential $V(\Phi)$ is constrained by the microscopic symmetries of the square lattice, and we present only a subset of the possible terms.
The manner in which these microscopic symmetries are embedded in the continuum theory may be derived by starting from the original lattice model, and these transformation properties are given in Table~\ref{tab:symmetryTrans}. The action of the $\operatorname{SU}(2)$ spin rotation symmetry requires a more careful analysis and is described below. All three Higgs fields transform trivially under this $\operatorname{SU}(2)$ symmetry. Our choice of representing the fermionic degrees of freedom in terms of Dirac fermions obfuscates the full $\operatorname{SU}(2)$ spin rotation symmetry, as rotations around the $x$ or $y$ axis involve charge conjugation. However, the $\operatorname{U}(1)$ subgroup corresponding to rotations around the $z$ axis is simply given by a phase shift in $\psi$, $\psi \rightarrow e^{i\theta} \psi$. The full $\operatorname{SU}(2)$ rotation symmetry may be made explicit by writing the theory in terms of Majorana fermions~\cite{Wang17}, but this will not be necessary for our purposes.
\begin{table}
\begin{center}
\begin{tabular}{c||c|c|c|c|c|c}
& $T_x$ & $T_y$ & $P_x$ & $P_y$ & $\mathcal{T}$ & $R_{\pi/2}$
\\
\hline\hline
$\Phi_1^a$ & $-$ & $+$ & $-$ & $-$ & $-$ & $-\Phi_2^a$
\\
$\Phi_2^a$ & $+$ & $-$ & $-$ & $-$ & $-$ & $-\Phi_1^a$
\\
$\Phi_3^a$ &$-$ & $-$& $+$ & $+$ & $+$ & $-$
\\
$\psi$ & $\mu^x \psi$ & $\mu^z \psi$ & $\gamma^x \mu^z \psi$ & $- \gamma^y \mu^x \psi$ & $\gamma^0 \mu^y \psi$ & $e^{i\pi \gamma^0 / 4} e^{-i \pi \mu^y / 4} \psi$
\end{tabular}
\end{center}
\caption{Listed are the microscopic symmetries of the square lattice and their action on the continuum fields. $T_i$ ($P_i$) indicate translations (reflections) along the $i$'th axis, $\mathcal{T}$ is time-reversal symmetry, and $R_{\pi/2}$ is a $\pi/2$ rotation. We omit the action of $\operatorname{SU}(2)$ spin rotation symmetry as its action on the fermions $\psi$ is non-trivial and described in the main text.}
\label{tab:symmetryTrans}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{mfPhaseDiagram.png}
\caption{Mean field phase diagram of our low energy theory obtained by minimization of the Higgs potential in Eq.~(\ref{eq:fullLocalLagrangian}). Dashed (solid red) lines indicate second (first) order transitions in mean field theory. We assume the SU(2) $\pi$-flux gauge theory confines to a N\'eel state, the U(1) staggered flux gauge theory confines to a VBS state, except at their deconfined critical boundaries to Wen's stable, gapless $\mathbb{Z}_2$ spin liquid Z2A$zz13$.
The dotted blue line indicates a possible trajectory of the square lattice antiferromagnet with increasing $J_2/J_1$. }
\label{fig:mfPhaseDiagram}
\end{figure}
On a mean-field level, this theory admits three phases, illustrated in Fig.~\ref{fig:mfPhaseDiagram}. When all three Higgs fields are uncondensed, we recover the $\pi$-flux phase with $\operatorname{SU}(2)$ gauge group. When $\Phi_3$ condenses, we obtain the $\operatorname{U}(1)$ staggered flux phase. The condensation of $\Phi_1$ and $\Phi_2$ yields the gapless spin liquid Z2A$zz$13. The masses of $\Phi_1$ and $\Phi_2$ are fixed to be equal by the microscopic square lattice symmetry, so both condense simultaneously. Furthermore, the symmetry-allowed cubic term $\epsilon_{abc} \Phi_1^a \Phi_2^b \Phi_3^c$ forces $\Phi_3$ to condense along with $\Phi_1$, $\Phi_2$. Our conjectured trajectory of the $J_1$-$J_2$ model as a function of $\frac{J_2}{J_1}$ is shown by the dotted blue line in Fig.~\ref{fig:mfPhaseDiagram}, where transitions into either the staggered flux or $\pi$-flux phases drive the N\'eel or VBS ordering. Our theory may also be compatible with the inclusion of an antiferromagnetic third-nearest-neighbor $J_3$ term, which has been shown numerically~\cite{Liu2021} to compete against the spin liquid phase, eventually leading to a direct N\'eel/VBS transition. This phenomenon can be described in our theory by a deformation of the ${J_2}/{J_1}$ path to a trajectory in the $\pi$-flux phase; this deformation introduces first-order transitions near the tricritical point.
The transition from the $\pi$-flux phase to the $\mathbb{Z}_2$ phase was studied in~\cite{Shackleton21}, and the focus of our work will be the $\operatorname{U}(1)$ to $\mathbb{Z}_2$ transition.
Before proceeding with our analysis, we briefly summarize the results of our study of the $\operatorname{SU}(2)$ to $\mathbb{Z}_2$ transition. The primary order parameters for this theory are the masses of $\Phi_{1, 2}$, so we neglect fluctuations of $\Phi_3$. This critical theory is studied in a $1/N_f$ expansion, with $4 N_f$ the total number of fermions. Due to the anisotropic couplings of the Higgs fields, the leading-order effective propagators for $\Phi_{1, 2}$ are divergent along a one-dimensional subspace in momentum space. These lines of zero modes may be thought of as a consequence of an emergent subsystem symmetry. More details on this perspective are presented in Appendix~\ref{app:subsystemSymmetry}. Although these divergences are lifted by higher-order corrections, the leading-order effective action is non-local and more relevant at long distances than the bare Higgs kinetic term. Performing a standard momentum-shell renormalization group study of this leading-order theory is ill-defined and necessitates the inclusion of the irrelevant bare Higgs kinetic terms, which become ``dangerously irrelevant'' due to the singular nature of the leading-order theory. At one-loop order, these features lead to $\log^2$ divergences, rather than the more standard single-logarithm divergences found in conventional field theories. Re-exponentiating these corrections, we predict that the universal properties of correlation functions at criticality, rather than being power law, are instead $r^{-\alpha} \exp \left( {\beta \ln^2 r} \right)$, with $\beta= -{6}/{\pi^2}$ for the VBS correlator and $-{12}/{\pi^2}$ for the N\'eel correlator, and $\alpha$ being some \textit{non-universal} coefficient.
For the remainder of our work, we study the $\operatorname{U}(1)$ to $\mathbb{Z}_2$ transition, which is driven by the condensation of a charge-2 complex scalar Higgs field. Upon calculating the large-$N_f$ effective action, we perform a renormalization group (RG) analysis to determine the fixed point of our critical theory. The primary observables of interest that we will study are the dynamical critical exponent $z$, which determines the difference between spatial and temporal scaling, and correlations of the N\'eel and VBS order parameters, which are given by fermion bilinears in our continuum theory.
We find that the presence of the massless Higgs fields strongly modify the critical behavior from the theory of massless QED which describes the pure staggered flux phase without Higgs fields. The staggered flux phase normally possesses an $\operatorname{SU}(4)$ symmetry~\cite{Hermele05}, which relates various physical order parameters including N\'eel and VBS. Moreover, the staggered flux phase has an emergent Lorentz invariance at low energy, as the only velocity anisotropy term allowed by the square lattice symmetries is irrelevant in a $1/N_f$ expansion. The critical Higgs fields explicitly break the $\operatorname{SU}(4)$ symmetry and generate a non-zero velocity anisotropy, which further breaks the $\operatorname{U}(1)$ spatial rotation symmetry down to the $C_4$ symmetry present in the microscopic model. While correlation functions still have power law decay as in more traditional critical theories, the angular profiles of the correlation functions are modified by the velocity anisotropy; we calculate the tree-level effects of these modifications for the N\'eel and VBS correlation functions.
Recall that in the absence of critical Higgs fields, we postulate that the staggered flux phase is unstable to monopole proliferation. A key assumption in our analysis of the critical theory is that monopoles are rendered irrelevant due to the presence of critical Higgs fields and the theory may be studied using a \textit{non-compact} $\operatorname{U}(1)$ gauge field. Indeed, such an assumption is similar to the analysis of earlier studies of deconfined criticality between N\'eel and VBS orders~\cite{senthil1, senthil2}. This assumption is elaborated further in Section~\ref{sec:monopoles}, where we note that in addition to the critical Higgs fields, the presence of a non-zero anisotropy in the fermion dispersion relation can also affect the relevance of monopoles at criticality. In a large-$N_f$ expansion, this has an $\order{N_f}$ effect on the scaling dimension of the monopole, in contrast to the Higgs screening, which is $\order{1}$.
\section{Large \texorpdfstring{$N_f$}{Nf} effective action}
\label{sec:largeN}
Our goal is to study the $\operatorname{U}(1)$ to $\mathbb{Z}_2$ transition of Eq.~(\ref{eq:fullLocalLagrangian}). Both phases have the Higgs field $\langle \Phi_3^a \rangle \neq 0$, so we fix $\Phi_3^a = \delta_{a z} \Phi$, $\Phi \neq 0$. In this theory, the $\operatorname{SU}(2)$ gauge symmetry is broken to $\operatorname{U}(1)$, so we only consider $A_\mu \equiv A_\mu^z$. It is important to include the consequences of the cubic Higgs term, whose sign determines the form of the low energy complex Higgs. We choose a gauge where $w \Phi < 0$, and the low energy behavior can be described in terms of a single complex Higgs
\begin{equation}
\begin{aligned}
\mathcal{H} &= \frac{1}{2} \left( \Phi_1^x + \Phi_2^y + i (\Phi_1^y - \Phi_2^x) \right)\,. \\
\label{eq:lowEnergyHiggs}
\end{aligned}
\end{equation}
This field transforms as a charge 2 Higgs field under the unbroken $\operatorname{U}(1)$ symmetry, as desired. Other linear combinations of $\Phi_{1, 2}$ are massive and can be neglected for the critical theory. As we will clarify later, we must also assume $w < 0$ in order for our transition to be continuous - if $w$ is positive, the aforementioned massive Higgs fields will turn out to have a negative mass at the $\order{1/N_f}$ fixed point, which will lead to a first-order transition.
With these definitions, the Lagrangian that describes the $\operatorname{U}(1) \rightarrow \mathbb{Z}_2$ transition is
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text{sf}} &= \mathcal{L}_{\psi} + \mathcal{L}_{\mathcal{H}} + \mathcal{L}_{\mathcal{H} \psi}
\\
\mathcal{L}_{\psi} &= i\bar{\psi} \gamma^\mu D_\mu \psi + \Phi \, \bar{\psi} \mu^y \sigma^z \left(\gamma^y D_x + \gamma^x D_y \right) \psi \,.
\\
\mathcal{L}_{\mathcal{H}} &= s |\mathcal{H}|^2 + \partial_\mu \mathcal{H}^* \partial^\mu \mathcal{H} + u |\mathcal{H}|^4
\\
\mathcal{L}_{\mathcal{H} \psi} &= \mathcal{H} \bar{\psi} \left(\mu^z \gamma^x + i \mu^x \gamma^y \right) \sigma^- \psi
+ \mathcal{H}^\ast \bar{\psi} \left(\mu^z \gamma^x - i \mu^x \gamma^y \right) \sigma^+ \psi\,.
\label{eq:sfLagrangian}
\end{aligned}
\end{equation}
with $D_\mu = \partial_\mu - i A_\mu \sigma^z$ and $\sigma^\pm \equiv \sigma^x \pm i \sigma^y$. This Lagrangian may contain higher-order terms, but we omit these as they will turn out to be irrelevant in a $1/N_f$ expansion. The term proportional to $\Phi$ is a modification to the Lorentz-invariant fermion propagator $1 / \slashed{p}$ allowed by the projective symmetry group of the staggered flux phase. In the absence of the critical Higgs field, this velocity anisotropy has a stable fixed point value of $\Phi=0$~\cite{Franz2002, Vafek2002, Hermele05}. In our case, terms of this form are spontaneously generated at one-loop order by the critical Higgs field, and $\Phi$ acquires a non-zero value at the fixed point.
In order to study our critical theory, we proceed in a $1 / N_f $ expansion, with $N_f$ the fermion number. Since our theory only makes sense when the number of fermions $N$ is a multiple of $4$, we define $4 N_f = N$; in other words, we take $N_f = 1$ to correspond to our physical theory. At leading order in ${1}/{N_f}$, our effective bosonic action takes the form
\begin{equation}
\begin{aligned}
\label{eq:effectiveAction}
\frac{S_{b}}{N_f} &= \int_k \left[s + \Gamma(k)\right]\mathcal{H}^*(-k) \mathcal{H}(k) + \frac{1}{2} \Pi_{\mu\nu}(k) A_\mu(-k) A_\nu(k) \\
\end{aligned}
\end{equation}
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{higgsEffectivePropagator.pdf}
\includegraphics[width=0.45\textwidth]{gaugeEffectivePropagator.pdf}
\caption{The effective action for the Higgs boson (left) and $\operatorname{U}(1)$ gauge field (right) are generated by
the fermions at leading order in a $1/N_f$ expansion.}
\label{fig:effectiveActionDiagrams}
\end{figure}
where the inverse propagators $\Gamma$, $\Pi$ are generated by the one-loop fermion diagrams shown in Fig.~\ref{fig:effectiveActionDiagrams}. Note that we have taken the bare Higgs mass to scale with $N_f$, although we will be interested in the critical theory where we tune the Higgs mass to zero.
To calculate the effective propagators, we need the fermion propagator, which receives corrections to its Lorentz-invariant value of $1/\slashed{p}$ due to a non-zero $\Phi$. This may be treated perturbatively in $\Phi$, but the existence of a stable fixed point turns out to not be viewable at leading order, so we instead proceed with a non-perturbative treatment of $\Phi$. We include further details of this calculation in Appendix~\ref{app:effectiveAction}, and cite the results in the main text. Defining the variables
\begin{equation}
\begin{aligned}
k_{x, \pm} \equiv k_x \pm \Phi k_y \,, \quad k_{y, \pm} \equiv k_y \pm \Phi k_x\,, \quad \abs{k_{\pm}} \equiv \sqrt{k_0^2 + k_{x, \pm}^2 + k_{y, \pm}^2}
\end{aligned}
\end{equation}
the effective inverse Higgs propagator (obtained from the $\Phi$-dependent fermion propagator) is
\begin{equation}
\begin{aligned}
\Gamma(k) &= \frac{1}{16 N_f (1-\Phi^2) } \left[ \frac{ k_+^2 + k_0^2 + 2 k_{x, +} k_{y, +}}{\abs{k_+}} + \frac{ k_-^2 + k_0^2 - 2 k_{x, -} k_{y, -}}{\abs{k_-}} \right]\,.
\label{eq:higgsPropIntegrand}
\end{aligned}
\end{equation}
Likewise, we need the general form of the effective gauge boson propagator. The presence of a non-zero $\Phi$ modifies the gauge coupling, and hence non-Lorentz-invariant corrections arise both from $\Phi$-dependent modifications to the fermion propagator as well as $\order{\Phi}$ vertices. We separate this calculation into three pieces. The first correction comes from using the $\order{\Phi^0}$ vertices, but with the full fermion propagator. This one-loop term contributes
\begin{equation}
\begin{aligned}
\Pi_{\mu\nu}^{(1)}(k) &= \sum_{a = \pm} \frac{k_{a}^2}{8 N_f (1-\Phi^2) \abs{k_{a}}} \left( \delta_{\mu\nu} - \frac{k_{\mu a} k_{\nu a}}{k_{a}^2} \right)\,.
\end{aligned}
\end{equation}
The second correction comes from using one $\order{\Phi}$ vertex, which gives the contribution
\begin{equation}
\begin{aligned}
\Pi_{\mu x}^{(2)}(k) &= \Pi^{(2)}_{x \mu}(k) =\sum_{a = \pm} \frac{a \Phi k_{a}^2}{8 N_f (1-\Phi^2) \abs{k_{a}}} \left( \delta_{\mu y} - \frac{k_{\mu a} k_{y a}}{k_{a}^2} \right)
\\
\Pi_{\mu y}^{(2)}(k) &= \Pi_{y \mu}^{(2)}(k) =\sum_{a = \pm} \frac{a \Phi k_{a}^2}{8 N_f (1-\Phi^2) \abs{k_{a}}} \left( \delta_{\mu x} - \frac{k_{\mu a} k_{x a}}{k_{a}^2} \right) \,;
\end{aligned}
\end{equation}
There is also an extra factor of $2$ in $\Pi^{(2)}_{xx, yy}$ due to the two possible vertex orderings. Finally, the third correction comes from using two $\order{\Phi}$ vertices,
\begin{equation}
\begin{aligned}
\Pi_{xx}^{(3)}(k) &= \sum_{a=\pm} \frac{\Phi^2 k_{a}^2}{8 N_f (1-\Phi^2) \abs{k_{a}}} \left( 1 - \frac{k_{y a}^2}{k_{a}^2} \right)
\\
\Pi_{yy}^{(3)}(k) &=\sum_{a=\pm} \frac{\Phi^2 k_{a}^2}{8 N_f (1-\Phi^2) \abs{k_{a}}} \left( 1 - \frac{k_{x a}^2}{k_{a}^2} \right)
\\
\Pi_{yx}^{(3)}(k) &= \Pi^{(3)}_{xy}(k) = \sum_{a=\pm}-\frac{\Phi^2 k_{x a} k_{y a}}{8 N_f (1-\Phi^2) \abs{k_{a}}}
\end{aligned}
\end{equation}
We verify that the combined inverse propagator $\Pi_{\mu\nu}(k) = \sum_{i=1,2,3}\Pi^{(i)}_{\mu\nu}(k)$ annihilates the vector $(k_0, k_x, k_y)$, as required by gauge invariance. Note that $\Pi_{\mu\nu}$ requires a gauge fixing term in order to be invertable. Followin Ref.~\cite{Hermele05}, we add the following non-local gauge fixing term to the Lagrangian
\begin{equation}
\begin{aligned}
\frac{1}{4 \xi \abs{k}} A_\mu k^\mu k^\nu A_\nu\,.
\end{aligned}
\end{equation}
All gauge-invariant observables have been checked to ensure they are independent of the choice of $\xi$.
\section{Renormalization group analysis}
\label{sec:rg}
We perform a renormalization group (RG) analysis of the $\order{1 / N_f}$ effective theory. We are interested in studying the behavior of this theory under the rescaling
\begin{equation}
\begin{aligned}
k &= k' e^{-\ell}
\\
\omega &= \omega' e^{-z \ell} \\
\end{aligned}
\end{equation}
We also define a rescaling of the fermion fields
\begin{equation}
\begin{aligned}
\psi(k, \omega) &= \psi'(k', \omega') e^{\frac{\ell}{2} (2+2z - \eta_f)}
\end{aligned}
\end{equation}
The Higgs and gauge fields must also be suitably rescaled, although the anomalous dimensions of these fields will not be needed to calculate our observables of interest. In the absence of a standard boson kinetic term at leading order in ${1}/{N_f}$, we define the scaling of the boson field by performing our RG such that the Yukawa coupling remains fixed under RG.
\subsection{Fermion self-energy}
We first evaluate the $\order{1 / N_f}$ contributions to the fermion self-energy, which come from both gauge and Higgs one-loop diagrams. The self-energy is UV divergent and requires a UV cutoff $\Lambda$. The logarithmic derivative of the fermion self-energy with respect to this cutoff takes the general form
\begin{equation}
\begin{aligned}
\Lambda \dv{\Lambda} \Sigma(k) = C_0 k_0 \gamma^0 + C_1 (k_x \gamma^x + k_y \gamma^y) + C_2 \Phi \mu^y \sigma^z (k_x \gamma^y + k_y \gamma^x)
\label{eq:logDerivativeSelfEnergy}
\end{aligned}
\end{equation}
for constants $C_{0, 1, 2}$. One must verify that only these terms are generated at one-loop order, which we have done.
In order to calculate the constants $C_i$, we will use the momentum-shell RG approach outlined in Ref.~\cite{Huh08}. The regularized one-loop contribution to the self-energy schematically takes the form
\begin{equation}
\begin{aligned}
\Sigma(k) = \int \frac{\dd[3]{p}}{(2\pi)^3} F(p+k) G(p) \mathcal{K}\left( \frac{p^2}{\Lambda^2} \right) \mathcal{K}\left( \frac{(k+p)^2}{\Lambda^2} \right)
\end{aligned}
\end{equation}
where $F$ and $G$ are homogeneous functions of the three-momenta, and $\mathcal{K}(y)$ serves as a UV cutoff with the property that $\mathcal{K}(0) = 1$ and $\mathcal{K}(y)$ falls off rapidly for large $y$. In our calculations, we take $F$ to be the fermion propagator, and $G$ to be the boson propagator (either Higgs or gauge), along with vertex coefficients. The fact that $F$ and $G$ are homogeneous functions allows us to remove the explicit dependence on $\mathcal{K}$ upon taking the logarithmic derivative and integrating by parts. We refer to Appendix~\ref{app:rg} for an explicit derivation of this, and state the result here - the logarithmic derivative of the self-energy takes the form
\begin{equation}
\begin{aligned}
\Lambda \dv{\Lambda} \Sigma(k) &= \frac{k_\lambda}{8\pi^3 N_f} \int_0^{2\pi} \dd{\phi} \int_0^\pi \sin\theta \dd{\theta} \pdv{F(\hat{p})}{p_\lambda} G(\hat{p})\,.
\label{eq:rgConvergentIntegrals}
\end{aligned}
\end{equation}
where $\hat{p} \equiv (\cos\theta, \sin\theta\sin\phi, \sin\theta \cos\phi)$. The resulting integrals in Eq.~(\ref{eq:rgConvergentIntegrals}) are fully convergent and may be evaluated numerically, from which we can extract the coefficients $C_{0, 1, 2}$.
\subsection{Fixed points}
The RG equations for the velocity anisotropy $\Phi$ are
\begin{equation}
\begin{aligned}
\dv{\Phi}{\ell} = (C_1 - C_2) \Phi\,.
\label{eq:rgFlow}
\end{aligned}
\end{equation}
In the absence of the Higgs field, $\Phi$ has a stable fixed point at $\Phi=0$. The gauge field contribution to this equation has been calculated to leading order in $\Phi$~\cite{Hermele05}, and we verify agreement with this result.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{RG.pdf}
\caption{An evaluation of the RG flow of $\Phi$, showing a stable fixed point at $\Phi_c \approx 0.46$.}
\label{fig:rgFlow}
\end{figure}
The evaluation of Eq.~(\ref{eq:rgFlow}) is plotted in Fig.~\ref{fig:rgFlow}. A stable fixed
point is found at $\Phi_c \approx 0.45765$. At this point, the dynamical critical exponent $z$
is given by
\begin{equation}
\begin{aligned}
z = 1 - C_0 + C_1 = 1 + \frac{0.225}{N_f} + \order{1 / N_f^2}\,.
\end{aligned}
\end{equation}
Recall that when we derived this critical $\operatorname{U}(1) \rightarrow \mathbb{Z}_2$ theory as a component of a parent $\operatorname{SU}(2)$ theory, we made a gauge choice such that $w \Phi < 0$, where $w$ is the coefficient of the symmetry-allowed cubic term, $w \epsilon_{abc} \Phi_1^a \Phi_2^b \Phi_3^c$. When $\langle \Phi_3^a \rangle = \Phi \delta_{az}$, we can diagonalize this term to yield a mass, $w \Phi (\mathcal{H}^* \mathcal{H} - \mathcal{M}^* \mathcal{M})$, where $\mathcal{H}$ is the combination of $\Phi_{1, 2}^{x, y}$ given in Eq.~\ref{eq:lowEnergyHiggs} and $\mathcal{M}$ is a charge-2 Higgs field of a similar form but with $x \leftrightarrow y$. If we assume $w > 0$, $\Phi < 0$, then $\mathcal{H}$ will become massless first, but the fixed-point value of $\Phi$ gives a negative mass to $\mathcal{M}$, leading to a first-order transition driven by the condensation of $\mathcal{M}$. As a consequence, we must fix our parent $\operatorname{SU}(2)$ theory to have $w < 0$ in order to yield a continuous transition. If we had made a gauge choice such that $w \Phi > 0$, then our theory would have been driven by the condensation of $\mathcal{M}$ rather than $\mathcal{H}$; this still leads to the gapless $\mathbb{Z}_2$ spin liquid Z2Azz13, and all gauge-invariant observables at the critical point remain the same, although the sign of $\Phi_c$ changes.
\subsection{N\'eel and VBS order parameter corrections}
We now calculate the vertex corrections to the N\'eel and VBS order parameters. These order parameters are given by fermion bilinears and can be identified based on the action of the microscopic square lattice symmetries on the fermions. The VBS order parameter is given by the bilinears $\bar{\psi} \mu^{z, x} \psi$. As mentioned previously, our particular representation obfuscates the full $\operatorname{SU}(2)$ action of spin rotation symmetry; however, the $\operatorname{U}(1)$ subgroup generated by rotations around the $z$-axis is given by the global $\operatorname{U}(1)$ symmetry $\psi \rightarrow e^{i\theta}\psi$ (recall that this is not the $\operatorname{U}(1)$ gauge symmetry, which acts as $e^{i \theta \sigma^z}$). As a consequence, we focus on the $z$-component of the N\'eel order parameter, which is given by $\bar{\psi} \mu^y \psi$. The two-point correlation functions of these bilinears are obtained by coupling them to external sources $J_{\text{VBS, N\'eel}}$ and calculating $\order{N_f^{-1}}$ vertex corrections, illustrated in Fig.~\ref{fig:bilinearDiagrams}. Of note are $\order{N_f^{-1}}$ two-loop corrections, the form of which were first found in Ref.~\cite{Hermele05}. These two-loop corrections to the N\'eel and VBS order parameters vanish in the pure staggered flux phase - this follows immediately from taking the trace over the internal fermion loop and noting that $\Tr \mu^i = 0$. We verify that these diagrams remain zero upon the inclusion of both Higgs fields and velocity anisotropy, although this identity is less readily apparent.
\begin{figure}
\centering
\includegraphics[width=1.5in]{feynmanHiggsVertex.pdf}
\includegraphics[width=1.5in]{feynmanGaugeVertex.pdf}
\includegraphics[width=2.3in]{twoLoopFeynmanDiagonalFermion.png}
\caption{The $\mathcal{O}(N_f^{-1})$ vertex corrections which contribute to the renormalization of the N\'eel and VBS order parameters.
The order parameter receives corrections at one-loop order from the Higgs fields (left) and the gauge boson (center).
An additional two-loop $\mathcal{O}(N_f^{-1})$ contribution (right) is possible - the diagram shown involves two intermediate
Higgs propagators, but additional diagrams with gauge propagators or mixed gauge/Higgs propagators are possible. These diagrams vanish exactly upon performing the trace over the internal fermion indices.}
\label{fig:bilinearDiagrams}
\end{figure}
\subsubsection{One-loop vertex corrections}
We first outline the procedure from Ref.~\cite{Huh08} for calculating the logarithmic corrections to the vertex functions. At zero external momenta, our one-loop vertex corrections schematically take the form
\begin{equation}
\begin{aligned}
\Xi_i = \frac{1}{N_f} \int \frac{\dd[3]{p}}{(2\pi)^3} H_i(p) \mathcal{K}^3\left( \frac{p^2}{\Lambda^2} \right)
\end{aligned}
\end{equation}
where $H_i(p)$, $i = x, y, z$, is a homogeneous function of $p$ and illustrated in Fig.~\ref{fig:bilinearDiagrams}. The index $i$ indicates whether the $J$ vertex includes a factor of $\mu^{x, z}$ (VBS order parameter) or $\mu^y$ (N\'eel order parameter). Once again, we can take the logarithmic derivative and remove the explicit cutoff dependence, leading to the equation
\begin{equation}
\begin{aligned}
\Lambda \dv{\Lambda} \Xi_i = \frac{1}{8\pi^3 N_f} \int_0^{2\pi} \dd{\phi} \int_0^\pi \sin\theta \dd{\theta} H_i(\hat{p}) \equiv B_i \mu^i\,.
\end{aligned}
\end{equation}
The $B_i$'s are not gauge-invariant by themselves, and must be combined with the self-energy to get a gauge-invariant quantity, which at the fixed-point value $\Phi_c$ gives
\begin{equation}
\begin{aligned}
\eta_{\text{VBS}} &= B_{x, z} + C_0 \approx 0.06468 N_f^{-1} + \order{N_f^{-2}}\,,
\\
\eta_{\text{N\'eel}} &= B_{y} + C_0 \approx -0.01634 N_f^{-1} + \order{N_f^{-2}}\,.
\end{aligned}
\end{equation}
The N\'eel and VBS correlators in momentum space have the scaling form
\begin{equation}
\begin{aligned}
G_{\text{N\'eel}}(k, \omega) &= G_{\text{N\'eel}}(a k, a^z \omega) a^{2 \eta_{\text{N\'eel}} - 1}\,,
\\
G_{\text{VBS}}(k, \omega) &= G_{\text{VBS}}(a k, a^z \omega) a^{2 \eta_{\text{VBS}} - 1}\,.
\end{aligned}
\end{equation}
Making a Fourier transform to real space, the equal-time N\'eel and VBS correlators have the power law decay
\begin{equation}
\begin{aligned}
G_{\text{N\'eel}}(r) &\sim \frac{1}{r^{3 + z - 2 \eta_{\text{N\'eel}}}}
\\
G_{\text{VBS}}(r) &\sim \frac{1}{r^{3 + z - 2 \eta_{\text{VBS}}}}\
\end{aligned}
\end{equation}
Note that both the anomalous dimensions for the N\'eel and VBS correlators are quite small. This is a rather surprising result and does not seem to be due to any particular small parameter. The magnitude of these anomalous dimensions do not decrease upon increasing the numerical precision of our integration, so we believe them to be small but not identically zero. We find that the gauge fluctuations generally enhance N\'eel and VBS correlations, whereas Higgs fluctuations suppress them - the combined result is the stated anomalous dimensions. As such, we cannot make a strong statement regarding which ordering the unstable $\operatorname{U}(1)$ phase will prefer, as neither the N\'eel nor VBS order parameter show exceptionally enhanced correlations. Higher-order corrections may show a clearer preference to either N\'eel or VBS ordering.
\subsection{Tree-level effect of velocity anisotropy on correlation functions}
As we have emphasized, one of the key features of this critical theory is that the emergent Lorentz invariance of the staggered flux phase is broken by the presence of critical Higgs fields, leading to a non-zero value of the symmetry-allowed velocity anisotropy term. This anisotropy term also has the effect of breaking the emergent $\operatorname{SU}(4)$ flavor symmetry. We refer to Ref.~\cite{Hermele05} for a more extensive study of the intertwining physical order parameters of the $\operatorname{SU}(4)$ theory and which relations hold in the presence of the velocity anisotropy - for our purposes, we note that the emergent $\operatorname{SO}(5) \subset \operatorname{SU}(4)$ symmetry that relates the N\'eel and VBS order parameters is broken down to the microscopic $\operatorname{SO}(3) \times C_4$. At tree-level, the scaling dimensions of the two order parameters are still the same, but the angular profile of their correlation functions are modified due to the velocity anisotropy. This lack of an emergent $\operatorname{SO}(2)$ spatial rotation symmetry in the N\'eel and VBS correlation functions may be useful as a numerical probe of the critical behavior, so we study the angular profile in more detail.
We analytically compute the spatial profile of the N\'eel order parameter at tree level. This calculation turns out to be feasible non-perturbatively in the velocity anisotropy $\Phi$. The VBS correlator is more difficult to study non-perturbatively in the velocity anisotropy, and we will later compute corrections to leading order in $\Phi$.
The two-point function in momentum space is given by, with $Q(p)$ the fermion propagator,
\begin{equation}
\begin{aligned}
G_{\text{N\'eel}}(k) &= - \int \frac{\dd[3]{p}}{(2\pi)^3} \Tr\left[ Q(p) \mu^y Q(p + k) \mu^y \right]
\\
&= - \frac{2}{1-\Phi^{2}} \sum_{a=\pm}\int \frac{\dd[3]{p}}{(2\pi)^3} \frac{p_0 (p_0 + k_0) + a p_x (p_x + k_{x, a}) + a p_y (p_y + k_{y, a})}{p^2 (p + k_{\pm})^2}
\\
&= \frac{1}{8(1-\Phi^2)}\left( \abs{k_+} + \abs{k_-} \right)\,.
\end{aligned}
\end{equation}
As before, we define $k_{\pm} = (k_0, k_x \pm \Phi k_y, k_y \pm \Phi k_x)$. The Fourier transform
\begin{equation}
\begin{aligned}
\int \frac{\dd[3]{k}}{(2\pi)^3} e^{i k \cdot r} \abs{k_\pm}
\end{aligned}
\end{equation}
can be performed by a change of variables to give
\begin{equation}
\begin{aligned}
G_{\text{N\'eel}}(r) &\sim \sum_{a=\pm} f_{a}(\theta, \phi)^{-2} \frac{1}{r^4}
\end{aligned}
\end{equation}
where we change to spherical coordinates, $t = r \cos\theta$, $x = r \sin\theta\sin\phi$, $y = r \sin\theta \cos\phi$, and
\begin{equation}
\begin{aligned}
f_{\pm}(\theta, \phi) &= 1 + \frac{\sin^2\theta\left( \pm 2 \Phi \sin 2 \phi + 3 \Phi^2 + \Phi^4 \right)}{(1 - \Phi^2)^2}\,.
\end{aligned}
\end{equation}
Therefore,
\begin{equation}
\begin{aligned}
G_{\text{N\'eel}}(r) \sim g(\theta, \phi) \frac{1}{r^4}
\end{aligned}
\end{equation}
where $g(\theta, \phi) = {1}/{f_+(\theta, \phi)^2} + {1}/{f_-(\theta, \phi)^2}$ is plotted in Fig.~\ref{fig:spatialDependence}. We note the enhanced correlations of the N\'eel order parameter along the diagonals, which holds true for generic
values of $\Phi$.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{angularDep}
\caption{Plotted are the angular profiles of the equal-time N\'eel and VBS correlation functions in real space, at the fixed point value of velocity anisotropy $\Phi_c$. The N\'eel order parameter shows enhanced correlations along the diagonals, whereas the VBS correlations are more enhanced along the $x$ and $y$ directions. Note that we only plot the absolute value of the correlation function, and the signs of the perturbative N\'eel and VBS correlators flip when moving from the $x$ and $y$ axes to the diagonals. As this feature is not present in the non-perturbative N\'eel correlator, it is possible that this feature similarly vanishes at higher orders for the VBS correlator. }
\label{fig:spatialDependence}
\end{figure}
An analogous calculation of the VBS order parameter is less analytically tractable, as the one-loop integral cannot be made isotropic by a coordinate transformation. As such, we resort to a perturbative study of the velocity anisotropy. This gives
\begin{equation}
\begin{aligned}
G_{\text{VBS}}(k) \sim \abs{k} - \Phi^2 \left[ 2 \abs{k} + \frac{k_x^2 k_y^2 - k_0^2 k^2}{\abs{k}^3} \right] + \order{\Phi^4}
\end{aligned}
\end{equation}
in momentum space, or
\begin{equation}
\begin{aligned}
G_{\text{VBS}}(r) \sim \frac{1}{r^4} \left[ 8 + \Phi^2 (\cos 2\theta (40 + 12 \cos 4 \phi) + \cos 4 \theta (6 - 3 \cos 4 \phi) - 18 \cos 4 \phi - 14) \right]
\end{aligned}
\end{equation}
in real space. The equal-time VBS correlation function is plotted in Fig.~\ref{fig:spatialDependence}, showing enhanced correlations along the cardinal directions. Note that the correlation function changes sign on the diagonals - this is an unusual feature, and would seemingly indicate lines in real space where the VBS correlator vanishes. This feature is also present in the $\order{\Phi^2}$ corrections to the N\'eel correlator but ultimately vanishes in the non-perturbative result, so this result may only be an artifact of the perturbative expansion. Further details on this calculation can be found in Appendix~\ref{app:ft}.
\section{Monopoles}
\label{sec:monopoles}
On the square lattice, there exists a monopole operator in the staggered flux phase - the \textit{trivial monopole} - that is invariant under all square lattice symmetries, and hence is an allowed perturbation. To leading order, the scaling dimension of the monopole operator scales with the number of fermions and becomes irrelevant for $N_f \geq 3$. Hence, the staggered flux phase by itself is unstable to monopole proliferation - this is the mechanism which we claim gives rise to ordered phases in the staggered flux phase, as condensation of the trivial monopole is conjectured to lead to a fermion chiral mass generation corresponding to either N\'eel or VBS order ~\cite{Song1, Song2}. Our calculations of the critical theory rely on the assumption that the presence of massless scalar fields screens monopoles and renders them irrelevant at the critical point. Here, we draw attention to an additional contribution to the monopole scaling dimension, which is the non-zero anisotropy in the spinon dispersion. Prior calculations of monopole scaling dimensions in $\text{QED}_3$ assume a Lorentz-invariant action for the fermions, which is natural in pure $\text{QED}_3$ given that velocity anisotropy terms are irrelevant in a ${1}/{N_f}$ expansion. However, as we have shown, the presence of critical Higgs fields can give rise to a non-zero velocity anisotropy at the critical point. An important question is whether this anisotropy increases or decreases the monopole scaling dimension. In contrast with the direct modification arising from the critical fields, which is $\order{1}$, the effect of the anisotropy on the monopole scaling dimension is $\order{N_f}$. Such a modification, if calculated perturbatively in $\Phi$, arises at $\order{\Phi^2}$ - this is still an appreciable shift given the relatively large anisotropy $\Phi_c \approx 0.46$. Previous works have studied the effects of a spin Hall mass on the monopole scaling dimension~\cite{Dupuis2019,Dupuis2021b}, although this perturbation is more tractable as the spin Hall mass is diagonal in the basis of spinor monopole harmonics. In Appendix~\ref{app:monopole}, we outline the structure of a perturbative calculation for calculating the $\order{\Phi^2}$ corrections to the monopole scaling dimension. An important observation which makes this calculation tractable is that, while the saddle-point monopole gauge configuration will not take the form of the rotationally-invariant Dirac monopole, corrections to the scaling dimension arising from this difference only arise at higher orders in $\Phi$; hence, to lowest non-trivial order, one can assume a Dirac monopole background. This calculation ultimately yields a divergent summation of terms involving Wigner $3$-$j$ symbols; we leave for future work further study of how to properly regularize this calculation.
Additionally, we briefly comment on the relation between this velocity anisotropy and the monopole quantum numbers. Prior studies on the effects of a spin Hall mass~\cite{Dupuis2019, Dupuis2021b} have found that the presence of such a term induces a spin polarization on the monopoles. Each fermion flavor has a zero mode in the presence of a monopole, and half of these zero modes must be filled in order to maintain gauge neutrality of the monopole. The presence of a spin Hall mass polarizes these zero modes, which in turn causes a splitting in the scaling dimension of the monopoles, with the most-relevant monopole being spin polarized. One may wonder whether a similar valley polarization can arise due to our velocity anisotropy term due to the presence of a $\mu^y$ in the anisotropy; however, we check that the first-order energy splitting of the fermion zero modes due to the velocity anisotropy vanishes. Higher-order corrections including corrections from Higgs and gauge fluctuations will in general break the six-fold degeneracy of monopole scaling dimensions; in particular, the monopoles with N\'eel and VBS quantum numbers will have different scaling dimensions, which may cause a preference towards a particular type of symmetry-breaking in the staggered flux phase. Further study of the spectrum of monopoles at this critical point may be useful for determining the IR fate of the proximate staggered flux phase - our observation is that this behavior is more complicated than a simple valley polarization of the monopoles.
\section{Conclusions and future directions}
We have presented a large-$N_f$ analysis of a deconfined critical theory separating a gapless $\mathbb{Z}_2$ spin liquid Z2Azz13 from the $\operatorname{U}(1)$ staggered flux phase, the latter of which we assume to be unstable to either N\'eel or VBS ordering on the square lattice. This completes the large-$N$ study of the phase diagram shown in Fig.~\ref{fig:mfPhaseDiagram}, where the gapless $\mathbb{Z}_2$ spin liquid may emerge as a Higgsed phase from either a $\operatorname{U}(1)$ or $\operatorname{SU}(2)$ gauge theory. Both of these parent gauge theories are conjectured to be unstable on the square lattice, and hence we propose the trajectory through the phase diagram as shown in Fig.~\ref{fig:mfPhaseDiagram} as a description of the $J_1$-$J_2$ square lattice antiferromagnet, where numerical studies suggest a gapless $\mathbb{Z}_2$ spin liquid emerging between N\'eel and VBS phases.
Our calculations yield several predictions which may be investigated by future numerical studies. One of the most striking features of both the $\operatorname{SU}(2) \rightarrow \mathbb{Z}_2$ and $\operatorname{U}(1) \rightarrow \mathbb{Z}_2$ transitions is the lack of Lorentz invariance, spatial rotation invariance (aside from the discrete $C_4$ rotational symmetry), and in the case of the $\operatorname{SU}(2) \rightarrow \mathbb{Z}_2$ transition, even a lack of traditional scale invariance. The lack of scale invariance takes the form of correlation functions decaying as $e^{-\ln^2(r)}$ rather than power law, the difference of which is difficult to detect for small system sizes. As such, it may be more promising to search for lack of Lorentz invariance ($z \neq 1$), or lack of a full $\operatorname{SO}(2)$ spatial rotation invariance of correlation functions. We draw attention to the angular profiles of the N\'eel and VBS correlation functions shown in Fig.~\ref{fig:spatialDependence}, which come from a mean-field description of the staggered flux state with the inclusion of a symmetry-allowed velocity anisotropy and predict enhanced N\'eel correlations along the diagonals, and enhanced VBS correlations along the cardinal directions.
\section{Acknowledgements}
We are grateful to Zheng-Cheng Gu, Patrick Ledwith, and Alex Thomson for insightful discussions. This research was supported by the National Science Foundation under Grant No. DMR-2002850. This work was also supported by the Simons Collaboration on Ultra- Quantum Matter, which is a grant from the Simons Foundation (651440, S.S.).
| b1ca2e00a669aa43a1dfc8fa28474d7dec1ee243 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Concise historical background and motivation}
On one hand, the approximation of $2\pi$-periodic and integrable functions by their Fourier series in the H\"{o}lder metric has been studied widely and consistently in many papers. Das at al. studied the degree of approximation of functions by matrix means of their Fourier series in the generalized H\"{o}lder metric \cite{DGR}, generalizing many previous known results. Again, Das at al. \cite{DNR} studied the rate of convergence problem of Fourier series in a new Banach space of functions conceived as a generalization of the spaces introduced by Pr\"{o}ssdorf \cite{SP} and Leindler \cite{L1}. Afterwards, Nayak at al. \cite{NDR} studied the rate of convergence problem of the Fourier series by delayed arithmetic mean in the generalized H\"{o}lder metric space which was earlier introduced in \cite{DNR} and obtaining a sharper estimate of Jackson's order which is the main objective of their result. Moreover, De\v{g}er \cite{D} determined the degree of approximation of functions by matrix means of their Fourier series in the same space of functions introduced in \cite{DNR}. In particular, he extended some results of Leindler \cite{L} and some other results by weakening the monotonicity conditions in the results obtained by Singh and Sonker \cite{SS} for some classes of numerical sequences introduced by Mohapatra and Szal \cite{MS}. Leindler's results obtained in \cite{L}, are generalized in \cite{XhK} by first author of the present paper, for functions from a Banach space, mainly using the generalized N\"{o}rlund and Riesz means. Very recently, Kim \cite{KIM1} presented a generalized result of a particular case of a result obtained previously in \cite{NDR}. In Kim's result is treated the degree of approximation of functions in the same generalized H\"{o}lder metric, but using the so-called even-type delayed arithmetic mean of Fourier series.
On the other hand, results on approximation of bivariate integrable functions and $2\pi$-periodic in each variable by their double Fourier series in the H\"{o}lder metric, the interested reader can find in \cite{UD2}, \cite{XhK2} and \cite{NH}. In all results reported in these papers, we came across to the degree of approximation of functions by various means of their double Fourier series and in which the quantity of the form $\mathcal{O}{(\log n)}$ appears. Involving such quantity produces a degree of approximation which is not of Jackson's order. This weakness motivated us to consider some means of double Fourier series which will overshoots it. Whence, we are going to investigate the degree of approximation of bivariate integrable functions and $2\pi$-periodic in each variable by their double Fourier series in the generalized H\"{o}lder metric, by motivation of removing the quantities of the form $\mathcal{O}{(\log n)}$ and obtaining the degree of approximation of Jackson's order, which is the aim of the paper.
Closing this section, for comparing of two quantities $u$ and $v>0$, throughout this paper we write $u=\mathcal{O}(v)$, whenever there exists a positive constant $c$ such that $u\leq cv$.
\section{Introduction and preliminaries}
By $L_p(T^2)$, $p\geq 1$, we denote the space of all functions $f(x,y)$
integrable with $p$-power on $T^2:=(0,2\pi)\times (0,2\pi)$, and with norm
\begin{equation*}
\|f\|_p:=\left(\frac{1}{(2\pi)^2}\int_{0}^{2\pi}\int_{0}^{2\pi}|f(x,y)|^p
dxdy\right)^{1/p}.
\end{equation*}
If $f\in L_p(T^2)$, $p\geq 1$, then $\omega _i$, $(i=1,2)$ are considered as
moduli of continuity if $\omega _i$ are two positive non-decreasing
continuous functions on $[0,2\pi]$ with properties
\begin{enumerate}
\item[(i)] $\omega _i(0)=0$,
\item[(ii)] $\omega _i(\delta_1+\delta_2)\leq \omega _i(\delta_1)+\omega
_i(\delta_2)$,
\item[(iii)] $\omega _{i}(\lambda \delta )\leq (\lambda +1)\omega
_{i}(\delta )$, $\lambda \geq 0$.
\end{enumerate}
We define the space $H_{p}^{(\omega_1, \omega_2)}$ by
\begin{equation*}
H_{p}^{(\omega_1, \omega_2 )}:=\left\{f\in L^{p}(T^2), p\geq 1:
A(f;\omega_1, \omega_2 )<\infty \right\},
\end{equation*}
where
\begin{equation*}
A(f;\omega_1, \omega_2 ):=\sup_{t_1\neq 0,\,\,t_2\neq 0 }\frac{\|f(x +t_1,y
+t_2)-f(x,y)\|_{p}}{\omega_1 (|t_1|)+\omega_2 (|t_2|)}
\end{equation*}
and the norm in the space $H_{p}^{(\omega_1, \omega_2)}$ is defined by
\begin{equation*}
\|f\|_{p}^{(\omega_1, \omega_2)}:=\|f\|_{p}+A(f;\omega_1, \omega_2 ).
\end{equation*}
If $\omega_1$, $\omega_2$, $v_1$ and $v_2$ are moduli of continuity so that
the two-variable function $\frac{\omega_1 (t_1)+\omega_2 (t_2)}
v_1(t_1)+v_2(t_2)}$ has a maximum $M$ on $T^2$, then it is easy to see that
\begin{equation*}
\|f\|_{p}^{(v_1,v_2)}\leq \max \left(1,M\right)\|f\|_{p}^{(\omega_1,
\omega_2)},
\end{equation*}
which shows that in this case, for the given spaces $H_{p}^{(\omega_1,
\omega_2)}$ and $H_{p}^{(v_1, v_2)}$ we have
\begin{equation*}
H_{p}^{(\omega_1, \omega_2)}\subseteq H_{p}^{(v_1, v_2)}\subseteq L_p \quad
(p\geq 1).
\end{equation*}
We write
\begin{equation*}
\Omega_p(\delta_1,\delta_2;f):=\sup_{0\leq h_1\leq \delta_1; 0\leq h_2\leq
\delta_2}\|f(x+h_1,y+h_2)-f(x,y)\|_{p}
\end{equation*}
for the integral modulus of continuity of $f(x,y)$, and whenever
\begin{equation*}
\|f(x +t_1,y +t_2)-f(x,y)\|_{p}=\mathcal{O}\left(\omega_1 (|t_1|)+\omega_2
(|t_2|)\right)
\end{equation*}
we write $f\in \text{Lip}(\omega_1,\omega_2,p)$, that is
\begin{equation*}
\text{Lip}(\omega_1,\omega_2,p)=\left\{f\in L_p(T^2): \|f(x +t_1,y
+t_2)-f(x,y)\|_{p}=\mathcal{O}\left(\omega_1 (|t_1|)+\omega_2
(|t_2|)\right)\right\}.
\end{equation*}
Clearly, for $\omega_1(t_1)=\mathcal{O}\left(t_1^{\alpha}\right)$ and
\omega_2(t_2)=\mathcal{O}\left(t_2^{\beta}\right)$, $0<\alpha \leq 1$,
0<\beta \leq 1$, the class $\text{Lip}(\omega_1,\omega_2,p)$ reduces to the
class $\text{Lip}(\alpha,\beta,p)$, that is
\begin{equation*}
\text{Lip}(\alpha,\beta, p)=\left\{f\in L^p(T^2): \|f(x +t_1,y
+t_2)-f(x,y)\|_{p}=\mathcal{O}\left(t_1^{\alpha}\right)+\mathcal{O
\left(t_2^{\beta}\right)\right\}.
\end{equation*}
Then for $1 \geq \alpha \geq \gamma \geq 0$ and $1 \geq \beta \geq \delta
\geq 0$, by noting $\frac{t_1^{\alpha}+t_2^{\beta}}{t_1^{\gamma}+t_2^{\delta
}$ is bounded on $T^2$, we have
\begin{equation*}
\text{Lip}(\alpha,\beta,p)\subseteq \text{Lip}(\gamma,\delta,p) \subseteq
L_p \quad (p\geq 1).
\end{equation*}
If $f(x,y)$ is a continuous function periodic in both variables with period
2\pi$ and $p=\infty$, then the class $\text{Lip}(\alpha,\beta,p)$ reduces to
the H\"older class $\text{H}_{(\alpha,\beta)}$ (also called Lipschitz
class), that is
\begin{equation*}
\text{H}_{(\alpha,\beta)}=\left\{f: |f(x +t_1,y +t_2)-f(x,y)|=\mathcal{O
\left(t_1^{\alpha}\right)+\mathcal{O}\left(t_2^{\beta}\right)\right\}.
\end{equation*}
It is verified that $\text{H}_{(\alpha,\beta)}$ is a Banach space (see \cit
{UD2}) with the norm $\|f \|_{\alpha,\beta}$ defined by
\begin{equation*}
\|f \|_{\alpha,\beta}=\|f\|_{C}+\sup_{t_1\neq 0,\,\,t_2\neq 0 }\frac{|f(x
+t_1,y +t_2)-f(x,y)|}{|t_1|^{\alpha}+|t_2|^{\beta}},
\end{equation*}
where
\begin{equation*}
\|f\|_{C}=\sup_{(x,y)\in T^2}|f(x,y)|.
\end{equation*}
Let $f(x,y)\in L^p(T^2)$ be a $2\pi$-periodic function with respect to each
variable, with its Fourier series
\begin{align*}
& f(x,y)\sim
\sum_{k=0}^{\infty}\sum_{\ell=0}^{\infty}\lambda_{k,\ell}(a_{k,\ell} \cos kx
\cos \ell y+b_{k,\ell} \sin kx \cos \ell y \\
& \hspace{3cm}+c_{k,\ell} \cos kx \sin \ell y+d_{k,\ell} \sin kx \sin \ell
y)
\end{align*}
at the point $(x,y)$, where
\begin{align*}
\lambda _{k,\ell} & = \left\{
\begin{array}{rcl}
\frac{1}{4}, & \mbox{if} & k=\ell=0, \\
\frac{1}{2}, & \mbox{if} & k=0, \,\ell>0;\,\,\,\mbox{or} \,\,\, k>0,
\,\ell=0, \\
1, & \mbox{if} & k, \,\ell>0
\end{array
\right. \\
a_{k,\ell} & = \frac{1}{\pi^{2}}\int_{-\pi}^{\pi}\int_{-\pi}^{\pi}f(u, v)
\cos ku \cos \ell v \,du dv, \\
b_{k,\ell} & = \frac{1}{\pi^{2}}\int_{-\pi}^{\pi}\int_{-\pi}^{\pi}f(u, v)
\sin ku \cos \ell v \,du dv, \\
c_{k,\ell} & = \frac{1}{\pi^{2}}\int_{-\pi}^{\pi}\int_{-\pi}^{\pi}f(u, v)
\cos ku \sin \ell v \,du dv, \\
d_{k,\ell} & = \frac{1}{\pi^{2}}\int_{-\pi}^{\pi}\int_{-\pi}^{\pi}f(u,
v)\sin ku \sin \ell v\, du dv,\quad k,\,\ell\in \mathbb{N}\cup\{0\},
\end{align*}
and whose partial sums are
\begin{align*}
& S_{n,m}(f;x,y)= \sum_{k=0}^{n}\sum_{\ell=0}^{m}\lambda_{k,\ell}(a_{k,\ell}
\cos kx \cos \ell y+b_{k,\ell} \sin kx \cos \ell y \\
& \hspace{3.8cm}+c_{k,\ell} \cos kx \sin \ell y+d_{k,\ell} \sin kx \sin \ell
y),\quad n,m\geq 0 .
\end{align*}
To reveal our intention, we recall some other notations and notions.
Let $\sum_{k=0}^{\infty}\sum_{\ell=0}^{\infty}u_{k,\ell}$ be an infinite
double series with its sequence of arithmetic mean $\{\sigma_{m,n}\}$, where
\begin{equation*}
\sigma_{m,n}=\frac{1}{(m+1)(n+1)}\sum_{k=0}^{m}\sum_{\ell=0}^{n}U_{k,\ell}
\end{equation*}
and $U_{k,\ell}:=\sum_{i=0}^{k}\sum_{j=0}^{\ell}u_{i,j}.$
We define the \textit{Double Delayed Arithmetic Mean} $\sigma_{m,k;n,\ell}$
by (see \cite{FW}):
\begin{align} \label{eq1}
\begin{split}
\sigma_{m,k;n,\ell}:=&\left(1+\frac{m}{k}\right)\left(1+\frac{n}{\ell
\right)\sigma_{m+k-1,n+\ell-1}-\left(1+\frac{m}{k}\right)\frac{n}{\ell
\sigma_{m+k-1,n-1} \\
& -\frac{m}{k}\left(1+\frac{n}{\ell}\right)\sigma_{m-1,n+\ell-1}+\frac{mn}
k\ell}\sigma_{m-1,n-1},
\end{split}
\end{align}
where $k$ and $\ell$ are positive integers.
If $k$ tends to $\infty$ with $m$ in such a way that $\frac{m}{k}$ is
bounded, and $\ell$ tends to $\infty$ with $n$ in such a way that $\frac{n}
\ell}$ is also bounded, then $\sigma_{m,k;n,\ell}$ defines a method of
summability which is at least as strong as the well-known $(C,1,1)$
summablity. This means that if $\sigma_{m,n}\to \mu$, then
\sigma_{m,k;n,\ell}\to \mu$ as well. This important fact follows from (\re
{eq1}) if we set $\sigma_{m,n}=\mu+\xi_{m,n}$, where $\xi_{m,n}\to 0$ as
m,n\to \infty$. Introducing this mean we expect to be useful in
applications, particularly in approximating of $2\pi$ periodic functions in
two variables.
We note that for $k=\ell=1$ we obtain $\sigma_{m,1;n,1}=U_{m,n}$, while for
m=n=0$ we get $\sigma_{0,k;0,\ell}=\sigma_{k-1,\ell-1}$. Moreover, for $k=m$
and $\ell=n$, we get
\begin{align} \label{eq2}
\sigma_{m,m;n,n}=4\sigma_{2m-1,2n-1}-2\sigma_{2m-1,n-1}-2\sigma_{m-1,2n-1}
\sigma_{m-1,n-1}
\end{align}
that we name as the \textit{first type} Double Delayed Arithmetic Mean.
However, in this research paper we will take $k=2m$ and $\ell=2n$ in the
Double Delayed Arithmetic Mean $\sigma_{m,k;n,\ell}$ to obtain
\begin{align} \label{eq3}
\sigma_{m,2m;n,2n}=\frac{1}{4}\left(9\sigma_{3m-1,3n-1}-3\sigma_{3m-1,n-1}-
\sigma_{m-1,3n-1}+\sigma_{m-1,n-1}\right).
\end{align}
We name these particular sums as the \textit{second type} Double Delayed
Arithmetic Mean.
By $\sigma _{m,n}(f;x,y)$ and $\sigma _{m,2m;n,2n}(f;x,y)$ we denote the
arithmetic mean and the second type Double Delayed Arithmetic Mean for
S_{k,\ell }(f;x,y)$, respectively. It is well-known (see e.g. \cite[page 4
{KIM}) that the double Fej\'{e}r kernel is
\begin{equation*}
F_{m,n}(t_{1},t_{2}):=\frac{4}{(m+1)(n+1)}\left( \frac{\sin \frac{(m+1)t_{1
}{2}\sin \frac{(n+1)t_{2}}{2}}{4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}
\right) ^{2},
\end{equation*
which we will use it in its equivalent form
\begin{equation}\label{eq4}
F_{m,n}(t_{1},t_{2}):=\frac{1}{(m+1)(n+1)}\frac{(1-\cos (m+1)t_{1})(1-\cos
(n+1)t_{2})}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}},
\end{equation
while the arithmetic mean $\sigma _{m,n}(f;x,y)$ is
\begin{equation}\label{eq5}
\sigma _{m,n}(f;x,y)=\frac{1}{\pi ^{2}}\int_{0}^{\pi }\int_{0}^{\pi
}h_{x,y}(t_{1},t_{2})F_{m,n}(t_{1},t_{2})\,dt_{1}dt_{2},
\end{equation
where
\begin{equation*}
h_{x,y}(t_{1},t_{2}):=f(x+t_{1},y+t_{2})+f(x-t_{1},y+t_{2})+f(x+t_{1},y-t_{2})+f(x-t_{1},y-t_{2}).
\end{equation*
Furthermore, using (\ref{eq4}) successively, we have
\begin{align*}
F_{3m-1,3n-1}(t_{1},t_{2})& =\frac{(1-\cos (3mt_{1}))(1-\cos (3nt_{2}))}
9mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}, \\
F_{3m-1,n-1}(t_{1},t_{2})& =\frac{(1-\cos (3mt_{1}))(1-\cos (nt_{2}))}
3mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}, \\
F_{m-1,3n-1}(t_{1},t_{2})& =\frac{(1-\cos (mt_{1}))(1-\cos (3nt_{2}))}
3mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}, \\
F_{m-1,n-1}(t_{1},t_{2})& =\frac{(1-\cos (mt_{1}))(1-\cos (nt_{2}))}
mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}.
\end{align*
Thus, we can write
\begin{equation}
\begin{split}
F_{m,2m;n,2n}(t_{1},t_{2}):=& \frac{1}{4}\left(
9F_{3m-1,3n-1}(t_{1},t_{2})-3F_{3m-1,n-1}(t_{1},t_{2})\right. \\
& \quad \left. -3F_{m-1,3n-1}(t_{1},t_{2})+F_{m-1,n-1}(t_{1},t_{2})\right) \\
=& \frac{1}{4mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}
\left( (1-\cos (3mt_{1}))(1-\cos (3nt_{2}))\right. \\
& \left. -(1-\cos (3mt_{1}))(1-\cos (nt_{2}))-(1-\cos (mt_{1}))(1-\cos
(3nt_{2}))\right. \\
& \left. +(1-\cos (mt_{1}))(1-\cos (nt_{2}))\right) \\
=& \frac{1}{4mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}
\left( \cos (mt_{1})-\cos (3mt_{1})\right) \left( \cos (nt_{2})-\cos
(3nt_{2})\right) \\
=& \frac{S(t_{1},t_{2})}{mn\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2
\right) ^{2}},
\end{split}
\label{eq6}
\end{equation
wher
\begin{equation*}
S(t_{1},t_{2}):=\sin 2mt_{1}\sin mt_{1}\sin 2nt_{2}\sin nt_{2}.
\end{equation*}
Therefore, using (\ref{eq5}) and (\ref{eq6}), we get the second type Double
Delayed Arithmetic Mean as
\begin{equation*}
\sigma_{m,2m;n,2n}(f;x,y)=\frac{1}{mn\pi^2}\int_{0}^{\pi}\int_{0}^
\pi}h_{x,y}(t_1,t_2)\frac{S(t_1,t_2)}{\left(4\sin \frac{t_1}{2}\sin \frac{t_
}{2}\right)^2}dt_1dt_2.
\end{equation*}
Throughout this paper we agree to put: $h_1:=h_1(m)=\frac{\pi}{m}$,
h_2:=h_2(n)=\frac{\pi}{n}$,
\begin{align*}
\phi_{x,y}(t_1,t_2) & := \frac{1}{4} ( f(x+t_1,y+t_2) + f(x-t_1,y+t_2) \\
& \hspace{1cm} + f(x+t_1,y-t_2) + f(x-t_1,y-t_2) - 4 f(x,y) ), \\
H_{x,z_1,y,z_2}(t_1,t_2) & :=\phi_{x+z_1,y+z_2}(t_1,t_2)-\phi_{x,y}(t_1,t_2),
\end{align*}
and
\begin{align*}
D_{m,n}(x,y):=\sigma_{m,2m;n,2n}(f;x,y)-f(x,y).
\end{align*}
In order to achieve to our aim, we need some helpful lemmas given in next section.
\section{Auxiliary Results}
\begin{lemma}
(The generalized Minkowski inequality \cite[p. 21]{SMN})\label{le1} For a
function $g(x_{1},y_{1})$ given on a measurable set $E:=E_{1}\times
E_{2}\subset \mathbb{R}_{2}$, where $x_{1}=(x_{1},x_{2},\dots ,x_{m})$ and
y_{1}=(x_{m+1},x_{m+2},\dots ,x_{n})$, the following inequality holds:
\begin{equation*}
\left( \int_{E_{1}}\left\vert \int_{E_{2}}g(x_{1},y_{1})dy_{1}\right\vert
^{p}dx_{1}\right) ^{\frac{1}{p}}\leq \int_{E_{2}}\left(
\int_{E_{1}}\left\vert g(x_{1},y_{1})\right\vert ^{p}dy_{1}\right) ^{\frac{
}{p}}dx_{1},
\end{equation*
for those values of $p$ for which the right-hand side of this inequality is
finite.
\end{lemma}
\begin{lemma}
\label{le2} Let $\omega_1$, $\omega_2$, $v_1$ and $v_2$ be moduli of
continuity so that $\frac{\omega_1 (t_1)}{v_1(t_1)}$ is non-decreasing in
t_1$, $\frac{\omega_2 (t_2)}{v_2(t_2)}$ is non-decreasing in $t_2$, and
f\in H_{p}^{(\omega_1, \omega_2)}$. Then for $0< t_1\leq \pi$, $0< t_2\leq
\pi$, and $p\geq 1$,
\begin{enumerate}
\item[(i)] $\|\phi_{x,y}(t_1,t_2)\|_p=\mathcal{O}\left(\omega_1
(t_1)+\omega_2 (t_2)\right)$,
\item[(ii)] $\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p=\mathcal{O}\left(\omega_1
(t_1)+\omega_2 (t_2)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_2
(|z_2|)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p=\mathcal{O}\left(\omega_1
(t_1+h_1)+\omega_2 (t_2)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p=\mathcal{O}\left(\omega_1
(|z_1|)+\omega_2 (|z_2|)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p=\mathcal{O}\left(\omega_1 (t_1)+\omega_2
(t_2+h_2)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p=\mathcal{O}\left(\omega_1
(|z_1|)+\omega_2 (|z_2|)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p=\mathcal{O}\left(\omega_1
(t_1+h_1)+\omega_2 (t_2+h_2)\right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p=\mathcal{O}\left(\omega_1
(|z_1|)+\omega_2 (|z_2|)\right)$.
\end{enumerate}
Moreover, if $\omega_1 = \omega_2$ and $v_1=v_2$, then
\begin{enumerate}
\item[(iii)] $\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p=\mathcal{O}\left( (v_1 (|z_1|)
+ (v_1 (|z_2|)) \left( \frac{\omega_1 (t_1)}{v_1 (t_1)} + \frac{\omega_1
(t_2)}{v_1 (t_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p=\mathcal{O}\left( (v_1 (|z_1|) + (v_1
(|z_2|)) \left( \frac{\omega_1 (t_1+h_1)}{v_1 (t_1+h_1)} + \frac{\omega_1
(t_2)}{v_1 (t_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p=\mathcal{O}\left( (v_1 (|z_1|) + (v_1
(|z_2|)) \left( \frac{\omega_1 (t_1)}{v_1 (t_1)} + \frac{\omega_1 (t_2+h_2)}
v_1 (t_2+h_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p=\mathcal{O}\left( (v_1 (|z_1|) +
(v_1 (|z_2|)) \left( \frac{\omega_1 (t_1+h_1)}{v_1 (t_1+h_1)} + \frac
\omega_1 (t_2+h_2)}{v_1 (t_2+h_2)}\right) \right)$.
\item[(iv)] $\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p
\newline
=\mathcal{O}\left( (v_1 (|z_1|) + (v_1 (|z_2|)) \left( \frac{\omega_1 (h_1)}
v_1 (h_1)} + \frac{\omega_1 (h_2)}{v_1 (h_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p\newline
=\mathcal{O}\left( (v_1 (|z_1|) + (v_1 (|z_2|)) \left( \frac{\omega_1 (h_1)}
v_1 (h_1)} + \frac{\omega_1 (h_2)}{v_1 (h_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2+h_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p\newline
=\mathcal{O}\left( (v_1 (|z_1|) + (v_1 (|z_2|)) \left( \frac{\omega_1 (h_1)}
v_1 (h_1)} + \frac{\omega_1 (h_2)}{v_1 (h_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1+h_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p\newline
=\mathcal{O}\left( (v_1 (|z_1|) + (v_1 (|z_2|)) \left( \frac{\omega_1 (h_1)}
v_1 (h_1)} + \frac{\omega_1 (h_2)}{v_1 (h_2)}\right) \right)$, \newline
$\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2) -
H_{x,z_1,y,z_2}(t_1,t_2+h_2) + H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2) \|_p
\mathcal{O}\left( (v_1 (|z_1|) + (v_1 (|z_2|)) \left( \frac{\omega_1 (h_1)}
v_1 (h_1)} + \frac{\omega_1 (h_2)}{v_1 (h_2)}\right) \right)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Part (i). We calculate as
\begin{equation*}
\begin{split}
\|\phi_{x,y}(t_1,t_2)\|_p&\leq \frac{1}{4}\left\{\|f(x+t_1,y+t_2)-f(x,y
\|_p+\|f(x,y)-f(x-t_1,y+t_2)\|_p \right. \\
&\quad + \left.
\|f(x,y)-f(x+t_1,y-t_2)\|_p+\|f(x,y)-f(x-t_1,y-t_2)\|_p\right\} \\
&=\mathcal{O}\left(\omega_1 (t_1)+\omega_2 (t_2)\right).
\end{split
\end{equation*}
Part (ii). Since $f\in \text{Lip}(\omega_1,\omega_2,p)$, we have
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=
\|\phi_{x+z_1,y+z_2}(t_1,t_2)-\phi_{x,y}(t_1,t_2)\|_p \\
&\leq\frac{1}{4}\left\{\|f(x+z_1+t_1,y+z_2+t_2)-f(x+z_1,y+z_2)\|_p \right. \\
&\quad +\|f(x+z_1-t_1,y+z_2+t_2)-f(x+z_1,y+z_2)\|_p \\
&\quad +\|f(x+z_1+t_1,y+z_2-t_2)-f(x+z_1,y+z_2)\|_p \\
&\quad +\|f(x+z_1-t_1,y+z_2-t_2)-f(x+z_1,y+z_2)\|_p \\
&\quad +\|f(x+t_1,y+t_2)-f(x,y)\|_p \\
&\quad +\|f(x-t_1,y+t_2)-f(x,y)\|_p \\
&\quad +\|f(x+t_1,y-t_2)-f(x,y)\|_p \\
&\quad + \left. \|f(x-t_1,y-t_2)-f(x,y)\|_p \right\} \\
&=\mathcal{O}\left(\omega_1 (t_1)+\omega_2 (t_2)\right).
\end{split
\end{equation*}
For the second part, a similar reasoning yields
\begin{equation} \label{eq9}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=
\|\phi_{x+z_1,y+z_2}(t_1,t_2)-\phi_{x,y}(t_1,t_2)\|_p \\
&\leq\frac{1}{4}\left\{\|f(x+z_1+t_1,y+z_2+t_2)-f(x+t_1,y+t_2)\|_p \right. \\
&\quad +\|f(x+z_1-t_1,y+z_1+t_2)-f(x-t_1,y+t_2)\|_p \\
&\quad +\|f(x+z_1+t_1,y+z_2-t_2)-f(x+t_1,y-t_2)\|_p \\
&\quad + \left. \|f(x+z_1-t_1,y+z_2-t_2)-f(x-t_1,y-t_2)\|_p\right\} \\
&\quad +\|f(x+z_1,y+z_2)-f(x,y)\|_p \\
& =\mathcal{O}\left(\omega_1 (|z_1|)+\omega_2 (|z_2|)\right).
\end{split
\end{equation}
The other relations can be verified in a very same way. We omit their
proofs.
Part (iii). Using part (ii) and the fact $v_1(t_1)$ is non-decreasing, in
case of $t_1\leq |z_1|$ and $t_2\leq |z_2|$, we get
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=\mathcal{O}\left(\omega_1 (t_1)+\omega_1
(t_2)\right) \\
&=\mathcal{O}\left( v_1 (t_1) \frac{\omega_1 (t_1)}{v_1 (t_1)} + v_1 (t_2
\frac{\omega_1 (t_2)}{v_1 (t_2)}\right) \\
&=\mathcal{O}\left( v_1 (|z_1|) \frac{\omega_1 (t_1)}{v_1 (t_1)} + v_1
(|z_2|)\frac{\omega_1 (t_2)}{v_1 (t_2)}\right).
\end{split
\end{equation*}
Since $\frac{\omega_1 (t_1)}{v_1(t_1)}$ is non-decreasing, and we have the
second part of (ii), then for $t_1\geq |z_1|$ and $t_2\geq |z_2|$, we also
obtain
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_1
(|z_2|)\right) \\
&=\mathcal{O}\left(v_1 (|z_1|) \frac{\omega_1 (|z_1|)}{v_1 (|z_1|)} + v_1
(|z_2|)\frac{\omega_1 (|z_2|)}{v_1 (|z_2|)}\right) \\
&=\mathcal{O}\left(v_1 (|z_1|) \frac{\omega_1 (t_1)}{v_1 (t_1)} + v_1 (|z_2|
\frac{\omega_1 (t_2)}{v_1 (t_2)}\right).
\end{split
\end{equation*}
For $t_1\leq |z_1|$ and $t_2\geq |z_2|$, consider two possibilities. If
|z_1| \geq t_2$, then
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=\mathcal{O}\left(\omega_1 (t_1)+\omega_1
(t_2)\right) \\
&=\mathcal{O}\left( v_1 (t_1) \frac{\omega_1 (t_1)}{v_1 (t_1)} + v_1 (t_2)
\frac{\omega_1 (t_2)}{v_1 (t_2)} \right) \\
& = \mathcal{O}\left( v_1 (|z_1|) \left( \frac{\omega_1 (t_1)}{v_1 (t_1)} +
\frac{\omega_1 (t_2)}{v_1 (t_2)}\right) \right).
\end{split
\end{equation*}
If $t_2 \geq |z_1|$, then
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p&=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_1
(|z_2|)\right) \\
&=\mathcal{O}\left(v_1 (|z_1|) \frac{\omega_1 (|z_1|)}{v_1 (|z_1|)} + v_1
(|z_2|)\frac{\omega_1 (|z_2|)}{v_1 (|z_2|)}\right) \\
&=\mathcal{O}\left( (v_1 (|z_1|) + v_1 (|z_2|)) \frac{\omega_1 (t_2)}{v_1
(t_2)}\right).
\end{split
\end{equation*}
For $t_1\geq |z_1|$ and $t_2\leq |z_2|$, we can get
\begin{equation*}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p = \mathcal{O}\left( v_1 (|z_2|) \left( \frac
\omega_1 (t_1)}{v_1 (t_1)} + \frac{\omega_1 (t_2)}{v_1 (t_2)}\right) \right)
\end{equation*}
in case of $|z_2| \geq t_1$, and
\begin{equation*}
\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p=\mathcal{O}\left( (v_1 (|z_1|) + v_1 (|z_2|))
\frac{\omega_1 (t_1)}{v_1 (t_1)}\right)
\end{equation*}
in case of $t_1 \geq |z_2|$ similarly as before. This shows the validity of
the first inequality in part (iii). The other relations can be verified in a
very same way.
Part (iv). We have
\begin{equation*}
\begin{split}
&\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p \\
&\leq\|\phi_{x,y}(t_1+h_1,t_2)-\phi_{x,y}(t_1,t_2)\|_p \\
&\quad +\|\phi_{x+z_1,y+z_2}(t_1+h_1,t_2)-\phi_{x+z_1,y+z_2}(t_1,t_2)\|_p \\
&=\mathcal{O}\left(\omega_1 (h_1)\right) = \mathcal{O}\left(\omega_1 (h_1) +
\omega_1 (h_2)\right)
\end{split
\end{equation*}
and
\begin{equation*}
\begin{split}
&\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p \\
&\leq\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p + \|H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p \\
&=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_2 (|z_2|)\right),
\end{split
\end{equation*}
while
\begin{equation*}
\begin{split}
&\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p \\
&\leq\|\phi_{x,y}(t_1,t_2+h_2)-\phi_{x,y}(t_1,t_2)\|_p \\
&\quad +\|\phi_{x+z_1,y+z_2}(t_1,t_2+h_2)-\phi_{x+z_1,y+z_2}(t_1,t_2)\|_p \\
&=\mathcal{O}\left(\omega_1 (h_2)\right) = \mathcal{O}\left(\omega_1 (h_1) +
\omega_1 (h_2)\right)
\end{split
\end{equation*}
and
\begin{equation*}
\begin{split}
&\|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p \\
&\leq\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p + \|H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p \\
&=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_2 (|z_2|)\right),
\end{split
\end{equation*}
furthermore
\begin{equation*}
\begin{split}
\| H_{x,z_1,y,z_2}(t_1,t_2)- &H_{x,z_1,y,z_2}(t_1+h_1,t_2)-
H_{x,z_1,y,z_2}(t_1,t_2+h_2)\\
&\quad + H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2) \|_p \\
&\leq \|H_{x,z_1,y,z_2}(t_1,t_2)-H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p \\
&\quad + \| H_{x,z_1,y,z_2}(t_1,t_2+h_2) -
H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p \\
&= \mathcal{O}\left(\omega_1 (h_1) + \omega_1 (h_2)\right)
\end{split
\end{equation*}
and
\begin{equation*}
\begin{split}
\|H_{x,z_1,y,z_2}(t_1,t_2)&-H_{x,z_1,y,z_2}(t_1+h_1,t_2) -
H_{x,z_1,y,z_2}(t_1,t_2+h_2)\\
&\quad + H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p \\
&\leq\|H_{x,z_1,y,z_2}(t_1,t_2)\|_p + \|H_{x,z_1,y,z_2}(t_1+h_1,t_2)\|_p \\
&\quad + \| H_{x,z_1,y,z_2}(t_1,t_2+h_2)\|_p + \|
H_{x,z_1,y,z_2}(t_1+h_1,t_2+h_2)\|_p \\
&=\mathcal{O}\left(\omega_1 (|z_1|)+\omega_2 (|z_2|)\right).
\end{split
\end{equation*}
From analogous estimations, we can obtain part (iv) by considering the four
cases $h_1\leq |z_1|$ and $h_2\leq |z_2|$; $h_1\geq |z_1|$ and $h_2\geq
|z_2| $; $h_1\leq |z_1|$ and $h_2\geq |z_2|$; $h_1\geq |z_1|$ and $h_2\leq
|z_2|$, respectively, as in part (iii). We omit the details.
\end{proof}
\section{Main Results}
We prove the following statement.
\begin{theorem}
\label{the01} Let $\omega $ and $v$ be moduli of continuity so that $\frac
\omega (t)}{v(t)}$ is non-decreasing in $t$. If $f\in H_{p}^{(\omega ,\omega
)}$, $p\geq 1$, then
\begin{equation*}
\Vert \sigma _{m,2m;n,2n}(f)-f\Vert _{p}^{(v,v)}= \mathcal{O} \left( \frac{\omega (h_{1}
}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) ,
\end{equation*
where $h_{1}=\frac{\pi }{m}$, $h_{2}=\frac{\pi }{n}$ for $m,n\in
\mathbb{N}
$.
\end{theorem}
\allowdisplaybreaks{\begin{proof}
Using the equality
\begin{equation}
\int_{0}^{\pi }\int_{0}^{\pi }\frac{\sin 2mt_{1}\sin mt_{1}\sin 2nt_{2}\sin
nt_{2}}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2}=\frac{mn\pi ^{2}}{4} \label{eq11}
\end{equation
we have
\begin{equation}
\begin{split}
D_{m,n}(x,y)& =\sigma _{m,2m;n,2n}(f;x,y)-f(x,y) \\
& =\frac{4}{mn\pi ^{2}}\int_{0}^{\pi }\int_{0}^{\pi }\phi _{x,y}(t_{1},t_{2}
\frac{S(t_{1},t_{2})}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2
\right) ^{2}}dt_{1}dt_{2},
\end{split}
\label{eq12}
\end{equation
where
\begin{equation*}
\begin{split}
\phi _{x,y}(t_{1},t_{2})=& \frac{1}{4}\big
f(x+t_{1},y+t_{2})+f(x-t_{1},y+t_{2}) \\
& \quad +f(x+t_{1},y-t_{2})+f(x-t_{1},y-t_{2})-4f(x,y)\big]
\end{split
\end{equation*
and
\begin{equation*}
S(t_{1},t_{2})=\sin 2mt_{1}\sin mt_{1}\sin 2nt_{2}\sin nt_{2}.
\end{equation*}
By definition, we have
\begin{equation}
\Vert D_{m,n}\Vert _{p}^{(v,v)}:=\Vert D_{m,n}\Vert _{p}+\sup_{z_{1}\neq
0,\,\,z_{2}\neq 0}\frac{\Vert D_{m,n}(x+z_{1},y+z_{2})-D_{m,n}(x,y)\Vert _{p
}{v(|z_{1}|)+v(|z_{2}|)}. \label{eq13}
\end{equation
Now, we can write
\begin{equation}
\begin{split}
& D_{m,n}(x+z_{1},y+z_{2})-D_{m,n}(x,y) \\
& =\frac{4}{mn\pi ^{2}}\int_{0}^{\pi }\int_{0}^{\pi }\left[ \phi
_{x+z_{1},y+z_{2}}(t_{1},t_{2})-\phi _{x,y}(t_{1},t_{2})\right] \frac
S(t_{1},t_{2})}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2} \\
& =\frac{4}{mn\pi ^{2}}\int_{0}^{\pi }\int_{0}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}{\left( 4\sin \frac
t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& =\frac{4}{mn\pi ^{2}}\int_{0}^{h_{1}
\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}{\left(
4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}
\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{4}{mn\pi ^{2}}\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}{\left( 4\sin \frac
t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}{\left( 4\sin \frac
t_{1}}{2}\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2}:=\sum_{r=1}^{4}J_{r},
\end{split}
\label{eq14}
\end{equation}
Using Lemma \ref{le1} for $p\geq 1$, by the estimate
\begin{equation*}
\left\vert \frac{S(t_{1},t_{2})}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2
}{2}\right) ^{2}}\right\vert =\mathcal{O}\left( m^{2}n^{2}\right) ,\quad
0<t_{1}\leq \pi ,0<t_{2}\leq \pi ,
\end{equation*
and Lemma \ref{le2}, we have
\begin{equation}
\begin{split}
\Vert J_{1}\Vert _{p}& \leq \frac{4}{mn\pi ^{2}}\int_{0}^{h_{1}
\int_{0}^{h_{2}}\Vert H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\Vert _{p}\left\vert
\frac{S(t_{1},t_{2})}{\left( 4\sin \frac{t_{1}}{2}\sin \frac{t_{2}}{2
\right) ^{2}}\right\vert dt_{1}dt_{2} \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right)
mn\int_{0}^{h_{1}}\int_{0}^{h_{2}}\left( \frac{\omega (t_{1})}{v(t_{1})}
\frac{\omega (t_{2})}{v(t_{2})}\right) dt_{1}dt_{2}\right) \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) mn\left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\int_{0}^{h_{1}}\int_{0}^{h_{2}}dt_{1}dt_{2}\right) \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{split}
\label{eq15}
\end{equation}
The quantity $J_{2}$ can be written as follows:
\begin{equation}
\begin{split}
J_{2}& =\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\sin 2mt_{1}\sin mt_{1}\\%
&\quad \times \frac{\sin 2nt_{2}\sin nt_{2}}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
\left( \frac{1}{\left( 2\sin \frac{t_{1}}{2}\right) ^{2}}-\frac{1}{t_{1}^{2}
\right) dt_{1}dt_{2}\\
&\quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}
t_{1}^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2}:=J_{21}+J_{22}.
\end{split}
\label{eq16}
\end{equation
Since the function
\begin{equation*}
\frac{1}{\left( 2\sin \frac{t_{1}}{2}\right) ^{2}}-\frac{1}{t_{1}^{2}}
\end{equation*
is bounded for $0<t_{1}\leq \pi $ and
\begin{equation*}
\left\vert \frac{\sin 2nt_{2}\sin nt_{2}}{\left( 2\sin \frac{t_{2}}{2
\right) ^{2}}\right\vert =\mathcal{O}\left( n^{2}\right) ,\quad 0<t_{2}\leq
\pi ,
\end{equation*
then by Lemma \ref{le1} with $p\geq 1$ and Lemma \ref{le2}, we get
\begin{equation}
\begin{split}
\Vert J_{21}\Vert _{p}& =\mathcal{O}\left( \frac{n}{m}\right)
\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\Vert
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\Vert _{p}dt_{1}dt_{2} \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \frac{n}{m
\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\left( \frac{\omega (t_{1})}{v(t_{1})}
\frac{\omega (t_{2})}{v(t_{2})}\right) dt_{1}dt_{2}\right) \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \frac{n}{m}\left(
\frac{\omega (\pi )}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}dt_{1}dt_{2}\right) \\
& =\mathcal{O}\left( \frac{1}{m}\left( v(|z_{1}|)+v(|z_{2}|)\right) \left(
\frac{\omega (\pi )}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right)
.
\end{split}
\label{eq17}
\end{equation
It is clear that for $m\in
\mathbb{N}
\begin{equation*}
\frac{1}{m}\omega \left( \pi \right) \leq 2\omega \left( \frac{\pi }{m
\right) ,
\end{equation*
whence
\begin{equation*}
\left\Vert J_{21}\right\Vert _{p}=O\left( \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{equation*
Further, substituting $t_{1}+h_{1}$ in place of $t_{1}$ in
\begin{equation}
J_{22}=\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}
t_{1}^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \label{eq18}
\end{equation
we obtain
\begin{equation}
J_{22}=-\frac{4}{mn\pi ^{2}}\int_{0}^{\pi
-h_{1}}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac
S(t_{1},t_{2})}{(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2}, \label{eq19}
\end{equation
Hence, from (\ref{eq18}) and (\ref{eq19}) we get
\begin{equation*}
\begin{split}
J_{22}& =\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}
t_{1}^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{2}{mn\pi ^{2}}\int_{0}^{\pi
-h_{1}}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac
S(t_{1},t_{2})}{(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2}\\
& =\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_{1},t_{2})}
t_{1}^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{2}{mn\pi ^{2}}\int_{0}^{h_{1}
\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{2}{mn\pi ^{2}}\int_{\pi -h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2}
\end{split}
\end{equation*}
\begin{equation}
\begin{split}
& =\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\left( \frac
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})}{t_{1}^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{(t_{1}+h_{1})^{2}}\right) \frac{S(t_{1},t_{2})}{\left( 2\sin \frac{t_{2}}{2
\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{2}{mn\pi ^{2}}\int_{0}^{h_{1}
\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{2}{mn\pi ^{2}}\int_{\pi -h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
(t_{1}+h_{1})^{2}\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
dt_{1}dt_{2}:=\sum_{s=1}^{3}J_{22}^{(s)}.
\end{split}
\label{eq20}
\end{equation
Using the inequalities $\frac{2}{\pi }\beta \leq \sin \beta $ for $\beta \in
(0,\pi /2)$, $\sin \beta \leq \beta $ for $\beta \in (0,\pi )$, Lemma \re
{le1} and Lemma \ref{le2}, we have
\begin{equation}
\begin{split}
\Vert J_{22}^{(2)}\Vert _{p}& =\mathcal{O}\left( \frac{1}{mn}\right)
\int_{0}^{h_{1}}\int_{0}^{h_{2}}\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\Vert _{p}\frac{(mt_{1}nt_{2})^{2}}
(t_{1}+h_{1})^{2}\left( \frac{t_{2}}{\pi }\right) ^{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( mn\right)
\int_{0}^{h_{1}}\int_{0}^{h_{2}}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega
(t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
dt_{1}dt_{2} \\
& =\mathcal{O}\left( mn\right) (v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega
(2h_{1})}{v(2h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\int_{0}^{h_{1}}\int_{0}^{h_{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( mn\right) (v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) h_{1}h_{2} \\
& =\mathcal{O}\left( (v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{split}
\label{eq21}
\end{equation
Moreover, the inequalities $|\sin \beta |\leq 1$ for $\beta \in (\pi
-h_{1},\pi )$, $\sin \beta \leq \beta $ for $\beta \in (0,\pi )$, Lemma \re
{le1}, and Lemma \ref{le2} imply
\begin{equation}
\begin{split}
\Vert J_{22}^{(3)}\Vert _{p}& =\mathcal{O}\left( \frac{1}{mn}\right)
\int_{\pi -h_{1}}^{\pi }\int_{0}^{h_{2}}\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\Vert _{p}\frac{(nt_{2})^{2}}
(t_{1}+h_{1})^{2}\left( \frac{t_{2}}{\pi }\right) ^{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( \frac{n}{m}\right) \int_{\pi -h_{1}}^{\pi
}\!\int_{0}^{h_{2}}(v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2})}{v(t_{2})}\right) \frac{1}
(t_{1}+h_{1})^{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( \frac{n}{m}\right) (v(|z_{1}|)+v(|z_{2}|))\int_{\pi
}^{\pi +h_{1}}\int_{0}^{h_{2}}\left( \frac{\omega (\theta _{1})}{v(\theta
_{1})}+\frac{\omega (t_{2})}{v(t_{2})}\right) \frac{1}{\theta _{1}^{2}
d\theta _{1}dt_{2} \\
& =\mathcal{O}\left( \frac{n}{m}\right) (v(|z_{1}|)+v(|z_{2}|))\left( \frac
\omega (\pi +h_{1})}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\int_{\pi }^{\pi +h_{1}}\int_{0}^{h_{2}}\frac{d\theta _{1}dt_{2}}{\theta
_{1}^{2}} \\
& =\mathcal{O}\left( \frac{n}{m}\right) (v(|z_{1}|)+v(|z_{2}|))\left( \frac
\omega (\pi )+\omega (h_{1})}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})
\right) \frac{h_{1}h_{2}}{\pi (\pi +h_{1})} \\
& =\mathcal{O}\left( \frac{1}{m^{2}}\right) (v(|z_{1}|)+v(|z_{2}|))\left(
\frac{\omega (\pi )}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})}\right) \\
& =\mathcal{O}\left( \frac{1}{m}(v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{split}
\label{eq22}
\end{equation
For $J_{22}^{(1)}$ we can write
\begin{equation}
\begin{split}
J_{22}^{(1)}& =\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\Bigg
\frac{H_{x,z_{1},y,z_{2}}(t_{1},t_{2})}{t_{1}^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{t_{1}^{2}} \\
& \qquad +\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{t_{1}^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{(t_{1}+h_{1})^{2}}\Bigg) \\
& \qquad \times \frac{S(t_{1},t_{2})}{\left( 2\sin \frac{t_{2}}{2}\right)
^{2}}dt_{1}dt_{2} \\
& =\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\left(
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}
\right) \\
& \qquad \times \frac{S(t_{1},t_{2})}{t_{1}^{2}\left( 2\sin \frac{t_{2}}{2
\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}
\left( 2\sin \frac{t_{2}}{2}\right) ^{2}} \\
& \qquad \times \left( \frac{1}{t_{1}^{2}}-\frac{1}{(t_{1}+h_{1})^{2}
\right) dt_{1}dt_{2}:=J_{221}^{(1)}+J_{222}^{(1)}.
\end{split}
\label{eq23}
\end{equation
Then, using Lemma \ref{le1}, and Lemma \ref{le2}, we have
\begin{equation}
\begin{split}
\Vert J_{221}^{(1)}\Vert _{p}& =\mathcal{O}\left( \frac{1}{mn}\right)
\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}
\right\Vert _{p}\frac{(nt_{2})^{2}}{t_{1}^{2}\left( 2\frac{t_{2}}{\pi
\right) ^{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( \frac{n}{m}\right) \int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}(v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1}
}+\frac{\omega (h_{2})}{v(h_{2})}\right) \frac{dt_{1}dt_{2}}{t_{1}^{2}} \\
& =\mathcal{O}\left( \frac{n}{m}\right) (v(|z_{1}|)+v(|z_{2}|))\left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \frac{h_{2
}{h_{1}} \\
& =\mathcal{O}\left( (v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{split}
\label{eq24}
\end{equation
With similar reasoning we obtain
\begin{equation}
\begin{split}
\Vert J_{222}^{(1)}\Vert _{p}& =\mathcal{O}\left( \frac{1}{mn}\right)
\int_{h_{1}}^{\pi }\int_{0}^{h_{2}}\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\right\Vert _{p}\frac{(nt_{2})^{2}}
\left( 2\frac{t_{2}}{\pi }\right) ^{2}}\frac{h_{1}(2t_{1}+h_{1})}
t_{1}^{2}(t_{1}+h_{1})^{2}}dt_{1}dt_{2} \\
& =\mathcal{O}\left( \frac{n}{m}\right) \int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2})}{v(t_{2})}\right) \frac
h_{1}dt_{1}dt_{2}}{t_{1}^{2}(t_{1}+h_{1})} \\
& =\mathcal{O}\left( \frac{n}{m^{2}}\right)
(v(|z_{1}|)+v(|z_{2}|))\int_{h_{1}}^{\pi }\left( \frac{\omega (t_{1}+h_{1})}
(t_{1}+h_{1})v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})t_{1}}\right) \frac
h_{2}dt_{1}}{t_{1}^{2}} \\
& =\mathcal{O}\left( \frac{1}{m^{2}}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (2h_{1})}{2h_{1}v(h_{1})
\int\limits_{h_{1}}^{\pi }\frac{dt_{1}}{t_{1}^{2}}+\frac{\omega (h_{2})}
v(h_{2})}\int\limits_{h_{1}}^{\pi }\frac{dt_{1}}{t_{1}^{3}}\right) \\
& =\mathcal{O}\left( \frac{1}{m^{2}}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( m\frac{\omega (h_{1})}{2h_{1}v(t_{1})
+m^{2}\frac{\omega (h_{2})}{v(h_{2})}\right) \\
& =\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{split}
\label{eq25}
\end{equation
So, from (\ref{eq23}), (\ref{eq24}) and (\ref{eq25}), we have
\begin{equation}
\Vert J_{22}^{(1)}\Vert _{p}=\mathcal{O}\left( (v(|z_{1}|)+v(|z_{2}|))\left(
\frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\right) . \label{eq26}
\end{equation
Now, taking into account (\ref{eq20}), (\ref{eq21}), (\ref{eq22}) and (\re
{eq26}), we have
\begin{equation}
\Vert J_{22}\Vert _{p}=\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})
\right) \right) . \label{eq27}
\end{equation
Whence, using (\ref{eq16}), (\ref{eq17}) and (\ref{eq27}), we obtain
\begin{equation}
\Vert J_{2}\Vert _{p}=\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})
\right) \right) . \label{eq28}
\end{equation}
By analogy, we conclude that
\begin{equation}
\Vert J_{3}\Vert _{p}=\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})
\right) \right) . \label{eq29}
\end{equation}
Finally, let us estimate the quantity $J_{4}$. Indeed, $J_{4}$ can be
rewritten as
\begin{align*}
J_{4}& = \frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_1,t_2) \\
&\qquad \times \left( \frac{1}{\left( 2\sin \frac{t_{1}}{2}\right) ^{2}}-\frac{1}
t_{1}{}^{2}}\right) \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
-\frac{1}{t_{2}{}^{2}}\right) dt_{1}dt_{2}\\
&\quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_1,t_2) \\
& \qquad \times \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{t_{1}{}^{2}}dt_{1}dt_{2}\\
& \quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_1,t_2) \\
& \qquad \times \left( \frac{1}{\left( 2\sin \frac{t_{1}}{2}\right) ^{2}}-\frac{1}
t_{1}{}^{2}}\right) \frac{1}{t_{2}{}^{2}}dt_{1}dt_{2}\\
& \quad +\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_1,t_2)}{\left( t_{1}t_{2}\right)
^{2}}dt_{1}dt_{2} \\
&:= J_{41}+J_{42}+J_{43}+J_{44}.
\end{align*
The boundedness of the function
\begin{equation*}
\left( \frac{1}{\left( 2\sin \frac{t_{1}}{2}\right) ^{2}}-\frac{1}
t_{1}{}^{2}}\right) \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}
-\frac{1}{t_{2}{}^{2}}\right)
\end{equation*
for $0<t_{1},t_{2}\leq \pi $, Lemma \ref{le1} and Lemma \ref{le2}, implies
\begin{align*}
\Vert J_{41}\Vert _{p}&=\mathcal{O}\left( \frac{1}{mn}\right)
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Vert
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\Vert _{p}dt_{1}dt_{2}\\
&=\mathcal{O}\left( \frac{1}{mn}\right) \int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}(v(|z_{1}|)+v(|z_{2}|))\left( \frac{\omega (t_{1})}{v(t_{1})}+\frac{\omega
(t_{2})}{v(t_{2})}\right) dt_{1}dt_{2}\\
&=\mathcal{O}\left( \frac{1}{mn}\right) (v(|z_{1}|)+v(|z_{2}|))\frac{\omega
\left( \pi \right) }{v\left( \pi \right) } \\
&=\mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{align*
Substituting $t_{1}$ with $t_{1}+h_{1}$ in
\begin{align*}
&&\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_1,t_2) \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{t_{1}{}^{2}}dt_{1}dt_{2}
\end{align*
we obtain
\begin{align*}
J_{42} &= -\frac{4}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})S(t_1,t_2) \\
& \qquad \times \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{\left( t_{1}+h_{1}\right) {}^{2}}dt_{1}dt_{2}.
\end{align*
Hence
\begin{align*}
J_{42} &= \frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_{1},t_{2}) \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{t_{1}^{2}}dt_{1}dt_{2}\\
& \quad -\frac{2}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})S(t_{1},t_{2}) \\
& \qquad \times \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{\left( t_{1}+h_{1}\right) {}^{2}}dt_{1}dt_{2}\\
&= \frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})S(t_{1},t_{2}) \left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}
t_{2}{}^{2}}\right) \frac{1}{t_{1}^{2}}dt_{1}dt_{2}\\
& \quad -\frac{2}{mn\pi ^{2}}\left( \int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}+\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }-\int_{\pi -h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\right) H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}) \\
& \qquad \times S(t_{1},t_{2})\left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right)
^{2}}-\frac{1}{t_{2}{}^{2}}\right) \frac{1}{\left( t_{1}+h_{1}\right) {}^{2}
dt_{1}dt_{2}\\
&= \frac{2}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\left( \frac
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})}{t_{1}{}^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{\left( t_{1}+h_{1}\right) {}^{2}
\right) \\
& \qquad \times S(t_{1},t_{2})\left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right)
^{2}}-\frac{1}{t_{2}{}^{2}}\right) \frac{1}{t_{1}{}^{2}}dt_{1}dt_{2}\\
& \quad -\frac{2}{mn\pi ^{2}}\left( \int_{0}^{h_{1}}\int_{h_{2}}^{\pi }-\int_{\pi
-h_{1}}^{\pi }\int_{h_{2}}^{\pi }\right)
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}) \\
& \qquad \times S(t_{1},t_{2})\left( \frac{1}{\left( 2\sin \frac{t_{2}}{2}\right)
^{2}}-\frac{1}{t_{2}{}^{2}}\right) \frac{1}{\left( t_{1}+h_{1}\right) {}^{2}
dt_{1}dt_{2} \\
&:= \sum\limits_{s=1}^{3}J_{42}^{\left( s\right) }.
\end{align*
Since the function
\begin{equation*}
\frac{1}{\left( 2\sin \frac{t_{2}}{2}\right) ^{2}}-\frac{1}{t_{2}^{2}}
\end{equation*
is bounded for $0<t_{1}\leq \pi $, using Lemma \ref{le1} with $p\geq 1$ and
Lemma \ref{le2}, we hav
\begin{align*}
\left\Vert J_{42}^{\left( 2\right) }\right\Vert _{p} &= \mathcal{O}\left(
\frac{1}{mn}\right) \int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2})}{v(t_{2})}\right) \frac
m^{2}t_{1}^{2}dt_{1}dt_{2}}{(t_{1}+h_{1})^{2}} \\
&= \mathcal{O}\left( \frac{m}{n}\right) \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (2h_{1})}{v(2h_{1})}+\frac{\omega (\pi )}{v(\pi )
\right) h_{1} \\
&= \mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) ,
\end{align*
\begin{align*}
\left\Vert J_{42}^{\left( 3\right) }\right\Vert _{p} &= \mathcal{O}\left(
\frac{1}{mn}\right) \int_{\pi -h_{1}}^{\pi }\int_{h_{2}}^{\pi }\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2})}{v(t_{2})}\right) \frac{dt_{1}dt_{2}}
(t_{1}+h_{1})^{2}} \\
&= \mathcal{O}\left( \frac{1}{mn}\right) \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (\pi )}{v(\pi )}\right) h_{1} \\
&= \mathcal{O}\left( \frac{1}{m}\left( v(|z_{1}|)+v(|z_{2}|)\right) \left(
\frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\right)
\end{align*
an
\begin{align*}
\left\Vert J_{42}^{\left( 1\right) }\right\Vert _{p} &= \mathcal{O}\left(
\frac{1}{mn}\right) \int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}
\right\Vert _{p}\frac{1}{t_{1}^{2}}dt_{1}dt_{2} \\
& \quad +\mathcal{O}\left( \frac{1}{mn}\right) \int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\right\Vert _{p}\frac
h_{1}(2t_{1}+h_{1})}{t_{1}^{2}\left( t_{1}+h_{1}\right) ^{2}}dt_{1}dt_{2} \\
&= \mathcal{O}\left( \frac{1}{mn}\right) \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})
\right) \int_{h_{1}}^{\pi }\frac{1}{t_{1}^{2}}dt_{1} \\
& \quad +\mathcal{O}\left( \frac{1}{m^{2}n}\right) \int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\left( v(|z_{1}|)+v(|z_{2}|)\right) \\
& \qquad \times \left( \frac{\omega (t_{1}+h_{1})}{\left( t_{1}+h_{1}\right)
v(t_{1}+h_{1})}+\frac{\omega (t_{2})}{t_{1}v(t_{2})}\right) \frac{1}
t_{1}^{2}}dt_{1}dt_{2} \\
&= \mathcal{O}\left( \frac{1}{n}\right) \left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\\
& \quad +\mathcal{O}\left( \frac{1}{m^{2}n}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (2h_{1})}{2h_{1}v(h_{1})
\int_{h_{1}}^{\pi }\frac{1}{t_{1}^{2}}dt_{1}+\frac{\omega (\pi )}{v(\pi )
\int_{h_{1}}^{\pi }\frac{1}{t_{1}^{3}}dt_{1}\right) \\
&= \mathcal{O}\left( \left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \right) .
\end{align*
The quantity $J_{43}$ can be estimated analogously to $J_{42}.$
Substituting in
\begin{equation*}
J_{44}=\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_1,t_2)}{\left( t_{1}t_{2}\right)
^{2}}dt_{1}dt_{2}
\end{equation*
$t_{1}$ with $t_{1}+h_{1}$, $t_{2}$ with $t_{2}+h_{2}$ and $t_{1}$ with
t_{1}+h_{1}$, $t_{2}$ with $t_{2}+h_{2}$, respectively, we get
\begin{equation*}
J_{44}=-\frac{4}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2},
\end{equation*
\begin{equation*}
J_{44}=-\frac{4}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2},
\end{equation*
\begin{equation*}
J_{44}=\frac{4}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}.
\end{equation*}
Henc
\begin{align*}
J_{44}& =\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_1,t_2)}{\left( t_{1}t_{2}\right)
^{2}}dt_{1}dt_{2}\\
& \quad -\frac{1}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2}\\
& \quad -\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\\
& \quad +\frac{1}{mn\pi ^{2}}\int_{0}^{\pi -h_{1}}\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\\
& =\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1},t_{2})\frac{S(t_1,t_2)}{\left( t_{1}t_{2}\right)
^{2}}dt_{1}dt_{2}\\
& \quad -\frac{1}{mn\pi ^{2}}\left( \int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}+\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }-\int_{\pi -h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\right) \\
& \quad \qquad H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{1}{mn\pi ^{2}}\left( \int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}+\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }-\int_{h_{1}}^{\pi
}\int_{\pi -h_{2}}^{\pi }\right) \\
& \quad \qquad H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\\
& \quad +\frac{1}{mn\pi ^{2}}\left(
\int_{0}^{h_{1}}\int_{0}^{h_{2}}+\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}-\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi }+\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}+\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\right.\\
& \quad \qquad \qquad \left. -\int_{h_{1}}^{\pi }\int_{\pi -h_{2}}^{\pi }-\int_{\pi -h_{1}}^{\pi
}\int_{0}^{h_{2}}-\int_{\pi -h_{1}}^{\pi }\int_{h_{2}}^{\pi }+\int_{\pi
-h_{1}}^{\pi }\int_{\pi -h_{2}}^{\pi }\right)\\
& \quad \qquad \qquad \quad H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\\
& =\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\left( \frac
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})}{\left( t_{1}t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2
}\right.\\
& \quad \qquad \quad \left. -\frac{H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}+\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\right) S(t_1,t_2) dt_{1}dt_{2}\\
& \quad -\frac{1}{mn\pi ^{2}}\left( \int_{0}^{h_{1}}\int_{h_{2}}^{\pi }-\int_{\pi
-h_{1}}^{\pi }\int_{h_{2}}^{\pi }\right)
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2}\\
& \quad -\frac{1}{mn\pi ^{2}}\left( \int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}-\int_{h_{1}}^{\pi }\int_{\pi -h_{2}}^{\pi }\right)
H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\\
& \quad +\frac{1}{mn\pi ^{2}}\left(
\int_{0}^{h_{1}}\int_{0}^{h_{2}}+\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}-\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi }+\int_{h_{1}}^{\pi
}\int_{0}^{h_{2}}\right.\\
& \quad \qquad \qquad \left. -\int_{h_{1}}^{\pi }\int_{\pi -h_{2}}^{\pi }-\int_{\pi -h_{1}}^{\pi
}\int_{0}^{h_{2}}-\int_{\pi -h_{1}}^{\pi }\int_{h_{2}}^{\pi }+\int_{\pi
-h_{1}}^{\pi }\int_{\pi -h_{2}}^{\pi }\right)\\
& \quad \qquad \qquad \quad H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_1,t_2)}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
& :=\sum \limits_{l=1}^{13}J_{44}^{\left( l\right) }.
\end{align*}
We start with the estimate of $J_{44}^{\left( 1\right) }$.
Indeed, we have
\begin{align*}
J_{44}^{\left( 1\right) } &= \frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\Bigg(\frac{H_{x,z_{1},y,z_{2}}(t_{1},t_{2})}{\left(
t_{1}t_{2}\right) ^{2}}-\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}} \\
& \quad \qquad -\frac{H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}+\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\Bigg)S(t_{1},t_{2})dt_{1}dt_{2} \\
&= \frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Bigg
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}) \\
&\quad \qquad -H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})+H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\Bigg)\frac{S(t_{1},t_{2})dt_{1}dt_{2}}{\left( t_{1}t_{2}\right) ^{2}} \\
&\quad +\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Bigg
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\Bigg) \\
&\qquad \times \Bigg(\frac{1}{\left( t_{1}t_{2}\right) ^{2}}-\frac{1}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}\Bigg)S(t_{1},t_{2})dt_{1}dt_{2} \\
&\quad +\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Bigg
H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\Bigg) \\
&\qquad \times \Bigg(\frac{1}{\left( t_{1}t_{2}\right) ^{2}}-\frac{1}{\left(
t_{1}(t_{2}+h_{2})\right) ^{2}}\Bigg)S(t_{1},t_{2})dt_{1}dt_{2} \\
&\quad +\frac{1}{mn\pi ^{2}}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\Bigg(\frac{1}{(t_{1}+h_{1})^{2
}-\frac{1}{t_{1}^{2}}\Bigg) \\
&\qquad \times \Bigg(\frac{1}{(t_{2}+h_{2})^{2}}-\frac{1}{t_{2}^{2}}\Bigg
S(t_{1},t_{2})dt_{1}dt_{2} \\
&:= J_{44}^{\left( 1,1\right) }+J_{44}^{\left( 1,2\right) }+J_{44}^{\left(
1,3\right) }+J_{44}^{\left( 1,4\right) }.
\end{align*}
It follows from Lemma \ref{le2} (iv) and $|S(t_{1},t_{2})|\leq 1$ for
t_{1}\in (h_{1},\pi )$, $t_{2}\in (h_{2},\pi )$ that
\begin{align*}
\Vert J_{44}^{\left( 1,1\right) }\Vert _{p} &= \frac{\mathcal{O}(1)}{mn
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Big\
H_{x,z_{1},y,z_{2}}(t_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}) \\
& \quad -H_{x,z_{1},y,z_{2}}(t_{1},t_{2}+h_{2})+H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\Big\|_{p}\frac{|S(t_{1},t_{2})|dt_{1}dt_{2}}{\left( t_{1}t_{2}\right) ^{2}}
\\
&= \frac{\mathcal{O}(1)}{mn}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\frac{dt_{1}dt_{2}}{\left( t_{1}t_{2}\right) ^{2}} \\
&= \mathcal{O}(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Moreover, from Lemma \ref{le2} (iii) and using the same arguments as above,
we obtain
\begin{align*}
\Vert J_{44}^{\left( 1,2\right) }\Vert _{p} &= \frac{\mathcal{O}(1)}{mn
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Big\
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\Big\|_{p} \\
& \quad \times \frac{1}{t_{2}^{2}}\Bigg(\frac{1}{t_{1}^{2}}-\frac{1}{\left(
t_{1}+h_{1}\right) ^{2}}\Bigg)|S(t_{1},t_{2})|dt_{1}dt_{2} \\
&= \frac{\mathcal{O}(1)}{mn}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\frac{h_{1}\left( 2t_{1}+h_{1}\right) }
t_{2}^{2}t_{1}^{2}\left( t_{1}+h_{1}\right) ^{2}}dt_{1}dt_{2} \\
&= \frac{\mathcal{O}(1)}{m^{2}n}\left( (v(|z_{1}|)+(v(|z_{2}|)\right) \left(
\frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\frac{1}{t_{2}^{2}t_{1}^{3}
dt_{1}dt_{2} \\
&= \mathcal{O}(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Similarly, we obtain
\begin{equation*}
\Vert J_{44}^{\left( 1,3\right) }\Vert _{p}=\mathcal{O
(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) .
\end{equation*}
To estimate $\Vert J_{44}^{\left( 1,4\right) }\Vert _{p}$ we use Lemma \re
{le1} and Lemma \ref{le2} (iii) in order to get
\begin{align*}
\Vert J_{44}^{\left( 1,4\right) }\Vert _{p} &= \frac{\mathcal{O}(1)}{mn
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Big\
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\Big\|_{p} \\
& \quad \times \Bigg(\frac{1}{(t_{1}+h_{1})^{2}}-\frac{1}{t_{1}^{2}}\Bigg)\Bigg
\frac{1}{(t_{2}+h_{2})^{2}}-\frac{1}{t_{2}^{2}}\Bigg
|S(t_{1},t_{2})|dt_{1}dt_{2} \\
&= \frac{\mathcal{O}(1)}{mn}\left( (v(|z_{1}|)+(v(|z_{2}|)\right)
\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Bigg(\frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2}+h_{2})}{v(t_{2}+h_{2})}\Bigg) \\
& \quad \times \frac{h_{1}h_{2}\left( 2t_{1}+h_{1}\right) \left(
2t_{2}+h_{2}\right) }{\left( t_{1}t_{2}\left( t_{1}+h_{1}\right) \left(
t_{2}+h_{2}\right) \right) ^{2}}dt_{1}dt_{2} \\
&= \frac{\mathcal{O}(1)}{\left( mn\right) ^{2}}\left(
(v(|z_{1}|)+(v(|z_{2}|)\right) \\
& \quad \times \int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\Bigg(\frac{\omega
(t_{1}+h_{1})}{\left( t_{1}+h_{1}\right) v(h_{1})t_{2}^{3}t_{1}^{2}}+\frac
\omega (t_{2}+h_{2})}{\left( t_{2}+h_{2}\right) v(h_{2})t_{1}^{3}t_{2}^{2}
\Bigg)dt_{1}dt_{2} \\
&= \frac{\mathcal{O}(1)}{\left( mn\right) ^{2}}\left(
(v(|z_{1}|)+(v(|z_{2}|)\right) \\
& \quad \times \Bigg(\frac{\omega (2h_{1})}{2h_{1}v(h_{1})}\int_{h_{1}}^{\pi
}\int_{h_{2}}^{\pi }\frac{1}{t_{2}^{3}t_{1}^{2}}dt_{1}dt_{2}+\frac{\omega
(2h_{2})}{2h_{2}v(h_{2})}\int_{h_{1}}^{\pi }\int_{h_{2}}^{\pi }\frac{1}
t_{2}^{2}t_{1}^{3}}dt_{1}dt_{2}\Bigg) \\
&= \mathcal{O}(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Whence, using the above estimates, we have
\begin{equation}
\Vert J_{44}^{\left( 1\right) }\Vert _{p}=\mathcal{O
(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb1}
\end{equation}
Replacing $t_{2}$ with $t_{2}+h_{2}$ in
\begin{equation*}
J_{44}^{\left( 2\right) }=\frac{1}{mn\pi ^{2}}\int_{0}^{h_{1}}\int_{h_{2}}^
\pi }H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2}
\end{equation*
we get
\begin{equation*}
J_{44}^{\left( 2\right) }=-\frac{1}{mn\pi ^{2}}\int_{0}^{h_{1}}\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}
\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}
\end{equation*
and after adding them we obtain
\begin{align*}
J_{44}^{\left( 2\right) } &= \frac{1}{2mn\pi ^{2}}\bigg(\int_{0}^{h_{1}
\int_{h_{2}}^{\pi }H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac{S(t_{1},t_{2}
}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\int_{0}^{h_{1}}\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}
\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\bigg) \\
&= \frac{1}{2mn\pi ^{2}}\bigg(\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\bigg(\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2
}-\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\bigg) \\
& \qquad \times
S(t_{1},t_{2})dt_{1}dt_{2}\\
& \quad -\int_{0}^{h_{1}} \int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
& \quad +\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\bigg) \\
&:= \sum\limits_{s=1}^{3}J_{44s}^{\left( 2\right) }.
\end{align*}
Since
\begin{equation*}
|\sin 2mt_{1}\sin mt_{1}\sin 2nt_{2}\sin nt_{2}|\leq
4(mnt_{1}t_{2})^{2}\quad \text{for}\quad 0<t_{1}<\pi ,0<t_{2}<\pi ,
\end{equation*
then applying Lemma \ref{le1} and Lemma \ref{le2} (iii) we have
\begin{align*}
\Vert J_{442}^{\left( 2\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{0}^{h_{1}}\int_{0}^{h_{2}}\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
(mnt_{1}t_{2})^{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}
dt_{1}dt_{2} \\
&= \mathcal{O}(mn)\int_{0}^{h_{1}}\int_{0}^{h_{2}}\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2}+h_{2})}{v(t_{2}+h_{2})}\right)
dt_{1}dt_{2} \\
&= \mathcal{O}(mn)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(2h_{1})}{v(2h_{1})}+\frac{\omega (2h_{2})}{v(2h_{2})}\right) h_{1}h_{2} \\
&= \mathcal{O}(1)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Because of
\begin{equation*}
|\sin 2mt_{1}\sin mt_{1}|\leq 2(mt_{1})^{2}\quad \text{for}\quad 0<t_{1}<\pi
\end{equation*
and
\begin{equation*}
|\sin 2nt_{2}\sin nt_{2}|\leq 1\quad \text{for}\quad \pi -h_{2}<t_{2}<\pi ,
\end{equation*
it follows that
\begin{align*}
\Vert J_{443}^{\left( 2\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
(mt_{1})^{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{m}{n}\right)\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi }\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}+\frac{\omega (t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac
dt_{1}dt_{2}}{(t_{2}+h_{2})^{2}} \\
&= \mathcal{O}\left(\frac{m}{n}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (2h_{1})}{v(2h_{1})}h_{1}h_{2}+h_{1}\int_{\pi }^{\pi +h_{2}}\frac
\omega (\theta _{2})}{v(\theta _{2})}\frac{d\theta _{2}}{\theta _{2}^2}\right)
\\
&= \mathcal{O}\left(\frac{m}{n}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (h_{1})}{v(h_{1})}h_{1}h_{2}+\frac{\omega (\pi +h_{2})}{v(\pi +h_{2})
\frac{h_{1}h_{2}}{\pi (\pi +h_{2})}\right) \\
&= \mathcal{O}\left( \frac{1}{n^{2}}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (\pi )}{v(\pi )}\right) \\
&= \mathcal{O}\left(\frac{1}{n}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Now we write
\begin{align*}
J_{441}^{\left( 2\right) } &= \frac{1}{2mn\pi ^{2}}\int_{0}^{h_{1}
\int_{h_{2}}^{\pi }\bigg(\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}
\left( (t_{1}+h_{1})t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}} \\
& \quad +\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\bigg)S(t_{1},t_{2})dt_{1}dt_{2} \\
&= \frac{1}{2mn\pi ^{2}}\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\bigg
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\bigg) \\
&\qquad \times \frac{S(t_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}
dt_{1}dt_{2} \\
&\quad +\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}{\left(
t_{1}+h_{1}\right) ^{2}} \bigg(\frac{1}{t_{2}^{2}}-\frac{1}{\left( t_{2}+h_{2}\right) ^{2}
\bigg)dt_{1}dt_{2} \\
& :=J_{441}^{\left( 21\right) }+J_{441}^{\left( 22\right) }.
\end{align*}
For $\Vert J_{441}^{\left( 21\right) }\Vert _{p}$ we have
\begin{align*}
\Vert J_{441}^{\left( 21\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}
\\
& \quad \times \frac{(mt_{1})^{2}}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}
dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{m}{n}\right)\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega
(h_{2})}{v(h_{2})}\right) \frac{dt_{1}dt_{2}}{t_{2}^{2}} \\
&= \mathcal{O}(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) ,
\end{align*
while for $\Vert J_{441}^{\left( 22\right) }\Vert _{p}$, we get
\begin{align*}
\Vert J_{441}^{\left( 22\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p} \\
& \quad \times \frac{(mt_{1})^{2}}{\left( t_{1}+h_{1}\right) ^{2}}\bigg(\frac{1}
t_{2}^{2}}-\frac{1}{\left( t_{2}+h_{2}\right) ^{2}}\bigg)dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{m}{n}\right)\int_{0}^{h_{1}}\int_{h_{2}}^{\pi
}(v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{h_{2}(2t_{2}+h_{2})}
t_{2}^{2}\left( t_{2}+h_{2}\right) ^{2}}dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{m}{n^{2}}\right)(v(|z_{1}|)+(v(|z_{2}|))\bigg
\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1}
}\frac{1}{t_{2}^{3}}dt_{1}dt_{2} \\
& \quad +\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\frac{\omega (t_{2}+h_{2})}{\left(
t_{2}+h_{2}\right) v(h_{2})}\frac{1}{t_{2}^{2}}dt_{1}dt_{2}\bigg) \\
&= \mathcal{O}\left(\frac{m}{n^{2}}\right)(v(|z_{1}|)+(v(|z_{2}|))\bigg(\frac{\omega
(2h_{1})}{v(h_{1})}\int_{h_{2}}^{\pi }\frac{1}{t_{2}^{3}}dt_{2}
+\frac{\omega (2h_{2})}{2h_{2}v(h_{2})}\int_{h_{2}}^{\pi }\frac{1}
t_{2}^{2}}dt_{2}\bigg)h_{1} \\
&= \mathcal{O}(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}
v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Whence,
\begin{equation*}
\Vert J_{441}^{\left( 2\right) }\Vert _{p}=\mathcal{O
(1)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) .
\end{equation*}
Thus, we obtain
\begin{equation}
\Vert J_{44}^{\left( 2\right) }\Vert _{p}=\mathcal{O}(1)\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb2}
\end{equation}
By analogy, we find that
\begin{equation}
\Vert J_{44}^{\left( 4\right) }\Vert _{p}=\mathcal{O}(1)\left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb3}
\end{equation}
Once more, replacing $t_{2}$ with $t_{2}+h_{2}$ in
\begin{equation*}
J_{44}^{\left( 3\right) }=\frac{1}{mn\pi ^{2}}\int_{\pi -h_{1}}^{\pi
}\!\!\int_{h_{2}}^{\pi }H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2}
\end{equation*
we get
\begin{equation*}
J_{44}^{\left( 3\right) }=-\frac{1}{mn\pi ^{2}}\int_{\pi -h_{1}}^{\pi
}\!\!\int_{0}^{\pi -h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}
\end{equation*
and adding them side by side, we have
\begin{align*}
J_{44}^{\left( 3\right) } &= \frac{1}{2mn\pi ^{2}}\bigg(\int_{\pi
-h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi }H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}
\frac{S(t_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\int_{\pi -h_{1}}^{\pi }\!\!\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}
\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\bigg) \\
&= \frac{1}{2mn\pi ^{2}}\bigg(\int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi
\bigg(\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\bigg) \\
&\quad \times S(t_{1},t_{2})dt_{1}dt_{2} \!- \!\int_{\pi -h_{1}}^{\pi
}\!\!\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
&\quad +\int_{\pi -h_{1}}^{\pi }\!\!\int_{\pi -h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac{S(t_{1},t_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2}\bigg) \\
&:= \sum\limits_{s=1}^{3}J_{44s}^{\left( 3\right) }
\end{align*}
Since
\begin{equation*}
|\sin 2mt_{1}\sin mt_{1}|\leq 1,\quad \text{for}\quad \pi -h_{1}<t_{1}<\pi
\end{equation*
and
\begin{equation*}
|\sin 2nt_{2}\sin nt_{2}|\leq 2(nt_{2})^{2}\quad \text{for}\quad 0<t_{2}<\pi
,
\end{equation*
then applying Lemma \ref{le1} and Lemma \ref{le2} (iii), we have
\begin{align*}
\Vert J_{442}^{\left( 3\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{0}^{h_{2}}\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
(nt_{2})^{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{n}{m}\right)\int_{\pi -h_{1}}^{\pi
}\!\!\int_{0}^{h_{2}}\left( v(|z_{1}|)+v(|z_{2}|)\right)
\left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{dt_{1}dt_{2}}{\left(
t_{1}+h_{1}\right) ^{2}} \\
&= \mathcal{O}\left(\frac{n}{m}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \\
& \quad \times h_{2}\left( \int_{\pi -h_{1}}^{\pi }\frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}\frac{dt_{1}}{\left( t_{1}+h_{1}\right) ^{2}}+\frac{\omega
(2h_{2})}{v(2h_{2})}\int_{\pi -h_{1}}^{\pi }\frac{dt_{1}}{\left(
t_{1}+h_{1}\right) ^{2}}\right) \\
&= \mathcal{O}\left(\frac{1}{m}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left(
\int_{\pi }^{\pi +h_{1}}\frac{\omega (\theta _{1})}{v(\theta _{1})}\frac
d\theta _{1}}{\theta _{1}^{2}}+\frac{\omega (2h_{2})}{v(2h_{2})}\int_{\pi
}^{\pi +h_{1}}\frac{d\theta _{1}}{\theta _{1}^{2}}\right) \\
&= \mathcal{O}\left(\frac{1}{m}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (\pi +h_{1})}{v(\pi +h_{1})}+\frac{\omega (2h_{2})}{v(2h_{2})}\right)
h_{1} \\
&= \mathcal{O}\left( \frac{1}{m^{2}}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac{\omega (\pi )}{v(\pi )}+\frac
\omega (h_{2})}{v(h_{2})}\right) \\
&= \mathcal{O}\left(\frac{1}{m}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Furthermore, the inequality
\begin{equation*}
|\sin 2mt_{1}\sin mt_{1}\sin 2nt_{2}\sin nt_{2}|\leq 1\quad \text{for}\quad
\pi -h_{1}<t_{1}<\pi ,\quad \pi -h_{2}<t_{2}<\pi ,
\end{equation*
implies
\begin{align*}
\Vert J_{443}^{\left( 3\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{\pi -h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{\pi
-h_{2}}^{\pi }\left( v(|z_{1}|)+v(|z_{2}|)\right) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{dt_{1}dt_{2}}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (\pi +h_{1})}{v(\pi +h_{1})}+\frac{\omega (\pi +h_{2})}{v(\pi +h_{2})
\right) \\
& \quad \times \int_{\pi -h_{1}}^{\pi }\!\!\int_{\pi -h_{2}}^{\pi }\frac
dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \frac{\omega
(2\pi )}{v(2\pi )}\int_{\pi }^{\pi +h_{1}}\!\!\int_{\pi }^{\pi +h_{2}}\frac
dt_{1}dt_{2}}{\left( t_{1}t_{2}\right) ^{2}} \\
&= \mathcal{O}\left( \frac{1}{\left( mn\right) ^{2}}\right) \left(
v(|z_{1}|)+v(|z_{2}|)\right) \frac{\omega (\pi )}{v(\pi )} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)\left( v(|z_{1}|)+v(|z_{2}|)\right) \left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
It is obvious that
\begin{align*}
J_{441}^{\left( 3\right) } &= \frac{1}{2mn\pi ^{2}}\int_{\pi -h_{1}}^{\pi
}\!\!\int_{h_{2}}^{\pi }\bigg(\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})}
\left( (t_{1}+h_{1})t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}} \\
& \quad +\frac{H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})t_{2}\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}\bigg)S(t_{1},t_{2})dt_{1}dt_{2} \\
&= \int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi }\bigg
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}
\bigg) \\
& \quad \times \frac{S(t_{1},t_{2})}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}
dt_{1}dt_{2}+\int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2}) \\
& \quad \times \frac{S(t_{1},t_{2})}{\left( t_{1}+h_{1}\right) ^{2}}\bigg(\frac{1}
t_{2}^{2}}-\frac{1}{\left( t_{2}+h_{2}\right) ^{2}}\bigg
dt_{1}dt_{2}:=J_{441}^{\left( 31\right) }+J_{441}^{\left( 32\right) }.
\end{align*}
For $\Vert J_{441}^{\left( 31\right) }\Vert _{p}$, we have
\begin{align*}
\Vert J_{441}^{\left( 31\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2})-H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}
\\
& \quad \times \frac{dt_{1}dt_{2}}{\left( (t_{1}+h_{1})t_{2}\right) ^{2}} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) \int_{\pi
-h_{1}}^{\pi }\frac{dt_{1}}{\left( t_{1}+h_{1}\right) ^{2}}\int_{h_{2}}^{\pi
}\frac{dt_{2}}{t_{2}^{2}} \\
&= \mathcal{O}\left( \frac{1}{m^{2}}\right) (v(|z_{1}|)+(v(|z_{2}|))\left(
\frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) ,
\end{align*
as far as for $\Vert J_{441}^{\left( 32\right) }\Vert _{p}$ we obtain
\begin{align*}
\Vert J_{441}^{\left( 32\right) }\Vert _{p} &= \mathcal{O}\left(\frac{1}{mn
\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p} \\
& \quad \times \frac{1}{\left( t_{1}+h_{1}\right) ^{2}}\bigg(\frac{1}{t_{2}^{2}}
\frac{1}{\left( t_{2}+h_{2}\right) ^{2}}\bigg)dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{1}{mn}\right)\int_{\pi -h_{1}}^{\pi }\!\!\int_{h_{2}}^{\pi
}(v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{1}{\left( t_{1}+h_{1}\right) ^{2
}\frac{h_{2}(2t_{2}+h_{2})}{t_{2}^{2}\left( t_{2}+h_{2}\right) ^{2}
dt_{1}dt_{2} \\
&= \mathcal{O}\left(\frac{1}{mn^{2}}\right)(v(|z_{1}|)+(v(|z_{2}|))\bigg(\frac{\omega
(\pi +h_{1})}{v(\pi +h_{1})}\int_{h_{2}}^{\pi }\frac{dt_{2}}{t_{2}^{3}}
+\frac{\omega (2h_{2})}{2h_{2}v(h_{2})}\int_{h_{2}}^{\pi }\frac{dt_{2}}
t_{2}^{2}}\bigg)h_{1} \\
&= \mathcal{O}\left( \frac{1}{m^{2}}\right) (v(|z_{1}|)+(v(|z_{2}|))\left(
\frac{\omega (\pi )}{v(\pi )}+\frac{\omega (h_{2})}{v(h_{2})}\right) \\
&= \mathcal{O}\left( \frac{1}{m}\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{align*}
Hence, we obtain
\begin{equation*}
\Vert J_{441}^{\left( 3\right) }\Vert _{p}=\mathcal{O}\left( \frac{1}{m
\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) .
\end{equation*}
Consequently, we get
\begin{equation}
\Vert J_{44}^{\left( 3\right) }\Vert _{p}=\mathcal{O}\left( \frac{1}{m
\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb4}
\end{equation}
By the same way, we obtain
\begin{equation}
\Vert J_{44}^{\left( 5\right) }\Vert _{p}=\mathcal{O}\left( \frac{1}{n
\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb5}
\end{equation}
For $J_{44}^{\left( 6\right) }$, we have
\begin{align}\label{eqb6}
\begin{split}
\Vert J_{44}^{\left( 6\right) }\Vert _{p}& = \mathcal{O}\left( \frac{1}{mn
\right) \int_{0}^{h_{1}}\!\!\int_{0}^{h_{2}}\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
(mnt_{1}t_{2})^{2}dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2
}\\
&= \mathcal{O}\left( mn\right)
\int_{0}^{h_{1}}\!\!\int_{0}^{h_{2}}(v(|z_{1}|)+(v(|z_{2}|))\left( \frac
\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega (t_{2}+h_{2})}
v(t_{2}+h_{2})}\right) dt_{1}dt_{2} \\
&= \mathcal{O}\left( 1\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega
(h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{split}
\end{align}
Similarly as in the estimation of $J_{44}^{\left( 2\right) }$, we hav
\begin{align*}
J_{44}^{\left( 7\right) } &= \frac{1}{2mn\pi ^{2}}\int_{0}^{h_{1}
\int_{h_{2}}^{\pi }H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}dt_{1}dt_{2} \\
& \quad -\frac{1}{2mn\pi ^{2}}\int_{0}^{h_{1}}\int_{0}^{\pi
-h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+2h_{2})\frac{S(t_{1},t_{2})}
\left( (t_{1}+h_{1})(t_{2}+2h_{2})\right) ^{2}}dt_{1}dt_{2} \\
&= \frac{1}{2mn\pi ^{2}}\bigg(\int_{0}^{h_{1}}\int_{h_{2}}^{\pi }\bigg(\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}}-\frac
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+2h_{2})}{\left(
(t_{1}+h_{1})(t_{2}+2h_{2})\right) ^{2}}\bigg) \\
&\qquad \times
S(t_{1},t_{2})dt_{1}dt_{2} \\
& \quad -\int_{0}^{h_{1}
\int_{0}^{h_{2}}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+2h_{2})\frac
S(t_{1},t_{2})}{\left( (t_{1}+h_{1})(t_{2}+2h_{2})\right) ^{2}}dt_{1}dt_{2}
\\
& \quad +\int_{0}^{h_{1}}\int_{\pi -h_{2}}^{\pi
}H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+2h_{2})\frac{S(t_{1},t_{2})}{\left(
(t_{1}+h_{1})(t_{2}+2h_{2})\right) ^{2}}dt_{1}dt_{2}\bigg)
\end{align*
and consequentl
\begin{equation}
\Vert J_{44}^{\left( 7\right) }\Vert _{p}=\mathcal{O}\left( 1\right)
(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega
(h_{2})}{v(h_{2})}\right) . \label{eqb7}
\end{equation}
Analogously,
\begin{equation}
\Vert J_{44}^{\left( 9\right) }\Vert _{p}=\mathcal{O}\left( 1\right)
(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega
(h_{2})}{v(h_{2})}\right) . \label{eqb8}
\end{equation}
For $J_{44}^{\left( 8\right) }$, we have
\begin{align}\label{eqb9}
\begin{split}
\Vert J_{44}^{\left( 8\right) }\Vert _{p} &= \mathcal{O}\left( \frac{1}{mn
\right) \int_{0}^{h_{1}}\!\!\int_{\pi -h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
(mt_{1})^{2}dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left( \frac{m}{n}\right) \int_{0}^{h_{1}}\!\!\int_{\pi
-h_{2}}^{\pi }(v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{dt_{1}dt_{2}}{\left(
t_{2}+h_{2}\right) ^{2}} \\
&= \mathcal{O}\left( \frac{m}{n}\right) (v(|z_{1}|)+(v(|z_{2}|))
\left( \frac{\omega (h_{1})}{v(h_{1})}h_{2}+\int_{\pi -h_{2}}^{\pi
\frac{\omega (t_{2}+h_{2})}{v(t_{2}+h_{2})}\frac{dt_{2}}{\left(
t_{2}+h_{2}\right) ^{2}}\right) h_{1} \\
&= \mathcal{O}\left( \frac{1}{n}\right) (v(|z_{1}|)+(v(|z_{2}|))
\left( \frac{\omega (h_{1})}{v(h_{1})}h_{2}+\int_{\pi }^{\pi +h_{2}
\frac{\omega (\theta _{2})}{v(\theta _{2})}\frac{d\theta _{2}}{\theta
_{2}^{2}}\right)\\
&= \mathcal{O}\left( \frac{1}{n^{2}}\right) (v(|z_{1}|)+(v(|z_{2}|))\left(
\frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega (\pi )}{v(\pi )}\right) \\
&= \mathcal{O}\left( \frac{1}{n}\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{split}
\end{align}
By analogy, we also get
\begin{equation}
\Vert J_{44}^{\left( 11\right) }\Vert _{p}=\mathcal{O}\left( \frac{1}{m
\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) . \label{eqb10}
\end{equation}
Now we need to give the upper bound for $\Vert J_{44}^{\left( 10\right)
}\Vert _{p}$. We have
\begin{align}\label{eqb11}
\begin{split}
\Vert J_{44}^{\left( 10\right) }\Vert _{p} &= \mathcal{O}\left( \frac{1}{mn
\right) \int_{h_{1}}^{\pi }\!\!\int_{\pi -h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left( \frac{1}{mn}\right) \int_{h_{1}}^{\pi }\!\!\int_{\pi
-h_{2}}^{\pi }(v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{dt_{1}dt_{2}}{\left(
t_{1}+h_{1}\right) ^{2}}\\
&= \mathcal{O}\left( \frac{1}{mn}\right) (v(|z_{1}|)+(v(|z_{2}|))\left(
\frac{\omega (\pi )}{v(\pi )}\frac{m}{n}+m\int_{\pi }^{\pi +h_{2}}\frac
\omega (u_{2})du_{2}}{v(u_{2})u_{2}^{2}}\right) \\
&= \mathcal{O}\left( \frac{1}{n^{2}}\right) (v(|z_{1}|)+(v(|z_{2}|))\frac
\omega (\pi )}{v(\pi )} \\
&= \mathcal{O}\left( \frac{1}{n}\right) (v(|z_{1}|)+(v(|z_{2}|))\frac{\omega
(h_{2})}{v(h_{2})}.
\end{split}
\end{align}
By analogy, we obtain
\begin{equation}
\Vert J_{44}^{\left( 12\right) }\Vert _{p}=\mathcal{O}\left( \frac{1}{m
\right) (v(|z_{1}|)+(v(|z_{2}|))\frac{\omega (h_{1})}{v(h_{1})}.
\label{eqb12}
\end{equation}
Finally, we hav
\begin{align}\label{eqb13}
\begin{split}
\Vert J_{44}^{\left( 13\right) }\Vert _{p} &= \mathcal{O}\left( \frac{1}{mn
\right) \int_{\pi -h_{1}}^{\pi }\!\!\int_{\pi -h_{2}}^{\pi }\left\Vert
H_{x,z_{1},y,z_{2}}(t_{1}+h_{1},t_{2}+h_{2})\right\Vert _{p}\frac
dt_{1}dt_{2}}{\left( (t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left( \frac{1}{mn}\right) \int_{\pi -h_{1}}^{\pi
}\!\!\int_{\pi -h_{2}}^{\pi }(v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( \frac{\omega (t_{1}+h_{1})}{v(t_{1}+h_{1})}+\frac{\omega
(t_{2}+h_{2})}{v(t_{2}+h_{2})}\right) \frac{dt_{1}dt_{2}}{\left(
(t_{1}+h_{1})(t_{2}+h_{2})\right) ^{2}} \\
&= \mathcal{O}\left( \frac{1}{mn}\right) (v(|z_{1}|)+(v(|z_{2}|)) \\
& \quad \times \left( h_{2}\int_{\pi -h_{1}}^{\pi }\frac{\omega (t_{1}+h_{1})}
v(t_{1}+h_{1})}\frac{dt_{1}}{(t_{1}+h_{1})^{2}}+h_{1}\int_{\pi -h_{2}}^{\pi
\frac{\omega (t_{2}+h_{2})}{v(t_{2}+h_{2})}\frac{dt_{2}}{(t_{2}+h_{2})^{2}
\right) \\
&= \mathcal{O}\left( \frac{1}{mn}\right) (v(|z_{1}|)+(v(|z_{2}|))\left(
h_{2}\int_{\pi }^{\pi +h_{1}}\frac{\omega (\theta _{1})}{v(\theta _{1})
\frac{d\theta _{1}}{\theta _{1}^{2}}+h_{1}\int_{\pi }^{\pi +h_{2}}\frac
\omega (\theta _{2})}{v(\theta _{2})}\frac{d\theta _{2}}{\theta _{2}^{2}
\right) \\
&= \mathcal{O}\left( \frac{1}{\left( mn\right) ^{2}}\right)
(v(|z_{1}|)+(v(|z_{2}|))\frac{\omega (\pi )}{v(\pi )}\\
& = \mathcal{O}\left( \frac{1}{mn}\right) (v(|z_{1}|)+(v(|z_{2}|))\left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right) .
\end{split}
\end{align}
Whence, using \eqref{eqb1}-\eqref{eqb13}, we obtain
\begin{equation*}
\left\Vert J_{44}\right\Vert _{p}=\mathcal{O}\left( 1\right)
(v(|z_{1}|)+(v(|z_{2}|))\left( \frac{\omega (h_{1})}{v(h_{1})}+\frac{\omega
(h_{2})}{v(h_{2})}\right) .
\end{equation*
Combining partial estimates, we ge
\begin{equation*}
\sup_{z_{1}\neq 0,\,\,z_{2}\neq 0}\frac{\Vert
D_{m,n}(x+z_{1},y+z_{2})-D_{m,n}(x,y)\Vert _{p}}{v(|z_{1}|)+v(|z_{2}|)}
\mathcal{O}\left( 1\right) \left( \frac{\omega (h_{1})}{v(h_{1})}+\frac
\omega (h_{2})}{v(h_{2})}\right) .
\end{equation*
Procceding as in the similar lines, it can be proved tha
\begin{equation*}
\Vert D_{m,n}\Vert _{p}=\mathcal{O}\left( 1\right) \left( \omega
(h_{1})+\omega (h_{2})\right) .
\end{equation*
Henc
\begin{equation*}
\Vert D_{m,n}\Vert _{p}^{(v,v)}=\mathcal{O}\left( 1\right) \left( \frac
\omega (h_{1})}{v(h_{1})}+\frac{\omega (h_{2})}{v(h_{2})}\right)
\end{equation*
and this ends our proof.
\end{proof}}
Now we can finish this section by deriving some particular results. For this, we need to
specialize the functions $\omega(t)$ and $v(t)$ in our theorem. In fact, taking $\omega(t)=t^{\alpha}$ and $v(t)=t^{\beta}$, $0\leq \beta< \alpha\leq 1$, in Theorem \ref{the01} we get:
\begin{corollary}
If $f\in \text{Lip}(\alpha,\beta, p)$, $p\geq 1$, $0\leq \beta< \alpha\leq 1$, then
\begin{equation*}
\Vert \sigma _{m,2m;n,2n}(f)-f\Vert _{p}^{(v,v)}= \mathcal{O} \left( \frac{1
}{m^{\alpha -\beta}}+\frac{1
}{n^{\alpha -\beta}}\right)
\end{equation*
for all $m,n\in
\mathbb{N}
$.
\end{corollary}
For $p=\infty$ in above corollary, we obtain the following.
\begin{corollary}
If $f\in \text{H}_{(\alpha,\beta)}$, $0\leq \beta< \alpha\leq 1$, then
\begin{equation*}
\Vert \sigma _{m,2m;n,2n}(f)-f\Vert _{\alpha,\beta}= \mathcal{O} \left( \frac{1
}{m^{\alpha -\beta}}+\frac{1
}{n^{\alpha -\beta}}\right)
\end{equation*
for all $m,n\in
\mathbb{N}
$.
\end{corollary}
\section{A generalization of Theorem \ref{the01}}
We denote by $W(L^p((-\pi,\pi)^2); \beta_1, \beta_2)$ the weighted space $L^p((-\pi,\pi)^2)$ with weight function $\left|\sin \left(\frac{x}{2}\right)\right|^{\beta_1 p}\left|\sin \left(\frac{y}{2}\right)\right|^{\beta_2 p}$, ($\beta_1,\beta_2\geq 0$), and endowed with norm
$$\|f\|_{p;\beta_1,\beta_2}:=\left(\frac{1}{(2\pi)^2}\int_{-\pi}^{\pi}\int_{-\pi}^{\pi}\left|f(x,y)\right|^p\left|\sin \left(\frac{x}{2}\right)\right|^{\beta_1 p}\left|\sin \left(\frac{y}{2}\right)\right|^{\beta_2 p}dxdy\right)^p$$
for $1\leq p<\infty$, and
$$\|f\|_{p;\beta_1,\beta_2}:=\text{ess}\!\!\!\!\!\sup_{-\pi\leq x,y \leq \pi }\left\{\left|f(x,y)\right|\left|\sin \left(\frac{x}{2}\right)\right|^{\beta_1 }\left|\sin \left(\frac{y}{2}\right)\right|^{\beta_2 }\right\}$$
for $p=\infty$ (for $\beta_1=0,\beta_2=0$ see \cite{US}).
Acting accordingly, we define the space $H_{p;\beta_1,\beta_2}^{(\omega_1, \omega_2)}$ by
\begin{equation*}
H_{p;\beta_1,\beta_2}^{(\omega_1, \omega_2 )}:=\left\{f\in W(L^p((-\pi ,\pi )^2); \beta_1, \beta_2), p\geq 1:
A(f;\omega_1 , \omega_2 ;\beta_1 ,\beta_2 )<\infty \right\},
\end{equation*}
where
\begin{equation*}
A(f; \omega_1 , \omega_2 ;\beta_1 ,\beta_2 ):=\sup_{t_1\neq 0,\,\,t_2\neq 0 }\frac{\|f(x +t_1,y
+t_2)-f(x,y)\|_{p ;\beta_1 ,\beta_2 }}{\omega_1 (|t_1|)+\omega_2 (|t_2|)}
\end{equation*}
and the norm in the space $H_{p ;\beta_1 ,\beta_2 }^{(\omega_1, \omega_2)}$ is defined by
\begin{equation*}
\|f\|_{p ;\beta_1 ,\beta_2 }^{(\omega_1, \omega_2)}:=\|f\|_{p ;\beta_1 ,\beta_2 }+A(f;\omega_1, \omega_2 ;\beta_1 ,\beta_2 ).
\end{equation*}
We take $k=rm$ and $\ell =qn$ in (\ref{eq1}), to obtain
\begin{align*}
\sigma_{m,rm;n,qn}:=&\left( 1+\frac{1}{r}\right)\left( 1+\frac{1}{q
\right)\sigma_{m(r+1)-1,n(q+1)-1}-\left( 1+\frac{1}{r}\right)\frac{1}{q
\sigma_{m(r+1)-1,n-1} \nonumber \\
& -\frac{1}{r}\left( 1+\frac{1}{q}\right)\sigma_{m-1,n(q+1)-1}+\frac{1}
rq}\sigma_{m-1,n-1},
\end{align*}
where $r,q\in \{2,4,6,\dots\}$.
Hence, we name the mean $\sigma _{m,rm;n,qn}(f;x,y)$ as {\it Double Even-Type Delayed Arithmetic Mean} for
S_{k,\ell }(f;x,y)$, which can be represented in its integral form
\begin{equation*}
\sigma_{m,rm;n,qn}(f;x,y)=\frac{1}{mn\pi^2}\int_{0}^{\pi}\int_{0}^
\pi}h_{x,y}(t_1,t_2)L_{m,rm;n,qn}(t_1,t_2)dt_1dt_2,
\end{equation*}
where
$$L_{m,rm;n,qn}(t_1,t_2):=\frac{4}{rq}\frac{\sin \frac{\left(r+2\right)mt_1}{2}\sin \frac{rmt_1}{2}\sin \frac{\left(q+2\right)nt_2}{2}\sin \frac{qnt_2}{2}}{\left( 4\sin \frac{t_1}{2}\sin \frac{t_
}{2}\right)^2}.$$
Further, we establish a more general theorem on the degree of approximation of function $f$ belonging to $H_{p ;\beta_1 ,\beta_2 }^{(\omega, \omega)}$ with norm $\|f\|_{p ;\beta_1 ,\beta_2 }^{(v,v)}$ by Double Even-Type Delayed Arithmetic Mean $\sigma_{m,rm;n,qn}(f;x,y)$. It represent a twofold generalization of Theorem \ref{the01}, both in context of the considered space and the mean.
\begin{theorem}
\label{the05} Let $\omega $ and $v$ be moduli of continuity so that $\frac
\omega (t)}{v(t)}$ is non-decreasing in $t$. If $f\in H_{p ;\beta_1 ,\beta_2 }^{(\omega, \omega)}$, $p\geq 1$, then
\begin{equation*}
\Vert \sigma_{m,rm;n,qn}(f)-f\Vert _{p ;\beta_1 ,\beta_2 }^{(v,v)}= \mathcal{O} \left( r\frac{\omega (h_{1}
}{v(h_{1})}+q\frac{\omega (h_{2})}{v(h_{2})}\right) ,
\end{equation*
where $h_{1}=\frac{\pi }{m}$, $h_{2}=\frac{\pi }{n}$, $m,n\in
\mathbb{N}
$, and $r,q\in \{2,4,6,\dots\}$.
\end{theorem}
\begin{proof}
The proof can be done in the same lines as the proof of Theorem \ref{the01}. We omit the details.
\end{proof}
\begin{remark}
For $\beta_1 =0$, $\beta_2=0$, $r=2$ and $q=2$, Theorem \ref{the05} reduces exactly to Theorem \ref{the01}.
\end{remark}
\section{Conclusions}
For one dimension, a degree of approximation of a function in the space $H_{p}^{(\omega)}$ has been given by several authors, see \cite{DGR,DNR,D,UD2,KIM1,XhK,L,NDR,SS} and some other references already mentioned here. Inspired by these papers, we have given two corresponding results using the second (even) type double delayed arithmetic means of the Fourier series of a function from the space $H_{p}^{(\omega , \omega)}$ ($H_{p ;\beta_1 ,\beta_2 }^{(\omega, \omega)}$).
| 649b0593d7cebe4c7eae613c239fdc30ff0d096f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Color transparency is the amazing prediction of QCD that initial- and final-state interactions are reduced in high-momentum transfer, high-energy coherent reactions~\cite{Frankfurt:1994hf,Jain:1995dd,Ashery:2006zw,Dutta:2012ii}.
Strong interactions are strong: when hadrons hit nuclei they generally break up the nucleus or themselves. Indeed, a well-known semi-classical formula states that the intensity of a beam of hadrons falls exponentially with the penetration distance through nuclei. This effect is known as absorption.
It is remarkable that QCD indicates
that, under certain specific conditions, the strong interactions can effectively be turned off and hadronic systems can move freely through a nuclear medium.\\
This phenomenon is based on three requirements:
\begin{itemize}
\item High momentum transfer coherent reactions are dominated by point-like color-singlet components of the struck hadron wave function, denoted as PLCs. This statement was a prediction intialy based on perturbative QCD (pQCD). For example, early pQCD calculations~\cite{Farrar:1979aw,Efremov:1979qk,Lepage:1979za,Lepage:1979zb,PhysRevD.21.1636,DUNCAN1980159,PhysRevD.21.1636},
of the pion elastic electromagnetic form factor (see Fig.~1a) were interpreted \cite{Mueller:1982bq} in the following manner:
A high-momentum, $Q$, photon hits one of the partons that greatly increases the relative momentum to $Q$. The system can only stay together only by exchanging a gluon carrying that momentum. That gluon has a range of only $1/Q$ so that the partons must be close together, making a small-sized system or point-like configuration (PLC).
The validity of pQCD for computation of electromagnetic form factors is a sufficient but not necessary condition for involvement of s PLC
\cite{Frankfurt:1993es} .\\
\item Small objects have small cross sections. It has been widely reported that the imaginary part of the forward scattering amplitude, $f$ of
a rapidly moving color singlet object is proportional to the square of the transverse separation, $b$, between positive and. negative color charges.
The dominant contribution to $f$, at high energies is given by two-gluon exchange \cite{Low:1975sv,Nussinov:1975mw,Gunion:1976iy}, and the remarkable feature is that, in the limit that $b$ approaches 0, $f$ vanishes because color-singlet point particles do not exchange colored gluons. This feature is expressed concisely as $\lim_{b\to0}\s(b^2)\propto b^2.$ This reduced interaction, caused by interference between emission by quarks of different colors in coherent processes, is the basic ingredient behind QCD factorization proofs, is widely used \cite{Donnachie:2002en} and not questioned.\\
\item A PLC, once created, will expand as it moves. This is because a PLC is not an eigenstate of the Hamiltonian. The expansion effect is diminished if the PLC moves with sufficiently high momentum.\\
\end{itemize}
If all three requirements are satisfied for a given coherent process the effects of color transparency will be evident.\\
The second and third items are based on many calculations, many experiments and basic principles of quantum mechanics. The interesting dynamical question is the validity or lack thereof of the idea the PLCs are involved. This question is intimately connected with the origin of hadronic electromagnetic form mechanism at high momentum transfer. In the perturbative QCD mechanism, the large momentum transfer is taken up by the exchange of high-momentum gluons. For this to occur, all of the partons must be at the same transverse spatial location.
Although it is natural to suppose that PLCs dominate coherent high-momentum transfer
processes, it is far from obvious that this is the case~\cite{PhysRevD.21.1636,DUNCAN1980159,PhysRevD.21.1636,Feynman:1973xc}. For example,
the momentum transfer can involve a single quark of high momentum. This is the Feynman mechanism, see Fig.~1, (first stated by Drell \& Yan \cite{Drell:1969km}). The spectator system is not required to shrink to a small size and color transparency effects involving protons would not be expected to occur. In this picture, a reasonably valid relation between elastic and deep inelastic scattering is obtained. See also~\cite{West:1970av}. A more recent example of a process that favors the Feynman mechanism appears in~\cite{PhysRevD.58.114008}. See also~\cite{ Burkardt:2003mb}. Note that
Feynman remarked, ``if a system is made of 3 particles, the large $Q^2$ behavior depends not on the singularity when just two come together, but rather when all three are on top of one another". Furthermore, ``such pictures are too simple and inadequate".\\
\begin{figure}[h] \label{Diags}
\includegraphics[width=.5\textwidth]{Diags.pdf}
\caption{ High momentum transfer reaction mechanisms. Top picture: a pQCD mechanism. Other diagrams of the same order are not shown. Middle picture: Initial state in the Feynman mechanism. Bottom picture: final state in the Feynman mechanism. The final state has a good overlap with the turned around version of the initial state.}
\end{figure}
Models of Generalized Parton Distributions (GPDS)~\cite{ Ji:2004gf} access both the longitudinal and transverse structure of nucleons so that measurements thereof can distinguish the different mechanisms. It has been said~\cite{Diehl:2003ny} that GPDs parameterize soft dynamics akin to the Feynman mechanism. Specific models of GPDs, for example~\cite{deTeramond:2018ecg} also favor the Feynman mechanism. More generally, a review of the history teaches us ~\cite{Belitsky:2005qn} that there are only two proposals for the mechanism responsible for high momentum elastic reactions. \\
Frankfurt, Miller \& Strikman (FMS) introduced~\cite{Frankfurt:1993es} a criteria to determine whether or not a given model of a hadronic wave function admits the existence of a PLC. They found that a PLC could arise from non-perturbative effects as well as from perturbative QCD. The aim of this paper is to use the FMS criteria to see if the relativistic light-front wave functions obtained from light front holographic QCD (see the review \cite{Brodsky:2014yha}) admit the existence of a PLC. \\
The first step is to discuss the FMS criterion.
The idea is that a PLC is originated via a hard probing interaction $T_H$ involving nucleons initially bound in a nucleus.
The soft interactions between the PLC and the surrounding medium are proportional to the square of the transverse separation distance \cite{Low:1975sv,Nussinov:1975mw,Gunion:1976iy} between constituents, $b^2=\sum_{i< j}(b_i-b_j)^2,$ where the constituents are labelled $i,j\cdots$ in first-quantized notation. Consider a high-momentum transfer process on a nucleon. Denote the initial nucleon state as $|\psi(p)\rangle$ and the final state as $|\psi(p+q)\rangle$. Then represent the high momentum transfer operator as $T_H(q)|\psi\rangle$. With this notation the form factor $F(Q^2)$ is given by the matrix element
\begin{eqnarray} F(Q^2)=\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta\psi(p+q)|T_H(q)|\psi(p)\rangle.\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The key question is whether the state $T_H(q)|\psi(p)\ra$ is a PLC that does not interact with the medium. The interaction with the surrounding medium is proportional to $b^2$, the square of the transverse separation between a struck parton and the remainder of the system. The first-order term in the interaction is proportional to the matrix element $ \langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta\psi(p+q)|b^2 T_H(q)|\psi(p)\rangle$. This term is small if the operator $T_H$ produces a PLC. The relevant comparison is with the form factor $F(Q^2) $ that is the process amplitude in the absence of final state interactions. Thus FMS defined the quantity $b^2(Q^2)$ as
\begin{eqnarray} b^2(Q^2)={\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta\psi(p+q)|b^2T_H(q)|\psi(p)\rangle\over \langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta\psi(p+q)|T_H(q)|\psi(p)\rangle}\equiv{F_{b^2}(Q^2)\over F(Q^2)}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
If $b^2(Q^2)=b^2(0)$ final state interactions of normal magnitudes occur. If $b^2(Q^2)$ drops with increasing values of $Q^2$, then the model wave function is said to admit the existence of a PLC.\\
Now turn to evaluating $b^2(Q^2) $
for wave functions obtained from holographic techniques used represent relativistic light front wave functionst, as discussed in the review~\cite{Brodsky:2014yha}.
I briefly discuss that approach. Light-front quantization is a relativistic, frame-independent approach to describing the constituent structure of hadrons. The assumed simple structure of the light-front (LF) vacuum allows a definition of the partonic content of a hadron in QCD and of hadronic light-front theory~\cite{Brodsky:1997de}. The spectrum and light-front wave functions of relativistic bound states are obtained in principle from the eigenvalue equation $H_{LF} \vert \psi \rangle = M^2 \vert \psi \rangle$ that becomes an infinite set of coupled integral equations for the LF components.
This provides a quantum-mechanical probabilistic interpretation of the structure of hadronic states in terms of their constituents at the same light-front time $x^+ = x^0 + x^3$, the time marked by the front of a light wave~\cite{Dirac:1949cp}. The matrix diagonalization~\cite{Brodsky:1997de} of the frame-independent LF Hamiltonian eigenvalue equation in four-dimensional space-time has not been achieved, so other methods and approximations ~\cite{Brodsky:2014yha} are needed to understand the nature of relativistic bound states in the strong-coupling regime of QCD.\\
To a first semiclassical approximation, where quantum loops and quark masses are not included, the relativistic bound-state equation for light hadrons can be reduced to an effective LF Schr\"odinger equation. The technique is to identify the invariant mass of the constituents as a key dynamical variable,
$\zeta$, which measures the separation of the partons within the hadron at equal light-front time~\cite{deTeramond:2008ht}. Thus, the multi-parton problem in QCD is reduced, in a first semi-classical approximation, to an effective one dimensional quantum field theory by properly identifying the key dynamical variable. In this approach the complexities of the strong interaction dynamics are hidden in an effective potential. \\
It is remarkable that in the semiclassical approximation described above, the light-front Hamiltonian has a structure which matches exactly the eigenvalue equations in AdS space~\cite{Brodsky:2014yha}. This offers the possibility to explicitly connect the AdS wave function $\Phi(z)$ to the internal constituent structure of hadrons. In fact, one can obtain the AdS wave equations by starting from the semiclassical approximation to light-front QCD in physical space-time. This connection yields a relation between the coordinate $z$ of AdS space with the impact LF variable $\zeta$~\cite{deTeramond:2008ht}, thus giving the holographic variable $z$ a precise definition and intuitive meaning in light-front QCD.\\
Light-front holographic methods were originally introduced~\cite{Brodsky:2006uqa,Brodsky:2007hb} by matching the electromagnetic current matrix elements in AdS space~\cite{Polchinski:2002jw} with the corresponding expression derived from light-front quantization in physical space-time~\cite{Drell:1969km,West:1970av}. It was also shown that one obtains identical holographic mapping using the matrix elements of the energy-momentum tensor~\cite{Brodsky:2008pf} by perturbing the AdS metric around its static solution~\cite{Abidin:2008ku}, thus establishing a precise relation between wave functions in AdS space and the light-front wave functions describing the internal structure of hadrons.\\
The light front wave functions that arise out of this light front holographic approach provide a new way to study old problems that require the use of relativistic-confining quark models. The study of the existence of a PLC by evaluating $b^2(Q^2)$ is an excellent example of such a problem. \\
I evaluate two examples. The first~\cite{Brodsky:2007hb} is an early representation of the pion wave function as a $q\bar q$ system.
\begin{eqnarray} \psi(x,b)={\k\over \sqrt{\pi}}\sqrt{x(1-x)}e^{-b^2\k^2x(1-x)/2}.\label{wf}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
I use the normalization $1=\int dxd^2b|\psi(x,b)|^2$ throughout this paper. The wave function of \eq{wf} appears in many models.
The form factor is given by
\begin{eqnarray} F(Q^2)=\int dxd^2b e^{i \bfQ \cdot\bfb(1-x)}|\psi(x,b)|^2,\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q} and evaluation yields
\begin{eqnarray} F(Q^2)=1-e^{Q^2/4\k^2}\G(0,Q^2/4\k^2),\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
with $\G$ being the incomplete Gamma function. This form factor falls asymptotically as $\sim 1/Q^2$, providing a nice example of how a non-perturbative wave function can yield a power-law falloff. The limit for $Q^2\to0 $ is also interesting:
\begin{eqnarray}
\lim_{Q^2\to0}F(Q^2)=1 +(\gamma}\def\m{\mu}\def\t{\tau}\def\ve{\varepsilon}\def\l{\lambda}\def\s{\sigma}\def\n{\nu}\def\o{\omega_E +\log(Q^2/4\k^2))Q^2/4\k^2+\cdots\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
because of the appearance of the $\log$.\\
The quantity $F_{b^2}(Q^2)$ is obtained by inserting a factor $b^2$ into the integrand:
\begin{eqnarray}& F_{b^2}(Q^2)=\int dxd^2b \,b^2\, e^{i \bfQ \cdot\bfb(1-x)}|\psi(x,b)|^2\nonumber\\
&=-\nabla_Q^2\int {dx\over (1-x)^2}d^2b e^{i \bfQ \cdot\bfb(1-x)}|\psi(x,b)|^2\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
Observe that $F_{b^2}$ is not simply obtained by differentiating the form factor with respect to $Q^2$. This is because of the factor $1-x$ that appears in the exponential function.
Evaluation of the integral over the transverse coordinates yields
\begin{eqnarray}
F_{b^2}(Q^2)={1\over \k^2}\int {dx \,e^{-{Q^2\over 4\k^2}}\over x^2(1-x)}(x+{Q^2\over 4\k^2}(1-x)).
\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The value of $F_{b^2}(Q^2)$ is infinite for all values of $Q^2$ because of the divergence (related to the $\log Q^2$ term in the form factor) as $x$ approaches unity. This shows that the simple wave function of~\cite{Brodsky:2007hb} is not suitable for use in evaluating high energy forward cross sections for pion-nucleus interactions.
The model given in \eq{wf} is very simple, so the next step is to use the universal light front wave functions of Ref.~\cite{deTeramond:2018ecg}.
This work presented a universal description of generalized parton distributions obtained from Light-Front Holographic QCD, and I use their models for light-front wave functions, presented as functions of the number $\t$ of constituents of a Fock space component. Regge behavior at small $x$ and inclusive counting rules as $x\rightarrow1$ are incorporated. Nucleon and pion valence quark distribution functions are obtained in precise agreement with global fits. The model is described by the
quark distribution $q_\tau(x)$ and the profile function $f(x)$ with
\begin{eqnarray} \label{qx} &
q_\tau(x) = \frac{1}{N_\tau} \big(1- w(x)\big)^{\tau-2}\, w(x)^{- {\frac{1}{2}}}\defx^-}\newcommand{\mbf}[1]{\mathbf{#1}}\newcommand{\req}[1]{(\ref{#1})}\def\tz{{\tilde z}{x^-}\newcommand{\mbf}[1]{\mathbf{#1}}\newcommand{\req}[1]{(\ref{#1})}\def\tz{{\tilde z}}\, w'(x), \\&
\label{fax}
f(x) = \frac{1}{4 \lambda}\left[ (1-x) \log\left(\frac{1}{x}\right) + a (1 - x)^2 \right],
\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
and $w(x) = x^{1-x} e^{-a (1-x)^2}.$
The value of the universal scale $\lambda$ is fixed from the $\rho$ mass: $\sqrt{\lambda} = \kappa = m_\rho/ \sqrt{2} = 0.548$ GeV~\cite{Brodsky:2014yha,Brodsky:2016yod}. The flavor-independent parameter $a = 0.531 \pm 0.037$. The $u$ and $d$ quark distributions of the proton are given by a linear superposition of $q_3$ and $q_4$ while those of the pion are obtained from $q_2$ and $q_4$.
Ref.~\cite{deTeramond:2018ecg} also presents
the universal light front wave function:
\begin{eqnarray} \label{LFWFb}
\psi_{\rm eff}^{(\tau)}(x, \mbf{b} ) = \frac{1}{2 \sqrt{\pi}} \sqrt{\frac{q_\tau(x)}{f(x)}}
(1-x) e^{ - \frac{(1-x)^2 }{8 f(x) } \, \mbf{b}^2},
\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
in the transverse impact space representation. The transverse coordinate $b$ again represents the relative distance between a struck parton and the spectator system.\\
The form factor for a given value of $\t$ is given by
\begin{eqnarray} F^{(\t)}(Q^2)=\int dx q_\t(x)e^{-Q^2 f(x)}, \end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
and $F^{(\t)}_{b^2}{Q^2}$, obtained by inserting a factor $b^2$ into the above integrand, is given by:
\begin{eqnarray} &F^{(\t)}_{b^2}{Q^2}=\int dx {q_\t(x)\over 4f(x)}e^{-Q^2 f(x)}(1-{Q^2\over 16f(x)}(1-x)^2).\nonumber\\&\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
Consider first the case of $\t=2$.
The use of \eq{qx} and \eq{fax} shows that
\begin{eqnarray}
\lim_{x\to1}{q_2(x)\over f(x)}={4\l\over (1-x)}+\cdots.\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
Thus the same divergence that haunted the wave function of \eq{wf} reappears for the more sophisticated $\t=2$ wave function of
\cite{deTeramond:2018ecg}
The functions $b^2_\t(Q^2)= F^{(\t)}_{b^2}{Q^2}/F^{(\t)}{Q^2}$ for $\t=3,4$ are shown in Fig.~2.
\begin{figure}[h]
\includegraphics[width=5.398cm,height=4.6cm]{bsq.pdf}
\caption{$b^2_\t(Q^2)$ in units of GeV$^{-2}$. The numbers refer to the value of $\t$, the number of constituents in the Fock state.}\label{bsqz}\end{figure}\\
Observe that $b^2_\t(Q^2)$ rises with increasing values of $Q^2$, so that these wave functions do not admit the existence of PLCs. Furthermore, observe the surprising effect that constituents with larger number of partons have smaller values of $b^2(Q^2)$ and so interact less strongly with a surrounding medium.\\
The summary of this work is that light front holographic wave functions do not contain a PLC, so they do not predict the appearance of color transparency no matter how large the value of $Q^2$. Adding a perturbative QCD tail to the momentum-space wave function
could change this result. The wave functions of light front holographic QCD are suitable for describing the soft dynamics involved the time evolution of a wave packet and were used~\cite{Caplow-Munro:2021xwi} to interpret the recent striking experimental finding~\cite{Bhetuwal:2020jes}, that color transparency does not occur in $(e,e'p)$ reactions with $Q^2$, up to 14.2 GeV$^2$. The result is that the Feynman mechanism is responsible for the proton form factor at high momentum transfer.
\section*{Acknowledgements}
This work was supported by the U. S. Department of Energy Office of Science, Office of Nuclear Physics under Award Number DE-FG02-97ER- 41014. I thank the organizers R. Dupre, D. Dutta, M. Sargsian and M. Strikman of the 2021 workshop on ``The Future of Color Transparency and Hadronization Studies at Jefferson Lab and Beyond" for re-stimulating my interest in this subject. I thank S. J. Brodsky for making this paper necessary.
\input{CT22.bbl}
\end{document}
The starting point for the present calculations is the relativistic light-front wave equation for hadronic light front wave function $\phi(\z)$ that encodes the dynamical properties. Consider a hadronic bound state of two constituents, each of vanishing mass.
The wave equation for a hadronic bound state of two constituents is given by:
\begin{eqnarray}
\bigg(-{d^2\over d\z^2}-{1-4L^2\over 4\z^2}+U(\z,J)\bigg)\phi(\z)=M^2\phi(\z)
\label{we} \end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The eigenmodes of this equation determine the hadronic mass spectrum and represent the probability amplitude to find the partons at transverse impact separation $\z$. The connection between $\phi$ and wave functions $\Phi$ in a higher-dimensional space asymptotic to anti-de-Sitter space is explained in Ref.~\cite{Brodsky:2014yha}.
In the soft-wall model ~\cite{Karch:2006pv} the potential is harmonic and holographically related to a unique dilaton profile the follows from the requirements of conformal invariance. The effective potential is given~\cite{Brodsky:2014yha} by
\begin{eqnarray} U(\zeta,J) = \kappa^4\zeta^2+2\k^2(J-1)\label{sw}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
where $ \kappa $ is the strength of the confinement, and $J$ is the total angular momentum.
This form was used in ~\cite{PhysRevLett.102.081601,PhysRevD.91.045040,PhysRevD.91.085016}, in particular to obtain relations between the baryon and meson spectrum~\cite{PhysRevD.91.085016}, with a universal value of $\k=0.53$ GeV. The results are specified here to: $U(\z)=\k^4 \z^2 +C_H$, with $C_\pi=-2\k^2,\,C_\r=0,\,C_N=2\k^2$. We shall see that these different values have profound effects on the ability to observe color transparency.
The associated eigenvalues are :
$M^2_{nL} = 4\kappa^2(n+{J+L_M\over 2}) $ for mesons, and for baryons
$ M^2_{nL} = 4\kappa^2(n+L_B+1) $
with $L_M=L_B +1$.\\
The next step is to convert \eq{sw} to a light front 3-dimensional wave equation in coordinate space.
This is done by using
$ \phi(\z)\to \sqrt{\z}\psi(x,\bfb)= \sqrt{z}e^{iL\phi}\psi_L(x,b)$ with the replacement $\z\to b \sqrt{x(1-x)}$,
so that the equation for a meson is given by:
\begin{eqnarray} \bigg(-{1\over x(1-x) }\nabla_b^2+U(b^2(x(1-x),J)\bigg)\psi(x,\bfb)=M^2 \psi(x,\bfb)\label{lfe}
\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The equation for the nucleon is of the same form, but there are two components, denoted $\psi_{\pm}$. The smaller-sized of these two components
is $\psi_+$. Therefore that is the relevant one whose time development is studied here. A good description of electromagnetic form factors and parton distributions is obtained~\cite{Brodsky:2016yod}.\\
Note that the Hamiltonian contains no operator that changes the value of $x$. The longitudinal degree of freedom is fixed in this model. However, the known meson and baryon spectra are well reproduced, which is the central necessity for obtaining a reasonable treatment of the expansion of the PLC wave packet.
The light-front wave functions obtained by solving \eq{lfe}
are given by:
\begin{eqnarray} \psi_{nL_z}(\bfb,x) = {\sqrt{2} }\sqrt{n!\over (n+L)!} {e^{i L_z \varphi} \over\sqrt{2\pi}}e^{-\k^2b^2 x(1-x)/2}(\sqrt{x(1-x)}b)^{L} L_n^{L}(\k^2 b^2 x(1-x)),\label{wf}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
where $ b $ is the transverse separation, $ \varphi $ is the transverse angle, and $ x $ is the plus-momentum ratio of one of the quarks.\\
The average value of $b^2$ in the ground state is given by $\bar b^2\equiv\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \Psi_{00}|b^2|\Psi_{00}\ra={1\over \k^2 x(1-x)}.$ This quantity can be very large if $x$ approaches 0 or 1. However, there would be an $x$-dependent wave function that vanishes at those endpoints in a more complete treatment. Here we are concerned with studying how a small-sized wave packet, or PLC expands as a function of time. This means that cases with $x$ near 0 or 1 are not under consideration. Moreover, if $x$ is near the endpoints, the Feynman mechanism is the dominant contributor to the form factor, so that the conclusion of this paper would be reached without the need to study expansion.\\
It is necessary to obtain the time-development operator for a system that is moving in the $z$ direction with a given large value of $P^+$.
This will be done using the Feynman path integral technique~\cite{Feynman:100771}.
The first step is to regard the right-hand-side of \eq{sw} as an operator $P^-P^+$, to obtain the LF Schroedinger equation
\begin{eqnarray} 2i{\partial\over \partial \t}\Psi=P^- \Psi={1\over P^+} \bigg(-{1\over x(1-x) }\bfnabla_b^2+U(b^2(x(1-x),J)\bigg)\ \Psi ,\label{lfh}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
with the light-front time variable: $\tau\equiv x^+$.
The factor of 2 arises from the definition: $ x^\pm=x^0\pm x^3$, so that $P\cdot x={1/2}(P^-x^++P^+x^-) -{\bf P}}\def\wt{{2\k^2\over P^+}t}\def\bfQ{{\bf Q}}\def\bfb{{\bf b}\cdot{\bf x}$. The vectors written in bold face are two-dimensional vectors often written with the subscript $\perp$.\\
The time development of a PLC that moves at high speed is studied here.
The longitudinal momentum fraction, $x$, is taken as fixed in the present model. This is a good approximation because interactions proportional to $b^2$ do not change the value of $x$~\cite{Donnachie:2002en,PhysRevLett.68.17}.\\
One could construct the time-development operator, $K$, using the stated wave functions of \eq{wf}. Instead, it is much more efficient to use the Feynman path integral formalism~\cite{Feynman:100771}, in which $U$ is the path integral of the exponential of $i$ times the classical action.
In this formalism the Lagrangian is expressed in terms of velocities that are defined as partial derivatives of the Hamiltonian with respect to momenta. The entire formalism is based on a Hamiltonian operator that is $i$ times the partial derivative with respect to time. For this purpose, it is useful to define a time $t\equiv \t/2$. For a system moving in the laboratory with speed close to that of light, $x^3\approx x^0$, so that $t$ is approximately the laboratory time. \\
The next step is to identify a Lagrangian that corresponds to the above Hamiltonian. The operator $\bfp= -i \bfnabla_b $
so that by Hamilton's equations ${\partial\over \partial t}\bfb= \bfv={\partial H\over \partial \bfp}=2\bfp/(P^+x(1-x))$. The following expression results:
\begin{eqnarray} L={P^+\over4}x(1-x)\bfv^2-{1\over P^+} U(\zeta,J)\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The formalism can be tested by taking the limit of a free particle, $U(\zeta,J)\to 0$, using the well-known expression for the time development of a plane wave, starting at $t=0$. At later times, the initial plane wave acquires time-dependence of
$e^{-i P^-t}= e^{-i P^-\tau/2}$, which is the correct form.\\
Having specified the
Lagrangian, the known result \cite{Feynman:100771}
for harmonic oscillator potentials lets us obtain the time development operator $K$ for times between 0 and $t$:
\begin{eqnarray}
\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \bfb,t|K|\bfb',0\ra=\bigg[{x(1-x)\k^2\over 2\pi i \sin {2\k^2t\over P^+}}\bigg]\exp\bigg[i\big({x(1-x)\k^2\over 2 \sin{2\k^2t\over P^+}}\big)((b^2+{b'}^2)\cos {2\k^2t\over P^+}-2\bfb\cdot\bfb')\bigg]\exp{(-i{C_H \over P^+}t)}.\label{U}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The term $C_H$ accounts for the constant term in $U$, and $C_\pi=-2\k^2,\,C_\r=0,C_N=2\k^2$.
One may check that the above expression becomes $\d(\bfb-\bfb') $ in the limit that $t\to0$. The same limit is achieved if $P^+\to\infty$ in which
the separation between partons in the PLC is unchanged as it moves through the nucleus. This fixed transverse separation configuration is said to be frozen. Note that the presence of the term $C_H$ shows that the expansion process depends on the hadron that is involved. This non-universality of the
expansion process that arises from a unified treatment of meson and baryon spectra is the first new result presented here. See also
Ref.~\cite{FARRAR1990125}.\\
Given the time-development operator of \eq{U}, we may study the time development of a PLC, potentially formed in a high momentum transfer (hard) reaction. A measure of expansion was introduced in Ref.~\cite{Frankfurt:1993es}. The idea is that a PLC is originated via a hard interaction involving nucleons initially bound in a nucleus.
The soft interactions between the PLC and the surrounding medium is proportional to the square of the transverse separation distance \cite{Low:1975sv,Nussinov:1975mw,Gunion:1976iy}.
We wish to compare the
relative effects
of escaping with an interaction to that of escaping without an interaction. This is given by a ratio defined as $b^2(t)$.
The effective size is given by the ratio of matrix elements:
\begin{eqnarray} b^2(t)\equiv {\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \Psi_{00}| b^2 K(t)|\rm PLC\ra\over \langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \Psi_{00}|\rm PLC\ra},\label{bdef}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q} where $\Psi_{00}$ is the ground state wave function of \eq{wf}, and $K(t)$ is the time-development operator of \eq{U}. This expression is first-order in a final-state interaction. A complete multiple scattering series was presented in~\cite{PhysRevD.47.1865}. The term $i\s {b^2(t)\over \langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta b^2 \ra}$ has been thought of as scattering amplitude that varies along the path length, $\ell$, of an outgoing PLC, with $t=\ell$~\cite{Frankfurt:1988nt}.\\
Evaluations of $b^2(t)$ proceed by taking a simple form for the PLC: $\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \bfb|\rm PLC\ra=e^{-\gamma}\def\m{\mu}\def\t{\tau}\def\ve{\varepsilon}\def\l{\lambda}\def\s{\sigma}\def\n{\nu}\def\o{\omega b^2}$. Evaluation of \eq{bdef} is then straightforward because Gaussians are involved.
The result is that
\begin{eqnarray} b_H^2(t)={2\over\k^2} {i\over x(1-x)}\sin{(\wt)}\exp{[-i({4\k^2\over P^+}+{C_H\over P^+})t]} +{2\over\k^2x(1-x)+2\gamma}\def\m{\mu}\def\t{\tau}\def\ve{\varepsilon}\def\l{\lambda}\def\s{\sigma}\def\n{\nu}\def\o{\omega}\exp{[-i({6\k^2\over P^+}+{C_H\over P^+})t}].\label{res}\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
The time dependence is seen to be independent of the value of $x$, and the results are periodic with an angular frequency of
$\o\equiv 2\k^2/P^+$. As a result, the second term vanishes smoothly for all values of $t$ as $\gamma}\def\m{\mu}\def\t{\tau}\def\ve{\varepsilon}\def\l{\lambda}\def\s{\sigma}\def\n{\nu}\def\o{\omega$ approaches infinity, the point-like limit.
In that case, the first term of \eq{res} accurately describes the time development and
is independent of the detailed form of $\langle}\def\ra{\rangle}\def\d{\delta}\def\bfr{{\bf r}}\def\k{\kappa}\def\G{\Gamma}\def\z{\zeta}\def\b{\beta \bfb|\rm PLC\ra$, provided the size parameter of the PLC is sufficiently small.
The expression for $b^2_H(t)$ is complex-valued (as has been known for a long time~\cite{Jennings:1990ma}), but only the real part contributes to
the absorption of the outgoing ejected proton. Therefore, we take the limit $\gamma}\def\m{\mu}\def\t{\tau}\def\ve{\varepsilon}\def\l{\lambda}\def\s{\sigma}\def\n{\nu}\def\o{\omega\to\infty$ and use the
real part of $b^2_H$. This and using the stated values of $C_H$ yields the results:
\begin{eqnarray}&{b^2_\pi\over 2\bar b^2}= \sin^2(\wt) \\%\sin(2\wt)\\
&{b^2_\r\over 2 \bar b^2}= \sin(\wt) \sin({4\k^2\over P^+}t)\\% \sin(2\wt)\\
&{b^2_N\over 2\bar b^2}= \sin(\wt)\sin({6\k^2\over P^+}t).
\end{eqnarray}}\def\a{\alpha}\newcommand{\bfk}{{\bf k}}\newcommand{{\bf q}}\def\bfp{{\bf p}}{{\bf q}
These expressions contain the remarkable result that the effects of confinement cause the expansion time to be very different for pions, rho mesons and nucleons. The oscillations inherent in the above results are of no consequence. Once the first maximum is hit the system is strongly absorbed and the formulae are no longer relevant.\\
\begin{figure}[h] \label{pion}
\includegraphics[width=.4\textwidth]{PiFig.pdf}
\caption{${b^2_\pi\over 2\bar b^2}$. Solid (red) $P_\pi^+=5.5 $ GeV, Dashed (blue) $P_\pi^+=8.8 $ GeV, Dot-dashed (green) $P_\pi^+=100 $ GeV. $t$ is in units of fm}
\label{rho}
\includegraphics[width=.4\textwidth]{RhoFig.pdf}
\caption{${b_\r^2\over 2\bar b^2}$. Solid (red) $P_\r^+=6 $ GeV, Dashed (blue) $P_\r^+=12 $ GeV. $t$ is in units of fm}\includegraphics[width=.4
\textwidth]{ProtonFig.pdf}
\caption{${b_N^2\over 2\bar b^2}$ Solid (red) $P_N^+=8 $ GeV, Dashed (blue) $P_N+=14 $ GeV. $t$ is in units of fm. } \end{figure}
The results of evaluations are shown in Figs.~2-4.
The pionic results of Fig~2 are presented for the upper and lower values of $P_\pi^+$ of Ref. \cite{Clasie:2007aa} along with 100 GeV/c. It is useful to define the expansion time, $t_H$ as the {\it lowest} time $t_E$ for which $b^2_H$ hits its maximum value. For the pion this is
$t_\pi = 1.57{P^+\over 2\k^2} =0.552 {P^+\over \rm GeV} \rm fm.$ This time increases from about 2 fm to 6 fm in the kinematic range of Ref.~\cite{Clasie:2007aa} . This increase is large enough so that the observed increase in the measured transparency can reasonably be ascribed to the effects of color transparency. The FermiLab experiment was done at 500 GeV, but the results shown here are for 100 GeV. This is done only to allow a non-zero value to be shown. It is very clear that, in the FermiLab experiment, the PLC goes through an entire nucleus without expanding.\\
The results for the expansion of a $\r$ PLC are presented in Fig.~3. For the experiment of~\cite{ElFassi:2012nr} $t_\r= .955{P^+\over 2\k^2}=0.335{P^+\over \rm GeV} \rm fm$ is only about 2 fm for all of its kinematic variables. This indicates that the observed rise in the transparency, while consistent with theoretical predictions, may not be due to color transparency. However, a new experiment that doubles the value of $P^+$ could lead to observation of the effects of color transparency.\\
The key results here are for the proton as shown in Fig.~4.
The proton expansion time is given by $t_p = 0.659{P^+\over 2\k^2} =0.231{P^+\over \rm GeV} \rm fm.$ In the $(e,e'p)$ experiment on a $^{12}$C target~\cite{Bhetuwal:2020jes}, the value of $Q^2$ ranges between 8 and 14.2 GeV$^2$. The photon energy $\nu= Q^2/(2M)$, and $P^+_p\approx 2\nu$. Thus $P^+$ ranges from about 8 to 14 GeV. As shown in Fig.~3, $t_p$ ranges between about 2 to about 3 fm. This is large enough to observe the effects of color transparency because the radius of $^{12}$C is only about 2.4 fm. If a PLC had been formed, it would have made it out of the nucleus without being absorbed. Thus we are led to the inevitable conclusion that a PLC was not formed. This means that all of the quarks are not on top of each other during the high-momentum transfer process, so that the Feynman mechanism is the only one left standing.\\
The conclusion that the Feynman mechanism is responsible for the proton electromagnetic form factor raises a number of interesting questions regarding its implications. The first is whether this Feynman dominance occurs for all hadrons, in particular the pion. The pion electromagnetic form factor has been studied extensively via perturbative and non-perturbative means, including lattice QCD calculations. In the analytic examples, both the perturbative QCD and light front holography calculations lead to an asymptotic form factor $\sim 1/Q^2$. The PLC is part of the former calculations. The Feynman mechanism does dominate in the latter calculations at infinite momentum~\cite{Sheckler:2020fbt}. However, values of $x$ near 1/2 dominate the light front wave function~\cite{Brodsky:2007hb} for the relative momenta involved in the FermiLab experiment. Thus the result that the astounding $A$-dependence observed is due to color transparency is not challenged.\\
Another point concerns the relation with the EMC effect, the nuclear suppression of high quark structure function at large values of Bjorken $x$. Frankfurt \& Strikman suggested~\cite{Frankfurt:1985cv,Frankfurt:1988nt} that this suppression is due to the reduction of the nucleon's PLC component caused by the attractive nature between the nucleon's non-PLC component and the residual nucleus. This idea has been followed up in many papers, including~\cite{Frank:1995pv,CiofidegliAtti:2007ork,Hen:2016kwk,Miller:2020eyc}, and an excellent interpretation along with a qualitative description of nuclear structure functions in the valence region has been obtained. The question arises: does the lack of PLC dominance of the electromagnetic form factor cast doubt on the basic idea that the different nuclear interactions of different-sized components of the nucleon wave function is responsible for the EMC effect. The answer is no, because the having a dominant PLC is not necessary. The only necessity is that different sized components of the nucleon wave function interact differently with the nucleus. A first application of this idea that uses a light front holographic treatment~\cite{deTeramond:2018ecg} of the nucleon wave function was presented in ~\cite{Miller:2020eyc}. The two different components are configurations of three and four partons. Further development is proceeding.\\
The results of Figs.~2-4 arise from using a specific model, based on a semi-classical approximation, in which quantum loops and quark masses are not included. For example, situations in which the starting configuration includes gluonic components are not included. In that case, the high-momentum transfer process proceeds via a mechanism that is not the Feynman mechanism. This would correspond to a new, unpublished mechanism that is neither perturbative QCD nor the Feynman mechanism.\\
The present results provide a new way to compute the effects of the spatial expansion of a putative PLC. The formalism of the light front holographic model and the resulting expansion effects, combined with the experimental result~\cite{Bhetuwal:2020jes} that color transparency does not occur in reactions with momentum transfer up to 14.2 GeV$^2$, leads to an answer to a decades old question: the Feynman mechanism dominates the high-momentum transfer values of the proton electromagnetic form factor.
\section*{Acknowledgements}
This work was supported by the U. S. Department of Energy Office of Science, Office of Nuclear Physics under Award Number DE-FG02-97ER- 41014.
| d851de8128d15e8342a776ebd663ceabb848c0e2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |